freebsd-nq/module/zfs/zio.c

3727 lines
106 KiB
C
Raw Normal View History

2008-11-20 20:01:55 +00:00
/*
* CDDL HEADER START
*
* The contents of this file are subject to the terms of the
* Common Development and Distribution License (the "License").
* You may not use this file except in compliance with the License.
*
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
* or http://www.opensolaris.org/os/licensing.
* See the License for the specific language governing permissions
* and limitations under the License.
*
* When distributing Covered Code, include this CDDL HEADER in each
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
* If applicable, add the following below this CDDL HEADER, with the
* fields enclosed by brackets "[]" replaced with your own identifying
* information: Portions Copyright [yyyy] [name of copyright owner]
*
* CDDL HEADER END
*/
/*
* Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
* Copyright (c) 2011, 2016 by Delphix. All rights reserved.
* Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
2008-11-20 20:01:55 +00:00
*/
Illumos 5027 - zfs large block support 5027 zfs large block support Reviewed by: Alek Pinchuk <pinchuk.alek@gmail.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Josef 'Jeff' Sipek <josef.sipek@nexenta.com> Reviewed by: Richard Elling <richard.elling@richardelling.com> Reviewed by: Saso Kiselkov <skiselkov.ml@gmail.com> Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov> Approved by: Dan McDonald <danmcd@omniti.com> References: https://www.illumos.org/issues/5027 https://github.com/illumos/illumos-gate/commit/b515258 Porting Notes: * Included in this patch is a tiny ISP2() cleanup in zio_init() from Illumos 5255. * Unlike the upstream Illumos commit this patch does not impose an arbitrary 128K block size limit on volumes. Volumes, like filesystems, are limited by the zfs_max_recordsize=1M module option. * By default the maximum record size is limited to 1M by the module option zfs_max_recordsize. This value may be safely increased up to 16M which is the largest block size supported by the on-disk format. At the moment, 1M blocks clearly offer a significant performance improvement but the benefits of going beyond this for the majority of workloads are less clear. * The illumos version of this patch increased DMU_MAX_ACCESS to 32M. This was determined not to be large enough when using 16M blocks because the zfs_make_xattrdir() function will fail (EFBIG) when assigning a TX. This was immediately observed under Linux because all newly created files must have a security xattr created and that was failing. Therefore, we've set DMU_MAX_ACCESS to 64M. * On 32-bit platforms a hard limit of 1M is set for blocks due to the limited virtual address space. We should be able to relax this one the ABD patches are merged. Ported-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #354
2014-11-03 20:15:08 +00:00
#include <sys/sysmacros.h>
2008-11-20 20:01:55 +00:00
#include <sys/zfs_context.h>
#include <sys/fm/fs/zfs.h>
#include <sys/spa.h>
#include <sys/txg.h>
#include <sys/spa_impl.h>
#include <sys/vdev_impl.h>
#include <sys/zio_impl.h>
#include <sys/zio_compress.h>
#include <sys/zio_checksum.h>
#include <sys/dmu_objset.h>
#include <sys/arc.h>
#include <sys/ddt.h>
#include <sys/blkptr.h>
#include <sys/zfeature.h>
Add -lhHpw options to "zpool iostat" for avg latency, histograms, & queues Update the zfs module to collect statistics on average latencies, queue sizes, and keep an internal histogram of all IO latencies. Along with this, update "zpool iostat" with some new options to print out the stats: -l: Include average IO latencies stats: total_wait disk_wait syncq_wait asyncq_wait scrub read write read write read write read write wait ----- ----- ----- ----- ----- ----- ----- ----- ----- - 41ms - 2ms - 46ms - 4ms - - 5ms - 1ms - 1us - 4ms - - 5ms - 1ms - 1us - 4ms - - - - - - - - - - - 49ms - 2ms - 47ms - - - - - - - - - - - - - 2ms - 1ms - - - 1ms - ----- ----- ----- ----- ----- ----- ----- ----- ----- 1ms 1ms 1ms 413us 16us 25us - 5ms - 1ms 1ms 1ms 413us 16us 25us - 5ms - 2ms 1ms 2ms 412us 26us 25us - 5ms - - 1ms - 413us - 25us - 5ms - - 1ms - 460us - 29us - 5ms - 196us 1ms 196us 370us 7us 23us - 5ms - ----- ----- ----- ----- ----- ----- ----- ----- ----- -w: Print out latency histograms: sdb total disk sync_queue async_queue latency read write read write read write read write scrub ------- ------ ------ ------ ------ ------ ------ ------ ------ ------ 1ns 0 0 0 0 0 0 0 0 0 ... 33us 0 0 0 0 0 0 0 0 0 66us 0 0 107 2486 2 788 12 12 0 131us 2 797 359 4499 10 558 184 184 6 262us 22 801 264 1563 10 286 287 287 24 524us 87 575 71 52086 15 1063 136 136 92 1ms 152 1190 5 41292 4 1693 252 252 141 2ms 245 2018 0 50007 0 2322 371 371 220 4ms 189 7455 22 162957 0 3912 6726 6726 199 8ms 108 9461 0 102320 0 5775 2526 2526 86 17ms 23 11287 0 37142 0 8043 1813 1813 19 34ms 0 14725 0 24015 0 11732 3071 3071 0 67ms 0 23597 0 7914 0 18113 5025 5025 0 134ms 0 33798 0 254 0 25755 7326 7326 0 268ms 0 51780 0 12 0 41593 10002 10002 0 537ms 0 77808 0 0 0 64255 13120 13120 0 1s 0 105281 0 0 0 83805 20841 20841 0 2s 0 88248 0 0 0 73772 14006 14006 0 4s 0 47266 0 0 0 29783 17176 17176 0 9s 0 10460 0 0 0 4130 6295 6295 0 17s 0 0 0 0 0 0 0 0 0 34s 0 0 0 0 0 0 0 0 0 69s 0 0 0 0 0 0 0 0 0 137s 0 0 0 0 0 0 0 0 0 ------------------------------------------------------------------------------- -h: Help -H: Scripted mode. Do not display headers, and separate fields by a single tab instead of arbitrary space. -q: Include current number of entries in sync & async read/write queues, and scrub queue: syncq_read syncq_write asyncq_read asyncq_write scrubq_read pend activ pend activ pend activ pend activ pend activ ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 0 0 0 0 78 29 0 0 0 0 0 0 0 0 78 29 0 0 0 0 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 0 0 0 0 0 0 0 0 0 0 ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 0 0 227 394 0 19 0 0 0 0 0 0 227 394 0 19 0 0 0 0 0 0 108 98 0 19 0 0 0 0 0 0 19 98 0 0 0 0 0 0 0 0 78 98 0 0 0 0 0 0 0 0 19 88 0 0 0 0 0 0 ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- -p: Display numbers in parseable (exact) values. Also, update iostat syntax to allow the user to specify specific vdevs to show statistics for. The three options for choosing pools/vdevs are: Display a list of pools: zpool iostat ... [pool ...] Display a list of vdevs from a specific pool: zpool iostat ... [pool vdev ...] Display a list of vdevs from any pools: zpool iostat ... [vdev ...] Lastly, allow zpool command "interval" value to be floating point: zpool iostat -v 0.5 Signed-off-by: Tony Hutter <hutter2@llnl.gov Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #4433
2016-02-29 18:05:23 +00:00
#include <sys/time.h>
#include <sys/trace_zio.h>
2008-11-20 20:01:55 +00:00
/*
* ==========================================================================
* I/O type descriptions
* ==========================================================================
*/
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
const char *zio_type_name[ZIO_TYPES] = {
"z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl"
};
2008-11-20 20:01:55 +00:00
/*
* ==========================================================================
* I/O kmem caches
* ==========================================================================
*/
kmem_cache_t *zio_cache;
2009-02-18 20:51:31 +00:00
kmem_cache_t *zio_link_cache;
2008-11-20 20:01:55 +00:00
kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT];
int zio_delay_max = ZIO_DELAY_MAX;
2008-11-20 20:01:55 +00:00
#define ZIO_PIPELINE_CONTINUE 0x100
#define ZIO_PIPELINE_STOP 0x101
#define BP_SPANB(indblkshift, level) \
(((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT)))
#define COMPARE_META_LEVEL 0x80000000ul
/*
* The following actions directly effect the spa's sync-to-convergence logic.
* The values below define the sync pass when we start performing the action.
* Care should be taken when changing these values as they directly impact
* spa_sync() performance. Tuning these values may introduce subtle performance
* pathologies and should only be done in the context of performance analysis.
* These tunables will eventually be removed and replaced with #defines once
* enough analysis has been done to determine optimal values.
*
* The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that
* regular blocks are not deferred.
*/
int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */
int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */
int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */
2008-11-20 20:01:55 +00:00
/*
* An allocating zio is one that either currently has the DVA allocate
* stage set or will have it later in its lifetime.
2008-11-20 20:01:55 +00:00
*/
#define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE)
Add missing ZFS tunables This commit adds module options for all existing zfs tunables. Ideally the average user should never need to modify any of these values. However, in practice sometimes you do need to tweak these values for one reason or another. In those cases it's nice not to have to resort to rebuilding from source. All tunables are visable to modinfo and the list is as follows: $ modinfo module/zfs/zfs.ko filename: module/zfs/zfs.ko license: CDDL author: Sun Microsystems/Oracle, Lawrence Livermore National Laboratory description: ZFS srcversion: 8EAB1D71DACE05B5AA61567 depends: spl,znvpair,zcommon,zunicode,zavl vermagic: 2.6.32-131.0.5.el6.x86_64 SMP mod_unload modversions parm: zvol_major:Major number for zvol device (uint) parm: zvol_threads:Number of threads for zvol device (uint) parm: zio_injection_enabled:Enable fault injection (int) parm: zio_bulk_flags:Additional flags to pass to bulk buffers (int) parm: zio_delay_max:Max zio millisec delay before posting event (int) parm: zio_requeue_io_start_cut_in_line:Prioritize requeued I/O (bool) parm: zil_replay_disable:Disable intent logging replay (int) parm: zfs_nocacheflush:Disable cache flushes (bool) parm: zfs_read_chunk_size:Bytes to read per chunk (long) parm: zfs_vdev_max_pending:Max pending per-vdev I/Os (int) parm: zfs_vdev_min_pending:Min pending per-vdev I/Os (int) parm: zfs_vdev_aggregation_limit:Max vdev I/O aggregation size (int) parm: zfs_vdev_time_shift:Deadline time shift for vdev I/O (int) parm: zfs_vdev_ramp_rate:Exponential I/O issue ramp-up rate (int) parm: zfs_vdev_read_gap_limit:Aggregate read I/O over gap (int) parm: zfs_vdev_write_gap_limit:Aggregate write I/O over gap (int) parm: zfs_vdev_scheduler:I/O scheduler (charp) parm: zfs_vdev_cache_max:Inflate reads small than max (int) parm: zfs_vdev_cache_size:Total size of the per-disk cache (int) parm: zfs_vdev_cache_bshift:Shift size to inflate reads too (int) parm: zfs_scrub_limit:Max scrub/resilver I/O per leaf vdev (int) parm: zfs_recover:Set to attempt to recover from fatal errors (int) parm: spa_config_path:SPA config file (/etc/zfs/zpool.cache) (charp) parm: zfs_zevent_len_max:Max event queue length (int) parm: zfs_zevent_cols:Max event column width (int) parm: zfs_zevent_console:Log events to the console (int) parm: zfs_top_maxinflight:Max I/Os per top-level (int) parm: zfs_resilver_delay:Number of ticks to delay resilver (int) parm: zfs_scrub_delay:Number of ticks to delay scrub (int) parm: zfs_scan_idle:Idle window in clock ticks (int) parm: zfs_scan_min_time_ms:Min millisecs to scrub per txg (int) parm: zfs_free_min_time_ms:Min millisecs to free per txg (int) parm: zfs_resilver_min_time_ms:Min millisecs to resilver per txg (int) parm: zfs_no_scrub_io:Set to disable scrub I/O (bool) parm: zfs_no_scrub_prefetch:Set to disable scrub prefetching (bool) parm: zfs_txg_timeout:Max seconds worth of delta per txg (int) parm: zfs_no_write_throttle:Disable write throttling (int) parm: zfs_write_limit_shift:log2(fraction of memory) per txg (int) parm: zfs_txg_synctime_ms:Target milliseconds between tgx sync (int) parm: zfs_write_limit_min:Min tgx write limit (ulong) parm: zfs_write_limit_max:Max tgx write limit (ulong) parm: zfs_write_limit_inflated:Inflated tgx write limit (ulong) parm: zfs_write_limit_override:Override tgx write limit (ulong) parm: zfs_prefetch_disable:Disable all ZFS prefetching (int) parm: zfetch_max_streams:Max number of streams per zfetch (uint) parm: zfetch_min_sec_reap:Min time before stream reclaim (uint) parm: zfetch_block_cap:Max number of blocks to fetch at a time (uint) parm: zfetch_array_rd_sz:Number of bytes in a array_read (ulong) parm: zfs_pd_blks_max:Max number of blocks to prefetch (int) parm: zfs_dedup_prefetch:Enable prefetching dedup-ed blks (int) parm: zfs_arc_min:Min arc size (ulong) parm: zfs_arc_max:Max arc size (ulong) parm: zfs_arc_meta_limit:Meta limit for arc size (ulong) parm: zfs_arc_reduce_dnlc_percent:Meta reclaim percentage (int) parm: zfs_arc_grow_retry:Seconds before growing arc size (int) parm: zfs_arc_shrink_shift:log2(fraction of arc to reclaim) (int) parm: zfs_arc_p_min_shift:arc_c shift to calc min/max arc_p (int)
2011-05-03 22:09:28 +00:00
int zio_requeue_io_start_cut_in_line = 1;
#ifdef ZFS_DEBUG
int zio_buf_debug_limit = 16384;
#else
int zio_buf_debug_limit = 0;
#endif
2008-11-20 20:01:55 +00:00
static inline void __zio_execute(zio_t *zio);
2008-11-20 20:01:55 +00:00
void
zio_init(void)
{
size_t c;
vmem_t *data_alloc_arena = NULL;
zio_cache = kmem_cache_create("zio_cache",
sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
2009-02-18 20:51:31 +00:00
zio_link_cache = kmem_cache_create("zio_link_cache",
sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
2008-11-20 20:01:55 +00:00
/*
* For small buffers, we want a cache for each multiple of
Illumos 5027 - zfs large block support 5027 zfs large block support Reviewed by: Alek Pinchuk <pinchuk.alek@gmail.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Josef 'Jeff' Sipek <josef.sipek@nexenta.com> Reviewed by: Richard Elling <richard.elling@richardelling.com> Reviewed by: Saso Kiselkov <skiselkov.ml@gmail.com> Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov> Approved by: Dan McDonald <danmcd@omniti.com> References: https://www.illumos.org/issues/5027 https://github.com/illumos/illumos-gate/commit/b515258 Porting Notes: * Included in this patch is a tiny ISP2() cleanup in zio_init() from Illumos 5255. * Unlike the upstream Illumos commit this patch does not impose an arbitrary 128K block size limit on volumes. Volumes, like filesystems, are limited by the zfs_max_recordsize=1M module option. * By default the maximum record size is limited to 1M by the module option zfs_max_recordsize. This value may be safely increased up to 16M which is the largest block size supported by the on-disk format. At the moment, 1M blocks clearly offer a significant performance improvement but the benefits of going beyond this for the majority of workloads are less clear. * The illumos version of this patch increased DMU_MAX_ACCESS to 32M. This was determined not to be large enough when using 16M blocks because the zfs_make_xattrdir() function will fail (EFBIG) when assigning a TX. This was immediately observed under Linux because all newly created files must have a security xattr created and that was failing. Therefore, we've set DMU_MAX_ACCESS to 64M. * On 32-bit platforms a hard limit of 1M is set for blocks due to the limited virtual address space. We should be able to relax this one the ABD patches are merged. Ported-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #354
2014-11-03 20:15:08 +00:00
* SPA_MINBLOCKSIZE. For larger buffers, we want a cache
* for each quarter-power of 2.
2008-11-20 20:01:55 +00:00
*/
for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
size_t size = (c + 1) << SPA_MINBLOCKSHIFT;
size_t p2 = size;
size_t align = 0;
size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0;
2008-11-20 20:01:55 +00:00
Illumos 5027 - zfs large block support 5027 zfs large block support Reviewed by: Alek Pinchuk <pinchuk.alek@gmail.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Josef 'Jeff' Sipek <josef.sipek@nexenta.com> Reviewed by: Richard Elling <richard.elling@richardelling.com> Reviewed by: Saso Kiselkov <skiselkov.ml@gmail.com> Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov> Approved by: Dan McDonald <danmcd@omniti.com> References: https://www.illumos.org/issues/5027 https://github.com/illumos/illumos-gate/commit/b515258 Porting Notes: * Included in this patch is a tiny ISP2() cleanup in zio_init() from Illumos 5255. * Unlike the upstream Illumos commit this patch does not impose an arbitrary 128K block size limit on volumes. Volumes, like filesystems, are limited by the zfs_max_recordsize=1M module option. * By default the maximum record size is limited to 1M by the module option zfs_max_recordsize. This value may be safely increased up to 16M which is the largest block size supported by the on-disk format. At the moment, 1M blocks clearly offer a significant performance improvement but the benefits of going beyond this for the majority of workloads are less clear. * The illumos version of this patch increased DMU_MAX_ACCESS to 32M. This was determined not to be large enough when using 16M blocks because the zfs_make_xattrdir() function will fail (EFBIG) when assigning a TX. This was immediately observed under Linux because all newly created files must have a security xattr created and that was failing. Therefore, we've set DMU_MAX_ACCESS to 64M. * On 32-bit platforms a hard limit of 1M is set for blocks due to the limited virtual address space. We should be able to relax this one the ABD patches are merged. Ported-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #354
2014-11-03 20:15:08 +00:00
#ifdef _ILP32
/*
* Cache size limited to 1M on 32-bit platforms until ARC
* buffers no longer require virtual address space.
*/
if (size > zfs_max_recordsize)
break;
#endif
while (!ISP2(p2))
2008-11-20 20:01:55 +00:00
p2 &= p2 - 1;
#ifndef _KERNEL
/*
* If we are using watchpoints, put each buffer on its own page,
* to eliminate the performance overhead of trapping to the
* kernel when modifying a non-watched buffer that shares the
* page with a watched buffer.
*/
if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE))
continue;
/*
* Here's the problem - on 4K native devices in userland on
* Linux using O_DIRECT, buffers must be 4K aligned or I/O
* will fail with EINVAL, causing zdb (and others) to coredump.
* Since userland probably doesn't need optimized buffer caches,
* we just force 4K alignment on everything.
*/
align = 8 * SPA_MINBLOCKSIZE;
#else
2008-11-20 20:01:55 +00:00
if (size <= 4 * SPA_MINBLOCKSIZE) {
align = SPA_MINBLOCKSIZE;
} else if (IS_P2ALIGNED(size, p2 >> 2)) {
Illumos 5027 - zfs large block support 5027 zfs large block support Reviewed by: Alek Pinchuk <pinchuk.alek@gmail.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Josef 'Jeff' Sipek <josef.sipek@nexenta.com> Reviewed by: Richard Elling <richard.elling@richardelling.com> Reviewed by: Saso Kiselkov <skiselkov.ml@gmail.com> Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov> Approved by: Dan McDonald <danmcd@omniti.com> References: https://www.illumos.org/issues/5027 https://github.com/illumos/illumos-gate/commit/b515258 Porting Notes: * Included in this patch is a tiny ISP2() cleanup in zio_init() from Illumos 5255. * Unlike the upstream Illumos commit this patch does not impose an arbitrary 128K block size limit on volumes. Volumes, like filesystems, are limited by the zfs_max_recordsize=1M module option. * By default the maximum record size is limited to 1M by the module option zfs_max_recordsize. This value may be safely increased up to 16M which is the largest block size supported by the on-disk format. At the moment, 1M blocks clearly offer a significant performance improvement but the benefits of going beyond this for the majority of workloads are less clear. * The illumos version of this patch increased DMU_MAX_ACCESS to 32M. This was determined not to be large enough when using 16M blocks because the zfs_make_xattrdir() function will fail (EFBIG) when assigning a TX. This was immediately observed under Linux because all newly created files must have a security xattr created and that was failing. Therefore, we've set DMU_MAX_ACCESS to 64M. * On 32-bit platforms a hard limit of 1M is set for blocks due to the limited virtual address space. We should be able to relax this one the ABD patches are merged. Ported-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #354
2014-11-03 20:15:08 +00:00
align = MIN(p2 >> 2, PAGESIZE);
2008-11-20 20:01:55 +00:00
}
#endif
2008-11-20 20:01:55 +00:00
if (align != 0) {
char name[36];
(void) sprintf(name, "zio_buf_%lu", (ulong_t)size);
zio_buf_cache[c] = kmem_cache_create(name, size,
align, NULL, NULL, NULL, NULL, NULL, cflags);
2008-11-20 20:01:55 +00:00
(void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size);
zio_data_buf_cache[c] = kmem_cache_create(name, size,
Improve meta data performance Profiling the system during meta data intensive workloads such as creating/removing millions of files, revealed that the system was cpu bound. A large fraction of that cpu time was being spent waiting on the virtual address space spin lock. It turns out this was caused by certain heavily used kmem_caches being backed by virtual memory. By default a kmem_cache will dynamically determine the type of memory used based on the object size. For large objects virtual memory is usually preferable and for small object physical memory is a better choice. See the spl_slab_alloc() function for a longer discussion on this. However, there is a certain amount of gray area when defining a 'large' object. For the following caches it turns out they were just over the line: * dnode_cache * zio_cache * zio_link_cache * zio_buf_512_cache * zfs_data_buf_512_cache Now because we know there will be a lot of churn in these caches, and because we know the slabs will still be reasonably sized. We can safely request with the KMC_KMEM flag that the caches be backed with physical memory addresses. This entirely avoids the need to serialize on the virtual address space lock. As a bonus this also reduces our vmalloc usage which will be good for 32-bit kernels which have a very small virtual address space. It will also probably be good for interactive performance since unrelated processes could also block of this same global lock. Finally, we may see less cpu time being burned in the arc_reclaim and txg_sync_threads. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #258
2011-11-01 23:56:48 +00:00
align, NULL, NULL, NULL, NULL,
data_alloc_arena, cflags);
2008-11-20 20:01:55 +00:00
}
}
while (--c != 0) {
ASSERT(zio_buf_cache[c] != NULL);
if (zio_buf_cache[c - 1] == NULL)
zio_buf_cache[c - 1] = zio_buf_cache[c];
ASSERT(zio_data_buf_cache[c] != NULL);
if (zio_data_buf_cache[c - 1] == NULL)
zio_data_buf_cache[c - 1] = zio_data_buf_cache[c];
}
zio_inject_init();
lz4_init();
2008-11-20 20:01:55 +00:00
}
void
zio_fini(void)
{
size_t c;
kmem_cache_t *last_cache = NULL;
kmem_cache_t *last_data_cache = NULL;
for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) {
Illumos 5027 - zfs large block support 5027 zfs large block support Reviewed by: Alek Pinchuk <pinchuk.alek@gmail.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Josef 'Jeff' Sipek <josef.sipek@nexenta.com> Reviewed by: Richard Elling <richard.elling@richardelling.com> Reviewed by: Saso Kiselkov <skiselkov.ml@gmail.com> Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov> Approved by: Dan McDonald <danmcd@omniti.com> References: https://www.illumos.org/issues/5027 https://github.com/illumos/illumos-gate/commit/b515258 Porting Notes: * Included in this patch is a tiny ISP2() cleanup in zio_init() from Illumos 5255. * Unlike the upstream Illumos commit this patch does not impose an arbitrary 128K block size limit on volumes. Volumes, like filesystems, are limited by the zfs_max_recordsize=1M module option. * By default the maximum record size is limited to 1M by the module option zfs_max_recordsize. This value may be safely increased up to 16M which is the largest block size supported by the on-disk format. At the moment, 1M blocks clearly offer a significant performance improvement but the benefits of going beyond this for the majority of workloads are less clear. * The illumos version of this patch increased DMU_MAX_ACCESS to 32M. This was determined not to be large enough when using 16M blocks because the zfs_make_xattrdir() function will fail (EFBIG) when assigning a TX. This was immediately observed under Linux because all newly created files must have a security xattr created and that was failing. Therefore, we've set DMU_MAX_ACCESS to 64M. * On 32-bit platforms a hard limit of 1M is set for blocks due to the limited virtual address space. We should be able to relax this one the ABD patches are merged. Ported-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #354
2014-11-03 20:15:08 +00:00
#ifdef _ILP32
/*
* Cache size limited to 1M on 32-bit platforms until ARC
* buffers no longer require virtual address space.
*/
if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize)
break;
#endif
2008-11-20 20:01:55 +00:00
if (zio_buf_cache[c] != last_cache) {
last_cache = zio_buf_cache[c];
kmem_cache_destroy(zio_buf_cache[c]);
}
zio_buf_cache[c] = NULL;
if (zio_data_buf_cache[c] != last_data_cache) {
last_data_cache = zio_data_buf_cache[c];
kmem_cache_destroy(zio_data_buf_cache[c]);
}
zio_data_buf_cache[c] = NULL;
}
2009-02-18 20:51:31 +00:00
kmem_cache_destroy(zio_link_cache);
2008-11-20 20:01:55 +00:00
kmem_cache_destroy(zio_cache);
zio_inject_fini();
lz4_fini();
2008-11-20 20:01:55 +00:00
}
/*
* ==========================================================================
* Allocate and free I/O buffers
* ==========================================================================
*/
/*
* Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a
* crashdump if the kernel panics, so use it judiciously. Obviously, it's
* useful to inspect ZFS metadata, but if possible, we should avoid keeping
* excess / transient data in-core during a crashdump.
*/
void *
zio_buf_alloc(size_t size)
{
size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
2008-11-20 20:01:55 +00:00
return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE));
2008-11-20 20:01:55 +00:00
}
/*
* Use zio_data_buf_alloc to allocate data. The data will not appear in a
* crashdump if the kernel panics. This exists so that we will limit the amount
* of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount
* of kernel heap dumped to disk when the kernel panics)
*/
void *
zio_data_buf_alloc(size_t size)
{
size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
2008-11-20 20:01:55 +00:00
return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE));
2008-11-20 20:01:55 +00:00
}
Fix vdev_queue_aggregate() deadlock This deadlock may manifest itself in slightly different ways but at the core it is caused by a memory allocation blocking on file- system reclaim in the zio pipeline. This is normally impossible because zio_execute() disables filesystem reclaim by setting PF_FSTRANS on the thread. However, kmem cache allocations may still indirectly block on file system reclaim while holding the critical vq->vq_lock as shown below. To resolve this issue zio_buf_alloc_flags() is introduced which allocation flags to be passed. This can then be used in vdev_queue_aggregate() with KM_NOSLEEP when allocating the aggregate IO buffer. Since aggregating the IO is purely a performance optimization we want this to either succeed or fail quickly. Trying too hard to allocate this memory under the vq->vq_lock can negatively impact performance and result in this deadlock. * z_wr_iss zio_vdev_io_start vdev_queue_io -> Takes vq->vq_lock vdev_queue_io_to_issue vdev_queue_aggregate zio_buf_alloc -> Waiting on spl_kmem_cache process * z_wr_int zio_vdev_io_done vdev_queue_io_done mutex_lock -> Waiting on vq->vq_lock held by z_wr_iss * txg_sync spa_sync dsl_pool_sync zio_wait -> Waiting on zio being handled by z_wr_int * spl_kmem_cache spl_cache_grow_work kv_alloc spl_vmalloc ... evict zpl_evict_inode zfs_inactive dmu_tx_wait txg_wait_open -> Waiting on txg_sync Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Chunwei Chen <david.chen@osnexus.com> Signed-off-by: Tim Chase <tim@chase2k.com> Closes #3808 Closes #3867
2015-12-16 19:28:15 +00:00
/*
* Use zio_buf_alloc_flags when specific allocation flags are needed. e.g.
* passing KM_NOSLEEP when it is acceptable for an allocation to fail.
*/
void *
zio_buf_alloc_flags(size_t size, int flags)
{
size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
return (kmem_cache_alloc(zio_buf_cache[c], flags));
}
2008-11-20 20:01:55 +00:00
void
zio_buf_free(void *buf, size_t size)
{
size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
2008-11-20 20:01:55 +00:00
kmem_cache_free(zio_buf_cache[c], buf);
}
void
zio_data_buf_free(void *buf, size_t size)
{
size_t c = (size - 1) >> SPA_MINBLOCKSHIFT;
VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT);
2008-11-20 20:01:55 +00:00
kmem_cache_free(zio_data_buf_cache[c], buf);
}
/*
* ==========================================================================
* Push and pop I/O transform buffers
* ==========================================================================
*/
OpenZFS 6950 - ARC should cache compressed data Authored by: George Wilson <george.wilson@delphix.com> Reviewed by: Prakash Surya <prakash.surya@delphix.com> Reviewed by: Dan Kimmel <dan.kimmel@delphix.com> Reviewed by: Matt Ahrens <mahrens@delphix.com> Reviewed by: Paul Dagnelie <pcd@delphix.com> Reviewed by: Tom Caputi <tcaputi@datto.com> Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov> Ported by: David Quigley <david.quigley@intel.com> This review covers the reading and writing of compressed arc headers, sharing data between the arc_hdr_t and the arc_buf_t, and the implementation of a new dbuf cache to keep frequently access data uncompressed. I've added a new member to l1 arc hdr called b_pdata. The b_pdata always hangs off the arc_buf_hdr_t (if an L1 hdr is in use) and points to the physical block for that DVA. The physical block may or may not be compressed. If compressed arc is enabled and the block on-disk is compressed, then the b_pdata will match the block on-disk and remain compressed in memory. If the block on disk is not compressed, then neither will the b_pdata. Lastly, if compressed arc is disabled, then b_pdata will always be an uncompressed version of the on-disk block. Typically the arc will cache only the arc_buf_hdr_t and will aggressively evict any arc_buf_t's that are no longer referenced. This means that the arc will primarily have compressed blocks as the arc_buf_t's are considered overhead and are always uncompressed. When a consumer reads a block we first look to see if the arc_buf_hdr_t is cached. If the hdr is cached then we allocate a new arc_buf_t and decompress the b_pdata contents into the arc_buf_t's b_data. If the hdr already has a arc_buf_t, then we will allocate an additional arc_buf_t and bcopy the uncompressed contents from the first arc_buf_t to the new one. Writing to the compressed arc requires that we first discard the b_pdata since the physical block is about to be rewritten. The new data contents will be passed in via an arc_buf_t (uncompressed) and during the I/O pipeline stages we will copy the physical block contents to a newly allocated b_pdata. When an l2arc is inuse it will also take advantage of the b_pdata. Now the l2arc will always write the contents of b_pdata to the l2arc. This means that when compressed arc is enabled that the l2arc blocks are identical to those stored in the main data pool. This provides a significant advantage since we can leverage the bp's checksum when reading from the l2arc to determine if the contents are valid. If the compressed arc is disabled, then we must first transform the read block to look like the physical block in the main data pool before comparing the checksum and determining it's valid. OpenZFS-issue: https://www.illumos.org/issues/6950 OpenZFS-commit: https://github.com/openzfs/openzfs/commit/7fc10f0 Issue #5078
2016-06-02 04:04:53 +00:00
void
zio_push_transform(zio_t *zio, void *data, uint64_t size, uint64_t bufsize,
zio_transform_func_t *transform)
2008-11-20 20:01:55 +00:00
{
zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP);
2008-11-20 20:01:55 +00:00
zt->zt_orig_data = zio->io_data;
zt->zt_orig_size = zio->io_size;
2008-11-20 20:01:55 +00:00
zt->zt_bufsize = bufsize;
zt->zt_transform = transform;
2008-11-20 20:01:55 +00:00
zt->zt_next = zio->io_transform_stack;
zio->io_transform_stack = zt;
zio->io_data = data;
zio->io_size = size;
}
OpenZFS 6950 - ARC should cache compressed data Authored by: George Wilson <george.wilson@delphix.com> Reviewed by: Prakash Surya <prakash.surya@delphix.com> Reviewed by: Dan Kimmel <dan.kimmel@delphix.com> Reviewed by: Matt Ahrens <mahrens@delphix.com> Reviewed by: Paul Dagnelie <pcd@delphix.com> Reviewed by: Tom Caputi <tcaputi@datto.com> Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov> Ported by: David Quigley <david.quigley@intel.com> This review covers the reading and writing of compressed arc headers, sharing data between the arc_hdr_t and the arc_buf_t, and the implementation of a new dbuf cache to keep frequently access data uncompressed. I've added a new member to l1 arc hdr called b_pdata. The b_pdata always hangs off the arc_buf_hdr_t (if an L1 hdr is in use) and points to the physical block for that DVA. The physical block may or may not be compressed. If compressed arc is enabled and the block on-disk is compressed, then the b_pdata will match the block on-disk and remain compressed in memory. If the block on disk is not compressed, then neither will the b_pdata. Lastly, if compressed arc is disabled, then b_pdata will always be an uncompressed version of the on-disk block. Typically the arc will cache only the arc_buf_hdr_t and will aggressively evict any arc_buf_t's that are no longer referenced. This means that the arc will primarily have compressed blocks as the arc_buf_t's are considered overhead and are always uncompressed. When a consumer reads a block we first look to see if the arc_buf_hdr_t is cached. If the hdr is cached then we allocate a new arc_buf_t and decompress the b_pdata contents into the arc_buf_t's b_data. If the hdr already has a arc_buf_t, then we will allocate an additional arc_buf_t and bcopy the uncompressed contents from the first arc_buf_t to the new one. Writing to the compressed arc requires that we first discard the b_pdata since the physical block is about to be rewritten. The new data contents will be passed in via an arc_buf_t (uncompressed) and during the I/O pipeline stages we will copy the physical block contents to a newly allocated b_pdata. When an l2arc is inuse it will also take advantage of the b_pdata. Now the l2arc will always write the contents of b_pdata to the l2arc. This means that when compressed arc is enabled that the l2arc blocks are identical to those stored in the main data pool. This provides a significant advantage since we can leverage the bp's checksum when reading from the l2arc to determine if the contents are valid. If the compressed arc is disabled, then we must first transform the read block to look like the physical block in the main data pool before comparing the checksum and determining it's valid. OpenZFS-issue: https://www.illumos.org/issues/6950 OpenZFS-commit: https://github.com/openzfs/openzfs/commit/7fc10f0 Issue #5078
2016-06-02 04:04:53 +00:00
void
zio_pop_transforms(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
zio_transform_t *zt;
while ((zt = zio->io_transform_stack) != NULL) {
if (zt->zt_transform != NULL)
zt->zt_transform(zio,
zt->zt_orig_data, zt->zt_orig_size);
2008-11-20 20:01:55 +00:00
if (zt->zt_bufsize != 0)
zio_buf_free(zio->io_data, zt->zt_bufsize);
2008-11-20 20:01:55 +00:00
zio->io_data = zt->zt_orig_data;
zio->io_size = zt->zt_orig_size;
zio->io_transform_stack = zt->zt_next;
2008-11-20 20:01:55 +00:00
kmem_free(zt, sizeof (zio_transform_t));
2008-11-20 20:01:55 +00:00
}
}
/*
* ==========================================================================
* I/O transform callbacks for subblocks and decompression
* ==========================================================================
*/
static void
zio_subblock(zio_t *zio, void *data, uint64_t size)
{
ASSERT(zio->io_size > size);
if (zio->io_type == ZIO_TYPE_READ)
bcopy(zio->io_data, data, size);
}
static void
zio_decompress(zio_t *zio, void *data, uint64_t size)
{
if (zio->io_error == 0 &&
zio_decompress_data(BP_GET_COMPRESS(zio->io_bp),
zio->io_data, data, zio->io_size, size) != 0)
zio->io_error = SET_ERROR(EIO);
}
/*
* ==========================================================================
* I/O parent/child relationships and pipeline interlocks
* ==========================================================================
*/
2009-02-18 20:51:31 +00:00
/*
* NOTE - Callers to zio_walk_parents() and zio_walk_children must
* continue calling these functions until they return NULL.
* Otherwise, the next caller will pick up the list walk in
* some indeterminate state. (Otherwise every caller would
* have to pass in a cookie to keep the state represented by
* io_walk_link, which gets annoying.)
*/
zio_t *
zio_walk_parents(zio_t *cio)
{
zio_link_t *zl = cio->io_walk_link;
list_t *pl = &cio->io_parent_list;
2009-02-18 20:51:31 +00:00
zl = (zl == NULL) ? list_head(pl) : list_next(pl, zl);
cio->io_walk_link = zl;
if (zl == NULL)
return (NULL);
ASSERT(zl->zl_child == cio);
return (zl->zl_parent);
}
zio_t *
zio_walk_children(zio_t *pio)
{
zio_link_t *zl = pio->io_walk_link;
list_t *cl = &pio->io_child_list;
zl = (zl == NULL) ? list_head(cl) : list_next(cl, zl);
pio->io_walk_link = zl;
if (zl == NULL)
return (NULL);
ASSERT(zl->zl_parent == pio);
return (zl->zl_child);
}
zio_t *
zio_unique_parent(zio_t *cio)
{
zio_t *pio = zio_walk_parents(cio);
VERIFY(zio_walk_parents(cio) == NULL);
return (pio);
}
void
zio_add_child(zio_t *pio, zio_t *cio)
{
zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP);
int w;
2009-02-18 20:51:31 +00:00
/*
* Logical I/Os can have logical, gang, or vdev children.
* Gang I/Os can have gang or vdev children.
* Vdev I/Os can only have vdev children.
* The following ASSERT captures all of these constraints.
*/
ASSERT(cio->io_child_type <= pio->io_child_type);
zl->zl_parent = pio;
zl->zl_child = cio;
mutex_enter(&cio->io_lock);
mutex_enter(&pio->io_lock);
2009-02-18 20:51:31 +00:00
ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0);
for (w = 0; w < ZIO_WAIT_TYPES; w++)
2009-02-18 20:51:31 +00:00
pio->io_children[cio->io_child_type][w] += !cio->io_state[w];
list_insert_head(&pio->io_child_list, zl);
list_insert_head(&cio->io_parent_list, zl);
pio->io_child_count++;
cio->io_parent_count++;
mutex_exit(&pio->io_lock);
2009-02-18 20:51:31 +00:00
mutex_exit(&cio->io_lock);
}
2008-11-20 20:01:55 +00:00
static void
2009-02-18 20:51:31 +00:00
zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl)
{
2009-02-18 20:51:31 +00:00
ASSERT(zl->zl_parent == pio);
ASSERT(zl->zl_child == cio);
2009-02-18 20:51:31 +00:00
mutex_enter(&cio->io_lock);
mutex_enter(&pio->io_lock);
2009-02-18 20:51:31 +00:00
list_remove(&pio->io_child_list, zl);
list_remove(&cio->io_parent_list, zl);
pio->io_child_count--;
cio->io_parent_count--;
mutex_exit(&pio->io_lock);
2009-02-18 20:51:31 +00:00
mutex_exit(&cio->io_lock);
kmem_cache_free(zio_link_cache, zl);
}
static boolean_t
zio_wait_for_children(zio_t *zio, enum zio_child child, enum zio_wait_type wait)
2008-11-20 20:01:55 +00:00
{
uint64_t *countp = &zio->io_children[child][wait];
boolean_t waiting = B_FALSE;
mutex_enter(&zio->io_lock);
ASSERT(zio->io_stall == NULL);
if (*countp != 0) {
zio->io_stage >>= 1;
zio->io_stall = countp;
waiting = B_TRUE;
}
mutex_exit(&zio->io_lock);
return (waiting);
}
2008-11-20 20:01:55 +00:00
__attribute__((always_inline))
static inline void
zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait)
{
uint64_t *countp = &pio->io_children[zio->io_child_type][wait];
int *errorp = &pio->io_child_error[zio->io_child_type];
2008-11-20 20:01:55 +00:00
mutex_enter(&pio->io_lock);
if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
*errorp = zio_worst_error(*errorp, zio->io_error);
pio->io_reexecute |= zio->io_reexecute;
ASSERT3U(*countp, >, 0);
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
(*countp)--;
if (*countp == 0 && pio->io_stall == countp) {
pio->io_stall = NULL;
mutex_exit(&pio->io_lock);
__zio_execute(pio);
} else {
mutex_exit(&pio->io_lock);
2008-11-20 20:01:55 +00:00
}
}
static void
zio_inherit_child_errors(zio_t *zio, enum zio_child c)
{
if (zio->io_child_error[c] != 0 && zio->io_error == 0)
zio->io_error = zio->io_child_error[c];
}
2008-11-20 20:01:55 +00:00
/*
* ==========================================================================
* Create the various types of I/O (read, write, free, etc)
2008-11-20 20:01:55 +00:00
* ==========================================================================
*/
static zio_t *
zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
void *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done,
void *private, zio_type_t type, zio_priority_t priority,
enum zio_flag flags, vdev_t *vd, uint64_t offset,
const zbookmark_phys_t *zb, enum zio_stage stage,
enum zio_stage pipeline)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE);
ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0);
ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0);
ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER));
ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER));
ASSERT(vd || stage == ZIO_STAGE_OPEN);
2008-11-20 20:01:55 +00:00
IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW) != 0);
zio = kmem_cache_alloc(zio_cache, KM_SLEEP);
bzero(zio, sizeof (zio_t));
Identify locks flagged by lockdep When running a kernel with CONFIG_LOCKDEP=y, lockdep reports possible recursive locking in some cases and possible circular locking dependency in others, within the SPL and ZFS modules. This patch uses a mutex type defined in SPL, MUTEX_NOLOCKDEP, to mark such mutexes when they are initialized. This mutex type causes attempts to take or release those locks to be wrapped in lockdep_off() and lockdep_on() calls to silence the dependency checker and allow the use of lock_stats to examine contention. For RW locks, it uses an analogous lock type, RW_NOLOCKDEP. The goal is that these locks are ultimately changed back to type MUTEX_DEFAULT or RW_DEFAULT, after the locks are annotated to reflect their relationship (e.g. z_name_lock below) or any real problem with the lock dependencies are fixed. Some of the affected locks are: tc_open_lock: ============= This is an array of locks, all with same name, which txg_quiesce must take all of in order to move txg to next state. All default to the same lockdep class, and so to lockdep appears recursive. zp->z_name_lock: ================ In zfs_rmdir, dzp = znode for the directory (input to zfs_dirent_lock) zp = znode for the entry being removed (output of zfs_dirent_lock) zfs_rmdir()->zfs_dirent_lock() takes z_name_lock in dzp zfs_rmdir() takes z_name_lock in zp Since both dzp and zp are type znode_t, the locks have the same default class, and lockdep considers it a possible recursive lock attempt. l->l_rwlock: ============ zap_expand_leaf() sometimes creates two new zap leaf structures, via these call paths: zap_deref_leaf()->zap_get_leaf_byblk()->zap_leaf_open() zap_expand_leaf()->zap_create_leaf()->zap_expand_leaf()->zap_create_leaf() Because both zap_leaf_open() and zap_create_leaf() initialize l->l_rwlock in their (separate) leaf structures, the lockdep class is the same, and the linux kernel believes these might both be the same lock, and emits a possible recursive lock warning. Signed-off-by: Olaf Faaland <faaland1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #3895
2015-10-15 20:08:27 +00:00
mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL);
cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL);
list_create(&zio->io_parent_list, sizeof (zio_link_t),
offsetof(zio_link_t, zl_parent_node));
list_create(&zio->io_child_list, sizeof (zio_link_t),
offsetof(zio_link_t, zl_child_node));
2009-02-18 20:51:31 +00:00
if (vd != NULL)
zio->io_child_type = ZIO_CHILD_VDEV;
else if (flags & ZIO_FLAG_GANG_CHILD)
zio->io_child_type = ZIO_CHILD_GANG;
else if (flags & ZIO_FLAG_DDT_CHILD)
zio->io_child_type = ZIO_CHILD_DDT;
else
zio->io_child_type = ZIO_CHILD_LOGICAL;
2008-11-20 20:01:55 +00:00
if (bp != NULL) {
zio->io_bp = (blkptr_t *)bp;
2008-11-20 20:01:55 +00:00
zio->io_bp_copy = *bp;
zio->io_bp_orig = *bp;
if (type != ZIO_TYPE_WRITE ||
zio->io_child_type == ZIO_CHILD_DDT)
zio->io_bp = &zio->io_bp_copy; /* so caller can free */
2009-07-02 22:44:48 +00:00
if (zio->io_child_type == ZIO_CHILD_LOGICAL)
zio->io_logical = zio;
2009-07-02 22:44:48 +00:00
if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp))
pipeline |= ZIO_GANG_STAGES;
2008-11-20 20:01:55 +00:00
}
zio->io_spa = spa;
zio->io_txg = txg;
2008-11-20 20:01:55 +00:00
zio->io_done = done;
zio->io_private = private;
zio->io_type = type;
zio->io_priority = priority;
zio->io_vd = vd;
zio->io_offset = offset;
zio->io_orig_data = zio->io_data = data;
zio->io_orig_size = zio->io_size = psize;
zio->io_lsize = lsize;
zio->io_orig_flags = zio->io_flags = flags;
zio->io_orig_stage = zio->io_stage = stage;
zio->io_orig_pipeline = zio->io_pipeline = pipeline;
2008-11-20 20:01:55 +00:00
2009-02-18 20:51:31 +00:00
zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY);
zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE);
if (zb != NULL)
zio->io_bookmark = *zb;
if (pio != NULL) {
if (zio->io_logical == NULL)
2008-11-20 20:01:55 +00:00
zio->io_logical = pio->io_logical;
2009-07-02 22:44:48 +00:00
if (zio->io_child_type == ZIO_CHILD_GANG)
zio->io_gang_leader = pio->io_gang_leader;
zio_add_child(pio, zio);
2008-11-20 20:01:55 +00:00
}
taskq_init_ent(&zio->io_tqent);
2008-11-20 20:01:55 +00:00
return (zio);
}
static void
zio_destroy(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
list_destroy(&zio->io_parent_list);
list_destroy(&zio->io_child_list);
mutex_destroy(&zio->io_lock);
cv_destroy(&zio->io_cv);
kmem_cache_free(zio_cache, zio);
2008-11-20 20:01:55 +00:00
}
zio_t *
2009-02-18 20:51:31 +00:00
zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done,
void *private, enum zio_flag flags)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
2009-02-18 20:51:31 +00:00
ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE);
2008-11-20 20:01:55 +00:00
return (zio);
}
zio_t *
zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags)
2008-11-20 20:01:55 +00:00
{
2009-02-18 20:51:31 +00:00
return (zio_null(NULL, spa, NULL, done, private, flags));
2008-11-20 20:01:55 +00:00
}
void
zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp)
{
int i;
if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) {
zfs_panic_recover("blkptr at %p has invalid TYPE %llu",
bp, (longlong_t)BP_GET_TYPE(bp));
}
if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS ||
BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) {
zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu",
bp, (longlong_t)BP_GET_CHECKSUM(bp));
}
if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS ||
BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) {
zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu",
bp, (longlong_t)BP_GET_COMPRESS(bp));
}
if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) {
zfs_panic_recover("blkptr at %p has invalid LSIZE %llu",
bp, (longlong_t)BP_GET_LSIZE(bp));
}
if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) {
zfs_panic_recover("blkptr at %p has invalid PSIZE %llu",
bp, (longlong_t)BP_GET_PSIZE(bp));
}
if (BP_IS_EMBEDDED(bp)) {
if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) {
zfs_panic_recover("blkptr at %p has invalid ETYPE %llu",
bp, (longlong_t)BPE_GET_ETYPE(bp));
}
}
/*
* Pool-specific checks.
*
* Note: it would be nice to verify that the blk_birth and
* BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze()
* allows the birth time of log blocks (and dmu_sync()-ed blocks
* that are in the log) to be arbitrarily large.
*/
for (i = 0; i < BP_GET_NDVAS(bp); i++) {
uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]);
vdev_t *vd;
uint64_t offset, asize;
if (vdevid >= spa->spa_root_vdev->vdev_children) {
zfs_panic_recover("blkptr at %p DVA %u has invalid "
"VDEV %llu",
bp, i, (longlong_t)vdevid);
continue;
}
vd = spa->spa_root_vdev->vdev_child[vdevid];
if (vd == NULL) {
zfs_panic_recover("blkptr at %p DVA %u has invalid "
"VDEV %llu",
bp, i, (longlong_t)vdevid);
continue;
}
if (vd->vdev_ops == &vdev_hole_ops) {
zfs_panic_recover("blkptr at %p DVA %u has hole "
"VDEV %llu",
bp, i, (longlong_t)vdevid);
continue;
}
if (vd->vdev_ops == &vdev_missing_ops) {
/*
* "missing" vdevs are valid during import, but we
* don't have their detailed info (e.g. asize), so
* we can't perform any more checks on them.
*/
continue;
}
offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
asize = DVA_GET_ASIZE(&bp->blk_dva[i]);
if (BP_IS_GANG(bp))
asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
if (offset + asize > vd->vdev_asize) {
zfs_panic_recover("blkptr at %p DVA %u has invalid "
"OFFSET %llu",
bp, i, (longlong_t)offset);
}
}
}
2008-11-20 20:01:55 +00:00
zio_t *
zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp,
void *data, uint64_t size, zio_done_func_t *done, void *private,
zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
zfs_blkptr_verify(spa, bp);
zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp,
data, size, size, done, private,
ZIO_TYPE_READ, priority, flags, NULL, 0, zb,
ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE);
2008-11-20 20:01:55 +00:00
return (zio);
}
2008-11-20 20:01:55 +00:00
zio_t *
zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp,
void *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp,
zio_done_func_t *ready, zio_done_func_t *children_ready,
zio_done_func_t *physdone, zio_done_func_t *done,
void *private, zio_priority_t priority, enum zio_flag flags,
const zbookmark_phys_t *zb)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF &&
zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS &&
zp->zp_compress >= ZIO_COMPRESS_OFF &&
zp->zp_compress < ZIO_COMPRESS_FUNCTIONS &&
DMU_OT_IS_VALID(zp->zp_type) &&
zp->zp_level < 32 &&
zp->zp_copies > 0 &&
zp->zp_copies <= spa_max_replication(spa));
2008-11-20 20:01:55 +00:00
zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private,
ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ?
ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE);
2008-11-20 20:01:55 +00:00
zio->io_ready = ready;
zio->io_children_ready = children_ready;
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
zio->io_physdone = physdone;
zio->io_prop = *zp;
2008-11-20 20:01:55 +00:00
/*
* Data can be NULL if we are going to call zio_write_override() to
* provide the already-allocated BP. But we may need the data to
* verify a dedup hit (if requested). In this case, don't try to
* dedup (just take the already-allocated BP verbatim).
*/
if (data == NULL && zio->io_prop.zp_dedup_verify) {
zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE;
}
2008-11-20 20:01:55 +00:00
return (zio);
}
zio_t *
zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, void *data,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
uint64_t size, zio_done_func_t *done, void *private,
zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
zio = zio_create(pio, spa, txg, bp, data, size, size, done, private,
ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb,
ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE);
2008-11-20 20:01:55 +00:00
return (zio);
}
void
zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite)
{
ASSERT(zio->io_type == ZIO_TYPE_WRITE);
ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa));
/*
* We must reset the io_prop to match the values that existed
* when the bp was first written by dmu_sync() keeping in mind
* that nopwrite and dedup are mutually exclusive.
*/
zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup;
zio->io_prop.zp_nopwrite = nopwrite;
zio->io_prop.zp_copies = copies;
zio->io_bp_override = bp;
}
void
zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp)
{
/*
* The check for EMBEDDED is a performance optimization. We
* process the free here (by ignoring it) rather than
* putting it on the list and then processing it in zio_free_sync().
*/
if (BP_IS_EMBEDDED(bp))
return;
metaslab_check_free(spa, bp);
/*
* Frees that are for the currently-syncing txg, are not going to be
* deferred, and which will not need to do a read (i.e. not GANG or
* DEDUP), can be processed immediately. Otherwise, put them on the
* in-memory list for later processing.
*/
if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) ||
txg != spa->spa_syncing_txg ||
spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) {
bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp);
} else {
VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0)));
}
}
2008-11-20 20:01:55 +00:00
zio_t *
zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
enum zio_flag flags)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
enum zio_stage stage = ZIO_FREE_PIPELINE;
2008-11-20 20:01:55 +00:00
ASSERT(!BP_IS_HOLE(bp));
ASSERT(spa_syncing_txg(spa) == txg);
ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free);
2008-11-20 20:01:55 +00:00
if (BP_IS_EMBEDDED(bp))
return (zio_null(pio, spa, NULL, NULL, NULL, 0));
metaslab_check_free(spa, bp);
arc_freed(spa, bp);
/*
* GANG and DEDUP blocks can induce a read (for the gang block header,
* or the DDT), so issue them asynchronously so that this thread is
* not tied up.
*/
if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp))
stage |= ZIO_STAGE_ISSUE_ASYNC;
zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW,
flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage);
2008-11-20 20:01:55 +00:00
return (zio);
}
zio_t *
zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp,
zio_done_func_t *done, void *private, enum zio_flag flags)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
dprintf_bp(bp, "claiming in txg %llu", txg);
if (BP_IS_EMBEDDED(bp))
return (zio_null(pio, spa, NULL, NULL, NULL, 0));
2008-11-20 20:01:55 +00:00
/*
* A claim is an allocation of a specific block. Claims are needed
* to support immediate writes in the intent log. The issue is that
* immediate writes contain committed data, but in a txg that was
* *not* committed. Upon opening the pool after an unclean shutdown,
* the intent log claims all blocks that contain immediate write data
* so that the SPA knows they're in use.
*
* All claims *must* be resolved in the first txg -- before the SPA
* starts allocating blocks -- so that nothing is allocated twice.
* If txg == 0 we just verify that the block is claimable.
2008-11-20 20:01:55 +00:00
*/
ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, spa_first_txg(spa));
ASSERT(txg == spa_first_txg(spa) || txg == 0);
ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */
2008-11-20 20:01:55 +00:00
zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp),
BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW,
flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE);
2008-11-20 20:01:55 +00:00
return (zio);
}
zio_t *
zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
zio_done_func_t *done, void *private, enum zio_flag flags)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
int c;
if (vd->vdev_children == 0) {
zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL,
2008-11-20 20:01:55 +00:00
ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE);
zio->io_cmd = cmd;
} else {
2009-02-18 20:51:31 +00:00
zio = zio_null(pio, spa, NULL, NULL, NULL, flags);
2008-11-20 20:01:55 +00:00
for (c = 0; c < vd->vdev_children; c++)
zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
done, private, flags));
2008-11-20 20:01:55 +00:00
}
return (zio);
}
zio_t *
zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
void *data, int checksum, zio_done_func_t *done, void *private,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
zio_priority_t priority, enum zio_flag flags, boolean_t labels)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
ASSERT(vd->vdev_children == 0);
ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
ASSERT3U(offset + size, <=, vd->vdev_psize);
2008-11-20 20:01:55 +00:00
zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd,
offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE);
2008-11-20 20:01:55 +00:00
zio->io_prop.zp_checksum = checksum;
2008-11-20 20:01:55 +00:00
return (zio);
}
zio_t *
zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size,
void *data, int checksum, zio_done_func_t *done, void *private,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
zio_priority_t priority, enum zio_flag flags, boolean_t labels)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
ASSERT(vd->vdev_children == 0);
ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE ||
offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE);
ASSERT3U(offset + size, <=, vd->vdev_psize);
2008-11-20 20:01:55 +00:00
zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done,
private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd,
offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE);
2008-11-20 20:01:55 +00:00
zio->io_prop.zp_checksum = checksum;
2008-11-20 20:01:55 +00:00
if (zio_checksum_table[checksum].ci_eck) {
2008-11-20 20:01:55 +00:00
/*
* zec checksums are necessarily destructive -- they modify
* the end of the write buffer to hold the verifier/checksum.
2008-11-20 20:01:55 +00:00
* Therefore, we must make a local copy in case the data is
* being written to multiple places in parallel.
2008-11-20 20:01:55 +00:00
*/
void *wbuf = zio_buf_alloc(size);
2008-11-20 20:01:55 +00:00
bcopy(data, wbuf, size);
zio_push_transform(zio, wbuf, size, size, NULL);
2008-11-20 20:01:55 +00:00
}
return (zio);
}
/*
* Create a child I/O to do some work for us.
2008-11-20 20:01:55 +00:00
*/
zio_t *
zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
void *data, uint64_t size, int type, zio_priority_t priority,
enum zio_flag flags, zio_done_func_t *done, void *private)
2008-11-20 20:01:55 +00:00
{
enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE;
zio_t *zio;
ASSERT(vd->vdev_parent ==
(pio->io_vd ? pio->io_vd : pio->io_spa->spa_root_vdev));
2008-11-20 20:01:55 +00:00
if (type == ZIO_TYPE_READ && bp != NULL) {
/*
* If we have the bp, then the child should perform the
* checksum and the parent need not. This pushes error
* detection as close to the leaves as possible and
* eliminates redundant checksums in the interior nodes.
*/
pipeline |= ZIO_STAGE_CHECKSUM_VERIFY;
pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2008-11-20 20:01:55 +00:00
}
if (vd->vdev_children == 0)
offset += VDEV_LABEL_START_SIZE;
flags |= ZIO_VDEV_CHILD_FLAGS(pio) | ZIO_FLAG_DONT_PROPAGATE;
/*
* If we've decided to do a repair, the write is not speculative --
* even if the original read was.
*/
if (flags & ZIO_FLAG_IO_REPAIR)
flags &= ~ZIO_FLAG_SPECULATIVE;
zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size,
done, private, type, priority, flags, vd, offset, &pio->io_bookmark,
ZIO_STAGE_VDEV_IO_START >> 1, pipeline);
2008-11-20 20:01:55 +00:00
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
zio->io_physdone = pio->io_physdone;
if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL)
zio->io_logical->io_phys_children++;
return (zio);
2008-11-20 20:01:55 +00:00
}
zio_t *
zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, void *data, uint64_t size,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
int type, zio_priority_t priority, enum zio_flag flags,
zio_done_func_t *done, void *private)
2008-11-20 20:01:55 +00:00
{
zio_t *zio;
2008-11-20 20:01:55 +00:00
ASSERT(vd->vdev_ops->vdev_op_leaf);
2008-11-20 20:01:55 +00:00
zio = zio_create(NULL, vd->vdev_spa, 0, NULL,
data, size, size, done, private, type, priority,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED,
vd, offset, NULL,
ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE);
2008-11-20 20:01:55 +00:00
return (zio);
2008-11-20 20:01:55 +00:00
}
void
zio_flush(zio_t *zio, vdev_t *vd)
2008-11-20 20:01:55 +00:00
{
zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE,
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
NULL, NULL,
ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY));
2008-11-20 20:01:55 +00:00
}
void
zio_shrink(zio_t *zio, uint64_t size)
{
ASSERT(zio->io_executor == NULL);
ASSERT(zio->io_orig_size == zio->io_size);
ASSERT(size <= zio->io_size);
/*
* We don't shrink for raidz because of problems with the
* reconstruction when reading back less than the block size.
* Note, BP_IS_RAIDZ() assumes no compression.
*/
ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
if (!BP_IS_RAIDZ(zio->io_bp)) {
/* we are not doing a raw write */
ASSERT3U(zio->io_size, ==, zio->io_lsize);
zio->io_orig_size = zio->io_size = zio->io_lsize = size;
}
}
2008-11-20 20:01:55 +00:00
/*
* ==========================================================================
* Prepare to read and write logical blocks
2008-11-20 20:01:55 +00:00
* ==========================================================================
*/
2008-11-20 20:01:55 +00:00
static int
zio_read_bp_init(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
blkptr_t *bp = zio->io_bp;
2008-11-20 20:01:55 +00:00
2009-01-15 21:59:39 +00:00
if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF &&
2009-07-02 22:44:48 +00:00
zio->io_child_type == ZIO_CHILD_LOGICAL &&
!(zio->io_flags & ZIO_FLAG_RAW)) {
uint64_t psize =
BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp);
void *cbuf = zio_buf_alloc(psize);
zio_push_transform(zio, cbuf, psize, psize, zio_decompress);
2008-11-20 20:01:55 +00:00
}
if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) {
zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
decode_embedded_bp_compressed(bp, zio->io_data);
} else {
ASSERT(!BP_IS_EMBEDDED(bp));
}
if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0)
zio->io_flags |= ZIO_FLAG_DONT_CACHE;
if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP)
zio->io_flags |= ZIO_FLAG_DONT_CACHE;
if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL)
zio->io_pipeline = ZIO_DDT_READ_PIPELINE;
return (ZIO_PIPELINE_CONTINUE);
2008-11-20 20:01:55 +00:00
}
static int
zio_write_bp_init(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
spa_t *spa = zio->io_spa;
zio_prop_t *zp = &zio->io_prop;
enum zio_compress compress = zp->zp_compress;
2008-11-20 20:01:55 +00:00
blkptr_t *bp = zio->io_bp;
uint64_t lsize = zio->io_lsize;
uint64_t psize = zio->io_size;
int pass = 1;
2008-11-20 20:01:55 +00:00
EQUIV(lsize != psize, (zio->io_flags & ZIO_FLAG_RAW) != 0);
/*
* If our children haven't all reached the ready stage,
* wait for them and then repeat this pipeline stage.
*/
if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_READY))
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
if (!IO_IS_ALLOCATING(zio))
return (ZIO_PIPELINE_CONTINUE);
2008-11-20 20:01:55 +00:00
if (zio->io_children_ready != NULL) {
/*
* Now that all our children are ready, run the callback
* associated with this zio in case it wants to modify the
* data to be written.
*/
ASSERT3U(zp->zp_level, >, 0);
zio->io_children_ready(zio);
}
ASSERT(zio->io_child_type != ZIO_CHILD_DDT);
if (zio->io_bp_override) {
ASSERT(bp->blk_birth != zio->io_txg);
ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0);
*bp = *zio->io_bp_override;
zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
if (BP_IS_EMBEDDED(bp))
return (ZIO_PIPELINE_CONTINUE);
/*
* If we've been overridden and nopwrite is set then
* set the flag accordingly to indicate that a nopwrite
* has already occurred.
*/
if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) {
ASSERT(!zp->zp_dedup);
zio->io_flags |= ZIO_FLAG_NOPWRITE;
return (ZIO_PIPELINE_CONTINUE);
}
ASSERT(!zp->zp_nopwrite);
if (BP_IS_HOLE(bp) || !zp->zp_dedup)
return (ZIO_PIPELINE_CONTINUE);
ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup ||
zp->zp_dedup_verify);
if (BP_GET_CHECKSUM(bp) == zp->zp_checksum) {
BP_SET_DEDUP(bp, 1);
zio->io_pipeline |= ZIO_STAGE_DDT_WRITE;
return (ZIO_PIPELINE_CONTINUE);
}
zio->io_bp_override = NULL;
BP_ZERO(bp);
}
2008-11-20 20:01:55 +00:00
if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) {
/*
* We're rewriting an existing block, which means we're
* working on behalf of spa_sync(). For spa_sync() to
* converge, it must eventually be the case that we don't
* have to allocate new blocks. But compression changes
* the blocksize, which forces a reallocate, and makes
* convergence take longer. Therefore, after the first
* few passes, stop compressing to ensure convergence.
*/
pass = spa_sync_pass(spa);
ASSERT(zio->io_txg == spa_syncing_txg(spa));
ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
ASSERT(!BP_GET_DEDUP(bp));
2008-11-20 20:01:55 +00:00
if (pass >= zfs_sync_pass_dont_compress)
compress = ZIO_COMPRESS_OFF;
2008-11-20 20:01:55 +00:00
/* Make sure someone doesn't change their mind on overwrites */
ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp),
spa_max_replication(spa)) == BP_GET_NDVAS(bp));
}
2008-11-20 20:01:55 +00:00
/* If it's a compressed write that is not raw, compress the buffer. */
if (compress != ZIO_COMPRESS_OFF && psize == lsize) {
void *cbuf = zio_buf_alloc(lsize);
psize = zio_compress_data(compress, zio->io_data, cbuf, lsize);
if (psize == 0 || psize == lsize) {
compress = ZIO_COMPRESS_OFF;
zio_buf_free(cbuf, lsize);
} else if (!zp->zp_dedup && psize <= BPE_PAYLOAD_SIZE &&
zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) &&
spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) {
encode_embedded_bp_compressed(bp,
cbuf, compress, lsize, psize);
BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA);
BP_SET_TYPE(bp, zio->io_prop.zp_type);
BP_SET_LEVEL(bp, zio->io_prop.zp_level);
zio_buf_free(cbuf, lsize);
bp->blk_birth = zio->io_txg;
zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
ASSERT(spa_feature_is_active(spa,
SPA_FEATURE_EMBEDDED_DATA));
return (ZIO_PIPELINE_CONTINUE);
} else {
/*
* Round up compressed size up to the ashift
* of the smallest-ashift device, and zero the tail.
* This ensures that the compressed size of the BP
* (and thus compressratio property) are correct,
* in that we charge for the padding used to fill out
* the last sector.
*/
size_t rounded;
ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
rounded = (size_t)P2ROUNDUP(psize,
1ULL << spa->spa_min_ashift);
if (rounded >= lsize) {
compress = ZIO_COMPRESS_OFF;
zio_buf_free(cbuf, lsize);
psize = lsize;
} else {
bzero((char *)cbuf + psize, rounded - psize);
psize = rounded;
zio_push_transform(zio, cbuf,
psize, lsize, NULL);
}
}
} else {
ASSERT3U(psize, !=, 0);
}
2008-11-20 20:01:55 +00:00
/*
* The final pass of spa_sync() must be all rewrites, but the first
* few passes offer a trade-off: allocating blocks defers convergence,
* but newly allocated blocks are sequential, so they can be written
* to disk faster. Therefore, we allow the first few passes of
* spa_sync() to allocate new blocks, but force rewrites after that.
* There should only be a handful of blocks after pass 1 in any case.
*/
if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg &&
BP_GET_PSIZE(bp) == psize &&
pass >= zfs_sync_pass_rewrite) {
enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES;
ASSERT(psize != 0);
zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages;
zio->io_flags |= ZIO_FLAG_IO_REWRITE;
} else {
BP_ZERO(bp);
zio->io_pipeline = ZIO_WRITE_PIPELINE;
}
2008-11-20 20:01:55 +00:00
if (psize == 0) {
if (zio->io_bp_orig.blk_birth != 0 &&
spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) {
BP_SET_LSIZE(bp, lsize);
BP_SET_TYPE(bp, zp->zp_type);
BP_SET_LEVEL(bp, zp->zp_level);
BP_SET_BIRTH(bp, zio->io_txg, 0);
}
zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
} else {
ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER);
BP_SET_LSIZE(bp, lsize);
BP_SET_TYPE(bp, zp->zp_type);
BP_SET_LEVEL(bp, zp->zp_level);
BP_SET_PSIZE(bp, psize);
BP_SET_COMPRESS(bp, compress);
BP_SET_CHECKSUM(bp, zp->zp_checksum);
BP_SET_DEDUP(bp, zp->zp_dedup);
BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER);
if (zp->zp_dedup) {
ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE;
}
if (zp->zp_nopwrite) {
ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
zio->io_pipeline |= ZIO_STAGE_NOP_WRITE;
}
}
return (ZIO_PIPELINE_CONTINUE);
}
static int
zio_free_bp_init(zio_t *zio)
{
blkptr_t *bp = zio->io_bp;
if (zio->io_child_type == ZIO_CHILD_LOGICAL) {
if (BP_GET_DEDUP(bp))
zio->io_pipeline = ZIO_DDT_FREE_PIPELINE;
}
2008-11-20 20:01:55 +00:00
return (ZIO_PIPELINE_CONTINUE);
}
/*
* ==========================================================================
* Execute the I/O pipeline
* ==========================================================================
*/
static void
Illumos #3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock contention 3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock is piping hot Reviewed by: Matthew Ahrens <mahrens@delphix.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Gordon Ross <gordon.ross@nexenta.com> Approved by: Richard Lowe <richlowe@richlowe.net> References: illumos/illumos-gate@ec94d32 https://illumos.org/issues/3581 Notes for Linux port: Earlier commit 08d08eb reduced contention on this taskq lock by simply reducing the number of z_fr_iss threads from 100 to one-per-CPU. We also optimized the taskq implementation in zfsonlinux/spl@3c6ed54. These changes significantly improved unlink performance to acceptable levels. This patch further reduces time spent spinning on this lock by randomly dispatching the work items over multiple independent task queues. The Illumos ZFS developers stated that this lock contention only arose after "3329 spa_sync() spends 10-20% of its time in spa_free_sync_cb()" was landed. It's not clear if 3329 affects the Linux port or not. I didn't see spa_free_sync_cb() show up in oprofile sessions while unlinking large files, but I may just not have used the right test case. I tested unlinking a 1 TB of data with and without the patch and didn't observe a meaningful difference in elapsed time. However, oprofile showed that the percent time spent in taskq_thread() was reduced from about 16% to about 5%. Aside from a possible slight performance benefit this may be worth landing if only for the sake of maintaining consistency with upstream. Ported-by: Ned Bass <bass6@llnl.gov> Closes #1327
2013-05-06 19:24:30 +00:00
zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline)
2008-11-20 20:01:55 +00:00
{
spa_t *spa = zio->io_spa;
zio_type_t t = zio->io_type;
int flags = (cutinline ? TQ_FRONT : 0);
2008-11-20 20:01:55 +00:00
/*
2009-07-02 22:44:48 +00:00
* If we're a config writer or a probe, the normal issue and
* interrupt threads may all be blocked waiting for the config lock.
* In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL.
2008-11-20 20:01:55 +00:00
*/
2009-07-02 22:44:48 +00:00
if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE))
t = ZIO_TYPE_NULL;
2008-11-20 20:01:55 +00:00
/*
* A similar issue exists for the L2ARC write thread until L2ARC 2.0.
2008-11-20 20:01:55 +00:00
*/
if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux)
t = ZIO_TYPE_NULL;
2008-11-20 20:01:55 +00:00
/*
Illumos #3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock contention 3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock is piping hot Reviewed by: Matthew Ahrens <mahrens@delphix.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Gordon Ross <gordon.ross@nexenta.com> Approved by: Richard Lowe <richlowe@richlowe.net> References: illumos/illumos-gate@ec94d32 https://illumos.org/issues/3581 Notes for Linux port: Earlier commit 08d08eb reduced contention on this taskq lock by simply reducing the number of z_fr_iss threads from 100 to one-per-CPU. We also optimized the taskq implementation in zfsonlinux/spl@3c6ed54. These changes significantly improved unlink performance to acceptable levels. This patch further reduces time spent spinning on this lock by randomly dispatching the work items over multiple independent task queues. The Illumos ZFS developers stated that this lock contention only arose after "3329 spa_sync() spends 10-20% of its time in spa_free_sync_cb()" was landed. It's not clear if 3329 affects the Linux port or not. I didn't see spa_free_sync_cb() show up in oprofile sessions while unlinking large files, but I may just not have used the right test case. I tested unlinking a 1 TB of data with and without the patch and didn't observe a meaningful difference in elapsed time. However, oprofile showed that the percent time spent in taskq_thread() was reduced from about 16% to about 5%. Aside from a possible slight performance benefit this may be worth landing if only for the sake of maintaining consistency with upstream. Ported-by: Ned Bass <bass6@llnl.gov> Closes #1327
2013-05-06 19:24:30 +00:00
* If this is a high priority I/O, then use the high priority taskq if
* available.
*/
if (zio->io_priority == ZIO_PRIORITY_NOW &&
Illumos #3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock contention 3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock is piping hot Reviewed by: Matthew Ahrens <mahrens@delphix.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Gordon Ross <gordon.ross@nexenta.com> Approved by: Richard Lowe <richlowe@richlowe.net> References: illumos/illumos-gate@ec94d32 https://illumos.org/issues/3581 Notes for Linux port: Earlier commit 08d08eb reduced contention on this taskq lock by simply reducing the number of z_fr_iss threads from 100 to one-per-CPU. We also optimized the taskq implementation in zfsonlinux/spl@3c6ed54. These changes significantly improved unlink performance to acceptable levels. This patch further reduces time spent spinning on this lock by randomly dispatching the work items over multiple independent task queues. The Illumos ZFS developers stated that this lock contention only arose after "3329 spa_sync() spends 10-20% of its time in spa_free_sync_cb()" was landed. It's not clear if 3329 affects the Linux port or not. I didn't see spa_free_sync_cb() show up in oprofile sessions while unlinking large files, but I may just not have used the right test case. I tested unlinking a 1 TB of data with and without the patch and didn't observe a meaningful difference in elapsed time. However, oprofile showed that the percent time spent in taskq_thread() was reduced from about 16% to about 5%. Aside from a possible slight performance benefit this may be worth landing if only for the sake of maintaining consistency with upstream. Ported-by: Ned Bass <bass6@llnl.gov> Closes #1327
2013-05-06 19:24:30 +00:00
spa->spa_zio_taskq[t][q + 1].stqs_count != 0)
q++;
ASSERT3U(q, <, ZIO_TASKQ_TYPES);
/*
* NB: We are assuming that the zio can only be dispatched
* to a single taskq at a time. It would be a grievous error
* to dispatch the zio to another taskq at the same time.
*/
ASSERT(taskq_empty_ent(&zio->io_tqent));
Illumos #3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock contention 3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock is piping hot Reviewed by: Matthew Ahrens <mahrens@delphix.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Gordon Ross <gordon.ross@nexenta.com> Approved by: Richard Lowe <richlowe@richlowe.net> References: illumos/illumos-gate@ec94d32 https://illumos.org/issues/3581 Notes for Linux port: Earlier commit 08d08eb reduced contention on this taskq lock by simply reducing the number of z_fr_iss threads from 100 to one-per-CPU. We also optimized the taskq implementation in zfsonlinux/spl@3c6ed54. These changes significantly improved unlink performance to acceptable levels. This patch further reduces time spent spinning on this lock by randomly dispatching the work items over multiple independent task queues. The Illumos ZFS developers stated that this lock contention only arose after "3329 spa_sync() spends 10-20% of its time in spa_free_sync_cb()" was landed. It's not clear if 3329 affects the Linux port or not. I didn't see spa_free_sync_cb() show up in oprofile sessions while unlinking large files, but I may just not have used the right test case. I tested unlinking a 1 TB of data with and without the patch and didn't observe a meaningful difference in elapsed time. However, oprofile showed that the percent time spent in taskq_thread() was reduced from about 16% to about 5%. Aside from a possible slight performance benefit this may be worth landing if only for the sake of maintaining consistency with upstream. Ported-by: Ned Bass <bass6@llnl.gov> Closes #1327
2013-05-06 19:24:30 +00:00
spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio,
flags, &zio->io_tqent);
}
2008-11-20 20:01:55 +00:00
static boolean_t
Illumos #3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock contention 3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock is piping hot Reviewed by: Matthew Ahrens <mahrens@delphix.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Gordon Ross <gordon.ross@nexenta.com> Approved by: Richard Lowe <richlowe@richlowe.net> References: illumos/illumos-gate@ec94d32 https://illumos.org/issues/3581 Notes for Linux port: Earlier commit 08d08eb reduced contention on this taskq lock by simply reducing the number of z_fr_iss threads from 100 to one-per-CPU. We also optimized the taskq implementation in zfsonlinux/spl@3c6ed54. These changes significantly improved unlink performance to acceptable levels. This patch further reduces time spent spinning on this lock by randomly dispatching the work items over multiple independent task queues. The Illumos ZFS developers stated that this lock contention only arose after "3329 spa_sync() spends 10-20% of its time in spa_free_sync_cb()" was landed. It's not clear if 3329 affects the Linux port or not. I didn't see spa_free_sync_cb() show up in oprofile sessions while unlinking large files, but I may just not have used the right test case. I tested unlinking a 1 TB of data with and without the patch and didn't observe a meaningful difference in elapsed time. However, oprofile showed that the percent time spent in taskq_thread() was reduced from about 16% to about 5%. Aside from a possible slight performance benefit this may be worth landing if only for the sake of maintaining consistency with upstream. Ported-by: Ned Bass <bass6@llnl.gov> Closes #1327
2013-05-06 19:24:30 +00:00
zio_taskq_member(zio_t *zio, zio_taskq_type_t q)
{
kthread_t *executor = zio->io_executor;
spa_t *spa = zio->io_spa;
zio_type_t t;
2008-11-20 20:01:55 +00:00
Illumos #3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock contention 3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock is piping hot Reviewed by: Matthew Ahrens <mahrens@delphix.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Gordon Ross <gordon.ross@nexenta.com> Approved by: Richard Lowe <richlowe@richlowe.net> References: illumos/illumos-gate@ec94d32 https://illumos.org/issues/3581 Notes for Linux port: Earlier commit 08d08eb reduced contention on this taskq lock by simply reducing the number of z_fr_iss threads from 100 to one-per-CPU. We also optimized the taskq implementation in zfsonlinux/spl@3c6ed54. These changes significantly improved unlink performance to acceptable levels. This patch further reduces time spent spinning on this lock by randomly dispatching the work items over multiple independent task queues. The Illumos ZFS developers stated that this lock contention only arose after "3329 spa_sync() spends 10-20% of its time in spa_free_sync_cb()" was landed. It's not clear if 3329 affects the Linux port or not. I didn't see spa_free_sync_cb() show up in oprofile sessions while unlinking large files, but I may just not have used the right test case. I tested unlinking a 1 TB of data with and without the patch and didn't observe a meaningful difference in elapsed time. However, oprofile showed that the percent time spent in taskq_thread() was reduced from about 16% to about 5%. Aside from a possible slight performance benefit this may be worth landing if only for the sake of maintaining consistency with upstream. Ported-by: Ned Bass <bass6@llnl.gov> Closes #1327
2013-05-06 19:24:30 +00:00
for (t = 0; t < ZIO_TYPES; t++) {
spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
uint_t i;
for (i = 0; i < tqs->stqs_count; i++) {
if (taskq_member(tqs->stqs_taskq[i], executor))
return (B_TRUE);
}
}
2008-11-20 20:01:55 +00:00
return (B_FALSE);
}
2008-11-20 20:01:55 +00:00
static int
zio_issue_async(zio_t *zio)
{
zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
}
void
zio_interrupt(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE);
}
2008-11-20 20:01:55 +00:00
void
zio_delay_interrupt(zio_t *zio)
{
/*
* The timeout_generic() function isn't defined in userspace, so
* rather than trying to implement the function, the zio delay
* functionality has been disabled for userspace builds.
*/
#ifdef _KERNEL
/*
* If io_target_timestamp is zero, then no delay has been registered
* for this IO, thus jump to the end of this function and "skip" the
* delay; issuing it directly to the zio layer.
*/
if (zio->io_target_timestamp != 0) {
hrtime_t now = gethrtime();
if (now >= zio->io_target_timestamp) {
/*
* This IO has already taken longer than the target
* delay to complete, so we don't want to delay it
* any longer; we "miss" the delay and issue it
* directly to the zio layer. This is likely due to
* the target latency being set to a value less than
* the underlying hardware can satisfy (e.g. delay
* set to 1ms, but the disks take 10ms to complete an
* IO request).
*/
DTRACE_PROBE2(zio__delay__miss, zio_t *, zio,
hrtime_t, now);
zio_interrupt(zio);
} else {
taskqid_t tid;
hrtime_t diff = zio->io_target_timestamp - now;
clock_t expire_at_tick = ddi_get_lbolt() +
NSEC_TO_TICK(diff);
DTRACE_PROBE3(zio__delay__hit, zio_t *, zio,
hrtime_t, now, hrtime_t, diff);
if (NSEC_TO_TICK(diff) == 0) {
/* Our delay is less than a jiffy - just spin */
zfs_sleep_until(zio->io_target_timestamp);
} else {
/*
* Use taskq_dispatch_delay() in the place of
* OpenZFS's timeout_generic().
*/
tid = taskq_dispatch_delay(system_taskq,
(task_func_t *) zio_interrupt,
zio, TQ_NOSLEEP, expire_at_tick);
if (!tid) {
/*
* Couldn't allocate a task. Just
* finish the zio without a delay.
*/
zio_interrupt(zio);
}
}
}
return;
}
#endif
DTRACE_PROBE1(zio__delay__skip, zio_t *, zio);
zio_interrupt(zio);
}
/*
* Execute the I/O pipeline until one of the following occurs:
* (1) the I/O completes; (2) the pipeline stalls waiting for
* dependent child I/Os; (3) the I/O issues, so we're waiting
* for an I/O completion interrupt; (4) the I/O is delegated by
* vdev-level caching or aggregation; (5) the I/O is deferred
* due to vdev-level queueing; (6) the I/O is handed off to
* another thread. In all cases, the pipeline stops whenever
* there's no CPU work; it never burns a thread in cv_wait_io().
*
* There's no locking on io_stage because there's no legitimate way
* for multiple threads to be attempting to process the same I/O.
*/
static zio_pipe_stage_t *zio_pipeline[];
2008-11-20 20:01:55 +00:00
/*
* zio_execute() is a wrapper around the static function
* __zio_execute() so that we can force __zio_execute() to be
* inlined. This reduces stack overhead which is important
* because __zio_execute() is called recursively in several zio
* code paths. zio_execute() itself cannot be inlined because
* it is externally visible.
*/
void
zio_execute(zio_t *zio)
{
fstrans_cookie_t cookie;
cookie = spl_fstrans_mark();
__zio_execute(zio);
spl_fstrans_unmark(cookie);
}
/*
* Used to determine if in the current context the stack is sized large
* enough to allow zio_execute() to be called recursively. A minimum
* stack size of 16K is required to avoid needing to re-dispatch the zio.
*/
boolean_t
zio_execute_stack_check(zio_t *zio)
{
#if !defined(HAVE_LARGE_STACKS)
dsl_pool_t *dp = spa_get_dsl(zio->io_spa);
/* Executing in txg_sync_thread() context. */
if (dp && curthread == dp->dp_tx.tx_sync_thread)
return (B_TRUE);
/* Pool initialization outside of zio_taskq context. */
if (dp && spa_is_initializing(dp->dp_spa) &&
!zio_taskq_member(zio, ZIO_TASKQ_ISSUE) &&
!zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH))
return (B_TRUE);
#endif /* HAVE_LARGE_STACKS */
return (B_FALSE);
}
__attribute__((always_inline))
static inline void
__zio_execute(zio_t *zio)
{
zio->io_executor = curthread;
2008-11-20 20:01:55 +00:00
while (zio->io_stage < ZIO_STAGE_DONE) {
enum zio_stage pipeline = zio->io_pipeline;
enum zio_stage stage = zio->io_stage;
int rv;
2008-11-20 20:01:55 +00:00
ASSERT(!MUTEX_HELD(&zio->io_lock));
ASSERT(ISP2(stage));
ASSERT(zio->io_stall == NULL);
2008-11-20 20:01:55 +00:00
do {
stage <<= 1;
} while ((stage & pipeline) == 0);
ASSERT(stage <= ZIO_STAGE_DONE);
2008-11-20 20:01:55 +00:00
/*
* If we are in interrupt context and this pipeline stage
* will grab a config lock that is held across I/O,
* or may wait for an I/O that needs an interrupt thread
* to complete, issue async to avoid deadlock.
*
* For VDEV_IO_START, we cut in line so that the io will
* be sent to disk promptly.
2008-11-20 20:01:55 +00:00
*/
if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL &&
zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) {
boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
zio_requeue_io_start_cut_in_line : B_FALSE;
zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
return;
}
/*
* If the current context doesn't have large enough stacks
* the zio must be issued asynchronously to prevent overflow.
*/
if (zio_execute_stack_check(zio)) {
boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ?
zio_requeue_io_start_cut_in_line : B_FALSE;
zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut);
return;
2008-11-20 20:01:55 +00:00
}
zio->io_stage = stage;
rv = zio_pipeline[highbit64(stage) - 1](zio);
2008-11-20 20:01:55 +00:00
if (rv == ZIO_PIPELINE_STOP)
return;
2008-11-20 20:01:55 +00:00
ASSERT(rv == ZIO_PIPELINE_CONTINUE);
}
2008-11-20 20:01:55 +00:00
}
/*
* ==========================================================================
* Initiate I/O, either sync or async
* ==========================================================================
*/
int
zio_wait(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
int error;
2008-11-20 20:01:55 +00:00
ASSERT(zio->io_stage == ZIO_STAGE_OPEN);
ASSERT(zio->io_executor == NULL);
2008-11-20 20:01:55 +00:00
zio->io_waiter = curthread;
2008-11-20 20:01:55 +00:00
__zio_execute(zio);
2008-11-20 20:01:55 +00:00
mutex_enter(&zio->io_lock);
while (zio->io_executor != NULL)
cv_wait_io(&zio->io_cv, &zio->io_lock);
mutex_exit(&zio->io_lock);
2008-11-20 20:01:55 +00:00
error = zio->io_error;
zio_destroy(zio);
2008-11-20 20:01:55 +00:00
return (error);
}
2008-11-20 20:01:55 +00:00
void
zio_nowait(zio_t *zio)
{
ASSERT(zio->io_executor == NULL);
2008-11-20 20:01:55 +00:00
2009-02-18 20:51:31 +00:00
if (zio->io_child_type == ZIO_CHILD_LOGICAL &&
zio_unique_parent(zio) == NULL) {
zio_t *pio;
2008-11-20 20:01:55 +00:00
/*
* This is a logical async I/O with no parent to wait for it.
2009-07-02 22:44:48 +00:00
* We add it to the spa_async_root_zio "Godfather" I/O which
* will ensure they complete prior to unloading the pool.
2008-11-20 20:01:55 +00:00
*/
spa_t *spa = zio->io_spa;
kpreempt_disable();
pio = spa->spa_async_zio_root[CPU_SEQID];
kpreempt_enable();
2009-07-02 22:44:48 +00:00
zio_add_child(pio, zio);
}
2008-11-20 20:01:55 +00:00
__zio_execute(zio);
}
2008-11-20 20:01:55 +00:00
/*
* ==========================================================================
* Reexecute or suspend/resume failed I/O
* ==========================================================================
*/
2008-11-20 20:01:55 +00:00
static void
zio_reexecute(zio_t *pio)
{
2009-02-18 20:51:31 +00:00
zio_t *cio, *cio_next;
int c, w;
2009-02-18 20:51:31 +00:00
ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL);
ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN);
2009-07-02 22:44:48 +00:00
ASSERT(pio->io_gang_leader == NULL);
ASSERT(pio->io_gang_tree == NULL);
2008-11-20 20:01:55 +00:00
pio->io_flags = pio->io_orig_flags;
pio->io_stage = pio->io_orig_stage;
pio->io_pipeline = pio->io_orig_pipeline;
pio->io_reexecute = 0;
pio->io_flags |= ZIO_FLAG_REEXECUTED;
pio->io_error = 0;
for (w = 0; w < ZIO_WAIT_TYPES; w++)
2009-02-18 20:51:31 +00:00
pio->io_state[w] = 0;
for (c = 0; c < ZIO_CHILD_TYPES; c++)
pio->io_child_error[c] = 0;
2008-11-20 20:01:55 +00:00
if (IO_IS_ALLOCATING(pio))
BP_ZERO(pio->io_bp);
2008-11-20 20:01:55 +00:00
/*
* As we reexecute pio's children, new children could be created.
2009-02-18 20:51:31 +00:00
* New children go to the head of pio's io_child_list, however,
* so we will (correctly) not reexecute them. The key is that
2009-02-18 20:51:31 +00:00
* the remainder of pio's io_child_list, from 'cio_next' onward,
* cannot be affected by any side effects of reexecuting 'cio'.
*/
2009-02-18 20:51:31 +00:00
for (cio = zio_walk_children(pio); cio != NULL; cio = cio_next) {
cio_next = zio_walk_children(pio);
mutex_enter(&pio->io_lock);
for (w = 0; w < ZIO_WAIT_TYPES; w++)
2009-02-18 20:51:31 +00:00
pio->io_children[cio->io_child_type][w]++;
mutex_exit(&pio->io_lock);
2009-02-18 20:51:31 +00:00
zio_reexecute(cio);
2008-11-20 20:01:55 +00:00
}
/*
* Now that all children have been reexecuted, execute the parent.
2009-07-02 22:44:48 +00:00
* We don't reexecute "The Godfather" I/O here as it's the
* responsibility of the caller to wait on him.
*/
2009-07-02 22:44:48 +00:00
if (!(pio->io_flags & ZIO_FLAG_GODFATHER))
__zio_execute(pio);
2008-11-20 20:01:55 +00:00
}
void
zio_suspend(spa_t *spa, zio_t *zio)
2008-11-20 20:01:55 +00:00
{
if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC)
fm_panic("Pool '%s' has encountered an uncorrectable I/O "
"failure and the failure mode property for this pool "
"is set to panic.", spa_name(spa));
2008-11-20 20:01:55 +00:00
cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O "
"failure and has been suspended.\n", spa_name(spa));
zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, NULL, 0, 0);
2008-11-20 20:01:55 +00:00
mutex_enter(&spa->spa_suspend_lock);
2008-11-20 20:01:55 +00:00
if (spa->spa_suspend_zio_root == NULL)
2009-07-02 22:44:48 +00:00
spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL,
ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
ZIO_FLAG_GODFATHER);
2008-11-20 20:01:55 +00:00
spa->spa_suspended = B_TRUE;
2008-11-20 20:01:55 +00:00
if (zio != NULL) {
2009-07-02 22:44:48 +00:00
ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
ASSERT(zio != spa->spa_suspend_zio_root);
ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2009-02-18 20:51:31 +00:00
ASSERT(zio_unique_parent(zio) == NULL);
ASSERT(zio->io_stage == ZIO_STAGE_DONE);
zio_add_child(spa->spa_suspend_zio_root, zio);
}
2008-11-20 20:01:55 +00:00
mutex_exit(&spa->spa_suspend_lock);
}
2008-11-20 20:01:55 +00:00
2009-07-02 22:44:48 +00:00
int
zio_resume(spa_t *spa)
{
2009-07-02 22:44:48 +00:00
zio_t *pio;
2008-11-20 20:01:55 +00:00
/*
* Reexecute all previously suspended i/o.
2008-11-20 20:01:55 +00:00
*/
mutex_enter(&spa->spa_suspend_lock);
spa->spa_suspended = B_FALSE;
cv_broadcast(&spa->spa_suspend_cv);
pio = spa->spa_suspend_zio_root;
spa->spa_suspend_zio_root = NULL;
mutex_exit(&spa->spa_suspend_lock);
if (pio == NULL)
2009-07-02 22:44:48 +00:00
return (0);
2008-11-20 20:01:55 +00:00
2009-07-02 22:44:48 +00:00
zio_reexecute(pio);
return (zio_wait(pio));
}
void
zio_resume_wait(spa_t *spa)
{
mutex_enter(&spa->spa_suspend_lock);
while (spa_suspended(spa))
cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock);
mutex_exit(&spa->spa_suspend_lock);
2008-11-20 20:01:55 +00:00
}
/*
* ==========================================================================
* Gang blocks.
*
* A gang block is a collection of small blocks that looks to the DMU
* like one large block. When zio_dva_allocate() cannot find a block
* of the requested size, due to either severe fragmentation or the pool
* being nearly full, it calls zio_write_gang_block() to construct the
* block from smaller fragments.
*
* A gang block consists of a gang header (zio_gbh_phys_t) and up to
* three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like
* an indirect block: it's an array of block pointers. It consumes
* only one sector and hence is allocatable regardless of fragmentation.
* The gang header's bps point to its gang members, which hold the data.
*
* Gang blocks are self-checksumming, using the bp's <vdev, offset, txg>
* as the verifier to ensure uniqueness of the SHA256 checksum.
* Critically, the gang block bp's blk_cksum is the checksum of the data,
* not the gang header. This ensures that data block signatures (needed for
* deduplication) are independent of how the block is physically stored.
*
* Gang blocks can be nested: a gang member may itself be a gang block.
* Thus every gang block is a tree in which root and all interior nodes are
* gang headers, and the leaves are normal blocks that contain user data.
* The root of the gang tree is called the gang leader.
*
* To perform any operation (read, rewrite, free, claim) on a gang block,
* zio_gang_assemble() first assembles the gang tree (minus data leaves)
* in the io_gang_tree field of the original logical i/o by recursively
* reading the gang leader and all gang headers below it. This yields
* an in-core tree containing the contents of every gang header and the
* bps for every constituent of the gang block.
*
* With the gang tree now assembled, zio_gang_issue() just walks the gang tree
* and invokes a callback on each bp. To free a gang block, zio_gang_issue()
* calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp.
* zio_claim_gang() provides a similarly trivial wrapper for zio_claim().
* zio_read_gang() is a wrapper around zio_read() that omits reading gang
* headers, since we already have those in io_gang_tree. zio_rewrite_gang()
* performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite()
* of the gang header plus zio_checksum_compute() of the data to update the
* gang header's blk_cksum as described above.
*
* The two-phase assemble/issue model solves the problem of partial failure --
* what if you'd freed part of a gang block but then couldn't read the
* gang header for another part? Assembling the entire gang tree first
* ensures that all the necessary gang header I/O has succeeded before
* starting the actual work of free, claim, or write. Once the gang tree
* is assembled, free and claim are in-memory operations that cannot fail.
*
* In the event that a gang write fails, zio_dva_unallocate() walks the
* gang tree to immediately free (i.e. insert back into the space map)
* everything we've allocated. This ensures that we don't get ENOSPC
* errors during repeated suspend/resume cycles due to a flaky device.
*
* Gang rewrites only happen during sync-to-convergence. If we can't assemble
* the gang tree, we won't modify the block, so we can safely defer the free
* (knowing that the block is still intact). If we *can* assemble the gang
* tree, then even if some of the rewrites fail, zio_dva_unallocate() will free
* each constituent bp and we can allocate a new block on the next sync pass.
*
* In all cases, the gang tree allows complete recovery from partial failure.
2008-11-20 20:01:55 +00:00
* ==========================================================================
*/
static zio_t *
zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
2008-11-20 20:01:55 +00:00
{
if (gn != NULL)
return (pio);
2008-11-20 20:01:55 +00:00
return (zio_read(pio, pio->io_spa, bp, data, BP_GET_PSIZE(bp),
NULL, NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio),
&pio->io_bookmark));
}
zio_t *
zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
{
zio_t *zio;
if (gn != NULL) {
zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
gn->gn_gbh, SPA_GANGBLOCKSIZE, NULL, NULL, pio->io_priority,
ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2008-11-20 20:01:55 +00:00
/*
* As we rewrite each gang header, the pipeline will compute
* a new gang block header checksum for it; but no one will
* compute a new data checksum, so we do that here. The one
* exception is the gang leader: the pipeline already computed
* its data checksum because that stage precedes gang assembly.
* (Presently, nothing actually uses interior data checksums;
* this is just good hygiene.)
2008-11-20 20:01:55 +00:00
*/
2009-07-02 22:44:48 +00:00
if (gn != pio->io_gang_leader->io_gang_tree) {
zio_checksum_compute(zio, BP_GET_CHECKSUM(bp),
data, BP_GET_PSIZE(bp));
}
/*
* If we are here to damage data for testing purposes,
* leave the GBH alone so that we can detect the damage.
*/
if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE)
zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
2008-11-20 20:01:55 +00:00
} else {
zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp,
data, BP_GET_PSIZE(bp), NULL, NULL, pio->io_priority,
ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2008-11-20 20:01:55 +00:00
}
return (zio);
}
2008-11-20 20:01:55 +00:00
/* ARGSUSED */
zio_t *
zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
{
return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp,
ZIO_GANG_CHILD_FLAGS(pio)));
2008-11-20 20:01:55 +00:00
}
/* ARGSUSED */
zio_t *
zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, void *data)
2008-11-20 20:01:55 +00:00
{
return (zio_claim(pio, pio->io_spa, pio->io_txg, bp,
NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio)));
}
static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = {
NULL,
zio_read_gang,
zio_rewrite_gang,
zio_free_gang,
zio_claim_gang,
NULL
};
2008-11-20 20:01:55 +00:00
static void zio_gang_tree_assemble_done(zio_t *zio);
2008-11-20 20:01:55 +00:00
static zio_gang_node_t *
zio_gang_node_alloc(zio_gang_node_t **gnpp)
{
zio_gang_node_t *gn;
2008-11-20 20:01:55 +00:00
ASSERT(*gnpp == NULL);
2008-11-20 20:01:55 +00:00
gn = kmem_zalloc(sizeof (*gn), KM_SLEEP);
gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE);
*gnpp = gn;
2008-11-20 20:01:55 +00:00
return (gn);
2008-11-20 20:01:55 +00:00
}
static void
zio_gang_node_free(zio_gang_node_t **gnpp)
2008-11-20 20:01:55 +00:00
{
zio_gang_node_t *gn = *gnpp;
int g;
2008-11-20 20:01:55 +00:00
for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
ASSERT(gn->gn_child[g] == NULL);
zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE);
kmem_free(gn, sizeof (*gn));
*gnpp = NULL;
2008-11-20 20:01:55 +00:00
}
static void
zio_gang_tree_free(zio_gang_node_t **gnpp)
2008-11-20 20:01:55 +00:00
{
zio_gang_node_t *gn = *gnpp;
int g;
2008-11-20 20:01:55 +00:00
if (gn == NULL)
return;
2008-11-20 20:01:55 +00:00
for (g = 0; g < SPA_GBH_NBLKPTRS; g++)
zio_gang_tree_free(&gn->gn_child[g]);
2008-11-20 20:01:55 +00:00
zio_gang_node_free(gnpp);
2008-11-20 20:01:55 +00:00
}
static void
2009-07-02 22:44:48 +00:00
zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp)
2008-11-20 20:01:55 +00:00
{
zio_gang_node_t *gn = zio_gang_node_alloc(gnpp);
2009-07-02 22:44:48 +00:00
ASSERT(gio->io_gang_leader == gio);
ASSERT(BP_IS_GANG(bp));
2008-11-20 20:01:55 +00:00
2009-07-02 22:44:48 +00:00
zio_nowait(zio_read(gio, gio->io_spa, bp, gn->gn_gbh,
SPA_GANGBLOCKSIZE, zio_gang_tree_assemble_done, gn,
2009-07-02 22:44:48 +00:00
gio->io_priority, ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark));
}
2008-11-20 20:01:55 +00:00
static void
zio_gang_tree_assemble_done(zio_t *zio)
{
2009-07-02 22:44:48 +00:00
zio_t *gio = zio->io_gang_leader;
zio_gang_node_t *gn = zio->io_private;
blkptr_t *bp = zio->io_bp;
int g;
2008-11-20 20:01:55 +00:00
2009-07-02 22:44:48 +00:00
ASSERT(gio == zio_unique_parent(zio));
ASSERT(zio->io_child_count == 0);
2008-11-20 20:01:55 +00:00
if (zio->io_error)
return;
2008-11-20 20:01:55 +00:00
if (BP_SHOULD_BYTESWAP(bp))
byteswap_uint64_array(zio->io_data, zio->io_size);
2008-11-20 20:01:55 +00:00
ASSERT(zio->io_data == gn->gn_gbh);
ASSERT(zio->io_size == SPA_GANGBLOCKSIZE);
ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2008-11-20 20:01:55 +00:00
for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
if (!BP_IS_GANG(gbp))
continue;
2009-07-02 22:44:48 +00:00
zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]);
}
2008-11-20 20:01:55 +00:00
}
static void
zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, void *data)
2008-11-20 20:01:55 +00:00
{
2009-07-02 22:44:48 +00:00
zio_t *gio = pio->io_gang_leader;
zio_t *zio;
int g;
2008-11-20 20:01:55 +00:00
ASSERT(BP_IS_GANG(bp) == !!gn);
2009-07-02 22:44:48 +00:00
ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp));
ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree);
2008-11-20 20:01:55 +00:00
/*
* If you're a gang header, your data is in gn->gn_gbh.
* If you're a gang member, your data is in 'data' and gn == NULL.
*/
2009-07-02 22:44:48 +00:00
zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data);
2008-11-20 20:01:55 +00:00
if (gn != NULL) {
ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC);
2008-11-20 20:01:55 +00:00
for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g];
if (BP_IS_HOLE(gbp))
continue;
zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data);
data = (char *)data + BP_GET_PSIZE(gbp);
}
2008-11-20 20:01:55 +00:00
}
2009-07-02 22:44:48 +00:00
if (gn == gio->io_gang_tree)
ASSERT3P((char *)gio->io_data + gio->io_size, ==, data);
2008-11-20 20:01:55 +00:00
if (zio != pio)
zio_nowait(zio);
2008-11-20 20:01:55 +00:00
}
static int
zio_gang_assemble(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
blkptr_t *bp = zio->io_bp;
2008-11-20 20:01:55 +00:00
2009-07-02 22:44:48 +00:00
ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL);
ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
zio->io_gang_leader = zio;
2008-11-20 20:01:55 +00:00
zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree);
2008-11-20 20:01:55 +00:00
return (ZIO_PIPELINE_CONTINUE);
}
static int
zio_gang_issue(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
blkptr_t *bp = zio->io_bp;
2008-11-20 20:01:55 +00:00
if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE))
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
2009-07-02 22:44:48 +00:00
ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio);
ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
2008-11-20 20:01:55 +00:00
if (zio->io_child_error[ZIO_CHILD_GANG] == 0)
2009-07-02 22:44:48 +00:00
zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_data);
else
2009-07-02 22:44:48 +00:00
zio_gang_tree_free(&zio->io_gang_tree);
2008-11-20 20:01:55 +00:00
zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2008-11-20 20:01:55 +00:00
return (ZIO_PIPELINE_CONTINUE);
}
static void
zio_write_gang_member_ready(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
2009-02-18 20:51:31 +00:00
zio_t *pio = zio_unique_parent(zio);
2008-11-20 20:01:55 +00:00
dva_t *cdva = zio->io_bp->blk_dva;
dva_t *pdva = pio->io_bp->blk_dva;
uint64_t asize;
int d;
ASSERTV(zio_t *gio = zio->io_gang_leader);
2008-11-20 20:01:55 +00:00
if (BP_IS_HOLE(zio->io_bp))
return;
ASSERT(BP_IS_HOLE(&zio->io_bp_orig));
ASSERT(zio->io_child_type == ZIO_CHILD_GANG);
ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies);
ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp));
ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp));
2008-11-20 20:01:55 +00:00
ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp));
mutex_enter(&pio->io_lock);
for (d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) {
2008-11-20 20:01:55 +00:00
ASSERT(DVA_GET_GANG(&pdva[d]));
asize = DVA_GET_ASIZE(&pdva[d]);
asize += DVA_GET_ASIZE(&cdva[d]);
DVA_SET_ASIZE(&pdva[d], asize);
}
mutex_exit(&pio->io_lock);
}
static int
zio_write_gang_block(zio_t *pio)
2008-11-20 20:01:55 +00:00
{
spa_t *spa = pio->io_spa;
blkptr_t *bp = pio->io_bp;
2009-07-02 22:44:48 +00:00
zio_t *gio = pio->io_gang_leader;
zio_t *zio;
zio_gang_node_t *gn, **gnpp;
2008-11-20 20:01:55 +00:00
zio_gbh_phys_t *gbh;
uint64_t txg = pio->io_txg;
uint64_t resid = pio->io_size;
uint64_t lsize;
int copies = gio->io_prop.zp_copies;
int gbh_copies = MIN(copies + 1, spa_max_replication(spa));
zio_prop_t zp;
int g, error;
2008-11-20 20:01:55 +00:00
error = metaslab_alloc(spa, spa_normal_class(spa), SPA_GANGBLOCKSIZE,
bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp,
METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER);
2008-11-20 20:01:55 +00:00
if (error) {
pio->io_error = error;
2008-11-20 20:01:55 +00:00
return (ZIO_PIPELINE_CONTINUE);
}
2009-07-02 22:44:48 +00:00
if (pio == gio) {
gnpp = &gio->io_gang_tree;
} else {
gnpp = pio->io_private;
ASSERT(pio->io_ready == zio_write_gang_member_ready);
2008-11-20 20:01:55 +00:00
}
gn = zio_gang_node_alloc(gnpp);
gbh = gn->gn_gbh;
bzero(gbh, SPA_GANGBLOCKSIZE);
2008-11-20 20:01:55 +00:00
/*
* Create the gang header.
*/
zio = zio_rewrite(pio, spa, txg, bp, gbh, SPA_GANGBLOCKSIZE, NULL, NULL,
pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark);
2008-11-20 20:01:55 +00:00
/*
* Create and nowait the gang children.
*/
for (g = 0; resid != 0; resid -= lsize, g++) {
lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g),
SPA_MINBLOCKSIZE);
ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid);
2009-07-02 22:44:48 +00:00
zp.zp_checksum = gio->io_prop.zp_checksum;
zp.zp_compress = ZIO_COMPRESS_OFF;
zp.zp_type = DMU_OT_NONE;
zp.zp_level = 0;
zp.zp_copies = gio->io_prop.zp_copies;
zp.zp_dedup = B_FALSE;
zp.zp_dedup_verify = B_FALSE;
zp.zp_nopwrite = B_FALSE;
zio_nowait(zio_write(zio, spa, txg, &gbh->zg_blkptr[g],
(char *)pio->io_data + (pio->io_size - resid), lsize,
lsize, &zp, zio_write_gang_member_ready, NULL, NULL, NULL,
&gn->gn_child[g], pio->io_priority,
ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark));
}
2008-11-20 20:01:55 +00:00
/*
* Set pio's pipeline to just wait for zio to finish.
2008-11-20 20:01:55 +00:00
*/
pio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 13:20:20 +00:00
/*
* We didn't allocate this bp, so make sure it doesn't get unmarked.
*/
pio->io_flags &= ~ZIO_FLAG_FASTWRITE;
zio_nowait(zio);
return (ZIO_PIPELINE_CONTINUE);
2008-11-20 20:01:55 +00:00
}
/*
* The zio_nop_write stage in the pipeline determines if allocating
* a new bp is necessary. By leveraging a cryptographically secure checksum,
* such as SHA256, we can compare the checksums of the new data and the old
* to determine if allocating a new block is required. The nopwrite
* feature can handle writes in either syncing or open context (i.e. zil
* writes) and as a result is mutually exclusive with dedup.
*/
static int
zio_nop_write(zio_t *zio)
{
blkptr_t *bp = zio->io_bp;
blkptr_t *bp_orig = &zio->io_bp_orig;
zio_prop_t *zp = &zio->io_prop;
ASSERT(BP_GET_LEVEL(bp) == 0);
ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE));
ASSERT(zp->zp_nopwrite);
ASSERT(!zp->zp_dedup);
ASSERT(zio->io_bp_override == NULL);
ASSERT(IO_IS_ALLOCATING(zio));
/*
* Check to see if the original bp and the new bp have matching
* characteristics (i.e. same checksum, compression algorithms, etc).
* If they don't then just continue with the pipeline which will
* allocate a new bp.
*/
if (BP_IS_HOLE(bp_orig) ||
!zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_dedup ||
BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) ||
BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) ||
BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) ||
zp->zp_copies != BP_GET_NDVAS(bp_orig))
return (ZIO_PIPELINE_CONTINUE);
/*
* If the checksums match then reset the pipeline so that we
* avoid allocating a new bp and issuing any I/O.
*/
if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) {
ASSERT(zio_checksum_table[zp->zp_checksum].ci_dedup);
ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig));
ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig));
ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF);
ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop,
sizeof (uint64_t)) == 0);
*bp = *bp_orig;
zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
zio->io_flags |= ZIO_FLAG_NOPWRITE;
}
return (ZIO_PIPELINE_CONTINUE);
}
2008-11-20 20:01:55 +00:00
/*
* ==========================================================================
* Dedup
2008-11-20 20:01:55 +00:00
* ==========================================================================
*/
static void
zio_ddt_child_read_done(zio_t *zio)
{
blkptr_t *bp = zio->io_bp;
ddt_entry_t *dde = zio->io_private;
ddt_phys_t *ddp;
zio_t *pio = zio_unique_parent(zio);
mutex_enter(&pio->io_lock);
ddp = ddt_phys_select(dde, bp);
if (zio->io_error == 0)
ddt_phys_clear(ddp); /* this ddp doesn't need repair */
if (zio->io_error == 0 && dde->dde_repair_data == NULL)
dde->dde_repair_data = zio->io_data;
else
zio_buf_free(zio->io_data, zio->io_size);
mutex_exit(&pio->io_lock);
}
static int
zio_ddt_read_start(zio_t *zio)
{
blkptr_t *bp = zio->io_bp;
int p;
ASSERT(BP_GET_DEDUP(bp));
ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
if (zio->io_child_error[ZIO_CHILD_DDT]) {
ddt_t *ddt = ddt_select(zio->io_spa, bp);
ddt_entry_t *dde = ddt_repair_start(ddt, bp);
ddt_phys_t *ddp = dde->dde_phys;
ddt_phys_t *ddp_self = ddt_phys_select(dde, bp);
blkptr_t blk;
ASSERT(zio->io_vsd == NULL);
zio->io_vsd = dde;
if (ddp_self == NULL)
return (ZIO_PIPELINE_CONTINUE);
for (p = 0; p < DDT_PHYS_TYPES; p++, ddp++) {
if (ddp->ddp_phys_birth == 0 || ddp == ddp_self)
continue;
ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp,
&blk);
zio_nowait(zio_read(zio, zio->io_spa, &blk,
zio_buf_alloc(zio->io_size), zio->io_size,
zio_ddt_child_read_done, dde, zio->io_priority,
ZIO_DDT_CHILD_FLAGS(zio) | ZIO_FLAG_DONT_PROPAGATE,
&zio->io_bookmark));
}
return (ZIO_PIPELINE_CONTINUE);
}
zio_nowait(zio_read(zio, zio->io_spa, bp,
zio->io_data, zio->io_size, NULL, NULL, zio->io_priority,
ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark));
return (ZIO_PIPELINE_CONTINUE);
}
static int
zio_ddt_read_done(zio_t *zio)
{
blkptr_t *bp = zio->io_bp;
if (zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE))
return (ZIO_PIPELINE_STOP);
ASSERT(BP_GET_DEDUP(bp));
ASSERT(BP_GET_PSIZE(bp) == zio->io_size);
ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
if (zio->io_child_error[ZIO_CHILD_DDT]) {
ddt_t *ddt = ddt_select(zio->io_spa, bp);
ddt_entry_t *dde = zio->io_vsd;
if (ddt == NULL) {
ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE);
return (ZIO_PIPELINE_CONTINUE);
}
if (dde == NULL) {
zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1;
zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE);
return (ZIO_PIPELINE_STOP);
}
if (dde->dde_repair_data != NULL) {
bcopy(dde->dde_repair_data, zio->io_data, zio->io_size);
zio->io_child_error[ZIO_CHILD_DDT] = 0;
}
ddt_repair_done(ddt, dde);
zio->io_vsd = NULL;
}
ASSERT(zio->io_vsd == NULL);
return (ZIO_PIPELINE_CONTINUE);
}
static boolean_t
zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde)
{
spa_t *spa = zio->io_spa;
int p;
ASSERT0(zio->io_flags & ZIO_FLAG_RAW);
/*
* Note: we compare the original data, not the transformed data,
* because when zio->io_bp is an override bp, we will not have
* pushed the I/O transforms. That's an important optimization
* because otherwise we'd compress/encrypt all dmu_sync() data twice.
*/
for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
zio_t *lio = dde->dde_lead_zio[p];
if (lio != NULL) {
return (lio->io_orig_size != zio->io_orig_size ||
bcmp(zio->io_orig_data, lio->io_orig_data,
zio->io_orig_size) != 0);
}
}
for (p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) {
ddt_phys_t *ddp = &dde->dde_phys[p];
if (ddp->ddp_phys_birth != 0) {
arc_buf_t *abuf = NULL;
arc_flags_t aflags = ARC_FLAG_WAIT;
blkptr_t blk = *zio->io_bp;
int error;
ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth);
ddt_exit(ddt);
error = arc_read(NULL, spa, &blk,
arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ,
ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
&aflags, &zio->io_bookmark);
if (error == 0) {
if (arc_buf_size(abuf) != zio->io_orig_size ||
bcmp(abuf->b_data, zio->io_orig_data,
zio->io_orig_size) != 0)
error = SET_ERROR(EEXIST);
OpenZFS 6950 - ARC should cache compressed data Authored by: George Wilson <george.wilson@delphix.com> Reviewed by: Prakash Surya <prakash.surya@delphix.com> Reviewed by: Dan Kimmel <dan.kimmel@delphix.com> Reviewed by: Matt Ahrens <mahrens@delphix.com> Reviewed by: Paul Dagnelie <pcd@delphix.com> Reviewed by: Tom Caputi <tcaputi@datto.com> Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov> Ported by: David Quigley <david.quigley@intel.com> This review covers the reading and writing of compressed arc headers, sharing data between the arc_hdr_t and the arc_buf_t, and the implementation of a new dbuf cache to keep frequently access data uncompressed. I've added a new member to l1 arc hdr called b_pdata. The b_pdata always hangs off the arc_buf_hdr_t (if an L1 hdr is in use) and points to the physical block for that DVA. The physical block may or may not be compressed. If compressed arc is enabled and the block on-disk is compressed, then the b_pdata will match the block on-disk and remain compressed in memory. If the block on disk is not compressed, then neither will the b_pdata. Lastly, if compressed arc is disabled, then b_pdata will always be an uncompressed version of the on-disk block. Typically the arc will cache only the arc_buf_hdr_t and will aggressively evict any arc_buf_t's that are no longer referenced. This means that the arc will primarily have compressed blocks as the arc_buf_t's are considered overhead and are always uncompressed. When a consumer reads a block we first look to see if the arc_buf_hdr_t is cached. If the hdr is cached then we allocate a new arc_buf_t and decompress the b_pdata contents into the arc_buf_t's b_data. If the hdr already has a arc_buf_t, then we will allocate an additional arc_buf_t and bcopy the uncompressed contents from the first arc_buf_t to the new one. Writing to the compressed arc requires that we first discard the b_pdata since the physical block is about to be rewritten. The new data contents will be passed in via an arc_buf_t (uncompressed) and during the I/O pipeline stages we will copy the physical block contents to a newly allocated b_pdata. When an l2arc is inuse it will also take advantage of the b_pdata. Now the l2arc will always write the contents of b_pdata to the l2arc. This means that when compressed arc is enabled that the l2arc blocks are identical to those stored in the main data pool. This provides a significant advantage since we can leverage the bp's checksum when reading from the l2arc to determine if the contents are valid. If the compressed arc is disabled, then we must first transform the read block to look like the physical block in the main data pool before comparing the checksum and determining it's valid. OpenZFS-issue: https://www.illumos.org/issues/6950 OpenZFS-commit: https://github.com/openzfs/openzfs/commit/7fc10f0 Issue #5078
2016-06-02 04:04:53 +00:00
arc_buf_destroy(abuf, &abuf);
}
ddt_enter(ddt);
return (error != 0);
}
}
return (B_FALSE);
}
static void
zio_ddt_child_write_ready(zio_t *zio)
{
int p = zio->io_prop.zp_copies;
ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
ddt_entry_t *dde = zio->io_private;
ddt_phys_t *ddp = &dde->dde_phys[p];
zio_t *pio;
if (zio->io_error)
return;
ddt_enter(ddt);
ASSERT(dde->dde_lead_zio[p] == zio);
ddt_phys_fill(ddp, zio->io_bp);
while ((pio = zio_walk_parents(zio)) != NULL)
ddt_bp_fill(ddp, pio->io_bp, zio->io_txg);
ddt_exit(ddt);
}
static void
zio_ddt_child_write_done(zio_t *zio)
{
int p = zio->io_prop.zp_copies;
ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp);
ddt_entry_t *dde = zio->io_private;
ddt_phys_t *ddp = &dde->dde_phys[p];
ddt_enter(ddt);
ASSERT(ddp->ddp_refcnt == 0);
ASSERT(dde->dde_lead_zio[p] == zio);
dde->dde_lead_zio[p] = NULL;
if (zio->io_error == 0) {
while (zio_walk_parents(zio) != NULL)
ddt_phys_addref(ddp);
} else {
ddt_phys_clear(ddp);
}
ddt_exit(ddt);
}
static void
zio_ddt_ditto_write_done(zio_t *zio)
{
int p = DDT_PHYS_DITTO;
blkptr_t *bp = zio->io_bp;
ddt_t *ddt = ddt_select(zio->io_spa, bp);
ddt_entry_t *dde = zio->io_private;
ddt_phys_t *ddp = &dde->dde_phys[p];
ddt_key_t *ddk = &dde->dde_key;
ASSERTV(zio_prop_t *zp = &zio->io_prop);
ddt_enter(ddt);
ASSERT(ddp->ddp_refcnt == 0);
ASSERT(dde->dde_lead_zio[p] == zio);
dde->dde_lead_zio[p] = NULL;
if (zio->io_error == 0) {
ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum));
ASSERT(zp->zp_copies < SPA_DVAS_PER_BP);
ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp));
if (ddp->ddp_phys_birth != 0)
ddt_phys_free(ddt, ddk, ddp, zio->io_txg);
ddt_phys_fill(ddp, bp);
}
ddt_exit(ddt);
}
static int
zio_ddt_write(zio_t *zio)
{
spa_t *spa = zio->io_spa;
blkptr_t *bp = zio->io_bp;
uint64_t txg = zio->io_txg;
zio_prop_t *zp = &zio->io_prop;
int p = zp->zp_copies;
int ditto_copies;
zio_t *cio = NULL;
zio_t *dio = NULL;
ddt_t *ddt = ddt_select(spa, bp);
ddt_entry_t *dde;
ddt_phys_t *ddp;
ASSERT(BP_GET_DEDUP(bp));
ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum);
ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override);
ASSERT0(zio->io_flags & ZIO_FLAG_RAW);
ddt_enter(ddt);
dde = ddt_lookup(ddt, bp, B_TRUE);
ddp = &dde->dde_phys[p];
if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) {
/*
* If we're using a weak checksum, upgrade to a strong checksum
* and try again. If we're already using a strong checksum,
* we can't resolve it, so just convert to an ordinary write.
* (And automatically e-mail a paper to Nature?)
*/
if (!zio_checksum_table[zp->zp_checksum].ci_dedup) {
zp->zp_checksum = spa_dedup_checksum(spa);
zio_pop_transforms(zio);
zio->io_stage = ZIO_STAGE_OPEN;
BP_ZERO(bp);
} else {
zp->zp_dedup = B_FALSE;
}
zio->io_pipeline = ZIO_WRITE_PIPELINE;
ddt_exit(ddt);
return (ZIO_PIPELINE_CONTINUE);
}
ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp);
ASSERT(ditto_copies < SPA_DVAS_PER_BP);
if (ditto_copies > ddt_ditto_copies_present(dde) &&
dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) {
zio_prop_t czp = *zp;
czp.zp_copies = ditto_copies;
/*
* If we arrived here with an override bp, we won't have run
* the transform stack, so we won't have the data we need to
* generate a child i/o. So, toss the override bp and restart.
* This is safe, because using the override bp is just an
* optimization; and it's rare, so the cost doesn't matter.
*/
if (zio->io_bp_override) {
zio_pop_transforms(zio);
zio->io_stage = ZIO_STAGE_OPEN;
zio->io_pipeline = ZIO_WRITE_PIPELINE;
zio->io_bp_override = NULL;
BP_ZERO(bp);
ddt_exit(ddt);
return (ZIO_PIPELINE_CONTINUE);
}
dio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL,
NULL, zio_ddt_ditto_write_done, dde, zio->io_priority,
ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
zio_push_transform(dio, zio->io_data, zio->io_size, 0, NULL);
dde->dde_lead_zio[DDT_PHYS_DITTO] = dio;
}
if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) {
if (ddp->ddp_phys_birth != 0)
ddt_bp_fill(ddp, bp, txg);
if (dde->dde_lead_zio[p] != NULL)
zio_add_child(zio, dde->dde_lead_zio[p]);
else
ddt_phys_addref(ddp);
} else if (zio->io_bp_override) {
ASSERT(bp->blk_birth == txg);
ASSERT(BP_EQUAL(bp, zio->io_bp_override));
ddt_phys_fill(ddp, bp);
ddt_phys_addref(ddp);
} else {
cio = zio_write(zio, spa, txg, bp, zio->io_orig_data,
zio->io_orig_size, zio->io_orig_size, zp,
zio_ddt_child_write_ready, NULL, NULL,
zio_ddt_child_write_done, dde, zio->io_priority,
ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark);
zio_push_transform(cio, zio->io_data, zio->io_size, 0, NULL);
dde->dde_lead_zio[p] = cio;
}
ddt_exit(ddt);
if (cio)
zio_nowait(cio);
if (dio)
zio_nowait(dio);
return (ZIO_PIPELINE_CONTINUE);
}
ddt_entry_t *freedde; /* for debugging */
static int
zio_ddt_free(zio_t *zio)
{
spa_t *spa = zio->io_spa;
blkptr_t *bp = zio->io_bp;
ddt_t *ddt = ddt_select(spa, bp);
ddt_entry_t *dde;
ddt_phys_t *ddp;
ASSERT(BP_GET_DEDUP(bp));
ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
ddt_enter(ddt);
freedde = dde = ddt_lookup(ddt, bp, B_TRUE);
if (dde) {
ddp = ddt_phys_select(dde, bp);
if (ddp)
ddt_phys_decref(ddp);
}
ddt_exit(ddt);
return (ZIO_PIPELINE_CONTINUE);
}
/*
* ==========================================================================
* Allocate and free blocks
* ==========================================================================
*/
2008-11-20 20:01:55 +00:00
static int
zio_dva_allocate(zio_t *zio)
{
spa_t *spa = zio->io_spa;
metaslab_class_t *mc = spa_normal_class(spa);
2008-11-20 20:01:55 +00:00
blkptr_t *bp = zio->io_bp;
int error;
int flags = 0;
2008-11-20 20:01:55 +00:00
2009-07-02 22:44:48 +00:00
if (zio->io_gang_leader == NULL) {
ASSERT(zio->io_child_type > ZIO_CHILD_GANG);
zio->io_gang_leader = zio;
}
2008-11-20 20:01:55 +00:00
ASSERT(BP_IS_HOLE(bp));
ASSERT0(BP_GET_NDVAS(bp));
ASSERT3U(zio->io_prop.zp_copies, >, 0);
ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa));
2008-11-20 20:01:55 +00:00
ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp));
/*
* The dump device does not support gang blocks so allocation on
* behalf of the dump device (i.e. ZIO_FLAG_NODATA) must avoid
* the "fast" gang feature.
*/
flags |= (zio->io_flags & ZIO_FLAG_NODATA) ? METASLAB_GANG_AVOID : 0;
flags |= (zio->io_flags & ZIO_FLAG_GANG_CHILD) ?
METASLAB_GANG_CHILD : 0;
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 13:20:20 +00:00
flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0;
error = metaslab_alloc(spa, mc, zio->io_size, bp,
zio->io_prop.zp_copies, zio->io_txg, NULL, flags);
2008-11-20 20:01:55 +00:00
if (error) {
spa_dbgmsg(spa, "%s: metaslab allocation failure: zio %p, "
"size %llu, error %d", spa_name(spa), zio, zio->io_size,
error);
if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE)
return (zio_write_gang_block(zio));
2008-11-20 20:01:55 +00:00
zio->io_error = error;
}
return (ZIO_PIPELINE_CONTINUE);
}
static int
zio_dva_free(zio_t *zio)
{
metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE);
2008-11-20 20:01:55 +00:00
return (ZIO_PIPELINE_CONTINUE);
}
static int
zio_dva_claim(zio_t *zio)
{
int error;
error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg);
if (error)
zio->io_error = error;
2008-11-20 20:01:55 +00:00
return (ZIO_PIPELINE_CONTINUE);
}
/*
* Undo an allocation. This is used by zio_done() when an I/O fails
* and we want to give back the block we just allocated.
* This handles both normal blocks and gang blocks.
*/
static void
zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp)
{
int g;
ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp));
ASSERT(zio->io_bp_override == NULL);
if (!BP_IS_HOLE(bp))
metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE);
if (gn != NULL) {
for (g = 0; g < SPA_GBH_NBLKPTRS; g++) {
zio_dva_unallocate(zio, gn->gn_child[g],
&gn->gn_gbh->zg_blkptr[g]);
}
}
}
/*
* Try to allocate an intent log block. Return 0 on success, errno on failure.
*/
int
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 13:20:20 +00:00
zio_alloc_zil(spa_t *spa, uint64_t txg, blkptr_t *new_bp, uint64_t size,
boolean_t use_slog)
{
int error = 1;
ASSERT(txg > spa_syncing_txg(spa));
/*
* ZIL blocks are always contiguous (i.e. not gang blocks) so we
* set the METASLAB_GANG_AVOID flag so that they don't "fast gang"
* when allocating them.
*/
if (use_slog) {
error = metaslab_alloc(spa, spa_log_class(spa), size,
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 13:20:20 +00:00
new_bp, 1, txg, NULL,
METASLAB_FASTWRITE | METASLAB_GANG_AVOID);
}
if (error) {
error = metaslab_alloc(spa, spa_normal_class(spa), size,
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 13:20:20 +00:00
new_bp, 1, txg, NULL,
METASLAB_FASTWRITE);
}
if (error == 0) {
BP_SET_LSIZE(new_bp, size);
BP_SET_PSIZE(new_bp, size);
BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF);
BP_SET_CHECKSUM(new_bp,
spa_version(spa) >= SPA_VERSION_SLIM_ZIL
? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG);
BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG);
BP_SET_LEVEL(new_bp, 0);
BP_SET_DEDUP(new_bp, 0);
BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER);
}
return (error);
}
/*
* Free an intent log block.
*/
void
zio_free_zil(spa_t *spa, uint64_t txg, blkptr_t *bp)
{
ASSERT(BP_GET_TYPE(bp) == DMU_OT_INTENT_LOG);
ASSERT(!BP_IS_GANG(bp));
zio_free(spa, txg, bp);
}
2008-11-20 20:01:55 +00:00
/*
* ==========================================================================
* Read and write to physical devices
* ==========================================================================
*/
/*
* Issue an I/O to the underlying vdev. Typically the issue pipeline
* stops after this stage and will resume upon I/O completion.
* However, there are instances where the vdev layer may need to
* continue the pipeline when an I/O was not issued. Since the I/O
* that was sent to the vdev layer might be different than the one
* currently active in the pipeline (see vdev_queue_io()), we explicitly
* force the underlying vdev layers to call either zio_execute() or
* zio_interrupt() to ensure that the pipeline continues with the correct I/O.
*/
2008-11-20 20:01:55 +00:00
static int
zio_vdev_io_start(zio_t *zio)
{
vdev_t *vd = zio->io_vd;
uint64_t align;
spa_t *spa = zio->io_spa;
Add -lhHpw options to "zpool iostat" for avg latency, histograms, & queues Update the zfs module to collect statistics on average latencies, queue sizes, and keep an internal histogram of all IO latencies. Along with this, update "zpool iostat" with some new options to print out the stats: -l: Include average IO latencies stats: total_wait disk_wait syncq_wait asyncq_wait scrub read write read write read write read write wait ----- ----- ----- ----- ----- ----- ----- ----- ----- - 41ms - 2ms - 46ms - 4ms - - 5ms - 1ms - 1us - 4ms - - 5ms - 1ms - 1us - 4ms - - - - - - - - - - - 49ms - 2ms - 47ms - - - - - - - - - - - - - 2ms - 1ms - - - 1ms - ----- ----- ----- ----- ----- ----- ----- ----- ----- 1ms 1ms 1ms 413us 16us 25us - 5ms - 1ms 1ms 1ms 413us 16us 25us - 5ms - 2ms 1ms 2ms 412us 26us 25us - 5ms - - 1ms - 413us - 25us - 5ms - - 1ms - 460us - 29us - 5ms - 196us 1ms 196us 370us 7us 23us - 5ms - ----- ----- ----- ----- ----- ----- ----- ----- ----- -w: Print out latency histograms: sdb total disk sync_queue async_queue latency read write read write read write read write scrub ------- ------ ------ ------ ------ ------ ------ ------ ------ ------ 1ns 0 0 0 0 0 0 0 0 0 ... 33us 0 0 0 0 0 0 0 0 0 66us 0 0 107 2486 2 788 12 12 0 131us 2 797 359 4499 10 558 184 184 6 262us 22 801 264 1563 10 286 287 287 24 524us 87 575 71 52086 15 1063 136 136 92 1ms 152 1190 5 41292 4 1693 252 252 141 2ms 245 2018 0 50007 0 2322 371 371 220 4ms 189 7455 22 162957 0 3912 6726 6726 199 8ms 108 9461 0 102320 0 5775 2526 2526 86 17ms 23 11287 0 37142 0 8043 1813 1813 19 34ms 0 14725 0 24015 0 11732 3071 3071 0 67ms 0 23597 0 7914 0 18113 5025 5025 0 134ms 0 33798 0 254 0 25755 7326 7326 0 268ms 0 51780 0 12 0 41593 10002 10002 0 537ms 0 77808 0 0 0 64255 13120 13120 0 1s 0 105281 0 0 0 83805 20841 20841 0 2s 0 88248 0 0 0 73772 14006 14006 0 4s 0 47266 0 0 0 29783 17176 17176 0 9s 0 10460 0 0 0 4130 6295 6295 0 17s 0 0 0 0 0 0 0 0 0 34s 0 0 0 0 0 0 0 0 0 69s 0 0 0 0 0 0 0 0 0 137s 0 0 0 0 0 0 0 0 0 ------------------------------------------------------------------------------- -h: Help -H: Scripted mode. Do not display headers, and separate fields by a single tab instead of arbitrary space. -q: Include current number of entries in sync & async read/write queues, and scrub queue: syncq_read syncq_write asyncq_read asyncq_write scrubq_read pend activ pend activ pend activ pend activ pend activ ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 0 0 0 0 78 29 0 0 0 0 0 0 0 0 78 29 0 0 0 0 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 0 0 0 0 0 0 0 0 0 0 ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 0 0 227 394 0 19 0 0 0 0 0 0 227 394 0 19 0 0 0 0 0 0 108 98 0 19 0 0 0 0 0 0 19 98 0 0 0 0 0 0 0 0 78 98 0 0 0 0 0 0 0 0 19 88 0 0 0 0 0 0 ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- -p: Display numbers in parseable (exact) values. Also, update iostat syntax to allow the user to specify specific vdevs to show statistics for. The three options for choosing pools/vdevs are: Display a list of pools: zpool iostat ... [pool ...] Display a list of vdevs from a specific pool: zpool iostat ... [pool vdev ...] Display a list of vdevs from any pools: zpool iostat ... [vdev ...] Lastly, allow zpool command "interval" value to be floating point: zpool iostat -v 0.5 Signed-off-by: Tony Hutter <hutter2@llnl.gov Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #4433
2016-02-29 18:05:23 +00:00
zio->io_delay = 0;
ASSERT(zio->io_error == 0);
ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0);
2008-11-20 20:01:55 +00:00
if (vd == NULL) {
if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
spa_config_enter(spa, SCL_ZIO, zio, RW_READER);
2008-11-20 20:01:55 +00:00
/*
* The mirror_ops handle multiple DVAs in a single BP.
*/
vdev_mirror_ops.vdev_op_io_start(zio);
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
}
/*
* We keep track of time-sensitive I/Os so that the scan thread
* can quickly react to certain workloads. In particular, we care
* about non-scrubbing, top-level reads and writes with the following
* characteristics:
* - synchronous writes of user data to non-slog devices
* - any reads of user data
* When these conditions are met, adjust the timestamp of spa_last_io
* which allows the scan thread to adjust its workload accordingly.
*/
if (!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && zio->io_bp != NULL &&
vd == vd->vdev_top && !vd->vdev_islog &&
zio->io_bookmark.zb_objset != DMU_META_OBJSET &&
zio->io_txg != spa_syncing_txg(spa)) {
uint64_t old = spa->spa_last_io;
uint64_t new = ddi_get_lbolt64();
if (old != new)
(void) atomic_cas_64(&spa->spa_last_io, old, new);
}
align = 1ULL << vd->vdev_top->vdev_ashift;
if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) &&
P2PHASE(zio->io_size, align) != 0) {
/* Transform logical writes to be a full physical block size. */
2008-11-20 20:01:55 +00:00
uint64_t asize = P2ROUNDUP(zio->io_size, align);
char *abuf = zio_buf_alloc(asize);
ASSERT(vd == vd->vdev_top);
2008-11-20 20:01:55 +00:00
if (zio->io_type == ZIO_TYPE_WRITE) {
bcopy(zio->io_data, abuf, zio->io_size);
bzero(abuf + zio->io_size, asize - zio->io_size);
}
zio_push_transform(zio, abuf, asize, asize, zio_subblock);
2008-11-20 20:01:55 +00:00
}
/*
* If this is not a physical io, make sure that it is properly aligned
* before proceeding.
*/
if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) {
ASSERT0(P2PHASE(zio->io_offset, align));
ASSERT0(P2PHASE(zio->io_size, align));
} else {
/*
* For physical writes, we allow 512b aligned writes and assume
* the device will perform a read-modify-write as necessary.
*/
ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE));
ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE));
}
VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa));
2009-01-15 21:59:39 +00:00
/*
* If this is a repair I/O, and there's no self-healing involved --
* that is, we're just resilvering what we expect to resilver --
* then don't do the I/O unless zio's txg is actually in vd's DTL.
* This prevents spurious resilvering with nested replication.
* For example, given a mirror of mirrors, (A+B)+(C+D), if only
* A is out of date, we'll read from C+D, then use the data to
* resilver A+B -- but we don't actually want to resilver B, just A.
* The top-level mirror has no way to know this, so instead we just
* discard unnecessary repairs as we work our way down the vdev tree.
* The same logic applies to any form of nested replication:
* ditto + mirror, RAID-Z + replacing, etc. This covers them all.
*/
if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) &&
!(zio->io_flags & ZIO_FLAG_SELF_HEAL) &&
zio->io_txg != 0 && /* not a delegated i/o */
!vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) {
ASSERT(zio->io_type == ZIO_TYPE_WRITE);
zio_vdev_io_bypass(zio);
return (ZIO_PIPELINE_CONTINUE);
}
2008-11-20 20:01:55 +00:00
if (vd->vdev_ops->vdev_op_leaf &&
(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) {
if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio))
2009-02-18 20:51:31 +00:00
return (ZIO_PIPELINE_CONTINUE);
if ((zio = vdev_queue_io(zio)) == NULL)
return (ZIO_PIPELINE_STOP);
if (!vdev_accessible(vd, zio)) {
zio->io_error = SET_ERROR(ENXIO);
zio_interrupt(zio);
return (ZIO_PIPELINE_STOP);
}
}
Add -lhHpw options to "zpool iostat" for avg latency, histograms, & queues Update the zfs module to collect statistics on average latencies, queue sizes, and keep an internal histogram of all IO latencies. Along with this, update "zpool iostat" with some new options to print out the stats: -l: Include average IO latencies stats: total_wait disk_wait syncq_wait asyncq_wait scrub read write read write read write read write wait ----- ----- ----- ----- ----- ----- ----- ----- ----- - 41ms - 2ms - 46ms - 4ms - - 5ms - 1ms - 1us - 4ms - - 5ms - 1ms - 1us - 4ms - - - - - - - - - - - 49ms - 2ms - 47ms - - - - - - - - - - - - - 2ms - 1ms - - - 1ms - ----- ----- ----- ----- ----- ----- ----- ----- ----- 1ms 1ms 1ms 413us 16us 25us - 5ms - 1ms 1ms 1ms 413us 16us 25us - 5ms - 2ms 1ms 2ms 412us 26us 25us - 5ms - - 1ms - 413us - 25us - 5ms - - 1ms - 460us - 29us - 5ms - 196us 1ms 196us 370us 7us 23us - 5ms - ----- ----- ----- ----- ----- ----- ----- ----- ----- -w: Print out latency histograms: sdb total disk sync_queue async_queue latency read write read write read write read write scrub ------- ------ ------ ------ ------ ------ ------ ------ ------ ------ 1ns 0 0 0 0 0 0 0 0 0 ... 33us 0 0 0 0 0 0 0 0 0 66us 0 0 107 2486 2 788 12 12 0 131us 2 797 359 4499 10 558 184 184 6 262us 22 801 264 1563 10 286 287 287 24 524us 87 575 71 52086 15 1063 136 136 92 1ms 152 1190 5 41292 4 1693 252 252 141 2ms 245 2018 0 50007 0 2322 371 371 220 4ms 189 7455 22 162957 0 3912 6726 6726 199 8ms 108 9461 0 102320 0 5775 2526 2526 86 17ms 23 11287 0 37142 0 8043 1813 1813 19 34ms 0 14725 0 24015 0 11732 3071 3071 0 67ms 0 23597 0 7914 0 18113 5025 5025 0 134ms 0 33798 0 254 0 25755 7326 7326 0 268ms 0 51780 0 12 0 41593 10002 10002 0 537ms 0 77808 0 0 0 64255 13120 13120 0 1s 0 105281 0 0 0 83805 20841 20841 0 2s 0 88248 0 0 0 73772 14006 14006 0 4s 0 47266 0 0 0 29783 17176 17176 0 9s 0 10460 0 0 0 4130 6295 6295 0 17s 0 0 0 0 0 0 0 0 0 34s 0 0 0 0 0 0 0 0 0 69s 0 0 0 0 0 0 0 0 0 137s 0 0 0 0 0 0 0 0 0 ------------------------------------------------------------------------------- -h: Help -H: Scripted mode. Do not display headers, and separate fields by a single tab instead of arbitrary space. -q: Include current number of entries in sync & async read/write queues, and scrub queue: syncq_read syncq_write asyncq_read asyncq_write scrubq_read pend activ pend activ pend activ pend activ pend activ ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 0 0 0 0 78 29 0 0 0 0 0 0 0 0 78 29 0 0 0 0 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 0 0 0 0 0 0 0 0 0 0 ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 0 0 227 394 0 19 0 0 0 0 0 0 227 394 0 19 0 0 0 0 0 0 108 98 0 19 0 0 0 0 0 0 19 98 0 0 0 0 0 0 0 0 78 98 0 0 0 0 0 0 0 0 19 88 0 0 0 0 0 0 ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- -p: Display numbers in parseable (exact) values. Also, update iostat syntax to allow the user to specify specific vdevs to show statistics for. The three options for choosing pools/vdevs are: Display a list of pools: zpool iostat ... [pool ...] Display a list of vdevs from a specific pool: zpool iostat ... [pool vdev ...] Display a list of vdevs from any pools: zpool iostat ... [vdev ...] Lastly, allow zpool command "interval" value to be floating point: zpool iostat -v 0.5 Signed-off-by: Tony Hutter <hutter2@llnl.gov Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #4433
2016-02-29 18:05:23 +00:00
zio->io_delay = gethrtime();
vd->vdev_ops->vdev_op_io_start(zio);
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
}
static int
zio_vdev_io_done(zio_t *zio)
{
vdev_t *vd = zio->io_vd;
vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops;
boolean_t unexpected_error = B_FALSE;
2008-11-20 20:01:55 +00:00
if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE);
Add -lhHpw options to "zpool iostat" for avg latency, histograms, & queues Update the zfs module to collect statistics on average latencies, queue sizes, and keep an internal histogram of all IO latencies. Along with this, update "zpool iostat" with some new options to print out the stats: -l: Include average IO latencies stats: total_wait disk_wait syncq_wait asyncq_wait scrub read write read write read write read write wait ----- ----- ----- ----- ----- ----- ----- ----- ----- - 41ms - 2ms - 46ms - 4ms - - 5ms - 1ms - 1us - 4ms - - 5ms - 1ms - 1us - 4ms - - - - - - - - - - - 49ms - 2ms - 47ms - - - - - - - - - - - - - 2ms - 1ms - - - 1ms - ----- ----- ----- ----- ----- ----- ----- ----- ----- 1ms 1ms 1ms 413us 16us 25us - 5ms - 1ms 1ms 1ms 413us 16us 25us - 5ms - 2ms 1ms 2ms 412us 26us 25us - 5ms - - 1ms - 413us - 25us - 5ms - - 1ms - 460us - 29us - 5ms - 196us 1ms 196us 370us 7us 23us - 5ms - ----- ----- ----- ----- ----- ----- ----- ----- ----- -w: Print out latency histograms: sdb total disk sync_queue async_queue latency read write read write read write read write scrub ------- ------ ------ ------ ------ ------ ------ ------ ------ ------ 1ns 0 0 0 0 0 0 0 0 0 ... 33us 0 0 0 0 0 0 0 0 0 66us 0 0 107 2486 2 788 12 12 0 131us 2 797 359 4499 10 558 184 184 6 262us 22 801 264 1563 10 286 287 287 24 524us 87 575 71 52086 15 1063 136 136 92 1ms 152 1190 5 41292 4 1693 252 252 141 2ms 245 2018 0 50007 0 2322 371 371 220 4ms 189 7455 22 162957 0 3912 6726 6726 199 8ms 108 9461 0 102320 0 5775 2526 2526 86 17ms 23 11287 0 37142 0 8043 1813 1813 19 34ms 0 14725 0 24015 0 11732 3071 3071 0 67ms 0 23597 0 7914 0 18113 5025 5025 0 134ms 0 33798 0 254 0 25755 7326 7326 0 268ms 0 51780 0 12 0 41593 10002 10002 0 537ms 0 77808 0 0 0 64255 13120 13120 0 1s 0 105281 0 0 0 83805 20841 20841 0 2s 0 88248 0 0 0 73772 14006 14006 0 4s 0 47266 0 0 0 29783 17176 17176 0 9s 0 10460 0 0 0 4130 6295 6295 0 17s 0 0 0 0 0 0 0 0 0 34s 0 0 0 0 0 0 0 0 0 69s 0 0 0 0 0 0 0 0 0 137s 0 0 0 0 0 0 0 0 0 ------------------------------------------------------------------------------- -h: Help -H: Scripted mode. Do not display headers, and separate fields by a single tab instead of arbitrary space. -q: Include current number of entries in sync & async read/write queues, and scrub queue: syncq_read syncq_write asyncq_read asyncq_write scrubq_read pend activ pend activ pend activ pend activ pend activ ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 0 0 0 0 78 29 0 0 0 0 0 0 0 0 78 29 0 0 0 0 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 0 0 0 0 0 0 0 0 0 0 ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 0 0 227 394 0 19 0 0 0 0 0 0 227 394 0 19 0 0 0 0 0 0 108 98 0 19 0 0 0 0 0 0 19 98 0 0 0 0 0 0 0 0 78 98 0 0 0 0 0 0 0 0 19 88 0 0 0 0 0 0 ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- -p: Display numbers in parseable (exact) values. Also, update iostat syntax to allow the user to specify specific vdevs to show statistics for. The three options for choosing pools/vdevs are: Display a list of pools: zpool iostat ... [pool ...] Display a list of vdevs from a specific pool: zpool iostat ... [pool vdev ...] Display a list of vdevs from any pools: zpool iostat ... [vdev ...] Lastly, allow zpool command "interval" value to be floating point: zpool iostat -v 0.5 Signed-off-by: Tony Hutter <hutter2@llnl.gov Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #4433
2016-02-29 18:05:23 +00:00
if (zio->io_delay)
zio->io_delay = gethrtime() - zio->io_delay;
if (vd != NULL && vd->vdev_ops->vdev_op_leaf) {
vdev_queue_io_done(zio);
if (zio->io_type == ZIO_TYPE_WRITE)
vdev_cache_write(zio);
if (zio_injection_enabled && zio->io_error == 0)
2009-07-02 22:44:48 +00:00
zio->io_error = zio_handle_device_injection(vd,
zio, EIO);
if (zio_injection_enabled && zio->io_error == 0)
zio->io_error = zio_handle_label_injection(zio, EIO);
if (zio->io_error) {
if (!vdev_accessible(vd, zio)) {
zio->io_error = SET_ERROR(ENXIO);
} else {
unexpected_error = B_TRUE;
}
}
}
ops->vdev_op_io_done(zio);
2008-11-20 20:01:55 +00:00
if (unexpected_error)
2009-02-18 20:51:31 +00:00
VERIFY(vdev_probe(vd, zio) == NULL);
2008-11-20 20:01:55 +00:00
return (ZIO_PIPELINE_CONTINUE);
2008-11-20 20:01:55 +00:00
}
/*
* For non-raidz ZIOs, we can just copy aside the bad data read from the
* disk, and use that to finish the checksum ereport later.
*/
static void
zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr,
const void *good_buf)
{
/* no processing needed */
zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE);
}
/*ARGSUSED*/
void
zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored)
{
void *buf = zio_buf_alloc(zio->io_size);
bcopy(zio->io_data, buf, zio->io_size);
zcr->zcr_cbinfo = zio->io_size;
zcr->zcr_cbdata = buf;
zcr->zcr_finish = zio_vsd_default_cksum_finish;
zcr->zcr_free = zio_buf_free;
}
2008-11-20 20:01:55 +00:00
static int
zio_vdev_io_assess(zio_t *zio)
{
vdev_t *vd = zio->io_vd;
if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE))
return (ZIO_PIPELINE_STOP);
if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER))
spa_config_exit(zio->io_spa, SCL_ZIO, zio);
if (zio->io_vsd != NULL) {
zio->io_vsd_ops->vsd_free(zio);
zio->io_vsd = NULL;
2008-11-20 20:01:55 +00:00
}
if (zio_injection_enabled && zio->io_error == 0)
2008-11-20 20:01:55 +00:00
zio->io_error = zio_handle_fault_injection(zio, EIO);
/*
* If the I/O failed, determine whether we should attempt to retry it.
*
* On retry, we cut in line in the issue queue, since we don't want
* compression/checksumming/etc. work to prevent our (cheap) IO reissue.
2008-11-20 20:01:55 +00:00
*/
if (zio->io_error && vd == NULL &&
!(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) {
ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */
ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */
2008-11-20 20:01:55 +00:00
zio->io_error = 0;
zio->io_flags |= ZIO_FLAG_IO_RETRY |
ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE;
zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1;
zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE,
zio_requeue_io_start_cut_in_line);
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
}
/*
* If we got an error on a leaf device, convert it to ENXIO
* if the device is not accessible at all.
*/
if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf &&
!vdev_accessible(vd, zio))
zio->io_error = SET_ERROR(ENXIO);
/*
* If we can't write to an interior vdev (mirror or RAID-Z),
* set vdev_cant_write so that we stop trying to allocate from it.
*/
if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE &&
vd != NULL && !vd->vdev_ops->vdev_op_leaf) {
vd->vdev_cant_write = B_TRUE;
}
if (zio->io_error)
zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
Illumos #4045 write throttle & i/o scheduler performance work 4045 zfs write throttle & i/o scheduler performance work 1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync read, sync write, async read, async write, and scrub/resilver. The scheduler issues a number of concurrent i/os from each class to the device. Once a class has been selected, an i/o is selected from this class using either an elevator algorithem (async, scrub classes) or FIFO (sync classes). The number of concurrent async write i/os is tuned dynamically based on i/o load, to achieve good sync i/o latency when there is not a high load of writes, and good write throughput when there is. See the block comment in vdev_queue.c (reproduced below) for more details. 2. The write throttle (dsl_pool_tempreserve_space() and txg_constrain_throughput()) is rewritten to produce much more consistent delays when under constant load. The new write throttle is based on the amount of dirty data, rather than guesses about future performance of the system. When there is a lot of dirty data, each transaction (e.g. write() syscall) will be delayed by the same small amount. This eliminates the "brick wall of wait" that the old write throttle could hit, causing all transactions to wait several seconds until the next txg opens. One of the keys to the new write throttle is decrementing the amount of dirty data as i/o completes, rather than at the end of spa_sync(). Note that the write throttle is only applied once the i/o scheduler is issuing the maximum number of outstanding async writes. See the block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for more details. This diff has several other effects, including: * the commonly-tuned global variable zfs_vdev_max_pending has been removed; use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead. * the size of each txg (meaning the amount of dirty data written, and thus the time it takes to write out) is now controlled differently. There is no longer an explicit time goal; the primary determinant is amount of dirty data. Systems that are under light or medium load will now often see that a txg is always syncing, but the impact to performance (e.g. read latency) is minimal. Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this. * zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression, checksum, etc. This improves latency by not allowing these CPU-intensive tasks to consume all CPU (on machines with at least 4 CPU's; the percentage is rounded up). --matt APPENDIX: problems with the current i/o scheduler The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem with this is that if there are always i/os pending, then certain classes of i/os can see very long delays. For example, if there are always synchronous reads outstanding, then no async writes will be serviced until they become "past due". One symptom of this situation is that each pass of the txg sync takes at least several seconds (typically 3 seconds). If many i/os become "past due" (their deadline is in the past), then we must service all of these overdue i/os before any new i/os. This happens when we enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in the future. If we can't complete all the i/os in 2.5 seconds (e.g. because there were always reads pending), then these i/os will become past due. Now we must service all the "async" writes (which could be hundreds of megabytes) before we service any reads, introducing considerable latency to synchronous i/os (reads or ZIL writes). Notes on porting to ZFS on Linux: - zio_t gained new members io_physdone and io_phys_children. Because object caches in the Linux port call the constructor only once at allocation time, objects may contain residual data when retrieved from the cache. Therefore zio_create() was updated to zero out the two new fields. - vdev_mirror_pending() relied on the depth of the per-vdev pending queue (vq->vq_pending_tree) to select the least-busy leaf vdev to read from. This tree has been replaced by vq->vq_active_tree which is now used for the same purpose. - vdev_queue_init() used the value of zfs_vdev_max_pending to determine the number of vdev I/O buffers to pre-allocate. That global no longer exists, so we instead use the sum of the *_max_active values for each of the five I/O classes described above. - The Illumos implementation of dmu_tx_delay() delays a transaction by sleeping in condition variable embedded in the thread (curthread->t_delay_cv). We do not have an equivalent CV to use in Linux, so this change replaced the delay logic with a wrapper called zfs_sleep_until(). This wrapper could be adopted upstream and in other downstream ports to abstract away operating system-specific delay logic. - These tunables are added as module parameters, and descriptions added to the zfs-module-parameters.5 man page. spa_asize_inflation zfs_deadman_synctime_ms zfs_vdev_max_active zfs_vdev_async_write_active_min_dirty_percent zfs_vdev_async_write_active_max_dirty_percent zfs_vdev_async_read_max_active zfs_vdev_async_read_min_active zfs_vdev_async_write_max_active zfs_vdev_async_write_min_active zfs_vdev_scrub_max_active zfs_vdev_scrub_min_active zfs_vdev_sync_read_max_active zfs_vdev_sync_read_min_active zfs_vdev_sync_write_max_active zfs_vdev_sync_write_min_active zfs_dirty_data_max_percent zfs_delay_min_dirty_percent zfs_dirty_data_max_max_percent zfs_dirty_data_max zfs_dirty_data_max_max zfs_dirty_data_sync zfs_delay_scale The latter four have type unsigned long, whereas they are uint64_t in Illumos. This accommodates Linux's module_param() supported types, but means they may overflow on 32-bit architectures. The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most likely to overflow on 32-bit systems, since they express physical RAM sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to 2^32 which does overflow. To resolve that, this port instead initializes it in arc_init() to 25% of physical RAM, and adds the tunable zfs_dirty_data_max_max_percent to override that percentage. While this solution doesn't completely avoid the overflow issue, it should be a reasonable default for most systems, and the minority of affected systems can work around the issue by overriding the defaults. - Fixed reversed logic in comment above zfs_delay_scale declaration. - Clarified comments in vdev_queue.c regarding when per-queue minimums take effect. - Replaced dmu_tx_write_limit in the dmu_tx kstat file with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts how many times a transaction has been delayed because the pool dirty data has exceeded zfs_delay_min_dirty_percent. The latter counts how many times the pool dirty data has exceeded zfs_dirty_data_max (which we expect to never happen). - The original patch would have regressed the bug fixed in zfsonlinux/zfs@c418410, which prevented users from setting the zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE. A similar fix is added to vdev_queue_aggregate(). - In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the heap instead of the stack. In Linux we can't afford such large structures on the stack. Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Adam Leventhal <ahl@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Ned Bass <bass6@llnl.gov> Reviewed by: Brendan Gregg <brendan.gregg@joyent.com> Approved by: Robert Mustacchi <rm@joyent.com> References: http://www.illumos.org/issues/4045 illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e Ported-by: Ned Bass <bass6@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #1913
2013-08-29 03:01:20 +00:00
if (vd != NULL && vd->vdev_ops->vdev_op_leaf &&
zio->io_physdone != NULL) {
ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED));
ASSERT(zio->io_child_type == ZIO_CHILD_VDEV);
zio->io_physdone(zio->io_logical);
}
2008-11-20 20:01:55 +00:00
return (ZIO_PIPELINE_CONTINUE);
}
void
zio_vdev_io_reissue(zio_t *zio)
{
ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
ASSERT(zio->io_error == 0);
zio->io_stage >>= 1;
2008-11-20 20:01:55 +00:00
}
void
zio_vdev_io_redone(zio_t *zio)
{
ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE);
zio->io_stage >>= 1;
2008-11-20 20:01:55 +00:00
}
void
zio_vdev_io_bypass(zio_t *zio)
{
ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START);
ASSERT(zio->io_error == 0);
zio->io_flags |= ZIO_FLAG_IO_BYPASS;
zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1;
2008-11-20 20:01:55 +00:00
}
/*
* ==========================================================================
* Generate and verify checksums
* ==========================================================================
*/
static int
zio_checksum_generate(zio_t *zio)
{
blkptr_t *bp = zio->io_bp;
enum zio_checksum checksum;
2008-11-20 20:01:55 +00:00
if (bp == NULL) {
/*
* This is zio_write_phys().
* We're either generating a label checksum, or none at all.
*/
checksum = zio->io_prop.zp_checksum;
2008-11-20 20:01:55 +00:00
if (checksum == ZIO_CHECKSUM_OFF)
return (ZIO_PIPELINE_CONTINUE);
ASSERT(checksum == ZIO_CHECKSUM_LABEL);
} else {
if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) {
ASSERT(!IO_IS_ALLOCATING(zio));
checksum = ZIO_CHECKSUM_GANG_HEADER;
} else {
checksum = BP_GET_CHECKSUM(bp);
}
}
2008-11-20 20:01:55 +00:00
zio_checksum_compute(zio, checksum, zio->io_data, zio->io_size);
2008-11-20 20:01:55 +00:00
return (ZIO_PIPELINE_CONTINUE);
}
static int
zio_checksum_verify(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
zio_bad_cksum_t info;
blkptr_t *bp = zio->io_bp;
int error;
2008-11-20 20:01:55 +00:00
ASSERT(zio->io_vd != NULL);
if (bp == NULL) {
/*
* This is zio_read_phys().
* We're either verifying a label checksum, or nothing at all.
*/
if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF)
return (ZIO_PIPELINE_CONTINUE);
2008-11-20 20:01:55 +00:00
ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL);
}
2008-11-20 20:01:55 +00:00
if ((error = zio_checksum_error(zio, &info)) != 0) {
zio->io_error = error;
if (error == ECKSUM &&
!(zio->io_flags & ZIO_FLAG_SPECULATIVE)) {
zfs_ereport_start_checksum(zio->io_spa,
zio->io_vd, zio, zio->io_offset,
zio->io_size, NULL, &info);
}
2008-11-20 20:01:55 +00:00
}
return (ZIO_PIPELINE_CONTINUE);
}
/*
* Called by RAID-Z to ensure we don't compute the checksum twice.
*/
void
zio_checksum_verified(zio_t *zio)
{
zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY;
2008-11-20 20:01:55 +00:00
}
/*
* ==========================================================================
* Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other.
* An error of 0 indicates success. ENXIO indicates whole-device failure,
* which may be transient (e.g. unplugged) or permament. ECKSUM and EIO
* indicate errors that are specific to one I/O, and most likely permanent.
* Any other error is presumed to be worse because we weren't expecting it.
* ==========================================================================
2008-11-20 20:01:55 +00:00
*/
int
zio_worst_error(int e1, int e2)
2008-11-20 20:01:55 +00:00
{
static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO };
int r1, r2;
for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++)
if (e1 == zio_error_rank[r1])
break;
2008-11-20 20:01:55 +00:00
for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++)
if (e2 == zio_error_rank[r2])
break;
return (r1 > r2 ? e1 : e2);
2008-11-20 20:01:55 +00:00
}
/*
* ==========================================================================
* I/O completion
2008-11-20 20:01:55 +00:00
* ==========================================================================
*/
static int
zio_ready(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
blkptr_t *bp = zio->io_bp;
2009-02-18 20:51:31 +00:00
zio_t *pio, *pio_next;
2008-11-20 20:01:55 +00:00
if (zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_READY) ||
zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_READY))
2009-07-02 22:44:48 +00:00
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
2009-07-02 22:44:48 +00:00
if (zio->io_ready) {
ASSERT(IO_IS_ALLOCATING(zio));
ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) ||
(zio->io_flags & ZIO_FLAG_NOPWRITE));
ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0);
2008-11-20 20:01:55 +00:00
zio->io_ready(zio);
}
2008-11-20 20:01:55 +00:00
if (bp != NULL && bp != &zio->io_bp_copy)
zio->io_bp_copy = *bp;
2008-11-20 20:01:55 +00:00
if (zio->io_error)
zio->io_pipeline = ZIO_INTERLOCK_PIPELINE;
2008-11-20 20:01:55 +00:00
2009-02-18 20:51:31 +00:00
mutex_enter(&zio->io_lock);
zio->io_state[ZIO_WAIT_READY] = 1;
pio = zio_walk_parents(zio);
mutex_exit(&zio->io_lock);
/*
* As we notify zio's parents, new parents could be added.
* New parents go to the head of zio's io_parent_list, however,
* so we will (correctly) not notify them. The remainder of zio's
* io_parent_list, from 'pio_next' onward, cannot change because
* all parents must wait for us to be done before they can be done.
*/
for (; pio != NULL; pio = pio_next) {
pio_next = zio_walk_parents(zio);
zio_notify_parent(pio, zio, ZIO_WAIT_READY);
2009-02-18 20:51:31 +00:00
}
2008-11-20 20:01:55 +00:00
if (zio->io_flags & ZIO_FLAG_NODATA) {
if (BP_IS_GANG(bp)) {
zio->io_flags &= ~ZIO_FLAG_NODATA;
} else {
ASSERT((uintptr_t)zio->io_data < SPA_MAXBLOCKSIZE);
zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES;
}
}
if (zio_injection_enabled &&
zio->io_spa->spa_syncing_txg == zio->io_txg)
zio_handle_ignored_writes(zio);
return (ZIO_PIPELINE_CONTINUE);
2008-11-20 20:01:55 +00:00
}
static int
zio_done(zio_t *zio)
2008-11-20 20:01:55 +00:00
{
2009-02-18 20:51:31 +00:00
zio_t *pio, *pio_next;
int c, w;
2008-11-20 20:01:55 +00:00
/*
2009-07-02 22:44:48 +00:00
* If our children haven't all completed,
* wait for them and then repeat this pipeline stage.
*/
if (zio_wait_for_children(zio, ZIO_CHILD_VDEV, ZIO_WAIT_DONE) ||
zio_wait_for_children(zio, ZIO_CHILD_GANG, ZIO_WAIT_DONE) ||
zio_wait_for_children(zio, ZIO_CHILD_DDT, ZIO_WAIT_DONE) ||
zio_wait_for_children(zio, ZIO_CHILD_LOGICAL, ZIO_WAIT_DONE))
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
for (c = 0; c < ZIO_CHILD_TYPES; c++)
for (w = 0; w < ZIO_WAIT_TYPES; w++)
ASSERT(zio->io_children[c][w] == 0);
if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) {
ASSERT(zio->io_bp->blk_pad[0] == 0);
ASSERT(zio->io_bp->blk_pad[1] == 0);
ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy,
sizeof (blkptr_t)) == 0 ||
(zio->io_bp == zio_unique_parent(zio)->io_bp));
if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) &&
zio->io_bp_override == NULL &&
!(zio->io_flags & ZIO_FLAG_IO_REPAIR)) {
ASSERT(!BP_SHOULD_BYTESWAP(zio->io_bp));
ASSERT3U(zio->io_prop.zp_copies, <=,
BP_GET_NDVAS(zio->io_bp));
ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 ||
(BP_COUNT_GANG(zio->io_bp) ==
BP_GET_NDVAS(zio->io_bp)));
}
if (zio->io_flags & ZIO_FLAG_NOPWRITE)
VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig));
}
/*
* If there were child vdev/gang/ddt errors, they apply to us now.
*/
zio_inherit_child_errors(zio, ZIO_CHILD_VDEV);
zio_inherit_child_errors(zio, ZIO_CHILD_GANG);
zio_inherit_child_errors(zio, ZIO_CHILD_DDT);
/*
* If the I/O on the transformed data was successful, generate any
* checksum reports now while we still have the transformed data.
*/
if (zio->io_error == 0) {
while (zio->io_cksum_report != NULL) {
zio_cksum_report_t *zcr = zio->io_cksum_report;
uint64_t align = zcr->zcr_align;
uint64_t asize = P2ROUNDUP(zio->io_size, align);
char *abuf = zio->io_data;
if (asize != zio->io_size) {
abuf = zio_buf_alloc(asize);
bcopy(zio->io_data, abuf, zio->io_size);
bzero(abuf+zio->io_size, asize-zio->io_size);
}
zio->io_cksum_report = zcr->zcr_next;
zcr->zcr_next = NULL;
zcr->zcr_finish(zcr, abuf);
zfs_ereport_free_checksum(zcr);
if (asize != zio->io_size)
zio_buf_free(abuf, asize);
}
}
zio_pop_transforms(zio); /* note: may set zio->io_error */
vdev_stat_update(zio, zio->io_size);
/*
* If this I/O is attached to a particular vdev is slow, exceeding
* 30 seconds to complete, post an error described the I/O delay.
* We ignore these errors if the device is currently unavailable.
*/
Add -lhHpw options to "zpool iostat" for avg latency, histograms, & queues Update the zfs module to collect statistics on average latencies, queue sizes, and keep an internal histogram of all IO latencies. Along with this, update "zpool iostat" with some new options to print out the stats: -l: Include average IO latencies stats: total_wait disk_wait syncq_wait asyncq_wait scrub read write read write read write read write wait ----- ----- ----- ----- ----- ----- ----- ----- ----- - 41ms - 2ms - 46ms - 4ms - - 5ms - 1ms - 1us - 4ms - - 5ms - 1ms - 1us - 4ms - - - - - - - - - - - 49ms - 2ms - 47ms - - - - - - - - - - - - - 2ms - 1ms - - - 1ms - ----- ----- ----- ----- ----- ----- ----- ----- ----- 1ms 1ms 1ms 413us 16us 25us - 5ms - 1ms 1ms 1ms 413us 16us 25us - 5ms - 2ms 1ms 2ms 412us 26us 25us - 5ms - - 1ms - 413us - 25us - 5ms - - 1ms - 460us - 29us - 5ms - 196us 1ms 196us 370us 7us 23us - 5ms - ----- ----- ----- ----- ----- ----- ----- ----- ----- -w: Print out latency histograms: sdb total disk sync_queue async_queue latency read write read write read write read write scrub ------- ------ ------ ------ ------ ------ ------ ------ ------ ------ 1ns 0 0 0 0 0 0 0 0 0 ... 33us 0 0 0 0 0 0 0 0 0 66us 0 0 107 2486 2 788 12 12 0 131us 2 797 359 4499 10 558 184 184 6 262us 22 801 264 1563 10 286 287 287 24 524us 87 575 71 52086 15 1063 136 136 92 1ms 152 1190 5 41292 4 1693 252 252 141 2ms 245 2018 0 50007 0 2322 371 371 220 4ms 189 7455 22 162957 0 3912 6726 6726 199 8ms 108 9461 0 102320 0 5775 2526 2526 86 17ms 23 11287 0 37142 0 8043 1813 1813 19 34ms 0 14725 0 24015 0 11732 3071 3071 0 67ms 0 23597 0 7914 0 18113 5025 5025 0 134ms 0 33798 0 254 0 25755 7326 7326 0 268ms 0 51780 0 12 0 41593 10002 10002 0 537ms 0 77808 0 0 0 64255 13120 13120 0 1s 0 105281 0 0 0 83805 20841 20841 0 2s 0 88248 0 0 0 73772 14006 14006 0 4s 0 47266 0 0 0 29783 17176 17176 0 9s 0 10460 0 0 0 4130 6295 6295 0 17s 0 0 0 0 0 0 0 0 0 34s 0 0 0 0 0 0 0 0 0 69s 0 0 0 0 0 0 0 0 0 137s 0 0 0 0 0 0 0 0 0 ------------------------------------------------------------------------------- -h: Help -H: Scripted mode. Do not display headers, and separate fields by a single tab instead of arbitrary space. -q: Include current number of entries in sync & async read/write queues, and scrub queue: syncq_read syncq_write asyncq_read asyncq_write scrubq_read pend activ pend activ pend activ pend activ pend activ ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 0 0 0 0 78 29 0 0 0 0 0 0 0 0 78 29 0 0 0 0 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 0 0 0 0 0 0 0 0 0 0 - - - - - - - - - - 0 0 0 0 0 0 0 0 0 0 ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- 0 0 227 394 0 19 0 0 0 0 0 0 227 394 0 19 0 0 0 0 0 0 108 98 0 19 0 0 0 0 0 0 19 98 0 0 0 0 0 0 0 0 78 98 0 0 0 0 0 0 0 0 19 88 0 0 0 0 0 0 ----- ----- ----- ----- ----- ----- ----- ----- ----- ----- -p: Display numbers in parseable (exact) values. Also, update iostat syntax to allow the user to specify specific vdevs to show statistics for. The three options for choosing pools/vdevs are: Display a list of pools: zpool iostat ... [pool ...] Display a list of vdevs from a specific pool: zpool iostat ... [pool vdev ...] Display a list of vdevs from any pools: zpool iostat ... [vdev ...] Lastly, allow zpool command "interval" value to be floating point: zpool iostat -v 0.5 Signed-off-by: Tony Hutter <hutter2@llnl.gov Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Closes #4433
2016-02-29 18:05:23 +00:00
if (zio->io_delay >= MSEC2NSEC(zio_delay_max)) {
if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd))
zfs_ereport_post(FM_EREPORT_ZFS_DELAY, zio->io_spa,
zio->io_vd, zio, 0, 0);
}
if (zio->io_error) {
/*
* If this I/O is attached to a particular vdev,
* generate an error message describing the I/O failure
* at the block level. We ignore these errors if the
* device is currently unavailable.
*/
if (zio->io_error != ECKSUM && zio->io_vd != NULL &&
!vdev_is_dead(zio->io_vd))
zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa,
zio->io_vd, zio, 0, 0);
2008-11-20 20:01:55 +00:00
if ((zio->io_error == EIO || !(zio->io_flags &
(ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) &&
zio == zio->io_logical) {
/*
* For logical I/O requests, tell the SPA to log the
* error and generate a logical data ereport.
*/
spa_log_error(zio->io_spa, zio);
zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa,
NULL, zio, 0, 0);
}
}
2008-11-20 20:01:55 +00:00
if (zio->io_error && zio == zio->io_logical) {
/*
* Determine whether zio should be reexecuted. This will
* propagate all the way to the root via zio_notify_parent().
*/
ASSERT(zio->io_vd == NULL && zio->io_bp != NULL);
ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
if (IO_IS_ALLOCATING(zio) &&
!(zio->io_flags & ZIO_FLAG_CANFAIL)) {
if (zio->io_error != ENOSPC)
zio->io_reexecute |= ZIO_REEXECUTE_NOW;
else
zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
}
if ((zio->io_type == ZIO_TYPE_READ ||
zio->io_type == ZIO_TYPE_FREE) &&
!(zio->io_flags & ZIO_FLAG_SCAN_THREAD) &&
zio->io_error == ENXIO &&
spa_load_state(zio->io_spa) == SPA_LOAD_NONE &&
spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE)
zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute)
zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND;
/*
* Here is a possibly good place to attempt to do
* either combinatorial reconstruction or error correction
* based on checksums. It also might be a good place
* to send out preliminary ereports before we suspend
* processing.
*/
2008-11-20 20:01:55 +00:00
}
/*
* If there were logical child errors, they apply to us now.
* We defer this until now to avoid conflating logical child
* errors with errors that happened to the zio itself when
* updating vdev stats and reporting FMA events above.
2008-11-20 20:01:55 +00:00
*/
zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL);
2008-11-20 20:01:55 +00:00
if ((zio->io_error || zio->io_reexecute) &&
IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio &&
!(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)))
zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp);
2009-07-02 22:44:48 +00:00
zio_gang_tree_free(&zio->io_gang_tree);
/*
* Godfather I/Os should never suspend.
*/
if ((zio->io_flags & ZIO_FLAG_GODFATHER) &&
(zio->io_reexecute & ZIO_REEXECUTE_SUSPEND))
zio->io_reexecute = 0;
if (zio->io_reexecute) {
/*
* This is a logical I/O that wants to reexecute.
*
* Reexecute is top-down. When an i/o fails, if it's not
* the root, it simply notifies its parent and sticks around.
* The parent, seeing that it still has children in zio_done(),
* does the same. This percolates all the way up to the root.
* The root i/o will reexecute or suspend the entire tree.
*
* This approach ensures that zio_reexecute() honors
* all the original i/o dependency relationships, e.g.
* parents not executing until children are ready.
*/
ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL);
2008-11-20 20:01:55 +00:00
2009-07-02 22:44:48 +00:00
zio->io_gang_leader = NULL;
2009-02-18 20:51:31 +00:00
mutex_enter(&zio->io_lock);
zio->io_state[ZIO_WAIT_DONE] = 1;
mutex_exit(&zio->io_lock);
2009-07-02 22:44:48 +00:00
/*
* "The Godfather" I/O monitors its children but is
* not a true parent to them. It will track them through
* the pipeline but severs its ties whenever they get into
* trouble (e.g. suspended). This allows "The Godfather"
* I/O to return status without blocking.
*/
for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
zio_link_t *zl = zio->io_walk_link;
pio_next = zio_walk_parents(zio);
if ((pio->io_flags & ZIO_FLAG_GODFATHER) &&
(zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) {
zio_remove_child(pio, zio, zl);
zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
}
}
2009-02-18 20:51:31 +00:00
if ((pio = zio_unique_parent(zio)) != NULL) {
/*
* We're not a root i/o, so there's nothing to do
* but notify our parent. Don't propagate errors
* upward since we haven't permanently failed yet.
*/
2009-07-02 22:44:48 +00:00
ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER));
zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE;
zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
} else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) {
/*
* We'd fail again if we reexecuted now, so suspend
* until conditions improve (e.g. device comes online).
*/
zio_suspend(zio->io_spa, zio);
} else {
/*
* Reexecution is potentially a huge amount of work.
* Hand it off to the otherwise-unused claim taskq.
*/
ASSERT(taskq_empty_ent(&zio->io_tqent));
Illumos #3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock contention 3581 spa_zio_taskq[ZIO_TYPE_FREE][ZIO_TASKQ_ISSUE]->tq_lock is piping hot Reviewed by: Matthew Ahrens <mahrens@delphix.com> Reviewed by: George Wilson <george.wilson@delphix.com> Reviewed by: Christopher Siden <christopher.siden@delphix.com> Reviewed by: Gordon Ross <gordon.ross@nexenta.com> Approved by: Richard Lowe <richlowe@richlowe.net> References: illumos/illumos-gate@ec94d32 https://illumos.org/issues/3581 Notes for Linux port: Earlier commit 08d08eb reduced contention on this taskq lock by simply reducing the number of z_fr_iss threads from 100 to one-per-CPU. We also optimized the taskq implementation in zfsonlinux/spl@3c6ed54. These changes significantly improved unlink performance to acceptable levels. This patch further reduces time spent spinning on this lock by randomly dispatching the work items over multiple independent task queues. The Illumos ZFS developers stated that this lock contention only arose after "3329 spa_sync() spends 10-20% of its time in spa_free_sync_cb()" was landed. It's not clear if 3329 affects the Linux port or not. I didn't see spa_free_sync_cb() show up in oprofile sessions while unlinking large files, but I may just not have used the right test case. I tested unlinking a 1 TB of data with and without the patch and didn't observe a meaningful difference in elapsed time. However, oprofile showed that the percent time spent in taskq_thread() was reduced from about 16% to about 5%. Aside from a possible slight performance benefit this may be worth landing if only for the sake of maintaining consistency with upstream. Ported-by: Ned Bass <bass6@llnl.gov> Closes #1327
2013-05-06 19:24:30 +00:00
spa_taskq_dispatch_ent(zio->io_spa,
ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE,
(task_func_t *)zio_reexecute, zio, 0,
&zio->io_tqent);
}
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
}
ASSERT(zio->io_child_count == 0);
ASSERT(zio->io_reexecute == 0);
ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL));
2008-11-20 20:01:55 +00:00
/*
* Report any checksum errors, since the I/O is complete.
*/
while (zio->io_cksum_report != NULL) {
zio_cksum_report_t *zcr = zio->io_cksum_report;
zio->io_cksum_report = zcr->zcr_next;
zcr->zcr_next = NULL;
zcr->zcr_finish(zcr, NULL);
zfs_ereport_free_checksum(zcr);
}
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 13:20:20 +00:00
if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp &&
!BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) &&
!(zio->io_flags & ZIO_FLAG_NOPWRITE)) {
Add FASTWRITE algorithm for synchronous writes. Currently, ZIL blocks are spread over vdevs using hint block pointers managed by the ZIL commit code and passed to metaslab_alloc(). Spreading log blocks accross vdevs is important for performance: indeed, using mutliple disks in parallel decreases the ZIL commit latency, which is the main performance metric for synchronous writes. However, the current implementation suffers from the following issues: 1) It would be best if the ZIL module was not aware of such low-level details. They should be handled by the ZIO and metaslab modules; 2) Because the hint block pointer is managed per log, simultaneous commits from multiple logs might use the same vdevs at the same time, which is inefficient; 3) Because dmu_write() does not honor the block pointer hint, indirect writes are not spread. The naive solution of rotating the metaslab rotor each time a block is allocated for the ZIL or dmu_sync() doesn't work in practice because the first ZIL block to be written is actually allocated during the previous commit. Consequently, when metaslab_alloc() decides the vdev for this block, it will do so while a bunch of other allocations are happening at the same time (from dmu_sync() and other ZILs). This means the vdev for this block is chosen more or less at random. When the next commit happens, there is a high chance (especially when the number of blocks per commit is slightly less than the number of the disks) that one disk will have to write two blocks (with a potential seek) while other disks are sitting idle, which defeats spreading and increases the commit latency. This commit introduces a new concept in the metaslab allocator: fastwrites. Basically, each top-level vdev maintains a counter indicating the number of synchronous writes (from dmu_sync() and the ZIL) which have been allocated but not yet completed. When the metaslab is called with the FASTWRITE flag, it will choose the vdev with the least amount of pending synchronous writes. If there are multiple vdevs with the same value, the first matching vdev (starting from the rotor) is used. Once metaslab_alloc() has decided which vdev the block is allocated to, it updates the fastwrite counter for this vdev. The rationale goes like this: when an allocation is done with FASTWRITE, it "reserves" the vdev until the data is written. Until then, all future allocations will naturally avoid this vdev, even after a full rotation of the rotor. As a result, pending synchronous writes at a given point in time will be nicely spread over all vdevs. This contrasts with the previous algorithm, which is based on the implicit assumption that blocks are written instantaneously after they're allocated. metaslab_fastwrite_mark() and metaslab_fastwrite_unmark() are used to manually increase or decrease fastwrite counters, respectively. They should be used with caution, as there is no per-BP tracking of fastwrite information, so leaks and "double-unmarks" are possible. There is, however, an assert in the vdev teardown code which will fire if the fastwrite counters are not zero when the pool is exported or the vdev removed. Note that as stated above, marking is also done implictly by metaslab_alloc(). ZIO also got a new FASTWRITE flag; when it is used, ZIO will pass it to the metaslab when allocating (assuming ZIO does the allocation, which is only true in the case of dmu_sync). This flag will also trigger an unmark when zio_done() fires. A side-effect of the new algorithm is that when a ZIL stops being used, its last block can stay in the pending state (allocated but not yet written) for a long time, polluting the fastwrite counters. To avoid that, I've implemented a somewhat crude but working solution which unmarks these pending blocks in zil_sync(), thus guaranteeing that linguering fastwrites will get pruned at each sync event. The best performance improvements are observed with pools using a large number of top-level vdevs and heavy synchronous write workflows (especially indirect writes and concurrent writes from multiple ZILs). Real-life testing shows a 200% to 300% performance increase with indirect writes and various commit sizes. Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Issue #1013
2012-06-27 13:20:20 +00:00
metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp);
}
2009-02-18 20:51:31 +00:00
/*
* It is the responsibility of the done callback to ensure that this
* particular zio is no longer discoverable for adoption, and as
* such, cannot acquire any new parents.
*/
if (zio->io_done)
zio->io_done(zio);
2008-11-20 20:01:55 +00:00
2009-02-18 20:51:31 +00:00
mutex_enter(&zio->io_lock);
zio->io_state[ZIO_WAIT_DONE] = 1;
mutex_exit(&zio->io_lock);
2008-11-20 20:01:55 +00:00
2009-02-18 20:51:31 +00:00
for (pio = zio_walk_parents(zio); pio != NULL; pio = pio_next) {
zio_link_t *zl = zio->io_walk_link;
pio_next = zio_walk_parents(zio);
zio_remove_child(pio, zio, zl);
zio_notify_parent(pio, zio, ZIO_WAIT_DONE);
}
2008-11-20 20:01:55 +00:00
if (zio->io_waiter != NULL) {
mutex_enter(&zio->io_lock);
zio->io_executor = NULL;
cv_broadcast(&zio->io_cv);
mutex_exit(&zio->io_lock);
} else {
zio_destroy(zio);
}
2008-11-20 20:01:55 +00:00
return (ZIO_PIPELINE_STOP);
2008-11-20 20:01:55 +00:00
}
/*
* ==========================================================================
* I/O pipeline definition
* ==========================================================================
2008-11-20 20:01:55 +00:00
*/
static zio_pipe_stage_t *zio_pipeline[] = {
NULL,
zio_read_bp_init,
zio_free_bp_init,
zio_issue_async,
zio_write_bp_init,
zio_checksum_generate,
zio_nop_write,
zio_ddt_read_start,
zio_ddt_read_done,
zio_ddt_write,
zio_ddt_free,
zio_gang_assemble,
zio_gang_issue,
zio_dva_allocate,
zio_dva_free,
zio_dva_claim,
zio_ready,
zio_vdev_io_start,
zio_vdev_io_done,
zio_vdev_io_assess,
zio_checksum_verify,
zio_done
};
/*
* Compare two zbookmark_phys_t's to see which we would reach first in a
* pre-order traversal of the object tree.
*
* This is simple in every case aside from the meta-dnode object. For all other
* objects, we traverse them in order (object 1 before object 2, and so on).
* However, all of these objects are traversed while traversing object 0, since
* the data it points to is the list of objects. Thus, we need to convert to a
* canonical representation so we can compare meta-dnode bookmarks to
* non-meta-dnode bookmarks.
*
* We do this by calculating "equivalents" for each field of the zbookmark.
* zbookmarks outside of the meta-dnode use their own object and level, and
* calculate the level 0 equivalent (the first L0 blkid that is contained in the
* blocks this bookmark refers to) by multiplying their blkid by their span
* (the number of L0 blocks contained within one block at their level).
* zbookmarks inside the meta-dnode calculate their object equivalent
* (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use
* level + 1<<31 (any value larger than a level could ever be) for their level.
* This causes them to always compare before a bookmark in their object
* equivalent, compare appropriately to bookmarks in other objects, and to
* compare appropriately to other bookmarks in the meta-dnode.
*/
int
zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2,
const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2)
{
/*
* These variables represent the "equivalent" values for the zbookmark,
* after converting zbookmarks inside the meta dnode to their
* normal-object equivalents.
*/
uint64_t zb1obj, zb2obj;
uint64_t zb1L0, zb2L0;
uint64_t zb1level, zb2level;
if (zb1->zb_object == zb2->zb_object &&
zb1->zb_level == zb2->zb_level &&
zb1->zb_blkid == zb2->zb_blkid)
return (0);
/*
* BP_SPANB calculates the span in blocks.
*/
zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level);
zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level);
if (zb1->zb_object == DMU_META_DNODE_OBJECT) {
zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
zb1L0 = 0;
zb1level = zb1->zb_level + COMPARE_META_LEVEL;
} else {
zb1obj = zb1->zb_object;
zb1level = zb1->zb_level;
}
if (zb2->zb_object == DMU_META_DNODE_OBJECT) {
zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT));
zb2L0 = 0;
zb2level = zb2->zb_level + COMPARE_META_LEVEL;
} else {
zb2obj = zb2->zb_object;
zb2level = zb2->zb_level;
}
/* Now that we have a canonical representation, do the comparison. */
if (zb1obj != zb2obj)
return (zb1obj < zb2obj ? -1 : 1);
else if (zb1L0 != zb2L0)
return (zb1L0 < zb2L0 ? -1 : 1);
else if (zb1level != zb2level)
return (zb1level > zb2level ? -1 : 1);
/*
* This can (theoretically) happen if the bookmarks have the same object
* and level, but different blkids, if the block sizes are not the same.
* There is presently no way to change the indirect block sizes
*/
return (0);
}
/*
* This function checks the following: given that last_block is the place that
* our traversal stopped last time, does that guarantee that we've visited
* every node under subtree_root? Therefore, we can't just use the raw output
* of zbookmark_compare. We have to pass in a modified version of
* subtree_root; by incrementing the block id, and then checking whether
* last_block is before or equal to that, we can tell whether or not having
* visited last_block implies that all of subtree_root's children have been
* visited.
*/
boolean_t
zbookmark_subtree_completed(const dnode_phys_t *dnp,
const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block)
{
zbookmark_phys_t mod_zb = *subtree_root;
mod_zb.zb_blkid++;
ASSERT(last_block->zb_level == 0);
/* The objset_phys_t isn't before anything. */
if (dnp == NULL)
return (B_FALSE);
/*
* We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the
* data block size in sectors, because that variable is only used if
* the bookmark refers to a block in the meta-dnode. Since we don't
* know without examining it what object it refers to, and there's no
* harm in passing in this value in other cases, we always pass it in.
*
* We pass in 0 for the indirect block size shift because zb2 must be
* level 0. The indirect block size is only used to calculate the span
* of the bookmark, but since the bookmark must be level 0, the span is
* always 1, so the math works out.
*
* If you make changes to how the zbookmark_compare code works, be sure
* to make sure that this code still works afterwards.
*/
return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift,
1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb,
last_block) <= 0);
}
#if defined(_KERNEL) && defined(HAVE_SPL)
EXPORT_SYMBOL(zio_type_name);
EXPORT_SYMBOL(zio_buf_alloc);
EXPORT_SYMBOL(zio_data_buf_alloc);
Fix vdev_queue_aggregate() deadlock This deadlock may manifest itself in slightly different ways but at the core it is caused by a memory allocation blocking on file- system reclaim in the zio pipeline. This is normally impossible because zio_execute() disables filesystem reclaim by setting PF_FSTRANS on the thread. However, kmem cache allocations may still indirectly block on file system reclaim while holding the critical vq->vq_lock as shown below. To resolve this issue zio_buf_alloc_flags() is introduced which allocation flags to be passed. This can then be used in vdev_queue_aggregate() with KM_NOSLEEP when allocating the aggregate IO buffer. Since aggregating the IO is purely a performance optimization we want this to either succeed or fail quickly. Trying too hard to allocate this memory under the vq->vq_lock can negatively impact performance and result in this deadlock. * z_wr_iss zio_vdev_io_start vdev_queue_io -> Takes vq->vq_lock vdev_queue_io_to_issue vdev_queue_aggregate zio_buf_alloc -> Waiting on spl_kmem_cache process * z_wr_int zio_vdev_io_done vdev_queue_io_done mutex_lock -> Waiting on vq->vq_lock held by z_wr_iss * txg_sync spa_sync dsl_pool_sync zio_wait -> Waiting on zio being handled by z_wr_int * spl_kmem_cache spl_cache_grow_work kv_alloc spl_vmalloc ... evict zpl_evict_inode zfs_inactive dmu_tx_wait txg_wait_open -> Waiting on txg_sync Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Chunwei Chen <david.chen@osnexus.com> Signed-off-by: Tim Chase <tim@chase2k.com> Closes #3808 Closes #3867
2015-12-16 19:28:15 +00:00
EXPORT_SYMBOL(zio_buf_alloc_flags);
EXPORT_SYMBOL(zio_buf_free);
EXPORT_SYMBOL(zio_data_buf_free);
module_param(zio_delay_max, int, 0644);
Add missing ZFS tunables This commit adds module options for all existing zfs tunables. Ideally the average user should never need to modify any of these values. However, in practice sometimes you do need to tweak these values for one reason or another. In those cases it's nice not to have to resort to rebuilding from source. All tunables are visable to modinfo and the list is as follows: $ modinfo module/zfs/zfs.ko filename: module/zfs/zfs.ko license: CDDL author: Sun Microsystems/Oracle, Lawrence Livermore National Laboratory description: ZFS srcversion: 8EAB1D71DACE05B5AA61567 depends: spl,znvpair,zcommon,zunicode,zavl vermagic: 2.6.32-131.0.5.el6.x86_64 SMP mod_unload modversions parm: zvol_major:Major number for zvol device (uint) parm: zvol_threads:Number of threads for zvol device (uint) parm: zio_injection_enabled:Enable fault injection (int) parm: zio_bulk_flags:Additional flags to pass to bulk buffers (int) parm: zio_delay_max:Max zio millisec delay before posting event (int) parm: zio_requeue_io_start_cut_in_line:Prioritize requeued I/O (bool) parm: zil_replay_disable:Disable intent logging replay (int) parm: zfs_nocacheflush:Disable cache flushes (bool) parm: zfs_read_chunk_size:Bytes to read per chunk (long) parm: zfs_vdev_max_pending:Max pending per-vdev I/Os (int) parm: zfs_vdev_min_pending:Min pending per-vdev I/Os (int) parm: zfs_vdev_aggregation_limit:Max vdev I/O aggregation size (int) parm: zfs_vdev_time_shift:Deadline time shift for vdev I/O (int) parm: zfs_vdev_ramp_rate:Exponential I/O issue ramp-up rate (int) parm: zfs_vdev_read_gap_limit:Aggregate read I/O over gap (int) parm: zfs_vdev_write_gap_limit:Aggregate write I/O over gap (int) parm: zfs_vdev_scheduler:I/O scheduler (charp) parm: zfs_vdev_cache_max:Inflate reads small than max (int) parm: zfs_vdev_cache_size:Total size of the per-disk cache (int) parm: zfs_vdev_cache_bshift:Shift size to inflate reads too (int) parm: zfs_scrub_limit:Max scrub/resilver I/O per leaf vdev (int) parm: zfs_recover:Set to attempt to recover from fatal errors (int) parm: spa_config_path:SPA config file (/etc/zfs/zpool.cache) (charp) parm: zfs_zevent_len_max:Max event queue length (int) parm: zfs_zevent_cols:Max event column width (int) parm: zfs_zevent_console:Log events to the console (int) parm: zfs_top_maxinflight:Max I/Os per top-level (int) parm: zfs_resilver_delay:Number of ticks to delay resilver (int) parm: zfs_scrub_delay:Number of ticks to delay scrub (int) parm: zfs_scan_idle:Idle window in clock ticks (int) parm: zfs_scan_min_time_ms:Min millisecs to scrub per txg (int) parm: zfs_free_min_time_ms:Min millisecs to free per txg (int) parm: zfs_resilver_min_time_ms:Min millisecs to resilver per txg (int) parm: zfs_no_scrub_io:Set to disable scrub I/O (bool) parm: zfs_no_scrub_prefetch:Set to disable scrub prefetching (bool) parm: zfs_txg_timeout:Max seconds worth of delta per txg (int) parm: zfs_no_write_throttle:Disable write throttling (int) parm: zfs_write_limit_shift:log2(fraction of memory) per txg (int) parm: zfs_txg_synctime_ms:Target milliseconds between tgx sync (int) parm: zfs_write_limit_min:Min tgx write limit (ulong) parm: zfs_write_limit_max:Max tgx write limit (ulong) parm: zfs_write_limit_inflated:Inflated tgx write limit (ulong) parm: zfs_write_limit_override:Override tgx write limit (ulong) parm: zfs_prefetch_disable:Disable all ZFS prefetching (int) parm: zfetch_max_streams:Max number of streams per zfetch (uint) parm: zfetch_min_sec_reap:Min time before stream reclaim (uint) parm: zfetch_block_cap:Max number of blocks to fetch at a time (uint) parm: zfetch_array_rd_sz:Number of bytes in a array_read (ulong) parm: zfs_pd_blks_max:Max number of blocks to prefetch (int) parm: zfs_dedup_prefetch:Enable prefetching dedup-ed blks (int) parm: zfs_arc_min:Min arc size (ulong) parm: zfs_arc_max:Max arc size (ulong) parm: zfs_arc_meta_limit:Meta limit for arc size (ulong) parm: zfs_arc_reduce_dnlc_percent:Meta reclaim percentage (int) parm: zfs_arc_grow_retry:Seconds before growing arc size (int) parm: zfs_arc_shrink_shift:log2(fraction of arc to reclaim) (int) parm: zfs_arc_p_min_shift:arc_c shift to calc min/max arc_p (int)
2011-05-03 22:09:28 +00:00
MODULE_PARM_DESC(zio_delay_max, "Max zio millisec delay before posting event");
module_param(zio_requeue_io_start_cut_in_line, int, 0644);
MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O");
module_param(zfs_sync_pass_deferred_free, int, 0644);
MODULE_PARM_DESC(zfs_sync_pass_deferred_free,
"Defer frees starting in this pass");
module_param(zfs_sync_pass_dont_compress, int, 0644);
MODULE_PARM_DESC(zfs_sync_pass_dont_compress,
"Don't compress starting in this pass");
module_param(zfs_sync_pass_rewrite, int, 0644);
MODULE_PARM_DESC(zfs_sync_pass_rewrite,
"Rewrite new bps starting in this pass");
#endif