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/4045illumos/illumos-gate@69962b5647
Ported-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#1913
3742 zfs comments need cleaner, more consistent style
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Eric Schrock <eric.schrock@delphix.com>
Approved by: Christopher Siden <christopher.siden@delphix.com>
References:
https://www.illumos.org/issues/3742illumos/illumos-gate@f717074149
Ported-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #1775
Porting notes:
1. The change to zfs_vfsops.c was dropped because it involves
zfs_mount_label_policy, which does not exist in the Linux port.
3741 zfs needs better comments
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Eric Schrock <eric.schrock@delphix.com>
Approved by: Christopher Siden <christopher.siden@delphix.com>
References:
https://www.illumos.org/issues/3741illumos/illumos-gate@3e30c24aee
Ported-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #1775
3582 zfs_delay() should support a variable resolution
3584 DTrace sdt probes for ZFS txg states
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: Dan McDonald <danmcd@nexenta.com>
Reviewed by: Richard Elling <richard.elling@dey-sys.com>
Approved by: Garrett D'Amore <garrett@damore.org>
References:
https://www.illumos.org/issues/3582illumos/illumos-gate@0689f76
Ported by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #1775
3642 dsl_scan_active() should not issue I/O to determine if async
destroying is active
3643 txg_delay should not hold the tc_lock
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Approved by: Gordon Ross <gwr@nexenta.com>
References:
https://www.illumos.org/issues/3642https://www.illumos.org/issues/3643illumos/illumos-gate@4a92375985
Ported-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #1775
Porting Notes:
1. The alignment assumptions for the tx_cpu structure assume that
a kmutex_t is 8 bytes. This isn't true under Linux but tc_pad[]
was adjusted anyway for consistency since this structure was
never carefully aligned in ZoL. If careful alignment does impact
performance significantly this should be reworked to be portable.
This change is an attempt to add visibility in to how txgs are being
formed on a system, in real time. To do this, a list was added to the
in memory SPA data structure for a pool, with each element on the list
corresponding to txg. These entries are then exported through the kstat
interface, which can then be interpreted in userspace.
For each txg, the following information is exported:
* Unique txg number (uint64_t)
* The time the txd was born (hrtime_t)
(*not* wall clock time; relative to the other entries on the list)
* The current txg state ((O)pen/(Q)uiescing/(S)yncing/(C)ommitted)
* The number of reserved bytes for the txg (uint64_t)
* The number of bytes read during the txg (uint64_t)
* The number of bytes written during the txg (uint64_t)
* The number of read operations during the txg (uint64_t)
* The number of write operations during the txg (uint64_t)
* The time the txg was closed (hrtime_t)
* The time the txg was quiesced (hrtime_t)
* The time the txg was synced (hrtime_t)
Note that while the raw kstat now stores relative hrtimes for the
open, quiesce, and sync times. Those relative times are used to
calculate how long each state took and these deltas and printed by
output handlers.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
A deadlock was accidentally introduced by commit e95853a which
can occur when the system is under memory pressure. What happens
is that while the txg_quiesce thread is holding the tx->tx_cpu
locks it enters memory reclaim. In the context of this memory
reclaim it then issues synchronous I/O to a ZVOL swap device.
Because the txg_quiesce thread is holding the tx->tx_cpu locks
a new txg cannot be opened to handle the I/O. Deadlock.
The fix is straight forward. Move the memory allocation outside
the critical region where the tx->tx_cpu locks are held. And for
good measure change the offending allocation to KM_PUSHPAGE to
ensure it never attempts to issue I/O during reclaim.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #1274
Create a kstat file which contains useful statistics about the
last N txgs processed. This can be helpful when analyzing pool
performance. The new KSTAT_TYPE_TXG type was added for this
purpose and it tracks the following statistics per-txg.
txg - Unique txg number
state - State (O)pen/(Q)uiescing/(S)yncing/(C)ommitted
birth; - Creation time
nread - Bytes read
nwritten; - Bytes written
reads - IOPs read
writes - IOPs write
open_time; - Length in nanoseconds the txg was open
quiesce_time - Length in nanoseconds the txg was quiescing
sync_time; - Length in nanoseconds the txg was syncing
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Allow the zfs_txg_timeout variable to be dynamically tuned at run
time. By pulling it down out of the variable declaration it will
be evaluted each time through the loop.
The zfs_txg_timeout variable is now declared extern in a the common
sys/txg.h header rather than locally in dsl_scan.c. This prevents
potential type mismatches if the global variable needs to be used
elsewhere.
Move the module_param() code in to the same source file where
zfs_txg_timeout is declared. This is the most logical location.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Differences between how paging is done on Solaris and Linux can cause
deadlocks if KM_SLEEP is used in any the following contexts.
* The txg_sync thread
* The zvol write/discard threads
* The zpl_putpage() VFS callback
This is because KM_SLEEP will allow for direct reclaim which may result
in the VM calling back in to the filesystem or block layer to write out
pages. If a lock is held over this operation the potential exists to
deadlock the system. To ensure forward progress all memory allocations
in these contexts must us KM_PUSHPAGE which disables performing any I/O
to accomplish the memory allocation.
Previously, this behavior was acheived by setting PF_MEMALLOC on the
thread. However, that resulted in unexpected side effects such as the
exhaustion of pages in ZONE_DMA. This approach touchs more of the zfs
code, but it is more consistent with the right way to handle these cases
under Linux.
This is patch lays the ground work for being able to safely revert the
following commits which used PF_MEMALLOC:
21ade34 Disable direct reclaim for z_wr_* threads
cfc9a5c Fix zpl_writepage() deadlock
eec8164 Fix ASSERTION(!dsl_pool_sync_context(tx->tx_pool))
Signed-off-by: Richard Yao <ryao@cs.stonybrook.edu>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #726
The txg_sync(), zfs_putpage(), zvol_write(), and zvol_discard()
call paths must only use KM_PUSHPAGE to avoid potential deadlocks
during direct reclaim.
This patch annotates these call paths so any accidental use of
KM_SLEEP will be quickly detected. In the interest of stability
if debugging is disabled the offending allocation will have its
GFP flags automatically corrected. When debugging is enabled
any misuse will be treated as a fatal error.
This patch is entirely for debugging. We should be careful to
NOT become dependant on it fixing up the incorrect allocations.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Commit eec8164771 worked around an issue
involving direct reclaim through the use of PF_MEMALLOC. Since we
are reworking thing to use KM_PUSHPAGE so that swap works, we revert
this patch in favor of the use of KM_PUSHPAGE in the affected areas.
Signed-off-by: Richard Yao <ryao@cs.stonybrook.edu>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #726
After surveying the code, the few places where smp_processor_id is used
were deemed to be safe to use with a preempt enabled kernel. As such, no
core logic had to be changed. These smp_processor_id call sites are simply
are wrapped in kpreempt_disable and kpreempt_enabled to prevent the
Linux kernel from emitting scary warnings.
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <ryao@cs.stonybrook.edu>
Issue #83
Keep counters for the various reasons that a thread may end up
in txg_wait_open() waiting on a new txg. This can be useful
when attempting to determine why a particular workload is
under performing.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
The function txg_delay() is used to delay txg (transaction group)
threads in ZFS. The timeout value for this function is calculated
using:
int timeout = ddi_get_lbolt() + ticks;
Later, the actual wait is performed:
while (ddi_get_lbolt() < timeout &&
tx->tx_syncing_txg < txg-1 && !txg_stalled(dp))
(void) cv_timedwait(&tx->tx_quiesce_more_cv, &tx->tx_sync_lock,
timeout - ddi_get_lbolt());
The ddi_get_lbolt() function returns current uptime in clock ticks
and is typed as clock_t. The clock_t type on 64-bit architectures
is int64_t.
The "timeout" variable will overflow depending on the tick frequency
(e.g. for 1000 it will overflow in 28.855 days). This will make the
expression "ddi_get_lbolt() < timeout" always false - txg threads will
not be delayed anymore at all. This leads to a slowdown in ZFS writes.
The attached patch initializes timeout as clock_t to match the return
value of ddi_get_lbolt().
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #352
Disable the normal reclaim path for the txg_sync thread. This
ensures the thread will never enter dmu_tx_assign() which can
otherwise occur due to direct reclaim. If this is allowed to
happen the system can deadlock. Direct reclaim call path:
->shrink_icache_memory->prune_icache->dispose_list->
clear_inode->zpl_clear_inode->zfs_inactive->dmu_tx_assign
Kernel threads which sleep uninterruptibly on Linux are marked in the (D)
state. These threads are usually in the process of performing IO and are
thus counted against the load average. The txg_quiesce and txg_sync threads
were always sleeping uninterruptibly and thus inflating the load average.
This change makes them sleep interruptibly. Some care is required however
because these threads may now be woken early by signals. In this case the
callers are all careful to check that the required conditions are met after
waking up. If we're woken early due to a signal they will simply go back
to sleep. In this case these changes are safe.
Closes#175
This adds an API to wait for pending commit callbacks of already-synced
transactions to finish processing. This is needed by the DMU-OSD in
Lustre during device finalization when some callbacks may still not be
called, this leads to non-zero reference count errors. See lustre.org
bug 23931.
The upstream commit cb code had a few bugs:
1) The arguments of the list_move_tail() call in txg_dispatch_callbacks()
were reversed by mistake. This caused the commit callbacks to not be
called at all.
2) ztest had a bug in ztest_dmu_commit_callbacks() where "error" was not
initialized correctly. This seems to have caused the test to always take
the simulated error code path, which made ztest unable to detect whether
commit cbs were being called for transactions that successfuly complete.
3) ztest had another bug in ztest_dmu_commit_callbacks() where the commit
cb threshold was not being compared correctly.
4) The commit cb taskq was using 'max_ncpus * 2' as the maxalloc argument
of taskq_create(), which could have caused unnecessary delays in the txg
sync thread.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
