6569 large file delete can starve out write ops
illumos/illumos-gate@ff5177ee8bff5177ee8bhttps://www.illumos.org/issues/6569
The core issue I've found is that there is no throttle for how many
deletes get assigned to one TXG. As a results when deleting large files
we end up filling consecutive TXGs with deletes/frees, then write
throttling other (more important) ops.
There is an easy test case for this problem. Try deleting several
large files (at least 1/2 TB) while you do write ops on the same
pool. What we've seen is performance of these write ops (let's
call it sideload I/O) would drop to zero.
More specifically the problem is that dmu_free_long_range_impl()
can/will fill up all of the dirty data in the pool "instantly",
before many of the sideload ops can get in. So sideload
performance will be impacted until all the files are freed.
The solution we have tested at Nexenta (with positive results)
creates a relatively simple throttle for how many "free" ops we let
into one TXG.
However this solution exposes other problems that should also be
addressed. If we are to slow down freeing of data that means one
has to wait even longer (assuming vnode ref count of 1) to get shell
back after an rm or for NFS thread to finish the free-ing op.
To avoid this the proposed solution is to call zfs_inactive() async
for "large" files. Async freeing then begs for the reclaimed space
to be accounted for in the zpool's "freeing" prop.
The other issue with having a longer delete is the inability to
export/unmount for a longer period of time. The proposed solution
is to interrupt freeing of blocks when a fs is unmounted.
Author: Alek Pinchuk <alek@nexenta.com>
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed by: Sanjay Nadkarni <sanjay.nadkarni@nexenta.com>
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Approved by: Dan McDonald <danmcd@omniti.com>
Reviewed by: avg
Differential Revision: D9008
the fifth argument to functions being traced, however there was an error
where the userspace stack was being used. This may be invalid leading to
a kernel panic if this address is unmapped.
Submitted by: Graeme Jenkinson <graeme.jenkinson@cl.cam.ac.uk>
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D9229
It was a temporary change to ease an import of native atomic_cas primitives.
Instead, atomic_fcmpset was devised with different semantics. See r311168.
At least on FreeBSD there are no legal way to access media or get its
size without opening device/provider first. Postponing this caching
allows to skip several disk seeks per ZVOL/snapshot during import.
For HDD pool with 1 ZVOL in dev mode with 1000 snapshots this reduces
pool import time from 40 to 10 seconds.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
These functions may be called in DTrace probe context, so they cannot be
safely traced. Moreover, they are currently only used by DTrace, so their
corresponding FBT probes are not particularly useful.
MFC after: 2 weeks
Original commit "7090 zfs should improve allocation order" declares alloc
queue sorted by time and offset. But in practice io_offset is always zero,
so sorting happened only by time, while order of writes with equal time was
completely random. On Illumos this did not affected much thanks to using
high resolution timestamps. On FreeBSD due to using much faster but low
resolution timestamps it caused bad data placement on disks, affecting
further read performance.
This change switches zio_timestamp_compare() from comparing uninitialized
io_offset to really populated io_bookmark values. I haven't decided yet
what to do with timestampts, but on simple tests this change gives the
same peformance results by just making code to work as declared.
MFC after: 1 week
On FreeBSD the sense of rw_write_held() and rw_iswriter() were reversed,
probably due to a cut and paste error. Using rw_iswriter() would cause
the kernel to panic.
Reviewed by: markj
MFC after: 2 weeks
Sponsored by: DARPA, AFRL
Differential Revision: https://reviews.freebsd.org/D8718
Note: there was a merge conflict resolved by me.
illumos/illumos-gate@43297f973a43297f973ahttps://www.illumos.org/issues/3821
We recently had nodes with some of the latest zfs bits panic on us in a
rollback-heavy environment. The following is from my preliminary analysis:
Let's look at where we died:
> $C
ffffff01ea6b9a10 taskq_dispatch+0x3a(0, fffffffff7d20450, ffffff5551dea920, 1)
ffffff01ea6b9a60 zil_clean+0xce(ffffff4b7106c080, 7e0f1)
ffffff01ea6b9aa0 dsl_pool_sync_done+0x47(ffffff4313065680, 7e0f1)
ffffff01ea6b9b70 spa_sync+0x55f(ffffff4310c1d040, 7e0f1)
ffffff01ea6b9c20 txg_sync_thread+0x20f(ffffff4313065680)
ffffff01ea6b9c30 thread_start+8()
If we dig in we can find that this dataset corresponds to a zone:
> ffffff4b7106c080::print zilog_t zl_os->os_dsl_dataset->ds_dir->dd_myname
zl_os->os_dsl_dataset->ds_dir->dd_myname = [ "8ffce16a-13c2-4efa-a233-
9e378e89877b" ]
Okay so we have a null taskq pointer. That only happens during the calls to
zil_open and zil_close. If we poke around we can see that we're actually in
midst of a rollback:
> ::pgrep zfs | ::printf "0x%x %s\\n" proc_t . p_user.u_psargs
0xffffff43262800a0 zfs rollback zones/15714eb6-f5ea-469f-ac6d-
4b8ab06213c2@marlin_init
0xffffff54e22a1028 zfs rollback zones/8ffce16a-13c2-4efa-a233-
9e378e89877b@marlin_init
0xffffff4362f3a058 zfs rollback zones/0ddb8e49-ca7e-42e1-8fdc-
4ac4ba8fe9f8@marlin_init
0xffffff5748e8d020 zfs rollback zones/426357b5-832d-4430-953e-
10cd45ff8e9f@marlin_init
0xffffff436b867008 zfs rollback zones/8f36bf37-8a9c-4a44-995c-
6d1b2751e6f5@marlin_init
0xffffff4381ad4090 zfs rollback zones/6c8eca18-fbd6-46dd-ac24-
2ed45cd0da70@marlin_init
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Dan Kimmel <dan.kimmel@delphix.com>
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed by: Andriy Gapon <avg@FreeBSD.org>
Approved by: Richard Lowe <richlowe@richlowe.net>
Author: George Wilson <george.wilson@delphix.com>
MFC after: 3 weeks
illumos/illumos-gate@90f2c094b390f2c094b3https://www.illumos.org/issues/7181
zfsvfs_setup() is called in both zfs_mount and zfs_resume_fs paths.
dmu_objset_set_user(zfsvfs->z_os, zfsvfs) is called early in zfsvfs_setup()
before the setup is actually completed,
thus an under-constructed zfsvfs becomes visible.
Additionally, there is nothing to serialize the two call paths. As a result two
threads can step on each other's toes.
assertion failed: zilog->zl_clean_taskq == NULL, file:
../../common/fs/zfs/zil.c, line: 1772
> $c
vpanic()
0xfffffffffbdf6928()
zil_open+0x45(ffffff1bbc5dd000, fffffffff7993880)
zfsvfs_setup+0x84(ffffffb378d77000, 0)
zfs_resume_fs+0x132(ffffffb378d77000, ffffffb37ddcf000)
zfs_ioc_rollback+0x96(ffffffb37ddcf000, ffffff01dcdc4cd0, ffffff01aa091000)
zfsdev_ioctl+0x215(10a00000000, 5a19, 80465f8, 100003, ffffff01ab318368,
ffffff0004b59e58)
cdev_ioctl+0x39(10a00000000, 5a19, 80465f8, 100003, ffffff01ab318368,
ffffff0004b59e58)
spec_ioctl+0x60(ffffff0197737700, 5a19, 80465f8, 100003,
ffffff01ab318368, ffffff0004b59e58)
fop_ioctl+0x55(ffffff0197737700, 5a19, 80465f8, 100003,
ffffff01ab318368, ffffff0004b59e58)
ioctl+0x9b(7, 5a19, 80465f8)
sys_syscall32+0x1f7()
> ffffff1bbc5dd000::print objset_t os_zil
os_zil = 0xffffff1c053cf7c0
> 0xffffff1c053cf7c0::print zilog_t zl_clean_taskq
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Approved by: Gordon Ross <gordon.w.ross@gmail.com>
Author: Andriy Gapon <andriy.gapon@clusterhq.com>
MFC after: 2 weeks
already free blocks
7199 dsl_dataset_rollback_sync may try to free already free blocks
7200 no blocks must be born in a txg after a snaphot is created
illumos/illumos-gate@bfaed0b91ebfaed0b91ehttps://www.illumos.org/issues/7199
dsl_dataset_rollback_sync may try to free already freed blocks when it calls
dsl_destroy_head_sync_impl to destroy a temporary clone.
That happens if a snapshot to which we are rolling back and from which the
clone is created has some ZIL records.
https://www.illumos.org/issues/7200
No new blocks must be born in a dataset in the same TXG after a snapshot of the
dataset is taken.
Those blocks would have the same blk_birth as the dataset's ds_prev_snap_txg
and as such they would be presumed to belong o the snapshot while in fact they
do not.
All the datasets must be clean before sync tasks are run, so the described
scenario may happen only if one of the sync tasks dirties the dataset and
another sync task takes its snapshot.
Then, there will be another sync pass because of the dirty data and the new
blocks will be born in the same TXG when the data is written out.
It seems that almost all of the existing sync tasks modify only MOS and do not
dirty any objsets.
The only exception that I've been able to identify so far is the rollback which
can modify an objset when it zeroes out the objset's ZIL.
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Approved by: Gordon Ross <gordon.w.ross@gmail.com>
Author: Andriy Gapon <andriy.gapon@clusterhq.com>
MFC after: 3 weeks
and zfs_ioc_rename
illumos/illumos-gate@690041b9ca690041b9cahttps://www.illumos.org/issues/7180
If a filesystem is not unmounted while the rename is being performed, then, for
example, a concurrect zfs rollback may call zfs_suspend_fs followed by
zfs_resume_fs on the same filesystem.
The latter takes the filesystem's name as an argument. If the filesystem name
changes as a result of the rename, then dmu_objset_hold(osname, zfsvfs, &os)
call in zfs_resume_fs would fail resulting in a kernel panic.
So far I have been able to reproduce this problem on FreeBSD where zfs rename
has -u option that skips the unmounting before doing the renaming.
But I think that in theory the same problem can occur on illumos as well,
because the unmounting is done in userland before invoking the rename ioctl and
there could be a race with, e.g., zfs mount.
panic: solaris assert: dmu_objset_hold(osname, zfsvfs, &zfsvfs->z_os) == 0 (0x2
== 0x0), file: /usr/devel/svn/head/sys/cddl/contrib/opensolaris/uts/common/fs/
zfs/zfs_vfsops.c, line: 2210
KDB: stack backtrace:
db_trace_self_wrapper() at db_trace_self_wrapper+0x2b/frame 0xfffffe004df30710
vpanic() at vpanic+0x182/frame 0xfffffe004df30790
panic() at panic+0x43/frame 0xfffffe004df307f0
assfail3() at assfail3+0x2c/frame 0xfffffe004df30810
zfs_resume_fs() at zfs_resume_fs+0xb9/frame 0xfffffe004df30860
zfs_ioc_rollback() at zfs_ioc_rollback+0x61/frame 0xfffffe004df308a0
zfsdev_ioctl() at zfsdev_ioctl+0x65c/frame 0xfffffe004df30940
devfs_ioctl_f() at devfs_ioctl_f+0x156/frame 0xfffffe004df309a0
kern_ioctl() at kern_ioctl+0x246/frame 0xfffffe004df30a00
sys_ioctl() at sys_ioctl+0x171/frame 0xfffffe004df30ae0
amd64_syscall() at amd64_syscall+0x2db/frame 0xfffffe004df30bf0
Xfast_syscall() at Xfast_syscall+0xfb/frame 0xfffffe004df30bf0
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Approved by: Richard Lowe <richlowe@richlowe.net>
MFC after: 2 weeks
It was very wrong to look at the vnode and znode internals without
having locked the vnode first.
Reported by: pho
Tested by: pho
MFC after: 1 week
X-MFC with: r308887
longer used. More precisely, they are always zero because the code that
decremented and incremented them no longer exists.
Bump __FreeBSD_version to mark this change.
Reviewed by: kib, markj
Sponsored by: Dell EMC Isilon
Differential Revision: https://reviews.freebsd.org/D8583
The idea was to avoid a false assertion in zfs_lock, but it was
implemented very dangerously and incorrectly.
Reported by: pho
Tested by: pho
MFC after: 1 week
due to zl_itx_list_sz not updated when async itx'es upgraded to sync.
Actually because of other changes about that time zl_itx_list_sz is not
really required to implement the functionality, so this patch removes
some unneeded broken code and variables.
Original idea of zil_slog_limit was to reduce chance of SLOG abuse by
single heavy logger, that increased latency for other (more latency critical)
loggers, by pushing heavy log out into the main pool instead of SLOG. Beside
huge latency increase for heavy writers, this implementation caused double
write of all data, since the log records were explicitly prepared for SLOG.
Since we now have I/O scheduler, I've found it can be much more efficient
to reduce priority of heavy logger SLOG writes from ZIO_PRIORITY_SYNC_WRITE
to ZIO_PRIORITY_ASYNC_WRITE, while still leave them on SLOG.
Existing ZIL implementation had problem with space efficiency when it
has to write large chunks of data into log blocks of limited size. In some
cases efficiency stopped to almost as low as 50%. In case of ZIL stored on
spinning rust, that also reduced log write speed in half, since head had to
uselessly fly over allocated but not written areas. This change improves
the situation by offloading problematic operations from z*_log_write() to
zil_lwb_commit(), which knows real situation of log blocks allocation and
can split large requests into pieces much more efficiently. Also as side
effect it removes one of two data copy operations done by ZIL code WR_COPIED
case.
While there, untangle and unify code of z*_log_write() functions.
Also zfs_log_write() alike to zvol_log_write() can now handle writes crossing
block boundary, that may also improve efficiency if ZPL is made to do that.
Sponsored by: iXsystems, Inc.
not remove user-space visible fields from vm_cnt or all of the references to
cached pages from comments. Those changes will come later.)
Reviewed by: kib, markj
Tested by: pho
Sponsored by: Dell EMC Isilon
Differential Revision: https://reviews.freebsd.org/D8497
These are FreeBSD-specific and were added in r178576 to provide the ability
to pretty-print instances of compound types. However, the print action has
long since been augmented to provide this functionality with a simpler
interface.
Discussed with: gnn
Differential Revision: https://reviews.freebsd.org/D8478
Before this an earlier writes to a ZVOL opened without FSYNC could get to
ZIL after later writes to the same ZVOL opened with FSYNC. Fix this by
replicating functionality of ZPL (zv_sync_cnt equivalent to z_sync_cnt),
marking all log records sync if anybody opened the ZVOL with FSYNC.
MFC after: 2 weeks
(gpt)zfsboot will read one-time boot directives from a special ZFS pool
area. The area was previously described as "Boot Block Header", but
currently it is know as Pad2, marked as reserved and is zeroed out on
pool creation. The new code interprets data in this area, if any, using
the same format as boot.config. The area is immediately wiped out.
Failure to parse the directives results in a reboot right after the
cleanup. Otherwise the boot sequence proceeds as usual.
zfsbootcfg writes zfsboot arguments specified on its command line to the
Pad2 area of a disk identified by vfs.zfs.boot.primary_pool and
vfs.zfs.boot.primary_vdev kenv variables that are set by loader during
boot. Please see the manual page for more.
Thanks to all who reviewed, contributed and made suggestions! There are
many potential improvements to the feature, please see the review for
details.
Reviewed by: wblock (docs)
Discussed with: jhb, tsoome
MFC after: 3 weeks
Relnotes: yes
Differential Revision: https://reviews.freebsd.org/D7612
It allows to avoid extra GEOM providers flapping without significant need.
Since GEOM got resize support, we don't need to reopen provider to get new
size. If provider was orphaned and no longer valid, ZFS should already
know that, and in such case reopen should be done in full as expected.
MFC after: 2 weeks
In case of vdev detach, causing top level mirror vdev destruction, leaf
vdev changes its GUID to one of the destroyed mirror, that creates race
condition when GUID in vdev label may not match one in the pool config.
This change replicates logic nuance of vdev_validate() by adding special
exception, matching the vdev GUID against the top level vdev GUID.
Since this exception is not completely reliable (may give false positives
if we fail to erase label on detached vdev), use it only as last resort.
Quick way to reproduce this scenario now is detach vdev from a pool with
enabled autoextend. During vdev detach autoextend logic tries to reopen
remaining vdev, that always fails now since in-memory configuration is
already updated, while on-disk labels are not yet.
MFC after: 2 weeks
illumos/illumos-gate@260af64db7260af64db7https://www.illumos.org/issues/3746
From the original change log:
It was possible for a reference to be added even with the lock held, and
for references added just after a lock release to be lost.
This bug was also independently found and reported in wesunsolve.net
issues 6985013 6995524.
In zrl_add(), always use an atomic operation to update the refcount.
The mutex in the ZRL only guarantees that wakeups occur for waiters on the
lock. It offers no protection against concurrent updates of the refcount.
The only refcount transition that is safe to perform without an atomic
operation is from ZRL_LOCKED back to 0, since this can only be performed
by the thread which has the ZRL locked.
Authored by: Will Andrews <will@freebsd.org>
Reviewed by: Boris Protopopov <bprotopopov@hotmail.com>
Reviewed by: Pavel Zakharov <pavel.zakha@gmail.com>
Reviewed by: Yuri Pankov <yuri.pankov@gmail.com>
Reviewed by: Justin T. Gibbs <gibbs@scsiguy.com>
Approved by: Matt Ahrens <mahrens@delphix.com>
Author: Youzhong Yang <yyang@mathworks.com>
PR: 204037
MFC after: 1 week
These allocations can reach up to 128KB, while FreeBSD kernel allocator
can cache allocations only up to 64KB. To avoid expensive allocations
for each large ZIL write use caching zio_buf_alloc() allocator instead.
To make it possible de-inline few instances of zil_itx_destroy().
6988 spa_sync() spends half its time in dmu_objset_do_userquota_updates
Using a benchmark which creates 2 million files in one TXG, I observe
that the thread running spa_sync() is on CPU almost the entire time we
are syncing, and therefore can be a performance bottleneck. About 50% of
the time in spa_sync() is in dmu_objset_do_userquota_updates().
The problem is that dmu_objset_do_userquota_updates() calls
zap_increment_int(DMU_USERUSED_OBJECT) once for every file that was
modified (or created). In this benchmark, all the files are owned by the
same user/group, so all 2 million calls to zap_increment_int() are
modifying the same entry in the zap. The same issue exists for the
DMU_GROUPUSED_OBJECT.
We should keep an in-memory map from user to space delta while we are
syncing, and when we finish, iterate over the in-memory map and modify
the ZAP once per entry. This reduces the number of calls to
zap_increment_int() from "number of objects modified" to "number of
owners/groups of modified files".
This reduced the time spent in spa_sync() in the file create benchmark
by ~33%, from 11 seconds to 7 seconds.
Closes#107
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Steve Gonczi <steve.gonczi@delphix.com>
Reviewed by: Ned Bass <bass6@llnl.gov>
Reviewed by: Jinshan Xiong <jinshan.xiong@intel.com>
Author: Matthew Ahrens <mahrens@delphix.com>
openzfs/openzfs@5fc46359c5
5120 zfs should allow large block/gzip/raidz boot pool (loader project)
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Yuri Pankov <yuri.pankov@nexenta.com>
Reviewed by: Andrew Stormont <andyjstormont@gmail.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Approved by: Robert Mustacchi <rm@joyent.com>
Author: Toomas Soome <tsoome@me.com>
openzfs/openzfs@c8811bd3e2
FreeBSD still does not support booting from gzip-compressed datasets,
so keep one chunk of this commit out.
Suppose that we have an exclusively busy page, and a thread which can
accept shared-busy page. In this case, typical code waiting for the
page xbusy state to pass is
again:
VM_OBJECT_WLOCK(object);
...
if (vm_page_xbusied(m)) {
vm_page_lock(m);
VM_OBJECT_WUNLOCK(object); <---1
vm_page_busy_sleep(p, "vmopax");
goto again;
}
Suppose that the xbusy state owner locked the object, unbusied the
page and unlocked the object after we are at the line [1], but before we
executed the load of the busy_lock word in vm_page_busy_sleep(). If it
happens that there is still no waiters recorded for the busy state,
the xbusy owner did not acquired the page lock, so it proceeded.
More, suppose that some other thread happen to share-busy the page
after xbusy state was relinquished but before the m->busy_lock is read
in vm_page_busy_sleep(). Again, that thread only needs vm_object lock
to proceed. Then, vm_page_busy_sleep() reads busy_lock value equal to
the VPB_SHARERS_WORD(1).
In this case, all tests in vm_page_busy_sleep(9) pass and we are going
to sleep, despite the page being share-busied.
Update check for m->busy_lock == VPB_UNBUSIED in vm_page_busy_sleep(9)
to also accept shared-busy state if we only wait for the xbusy state to
pass.
Merge sequential if()s with the same 'then' clause in
vm_page_busy_sleep().
Note that the current code does not share-busy pages from parallel
threads, the only way to have more that one sbusy owner is right now
is to recurse.
Reported and tested by: pho (previous version)
Reviewed by: alc, markj
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D8196
Other uses of cache_purgevfs() do rely on the cache purge for correct
operations, when paths are invalidated without unmount.
Reported and tested by: jkim
Discussed with: mjg
Sponsored by: The FreeBSD Foundation
This should allow vn_fullpath() to work even when vfs name cache is
disabled for zfs, which is the case when zfs properties like
casesensitivity and normalization are set non-default values.
The new code should be 100% reliable for directories and "mostly"
reliable for files, that is, when hardlinks across directories are
not used.
Reported by: Frederic Chardon <chardon.frederic@gmail.com>
Reviewed by: kib (vfs contract)
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D8146
For the extended attributes the order between z_teardown_lock and the
vnode lock is different.
The bug was triggered only with DIAGNOSTIC turned on.
This fix is developed in cooperation with avos.
PR: 213112
Reported by: avos
Tested by: avos
MFC after: 1 week
