This reverts commit 9dcb971983
which was originally introduced to debug occasional slow I/Os.
These I/Os would complete eventually but were observed to take
several 100 seconds.
The root cause of this issue was the CFQ scheduler which can,
under certain conditions, excessively delay an I/O from being
issued to the device. This issue was mitigated somewhat by
commit 84daaddedb which ensures
the I/O elevator gets changed even for DM style devices.
This change isn't in any way harmful but it does conflict with
a required change to properly account from I/O wait time.
Because Linux does not export the io_schedule_timeout() function
we must instead rely on io_schedule() via cv_wait_io().
The additional debugging information which was added to the
delay event has been intentionally left in place.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
There have been reports of ZFS deadlocking due to what appears to
be a lost IO. This patch addes some debugging to determine the
exact state of the IO which neither 1) completed, 2) failed, or
3) timed out after zio_delay_max (30) seconds.
This information will be logged using the ZFS FMA infrastructure
as a 'delay' event and posted to the internal zevent log. By
default the last 64 events will be kept in the log but the limit
is configurable via the zfs_zevent_len_max module option.
To dump the contents of the log use the 'zpool events -v' command
and look for the resource.fs.zfs.delay event. It will include
various information about the pool, vdev, and zio which may shed
some light on the issue.
In the context of this change the 120 second kernel blocked thread
watchdog has been disabled for synchronous IOs.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #930
This reverts commit a5c20e2a0a which
accidentally introduced a regression for real 4k sector devices.
See issue #1065 for details.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #1065
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
Currently, the size of read and write requests on vdevs is aligned
according to the vdev's ashift, allocating a new ZIO buffer and padding
if need be.
This makes sense for write requests to prevent read/modify/write if the
write happens to be smaller than the device's internal block size.
For reads however, the rationale is less clear. It seems that the
original code aligns reads because, on Solaris, device drivers will
outright refuse unaligned requests.
We don't have that issue on Linux. Indeed, Linux block devices are able
to accept requests of any size, and take care of alignment issues
themselves.
As a result, there's no point in enforcing alignment for read requests
on Linux. This is a nice optimization opportunity for two reasons:
- We remove a memory allocation in a heavily-used code path;
- The request gets aligned in the lowest layer possible, which shrinks
the path that the additional, useless padding data has to travel.
For example, when using 4k-sector drives that lie about their sector
size, using 512b read requests instead of 4k means that there will
be less data traveling down the ATA/SCSI interface, even though the
drive actually reads 4k from the platter.
The only exception is raidz, because raidz needs to read the whole
allocated block for parity.
This patch removes alignment enforcement for read requests, except on
raidz. Note that we also remove an assertion that checks that we're
aligning a top-level vdev I/O, because that's not the case anymore for
repair writes that results from failed reads.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#1022
Buffers for the ARC are normally backed by the SPL virtual slab.
However, if memory is low, AND no slab objects are available,
AND a new slab cannot be quickly constructed a new emergency
object will be directly allocated.
These objects can be as large as order 5 on a system with 4k
pages. And because they are allocated with KM_PUSHPAGE, to
avoid a potential deadlock, they are not allowed to initiate I/O
to satisfy the allocation. This can result in the occasional
allocation failure.
However, since these allocations are allowed to block and
perform operations such as memory compaction they will eventually
succeed. Since this is not unexpected (just unlikely) behavior
this patch disables the warning for the allocation failure.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #465
The vdev queue layer may require a small number of buffers
when attempting to create aggregate I/O requests. Rather than
attempting to allocate them from the global zio buffers, which
is slow under memory pressure, it makes sense to pre-allocate
them because...
1) These buffers are short lived. They are only required for
the life of a single I/O at which point they can be used by
the next I/O.
2) The maximum number of concurrent buffers needed by a vdev is
small. It's roughly limited by the zfs_vdev_max_pending tunable
which defaults to 10.
By keeping a small list of these buffer per-vdev we can ensure
one is always available when we need it. This significantly
reduces contention on the vq->vq_lock, because we no longer
need to perform a slow allocation under this lock. This is
particularly important when memory is already low on the system.
It would probably be wise to extend the use of these buffers beyond
aggregate I/O and in to the raidz implementation. The inability
to quickly allocate buffer for the parity stripes could result in
similiar problems.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed by: Matt Ahrens <matt@delphix.com>
Reviewed by: Eric Schrock <eric.schrock@delphix.com>
Reviewed by: Robert Mustacchi <rm@joyent.com>
Reviewed by: Bill Pijewski <wdp@joyent.com>
Reviewed by: Richard Elling <richard.elling@richardelling.com>
Reviewed by: Steve Gonczi <gonczi@comcast.net>
Reviewed by: Garrett D'Amore <garrett.damore@gmail.com>
Reviewed by: Dan McDonald <danmcd@nexenta.com>
Reviewed by: Albert Lee <trisk@nexenta.com>
Approved by: Eric Schrock <eric.schrock@delphix.com>
Refererces to Illumos issue:
https://www.illumos.org/issues/1909
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#680
Add a standard zio constructor and destructor. Normally, this is
done to reduce to cost of allocating a new structure by reducing
expensive operations such as memory allocations. However, in this
case none of the operations moved out of zio_create() were really
very expensive.
This change was principly made as a debug patch (and workaround)
for a zio_destroy() race. The is good evidence that zio_create()
is reinitializing a mutex which is really still in use by another
thread. This would completely explain the observed symptoms in
the issue report.
This patch doesn't fix the root cause of the race, but it should
make it less likely by only initializing the mutex once in the
constructor. Also, this particular flaw might have gone unnoticed
in other zfs implementations due to the specific implementation
details of Linux ticket spinlocks.
Once the real root cause is determined and resolved this change
can be safely reverted. Until then this should help workaround
the issue.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #496
This patch was slightly flawed and allowed for zio->io_logical
to potentially not be reinitialized for a new zio. This could
lead to assertion failures in specific cases when debugging is
enabled (--enable-debug) and I/O errors are encountered. It
may also have caused problems when issues logical I/Os.
Since we want to make sure this workaround can be easily removed
in the future (when we have the real fix). I'm reverting this
change and applying a new version of the patch which includes
the zio->io_logical fix.
This reverts commit 2c6d0b1e07.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #602
Issue #604
Add a standard zio constructor and destructor. Normally, this is
done to reduce to cost of allocating a new structure by reducing
expensive operations such as memory allocations. However, in this
case none of the operations moved out of zio_create() were really
very expensive.
This change was principly made as a debug patch (and workaround)
for a zio_destroy() race. The is good evidence that zio_create()
is reinitializing a mutex which is really still in use by another
thread. This would completely explain the observed symptoms in
the issue report.
This patch doesn't fix the root cause of the race, but it should
make it less likely by only initializing the mutex once in the
constructor. Also, this particular flaw might have gone unnoticed
in other zfs implementations due to the specific implementation
details of Linux ticket spinlocks.
Once the real root cause is determined and resolved this change
can be safely reverted. Until then this should help workaround
the issue.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #496
It has been observed that some of the hottest locks are those
of the zio taskqs. Contention on these locks can limit the
rate at which zios are dispatched which limits performance.
This upstream change from Illumos uses new interface to the
taskqs which allow them to utilize a prealloc'ed taskq_ent_t.
This removes the need to perform an allocation at dispatch
time while holding the contended lock. This has the effect
of improving system performance.
Reviewed by: Albert Lee <trisk@nexenta.com>
Reviewed by: Richard Lowe <richlowe@richlowe.net>
Reviewed by: Alexey Zaytsev <alexey.zaytsev@nexenta.com>
Reviewed by: Jason Brian King <jason.brian.king@gmail.com>
Reviewed by: George Wilson <gwilson@zfsmail.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Approved by: Gordon Ross <gwr@nexenta.com>
References to Illumos issue:
https://www.illumos.org/issues/734
Ported-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#482
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
Today zfs tries to allocate blocks evenly across all devices.
This means when devices are imbalanced zfs will use lots of
CPU searching for space on devices which tend to be pretty
full. It should instead fail quickly on the full LUNs and
move onto devices which have more availability.
Reviewed by: Eric Schrock <Eric.Schrock@delphix.com>
Reviewed by: Matt Ahrens <Matt.Ahrens@delphix.com>
Reviewed by: Adam Leventhal <Adam.Leventhal@delphix.com>
Reviewed by: Albert Lee <trisk@nexenta.com>
Reviewed by: Gordon Ross <gwr@nexenta.com>
Approved by: Garrett D'Amore <garrett@nexenta.com>
References to Illumos issue and patch:
- https://www.illumos.org/issues/510
- https://github.com/illumos/illumos-gate/commit/5ead3ed965
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #340
The majority of the recursive operations performed by the dsl
are done either in the context of the tgx_sync_thread or during
pool import. It is these recursive operations which contribute
greatly to the stack depth. When this recursion is coupled with
a synchronous I/O in the same context overflow becomes possible.
Previously to handle this case I have focused on keeping the
individual stack frames as light as possible. This is a good
idea as long as it can be done in a way which doesn't overly
complicate the code. However, there is a better solution.
If we treat all zio's issued by the tgx_sync_thread as async then
we can use the tgx_sync_thread stack for the recursive parts, and
the zio_* threads for the I/O parts. This effectively doubles our
available stack space with the only drawback being a small delay
to schedule the I/O. However, in practice the scheduling time
is so much smaller than the actual I/O time this isn't an issue.
Another benefit of making the zio async is that the zio pipeline
is now parallel. That should mean for CPU intensive pipelines
such as compression or dedup performance may be improved.
With this change in place the worst case stack usage observed so
far is 6902 bytes. This is still higher than I'd like but
significantly improved. Additional changes to specific functions
should improve this further. This change allows us to revent
commit 6656bf5 which did some horrible things to the recursive
traverse_visitbp() callpath in the name of saving stack.
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)
It used to be the case that all KM_SLEEP allocations were GFS_NOFS.
Unfortunately this often resulted in the kernel being unable to
reclaim the ARC, inode, and dentry caches in a timely manor.
The fix was to make KM_SLEEP a GFP_KERNEL allocation in the SPL.
However, this increases the posibility of deadlocking the system
on a zfs write thread. If a zfs write thread attempts to perform
an allocation it may trigger synchronous reclaim. This reclaim
may attempt to flush dirty data/inode to disk to free memory.
Unforunately, this write cannot finish because the write thread
which would handle it is holding the previous transaction open.
Deadlock.
To avoid this all allocations in the zfs write thread path must
use KM_PUSHPAGE which prohibits synchronous reclaim for that
thread. In this way forward progress in ensured. The risk
with this change is I missed updating an allocation for the
write threads leaving an increased posibility of deadlock. If
any deadlocks remain they will be unlikely but we'll have to
make sure they all get fixed.
Linux kernel thread names are expected to be short. This change shortens
the zio thread names to 10 characters leaving a few chracters to append
the /<cpuid> to which the thread is bound. For example: z_wr_iss/0.
While there is no right maximum timeout for a disk IO we can start
laying the ground work to measure how long they do take in practice.
This change simply measures the IO time and if it exceeds 30s an
event is posted for 'zpool events'.
This value was carefully selected because for sd devices it implies
that at least one timeout (SD_TIMEOUT) has occured. Unfortunately,
even with FAILFAST set we may retry and request and not get an
error. This behavior is strongly dependant on the device driver
and how it is hooked in to the scsi error handling stack. However
by setting the limit at 30s we can log the event even if no error
was returned.
Slightly longer term we can start recording these delays perhaps
as a simple power-of-two histrogram. This histogram can then be
reported as part of the 'zpool status' command when given an command
line option.
None of this code changes the internal behavior of ZFS. Currently
it is simply for reporting excessively long delays.
Implement zio_execute() as 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.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Eliminated local variables pointing to members of the zio struct.
Just refer to the struct members directly. This saved about 32 bytes per
call, but this function can be called recurisvely up to 19 levels deep,
so we potentially save up to 608 bytes.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Decrease stack usage for various call paths by forcing certain
functions to be inlined. By inlining the functions the overhead
of a new stack frame is removed at the cost of increased code size.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
The zio_taskq_dispatch() function may be called at interrupt time
and it is critical that we never sleep.
Additionally, wrap taskq_dispatch() in a while loop because it may
fail. This is non optimal but is OK for now.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Fix non-c90 compliant code, for the most part these changes
simply deal with where a particular variable is declared.
Under c90 it must alway be done at the very start of a block.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
