Authored by: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: John Kennedy <jwk404@gmail.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed by: Igor Kozhukhov <igor@dilos.org>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Approved by: Robert Mustacchi <rm@joyent.com>
Ported-by: Prakash Surya <prakash.surya@delphix.com>
PROBLEM
=======
There's a race condition that exists if `zil_free_lwb` races with either
`zil_commit_waiter_timeout` and/or `zil_lwb_flush_vdevs_done`.
Here's an example panic due to this bug:
> ::status
debugging crash dump vmcore.0 (64-bit) from ip-10-110-205-40
operating system: 5.11 dlpx-5.2.2.0_2017-12-04-17-28-32b6ba51fb (i86pc)
image uuid: 4af0edfb-e58e-6ed8-cafc-d3e9167c7513
panic message:
BAD TRAP: type=e (#pf Page fault) rp=ffffff0010555970 addr=60 occurred in module "zfs" due to a NULL pointer dereference
dump content: kernel pages only
> $c
zio_shrink+0x12()
zil_lwb_write_issue+0x30d(ffffff03dcd15cc0, ffffff03e0730e20)
zil_commit_waiter_timeout+0xa2(ffffff03dcd15cc0, ffffff03d97ffcf8)
zil_commit_waiter+0xf3(ffffff03dcd15cc0, ffffff03d97ffcf8)
zil_commit+0x80(ffffff03dcd15cc0, 9a9)
zfs_write+0xc34(ffffff03dc38b140, ffffff0010555e60, 40, ffffff03e00fb758, 0)
fop_write+0x5b(ffffff03dc38b140, ffffff0010555e60, 40, ffffff03e00fb758, 0)
write+0x250(42, fffffd7ff4832000, 2000)
sys_syscall+0x177()
If there's an outstanding lwb that's in `zil_commit_waiter_timeout`
waiting to timeout, waiting on it's waiter's CV, we must be sure not to
call `zil_free_lwb`. If we end up calling `zil_free_lwb`, then that LWB
may be freed and can result in a use-after-free situation where the
stale lwb pointer stored in the `zil_commit_waiter_t` structure of the
thread waiting on the waiter's CV is used.
A similar situation can occur if an lwb is issued to disk, and thus in
the `LWB_STATE_ISSUED` state, and `zil_free_lwb` is called while the
disk is servicing that lwb. In this situation, the lwb will be freed by
`zil_free_lwb`, which will result in a use-after-free situation when the
lwb's zio completes, and `zil_lwb_flush_vdevs_done` is called.
This race condition is prevented in `zil_close` by calling `zil_commit`
before `zil_free_lwb` is called, which will ensure all outstanding (i.e.
all lwb's in the `LWB_STATE_OPEN` and/or `LWB_STATE_ISSUED` states)
reach the `LWB_STATE_DONE` state before the lwb's are freed
(`zil_commit` will not return untill all the lwb's are
`LWB_STATE_DONE`).
Further, this race condition is prevented in `zil_sync` by only calling
`zil_free_lwb` for lwb's that do not have their `lwb_buf` pointer set.
All lwb's not in the `LWB_STATE_DONE` state will have a non-null value
for this pointer; the pointer is only cleared in
`zil_lwb_flush_vdevs_done`, at which point the lwb's state will be
changed to `LWB_STATE_DONE`.
This race *is* present in `zil_suspend`, leading to this bug.
At first glance, it would appear as though this would not be true
because `zil_suspend` will call `zil_commit`, just like `zil_close`, but
the problem is that `zil_suspend` will set the zilog's `zl_suspend`
field prior to calling `zil_commit`. Further, in `zil_commit`, if
`zl_suspend` is set, `zil_commit` will take a special branch of logic
and use `txg_wait_synced` instead of performing the normal `zil_commit`
logic.
This call to `txg_wait_synced` might be good enough for the data to
reach disk safely before it returns, but it does not ensure that all
outstanding lwb's reach the `LWB_STATE_DONE` state before it returns.
This is because, if there's an lwb "stuck" in
`zil_commit_waiter_timeout`, waiting for it's lwb to timeout, it will
maintain a non-null value for it's `lwb_buf` field and thus `zil_sync`
will not free that lwb. Thus, even though the lwb's data is already on
disk, the lwb will be left lingering, waiting on the CV, and will
eventually timeout and be issued to disk even though the write is
unnecessary.
So, after `zil_commit` is called from `zil_suspend`, we incorrectly
assume that there are not outstanding lwb's, and proceed to free all
lwb's found on the zilog's lwb list. As a result, we free the lwb that
will later be used `zil_commit_waiter_timeout`.
SOLUTION
========
The solution to this, is to ensure all outstanding lwb's complete before
calling `zil_free_lwb` via `zil_destroy` in `zil_suspend`. This patch
accomplishes this goal by forcing the normal `zil_commit` logic when
called from `zil_sync`.
Now, `zil_suspend` will call `zil_commit_impl` which will always use the
normal logic of waiting/issuing lwb's to disk before it returns. As a
result, any lwb's outstanding when `zil_commit_impl` is called will be
guaranteed to reach the `LWB_STATE_DONE` state by the time it returns.
Further, no new lwb's will be created via `zil_commit` since the zilog's
`zl_suspend` flag will be set. This will force all new callers of
`zil_commit` to use `txg_wait_synced` instead of creating and issuing
new lwb's.
Thus, all lwb's left on the zilog's lwb list when `zil_destroy` is
called will be in the `LWB_STATE_DONE` state, and we'll avoid this race
condition.
OpenZFS-issue: https://www.illumos.org/issues/8909
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/ece62b6f8dCloses#6940
This is a purely cosmetic change. The zilog's "zl_writer_lock" field is
being renamed to "zl_issuer_lock" to try and make the code easier to
understand; no other changes are made.
Authored by: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: C Fraire <cfraire@me.com>
Approved by: Dan McDonald <danmcd@joyent.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Ported-by: Giuseppe Di Natale <dinatale2@llnl.gov>
OpenZFS-issue: https://www.illumos.org/issues/8603
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/2daf06546bCloses#6927
Authored by: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Approved by: Dan McDonald <danmcd@joyent.com>
Ported-by: Prakash Surya <prakash.surya@delphix.com>
Problem
=======
The current implementation of zil_commit() can introduce significant
latency, beyond what is inherent due to the latency of the underlying
storage. The additional latency comes from two main problems:
1. When there's outstanding ZIL blocks being written (i.e. there's
already a "writer thread" in progress), then any new calls to
zil_commit() will block waiting for the currently oustanding ZIL
blocks to complete. The blocks written for each "writer thread" is
coined a "batch", and there can only ever be a single "batch" being
written at a time. When a batch is being written, any new ZIL
transactions will have to wait for the next batch to be written,
which won't occur until the current batch finishes.
As a result, the underlying storage may not be used as efficiently
as possible. While "new" threads enter zil_commit() and are blocked
waiting for the next batch, it's possible that the underlying
storage isn't fully utilized by the current batch of ZIL blocks. In
that case, it'd be better to allow these new threads to generate
(and issue) a new ZIL block, such that it could be serviced by the
underlying storage concurrently with the other ZIL blocks that are
being serviced.
2. Any call to zil_commit() must wait for all ZIL blocks in its "batch"
to complete, prior to zil_commit() returning. The size of any given
batch is proportional to the number of ZIL transaction in the queue
at the time that the batch starts processing the queue; which
doesn't occur until the previous batch completes. Thus, if there's a
lot of transactions in the queue, the batch could be composed of
many ZIL blocks, and each call to zil_commit() will have to wait for
all of these writes to complete (even if the thread calling
zil_commit() only cared about one of the transactions in the batch).
To further complicate the situation, these two issues result in the
following side effect:
3. If a given batch takes longer to complete than normal, this results
in larger batch sizes, which then take longer to complete and
further drive up the latency of zil_commit(). This can occur for a
number of reasons, including (but not limited to): transient changes
in the workload, and storage latency irregularites.
Solution
========
The solution attempted by this change has the following goals:
1. no on-disk changes; maintain current on-disk format.
2. modify the "batch size" to be equal to the "ZIL block size".
3. allow new batches to be generated and issued to disk, while there's
already batches being serviced by the disk.
4. allow zil_commit() to wait for as few ZIL blocks as possible.
5. use as few ZIL blocks as possible, for the same amount of ZIL
transactions, without introducing significant latency to any
individual ZIL transaction. i.e. use fewer, but larger, ZIL blocks.
In theory, with these goals met, the new allgorithm will allow the
following improvements:
1. new ZIL blocks can be generated and issued, while there's already
oustanding ZIL blocks being serviced by the storage.
2. the latency of zil_commit() should be proportional to the underlying
storage latency, rather than the incoming synchronous workload.
Porting Notes
=============
Due to the changes made in commit 119a394ab0, the lifetime of an itx
structure differs than in OpenZFS. Specifically, the itx structure is
kept around until the data associated with the itx is considered to be
safe on disk; this is so that the itx's callback can be called after the
data is committed to stable storage. Since OpenZFS doesn't have this itx
callback mechanism, it's able to destroy the itx structure immediately
after the itx is committed to an lwb (before the lwb is written to
disk).
To support this difference, and to ensure the itx's callbacks can still
be called after the itx's data is on disk, a few changes had to be made:
* A list of itxs was added to the lwb structure. This list contains
all of the itxs that have been committed to the lwb, such that the
callbacks for these itxs can be called from zil_lwb_flush_vdevs_done(),
after the data for the itxs is committed to disk.
* A list of itxs was added on the stack of the zil_process_commit_list()
function; the "nolwb_itxs" list. In some circumstances, an itx may
not be committed to an lwb (e.g. if allocating the "next" ZIL block
on disk fails), so this list is used to keep track of which itxs
fall into this state, such that their callbacks can be called after
the ZIL's writer pipeline is "stalled".
* The logic to actually call the itx's callback was moved into the
zil_itx_destroy() function. Since all consumers of zil_itx_destroy()
were effectively performing the same logic (i.e. if callback is
non-null, call the callback), it seemed like useful code cleanup to
consolidate this logic into a single function.
Additionally, the existing Linux tracepoint infrastructure dealing with
the ZIL's probes and structures had to be updated to reflect these code
changes. Specifically:
* The "zil__cw1" and "zil__cw2" probes were removed, so they had to be
removed from "trace_zil.h" as well.
* Some of the zilog structure's fields were removed, which affected
the tracepoint definitions of the structure.
* New tracepoints had to be added for the following 3 new probes:
* zil__process__commit__itx
* zil__process__normal__itx
* zil__commit__io__error
OpenZFS-issue: https://www.illumos.org/issues/8585
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/5d95a3aCloses#6566
8558 lwp_create() returns EAGAIN on system with more than 80K ZFS filesystems
On a system with more than 80K ZFS filesystems, we've seen cases
where lwp_create() will start to fail by returning EAGAIN. The
problem being, for each of those 80K ZFS filesystems, a taskq will
be created for each dataset as part of the ZIL for each dataset.
Porting Notes:
- The new nomem taskq kstat was dropped.
- Added module options and documentation for new tunings
zfs_zil_clean_taskq_nthr_pct, zfs_zil_clean_taskq_minalloc,
zfs_zil_clean_taskq_maxalloc, and zfs_sync_taskq_batch_pct.
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Sebastien Roy <sebastien.roy@delphix.com>
Approved by: Robert Mustacchi <rm@joyent.com>
Authored by: Prakash Surya <prakash.surya@delphix.com>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Chris Dunlop <chris@onthe.net.au>
Ported-by: Brian Behlendorf <behlendorf1@llnl.gov>
OpenZFS-issue: https://www.illumos.org/issues/8558
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/216d772
8602 remove unused "dp_early_sync_tasks" field from "dsl_pool" structure
Reviewed by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Approved by: Robert Mustacchi <rm@joyent.com>
Authored by: Prakash Surya <prakash.surya@delphix.com>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Chris Dunlop <chris@onthe.net.au>
Ported-by: Brian Behlendorf <behlendorf1@llnl.gov>
OpenZFS-issue: https://www.illumos.org/issues/8602
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/2bcb545Closes#6779
- After some ZIL changes 6 years ago zil_slog_limit got partially broken
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.
Authored by: Alexander Motin <mav@FreeBSD.org>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: Andriy Gapon <avg@FreeBSD.org>
Reviewed by: Steven Hartland <steven.hartland@multiplay.co.uk>
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed by: Richard Elling <Richard.Elling@RichardElling.com>
Approved by: Robert Mustacchi <rm@joyent.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <ryao@gentoo.org>
Ported-by: Giuseppe Di Natale <dinatale2@llnl.gov>
OpenZFS-issue: https://www.illumos.org/issues/7578
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/aeb13acCloses#6191
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/5027https://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
The vast majority of these changes are in Linux specific code.
They are the result of not having an automated style checker to
validate the code when it was originally written. Others were
caused when the common code was slightly adjusted for Linux.
This patch contains no functional changes. It only refreshes
the code to conform to style guide.
Everyone submitting patches for inclusion upstream should now
run 'make checkstyle' and resolve any warning prior to opening
a pull request. The automated builders have been updated to
fail a build if when 'make checkstyle' detects an issue.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#1821
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
One of the neat tricks an autoconf style project is capable of
is allow configurion/building in a directory other than the
source directory. The major advantage to this is that you can
build the project various different ways while making changes
in a single source tree.
For example, this project is designed to work on various different
Linux distributions each of which work slightly differently. This
means that changes need to verified on each of those supported
distributions perferably before the change is committed to the
public git repo.
Using nfs and custom build directories makes this much easier.
I now have a single source tree in nfs mounted on several different
systems each running a supported distribution. When I make a
change to the source base I suspect may break things I can
concurrently build from the same source on all the systems each
in their own subdirectory.
wget -c http://github.com/downloads/behlendorf/zfs/zfs-x.y.z.tar.gz
tar -xzf zfs-x.y.z.tar.gz
cd zfs-x-y-z
------------------------- run concurrently ----------------------
<ubuntu system> <fedora system> <debian system> <rhel6 system>
mkdir ubuntu mkdir fedora mkdir debian mkdir rhel6
cd ubuntu cd fedora cd debian cd rhel6
../configure ../configure ../configure ../configure
make make make make
make check make check make check make check
This change also moves many of the include headers from individual
incude/sys directories under the modules directory in to a single
top level include directory. This has the advantage of making
the build rules cleaner and logically it makes a bit more sense.
Prakash Surya
Brian Behlendorf
Giuseppe Di Natale
Matthew Ahrens
Michael Kjorling
Etienne Dechamps