Fix regression seen in issue #11545 where checksum errors
where not being counted or showing up in a zpool event.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Don Brady <don.brady@delphix.com>
Closes#11609
Mixing ZIL and normal allocations has several problems:
1. The ZIL allocations are allocated, written to disk, and then a few
seconds later freed. This leaves behind holes (free segments) where the
ZIL blocks used to be, which increases fragmentation, which negatively
impacts performance.
2. When under moderate load, ZIL allocations are of 128KB. If the pool
is fairly fragmented, there may not be many free chunks of that size.
This causes ZFS to load more metaslabs to locate free segments of 128KB
or more. The loading happens synchronously (from zil_commit()), and can
take around a second even if the metaslab's spacemap is cached in the
ARC. All concurrent synchronous operations on this filesystem must wait
while the metaslab is loading. This can cause a significant performance
impact.
3. If the pool is very fragmented, there may be zero free chunks of
128KB or more. In this case, the ZIL falls back to txg_wait_synced(),
which has an enormous performance impact.
These problems can be eliminated by using a dedicated log device
("slog"), even one with the same performance characteristics as the
normal devices.
This change sets aside one metaslab from each top-level vdev that is
preferentially used for ZIL allocations (vdev_log_mg,
spa_embedded_log_class). From an allocation perspective, this is
similar to having a dedicated log device, and it eliminates the
above-mentioned performance problems.
Log (ZIL) blocks can be allocated from the following locations. Each
one is tried in order until the allocation succeeds:
1. dedicated log vdevs, aka "slog" (spa_log_class)
2. embedded slog metaslabs (spa_embedded_log_class)
3. other metaslabs in normal vdevs (spa_normal_class)
The space required for the embedded slog metaslabs is usually between
0.5% and 1.0% of the pool, and comes out of the existing 3.2% of "slop"
space that is not available for user data.
On an all-ssd system with 4TB storage, 87% fragmentation, 60% capacity,
and recordsize=8k, testing shows a ~50% performance increase on random
8k sync writes. On even more fragmented systems (which hit problem #3
above and call txg_wait_synced()), the performance improvement can be
arbitrarily large (>100x).
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Don Brady <don.brady@delphix.com>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#11389
The `abd_get_offset_*()` routines create an abd_t that references
another abd_t, and doesn't allocate any pages/buffers of its own. In
some workloads, these routines may be called frequently, to create many
abd_t's representing small pieces of a single large abd_t. In
particular, the upcoming RAIDZ Expansion project makes heavy use of
these routines.
This commit adds the ability for the caller to allocate and provide the
abd_t struct to a variant of `abd_get_offset_*()`. This eliminates the
cost of allocating the abd_t and performing the accounting associated
with it (`abdstat_struct_size`). The RAIDZ/DRAID code uses this for
the `rc_abd`, which references the zio's abd. The upcoming RAIDZ
Expansion project will leverage this infrastructure to increase
performance of reads post-expansion by around 50%.
Additionally, some of the interfaces around creating and destroying
abd_t's are cleaned up. Most significantly, the distinction between
`abd_put()` and `abd_free()` is eliminated; all types of abd_t's are
now disposed of with `abd_free()`.
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Issue #8853Closes#11439
On a system with very high fragmentation, we may need to do lots of gang
allocations (e.g. most indirect block allocations (~50KB) may need to
gang). Before failing a "normal" allocation and resorting to ganging, we
try every metaslab. This has the impact of loading every metaslab (not
a huge deal since we now typically keep all metaslabs loaded), and also
iterating over every metaslab for every failing allocation. If there are
many metaslabs (more than the typical ~200, e.g. due to vdev expansion
or very large vdevs), the CPU cost of this iteration can be very
impactful. This iteration is done with the mg_lock held, creating long
hold times and high lock contention for concurrent allocations,
ultimately causing long txg sync times and poor application performance.
To address this, this commit changes the behavior of "normal" (not
try_hard, not ZIL) allocations. These will now only examine the 100
best metaslabs (as determined by their ms_weight). If none of these
have a large enough free segment, then the allocation will fail and
we'll fall back on ganging.
To accomplish this, we will now (normally) gang before doing a
`try_hard` allocation. Non-try_hard allocations will only examine the
100 best metaslabs of each vdev. In summary, we will first try normal
allocation. If that fails then we will do a gang allocation. If that
fails then we will do a "try hard" gang allocation. If that fails then
we will have a multi-layer gang block.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#11327
Metaslab rotor and aliquot are used to distribute workload between
vdevs while keeping some locality for logically adjacent blocks. Once
multiple allocators were introduced to separate allocation of different
objects it does not make much sense for different allocators to write
into different metaslabs of the same metaslab group (vdev) same time,
competing for its resources. This change makes each allocator choose
metaslab group independently, colliding with others only sporadically.
Test including simultaneous write into 4 files with recordsize of 4KB
on a striped pool of 30 disks on a system with 40 logical cores show
reduction of vdev queue lock contention from 54 to 27% due to better
load distribution. Unfortunately it won't help much ZVOLs yet since
only one dataset/ZVOL is synced at a time, and so for the most part
only one allocator is used, but it may improve later.
While there, to reduce the number of pointer dereferences change
per-allocator storage for metaslab classes and groups from several
separate malloc()'s to variable length arrays at the ends of the
original class and group structures.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#11288
This patch adds a new top-level vdev type called dRAID, which stands
for Distributed parity RAID. This pool configuration allows all dRAID
vdevs to participate when rebuilding to a distributed hot spare device.
This can substantially reduce the total time required to restore full
parity to pool with a failed device.
A dRAID pool can be created using the new top-level `draid` type.
Like `raidz`, the desired redundancy is specified after the type:
`draid[1,2,3]`. No additional information is required to create the
pool and reasonable default values will be chosen based on the number
of child vdevs in the dRAID vdev.
zpool create <pool> draid[1,2,3] <vdevs...>
Unlike raidz, additional optional dRAID configuration values can be
provided as part of the draid type as colon separated values. This
allows administrators to fully specify a layout for either performance
or capacity reasons. The supported options include:
zpool create <pool> \
draid[<parity>][:<data>d][:<children>c][:<spares>s] \
<vdevs...>
- draid[parity] - Parity level (default 1)
- draid[:<data>d] - Data devices per group (default 8)
- draid[:<children>c] - Expected number of child vdevs
- draid[:<spares>s] - Distributed hot spares (default 0)
Abbreviated example `zpool status` output for a 68 disk dRAID pool
with two distributed spares using special allocation classes.
```
pool: tank
state: ONLINE
config:
NAME STATE READ WRITE CKSUM
slag7 ONLINE 0 0 0
draid2:8d:68c:2s-0 ONLINE 0 0 0
L0 ONLINE 0 0 0
L1 ONLINE 0 0 0
...
U25 ONLINE 0 0 0
U26 ONLINE 0 0 0
spare-53 ONLINE 0 0 0
U27 ONLINE 0 0 0
draid2-0-0 ONLINE 0 0 0
U28 ONLINE 0 0 0
U29 ONLINE 0 0 0
...
U42 ONLINE 0 0 0
U43 ONLINE 0 0 0
special
mirror-1 ONLINE 0 0 0
L5 ONLINE 0 0 0
U5 ONLINE 0 0 0
mirror-2 ONLINE 0 0 0
L6 ONLINE 0 0 0
U6 ONLINE 0 0 0
spares
draid2-0-0 INUSE currently in use
draid2-0-1 AVAIL
```
When adding test coverage for the new dRAID vdev type the following
options were added to the ztest command. These options are leverages
by zloop.sh to test a wide range of dRAID configurations.
-K draid|raidz|random - kind of RAID to test
-D <value> - dRAID data drives per group
-S <value> - dRAID distributed hot spares
-R <value> - RAID parity (raidz or dRAID)
The zpool_create, zpool_import, redundancy, replacement and fault
test groups have all been updated provide test coverage for the
dRAID feature.
Co-authored-by: Isaac Huang <he.huang@intel.com>
Co-authored-by: Mark Maybee <mmaybee@cray.com>
Co-authored-by: Don Brady <don.brady@delphix.com>
Co-authored-by: Matthew Ahrens <mahrens@delphix.com>
Co-authored-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mmaybee@cray.com>
Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#10102
It is a leftover from illumos always set to NULL and introducing a
spurious difference between zio_buf and zio_data_buf.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Closes#11188
Current CPU_SEQID users don't care about possibly changing CPU ID, but
enclose it within kpreempt disable/enable in order to fend off warnings
from Linux's CONFIG_DEBUG_PREEMPT.
There is no need to do it. The expected way to get CPU ID while allowing
for migration is to use raw_smp_processor_id.
In order to make this future-proof this patch keeps CPU_SEQID as is and
introduces CPU_SEQID_UNSTABLE instead, to make it clear that consumers
explicitly want this behavior.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Matt Macy <mmacy@FreeBSD.org>
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Closes#11142
Refer to the correct section or alternative for FreeBSD and Linux.
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes#11132
Duplicate io and checksum ereport events can misrepresent that
things are worse than they seem. Ideally the zpool events and the
corresponding vdev stat error counts in a zpool status should be
for unique errors -- not the same error being counted over and over.
This can be demonstrated in a simple example. With a single bad
block in a datafile and just 5 reads of the file we end up with a
degraded vdev, even though there is only one unique error in the pool.
The proposed solution to the above issue, is to eliminate duplicates
when posting events and when updating vdev error stats. We now save
recent error events of interest when posting events so that we can
easily check for duplicates when posting an error.
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Don Brady <don.brady@delphix.com>
Closes#10861
use (void) to silence analyzers.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Toomas Soome <tsoome@me.com>
Closes#10857
This PR adds two new compression types, based on ZStandard:
- zstd: A basic ZStandard compression algorithm Available compression.
Levels for zstd are zstd-1 through zstd-19, where the compression
increases with every level, but speed decreases.
- zstd-fast: A faster version of the ZStandard compression algorithm
zstd-fast is basically a "negative" level of zstd. The compression
decreases with every level, but speed increases.
Available compression levels for zstd-fast:
- zstd-fast-1 through zstd-fast-10
- zstd-fast-20 through zstd-fast-100 (in increments of 10)
- zstd-fast-500 and zstd-fast-1000
For more information check the man page.
Implementation details:
Rather than treat each level of zstd as a different algorithm (as was
done historically with gzip), the block pointer `enum zio_compress`
value is simply zstd for all levels, including zstd-fast, since they all
use the same decompression function.
The compress= property (a 64bit unsigned integer) uses the lower 7 bits
to store the compression algorithm (matching the number of bits used in
a block pointer, as the 8th bit was borrowed for embedded block
pointers). The upper bits are used to store the compression level.
It is necessary to be able to determine what compression level was used
when later reading a block back, so the concept used in LZ4, where the
first 32bits of the on-disk value are the size of the compressed data
(since the allocation is rounded up to the nearest ashift), was
extended, and we store the version of ZSTD and the level as well as the
compressed size. This value is returned when decompressing a block, so
that if the block needs to be recompressed (L2ARC, nop-write, etc), that
the same parameters will be used to result in the matching checksum.
All of the internal ZFS code ( `arc_buf_hdr_t`, `objset_t`,
`zio_prop_t`, etc.) uses the separated _compress and _complevel
variables. Only the properties ZAP contains the combined/bit-shifted
value. The combined value is split when the compression_changed_cb()
callback is called, and sets both objset members (os_compress and
os_complevel).
The userspace tools all use the combined/bit-shifted value.
Additional notes:
zdb can now also decode the ZSTD compression header (flag -Z) and
inspect the size, version and compression level saved in that header.
For each record, if it is ZSTD compressed, the parameters of the decoded
compression header get printed.
ZSTD is included with all current tests and new tests are added
as-needed.
Per-dataset feature flags now get activated when the property is set.
If a compression algorithm requires a feature flag, zfs activates the
feature when the property is set, rather than waiting for the first
block to be born. This is currently only used by zstd but can be
extended as needed.
Portions-Sponsored-By: The FreeBSD Foundation
Co-authored-by: Allan Jude <allanjude@freebsd.org>
Co-authored-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Sebastian Gottschall <s.gottschall@dd-wrt.com>
Co-authored-by: Kjeld Schouten-Lebbing <kjeld@schouten-lebbing.nl>
Co-authored-by: Michael Niewöhner <foss@mniewoehner.de>
Signed-off-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Allan Jude <allanjude@freebsd.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Sebastian Gottschall <s.gottschall@dd-wrt.com>
Signed-off-by: Kjeld Schouten-Lebbing <kjeld@schouten-lebbing.nl>
Signed-off-by: Michael Niewöhner <foss@mniewoehner.de>
Closes#6247Closes#9024Closes#10277Closes#10278
Mark functions used only in the same translation unit as static. This
only includes functions that do not have a prototype in a header file
either.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Arvind Sankar <nivedita@alum.mit.edu>
Closes#10470
The strcpy() and sprintf() functions are deprecated on some platforms.
Care is needed to ensure correct size is used. If some platforms
miss snprintf, we can add a #define to sprintf, likewise strlcpy().
The biggest change is adding a size parameter to zfs_id_to_fuidstr().
The various *_impl_get() functions are only used on linux and have
not yet been updated.
Reviewed by: Sean Eric Fagan <sef@ixsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Jorgen Lundman <lundman@lundman.net>
Closes#10400
Make the cityhash code compile into libzfs, in preparation for the new
"zstream" command.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10152
Also dprintf_bp() in case BLK_VERIFY_HALT of zfs_blkptr_verify_log()
since dprintf_bp() in zfs_blkptr_verify() will never be executed.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Paul Zuchowski <pzuchowski@datto.com>
Signed-off-by: Justin Keogh <commits@v6y.net>
Closes#10086
__zio_execute() calls zio_taskq_member() to determine if we are running
in a zio interrupt taskq, in which case we may need to switch to
processing this zio in a zio issue taskq. The call to
zio_taskq_member() can become a performance bottleneck when we are
processing a high rate of zio's.
zio_taskq_member() calls taskq_member() on each of the zio interrupt
taskqs, of which there are 21. This is slow because each call to
taskq_member() does tsd_get(taskq_tsd), which on Linux is relatively
slow.
This commit improves the performance of zio_taskq_member() by having it
cache the value of tsd_get(taskq_tsd), reducing the number of those
calls to 1/21th of the current behavior.
In a test case running `zfs send -c >/dev/null` of a filesystem with
small blocks (average 2.5KB/block), zio_taskq_member() was using 6.7% of
one CPU, and with this change it is reduced to 1.3%. Overall time to
perform the `zfs send` reduced by 10% (~150,000 block/sec to ~165,000
blocks/sec).
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10070
When "zfs destroy" is run, it completes quickly, and in the background
we locate the blocks to free and free them. This background activity
can be observed with `zpool get freeing` and `zpool wait -t free ...`.
This background activity is processed by a single thread (the spa_sync
thread) which calls zio_free() on each of the blocks to free. With even
modest storage performance, the CPU consumption of zio_free() can be the
performance bottleneck.
Performance of zio_free() can be improved by not actually creating a
zio_t in the common case (non-dedup, non-gang), instead calling
metaslab_free() directly. This avoids the CPU cost of allocating the
zio_t, and more importantly the cost of adding and later removing this
zio_t from the parent zio's child list.
The result is that performance of background freeing more than doubles,
from 0.6 million blocks per second to 1.3 million blocks per second.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10034
zdb -R :b fails due to the indirect block being compressed,
and the 'b' and 'd' flag not working in tandem when specified.
Fix the flag parsing code and create a zfs test for zdb -R
block display. Also fix the zio flags where the dotted notation
for the vdev portion of DVA (i.e. 0.0:offset:length) fails.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Zuchowski <pzuchowski@datto.com>
Closes#9640Closes#9729
With recent SPL changes there is no longer any need for a per
platform version.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#9860
This change allows us to align the code dump logic across platforms.
Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Don Brady <don.brady@delphix.com>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#9691
Remove the ASSERTV macro and handle suppressing unused
compiler warnings for variables only in ASSERTs using the
__attribute__((unused)) compiler annotation. The annotation
is understood by both gcc and clang.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#9671
This change modifies some of the infrastructure for enabling the use of
the DTRACE_PROBE* macros, such that we can use tehm in the "spl" module.
Currently, when the DTRACE_PROBE* macros are used, they get expanded to
create new functions, and these dynamically generated functions become
part of the "zfs" module.
Since the "spl" module does not depend on the "zfs" module, the use of
DTRACE_PROBE* in the "spl" module would result in undefined symbols
being used in the "spl" module. Specifically, DTRACE_PROBE* would turn
into a function call, and the function being called would be a symbol
only contained in the "zfs" module; which results in a linker and/or
runtime error.
Thus, this change adds the necessary logic to the "spl" module, to
mirror the tracing functionality available to the "zfs" module. After
this change, we'll have a "trace_zfs.h" header file which defines the
probes available only to the "zfs" module, and a "trace_spl.h" header
file which defines the probes available only to the "spl" module.
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Prakash Surya <prakash.surya@delphix.com>
Closes#9525
FreeBSD has its own implementation as do other platforms.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#9439
Move Linux specific tracing headers and source to platform directories
and update the build system.
Reviewed-by: Allan Jude <allanjude@freebsd.org>
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#9290
Adds ZFS_MODULE_PARAM to abstract module parameter
setting to operating systems other than Linux.
Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
Closes#9230
When we check the vdev of the blkptr in zfs_blkptr_verify, we can run
into a race condition where that vdev is temporarily unavailable. This
happens when a device removal operation and the old vdev_t has been
removed from the array, but the new indirect vdev has not yet been
inserted.
We hold the spa_config_lock while doing our sensitive verification.
To ensure that we don't deadlock, we only grab the lock if we don't
have config_writer held. In addition, I had to const the tags of the
refcounts and the spa_config_lock arguments.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes#9112
= Motivation
At Delphix we've seen a lot of customer systems where fragmentation
is over 75% and random writes take a performance hit because a lot
of time is spend on I/Os that update on-disk space accounting metadata.
Specifically, we seen cases where 20% to 40% of sync time is spend
after sync pass 1 and ~30% of the I/Os on the system is spent updating
spacemaps.
The problem is that these pools have existed long enough that we've
touched almost every metaslab at least once, and random writes
scatter frees across all metaslabs every TXG, thus appending to
their spacemaps and resulting in many I/Os. To give an example,
assuming that every VDEV has 200 metaslabs and our writes fit within
a single spacemap block (generally 4K) we have 200 I/Os. Then if we
assume 2 levels of indirection, we need 400 additional I/Os and
since we are talking about metadata for which we keep 2 extra copies
for redundancy we need to triple that number, leading to a total of
1800 I/Os per VDEV every TXG.
We could try and decrease the number of metaslabs so we have less
I/Os per TXG but then each metaslab would cover a wider range on
disk and thus would take more time to be loaded in memory from disk.
In addition, after it's loaded, it's range tree would consume more
memory.
Another idea would be to just increase the spacemap block size
which would allow us to fit more entries within an I/O block
resulting in fewer I/Os per metaslab and a speedup in loading time.
The problem is still that we don't deal with the number of I/Os
going up as the number of metaslabs is increasing and the fact
is that we generally write a lot to a few metaslabs and a little
to the rest of them. Thus, just increasing the block size would
actually waste bandwidth because we won't be utilizing our bigger
block size.
= About this patch
This patch introduces the Log Spacemap project which provides the
solution to the above problem while taking into account all the
aforementioned tradeoffs. The details on how it achieves that can
be found in the references sections below and in the code (see
Big Theory Statement in spa_log_spacemap.c).
Even though the change is fairly constraint within the metaslab
and lower-level SPA codepaths, there is a side-change that is
user-facing. The change is that VDEV IDs from VDEV holes will no
longer be reused. To give some background and reasoning for this,
when a log device is removed and its VDEV structure was replaced
with a hole (or was compacted; if at the end of the vdev array),
its vdev_id could be reused by devices added after that. Now
with the pool-wide space maps recording the vdev ID, this behavior
can cause problems (e.g. is this entry referring to a segment in
the new vdev or the removed log?). Thus, to simplify things the
ID reuse behavior is gone and now vdev IDs for top-level vdevs
are truly unique within a pool.
= Testing
The illumos implementation of this feature has been used internally
for a year and has been in production for ~6 months. For this patch
specifically there don't seem to be any regressions introduced to
ZTS and I have been running zloop for a week without any related
problems.
= Performance Analysis (Linux Specific)
All performance results and analysis for illumos can be found in
the links of the references. Redoing the same experiments in Linux
gave similar results. Below are the specifics of the Linux run.
After the pool reached stable state the percentage of the time
spent in pass 1 per TXG was 64% on average for the stock bits
while the log spacemap bits stayed at 95% during the experiment
(graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png).
Sync times per TXG were 37.6 seconds on average for the stock
bits and 22.7 seconds for the log spacemap bits (related graph:
sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result
the log spacemap bits were able to push more TXGs, which is also
the reason why all graphs quantified per TXG have more entries for
the log spacemap bits.
Another interesting aspect in terms of txg syncs is that the stock
bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8,
and 20% reach 9. The log space map bits reached sync pass 4 in 79%
of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This
emphasizes the fact that not only we spend less time on metadata
but we also iterate less times to convergence in spa_sync() dirtying
objects.
[related graphs:
stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png
lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png]
Finally, the improvement in IOPs that the userland gains from the
change is approximately 40%. There is a consistent win in IOPS as
you can see from the graphs below but the absolute amount of
improvement that the log spacemap gives varies within each minute
interval.
sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png
sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png
= Porting to Other Platforms
For people that want to port this commit to other platforms below
is a list of ZoL commits that this patch depends on:
Make zdb results for checkpoint tests consistent
db587941c5
Update vdev_is_spacemap_addressable() for new spacemap encoding
419ba59145
Simplify spa_sync by breaking it up to smaller functions
8dc2197b7b
Factor metaslab_load_wait() in metaslab_load()
b194fab0fb
Rename range_tree_verify to range_tree_verify_not_present
df72b8bebe
Change target size of metaslabs from 256GB to 16GB
c853f382db
zdb -L should skip leak detection altogether
21e7cf5da8
vs_alloc can underflow in L2ARC vdevs
7558997d2f
Simplify log vdev removal code
6c926f426a
Get rid of space_map_update() for ms_synced_length
425d3237ee
Introduce auxiliary metaslab histograms
928e8ad47d
Error path in metaslab_load_impl() forgets to drop ms_sync_lock
8eef997679
= References
Background, Motivation, and Internals of the Feature
- OpenZFS 2017 Presentation:
youtu.be/jj2IxRkl5bQ
- Slides:
slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project
Flushing Algorithm Internals & Performance Results
(Illumos Specific)
- Blogpost:
sdimitro.github.io/post/zfs-lsm-flushing/
- OpenZFS 2018 Presentation:
youtu.be/x6D2dHRjkxw
- Slides:
slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm
Upstream Delphix Issues:
DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320
DLPX-63385
Reviewed-by: Sean Eric Fagan <sef@ixsystems.com>
Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes#8442
After device removal, performing nopwrites on a dmu_sync-ed block
will result in a panic. This panic can show up in two ways:
1. an attempt to issue an IOCTL in vdev_indirect_io_start()
2. a failed comparison of zio->io_bp and zio->io_bp_orig in
zio_done()
To resolve both of these panics, nopwrites of blocks on indirect
vdevs should be ignored and new allocations should be performed on
concrete vdevs.
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Don Brady <don.brady@delphix.com>
Signed-off-by: George Wilson <gwilson@delphix.com>
Closes#8957
With the addition of BP_EMBEDDED_TYPE_REDACTED in 30af21b0 a couple of
codepaths make wrong assumptions and could potentially result in errors.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Chris Dunlop <chris@onthe.net.au>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: loli10K <ezomori.nozomu@gmail.com>
Closes#8951
The logic to handle strong checksum collisions where the data doesn't
match is incorrect. It is not clearing the dedup bit of the blkptr,
which can cause a panic later in zio_ddt_free() due to the dedup table
not matching what is in the blkptr.
Reviewed-by: Tom Caputi <tcaputi@datto.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
External-issue: DLPX-48097
Closes#8936
If dedup is in use, the `dedupditto` property can be set, causing ZFS to
keep an extra copy of data that is referenced many times (>100x). The
idea was that this data is more important than other data and thus we
want to be really sure that it is not lost if the disk experiences a
small amount of random corruption.
ZFS (and system administrators) rely on the pool-level redundancy to
protect their data (e.g. mirroring or RAIDZ). Since the user/sysadmin
doesn't have control over what data will be offered extra redundancy by
dedupditto, this extra redundancy is not very useful. The bulk of the
data is still vulnerable to loss based on the pool-level redundancy.
For example, if particle strikes corrupt 0.1% of blocks, you will either
be saved by mirror/raidz, or you will be sad. This is true even if
dedupditto saved another 0.01% of blocks from being corrupted.
Therefore, the dedupditto functionality is rarely enabled (i.e. the
property is rarely set), and it fulfills its promise of increased
redundancy even more rarely.
Additionally, this feature does not work as advertised (on existing
releases), because scrub/resilver did not repair the extra (dedupditto)
copy (see https://github.com/zfsonlinux/zfs/pull/8270).
In summary, this seldom-used feature doesn't work, and even if it did it
wouldn't provide useful data protection. It has a non-trivial
maintenance burden (again see https://github.com/zfsonlinux/zfs/pull/8270).
We should remove the dedupditto functionality. For backwards
compatibility with the existing CLI, "zpool set dedupditto" will still
"succeed" (exit code zero), but won't have any effect. For backwards
compatibility with existing pools that had dedupditto enabled at some
point, the code will still be able to understand dedupditto blocks and
free them when appropriate. However, ZFS won't write any new dedupditto
blocks.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Alek Pinchuk <apinchuk@datto.com>
Issue #8270Closes#8310
Redacted send/receive allows users to send subsets of their data to
a target system. One possible use case for this feature is to not
transmit sensitive information to a data warehousing, test/dev, or
analytics environment. Another is to save space by not replicating
unimportant data within a given dataset, for example in backup tools
like zrepl.
Redacted send/receive is a three-stage process. First, a clone (or
clones) is made of the snapshot to be sent to the target. In this
clone (or clones), all unnecessary or unwanted data is removed or
modified. This clone is then snapshotted to create the "redaction
snapshot" (or snapshots). Second, the new zfs redact command is used
to create a redaction bookmark. The redaction bookmark stores the
list of blocks in a snapshot that were modified by the redaction
snapshot(s). Finally, the redaction bookmark is passed as a parameter
to zfs send. When sending to the snapshot that was redacted, the
redaction bookmark is used to filter out blocks that contain sensitive
or unwanted information, and those blocks are not included in the send
stream. When sending from the redaction bookmark, the blocks it
contains are considered as candidate blocks in addition to those
blocks in the destination snapshot that were modified since the
creation_txg of the redaction bookmark. This step is necessary to
allow the target to rehydrate data in the case where some blocks are
accidentally or unnecessarily modified in the redaction snapshot.
The changes to bookmarks to enable fast space estimation involve
adding deadlists to bookmarks. There is also logic to manage the
life cycles of these deadlists.
The new size estimation process operates in cases where previously
an accurate estimate could not be provided. In those cases, a send
is performed where no data blocks are read, reducing the runtime
significantly and providing a byte-accurate size estimate.
Reviewed-by: Dan Kimmel <dan.kimmel@delphix.com>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Prashanth Sreenivasa <pks@delphix.com>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: George Wilson <george.wilson@delphix.com>
Reviewed-by: Chris Williamson <chris.williamson@delphix.com>
Reviewed-by: Pavel Zhakarov <pavel.zakharov@delphix.com>
Reviewed-by: Sebastien Roy <sebastien.roy@delphix.com>
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes#7958
Starting in sync pass 5 (zfs_sync_pass_dont_compress), we disable
compression (including of metadata). Ostensibly this helps the sync
passes to converge (i.e. for a sync pass to not need to allocate
anything because it is 100% overwrites).
However, in practice it increases the average number of sync passes,
because when we turn compression off, a lot of block's size will change
and thus we have to re-allocate (not overwrite) them. It also increases
the number of 128KB allocations (e.g. for indirect blocks and spacemaps)
because these will not be compressed. The 128K allocations are
especially detrimental to performance on highly fragmented systems,
which may have very few free segments of this size, and may need to load
new metaslabs to satisfy 128K allocations.
We should increase zfs_sync_pass_dont_compress. In practice on a highly
fragmented system we see a few 5-pass txg's, a tiny number of 6-pass
txg's, and no txg's with more than 6 passes.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Elling <Richard.Elling@RichardElling.com>
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed-by: George Wilson <george.wilson@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
External-issue: DLPX-63431
Closes#8892
Scatter ABD's are allocated from a number of pages. In contrast to
linear ABD's, these pages are disjoint in the kernel's virtual address
space, so they can't be accessed as a contiguous buffer. Therefore
routines that need a linear buffer (e.g. abd_borrow_buf() and friends)
must allocate a separate linear buffer (with zio_buf_alloc()), and copy
the contents of the pages to/from the linear buffer. This can have a
measurable performance overhead on some workloads.
https://github.com/zfsonlinux/zfs/commit/87c25d567fb7969b44c7d8af63990e
("abd_alloc should use scatter for >1K allocations") increased the use
of scatter ABD's, specifically switching 1.5K through 4K (inclusive)
buffers from linear to scatter. For workloads that access blocks whose
compressed sizes are in this range, that commit introduced an additional
copy into the read code path. For example, the
sequential_reads_arc_cached tests in the test suite were reduced by
around 5% (this is doing reads of 8K-logical blocks, compressed to 3K,
which are cached in the ARC).
This commit treats single-chunk scattered buffers as linear buffers,
because they are contiguous in the kernel's virtual address space.
All single-page (4K) ABD's can be represented this way. Some multi-page
ABD's can also be represented this way, if we were able to allocate a
single "chunk" (higher-order "page" which represents a power-of-2 series
of physically-contiguous pages). This is often the case for 2-page (8K)
ABD's.
Representing a single-entry scatter ABD as a linear ABD has the
performance advantage of avoiding the copy (and allocation) in
abd_borrow_buf_copy / abd_return_buf_copy. A performance increase of
around 5% has been observed for ARC-cached reads (of small blocks which
can take advantage of this), fixing the regression introduced by
87c25d567.
Note that this optimization is only possible because all physical memory
is always mapped into the kernel's address space. This is not the case
for HIGHMEM pages, so the optimization can not be made on 32-bit
systems.
Reviewed-by: Chunwei Chen <tuxoko@gmail.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#8580
There are several places where we use zfs_dbgmsg and %p to
print pointers. In the Linux kernel, these values obfuscated
to prevent information leaks which means the pointers aren't
very useful for debugging crash dumps. We decided to restrict
the permissions of dbgmsg (and some other kstats while we were
at it) and print pointers with %px in zfs_dbgmsg as well as
spl_dumpstack
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Gallagher <john.gallagher@delphix.com>
Signed-off-by: sara hartse <sara.hartse@delphix.com>
Closes#8467Closes#8476
UNMAP/TRIM support is a frequently-requested feature to help
prevent performance from degrading on SSDs and on various other
SAN-like storage back-ends. By issuing UNMAP/TRIM commands for
sectors which are no longer allocated the underlying device can
often more efficiently manage itself.
This TRIM implementation is modeled on the `zpool initialize`
feature which writes a pattern to all unallocated space in the
pool. The new `zpool trim` command uses the same vdev_xlate()
code to calculate what sectors are unallocated, the same per-
vdev TRIM thread model and locking, and the same basic CLI for
a consistent user experience. The core difference is that
instead of writing a pattern it will issue UNMAP/TRIM commands
for those extents.
The zio pipeline was updated to accommodate this by adding a new
ZIO_TYPE_TRIM type and associated spa taskq. This new type makes
is straight forward to add the platform specific TRIM/UNMAP calls
to vdev_disk.c and vdev_file.c. These new ZIO_TYPE_TRIM zios are
handled largely the same way as ZIO_TYPE_READs or ZIO_TYPE_WRITEs.
This makes it possible to largely avoid changing the pipieline,
one exception is that TRIM zio's may exceed the 16M block size
limit since they contain no data.
In addition to the manual `zpool trim` command, a background
automatic TRIM was added and is controlled by the 'autotrim'
property. It relies on the exact same infrastructure as the
manual TRIM. However, instead of relying on the extents in a
metaslab's ms_allocatable range tree, a ms_trim tree is kept
per metaslab. When 'autotrim=on', ranges added back to the
ms_allocatable tree are also added to the ms_free tree. The
ms_free tree is then periodically consumed by an autotrim
thread which systematically walks a top level vdev's metaslabs.
Since the automatic TRIM will skip ranges it considers too small
there is value in occasionally running a full `zpool trim`. This
may occur when the freed blocks are small and not enough time
was allowed to aggregate them. An automatic TRIM and a manual
`zpool trim` may be run concurrently, in which case the automatic
TRIM will yield to the manual TRIM.
Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Reviewed-by: Tim Chase <tim@chase2k.com>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: George Wilson <george.wilson@delphix.com>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Contributions-by: Saso Kiselkov <saso.kiselkov@nexenta.com>
Contributions-by: Tim Chase <tim@chase2k.com>
Contributions-by: Chunwei Chen <tuxoko@gmail.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#8419Closes#598
The number of IO and checksum events should match the number of errors
seen in zpool status. Previously there was a mismatch between the
two counts because zpool status would only count unrecovered errors,
while zpool events would get an event for *all* errors (recovered or
not). This lead to situations where disks could be faulted for
"too many errors", while at the same time showing zero errors in zpool
status.
This fixes the zpool status error counters to increment at the same
times we post the error events.
Reviewed-by: Tom Caputi <tcaputi@datto.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Closes#4851Closes#7817
If we hit the (NSEC_TO_TICK(diff) == 0) condition in
zio_delay_interrupt, zio_interrupt is never called and the
zio does not progress.
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: sara hartse <sara.hartse@delphix.com>
Closes#8404
Add the zio_deadman_log_all tunable to print all zios in
zio_deadman_impl(). Also, in all cases, display the depth of the
zio relative to the original parent zio. This is meant to be used by
developers to gain diagnostic information for hangs which don't involve
fully set-up zio trees or are otherwise stuck or hung in an early stage.
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: loli10K <ezomori.nozomu@gmail.com>
Signed-off-by: Tim Chase <tim@chase2k.com>
Closes#8362
Some minor spelling mistakes and typos. No functional changes.
Reviewed-by: Neal Gompa <ngompa@datto.com>
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Giuseppe Di Natale <guss80@gmail.com>
Reviewed-by: bunder2015 <omfgbunder@gmail.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#8272
This patch adds a new slow I/Os (-s) column to zpool status to show the
number of VDEV slow I/Os. This is the number of I/Os that didn't
complete in zio_slow_io_ms milliseconds. It also adds a new parsable
(-p) flag to display exact values.
NAME STATE READ WRITE CKSUM SLOW
testpool ONLINE 0 0 0 -
mirror-0 ONLINE 0 0 0 -
loop0 ONLINE 0 0 0 20
loop1 ONLINE 0 0 0 0
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Closes#7756Closes#6885
Recent changes in the Linux kernel made it necessary to prefix
the refcount_add() function with zfs_ due to a name collision.
To bring the other functions in line with that and to avoid future
collisions, prefix the other refcount functions as well.
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tim Schumacher <timschumi@gmx.de>
Closes#7963
Allocation Classes add the ability to have allocation classes in a
pool that are dedicated to serving specific block categories, such
as DDT data, metadata, and small file blocks. A pool can opt-in to
this feature by adding a 'special' or 'dedup' top-level VDEV.
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Alek Pinchuk <apinchuk@datto.com>
Reviewed-by: Håkan Johansson <f96hajo@chalmers.se>
Reviewed-by: Andreas Dilger <andreas.dilger@chamcloud.com>
Reviewed-by: DHE <git@dehacked.net>
Reviewed-by: Richard Elling <Richard.Elling@RichardElling.com>
Reviewed-by: Gregor Kopka <gregor@kopka.net>
Reviewed-by: Kash Pande <kash@tripleback.net>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Don Brady <don.brady@delphix.com>
Closes#5182
Assertion failed in arc_buf_destroy() when concurrently reading
block with checksum error.
Porting notes:
* The ability to zinject decompression errors has been added, but
this only works at the zio_decompress() level, where we have all
of the info we need to match against the user's zinject options.
* The decompress_fault test has been added to test the new zinject
functionality
* We attempted to set zio_decompress_fail_fraction to (1 << 18) in
ztest for further test coverage. Although this did uncover a few
low priority issues, this unfortuantely also causes ztest to
ASSERT in many locations where the code is working correctly since
it is designed to fail on IO errors. Developers can manually set
this variable with the '-o' option to find and debug issues.
Authored by: Matt Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Paul Dagnelie <pcd@delphix.com>
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Approved by: Matt Ahrens <mahrens@delphix.com>
Ported-by: Tom Caputi <tcaputi@datto.com>
OpenZFS-issue: https://illumos.org/issues/9403
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/fa98e487a9Closes#7822
When doing a read from disk, ZFS creates 3 ZIO's: a zio_null(), the
logical zio_read(), and then a physical zio. Currently, each of these
results in a separate taskq_dispatch(zio_execute).
On high-read-iops workloads, this causes a significant performance
impact. By processing all 3 ZIO's in a single taskq entry, we reduce the
overhead on taskq locking and context switching. We accomplish this by
allowing zio_done() to return a "next zio to execute" to zio_execute().
This results in a ~12% performance increase for random reads, from
96,000 iops to 108,000 iops (with recordsize=8k, on SSD's).
Reviewed by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed by: George Wilson <george.wilson@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
External-issue: DLPX-59292
Closes#7736
Overview
========
We parallelize the allocation process by creating the concept of
"allocators". There are a certain number of allocators per metaslab
group, defined by the value of a tunable at pool open time. Each
allocator for a given metaslab group has up to 2 active metaslabs; one
"primary", and one "secondary". The primary and secondary weight mean
the same thing they did in in the pre-allocator world; primary metaslabs
are used for most allocations, secondary metaslabs are used for ditto
blocks being allocated in the same metaslab group. There is also the
CLAIM weight, which has been separated out from the other weights, but
that is less important to understanding the patch. The active metaslabs
for each allocator are moved from their normal place in the metaslab
tree for the group to the back of the tree. This way, they will not be
selected for use by other allocators searching for new metaslabs unless
all the passive metaslabs are unsuitable for allocations. If that does
happen, the allocators will "steal" from each other to ensure that IOs
don't fail until there is truly no space left to perform allocations.
In addition, the alloc queue for each metaslab group has been broken
into a separate queue for each allocator. We don't want to dramatically
increase the number of inflight IOs on low-end systems, because it can
significantly increase txg times. On the other hand, we want to ensure
that there are enough IOs for each allocator to allow for good
coalescing before sending the IOs to the disk. As a result, we take a
compromise path; each allocator's alloc queue max depth starts at a
certain value for every txg. Every time an IO completes, we increase the
max depth. This should hopefully provide a good balance between the two
failure modes, while not dramatically increasing complexity.
We also parallelize the spa_alloc_tree and spa_alloc_lock, which cause
very similar contention when selecting IOs to allocate. This
parallelization uses the same allocator scheme as metaslab selection.
Performance Results
===================
Performance improvements from this change can vary significantly based
on the number of CPUs in the system, whether or not the system has a
NUMA architecture, the speed of the drives, the values for the various
tunables, and the workload being performed. For an fio async sequential
write workload on a 24 core NUMA system with 256 GB of RAM and 8 128 GB
SSDs, there is a roughly 25% performance improvement.
Future Work
===========
Analysis of the performance of the system with this patch applied shows
that a significant new bottleneck is the vdev disk queues, which also
need to be parallelized. Prototyping of this change has occurred, and
there was a performance improvement, but more work needs to be done
before its stability has been verified and it is ready to be upstreamed.
Authored by: Paul Dagnelie <pcd@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Serapheim Dimitropoulos <serapheim.dimitro@delphix.com>
Reviewed by: Alexander Motin <mav@FreeBSD.org>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Approved by: Gordon Ross <gwr@nexenta.com>
Ported-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Porting Notes:
* Fix reservation test failures by increasing tolerance.
OpenZFS-issue: https://illumos.org/issues/9112
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/3f3cc3c3Closes#7682
Details about the motivation of this feature and its usage can
be found in this blogpost:
https://sdimitro.github.io/post/zpool-checkpoint/
A lightning talk of this feature can be found here:
https://www.youtube.com/watch?v=fPQA8K40jAM
Implementation details can be found in big block comment of
spa_checkpoint.c
Side-changes that are relevant to this commit but not explained
elsewhere:
* renames members of "struct metaslab trees to be shorter without
losing meaning
* space_map_{alloc,truncate}() accept a block size as a
parameter. The reason is that in the current state all space
maps that we allocate through the DMU use a global tunable
(space_map_blksz) which defauls to 4KB. This is ok for metaslab
space maps in terms of bandwirdth since they are scattered all
over the disk. But for other space maps this default is probably
not what we want. Examples are device removal's vdev_obsolete_sm
or vdev_chedkpoint_sm from this review. Both of these have a
1:1 relationship with each vdev and could benefit from a bigger
block size.
Porting notes:
* The part of dsl_scan_sync() which handles async destroys has
been moved into the new dsl_process_async_destroys() function.
* Remove "VERIFY(!(flags & FWRITE))" in "kernel.c" so zhack can write
to block device backed pools.
* ZTS:
* Fix get_txg() in zpool_sync_001_pos due to "checkpoint_txg".
* Don't use large dd block sizes on /dev/urandom under Linux in
checkpoint_capacity.
* Adopt Delphix-OS's setting of 4 (spa_asize_inflation =
SPA_DVAS_PER_BP + 1) for the checkpoint_capacity test to speed
its attempts to fill the pool
* Create the base and nested pools with sync=disabled to speed up
the "setup" phase.
* Clear labels in test pool between checkpoint tests to avoid
duplicate pool issues.
* The import_rewind_device_replaced test has been marked as "known
to fail" for the reasons listed in its DISCLAIMER.
* New module parameters:
zfs_spa_discard_memory_limit,
zfs_remove_max_bytes_pause (not documented - debugging only)
vdev_max_ms_count (formerly metaslabs_per_vdev)
vdev_min_ms_count
Authored by: Serapheim Dimitropoulos <serapheim.dimitro@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: John Kennedy <john.kennedy@delphix.com>
Reviewed by: Dan Kimmel <dan.kimmel@delphix.com>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Approved by: Richard Lowe <richlowe@richlowe.net>
Ported-by: Tim Chase <tim@chase2k.com>
Signed-off-by: Tim Chase <tim@chase2k.com>
OpenZFS-issue: https://illumos.org/issues/9166
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/7159fdb8Closes#7570
