]> granicus.if.org Git - zfs/log
zfs
5 years agoFix typos in man/
Andrea Gelmini [Fri, 30 Aug 2019 16:41:35 +0000 (18:41 +0200)]
Fix typos in man/

Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Andrea Gelmini <andrea.gelmini@gelma.net>
Closes #9233

5 years agoFix typos in config/
Andrea Gelmini [Fri, 30 Aug 2019 16:40:30 +0000 (18:40 +0200)]
Fix typos in config/

Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Andrea Gelmini <andrea.gelmini@gelma.net>
Closes #9232

5 years agoFix refquota_007_neg.ksh
Igor K [Fri, 30 Aug 2019 16:32:25 +0000 (19:32 +0300)]
Fix refquota_007_neg.ksh

Must use 'zfs' instead of '$ZFS' which is undefined.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Igor Kozhukhov <igor@dilos.org>
Closes #9257

5 years agoPrevent metaslab_sync panic due to spa_final_dirty_txg
Paul Dagnelie [Fri, 30 Aug 2019 16:28:31 +0000 (09:28 -0700)]
Prevent metaslab_sync panic due to spa_final_dirty_txg

If a pool enables the SPACEMAP_HISTOGRAM feature shortly before being
exported, we can enter a situation that causes a kernel panic. Any metaslabs
that are loaded during the final dirty txg and haven't already been condensed
will cause metaslab_sync to proceed after the final dirty txg so that the
condense can be performed, which there are assertions to prevent. Because of
the nature of this issue, there are a number of ways we can enter this
state. Rather than try to prevent each of them one by one, potentially missing
some edge cases, we instead cut it off at the point of intersection; by
preventing metaslab_sync from proceeding if it would only do so to perform a
condense and we're past the final dirty txg, we preserve the utility of the
existing asserts while preventing this particular issue.

Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes #9185
Closes #9186
Closes #9231
Closes #9253

5 years agoetc/init.d/zfs-functions.in: remove arch warning
Georgy Yakovlev [Thu, 29 Aug 2019 20:14:48 +0000 (12:14 -0800)]
etc/init.d/zfs-functions.in: remove arch warning

Remove the x86_64 warning, it's no longer the case that this is the
only supported architecture.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Georgy Yakovlev <gyakovlev@gentoo.org>
Closes: #9177
5 years agoSimplify deleting partitions in libtest
Ryan Moeller [Thu, 29 Aug 2019 20:11:29 +0000 (16:11 -0400)]
Simplify deleting partitions in libtest

Eliminate unnecessary code duplication. We can use a for-loop instead
of a while-loop. There is no need to echo $DISKSARRAY in a subshell or
return 0. Declare all variables with typeset.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
Closes #9224

5 years agoUse compatible arg order in tests
Ryan Moeller [Thu, 29 Aug 2019 18:03:09 +0000 (14:03 -0400)]
Use compatible arg order in tests

BSD getopt() and getopt_long() want options before arguments.
Reorder arguments to zfs/zpool in tests to put all the options first.

Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
Closes #9228

5 years agoKeep more metaslabs loaded
Paul Dagnelie [Thu, 29 Aug 2019 17:20:36 +0000 (10:20 -0700)]
Keep more metaslabs loaded

With the other metaslab changes loaded onto a system, we can
significantly reduce the memory usage of each loaded metaslab and
unload them on demand if there is memory pressure. However, none
of those changes actually result in us keeping more metaslabs loaded.
If we don't keep more metaslabs loaded, we will still have to wait
for demand-loading to finish when no loaded metaslab can satisfy our
allocation, which can cause ZIL performance issues. In addition,
performance is traditionally measured by IOs per unit time, while
unloading is currently done on a txg-count basis. Txgs can take a
widely varying range of times, from tenths of a second to several
seconds. This can result in confusing, hard to predict behavior.

This change simply adds a time-based component to metaslab unloading.
A metaslab will remain loaded for one minute and 8 txgs (by default)
after it was last used, unless it is evicted due to memory pressure.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
External-issue: DLPX-65016
External-issue: DLPX-65047
Closes #9197

5 years agozfs_handle used after being closed/freed in change_one callback
Pavel Zakharov [Wed, 28 Aug 2019 22:02:58 +0000 (18:02 -0400)]
zfs_handle used after being closed/freed in change_one callback

This is a typical case of use after free. We would call zfs_close(zhp)
which would free the handle, and then call zfs_iter_children() on that
handle later.  This change ensures that the zfs_handle is only closed
when we are ready to return.

Running `zfs inherit -r sharenfs pool` was failing with an error
code without any error messages. After some debugging I've pinpointed
the issue to be memory corruption, which would cause zfs to try to
issue an ioctl to the wrong device and receive ENOTTY.

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Sebastien Roy <sebastien.roy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alek Pinchuk <apinchuk@datto.com>
Signed-off-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Issue #7967
Closes #9165

5 years agoUse smaller default slack/delta value for schedule_hrtimeout_range()
Tony Nguyen [Wed, 28 Aug 2019 21:56:54 +0000 (15:56 -0600)]
Use smaller default slack/delta value for schedule_hrtimeout_range()

For interrupt coalescing, cv_timedwait_hires() uses a 100us slack/delta
for calls to schedule_hrtimeout_range(). This 100us slack can be costly
for small writes.

This change improves small write performance by passing resolution `res`
parameter to schedule_hrtimeout_range() to be used as delta/slack. A new
tunable `spl_schedule_hrtimeout_slack_us` is added to preserve old
behavior when desired.

Performance observations on 8K recordsize filesystem:
- 8K random writes at 1-64 threads, up to 60% improvement for one thread
  and smaller gains as thread count increases. At >64 threads, 2-5%
  decrease in performance was observed.
- 8K sequential writes, similar 60% improvement for one thread and
  leveling out around 64 threads. At >64 threads, 5-10% decrease in
  performance was observed.
- 128K sequential write sees 1-5 for the 128K. No observed regression at
  high thread count.

Testing done on Ubuntu 18.04 with 4.15 kernel, 8vCPUs and SSD storage on
VMware ESX.

Reviewed-by: Richard Elling <Richard.Elling@RichardElling.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Signed-off-by: Tony Nguyen <tony.nguyen@delphix.com>
Closes #9217

5 years agoZTS: Temporarily disable several upgrade tests
Brian Behlendorf [Wed, 28 Aug 2019 21:52:08 +0000 (14:52 -0700)]
ZTS: Temporarily disable several upgrade tests

Until issues #9185 and #9186 have been resolved the following zpool
upgrade tests are being disabled to prevent CI failures.

  zpool_upgrade_002_pos,
  zpool_upgrade_003_pos,
  zpool_upgrade_004_pos,
  zpool_upgrade_007_pos,
  zpool_upgrade_008_pos

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #9185
Issue #9186
Closes #9225

5 years agoTag ABD pages for exclusion in kernel crash dumps
Don Brady [Wed, 28 Aug 2019 17:44:46 +0000 (11:44 -0600)]
Tag ABD pages for exclusion in kernel crash dumps

Tag the ABD data pages so that they can be identified for exclusion
from kernel crash dumps. Eliminating the zfs file data allows for
significantly smaller crash dump files. Note that ZFS in illumos has
always excluded the zfs data pages from a kernel crash dump.

This change tags ARC scatter data pages so they can be identified from
the makedumpfile(8) command. That command is used to create smaller
dump files by ignoring some memory regions and using compression. It
already filters file data from the VFS page cache and will now be able
to exclude ZFS file data pages from the dump file.

A corresponding change to makeumpfile(8) is required to identify ZFS
data pages.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: Don Brady <don.brady@delphix.com>
Closes #8899

5 years agoFix zil replay panic when TX_REMOVE followed by TX_CREATE
Chunwei Chen [Wed, 28 Aug 2019 17:42:02 +0000 (10:42 -0700)]
Fix zil replay panic when TX_REMOVE followed by TX_CREATE

If TX_REMOVE is followed by TX_CREATE on the same object id, we need to
make sure the object removal is completely finished before creation. The
current implementation relies on dnode_hold_impl with
DNODE_MUST_BE_ALLOCATED returning ENOENT. While this check seems to work
fine before, in current version it does not guarantee the object removal
is completed.

We fix this by checking if DNODE_MUST_BE_FREE returns successful
instead. Also add test and remove dead code in dnode_hold_impl.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tom Caputi <tcaputi@datto.com>
Signed-off-by: Chunwei Chen <david.chen@nutanix.com>
Closes #7151
Closes #8910
Closes #9123
Closes #9145

5 years agoPrefer `for (;;)` to `while (TRUE)`
Ryan Moeller [Wed, 28 Aug 2019 17:38:40 +0000 (13:38 -0400)]
Prefer `for (;;)` to `while (TRUE)`

Defining a special constant to make an infinite loop is excessive,
especially when the name clashes with symbols commonly defined on
some platforms (ie FreeBSD).

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: John Kennedy <john.kennedy@delphix.com
Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
Closes #9219

5 years agozfs_ioc_snapshot: check user-prop permissions on snapshotted datasets
Andriy Gapon [Tue, 27 Aug 2019 20:45:53 +0000 (23:45 +0300)]
zfs_ioc_snapshot: check user-prop permissions on snapshotted datasets

Previously, the permissions were checked on the pool which was obviously
incorrect.

After this change, zfs_check_userprops() only validates the properties
without any permission checks.  The permissions are checked individually
for each snapshotted dataset.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Signed-off-by: Andriy Gapon <avg@FreeBSD.org>
Closes #9179
Closes #9180

5 years agoFix Plymouth passphrase prompt in initramfs script
Richard Allen [Tue, 27 Aug 2019 20:44:02 +0000 (21:44 +0100)]
Fix Plymouth passphrase prompt in initramfs script

Entering the ZFS encryption passphrase under Plymouth wasn't working
because in the ZFS initrd script, Plymouth was calling zfs via
"--command", which wasn't passing through the filesystem argument to
zfs load-key properly (it was passing through the single quotes around
the filesystem name intended to handle spaces literally,
which zfs load-key couldn't understand).

Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Garrett Fields <ghfields@gmail.com>
Signed-off-by: Richard Allen <belperite@gmail.com>
Issue #9193
Closes #9202

5 years agoFix deadlock in 'zfs rollback'
Tom Caputi [Tue, 27 Aug 2019 16:55:51 +0000 (12:55 -0400)]
Fix deadlock in 'zfs rollback'

Currently, the 'zfs rollback' code can end up deadlocked due to
the way the kernel handles unreferenced inodes on a suspended fs.
Essentially, the zfs_resume_fs() code path may cause zfs to spawn
new threads as it reinstantiates the suspended fs's zil. When a
new thread is spawned, the kernel may attempt to free memory for
that thread by freeing some unreferenced inodes. If it happens to
select inodes that are a a part of the suspended fs a deadlock
will occur because freeing inodes requires holding the fs's
z_teardown_inactive_lock which is still held from the suspend.

This patch corrects this issue by adding an additional reference
to all inodes that are still present when a suspend is initiated.
This prevents them from being freed by the kernel for any reason.

Reviewed-by: Alek Pinchuk <apinchuk@datto.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tom Caputi <tcaputi@datto.com>
Closes #9203

5 years agoRestore :: in Makefile.am
Ryan Moeller [Mon, 26 Aug 2019 18:48:31 +0000 (14:48 -0400)]
Restore :: in Makefile.am

The double-colon looked like a typo, but it's actually an obscure
feature. Rules with :: may appear multiple times and are run
independently of one another in the order they appear. The use of ::
for distclean-local was conventional, not accidental.

Add comments to indicate the intentional use of double-colon rules.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
Closes #9210

5 years agoAdd regression test for "zpool list -p"
Paul Dagnelie [Mon, 26 Aug 2019 01:33:03 +0000 (18:33 -0700)]
Add regression test for "zpool list -p"

Other than this test, zpool list -p is not well tested by any of the
automated tests.  Add a test for zpool list -p.

Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes #9134

5 years agoSplit argument list, satisfy shellcheck SC2086
Ryan Moeller [Mon, 26 Aug 2019 01:30:39 +0000 (21:30 -0400)]
Split argument list, satisfy shellcheck SC2086

Split the arguments for ${TEST_RUNNER} across multiple lines for
clarity. Also added quotes in the message to match the invoked command.

Unquoted variables in argument lists are subject to splitting. In this
particular case we can't quote the variable because it is an optional
argument. Use the method suggested in the description linked below,
instead.

The technique is to use an unquoted variable with an alternate value.

https://github.com/koalaman/shellcheck/wiki/SC2086

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Giuseppe Di Natale <guss80@gmail.com>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
Closes #9212

5 years agoZTS: Fix in-tree dbufstats test case
Brian Behlendorf [Fri, 23 Aug 2019 00:37:48 +0000 (17:37 -0700)]
ZTS: Fix in-tree dbufstats test case

Commit a887d653 updated the dbufstats such that escalated privileges
are required.  Since all tests under cli_user are run with normal
privileges move this test case to a location where it will be run
required privileges.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Reviewed-by: Michael Niewöhner <foss@mniewoehner.de>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #9118
Closes #9196

5 years agoFix install error introduced by #9089 (#9205)
Brian Behlendorf [Fri, 23 Aug 2019 00:31:38 +0000 (17:31 -0700)]
Fix install error introduced by #9089 (#9205)

Signed-off-by: Paul Dagnelie <pcd@delphix.com>
5 years agoMake slog test setup more robust
Ryan Moeller [Fri, 23 Aug 2019 00:26:51 +0000 (20:26 -0400)]
Make slog test setup more robust

The slog tests fail when attempting to create pools using file vdevs
that already exist from previous test runs. Remove these files in the
setup for the test.

Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
Closes #9194

5 years agozfs-mount-genrator: dependencies should be space-separated
yshui [Fri, 23 Aug 2019 00:11:17 +0000 (01:11 +0100)]
zfs-mount-genrator: dependencies should be space-separated

Reviewed-by: Antonio Russo <antonio.e.russo@gmail.com>
Reviewed-by: Richard Laager <rlaager@wiktel.com>
Signed-off-by: Yuxuan Shui <yshuiv7@gmail.com>
Closes #9174

5 years agoFix install error introduced by #9089
Paul Dagnelie [Thu, 22 Aug 2019 19:01:41 +0000 (12:01 -0700)]
Fix install error introduced by #9089

Signed-off-by: Paul Dagnelie <pcd@delphix.com>
5 years agoZTS: Use decimal values when setting tunables
Brian Behlendorf [Thu, 22 Aug 2019 17:36:57 +0000 (10:36 -0700)]
ZTS: Use decimal values when setting tunables

The mdb_set_uint32 function requires that the values passed in be
decimal.  This was overlooked initially because the matching Linux
function accepts both decimal and hexadecimal values.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed by: Sara Hartse <sara.hartse@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Igor Kozhukhov <igor@dilos.org>
Closes #9125
Closes #9195

5 years agoFix automake program name transformations (#9190)
Brian Behlendorf [Thu, 22 Aug 2019 16:53:45 +0000 (09:53 -0700)]
Fix automake program name transformations (#9190)

Automake can perform program name transformations at install time.
However, arc_summary has its own name transformation taking place,
which interferes with the automake transforms. The automake transforms
must be taken into account in order to resolve the conflict.

Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
5 years agoDocument ZFS_DKMS_ENABLE_DEBUGINFO in userland configuration
Mauricio Faria de Oliveira [Thu, 22 Aug 2019 16:48:48 +0000 (13:48 -0300)]
Document ZFS_DKMS_ENABLE_DEBUGINFO in userland configuration

Document the ZFS_DKMS_ENABLE_DEBUGINFO option in the userland
configuration file, as done with the other ZFS_DKMS_* options.

It has been introduced with commit e45c1734a665 ("dkms: Enable
debuginfo option to be set with zfs sysconfig file") but isn't
mentioned anywhere other than the 'dkms.conf' file (generated).

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Mauricio Faria de Oliveira <mfo@canonical.com>
Closes #9191

5 years agoDedup IOC enum values in libzfs_input_check
Ryan Moeller [Thu, 22 Aug 2019 16:46:09 +0000 (12:46 -0400)]
Dedup IOC enum values in libzfs_input_check

Reuse enum value ZFS_IOC_BASE for `('Z' << 8)`.
This is helpful on FreeBSD where ZFS_IOC_BASE has a different value and
`('Z' << 8)` is wrong.

Reviewed-by: Chris Dunlop <chris@onthe.net.au>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
Closes #9188

5 years agoEnhance ioctl number checks
Ryan Moeller [Thu, 22 Aug 2019 16:44:11 +0000 (12:44 -0400)]
Enhance ioctl number checks

When checking ZFS_IOC_* numbers, print which numbers are wrong rather
than silently failing.

Reviewed-by: Chris Dunlop <chris@onthe.net.au>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
Closes #9187

5 years agoZTS: Fix vdev_zaps_005_pos on CentOS 6
Brian Behlendorf [Thu, 22 Aug 2019 15:53:44 +0000 (08:53 -0700)]
ZTS: Fix vdev_zaps_005_pos on CentOS 6

The ancient version of blkid (v2.17.2) used in CentOS 6 will not
detect the newly created pool unless it has been written to.
Force a pool sync so `zpool import` will detect the newly created
pool.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #9199

5 years agoLinux 5.3: Fix switch() fall though compiler errors
Tony Hutter [Wed, 21 Aug 2019 16:29:23 +0000 (09:29 -0700)]
Linux 5.3: Fix switch() fall though compiler errors

Fix some switch() fall-though compiler errors:

    abd.c:1504:9: error: this statement may fall through

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Closes #9170

5 years agoMinor cleanup in Makefile.am
Ryan Moeller [Wed, 21 Aug 2019 16:01:59 +0000 (12:01 -0400)]
Minor cleanup in Makefile.am

Split long lines where adding license info to dist archive.

Remove extra colon from target line.

Reviewed-by: Chris Dunlop <chris@onthe.net.au>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
Closes #9189

5 years ago zfs-functions.in: in_mtab() always returns 1
Alexey Smirnoff [Tue, 20 Aug 2019 23:26:19 +0000 (23:26 +0000)]
 zfs-functions.in: in_mtab() always returns 1

$fs used with the wrong sed command where should be $mntpnt instead
to match a variable exported by read_mtab()

The fix is mostly to reuse the sed command found in read_mtab()

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Michael Niewöhner <foss@mniewoehner.de>
Signed-off-by: Alexey Smirnoff <fling@member.fsf.org>
Closes #9168

5 years agoFix automake program name transformations
Ryan Moeller [Tue, 20 Aug 2019 21:45:26 +0000 (17:45 -0400)]
Fix automake program name transformations

Automake can perform program name transformations at install time.
However, arc_summary has its own name transformation taking place,
which interferes with the automake transforms. The automake transforms
must be taken into account in order to resolve the conflict.

Signed-off-by: Ryan Moeller <ryan@ixsystems.com>
5 years agoAdd fast path for zfs_ioc_space_snaps() handling of empty_bpobj
Matthew Ahrens [Tue, 20 Aug 2019 18:34:52 +0000 (11:34 -0700)]
Add fast path for zfs_ioc_space_snaps() handling of empty_bpobj

When there are many snapshots, calls to zfs_ioc_space_snaps() (e.g. from
`zfs destroy -nv pool/fs@snap1%snap10000`) can be very slow, resulting
in poor performance because we are holding the dp_config_rwlock the
entire time, blocking spa_sync() from continuing.  With around ten
thousand snapshots, we've seen up to 500 seconds in this ioctl,
iterating over up to 50,000,000 bpobjs, ~99% of which are the empty
bpobj.

By creating a fast path for zfs_ioc_space_snaps() handling of the
empty_bpobj, we can achieve a ~5x performance improvement of this ioctl
(when there are many snapshots, and the deadlist is mostly
empty_bpobj's).

Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
External-issue: DLPX-58348
Closes #8744

5 years agoFix lockdep circular locking false positive involving sa_lock
jdike [Mon, 19 Aug 2019 23:04:26 +0000 (19:04 -0400)]
Fix lockdep circular locking false positive involving sa_lock

There are two different deadlock scenarios, but they share a common
link, which is
thread 1 holding sa_lock and trying to get zap->zap_rwlock:
    zap_lockdir_impl+0x858/0x16c0 [zfs]
    zap_lockdir+0xd2/0x100 [zfs]
    zap_lookup_norm+0x7f/0x100 [zfs]
    zap_lookup+0x12/0x20 [zfs]
    sa_setup+0x902/0x1380 [zfs]
    zfsvfs_init+0x3d6/0xb20 [zfs]
    zfsvfs_create+0x5dd/0x900 [zfs]
    zfs_domount+0xa3/0xe20 [zfs]

and thread 2 trying to get sa_lock, either in sa_setup:
   sa_setup+0x742/0x1380 [zfs]
   zfsvfs_init+0x3d6/0xb20 [zfs]
   zfsvfs_create+0x5dd/0x900 [zfs]
   zfs_domount+0xa3/0xe20 [zfs]
or in sa_build_index:
   sa_build_index+0x13d/0x790 [zfs]
   sa_handle_get_from_db+0x368/0x500 [zfs]
   zfs_znode_sa_init.isra.0+0x24b/0x330 [zfs]
   zfs_znode_alloc+0x3da/0x1a40 [zfs]
   zfs_zget+0x39a/0x6e0 [zfs]
   zfs_root+0x101/0x160 [zfs]
   zfs_domount+0x91f/0xea0 [zfs]

From there, there are different locking paths back to something
holding zap->zap_rwlock.

The deadlock scenarios involve multiple different ZFS filesystems
being mounted.  sa_lock is common to these scenarios, and the sa
struct involved is private to a mount.  Therefore, these must be
referring to different sa_lock instances and these deadlocks can't
occur in practice.

The fix, from Brian Behlendorf, is to remove sa_lock from lockdep
coverage by initializing it with MUTEX_NOLOCKDEP.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Jeff Dike <jdike@akamai.com>
Closes #9110

5 years agoLinux 5.3 compat: Makefile subdir-m no longer supported
Dominic Pearson [Mon, 19 Aug 2019 22:22:52 +0000 (00:22 +0200)]
Linux 5.3 compat: Makefile subdir-m no longer supported

Uses obj-m instead, due to kernel changes.

See LKML: Masahiro Yamada, Tue, 6 Aug 2019 19:03:23 +0900

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Dominic Pearson <dsp@technoanimal.net>
Closes #9169

5 years agoSet "none" scheduler if available (initramfs)
colmbuckley [Mon, 19 Aug 2019 22:11:47 +0000 (23:11 +0100)]
Set "none" scheduler if available (initramfs)

Existing zfs initramfs script logic will attempt to set the 'noop'
scheduler if it's available on the vdev block devices. Newer kernels
have the similar 'none' scheduler on multiqueue devices; this change
alters the initramfs script logic to also attempt to set this scheduler
if it's available.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Garrett Fields <ghfields@gmail.com>
Reviewed-by: Richard Laager <rlaager@wiktel.com>
Signed-off-by: Colm Buckley <colm@tuatha.org>
Closes #9042

5 years agoAdd more refquota tests
Paul Dagnelie [Mon, 19 Aug 2019 22:06:53 +0000 (15:06 -0700)]
Add more refquota tests

It used to be possible for zfs receive (and other operations related
to clone swap) to bypass refquotas. This can cause a number of issues,
and there should be an automated test for it.

Added tests for rollback and receive not overriding refquota.

Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes #9139

5 years agoCap metaslab memory usage
Paul Dagnelie [Fri, 16 Aug 2019 15:08:21 +0000 (08:08 -0700)]
Cap metaslab memory usage

On systems with large amounts of storage and high fragmentation, a huge
amount of space can be used by storing metaslab range trees. Since
metaslabs are only unloaded during a txg sync, and only if they have
been inactive for 8 txgs, it is possible to get into a state where all
of the system's memory is consumed by range trees and metaslabs, and
txgs cannot sync. While ZFS knows how to evict ARC data when needed,
it has no such mechanism for range tree data. This can result in boot
hangs for some system configurations.

First, we add the ability to unload metaslabs outside of syncing
context. Second, we store a multilist of all loaded metaslabs, sorted
by their selection txg, so we can quickly identify the oldest
metaslabs.  We use a multilist to reduce lock contention during heavy
write workloads. Finally, we add logic that will unload a metaslab
when we're loading a new metaslab, if we're using more than a certain
fraction of the available memory on range trees.

Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Sebastien Roy <sebastien.roy@delphix.com>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes #9128

5 years agoinitramfs: fixes for (debian) initramfs
Michael Niewöhner [Fri, 16 Aug 2019 15:02:32 +0000 (17:02 +0200)]
initramfs: fixes for (debian) initramfs

* contrib/initramfs: include /etc/default/zfs and /etc/zfs/zfs-functions
At least debian needs /etc/default/zfs and /etc/zfs/zfs-functions for
its initramfs. Include both in build when initramfs is configured.

* contrib/initramfs: include 60-zvol.rules and zvol_id
Include 60-zvol.rules and zvol_id and set udev as predependency instead
of debians zdev. This makes debians additional zdev hook unneeded.

* Correct initconfdir substitution for some distros
Not every Linux distro is using @sysconfdir@/default but @initconfdir@
which is already determined by configure. Let's use it.

* systemd: prevent possible conflict between systemd and sysvinit
Systemd will not load a sysvinit service if a unit exists with the same
name. This prevents conflicts between sysvinit and systemd.
In ZFS there is one sysvinit service that does not have a systemd
service but a target counterpart, zfs-import.target.
Usually it does not make any sense to install both but it is possisble.
Let's prevent any conflict by masking zfs-import.service by default.
This does not harm even if init.d/zfs-import does not exist.

Reviewed-by: Chris Wedgwood <cw@f00f.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Tested-by: Alex Ingram <reimu@reimuhakurei.net>
Tested-by: Dreamcat4 <dreamcat4@gmail.com>
Signed-off-by: Michael Niewöhner <foss@mniewoehner.de>
Closes #7904
Closes #9089

5 years agodmu_tx_wait() hang likely due to cv_signal() in dsl_pool_dirty_delta()
Serapheim Dimitropoulos [Thu, 15 Aug 2019 23:53:53 +0000 (16:53 -0700)]
dmu_tx_wait() hang likely due to cv_signal() in dsl_pool_dirty_delta()

Even though the bug's writeup (Github issue #9136) is very detailed,
we still don't know exactly how we got to that state, thus I wasn't
able to reproduce the bug. That said, we can make an educated guess
combining the information on filled issue with the code.

From the fact that `dp_dirty_total` was 0 (which is less than
`zfs_dirty_data_max`) we know that there was one thread that set it to
0 and then signaled one of the waiters of `dp_spaceavail_cv` [see
`dsl_pool_dirty_delta()` which is also the only place that
`dp_dirty_total` is changed].  Thus, the only logical explaination
then for the bug being hit is that the waiter that just got awaken
didn't go through `dsl_pool_dirty_data()`. Given that this function
is only called by `dsl_pool_dirty_space()` or `dsl_pool_undirty_space()`
I can only think of two possible ways of the above scenario happening:

[1] The waiter didn't call into any of the two functions - which I
    find highly unlikely (i.e. why wait on `dp_spaceavail_cv` to begin
    with?).
[2] The waiter did call in one of the above function but it passed 0 as
    the space/delta to be dirtied (or undirtied) and then the callee
    returned immediately (e.g both `dsl_pool_dirty_space()` and
    `dsl_pool_undirty_space()` return immediately when space is 0).

In any case and no matter how we got there, the easy fix would be to
just broadcast to all waiters whenever `dp_dirty_total` hits 0. That
said and given that we've never hit this before, it would make sense
to think more on why the above situation occured.

Attempting to mimic what Prakash was doing in the issue filed, I
created a dataset with `sync=always` and started doing contiguous
writes in a file within that dataset. I observed with DTrace that even
though we update the pool's dirty data accounting when we would dirty
stuff, the accounting wouldn't be decremented incrementally as we were
done with the ZIOs of those writes (the reason being that
`dbuf_write_physdone()` isn't be called as we go through the override
code paths, and thus `dsl_pool_undirty_space()` is never called). As a
result we'd have to wait until we get to `dsl_pool_sync()` where we
zero out all dirty data accounting for the pool and the current TXG's
metadata.

In addition, as Matt noted and I later verified, the same issue would
arise when using dedup.

In both cases (sync & dedup) we shouldn't have to wait until
`dsl_pool_sync()` zeros out the accounting data. According to the
comment in that part of the code, the reasons why we do the zeroing,
have nothing to do with what we observe:
````
/*
 * We have written all of the accounted dirty data, so our
 * dp_space_towrite should now be zero.  However, some seldom-used
 * code paths do not adhere to this (e.g. dbuf_undirty(), also
 * rounding error in dbuf_write_physdone).
 * Shore up the accounting of any dirtied space now.
 */
dsl_pool_undirty_space(dp, dp->dp_dirty_pertxg[txg & TXG_MASK], txg);
````

Ideally what we want to do is to undirty in the accounting exactly what
we dirty (I use the word ideally as we can still have rounding errors).
This would make the behavior of the system more clear and predictable.

Another interesting issue that I observed with DTrace was that we
wouldn't update any of the pool's dirty data accounting whenever we
would dirty and/or undirty MOS data. In addition, every time we would
change the size of a dbuf through `dbuf_new_size()` we wouldn't update
the accounted space dirtied in the appropriate dirty record, so when
ZIOs are done we would undirty less that we dirtied from the pool's
accounting point of view.

For the first two issues observed (sync & dedup) this patch ensures
that we still update the pool's accounting when we undirty data,
regardless of the write being physical or not.

For changes in the MOS, we first ensure to zero out the pool's dirty
data accounting in `dsl_pool_sync()` after we synced the MOS. Then we
can go ahead and enable the update of the pool's dirty data accounting
wheneve we change MOS data.

Another fix is that we now update the accounting explicitly for
counting errors in `dbuf_write_done()`.

Finally, `dbuf_new_size()` updates the accounted space of the
appropriate dirty record correctly now.

The problem is that we still don't know how the bug came up in the
issue filled. That said the issues fixed seem to be very relevant, so
instead of going with the broadcasting solution right away,
I decided to leave this patch as is.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
External-issue: DLPX-47285
Closes #9137

5 years agoImprove write performance by using dmu_read_by_dnode()
Tony Nguyen [Thu, 15 Aug 2019 23:36:24 +0000 (17:36 -0600)]
Improve write performance by using dmu_read_by_dnode()

In zfs_log_write(), we can use dmu_read_by_dnode() rather than
dmu_read() thus avoiding unnecessary dnode_hold() calls.

We get a 2-5% performance gain for large sequential_writes tests, >=128K
writes to files with recordsize=8K.

Testing done on Ubuntu 18.04 with 4.15 kernel, 8vCPUs and SSD storage on
VMware ESX.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tony Nguyen <tony.nguyen@delphix.com>
Closes #9156

5 years agoAssert that a dnode's bonuslen never exceeds its recorded size
Serapheim Dimitropoulos [Thu, 15 Aug 2019 14:44:57 +0000 (07:44 -0700)]
Assert that a dnode's bonuslen never exceeds its recorded size

This patch introduces an assertion that can catch pitfalls in
development where there is a mismatch between the size of
reads and writes between a *_phys structure and its respective
in-core structure when bonus buffers are used.

This debugging-aid should be complementary to the verification
done by ztest in ztest_verify_dnode_bt().

A side to this patch is that we now clear out any extra bytes
past a bonus buffer's new size when the buffer is shrinking.

Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tom Caputi <tcaputi@datto.com>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes #8348

5 years agoMake txg_wait_synced conditional in zfsvfs_teardown
Paul Zuchowski [Thu, 15 Aug 2019 14:27:13 +0000 (10:27 -0400)]
Make txg_wait_synced conditional in zfsvfs_teardown

The call to txg_wait_synced in zfsvfs_teardown should
be made conditional on the objset having dirty data.
This can prevent unnecessary txg_wait_synced during
some unmount operations.

Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Zuchowski <pzuchowski@datto.com>
Closes #9115

5 years agoPrevent race in blkptr_verify against device removal
Paul Dagnelie [Wed, 14 Aug 2019 03:24:43 +0000 (20:24 -0700)]
Prevent race in blkptr_verify against device removal

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

5 years agoFix out-of-order ZIL txtype lost on hardlinked files
Chunwei Chen [Wed, 14 Aug 2019 03:21:27 +0000 (20:21 -0700)]
Fix out-of-order ZIL txtype lost on hardlinked files

We should only call zil_remove_async when an object is removed. However,
in current implementation, it is called whenever TX_REMOVE is called. In
the case of hardlinked file, every unlink will generate TX_REMOVE and
causing operations to be dropped even when the object is not removed.

We fix this by only calling zil_remove_async when the file is fully
unlinked.

Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Signed-off-by: Chunwei Chen <david.chen@nutanix.com>
Closes #8769
Closes #9061

5 years agoFix device expansion when VM is powered off
Prakash Surya [Wed, 14 Aug 2019 03:18:53 +0000 (20:18 -0700)]
Fix device expansion when VM is powered off

When running on an ESXi based VM, I've found that "zpool online -e" will
not expand the zpool, if the disk was expanded in ESXi while the VM was
powered off.

For example, take the following scenario:

 1. VM running on top of VMware ESXi
 2. ZFS pool created with a given device "sda" of size 8GB
 3. VM powered off
 4. Device "sda" size expanded to 16GB
 5. VM powered on
 6. "zpool online -e" used on device "sda"

In this situation, after (2) the zpool will be roughly 8GB in size.
After (6), the expectation is the zpool's size will expand to roughly
16GB in size; i.e. expand to the new size of the "sda" device.
Unfortunately, I've seen that after (6), the zpool size does not change.

What's happening is after (5), the EFI label of the "sda" device will be
such that fields "efi_last_u_lba", "efi_last_lba", and "efi_altern_lba"
all reflect the new size of the disk; i.e. "33554398", "33554431", and
"33554431" respectively.

Thus, the check that we perform in "efi_use_whole_disk":

    if ((efi_label->efi_altern_lba == 1) || (efi_label->efi_altern_lba
        >= efi_label->efi_last_lba)) {

This will return true, and then we return from the function without
having expanded the size of the zpool/device.

In contrast, if we remove steps (3) and (5) in the sequence above, i.e.
the device is expanded while the VM is powered on, things change. In
that case, the fields "efi_last_u_lba" and "efi_altern_lba" do not
change (i.e. they still reflect the old 8GB device size), but the
"efi_last_lba" field does change (i.e. it now reflects the new 16GB
device size). Thus, when we evaluate the same conditional in
"efi_use_whole_disk", it'll return false, so the zpool is expanded.

Taking all of this into account, this PR updates "efi_use_whole_disk" to
properly expand the zpool when the underlying disk is expanded while the
VM is powered off.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Don Brady <don.brady@delphix.com>
Signed-off-by: Prakash Surya <prakash.surya@delphix.com>
Closes #9111

5 years agoMark dsl_livelist_should_disable() static
Allan Jude [Wed, 14 Aug 2019 03:16:23 +0000 (23:16 -0400)]
Mark dsl_livelist_should_disable() static

This function is not used outside of dsl_dataset.c

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed by: Sara Hartse <sara.hartse@delphix.com>
Signed-off-by: Allan Jude <allanjude@freebsd.org>
Closes #9154

5 years agospa_load_verify() may consume too much memory
George Wilson [Tue, 13 Aug 2019 14:11:57 +0000 (08:11 -0600)]
spa_load_verify() may consume too much memory

When a pool is imported it will scan the pool to verify the integrity
of the data and metadata. The amount it scans will depend on the
import flags provided. On systems with small amounts of memory or
when importing a pool from the crash kernel, it's possible for
spa_load_verify to issue too many I/Os that it consumes all the memory
of the system resulting in an OOM message or a hang.

To prevent this, we limit the amount of memory that the initial pool
scan can consume. This change will, by default, use 1/16th of the ARC
for scan I/Os to prevent running the system out of memory during import.

Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Signed-off-by: George Wilson george.wilson@delphix.com
External-issue: DLPX-65237
External-issue: DLPX-65238
Closes #9146

5 years agoChange boolean-like uint8_t fields in znode_t to boolean_t
Tomohiro Kusumi [Tue, 13 Aug 2019 13:58:02 +0000 (22:58 +0900)]
Change boolean-like uint8_t fields in znode_t to boolean_t

Given znode_t is an in-core structure, it's more readable to have
them as boolean. Also co-locate existing boolean fields with them
for space efficiency (expecting 8 booleans to be packed/aligned).

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Closes #9092

5 years agoDrop KMC_NOEMERGENCY
Richard Yao [Tue, 13 Aug 2019 13:46:12 +0000 (09:46 -0400)]
Drop KMC_NOEMERGENCY

This is not implemented. If it were implemented, using it would risk
deadlocks on pre-3.18 kernels. Lets just drop it.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Michael Niewöhner <foss@mniewoehner.de>
Signed-off-by: Richard Yao <ryao@gentoo.org>
Closes #9119

5 years agoIntroduce getting holds and listing bookmarks through ZCP
Serapheim Dimitropoulos [Mon, 12 Aug 2019 17:02:34 +0000 (10:02 -0700)]
Introduce getting holds and listing bookmarks through ZCP

Consumers of ZFS Channel Programs can now list bookmarks,
and get holds from datasets. A minor-refactoring was also
applied to distinguish between user and system properties
in ZCP.

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Ported-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Signed-off-by: Dan Kimmel <dan.kimmel@delphix.com>
OpenZFS-issue: https://illumos.org/issues/8862
Closes #7902

5 years agoSort log spacemap tunables in alphabetical order
Serapheim Dimitropoulos [Mon, 12 Aug 2019 16:49:07 +0000 (09:49 -0700)]
Sort log spacemap tunables in alphabetical order

Beside the whole commit being a nit in reality it should
bring the diffs of the spa_log_spacemap.c source file
between ZoL and delphix/zfs to 0.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Chris Dunlop <chris@onthe.net.au>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes #9143

5 years agoMetaslab max_size should be persisted while unloaded
Paul Dagnelie [Mon, 5 Aug 2019 21:34:27 +0000 (14:34 -0700)]
Metaslab max_size should be persisted while unloaded

When we unload metaslabs today in ZFS, the cached max_size value is
discarded. We instead use the histogram to determine whether or not we
think we can satisfy an allocation from the metaslab. This can result in
situations where, if we're doing I/Os of a size not aligned to a
histogram bucket, a metaslab is loaded even though it cannot satisfy the
allocation we think it can. For example, a metaslab with 16 entries in
the 16k-32k bucket may have entirely 16kB entries. If we try to allocate
a 24kB buffer, we will load that metaslab because we think it should be
able to handle the allocation. Doing so is expensive in CPU time, disk
reads, and average IO latency. This is exacerbated if the write being
attempted is a sync write.

This change makes ZFS cache the max_size after the metaslab is
unloaded. If we ever get a free (or a coalesced group of frees) larger
than the max_size, we will update it. Otherwise, we leave it as is. When
attempting to allocate, we use the max_size as a lower bound, and
respect it unless we are in try_hard. However, we do age the max_size
out at some point, since we expect the actual max_size to increase as we
do more frees. A more sophisticated algorithm here might be helpful, but
this works reasonably well.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes #9055

5 years agoDon't wakeup unnecessarily in 'zpool events -f'
DeHackEd [Mon, 5 Aug 2019 18:35:47 +0000 (14:35 -0400)]
Don't wakeup unnecessarily in 'zpool events -f'

ZED can prevent CPU's from properly sleeping.

Rather than periodically waking up in the zevents code, just go to sleep and wait for a wakeup.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: DHE <git@dehacked.net>
Closes #9091

5 years agoTest cancelling a removal in ZTS
Serapheim Dimitropoulos [Mon, 5 Aug 2019 17:50:20 +0000 (10:50 -0700)]
Test cancelling a removal in ZTS

This patch adds a new test that sanity checks cancelling a removal.

Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes #9101

5 years agolockdep false positive - move txg_kick() outside of ->dp_lock
jdike [Wed, 31 Jul 2019 21:53:39 +0000 (17:53 -0400)]
lockdep false positive - move txg_kick() outside of ->dp_lock

This fixes a lockdep warning by breaking a link between ->tx_sync_lock
and ->dp_lock.

The deadlock envisioned by lockdep is this:
    thread 1 holds db->db_mtx and tries to get dp->dp_lock:
dsl_pool_dirty_space+0x70/0x2d0 [zfs]
dbuf_dirty+0x778/0x31d0 [zfs]

    thread 2 holds bpo->bpo_lock and tries to get db->db_mtx:
        dmu_buf_will_dirty_impl
        dmu_buf_will_dirty+0x6b/0x6c0 [zfs]
        bpobj_iterate_impl+0xbe6/0x1410 [zfs]

    thread 3 holds tx->tx_sync_lock and tries to get bpo->bpo_lock:
        bpobj_space+0x63/0x470 [zfs]
        dsl_scan_active+0x340/0x3d0 [zfs]
        txg_sync_thread+0x3f2/0x1370 [zfs]

    thread 4 holds dp->dp_lock and tries to get tx->tx_sync_lock
       txg_kick+0x61/0x420 [zfs]
       dsl_pool_need_dirty_delay+0x1c7/0x3f0 [zfs]

This patch is orginally from Brian Behlendorf and slightly simplified
by me.

It breaks this cycle in thread 4 by moving the call from
dsl_pool_need_dirty_delay to txg_kick outside the section controlled
by dp->dp_lock.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Signed-off-by: Jeff Dike <jdike@akamai.com>
Closes #9094

5 years agoList log_spacemap feature in zpool-features.5 manual
Serapheim Dimitropoulos [Wed, 31 Jul 2019 16:29:01 +0000 (09:29 -0700)]
List log_spacemap feature in zpool-features.5 manual

Update zpool-features.5 manpage to describe the log_spacemap feature.

Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes #9096

5 years agoAdd channel program for property based snapshots
Clint Armstrong [Tue, 30 Jul 2019 23:02:19 +0000 (19:02 -0400)]
Add channel program for property based snapshots

Channel programs that many users find useful should be included with zfs
in the /contrib directory. This is the first of these contributions. A
channel program to recursively take snapshots of datasets with the
property com.sun:auto-snapshot=true.

Reviewed-by: Kash Pande <kash@tripleback.net>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Clint Armstrong <clint@clintarmstrong.net>
Closes #8443
Closes #9050

5 years ago9072 handle error of zap_cursor_retrieve() for log spacemap zap
Serapheim Dimitropoulos [Tue, 30 Jul 2019 20:20:01 +0000 (13:20 -0700)]
9072 handle error of zap_cursor_retrieve() for log spacemap zap

In spa_ld_log_sm_metadata(), it is possible for zap_cursor_retrieve()
to return errors other than the expected ENOENT (e.g. when we are at
the end of the zap). Ensure that these error cases are handled
correctly by the import path.

Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed by: Sara Hartse <sara.hartse@delphix.com>
Reviewed by: Matt Ahrens <matt@delphix.com>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes #9074

5 years agomismerged log spacemap comment for metaslab_verify_weight_and_frag
Serapheim Dimitropoulos [Tue, 30 Jul 2019 17:13:44 +0000 (10:13 -0700)]
mismerged log spacemap comment for metaslab_verify_weight_and_frag

When the log spacemap commit was merged in ZoL, the
metaslab_verify_unflushed_changes() debugging function
was deleted as the feature was pretty much stable by
then. Unfortunately though there was a reference to
it from a comment in metaslab_verify_weight_and_frag().

This patch deletes the reference and pastes that
comment as is.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes #9097

5 years agoinstall path fixes
Michael Niewöhner [Tue, 30 Jul 2019 17:06:09 +0000 (19:06 +0200)]
install path fixes

* rpm: correct pkgconfig path

pkconfig files get installed to $datarootdir/pkgconfig but rpm expects
them to be at $datadir. This works when $datarootdir==$datadir which is
the case most of the time but will fail when they differ.

* install: make initramfs-tools path static

Since initramfs-tools' path is nothing we can control as it is an
external package it does not make any sense to install zfs additions
anywhere else. Simply use /usr/share/initramfs-tools as path.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Laager <rlaager@wiktel.com>
Signed-off-by: Michael Niewöhner <foss@mniewoehner.de>
Closes #9087

5 years agoIncrease default zcmd allocation to 256K
Michael Niewöhner [Tue, 30 Jul 2019 16:59:38 +0000 (18:59 +0200)]
Increase default zcmd allocation to 256K

When creating hundreds of clones (for example using containers with
LXD) cloning slows down as the number of clones increases over time.
The reason for this is that the fetching of the clone information
using a small zcmd buffer requires two ioctl calls, one to determine
the size and a second to return the data. However, this requires
gathering the data twice, once to determine the size and again to
populate the zcmd buffer to return it to userspace.
These are expensive ioctl() calls, so instead, make the default buffer
size much larger: 256K.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Michael Niewöhner <foss@mniewoehner.de>
Closes #9084

5 years agoImprove performance by using dmu_tx_hold_*_by_dnode()
Matthew Ahrens [Tue, 30 Jul 2019 16:18:30 +0000 (09:18 -0700)]
Improve performance by using dmu_tx_hold_*_by_dnode()

In zfs_write() and dmu_tx_hold_sa(), we can use dmu_tx_hold_*_by_dnode()
instead of dmu_tx_hold_*(), since we already have a dbuf from the target
dnode in hand.  This eliminates some calls to dnode_hold(), which can be
expensive.  This is especially impactful if several threads are
accessing objects that are in the same block of dnodes, because they
will contend for that dbuf's lock.

We are seeing 10-20% performance wins for the sequential_writes tests in
the performance test suite, when doing >=128K writes to files with
recordsize=8K.

This also removes some unnecessary casts that are in the area.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #9081

5 years agoRevert "Develop tests for issues #5866 and #8858"
Brian Behlendorf [Mon, 29 Jul 2019 19:46:56 +0000 (12:46 -0700)]
Revert "Develop tests for issues #5866 and #8858"

This reverts commit 693c1fc478cc8118dd0168c4815c0ae3be41c9c3.  This
change resulted in a kmem leak being observed in existing code which
needs to be identified and addressed.

Reviewed-by: Paul Zuchowski <pzuchowski@datto.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #8978
Closes #9090

5 years agoFix channel programs on s390x
Brian Behlendorf [Mon, 29 Jul 2019 01:15:26 +0000 (18:15 -0700)]
Fix channel programs on s390x

When adapting the original sources for s390x the JMP_BUF_CNT was
mistakenly halved due to an incorrect assumption of the size of
a unsigned long.  They are 8 bytes for the s390x architecture.
Increase JMP_BUF_CNT accordingly.

Authored-by: Don Brady <don.brady@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reported-by: Colin Ian King <canonical.com>
Tested-by: Colin Ian King <canonical.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #8992
Closes #9080

5 years agoRace between zfs-share and zfs-mount services
George Wilson [Mon, 29 Jul 2019 01:13:56 +0000 (21:13 -0400)]
Race between zfs-share and zfs-mount services

When a system boots the zfs-mount.service and the
zfs-share.service can start simultaneously. What may be
unclear is that sharing a filesystem will first mount
the filesystem if it's not already mounted. This means
that both service can race to mount the same fileystem.
This race can result in a SEGFAULT or EBUSY conditions.

This change explicitly defines the start ordering between the
two services such that the zfs-mount.service is solely
responsible for mounting filesystems eliminating the race
between "zfs mount -a" and "zfs share -a" commands.

Reviewed-by: Sebastien Roy <sebastien.roy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Wilson <george.wilson@delphix.com>
Closes #9083

5 years agoDevelop tests for issues #5866 and #8858
Paul Zuchowski [Sat, 27 Jul 2019 00:52:13 +0000 (20:52 -0400)]
Develop tests for issues #5866 and #8858

Provide zfstest coverage for these two issues which
were a panic accessing extended attributes and
a problem comparing 64 bit and 32 bit generation
numbers.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Zuchowski <pzuchowski@datto.com>
Issue #5866
Issue #8858
Closes #8978

5 years agoImplement secpolicy_vnode_setid_retain()
Tomohiro Kusumi [Fri, 26 Jul 2019 20:52:30 +0000 (05:52 +0900)]
Implement secpolicy_vnode_setid_retain()

Don't unconditionally return 0 (i.e. retain SUID/SGID).
Test CAP_FSETID capability.

https://github.com/pjd/pjdfstest/blob/master/tests/chmod/12.t
which expects SUID/SGID to be dropped on write(2) by non-owner fails
without this. Most filesystems make this decision within VFS by using
a generic file write for fops.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Closes #9035
Closes #9043

5 years agozed crashes when devid not present
Matthew Ahrens [Fri, 26 Jul 2019 19:07:48 +0000 (12:07 -0700)]
zed crashes when devid not present

zed core dumps due to a NULL pointer in zfs_agent_iter_vdev(). The
gs_devid is NULL, but the nvl has a "devid" entry.

zfs_agent_post_event() checks that ZFS_EV_VDEV_GUID or DEV_IDENTIFIER is
present in nvl, but then later it and zfs_agent_iter_vdev() assume that
DEV_IDENTIFIER is present and thus gs_devid is set.

Typically this is not a problem because usually either all vdevs have
devid's, or none of them do. Since zfs_agent_iter_vdev() first checks if
the vdev has devid before dereferencing gs_devid, the problem isn't
typically encountered. However, if some vdevs have devid's and some do
not, then the problem is easily reproduced.  This can happen if the pool
has been moved from a system that has devid's to one that does not.

The fix is for zfs_agent_iter_vdev() to only try to match the devid's if
both nvl and gsp have devid's present.

Reviewed-by: Prashanth Sreenivasa <pks@delphix.com>
Reviewed-by: Don Brady <don.brady@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: loli10K <ezomori.nozomu@gmail.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
External-issue: DLPX-65090
Closes #9054
Closes #9060

5 years agoFast Clone Deletion
Sara Hartse [Fri, 26 Jul 2019 17:54:14 +0000 (10:54 -0700)]
Fast Clone Deletion

Deleting a clone requires finding blocks are clone-only, not shared
with the snapshot. This was done by traversing the entire block tree
which results in a large performance penalty for sparsely
written clones.

This is new method keeps track of clone blocks when they are
modified in a "Livelist" so that, when it’s time to delete,
the clone-specific blocks are already at hand.

We see performance improvements because now deletion work is
proportional to the number of clone-modified blocks, not the size
of the original dataset.

Reviewed-by: Sean Eric Fagan <sef@ixsystems.com>
Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Signed-off-by: Sara Hartse <sara.hartse@delphix.com>
Closes #8416

5 years agoDon't directly cast unsigned long to void*
Tomohiro Kusumi [Thu, 25 Jul 2019 18:59:20 +0000 (03:59 +0900)]
Don't directly cast unsigned long to void*

Cast to uintptr_t first for portability on integer to/from pointer
conversion.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Closes #9065

5 years agoReplace zf_rwlock with a mutex
Matthew Ahrens [Thu, 25 Jul 2019 18:57:58 +0000 (11:57 -0700)]
Replace zf_rwlock with a mutex

The rwlock implementation on linux does not perform as well as mutexes.
We can realize a performance benefit by replacing the zf_rwlock with a
mutex.  Local microbenchmarks show ~50% improvement, and over NFS we see
~5% improvement on several of the ZFS Performance Tests cases,
especially randwrite and seq_write.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #9062

5 years agoFix module_param() type for zfs_read_chunk_size
Tomohiro Kusumi [Fri, 19 Jul 2019 18:23:56 +0000 (03:23 +0900)]
Fix module_param() type for zfs_read_chunk_size

zfs_read_chunk_size is unsigned long.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Closes #9051

5 years agoMove some tests to cli_user/zpool_status
Tony Hutter [Fri, 19 Jul 2019 18:21:54 +0000 (11:21 -0700)]
Move some tests to cli_user/zpool_status

The tests in tests/functional/cli_root/zpool_status should all require
root. However, linux.run has "user =" specified for those tests, which
means they run as a normal user.  When I removed that line to run them
as root, the following tests did not pass:

zpool_status_003_pos
zpool_status_-c_disable
zpool_status_-c_homedir
zpool_status_-c_searchpath

These tests need to be run as a normal user.  To fix this, move these
tests to a new tests/functional/cli_user/zpool_status directory.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Giuseppe Di Natale <guss80@gmail.com>
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Closes #9057

5 years agoTricky semantics of ms_max_size in metaslab_should_allocate()
Serapheim Dimitropoulos [Fri, 19 Jul 2019 18:19:50 +0000 (11:19 -0700)]
Tricky semantics of ms_max_size in metaslab_should_allocate()

metaslab_should_allocate() is used in two places:
[1] When trying to select a metaslab to allocate from
[2] When trying to allocate from a metaslab

In [2] we always expect the metaslab to be loaded, and after
the refactoring of the log spacemap changes, whenever we load
a metaslab we set ms_max_size to the biggest range in the
ms_allocatable tree. Thus, when it is used in [2], if that
field is 0, it means that the metaslab doesn't have any
segments that can be used for allocations now (though it may
have some free space but that space can be in the freeing,
freed, or deferred trees).

In [1] a metaslab can be loaded or unloaded at which point 0
can either mean the metaslab doesn't have any space or the
metaslab is just not loaded thus we go ahead and try to make
an estimation based on its weight.

The issue here is when we call the above function for [2] and
the metaslab doesn't have any allocatable space, we still go
ahead and check its ms_weight which may be out of date because
we haven't ran metaslab_sync_done() yet. At that point we are
allowing an allocation to be attempted even though we know
there is no range that is allocatable.

This patch fixes this issue by explicitly checking if the
metaslab is loaded and if it is, the ms_max_size is used.

Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes #9045

5 years agoRace condition between spa async threads and export
Serapheim Dimitropoulos [Thu, 18 Jul 2019 20:02:33 +0000 (13:02 -0700)]
Race condition between spa async threads and export

In the past we've seen multiple race conditions that have
to do with open-context threads async threads and concurrent
calls to spa_export()/spa_destroy() (including the one
referenced in issue #9015).

This patch ensures that only one thread can execute the
main body of spa_export_common() at a time, with subsequent
threads returning with a new error code created just for
this situation, eliminating this way any race condition
bugs introduced by concurrent calls to this function.

Reviewed by: Matt Ahrens <matt@delphix.com>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes #9015
Closes #9044

5 years agohdr_recl calls zthr_wakeup() on destroyed zthr
Serapheim Dimitropoulos [Thu, 18 Jul 2019 19:55:29 +0000 (12:55 -0700)]
hdr_recl calls zthr_wakeup() on destroyed zthr

There exists a race condition were hdr_recl() calls
zthr_wakeup() on a destroyed zthr. The timeline is the
following:

[1] hdr_recl() runs first and goes intro zthr_wakeup()
    because arc_initialized is set.
[2] arc_fini() is called by another thread, zeroes
    that flag, destroying the zthr, and goes into
    buf_init().
[3] hdr_recl() tries to enter the destroyed mutex
    and we blow up.

This patch ensures that the ARC's zthrs are not offloaded
any new work once arc_initialized is set and then destroys
them after all of the ARC state has been deleted.

Reviewed by: Matt Ahrens <matt@delphix.com>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes #9047

5 years agozdb: don't print log spacemap stats in pools without the feature
Serapheim Dimitropoulos [Thu, 18 Jul 2019 19:54:03 +0000 (12:54 -0700)]
zdb: don't print log spacemap stats in pools without the feature

Creating a pool with not features enabled and running
`zdb -mmmmmm on` it before the patch:

```
Log Space Maps in Pool:

Log Space Map Obsolete Entry Statistics:
0        valid entries out of 0        - txg 0
0        valid entries out of 0        - total
```

After this patch the above output goes away.

Reviewed by: Matt Ahrens <matt@delphix.com>
Reviewed by: Sara Hartse <sara.hartse@delphix.com>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes #9048

5 years agoFix wrong comment on zcr_blksz_{min,max}
Tomohiro Kusumi [Thu, 18 Jul 2019 19:48:46 +0000 (04:48 +0900)]
Fix wrong comment on zcr_blksz_{min,max}

These aren't tunable; illumos has this comment fixed in
"3742 zfs comments need cleaner, more consistent style",
so sync with that.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Closes #9052

5 years agoNew service that waits on zvol links to be created
Pavel Zakharov [Wed, 17 Jul 2019 22:33:05 +0000 (18:33 -0400)]
New service that waits on zvol links to be created

The zfs-volume-wait.service scans existing zvols and waits for their
links under /dev to be created. Any service that depends on zvol
links to be there should add a dependency on zfs-volumes.target.
By default, this target is not enabled.

Reviewed-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
Reviewed-by: Antonio Russo <antonio.e.russo@gmail.com>
Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: loli10K <ezomori.nozomu@gmail.com>
Reviewed-by: John Gallagher <john.gallagher@delphix.com>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Pavel Zakharov <pzakharov@delphix.com>
Closes #8975

5 years agoRetire unused spl_{mutex,rwlock}_{init_fini}
Brian Behlendorf [Fri, 12 Jul 2019 21:40:15 +0000 (14:40 -0700)]
Retire unused spl_{mutex,rwlock}_{init_fini}

These functions are unused and can be removed along
with the spl-mutex.c and spl-rwlock.c source files.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #9029

5 years agoLinux 5.3 compat: retire rw_tryupgrade()
Brian Behlendorf [Fri, 12 Jul 2019 21:06:36 +0000 (14:06 -0700)]
Linux 5.3 compat: retire rw_tryupgrade()

The Linux kernel's rwsem's have never provided an interface to
allow a reader to be upgraded to a writer.  Historically, this
functionality has been implemented by a SPL wrapper function.
However, this approach depends on internal knowledge of the
rw_semaphore and is therefore rather brittle.

Since the ZFS code must always be able to fallback to rw_exit()
and rw_enter() when an rw_tryupgrade() fails; this functionality
isn't critical.  Furthermore, the only potentially performance
sensitive consumer is dmu_zfetch() and no decrease in performance
was observed with this change applied.  See the PR comments for
additional testing details.

Therefore, it is being retired to make the build more robust and
to simplify the rwlock implementation.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #9029

5 years agoLinux 5.3 compat: rw_semaphore owner
Brian Behlendorf [Fri, 12 Jul 2019 20:27:24 +0000 (13:27 -0700)]
Linux 5.3 compat: rw_semaphore owner

Commit https://github.com/torvalds/linux/commit/94a9717b updated the
rwsem's owner field to contain additional flags describing the rwsem's
state.  Rather then update the wrappers to mask out these bits, the
code no longer relies on the owner stored by the kernel.  This does
increase the size of a krwlock_t but it makes the implementation
less sensitive to future kernel changes.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #9029

5 years agoFix lockdep recursive locking false positive in dbuf_destroy
jdike [Wed, 17 Jul 2019 16:18:24 +0000 (12:18 -0400)]
Fix lockdep recursive locking false positive in dbuf_destroy

lockdep reports a possible recursive lock in dbuf_destroy.

It is true that dbuf_destroy is acquiring the dn_dbufs_mtx
on one dnode while holding it on another dnode.  However,
it is impossible for these to be the same dnode because,
among other things,dbuf_destroy checks MUTEX_HELD before
acquiring the mutex.

This fix defines a class NESTED_SINGLE == 1 and changes
that lock to call mutex_enter_nested with a subclass of
NESTED_SINGLE.

In order to make the userspace code compile,
include/sys/zfs_context.h now defines mutex_enter_nested and
NESTED_SINGLE.

This is the lockdep report:

[  122.950921] ============================================
[  122.950921] WARNING: possible recursive locking detected
[  122.950921] 4.19.29-4.19.0-debug-d69edad5368c1166 #1 Tainted: G           O
[  122.950921] --------------------------------------------
[  122.950921] dbu_evict/1457 is trying to acquire lock:
[  122.950921] 0000000083e9cbcf (&dn->dn_dbufs_mtx){+.+.}, at: dbuf_destroy+0x3c0/0xdb0 [zfs]
[  122.950921]
               but task is already holding lock:
[  122.950921] 0000000055523987 (&dn->dn_dbufs_mtx){+.+.}, at: dnode_evict_dbufs+0x90/0x740 [zfs]
[  122.950921]
               other info that might help us debug this:
[  122.950921]  Possible unsafe locking scenario:

[  122.950921]        CPU0
[  122.950921]        ----
[  122.950921]   lock(&dn->dn_dbufs_mtx);
[  122.950921]   lock(&dn->dn_dbufs_mtx);
[  122.950921]
                *** DEADLOCK ***

[  122.950921]  May be due to missing lock nesting notation

[  122.950921] 1 lock held by dbu_evict/1457:
[  122.950921]  #0: 0000000055523987 (&dn->dn_dbufs_mtx){+.+.}, at: dnode_evict_dbufs+0x90/0x740 [zfs]
[  122.950921]
               stack backtrace:
[  122.950921] CPU: 0 PID: 1457 Comm: dbu_evict Tainted: G           O      4.19.29-4.19.0-debug-d69edad5368c1166 #1
[  122.950921] Hardware name: Supermicro H8SSL-I2/H8SSL-I2, BIOS 080011  03/13/2009
[  122.950921] Call Trace:
[  122.950921]  dump_stack+0x91/0xeb
[  122.950921]  __lock_acquire+0x2ca7/0x4f10
[  122.950921]  lock_acquire+0x153/0x330
[  122.950921]  dbuf_destroy+0x3c0/0xdb0 [zfs]
[  122.950921]  dbuf_evict_one+0x1cc/0x3d0 [zfs]
[  122.950921]  dbuf_rele_and_unlock+0xb84/0xd60 [zfs]
[  122.950921]  dnode_evict_dbufs+0x3a6/0x740 [zfs]
[  122.950921]  dmu_objset_evict+0x7a/0x500 [zfs]
[  122.950921]  dsl_dataset_evict_async+0x70/0x480 [zfs]
[  122.950921]  taskq_thread+0x979/0x1480 [spl]
[  122.950921]  kthread+0x2e7/0x3e0
[  122.950921]  ret_from_fork+0x27/0x50

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Jeff Dike <jdike@akamai.com>
Closes #8984

5 years agoFix CONFIG_X86_DEBUG_FPU build failure
Brian Behlendorf [Wed, 17 Jul 2019 16:14:36 +0000 (09:14 -0700)]
Fix CONFIG_X86_DEBUG_FPU build failure

When CONFIG_X86_DEBUG_FPU is defined the alternatives_patched symbol
is pulled in as a dependency which results in a build failure.  To
prevent this undefine CONFIG_X86_DEBUG_FPU to disable the WARN_ON_FPU()
macro and rely on WARN_ON_ONCE debugging checks which were previously
added.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #9041
Closes #9049

5 years agoAdd missing __GFP_HIGHMEM flag to vmalloc
Michael Niewöhner [Wed, 17 Jul 2019 16:09:22 +0000 (18:09 +0200)]
Add missing __GFP_HIGHMEM flag to vmalloc

Make use of __GFP_HIGHMEM flag in vmem_alloc, which is required for
some 32-bit systems to make use of full available memory.
While kernel versions >=4.12-rc1 add this flag implicitly, older
kernels do not.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Sebastian Gottschall <s.gottschall@dd-wrt.com>
Signed-off-by: Michael Niewöhner <foss@mniewoehner.de>
Closes #9031

5 years agoUse zfsctl_snapshot_hold() wrapper
Tomohiro Kusumi [Wed, 17 Jul 2019 16:07:53 +0000 (01:07 +0900)]
Use zfsctl_snapshot_hold() wrapper

zfs_refcount_*() are to be wrapped by zfsctl_snapshot_*() in this file.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Closes #9039

5 years agoMinor style cleanup
Brian Behlendorf [Wed, 17 Jul 2019 00:22:31 +0000 (17:22 -0700)]
Minor style cleanup

Resolve an assortment of style inconsistencies including
use of white space, typos, capitalization, and line wrapping.
There is no functional change.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #9030

5 years agoFix get_special_prop() build failure
Brian Behlendorf [Tue, 16 Jul 2019 21:14:12 +0000 (14:14 -0700)]
Fix get_special_prop() build failure

The cast of the size_t returned by strlcpy() to a uint64_t by the
VERIFY3U can result in a build failure when CONFIG_FORTIFY_SOURCE
is set.  This is due to the additional hardening.  Since the token
is expected to always fit in strval the VERIFY3U has been removed.
If somehow it doesn't, it will still be safely truncated.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Don Brady <don.brady@delphix.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #8999
Closes #9020

5 years agoAdd zfs create dryrun
Mike Gerdts [Tue, 16 Jul 2019 18:19:24 +0000 (13:19 -0500)]
Add zfs create dryrun

Adds the ability to sanity check zfs create arguments and to see the
value of any additional properties that will local to the dataset.  For
example, automation that may need to adjust quota on a parent filesystem
before creating a volume may call `zfs create -nP -V <size> <volume>` to
obtain the value of refreservation.  This adds the following options to
zfs create:

- -n dry-run (no-op)
- -v verbose
- -P parseable (implies verbose)

Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed-by: Jerry Jelinek <jerry.jelinek@joyent.com>
Signed-off-by: Mike Gerdts <mike.gerdts@joyent.com>
Closes #8974

5 years agoLog Spacemap Project
Serapheim Dimitropoulos [Tue, 16 Jul 2019 17:11:49 +0000 (10:11 -0700)]
Log Spacemap Project

= 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
db587941c5ff6dea01932bb78f70db63cf7f38ba

Update vdev_is_spacemap_addressable() for new spacemap encoding
419ba5914552c6185afbe1dd17b3ed4b0d526547

Simplify spa_sync by breaking it up to smaller functions
8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834

Factor metaslab_load_wait() in metaslab_load()
b194fab0fb6caad18711abccaff3c69ad8b3f6d3

Rename range_tree_verify to range_tree_verify_not_present
df72b8bebe0ebac0b20e0750984bad182cb6564a

Change target size of metaslabs from 256GB to 16GB
c853f382db731e15a87512f4ef1101d14d778a55

zdb -L should skip leak detection altogether
21e7cf5da89f55ce98ec1115726b150e19eefe89

vs_alloc can underflow in L2ARC vdevs
7558997d2f808368867ca7e5234e5793446e8f3f

Simplify log vdev removal code
6c926f426a26ffb6d7d8e563e33fc176164175cb

Get rid of space_map_update() for ms_synced_length
425d3237ee88abc53d8522a7139c926d278b4b7f

Introduce auxiliary metaslab histograms
928e8ad47d3478a3d5d01f0dd6ae74a9371af65e

Error path in metaslab_load_impl() forgets to drop ms_sync_lock
8eef997679ba54547f7d361553d21b3291f41ae7

= 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

5 years agoEnable zfs-mount-generator by default
Antonio Russo [Sun, 2 Jun 2019 20:29:37 +0000 (14:29 -0600)]
Enable zfs-mount-generator by default

Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Antonio Russo <antonio.e.russo@gmail.com>
Closes #8750
Closes #8848

5 years agosystemd encryption key support
Antonio Russo [Sun, 2 Jun 2019 12:57:10 +0000 (08:57 -0400)]
systemd encryption key support

Modify zfs-mount-generator to produce a dependency on new
zfs-import-key-*.service units, dynamically created at boot to call
zfs load-key for the encryption root, before attempting to mount any
encrypted datasets.

These units are created by zfs-mount-generator, and RequiresMountsFor on
the keyfile, if present, or call systemd-ask-password if a passphrase is
requested.

This patch includes suggestions from @Fabian-Gruenbichler, @ryanjaeb and
@rlaager, as well an adaptation of @rlaager's script to retry on
incorrect password entry.

Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Antonio Russo <antonio.e.russo@gmail.com>
Closes #8750
Closes #8848

5 years agoDrop redundant POSIX ACL check in zpl_init_acl()
Tomohiro Kusumi [Mon, 15 Jul 2019 23:26:52 +0000 (08:26 +0900)]
Drop redundant POSIX ACL check in zpl_init_acl()

ZFS_ACLTYPE_POSIXACL has already been tested in zpl_init_acl(),
so no need to test again on POSIX ACL access.

Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Closes #9009

5 years agoExport dnode symbols
Brian Behlendorf [Mon, 15 Jul 2019 23:11:55 +0000 (16:11 -0700)]
Export dnode symbols

External consumers such as Lustre require access to the dnode
interfaces in order to correctly manipulate dnodes.

Reviewed-by: James Simmons <uja.ornl@yahoo.com>
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #8994
Closes #9027

5 years agoEnsure dsl_destroy_head() decrypts objsets
Tom Caputi [Mon, 15 Jul 2019 23:08:42 +0000 (16:08 -0700)]
Ensure dsl_destroy_head() decrypts objsets

This patch corrects a small issue where the dsl_destroy_head()
code that runs when the async_destroy feature is disabled would
not properly decrypt the dataset before beginning processing.
If the dataset is not able to be decrypted, the optimization
code now simply does not run and the dataset is completely
destroyed in the DSL sync task.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tom Caputi <tcaputi@datto.com>
Closes #9021

5 years agoDisable unused pathname::pn_path* (unneeded in Linux)
Tomohiro Kusumi [Mon, 15 Jul 2019 20:57:56 +0000 (05:57 +0900)]
Disable unused pathname::pn_path* (unneeded in Linux)

struct pathname is originally from Solaris VFS, and it has been used
in ZoL to merely call VOP from Linux VFS interface without API change,
therefore pathname::pn_path* are unused and unneeded. Technically,
struct pathname is a wrapper for C string in ZoL.

Saves stack a bit on lookup and unlink.

(#if0'd members instead of removing since comments refer to them.)

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Elling <Richard.Elling@RichardElling.com>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Closes #9025