]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/log
FreeBSD/FreeBSD.git
4 years agoFix SC2086 note in zpool.d/smart
Ryan Moeller [Tue, 14 Apr 2020 20:18:23 +0000 (16:18 -0400)]
Fix SC2086 note in zpool.d/smart

./cmd/zpool/zpool.d/smart:78:32:
note: Double quote to prevent globbing and word splitting. [SC2086]

Reported by latest shellcheck on FreeBSD.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10194

4 years agosys/mnttab.h: include sys/stat.h for stat64
alaviss [Tue, 14 Apr 2020 18:47:40 +0000 (18:47 +0000)]
sys/mnttab.h: include sys/stat.h for stat64

Musl libc defined `stat64` as a macro, which causes the build to fail
upon compiling os/linux/getmntany.c due to conflicts between the forward
declaration and the implementation.

This commit fixes that by including <sys/stat.h> in "sys/mnttab.h"
directly.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Hiếu Lê <leorize+oss@disroot.org>
Closes #10195

4 years agoAdd FreeBSD support to OpenZFS
Matthew Macy [Tue, 14 Apr 2020 18:36:28 +0000 (11:36 -0700)]
Add FreeBSD support to OpenZFS

Add the FreeBSD platform code to the OpenZFS repository.  As of this
commit the source can be compiled and tested on FreeBSD 11 and 12.
Subsequent commits are now required to compile on FreeBSD and Linux.
Additionally, they must pass the ZFS Test Suite on FreeBSD which is
being run by the CI.  As of this commit 1230 tests pass on FreeBSD
and there are no unexpected failures.

Reviewed-by: Sean Eric Fagan <sef@ixsystems.com>
Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #898
Closes #8987

4 years agoFix allocation errors, detected using ASAN
Joao Carlos Mendes Luis [Mon, 13 Apr 2020 17:54:41 +0000 (14:54 -0300)]
Fix allocation errors, detected using ASAN

The test for VDEV_TYPE_INDIRECT is done after a memory allocation, and
could return from function without freeing it.  Since we don't need that
allocation yet, just postpone it.

Add a missing free() when buffer is no longer needed.

Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: João Carlos Mendes Luís <jonny@jonny.eng.br>
Closes #10193

4 years agoDisable user space reference tracking
Brian Behlendorf [Mon, 13 Apr 2020 17:51:44 +0000 (10:51 -0700)]
Disable user space reference tracking

The memory and cpu cost of reference count tracking with the current
implementation is significant.  For this reason it has always been
disabled by default for the kmods.  Apply this same default to user
space so ztest doesn't always incur this performance penalty.

Our intention is to re-enable this by default for ztest once the code
has been optimized.  Since we expect to at some point provide a FUSE
implementation we wouldn't want this enabled by default for libzpool.

Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #10189

4 years agoZTS: Fix and change testcase cache_010_neg
alex [Mon, 13 Apr 2020 17:50:41 +0000 (01:50 +0800)]
ZTS: Fix and change testcase cache_010_neg

Commit 379ca9c removed the requirement on aux devices to be block
devices only but the test case cache_010_neg was not updated, making it
fail consistently.

This change changes the test to check that cache devices _can_ be
anything that presents a block interface. The testcase is renamed to
cache_010_pos and the exceptions for known failure removed from the test
runner.

Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reported-by: Richard Elling <Richard.Elling@RichardElling.com>
Signed-off-by: Alex John <alex@stty.io>
Closes #10172

4 years agozvol_write() can use dmu_tx_hold_write_by_dnode()
Matthew Ahrens [Sat, 11 Apr 2020 04:14:01 +0000 (21:14 -0700)]
zvol_write() can use dmu_tx_hold_write_by_dnode()

We can improve the performance of writes to zvols by using
dmu_tx_hold_write_by_dnode() instead of dmu_tx_hold_write().  This
reduces lock contention on the first block of the dnode object, and also
reduces the amount of CPU needed.  The benefit will be highest with
multi-threaded async writes (i.e. writes that don't call zil_commit()).

Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #10184

4 years agoMinor `zstream redup` command fixes
Brian Behlendorf [Sat, 11 Apr 2020 04:10:09 +0000 (21:10 -0700)]
Minor `zstream redup` command fixes

* Fix uninitialized variable in `zstream redup` command.  The
  'rdt.ddt_count' variable is uninitialized because it was
  allocated from the stack and not globally.  Initialize it.
  This was reported by gcc when compiling with debugging enabled.

    zstream_redup.c:157:16: error: 'rdt.ddt_count' may be used
    uninitialized in this function [-Werror=maybe-uninitialized]

* Remove the cmd/zstreamdump/.gitignore file.  It's no longer
  needed now that the zstreamdump command is a script.

Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #10192

4 years agoAdd `zstream redup` command to convert deduplicated send streams
Matthew Ahrens [Fri, 10 Apr 2020 17:39:55 +0000 (10:39 -0700)]
Add `zstream redup` command to convert deduplicated send streams

Deduplicated send and receive is deprecated.  To ease migration to the
new dedup-send-less world, the commit adds a `zstream redup` utility to
convert deduplicated send streams to normal streams, so that they can
continue to be received indefinitely.

The new `zstream` command also replaces the functionality of
`zstreamdump`, by way of the `zstream dump` subcommand.  The
`zstreamdump` command is replaced by a shell script which invokes
`zstream dump`.

The way that `zstream redup` works under the hood is that as we read the
send stream, we build up a hash table which maps from `<GUID, object,
offset> -> <file_offset>`.

Whenever we see a WRITE record, we add a new entry to the hash table,
which indicates where in the stream file to find the WRITE record for
this block. (The key is `drr_toguid, drr_object, drr_offset`.)

For entries other than WRITE_BYREF, we pass them through unchanged
(except for the running checksum, which is recalculated).

For WRITE_BYREF records, we change them to WRITE records.  We find the
referenced WRITE record by looking in the hash table (for the record
with key `drr_refguid, drr_refobject, drr_refoffset`), and then reading
the record header and payload from the specified offset in the stream
file.  This is why the stream can not be a pipe.  The found WRITE record
replaces the WRITE_BYREF record, with its `drr_toguid`, `drr_object`,
and `drr_offset` fields changed to be the same as the WRITE_BYREF's
(i.e. we are writing the same logical block, but with the data supplied
by the previous WRITE record).

This algorithm requires memory proportional to the number of WRITE
records (same as `zfs send -D`), but the size per WRITE record is
relatively low (40 bytes, vs. 72 for `zfs send -D`).  A 1TB send stream
with 8KB blocks (`recordsize=8k`) would use around 5GB of RAM to
"redup".

Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #10124
Closes #10156

4 years agoPersistent L2ARC
George Amanakis [Fri, 10 Apr 2020 17:33:35 +0000 (13:33 -0400)]
Persistent L2ARC

This commit makes the L2ARC persistent across reboots. We implement
a light-weight persistent L2ARC metadata structure that allows L2ARC
contents to be recovered after a reboot. This significantly eases the
impact a reboot has on read performance on systems with large caches.

Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Saso Kiselkov <skiselkov@gmail.com>
Co-authored-by: Jorgen Lundman <lundman@lundman.net>
Co-authored-by: George Amanakis <gamanakis@gmail.com>
Ported-by: Yuxuan Shui <yshuiv7@gmail.com>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes #925
Closes #1823
Closes #2672
Closes #3744
Closes #9582

4 years agoDon't ignore zfs_arc_max below allmem/32
Ryan Moeller [Thu, 9 Apr 2020 22:39:48 +0000 (18:39 -0400)]
Don't ignore zfs_arc_max below allmem/32

Set arc_c_min before arc_c_max so that when zfs_arc_min is set lower
than the default allmem/32 zfs_arc_max can also be set lower.

Add warning messages when tunables are being ignored.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10157
Closes #10158

4 years agoAdd separate field for indicating that spa is in middle of split
Matthew Macy [Thu, 9 Apr 2020 16:59:31 +0000 (09:59 -0700)]
Add separate field for indicating that spa is in middle of split

By default it's not possible to open a device already owned by an
active vdev. It's necessary to make an exception to this for vdev
split. The FreeBSD platform code will make an exception if
spa_is splitting is set to to true.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes #10178

4 years agoLinux 5.7 compat: blk_alloc_queue()
Brian Behlendorf [Thu, 9 Apr 2020 16:16:46 +0000 (09:16 -0700)]
Linux 5.7 compat: blk_alloc_queue()

Commit https://github.com/torvalds/linux/commit/3d745ea5 simplified
the blk_alloc_queue() interface by updating it to take the request
queue as an argument.  Add a wrapper function which accepts the new
arguments and internally uses the available interfaces.

Other minor changes include increasing the Linux-Maximum to 5.6 now
that 5.6 has been released.  It was not bumped to 5.7 because this
release has not yet been finalized and is still subject to change.

Added local 'struct zvol_state_os *zso' variable to zvol_alloc.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #10181
Closes #10187

4 years agoUse vn_io_fault_uiomove on FreeBSD to avoid potential deadlock
Matthew Macy [Wed, 8 Apr 2020 17:30:27 +0000 (10:30 -0700)]
Use vn_io_fault_uiomove on FreeBSD to avoid potential deadlock

Added to prevent a possible deadlock, the following comments from
FreeBSD explain the issue.  The comment describing vn_io_fault_uiomove:

/*
 * Helper function to perform the requested uiomove operation using
 * the held pages for io->uio_iov[0].iov_base buffer instead of
 * copyin/copyout.  Access to the pages with uiomove_fromphys()
 * instead of iov_base prevents page faults that could occur due to
 * pmap_collect() invalidating the mapping created by
 * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or
 * object cleanup revoking the write access from page mappings.
 *
 * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove()
 * instead of plain uiomove().
 */

This used for vn_io_fault which has the following motivation:

/*
 * The vn_io_fault() is a wrapper around vn_read() and vn_write() to
 * prevent the following deadlock:
 *
 * Assume that the thread A reads from the vnode vp1 into userspace
 * buffer buf1 backed by the pages of vnode vp2.  If a page in buf1 is
 * currently not resident, then system ends up with the call chain
 *   vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] ->
 *     vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2)
 * which establishes lock order vp1->vn_lock, then vp2->vn_lock.
 * If, at the same time, thread B reads from vnode vp2 into buffer buf2
 * backed by the pages of vnode vp1, and some page in buf2 is not
 * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock.
 *
 * To prevent the lock order reversal and deadlock, vn_io_fault() does
 * not allow page faults to happen during VOP_READ() or VOP_WRITE().
 * Instead, it first tries to do the whole range i/o with pagefaults
 * disabled. If all pages in the i/o buffer are resident and mapped,
 * VOP will succeed (ignoring the genuine filesystem errors).
 * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do
 * i/o in chunks, with all pages in the chunk prefaulted and held
 * using vm_fault_quick_hold_pages().
 *
 * Filesystems using this deadlock avoidance scheme should use the
 * array of the held pages from uio, saved in the curthread->td_ma,
 * instead of doing uiomove().  A helper function
 * vn_io_fault_uiomove() converts uiomove request into
 * uiomove_fromphys() over td_ma array.
 *
 * Since vnode locks do not cover the whole i/o anymore, rangelocks
 * make the current i/o request atomic with respect to other i/os and
 * truncations.
 */

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes #10177

4 years agoFinish refactoring for ZFS_MODULE_PARAM_CALL
Ryan Moeller [Tue, 7 Apr 2020 17:06:22 +0000 (13:06 -0400)]
Finish refactoring for ZFS_MODULE_PARAM_CALL

Linux and FreeBSD have different parameters for tunable proc handler.
This has prevented FreeBSD from implementing the ZFS_MODULE_PARAM_CALL
macro.

To complete the sharing of ZFS_MODULE_PARAM_CALL declarations, create
per-platform definitions of the parameter list, ZFS_MODULE_PARAM_ARGS.

With the declarations wired up we discovered an incorrect scope prefix
for spa_slop_shift, so this is now fixed.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10179

4 years agolibzfs_pool: Remove unused check for ENOTBLK
alex [Tue, 7 Apr 2020 17:04:40 +0000 (01:04 +0800)]
libzfs_pool: Remove unused check for ENOTBLK

Commit 379ca9c removed the check on aux devices to be block devices also
changing zfs_ioctl(hdl, ZFS_IOC_VDEV_ADD, ...) and
zfs_ioctl(hdl, ZFS_IOC_POOL_CREATE, ...) to never set ENOTBLK. This
change removes the dangling check for ENOTBLK that will never trigger.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reported-by: Richard Elling <Richard.Elling@RichardElling.com>
Signed-off-by: Alex John <alex@stty.io>
Closes #10173

4 years agoZTS: Fix non-portable date format
Ryan Moeller [Mon, 6 Apr 2020 23:07:35 +0000 (19:07 -0400)]
ZTS: Fix non-portable date format

The delegate tests use `date(1)` to generate snapshot names, using
the format '%F-%T-%N' to get nanosecond resolution (since multiple
snapshots may be taken in the same second).  '%N' is not portable, and
causes tests to fail on FreeBSD.

Since the only purpose these timestamps serve is to create a unique
name, simply use $RANDOM instead.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10170

4 years agoAdd 'zfs wait' command
Paul Dagnelie [Wed, 1 Apr 2020 17:02:06 +0000 (10:02 -0700)]
Add 'zfs wait' command

Add a mechanism to wait for delete queue to drain.

When doing redacted send/recv, many workflows involve deleting files
that contain sensitive data. Because of the way zfs handles file
deletions, snapshots taken quickly after a rm operation can sometimes
still contain the file in question, especially if the file is very
large. This can result in issues for redacted send/recv users who
expect the deleted files to be redacted in the send streams, and not
appear in their clones.

This change duplicates much of the zpool wait related logic into a
zfs wait command, which can be used to wait until the internal
deleteq has been drained.  Additional wait activities may be added
in the future.

Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Gallagher <john.gallagher@delphix.com>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes #9707

4 years agoBugfix/fix uio partial copies
Fabio Scaccabarozzi [Wed, 1 Apr 2020 16:48:54 +0000 (18:48 +0200)]
Bugfix/fix uio partial copies

In zfs_write(), the loop continues to the next iteration without
accounting for partial copies occurring in uiomove_iov when
copy_from_user/__copy_from_user_inatomic return a non-zero status.
This results in "zfs: accessing past end of object..." in the
kernel log, and the write failing.

Account for partial copies and update uio struct before returning
EFAULT, leave a comment explaining the reason why this is done.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: ilbsmart <wgqimut@gmail.com>
Signed-off-by: Fabio Scaccabarozzi <fsvm88@gmail.com>
Closes #8673
Closes #10148

4 years agoImprove ZVOL sync write performance by using a taskq
Matthew Ahrens [Tue, 31 Mar 2020 17:50:44 +0000 (10:50 -0700)]
Improve ZVOL sync write performance by using a taskq

== Summary ==

Prior to this change, sync writes to a zvol are processed serially.
This commit makes zvols process concurrently outstanding sync writes in
parallel, similar to how reads and async writes are already handled.
The result is that the throughput of sync writes is tripled.

== Background ==

When a write comes in for a zvol (e.g. over iscsi), it is processed by
calling `zvol_request()` to initiate the operation.  ZFS is expected to
later call `BIO_END_IO()` when the operation completes (possibly from a
different thread).  There are a limited number of threads that are
available to call `zvol_request()` - one one per iscsi client (unless
using MC/S).  Therefore, to ensure good performance, the latency of
`zvol_request()` is important, so that many i/o operations to the zvol
can be processed concurrently.  In other words, if the client has
multiple outstanding requests to the zvol, the zvol should have multiple
outstanding requests to the storage hardware (i.e. issue multiple
concurrent `zio_t`'s).

For reads, and async writes (i.e. writes which can be acknowledged
before the data reaches stable storage), `zvol_request()` achieves low
latency by dispatching the bulk of the work (including waiting for i/o
to disk) to a taskq.  The taskq callback (`zvol_read()` or
`zvol_write()`) blocks while waiting for the i/o to disk to complete.
The `zvol_taskq` has 32 threads (by default), so we can have up to 32
concurrent i/os to disk in service of requests to zvols.

However, for sync writes (i.e. writes which must be persisted to stable
storage before they can be acknowledged, by calling `zil_commit()`),
`zvol_request()` does not use `zvol_taskq`.  Instead it blocks while
waiting for the ZIL write to disk to complete.  This has the effect of
serializing sync writes to each zvol.  In other words, each zvol will
only process one sync write at a time, waiting for it to be written to
the ZIL before accepting the next request.

The same issue applies to FLUSH operations, for which `zvol_request()`
calls `zil_commit()` directly.

== Description of change ==

This commit changes `zvol_request()` to use
`taskq_dispatch_ent(zvol_taskq)` for sync writes, and FLUSh operations.
Therefore we can have up to 32 threads (the taskq threads)
simultaneously calling `zil_commit()`, for a theoretical performance
improvement of up to 32x.

To avoid the locking issue described in the comment (which this commit
removes), we acquire the rangelock from the taskq callback (e.g.
`zvol_write()`) rather than from `zvol_request()`.  This applies to all
writes (sync and async), reads, and discard operations.  This means that
multiple simultaneously-outstanding i/o's which access the same block
can complete in any order.  This was previously thought to be incorrect,
but a review of the block device interface requirements revealed that
this is fine - the order is inherently not defined.  The shorter hold
time of the rangelock should also have a slight performance improvement.

For an additional slight performance improvement, we use
`taskq_dispatch_ent()` instead of `taskq_dispatch()`, which avoids a
`kmem_alloc()` and eliminates a failure mode.  This applies to all
writes (sync and async), reads, and discard operations.

== Performance results ==

We used a zvol as an iscsi target (server) for a Windows initiator
(client), with a single connection (the default - i.e. not MC/S).

We used `diskspd` to generate a workload with 4 threads, doing 1MB
writes to random offsets in the zvol.  Without this change we get
231MB/s, and with the change we get 728MB/s, which is 3.15x the original
performance.

We ran a real-world workload, restoring a MSSQL database, and saw
throughput 2.5x the original.

We saw more modest performance wins (typically 1.5x-2x) when using MC/S
with 4 connections, and with different number of client threads (1, 8,
32).

Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #10163

4 years agoReset l2ad_hand and l2ad_first in l2arc_evict
George Amanakis [Tue, 31 Mar 2020 17:46:48 +0000 (13:46 -0400)]
Reset l2ad_hand and l2ad_first in l2arc_evict

Increasing l2arc_write_size or l2arc_write_boost can result in
l2arc_write_buffers() not having enough space to perform its writes and
panic zio_write_phys().

Instead of resetting l2ad_hand to l2ad_start at the end of
l2arc_write_buffers() and not taking into account a possible
user-mediated increase of l2arc_write_max, we do this in l2arc_evict(),
right after l2arc_write_size() has run. If there is not enough space to
evict (ie we will exceed l2ad_end) we evict to the end of the device,
reset l2ad_hand to l2ad_start, set l2ad_first to 0 and iterate
l2arc_evict(). We avoid infinite iteration of l2arc_evict() by making
sure in l2arc_write_size() that l2ad_start + size does not exceed
l2ad_end.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes #10154

4 years agoZTS: Skip udev actions in zvol_misc when not Linux
Ryan Moeller [Tue, 31 Mar 2020 17:35:14 +0000 (13:35 -0400)]
ZTS: Skip udev actions in zvol_misc when not Linux

udev is only used on Linux.

Skip udev_wait and udev_cleanup when not on Linux.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10165

4 years agoLet default arc_c_max be platform dependent
Ryan Moeller [Fri, 27 Mar 2020 16:14:46 +0000 (12:14 -0400)]
Let default arc_c_max be platform dependent

Linux changed the default max ARC size to 1/2 of physical memory to
deal with shortcomings of the Linux SLUB allocator.  Other platforms
do not require the same logic.

Implement an arc_default_max() function to determine a default max ARC
size in platform code.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10155

4 years agoCompile cityhash code into libzfs
Matthew Ahrens [Fri, 27 Mar 2020 16:11:22 +0000 (09:11 -0700)]
Compile cityhash code into libzfs

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

4 years agoZTS: Wait for free space between quota tests
Ryan Moeller [Thu, 26 Mar 2020 17:48:19 +0000 (13:48 -0400)]
ZTS: Wait for free space between quota tests

And in removal tests, sync the specific pool we are waiting on.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10146

4 years agoRemove checks for null out value in encryption paths
Dirkjan Bussink [Thu, 26 Mar 2020 17:41:57 +0000 (18:41 +0100)]
Remove checks for null out value in encryption paths

These paths are never exercised, as the parameters given are always
different cipher and plaintext `crypto_data_t` pointers.

Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Attila Fueloep <attila@fueloep.org>
Signed-off-by: Dirkjan Bussink <d.bussink@gmail.com>
Closes #9661
Closes #10015

4 years agozfs_get: change time format string from %k to %H
alex [Thu, 26 Mar 2020 15:28:22 +0000 (23:28 +0800)]
zfs_get: change time format string from %k to %H

Issue #10090 reported that snapshots created between midnight and 1 AM
are missing a padded zero in the creation property

This change fixes the bug reported in issue #10090 where snapshots
created between midnight and 1 AM were missing a padded zero in the
creation timestamp output.

The leading zero was missing because the time format string used `%k`
which formats the hour as a decimal number from 0 to 23 where single
digits are preceded by blanks[0] and is fixed by changing it to `%H`
which formats the hour as 00-23.

The difference in output is as below

```
-Thu Mar 26  0:39 2020
+Thu Mar 26 00:39 2020
```

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Alex John <alex@stty.io>
Closes #10090
Closes #10153

4 years agoDeprecate deduplicated send streams
Matthew Ahrens [Wed, 18 Mar 2020 20:31:10 +0000 (13:31 -0700)]
Deprecate deduplicated send streams

Dedup send can only deduplicate over the set of blocks in the send
command being invoked, and it does not take advantage of the dedup table
to do so. This is a very common misconception among not only users, but
developers, and makes the feature seem more useful than it is. As a
result, many users are using the feature but not getting any benefit
from it.

Dedup send requires a nontrivial expenditure of memory and CPU to
operate, especially if the dataset(s) being sent is (are) not already
using a dedup-strength checksum.

Dedup send adds developer burden. It expands the test matrix when
developing new features, causing bugs in released code, and delaying
development efforts by forcing more testing to be done.

As a result, we are deprecating the use of `zfs send -D` and receiving
of such streams.  This change adds a warning to the man page, and also
prints the warning whenever dedup send or receive are used.

In a future release, we plan to:
1. remove the kernel code for generating deduplicated streams
2. make `zfs send -D` generate regular, non-deduplicated streams
3. remove the kernel code for receiving deduplicated streams
4. make `zfs receive` of deduplicated streams process them in userland
   to "re-duplicate" them, so that they can still be received.

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #7887
Closes #10117

4 years agoAvoid core dump on invalid redaction bookmark
Ryan Moeller [Wed, 18 Mar 2020 19:54:12 +0000 (15:54 -0400)]
Avoid core dump on invalid redaction bookmark

libzfs aborts and dumps core on EINVAL from the kernel when trying to
do a redacted send with a bookmark that is not a redaction bookmark.

Move redacted bookmark validation into libzfs.

Check if the bookmark given for redactions is actually a redaction
bookmark.  Print an error message and exit gracefully if it is not.

Don't abort on EINVAL in zfs_send_one.

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10138

4 years agoChanged decimals to integers in the arcstat script
Avatat [Wed, 18 Mar 2020 18:50:45 +0000 (19:50 +0100)]
Changed decimals to integers in the arcstat script

Changed interval value type from decimal to integer,
because of deprecation warning in Python 3.8 and above.
Also changed kstat values type from decimal to integer,
because all the values are integers.

Fixed behavior of arcstat when run without args.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Bartosz Zieba <bartosz@zieba.pro>
Closes #10132
Closes #10142

4 years agoFix zfs_rmnode() unlink / rollback issue
Brian Behlendorf [Wed, 18 Mar 2020 18:47:07 +0000 (11:47 -0700)]
Fix zfs_rmnode() unlink / rollback issue

If a has rollback has occurred while a file is open and unlinked.
Then when the file is closed post rollback it will not exist in the
rolled back version of the unlinked object.  Therefore, the call to
zap_remove_int() may correctly return ENOENT and should be allowed.

Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #6812
Closes #9739

4 years agoFix cstyle warnings
Brian Behlendorf [Tue, 17 Mar 2020 22:42:27 +0000 (15:42 -0700)]
Fix cstyle warnings

Fix minor cstyle warnings accidentally introduced by 7145123b.

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #10143

4 years agoSeparate warning for incomplete and corrupt streams
Paul Dagnelie [Tue, 17 Mar 2020 17:30:33 +0000 (10:30 -0700)]
Separate warning for incomplete and corrupt streams

This change adds a separate return code to zfs_ioc_recv that is used
for incomplete streams, in addition to the existing return code for
streams that contain corruption.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes #10122

4 years agoICP: gcm-avx: Support architectures lacking the MOVBE instruction
Attila Fülöp [Tue, 17 Mar 2020 17:24:38 +0000 (18:24 +0100)]
ICP: gcm-avx: Support architectures lacking the MOVBE instruction

There are a couple of x86_64 architectures which support all needed
features to make the accelerated GCM implementation work but the
MOVBE instruction. Those are mainly Intel Sandy- and Ivy-Bridge
and AMD Bulldozer, Piledriver, and Steamroller.

By using MOVBE only if available and replacing it with a MOV
followed by a BSWAP if not, those architectures now benefit from
the new GCM routines and performance is considerably better
compared to the original implementation.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Adam D. Moss <c@yotes.com>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Followup #9749
Closes #10029

4 years agoAdd option for forcible unmounting dataset while receiving snapshot.
Mariusz Zaborski [Tue, 17 Mar 2020 17:08:32 +0000 (18:08 +0100)]
Add option for forcible unmounting dataset while receiving snapshot.

Currently when the dataset is in use we can't receive snapshots.

    zfs send test/1@asd | zfs recv -FM test/2
    cannot unmount '/test/2': Device busy

This commits add option 'M' which attempts to forcibly unmount the
dataset.  Thanks to this we can enforce receiving snapshots in a
single step.

Note that this functionality is not supported on Linux because the
VFS will prevent active mounted filesystems from being unmounted,
even with the force option.  This is the intended VFS behavior.

Test cases were added to verify the expected behavior based on
the platform.

Discussed-with: Pawel Jakub Dawidek <pjd@FreeBSD.org>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Allan Jude <allanjude@freebsd.org>
External-issue: https://reviews.freebsd.org/D22306
Closes #9904

4 years agoZTS: Use default_cleanup_noexit where needed
Ryan Moeller [Tue, 17 Mar 2020 16:55:18 +0000 (12:55 -0400)]
ZTS: Use default_cleanup_noexit where needed

And add log_pass appropriately.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10136

4 years agoExit status 256+signum is actually baked in to ksh
Ryan Moeller [Tue, 17 Mar 2020 16:49:58 +0000 (12:49 -0400)]
Exit status 256+signum is actually baked in to ksh

While #10121 did fix the signal numbers for FreeBSD/Darwin, it
incorrectly changed the expected encoding of exit status for commands
that exited on a signal.  The encoding 256+signum is a feature of the
shell.  Only the signal numbers themselves are platform-dependent.

Always use the encoding 256+signum when checking exit status for
signal exits.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10137

4 years agolibzfs: Fix bounds checks for float parsing
Ryan Moeller [Mon, 16 Mar 2020 18:56:29 +0000 (14:56 -0400)]
libzfs: Fix bounds checks for float parsing

UINT64_MAX is not exactly representable as a double.

The closest representation is UINT64_MAX + 1, so we can use a >=
comparison instead of > for the bounds check.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10127

4 years agoImprove zfs receive performance by batching writes
Matthew Ahrens [Mon, 16 Mar 2020 18:51:56 +0000 (11:51 -0700)]
Improve zfs receive performance by batching writes

For each WRITE record in the stream, `zfs receive` creates a DMU
transaction (`dmu_tx_create()`) and writes this block's data into the
object.  If per-block overheads (as opposed to per-byte overheads)
dominate performance (as is often the case with small recordsize), the
per-dmu-transaction overheads can be significant.  For example, in some
workloads the `receieve_writer` thread is 100% on CPU, and more than
half of its CPU time is in these per-tx routines (e.g.
dmu_tx_hold_write, dmu_tx_assign, dmu_tx_commit).

To improve performance of `zfs receive`, this commit batches WRITE
records which are to nearby offsets of the same object, and uses one DMU
transaction to write them all.  By default the batch size is 1MB, which
for recordsize=8K reduces the number of DMU transactions by 128x for
full send streams (incrementals will depend on how "clumpy" the changed
blocks are).

This commit improves the performance of `dd if=stream | zfs recv`
from 78,800 blocks/sec to 98,100 blocks/sec (25% improvement).

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #10099

4 years agoRemove CI builder customization from TEST
Brian Behlendorf [Mon, 16 Mar 2020 17:46:03 +0000 (10:46 -0700)]
Remove CI builder customization from TEST

The default options are reasonable for all of the CI builders.

* TEST_XFSTESTS_SKIP=yes  - This is already the default.
* TEST_ZTEST_TIMEOUT=3600 - Increased ztest run time only increases
  code coverage by a small degree.  Default 900s runs are sufficient.
* Disabling certain tests on 32-bit builders is no longer needed.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Kjeld Schouten <kjeld@schouten-lebbing.nl>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #10129

4 years agoZTS: Update flaky tests in zts-report
Ryan Moeller [Fri, 13 Mar 2020 16:29:10 +0000 (12:29 -0400)]
ZTS: Update flaky tests in zts-report

Some tests which pass on FreeBSD but fail on Linux had been put in the
"maybe" set.  Move these back to "known" under an "if Linux" check so
the expected outcome is clear.

Add some tests that have been found to be flaky on FreeBSD stable/12
to the "maybe" set.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10120

4 years agodmu_objset_from_ds must be called with dp_config_rwlock held
Matthew Ahrens [Thu, 12 Mar 2020 17:55:02 +0000 (10:55 -0700)]
dmu_objset_from_ds must be called with dp_config_rwlock held

The normal lock order is that the dp_config_rwlock must be held before
the ds_opening_lock.  For example, dmu_objset_hold() does this.
However, dmu_objset_open_impl() is called with the ds_opening_lock held,
and if the dp_config_rwlock is not already held, it will attempt to
acquire it.  This may lead to deadlock, since the lock order is
reversed.

Looking at all the callers of dmu_objset_open_impl() (which is
principally the callers of dmu_objset_from_ds()), almost all callers
already have the dp_config_rwlock.  However, there are a few places in
the send and receive code paths that do not.  For example:
dsl_crypto_populate_key_nvlist, send_cb, dmu_recv_stream,
receive_write_byref, redact_traverse_thread.

This commit resolves the problem by requiring all callers ot
dmu_objset_from_ds() to hold the dp_config_rwlock.  In most cases, the
code has been restructured such that we call dmu_objset_from_ds()
earlier on in the send and receive processes, when we already have the
dp_config_rwlock, and save the objset_t until we need it in the middle
of the send or receive (similar to what we already do with the
dsl_dataset_t).  Thus we do not need to acquire the dp_config_rwlock in
many new places.

I also cleaned up code in dmu_redact_snap() and send_traverse_thread().

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Paul Zuchowski <pzuchowski@datto.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #9662
Closes #10115

4 years agoFix infinite scan on a pool with only special allocations
Alexander Motin [Thu, 12 Mar 2020 17:52:03 +0000 (13:52 -0400)]
Fix infinite scan on a pool with only special allocations

Attempt to run scrub or resilver on a new pool containing only special
allocations (special vdev added on creation) caused infinite loop
because of dsl_scan_should_clear() limiting memory usage to 5% of pool
size, which it calculated accounting only normal allocation class.

Addition of special and just in case dedup classes fixes the issue.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes #10106
Closes #8694

4 years agoZTS: Use correct signal numbers for status checks
Ryan Moeller [Thu, 12 Mar 2020 17:50:51 +0000 (13:50 -0400)]
ZTS: Use correct signal numbers for status checks

Different operating systems encode exit status in different ways.
The logapi shell library assumes the Solaris meaning of exit codes,
which is not correct on other platforms.

Define the needed constants according to the platform we are running
on and use those to decode process exit status.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10121

4 years agoZTS: Test boundary conditions in alloc_class_012
Ryan Moeller [Thu, 12 Mar 2020 17:50:01 +0000 (13:50 -0400)]
ZTS: Test boundary conditions in alloc_class_012

Issue #9142 describes an error in the checks for device removal that
can prevent removal of special allocation class vdevs in some
situations.

Enhance alloc_class/alloc_class_012_pos to check situations where this
bug occurs.

Update zts-report with knowledge of issue #9142.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10116
Issue #9142

4 years agoZTS: Wait for free space between write_dirs tests
Ryan Moeller [Thu, 12 Mar 2020 17:48:46 +0000 (13:48 -0400)]
ZTS: Wait for free space between write_dirs tests

Cleanup for write_dirs involves destroying a dataset filling a pool
and then recreating the dataset for the next test.  Due to the
asynchronous nature of free space accounting, recreating the dataset
can fail for lack of space, causing problems for the next test.

Add wait_freeing $TESTPOOL to wait for the space to be freed and then
sync_pool $TESTPOOL to update the space accounting before attempting
to recreate the test filesystem.

Only use a single disk to create the pool.  Make it a small file so it
does not take too long to fill.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10112

4 years agoPrevent race condition in dnode_dest (#10101)
John Poduska [Thu, 12 Mar 2020 17:25:56 +0000 (13:25 -0400)]
Prevent race condition in dnode_dest (#10101)

dnode_special_close() waits for the refcount of dn_holds to go to zero
without holding the dn_mtx. dnode_rele_and_unlock() does the final
remove to dn_holds with dn_mtx being held:

refs = zfs_refcount_remove(&dn->dn_holds, tag);
mutex_exit(&dn->dn_mtx);

So, there is a race condition after the remove until dn_mtx is
dropped. During that time, dnode_destroy() can get called, which ends
up in dnode_dest() calling mutex_destroy() and a panic since the lock
is still held.

This change adds a condvar to wait for the final dnode_rele_and_unlock()
to release the dn_mtx before calling dnode_destroy().

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: John Poduska <jpoduska@datto.com>
Closes #7814
Closes #10101

4 years agoPrevent deadlock in arc_read in Linux memory reclaim callback
Mark Roper [Thu, 12 Mar 2020 17:24:43 +0000 (13:24 -0400)]
Prevent deadlock in arc_read in Linux memory reclaim callback

Using zfs with Lustre, an arc_read can trigger kernel memory allocation
that in turn leads to a memory reclaim callback and a deadlock within a
single zfs process. This change uses spl_fstrans_mark and
spl_trans_unmark to prevent the reclaim attempt and the deadlock
(https://zfsonlinux.topicbox.com/groups/zfs-devel/T4db2c705ec1804ba).
The stack trace observed is:

    __schedule at ffffffff81610f2e
    schedule at ffffffff81611558
    schedule_preempt_disabled at ffffffff8161184a
    __mutex_lock at ffffffff816131e8
    arc_buf_destroy at ffffffffa0bf37d7 [zfs]
    dbuf_destroy at ffffffffa0bfa6fe [zfs]
    dbuf_evict_one at ffffffffa0bfaa96 [zfs]
    dbuf_rele_and_unlock at ffffffffa0bfa561 [zfs]
    dbuf_rele_and_unlock at ffffffffa0bfa32b [zfs]
    osd_object_delete at ffffffffa0b64ecc [osd_zfs]
    lu_object_free at ffffffffa06d6a74 [obdclass]
    lu_site_purge_objects at ffffffffa06d7fc1 [obdclass]
    lu_cache_shrink_scan at ffffffffa06d81b8 [obdclass]
    shrink_slab at ffffffff811ca9d8
    shrink_node at ffffffff811cfd94
    do_try_to_free_pages at ffffffff811cfe63
    try_to_free_pages at ffffffff811d01c4
    __alloc_pages_slowpath at ffffffff811be7f2
    __alloc_pages_nodemask at ffffffff811bf3ed
    new_slab at ffffffff81226304
    ___slab_alloc at ffffffff812272ab
    __slab_alloc at ffffffff8122740c
    kmem_cache_alloc at ffffffff81227578
    spl_kmem_cache_alloc at ffffffffa048a1fd [spl]
    arc_buf_alloc_impl at ffffffffa0befba2 [zfs]
    arc_read at ffffffffa0bf0924 [zfs]
    dbuf_read at ffffffffa0bf9083 [zfs]
    dmu_buf_hold_by_dnode at ffffffffa0c04869 [zfs]

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Mark Roper <markroper@gmail.com>
Closes #9987

4 years agozloop.sh should call ZDB with pool name
Olaf Faaland [Wed, 11 Mar 2020 17:02:23 +0000 (10:02 -0700)]
zloop.sh should call ZDB with pool name

Commit 54007c79 introduced an error, changing the final
argument to $ZDB from ztest to $ZTEST.  This argument
indicates the pool name, not the script, and so should
not have been changed.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Olaf Faaland <faaland1@llnl.gov>
Closes #10118

4 years agoZTS: Add a failsafe callback to run after each test
Ryan Moeller [Tue, 10 Mar 2020 18:00:56 +0000 (14:00 -0400)]
ZTS: Add a failsafe callback to run after each test

Tests that get killed do not have an opportunity to clean up.

There are many bad states this can leave the system in, but of
particular gravity is when zinject has been used to induce bad
behavior for one or more of the test disks.

Create a failsafe mechanism in test-runner.py that runs a callback
script after every test. The script is common to all tests so all
tests benefit from the protection.

Add an obligatory `zinject -c all` to clear all zinject state after
every test case is run.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10096

4 years agoImprove zfs send performance by bypassing the ARC
Matthew Ahrens [Tue, 10 Mar 2020 17:51:04 +0000 (10:51 -0700)]
Improve zfs send performance by bypassing the ARC

When doing a zfs send on a dataset with small recordsize (e.g. 8K),
performance is dominated by the per-block overheads.  This is especially
true with `zfs send --compressed`, which further reduces the amount of
data sent, for the same number of blocks.  Several threads are involved,
but the limiting factor is the `send_prefetch` thread, which is 100% on
CPU.

The main job of the `send_prefetch` thread is to issue zio's for the
data that will be needed by the main thread.  It does this by calling
`arc_read(ARC_FLAG_PREFETCH)`.  This has an immediate cost of creating
an arc_hdr, which takes around 14% of one CPU.  It also induces later
costs by other threads:

 * Since the data was only prefetched, dmu_send()->dmu_dump_write() will
   need to call arc_read() again to get the data.  This will have to
   look up the arc_hdr in the hash table and copy the data from the
   scatter ABD in the arc_hdr to a linear ABD in arc_buf.  This takes
   27% of one CPU.

 * dmu_dump_write() needs to arc_buf_destroy()  This takes 11% of one
   CPU.

 * arc_adjust() will need to evict this arc_hdr, taking about 50% of one
   CPU.

All of these costs can be avoided by bypassing the ARC if the data is
not already cached.  This commit changes `zfs send` to check for the
data in the ARC, and if it is not found then we directly call
`zio_read()`, reading the data into a linear ABD which is used by
dmu_dump_write() directly.

The performance improvement is best expressed in terms of how many
blocks can be processed by `zfs send` in one second.  This change
increases the metric by 50%, from ~100,000 to ~150,000.  When the amount
of data per block is small (e.g. 2KB), there is a corresponding
reduction in the elapsed time of `zfs send >/dev/null` (from 86 minutes
to 58 minutes in this test case).

In addition to improving the performance of `zfs send`, this change
makes `zfs send` not pollute the ARC cache.  In most cases the data will
not be reused, so this allows us to keep caching useful data in the MRU
(hit-once) part of the ARC.

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #10067

4 years agoZTS: Simplify some libtest functions
Ryan Moeller [Tue, 10 Mar 2020 17:44:14 +0000 (13:44 -0400)]
ZTS: Simplify some libtest functions

Don't echo the results of arithmetic expressions, it's not necessary.

Use hw.clockrate sysctl to get CPU freq instead of parsing dmesg.boot
for a line that might not even be there anymore.

Reduce bookkeeping in fill_fs, making it easier to follow.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10113

4 years agoFix zfs-functions packaging bug
Richard Laager [Thu, 6 Feb 2020 17:28:20 +0000 (09:28 -0800)]
Fix zfs-functions packaging bug

This fixes a bug where the generated zfs-functions was being included
along with original zfs-functions.in in the make dist tarball.  This
caused an unfortunate series of events during build/packaging that
resulted in the RPM-installed /etc/zfs/zfs-functions listing the
paths as:

ZFS="/usr/local/sbin/zfs"
ZED="/usr/local/sbin/zed"
ZPOOL="/usr/local/sbin/zpool"

When they should have been:

ZFS="/sbin/zfs"
ZED="/sbin/zed"
ZPOOL="/sbin/zpool"

This affects init.d (non-systemd) distros like CentOS 6.

/etc/default/zfs and /etc/zfs/zfs-functions are also used by the
initramfs, so they need to be built even when init.d support is not.
They have been moved to the (new) etc/default and (existing) etc/zfs
source directories, respectively.

Fixes: #9443
Co-authored-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Richard Laager <rlaager@wiktel.com>
4 years agoinitramfs: Eliminate substitutions
Richard Laager [Sun, 23 Feb 2020 00:09:55 +0000 (18:09 -0600)]
initramfs: Eliminate substitutions

These are now handled in zfs-functions, so this is all duplicative and
unnecessary.

Signed-off-by: Richard Laager <rlaager@wiktel.com>
4 years agoDelete built init scripts in make clean
Richard Laager [Thu, 6 Feb 2020 00:02:43 +0000 (18:02 -0600)]
Delete built init scripts in make clean

Previously, they were being deleted in make distclean.  This brings it
in line with the example:
https://www.gnu.org/software/automake/manual/html_node/Scripts.html

Signed-off-by: Richard Laager <rlaager@wiktel.com>
4 years agoMake init scripts depend on Makefile
Richard Laager [Thu, 6 Feb 2020 00:01:49 +0000 (18:01 -0600)]
Make init scripts depend on Makefile

This brings it in line with the example:
https://www.gnu.org/software/automake/manual/html_node/Scripts.html

This way, if the substitution code is changed, they should update.

Signed-off-by: Richard Laager <rlaager@wiktel.com>
4 years agoSystemd mount generator: don't fail keyload from file if already loaded
InsanePrawn [Mon, 9 Mar 2020 18:09:09 +0000 (19:09 +0100)]
Systemd mount generator: don't fail keyload from file if already loaded

Previously the generated keyload units for encryption roots with
keylocation=file://* didn't contain the code to detect if the key
was already loaded and would be marked failed in such situations.

Move the code to check whether the key is already loaded
from keylocation=prompt handling to general key loading code.

Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: InsanePrawn <insane.prawny@gmail.com>
Closes #10103

4 years agoZTS: Another round of changes for FreeBSD
Ryan Moeller [Fri, 6 Mar 2020 17:31:32 +0000 (12:31 -0500)]
ZTS: Another round of changes for FreeBSD

Highlights:
* is_linux -> is_illumos swaps
* make block_device_wait more clever when paths are given
* slightly optimize default_cleanup_noexit
* remove platform differences in user_run
* temporarily expect non-libfetch behavior for keylocation=/foo/bar
* fix sharenfs exceptions
* don't test multihost property
* fix misc broken platform checks
* clear zinjected faults in removal_resume_export callback

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10092

4 years agoChange default to overlay=on
Ryan Moeller [Fri, 6 Mar 2020 17:28:19 +0000 (12:28 -0500)]
Change default to overlay=on

Filesystems allow overlay mounts by default on FreeBSD and Linux.

Respect the native convention by switching the default to overlay=on,
while retaining the option to turn the property off for compatibility
with other operating systems' conventions.

Update documentation and tests accordingly.

Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10030

4 years agoZTS: Update zts-report exceptions for FreeBSD
Ryan Moeller [Fri, 6 Mar 2020 17:26:38 +0000 (12:26 -0500)]
ZTS: Update zts-report exceptions for FreeBSD

The new zfs_sync_trim_* tests are skipped on FreeBSD.
Both of the previously failing tests are now passing.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10105

4 years agoZTS: Speed up write_dirs cleanup
Brian Behlendorf [Wed, 4 Mar 2020 23:12:12 +0000 (15:12 -0800)]
ZTS: Speed up write_dirs cleanup

The write_dirs tests fill a filesystem with a bunch of files until it
is full.  In cleanup the files are truncated and removed individually.
These tests already take a while to run.

It is quicker and easier to destroy the whole dataset and create a new
one to replace it in the cleanup functions.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10098

4 years agoZTS: Add missing quotes
Brian Behlendorf [Wed, 4 Mar 2020 23:10:45 +0000 (15:10 -0800)]
ZTS: Add missing quotes

`default_setup` takes a disk list as the first argument and has
optional additional arguments that control secondary functionality.
A couple of test setups mistakenly call `default_setup $DISKS`.

Add quotes so the second and subsequent disks are correctly included
in the pool as vdevs rather than triggering unwanted behavior from
`default_setup`.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10097

4 years agoZTS: Add zts-report exceptions for FreeBSD
Brian Behlendorf [Wed, 4 Mar 2020 23:09:40 +0000 (15:09 -0800)]
ZTS: Add zts-report exceptions for FreeBSD

There are three tests we expect to fail only on FreeBSD.
* link_count never exits and eventually times out:
 - @amotin tells me this test is probably not applicable to us
 - Skip on FreeBSD
* userobj feature does not activate immediately after pool upgrade
 - low impact; we are aware of this issue
* removal does not appear to condense on export
 - low impact; we are aware of this issue

Additionally removal_with_zdb passes on FreeBSD, so it is moved to
"maybe".

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10093

4 years agozio: dprintf_bp() if errors > 0 in zfs_blkptr_verify()
Brian Behlendorf [Wed, 4 Mar 2020 23:08:41 +0000 (15:08 -0800)]
zio: dprintf_bp() if errors > 0 in zfs_blkptr_verify()

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

4 years agoZTS: Test the correct filesystem_limits behavior
Brian Behlendorf [Wed, 4 Mar 2020 23:07:52 +0000 (15:07 -0800)]
ZTS: Test the correct filesystem_limits behavior

See issue #8226: Property filesystem_limit does not work as documented

There have been previous attempts to fix the behavior on Linux, but so
far the issue is still open.  See PRs #8228, #8280.

The existing tests pass for the incorrect behavior.  This is a problem
on FreeBSD; we are failing the tests because we implement the feature
correctly.

I have adapted the tests based on the work by @loli10k in #8280 and
extended the changes to fix the snapshot_limit test as well.

Linux now fails these tests, so entries linking to the issue have been
added to the "maybe" group in zts-report.py.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10082

4 years agoAdd trim support to zpool wait
Brian Behlendorf [Wed, 4 Mar 2020 23:07:11 +0000 (15:07 -0800)]
Add trim support to zpool wait

Manual trims fall into the category of long-running pool activities
which people might want to wait synchronously for. This change adds
support to 'zpool wait' for waiting for manual trim operations to
complete. It also adds a '-w' flag to 'zpool trim' which can be used to
turn 'zpool trim' into a synchronous operation.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Signed-off-by: John Gallagher <john.gallagher@delphix.com>
Closes #10071

4 years agoImprove performance of zio_taskq_member
Matthew Ahrens [Tue, 3 Mar 2020 18:29:38 +0000 (10:29 -0800)]
Improve performance of zio_taskq_member

__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

4 years agoZTS: Provide for nested cleanup routines
Ryan Moeller [Tue, 3 Mar 2020 18:28:09 +0000 (13:28 -0500)]
ZTS: Provide for nested cleanup routines

Shared test library functions lack a simple way to ensure proper
cleanup in the event of a failure.  The `log_onexit` cleanup pattern
cannot be used in library functions because it uses one global
variable to store the cleanup command.

An example of where this is a serious issue is when a tunable that
artifically stalls kernel progress gets activated and then some check
fails.  Unless the caller knows about the tunable and sets it back,
the system will be left in a bad state.

To solve this problem, turn the global cleanup variable into a stack.
Provide push and pop functions to add additional cleanup steps and
remove them after it is safe again.

The first use of this new functionality is in attempt_during_removal,
which sets REMOVAL_SUSPEND_PROGRESS.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10080

4 years agoMake spa_history_zone platform-dependent in kernel
Ryan Moeller [Mon, 2 Mar 2020 17:43:30 +0000 (12:43 -0500)]
Make spa_history_zone platform-dependent in kernel

This function should only return "linux" on Linux.

Move the kernel part of the function out of common code.
Fix the tests for FreeBSD.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10079

4 years agoZTS: Change issue URL template to OpenZFS org
Ryan Moeller [Mon, 2 Mar 2020 17:42:22 +0000 (12:42 -0500)]
ZTS: Change issue URL template to OpenZFS org

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10081

4 years agoDon't open zfs control device exclusively
Matthew Macy [Fri, 28 Feb 2020 22:54:14 +0000 (14:54 -0800)]
Don't open zfs control device exclusively

With the FreeBSD platform changes that were made for #10073
it is no longer necessary on FreeBSD to open the control device
exclusively to get onexit callbacks invoked.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes #10076

4 years agoDon't call zrele on passed zp in zfs_xattr_owner_unlinked on FreeBSD
Matthew Macy [Fri, 28 Feb 2020 22:53:18 +0000 (14:53 -0800)]
Don't call zrele on passed zp in zfs_xattr_owner_unlinked on FreeBSD

FreeBSD has a somewhat more cumbersome locking and refcounting
protocol for the platform counterpart to znode. We need to not call
zrele on the passed zp, but do need to do so on any intermediate zp.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes #10075

4 years agoRe-share zfsdev_getminor and zfs_onexit_fd_hold
Matthew Macy [Fri, 28 Feb 2020 22:50:32 +0000 (14:50 -0800)]
Re-share zfsdev_getminor and zfs_onexit_fd_hold

By adding a zfs_file_private accessor to the common
interfaces and some extensions to FreeBSD platform
code it is now possible to share the implementations
for the aforementioned functions.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes #10073

4 years agoImprove zfs destroy performance with zio_t-free zio_free()
Matthew Ahrens [Fri, 28 Feb 2020 22:49:44 +0000 (14:49 -0800)]
Improve zfs destroy performance with zio_t-free zio_free()

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

4 years agoZTS: Fixup shebang in rsend_016, add to common.run
Ryan Moeller [Fri, 28 Feb 2020 17:48:29 +0000 (12:48 -0500)]
ZTS: Fixup shebang in rsend_016, add to common.run

All other ksh scripts use /bin/ksh in the shebang.

Make rsend_016_neg consistent with the rest of the suite.

The test also was absent from any runfiles. Add it to common.run.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10051

4 years agoZTS: Eliminate partitioning from zpool_add
Ryan Moeller [Fri, 28 Feb 2020 17:46:51 +0000 (12:46 -0500)]
ZTS: Eliminate partitioning from zpool_add

Use file vdevs if we are short on $DISKS.
Also fixed vol recursion for FreeBSD in 004.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10060

4 years agoFix CONFIG_MODULES=no Linux kernel config
Brian Behlendorf [Fri, 28 Feb 2020 17:23:48 +0000 (09:23 -0800)]
Fix CONFIG_MODULES=no Linux kernel config

When configuring as builtin (--enable-linux-builtin) for kernels
without loadable module support (CONFIG_MODULES=n) only the object
file is created.  Never a loadable kmod.

Update ZFS_LINUX_TRY_COMPILE to handle this in a manor similar to
the ZFS_LINUX_TEST_COMPILE_ALL macro.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #9887
Closes #10063

4 years agoLinux 5.5 compat: blkg_tryget()
Brian Behlendorf [Fri, 28 Feb 2020 16:58:39 +0000 (08:58 -0800)]
Linux 5.5 compat: blkg_tryget()

Commit https://github.com/torvalds/linux/commit/9e8d42a0f accidentally
converted the static inline function blkg_tryget() to GPL-only for
kernels built with CONFIG_PREEMPT_RCU=y and CONFIG_BLK_CGROUP=y.

Resolve the build issue by providing our own equivalent functionality
when needed which uses rcu_read_lock_sched() internally as before.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #9745
Closes #10072

4 years agoarc_summary: Make get_descriptions per platform
Ryan Moeller [Fri, 28 Feb 2020 01:15:06 +0000 (20:15 -0500)]
arc_summary: Make get_descriptions per platform

Linux uses modinfo to get tunables descriptions, FreeBSD has to use
sysctl.

Move the existing function definition so it is defined that way on
Linux, and add a definition in terms of sysctl for FreeBSD.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10062

4 years agopyzfs: Add constants for platform-specific errnos
Ryan Moeller [Fri, 28 Feb 2020 01:14:21 +0000 (20:14 -0500)]
pyzfs: Add constants for platform-specific errnos

FreeBSD doesn't have EBADE, ECHRNG, or ETIME.

Add constants for these and set them appropriately for the platform.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10061

4 years agoConsolidate arc_buf allocation checks
Matthew Macy [Fri, 28 Feb 2020 01:12:44 +0000 (17:12 -0800)]
Consolidate arc_buf allocation checks

The following check currently occurs in three separate locations
in dbuf.c.  This change consolidates those checks in to the
dbuf_alloc_arcbuf_from_arcbuf() function.

if (arc_is_encrypted(data)) {
...
} else if (compress_type != ZIO_COMPRESS_OFF) {
...
} else {
...
}

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes #10057

4 years agoZTS: Misc fixes for FreeBSD
Ryan Moeller [Thu, 27 Feb 2020 17:38:34 +0000 (12:38 -0500)]
ZTS: Misc fixes for FreeBSD

* Set geom debug flags in corrupt_blocks_at_level
* Use the right time zone for history tests
* Add missing commands.cfg entry for diskinfo
* Rewrite get_last_txg_synced to use zdb
* Don't check ulimits for sparse files
* Suspend removal before removing a vdev, not after

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10054

4 years agoZTS: Fix zfs_receive_004_neg
Matthew Ahrens [Thu, 27 Feb 2020 17:37:34 +0000 (09:37 -0800)]
ZTS: Fix zfs_receive_004_neg

`zfs recv` of an incremental stream that already exists is ignored, with
a message like:

    receiving incremental stream of pool/fs@incsnap into pool/fs@incsnap
    snap testpool/testfs@incsnap already exists; ignoring

And the command exits successfully (exit code 0).

The zfs_receive_004_neg test is expecting that a this case will fail,
with nonzero exit code.

The fix is to remove this specific command from the test case.  This
lets us check that the remaining commands do in fact fail.

Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #10055

4 years agoLinux 5.6 compat: time_t
Brian Behlendorf [Wed, 26 Feb 2020 21:18:07 +0000 (13:18 -0800)]
Linux 5.6 compat: time_t

As part of the Linux kernel's y2038 changes the time_t type has been
fully retired.  Callers are now required to use the time64_t type.

Rather than move to the new type, I've removed the few remaining
places where a time_t is used in the kernel code.  They've been
replaced with a uint64_t which is already how ZFS internally
handled these values.

Going forward we should work towards updating the remaining user
space time_t consumers to the 64-bit interfaces.

Reviewed-by: Matthew Macy <mmacy@freebsd.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #10052
Closes #10064

4 years agoLinux 5.6 compat: ktime_get_raw_ts64()
Brian Behlendorf [Wed, 26 Feb 2020 20:42:33 +0000 (12:42 -0800)]
Linux 5.6 compat: ktime_get_raw_ts64()

The getrawmonotonic() and getrawmonotonic64() interfaces have been
fully retired.  Update gethrtime() to use the replacement interface
ktime_get_raw_ts64() which was introduced in the 4.18 kernel.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #10052
Closes #10064

4 years agoRefactor dnode dirty context from dbuf_dirty
Matthew Macy [Thu, 27 Feb 2020 00:09:17 +0000 (16:09 -0800)]
Refactor dnode dirty context from dbuf_dirty

* Add dedicated donde_set_dirtyctx routine.
* Add empty dirty record on destroy assertion.
* Make much more extensive use of the SET_ERROR macro.

Reviewed-by: Will Andrews <wca@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes #9924

4 years agoZTS: Fix zfs_copies_002_pos
Ryan Moeller [Wed, 26 Feb 2020 22:29:13 +0000 (17:29 -0500)]
ZTS: Fix zfs_copies_002_pos

The function `get_used_prop` does not exist.

Use `get_prop used` instead.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10059

4 years agoZTS: Adapt casenorm tests for FreeBSD
Ryan Moeller [Wed, 26 Feb 2020 16:41:30 +0000 (11:41 -0500)]
ZTS: Adapt casenorm tests for FreeBSD

Several casenorm tests pass on FreeBSD but are expected to fail on
Linux.

Move the passing tests from "fail" to "maybe" so that passing on
FreeBSD is not unexpected.

Invert platform logic so FreeBSD doesn't use illumos-only zlook.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10050

4 years agoZTS: Misc fixes for FreeBSD
Ryan Moeller [Wed, 26 Feb 2020 00:23:27 +0000 (19:23 -0500)]
ZTS: Misc fixes for FreeBSD

* Check for mountd in is_shared to avoid timeout when not running
* Enhance robustness of some cleanup functions
* Simplify atime lookup
* Skip sharenfs validation for now
* Don't add mountpoint property to inheritance validation on FreeBSD

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10047

4 years agoAdd missing newline after zfs redact help message
Ryan Moeller [Wed, 26 Feb 2020 00:20:50 +0000 (19:20 -0500)]
Add missing newline after zfs redact help message

Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10045

4 years agoZTS: zed_start should not fail if zed is already running
Olaf Faaland [Wed, 26 Feb 2020 00:02:10 +0000 (16:02 -0800)]
ZTS: zed_start should not fail if zed is already running

zed_start may be called in places where zed is not
typically already running, but this is not a requirement
of the tests.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Olaf Faaland <faaland1@llnl.gov>
Closes #9974

4 years agoRemove dead code error handling from dsl_crypt.c
Matthew Macy [Tue, 25 Feb 2020 23:59:29 +0000 (15:59 -0800)]
Remove dead code error handling from dsl_crypt.c

Sleepable (KM_SLEEP) allocations cannot fail. Hence
error handling for them is not useful.

Reviewed-By: Tom Caputi <tcaputi@datto.com>
Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes #10031

4 years agoZTS: Move atime_003 to linux.run
Ryan Moeller [Tue, 25 Feb 2020 23:27:41 +0000 (18:27 -0500)]
ZTS: Move atime_003 to linux.run

This test verifies relatime behavior, which is only present on Linux.

Move the test to linux.run

Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10046

4 years agoUpdate README for OpenZFS
Matthew Ahrens [Tue, 25 Feb 2020 19:43:20 +0000 (11:43 -0800)]
Update README for OpenZFS

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #10053

4 years agoRemove zfs_getattr and convoff dead code
Dirkjan Bussink [Mon, 24 Feb 2020 23:38:23 +0000 (00:38 +0100)]
Remove zfs_getattr and convoff dead code

The `convoff` function is called only in one code path in `zfs_space`.
Each caller of `zfs_space` is called with a `flock64_t` that has
`l_whence` set to `SEEK_SET`. This means that `convoff` always results
in a no-op as the `bfp` parameter has `l_whence` set to `SEEK_SET` and
`int whence` is `SEEK_SET` as well.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Dirkjan Bussink <d.bussink@gmail.com>
Closes #10006

4 years agoZTS: Misc fixes for FreeBSD
Ryan Moeller [Mon, 24 Feb 2020 18:17:55 +0000 (13:17 -0500)]
ZTS: Misc fixes for FreeBSD

* Force UFS sync before snap in vol rollback tests
* rw is not a valid share option on FreeBSD, use ro instead
* zfs_unmount_nested: mountpoint is in the pool, rmdir *before* export
* Fix some more platform checks
* Fix disappearing group in delegate tests
* Don't try delegating for jailed, only root can set it

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10038

4 years agoRemove unused structs and members in dmu_send.c
Matthew Ahrens [Mon, 24 Feb 2020 17:50:14 +0000 (09:50 -0800)]
Remove unused structs and members in dmu_send.c

There are several structs (and members of structs) related to redaction,
which are no longer used.  This commit removes them.

Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #10039

4 years agoZTS: Eliminate partitioning from zpool_destroy
Ryan Moeller [Sat, 22 Feb 2020 00:00:23 +0000 (19:00 -0500)]
ZTS: Eliminate partitioning from zpool_destroy

The zpool destroy tests partition a single disk to create two pools.

This can be done using two disks and no partitioning instead.
And temporarily allow vol recursion for FreeBSD while in here.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10036

4 years agoZTS: Refactor is_shared, fix impl on FreeBSD
Ryan Moeller [Fri, 21 Feb 2020 23:59:20 +0000 (18:59 -0500)]
ZTS: Refactor is_shared, fix impl on FreeBSD

FreeBSD doesn't have a `share` command.  It does have showmount.

Split the separate platform impls out of is_shared_impl.
Dispatch to the correct platform impl function from is_shared.
Eliminate the use of is_shared_impl from tests.  is_shared works.

Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10037

4 years agoZTS: Move privilege tests to sunos.run
Ryan Moeller [Fri, 21 Feb 2020 16:52:44 +0000 (11:52 -0500)]
ZTS: Move privilege tests to sunos.run

These tests are unspported on FreeBSD and Linux for lack of pfexec.

Move the privilege tests to sunos.run and remove the platform checks.

Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes #10035