Adding O_DIRECT support to ZFS to bypass the ARC for writes/reads.
O_DIRECT support in ZFS will always ensure there is coherency between
buffered and O_DIRECT IO requests. This ensures that all IO requests,
whether buffered or direct, will see the same file contents at all
times. Just as in other FS's , O_DIRECT does not imply O_SYNC. While
data is written directly to VDEV disks, metadata will not be synced
until the associated TXG is synced.
For both O_DIRECT read and write request the offset and request sizes,
at a minimum, must be PAGE_SIZE aligned. In the event they are not,
then EINVAL is returned unless the direct property is set to always (see
below).
For O_DIRECT writes:
The request also must be block aligned (recordsize) or the write
request will take the normal (buffered) write path. In the event that
request is block aligned and a cached copy of the buffer in the ARC,
then it will be discarded from the ARC forcing all further reads to
retrieve the data from disk.
For O_DIRECT reads:
The only alignment restrictions are PAGE_SIZE alignment. In the event
that the requested data is in buffered (in the ARC) it will just be
copied from the ARC into the user buffer.
For both O_DIRECT writes and reads the O_DIRECT flag will be ignored in
the event that file contents are mmap'ed. In this case, all requests
that are at least PAGE_SIZE aligned will just fall back to the buffered
paths. If the request however is not PAGE_SIZE aligned, EINVAL will
be returned as always regardless if the file's contents are mmap'ed.
Since O_DIRECT writes go through the normal ZIO pipeline, the
following operations are supported just as with normal buffered writes:
Checksum
Compression
Encryption
Erasure Coding
There is one caveat for the data integrity of O_DIRECT writes that is
distinct for each of the OS's supported by ZFS.
FreeBSD - FreeBSD is able to place user pages under write protection so
any data in the user buffers and written directly down to the
VDEV disks is guaranteed to not change. There is no concern
with data integrity and O_DIRECT writes.
Linux - Linux is not able to place anonymous user pages under write
protection. Because of this, if the user decides to manipulate
the page contents while the write operation is occurring, data
integrity can not be guaranteed. However, there is a module
parameter `zfs_vdev_direct_write_verify` that controls the
if a O_DIRECT writes that can occur to a top-level VDEV before
a checksum verify is run before the contents of the I/O buffer
are committed to disk. In the event of a checksum verification
failure the write will return EIO. The number of O_DIRECT write
checksum verification errors can be observed by doing
`zpool status -d`, which will list all verification errors that
have occurred on a top-level VDEV. Along with `zpool status`, a
ZED event will be issues as `dio_verify` when a checksum
verification error occurs.
ZVOLs and dedup is not currently supported with Direct I/O.
A new dataset property `direct` has been added with the following 3
allowable values:
disabled - Accepts O_DIRECT flag, but silently ignores it and treats
the request as a buffered IO request.
standard - Follows the alignment restrictions outlined above for
write/read IO requests when the O_DIRECT flag is used.
always - Treats every write/read IO request as though it passed
O_DIRECT and will do O_DIRECT if the alignment restrictions
are met otherwise will redirect through the ARC. This
property will not allow a request to fail.
There is also a module parameter zfs_dio_enabled that can be used to
force all reads and writes through the ARC. By setting this module
parameter to 0, it mimics as if the direct dataset property is set to
disabled.
Reviewed-by: Brian Behlendorf <behlendorf@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Atkinson <batkinson@lanl.gov>
Co-authored-by: Mark Maybee <mark.maybee@delphix.com>
Co-authored-by: Matt Macy <mmacy@FreeBSD.org>
Co-authored-by: Brian Behlendorf <behlendorf@llnl.gov>
Closes#10018
The Linux abd_os.c serves double-duty as the userspace scatter abd
implementation, by carrying an emulation of kernel scatterlists. This
commit lifts common and userspace-specific parts out into a separate
abd_os.c for libzpool.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16253
Nothing ever checks it.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16253
Linux provides SLAB_RECLAIM_ACCOUNT and __GFP_RECLAIMABLE flags to
mark memory allocations that can be freed via shinker calls. It
should allow kernel to tune and group such allocations for lower
memory fragmentation and better reclamation under pressure.
This patch marks as reclaimable most of ARC memory, directly
evictable via ZFS shrinker, plus also dnode/znode/sa memory,
indirectly evictable via kernel's superblock shrinker.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Changed zfs_k(un)map_atomic to zfs_k(un)map_local
Signed-off-by: Jason Lee <jasonlee@lanl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Previously, abd_iter_page() would assume that every scatterlist would
contain a single page (compound or no), because that's all we ever
create in abd_alloc_chunks(). However, scatterlists can contain multiple
pages of arbitrary provenance, and if we get one of those, we'd get all
the math wrong.
This reworks things to handle multiple pages in a scatterlist, by
properly finding the right page within it for the given offset, and
understanding better where the end of the page is and not crossing it.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reported-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16108
Before 4.5 (specifically, torvalds/linux@ddc58f2), head and tail pages
in a compound page were refcounted separately. This means that using the
head page without taking a reference to it could see it cleaned up later
before we're finished with it. Specifically, bio_add_page() would take a
reference, and drop its reference after the bio completion callback
returns.
If the zio is executed immediately from the completion callback, this is
usually ok, as any data is referenced through the tail page referenced
by the ABD, and so becomes "live" that way. If there's a delay in zio
execution (high load, error injection), then the head page can be freed,
along with any dirty flags or other indicators that the underlying
memory is used. Later, when the zio completes and that memory is
accessed, its either unmapped and an unhandled fault takes down the
entire system, or it is mapped and we end up messing around in someone
else's memory. Both of these are very bad.
The solution on these older kernels is to take a reference to the head
page when we use it, and release it when we're done. There's not really
a sensible way under our current structure to do this; the "best" would
be to keep a list of head page references in the ABD, and release them
when the ABD is freed.
Since this additional overhead is totally unnecessary on 4.5+, where
head and tail pages share refcounts, I've opted to simply not use the
compound head in ABD page iteration there. This is theoretically less
efficient (though cleaning up head page references would add overhead),
but its safe, and we still get the other benefits of not mapping pages
before adding them to a bio and not mis-splitting pages.
There doesn't appear to be an obvious symbol name or config option we
can match on to discover this behaviour in configure (and the mm/page
APIs have changed a lot since then anyway), so I've gone with a simple
version check.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Closes#15533Closes#15588
This is just renaming the existing functions we're about to replace and
grouping them together to make the next commits easier to follow.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Closes#15533Closes#15588
The regular ABD iterators yield data buffers, so they have to map and
unmap pages into kernel memory. If the caller only wants to count
chunks, or can use page pointers directly, then the map/unmap is just
unnecessary overhead.
This adds adb_iterate_page_func, which yields unmapped struct page
instead.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Closes#15533Closes#15588
MAX_ORDER has been renamed to MAX_PAGE_ORDER. Rather than just
redefining it, instead define our own name and set it consistently from
the start.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/Closes#15805
Clang's static analyzer pointed out that if alloc_pages >= nr_pages
before the loop, the value of page will be undefined and will be used
anyway. This should not be possible, but as cleanup, we add an
assertion. We also recognize that the local variables should be unsigned
in the first place, so we make them unsigned. This is not enough to
avoid the need for the assertion, since there is still the case that
alloc_pages == nr_pages and nr_pages == 0, which the assertion
implicitly checks.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14456
Windows port frees memory that was alloc'd aligned in a different way
then alloc'd memory. So changing frees to be specific.
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Andrew Innes <andrew.c12@gmail.com>
Co-Authored-By: Jorgen Lundman <lundman@lundman.net>
Closes#14059
Clang's static analyzer complained that we could use after free here if
the inner loop ever iterated. That is a false positive, but upon
inspection, the userland abd_alloc_chunks() function never will put
multiple consecutive pages into a `struct scatterlist`, so there is no
need to loop. We delete the inner loop.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14042
This patch inserts the `static` keyword to non-global variables,
which where found by the analysis tool smatch.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tino Reichardt <milky-zfs@mcmilk.de>
Closes#13970
Refcount creation for abd_zero_scatter->abd_children is redundant in
abd_alloc_zero_scatter, as it has been done in abd_init_struct.
In addition, abd_children is undefined when ZFS_DEBUG is disabled, the
reference of abd_children in abd_alloc_zero_scatter breaks build of
libzpool when ZFS_DEBUG is disabled.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Ping Huang <huangping@smartx.com>
Closes#13429
On some architectures ZERO_PAGE is unavailable because it references
a GPL exported symbol of empty_zero_page. Originally e08b993 removed
the call to PAGE_ZERO(0) for assignment to the abd_zero_page. However,
a simple check can be done to avoid a kernel allocation and free for
the abd_zero_page if ZERO_PAGE is available.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Brian Atkinson <batkinson@lanl.gov>
Closes#13199
Evaluated every variable that lives in .data (and globals in .rodata)
in the kernel modules, and constified/eliminated/localised them
appropriately. This means that all read-only data is now actually
read-only data, and, if possible, at file scope. A lot of previously-
global-symbols became inlinable (and inlined!) constants. Probably
not in a big Wowee Performance Moment, but hey.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ahelenia Ziemiańska <nabijaczleweli@nabijaczleweli.xyz>
Closes#12899
It is wrong for arc_write_ready() to use zfs_abd_scatter_enabled to
decide whether to reallocate/copy the buffer, because the answer is
OS-specific and depends on the buffer size. Instead of that use
abd_size_alloc_linear(), moved into public header.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12425
Fix a leak of abd_t that manifested mostly when using
raidzN with at least as many columns as N (e.g. a
four-disk raidz2 but not a three-disk raidz2).
Sufficiently heavy raidz use would eventually run a system
out of memory.
Additionally:
* Switch abd_cache arena to FIRSTFIT, which empirically
improves perofrmance.
* Make abd_chunk_cache more performant and debuggable.
* Allocate the abd_zero_buf from abd_chunk_cache rather
than the heap.
* Don't try to reap non-existent qcaches in abd_cache arena.
* KM_PUSHPAGE->KM_SLEEP when allocating chunks from their
own arena
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Jorgen Lundman <lundman@lundman.net>
Co-authored-by: Sean Doran <smd@use.net>
Closes#12295
wmsum was designed exactly for cases like these with many updates
and rare reads. It allows to completely avoid atomic operations on
congested global variables.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#12172
It used to be required to pass a enum km_type to kmap_atomic() and
kunmap_atomic(), however this is no longer necessary and the wrappers
zfs_k(un)map_atomic removed these. This is confusing in the ABD code as
the struct abd_iter member iter_km no longer exists and the wrapper
macros simply compile them out.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Adam Moss <c@yotes.com>
Signed-off-by: Brian Atkinson <batkinson@lanl.gov>
Closes#11768
`__vdev_disk_physio()` uses `abd_nr_pages_off()` to allocate a bio with
a sufficient number of iovec's to process this zio (i.e.
`nr_iovecs`/`bi_max_vecs`). If there are not enough iovec's in the bio,
then additional bio's will be allocated. However, this is a sub-optimal
code path. In particular, it requires several abd calls (to
`abd_nr_pages_off()` and `abd_bio_map_off()`) which will have to walk
the constituents of the ABD (the pages or the gang children) because
they are looking for offsets > 0.
For gang ABD's, `abd_nr_pages_off()` returns the number of iovec's
needed for the first constituent, rather than the sum of all
constituents (within the requested range). This always under-estimates
the required number of iovec's, which causes us to always need several
bio's. The end result is that `__vdev_disk_physio()` is usually O(n^2)
for gang ABD's (and occasionally O(n^3), when more than 16 bio's are
needed).
This commit fixes `abd_nr_pages_off()`'s handling of gang ABD's, to
correctly determine how many iovec's are needed, by adding up the number
of iovec's for each of the gang children in the requested range.
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#11536
The `abd_get_offset_*()` routines create an abd_t that references
another abd_t, and doesn't allocate any pages/buffers of its own. In
some workloads, these routines may be called frequently, to create many
abd_t's representing small pieces of a single large abd_t. In
particular, the upcoming RAIDZ Expansion project makes heavy use of
these routines.
This commit adds the ability for the caller to allocate and provide the
abd_t struct to a variant of `abd_get_offset_*()`. This eliminates the
cost of allocating the abd_t and performing the accounting associated
with it (`abdstat_struct_size`). The RAIDZ/DRAID code uses this for
the `rc_abd`, which references the zio's abd. The upcoming RAIDZ
Expansion project will leverage this infrastructure to increase
performance of reads post-expansion by around 50%.
Additionally, some of the interfaces around creating and destroying
abd_t's are cleaned up. Most significantly, the distinction between
`abd_put()` and `abd_free()` is eliminated; all types of abd_t's are
now disposed of with `abd_free()`.
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Issue #8853Closes#11439
With both abd_size and abd_nents being uint_t it makes no sense for
abd_chunkcnt_for_bytes() to return size_t. Random mix of different
types used to count chunks looks bad and makes compiler more difficult
to optimize the code.
In particular on FreeBSD this change allows compiler to completely
optimize out abd_verify_scatter() when built without debug, removing
pointless 64-bit division and even more pointless empty loop.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#11279
The ARC caches data in scatter ABD's, which are collections of pages,
which are typically 4K. Therefore, the space used to cache each block
is rounded up to a multiple of 4K. The ABD subsystem tracks this wasted
memory in the `scatter_chunk_waste` kstat. However, the ARC's `size` is
not aware of the memory used by this round-up, it only accounts for the
size that it requested from the ABD subsystem.
Therefore, the ARC is effectively using more memory than it is aware of,
due to the `scatter_chunk_waste`. This impacts observability, e.g.
`arcstat` will show that the ARC is using less memory than it
effectively is. It also impacts how the ARC responds to memory
pressure. As the amount of `scatter_chunk_waste` changes, it appears to
the ARC as memory pressure, so it needs to resize `arc_c`.
If the sector size (`1<<ashift`) is the same as the page size (or
larger), there won't be any waste. If the (compressed) block size is
relatively large compared to the page size, the amount of
`scatter_chunk_waste` will be small, so the problematic effects are
minimal.
However, if using 512B sectors (`ashift=9`), and the (compressed) block
size is small (e.g. `compression=on` with the default `volblocksize=8k`
or a decreased `recordsize`), the amount of `scatter_chunk_waste` can be
very large. On a production system, with `arc_size` at a constant 50%
of memory, `scatter_chunk_waste` has been been observed to be 10-30% of
memory.
This commit adds `scatter_chunk_waste` to `arc_size`, and adds a new
`waste` field to `arcstat`. As a result, the ARC's memory usage is more
observable, and `arc_c` does not need to be adjusted as frequently.
Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10701
Apparently missed in the initial port integration was
the need to reap the abd_chunk_cache on FreeBSD. This
change addresses that oversight.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#10474
For MIPS architectures on Linux the ZERO_PAGE macro references
empty_zero_page, which is exported as a GPL symbol. The call to
ZERO_PAGE in abd_alloc_zero_scatter has been removed and a single
zero'd page is now allocated for each of the pages in abd_zero_scatter
in the kernel ABD code path.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Brian Atkinson <batkinson@lanl.gov>
Closes#10428
Adding the gang ABD type, which allows for linear and scatter ABDs to
be chained together into a single ABD.
This can be used to avoid doing memory copies to/from ABDs. An example
of this can be found in vdev_queue.c in the vdev_queue_aggregate()
function.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Brian <bwa@clemson.edu>
Co-authored-by: Mark Maybee <mmaybee@cray.com>
Signed-off-by: Brian Atkinson <batkinson@lanl.gov>
Closes#10069
Commit fc551d7 introduced the wrappers abd_enter_critical() and
abd_exit_critical() to mark critical sections. On Linux these are
implemented with the local_irq_save() and local_irq_restore() macros
which set the 'flags' argument when saving. By wrapping them with
a function the local variable is no longer set by the macro and is
no longer properly restored.
Convert abd_enter_critical() and abd_exit_critical() to macros to
resolve this issue and ensure the flags are properly restored.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#10332
Reorganizing ABD code base so OS-independent ABD code has been placed
into a common abd.c file. OS-dependent ABD code has been left in each
OS's ABD source files, and these source files have been renamed to
abd_os.
The OS-independent ABD code is now under:
module/zfs/abd.c
With the OS-dependent code in:
module/os/linux/zfs/abd_os.c
module/os/freebsd/zfs/abd_os.c
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Brian Atkinson <batkinson@lanl.gov>
Closes#10293
