When xlog_sync() rounds off the tail the iclog that is being
flushed, it manually subtracts that space from the grant heads. This
space is actually reserved by the transaction ticket that covers
the xlog_sync() call from xlog_write(), but we don't plumb the
ticket down far enough for it to account for the space consumed in
the current log ticket.
The grant heads are hot, so we really should be accounting this to
the ticket is we can, rather than adding thousands of extra grant
head updates every CIL commit.
Interestingly, this actually indicates a potential log space overrun
can occur when we force the log. By the time that xfs_log_force()
pushes out an active iclog and consumes the roundoff space, the
reservation for that roundoff space has been returned to the grant
heads and is no longer covered by a reservation. In theory the
roundoff added to log force on an already full log could push the
write head past the tail. In practice, the CIL commit that writes to
the log and needs the iclog pushed will have reserved space for
roundoff, so when it releases the ticket there will still be
physical space for the roundoff to be committed to the log, even
though it is no longer reserved. This roundoff won't be enough space
to allow a transaction to be woken if the log is full, so overruns
should not actually occur in practice.
That said, it indicates that we should not release the CIL context
log ticket until after we've released the commit iclog. It also
means that xlog_sync() still needs the direct grant head
manipulation if we don't provide it with a ticket. Log forces are
rare when we are in fast paths running 1.5 million transactions/s
that make the grant heads hot, so let's optimise the hot case and
pass CIL log tickets down to the xlog_sync() code.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Because the next change is going to require sorting log vectors, and
that requires arbitrary rearrangement of the list which cannot be
done easily with a single linked list.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
So that we can remove the cil_lock which is a global serialisation
point. We've already got ordering sorted, so all we need to do is
treat the CIL list like the busy extent list and reconstruct it
before the push starts.
This is what we're trying to avoid:
- 75.35% 1.83% [kernel] [k] xfs_log_commit_cil
- 46.35% xfs_log_commit_cil
- 41.54% _raw_spin_lock
- 67.30% do_raw_spin_lock
66.96% __pv_queued_spin_lock_slowpath
Which happens on a 32p system when running a 32-way 'rm -rf'
workload. After this patch:
- 20.90% 3.23% [kernel] [k] xfs_log_commit_cil
- 17.67% xfs_log_commit_cil
- 6.51% xfs_log_ticket_ungrant
1.40% xfs_log_space_wake
2.32% memcpy_erms
- 2.18% xfs_buf_item_committing
- 2.12% xfs_buf_item_release
- 1.03% xfs_buf_unlock
0.96% up
0.72% xfs_buf_rele
1.33% xfs_inode_item_format
1.19% down_read
0.91% up_read
0.76% xfs_buf_item_format
- 0.68% kmem_alloc_large
- 0.67% kmem_alloc
0.64% __kmalloc
0.50% xfs_buf_item_size
It kinda looks like the workload is running out of log space all
the time. But all the spinlock contention is gone and the
transaction commit rate has gone from 800k/s to 1.3M/s so the amount
of real work being done has gone up a *lot*.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Before we split the ordered CIL up into per cpu lists, we need a
mechanism to track the order of the items in the CIL. We need to do
this because there are rules around the order in which related items
must physically appear in the log even inside a single checkpoint
transaction.
An example of this is intents - an intent must appear in the log
before it's intent done record so that log recovery can cancel the
intent correctly. If we have these two records misordered in the
CIL, then they will not be recovered correctly by journal replay.
We also will not be able to move items to the tail of
the CIL list when they are relogged, hence the log items will need
some mechanism to allow the correct log item order to be recreated
before we write log items to the hournal.
Hence we need to have a mechanism for recording global order of
transactions in the log items so that we can recover that order
from un-ordered per-cpu lists.
Do this with a simple monotonic increasing commit counter in the CIL
context. Each log item in the transaction gets stamped with the
current commit order ID before it is added to the CIL. If the item
is already in the CIL, leave it where it is instead of moving it to
the tail of the list and instead sort the list before we start the
push work.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Now that we have the CIL percpu structures in place, implement the
space used counter as a per-cpu counter.
We have to be really careful now about ensuring that the checks and
updates run without arbitrary delays, which means they need to run
with pre-emption disabled. We do this by careful placement of
the get_cpu_ptr/put_cpu_ptr calls to access the per-cpu structures
for that CPU.
We need to be able to reliably detect that the CIL has reached
the hard limit threshold so we can take extra reservations for the
iclog headers when the space used overruns the original reservation.
hence we factor out xlog_cil_over_hard_limit() from
xlog_cil_push_background().
The global CIL space used is an atomic variable that is backed by
per-cpu aggregation to minimise the number of atomic updates we do
to the global state in the fast path. While we are under the soft
limit, we aggregate only when the per-cpu aggregation is over the
proportion of the soft limit assigned to that CPU. This means that
all CPUs can use all but one byte of their aggregation threshold
and we will not go over the soft limit.
Hence once we detect that we've gone over both a per-cpu aggregation
threshold and the soft limit, we know that we have only
exceeded the soft limit by one per-cpu aggregation threshold. Even
if all CPUs hit this at the same time, we can't be over the hard
limit, so we can run an aggregation back into the atomic counter
at this point and still be under the hard limit.
At this point, we will be over the soft limit and hence we'll
aggregate into the global atomic used space directly rather than the
per-cpu counters, hence providing accurate detection of hard limit
excursion for accounting and reservation purposes.
Hence we get the best of both worlds - lockless, scalable per-cpu
fast path plus accurate, atomic detection of hard limit excursion.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
The CIL push lock is highly contended on larger machines, becoming a
hard bottleneck that about 700,000 transaction commits/s on >16p
machines. To address this, start moving the CIL tracking
infrastructure to utilise per-CPU structures.
We need to track the space used, the amount of log reservation space
reserved to write the CIL, the log items in the CIL and the busy
extents that need to be completed by the CIL commit. This requires
a couple of per-cpu counters, an unordered per-cpu list and a
globally ordered per-cpu list.
Create a per-cpu structure to hold these and all the management
interfaces needed, as well as the hooks to handle hotplug CPUs.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
For every iclog that a CIL push will use up, we need to ensure we
have space reserved for the iclog header in each iclog. It is
extremely difficult to do this accurately with a per-cpu counter
without expensive summing of the counter in every commit. However,
we know what the maximum CIL size is going to be because of the
hard space limit we have, and hence we know exactly how many iclogs
we are going to need to write out the CIL.
We are constrained by the requirement that small transactions only
have reservation space for a single iclog header built into them.
At commit time we don't know how much of the current transaction
reservation is made up of iclog header reservations as calculated by
xfs_log_calc_unit_res() when the ticket was reserved. As larger
reservations have multiple header spaces reserved, we can steal
more than one iclog header reservation at a time, but we only steal
the exact number needed for the given log vector size delta.
As a result, we don't know exactly when we are going to steal iclog
header reservations, nor do we know exactly how many we are going to
need for a given CIL.
To make things simple, start by calculating the worst case number of
iclog headers a full CIL push will require. Record this into an
atomic variable in the CIL. Then add a byte counter to the log
ticket that records exactly how much iclog header space has been
reserved in this ticket by xfs_log_calc_unit_res(). This tells us
exactly how much space we can steal from the ticket at transaction
commit time.
Now, at transaction commit time, we can check if the CIL has a full
iclog header reservation and, if not, steal the entire reservation
the current ticket holds for iclog headers. This minimises the
number of times we need to do atomic operations in the fast path,
but still guarantees we get all the reservations we need.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
In the next patches we are going to make the CIL list itself
per-cpu, and so we cannot use list_empty() to check is the list is
empty. Replace the list_empty() checks with a flag in the CIL to
indicate we have committed at least one transaction to the CIL and
hence the CIL is not empty.
We need this flag to be an atomic so that we can clear it without
holding any locks in the commit fast path, but we also need to be
careful to avoid atomic operations in the fast path. Hence we use
the fact that test_bit() is not an atomic op to first check if the
flag is set and then run the atomic test_and_clear_bit() operation
to clear it and steal the initial unit reservation for the CIL
context checkpoint.
When we are switching to a new context in a push, we place the
setting of the XLOG_CIL_EMPTY flag under the xc_push_lock. THis
allows all the other places that need to check whether the CIL is
empty to use test_bit() and still be serialised correctly with the
CIL context swaps that set the bit.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Move the code that allocates and frees the buffer cancellation tables
used by log recovery into the file that actually uses the tables. This
is a precursor to some cleanups and a memory leak fix.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Because heap allocation of 64kB buffers will fail:
....
XFS: fs_mark(8414) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
XFS: fs_mark(8417) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
XFS: fs_mark(8409) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
XFS: fs_mark(8428) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
XFS: fs_mark(8430) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
XFS: fs_mark(8437) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
XFS: fs_mark(8433) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
XFS: fs_mark(8406) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
XFS: fs_mark(8412) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
XFS: fs_mark(8432) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
XFS: fs_mark(8424) possible memory allocation deadlock size 65768 in kmem_alloc (mode:0x2d40)
....
I'd use kvmalloc() instead, but....
- 48.19% xfs_attr_create_intent
- 46.89% xfs_attri_init
- kvmalloc_node
- 46.04% __kmalloc_node
- kmalloc_large_node
- 45.99% __alloc_pages
- 39.39% __alloc_pages_slowpath.constprop.0
- 38.89% __alloc_pages_direct_compact
- 38.71% try_to_compact_pages
- compact_zone_order
- compact_zone
- 21.09% isolate_migratepages_block
10.31% PageHuge
5.82% set_pfnblock_flags_mask
0.86% get_pfnblock_flags_mask
- 4.48% __reset_isolation_suitable
4.44% __reset_isolation_pfn
- 3.56% __pageblock_pfn_to_page
1.33% pfn_to_online_page
2.83% get_pfnblock_flags_mask
- 0.87% migrate_pages
0.86% compaction_alloc
0.84% find_suitable_fallback
- 6.60% get_page_from_freelist
4.99% clear_page_erms
- 1.19% _raw_spin_lock_irqsave
- do_raw_spin_lock
__pv_queued_spin_lock_slowpath
- 0.86% __vmalloc_node_range
0.65% __alloc_pages_bulk
.... this is just yet another reminder of how much kvmalloc() sucks.
So lift xlog_cil_kvmalloc(), rename it to xlog_kvmalloc() and use
that instead....
We also clean up the attribute name and value lengths as they no
longer need to be rounded out to sizes compatible with log vectors.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Dave Chinner <david@fromorbit.com>
5.18 w/ std=gnu11 compiled with gcc-5 wants flags stored in unsigned
fields to be unsigned.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Now that we account for log opheaders in the log item formatting
code, we don't actually use the aggregated count of log iovecs in
the CIL for anything. Remove it and the tracking code that
calculates it.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
So remove it from the interface and callers.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Just check that the offset in xlog_write_vec is smaller than the iclog
size and remove the expensive cycling through all iclogs.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Re-implement writing of a log vector that does not fit into the
current iclog. The iclog will already be in XLOG_STATE_WANT_SYNC
because xlog_get_iclog_space() will have reserved all the remaining
iclog space for us, hence we can simply iterate over the iovecs in
the log vector getting more iclog space until the entire log vector
is written.
Handling this partial write case separately means we do need to pass
unnecessary state around for the common, fast path case when the log
vector fits entirely within the current iclog. It isolates the
complexity and allows us to modify and improve the partial write
case without impacting the simple fast path.
This change includes several improvements incorporated from patches
written by Christoph Hellwig.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Turn ic_datap from a char into a void pointer given that it points
to arbitrary data.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
[dgc: also remove (char *) cast in xlog_alloc_log()]
Signed-off-by: Dave Chinner <david@fromorbit.com>
The caller of xlog_write() usually has a close accounting of the
aggregated vector length contained in the log vector chain passed to
xlog_write(). There is no need to iterate the chain to calculate he
length of the data in xlog_write_calculate_len() if the caller is
already iterating that chain to build it.
Passing in the vector length avoids doing an extra chain iteration,
which can be a significant amount of work given that large CIL
commits can have hundreds of thousands of vectors attached to the
chain.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
xlog_tic_add_region() is used to trace the regions being added to a
log ticket to provide information in the situation where a ticket
reservation overrun occurs. The information gathered is stored int
the ticket, and dumped if xlog_print_tic_res() is called.
For a front end struct xfs_trans overrun, the ticket only contains
reservation tracking information - the ticket is never handed to the
log so has no regions attached to it. The overrun debug information in this
case comes from xlog_print_trans(), which walks the items attached
to the transaction and dumps their attached formatted log vectors
directly. It also dumps the ticket state, but that only contains
reservation accounting and nothing else. Hence xlog_print_tic_res()
never dumps region or overrun information from this path.
xlog_tic_add_region() is actually called from xlog_write(), which
means it is being used to track the regions seen in a
CIL checkpoint log vector chain. In looking at CIL behaviour
recently, I've seen 32MB checkpoints regularly exceed 250,000
regions in the LV chain. The log ticket debug code can track *15*
regions. IOWs, if there is a ticket overrun in the CIL code, the
ticket region tracking code is going to be completely useless for
determining what went wrong. The only thing it can tell us is how
much of an overrun occurred, and we really don't need extra debug
information in the log ticket to tell us that.
Indeed, the main place we call xlog_tic_add_region() is also adding
up the number of regions and the space used so that xlog_write()
knows how much will be written to the log. This is exactly the same
information that log ticket is storing once we take away the useless
region tracking array. Hence xlog_tic_add_region() is not useful,
but can be called 250,000 times a CIL push...
Just strip all that debug "information" out of the of the log ticket
and only have it report reservation space information when an
overrun occurs. This also reduces the size of a log ticket down by
about 150 bytes...
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
We currently set the log ticket client ID when we reserve a
transaction. This client ID is only ever written to the log by
a CIL checkpoint or unmount records, and so anything using a high
level transaction allocated through xfs_trans_alloc() does not need
a log ticket client ID to be set.
For the CIL checkpoint, the client ID written to the journal is
always XFS_TRANSACTION, and for the unmount record it is always
XFS_LOG, and nothing else writes to the log. All of these operations
tell xlog_write() exactly what they need to write to the log (the
optype) and build their own opheaders for start, commit and unmount
records. Hence we no longer need to set the client id in either the
log ticket or the xfs_trans.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
Jan Kara reported a performance regression in dbench that he
bisected down to commit bad77c375e ("xfs: CIL checkpoint
flushes caches unconditionally").
Whilst developing the journal flush/fua optimisations this cache was
part of, it appeared to made a significant difference to
performance. However, now that this patchset has settled and all the
correctness issues fixed, there does not appear to be any
significant performance benefit to asynchronous cache flushes.
In fact, the opposite is true on some storage types and workloads,
where additional cache flushes that can occur from fsync heavy
workloads have measurable and significant impact on overall
throughput.
Local dbench testing shows little difference on dbench runs with
sync vs async cache flushes on either fast or slow SSD storage, and
no difference in streaming concurrent async transaction workloads
like fs-mark.
Fast NVME storage.
From `dbench -t 30`, CIL scale:
clients async sync
BW Latency BW Latency
1 935.18 0.855 915.64 0.903
8 2404.51 6.873 2341.77 6.511
16 3003.42 6.460 2931.57 6.529
32 3697.23 7.939 3596.28 7.894
128 7237.43 15.495 7217.74 11.588
512 5079.24 90.587 5167.08 95.822
fsmark, 32 threads, create w/ 64 byte xattr w/32k logbsize
create chown unlink
async 1m41s 1m16s 2m03s
sync 1m40s 1m19s 1m54s
Slower SATA SSD storage:
From `dbench -t 30`, CIL scale:
clients async sync
BW Latency BW Latency
1 78.59 15.792 83.78 10.729
8 367.88 92.067 404.63 59.943
16 564.51 72.524 602.71 76.089
32 831.66 105.984 870.26 110.482
128 1659.76 102.969 1624.73 91.356
512 2135.91 223.054 2603.07 161.160
fsmark, 16 threads, create w/32k logbsize
create unlink
async 5m06s 4m15s
sync 5m00s 4m22s
And on Jan's test machine:
5.18-rc8-vanilla 5.18-rc8-patched
Amean 1 71.22 ( 0.00%) 64.94 * 8.81%*
Amean 2 93.03 ( 0.00%) 84.80 * 8.85%*
Amean 4 150.54 ( 0.00%) 137.51 * 8.66%*
Amean 8 252.53 ( 0.00%) 242.24 * 4.08%*
Amean 16 454.13 ( 0.00%) 439.08 * 3.31%*
Amean 32 835.24 ( 0.00%) 829.74 * 0.66%*
Amean 64 1740.59 ( 0.00%) 1686.73 * 3.09%*
Performance and cache flush behaviour is restored to pre-regression
levels.
As such, we can now consider the async cache flush mechanism an
unnecessary exercise in premature optimisation and hence we can
now remove it and the infrastructure it requires completely.
Fixes: bad77c375e ("xfs: CIL checkpoint flushes caches unconditionally")
Reported-and-tested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
When we call xfs_forced_shutdown(), the caller often expects the
filesystem to be completely shut down when it returns. However,
if we have racing xfs_forced_shutdown() calls, the first caller sets
the mount shutdown flag then goes to shutdown the log. The second
caller sees the mount shutdown flag and returns immediately - it
does not wait for the log to be shut down.
Unfortunately, xfs_forced_shutdown() is used in some places that
expect it to completely shut down the filesystem before it returns
(e.g. xfs_trans_log_inode()). As such, returning before the log has
been shut down leaves us in a place where the transaction failed to
complete correctly but we still call xfs_trans_commit(). This
situation arises because xfs_trans_log_inode() does not return an
error and instead calls xfs_force_shutdown() to ensure that the
transaction being committed is aborted.
Unfortunately, we have a race condition where xfs_trans_commit()
needs to check xlog_is_shutdown() because it can't abort log items
before the log is shut down, but it needs to use xfs_is_shutdown()
because xfs_forced_shutdown() does not block waiting for the log to
shut down.
To fix this conundrum, first we make all calls to
xfs_forced_shutdown() block until the log is also shut down. This
means we can then safely use xfs_forced_shutdown() as a mechanism
that ensures the currently running transaction will be aborted by
xfs_trans_commit() regardless of the shutdown check it uses.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Now that we've gotten rid of the kmem_zone_t typedef, rename the
variables to _cache since that's what they are.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
Remove these typedefs by referencing kmem_cache directly.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Chandan Babu R <chandan.babu@oracle.com>
We only use the CIL workqueue in the CIL, so it makes no sense to
hang it off the xfs_mount and have to walk multiple pointers back up
to the mount when we have the CIL structures right there.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Because we use a single work structure attached to the CIL rather
than the CIL context, we can only queue a single work item at a
time. This results in the CIL being single threaded and limits
performance when it becomes CPU bound.
The design of the CIL is that it is pipelined and multiple commits
can be running concurrently, but the way the work is currently
implemented means that it is not pipelining as it was intended. The
critical work to switch the CIL context can take a few milliseconds
to run, but the rest of the CIL context flush can take hundreds of
milliseconds to complete. The context switching is the serialisation
point of the CIL, once the context has been switched the rest of the
context push can run asynchrnously with all other context pushes.
Hence we can move the work to the CIL context so that we can run
multiple CIL pushes at the same time and spread the majority of
the work out over multiple CPUs. We can keep the per-cpu CIL commit
state on the CIL rather than the context, because the context is
pinned to the CIL until the switch is done and we aggregate and
drain the per-cpu state held on the CIL during the context switch.
However, because we no longer serialise the CIL work, we can have
effectively unlimited CIL pushes in progress. We don't want to do
this - not only does it create contention on the iclogs and the
state machine locks, we can run the log right out of space with
outstanding pushes. Instead, limit the work concurrency to 4
concurrent works being processed at a time. This is enough
concurrency to remove the CIL from being a CPU bound bottleneck but
not enough to create new contention points or unbound concurrency
issues.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
The AIL pushing is stalling on log forces when it comes across
pinned items. This is happening on removal workloads where the AIL
is dominated by stale items that are removed from AIL when the
checkpoint that marks the items stale is committed to the journal.
This results is relatively few items in the AIL, but those that are
are often pinned as directories items are being removed from are
still being logged.
As a result, many push cycles through the CIL will first issue a
blocking log force to unpin the items. This can take some time to
complete, with tracing regularly showing push delays of half a
second and sometimes up into the range of several seconds. Sequences
like this aren't uncommon:
....
399.829437: xfsaild: last lsn 0x11002dd000 count 101 stuck 101 flushing 0 tout 20
<wanted 20ms, got 270ms delay>
400.099622: xfsaild: target 0x11002f3600, prev 0x11002f3600, last lsn 0x0
400.099623: xfsaild: first lsn 0x11002f3600
400.099679: xfsaild: last lsn 0x1100305000 count 16 stuck 11 flushing 0 tout 50
<wanted 50ms, got 500ms delay>
400.589348: xfsaild: target 0x110032e600, prev 0x11002f3600, last lsn 0x0
400.589349: xfsaild: first lsn 0x1100305000
400.589595: xfsaild: last lsn 0x110032e600 count 156 stuck 101 flushing 30 tout 50
<wanted 50ms, got 460ms delay>
400.950341: xfsaild: target 0x1100353000, prev 0x110032e600, last lsn 0x0
400.950343: xfsaild: first lsn 0x1100317c00
400.950436: xfsaild: last lsn 0x110033d200 count 105 stuck 101 flushing 0 tout 20
<wanted 20ms, got 200ms delay>
401.142333: xfsaild: target 0x1100361600, prev 0x1100353000, last lsn 0x0
401.142334: xfsaild: first lsn 0x110032e600
401.142535: xfsaild: last lsn 0x1100353000 count 122 stuck 101 flushing 8 tout 10
<wanted 10ms, got 10ms delay>
401.154323: xfsaild: target 0x1100361600, prev 0x1100361600, last lsn 0x1100353000
401.154328: xfsaild: first lsn 0x1100353000
401.154389: xfsaild: last lsn 0x1100353000 count 101 stuck 101 flushing 0 tout 20
<wanted 20ms, got 300ms delay>
401.451525: xfsaild: target 0x1100361600, prev 0x1100361600, last lsn 0x0
401.451526: xfsaild: first lsn 0x1100353000
401.451804: xfsaild: last lsn 0x1100377200 count 170 stuck 22 flushing 122 tout 50
<wanted 50ms, got 500ms delay>
401.933581: xfsaild: target 0x1100361600, prev 0x1100361600, last lsn 0x0
....
In each of these cases, every AIL pass saw 101 log items stuck on
the AIL (pinned) with very few other items being found. Each pass, a
log force was issued, and delay between last/first is the sleep time
+ the sync log force time.
Some of these 101 items pinned the tail of the log. The tail of the
log does slowly creep forward (first lsn), but the problem is that
the log is actually out of reservation space because it's been
running so many transactions that stale items that never reach the
AIL but consume log space. Hence we have a largely empty AIL, with
long term pins on items that pin the tail of the log that don't get
pushed frequently enough to keep log space available.
The problem is the hundreds of milliseconds that we block in the log
force pushing the CIL out to disk. The AIL should not be stalled
like this - it needs to run and flush items that are at the tail of
the log with minimal latency. What we really need to do is trigger a
log flush, but then not wait for it at all - we've already done our
waiting for stuff to complete when we backed off prior to the log
force being issued.
Even if we remove the XFS_LOG_SYNC from the xfs_log_force() call, we
still do a blocking flush of the CIL and that is what is causing the
issue. Hence we need a new interface for the CIL to trigger an
immediate background push of the CIL to get it moving faster but not
to wait on that to occur. While the CIL is pushing, the AIL can also
be pushing.
We already have an internal interface to do this -
xlog_cil_push_now() - but we need a wrapper for it to be used
externally. xlog_cil_force_seq() can easily be extended to do what
we need as it already implements the synchronous CIL push via
xlog_cil_push_now(). Add the necessary flags and "push current
sequence" semantics to xlog_cil_force_seq() and convert the AIL
pushing to use it.
One of the complexities here is that the CIL push does not guarantee
that the commit record for the CIL checkpoint is written to disk.
The current log force ensures this by submitting the current ACTIVE
iclog that the commit record was written to. We need the CIL to
actually write this commit record to disk for an async push to
ensure that the checkpoint actually makes it to disk and unpins the
pinned items in the checkpoint on completion. Hence we need to pass
down to the CIL push that we are doing an async flush so that it can
switch out the commit_iclog if necessary to get written to disk when
the commit iclog is finally released.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Because log recovery depends on strictly ordered start records as
well as strictly ordered commit records.
This is a zero day bug in the way XFS writes pipelined transactions
to the journal which is exposed by fixing the zero day bug that
prevents the CIL from pipelining checkpoints. This re-introduces
explicit concurrent commits back into the on-disk journal and hence
out of order start records.
The XFS journal commit code has never ordered start records and we
have relied on strict commit record ordering for correct recovery
ordering of concurrently written transactions. Unfortunately, root
cause analysis uncovered the fact that log recovery uses the LSN of
the start record for transaction commit processing. Hence, whilst
the commits are processed in strict order by recovery, the LSNs
associated with the commits can be out of order and so recovery may
stamp incorrect LSNs into objects and/or misorder intents in the AIL
for later processing. This can result in log recovery failures
and/or on disk corruption, sometimes silent.
Because this is a long standing log recovery issue, we can't just
fix log recovery and call it good. This still leaves older kernels
susceptible to recovery failures and corruption when replaying a log
from a kernel that pipelines checkpoints. There is also the issue
that in-memory ordering for AIL pushing and data integrity
operations are based on checkpoint start LSNs, and if the start LSN
is incorrect in the journal, it is also incorrect in memory.
Hence there's really only one choice for fixing this zero-day bug:
we need to strictly order checkpoint start records in ascending
sequence order in the log, the same way we already strictly order
commit records.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Now that we have a mechanism to guarantee that the callbacks
attached to an iclog are owned by the context that attaches them
until they drop their reference to the iclog via
xlog_state_release_iclog(), we can attach callbacks to the iclog at
any time we have an active reference to the iclog.
xlog_state_get_iclog_space() always guarantees that the commit
record will fit in the iclog it returns, so we can move this IO
callback setting to xlog_cil_set_ctx_write_state(), record the
commit iclog in the context and remove the need for the commit iclog
to be returned by xlog_write() altogether.
This, in turn, allows us to move the wakeup for ordered commit
record writes up into xlog_cil_set_ctx_write_state(), too, because
we have been guaranteed that this commit record will be physically
located in the iclog before any waiting commit record at a higher
sequence number will be granted iclog space.
This further cleans up the post commit record write processing in
the CIL push code, especially as xlog_state_release_iclog() will now
clean up the context when shutdown errors occur.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Pass the CIL context to xlog_write() rather than a pointer to a LSN
variable. Only the CIL checkpoint calls to xlog_write() need to know
about the start LSN of the writes, so rework xlog_write to directly
write the LSNs into the CIL context structure.
This removes the commit_lsn variable from xlog_cil_push_work(), so
now we only have to issue the commit record ordering wakeup from
there.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
It is only used by the CIL checkpoints, and is the counterpart to
start record formatting and writing that is already local to
xfs_log_cil.c.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
log->l_flags doesn't actually contain "flags" as such, it contains
operational state information that can change at runtime. For the
shutdown state, this at least should be an atomic bit because
it is read without holding locks in many places and so using atomic
bitops for the state field modifications makes sense.
This allows us to use things like test_and_set_bit() on state
changes (e.g. setting XLOG_TAIL_WARN) to avoid races in setting the
state when we aren't holding locks.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
We don't need an iclog state field to tell us the log has been shut
down. We can just check the xlog_is_shutdown() instead. The avoids
the need to have shutdown overwrite the current iclog state while
being active used by the log code and so having to ensure that every
iclog state check handles XLOG_STATE_IOERROR appropriately.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Make it less shouty and a static inline before adding more calls
through the log code.
Also convert internal log code that uses XFS_FORCED_SHUTDOWN(mount)
to use xlog_is_shutdown(log) as well.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
When there are no ongoing transactions and the log contents have been
checkpointed back into the filesystem, the log performs 'covering',
which is to say that it log a dummy transaction to record the fact that
the tail has caught up with the head. This is a good time to clear log
incompat feature flags, because they are flags that are temporarily set
to limit the range of kernels that can replay a dirty log.
Since it's possible that some other higher level thread is about to
start logging items protected by a log incompat flag, we create a rwsem
so that upper level threads can coordinate this with the log. It would
probably be more performant to use a percpu rwsem, but the ability to
/try/ taking the write lock during covering is critical, and percpu
rwsems do not provide that.
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Because I cannot tell if the NEED_FLUSH flag is being set correctly
by the log force and CIL push machinery without it.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
There is a race between the new CIL async data device metadata IO
completion cache flush and the log tail in the iclog the flush
covers being updated. This can be seen by repeating generic/482 in a
loop and eventually log recovery fails with a failures such as this:
XFS (dm-3): Starting recovery (logdev: internal)
XFS (dm-3): bad inode magic/vsn daddr 228352 #0 (magic=0)
XFS (dm-3): Metadata corruption detected at xfs_inode_buf_verify+0x180/0x190, xfs_inode block 0x37c00 xfs_inode_buf_verify
XFS (dm-3): Unmount and run xfs_repair
XFS (dm-3): First 128 bytes of corrupted metadata buffer:
00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000040: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000050: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000060: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00000070: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
XFS (dm-3): metadata I/O error in "xlog_recover_items_pass2+0x55/0xc0" at daddr 0x37c00 len 32 error 117
Analysis of the logwrite replay shows that there were no writes to
the data device between the FUA @ write 124 and the FUA at write @
125, but log recovery @ 125 failed. The difference was the one log
write @ 125 moved the tail of the log forwards from (1,8) to (1,32)
and so the inode create intent in (1,8) was not replayed and so the
inode cluster was zero on disk when replay of the first inode item
in (1,32) was attempted.
What this meant was that the journal write that occurred at @ 125
did not ensure that metadata completed before the iclog was written
was correctly on stable storage. The tail of the log moved forward,
so IO must have been completed between the two iclog writes. This
means that there is a race condition between the unconditional async
cache flush in the CIL push work and the tail LSN that is written to
the iclog. This happens like so:
CIL push work AIL push work
------------- -------------
Add to committing list
start async data dev cache flush
.....
<flush completes>
<all writes to old tail lsn are stable>
xlog_write
.... push inode create buffer
<start IO>
.....
xlog_write(commit record)
.... <IO completes>
log tail moves
xlog_assign_tail_lsn()
start_lsn == commit_lsn
<no iclog preflush!>
xlog_state_release_iclog
__xlog_state_release_iclog()
<writes *new* tail_lsn into iclog>
xlog_sync()
....
submit_bio()
<tail in log moves forward without flushing written metadata>
Essentially, this can only occur if the commit iclog is issued
without a cache flush. If the iclog bio is submitted with
REQ_PREFLUSH, then it will guarantee that all the completed IO is
one stable storage before the iclog bio with the new tail LSN in it
is written to the log.
IOWs, the tail lsn that is written to the iclog needs to be sampled
*before* we issue the cache flush that guarantees all IO up to that
LSN has been completed.
To fix this without giving up the performance advantage of the
flush/FUA optimisations (e.g. g/482 runtime halves with 5.14-rc1
compared to 5.13), we need to ensure that we always issue a cache
flush if the tail LSN changes between the initial async flush and
the commit record being written. THis requires sampling the tail_lsn
before we start the flush, and then passing the sampled tail LSN to
xlog_state_release_iclog() so it can determine if the the tail LSN
has changed while writing the checkpoint. If the tail LSN has
changed, then it needs to set the NEED_FLUSH flag on the iclog and
we'll issue another cache flush before writing the iclog.
Fixes: eef983ffea ("xfs: journal IO cache flush reductions")
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
The iclog callback chain has it's own lock. That was added way back
in 2008 by myself to alleviate severe lock contention on the
icloglock in commit 114d23aae5 ("[XFS] Per iclog callback chain
lock"). This was long before delayed logging took the icloglock out
of the hot transaction commit path and removed all contention on it.
Hence the separate ic_callback_lock doesn't serve any scalability
purpose anymore, and hasn't for close on a decade.
Further, we only attach callbacks to iclogs in one place where we
are already taking the icloglock soon after attaching the callbacks.
We also have to drop the icloglock to run callbacks and grab it
immediately afterwards again. So given that the icloglock is no
longer hot, making it cover callbacks again doesn't really change
the locking patterns very much at all.
We also need to extend the icloglock to cover callback addition to
fix a zero-day UAF in the CIL push code. This occurs when shutdown
races with xlog_cil_push_work() and the shutdown runs the callbacks
before the push releases the iclog. This results in the CIL context
structure attached to the iclog being freed by the callback before
the CIL push has finished referencing it, leading to UAF bugs.
Hence, to avoid this UAF, we need the callback attachment to be
atomic with post processing of the commit iclog and references to
the structures being attached to the iclog. This requires holding
the icloglock as that's the only way to serialise iclog state
against a shutdown in progress.
The result is we need to be using the icloglock to protect the
callback list addition and removal and serialise them with shutdown.
That makes the ic_callback_lock redundant and so it can be removed.
Fixes: 71e330b593 ("xfs: Introduce delayed logging core code")
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
For the DEBUGS!
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
In doing an investigation into AIL push stalls, I was looking at the
log force code to see if an async CIL push could be done instead.
This lead me to xfs_log_force_lsn() and looking at how it works.
xfs_log_force_lsn() is only called from inode synchronisation
contexts such as fsync(), and it takes the ip->i_itemp->ili_last_lsn
value as the LSN to sync the log to. This gets passed to
xlog_cil_force_lsn() via xfs_log_force_lsn() to flush the CIL to the
journal, and then used by xfs_log_force_lsn() to flush the iclogs to
the journal.
The problem is that ip->i_itemp->ili_last_lsn does not store a
log sequence number. What it stores is passed to it from the
->iop_committing method, which is called by xfs_log_commit_cil().
The value this passes to the iop_committing method is the CIL
context sequence number that the item was committed to.
As it turns out, xlog_cil_force_lsn() converts the sequence to an
actual commit LSN for the related context and returns that to
xfs_log_force_lsn(). xfs_log_force_lsn() overwrites it's "lsn"
variable that contained a sequence with an actual LSN and then uses
that to sync the iclogs.
This caused me some confusion for a while, even though I originally
wrote all this code a decade ago. ->iop_committing is only used by
a couple of log item types, and only inode items use the sequence
number it is passed.
Let's clean up the API, CIL structures and inode log item to call it
a sequence number, and make it clear that the high level code is
using CIL sequence numbers and not on-disk LSNs for integrity
synchronisation purposes.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Currently every journal IO is issued as REQ_PREFLUSH | REQ_FUA to
guarantee the ordering requirements the journal has w.r.t. metadata
writeback. THe two ordering constraints are:
1. we cannot overwrite metadata in the journal until we guarantee
that the dirty metadata has been written back in place and is
stable.
2. we cannot write back dirty metadata until it has been written to
the journal and guaranteed to be stable (and hence recoverable) in
the journal.
The ordering guarantees of #1 are provided by REQ_PREFLUSH. This
causes the journal IO to issue a cache flush and wait for it to
complete before issuing the write IO to the journal. Hence all
completed metadata IO is guaranteed to be stable before the journal
overwrites the old metadata.
The ordering guarantees of #2 are provided by the REQ_FUA, which
ensures the journal writes do not complete until they are on stable
storage. Hence by the time the last journal IO in a checkpoint
completes, we know that the entire checkpoint is on stable storage
and we can unpin the dirty metadata and allow it to be written back.
This is the mechanism by which ordering was first implemented in XFS
way back in 2002 by commit 95d97c36e5155075ba2eb22b17562cfcc53fcf96
("Add support for drive write cache flushing") in the xfs-archive
tree.
A lot has changed since then, most notably we now use delayed
logging to checkpoint the filesystem to the journal rather than
write each individual transaction to the journal. Cache flushes on
journal IO are necessary when individual transactions are wholly
contained within a single iclog. However, CIL checkpoints are single
transactions that typically span hundreds to thousands of individual
journal writes, and so the requirements for device cache flushing
have changed.
That is, the ordering rules I state above apply to ordering of
atomic transactions recorded in the journal, not to the journal IO
itself. Hence we need to ensure metadata is stable before we start
writing a new transaction to the journal (guarantee #1), and we need
to ensure the entire transaction is stable in the journal before we
start metadata writeback (guarantee #2).
Hence we only need a REQ_PREFLUSH on the journal IO that starts a
new journal transaction to provide #1, and it is not on any other
journal IO done within the context of that journal transaction.
The CIL checkpoint already issues a cache flush before it starts
writing to the log, so we no longer need the iclog IO to issue a
REQ_REFLUSH for us. Hence if XLOG_START_TRANS is passed
to xlog_write(), we no longer need to mark the first iclog in
the log write with REQ_PREFLUSH for this case. As an added bonus,
this ordering mechanism works for both internal and external logs,
meaning we can remove the explicit data device cache flushes from
the iclog write code when using external logs.
Given the new ordering semantics of commit records for the CIL, we
need iclogs containing commit records to issue a REQ_PREFLUSH. We
also require unmount records to do this. Hence for both
XLOG_COMMIT_TRANS and XLOG_UNMOUNT_TRANS xlog_write() calls we need
to mark the first iclog being written with REQ_PREFLUSH.
For both commit records and unmount records, we also want them
immediately on stable storage, so we want to also mark the iclogs
that contain these records to be marked REQ_FUA. That means if a
record is split across multiple iclogs, they are all marked REQ_FUA
and not just the last one so that when the transaction is completed
all the parts of the record are on stable storage.
And for external logs, unmount records need a pre-write data device
cache flush similar to the CIL checkpoint cache pre-flush as the
internal iclog write code does not do this implicitly anymore.
As an optimisation, when the commit record lands in the same iclog
as the journal transaction starts, we don't need to wait for
anything and can simply use REQ_FUA to provide guarantee #2. This
means that for fsync() heavy workloads, the cache flush behaviour is
completely unchanged and there is no degradation in performance as a
result of optimise the multi-IO transaction case.
The most notable sign that there is less IO latency on my test
machine (nvme SSDs) is that the "noiclogs" rate has dropped
substantially. This metric indicates that the CIL push is blocking
in xlog_get_iclog_space() waiting for iclog IO completion to occur.
With 8 iclogs of 256kB, the rate is appoximately 1 noiclog event to
every 4 iclog writes. IOWs, every 4th call to xlog_get_iclog_space()
is blocking waiting for log IO. With the changes in this patch, this
drops to 1 noiclog event for every 100 iclog writes. Hence it is
clear that log IO is completing much faster than it was previously,
but it is also clear that for large iclog sizes, this isn't the
performance limiting factor on this hardware.
With smaller iclogs (32kB), however, there is a substantial
difference. With the cache flush modifications, the journal is now
running at over 4000 write IOPS, and the journal throughput is
largely identical to the 256kB iclogs and the noiclog event rate
stays low at about 1:50 iclog writes. The existing code tops out at
about 2500 IOPS as the number of cache flushes dominate performance
and latency. The noiclog event rate is about 1:4, and the
performance variance is quite large as the journal throughput can
fall to less than half the peak sustained rate when the cache flush
rate prevents metadata writeback from keeping up and the log runs
out of space and throttles reservations.
As a result:
logbsize fsmark create rate rm -rf
before 32kb 152851+/-5.3e+04 5m28s
patched 32kb 221533+/-1.1e+04 5m24s
before 256kb 220239+/-6.2e+03 4m58s
patched 256kb 228286+/-9.2e+03 5m06s
The rm -rf times are included because I ran them, but the
differences are largely noise. This workload is largely metadata
read IO latency bound and the changes to the journal cache flushing
doesn't really make any noticable difference to behaviour apart from
a reduction in noiclog events from background CIL pushing.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
The CIL push is the only call to xlog_write that sets this variable
to true. The other callers don't need a start rec, and they tell
xlog_write what to do by passing the type of ophdr they need written
in the flags field. The need_start_rec parameter essentially tells
xlog_write to to write an extra ophdr with a XLOG_START_TRANS type,
so get rid of the variable to do this and pass XLOG_START_TRANS as
the flag value into xlog_write() from the CIL push.
$ size fs/xfs/xfs_log.o*
text data bss dec hex filename
27595 560 8 28163 6e03 fs/xfs/xfs_log.o.orig
27454 560 8 28022 6d76 fs/xfs/xfs_log.o.patched
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
To allow for iclog IO device cache flush behaviour to be optimised,
we first need to separate out the commit record iclog IO from the
rest of the checkpoint so we can wait for the checkpoint IO to
complete before we issue the commit record.
This separation is only necessary if the commit record is being
written into a different iclog to the start of the checkpoint as the
upcoming cache flushing changes requires completion ordering against
the other iclogs submitted by the checkpoint.
If the entire checkpoint and commit is in the one iclog, then they
are both covered by the one set of cache flush primitives on the
iclog and hence there is no need to separate them for ordering.
Otherwise, we need to wait for all the previous iclogs to complete
so they are ordered correctly and made stable by the REQ_PREFLUSH
that the commit record iclog IO issues. This guarantees that if a
reader sees the commit record in the journal, they will also see the
entire checkpoint that commit record closes off.
This also provides the guarantee that when the commit record IO
completes, we can safely unpin all the log items in the checkpoint
so they can be written back because the entire checkpoint is stable
in the journal.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
We don't need to look at the xfs_mount and superblock every time we
need to do an iclog roundoff calculation. The property is fixed for
the life of the log, so store the roundoff in the log at mount time
and use that everywhere.
On a debug build:
$ size fs/xfs/xfs_log.o.*
text data bss dec hex filename
27360 560 8 27928 6d18 fs/xfs/xfs_log.o.orig
27219 560 8 27787 6c8b fs/xfs/xfs_log.o.patched
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
xlog_ticket_alloc() is always called under NOFS context, except from
unmount path, which eitherway is holding many FS locks, so, there is no
need for its callers to keep passing allocation flags into it.
change xlog_ticket_alloc() to use default kmem_cache_zalloc(), remove
its alloc_flags argument, and always use GFP_NOFS | __GFP_NOFAIL flags.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
xlog_wait() on the CIL context can reference a freed context if the
waiter doesn't get scheduled before the CIL context is freed. This
can happen when a task is on the hard throttle and the CIL push
aborts due to a shutdown. This was detected by generic/019:
thread 1 thread 2
__xfs_trans_commit
xfs_log_commit_cil
<CIL size over hard throttle limit>
xlog_wait
schedule
xlog_cil_push_work
wake_up_all
<shutdown aborts commit>
xlog_cil_committed
kmem_free
remove_wait_queue
spin_lock_irqsave --> UAF
Fix it by moving the wait queue to the CIL rather than keeping it in
in the CIL context that gets freed on push completion. Because the
wait queue is now independent of the CIL context and we might have
multiple contexts in flight at once, only wake the waiters on the
push throttle when the context we are pushing is over the hard
throttle size threshold.
Fixes: 0e7ab7efe7 ("xfs: Throttle commits on delayed background CIL push")
Reported-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
In certain situations the background CIL push can be indefinitely
delayed. While we have workarounds from the obvious cases now, it
doesn't solve the underlying issue. This issue is that there is no
upper limit on the CIL where we will either force or wait for
a background push to start, hence allowing the CIL to grow without
bound until it consumes all log space.
To fix this, add a new wait queue to the CIL which allows background
pushes to wait for the CIL context to be switched out. This happens
when the push starts, so it will allow us to block incoming
transaction commit completion until the push has started. This will
only affect processes that are running modifications, and only when
the CIL threshold has been significantly overrun.
This has no apparent impact on performance, and doesn't even trigger
until over 45 million inodes had been created in a 16-way fsmark
test on a 2GB log. That was limiting at 64MB of log space used, so
the active CIL size is only about 3% of the total log in that case.
The concurrent removal of those files did not trigger the background
sleep at all.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
The current CIL size aggregation limit is 1/8th the log size. This
means for large logs we might be aggregating at least 250MB of dirty objects
in memory before the CIL is flushed to the journal. With CIL shadow
buffers sitting around, this means the CIL is often consuming >500MB
of temporary memory that is all allocated under GFP_NOFS conditions.
Flushing the CIL can take some time to do if there is other IO
ongoing, and can introduce substantial log force latency by itself.
It also pins the memory until the objects are in the AIL and can be
written back and reclaimed by shrinkers. Hence this threshold also
tends to determine the minimum amount of memory XFS can operate in
under heavy modification without triggering the OOM killer.
Modify the CIL space limit to prevent such huge amounts of pinned
metadata from aggregating. We can have 2MB of log IO in flight at
once, so limit aggregation to 16x this size. This threshold was
chosen as it little impact on performance (on 16-way fsmark) or log
traffic but pins a lot less memory on large logs especially under
heavy memory pressure. An aggregation limit of 8x had 5-10%
performance degradation and a 50% increase in log throughput for
the same workload, so clearly that was too small for highly
concurrent workloads on large logs.
This was found via trace analysis of AIL behaviour. e.g. insertion
from a single CIL flush:
xfs_ail_insert: old lsn 0/0 new lsn 1/3033090 type XFS_LI_INODE flags IN_AIL
$ grep xfs_ail_insert /mnt/scratch/s.t |grep "new lsn 1/3033090" |wc -l
1721823
$
So there were 1.7 million objects inserted into the AIL from this
CIL checkpoint, the first at 2323.392108, the last at 2325.667566 which
was the end of the trace (i.e. it hadn't finished). Clearly a major
problem.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Allison Collins <allison.henderson@oracle.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
xlog_write_done() is just a thin wrapper around xlog_commit_record(), so
they can be merged together easily.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
