I have been seeing occasional pauses in transaction throughput up to
30s long under heavy parallel workloads. The only notable thing was
that the xfsaild was trying to be active during the pauses, but
making no progress. It was running exactly 20 times a second (on the
50ms no-progress backoff), and the number of pushbuf events was
constant across this time as well. IOWs, the xfsaild appeared to be
stuck on buffers that it could not push out.
Further investigation indicated that it was trying to push out inode
buffers that were pinned and/or locked. The xfsbufd was also getting
woken at the same frequency (by the xfsaild, no doubt) to push out
delayed write buffers. The xfsbufd was not making any progress
because all the buffers in the delwri queue were pinned. This scan-
and-make-no-progress dance went one in the trace for some seconds,
before the xfssyncd came along an issued a log force, and then
things started going again.
However, I noticed something strange about the log force - there
were way too many IO's issued. 516 log buffers were written, to be
exact. That added up to 129MB of log IO, which got me very
interested because it's almost exactly 25% of the size of the log.
He delayed logging code is suppose to aggregate the minimum of 25%
of the log or 8MB worth of changes before flushing. That's what
really puzzled me - why did a log force write 129MB instead of only
8MB?
Essentially what has happened is that no CIL pushes had occurred
since the previous tail push which cleared out 25% of the log space.
That caused all the new transactions to block because there wasn't
log space for them, but they kick the xfsaild to push the tail.
However, the xfsaild was not making progress because there were
buffers it could not lock and flush, and the xfsbufd could not flush
them because they were pinned. As a result, both the xfsaild and the
xfsbufd could not move the tail of the log forward without the CIL
first committing.
The cause of the problem was that the background CIL push, which
should happen when 8MB of aggregated changes have been committed, is
being held off by the concurrent transaction commit load. The
background push does a down_write_trylock() which will fail if there
is a concurrent transaction commit holding the push lock in read
mode. With 8 CPUs all doing transactions as fast as they can, there
was enough concurrent transaction commits to hold off the background
push until tail-pushing could no longer free log space, and the halt
would occur.
It should be noted that there is no reason why it would halt at 25%
of log space used by a single CIL checkpoint. This bug could
definitely violate the "no transaction should be larger than half
the log" requirement and hence result in corruption if the system
crashed under heavy load. This sort of bug is exactly the reason why
delayed logging was tagged as experimental....
The fix is to start blocking background pushes once the threshold
has been exceeded. Rework the threshold calculations to keep the
amount of log space a CIL checkpoint can use to below that of the
AIL push threshold to avoid the problem completely.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Alex Elder <aelder@sgi.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Formatting items requires memory allocation when using delayed
logging. Currently that memory allocation is done while holding the
CIL context lock in read mode. This means that if memory allocation
takes some time (e.g. enters reclaim), we cannot push on the CIL
until the allocation(s) required by formatting complete. This can
stall CIL pushes for some time, and once a push is stalled so are
all new transaction commits.
Fix this splitting the item formatting into two steps. The first
step which does the allocation and memcpy() into the allocated
buffer is now done outside the CIL context lock, and only the CIL
insert is done inside the CIL context lock. This avoids the stall
issue.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Delayed logging adds some serialisation to the log force process to
ensure that it does not deference a bad commit context structure
when determining if a CIL push is necessary or not. It does this by
grabing the CIL context lock exclusively, then dropping it before
pushing the CIL if necessary. This causes serialisation of all log
forces and pushes regardless of whether a force is necessary or not.
As a result fsync heavy workloads (like dbench) can be significantly
slower with delayed logging than without.
To avoid this penalty, copy the current sequence from the context to
the CIL structure when they are swapped. This allows us to do
unlocked checks on the current sequence without having to worry
about dereferencing context structures that may have already been
freed. Hence we can remove the CIL context locking in the forcing
code and only call into the push code if the current context matches
the sequence we need to force.
By passing the sequence into the push code, we can check the
sequence again once we have the CIL lock held exclusive and abort if
the sequence has already been pushed. This avoids a lock round-trip
and unnecessary CIL pushes when we have racing push calls.
The result is that the regression in dbench performance goes away -
this change improves dbench performance on a ramdisk from ~2100MB/s
to ~2500MB/s. This compares favourably to not using delayed logging
which retuns ~2500MB/s for the same workload.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
When we commit a transaction using delayed logging, we need to
unlock the items in the transaciton before we unlock the CIL context
and allow it to be checkpointed. If we unlock them after we release
the CIl context lock, the CIL can checkpoint and complete before
we free the log items. This breaks stale buffer item unlock and
unpin processing as there is an implicit assumption that the unlock
will occur before the unpin.
Also, some log items need to store the LSN of the transaction commit
in the item (inodes and EFIs) and so can race with other transaction
completions if we don't prevent the CIL from checkpointing before
the unlock occurs.
Cc: <stable@kernel.org>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
By making this member a void pointer we can get rid of a lot of pointless
casts.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Dmapi support was never merged upstream, but we still have a lot of hooks
bloating XFS for it, all over the fast pathes of the filesystem.
This patch drops over 700 lines of dmapi overhead. If we'll ever get HSM
support in mainline at least the namespace events can be done much saner
in the VFS instead of the individual filesystem, so it's not like this
is much help for future work.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
With delayed logging, we can get inode allocation buffers in the
same transaction inode unlink buffers. We don't currently mark inode
allocation buffers in the log, so inode unlink buffers take
precedence over allocation buffers.
The result is that when they are combined into the same checkpoint,
only the unlinked inode chain fields are replayed, resulting in
uninitialised inode buffers being detected when the next inode
modification is replayed.
To fix this, we need to ensure that we do not set the inode buffer
flag in the buffer log item format flags if the inode allocation has
not already hit the log. To avoid requiring a change to log
recovery, we really need to make this a modification that relies
only on in-memory sate.
We can do this by checking during buffer log formatting (while the
CIL cannot be flushed) if we are still in the same sequence when we
commit the unlink transaction as the inode allocation transaction.
If we are, then we do not add the inode buffer flag to the buffer
log format item flags. This means the entire buffer will be
replayed, not just the unlinked fields. We do this while
CIL flusheѕ are locked out to ensure that we don't race with the
sequence numbers changing and hence fail to put the inode buffer
flag in the buffer format flags when we really need to.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
If we let the CIL grow without bound, it will grow large enough to violate
recovery constraints (must be at least one complete transaction in the log at
all times) or take forever to write out through the log buffers. Hence we need
a check during asynchronous transactions as to whether the CIL needs to be
pushed.
We track the amount of log space the CIL consumes, so it is relatively simple
to limit it on a pure size basis. Make the limit the minimum of just under half
the log size (recovery constraint) or 8MB of log space (which is an awful lot
of metadata).
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
The delayed logging code only changes in-memory structures and as
such can be enabled and disabled with a mount option. Add the mount
option and emit a warning that this is an experimental feature that
should not be used in production yet.
We also need infrastructure to track committed items that have not
yet been written to the log. This is what the Committed Item List
(CIL) is for.
The log item also needs to be extended to track the current log
vector, the associated memory buffer and it's location in the Commit
Item List. Extend the log item and log vector structures to enable
this tracking.
To maintain the current log format for transactions with delayed
logging, we need to introduce a checkpoint transaction and a context
for tracking each checkpoint from initiation to transaction
completion. This includes adding a log ticket for tracking space
log required/used by the context checkpoint.
To track all the changes we need an io vector array per log item,
rather than a single array for the entire transaction. Using the new
log vector structure for this requires two passes - the first to
allocate the log vector structures and chain them together, and the
second to fill them out. This log vector chain can then be passed
to the CIL for formatting, pinning and insertion into the CIL.
Formatting of the log vector chain is relatively simple - it's just
a loop over the iovecs on each log vector, but it is made slightly
more complex because we re-write the iovec after the copy to point
back at the memory buffer we just copied into.
This code also needs to pin log items. If the log item is not
already tracked in this checkpoint context, then it needs to be
pinned. Otherwise it is already pinned and we don't need to pin it
again.
The only other complexity is calculating the amount of new log space
the formatting has consumed. This needs to be accounted to the
transaction in progress, and the accounting is made more complex
becase we need also to steal space from it for log metadata in the
checkpoint transaction. Calculate all this at insert time and update
all the tickets, counters, etc correctly.
Once we've formatted all the log items in the transaction, attach
the busy extents to the checkpoint context so the busy extents live
until checkpoint completion and can be processed at that point in
time. Transactions can then be freed at this point in time.
Now we need to issue checkpoints - we are tracking the amount of log space
used by the items in the CIL, so we can trigger background checkpoints when the
space usage gets to a certain threshold. Otherwise, checkpoints need ot be
triggered when a log synchronisation point is reached - a log force event.
Because the log write code already handles chained log vectors, writing the
transaction is trivial, too. Construct a transaction header, add it
to the head of the chain and write it into the log, then issue a
commit record write. Then we can release the checkpoint log ticket
and attach the context to the log buffer so it can be called during
Io completion to complete the checkpoint.
We also need to allow for synchronising multiple in-flight
checkpoints. This is needed for two things - the first is to ensure
that checkpoint commit records appear in the log in the correct
sequence order (so they are replayed in the correct order). The
second is so that xfs_log_force_lsn() operates correctly and only
flushes and/or waits for the specific sequence it was provided with.
To do this we need a wait variable and a list tracking the
checkpoint commits in progress. We can walk this list and wait for
the checkpoints to change state or complete easily, an this provides
the necessary synchronisation for correct operation in both cases.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
Dave Chinner