The l_flushsema doesn't exactly have completion semantics, nor mutex
semantics. It's used as a list of tasks which are waiting to be notified
that a flush has completed. It was also being used in a way that was
potentially racy, depending on the semaphore implementation.
By using a sv_t instead of a semaphore we avoid the need for a separate
counter, since we know we just need to wake everything on the queue.
Original waitqueue implementation from Matthew Wilcox. Cleanup and
conversion to sv_t by Christoph Hellwig.
SGI-PV: 981507
SGI-Modid: xfs-linux-melb:xfs-kern:31059a
Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
We found this while experimenting with 2GiB xfs logs. The previous code
never assumed that xfs logs would ever get so large.
SGI-PV: 981502
SGI-Modid: xfs-linux-melb:xfs-kern:31058a
Signed-off-by: Michael Nishimoto <miken@agami.com>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
kmem_free() function takes (ptr, size) arguments but doesn't actually use
second one.
This patch removes size argument from all callsites.
SGI-PV: 981498
SGI-Modid: xfs-linux-melb:xfs-kern:31050a
Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
When we release the iclog, we do an atomic_dec_and_lock to determine if
we are the last reference and need to trigger update of log headers and
writeout. However, in xlog_state_get_iclog_space() we also need to
check if we have the last reference count there. If we do, we release
the log buffer, otherwise we decrement the reference count.
But the compare and decrement in xlog_state_get_iclog_space() is not
atomic, so both places can see a reference count of 2 and neither will
release the iclog. That leads to a filesystem hang.
Close the race by replacing the atomic_read() and atomic_dec() pair with
atomic_add_unless() to ensure that they are executed atomically.
Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Tim Shimmin <tes@sgi.com>
Tested-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Unmounting the log can fail. unlikely, but it can. Catch all the error
conditions an make sure it's propagated upwards.
SGI-PV: 980084
SGI-Modid: xfs-linux-melb:xfs-kern:30833a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Niv Sardi <xaiki@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
xfs_log_force() is declared to return an error, but we almost never check
it. We don't need to check it in most cases; if there's a log I/O error
then we'll be shutting down the filesystem anyway and that means we'll
catch the error somewhere else.
However, on certain calls we should be returning an error - sync
transactions, fsync, sync writes, etc. so this isn't a pure black and
white distinction. Hence make xfs_log_force() a void function that issues
a warning to the syslog on error, and call _xfs_log_force() in all the
places where we actually care about the error status returned.
SGI-PV: 980084
SGI-Modid: xfs-linux-melb:xfs-kern:30832a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Niv Sardi <xaiki@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Recent changes to xlog_state_release_iclog() placed the grant_lock inside
the icloglock. forced unmount of the log does this the opposite way
around, but does not depend on the order for correct working. Fix the
inversion by changing the order locks are gained in
xfs_log_force_umount().
SGI-PV: 979661
SGI-Modid: xfs-linux-melb:xfs-kern:30773a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
To reduce contention on the log in large CPU count, separate out different
parts of the xlog_t structure onto different cachelines. Move each lock
onto a different cacheline along with all the members that are
accessed/modified while that lock is held.
Also, move the debugging code into debug code.
SGI-PV: 978729
SGI-Modid: xfs-linux-melb:xfs-kern:30772a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
The ticket allocator is just a simple slab implementation internal to the
log. It requires the icloglock to be held when manipulating it and this
contributes to contention on that lock.
Just kill the entire allocator and use a memory zone instead. While there,
allow us to gracefully fail allocation with ENOMEM.
SGI-PV: 978729
SGI-Modid: xfs-linux-melb:xfs-kern:30771a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Rather than use the icloglock for protecting the iclog completion callback
chain, use a new per-iclog lock so that walking the callback chain doesn't
require holding a global lock.
This reduces contention on the icloglock during transaction commit and log
I/O completion by reducing the number of times we need to hold the global
icloglock during these operations.
SGI-PV: 978729
SGI-Modid: xfs-linux-melb:xfs-kern:30770a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Now that we update the log tail LSN less frequently on transaction
completion, we pass the contention straight to the global log state lock
(l_iclog_lock) during transaction completion.
We currently have to take this lock to decrement the iclog reference
count. there is a reference count on each iclog, so we need to take þhe
global lock for all refcount changes.
When large numbers of processes are all doing small trnasctions, the iclog
reference counts will be quite high, and the state change that absolutely
requires the l_iclog_lock is the except rather than the norm.
Change the reference counting on the iclogs to use atomic_inc/dec so that
we can use atomic_dec_and_lock during transaction completion and avoid the
need for grabbing the l_iclog_lock for every reference count decrement
except the one that matters - the last.
SGI-PV: 975671
SGI-Modid: xfs-linux-melb:xfs-kern:30505a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
When hundreds of processors attempt to commit transactions at the same
time, they can contend on the AIL lock when updating the tail LSN held in
the in-core log structure.
At the moment, the tail LSN is only needed when actually writing out an
iclog, so it really does not need to be updated on every single
transaction completion - only those that result in switching iclogs and
flushing them to disk.
The result is that we reduce the number of times we need to grab the AIL
lock and the log grant lock by up to two orders of magnitude on large
processor count machines. The problem has previously been hidden by AIL
lock contention walking the AIL list which was recently solved and
uncovered this issue.
SGI-PV: 975671
SGI-Modid: xfs-linux-melb:xfs-kern:30504a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Remove macro-to-small-function indirection from xfs_sb.h, and remove some
which are completely unused.
SGI-PV: 976035
SGI-Modid: xfs-linux-melb:xfs-kern:30528a
Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: Donald Douwsma <donaldd@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
remove beX_add functions and replace all uses with beX_add_cpu
Signed-off-by: Marcin Slusarz <marcin.slusarz@gmail.com>
Cc: Mark Fasheh <mark.fasheh@oracle.com>
Reviewed-by: Dave Chinner <dgc@sgi.com>
Cc: Timothy Shimmin <tes@sgi.com>
Cc: <linux-ext4@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When many hundreds to thousands of threads all try to do simultaneous
transactions and the log is in a tail-pushing situation (i.e. full), we
can get multiple threads walking the AIL list and contending on the AIL
lock.
The AIL push is, in effect, a simple I/O dispatch algorithm complicated by
the ordering constraints placed on it by the transaction subsystem. It
really does not need multiple threads to push on it - even when only a
single CPU is pushing the AIL, it can push the I/O out far faster that
pretty much any disk subsystem can handle.
So, to avoid contention problems stemming from multiple list walkers, move
the list walk off into another thread and simply provide a "target" to
push to. When a thread requires a push, it sets the target and wakes the
push thread, then goes to sleep waiting for the required amount of space
to become available in the log.
This mechanism should also be a lot fairer under heavy load as the waiters
will queue in arrival order, rather than queuing in "who completed a push
first" order.
Also, by moving the pushing to a separate thread we can do more
effectively overload detection and prevention as we can keep context from
loop iteration to loop iteration. That is, we can push only part of the
list each loop and not have to loop back to the start of the list every
time we run. This should also help by reducing the number of items we try
to lock and/or push items that we cannot move.
Note that this patch is not intended to solve the inefficiencies in the
AIL structure and the associated issues with extremely large list
contents. That needs to be addresses separately; parallel access would
cause problems to any new structure as well, so I'm only aiming to isolate
the structure from unbounded parallelism here.
SGI-PV: 972759
SGI-Modid: xfs-linux-melb:xfs-kern:30371a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
The BPCSHIFT based macros, btoc*, ctob*, offtoc* and ctooff are either not
used or don't need to be used. The NDPP, NDPP, NBBY macros don't need to
be used but instead are replaced directly by PAGE_SIZE and PAGE_CACHE_SIZE
where appropriate. Initial patch and motivation from Nicolas Kaiser.
SGI-PV: 971186
SGI-Modid: xfs-linux-melb:xfs-kern:30096a
Signed-off-by: Tim Shimmin <tes@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
This assert is bogus. We can have a forced shutdown occur
between the check for the XLOG_FORCED_SHUTDOWN and the ASSERT. Also, the
logging system shouldn't care about the state of XFS_FORCED_SHUTDOWN, it
should only check XLOG_FORCED_SHUTDOWN. The logging system has it's own
forced shutdown flag so, for the case of a forced shutdown that's not due
to a logging error, we can flush the log.
SGI-PV: 972985
SGI-Modid: xfs-linux-melb:xfs-kern:30029a
Signed-off-by: Niv Sardi <xaiki@sgi.com>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
These are mostly locking annotations, marking things static, casts where
needed and declaring stuff in header files.
SGI-PV: 971186
SGI-Modid: xfs-linux-melb:xfs-kern:30002a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Mostly trivial conversion with one exceptions: h_num_logops was kept in
native endian previously and only converted to big endian in xlog_sync,
but we always keep it big endian now. With todays cpus fast byteswap
instructions that's not an issue but the new variant keeps the code clean
and maintainable.
SGI-PV: 971186
SGI-Modid: xfs-linux-melb:xfs-kern:29821a
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
- the various assign lsn macros are replaced by a single inline,
xlog_assign_lsn, which is equivalent to ASSIGN_ANY_LSN_HOST except
for a more sane calling convention. ASSIGN_LSN_DISK is replaced
by xlog_assign_lsn and a manual bytespap, and ASSIGN_LSN by the same,
except we pass the cycle and block arguments explicitly instead of a
log paramter. The latter two variants only had 2, respectively one
user anyway.
- the GET_CYCLE is replaced by a xlog_get_cycle inline with exactly the
same calling conventions.
- GET_CLIENT_ID is replaced by xlog_get_client_id which leaves away
the unused arch argument. Instead of conditional defintions
depending on host endianess we now do an unconditional swap and shift
then, which generates equal code.
- the unused XLOG_SET macro is removed.
SGI-PV: 971186
SGI-Modid: xfs-linux-melb:xfs-kern:29820a
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
- the various assign lsn macros are replaced by a single inline,
xlog_assign_lsn, which is equivalent to ASSIGN_ANY_LSN_HOST except
for a more sane calling convention. ASSIGN_LSN_DISK is replaced
by xlog_assign_lsn and a manual bytespap, and ASSIGN_LSN by the same,
except we pass the cycle and block arguments explicitly instead of a
log paramter. The latter two variants only had 2, respectively one
user anyway.
- the GET_CYCLE is replaced by a xlog_get_cycle inline with exactly the
same calling conventions.
- GET_CLIENT_ID is replaced by xlog_get_client_id which leaves away
the unused arch argument. Instead of conditional defintions
depending on host endianess we now do an unconditional swap and shift
then, which generates equal code.
- the unused XLOG_SET macro is removed.
SGI-PV: 971186
SGI-Modid: xfs-linux-melb:xfs-kern:29819a
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
remove spinlock init abstraction macro in spin.h, remove the callers, and
remove the file. Move no-op spinlock_destroy to xfs_linux.h Cleanup
spinlock locals in xfs_mount.c
SGI-PV: 970382
SGI-Modid: xfs-linux-melb:xfs-kern:29751a
Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: Donald Douwsma <donaldd@sgi.com>
Signed-off-by: Lachlan McIlroy <lachlan@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Un-obfuscate GRANT_LOCK, remove GRANT_LOCK->mutex_lock->spin_lock macros,
call spin_lock directly, remove extraneous cookie holdover from old xfs
code, and change lock type to spinlock_t.
SGI-PV: 970382
SGI-Modid: xfs-linux-melb:xfs-kern:29741a
Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: Donald Douwsma <donaldd@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Un-obfuscate LOG_LOCK, remove LOG_LOCK->mutex_lock->spin_lock macros, call
spin_lock directly, remove extraneous cookie holdover from old xfs code,
and change lock type to spinlock_t.
SGI-PV: 970382
SGI-Modid: xfs-linux-melb:xfs-kern:29740a
Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: Donald Douwsma <donaldd@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
... or in the case of XLOG_TIC_ADD_OPHDR remove a useless macro entirely.
SGI-PV: 968563
SGI-Modid: xfs-linux-melb:xfs-kern:29511a
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
All flags are added to xfs_mount's m_flag instead. Note that the 32bit
inode flag was duplicated in both of them, but only cleared in the mount
when it was not nessecary due to the filesystem beeing small enough. Two
flags are still required here - one to indicate the mount option setting,
and one to indicate if it applies or not.
SGI-PV: 969608
SGI-Modid: xfs-linux-melb:xfs-kern:29507a
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Remove sizing of logbuf size & count based on physical memory; this was
never a very good gauge as it's looking at global memory, but deciding on
sizing per-filesystem; no account is made of the total number of
filesystems, for example.
For now just take the largest "default" case, as was set for machines with
>400MB - 8 x 32k buffers. This can always be tuned higher or lower with
mount options if necessary. Removes one more user of xfs_physmem.
SGI-PV: 968563
SGI-Modid: xfs-linux-melb:xfs-kern:29323a
Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
If the underlying block device suddenly stops supporting barriers, we need
to handle the -EOPNOTSUPP error in a sane manner rather than shutting
down the filesystem. If we get this error, clear the barrier flag, reissue
the I/O, and tell the world bad things are occurring.
SGI-PV: 964544
SGI-Modid: xfs-linux-melb:xfs-kern:28568a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Sparse now warns about comparing pointers to 0, so change all instance
where that happens to NULL instead.
SGI-PV: 968555
SGI-Modid: xfs-linux-melb:xfs-kern:29308a
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
When we have a couple of hundred transactions on the fly at once, they all
typically modify the on disk superblock in some way.
create/unclink/mkdir/rmdir modify inode counts, allocation/freeing modify
free block counts.
When these counts are modified in a transaction, they must eventually lock
the superblock buffer and apply the mods. The buffer then remains locked
until the transaction is committed into the incore log buffer. The result
of this is that with enough transactions on the fly the incore superblock
buffer becomes a bottleneck.
The result of contention on the incore superblock buffer is that
transaction rates fall - the more pressure that is put on the superblock
buffer, the slower things go.
The key to removing the contention is to not require the superblock fields
in question to be locked. We do that by not marking the superblock dirty
in the transaction. IOWs, we modify the incore superblock but do not
modify the cached superblock buffer. In short, we do not log superblock
modifications to critical fields in the superblock on every transaction.
In fact we only do it just before we write the superblock to disk every
sync period or just before unmount.
This creates an interesting problem - if we don't log or write out the
fields in every transaction, then how do the values get recovered after a
crash? the answer is simple - we keep enough duplicate, logged information
in other structures that we can reconstruct the correct count after log
recovery has been performed.
It is the AGF and AGI structures that contain the duplicate information;
after recovery, we walk every AGI and AGF and sum their individual
counters to get the correct value, and we do a transaction into the log to
correct them. An optimisation of this is that if we have a clean unmount
record, we know the value in the superblock is correct, so we can avoid
the summation walk under normal conditions and so mount/recovery times do
not change under normal operation.
One wrinkle that was discovered during development was that the blocks
used in the freespace btrees are never accounted for in the AGF counters.
This was once a valid optimisation to make; when the filesystem is full,
the free space btrees are empty and consume no space. Hence when it
matters, the "accounting" is correct. But that means the when we do the
AGF summations, we would not have a correct count and xfs_check would
complain. Hence a new counter was added to track the number of blocks used
by the free space btrees. This is an *on-disk format change*.
As a result of this, lazy superblock counters are a mkfs option and at the
moment on linux there is no way to convert an old filesystem. This is
possible - xfs_db can be used to twiddle the right bits and then
xfs_repair will do the format conversion for you. Similarly, you can
convert backwards as well. At some point we'll add functionality to
xfs_admin to do the bit twiddling easily....
SGI-PV: 964999
SGI-Modid: xfs-linux-melb:xfs-kern:28652a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
When setting the length of the iclogbuf to write out we should just be
changing the desired byte count rather completely reassociating the buffer
memory with the buffer. Reassociating the buffer memory changes the
apparent length of the buffer and hence when we free the buffer, we don't
free all the vmap()d space we originally allocated.
SGI-PV: 964983
SGI-Modid: xfs-linux-melb:xfs-kern:28640a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Don't reference the log buffer after running the callbacks as the callback
can trigger the log buffers to be freed during unmount.
SGI-PV: 964545
SGI-Modid: xfs-linux-melb:xfs-kern:28567a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Many block drivers (aoe, iscsi) really want refcountable pages in bios,
which is what almost everyone send down. XFS unfortunately has a few
places where it sends down buffers that may come from kmalloc, which
breaks them.
Fix the places that use kmalloc()d buffers.
SGI-PV: 964546
SGI-Modid: xfs-linux-melb:xfs-kern:28562a
Signed-Off-By: Christoph Hellwig <hch@infradead.org>
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
space for the unmount record - which becomes a problem in the freeze/thaw
scenario.
SGI-PV: 942533
SGI-Modid: xfs-linux-melb:xfs-kern:26815a
Signed-off-by: Tim Shimmin <tes@sgi.com>
one page.
SGI-PV: 955302
SGI-Modid: xfs-linux-melb:xfs-kern:26800a
Signed-off-by: Nathan Scott <nathans@sgi.com>
Signed-off-by: Tim Shimmin <tes@sgi.com>
flags from iclog buffers before submitting them for writing.
SGI-PV: 954772
SGI-Modid: xfs-linux-melb:xfs-kern:26605a
Signed-off-by: Nathan Scott <nathans@sgi.com>
get more useful error info on space for trans items
SGI-PV: 947110
SGI-Modid: xfs-linux-melb:xfs-kern:24886a
Signed-off-by: Tim Shimmin <tes@sgi.com>
Signed-off-by: Nathan Scott <nathans@sgi.com>
writes. In addition flush the disk cache on fsync if the sync cached
operation didn't sync the log to disk (this requires some additional
bookeping in the transaction and log code). If the device doesn't claim to
support barriers, the filesystem has an extern log volume or the trial
superblock write with barriers enabled failed we disable barriers and
print a warning. We should probably fail the mount completely, but that
could lead to nasty boot failures for the root filesystem. Not enabled by
default yet, needs more destructive testing first.
SGI-PV: 912426
SGI-Modid: xfs-linux:xfs-kern:198723a
Signed-off-by: Christoph Hellwig <hch@sgi.com>
Signed-off-by: Nathan Scott <nathans@sgi.com>
include all the components which make up the transaction in the ondisk
log. Having this incomplete has shown up as problems on IRIX when some v2
log changes went in. The symptom was the msg of "xfs_log_write:
reservation ran out. Need to up reservation" and was seen on synchronous
writes on files with lots of holes (and therefore lots of extents).
SGI-PV: 931457
SGI-Modid: xfs-linux:xfs-kern:23095a
Signed-off-by: Tim Shimmin <tes@sgi.com>
Signed-off-by: Nathan Scott <nathans@sgi.com>
Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.
Let it rip!