The btrfs IO submission threads try to service a bunch of devices with a small
number of threads. They do a congestion check to try and avoid waiting
on requests for a busy device.
The checks make sure we've sent a few requests down to a given device just so
that we aren't bouncing between busy devices without actually sending down
any IO. The counter used to decide if we can switch to the next device
is somewhat overloaded. It is also being used to decide if we've done
a good batch of requests between the WRITE_SYNC or regular priority lists.
It may get reset to zero often, leaving us hammering on a busy device
instead of moving on to another disk.
This commit adds a new counter for the number of bios sent while
servicing a device. It doesn't get reset or fiddled with. On
multi-device filesystems, this fixes IO stalls in streaming
write workloads.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Btrfs uses dedicated threads to submit bios when checksumming is on,
which allows us to make sure the threads dedicated to checksumming don't get
stuck waiting for requests. For each btrfs device, there are
two lists of bios. One list is for WRITE_SYNC bios and the other
is for regular priority bios.
The IO submission threads used to process all of the WRITE_SYNC bios first and
then switch to the regular bios. This commit makes sure we don't completely
starve the regular bios by rotating between the two lists.
WRITE_SYNC bios are still favored 2:1 over the regular bios, and this tries
to run in batches to avoid seeking. Benchmarking shows this eliminates
stalls during streaming buffered writes on both multi-device and
single device filesystems.
If the regular bios starve, the system can end up with a large amount of ram
pinned down in writeback pages. If we are a little more fair between the two
classes, we're able to keep throughput up and make progress on the bulk of
our dirty ram.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Once a metadata block has been written, it must be recowed, so the
btrfs dirty balancing call has a check to make sure a fair amount of metadata
was actually dirty before it started writing it back to disk.
A previous commit had changed the dirty tracking for metadata without
updating the btrfs dirty balancing checks. This commit switches it
to use the correct counter.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
The block allocator in SSD mode will try to find groups of free blocks
that are close together. This commit makes it loop less on a given
group size before bumping it.
The end result is that we are less likely to fill small holes in the
available free space, but we don't waste as much CPU building the
large cluster used by ssd mode.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
With the new back reference code, the cost of a balance has gone down
in terms of the number of back reference updates done. This commit
makes us more aggressively balance leaves and nodes as they become
less full.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
When the delayed reference code was added, some checks were added
to avoid extra balancing while the delayed references were being flushed.
This made for less efficient btrees, but it reduced the chances of
loops where no forward progress was made because the balances made
more delayed ref updates.
With the new dead root removal code and the mixed back references,
the extent allocation tree is no longer using precise back refs, and
the delayed reference updates don't carry the risk of looping forever
anymore. So, the balance avoidance is no longer required.
Signed-off-by: Chris Mason <chris.mason@oracle.com>
This commit introduces a new kind of back reference for btrfs metadata.
Once a filesystem has been mounted with this commit, IT WILL NO LONGER
BE MOUNTABLE BY OLDER KERNELS.
When a tree block in subvolume tree is cow'd, the reference counts of all
extents it points to are increased by one. At transaction commit time,
the old root of the subvolume is recorded in a "dead root" data structure,
and the btree it points to is later walked, dropping reference counts
and freeing any blocks where the reference count goes to 0.
The increments done during cow and decrements done after commit cancel out,
and the walk is a very expensive way to go about freeing the blocks that
are no longer referenced by the new btree root. This commit reduces the
transaction overhead by avoiding the need for dead root records.
When a non-shared tree block is cow'd, we free the old block at once, and the
new block inherits old block's references. When a tree block with reference
count > 1 is cow'd, we increase the reference counts of all extents
the new block points to by one, and decrease the old block's reference count by
one.
This dead tree avoidance code removes the need to modify the reference
counts of lower level extents when a non-shared tree block is cow'd.
But we still need to update back ref for all pointers in the block.
This is because the location of the block is recorded in the back ref
item.
We can solve this by introducing a new type of back ref. The new
back ref provides information about pointer's key, level and in which
tree the pointer lives. This information allow us to find the pointer
by searching the tree. The shortcoming of the new back ref is that it
only works for pointers in tree blocks referenced by their owner trees.
This is mostly a problem for snapshots, where resolving one of these
fuzzy back references would be O(number_of_snapshots) and quite slow.
The solution used here is to use the fuzzy back references in the common
case where a given tree block is only referenced by one root,
and use the full back references when multiple roots have a reference
on a given block.
This commit adds per subvolume red-black tree to keep trace of cached
inodes. The red-black tree helps the balancing code to find cached
inodes whose inode numbers within a given range.
This commit improves the balancing code by introducing several data
structures to keep the state of balancing. The most important one
is the back ref cache. It caches how the upper level tree blocks are
referenced. This greatly reduce the overhead of checking back ref.
The improved balancing code scales significantly better with a large
number of snapshots.
This is a very large commit and was written in a number of
pieces. But, they depend heavily on the disk format change and were
squashed together to make sure git bisect didn't end up in a
bad state wrt space balancing or the format change.
Signed-off-by: Yan Zheng <zheng.yan@oracle.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
There are some 'start = state->end + 1;' like code in set_extent_bit
and clear_extent_bit. They overflow when end == (u64)-1.
Signed-off-by: Yan Zheng <zheng.yan@oracle.com>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
There's a bug in the mxser kernel module that still appears in the
2.6.29.4 kernel.
mxser_get_ISA_conf takes a ioaddress as its first argument, by passing the
not of the ioaddr, you're effectively passing 0 which means it won't be
able to talk to an ISA card. I have tested this, and removing the !
fixes the problem.
Cc: "Peter Botha" <peterb@goldcircle.co.za>
Signed-off-by: Jiri Slaby <jirislaby@gmail.com>
Acked-by: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In commit code, we scan buffers attached to a transaction. During this
scan, we sometimes have to drop j_list_lock and then we recheck whether
the journal buffer head didn't get freed by journal_try_to_free_buffers().
But checking for buffer_jbd(bh) isn't enough because a new journal head
could get attached to our buffer head. So add a check whether the journal
head remained the same and whether it's still at the same transaction and
list.
This is a nasty bug and can cause problems like memory corruption (use after
free) or trigger various assertions in JBD code (observed).
Signed-off-by: Jan Kara <jack@suse.cz>
Cc: <stable@kernel.org>
Cc: <linux-ext4@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The recent ->lookup() deadlock correction required the directory inode
mutex to be dropped while waiting for expire completion. We were
concerned about side effects from this change and one has been identified.
I saw several error messages.
They cause autofs to become quite confused and don't really point to the
actual problem.
Things like:
handle_packet_missing_direct:1376: can't find map entry for (43,1827932)
which is usually totally fatal (although in this case it wouldn't be
except that I treat is as such because it normally is).
do_mount_direct: direct trigger not valid or already mounted
/test/nested/g3c/s1/ss1
which is recoverable, however if this problem is at play it can cause
autofs to become quite confused as to the dependencies in the mount tree
because mount triggers end up mounted multiple times. It's hard to
accurately check for this over mounting case and automount shouldn't need
to if the kernel module is doing its job.
There was one other message, similar in consequence of this last one but I
can't locate a log example just now.
When checking if a mount has already completed prior to adding a new mount
request to the wait queue we check if the dentry is hashed and, if so, if
it is a mount point. But, if a mount successfully completed while we
slept on the wait queue mutex the dentry must exist for the mount to have
completed so the test is not really needed.
Mounts can also be done on top of a global root dentry, so for the above
case, where a mount request completes and the wait queue entry has already
been removed, the hashed test returning false can cause an incorrect
callback to the daemon. Also, d_mountpoint() is not sufficient to check
if a mount has completed for the multi-mount case when we don't have a
real mount at the base of the tree.
Signed-off-by: Ian Kent <raven@themaw.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The massive nommu update (8feae131) resulted in these warnings:
ipc/shm.c: In function `sys_shmdt':
ipc/shm.c:974: warning: unused variable `size'
ipc/shm.c:972: warning: unused variable `next'
Signed-off-by: Mike Frysinger <vapier@gentoo.org>
Cc: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6:
cls_cgroup: Fix oops when user send improperly 'tc filter add' request
r8169: fix crash when large packets are received
* 'for-linus' of git://neil.brown.name/md:
md/raid5: fix bug in reshape code when chunk_size decreases.
md/raid5 - avoid deadlocks in get_active_stripe during reshape
md/raid5: use conf->raid_disks in preference to mddev->raid_disk
Due to commit 1cd96c242a ("block: WARN
in __blk_put_request() for potential bio leak"), BSG SMP requests get
the false warnings:
WARNING: at block/blk-core.c:1068 __blk_put_request+0x52/0xc0()
This sets rq->bio to NULL to avoid that false warnings.
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
These are defined as static cpumask_var_t so if MAXSMP is not used,
they are cleared already. Avoid surprises when MAXSMP is enabled.
Signed-off-by: Yinghai Lu <yinghai.lu@kernel.org>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
I found a bug in cls_cgroup_change() in cls_cgroup.c.
cls_cgroup_change() expected tca[TCA_OPTIONS] was set from user space properly,
but tc in iproute2-2.6.29-1 (which I used) didn't set it.
In the current source code of tc in git, it set tca[TCA_OPTIONS].
git://git.kernel.org/pub/scm/linux/kernel/git/shemminger/iproute2.git
If we always use a newest iproute2 in git when we use cls_cgroup,
we don't face this oops probably.
But I think, kernel shouldn't panic regardless of use program's behaviour.
Signed-off-by: Minoru Usui <usui@mxm.nes.nec.co.jp>
Signed-off-by: David S. Miller <davem@davemloft.net>
Michael Tokarev reported receiving a large packet could crash
a machine with RTL8169 NIC.
( original thread at http://lkml.org/lkml/2009/6/8/192 )
Problem is this driver tells that NIC frames up to 16383 bytes
can be received but provides skb to rx ring allocated with
smaller sizes (1536 bytes in case standard 1500 bytes MTU is used)
When a frame larger than what was allocated by driver is received,
dma transfert can occurs past the end of buffer and corrupt
kernel memory.
Fix is to tell to NIC what is the maximum size a frame can be.
This bug is very old, (before git introduction, linux-2.6.10), and
should be backported to stable versions.
Reported-by: Michael Tokarev <mjt@tls.msk.ru>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Tested-by: Michael Tokarev <mjt@tls.msk.ru>
Signed-off-by: David S. Miller <davem@davemloft.net>
Now that we support changing the chunksize, we calculate
"reshape_sectors" to be the max of number of sectors in old
and new chunk size.
However there is one please where we still use 'chunksize'
rather than 'reshape_sectors'.
This causes a reshape that reduces the size of chunks to freeze.
Signed-off-by: NeilBrown <neilb@suse.de>
md has functionality to 'quiesce' and array so that all pending
IO completed and no new IO starts. This is used to achieve a
stable state before making internal changes.
Currently this quiescing applies equally to normal IO, resync
IO, and reshape IO.
However there is a problem with applying it to reshape IO.
Reshape can have multiple 'stripe_heads' that must be active together.
If the quiesce come between allocating the first and the last of
such a collection, then we deadlock, as the last will not be allocated
until the quiesce is lifted, the quiesce will not be lifted until the
first (which has been allocated) gets used, and that first cannot be
used until the last is allocated.
It is not necessary to inhibit reshape IO when a quiesce is
requested. Those places in the code that require a full quiesce will
ensure the reshape thread is not running at all.
So allow reshape requests to get access to new stripe_heads without
being blocked by a 'quiesce'.
This only affects in-place reshapes (i.e. where the array does not
grow or shrink) and these are only newly supported. So this patch is
not needed in earlier kernels.
Signed-off-by: NeilBrown <neilb@suse.de>
mddev->raid_disks can be changed and any time by a request from
user-space. It is a suggestion as to what number of raid_disks is
desired.
conf->raid_disks can only be changed by the raid5 module with suitable
locks in place. It is a statement as to the current number of
raid_disks.
There are two places where the latter should be used, but the former
is used. This can lead to a crash when reshaping an array.
This patch changes to mddev-> to conf->
Signed-off-by: NeilBrown <neilb@suse.de>
Our async work synchronization was broken by "async: make sure
independent async domains can't accidentally entangle" (commit
d5a877e8dd), because it would report
the wrong lowest active async ID when there was both running and
pending async work.
This caused things like no being able to read the root filesystem,
resulting in missing console devices and inability to run 'init',
causing a boot-time panic.
This fixes it by properly returning the lowest pending async ID: if
there is any running async work, that will have a lower ID than any
pending work, and we should _not_ look at the pending work list.
There were alternative patches from Jaswinder and James, but this one
also cleans up the code by removing the pointless 'ret' variable and
the unnecesary testing for an empty list around 'for_each_entry()' (if
the list is empty, the for_each_entry() thing just won't execute).
Fixes-bug: http://bugzilla.kernel.org/show_bug.cgi?id=13474
Reported-and-tested-by: Chris Clayton <chris2553@googlemail.com>
Cc: Jaswinder Singh Rajput <jaswinder@kernel.org>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'upstream' of git://ftp.linux-mips.org/pub/scm/upstream-linus:
MIPS: Outline udelay and fix a few issues.
MIPS: ioctl.h: Fix headers_check warnings
MIPS: Cobalt: PCI bus is always required to obtain the board ID
MIPS: Kconfig: Remove "Support for" from Cavium system type
MIPS: Sibyte: Honor CONFIG_CMDLINE
SSB: BCM47xx: Export ssb_watchdog_timer_set
The previous patch submission had a I typo I didn't catch but Bartlomiej
noted. Guess this proves the point about any patch being risky late in an rc
Signed-off-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Outlining fixes the issue were on certain CPUs such as the R10000 family
the delay loop would need an extra cycle if it overlaps a cacheline
boundary.
The rewrite also fixes build errors with GCC 4.4 which was changed in
way incompatible with the kernel's inline assembly.
Relying on pure C for computation of the delay value removes the need for
explicit. The price we pay is a slight slowdown of the computation - to
be fixed on another day.
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
Make ioctl.h compatible with asm-generic/ioctl.h and userspace
fix the following 'make headers_check' warning:
usr/include/asm-mips/ioctl.h:64: extern's make no sense in userspace
Signed-off-by: Jaswinder Singh Rajput <jaswinderrajput@gmail.com>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
this patch export ssb_watchdog_timer_set to allow to use it in a Linux
watchdog driver.
Signed-off-by: Matthieu CASTET <castet.matthieu@free.fr>
Acked-by : Michael Buesch <mb@bu3sch.de>
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
* master.kernel.org:/home/rmk/linux-2.6-arm:
[ARM] 5543/1: arm: serial amba: add missing declaration in serial.h
[ARM] pxa: fix pxa27x_udc default pullup GPIO
[ARM] pxa/imote2: fix UCAM sensor board ADC model number
mx[23]: don't put clock lookups in __initdata
fix oops when using console=ttymxcN with N > 0
[ARM] ARMv7 errata: only apply fixes when running on applicable CPU
[ARM] 5534/1: kmalloc must return a cache line aligned buffer
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/drzeus/mmc:
sdhci-of: Fix the wrong accessor to HOSTVER register
mvsdio: fix config failure with some high speed SDHC cards
mvsdio: ignore high speed timing requests from the core
mmc/omap: Use disable_irq_nosync() from within irq handlers.
sdhci-of: Add fsl,esdhc as a valid compatible to bind against
mvsdio: allow automatic loading when modular
mxcmmc: Fix missing return value checking in DMA setup code.
mxcmmc : Reset the SDHC hardware if software timeout occurs.
omap_hsmmc: Trivial fix for a typo in comment
mxcmmc: decrease minimum frequency to make MMC cards work
Under CONFIG_MAXSMP, cpus_hardware_enabled is allocated from the heap and
not statically initialized. This causes a crash on reboot when kvm thinks
vmx is enabled on random nonexistent cpus and accesses nonexistent percpu
lists.
Fix by explicitly clearing the variable.
Cc: stable@kernel.org
Reported-and-tested-by: Yinghai Lu <yinghai@kernel.org>
Signed-off-by: Avi Kivity <avi@redhat.com>
This header is sometimes included in the uncompress stage to get
register values, but no <linux/amba/bus.h> can be included there.
So declare "struct amba_device" here before using it in a prototype.
Signed-off-by: Alessandro Rubini <rubini@unipv.it>
Acked-by: Andrea Gallo <andrea.gallo@stericsson.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
pdc202xx_reset() calls pdc202xx_reset_host() twice, for both channels, while
that function actually twiddles the single, shared software reset bit -- the
net effect is a duplicated reset and horrendous 4 second delay happening not
only on a channel reset but also when dma_lost_irq() and dma_clear() methods
are called. Fold pdc202xx_reset_host() into pdc202xx_reset(), fix printk(),
and move it before the actual reset...
Signed-off-by: Sergei Shtylyov <sshtylyov@ru.mvista.com>
Signed-off-by: Bartlomiej Zolnierkiewicz <bzolnier@gmail.com>
Commit ac95beedf8 (ide: add struct ide_port_ops
(take 2)) erroneously converted the driver's dma_timeout() and dma_lost_irq()
methods to call the driver's resetproc() method regardless of whether it was
defined for this specific controller while it hadn't been defined and hence
called for PDC20246. So the dma_clear() method, the successor of dma_timeout(),
shouldn't exist and the dma_lost_irq() method should be standard for PDC20246.
Signed-off-by: Sergei Shtylyov <sshtylyov@ru.mvista.com>
Signed-off-by: Bartlomiej Zolnierkiewicz <bzolnier@gmail.com>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jbarnes/pci-2.6:
x86/pci: fix mmconfig detection with 32bit near 4g
PCI: use fixed-up device class when configuring device
CONFIG_IMA=y inode activity leaks iint_cache and radix_tree_node objects
until the system runs out of memory. Nowhere is calling ima_inode_free()
a.k.a. ima_iint_delete(). Fix that by calling it from destroy_inode().
Signed-off-by: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs-2.6:
ext3/4 with synchronous writes gets wedged by Postfix
Fix nobh_truncate_page() to not pass stack garbage to get_block()
OK, that's probably the easiest way to do that, as much as I don't like it...
Since iget() et.al. will not accept I_FREEING (will wait to go away
and restart), and since we'd better have serialization between new/free
on fs data structures anyway, we can afford simply skipping I_FREEING
et.al. in insert_inode_locked().
We do that from new_inode, so it won't race with free_inode in any interesting
ways and it won't race with iget (of any origin; nfsd or in case of fs
corruption a lookup) since both still will wait for I_LOCK.
Reviewed-by: "Theodore Ts'o" <tytso@mit.edu>
Acked-by: Jan Kara <jack@suse.cz>
Tested-by: David Watson <dbwatson@ukfsn.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The nobh_truncate_page() function is used by ext2, exofs, and jfs. Of
these three, only ext2 and jfs's get_block() function pays attention
to bh->b_size --- which is normally always the filesystem blocksize
except when the get_block() function is called by either
mpage_readpage(), mpage_readpages(), or the direct I/O routines in
fs/direct_io.c.
Unfortunately, nobh_truncate_page() does not initialize map_bh before
calling the filesystem-supplied get_block() function. So ext2 and jfs
will try to calculate the number of blocks to map by taking stack
garbage and shifting it left by inode->i_blkbits. This should be
*mostly* harmless (except the filesystem will do some unnneeded work)
unless the stack garbage is less than filesystem's blocksize, in which
case maxblocks will be zero, and the attempt to find out whether or
not the filesystem has a hole at a given logical block will fail, and
the page cache entry might not get zero'ed out.
Also if the stack garbage in in map_bh->state happens to have the
BH_Mapped bit set, there could be an attempt to call readpage() on a
non-existent page, which could cause nobh_truncate_page() to return an
error when it should not.
Fix this by initializing map_bh->state and map_bh->size.
Fortunately, it's probably fairly unlikely that ext2 and jfs users
mount with nobh these days.
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: Dave Kleikamp <shaggy@linux.vnet.ibm.com>
Cc: linux-fsdevel@vger.kernel.org
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
