Currently it's possible to meet such too high latency results
with 'perf sched latency'.
-----------------------------------------------------------------------------------
Task | Runtime ms | Switches | Average delay ms | Maximum delay ms |
-----------------------------------------------------------------------------------
xfce4-panel | 0.222 ms | 2 | avg: 4718.345 ms | max: 9436.493 ms |
scsi_eh_3 | 3.962 ms | 36 | avg: 55.957 ms | max: 1977.829 ms |
The origin is on traces that are sometimes badly serialized across cpus.
For example the raw traces that raised such results for xfce4-panel:
(1) [init]-0 [000] 1494.663899990: sched_switch: task swapper:0 [140] (R) ==> xfce4-panel:4569 [120]
(2) xfce4-panel-4569 [000] 1494.663928373: sched_switch: task xfce4-panel:4569 [120] (S) ==> swapper:0 [140]
(3) Xorg-4276 [001] 1494.663860125: sched_wakeup: task xfce4-panel:4569 [120] success=1 [000]
(4) Xorg-4276 [001] 1504.098252756: sched_wakeup: task xfce4-panel:4569 [120] success=1 [000]
(5) perf-5219 [000] 1504.100353302: sched_switch: task perf:5219 [120] (S) ==> xfce4-panel:4569 [120]
The traces are processed in the order they arrive. Then in (2),
xfce4-panel sleeps, it is first waken up in (3) and eventually
scheduled in (5).
The latency reported is then 1504 - 1495 = 9 secs, as reported by perf
sched. But this is wrong, we are confident in the fact the traces are
nicely serialized while we should actually more trust the timestamps.
If we reorder by timestamps we get:
(1) Xorg-4276 [001] 1494.663860125: sched_wakeup: task xfce4-panel:4569 [120] success=1 [000]
(2) [init]-0 [000] 1494.663899990: sched_switch: task swapper:0 [140] (R) ==> xfce4-panel:4569 [120]
(3) xfce4-panel-4569 [000] 1494.663928373: sched_switch: task xfce4-panel:4569 [120] (S) ==> swapper:0 [140]
(4) Xorg-4276 [001] 1504.098252756: sched_wakeup: task xfce4-panel:4569 [120] success=1 [000]
(5) perf-5219 [000] 1504.100353302: sched_switch: task perf:5219 [120] (S) ==> xfce4-panel:4569 [120]
Now the trace make more sense, xfce4-panel is sleeping. Then it is
woken up in (1), scheduled in (2)
It goes to sleep in (3), woken up in (4) and scheduled in (5).
Now, latency captured between (1) and (2) is of 39 us.
And between (4) and (5) it is 2.1 ms.
Such pattern of bad serializing is the origin of the high latencies
reported by perf sched.
Basically, we need to check whether wake up time is higher than
schedule out time. If it's not the case, we need to tag the current
work atom as invalid.
Beside that, we may need to work later on a better ordering of the
traces given by the kernel.
After this patch:
xfce4-session | 0.221 ms | 1 | avg: 0.538 ms | max: 0.538 ms |
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add the quirk entry for HP dv6. Also add a workaround for the headphone
detection by setting hp_detect=1 beforehand. Without this, the driver
won't do auto-muting because BIOS doesn't give any HP pin but only a
line-out pin.
Signed-off-by: Takashi Iwai <tiwai@suse.de>
The cmpxchg used by PowerPC does the following:
({ \
__typeof__(*(ptr)) _o_ = (o); \
__typeof__(*(ptr)) _n_ = (n); \
(__typeof__(*(ptr))) __cmpxchg((ptr), (unsigned long)_o_, \
(unsigned long)_n_, sizeof(*(ptr))); \
})
This does a type check of *ptr to both o and n.
Unfortunately, the code in ring-buffer.c assigns longs to pointers
and pointers to longs and causes a warning on PowerPC:
ring_buffer.c: In function 'rb_head_page_set':
ring_buffer.c:704: warning: initialization makes pointer from integer without a cast
ring_buffer.c:704: warning: initialization makes pointer from integer without a cast
ring_buffer.c: In function 'rb_head_page_replace':
ring_buffer.c:797: warning: initialization makes integer from pointer without a cast
This patch adds the typecasts inside cmpxchg to annotate that a long is
being cast to a pointer and a pointer is being casted to a long and this
removes the PowerPC warnings.
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
It's possible that hp_detect is set even though no headphone pin is
detected. The driver issues, however, an unsol event only to hp_pins[0],
which can be invalid.
This patch adds the check of the valid pin to send an unsol event
at initialization and resume callbacks.
Signed-off-by: Takashi Iwai <tiwai@suse.de>
IDT92HD73xx and STAC927x codecs use GPIO0 bit as EAPD on many machines.
However, currently we don't set it unless the model is specified just
for safety reason. But, most machines do need this bit, so this safety
handling is rather annoying.
This patch enables GPIO0 setup as default for them. Many HP / Dell
laptops should work even without model override with this change.
Signed-off-by: Takashi Iwai <tiwai@suse.de>
Some people asked me questions like the following:
On Wed, 15 Jul 2009 13:11:21 +0200, Leon Woestenberg wrote:
> just wondering, any reasons why NILFS2 is one of the miscellaneous
> filesystems and, for example, btrfs, is not in Kconfig?
Actually, nilfs is NOT a filesystem came from other operating systems,
but a filesystem created purely for Linux. Nor is it a flash
filesystem but that for generic block devices.
So, this moves nilfs outside the misc category as I responded in LKML
"Re: Why does NILFS2 hide under Miscellaneous filesystems?"
(Message-Id: <20090716.002526.93465395.ryusuke@osrg.net>).
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
The nilfs_bmap_lookup() is now a wrapper function of
nilfs_bmap_lookup_at_level().
This moves the nilfs_bmap_lookup() to a header file converting it to
an inline function and gives an opportunity for optimization.
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
The current btree code is written so that btree functions call dat
operations via wrapper functions in bmap.c when they allocate, free,
or modify virtual block addresses.
This abstraction requires additional function calls and causes
frequent call of nilfs_bmap_get_dat() function since it is used in the
every wrapper function.
This removes the wrapper functions and makes them available from
btree.c and direct.c, which will increase the opportunity of
compiler optimization.
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
This is a preparation for the successive cleanup ("nilfs2: allow btree
to directly call dat operations").
This adds functions bundling a few operations to change an entry of
virtual block address on the dat file.
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
This gets rid of NILFS_CPFILE_GFP, NILFS_SUFILE_GFP, NILFS_DAT_GFP,
and NILFS_IFILE_GFP. All of these constants refer to NILFS_MDT_GFP,
and can be removed.
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
The btree path object is cleared just before it is freed.
This will remove the code doing the unnecessary clear operation.
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Even though many btree functions take a btree object as their first
argument, most of them are not used in their functions.
This sticky use of the btree argument is hurting code readability and
giving the possibility of inefficient code generation.
So, this removes the unnecessary btree arguments.
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
This is a re-revised patch to shorten freeze period.
This version include a fix of the bug Konishi-san mentioned last time.
When GC is runnning, GC moves live block to difference segments.
Copying live blocks into memory is done in a transaction,
however it is not necessarily to be in the transaction.
This patch will get the nilfs_ioctl_move_blocks() out from
transaction lock and put it before the transaction.
I ran sysbench fileio test against nilfs partition.
I copied some DVD/CD images and created snapshot to create live blocks
before starting the benchmark.
Followings are summary of rc8 and rc8 w/ the patch of per-request
statistics, which is min/max and avg. I ran each test three times and
bellow is average of those numers.
According to this benchmark result, average time is slightly degrated.
However, worstcase (max) result is significantly improved.
This can address a few seconds write freeze.
- random write per-request performance of rc8
min 0.843ms
max 680.406ms
avg 3.050ms
- random write per-request performance of rc8 w/ this patch
min 0.843ms -> 100.00%
max 380.490ms -> 55.90%
avg 3.233ms -> 106.00%
- sequential write per-request performance of rc8
min 0.736ms
max 774.343ms
avg 2.883ms
- sequential write per-request performance of rc8 w/ this patch
min 0.720ms -> 97.80%
max 644.280ms-> 83.20%
avg 3.130ms -> 108.50%
-----8<-----8<-----nilfs_cleanerd.conf-----8<-----8<-----
protection_period 150
selection_policy timestamp # timestamp in ascend order
nsegments_per_clean 2
cleaning_interval 2
retry_interval 60
use_mmap
log_priority info
-----8<-----8<-----nilfs_cleanerd.conf-----8<-----8<-----
Signed-off-by: Jiro SEKIBA <jir@unicus.jp>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
nilfs2: Add more safeguard routines and protections in mount process,
which also makes nilfs2 report consistency error messages when
checkpoint number is invalid.
Signed-off-by: Zhu Yanhai <zhu.yanhai@gmail.com>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
nilfs2: In procedure 'nilfs_get_sb()', when a nilfs filesysttem is
mounted for the first time, local variable 'nilfs->ns_last_cno' is
used before loading the latest checkpoint number from disk (in
'nilfs_fill_super'). 'nilfs->ns_last_cno' is assigned to 'sd.cno', but
'sd.cno' has never been used in the procedure.
Signed-off-by: Zhang Qiang <zhangqiang.buaa@gmail.com>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Alberto Bertogli advised me about bio_alloc() use in nilfs:
On Sat, 13 Jun 2009 22:52:40 -0300, Alberto Bertogli wrote:
> By the way, those bio_alloc()s are using GFP_NOWAIT but it looks
> like they could use at least GFP_NOIO or GFP_NOFS, since the caller
> can (and sometimes do) sleep. The only caller is nilfs_submit_bh(),
> which calls nilfs_submit_seg_bio() which can sleep calling
> wait_for_completion().
This takes in the comment and replaces the use of GFP_NOWAIT flag with
GFP_NOIO.
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
This removes nilfs_write_super and commit super block in nilfs
internal thread, instead of periodic write_super callback.
VFS layer calls ->write_super callback periodically. However,
it looks like that calling back is ommited when disk I/O is busy.
And when cleanerd (nilfs GC) is runnig, disk I/O tend to be busy thus
nilfs superblock is not synchronized as nilfs designed.
To avoid it, syncing superblock by nilfs thread instead of pdflush.
Signed-off-by: Jiro SEKIBA <jir@unicus.jp>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Separate conditions that check if syncing super block and alternative
super block are required as inline functions to reuse the conditions.
Signed-off-by: Jiro SEKIBA <jir@unicus.jp>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
This fixes disorder of nilfs_write_super in nilfs_sync_fs. Commiting
super block must be the end of the function so that every changes are
reflected.
->sync_fs() is not called frequently so this makes nilfs_sync_fs call
nilfs_commit_super instead of nilfs_write_super.
Signed-off-by: Jiro SEKIBA <jir@unicus.jp>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
This removes redundant super block commit.
nilfs_write_super will call nilfs_commit_super to store super block
into block device. However, nilfs_put_super will call
nilfs_commit_super right after calling nilfs_write_super. So calling
nilfs_write_super in nilfs_put_super would be redundant.
Signed-off-by: Jiro SEKIBA <jir@unicus.jp>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
This is a patch to display mount options in procfs.
Mount options will show up in the /proc/mounts as other fs does.
...
/dev/sda6 /mnt nilfs2 ro,relatime,barrier=off,cp=3,order=strict 0 0
...
Signed-off-by: Jiro SEKIBA <jir@unicus.jp>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
The current metadata file code skips disk address lookup for its data
block if the buffer has a mapped flag.
This has a potential risk to cause read request to be performed
against the stale block address that GC moved, and it may lead to meta
data corruption. The mapped flag is safe if the buffer has an
uptodate flag, otherwise it may prevent necessary update of disk
address in the next read.
This will avoid the potential problem by ensuring disk address lookup
before reading metadata block even for buffers with the mapped flag.
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
will get rid of nilfs_get_writer() and nilfs_put_writer() pair used to
retain a writable FS-instance for a period.
The pair functions were making up some kind of recursive lock with a
mutex, but they became overkill since the commit
201913ed74. Furthermore, they caused
the following lockdep warning because the mutex can be released by a
task which didn't lock it:
=====================================
[ BUG: bad unlock balance detected! ]
-------------------------------------
kswapd0/422 is trying to release lock (&nilfs->ns_writer_mutex) at:
[<c1359ff5>] mutex_unlock+0x8/0xa
but there are no more locks to release!
other info that might help us debug this:
no locks held by kswapd0/422.
stack backtrace:
Pid: 422, comm: kswapd0 Not tainted 2.6.31-rc4-nilfs #51
Call Trace:
[<c1358f97>] ? printk+0xf/0x18
[<c104fea7>] print_unlock_inbalance_bug+0xcc/0xd7
[<c11578de>] ? prop_put_global+0x3/0x35
[<c1050195>] lock_release+0xed/0x1dc
[<c1359ff5>] ? mutex_unlock+0x8/0xa
[<c1359f83>] __mutex_unlock_slowpath+0xaf/0x119
[<c1359ff5>] mutex_unlock+0x8/0xa
[<d1284add>] nilfs_mdt_write_page+0xd8/0xe1 [nilfs2]
[<c1092653>] shrink_page_list+0x379/0x68d
[<c109171b>] ? isolate_pages_global+0xb4/0x18c
[<c1092bd2>] shrink_list+0x26b/0x54b
[<c10930be>] shrink_zone+0x20c/0x2a2
[<c10936b7>] kswapd+0x407/0x591
[<c1091667>] ? isolate_pages_global+0x0/0x18c
[<c1040603>] ? autoremove_wake_function+0x0/0x33
[<c10932b0>] ? kswapd+0x0/0x591
[<c104033b>] kthread+0x69/0x6e
[<c10402d2>] ? kthread+0x0/0x6e
[<c1003e33>] kernel_thread_helper+0x7/0x1a
This patch uses a reader/writer semaphore instead of the own lock and
kills this warning.
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Unlike on most other architectures ino_t is an unsigned int on s390.
So add an explicit cast to avoid this compile warning:
fs/nilfs2/recovery.c: In function 'recover_dsync_blocks':
fs/nilfs2/recovery.c:555: warning: format '%lu' expects type 'long unsigned int', but argument 3 has type 'ino_t'
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
The __nilfs_read_inode function is ignoring the error code returned
from nilfs_read_inode_common(), and wrongly delivers a success code
(zero) when it escapes from the function in erroneous cases.
This adds the missing error handling.
Signed-off-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
As early pci resume has already restored config for host
bridge and graphics device, don't need to restore it again,
This removes an original order hack for graphics device restore.
This fixed the resume hang issue found by Alan Stern on 845G,
caused by extra config restore on graphics device.
Cc: Stable Team <stable@kernel.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
Signed-off-by: Dave Airlie <airlied@linux.ie>
Add an option to multiplex counters output in the channel of
the group leader, ie: the first counter opened:
-M --multiplex
The effect is better serialized samples. This is especially
useful for tracepoint samples that need to be well serialized
for their post-processing.
Also make use of this option in 'perf sched'.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
blk_ioctl_discard duplicates large amounts of code from blkdev_issue_discard,
the only difference between the two is that blkdev_issue_discard needs to
send a barrier discard request and blk_ioctl_discard a non-barrier one,
and blk_ioctl_discard needs to wait on the request. To facilitates this
add a flags argument to blkdev_issue_discard to control both aspects of the
behaviour. This will be very useful later on for using the waiting
funcitonality for other callers.
Based on an earlier patch from Matthew Wilcox <matthew@wil.cx>.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
The commands are conceptually writes, and in the case of IDE and SCSI
commands actually are writes. They were only reads because we thought
that would interact better with the elevators. Now the elevators know
about discard requests, that advantage no longer exists.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Stacked devices do not. For now, just error out with -EINVAL. Later
we could make the limit apply on stacked devices too, for throttling
reasons.
This fixes
5a54cd13353bb3b88887604e2c980aa01e314309
and should go into 2.6.31 stable as well.
Cc: stable@kernel.org
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Implement blk_limits_io_opt() and make blk_queue_io_opt() a wrapper
around it. DM needs this to avoid poking at the queue_limits directly.
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Currently, there is a single in_flight counter measuring the number of
requests in the request_queue. But some monitoring tools would like to
know how many read requests and write requests are in progress. Split the
current in_flight counter into two seperate counters for read and write.
This information is exported as a sysfs attribute, as changing the
currently available stat files would break the existing tools.
Signed-off-by: Nikanth Karthikesan <knikanth@suse.de>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
BugLink: http://bugzilla.kernel.org/show_bug.cgi?id=13942
Bruno Premont noticed that aoe throws a BUG during umount of an XFS in
2.6.31:
[ 5259.349897] aoe: bi_io_vec is NULL
[ 5259.349940] ------------[ cut here ]------------
[ 5259.349958] kernel BUG at /usr/src/linux-2.6/drivers/block/aoe/aoeblk.c:177!
[ 5259.349990] invalid opcode: 0000 [#1]
The bio in question is a barrier. Jens Axboe suggested that such bios
need to be recognized and ended with -EOPNOTSUPP by any driver that
provides its own ->make_request_fn handler and does not handle
barriers.
In testing the changes below eliminate the BUG.
(Better would be real barrier support, something that Ed says he'll add
for later in the .32 cycle. For now, this at least gets rid of a bug
with crashing on an empty barrier. Jens)
Signed-off-by: Ed L. Cashin <ecashin@coraid.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
radeon_share.h was begining to give problem with include order in
respect of radeon.h. It's easier and also i think cleaner to move
what was in radeon_share.h into radeon.h. At the same time use the
extern keyword for function shared accross the module.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Move mtrr range and memory information printing to radeon_object_init,
this are memory information and initialization common to all GPU and
they better fit in this function. Will also prevent code duplication
with upcoming init path changes.
airlied: fixed warning introduced
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This convert r4xx to new init path it also fix few bugs.
Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Until now, trying to unlock the reiserfs write lock whereas the current
task doesn't hold it lead to a simple warning.
We should actually warn and panic in this case to avoid the user datas
to reach an unstable state.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jeff Mahoney <jeffm@suse.com>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Alexander Beregalov <a.beregalov@gmail.com>
Cc: Laurent Riffard <laurent.riffard@free.fr>
reiserfs_commit_write() is always called with the write lock held.
Thus the current calls to reiserfs_write_lock() in this function are
acquiring the lock recursively.
We can safely drop them.
This also solves further assumptions for this lock to be really
released while calling reiserfs_write_unlock().
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jeff Mahoney <jeffm@suse.com>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Alexander Beregalov <a.beregalov@gmail.com>
Cc: Laurent Riffard <laurent.riffard@free.fr>
reiserfs_mkdir() acquires the reiserfs lock, assuming it has been called
from the dir inodes callbacks, without the lock held.
But it can also be called from other internal sites such as
reiserfs_xattr_init() which already holds the lock. This recursive
locking leads to further wrong assumptions. For example, later calls
to reiserfs_mutex_lock_safe() won't actually unlock the reiserfs lock
the time we acquire a given mutex, creating unexpected lock inversions.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jeff Mahoney <jeffm@suse.com>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Alexander Beregalov <a.beregalov@gmail.com>
Cc: Laurent Riffard <laurent.riffard@free.fr>
reiserfs_xattr_init is called with the reiserfs write lock held, but
if the ".reiserfs_priv" entry is not created, we take the superblock
root directory inode mutex until .reiserfs_priv is created.
This creates a lock dependency inversion against other sites such as
reiserfs_file_release() which takes an inode mutex and the reiserfs
lock after.
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jeff Mahoney <jeffm@suse.com>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Alexander Beregalov <a.beregalov@gmail.com>
Cc: Laurent Riffard <laurent.riffard@free.fr>
When do_balance() balances the tree, a trick is performed to
provide the ability for other tree writers/readers to check whether
do_balance() is executing concurrently (requires CONFIG_REISERFS_CHECK).
This is done to protect concurrent accesses to the tree. The trick
is the following:
When do_balance is called, a unique global variable called cur_tb
takes a pointer to the current tree to be rebalanced.
Once do_balance finishes its work, cur_tb takes the NULL value.
Then, concurrent tree readers/writers just have to check the value
of cur_tb to ensure do_balance isn't executing concurrently.
If it is, then it proves that schedule() occured on do_balance(),
which then relaxed the bkl that protected the tree.
Now that the bkl has be turned into a mutex, this check is still
fine even though do_balance() becomes preemptible: the write lock
will not be automatically released on schedule(), so the tree is
still protected.
But this is only fine if we have a single reiserfs mountpoint.
Indeed, because the bkl is a global lock, it didn't allowed
concurrent executions between a tree reader/writer in a mount point
and a do_balance() on another tree from another mountpoint.
So assuming all these readers/writers weren't supposed to be
reentrant, the current check now sometimes detect false positives with
the current per-superblock mutex which allows this reentrancy.
This patch keeps the concurrent tree accesses check but moves it
per superblock, so that only trees from a same mount point are
checked to be not accessed concurrently.
[ Impact: fix spurious panic while running several reiserfs mount-points ]
Cc: Jeff Mahoney <jeffm@suse.com>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Alexander Beregalov <a.beregalov@gmail.com>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
While searching a pathname, an inode mutex can be acquired
in do_lookup() which calls reiserfs_lookup() which in turn
acquires the write lock.
On the other side reiserfs_fill_super() can acquire the write_lock
and then call reiserfs_lookup_privroot() which can acquire an
inode mutex (the root of the mount point).
So we theoretically risk an AB - BA lock inversion that could lead
to a deadlock.
As for other lock dependencies found since the bkl to mutex
conversion, the fix is to use reiserfs_mutex_lock_safe() which
drops the lock dependency to the write lock.
[ Impact: fix a possible deadlock with reiserfs ]
Cc: Jeff Mahoney <jeffm@suse.com>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Alexander Beregalov <a.beregalov@gmail.com>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
search_by_key() is the site which most requires the lock.
This is mostly because it is a very central function and also
because it releases/reaqcuires the write lock at least once each
time it is called.
Such release/reacquire creates a lot of contention in this place and
also opens more the window which let another thread changing the tree.
When it happens, the current path searching over the tree must be
retried from the beggining (the root) which is a wasteful and
time consuming recovery.
This patch factorizes two release/reacquire sequences:
- reading leaf nodes blocks
- reading current block
The latter immediately follows the former.
The whole sequence is safe as a single unlocked section because
we check just after if the tree has changed during these operations.
Cc: Jeff Mahoney <jeffm@suse.com>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Alexander Beregalov <a.beregalov@gmail.com>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
reiserfs_mutex_lock_safe() is a hack to avoid any dependency between
an internal reiserfs mutex and the write lock, it has been proposed
to follow the old bkl logic.
The code does the following:
while (!mutex_trylock(m)) {
reiserfs_write_unlock(s);
schedule();
reiserfs_write_lock(s);
}
It then imitate the implicit behaviour of the lock when it was
a Bkl and hadn't such dependency:
mutex_lock(m) {
if (fastpath)
let's go
else {
wait_for_mutex() {
schedule() {
unlock_kernel()
reacquire_lock_kernel()
}
}
}
}
The problem is that by using such explicit schedule(), we don't
benefit of the adaptive mutex spinning on owner.
The logic in use now is:
reiserfs_write_unlock(s);
mutex_lock(m); // -> possible adaptive spinning
reiserfs_write_lock(s);
[ Impact: restore the use of adaptive spinning mutexes in reiserfs ]
Cc: Jeff Mahoney <jeffm@suse.com>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Alexander Beregalov <a.beregalov@gmail.com>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
reiserfs_write_end() is a hot path in reiserfs.
We have two wasteful write lock lock/release inside that can be gathered
without changing the code logic.
This patch factorizes them out in a single protected section, reducing the
number of contentions inside.
[ Impact: reduce lock contention in a reiserfs hotpath ]
Cc: Jeff Mahoney <jeffm@suse.com>
Cc: Chris Mason <chris.mason@oracle.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Alexander Beregalov <a.beregalov@gmail.com>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>