Add a new flag DMF_MERGE_IS_OPTIONAL to struct mapped_device to indicate
whether the device can accept bios larger than the size its merge
function returns. When set, use this to send large bios to snapshots
which can split them if necessary. Snapshot I/O may be significantly
fragmented and this approach seems to improve peformance.
Before the patch, dm_set_device_limits restricted bio size to page size
if the underlying device had a merge function and the target didn't
provide a merge function. After the patch, dm_set_device_limits
restricts bio size to page size if the underlying device has a merge
function, doesn't have DMF_MERGE_IS_OPTIONAL flag and the target doesn't
provide a merge function.
The snapshot target can't provide a merge function because when the merge
function is called, it is impossible to determine where the bio will be
remapped. Previously this led us to impose a 4k limit, which we can
now remove if the snapshot store is located on a device without a merge
function. Together with another patch for optimizing full chunk writes,
it improves performance from 29MB/s to 40MB/s when writing to the
filesystem on snapshot store.
If the snapshot store is placed on a non-dm device with a merge function
(such as md-raid), device mapper still limits all bios to page size.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Remove 'discards_supported' from the dm_table structure. The same
information can be easily discovered from the table's target(s) in
dm_table_supports_discards().
Before this fix dm_table_supports_discards() would skip checking the
individual targets' 'discards_supported' flag if any one target in the
table didn't set num_discard_requests > 0. Now the per-target
'discards_supported' flag is effective at insuring the final DM device
advertises discard support. But, to be clear, targets that don't
support discards (!num_discard_requests) will not receive discard
requests.
Also DMWARN if a target sets 'discards_supported' override but forgets
to set 'num_discard_requests'.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
After the stack plugging introduction, these are called lockless.
Ensure that the counters are updated atomically.
Signed-off-by: Shaohua Li<shaohua.li@intel.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
MD and DM create a new bio_set for every metadevice. Each bio_set has an
integrity mempool attached regardless of whether the metadevice is
capable of passing integrity metadata. This is a waste of memory.
Instead we defer the allocation decision to MD and DM since we know at
metadevice creation time whether integrity passthrough is needed or not.
Automatic integrity mempool allocation can then be removed from
bioset_create() and we make an explicit integrity allocation for the
fs_bio_set.
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Acked-by: Mike Snitzer <snizer@redhat.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Code has been converted over to the new explicit on-stack plugging,
and delay users have been converted to use the new API for that.
So lets kill off the old plugging along with aops->sync_page().
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
This patch changes spin_lock_irq() to spin_lock() in dm_request_fn().
This patch is just a clean-up and no functional change.
The spin_lock_irq() was leftover from the early request-based dm code,
where map_request() used to enable interrupts.
Since current map_request() never enables interrupts, we can change it
to spin_lock() to match the prior spin_unlock().
Auditing through the dm and block-layer code called from
map_request(), I confirmed all functions save/restore interrupt
status, so no function returning with interrupts enabled.
Also I haven't observed any problem on my test environment which
uses scsi and lpfc driver after heavy I/O testing with occasional
path down/up.
Added BUG_ON() to detect breakage in future.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
kmirrord_wq, kcopyd_work and md->wq are created per dm instance and
serve only a single work item from the dm instance, so non-reentrant
workqueues would provide the same ordering guarantees as ordered ones
while allowing CPU affinity and use of the workqueues for other
purposes. Switch them to non-reentrant workqueues.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Convert all create[_singlethread]_work() users to the new
alloc[_ordered]_workqueue(). This conversion is mechanical and
doesn't introduce any behavior change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch replaces dm_mutex with _minor_lock in dm_blk_close()
and then removes it.
During the BKL conversion, commit 6e9624b8ca
(block: push down BKL into .open and .release) pushed lock_kernel()
down into dm_blk_open/close calls.
Commit 2a48fc0ab2
(block: autoconvert trivial BKL users to private mutex) converted it to a
local mutex, but _minor_lock is sufficient.
Signed-off-by: Milan Broz <mbroz@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
No longer needlessly hold md->bdev->bd_inode->i_mutex when changing the
size of a DM device. This additional locking is unnecessary because
i_size_write() is already protected by the existing critical section in
dm_swap_table(). DM already has a reference on md->bdev so the
associated bd_inode may be changed without lifetime concerns.
A negative side-effect of having held md->bdev->bd_inode->i_mutex was
that a concurrent DM device resize and flush (via fsync) would deadlock.
Dropping md->bdev->bd_inode->i_mutex eliminates this potential for
deadlock. The following reproducer no longer deadlocks:
https://www.redhat.com/archives/dm-devel/2009-July/msg00284.html
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Cc: stable@kernel.org
The "error" field in block_bio_complete is not assigned, leaving the memory area
uninitialized (keeping garbage data). Pass an additional tracepoint argument to
this event to initialize this field.
Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
CC: Steven Rostedt <rostedt@goodmis.org>
CC: Frederic Weisbecker <fweisbec@gmail.com>
CC: Ingo Molnar <mingo@elte.hu>
CC: Thomas Gleixner <tglx@linutronix.de>
CC: Li Zefan <lizf@cn.fujitsu.com>
CC: Alan.Brunelle@hp.com
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
* 'for-2.6.37/barrier' of git://git.kernel.dk/linux-2.6-block: (46 commits)
xen-blkfront: disable barrier/flush write support
Added blk-lib.c and blk-barrier.c was renamed to blk-flush.c
block: remove BLKDEV_IFL_WAIT
aic7xxx_old: removed unused 'req' variable
block: remove the BH_Eopnotsupp flag
block: remove the BLKDEV_IFL_BARRIER flag
block: remove the WRITE_BARRIER flag
swap: do not send discards as barriers
fat: do not send discards as barriers
ext4: do not send discards as barriers
jbd2: replace barriers with explicit flush / FUA usage
jbd2: Modify ASYNC_COMMIT code to not rely on queue draining on barrier
jbd: replace barriers with explicit flush / FUA usage
nilfs2: replace barriers with explicit flush / FUA usage
reiserfs: replace barriers with explicit flush / FUA usage
gfs2: replace barriers with explicit flush / FUA usage
btrfs: replace barriers with explicit flush / FUA usage
xfs: replace barriers with explicit flush / FUA usage
block: pass gfp_mask and flags to sb_issue_discard
dm: convey that all flushes are processed as empty
...
The block device drivers have all gained new lock_kernel
calls from a recent pushdown, and some of the drivers
were already using the BKL before.
This turns the BKL into a set of per-driver mutexes.
Still need to check whether this is safe to do.
file=$1
name=$2
if grep -q lock_kernel ${file} ; then
if grep -q 'include.*linux.mutex.h' ${file} ; then
sed -i '/include.*<linux\/smp_lock.h>/d' ${file}
else
sed -i 's/include.*<linux\/smp_lock.h>.*$/include <linux\/mutex.h>/g' ${file}
fi
sed -i ${file} \
-e "/^#include.*linux.mutex.h/,$ {
1,/^\(static\|int\|long\)/ {
/^\(static\|int\|long\)/istatic DEFINE_MUTEX(${name}_mutex);
} }" \
-e "s/\(un\)*lock_kernel\>[ ]*()/mutex_\1lock(\&${name}_mutex)/g" \
-e '/[ ]*cycle_kernel_lock();/d'
else
sed -i -e '/include.*\<smp_lock.h\>/d' ${file} \
-e '/cycle_kernel_lock()/d'
fi
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Rename __clone_and_map_flush to __clone_and_map_empty_flush for added
clarity.
Simplify logic associated with REQ_FLUSH conditionals.
Introduce a BUG_ON() and add a few more helpful comments to the code
so that it is clear that all flushes are empty.
Cleanup __split_and_process_bio() so that an empty flush isn't processed
by a 'sector_count' focused while loop.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Now queue_io() is called from dec_pending(), which may be called with
interrupts disabled, so queue_io() must not enable interrupts
unconditionally and must save/restore the current interrupts status.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Unlike REQ_HARDBARRIER, REQ_FLUSH/FUA doesn't mandate any ordering
against other bio's. This patch relaxes ordering around flushes.
* A flush bio is no longer deferred to workqueue directly. It's
processed like other bio's but __split_and_process_bio() uses
md->flush_bio as the clone source. md->flush_bio is initialized to
empty flush during md initialization and shared for all flushes.
* As a flush bio now travels through the same execution path as other
bio's, there's no need for dedicated error handling path either. It
can use the same error handling path in dec_pending(). Dedicated
error handling removed along with md->flush_error.
* When dec_pending() detects that a flush has completed, it checks
whether the original bio has data. If so, the bio is queued to the
deferred list w/ REQ_FLUSH cleared; otherwise, it's completed.
* As flush sequencing is handled in the usual issue/completion path,
dm_wq_work() no longer needs to handle flushes differently. Now its
only responsibility is re-issuing deferred bio's the same way as
_dm_request() would. REQ_FLUSH handling logic including
process_flush() is dropped.
* There's no reason for queue_io() and dm_wq_work() write lock
dm->io_lock. queue_io() now only uses md->deferred_lock and
dm_wq_work() read locks dm->io_lock.
* bio's no longer need to be queued on the deferred list while a flush
is in progress making DMF_QUEUE_IO_TO_THREAD unncessary. Drop it.
This avoids stalling the device during flushes and simplifies the
implementation.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
This patch converts request-based dm to support the new REQ_FLUSH/FUA.
The original request-based flush implementation depended on
request_queue blocking other requests while a barrier sequence is in
progress, which is no longer true for the new REQ_FLUSH/FUA.
In general, request-based dm doesn't have infrastructure for cloning
one source request to multiple targets, but the original flush
implementation had a special mostly independent path which can issue
flushes to multiple targets and sequence them. However, the
capability isn't currently in use and adds a lot of complexity.
Moreoever, it's unlikely to be useful in its current form as it
doesn't make sense to be able to send out flushes to multiple targets
when write requests can't be.
This patch rips out special flush code path and deals handles
REQ_FLUSH/FUA requests the same way as other requests. The only
special treatment is that REQ_FLUSH requests use the block address 0
when finding target, which is enough for now.
* added BUG_ON(!dm_target_is_valid(ti)) in dm_request_fn() as
suggested by Mike Snitzer
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Mike Snitzer <snitzer@redhat.com>
Tested-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
This patch converts bio-based dm to support REQ_FLUSH/FUA instead of
now deprecated REQ_HARDBARRIER.
* -EOPNOTSUPP handling logic dropped.
* Preflush is handled as before but postflush is dropped and replaced
with passing down REQ_FUA to member request_queues. This replaces
one array wide cache flush w/ member specific FUA writes.
* __split_and_process_bio() now calls __clone_and_map_flush() directly
for flushes and guarantees all FLUSH bio's going to targets are zero
` length.
* It's now guaranteed that all FLUSH bio's which are passed onto dm
targets are zero length. bio_empty_barrier() tests are replaced
with REQ_FLUSH tests.
* Empty WRITE_BARRIERs are replaced with WRITE_FLUSHes.
* Dropped unlikely() around REQ_FLUSH tests. Flushes are not unlikely
enough to be marked with unlikely().
* Block layer now filters out REQ_FLUSH/FUA bio's if the request_queue
doesn't support cache flushing. Advertise REQ_FLUSH | REQ_FUA
capability.
* Request based dm isn't converted yet. dm_init_request_based_queue()
resets flush support to 0 for now. To avoid disturbing request
based dm code, dm->flush_error is added for bio based dm while
requested based dm continues to use dm->barrier_error.
Lightly tested linear, stripe, raid1, snap and crypt targets. Please
proceed with caution as I'm not familiar with the code base.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: dm-devel@redhat.com
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Barrier is deemed too heavy and will soon be replaced by FLUSH/FUA
requests. Deprecate barrier. All REQ_HARDBARRIERs are failed with
-EOPNOTSUPP and blk_queue_ordered() is replaced with simpler
blk_queue_flush().
blk_queue_flush() takes combinations of REQ_FLUSH and FUA. If a
device has write cache and can flush it, it should set REQ_FLUSH. If
the device can handle FUA writes, it should also set REQ_FUA.
All blk_queue_ordered() users are converted.
* ORDERED_DRAIN is mapped to 0 which is the default value.
* ORDERED_DRAIN_FLUSH is mapped to REQ_FLUSH.
* ORDERED_DRAIN_FLUSH_FUA is mapped to REQ_FLUSH | REQ_FUA.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Boaz Harrosh <bharrosh@panasas.com>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Nick Piggin <npiggin@kernel.dk>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Jeremy Fitzhardinge <jeremy@xensource.com>
Cc: Chris Wright <chrisw@sous-sol.org>
Cc: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Cc: Geert Uytterhoeven <Geert.Uytterhoeven@sonycom.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Alasdair G Kergon <agk@redhat.com>
Cc: Pierre Ossman <drzeus@drzeus.cx>
Cc: Stefan Weinhuber <wein@de.ibm.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Update __clone_and_map_discard to loop across all targets in a DM
device's table when it processes a discard bio. If a discard crosses a
target boundary it must be split accordingly.
Update __issue_target_requests and __issue_target_request to allow a
cloned discard bio to have a custom start sector and size.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Split max_io_len_target_boundary out of max_io_len so that the discard
support can make use of it without duplicating max_io_len code.
Avoiding max_io_len's split_io logic enables DM's discard support to
submit the entire discard request to a target. But discards must still
be split on target boundaries.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Rename __flush_target to __issue_target_request now that it is used to
issue both flush and discard requests.
Introduce __issue_target_requests as a convenient wrapper to
__issue_target_request 'num_flush_requests' or 'num_discard_requests'
times per target.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Allow discards to be passed through to linear mappings if at least one
underlying device supports it. Discards will be forwarded only to
devices that support them.
A target that supports discards should set num_discard_requests to
indicate how many times each discard request must be submitted to it.
Verify table's underlying devices support discards prior to setting the
associated DM device as capable of discards (via QUEUE_FLAG_DISCARD).
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Reviewed-by: Joe Thornber <thornber@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
'target_request_nr' is a more generic name that reflects the fact that
it will be used for both flush and discard support.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Change bio-based mapped devices no longer to have a fully initialized
request_queue (request_fn, elevator, etc). This means bio-based DM
devices no longer register elevator sysfs attributes ('iosched/' tree
or 'scheduler' other than "none").
In contrast, a request-based DM device will continue to have a full
request_queue and will register elevator sysfs attributes. Therefore
a user can determine a DM device's type by checking if elevator sysfs
attributes exist.
First allocate a minimalist request_queue structure for a DM device
(needed for both bio and request-based DM).
Initialization of a full request_queue is deferred until it is known
that the DM device is request-based, at the end of the table load
sequence.
Factor DM device's request_queue initialization:
- common to both request-based and bio-based into dm_init_md_queue().
- specific to request-based into dm_init_request_based_queue().
The md->type_lock mutex is used to protect md->queue, in addition to
md->type, during table_load().
A DM device's first table_load will establish the immutable md->type.
But md->queue initialization, based on md->type, may fail at that time
(because blk_init_allocated_queue cannot allocate memory). Therefore
any subsequent table_load must (re)try dm_setup_md_queue independently of
establishing md->type.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Acked-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Determine whether a mapped device is bio-based or request-based when
loading its first (inactive) table and don't allow that to be changed
later.
This patch performs different device initialisation in each of the two
cases. (We don't think it's necessary to add code to support changing
between the two types.)
Allowed md->type transitions:
DM_TYPE_NONE to DM_TYPE_BIO_BASED
DM_TYPE_NONE to DM_TYPE_REQUEST_BASED
We now prevent table_load from replacing the inactive table with a
conflicting type of table even after an explicit table_clear.
Introduce 'type_lock' into the struct mapped_device to protect md->type
and to prepare for the next patch that will change the queue
initialization and allocate memory while md->type_lock is held.
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Acked-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
drivers/md/dm-ioctl.c | 15 +++++++++++++++
drivers/md/dm.c | 37 ++++++++++++++++++++++++++++++-------
drivers/md/dm.h | 5 +++++
include/linux/dm-ioctl.h | 4 ++--
4 files changed, 52 insertions(+), 9 deletions(-)
When processing barriers, skip the second flush if processing the bio
failed with -EOPNOTSUPP. This can happen with discard+barrier requests.
If the device doesn't support discard, there would be two useless
SYNCHRONIZE CACHE commands. The first dm_flush cannot be so easily
optimized out, so we leave it there.
Previously, -EOPNOTSUPP could be received in dec_pending only with empty
barriers and we ignored that error, assuming the device not supporting
cache flushes has cache always consistent. With the addition of discard
barriers, this -EOPNOTSUPP can also be generated by discards and we
must record it in md->barrier_error for process_barrier.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch separates the device deletion code from dm_put()
to make sure the deletion happens in the process context.
By this patch, device deletion always occurs in an ioctl (process)
context and dm_put() can be called in interrupt context.
As a result, the request-based dm's bad dm_put() usage pointed out
by Mikulas below disappears.
http://marc.info/?l=dm-devel&m=126699981019735&w=2
Without this patch, I confirmed there is a case to crash the system:
dm_put() => dm_table_destroy() => vfree() => BUG_ON(in_interrupt())
Some more backgrounds and details:
In request-based dm, a device opener can remove a mapped_device
while the last request is still completing, because bios in the last
request complete first and then the device opener can close and remove
the mapped_device before the last request completes:
CPU0 CPU1
=================================================================
<<INTERRUPT>>
blk_end_request_all(clone_rq)
blk_update_request(clone_rq)
bio_endio(clone_bio) == end_clone_bio
blk_update_request(orig_rq)
bio_endio(orig_bio)
<<I/O completed>>
dm_blk_close()
dev_remove()
dm_put(md)
<<Free md>>
blk_finish_request(clone_rq)
....
dm_end_request(clone_rq)
free_rq_clone(clone_rq)
blk_end_request_all(orig_rq)
rq_completed(md)
So request-based dm used dm_get()/dm_put() to hold md for each I/O
until its request completion handling is fully done.
However, the final dm_put() can call the device deletion code which
must not be run in interrupt context and may cause kernel panic.
To solve the problem, this patch moves the device deletion code,
dm_destroy(), to predetermined places that is actually deleting
the mapped_device in ioctl (process) context, and changes dm_put()
just to decrement the reference count of the mapped_device.
By this change, dm_put() can be used in any context and the symmetric
model below is introduced:
dm_create(): create a mapped_device
dm_destroy(): destroy a mapped_device
dm_get(): increment the reference count of a mapped_device
dm_put(): decrement the reference count of a mapped_device
dm_destroy() waits for all references of the mapped_device to disappear,
then deletes the mapped_device.
dm_destroy() uses active waiting with msleep(1), since deleting
the mapped_device isn't performance-critical task.
And since at this point, nobody opens the mapped_device and no new
reference will be taken, the pending counts are just for racing
completing activity and will eventually decrease to zero.
For the unlikely case of the forced module unload, dm_destroy_immediate(),
which doesn't wait and forcibly deletes the mapped_device, is also
introduced and used in dm_hash_remove_all(). Otherwise, "rmmod -f"
may be stuck and never return.
And now, because the mapped_device is deleted at this point, subsequent
accesses to the mapped_device may cause NULL pointer references.
Cc: stable@kernel.org
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch prevents access to mapped_device which is being deleted.
Currently, even after a mapped_device has been removed from the hash,
it could be accessed through idr_find() using minor number.
That could cause a race and NULL pointer reference below:
CPU0 CPU1
------------------------------------------------------------------
dev_remove(param)
down_write(_hash_lock)
dm_lock_for_deletion(md)
spin_lock(_minor_lock)
set_bit(DMF_DELETING)
spin_unlock(_minor_lock)
__hash_remove(hc)
up_write(_hash_lock)
dev_status(param)
md = find_device(param)
down_read(_hash_lock)
__find_device_hash_cell(param)
dm_get_md(param->dev)
md = dm_find_md(dev)
spin_lock(_minor_lock)
md = idr_find(MINOR(dev))
spin_unlock(_minor_lock)
dm_put(md)
free_dev(md)
dm_get(md)
up_read(_hash_lock)
__dev_status(md, param)
dm_put(md)
This patch fixes such problems.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Cc: stable@kernel.org
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
The open and release block_device_operations are currently
called with the BKL held. In order to change that, we must
first make sure that all drivers that currently rely
on this have no regressions.
This blindly pushes the BKL into all .open and .release
operations for all block drivers to prepare for the
next step. The drivers can subsequently replace the BKL
with their own locks or remove it completely when it can
be shown that it is not needed.
The functions blkdev_get and blkdev_put are the only
remaining users of the big kernel lock in the block
layer, besides a few uses in the ioctl code, none
of which need to serialize with blkdev_{get,put}.
Most of these two functions is also under the protection
of bdev->bd_mutex, including the actual calls to
->open and ->release, and the common code does not
access any global data structures that need the BKL.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Remove the current bio flags and reuse the request flags for the bio, too.
This allows to more easily trace the type of I/O from the filesystem
down to the block driver. There were two flags in the bio that were
missing in the requests: BIO_RW_UNPLUG and BIO_RW_AHEAD. Also I've
renamed two request flags that had a superflous RW in them.
Note that the flags are in bio.h despite having the REQ_ name - as
blkdev.h includes bio.h that is the only way to go for now.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Remove all the trivial wrappers for the cmd_type and cmd_flags fields in
struct requests. This allows much easier grepping for different request
types instead of unwinding through macros.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Set a new DM_UEVENT_GENERATED_FLAG when returning from ioctls to
indicate that a uevent was actually generated. This tells the userspace
caller that it may need to wait for the event to be processed.
Signed-off-by: Peter Rajnoha <prajnoha@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Free the dm_io structure before calling bio_endio() instead of after it,
to ensure that the io_pool containing it is not referenced after it is
freed.
This partially fixes a problem described here
https://www.redhat.com/archives/dm-devel/2010-February/msg00109.html
thread 1:
bio_endio(bio, io_error);
/* scheduling happens */
thread 2:
close the device
remove the device
thread 1:
free_io(md, io);
Thread 2, when removing the device, sees non-empty md->io_pool (because the
io hasn't been freed by thread 1 yet) and may crash with BUG in mempool_free.
Thread 1 may also crash, when freeing into a nonexisting mempool.
To fix this we must make sure that bio_endio() is the last call and
the md structure is not accessed afterwards.
There is another bio_endio in process_barrier, but it is called from the thread
and the thread is destroyed prior to freeing the mempools, so this call is
not affected by the bug.
A similar bug exists with module unloads - the module may be unloaded
immediately after bio_endio - but that is more difficult to fix.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@kernel.org
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Remove the dm_get() in dm_table_get_md() because dm_table_get_md() could
be called from presuspend/postsuspend, which are called while
mapped_device is in DMF_FREEING state, where dm_get() is not allowed.
Justification for that is the lifetime of both objects: As far as the
current dm design/implementation, mapped_device is never freed while
targets are doing something, because dm core waits for targets to become
quiet in dm_put() using presuspend/postsuspend. So targets should be
able to touch mapped_device without holding reference count of the
mapped_device, and we should allow targets to touch mapped_device even
if it is in DMF_FREEING state.
Backgrounds:
I'm trying to remove the multipath internal queue, since dm core now has
a generic queue for request-based dm. In the patch-set, the multipath
target wants to request dm core to start/stop queue. One of such
start/stop requests can happen during postsuspend() while the target
waits for pg-init to complete, because the target stops queue when
starting pg-init and tries to restart it when completing pg-init. Since
queue belongs to mapped_device, it involves calling dm_table_get_md()
and dm_put(). On the other hand, postsuspend() is called in dm_put()
for mapped_device which is in DMF_FREEING state, and that triggers
BUG_ON(DMF_FREEING) in the 2nd dm_put().
I had tried to solve this problem by changing only multipath not to
touch mapped_device which is in DMF_FREEING state, but I couldn't and I
came up with a question why we need dm_get() in dm_table_get_md().
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch fixes the problem that system may stall if target's ->map_rq
returns DM_MAPIO_REQUEUE in map_request().
E.g. stall happens on 1 CPU box when a dm-mpath device with queue_if_no_path
bounces between all-paths-down and paths-up on I/O load.
When target's ->map_rq returns DM_MAPIO_REQUEUE, map_request() requeues
the request and returns to dm_request_fn(). Then, dm_request_fn()
doesn't exit the I/O dispatching loop and continues processing
the requeued request again.
This map and requeue loop can be done with interrupt disabled,
so 1 CPU system can be stalled if this situation happens.
For example, commands below can stall my 1 CPU box within 1 minute or so:
# dmsetup table mp
mp: 0 2097152 multipath 1 queue_if_no_path 0 1 1 service-time 0 1 2 8:144 1 1
# while true; do dd if=/dev/mapper/mp of=/dev/null bs=1M count=100; done &
# while true; do \
> dmsetup message mp 0 "fail_path 8:144" \
> dmsetup suspend --noflush mp \
> dmsetup resume mp \
> dmsetup message mp 0 "reinstate_path 8:144" \
> done
To fix the problem above, this patch changes dm_request_fn() to exit
the I/O dispatching loop once if a request is requeued in map_request().
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Cc: stable@kernel.org
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch adds the exported dm_suspended() function so that targets
can check whether or not they are suspended.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Cc: Mike Anderson <andmike@linux.vnet.ibm.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch renames dm_suspended() to dm_suspended_md() and
keeps it internal to dm.
No functional change.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Cc: Mike Anderson <andmike@linux.vnet.ibm.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch moves DMF_SUSPENDED flag set before postsuspend.
No one should care about the ordering, because the flag set and
the postsuspend are protected by a single lock, md->suspend_lock,
and all strict flag-checkers take the lock.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Cc: Mike Anderson <andmike@linux.vnet.ibm.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
When swapping a new table into place, retain the old table until
its replacement is in place.
An old check for an empty table is removed because this is enforced
in populate_table().
__unbind() becomes redundant when followed by __bind().
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
When replacing a mapped device's table during a 'resume', delay the
destruction of the old table until the new one is successfully in place.
This will make it easier for a later patch to transfer internal state
information from the old table to the new one (something we do not currently
support) while giving us more options for reversion if a later part
of the operation fails.
Devices are always in the suspended state during dm_swap_table().
This patch reinforces the requirement that all I/O must have been
flushed from the table targets while in this state (including any in
workqueues). In the case of 'noflush' suspending, unprocessed
I/O should have been 'pushed back' to the dm core prior to this point,
for resubmission after the new table is in place.
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Add dm_deleting_md to check whether or not a given mapped
device is currently being deleted.
Signed-off-by: Mike Anderson <andmike@linux.vnet.ibm.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch adds barrier support for request-based dm.
CORE DESIGN
The design is basically same as bio-based dm, which emulates barrier
by mapping empty barrier bios before/after a barrier I/O.
But request-based dm has been using struct request_queue for I/O
queueing, so the block-layer's barrier mechanism can be used.
o Summary of the block-layer's behavior (which is depended by dm-core)
Request-based dm uses QUEUE_ORDERED_DRAIN_FLUSH ordered mode for
I/O barrier. It means that when an I/O requiring barrier is found
in the request_queue, the block-layer makes pre-flush request and
post-flush request just before and just after the I/O respectively.
After the ordered sequence starts, the block-layer waits for all
in-flight I/Os to complete, then gives drivers the pre-flush request,
the barrier I/O and the post-flush request one by one.
It means that the request_queue is stopped automatically by
the block-layer until drivers complete each sequence.
o dm-core
For the barrier I/O, treats it as a normal I/O, so no additional
code is needed.
For the pre/post-flush request, flushes caches by the followings:
1. Make the number of empty barrier requests required by target's
num_flush_requests, and map them (dm_rq_barrier()).
2. Waits for the mapped barriers to complete (dm_rq_barrier()).
If error has occurred, save the error value to md->barrier_error
(dm_end_request()).
(*) Basically, the first reported error is taken.
But -EOPNOTSUPP supersedes any error and DM_ENDIO_REQUEUE
follows.
3. Requeue the pre/post-flush request if the error value is
DM_ENDIO_REQUEUE. Otherwise, completes with the error value
(dm_rq_barrier_work()).
The pre/post-flush work above is done in the kernel thread (kdmflush)
context, since memory allocation which might sleep is needed in
dm_rq_barrier() but sleep is not allowed in dm_request_fn(), which is
an irq-disabled context.
Also, clones of the pre/post-flush request share an original, so
such clones can't be completed using the softirq context.
Instead, complete them in the context of underlying device drivers.
It should be safe since there is no I/O dispatching during
the completion of such clones.
For suspend, the workqueue of kdmflush needs to be flushed after
the request_queue has been stopped. Otherwise, the next flush work
can be kicked even after the suspend completes.
TARGET INTERFACE
No new interface is added.
Just use the existing num_flush_requests in struct target_type
as same as bio-based dm.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch moves dm_end_request() to make the next patch more readable.
No functional change.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch factors out the clone completion code, dm_done(),
from dm_softirq_done() in preparation for a subsequent patch.
No functional change.
dm_done() will be used in barrier completion, which can't use and
doesn't need softirq. The softirq_done callback needs to get a clone
from an original request but it can't in the case of barrier, where
an original request is shared by multiple clones. On the other hand,
the completion of barrier clones doesn't involve re-submitting requests,
which was the primary reason of the need for softirq.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch changes the counter for the number of in_flight I/Os
to md->pending from q->in_flight in preparation for a later patch.
No functional change.
Request-based dm used q->in_flight to count the number of in-flight
clones assuming the counter is always incremented for an in-flight
original request and original:clone is 1:1 relationship.
However, it this no longer true for barrier requests.
So use md->pending to count the number of in-flight clones.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
The semantics of bio-based dm were changed recently in the case of
suspend with "--nolockfs" but without "--noflush".
Before 2.6.30, I/Os submitted before the suspend invocation were always
flushed. From 2.6.30 onwards, I/Os submitted before the suspend
invocation might not be flushed. (For details, see
http://marc.info/?t=123994433400003&r=1&w=2)
This patch brings the behaviour of request-based dm into line with
bio-based dm, simplifying the code and preparing for a subsequent patch
that will wait for all in_flight I/Os to complete without stopping
request_queue and use dm_wait_for_completion() for it.
This change in semantics simplifies the suspend code as follows:
o Suspend is implemented as stopping request_queue
in request-based dm, and all I/Os are queued in the request_queue
even after suspend is invoked.
o In the old semantics, we had to track whether I/Os were
queued before or after the suspend invocation, so a special
barrier-like request called 'suspend marker' was introduced.
o With the new semantics, we don't need to flush any I/O
so we can remove the marker and the code related to the marker
handling and I/O flushing.
After removing this codes, the suspend sequence is now:
1. Flush all I/Os by lock_fs() if needed.
2. Stop dispatching any I/O by stopping the request_queue.
3. Wait for all in-flight I/Os to be completed or requeued.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch factors out the request cloning code in dm_prep_fn()
as clone_rq(). No functional change.
This patch is a preparation for a later patch in this series which needs to
make clones from an original barrier request.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch adds the gfp_mask argument to alloc_rq_tio().
No functional change.
This patch is a preparation for a later patch in this series which needs to
allocate tio (for barrier I/O) with different allocation flag (GFP_NOIO) from
the one in the normal I/O code path.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch changes the argument of map_request() to clone request
from original request. No functional change.
This patch is a preparation for PATCH 9, which needs to use
map_request() for clones sharing an original barrier request.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch adds md_in_flight() to get the number of in_flight I/Os.
No functional change.
This patch is a preparation for a later patch in this series, which
changes I/O counter to md->pending from q->in_flight in request-based dm.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Allocate "struct io" from a slab.
This patch changes dm-io, so that "struct io" is allocated from a slab cache.
It used to be allocated with kmalloc. Allocating from a slab will be needed
for the next patch, because it requires a special alignment of "struct io"
and kmalloc cannot meet this alignment.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Multiple instances of dec_pending() can run concurrently so a lock is
needed when it saves the first error code.
I have never experienced actual problem without locking and just found
this during code inspection while implementing the barrier support
patch for request-based dm.
This patch adds the locking.
I've done compile, boot and basic I/O testings.
Cc: stable@kernel.org
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Add missing del_gendisk() to error path when creation of workqueue fails.
Otherwice there is a resource leak and following warning is shown:
WARNING: at fs/sysfs/dir.c:487 sysfs_add_one+0xc5/0x160()
sysfs: cannot create duplicate filename '/devices/virtual/block/dm-0'
Cc: stable@kernel.org
Signed-off-by: Zdenek Kabelac <zkabelac@redhat.com>
Reviewed-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Commit a9327cac44 added seperate read
and write statistics of in_flight requests. And exported the number
of read and write requests in progress seperately through sysfs.
But Corrado Zoccolo <czoccolo@gmail.com> reported getting strange
output from "iostat -kx 2". Global values for service time and
utilization were garbage. For interval values, utilization was always
100%, and service time is higher than normal.
So this was reverted by commit 0f78ab9899
The problem was in part_round_stats_single(), I missed the following:
if (now == part->stamp)
return;
- if (part->in_flight) {
+ if (part_in_flight(part)) {
__part_stat_add(cpu, part, time_in_queue,
part_in_flight(part) * (now - part->stamp));
__part_stat_add(cpu, part, io_ticks, (now - part->stamp));
With this chunk included, the reported regression gets fixed.
Signed-off-by: Nikanth Karthikesan <knikanth@suse.de>
--
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>
Get rid of any functions that test for these bits and make callers
use bio_rw_flagged() directly. Then it is at least directly apparent
what variable and flag they check.
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
This patch removes DM's bio-based vs request-based conditional setting
of next_ordered. For bio-based DM the next_ordered check is no longer a
concern (as that check is now in the __make_request path). For
request-based DM the default of QUEUE_ORDERED_NONE is now appropriate.
bio-based DM was changed to work-around the previously misplaced
next_ordered check with this commit:
99360b4c18
request-based DM does not yet support barriers but reacted to the above
bio-based DM change with this commit:
5d67aa2366
The above changes are no longer needed given Neil Brown's recent fix to
put the next_ordered check in the __make_request path:
db64f680ba
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Cc: NeilBrown <neilb@suse.de>
Acked-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Acked-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch restores stacking ability to the block layer integrity
infrastructure by creating a set of dedicated bip slabs. Each bip slab
has an embedded bio_vec array at the end. This cuts down on memory
allocations and also simplifies the code compared to the original bvec
version. Only the largest bip slab is backed by a mempool. The pool is
contained in the bio_set so stacking drivers can ensure forward
progress.
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <axboe@carl.(none)>
This patch disables interrupt when taking map_lock to avoid
lockdep warnings in request-based dm.
request-based dm takes map_lock after taking queue_lock with
disabling interrupt:
spin_lock_irqsave(queue_lock)
q->request_fn() == dm_request_fn()
=> dm_get_table()
=> read_lock(map_lock)
while queue_lock could be (but isn't) taken in interrupt context.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Acked-by: Christof Schmitt <christof.schmitt@de.ibm.com>
Acked-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Request-based dm doesn't have barrier support yet.
So we need to set QUEUE_ORDERED_DRAIN only for bio-based dm.
Since the device type is decided at the first table loading time,
the flag set is deferred until then.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Acked-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch enables request-based dm.
o Request-based dm and bio-based dm coexist, since there are
some target drivers which are more fitting to bio-based dm.
Also, there are other bio-based devices in the kernel
(e.g. md, loop).
Since bio-based device can't receive struct request,
there are some limitations on device stacking between
bio-based and request-based.
type of underlying device
bio-based request-based
----------------------------------------------
bio-based OK OK
request-based -- OK
The device type is recognized by the queue flag in the kernel,
so dm follows that.
o The type of a dm device is decided at the first table binding time.
Once the type of a dm device is decided, the type can't be changed.
o Mempool allocations are deferred to at the table loading time, since
mempools for request-based dm are different from those for bio-based
dm and needed mempool type is fixed by the type of table.
o Currently, request-based dm supports only tables that have a single
target. To support multiple targets, we need to support request
splitting or prevent bio/request from spanning multiple targets.
The former needs lots of changes in the block layer, and the latter
needs that all target drivers support merge() function.
Both will take a time.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch adds core functions for request-based dm.
When struct mapped device (md) is initialized, md->queue has
an I/O scheduler and the following functions are used for
request-based dm as the queue functions:
make_request_fn: dm_make_request()
pref_fn: dm_prep_fn()
request_fn: dm_request_fn()
softirq_done_fn: dm_softirq_done()
lld_busy_fn: dm_lld_busy()
Actual initializations are done in another patch (PATCH 2).
Below is a brief summary of how request-based dm behaves, including:
- making request from bio
- cloning, mapping and dispatching request
- completing request and bio
- suspending md
- resuming md
bio to request
==============
md->queue->make_request_fn() (dm_make_request()) calls __make_request()
for a bio submitted to the md.
Then, the bio is kept in the queue as a new request or merged into
another request in the queue if possible.
Cloning and Mapping
===================
Cloning and mapping are done in md->queue->request_fn() (dm_request_fn()),
when requests are dispatched after they are sorted by the I/O scheduler.
dm_request_fn() checks busy state of underlying devices using
target's busy() function and stops dispatching requests to keep them
on the dm device's queue if busy.
It helps better I/O merging, since no merge is done for a request
once it is dispatched to underlying devices.
Actual cloning and mapping are done in dm_prep_fn() and map_request()
called from dm_request_fn().
dm_prep_fn() clones not only request but also bios of the request
so that dm can hold bio completion in error cases and prevent
the bio submitter from noticing the error.
(See the "Completion" section below for details.)
After the cloning, the clone is mapped by target's map_rq() function
and inserted to underlying device's queue using
blk_insert_cloned_request().
Completion
==========
Request completion can be hooked by rq->end_io(), but then, all bios
in the request will have been completed even error cases, and the bio
submitter will have noticed the error.
To prevent the bio completion in error cases, request-based dm clones
both bio and request and hooks both bio->bi_end_io() and rq->end_io():
bio->bi_end_io(): end_clone_bio()
rq->end_io(): end_clone_request()
Summary of the request completion flow is below:
blk_end_request() for a clone request
=> blk_update_request()
=> bio->bi_end_io() == end_clone_bio() for each clone bio
=> Free the clone bio
=> Success: Complete the original bio (blk_update_request())
Error: Don't complete the original bio
=> blk_finish_request()
=> rq->end_io() == end_clone_request()
=> blk_complete_request()
=> dm_softirq_done()
=> Free the clone request
=> Success: Complete the original request (blk_end_request())
Error: Requeue the original request
end_clone_bio() completes the original request on the size of
the original bio in successful cases.
Even if all bios in the original request are completed by that
completion, the original request must not be completed yet to keep
the ordering of request completion for the stacking.
So end_clone_bio() uses blk_update_request() instead of
blk_end_request().
In error cases, end_clone_bio() doesn't complete the original bio.
It just frees the cloned bio and gives over the error handling to
end_clone_request().
end_clone_request(), which is called with queue lock held, completes
the clone request and the original request in a softirq context
(dm_softirq_done()), which has no queue lock, to avoid a deadlock
issue on submission of another request during the completion:
- The submitted request may be mapped to the same device
- Request submission requires queue lock, but the queue lock
has been held by itself and it doesn't know that
The clone request has no clone bio when dm_softirq_done() is called.
So target drivers can't resubmit it again even error cases.
Instead, they can ask dm core for requeueing and remapping
the original request in that cases.
suspend
=======
Request-based dm uses stopping md->queue as suspend of the md.
For noflush suspend, just stops md->queue.
For flush suspend, inserts a marker request to the tail of md->queue.
And dispatches all requests in md->queue until the marker comes to
the front of md->queue. Then, stops dispatching request and waits
for the all dispatched requests to complete.
After that, completes the marker request, stops md->queue and
wake up the waiter on the suspend queue, md->wait.
resume
======
Starts md->queue.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Currently, device-mapper maintains a separate instance of 'struct
queue_limits' for each table of each device. When the configuration of
a device is to be changed, first its table is loaded and this structure
is populated, then the device is 'resumed' and the calculated
queue_limits are applied.
This places restrictions on how userspace may process related devices,
where it is often advantageous to 'load' tables for several devices
at once before 'resuming' them together. As the new queue_limits
only take effect after the 'resume', if they are changing and one
device uses another, the latter must be 'resumed' before the former
may be 'loaded'.
This patch moves the calculation of these queue_limits out of
the 'load' operation into 'resume'. Since we are no longer
pre-calculating this struct, we no longer need to maintain copies
within our dm structs.
dm_set_device_limits() now passes the 'start' of the device's
data area (aka pe_start) as the 'offset' to blk_stack_limits().
init_valid_queue_limits() is replaced by blk_set_default_limits().
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Cc: martin.petersen@oracle.com
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Add support for passing a 32 bit "cookie" into the kernel with the
DM_SUSPEND, DM_DEV_RENAME and DM_DEV_REMOVE ioctls. The (unsigned)
value of this cookie is returned to userspace alongside the uevents
issued by these ioctls in the variable DM_COOKIE.
This means the userspace process issuing these ioctls can be notified
by udev after udev has completed any actions triggered.
To minimise the interface extension, we pass the cookie into the
kernel in the event_nr field which is otherwise unused when calling
these ioctls. Incrementing the version number allows userspace to
determine in advance whether or not the kernel supports the cookie.
If the kernel does support this but userspace does not, there should
be no impact as the new variable will just get ignored.
Signed-off-by: Milan Broz <mbroz@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Pass empty barrier flushes to the targets in dm_flush().
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Introduce num_flush_requests for a target to set to say how many flush
instructions (empty barriers) it wants to receive. These are sent by
__clone_and_map_empty_barrier with map_info->flush_request going from 0
to (num_flush_requests - 1).
Old targets without flush support won't receive any flush requests.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Remove the check that the size of the cloned bio is not zero because a
subsequent patch needs to send zero-sized barriers down this path.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
If the underlying device doesn't support barriers and dm receives a
barrier, it waits until all requests on that device drain so it no
longer needs to report -EOPNOTSUPP to the caller.
This patch deals with the confusing situation when moving a volume from
one physical device to another triggers an EOPNOTSUPP on a volume that
didn't report it before.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
With the following patches, more than one error can occur during
processing. Change md->barrier_error so that only the first one is
recorded and returned to the caller.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
If barrier request was returned with DM_ENDIO_REQUEUE,
requeue it in dm_wq_work instead of dec_pending.
This allows us to correctly handle a situation when some targets
are asking for a requeue and other targets signal an error.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Make dm_flush return void.
The first error during flush is stored in md->barrier_error instead.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Fix a potential deadlock when creating multiple snapshots by holding a
reference to struct block_device for the whole lifecycle of every dm
device instead of obtaining it independently at each point it is needed.
bdget_disk() was called while the device was being suspended, in
dm_suspend(). However there could be other devices already suspended,
for example when creating additional snapshots of a device. bdget_disk()
can wait for IO and allocate memory resulting in waiting for the
already-suspended device - deadlock.
This patch changes the code so that it gets the reference to struct
block_device when struct mapped_device is allocated and initialized in
alloc_dev() where it is always OK to allocate memory or wait for I/O.
It drops the reference when it is destroyed in free_dev(). Thus there
is no call to bdget_disk() while any device is suspended.
Previously unlock_fs() was called only if bdev was held. Now it is
called unconditionally, but the superfluous calls are harmless because
it returns immediately if the filesystem was not previously frozen.
This patch also now allows the device size to be changed in a
noflush suspend because the bdev is held. This has no adverse effect.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Rename suspended_bdev to bdev.
This patch doesn't change any functionality, just renames the variable.
In the next patch, the variable will be used even for non-suspended device.
(Pre-requisite for the per-target barrier support patches.)
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
A bio that has two or more vector entries, size less than or equal to
page size, that crosses a stripe boundary of an underlying md device is
accepted by device mapper (it conforms to all its limits) but not by the
underlying device.
The fix is: If device mapper selects the one-page maximum request size,
it also needs to set its own q->merge_bvec_fn to reject any bios with
multiple vector entries that span more pages.
The problem was discovered in the following scenario:
* MD - RAID-0
* LV on the top of it (raid1, snapshot or striped with chunk
size/stripe larger than RAID-0 stripe)
* one of the logical volumes is exported to xen domU
* inside xen domU it is partitioned, the key point is that the partition
must be unaligned on page boundary (fdisk normally aligns the partition to
63 sectors which will trigger it)
* install the system on the partitioned disk in domU
This causes I/O failures in dom0.
Reference: https://bugzilla.redhat.com/show_bug.cgi?id=223947
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Do not process sysfs attributes when device is being destroyed.
Otherwise code can cause
BUG_ON(test_bit(DMF_FREEING, &md->flags));
in dm_put() call.
Cc: stable@kernel.org
Signed-off-by: Milan Broz <mbroz@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
When porting blktrace to tracepoints, we changed to trace/block.h
for trace prober declarations.
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
TRACE_EVENT is a more generic way to define tracepoints. Doing so adds
these new capabilities to this tracepoint:
- zero-copy and per-cpu splice() tracing
- binary tracing without printf overhead
- structured logging records exposed under /debug/tracing/events
- trace events embedded in function tracer output and other plugins
- user-defined, per tracepoint filter expressions
...
Cons:
- no dev_t info for the output of plug, unplug_timer and unplug_io events.
no dev_t info for getrq and sleeprq events if bio == NULL.
no dev_t info for rq_abort,...,rq_requeue events if rq->rq_disk == NULL.
This is mainly because we can't get the deivce from a request queue.
But this may change in the future.
- A packet command is converted to a string in TP_assign, not TP_print.
While blktrace do the convertion just before output.
Since pc requests should be rather rare, this is not a big issue.
- In blktrace, an event can have 2 different print formats, but a TRACE_EVENT
has a unique format, which means we have some unused data in a trace entry.
The overhead is minimized by using __dynamic_array() instead of __array().
I've benchmarked the ioctl blktrace vs the splice based TRACE_EVENT tracing:
dd dd + ioctl blktrace dd + TRACE_EVENT (splice)
1 7.36s, 42.7 MB/s 7.50s, 42.0 MB/s 7.41s, 42.5 MB/s
2 7.43s, 42.3 MB/s 7.48s, 42.1 MB/s 7.43s, 42.4 MB/s
3 7.38s, 42.6 MB/s 7.45s, 42.2 MB/s 7.41s, 42.5 MB/s
So the overhead of tracing is very small, and no regression when using
those trace events vs blktrace.
And the binary output of TRACE_EVENT is much smaller than blktrace:
# ls -l -h
-rw-r--r-- 1 root root 8.8M 06-09 13:24 sda.blktrace.0
-rw-r--r-- 1 root root 195K 06-09 13:24 sda.blktrace.1
-rw-r--r-- 1 root root 2.7M 06-09 13:25 trace_splice.out
Following are some comparisons between TRACE_EVENT and blktrace:
plug:
kjournald-480 [000] 303.084981: block_plug: [kjournald]
kjournald-480 [000] 303.084981: 8,0 P N [kjournald]
unplug_io:
kblockd/0-118 [000] 300.052973: block_unplug_io: [kblockd/0] 1
kblockd/0-118 [000] 300.052974: 8,0 U N [kblockd/0] 1
remap:
kjournald-480 [000] 303.085042: block_remap: 8,0 W 102736992 + 8 <- (8,8) 33384
kjournald-480 [000] 303.085043: 8,0 A W 102736992 + 8 <- (8,8) 33384
bio_backmerge:
kjournald-480 [000] 303.085086: block_bio_backmerge: 8,0 W 102737032 + 8 [kjournald]
kjournald-480 [000] 303.085086: 8,0 M W 102737032 + 8 [kjournald]
getrq:
kjournald-480 [000] 303.084974: block_getrq: 8,0 W 102736984 + 8 [kjournald]
kjournald-480 [000] 303.084975: 8,0 G W 102736984 + 8 [kjournald]
bash-2066 [001] 1072.953770: 8,0 G N [bash]
bash-2066 [001] 1072.953773: block_getrq: 0,0 N 0 + 0 [bash]
rq_complete:
konsole-2065 [001] 300.053184: block_rq_complete: 8,0 W () 103669040 + 16 [0]
konsole-2065 [001] 300.053191: 8,0 C W 103669040 + 16 [0]
ksoftirqd/1-7 [001] 1072.953811: 8,0 C N (5a 00 08 00 00 00 00 00 24 00) [0]
ksoftirqd/1-7 [001] 1072.953813: block_rq_complete: 0,0 N (5a 00 08 00 00 00 00 00 24 00) 0 + 0 [0]
rq_insert:
kjournald-480 [000] 303.084985: block_rq_insert: 8,0 W 0 () 102736984 + 8 [kjournald]
kjournald-480 [000] 303.084986: 8,0 I W 102736984 + 8 [kjournald]
Changelog from v2 -> v3:
- use the newly introduced __dynamic_array().
Changelog from v1 -> v2:
- use __string() instead of __array() to minimize the memory required
to store hex dump of rq->cmd().
- support large pc requests.
- add missing blk_fill_rwbs_rq() in block_rq_requeue TRACE_EVENT.
- some cleanups.
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
LKML-Reference: <4A2DF669.5070905@cn.fujitsu.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
Merge reason: tracing/core was on a .30-rc1 base and was missing out on
on a handful of tracing fixes present in .30-rc5-almost.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
It's used by DM and MD and generally useful, so move the bio list
helpers into bio.h.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Alasdair G Kergon <agk@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
Barriers are submitted to a worker thread that issues them in-order.
The thread is modified so that when it sees a barrier request it waits
for all pending IO before the request then submits the barrier and
waits for it. (We must wait, otherwise it could be intermixed with
following requests.)
Errors from the barrier request are recorded in a per-device barrier_error
variable. There may be only one barrier request in progress at once.
For now, the barrier request is converted to a non-barrier request when
sending it to the underlying device.
This patch guarantees correct barrier behavior if the underlying device
doesn't perform write-back caching. The same requirement existed before
barriers were supported in dm.
Bottom layer barrier support (sending barriers by target drivers) and
handling devices with write-back caches will be done in further patches.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Remove queue_io return value and a loop in dm_request.
IO may be submitted to a worker thread with queue_io(). queue_io() sets
DMF_QUEUE_IO_TO_THREAD so that all further IO is queued for the thread. When
the thread finishes its work, it clears DMF_QUEUE_IO_TO_THREAD and from this
point on, requests are submitted from dm_request again. This will be used
for processing barriers.
Remove the loop in dm_request. queue_io() can submit I/Os to the worker thread
even if DMF_QUEUE_IO_TO_THREAD was not set.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Rework shutting down on suspend and document the associated rules.
Drop write lock in __split_and_process_bio to allow more processing
concurrency.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Refactor the code in dm_request().
Require the new DMF_BLOCK_FOR_SUSPEND flag on readahead bios we will
discard so we don't drop such bios while processing a barrier.
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Split the DMF_BLOCK_IO flag into two.
DMF_BLOCK_IO_FOR_SUSPEND is set when I/O must be blocked while suspending a
device. DMF_QUEUE_IO_TO_THREAD is set when I/O must be queued to a
worker thread for later processing.
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Prepare for full barrier implementation: first remove the restricted support.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch provides support for data integrity passthrough in the device
mapper.
- If one or more component devices support integrity an integrity
profile is preallocated for the DM device.
- If all component devices have compatible profiles the DM device is
flagged as capable.
- Handle integrity metadata when splitting and cloning bios.
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Set queue ordered mode. It doesn't really matter what we set here
because we don't ever put any requests on the queue. But we need to set
something other than QUEUE_ORDERED_NONE so that __generic_make_request
passes barrier requests to us.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Mikulas Patocka
Mike Snitzer
Alasdair G Kergon
Shaohua Li
Martin K. Petersen
Jens Axboe
Kiyoshi Ueda
Tejun Heo
Milan Broz
Jeff Moyer
Linus Torvalds
Arnd Bergmann
FUJITA Tomonori
Christoph Hellwig
Peter Rajnoha
Jun'ichi Nomura
Mike Anderson
Zdenek Kabelac
Nikanth Karthikesan
Jens Axboe
Alexey Dobriyan
Li Zefan
Ingo Molnar
Alan D. Brunelle