This patch enables the sysfs to control I/O request merge
functionality in the plug list. While this control has been
implemented for the request queue, it was dismissed in the plug list.
Therefore, block layer merges requests together (or attempt to merge)
even if the merge capability was disable using sysfs nomerge parameter
value 2.
This limitation is directly affects functionality of io_submit()
system call. The system call enables user to submit a bunch of IO
requests from user space using struct iocb **ios input argument.
However, the unconditioned merging functionality in the plug list
potentially merges these requests together down the road. Therefore,
there is no way to distinguish between an application sending bunch of
sequential IOs and an application sending one big IO. Ultimately, all
requests generated by the former app merge within the plug list
together and looks similar to the second app.
While the merging functionality is a desirable feature to improve the
performance of IO subsystem for some applications, it is not useful
for other application like ours at all.
Signed-off-by: Alireza Haghdoost <alireza@cs.umn.edu>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Coding style modified.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The soft lockup below happens at the boot time of the system using dm
multipath and the udev rules to switch scheduler.
[ 356.127001] BUG: soft lockup - CPU#3 stuck for 22s! [sh:483]
[ 356.127001] RIP: 0010:[<ffffffff81072a7d>] [<ffffffff81072a7d>] lock_timer_base.isra.35+0x1d/0x50
...
[ 356.127001] Call Trace:
[ 356.127001] [<ffffffff81073810>] try_to_del_timer_sync+0x20/0x70
[ 356.127001] [<ffffffff8118b08a>] ? kmem_cache_alloc_node_trace+0x20a/0x230
[ 356.127001] [<ffffffff810738b2>] del_timer_sync+0x52/0x60
[ 356.127001] [<ffffffff812ece22>] cfq_exit_queue+0x32/0xf0
[ 356.127001] [<ffffffff812c98df>] elevator_exit+0x2f/0x50
[ 356.127001] [<ffffffff812c9f21>] elevator_change+0xf1/0x1c0
[ 356.127001] [<ffffffff812caa50>] elv_iosched_store+0x20/0x50
[ 356.127001] [<ffffffff812d1d09>] queue_attr_store+0x59/0xb0
[ 356.127001] [<ffffffff812143f6>] sysfs_write_file+0xc6/0x140
[ 356.127001] [<ffffffff811a326d>] vfs_write+0xbd/0x1e0
[ 356.127001] [<ffffffff811a3ca9>] SyS_write+0x49/0xa0
[ 356.127001] [<ffffffff8164e899>] system_call_fastpath+0x16/0x1b
This is caused by a race between md device initialization by multipathd and
shell script to switch the scheduler using sysfs.
- multipathd:
SyS_ioctl -> do_vfs_ioctl -> dm_ctl_ioctl -> ctl_ioctl -> table_load
-> dm_setup_md_queue -> blk_init_allocated_queue -> elevator_init
q->elevator = elevator_alloc(q, e); // not yet initialized
- sh -c 'echo deadline > /sys/$DEVPATH/queue/scheduler':
elevator_switch (in the call trace above)
struct elevator_queue *old = q->elevator;
q->elevator = elevator_alloc(q, new_e);
elevator_exit(old); // lockup! (*)
- multipathd: (cont.)
err = e->ops.elevator_init_fn(q); // init fails; q->elevator is modified
(*) When del_timer_sync() is called, lock_timer_base() will loop infinitely
while timer->base == NULL. In this case, as timer will never initialized,
it results in lockup.
This patch introduces acquisition of q->sysfs_lock around elevator_init()
into blk_init_allocated_queue(), to provide mutual exclusion between
initialization of the q->scheduler and switching of the scheduler.
This should fix this bugzilla:
https://bugzilla.redhat.com/show_bug.cgi?id=902012
Signed-off-by: Tomoki Sekiyama <tomoki.sekiyama@hds.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
crocode i2c_i801 i2c_core iTCO_wdt iTCO_vendor_support shpchp ioatdma dca be2net sg ses enclosure ext4 mbcache jbd2 sd_mod crc_t10dif ahci megaraid_sas(U) dm_mirror dm_region_hash dm_log dm_mod [last unloaded: scsi_wait_scan]
Pid: 491, comm: scsi_eh_0 Tainted: G W ---------------- 2.6.32-220.13.1.el6.x86_64 #1 IBM -[8722PAX]-/00D1461
RIP: 0010:[<ffffffff8124e424>] [<ffffffff8124e424>] blk_requeue_request+0x94/0xa0
RSP: 0018:ffff881057eefd60 EFLAGS: 00010012
RAX: ffff881d99e3e8a8 RBX: ffff881d99e3e780 RCX: ffff881d99e3e8a8
RDX: ffff881d99e3e8a8 RSI: ffff881d99e3e780 RDI: ffff881d99e3e780
RBP: ffff881057eefd80 R08: ffff881057eefe90 R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000000 R12: ffff881057f92338
R13: 0000000000000000 R14: ffff881057f92338 R15: ffff883058188000
FS: 0000000000000000(0000) GS:ffff880040200000(0000) knlGS:0000000000000000
CS: 0010 DS: 0018 ES: 0018 CR0: 000000008005003b
CR2: 00000000006d3ec0 CR3: 000000302cd7d000 CR4: 00000000000406b0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Process scsi_eh_0 (pid: 491, threadinfo ffff881057eee000, task ffff881057e29540)
Stack:
0000000000001057 0000000000000286 ffff8810275efdc0 ffff881057f16000
<0> ffff881057eefdd0 ffffffff81362323 ffff881057eefe20 ffffffff8135f393
<0> ffff881057e29af8 ffff8810275efdc0 ffff881057eefe78 ffff881057eefe90
Call Trace:
[<ffffffff81362323>] __scsi_queue_insert+0xa3/0x150
[<ffffffff8135f393>] ? scsi_eh_ready_devs+0x5e3/0x850
[<ffffffff81362a23>] scsi_queue_insert+0x13/0x20
[<ffffffff8135e4d4>] scsi_eh_flush_done_q+0x104/0x160
[<ffffffff8135fb6b>] scsi_error_handler+0x35b/0x660
[<ffffffff8135f810>] ? scsi_error_handler+0x0/0x660
[<ffffffff810908c6>] kthread+0x96/0xa0
[<ffffffff8100c14a>] child_rip+0xa/0x20
[<ffffffff81090830>] ? kthread+0x0/0xa0
[<ffffffff8100c140>] ? child_rip+0x0/0x20
Code: 00 00 eb d1 4c 8b 2d 3c 8f 97 00 4d 85 ed 74 bf 49 8b 45 00 49 83 c5 08 48 89 de 4c 89 e7 ff d0 49 8b 45 00 48 85 c0 75 eb eb a4 <0f> 0b eb fe 0f 1f 84 00 00 00 00 00 55 48 89 e5 0f 1f 44 00 00
RIP [<ffffffff8124e424>] blk_requeue_request+0x94/0xa0
RSP <ffff881057eefd60>
The RIP is this line:
BUG_ON(blk_queued_rq(rq));
After digging through the code, I think there may be a race between the
request completion and the timer handler running.
A timer is started for each request put on the device's queue (see
blk_start_request->blk_add_timer). If the request does not complete
before the timer expires, the timer handler (blk_rq_timed_out_timer)
will mark the request complete atomically:
static inline int blk_mark_rq_complete(struct request *rq)
{
return test_and_set_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
}
and then call blk_rq_timed_out. The latter function will call
scsi_times_out, which will return one of BLK_EH_HANDLED,
BLK_EH_RESET_TIMER or BLK_EH_NOT_HANDLED. If BLK_EH_RESET_TIMER is
returned, blk_clear_rq_complete is called, and blk_add_timer is again
called to simply wait longer for the request to complete.
Now, if the request happens to complete while this is going on, what
happens? Given that we know the completion handler will bail if it
finds the REQ_ATOM_COMPLETE bit set, we need to focus on the completion
handler running after that bit is cleared. So, from the above
paragraph, after the call to blk_clear_rq_complete. If the completion
sets REQ_ATOM_COMPLETE before the BUG_ON in blk_add_timer, we go boom
there (I haven't seen this in the cores). Next, if we get the
completion before the call to list_add_tail, then the timer will
eventually fire for an old req, which may either be freed or reallocated
(there is evidence that this might be the case). Finally, if the
completion comes in *after* the addition to the timeout list, I think
it's harmless. The request will be removed from the timeout list,
req_atom_complete will be set, and all will be well.
This will only actually explain the coredumps *IF* the request
structure was freed, reallocated *and* queued before the error handler
thread had a chance to process it. That is possible, but it may make
sense to keep digging for another race. I think that if this is what
was happening, we would see other instances of this problem showing up
as null pointer or garbage pointer dereferences, for example when the
request structure was not re-used. It looks like we actually do run
into that situation in other reports.
This patch moves the BUG_ON(test_bit(REQ_ATOM_COMPLETE,
&req->atomic_flags)); from blk_add_timer to the only caller that could
trip over it (blk_start_request). It then inverts the calls to
blk_clear_rq_complete and blk_add_timer in blk_rq_timed_out to address
the race. I've boot tested this patch, but nothing more.
Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Acked-by: Hannes Reinecke <hare@suse.de>
Cc: stable@kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We switched to plug mq_list for mq, but some code are still using old list.
Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The flush state machine takes in a struct request, which then is
submitted multiple times to the underling driver. The old block code
requeses the same request for each of those, so it does not have an
issue with tapping into the request pool. The new one on the other hand
allocates a new request for each of the actualy steps of the flush
sequence. If have already allocated all of the tags for IO, we will
fail allocating the flush request.
Set aside a reserved request just for flushes.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Linux currently has two models for block devices:
- The classic request_fn based approach, where drivers use struct
request units for IO. The block layer provides various helper
functionalities to let drivers share code, things like tag
management, timeout handling, queueing, etc.
- The "stacked" approach, where a driver squeezes in between the
block layer and IO submitter. Since this bypasses the IO stack,
driver generally have to manage everything themselves.
With drivers being written for new high IOPS devices, the classic
request_fn based driver doesn't work well enough. The design dates
back to when both SMP and high IOPS was rare. It has problems with
scaling to bigger machines, and runs into scaling issues even on
smaller machines when you have IOPS in the hundreds of thousands
per device.
The stacked approach is then most often selected as the model
for the driver. But this means that everybody has to re-invent
everything, and along with that we get all the problems again
that the shared approach solved.
This commit introduces blk-mq, block multi queue support. The
design is centered around per-cpu queues for queueing IO, which
then funnel down into x number of hardware submission queues.
We might have a 1:1 mapping between the two, or it might be
an N:M mapping. That all depends on what the hardware supports.
blk-mq provides various helper functions, which include:
- Scalable support for request tagging. Most devices need to
be able to uniquely identify a request both in the driver and
to the hardware. The tagging uses per-cpu caches for freed
tags, to enable cache hot reuse.
- Timeout handling without tracking request on a per-device
basis. Basically the driver should be able to get a notification,
if a request happens to fail.
- Optional support for non 1:1 mappings between issue and
submission queues. blk-mq can redirect IO completions to the
desired location.
- Support for per-request payloads. Drivers almost always need
to associate a request structure with some driver private
command structure. Drivers can tell blk-mq this at init time,
and then any request handed to the driver will have the
required size of memory associated with it.
- Support for merging of IO, and plugging. The stacked model
gets neither of these. Even for high IOPS devices, merging
sequential IO reduces per-command overhead and thus
increases bandwidth.
For now, this is provided as a potential 3rd queueing model, with
the hope being that, as it matures, it can replace both the classic
and stacked model. That would get us back to having just 1 real
model for block devices, leaving the stacked approach to dm/md
devices (as it was originally intended).
Contributions in this patch from the following people:
Shaohua Li <shli@fusionio.com>
Alexander Gordeev <agordeev@redhat.com>
Christoph Hellwig <hch@infradead.org>
Mike Christie <michaelc@cs.wisc.edu>
Matias Bjorling <m@bjorling.me>
Jeff Moyer <jmoyer@redhat.com>
Acked-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This reference count has been around since before git history, but the only
place where it's used is in blk_execute_rq, and ther it is entirely useless
as it is incremented before submitting the request and decremented in the
end_io handler before waking up the submitter thread.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We have officially run out of flags in a 32-bit space. Extend it
to 64-bit even on 32-bit archs.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull block IO fixes from Jens Axboe:
"After merge window, no new stuff this time only a collection of neatly
confined and simple fixes"
* 'for-3.12/core' of git://git.kernel.dk/linux-block:
cfq: explicitly use 64bit divide operation for 64bit arguments
block: Add nr_bios to block_rq_remap tracepoint
If the queue is dying then we only call the rq->end_io callout. This leaves bios setup on the request, because the caller assumes when the blk_execute_rq_nowait/blk_execute_rq call has completed that the rq->bios have been cleaned up.
bio-integrity: Fix use of bs->bio_integrity_pool after free
blkcg: relocate root_blkg setting and clearing
block: Convert kmalloc_node(...GFP_ZERO...) to kzalloc_node(...)
block: trace all devices plug operation
In func blk_queue_bio, if list of plug is empty,it will call
blk_trace_plug.
If process deal with a single device,it't ok.But if process deal with
multi devices,it only trace the first device.
Using request_count to judge, it can soleve this problem.
In addition, i modify the comment.
Signed-off-by: Jianpeng Ma <majianpeng@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When a medium error is detected the SCSI stack should return
ENODATA to the upper layers.
[jejb: fix whitespace error]
Signed-off-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
When the thin provisioning hard threshold is reached we
should return ENOSPC to inform upper layers about this fact.
Signed-off-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
Pull s390 updates from Martin Schwidefsky:
"This is the bulk of the s390 patches for the 3.11 merge window.
Notable enhancements are: the block timeout patches for dasd from
Hannes, and more work on the PCI support front. In addition some
cleanup and the usual bug fixing."
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/s390/linux: (42 commits)
s390/dasd: Fail all requests when DASD_FLAG_ABORTIO is set
s390/dasd: Add 'timeout' attribute
block: check for timeout function in blk_rq_timed_out()
block/dasd: detailed I/O errors
s390/dasd: Reduce amount of messages for specific errors
s390/dasd: Implement block timeout handling
s390/dasd: process all requests in the device tasklet
s390/dasd: make number of retries configurable
s390/dasd: Clarify comment
s390/hwsampler: Updated misleading member names in hws_data_entry
s390/appldata_net_sum: do not use static data
s390/appldata_mem: do not use static data
s390/vmwatchdog: do not use static data
s390/airq: simplify adapter interrupt code
s390/pci: remove per device debug attribute
s390/dma: remove gratuitous brackets
s390/facility: decompose test_facility()
s390/sclp: remove duplicated include from sclp_ctl.c
s390/irq: store interrupt information in pt_regs
s390/drivers: Cocci spatch "ptr_ret.spatch"
...
Pull workqueue changes from Tejun Heo:
"Surprisingly, Lai and I didn't break too many things implementing
custom pools and stuff last time around and there aren't any follow-up
changes necessary at this point.
The only change in this pull request is Viresh's patches to make some
per-cpu workqueues to behave as unbound workqueues dependent on a boot
param whose default can be configured via a config option. This leads
to higher processing overhead / lower bandwidth as more work items are
bounced across CPUs; however, it can lead to noticeable powersave in
certain configurations - ~10% w/ idlish constant workload on a
big.LITTLE configuration according to Viresh.
This is because per-cpu workqueues interfere with how the scheduler
perceives whether or not each CPU is idle by forcing pinned tasks on
them, which makes the scheduler's power-aware scheduling decisions
less effective.
Its effectiveness is likely less pronounced on homogenous
configurations and this type of optimization can probably be made
automatic; however, the changes are pretty minimal and the affected
workqueues are clearly marked, so it's an easy gain for some
configurations for the time being with pretty unintrusive changes."
* 'for-3.11' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq:
fbcon: queue work on power efficient wq
block: queue work on power efficient wq
PHYLIB: queue work on system_power_efficient_wq
workqueue: Add system wide power_efficient workqueues
workqueues: Introduce new flag WQ_POWER_EFFICIENT for power oriented workqueues
The DASD driver is using FASTFAIL as an equivalent to the
transport errors in SCSI. And the 'steal lock' function maps
roughly to a reservation error. So we should be returning the
appropriate error codes when completing a request.
Acked-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Stefan Weinhuber <wein@de.ibm.com>
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
In blk_post_runtime_resume, an autosuspend request will be initiated for
the device. Since we are holding the queue lock, we can't sleep and thus
we should use the async version to initiate an autosuspend, i.e.
pm_request_suspend instead of pm_runtime_suspend, which might sleep.
Signed-off-by: Aaron Lu <aaron.lu@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Block layer uses workqueues for multiple purposes. There is no real dependency
of scheduling these on the cpu which scheduled them.
On a idle system, it is observed that and idle cpu wakes up many times just to
service this work. It would be better if we can schedule it on a cpu which the
scheduler believes to be the most appropriate one.
This patch replaces normal workqueues with power efficient versions.
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull block core updates from Jens Axboe:
- Major bit is Kents prep work for immutable bio vecs.
- Stable candidate fix for a scheduling-while-atomic in the queue
bypass operation.
- Fix for the hang on exceeded rq->datalen 32-bit unsigned when merging
discard bios.
- Tejuns changes to convert the writeback thread pool to the generic
workqueue mechanism.
- Runtime PM framework, SCSI patches exists on top of these in James'
tree.
- A few random fixes.
* 'for-3.10/core' of git://git.kernel.dk/linux-block: (40 commits)
relay: move remove_buf_file inside relay_close_buf
partitions/efi.c: replace useless kzalloc's by kmalloc's
fs/block_dev.c: fix iov_shorten() criteria in blkdev_aio_read()
block: fix max discard sectors limit
blkcg: fix "scheduling while atomic" in blk_queue_bypass_start
Documentation: cfq-iosched: update documentation help for cfq tunables
writeback: expose the bdi_wq workqueue
writeback: replace custom worker pool implementation with unbound workqueue
writeback: remove unused bdi_pending_list
aoe: Fix unitialized var usage
bio-integrity: Add explicit field for owner of bip_buf
block: Add an explicit bio flag for bios that own their bvec
block: Add bio_alloc_pages()
block: Convert some code to bio_for_each_segment_all()
block: Add bio_for_each_segment_all()
bounce: Refactor __blk_queue_bounce to not use bi_io_vec
raid1: use bio_copy_data()
pktcdvd: Use bio_reset() in disabled code to kill bi_idx usage
pktcdvd: use bio_copy_data()
block: Add bio_copy_data()
...
This reverts commit 3a366e614d.
Wanlong Gao reports that it causes a kernel panic on his machine several
minutes after boot. Reverting it removes the panic.
Jens says:
"It's not quite clear why that is yet, so I think we should just revert
the commit for 3.9 final (which I'm assuming is pretty close).
The wifi is crap at the LSF hotel, so sending this email instead of
queueing up a revert and pull request."
Reported-by: Wanlong Gao <gaowanlong@cn.fujitsu.com>
Requested-by: Jens Axboe <axboe@kernel.dk>
Cc: Tejun Heo <tj@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Converts it to use bio_advance(), simplifying it quite a bit in the
process.
Note that req_bio_endio() now always calls bio_advance() - which means
it always loops over the biovec, not just on partial completions. Don't
expect it to affect performance, but worth noting.
Tested it by forcing partial updates, and dumping before and after on
various bio/bvec fields when doing a partial update.
Signed-off-by: Kent Overstreet <koverstreet@google.com>
CC: Jens Axboe <axboe@kernel.dk>
When a request is added:
If device is suspended or is suspending and the request is not a
PM request, resume the device.
When the last request finishes:
Call pm_runtime_mark_last_busy().
When pick a request:
If device is resuming/suspending, then only PM request is allowed
to go.
The idea and API is designed by Alan Stern and described here:
http://marc.info/?l=linux-scsi&m=133727953625963&w=2
Signed-off-by: Lin Ming <ming.m.lin@intel.com>
Signed-off-by: Aaron Lu <aaron.lu@intel.com>
Acked-by: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe
Alireza Haghdoost
Tomoki Sekiyama
Mikulas Patocka
Jeff Moyer
Shaohua Li
Christoph Hellwig
Christoph Hellwig
Linus Torvalds
Jianpeng Ma
Hannes Reinecke
Aaron Lu
Viresh Kumar
Kent Overstreet
Lin Ming