We've added a considerable amount of fixes for stalls and issues
with the blk-mq scheduling in the 4.11 series since forking
off the for-4.12/block branch. We need to do improvements on
top of that for 4.12, so pull in the previous fixes to make
our lives easier going forward.
Signed-off-by: Jens Axboe <axboe@fb.com>
To improve scalability, if hardware queues are shared, restart
a single hardware queue in round-robin fashion. Rename
blk_mq_sched_restart_queues() to reflect the new semantics.
Remove blk_mq_sched_mark_restart_queue() because this function
has no callers. Remove flag QUEUE_FLAG_RESTART because this
patch removes the code that uses this flag.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Introduce a function that runs a hardware queue unconditionally
after a delay. Note: there is already a function that stops and
restarts a hardware queue after a delay, namely blk_mq_delay_queue().
This function will be used in the next patch in this series.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.de>
Cc: Long Li <longli@microsoft.com>
Cc: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently only dm and md/raid5 bios trigger
trace_block_bio_complete(). Now that we have bio_chain() and
bio_inc_remaining(), it is not possible, in general, for a driver to
know when the bio is really complete. Only bio_endio() knows that.
So move the trace_block_bio_complete() call to bio_endio().
Now trace_block_bio_complete() pairs with trace_block_bio_queue().
Any bio for which a 'queue' event is traced, will subsequently
generate a 'complete' event.
There are a few cases where completion tracing is not wanted.
1/ If blk_update_request() has already generated a completion
trace event at the 'request' level, there is no point generating
one at the bio level too. In this case the bi_sector and bi_size
will have changed, so the bio level event would be wrong
2/ If the bio hasn't actually been queued yet, but is being aborted
early, then a trace event could be confusing. Some filesystems
call bio_endio() but do not want tracing.
3/ The bio_integrity code interposes itself by replacing bi_end_io,
then restoring it and calling bio_endio() again. This would produce
two identical trace events if left like that.
To handle these, we introduce a flag BIO_TRACE_COMPLETION and only
produce the trace event when this is set.
We address point 1 above by clearing the flag in blk_update_request().
We address point 2 above by only setting the flag when
generic_make_request() is called.
We address point 3 above by clearing the flag after generating a
completion event.
When bio_split() is used on a bio, particularly in blk_queue_split(),
there is an extra complication. A new bio is split off the front, and
may be handle directly without going through generic_make_request().
The old bio, which has been advanced, is passed to
generic_make_request(), so it will trigger a trace event a second
time.
Probably the best result when a split happens is to see a single
'queue' event for the whole bio, then multiple 'complete' events - one
for each component. To achieve this was can:
- copy the BIO_TRACE_COMPLETION flag to the new bio in bio_split()
- avoid generating a 'queue' event if BIO_TRACE_COMPLETION is already set.
This way, the split-off bio won't create a queue event, the original
won't either even if it re-submitted to generic_make_request(),
but both will produce completion events, each for their own range.
So if generic_make_request() is called (which generates a QUEUED
event), then bi_endio() will create a single COMPLETE event for each
range that the bio is split into, unless the driver has explicitly
requested it not to.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq_update_nr_hw_queues() used to remap hardware queues, which is the
behavior that drivers expect. However, commit 4e68a01142 changed
blk_mq_queue_reinit() to not remap queues for the case of CPU
hotplugging, inadvertently making blk_mq_update_nr_hw_queues() not remap
queues as well. This breaks, for example, NBD's multi-connection mode,
leaving the added hardware queues unused. Fix it by making
blk_mq_update_nr_hw_queues() explicitly remap the queues.
Fixes: 4e68a01142 ("blk-mq: don't redistribute hardware queues on a CPU hotplug event")
Reviewed-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
In elevator_switch(), if blk_mq_init_sched() fails, we attempt to fall
back to the original scheduler. However, at this point, we've already
torn down the original scheduler's tags, so this causes a crash. Doing
the fallback like the legacy elevator path is much harder for mq, so fix
it by just falling back to none, instead.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
If a new hardware queue is added at runtime, we don't allocate scheduler
tags for it, leading to a crash. This hooks up the scheduler framework
to blk_mq_{init,exit}_hctx() to make sure everything gets properly
initialized/freed.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Preparation cleanup for the next couple of fixes, push
blk_mq_sched_setup() and e->ops.mq.init_sched() into a helper.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
While dispatching requests, if we fail to get a driver tag, we mark the
hardware queue as waiting for a tag and put the requests on a
hctx->dispatch list to be run later when a driver tag is freed. However,
blk_mq_dispatch_rq_list() may dispatch requests from multiple hardware
queues if using a single-queue scheduler with a multiqueue device. If
blk_mq_get_driver_tag() fails, it doesn't update the hardware queue we
are processing. This means we end up using the hardware queue of the
previous request, which may or may not be the same as that of the
current request. If it isn't, the wrong hardware queue will end up
waiting for a tag, and the requests will be on the wrong dispatch list,
leading to a hang.
The fix is twofold:
1. Make sure we save which hardware queue we were trying to get a
request for in blk_mq_get_driver_tag() regardless of whether it
succeeds or not.
2. Make blk_mq_dispatch_rq_list() take a request_queue instead of a
blk_mq_hw_queue to make it clear that it must handle multiple
hardware queues, since I've already messed this up on a couple of
occasions.
This didn't appear in testing with nvme and mq-deadline because nvme has
more driver tags than the default number of scheduler tags. However,
with the blk_mq_update_nr_hw_queues() fix, it showed up with nbd.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Instead of bloating the generic struct request with it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
The block layer core sets blk_mq_queue_data.list but no block
drivers read that member. Hence remove it and also the code that
is used to set this member.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
Writeback throttling does not play well with CFQ since that also tries
to throttle async writes. As a result async writeback can get starved in
presence of readers. As an example take a benchmark simulating
postgreSQL database running over a standard rotating SATA drive. There
are 16 processes doing random reads from a huge file (2*machine memory),
1 process doing random writes to the huge file and calling fsync once
per 50000 writes and 1 process doing sequential 8k writes to a
relatively small file wrapping around at the end of the file and calling
fsync every 5 writes. Under this load read latency easily exceeds the
target latency of 75 ms (just because there are so many reads happening
against a relatively slow disk) and thus writeback is throttled to a
point where only 1 write request is allowed at a time. Blktrace data
then looks like:
8,0 1 0 8.347751764 0 m N cfq workload slice:40000000
8,0 1 0 8.347755256 0 m N cfq293A / set_active wl_class: 0 wl_type:0
8,0 1 0 8.347784100 0 m N cfq293A / Not idling. st->count:1
8,0 1 3814 8.347763916 5839 UT N [kworker/u9:2] 1
8,0 0 0 8.347777605 0 m N cfq293A / Not idling. st->count:1
8,0 1 0 8.347784100 0 m N cfq293A / Not idling. st->count:1
8,0 3 1596 8.354364057 0 C R 156109528 + 8 (6906954) [0]
8,0 3 0 8.354383193 0 m N cfq6196SN / complete rqnoidle 0
8,0 3 0 8.354386476 0 m N cfq schedule dispatch
8,0 3 0 8.354399397 0 m N cfq293A / Not idling. st->count:1
8,0 3 0 8.354404705 0 m N cfq293A / dispatch_insert
8,0 3 0 8.354409454 0 m N cfq293A / dispatched a request
8,0 3 0 8.354412527 0 m N cfq293A / activate rq, drv=1
8,0 3 1597 8.354414692 0 D W 145961400 + 24 (6718452) [swapper/0]
8,0 3 0 8.354484184 0 m N cfq293A / Not idling. st->count:1
8,0 3 0 8.354487536 0 m N cfq293A / slice expired t=0
8,0 3 0 8.354498013 0 m N / served: vt=5888102466265088 min_vt=5888074869387264
8,0 3 0 8.354502692 0 m N cfq293A / sl_used=6737519 disp=1 charge=6737519 iops=0 sect=24
8,0 3 0 8.354505695 0 m N cfq293A / del_from_rr
...
8,0 0 1810 8.354728768 0 C W 145961400 + 24 (314076) [0]
8,0 0 0 8.354746927 0 m N cfq293A / complete rqnoidle 0
...
8,0 1 3829 8.389886102 5839 G W 145962968 + 24 [kworker/u9:2]
8,0 1 3830 8.389888127 5839 P N [kworker/u9:2]
8,0 1 3831 8.389908102 5839 A W 145978336 + 24 <- (8,4) 44000
8,0 1 3832 8.389910477 5839 Q W 145978336 + 24 [kworker/u9:2]
8,0 1 3833 8.389914248 5839 I W 145962968 + 24 (28146) [kworker/u9:2]
8,0 1 0 8.389919137 0 m N cfq293A / insert_request
8,0 1 0 8.389924305 0 m N cfq293A / add_to_rr
8,0 1 3834 8.389933175 5839 UT N [kworker/u9:2] 1
...
8,0 0 0 9.455290997 0 m N cfq workload slice:40000000
8,0 0 0 9.455294769 0 m N cfq293A / set_active wl_class:0 wl_type:0
8,0 0 0 9.455303499 0 m N cfq293A / fifo=ffff880003166090
8,0 0 0 9.455306851 0 m N cfq293A / dispatch_insert
8,0 0 0 9.455311251 0 m N cfq293A / dispatched a request
8,0 0 0 9.455314324 0 m N cfq293A / activate rq, drv=1
8,0 0 2043 9.455316210 6204 D W 145962968 + 24 (1065401962) [pgioperf]
8,0 0 0 9.455392407 0 m N cfq293A / Not idling. st->count:1
8,0 0 0 9.455395969 0 m N cfq293A / slice expired t=0
8,0 0 0 9.455404210 0 m N / served: vt=5888958194597888 min_vt=5888941810597888
8,0 0 0 9.455410077 0 m N cfq293A / sl_used=4000000 disp=1 charge=4000000 iops=0 sect=24
8,0 0 0 9.455416851 0 m N cfq293A / del_from_rr
...
8,0 0 2045 9.455648515 0 C W 145962968 + 24 (332305) [0]
8,0 0 0 9.455668350 0 m N cfq293A / complete rqnoidle 0
...
8,0 1 4371 9.455710115 5839 G W 145978336 + 24 [kworker/u9:2]
8,0 1 4372 9.455712350 5839 P N [kworker/u9:2]
8,0 1 4373 9.455730159 5839 A W 145986616 + 24 <- (8,4) 52280
8,0 1 4374 9.455732674 5839 Q W 145986616 + 24 [kworker/u9:2]
8,0 1 4375 9.455737563 5839 I W 145978336 + 24 (27448) [kworker/u9:2]
8,0 1 0 9.455742871 0 m N cfq293A / insert_request
8,0 1 0 9.455747550 0 m N cfq293A / add_to_rr
8,0 1 4376 9.455756629 5839 UT N [kworker/u9:2] 1
So we can see a Q event for a write request, then IO is blocked by
writeback throttling and G and I events for the request happen only once
other writeback IO is completed. Thus CFQ always sees only one write
request. When it sees it, it queues the async queue behind all the read
queues and the async queue gets scheduled after about one second. When
it is scheduled, that one request gets dispatched and async queue is
expired as it has no more requests to submit. Overall we submit about
one write request per second.
Although this scheduling is beneficial for read latency, writes are
heavily starved and this causes large delays all over the system (due to
processes blocking on page lock, transaction starts, etc.). When
writeback throttling is disabled, write throughput is about one fifth of
a read throughput which roughly matches readers/writers ratio and
overall the system stalls are much shorter.
Mixing writeback throttling logic with CFQ throttling logic is always a
recipe for surprises as CFQ assumes it sees the big part of the picture
which is not necessarily true when writeback throttling is blocking
requests. So disable writeback throttling logic by default when CFQ is
used as an IO scheduler.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
In 4.10 I introduced a patch that associates the ioc priority with
each request in the block layer. This work was done in the single queue
block layer code. This patch unifies ioc priority to request mapping across
the single/multi queue block layers.
I have tested this patch with the null block device driver with the following
parameters.
null_blk queue_mode=2 irqmode=0 use_per_node_hctx=1 nr_devices=1
I have not seen a performance regression with this patch and I would appreciate
any feedback or additional testing.
I have also verified that io priorities are passed to the device when using
the SQ and MQ path to a SATA HDD that supports io priorities.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Adam Manzanares <adam.manzanares@wdc.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Commit a4d907b6a3 unified the single and multi queue request handlers,
but in the process, it also screwed up the locking balance and calls
blk_mq_try_issue_directly() with the ctx preempt lock held. This is a
problem for drivers that have set BLK_MQ_F_BLOCKING, since now they
can't reliably sleep.
While in there, protect against similar issues in the future, by adding
a might_sleep() trigger in the BLOCKING path for direct issue or queue
run.
Reported-by: Josef Bacik <josef@toxicpanda.com>
Tested-by: Josef Bacik <josef@toxicpanda.com>
Fixes: a4d907b6a3 ("blk-mq: streamline blk_mq_make_request")
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
trivial fix to spelling mistake in pr_err error message
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
In blk_mq_alloc_request_hctx, blk_mq_sched_get_request doesn't
get sw context so we don't need to put the context with
blk_mq_put_ctx. Unless, we will see preempt counter underflow.
Cc: Omar Sandoval <osandov@fb.com>
Signed-off-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
In blk_mq_alloc_request_hctx, blk_mq_sched_get_request doesn't
get sw context so we don't need to put the context with
blk_mq_put_ctx. Unless, we will see preempt counter underflow.
Cc: Omar Sandoval <osandov@fb.com>
Signed-off-by: Minchan Kim <minchan@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently we return true in blk_mq_dispatch_rq_list() if we queued IO
successfully, but we really want to return whether or not the we made
progress. Progress includes if we got an error return. If we don't,
this can lead to a hang in blk_mq_sched_dispatch_requests() when a
driver is draining IO by returning BLK_MQ_QUEUE_ERROR instead of
manually ending the IO in error and return BLK_MQ_QUEUE_OK.
Tested-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
When try to issue a request directly and we fail we will requeue the
request, but call blk_mq_end_request() as well. This leads to the
completed request being on a queuelist and getting ended twice, which
causes list corruption in schedulers and other shenanigans.
Signed-off-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Ming Lei <tom.leiming@gmail.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
blkg_conf_prep() currently calls blkg_lookup_create() while holding
request queue spinlock. This means allocating memory for struct
blkcg_gq has to be made non-blocking. This causes occasional -ENOMEM
failures in call paths like below:
pcpu_alloc+0x68f/0x710
__alloc_percpu_gfp+0xd/0x10
__percpu_counter_init+0x55/0xc0
cfq_pd_alloc+0x3b2/0x4e0
blkg_alloc+0x187/0x230
blkg_create+0x489/0x670
blkg_lookup_create+0x9a/0x230
blkg_conf_prep+0x1fb/0x240
__cfqg_set_weight_device.isra.105+0x5c/0x180
cfq_set_weight_on_dfl+0x69/0xc0
cgroup_file_write+0x39/0x1c0
kernfs_fop_write+0x13f/0x1d0
__vfs_write+0x23/0x120
vfs_write+0xc2/0x1f0
SyS_write+0x44/0xb0
entry_SYSCALL_64_fastpath+0x18/0xad
In the code path above, percpu allocator cannot call vmalloc() due to
queue spinlock.
A failure in this call path gives grief to tools which are trying to
configure io weights. We see occasional failures happen shortly after
reboots even when system is not under any memory pressure. Machines
with a lot of cpus are more vulnerable to this condition.
Do struct blkcg_gq allocations outside the queue spinlock to allow
blocking during memory allocations.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Tahsin Erdogan <tahsin@google.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
I inadvertently applied the v5 version of this patch, whereas
the agreed upon version was v5. Revert this one so we can apply
the right one.
This reverts commit 7fc6b87a9f.
CONFIG_DEBUG_TEST_DRIVER_REMOVE found a possible leak of q->rq_wb when a
request queue is reregistered. This has been a problem since wbt was
introduced, but the WARN_ON(!list_empty(&stats->callbacks)) in the
blk-stat rework exposed it. Fix it by cleaning up wbt when we unregister
the queue.
Fixes: 87760e5eef ("block: hook up writeback throttling")
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_alloc_queue_node() already allocates q->stats, so
blk_mq_init_allocated_queue() is overwriting it with a new allocation.
Fixes: a83b576c9c ("block: fix stacked driver stats init and free")
Reviewed-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Now that the remaining drivers have been converted to one request queue
per gendisk, let's warn if a request queue gets registered more than
once. This will catch future drivers which might do it inadvertently or
any old drivers that I may have missed.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Before commit 780db2071a(blk-mq: decouble blk-mq freezing
from generic bypassing), the dying flag is checked before
entering queue, and Tejun converts the checking into .mq_freeze_depth,
and assumes the counter is increased just after dying flag
is set. Unfortunately we doesn't do that in blk_set_queue_dying().
This patch calls blk_freeze_queue_start() in blk_set_queue_dying(),
so that we can block new I/O coming once the queue is set as dying.
Given blk_set_queue_dying() is always called in remove path
of block device, and queue will be cleaned up later, we don't
need to worry about undoing the counter.
Cc: Tejun Heo <tj@kernel.org>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
As the .q_usage_counter is used by both legacy and
mq path, we need to block new I/O if queue becomes
dead in blk_queue_enter().
So rename it and we can use this function in both
paths.
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Without the barrier, reading DEAD flag of .q_usage_counter
and reading .mq_freeze_depth may be reordered, then the
following wait_event_interruptible() may never return.
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This patch adds comment on two races related with
timeout handler:
- requeue from queue busy vs. timeout
- rq free & reallocation vs. timeout
Both the races themselves and current solution aren't
explicit enough, so add comments on them.
Cc: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
blkg_conf_prep() currently calls blkg_lookup_create() while holding
request queue spinlock. This means allocating memory for struct
blkcg_gq has to be made non-blocking. This causes occasional -ENOMEM
failures in call paths like below:
pcpu_alloc+0x68f/0x710
__alloc_percpu_gfp+0xd/0x10
__percpu_counter_init+0x55/0xc0
cfq_pd_alloc+0x3b2/0x4e0
blkg_alloc+0x187/0x230
blkg_create+0x489/0x670
blkg_lookup_create+0x9a/0x230
blkg_conf_prep+0x1fb/0x240
__cfqg_set_weight_device.isra.105+0x5c/0x180
cfq_set_weight_on_dfl+0x69/0xc0
cgroup_file_write+0x39/0x1c0
kernfs_fop_write+0x13f/0x1d0
__vfs_write+0x23/0x120
vfs_write+0xc2/0x1f0
SyS_write+0x44/0xb0
entry_SYSCALL_64_fastpath+0x18/0xad
In the code path above, percpu allocator cannot call vmalloc() due to
queue spinlock.
A failure in this call path gives grief to tools which are trying to
configure io weights. We see occasional failures happen shortly after
reboots even when system is not under any memory pressure. Machines
with a lot of cpus are more vulnerable to this condition.
Update blkg_create() function to temporarily drop the rcu and queue
locks when it is allowed by gfp mask.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Tahsin Erdogan <tahsin@google.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
One hard problem adding .low limit is to detect idle cgroup. If one
cgroup doesn't dispatch enough IO against its low limit, we must have a
mechanism to determine if other cgroups dispatch more IO. We added the
think time detection mechanism before, but it doesn't work for all
workloads. Here we add a latency based approach.
We already have mechanism to calculate latency threshold for each IO
size. For every IO dispatched from a cgorup, we compare its latency
against its threshold and record the info. If most IO latency is below
threshold (in the code I use 75%), the cgroup could be treated idle and
other cgroups can dispatch more IO.
Currently this latency target check is only for SSD as we can't
calcualte the latency target for hard disk. And this is only for cgroup
leaf node so far.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
User configures latency target, but the latency threshold for each
request size isn't fixed. For a SSD, the IO latency highly depends on
request size. To calculate latency threshold, we sample some data, eg,
average latency for request size 4k, 8k, 16k, 32k .. 1M. The latency
threshold of each request size will be the sample latency (I'll call it
base latency) plus latency target. For example, the base latency for
request size 4k is 80us and user configures latency target 60us. The 4k
latency threshold will be 80 + 60 = 140us.
To sample data, we calculate the order base 2 of rounded up IO sectors.
If the IO size is bigger than 1M, it will be accounted as 1M. Since the
calculation does round up, the base latency will be slightly smaller
than actual value. Also if there isn't any IO dispatched for a specific
IO size, we will use the base latency of smaller IO size for this IO
size.
But we shouldn't sample data at any time. The base latency is supposed
to be latency where disk isn't congested, because we use latency
threshold to schedule IOs between cgroups. If disk is congested, the
latency is higher, using it for scheduling is meaningless. Hence we only
do the sampling when block throttling is in the LOW limit, with
assumption disk isn't congested in such state. If the assumption isn't
true, eg, low limit is too high, calculated latency threshold will be
higher.
Hard disk is completely different. Latency depends on spindle seek
instead of request size. Currently this feature is SSD only, we probably
can use a fixed threshold like 4ms for hard disk though.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently there is no way to know the request size when the request is
finished. Next patch will need this info. We could add extra field to
record the size, but blk_issue_stat has enough space to record it, so
this patch just overloads blk_issue_stat. With this, we will have 49bits
to track time, which still is very long time.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Here we introduce per-cgroup latency target. The target determines how a
cgroup can afford latency increasement. We will use the target latency
to calculate a threshold and use it to schedule IO for cgroups. If a
cgroup's bandwidth is below its low limit but its average latency is
below the threshold, other cgroups can safely dispatch more IO even
their bandwidth is higher than their low limits. On the other hand, if
the first cgroup's latency is higher than the threshold, other cgroups
are throttled to their low limits. So the target latency determines how
we efficiently utilize free disk resource without sacifice of worload's
IO latency.
For example, assume 4k IO average latency is 50us when disk isn't
congested. A cgroup sets the target latency to 30us. Then the cgroup can
accept 50+30=80us IO latency. If the cgroupt's average IO latency is
90us and its bandwidth is below low limit, other cgroups are throttled
to their low limit. If the cgroup's average IO latency is 60us, other
cgroups are allowed to dispatch more IO. When other cgroups dispatch
more IO, the first cgroup's IO latency will increase. If it increases to
81us, we then throttle other cgroups.
User will configure the interface in this way:
echo "8:16 rbps=2097152 wbps=max latency=100 idle=200" > io.low
latency is in microsecond unit
By default, latency target is 0, which means to guarantee IO latency.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Last patch introduces a way to detect idle cgroup. We use it to make
upgrade/downgrade decision. And the new algorithm can detect completely
idle cgroup too, so we can delete the corresponding code.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Add interface to configure the threshold. The io.low interface will
like:
echo "8:16 rbps=2097152 wbps=max idle=2000" > io.low
idle is in microsecond unit.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
A cgroup gets assigned a low limit, but the cgroup could never dispatch
enough IO to cross the low limit. In such case, the queue state machine
will remain in LIMIT_LOW state and all other cgroups will be throttled
according to low limit. This is unfair for other cgroups. We should
treat the cgroup idle and upgrade the state machine to lower state.
We also have a downgrade logic. If the state machine upgrades because of
cgroup idle (real idle), the state machine will downgrade soon as the
cgroup is below its low limit. This isn't what we want. A more
complicated case is cgroup isn't idle when queue is in LIMIT_LOW. But
when queue gets upgraded to lower state, other cgroups could dispatch
more IO and this cgroup can't dispatch enough IO, so the cgroup is below
its low limit and looks like idle (fake idle). In this case, the queue
should downgrade soon. The key to determine if we should do downgrade is
to detect if cgroup is truely idle.
Unfortunately it's very hard to determine if a cgroup is real idle. This
patch uses the 'think time check' idea from CFQ for the purpose. Please
note, the idea doesn't work for all workloads. For example, a workload
with io depth 8 has disk utilization 100%, hence think time is 0, eg,
not idle. But the workload can run higher bandwidth with io depth 16.
Compared to io depth 16, the io depth 8 workload is idle. We use the
idea to roughly determine if a cgroup is idle.
We treat a cgroup idle if its think time is above a threshold (by
default 1ms for SSD and 100ms for HD). The idea is think time above the
threshold will start to harm performance. HD is much slower so a longer
think time is ok.
The patch (and the latter patches) uses 'unsigned long' to track time.
We convert 'ns' to 'us' with 'ns >> 10'. This is fast but loses
precision, should not a big deal.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
When cgroups all reach low limit, cgroups can dispatch more IO. This
could make some cgroups dispatch more IO but others not, and even some
cgroups could dispatch less IO than their low limit. For example, cg1
low limit 10MB/s, cg2 limit 80MB/s, assume disk maximum bandwidth is
120M/s for the workload. Their bps could something like this:
cg1/cg2 bps: T1: 10/80 -> T2: 60/60 -> T3: 10/80
At T1, all cgroups reach low limit, so they can dispatch more IO later.
Then cg1 dispatch more IO and cg2 has no room to dispatch enough IO. At
T2, cg2 only dispatches 60M/s. Since We detect cg2 dispatches less IO
than its low limit 80M/s, we downgrade the queue from LIMIT_MAX to
LIMIT_LOW, then all cgroups are throttled to their low limit (T3). cg2
will have bandwidth below its low limit at most time.
The big problem here is we don't know the maximum bandwidth of the
workload, so we can't make smart decision to avoid the situation. This
patch makes cgroup bandwidth change smooth. After disk upgrades from
LIMIT_LOW to LIMIT_MAX, we don't allow cgroups use all bandwidth upto
their max limit immediately. Their bandwidth limit will be increased
gradually to avoid above situation. So above example will became
something like:
cg1/cg2 bps: 10/80 -> 15/105 -> 20/100 -> 25/95 -> 30/90 -> 35/85 -> 40/80
-> 45/75 -> 22/98
In this way cgroups bandwidth will be above their limit in majority
time, this still doesn't fully utilize disk bandwidth, but that's
something we pay for sharing.
Scale up is linear. The limit scales up 1/2 .low limit every
throtl_slice after upgrade. The scale up will stop if the adjusted limit
hits .max limit. Scale down is exponential. We cut the scale value half
if a cgroup doesn't hit its .low limit. If the scale becomes 0, we then
fully downgrade the queue to LIMIT_LOW state.
Note this doesn't completely avoid cgroup running under its low limit.
The best way to guarantee cgroup doesn't run under its limit is to set
max limit. For example, if we set cg1 max limit to 40, cg2 will never
run under its low limit.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
cgroup could be assigned a limit, but doesn't dispatch enough IO, eg the
cgroup is idle. When this happens, the cgroup doesn't hit its limit, so
we can't move the state machine to higher level and all cgroups will be
throttled to their lower limit, so we waste bandwidth. Detecting idle
cgroup is hard. This patch handles a simple case, a cgroup doesn't
dispatch any IO. We ignore such cgroup's limit, so other cgroups can use
the bandwidth.
Please note this will be replaced with a more sophisticated algorithm
later, but this demonstrates the idea how we handle idle cgroups, so I
leave it here.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The throtl_slice is 100ms by default. This is a long time for SSD, a lot
of IO can run. To make cgroups have smoother throughput, we choose a
small value (20ms) for SSD.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
throtl_slice is important for blk-throttling. It's called slice
internally but it really is a time window blk-throttling samples data.
blk-throttling will make decision based on the samplings. An example is
bandwidth measurement. A cgroup's bandwidth is measured in the time
interval of throtl_slice.
A small throtl_slice meanse cgroups have smoother throughput but burn
more CPUs. It has 100ms default value, which is not appropriate for all
disks. A fast SSD can dispatch a lot of IOs in 100ms. This patch makes
it tunable.
Since throtl_slice isn't a time slice, the sysfs name
'throttle_sample_time' reflects its character better.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
cgroup could be throttled to a limit but when all cgroups cross high
limit, queue enters a higher state and so the group should be throttled
to a higher limit. It's possible the cgroup is sleeping because of
throttle and other cgroups don't dispatch IO any more. In this case,
nobody can trigger current downgrade/upgrade logic. To fix this issue,
we could either set up a timer to wakeup the cgroup if other cgroups are
idle or make sure this cgroup doesn't sleep too long. Setting up a timer
means we must change the timer very frequently. This patch chooses the
latter. Making cgroup sleep time not too big wouldn't change cgroup
bps/iops, but could make it wakeup more frequently, which isn't a big
issue because throtl_slice * 8 is already quite big.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
When queue state machine is in LIMIT_MAX state, but a cgroup is below
its low limit for some time, the queue should be downgraded to lower
state as one cgroup's low limit isn't met.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
When queue is in LIMIT_LOW state and all cgroups with low limit cross
the bps/iops limitation, we will upgrade queue's state to
LIMIT_MAX. To determine if a cgroup exceeds its limitation, we check if
the cgroup has pending request. Since cgroup is throttled according to
the limit, pending request means the cgroup reaches the limit.
If a cgroup has limit set for both read and write, we consider the
combination of them for upgrade. The reason is read IO and write IO can
interfere with each other. If we do the upgrade based in one direction
IO, the other direction IO could be severly harmed.
For a cgroup hierarchy, there are two cases. Children has lower low
limit than parent. Parent's low limit is meaningless. If children's
bps/iops cross low limit, we can upgrade queue state. The other case is
children has higher low limit than parent. Children's low limit is
meaningless. As long as parent's bps/iops (which is a sum of childrens
bps/iops) cross low limit, we can upgrade queue state.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
each queue will have a state machine. Initially queue is in LIMIT_LOW
state, which means all cgroups will be throttled according to their low
limit. After all cgroups with low limit cross the limit, the queue state
gets upgraded to LIMIT_MAX state.
For max limit, cgroup will use the limit configured by user.
For low limit, cgroup will use the minimal value between low limit and
max limit configured by user. If the minimal value is 0, which means the
cgroup doesn't configure low limit, we will use max limit to throttle
the cgroup and the cgroup is ready to upgrade to LIMIT_MAX
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Add low limit for cgroup and corresponding cgroup interface. To be
consistent with memcg, we allow users configure .low limit higher than
.max limit. But the internal logic always assumes .low limit is lower
than .max limit. So we add extra bps/iops_conf fields in throtl_grp for
userspace configuration. Old bps/iops fields in throtl_grp will be the
actual limit we use for throttling.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
As discussed in LSF, add configure option for the interface and mark it
as experimental, so people can try/test.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
We are going to support low/max limit, each cgroup will have 2 limits
after that. This patch prepares for the multiple limits change.
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Jens Axboe