blk_mq freezing is entangled with generic bypassing which bypasses
blkcg and io scheduler and lets IO requests fall through the block
layer to the drivers in FIFO order. This allows forward progress on
IOs with the advanced features disabled so that those features can be
configured or altered without worrying about stalling IO which may
lead to deadlock through memory allocation.
However, generic bypassing doesn't quite fit blk-mq. blk-mq currently
doesn't make use of blkcg or ioscheds and it maps bypssing to
freezing, which blocks request processing and drains all the in-flight
ones. This causes problems as bypassing assumes that request
processing is online. blk-mq works around this by conditionally
allowing request processing for the problem case - during queue
initialization.
Another weirdity is that except for during queue cleanup, bypassing
started on the generic side prevents blk-mq from processing new
requests but doesn't drain the in-flight ones. This shouldn't break
anything but again highlights that something isn't quite right here.
The root cause is conflating blk-mq freezing and generic bypassing
which are two different mechanisms. The only intersecting purpose
that they serve is during queue cleanup. Let's properly separate
blk-mq freezing from generic bypassing and simply use it where
necessary.
* request_queue->mq_freeze_depth is added and
blk_mq_[un]freeze_queue() now operate on this counter instead of
->bypass_depth. The replacement for QUEUE_FLAG_BYPASS isn't added
but the counter is tested directly. This will be further updated by
later changes.
* blk_mq_drain_queue() is dropped and "__" prefix is dropped from
blk_mq_freeze_queue(). Queue cleanup path now calls
blk_mq_freeze_queue() directly.
* blk_queue_enter()'s fast path condition is simplified to simply
check @q->mq_freeze_depth. Previously, the condition was
!blk_queue_dying(q) &&
(!blk_queue_bypass(q) || !blk_queue_init_done(q))
mq_freeze_depth is incremented right after dying is set and
blk_queue_init_done() exception isn't necessary as blk-mq doesn't
start frozen, which only leaves the blk_queue_bypass() test which
can be replaced by @q->mq_freeze_depth test.
This change simplifies the code and reduces confusion in the area.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Nicholas A. Bellinger <nab@linux-iscsi.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently, both blk_queue_bypass_start() and blk_mq_freeze_queue()
skip queue draining if bypass_depth was already above zero. The
assumption is that the one which bumped the bypass_depth should have
performed draining already; however, there's nothing which prevents a
new instance of bypassing/freezing from starting before the previous
one finishes draining. The current code may allow the later
bypassing/freezing instances to complete while there still are
in-flight requests which haven't finished draining.
Fix it by draining regardless of bypass_depth. We still skip draining
from blk_queue_bypass_start() while the queue is initializing to avoid
introducing excessive delays during boot. INIT_DONE setting is moved
above the initial blk_queue_bypass_end() so that bypassing attempts
can't slip inbetween.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Nicholas A. Bellinger <nab@linux-iscsi.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk-mq uses a percpu_counter to keep track of how many usages are in
flight. The percpu_counter is drained while freezing to ensure that
no usage is left in-flight after freezing is complete.
blk_mq_queue_enter/exit() and blk_mq_[un]freeze_queue() implement this
per-cpu gating mechanism; unfortunately, it contains a subtle bug -
smp_wmb() in blk_mq_queue_enter() doesn't prevent prevent the cpu from
fetching @q->bypass_depth before incrementing @q->mq_usage_counter and
if freezing happens inbetween the caller can slip through and freezing
can be complete while there are active users.
Use smp_mb() instead so that bypass_depth and mq_usage_counter
modifications and tests are properly interlocked.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Nicholas A. Bellinger <nab@linux-iscsi.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
Pull block fixes from Jens Axboe:
"A small collection of fixes/changes for the current series. This
contains:
- Removal of dead code from Gu Zheng.
- Revert of two bad fixes that went in earlier in this round, marking
things as __init that were not purely used from init.
- A fix for blk_mq_start_hw_queue() using the __blk_mq_run_hw_queue(),
which could place us wrongly. Make it use the non __ variant,
which handles cases where we are called from the wrong CPU set.
From me.
- A fix for drbd, which allocates discard requests without room for
the SCSI payload. From Lars Ellenberg.
- A fix for user-after-free in the blkcg code from Tejun.
- Addition of limiting gaps in SG lists, if the hardware needs it.
This is the last pre-req patch for blk-mq to enable the full NVMe
conversion. Could wait until 3.17, but it's simple enough so would
be nice to have everything we need for the NVMe port in the 3.17
release. From me"
* 'for-linus' of git://git.kernel.dk/linux-block:
drbd: fix NULL pointer deref in blk_add_request_payload
blk-mq: blk_mq_start_hw_queue() should use blk_mq_run_hw_queue()
block: add support for limiting gaps in SG lists
bio: remove unused macro bip_vec_idx()
Revert "block: add __init to elv_register"
Revert "block: add __init to blkcg_policy_register"
blkcg: fix use-after-free in __blkg_release_rcu() by making blkcg_gq refcnt an atomic_t
floppy: format block0 read error message properly
Currently it calls __blk_mq_run_hw_queue(), which depends on the
CPU placement being correct. This means it's not possible to call
blk_mq_start_hw_queues(q) from a context that is correct for all
queues, leading to triggering the
WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
in __blk_mq_run_hw_queue().
Reported-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Another restriction inherited for NVMe - those devices don't support
SG lists that have "gaps" in them. Gaps refers to cases where the
previous SG entry doesn't end on a page boundary. For NVMe, all SG
entries must start at offset 0 (except the first) and end on a page
boundary (except the last).
Signed-off-by: Jens Axboe <axboe@fb.com>
This reverts commit b5097e956a.
The original commit is buggy, we do use the registration functions
at runtime, for instance when loading IO schedulers through sysfs.
Reported-by: Damien Wyart <damien.wyart@gmail.com>
Hello,
So, this patch should do. Joe, Vivek, can one of you guys please
verify that the oops goes away with this patch?
Jens, the original thread can be read at
http://thread.gmane.org/gmane.linux.kernel/1720729
The fix converts blkg->refcnt from int to atomic_t. It does some
overhead but it should be minute compared to everything else which is
going on and the involved cacheline bouncing, so I think it's highly
unlikely to cause any noticeable difference. Also, the refcnt in
question should be converted to a perpcu_ref for blk-mq anyway, so the
atomic_t is likely to go away pretty soon anyway.
Thanks.
------- 8< -------
__blkg_release_rcu() may be invoked after the associated request_queue
is released with a RCU grace period inbetween. As such, the function
and callbacks invoked from it must not dereference the associated
request_queue. This is clearly indicated in the comment above the
function.
Unfortunately, while trying to fix a different issue, 2a4fd070ee
("blkcg: move bulk of blkcg_gq release operations to the RCU
callback") ignored this and added [un]locking of @blkg->q->queue_lock
to __blkg_release_rcu(). This of course can cause oops as the
request_queue may be long gone by the time this code gets executed.
general protection fault: 0000 [#1] SMP
CPU: 21 PID: 30 Comm: rcuos/21 Not tainted 3.15.0 #1
Hardware name: Stratus ftServer 6400/G7LAZ, BIOS BIOS Version 6.3:57 12/25/2013
task: ffff880854021de0 ti: ffff88085403c000 task.ti: ffff88085403c000
RIP: 0010:[<ffffffff8162e9e5>] [<ffffffff8162e9e5>] _raw_spin_lock_irq+0x15/0x60
RSP: 0018:ffff88085403fdf0 EFLAGS: 00010086
RAX: 0000000000020000 RBX: 0000000000000010 RCX: 0000000000000000
RDX: 000060ef80008248 RSI: 0000000000000286 RDI: 6b6b6b6b6b6b6b6b
RBP: ffff88085403fdf0 R08: 0000000000000286 R09: 0000000000009f39
R10: 0000000000020001 R11: 0000000000020001 R12: ffff88103c17a130
R13: ffff88103c17a080 R14: 0000000000000000 R15: 0000000000000000
FS: 0000000000000000(0000) GS:ffff88107fca0000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00000000006e5ab8 CR3: 000000000193d000 CR4: 00000000000407e0
Stack:
ffff88085403fe18 ffffffff812cbfc2 ffff88103c17a130 0000000000000000
ffff88103c17a130 ffff88085403fec0 ffffffff810d1d28 ffff880854021de0
ffff880854021de0 ffff88107fcaec58 ffff88085403fe80 ffff88107fcaec30
Call Trace:
[<ffffffff812cbfc2>] __blkg_release_rcu+0x72/0x150
[<ffffffff810d1d28>] rcu_nocb_kthread+0x1e8/0x300
[<ffffffff81091d81>] kthread+0xe1/0x100
[<ffffffff8163813c>] ret_from_fork+0x7c/0xb0
Code: ff 47 04 48 8b 7d 08 be 00 02 00 00 e8 55 48 a4 ff 5d c3 0f 1f 00 66 66 66 66 90 55 48 89 e5
+fa 66 66 90 66 66 90 b8 00 00 02 00 <f0> 0f c1 07 89 c2 c1 ea 10 66 39 c2 75 02 5d c3 83 e2 fe 0f
+b7
RIP [<ffffffff8162e9e5>] _raw_spin_lock_irq+0x15/0x60
RSP <ffff88085403fdf0>
The request_queue locking was added because blkcg_gq->refcnt is an int
protected with the queue lock and __blkg_release_rcu() needs to put
the parent. Let's fix it by making blkcg_gq->refcnt an atomic_t and
dropping queue locking in the function.
Given the general heavy weight of the current request_queue and blkcg
operations, this is unlikely to cause any noticeable overhead.
Moreover, blkcg_gq->refcnt is likely to be converted to percpu_ref in
the near future, so whatever (most likely negligible) overhead it may
add is temporary.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Joe Lawrence <joe.lawrence@stratus.com>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Link: http://lkml.kernel.org/g/alpine.DEB.2.02.1406081816540.17948@jlaw-desktop.mno.stratus.com
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@fb.com>
Pull block fixes from Jens Axboe:
"A smaller collection of fixes for the block core that would be nice to
have in -rc2. This pull request contains:
- Fixes for races in the wait/wakeup logic used in blk-mq from
Alexander. No issues have been observed, but it is definitely a
bit flakey currently. Alternatively, we may drop the cyclic
wakeups going forward, but that needs more testing.
- Some cleanups from Christoph.
- Fix for an oops in null_blk if queue_mode=1 and softirq completions
are used. From me.
- A fix for a regression caused by the chunk size setting. It
inadvertently used max_hw_sectors instead of max_sectors, which is
incorrect, and causes hangs on btrfs multi-disk setups (where hw
sectors apparently isn't set). From me.
- Removal of WQ_POWER_EFFICIENT in the kblockd creation. This was a
recent addition as well, but it actually breaks blk-mq which relies
on strict scheduling. If the workqueue power_efficient mode is
turned on, this breaks blk-mq. From Matias.
- null_blk module parameter description fix from Mike"
* 'for-linus' of git://git.kernel.dk/linux-block:
blk-mq: bitmap tag: fix races in bt_get() function
blk-mq: bitmap tag: fix race on blk_mq_bitmap_tags::wake_cnt
blk-mq: bitmap tag: fix races on shared ::wake_index fields
block: blk_max_size_offset() should check ->max_sectors
null_blk: fix softirq completions for queue_mode == 1
blk-mq: merge blk_mq_drain_queue and __blk_mq_drain_queue
blk-mq: properly drain stopped queues
block: remove WQ_POWER_EFFICIENT from kblockd
null_blk: fix name and description of 'queue_mode' module parameter
block: remove elv_abort_queue and blk_abort_flushes
This update fixes few issues in bt_get() function:
- list_empty(&wait.task_list) check is not protected;
- was_empty check is always true which results in *every* thread
entering the loop resets bt_wait_state::wait_cnt counter rather
than every bt->wake_cnt'th thread;
- 'bt_wait_state::wait_cnt' counter update is redundant, since
it also gets reset in bt_clear_tag() function;
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Ming Lei <tom.leiming@gmail.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Alexander Gordeev <agordeev@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This piece of code in bt_clear_tag() function is racy:
bs = bt_wake_ptr(bt);
if (bs && atomic_dec_and_test(&bs->wait_cnt)) {
atomic_set(&bs->wait_cnt, bt->wake_cnt);
wake_up(&bs->wait);
}
Since nothing prevents bt_wake_ptr() from returning the very
same 'bs' address on multiple CPUs, the following scenario is
possible:
CPU1 CPU2
---- ----
0. bs = bt_wake_ptr(bt); bs = bt_wake_ptr(bt);
1. atomic_dec_and_test(&bs->wait_cnt)
2. atomic_dec_and_test(&bs->wait_cnt)
3. atomic_set(&bs->wait_cnt, bt->wake_cnt);
If the decrement in [1] yields zero then for some amount of time
the decrement in [2] results in a negative/overflow value, which
is not expected. The follow-up assignment in [3] overwrites the
invalid value with the batch value (and likely prevents the issue
from being severe) which is still incorrect and should be a lesser.
Cc: Ming Lei <tom.leiming@gmail.com>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Alexander Gordeev <agordeev@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Fix racy updates of shared blk_mq_bitmap_tags::wake_index
and blk_mq_hw_ctx::wake_index fields.
Cc: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Alexander Gordeev <agordeev@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Pull NVMe update from Matthew Wilcox:
"Mostly bugfixes again for the NVMe driver. I'd like to call out the
exported tracepoint in the block layer; I believe Keith has cleared
this with Jens.
We've had a few reports from people who're really pounding on NVMe
devices at scale, hence the timeout changes (and new module
parameters), hotplug cpu deadlock, tracepoints, and minor performance
tweaks"
[ Jens hadn't seen that tracepoint thing, but is ok with it - it will
end up going away when mq conversion happens ]
* git://git.infradead.org/users/willy/linux-nvme: (22 commits)
NVMe: Fix START_STOP_UNIT Scsi->NVMe translation.
NVMe: Use Log Page constants in SCSI emulation
NVMe: Define Log Page constants
NVMe: Fix hot cpu notification dead lock
NVMe: Rename io_timeout to nvme_io_timeout
NVMe: Use last bytes of f/w rev SCSI Inquiry
NVMe: Adhere to request queue block accounting enable/disable
NVMe: Fix nvme get/put queue semantics
NVMe: Delete NVME_GET_FEAT_TEMP_THRESH
NVMe: Make admin timeout a module parameter
NVMe: Make iod bio timeout a parameter
NVMe: Prevent possible NULL pointer dereference
NVMe: Fix the buffer size passed in GetLogPage(CDW10.NUMD)
NVMe: Update data structures for NVMe 1.2
NVMe: Enable BUILD_BUG_ON checks
NVMe: Update namespace and controller identify structures to the 1.1a spec
NVMe: Flush with data support
NVMe: Configure support for block flush
NVMe: Add tracepoints
NVMe: Protect against badly formatted CQEs
...
If we need to drain a queue we need to run all queues, even if they
are marked stopped to make sure the driver has a chance to error out
on all queued requests.
This fixes surprise removal with scsi-mq.
Reported-by: Bart Van Assche <bvanassche@acm.org>
Tested-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk-mq issues async requests through kblockd. To issue a work request on
a specific CPU, kblockd_schedule_delayed_work_on is used. However, the
specific CPU choice may not be honored, if the power_efficient option
for workqueues is set. blk-mq requires that we have strict per-cpu
scheduling, so it wont work properly if kblockd is marked
POWER_EFFICIENT and power_efficient is set.
Remove the kblockd WQ_POWER_EFFICIENT flag to prevent this behavior.
This essentially reverts part of commit 695588f945, which added
the WQ_POWER_EFFICIENT marker to kblockd.
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
elv_abort_queue has no callers, and blk_abort_flushes is only called by
elv_abort_queue.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Pull block layer fixes from Jens Axboe:
"Final small batch of fixes to be included before -rc1. Some general
cleanups in here as well, but some of the blk-mq fixes we need for the
NVMe conversion and/or scsi-mq. The pull request contains:
- Support for not merging across a specified "chunk size", if set by
the driver. Some NVMe devices perform poorly for IO that crosses
such a chunk, so we need to support it generically as part of
request merging avoid having to do complicated split logic. From
me.
- Bump max tag depth to 10Ki tags. Some scsi devices have a huge
shared tag space. Before we failed with EINVAL if a too large tag
depth was specified, now we truncate it and pass back the actual
value. From me.
- Various blk-mq rq init fixes from me and others.
- A fix for enter on a dying queue for blk-mq from Keith. This is
needed to prevent oopsing on hot device removal.
- Fixup for blk-mq timer addition from Ming Lei.
- Small round of performance fixes for mtip32xx from Sam Bradshaw.
- Minor stack leak fix from Rickard Strandqvist.
- Two __init annotations from Fabian Frederick"
* 'for-linus' of git://git.kernel.dk/linux-block:
block: add __init to blkcg_policy_register
block: add __init to elv_register
block: ensure that bio_add_page() always accepts a page for an empty bio
blk-mq: add timer in blk_mq_start_request
blk-mq: always initialize request->start_time
block: blk-exec.c: Cleaning up local variable address returnd
mtip32xx: minor performance enhancements
blk-mq: ->timeout should be cleared in blk_mq_rq_ctx_init()
blk-mq: don't allow queue entering for a dying queue
blk-mq: bump max tag depth to 10K tags
block: add blk_rq_set_block_pc()
block: add notion of a chunk size for request merging
blkcg_policy_register is only called by
__init functions:
__init cfq_init
__init throtl_init
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: Jens Axboe <axboe@fb.com>
elv_register is only called by elevator init functions:
__init cfq_init
__init deadline_init
__init noop_init
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: Jens Axboe <axboe@fb.com>
With commit 762380ad93 added support for chunk sizes and no merging
across them, it broke the rule of always allowing adding of a single
page to an empty bio. So relax the restriction a bit to allow for that,
similarly to what we have always done.
This fixes a crash with mkfs.xfs and 512b sector sizes on NVMe.
Reported-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Pull cgroup updates from Tejun Heo:
"A lot of activities on cgroup side. Heavy restructuring including
locking simplification took place to improve the code base and enable
implementation of the unified hierarchy, which currently exists behind
a __DEVEL__ mount option. The core support is mostly complete but
individual controllers need further work. To explain the design and
rationales of the the unified hierarchy
Documentation/cgroups/unified-hierarchy.txt
is added.
Another notable change is css (cgroup_subsys_state - what each
controller uses to identify and interact with a cgroup) iteration
update. This is part of continuing updates on css object lifetime and
visibility. cgroup started with reference count draining on removal
way back and is now reaching a point where csses behave and are
iterated like normal refcnted objects albeit with some complexities to
allow distinguishing the state where they're being deleted. The css
iteration update isn't taken advantage of yet but is planned to be
used to simplify memcg significantly"
* 'for-3.16' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup: (77 commits)
cgroup: disallow disabled controllers on the default hierarchy
cgroup: don't destroy the default root
cgroup: disallow debug controller on the default hierarchy
cgroup: clean up MAINTAINERS entries
cgroup: implement css_tryget()
device_cgroup: use css_has_online_children() instead of has_children()
cgroup: convert cgroup_has_live_children() into css_has_online_children()
cgroup: use CSS_ONLINE instead of CGRP_DEAD
cgroup: iterate cgroup_subsys_states directly
cgroup: introduce CSS_RELEASED and reduce css iteration fallback window
cgroup: move cgroup->serial_nr into cgroup_subsys_state
cgroup: link all cgroup_subsys_states in their sibling lists
cgroup: move cgroup->sibling and ->children into cgroup_subsys_state
cgroup: remove cgroup->parent
device_cgroup: remove direct access to cgroup->children
memcg: update memcg_has_children() to use css_next_child()
memcg: remove tasks/children test from mem_cgroup_force_empty()
cgroup: remove css_parent()
cgroup: skip refcnting on normal root csses and cgrp_dfl_root self css
cgroup: use cgroup->self.refcnt for cgroup refcnting
...
This way will become consistent with non-mq case, also
avoid to update rq->deadline twice for mq.
The comment said: "We do this early, to ensure we are on
the right CPU.", but no percpu stuff is used in blk_add_timer(),
so it isn't necessary. Even when inserting from plug list, there
is no such guarantee at all.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The blk-mq core only initializes this if io stats are enabled, since
blk-mq only reads the field in that case. But drivers could
potentially use it internally, so ensure that we always set it to
the current time when the request is allocated.
Reported-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Address of local variable assigned to a function parameter
This was partly found using a static code analysis program called cppcheck.
Signed-off-by: Rickard Strandqvist <rickard_strandqvist@spectrumdigital.se>
Signed-off-by: Jens Axboe <axboe@fb.com>
printk is meant to be used with an associated log level. There are some
instances of printk scattered around the mm code where the log level is
missing. Add a log level and adhere to suggestions by
scripts/checkpatch.pl by moving to the pr_* macros.
Also add the typical pr_fmt definition so that print statements can be
easily traced back to the modules where they occur, correlated one with
another, etc. This will require the removal of some (now redundant)
prefixes on a few print statements.
Signed-off-by: Mitchel Humpherys <mitchelh@codeaurora.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
It'll be used in blk_mq_start_request() to set a potential timeout
for the request, so clear it to zero at alloc time to ensure that
we know if someone has set it or not.
Fixes random early timeouts on NVMe testing.
Signed-off-by: Jens Axboe <axboe@fb.com>
If the queue is going away, don't let new allocs or queueing
happen on it. Go through the normal wait process, and exit with
ENODEV in that case.
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
For some scsi-mq cases, the tag map can be huge. So increase the
max number of tags we support.
Additionally, don't fail with EINVAL if a user requests too many
tags. Warn that the tag depth has been adjusted down, and store
the new value inside the tag_set passed in.
Signed-off-by: Jens Axboe <axboe@fb.com>
With the optimizations around not clearing the full request at alloc
time, we are leaving some of the needed init for REQ_TYPE_BLOCK_PC
up to the user allocating the request.
Add a blk_rq_set_block_pc() that sets the command type to
REQ_TYPE_BLOCK_PC, and properly initializes the members associated
with this type of request. Update callers to use this function instead
of manipulating rq->cmd_type directly.
Includes fixes from Christoph Hellwig <hch@lst.de> for my half-assed
attempt.
Signed-off-by: Jens Axboe <axboe@fb.com>
Some drivers have different limits on what size a request should
optimally be, depending on the offset of the request. Similar to
dividing a device into chunks. Add a setting that allows the driver
to inform the block layer of such a chunk size. The block layer will
then prevent merging across the chunks.
This is needed to optimally support NVMe with a non-zero stripe size.
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq_tag_to_rq() needs to be able to tell if it should return
the original request, or the flush request if we are doing a flush
sequence. Clear the flush tag when IO completes for a flush, since
that is what we are comparing against.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
We currently pass in the hardware queue, and get the tags from there.
But from scsi-mq, with a shared tag space, it's a lot more convenient
to pass in the blk_mq_tags instead as the hardware queue isn't always
directly available. So instead of having to re-map to a given
hardware queue from rq->mq_ctx, just pass in the tags structure.
Signed-off-by: Jens Axboe <axboe@fb.com>
When the code was collapsed to avoid duplication, the recent patch
for ensuring that a queue is idled before free was dropped, which was
added by commit 19c5d84f14.
Add back the blk_mq_tag_idle(), to ensure we don't leak a reference
to an active queue when it is freed.
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq_map_request() can return NULL if we fail entering the queue
(dying, or removed), in which case it has already ended IO on the
bio. So nothing more to do, except just return.
Signed-off-by: Jens Axboe <axboe@fb.com>
'struct blk_mq_ctx' is __percpu, so add the annotation
and fix the sparse warning reported from Fengguang:
[block:for-linus 2/3] block/blk-mq.h:75:16: sparse: incorrect
type in initializer (different address spaces)
Reported-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq_put_ctx() has to be called before io_schedule() in
bt_get().
This patch fixes the problem by taking similar approach from
percpu_ida allocation for the situation.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Pull block driver changes from Jens Axboe:
"Now that the core bits are in, here's the pull request for the driver
related changes for 3.16. Nothing out of the ordinary here, mostly
business as usual. There are a few pulls of for-3.16/core into this
branch, which were done when the blk-mq was modified after the
mtip32xx conversion was put in.
The pull request contains:
- skd and cciss converted to use pci_enable_msix_exact(). From
Alexander Gordeev.
- A few mtip32xx fixes from Asai @ Micron.
- The conversion of mtip32xx from make_request_fn to blk-mq, and a
later small fix for that conversion on quiescing for non-queued IO.
From me.
- A fix for bsg to use an exported function to check whether this
driver is request based or not. Needed updating for blk-mq, which
is request based, but does not have a request_fn hook. From me.
- Small floppy bug fix from Jiri.
- A series of cleanups for the cdrom uniform layer from Joe Perches.
Gets rid of various old ugly macros, making the code conform more
to the modern coding style.
- A series of patches for drbd from the drbd crew (Lars Ellenberg and
Philipp Reisner).
- A use-after-free fix for null_blk from Ming Lei.
- Also from Ming Lei is a performance patch for virtio-blk, which can
net us a 3x win on kvm platforms where world notification is
expensive.
- Ming Lei also fixed a stall issue in virtio-blk, due to a race
between queue start/stop and resource limits.
- A small batch of fixes for xen-blk{back,front} from Olaf Hering and
Valentin Priescu"
* 'for-3.16/drivers' of git://git.kernel.dk/linux-block: (54 commits)
block: virtio_blk: don't hold spin lock during world switch
xen-blkback: defer freeing blkif to avoid blocking xenwatch
xen blkif.h: fix comment typo in discard-alignment
xen/blkback: disable discard feature if requested by toolstack
xen-blkfront: remove type check from blkfront_setup_discard
floppy: do not corrupt bio.bi_flags when reading block 0
mtip32xx: move error handling to service thread
virtio_blk: fix race between start and stop queue
mtip32xx: stop block hardware queues before quiescing IO
mtip32xx: blk_mq_init_queue() returns an ERR_PTR
mtip32xx: convert to use blk-mq
cdrom: Remove unnecessary prototype for cdrom_get_disc_info
cdrom: Remove unnecessary prototype for cdrom_mrw_exit
cdrom: Remove cdrom_count_tracks prototype
cdrom: Remove cdrom_get_next_writeable prototype
cdrom: Remove cdrom_get_last_written prototype
cdrom: Move mmc_ioctls above cdrom_ioctl to remove unnecessary prototype
cdrom: Remove unnecessary sanitize_format prototype
cdrom: Remove unnecessary check_for_audio_disc prototype
cdrom: Remove prototype for open_for_data
...
Pull block core updates from Jens Axboe:
"It's a big(ish) round this time, lots of development effort has gone
into blk-mq in the last 3 months. Generally we're heading to where
3.16 will be a feature complete and performant blk-mq. scsi-mq is
progressing nicely and will hopefully be in 3.17. A nvme port is in
progress, and the Micron pci-e flash driver, mtip32xx, is converted
and will be sent in with the driver pull request for 3.16.
This pull request contains:
- Lots of prep and support patches for scsi-mq have been integrated.
All from Christoph.
- API and code cleanups for blk-mq from Christoph.
- Lots of good corner case and error handling cleanup fixes for
blk-mq from Ming Lei.
- A flew of blk-mq updates from me:
* Provide strict mappings so that the driver can rely on the CPU
to queue mapping. This enables optimizations in the driver.
* Provided a bitmap tagging instead of percpu_ida, which never
really worked well for blk-mq. percpu_ida relies on the fact
that we have a lot more tags available than we really need, it
fails miserably for cases where we exhaust (or are close to
exhausting) the tag space.
* Provide sane support for shared tag maps, as utilized by scsi-mq
* Various fixes for IO timeouts.
* API cleanups, and lots of perf tweaks and optimizations.
- Remove 'buffer' from struct request. This is ancient code, from
when requests were always virtually mapped. Kill it, to reclaim
some space in struct request. From me.
- Remove 'magic' from blk_plug. Since we store these on the stack
and since we've never caught any actual bugs with this, lets just
get rid of it. From me.
- Only call part_in_flight() once for IO completion, as includes two
atomic reads. Hopefully we'll get a better implementation soon, as
the part IO stats are now one of the more expensive parts of doing
IO on blk-mq. From me.
- File migration of block code from {mm,fs}/ to block/. This
includes bio.c, bio-integrity.c, bounce.c, and ioprio.c. From me,
from a discussion on lkml.
That should describe the meat of the pull request. Also has various
little fixes and cleanups from Dave Jones, Shaohua Li, Duan Jiong,
Fengguang Wu, Fabian Frederick, Randy Dunlap, Robert Elliott, and Sam
Bradshaw"
* 'for-3.16/core' of git://git.kernel.dk/linux-block: (100 commits)
blk-mq: push IPI or local end_io decision to __blk_mq_complete_request()
blk-mq: remember to start timeout handler for direct queue
block: ensure that the timer is always added
blk-mq: blk_mq_unregister_hctx() can be static
blk-mq: make the sysfs mq/ layout reflect current mappings
blk-mq: blk_mq_tag_to_rq should handle flush request
block: remove dead code in scsi_ioctl:blk_verify_command
blk-mq: request initialization optimizations
block: add queue flag for disabling SG merging
block: remove 'magic' from struct blk_plug
blk-mq: remove alloc_hctx and free_hctx methods
blk-mq: add file comments and update copyright notices
blk-mq: remove blk_mq_alloc_request_pinned
blk-mq: do not use blk_mq_alloc_request_pinned in blk_mq_map_request
blk-mq: remove blk_mq_wait_for_tags
blk-mq: initialize request in __blk_mq_alloc_request
blk-mq: merge blk_mq_alloc_reserved_request into blk_mq_alloc_request
blk-mq: add helper to insert requests from irq context
blk-mq: remove stale comment for blk_mq_complete_request()
blk-mq: allow non-softirq completions
...
We have callers outside of the blk-mq proper (like timeouts) that
want to call __blk_mq_complete_request(), so rename the function
and put the decision code for whether to use ->softirq_done_fn
or blk_mq_endio() into __blk_mq_complete_request().
This also makes the interface more logical again.
blk_mq_complete_request() attempts to atomically mark the request
completed, and calls __blk_mq_complete_request() if successful.
__blk_mq_complete_request() then just ends the request.
Signed-off-by: Jens Axboe <axboe@fb.com>
Commit 07068d5b8e added a direct-to-hw-queue mode, but this mode
needs to remember to add the request timeout handler as well.
Without it, we don't track timeouts for these requests.
Signed-off-by: Jens Axboe <axboe@fb.com>
Commit f793aa5378 relaxed the timer addition a little too much.
If the timer isn't pending, we always need to add it.
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently blk-mq registers all the hardware queues in sysfs,
regardless of whether it uses them (e.g. they have CPU mappings)
or not. The unused hardware queues lack the cpux/ directories,
and the other sysfs entries (like active, pending, etc) are all
zeroes.
Change this so that sysfs correctly reflects the current mappings
of the hardware queues.
Signed-off-by: Jens Axboe <axboe@fb.com>
flush request is special, which borrows the tag from the parent
request. Hence blk_mq_tag_to_rq needs special handling to return
the flush request from the tag.
Signed-off-by: Shaohua Li <shli@fusionio.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
filter gets assigned the address of blk_default_cmd_filter on
entry to this function, so the !filter condition can never be true.
Signed-off-by: Dave Jones <davej@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
We currently clear a lot more than we need to, so make that a bit
more clever. Make some of the init dependent on features, like
only setting start_time if we are going to use it.
Signed-off-by: Jens Axboe <axboe@fb.com>
If devices are not SG starved, we waste a lot of time potentially
collapsing SG segments. Enough that 1.5% of the CPU time goes
to this, at only 400K IOPS. Add a queue flag, QUEUE_FLAG_NO_SG_MERGE,
which just returns the number of vectors in a bio instead of looping
over all segments and checking for collapsible ones.
Add a BLK_MQ_F_SG_MERGE flag so that drivers can opt-in on the sg
merging, if they so desire.
Signed-off-by: Jens Axboe <axboe@fb.com>
I don't think we've ever caught any bugs with this, and there's the
list poisoning for the plug lists to catch uninitialized cases.
So remove the magic member and save 8 bytes in the struct.
Signed-off-by: Jens Axboe <axboe@fb.com>
Pull in core changes (again), since we got rid of the alloc/free
hctx mq_ops hooks and mtip32xx then needed updating again.
Signed-off-by: Jens Axboe <axboe@fb.com>
There is no need for drivers to control hardware context allocation
now that we do the context to node mapping in common code.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
None of the blk-mq files have an explanatory comment at the top
for what that particular file does. Add that and add appropriate
copyright notices as well.
Signed-off-by: Jens Axboe <axboe@fb.com>
mtip32xx uses blk_mq_alloc_reserved_request(), so pull in the
core changes so we have a properly merged end result.
Signed-off-by: Jens Axboe <axboe@fb.com>
We now only have one caller left and can open code it there in a cleaner
way.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
We already do a non-blocking allocation in blk_mq_map_request, no need
to repeat it. Just call __blk_mq_alloc_request to wait directly.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
The current logic for blocking tag allocation is rather confusing, as we
first allocated and then free again a tag in blk_mq_wait_for_tags, just
to attempt a non-blocking allocation and then repeat if someone else
managed to grab the tag before us.
Instead change blk_mq_alloc_request_pinned to simply do a blocking tag
allocation itself and use the request we get back from it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Both callers if __blk_mq_alloc_request want to initialize the request, so
lift it into the common path.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Instead of having two almost identical copies of the same code just let
the callers pass in the reserved flag directly.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Both the cache flush state machine and the SCSI midlayer want to submit
requests from irq context, and the current per-request requeue_work
unfortunately causes corruption due to sharing with the csd field for
flushes. Replace them with a per-request_queue list of requests to
be requeued.
Based on an earlier test by Ming Lei.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reported-by: Ming Lei <tom.leiming@gmail.com>
Tested-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Right now we export two ways of completing a request:
1) blk_mq_complete_request(). This uses an IPI (if needed) and
completes through q->softirq_done_fn(). It also works with
timeouts.
2) blk_mq_end_io(). This completes inline, and ignores any timeout
state of the request.
Let blk_mq_complete_request() handle non-softirq_done_fn completions
as well, by just completing inline. If a driver has enough completion
ports to place completions correctly, it need not define a
mq_ops->complete() and we can avoid an indirect function call by
doing the completion inline.
Signed-off-by: Jens Axboe <axboe@fb.com>
Drivers currently have to figure this out on their own, and they
are missing information to do it properly. The ones that did
attempt to do it, do it wrong.
So just pass in the suggested node directly to the alloc
function.
Signed-off-by: Jens Axboe <axboe@fb.com>
The percpu counter is only used for blk-mq, so move
its allocation and free inside blk-mq, and don't
allocate it for legacy queue device.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq_exit_hw_queues() and blk_mq_free_hw_queues()
are introduced to avoid code duplication.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
hctx->ctx_map should have been freed inside blk_mq_free_queue().
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
__blkdev_issue_zeroout is only used in blk-lib.c
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: Jens Axboe <axboe@fb.com>
Without this we can leak the active_queues reference if a command is
freed while it is considered active.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently blk-mq uses the queue timeout for all requests. But
for some commands, drivers may want to set a specific timeout
for special requests. Allow this to be passed in through
request->timeout, and use it if set.
Signed-off-by: Jens Axboe <axboe@fb.com>
Export the blk-mq in-flight tag iterator for driver consumption.
This is particularly useful in exception paths or SRSI where
in-flight IOs need to be cancelled and/or reissued. The NVMe driver
conversion will use this.
Signed-off-by: Sam Bradshaw <sbradshaw@micron.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
We want slightly different behavior from them:
- On single queue devices, we currently use the per-process plug
for deferred IO and for merging.
- On multi queue devices, we don't use the per-process plug, but
we want to go straight to hardware for SYNC IO.
Split blk_mq_make_request() into a blk_sq_make_request() for single
queue devices, and retain blk_mq_make_request() for multi queue
devices. Then we don't need multiple checks for q->nr_hw_queues
in the request mapping.
Signed-off-by: Jens Axboe <axboe@fb.com>
Depending on the topology of the machine and the number of queues
exposed by a device, we can end up in a situation where some of
the hardware queues are unused (as in, they don't map to any
software queues). For this case, free up the memory used by the
request map, as we will not use it. This can be a substantial
amount of memory, depending on the number of queues vs CPUs and
the queue depth of the device.
Signed-off-by: Jens Axboe <axboe@fb.com>
Prepare this for the next patch which adds more smarts in the
plugging logic, so that we can save some memory.
Signed-off-by: Jens Axboe <axboe@fb.com>
In blk_mq_make_request(), do the blk_queue_nomerges() check
outside the call to blk_attempt_plug_merge() to eliminate
function call overhead when nomerges=2 (disabled)
Signed-off-by: Robert Elliott <elliott@hp.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_queue_make_requests() overwrites our set value for q->nr_requests,
turning it into the default of 128. Set this appropriately after
initializing queue values in blk_queue_make_request().
Signed-off-by: Jens Axboe <axboe@fb.com>
For request_fn based devices, the block layer exports a 'nr_requests'
file through sysfs to allow adjusting of queue depth on the fly.
Currently this returns -EINVAL for blk-mq, since it's not wired up.
Wire this up for blk-mq, so that it now also always dynamic
adjustments of the allowed queue depth for any given block device
managed by blk-mq.
Signed-off-by: Jens Axboe <axboe@fb.com>
Continue moving some of the block files that are scattered around.
bounce.c contains only code for bouncing the contents of a bio.
It's block proper code, not mm code.
Suggested-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Each hardware queue has a bitmap of software queues with pending
requests. When new IO is queued on a software queue, the bit is
set, and when IO is pruned on a hardware queue run, the bit is
cleared. This causes a lot of traffic. Switch this from the regular
BITS_PER_LONG bitmap to a sparser layout, similarly to what was
done for blk-mq tagging.
20% performance increase was observed for single threaded IO, and
about 15% performanc increase on multiple threads driving the
same device.
Signed-off-by: Jens Axboe <axboe@fb.com>
They really belong in block/, especially now since it's not in
drivers/block/ anymore. Additionally, the get_maintainer script
gets it wrong when in fs/.
Suggested-by: Christoph Hellwig <hch@infradead.org>
Acked-by: Al Viro <viro@ZenIV.linux.org.uk>
Signed-off-by: Jens Axboe <axboe@fb.com>
cgroup in general is moving towards using cgroup_subsys_state as the
fundamental structural component and css_parent() was introduced to
convert from using cgroup->parent to css->parent. It was quite some
time ago and we're moving forward with making css more prominent.
This patch drops the trivial wrapper css_parent() and let the users
dereference css->parent. While at it, explicitly mark fields of css
which are public and immutable.
v2: New usage from device_cgroup.c converted.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: "David S. Miller" <davem@davemloft.net>
Acked-by: Li Zefan <lizefan@huawei.com>
Cc: Vivek Goyal <vgoyal@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Johannes Weiner <hannes@cmpxchg.org>
This adds support for active queue tracking, meaning that the
blk-mq tagging maintains a count of active users of a tag set.
This allows us to maintain a notion of fairness between users,
so that we can distribute the tag depth evenly without starving
some users while allowing others to try unfair deep queues.
If sharing of a tag set is detected, each hardware queue will
track the depth of its own queue. And if this exceeds the total
depth divided by the number of active queues, the user is actively
throttled down.
The active queue count is done lazily to avoid bouncing that data
between submitter and completer. Each hardware queue gets marked
active when it allocates its first tag, and gets marked inactive
when 1) the last tag is cleared, and 2) the queue timeout grace
period has passed.
Signed-off-by: Jens Axboe <axboe@fb.com>
Convert all cftype->write_string() users to the new cftype->write()
which maps directly to kernfs write operation and has full access to
kernfs and cgroup contexts. The conversions are mostly mechanical.
* @css and @cft are accessed using of_css() and of_cft() accessors
respectively instead of being specified as arguments.
* Should return @nbytes on success instead of 0.
* @buf is not trimmed automatically. Trim if necessary. Note that
blkcg and netprio don't need this as the parsers already handle
whitespaces.
cftype->write_string() has no user left after the conversions and
removed.
While at it, remove unnecessary local variable @p in
cgroup_subtree_control_write() and stale comment about
CGROUP_LOCAL_BUFFER_SIZE in cgroup_freezer.c.
This patch doesn't introduce any visible behavior changes.
v2: netprio was missing from conversion. Converted.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Aristeu Rozanski <arozansk@redhat.com>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Acked-by: Li Zefan <lizefan@huawei.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Neil Horman <nhorman@tuxdriver.com>
Cc: "David S. Miller" <davem@davemloft.net>
Unlike the more usual refcnting, what css_tryget() provides is the
distinction between online and offline csses instead of protection
against upping a refcnt which already reached zero. cgroup is
planning to provide actual tryget which fails if the refcnt already
reached zero. Let's rename the existing trygets so that they clearly
indicate that they're onliness.
I thought about keeping the existing names as-are and introducing new
names for the planned actual tryget; however, given that each
controller participates in the synchronization of the online state, it
seems worthwhile to make it explicit that these functions are about
on/offline state.
Rename css_tryget() to css_tryget_online() and css_tryget_from_dir()
to css_tryget_online_from_dir(). This is pure rename.
v2: cgroup_freezer grew new usages of css_tryget(). Update
accordingly.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Acked-by: Michal Hocko <mhocko@suse.cz>
Acked-by: Li Zefan <lizefan@huawei.com>
Cc: Vivek Goyal <vgoyal@redhat.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Both nr_cache and nr_tags arn't needed for bitmap tag anymore.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The selected tag should be selected at random between 0 and
(depth - 1) with probability 1/depth, instead between 0 and
(depth - 2) with probability 1/(depth - 1).
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The barrier isn't necessary because both atomic_dec_and_test()
and wake_up() implicate one barrier.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The unlock memory barrier need to order access to req in free
path and clearing tag bit, otherwise either request free path
may see a allocated request, or initialized request in allocate
path might be modified by the ongoing free path.
Signed-off-by: Ming Lei <tom.leiming@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
We first check if we have inflight IO, then retrieve that
same number again. Usually this isn't that costly since the
chance of having the data dirtied in between is small, but
there's no reason for calling part_in_flight() twice.
Signed-off-by: Jens Axboe <axboe@fb.com>
Commit c6d600c6 opened up a small race where we could attempt to
account IO completion on a request, racing with IO start accounting.
Fix this up by ensuring that we've accounted for IO start before
inserting the request.
Signed-off-by: Jens Axboe <axboe@fb.com>
For best performance, spreading tags over multiple cachelines
makes the tagging more efficient on multicore systems. But since
we have 8 * sizeof(unsigned long) tags per cacheline, we don't
always get a nice spread.
Attempt to spread the tags over at least 4 cachelines, using fewer
number of bits per unsigned long if we have to. This improves
tagging performance in setups with 32-128 tags. For higher depths,
the spread is the same as before (BITS_PER_LONG tags per cacheline).
Signed-off-by: Jens Axboe <axboe@fb.com>
blk-mq currently uses percpu_ida for tag allocation. But that only
works well if the ratio between tag space and number of CPUs is
sufficiently high. For most devices and systems, that is not the
case. The end result if that we either only utilize the tag space
partially, or we end up attempting to fully exhaust it and run
into lots of lock contention with stealing between CPUs. This is
not optimal.
This new tagging scheme is a hybrid bitmap allocator. It uses
two tricks to both be SMP friendly and allow full exhaustion
of the space:
1) We cache the last allocated (or freed) tag on a per blk-mq
software context basis. This allows us to limit the space
we have to search. The key element here is not caching it
in the shared tag structure, otherwise we end up dirtying
more shared cache lines on each allocate/free operation.
2) The tag space is split into cache line sized groups, and
each context will start off randomly in that space. Even up
to full utilization of the space, this divides the tag users
efficiently into cache line groups, avoiding dirtying the same
one both between allocators and between allocator and freeer.
This scheme shows drastically better behaviour, both on small
tag spaces but on large ones as well. It has been tested extensively
to show better performance for all the cases blk-mq cares about.
Signed-off-by: Jens Axboe <axboe@fb.com>
This allows us to avoid a non-atomic memset over ->atomic_flags as well
as killing lots of duplicate initializations.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Right now we just pick the first CPU in the mask, but that can
easily overload that one. Add some basic batching and round-robin
all the entries in the mask instead.
Signed-off-by: Jens Axboe <axboe@fb.com>
