Currently when non-mq aware IO scheduler (BFQ, mq-deadline) is used for
a queue with multiple HW queues, the performance it rather bad. The
problem is that these IO schedulers use queue-wide locking and their
dispatch function does not respect the hctx it is passed in and returns
any request it finds appropriate. Thus locality of request access is
broken and dispatch from multiple CPUs just contends on IO scheduler
locks. For these IO schedulers there's little point in dispatching from
multiple CPUs. Instead dispatch always only from a single CPU to limit
contention.
Below is a comparison of dbench runs on XFS filesystem where the storage
is a raid card with 64 HW queues and to it attached a single rotating
disk. BFQ is used as IO scheduler:
clients MQ SQ MQ-Patched
Amean 1 39.12 (0.00%) 43.29 * -10.67%* 36.09 * 7.74%*
Amean 2 128.58 (0.00%) 101.30 * 21.22%* 96.14 * 25.23%*
Amean 4 577.42 (0.00%) 494.47 * 14.37%* 508.49 * 11.94%*
Amean 8 610.95 (0.00%) 363.86 * 40.44%* 362.12 * 40.73%*
Amean 16 391.78 (0.00%) 261.49 * 33.25%* 282.94 * 27.78%*
Amean 32 324.64 (0.00%) 267.71 * 17.54%* 233.00 * 28.23%*
Amean 64 295.04 (0.00%) 253.02 * 14.24%* 242.37 * 17.85%*
Amean 512 10281.61 (0.00%) 10211.16 * 0.69%* 10447.53 * -1.61%*
Numbers are times so lower is better. MQ is stock 5.10-rc6 kernel. SQ is
the same kernel with megaraid_sas.host_tagset_enable=0 so that the card
advertises just a single HW queue. MQ-Patched is a kernel with this
patch applied.
You can see multiple hardware queues heavily hurt performance in
combination with BFQ. The patch restores the performance.
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Introduce pointers for the blk_mq_tags regular and reserved bitmap tags,
with the goal of later being able to use a common shared tag bitmap across
all HW contexts in a set.
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Don Brace<don.brace@microsemi.com> #SCSI resv cmds patches used
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Move the blk_mq_bio_list_merge() into blk-merge.c and
rename it as a generic name.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
None of the I/O schedulers actually needs it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Daniel Wagner <dwagner@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
No code that occurs between blk_mq_get_ctx() and blk_mq_put_ctx() depends
on preemption being disabled for its correctness. Since removing the CPU
preemption calls does not measurably affect performance, simplify the
blk-mq code by removing the blk_mq_put_ctx() function and also by not
disabling preemption in blk_mq_get_ctx().
Cc: Hannes Reinecke <hare@suse.com>
Cc: Omar Sandoval <osandov@fb.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We only need the number of segments in the blk-mq submission path.
Remove the field from struct bio, and return it from a variant of
blk_queue_split instead of that it can passed as an argument to
those functions that need the value.
This also means we stop recounting segments except for cloning
and partial segments.
To keep the number of arguments in this how path down remove
pointless struct request_queue arguments from any of the functions
that had it and grew a nr_segs argument.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
All these files have some form of the usual GPLv2 boilerplate. Switch
them to use SPDX tags instead.
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
sbq_wake_ptr() checks sbq->ws_active to know if it needs to loop
the wait indexes or not. This requires the use of the sbitmap
waitqueue wrappers, but kyber doesn't use those for its domain
token waitqueue handling.
Convert kyber to use the helpers. This fixes a hang with waiting
for domain tokens.
Fixes: 5d2ee7122c ("sbitmap: optimize wakeup check")
Tested-by: Ming Lei <ming.lei@redhat.com>
Reported-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The mapping used to be dependent on just the CPU location, but
now it's a tuple of (type, cpu) instead. This is a prep patch
for allowing a single software queue to map to multiple hardware
queues. No functional changes in this patch.
This changes the software queue count to an unsigned short
to save a bit of space. We can still support 64K-1 CPUs,
which should be enough. Add a check to catch a wrap.
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This is a remnant of when we had ops for both SQ and MQ
schedulers. Now it's just MQ, so get rid of the union.
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This removes a bunch of core and elevator related code. On the core
front, we remove anything related to queue running, draining,
initialization, plugging, and congestions. We also kill anything
related to request allocation, merging, retrieval, and completion.
Remove any checking for single queue IO schedulers, as they no
longer exist. This means we can also delete a bunch of code related
to request issue, adding, completion, etc - and all the SQ related
ops and helpers.
Also kill the load_default_modules(), as all that did was provide
for a way to load the default single queue elevator.
Tested-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
NSEC_PER_SEC has type long, so 5 * NSEC_PER_SEC is calculated as a long.
However, 5 seconds is 5,000,000,000 nanoseconds, which overflows a
32-bit long. Make sure all of the targets are calculated as 64-bit
values.
Fixes: 6e25cb01ea ("kyber: implement improved heuristics")
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When debugging Kyber, it's really useful to know what latencies we've
been having, how the domain depths have been adjusted, and if we've
actually been throttling. Add three tracepoints, kyber_latency,
kyber_adjust, and kyber_throttled, to record that.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Kyber's current heuristics have a few flaws:
- It's based on the mean latency, but p99 latency tends to be more
meaningful to anyone who cares about latency. The mean can also be
skewed by rare outliers that the scheduler can't do anything about.
- The statistics calculations are purely time-based with a short window.
This works for steady, high load, but is more sensitive to outliers
with bursty workloads.
- It only considers the latency once an I/O has been submitted to the
device, but the user cares about the time spent in the kernel, as
well.
These are shortcomings of the generic blk-stat code which doesn't quite
fit the ideal use case for Kyber. So, this replaces the statistics with
a histogram used to calculate percentiles of total latency and I/O
latency, which we then use to adjust depths in a slightly more
intelligent manner:
- Sync and async writes are now the same domain.
- Discards are a separate domain.
- Domain queue depths are scaled by the ratio of the p99 total latency
to the target latency (e.g., if the p99 latency is double the target
latency, we will double the queue depth; if the p99 latency is half of
the target latency, we can halve the queue depth).
- We use the I/O latency to determine whether we should scale queue
depths down: we will only scale down if any domain's I/O latency
exceeds the target latency, which is an indicator of congestion in the
device.
These new heuristics are just as scalable as the heuristics they
replace.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The domain token sbitmaps are currently initialized to the device queue
depth or 256, whichever is larger, and immediately resized to the
maximum depth for that domain (256, 128, or 64 for read, write, and
other, respectively). The sbitmap is never resized larger than that, so
it's unnecessary to allocate a bitmap larger than the maximum depth.
Let's just allocate it to the maximum depth to begin with. This will use
marginally less memory, and more importantly, give us a more appropriate
number of bits per sbitmap word.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Commit 4bc6339a58 ("block: move blk_stat_add() to
__blk_mq_end_request()") consolidated some calls using ktime_get() so
we'd only need to call it once. Kyber's ->completed_request() hook also
calls ktime_get(), so let's move it to the same place, too.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, kyber is very unfriendly with merging. kyber depends
on ctx rq_list to do merging, however, most of time, it will not
leave any requests in ctx rq_list. This is because even if tokens
of one domain is used up, kyber will try to dispatch requests
from other domain and flush the rq_list there.
To improve this, we setup kyber_ctx_queue (kcq) which is similar
with ctx, but it has rq_lists for different domain and build same
mapping between kcq and khd as the ctx & hctx. Then we could merge,
insert and dispatch for different domains separately. At the same
time, only flush the rq_list of kcq when get domain token successfully.
Then if one domain token is used up, the requests could be left in
the rq_list of that domain and maybe merged with following io.
Following is my test result on machine with 8 cores and NVMe card
INTEL SSDPEKKR128G7
fio size=256m ioengine=libaio iodepth=64 direct=1 numjobs=8
seq/random
+------+---------------------------------------------------------------+
|patch?| bw(MB/s) | iops | slat(usec) | clat(usec) | merge |
+----------------------------------------------------------------------+
| w/o | 606/612 | 151k/153k | 6.89/7.03 | 3349.21/3305.40 | 0/0 |
+----------------------------------------------------------------------+
| w/ | 1083/616 | 277k/154k | 4.93/6.95 | 1830.62/3279.95 | 223k/3k |
+----------------------------------------------------------------------+
When set numjobs to 16, the bw and iops could reach 1662MB/s and 425k
on my platform.
Signed-off-by: Jianchao Wang <jianchao.w.wang@oracle.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We don't expect the async depth to be smaller than the wake batch
count for sbitmap, but just in case, inform sbitmap of what shallow
depth kyber may use.
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
struct blk_issue_stat squashes three things into one u64:
- The time the driver started working on a request
- The original size of the request (for the io.low controller)
- Flags for writeback throttling
It turns out that on x86_64, we have a 4 byte hole in struct request
which we can fill with the non-timestamp fields from blk_issue_stat,
simplifying things quite a bit.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When requeuing request, the domain token should have been freed
before re-inserting the request to io scheduler. Otherwise, the
assigned domain token will be leaked, and IO hang can be caused.
Cc: Paolo Valente <paolo.valente@linaro.org>
Cc: Omar Sandoval <osandov@fb.com>
Cc: stable@vger.kernel.org
Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Commit 8cf4666020 ("kyber: fix hang on domain token wait queue") fixed
a hang caused by leaving wait entries on the domain token wait queue
after the __sbitmap_queue_get() retry succeeded, making that wait entry
a "dud" which won't in turn wake more entries up. However, we can also
get a dud entry if kyber_get_domain_token() fails once but is then
called again and succeeds. This can happen if the hardware queue is
rerun for some other reason, or, more likely, kyber_dispatch_request()
tries the same domain twice.
The fix is to remove our entry from the wait queue whenever we
successfully get a token. The only complication is that we might be on
one of many wait queues in the struct sbitmap_queue, but that's easily
fixed by remembering which wait queue we were put on.
While we're here, only initialize the wait queue entry once instead of
on every wait, and use spin_lock_irq() instead of spin_lock_irqsave(),
since this is always called from process context with irqs enabled.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There may be request in sw queue, and not fetched to domain queue
yet, so check it in kyber_has_work().
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When we're getting a domain token, if we fail to get a token on our
first attempt, we put the current hardware queue on a wait queue and
then try again just in case a token was freed after our initial attempt
but before we got on the wait queue. If this second attempt succeeds, we
currently leave the hardware queue on the wait queue. Usually this is
okay; we'll just run the hardware queue one extra time when another
token is freed. However, if the hardware queue doesn't have any other
requests waiting, then when it it gets the extra wakeup, it won't have
anything to free and therefore won't wake up any other hardware queues.
If tokens are limited, then we won't make forward progress and the
device will hang.
Reported-by: Bin Zha <zhabin.zb@alibaba-inc.com>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull scheduler updates from Ingo Molnar:
"The main changes in this cycle were:
- Add the SYSTEM_SCHEDULING bootup state to move various scheduler
debug checks earlier into the bootup. This turns silent and
sporadically deadly bugs into nice, deterministic splats. Fix some
of the splats that triggered. (Thomas Gleixner)
- A round of restructuring and refactoring of the load-balancing and
topology code (Peter Zijlstra)
- Another round of consolidating ~20 of incremental scheduler code
history: this time in terms of wait-queue nomenclature. (I didn't
get much feedback on these renaming patches, and we can still
easily change any names I might have misplaced, so if anyone hates
a new name, please holler and I'll fix it.) (Ingo Molnar)
- sched/numa improvements, fixes and updates (Rik van Riel)
- Another round of x86/tsc scheduler clock code improvements, in hope
of making it more robust (Peter Zijlstra)
- Improve NOHZ behavior (Frederic Weisbecker)
- Deadline scheduler improvements and fixes (Luca Abeni, Daniel
Bristot de Oliveira)
- Simplify and optimize the topology setup code (Lauro Ramos
Venancio)
- Debloat and decouple scheduler code some more (Nicolas Pitre)
- Simplify code by making better use of llist primitives (Byungchul
Park)
- ... plus other fixes and improvements"
* 'sched-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (103 commits)
sched/cputime: Refactor the cputime_adjust() code
sched/debug: Expose the number of RT/DL tasks that can migrate
sched/numa: Hide numa_wake_affine() from UP build
sched/fair: Remove effective_load()
sched/numa: Implement NUMA node level wake_affine()
sched/fair: Simplify wake_affine() for the single socket case
sched/numa: Override part of migrate_degrades_locality() when idle balancing
sched/rt: Move RT related code from sched/core.c to sched/rt.c
sched/deadline: Move DL related code from sched/core.c to sched/deadline.c
sched/cpuset: Only offer CONFIG_CPUSETS if SMP is enabled
sched/fair: Spare idle load balancing on nohz_full CPUs
nohz: Move idle balancer registration to the idle path
sched/loadavg: Generalize "_idle" naming to "_nohz"
sched/core: Drop the unused try_get_task_struct() helper function
sched/fair: WARN() and refuse to set buddy when !se->on_rq
sched/debug: Fix SCHED_WARN_ON() to return a value on !CONFIG_SCHED_DEBUG as well
sched/wait: Disambiguate wq_entry->task_list and wq_head->task_list naming
sched/wait: Move bit_wait_table[] and related functionality from sched/core.c to sched/wait_bit.c
sched/wait: Split out the wait_bit*() APIs from <linux/wait.h> into <linux/wait_bit.h>
sched/wait: Re-adjust macro line continuation backslashes in <linux/wait.h>
...
So I've noticed a number of instances where it was not obvious from the
code whether ->task_list was for a wait-queue head or a wait-queue entry.
Furthermore, there's a number of wait-queue users where the lists are
not for 'tasks' but other entities (poll tables, etc.), in which case
the 'task_list' name is actively confusing.
To clear this all up, name the wait-queue head and entry list structure
fields unambiguously:
struct wait_queue_head::task_list => ::head
struct wait_queue_entry::task_list => ::entry
For example, this code:
rqw->wait.task_list.next != &wait->task_list
... is was pretty unclear (to me) what it's doing, while now it's written this way:
rqw->wait.head.next != &wait->entry
... which makes it pretty clear that we are iterating a list until we see the head.
Other examples are:
list_for_each_entry_safe(pos, next, &x->task_list, task_list) {
list_for_each_entry(wq, &fence->wait.task_list, task_list) {
... where it's unclear (to me) what we are iterating, and during review it's
hard to tell whether it's trying to walk a wait-queue entry (which would be
a bug), while now it's written as:
list_for_each_entry_safe(pos, next, &x->head, entry) {
list_for_each_entry(wq, &fence->wait.head, entry) {
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Rename:
wait_queue_t => wait_queue_entry_t
'wait_queue_t' was always a slight misnomer: its name implies that it's a "queue",
but in reality it's a queue *entry*. The 'real' queue is the wait queue head,
which had to carry the name.
Start sorting this out by renaming it to 'wait_queue_entry_t'.
This also allows the real structure name 'struct __wait_queue' to
lose its double underscore and become 'struct wait_queue_entry',
which is the more canonical nomenclature for such data types.
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
This patch makes sure we always allocate requests in the core blk-mq
code and use a common prepare_request method to initialize them for
both mq I/O schedulers. For Kyber and additional limit_depth method
is added that is called before allocating the request.
Also because none of the intializations can really fail the new method
does not return an error - instead the bfq finish method is hardened
to deal with the no-IOC case.
Last but not least this removes the abuse of RQF_QUEUE by the blk-mq
scheduling code as RQF_ELFPRIV is all that is needed now.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In order to allow for filtering of IO based on some other properties
of the request than direction we allow the bucket function to return
an int.
If the bucket callback returns a negative do no count it in the stats
accumulation.
Signed-off-by: Stephen Bates <sbates@raithlin.com>
Fixed up Kyber scheduler stat callback.
Signed-off-by: Jens Axboe <axboe@fb.com>
The Kyber I/O scheduler is an I/O scheduler for fast devices designed to
scale to multiple queues. Users configure only two knobs, the target
read and synchronous write latencies, and the scheduler tunes itself to
achieve that latency goal.
The implementation is based on "tokens", built on top of the scalable
bitmap library. Tokens serve as a mechanism for limiting requests. There
are two tiers of tokens: queueing tokens and dispatch tokens.
A queueing token is required to allocate a request. In fact, these
tokens are actually the blk-mq internal scheduler tags, but the
scheduler manages the allocation directly in order to implement its
policy.
Dispatch tokens are device-wide and split up into two scheduling
domains: reads vs. writes. Each hardware queue dispatches batches
round-robin between the scheduling domains as long as tokens are
available for that domain.
These tokens can be used as the mechanism to enable various policies.
The policy Kyber uses is inspired by active queue management techniques
for network routing, similar to blk-wbt. The scheduler monitors
latencies and scales the number of dispatch tokens accordingly. Queueing
tokens are used to prevent starvation of synchronous requests by
asynchronous requests.
Various extensions are possible, including better heuristics and ionice
support. The new scheduler isn't set as the default yet.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Jan Kara
John Garry
Baolin Wang
Christoph Hellwig
Bart Van Assche
Jens Axboe
Omar Sandoval
Jianchao Wang
Ming Lei
Linus Torvalds
Ingo Molnar
Stephen Bates