Consider a sync bfq_queue Q that remains empty while in service, and
suppose that, when this happens, there is a fair amount of already
in-flight I/O not belonging to Q. In such a situation, I/O dispatching
may need to be plugged (until new I/O arrives for Q), for the
following reason.
The drive may decide to serve in-flight non-Q's I/O requests before
Q's ones, thereby delaying the arrival of new I/O requests for Q
(recall that Q is sync). If I/O-dispatching is not plugged, then,
while Q remains empty, a basically uncontrolled amount of I/O from
other queues may be dispatched too, possibly causing the service of
Q's I/O to be delayed even longer in the drive. This problem gets more
and more serious as the speed and the queue depth of the drive grow,
because, as these two quantities grow, the probability to find no
queue busy but many requests in flight grows too.
If Q has the same weight and priority as the other queues, then the
above delay is unlikely to cause any issue, because all queues tend to
undergo the same treatment. So, since not plugging I/O dispatching is
convenient for throughput, it is better not to plug. Things change in
case Q has a higher weight or priority than some other queue, because
Q's service guarantees may simply be violated. For this reason,
commit 1de0c4cd9e ("block, bfq: reduce idling only in symmetric
scenarios") does plug I/O in such an asymmetric scenario. Plugging
minimizes the delay induced by already in-flight I/O, and enables Q to
recover the bandwidth it may lose because of this delay.
Yet the above commit does not cover the case of weight-raised queues,
for efficiency concerns. For weight-raised queues, I/O-dispatch
plugging is activated simply if not all bfq_queues are
weight-raised. But this check does not handle the case of in-flight
requests, because a bfq_queue may become non busy *before* all its
in-flight requests are completed.
This commit performs I/O-dispatch plugging for weight-raised queues if
there are some in-flight requests.
As a practical example of the resulting recover of control, under
write load on a Samsung SSD 970 PRO, gnome-terminal starts in 1.5
seconds after this fix, against 15 seconds before the fix (as a
reference, gnome-terminal takes about 35 seconds to start with any of
the other I/O schedulers).
Fixes: 1de0c4cd9e ("block, bfq: reduce idling only in symmetric scenarios")
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
-----BEGIN PGP SIGNATURE-----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=VSYT
-----END PGP SIGNATURE-----
Merge tag 'for-5.3/block-20190708' of git://git.kernel.dk/linux-block
Pull block updates from Jens Axboe:
"This is the main block updates for 5.3. Nothing earth shattering or
major in here, just fixes, additions, and improvements all over the
map. This contains:
- Series of documentation fixes (Bart)
- Optimization of the blk-mq ctx get/put (Bart)
- null_blk removal race condition fix (Bob)
- req/bio_op() cleanups (Chaitanya)
- Series cleaning up the segment accounting, and request/bio mapping
(Christoph)
- Series cleaning up the page getting/putting for bios (Christoph)
- block cgroup cleanups and moving it to where it is used (Christoph)
- block cgroup fixes (Tejun)
- Series of fixes and improvements to bcache, most notably a write
deadlock fix (Coly)
- blk-iolatency STS_AGAIN and accounting fixes (Dennis)
- Series of improvements and fixes to BFQ (Douglas, Paolo)
- debugfs_create() return value check removal for drbd (Greg)
- Use struct_size(), where appropriate (Gustavo)
- Two lighnvm fixes (Heiner, Geert)
- MD fixes, including a read balance and corruption fix (Guoqing,
Marcos, Xiao, Yufen)
- block opal shadow mbr additions (Jonas, Revanth)
- sbitmap compare-and-exhange improvemnts (Pavel)
- Fix for potential bio->bi_size overflow (Ming)
- NVMe pull requests:
- improved PCIe suspent support (Keith Busch)
- error injection support for the admin queue (Akinobu Mita)
- Fibre Channel discovery improvements (James Smart)
- tracing improvements including nvmetc tracing support (Minwoo Im)
- misc fixes and cleanups (Anton Eidelman, Minwoo Im, Chaitanya
Kulkarni)"
- Various little fixes and improvements to drivers and core"
* tag 'for-5.3/block-20190708' of git://git.kernel.dk/linux-block: (153 commits)
blk-iolatency: fix STS_AGAIN handling
block: nr_phys_segments needs to be zero for REQ_OP_WRITE_ZEROES
blk-mq: simplify blk_mq_make_request()
blk-mq: remove blk_mq_put_ctx()
sbitmap: Replace cmpxchg with xchg
block: fix .bi_size overflow
block: sed-opal: check size of shadow mbr
block: sed-opal: ioctl for writing to shadow mbr
block: sed-opal: add ioctl for done-mark of shadow mbr
block: never take page references for ITER_BVEC
direct-io: use bio_release_pages in dio_bio_complete
block_dev: use bio_release_pages in bio_unmap_user
block_dev: use bio_release_pages in blkdev_bio_end_io
iomap: use bio_release_pages in iomap_dio_bio_end_io
block: use bio_release_pages in bio_map_user_iov
block: use bio_release_pages in bio_unmap_user
block: optionally mark pages dirty in bio_release_pages
block: move the BIO_NO_PAGE_REF check into bio_release_pages
block: skd_main.c: Remove call to memset after dma_alloc_coherent
block: mtip32xx: Remove call to memset after dma_alloc_coherent
...
In reboot tests on several devices we were seeing a "use after free"
when slub_debug or KASAN was enabled. The kernel complained about:
Unable to handle kernel paging request at virtual address 6b6b6c2b
...which is a classic sign of use after free under slub_debug. The
stack crawl in kgdb looked like:
0 test_bit (addr=<optimized out>, nr=<optimized out>)
1 bfq_bfqq_busy (bfqq=<optimized out>)
2 bfq_select_queue (bfqd=<optimized out>)
3 __bfq_dispatch_request (hctx=<optimized out>)
4 bfq_dispatch_request (hctx=<optimized out>)
5 0xc056ef00 in blk_mq_do_dispatch_sched (hctx=0xed249440)
6 0xc056f728 in blk_mq_sched_dispatch_requests (hctx=0xed249440)
7 0xc0568d24 in __blk_mq_run_hw_queue (hctx=0xed249440)
8 0xc0568d94 in blk_mq_run_work_fn (work=<optimized out>)
9 0xc024c5c4 in process_one_work (worker=0xec6d4640, work=0xed249480)
10 0xc024cff4 in worker_thread (__worker=0xec6d4640)
Digging in kgdb, it could be found that, though bfqq looked fine,
bfqq->bic had been freed.
Through further digging, I postulated that perhaps it is illegal to
access a "bic" (AKA an "icq") after bfq_exit_icq() had been called
because the "bic" can be freed at some point in time after this call
is made. I confirmed that there certainly were cases where the exact
crashing code path would access the "bic" after bfq_exit_icq() had
been called. Sspecifically I set the "bfqq->bic" to (void *)0x7 and
saw that the bic was 0x7 at the time of the crash.
To understand a bit more about why this crash was fairly uncommon (I
saw it only once in a few hundred reboots), you can see that much of
the time bfq_exit_icq_fbqq() fully frees the bfqq and thus it can't
access the ->bic anymore. The only case it doesn't is if
bfq_put_queue() sees a reference still held.
However, even in the case when bfqq isn't freed, the crash is still
rare. Why? I tracked what happened to the "bic" after the exit
routine. It doesn't get freed right away. Rather,
put_io_context_active() eventually called put_io_context() which
queued up freeing on a workqueue. The freeing then actually happened
later than that through call_rcu(). Despite all these delays, some
extra debugging showed that all the hoops could be jumped through in
time and the memory could be freed causing the original crash. Phew!
To make a long story short, assuming it truly is illegal to access an
icq after the "exit_icq" callback is finished, this patch is needed.
Cc: stable@vger.kernel.org
Reviewed-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: Douglas Anderson <dianders@chromium.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Some debug code suggested by Paolo was tripping when I did reboot
stress tests. Specifically in bfq_bfqq_resume_state()
"bic->saved_wr_start_at_switch_to_srt" was later than the current
value of "jiffies". A bit of debugging showed that
"bic->saved_wr_start_at_switch_to_srt" was actually 0 and a bit more
debugging showed that was because we had run through the "unlikely"
case in the bfq_bfqq_save_state() function.
Let's init "saved_wr_start_at_switch_to_srt" in the unlikely case to
something sane.
NOTE: this fixes no known real-world errors.
Reviewed-by: Paolo Valente <paolo.valente@linaro.org>
Reviewed-by: Guenter Roeck <groeck@chromium.org>
Signed-off-by: Douglas Anderson <dianders@chromium.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
By mistake, there is a '&' instead of a '==' in the definition of the
macro BFQQ_TOTALLY_SEEKY. This commit replaces the wrong operator with
the correct one.
Fixes: 7074f076ff ("block, bfq: do not tag totally seeky queues as soft rt")
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Consider, on one side, a bfq_queue Q that remains empty while in
service, and, on the other side, the pending I/O of bfq_queues that,
according to their timestamps, have to be served after Q. If an
uncontrolled amount of I/O from the latter bfq_queues were dispatched
while Q is waiting for its new I/O to arrive, then Q's bandwidth
guarantees would be violated. To prevent this, I/O dispatch is plugged
until Q receives new I/O (except for a properly controlled amount of
injected I/O). Unfortunately, preemption breaks I/O-dispatch plugging,
for the following reason.
Preemption is performed in two steps. First, Q is expired and
re-scheduled. Second, the new bfq_queue to serve is chosen. The first
step is needed by the second, as the second can be performed only
after Q's timestamps have been properly updated (done in the
expiration step), and Q has been re-queued for service. This
dependency is a consequence of the way how BFQ's scheduling algorithm
is currently implemented.
But Q is not re-scheduled at all in the first step, because Q is
empty. As a consequence, an uncontrolled amount of I/O may be
dispatched until Q becomes non empty again. This breaks Q's service
guarantees.
This commit addresses this issue by re-scheduling Q even if it is
empty. This in turn breaks the assumption that all scheduled queues
are non empty. Then a few extra checks are now needed.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
BFQ enqueues the I/O coming from each process into a separate
bfq_queue, and serves bfq_queues one at a time. Each bfq_queue may be
served for at most timeout_sync milliseconds (default: 125 ms). This
service scheme is prone to the following inaccuracy.
While a bfq_queue Q1 is in service, some empty bfq_queue Q2 may
receive I/O, and, according to BFQ's scheduling policy, may become the
right bfq_queue to serve, in place of the currently in-service
bfq_queue. In this respect, postponing the service of Q2 to after the
service of Q1 finishes may delay the completion of Q2's I/O, compared
with an ideal service in which all non-empty bfq_queues are served in
parallel, and every non-empty bfq_queue is served at a rate
proportional to the bfq_queue's weight. This additional delay is equal
at most to the time Q1 may unjustly remain in service before switching
to Q2.
If Q1 and Q2 have the same weight, then this time is most likely
negligible compared with the completion time to be guaranteed to Q2's
I/O. In addition, first, one of the reasons why BFQ may want to serve
Q1 for a while is that this boosts throughput and, second, serving Q1
longer reduces BFQ's overhead. As a conclusion, it is usually better
not to preempt Q1 if both Q1 and Q2 have the same weight.
In contrast, as Q2's weight or priority becomes higher and higher
compared with that of Q1, the above delay becomes larger and larger,
compared with the I/O completion times that have to be guaranteed to
Q2 according to Q2's weight. So reducing this delay may be more
important than avoiding the costs of preempting Q1.
Accordingly, this commit preempts Q1 if Q2 has a higher weight or a
higher priority than Q1. Preemption causes Q1 to be re-scheduled, and
triggers a new choice of the next bfq_queue to serve. If Q2 really is
the next bfq_queue to serve, then Q2 will be set in service
immediately.
This change reduces the component of the I/O latency caused by the
above delay by about 80%. For example, on an (old) PLEXTOR PX-256M5
SSD, the maximum latency reported by fio drops from 15.1 to 3.2 ms for
a process doing sporadic random reads while another process is doing
continuous sequential reads.
Signed-off-by: Nicola Bottura <bottura.nicola95@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A bfq_queue Q may happen to be synchronized with another
bfq_queue Q2, i.e., the I/O of Q2 may need to be completed for Q to
receive new I/O. We call Q2 "waker queue".
If I/O plugging is being performed for Q, and Q is not receiving any
more I/O because of the above synchronization, then, thanks to BFQ's
injection mechanism, the waker queue is likely to get served before
the I/O-plugging timeout fires.
Unfortunately, this fact may not be sufficient to guarantee a high
throughput during the I/O plugging, because the inject limit for Q may
be too low to guarantee a lot of injected I/O. In addition, the
duration of the plugging, i.e., the time before Q finally receives new
I/O, may not be minimized, because the waker queue may happen to be
served only after other queues.
To address these issues, this commit introduces the explicit detection
of the waker queue, and the unconditional injection of a pending I/O
request of the waker queue on each invocation of
bfq_dispatch_request().
One may be concerned that this systematic injection of I/O from the
waker queue delays the service of Q's I/O. Fortunately, it doesn't. On
the contrary, next Q's I/O is brought forward dramatically, for it is
not blocked for milliseconds.
Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Until the base value for request service times gets finally computed
for a bfq_queue, the inject limit for that queue does depend on the
think-time state (short|long) of the queue. A timely update of the
think time then guarantees a quicker activation or deactivation of the
injection. Fortunately, the think time of a bfq_queue is updated in
the same code path as the inject limit; but after the inject limit.
This commits moves the update of the think time before the update of
the inject limit. For coherence, it moves the update of the seek time
too.
Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
I/O injection gets reduced if it increases the request service times
of the victim queue beyond a certain threshold. The threshold, in its
turn, is computed as a function of the base service time enjoyed by
the queue when it undergoes no injection.
As a consequence, for injection to work properly, the above base value
has to be accurate. In this respect, such a value may vary over
time. For example, it varies if the size or the spatial locality of
the I/O requests in the queue change. It is then important to update
this value whenever possible. This commit performs this update.
Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
One of the cases where the parameters for injection may be updated is
when there are no more in-flight I/O requests. The number of in-flight
requests is stored in the field bfqd->rq_in_driver of the descriptor
bfqd of the device. So, the controlled condition is
bfqd->rq_in_driver == 0.
Unfortunately, this is wrong because, the instruction that checks this
condition is in the code path that handles the completion of a
request, and, in particular, the instruction is executed before
bfqd->rq_in_driver is decremented in such a code path.
This commit fixes this issue by just replacing 0 with 1 in the
comparison.
Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Until the base value of the request service times gets finally
computed for a bfq_queue, the inject limit does depend on the
think-time state (short|long). The limit must be 0 or 1 if the think
time is deemed, respectively, as short or long. However, such a check
and possible limit update is performed only periodically, once per
second. So, to make the injection mechanism much more reactive, this
commit performs the update also every time the think-time state
changes.
In addition, in the following special case, this commit lets the
inject limit of a bfq_queue bfqq remain equal to 1 even if bfqq's
think time is short: bfqq's I/O is synchronized with that of some
other queue, i.e., bfqq may receive new I/O only after the I/O of the
other queue is completed. Keeping the inject limit to 1 allows the
blocking I/O to be served while bfqq is in service. And this is very
convenient both for bfqq and for the total throughput, as explained
in detail in the comments in bfq_update_has_short_ttime().
Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This option is entirely bfq specific, give it an appropinquate name.
Also make it depend on CONFIG_BFQ_GROUP_IOSCHED in Kconfig, as all
the functionality already does so anyway.
Acked-by: Tejun Heo <tj@kernel.org>
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Christoph Hellwig <hch@lst.de>
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 or later 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>
-----BEGIN PGP SIGNATURE-----
iQFSBAABCAA8FiEEq68RxlopcLEwq+PEeb4+QwBBGIYFAly8rGYeHHRvcnZhbGRz
QGxpbnV4LWZvdW5kYXRpb24ub3JnAAoJEHm+PkMAQRiGmZMH/1IRB0E1Qmzz8yzw
wj79UuRGYPqxDDSWW+wNc8sU4Ic7iYirn9APHAztCdQqsjmzU/OVLfSa3JhdBe5w
THo7pbGKBqEDcWnKfNk/21jXFNLZ1vr9BoQv2DGU2MMhHAyo/NZbalo2YVtpQPmM
OCRth5n+LzvH7rGrX7RYgWu24G9l3NMfgtaDAXBNXesCGFAjVRrdkU5CBAaabvtU
4GWh/nnutndOOLdByL3x+VZ3H3fIBnbNjcIGCglvvqzk7h3hrfGEl4UCULldTxcM
IFsfMUhSw1ENy7F6DHGbKIG90cdCJcrQ8J/ziEzjj/KLGALluutfFhVvr6YCM2J6
2RgU8CY=
=CfY1
-----END PGP SIGNATURE-----
Merge tag 'v5.1-rc6' into for-5.2/block
Pull in v5.1-rc6 to resolve two conflicts. One is in BFQ, in just a
comment, and is trivial. The other one is a conflict due to a later fix
in the bio multi-page work, and needs a bit more care.
* tag 'v5.1-rc6': (770 commits)
Linux 5.1-rc6
block: make sure that bvec length can't be overflow
block: kill all_q_node in request_queue
x86/cpu/intel: Lower the "ENERGY_PERF_BIAS: Set to normal" message's log priority
coredump: fix race condition between mmget_not_zero()/get_task_mm() and core dumping
mm/kmemleak.c: fix unused-function warning
init: initialize jump labels before command line option parsing
kernel/watchdog_hld.c: hard lockup message should end with a newline
kcov: improve CONFIG_ARCH_HAS_KCOV help text
mm: fix inactive list balancing between NUMA nodes and cgroups
mm/hotplug: treat CMA pages as unmovable
proc: fixup proc-pid-vm test
proc: fix map_files test on F29
mm/vmstat.c: fix /proc/vmstat format for CONFIG_DEBUG_TLBFLUSH=y CONFIG_SMP=n
mm/memory_hotplug: do not unlock after failing to take the device_hotplug_lock
mm: swapoff: shmem_unuse() stop eviction without igrab()
mm: swapoff: take notice of completion sooner
mm: swapoff: remove too limiting SWAP_UNUSE_MAX_TRIES
mm: swapoff: shmem_find_swap_entries() filter out other types
slab: store tagged freelist for off-slab slabmgmt
...
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A previous commit moved the shallow depth and BFQ depth map calculations
to be done at init time, moving it outside of the hotter IO path. This
potentially causes hangs if the users changes the depth of the scheduler
map, by writing to the 'nr_requests' sysfs file for that device.
Add a blk-mq-sched hook that allows blk-mq to inform the scheduler if
the depth changes, so that the scheduler can update its internal state.
Tested-by: Kai Krakow <kai@kaishome.de>
Reported-by: Paolo Valente <paolo.valente@linaro.org>
Fixes: f0635b8a41 ("bfq: calculate shallow depths at init time")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The function bfq_bfqq_expire() invokes the function
__bfq_bfqq_expire(), and the latter may free the in-service bfq-queue.
If this happens, then no other instruction of bfq_bfqq_expire() must
be executed, or a use-after-free will occur.
Basing on the assumption that __bfq_bfqq_expire() invokes
bfq_put_queue() on the in-service bfq-queue exactly once, the queue is
assumed to be freed if its refcounter is equal to one right before
invoking __bfq_bfqq_expire().
But, since commit 9dee8b3b05 ("block, bfq: fix queue removal from
weights tree") this assumption is false. __bfq_bfqq_expire() may also
invoke bfq_weights_tree_remove() and, since commit 9dee8b3b05
("block, bfq: fix queue removal from weights tree"), also
the latter function may invoke bfq_put_queue(). So __bfq_bfqq_expire()
may invoke bfq_put_queue() twice, and this is the actual case where
the in-service queue may happen to be freed.
To address this issue, this commit moves the check on the refcounter
of the queue right around the last bfq_put_queue() that may be invoked
on the queue.
Fixes: 9dee8b3b05 ("block, bfq: fix queue removal from weights tree")
Reported-by: Dmitrii Tcvetkov <demfloro@demfloro.ru>
Reported-by: Douglas Anderson <dianders@chromium.org>
Tested-by: Dmitrii Tcvetkov <demfloro@demfloro.ru>
Tested-by: Douglas Anderson <dianders@chromium.org>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Some of the comments in the bfq files had typos. This patch fixes them.
Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
bfq saves the state of a queue each time a merge occurs, to be
able to resume such a state when the queue is associated again
with its original process, on a split.
Unfortunately bfq does not save & restore also the weight of the
queue. If the weight is not correctly resumed when the queue is
recycled, then the weight of the recycled queue could differ
from the weight of the original queue.
This commit adds the missing save & resume of the weight.
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Francesco Pollicino <fra.fra.800@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The function "bfq_log_bfqq" prints the pid of the process
associated with the queue passed as input.
Unfortunately, if the queue is shared, then more than one process
is associated with the queue. The pid that gets printed in this
case is the pid of one of the associated processes.
Which process gets printed depends on the exact sequence of merge
events the queue underwent. So printing such a pid is rather
useless and above all is often rather confusing because it
reports a random pid between those of the associated processes.
This commit addresses this issue by printing SHARED instead of a pid
if the queue is shared.
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Francesco Pollicino <fra.fra.800@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If many bfq_queues belonging to the same group happen to be created
shortly after each other, then the processes associated with these
queues have typically a common goal. In particular, bursts of queue
creations are usually caused by services or applications that spawn
many parallel threads/processes. Examples are systemd during boot, or
git grep. If there are no other active queues, then, to help these
processes get their job done as soon as possible, the best thing to do
is to reach a high throughput. To this goal, it is usually better to
not grant either weight-raising or device idling to the queues
associated with these processes. And this is exactly what BFQ
currently does.
There is however a drawback: if, in contrast, some other queues are
already active, then the newly created queues must be protected from
the I/O flowing through the already existing queues. In this case, the
best thing to do is the opposite as in the other case: it is much
better to grant weight-raising and device idling to the newly-created
queues, if they deserve it. This commit addresses this issue by doing
so if there are already other active queues.
This change also helps eliminating false positives, which occur when
the newly-created queues do not belong to an actual large burst of
creations, but some background task (e.g., a service) happens to
trigger the creation of new queues in the middle, i.e., very close to
when the victim queues are created. These false positive may cause
total loss of control on process latencies.
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Sync random I/O is likely to be confused with soft real-time I/O,
because it is characterized by limited throughput and apparently
isochronous arrival pattern. To avoid false positives, this commits
prevents bfq_queues containing only random (seeky) I/O from being
tagged as soft real-time.
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
To boost throughput with a set of processes doing interleaved I/O
(i.e., a set of processes whose individual I/O is random, but whose
merged cumulative I/O is sequential), BFQ merges the queues associated
with these processes, i.e., redirects the I/O of these processes into a
common, shared queue. In the shared queue, I/O requests are ordered by
their position on the medium, thus sequential I/O gets dispatched to
the device when the shared queue is served.
Queue merging costs execution time, because, to detect which queues to
merge, BFQ must maintain a list of the head I/O requests of active
queues, ordered by request positions. Measurements showed that this
costs about 10% of BFQ's total per-request processing time.
Request processing time becomes more and more critical as the speed of
the underlying storage device grows. Yet, fortunately, queue merging
is basically useless on the very devices that are so fast to make
request processing time critical. To reach a high throughput, these
devices must have many requests queued at the same time. But, in this
configuration, the internal scheduling algorithms of these devices do
also the job of queue merging: they reorder requests so as to obtain
as much as possible a sequential I/O pattern. As a consequence, with
processes doing interleaved I/O, the throughput reached by one such
device is likely to be the same, with and without queue merging.
In view of this fact, this commit disables queue merging, and all
related housekeeping, for non-rotational devices with internal
queueing. The total, single-lock-protected, per-request processing
time of BFQ drops to, e.g., 1.9 us on an Intel Core i7-2760QM@2.40GHz
(time measured with simple code instrumentation, and using the
throughput-sync.sh script of the S suite [1], in performance-profiling
mode). To put this result into context, the total,
single-lock-protected, per-request execution time of the lightest I/O
scheduler available in blk-mq, mq-deadline, is 0.7 us (mq-deadline is
~800 LOC, against ~10500 LOC for BFQ).
Disabling merging provides a further, remarkable benefit in terms of
throughput. Merging tends to make many workloads artificially more
uneven, mainly because of shared queues remaining non empty for
incomparably more time than normal queues. So, if, e.g., one of the
queues in a set of merged queues has a higher weight than a normal
queue, then the shared queue may inherit such a high weight and, by
staying almost always active, may force BFQ to perform I/O plugging
most of the time. This evidently makes it harder for BFQ to let the
device reach a high throughput.
As a practical example of this problem, and of the benefits of this
commit, we measured again the throughput in the nasty scenario
considered in previous commit messages: dbench test (in the Phoronix
suite), with 6 clients, on a filesystem with journaling, and with the
journaling daemon enjoying a higher weight than normal processes. With
this commit, the throughput grows from ~150 MB/s to ~200 MB/s on a
PLEXTOR PX-256M5 SSD. This is the same peak throughput reached by any
of the other I/O schedulers. As such, this is also likely to be the
maximum possible throughput reachable with this workload on this
device, because I/O is mostly random, and the other schedulers
basically just pass I/O requests to the drive as fast as possible.
[1] https://github.com/Algodev-github/S
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Francesco Pollicino <fra.fra.800@gmail.com>
Signed-off-by: Alessio Masola <alessio.masola@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The processes associated with a bfq_queue, say Q, may happen to
generate their cumulative I/O at a lower rate than the rate at which
the device could serve the same I/O. This is rather probable, e.g., if
only one process is associated with Q and the device is an SSD. It
results in Q becoming often empty while in service. If BFQ is not
allowed to switch to another queue when Q becomes empty, then, during
the service of Q, there will be frequent "service holes", i.e., time
intervals during which Q gets empty and the device can only consume
the I/O already queued in its hardware queues. This easily causes
considerable losses of throughput.
To counter this problem, BFQ implements a request injection mechanism,
which tries to fill the above service holes with I/O requests taken
from other bfq_queues. The hard part in this mechanism is finding the
right amount of I/O to inject, so as to both boost throughput and not
break Q's bandwidth and latency guarantees. To this goal, the current
version of this mechanism measures the bandwidth enjoyed by Q while it
is being served, and tries to inject the maximum possible amount of
extra service that does not cause Q's bandwidth to decrease too
much.
This solution has an important shortcoming. For bandwidth measurements
to be stable and reliable, Q must remain in service for a much longer
time than that needed to serve a single I/O request. Unfortunately,
this does not hold with many workloads. This commit addresses this
issue by changing the way the amount of injection allowed is
dynamically computed. It tunes injection as a function of the service
times of single I/O requests of Q, instead of Q's
bandwidth. Single-request service times are evidently meaningful even
if Q gets very few I/O requests completed while it is in service.
As a testbed for this new solution, we measured the throughput reached
by BFQ for one of the nastiest workloads and configurations for this
scheduler: the workload generated by the dbench test (in the Phoronix
suite), with 6 clients, on a filesystem with journaling, and with the
journaling daemon enjoying a higher weight than normal processes.
With this commit, the throughput grows from ~100 MB/s to ~150 MB/s on
a PLEXTOR PX-256M5.
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Francesco Pollicino <fra.fra.800@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In most cases, it is detrimental for throughput to plug I/O dispatch
when the in-service bfq_queue becomes temporarily empty (plugging is
performed to wait for the possible arrival, soon, of new I/O from the
in-service queue). There is however a case where plugging is needed
for service guarantees. If a bfq_queue, say Q, has a higher weight
than some other active bfq_queue, and is sync, i.e., contains sync
I/O, then, to guarantee that Q does receive a higher share of the
throughput than other lower-weight queues, it is necessary to plug I/O
dispatch when Q remains temporarily empty while being served.
For this reason, BFQ performs I/O plugging when some active bfq_queue
has a higher weight than some other active bfq_queue. But this is
unnecessarily overkill. In fact, if the in-service bfq_queue actually
has a weight lower than or equal to the other queues, then the queue
*must not* be guaranteed a higher share of the throughput than the
other queues. So, not plugging I/O cannot cause any harm to the
queue. And can boost throughput.
Taking advantage of this fact, this commit does not plug I/O for sync
bfq_queues with a weight lower than or equal to the weights of the
other queues. Here is an example of the resulting throughput boost
with the dbench workload, which is particularly nasty for BFQ. With
the dbench test in the Phoronix suite, BFQ reaches its lowest total
throughput with 6 clients on a filesystem with journaling, in case the
journaling daemon has a higher weight than normal processes. Before
this commit, the total throughput was ~80 MB/sec on a PLEXTOR PX-256M5,
after this commit it is ~100 MB/sec.
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If a sync bfq_queue has a higher weight than some other queue, and
remains temporarily empty while in service, then, to preserve the
bandwidth share of the queue, it is necessary to plug I/O dispatching
until a new request arrives for the queue. In addition, a timeout
needs to be set, to avoid waiting for ever if the process associated
with the queue has actually finished its I/O.
Even with the above timeout, the device is however not fed with new
I/O for a while, if the process has finished its I/O. If this happens
often, then throughput drops and latencies grow. For this reason, the
timeout is kept rather low: 8 ms is the current default.
Unfortunately, such a low value may cause, on the opposite end, a
violation of bandwidth guarantees for a process that happens to issue
new I/O too late. The higher the system load, the higher the
probability that this happens to some process. This is a problem in
scenarios where service guarantees matter more than throughput. One
important case are weight-raised queues, which need to be granted a
very high fraction of the bandwidth.
To address this issue, this commit lower-bounds the plugging timeout
for weight-raised queues to 20 ms. This simple change provides
relevant benefits. For example, on a PLEXTOR PX-256M5S, with which
gnome-terminal starts in 0.6 seconds if there is no other I/O in
progress, the same applications starts in
- 0.8 seconds, instead of 1.2 seconds, if ten files are being read
sequentially in parallel
- 1 second, instead of 2 seconds, if, in parallel, five files are
being read sequentially, and five more files are being written
sequentially
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When a new I/O request arrives for a bfq_queue, say Q, bfq checks
whether that request is close to
(a) the head request of some other queue waiting to be served, or
(b) the last request dispatched for the in-service queue (in case Q
itself is not the in-service queue)
If a queue, say Q2, is found for which the above condition holds, then
bfq merges Q and Q2, to hopefully get a more sequential I/O in the
resulting merged queue, and thus a possibly higher throughput.
Case (b) is checked by comparing the new request for Q with the last
request dispatched, assuming that the latter necessarily belonged to the
in-service queue. Unfortunately, this assumption is no longer always
correct, since commit d0edc2473b ("block, bfq: inject other-queue I/O
into seeky idle queues on NCQ flash").
When the assumption does not hold, queues that must not be merged may be
merged, causing unexpected loss of control on per-queue service
guarantees.
This commit solves this problem by adding an extra field, which stores
the actual last request dispatched for the in-service queue, and by
using this new field to correctly check case (b).
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Writes tend to starve reads. bfq counters this problem by overcharging
writes with an inflated service w.r.t. the actual service (number of
sector written) they receive.
Yet his overcharging is useless, and actually causes unfairness in the
opposite direction, when bfq happens to be enforcing strong I/O control.
bfq does this enforcing when the scenario is asymmetric, i.e., when some
bfq_queue or group of bfq_queues is to be granted a different bandwidth
than some other bfq_queue or group of bfq_queues. So, in such a
scenario, this commit disables write overcharging.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The original commit is commit 1a1238a7dd ("cfq-iosched: improve hw_tag
detection") and has the following commit message:
If active queue hasn't enough requests and idle window opens, cfq will
not dispatch sufficient requests to hardware. In such situation, current
code will zero hw_tag. But this is because cfq doesn't dispatch enough
requests instead of hardware queue doesn't work. Don't zero hw_tag in
such case.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
bfq simple heuristic from cfq for detecting whether the drive performs
command queueing: check whether the average number of in-flight requests
is above a given threshold. Unfortunately this heuristic does fail to
detect queueing (on drives with queueing) if processes doing I/O are few
and issue I/O with a low depth.
To reduce false negatives, this commit lowers the threshold.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
bfq maintains an ordered list, through a red-black tree, of unique
weights of active bfq_queues. This list is used to detect whether there
are active queues with differentiated weights. The weight of a queue is
removed from the list when both the following two conditions become
true:
(1) the bfq_queue is flagged as inactive
(2) the has no in-flight request any longer;
Unfortunately, in the rare cases where condition (2) becomes true before
condition (1), the removal fails, because the function to remove the
weight of the queue (bfq_weights_tree_remove) is rightly invoked in the
path that deactivates the bfq_queue, but mistakenly invoked *before* the
function that actually performs the deactivation (bfq_deactivate_bfqq).
This commits moves the invocation of bfq_weights_tree_remove for
condition (1) to after bfq_deactivate_bfqq. As a consequence of this
move, it is necessary to add a further reference to the queue when the
weight of a queue is added, because the queue might otherwise be freed
before bfq_weights_tree_remove is invoked. This commit adds this
reference and makes all related modifications.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In bfq_update_peak_rate, to check whether an I/O request rq is
sequential, only the seek distance of rq w.r.t. the last request
dispatched is controlled. This is not sufficient for non-rotational
storage, where the size of rq is at least as relevant. This commit adds
the missing control.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
bfq detects the creation of multiple bfq_queues shortly after each
other, namely a burst of queue creations in the terminology used in the
code. If the burst is large, then no queue in the burst is granted
- either I/O-dispatch plugging when the queue remains temporarily idle
while in service;
- or weight raising, because it causes even longer plugging.
In fact, such a plugging tends to lower throughput, while these bursts
are typically due to applications or services that spawn multiple
processes, to reach a common goal as soon as possible. Examples are a
"git grep" or the booting of a system.
Unfortunately, disabling plugging may cause a loss of service guarantees
in asymmetric scenarios, i.e., if queue weights are differentiated or if
more than one group is active.
This commit addresses this issue by no longer disabling I/O-dispatch
plugging for queues in large bursts.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If the in-service bfq_queue is sync and remains temporarily idle, then
I/O dispatching (from other queues) may be plugged. It may be dome for
two reasons: either to boost throughput, or to preserve the bandwidth
share of the in-service queue. In the first case, if the I/O of the
in-service queue, when it finally arrives, consists only of one small
I/O request, then it makes sense to plug even the I/O of the in-service
queue. In fact, serving such a small request immediately is likely to
lower throughput instead of boosting it, whereas waiting a little bit is
likely to let that request grow, thanks to request merging, and become
more profitable in terms of throughput (this is likely to happen exactly
because the I/O of the queue has been detected to boost throughput).
On the opposite end, if I/O dispatching is being plugged only to
preserve the bandwidth of the in-service queue, then it would be better
not to plug also the I/O of the in-service queue, because such a
plugging is likely to cause only loss of bandwidth for the queue.
Unfortunately, no distinction is made between the two cases, and the I/O
of the in-service queue is always plugged in case just a small I/O
request arrives. This commit draws this missing distinction and does not
perform harmful plugging.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This is a preparatory commit for commits that need to check only one of
the two main reasons for idling. This change should also improve the
quality of the code a little bit, by splitting a function that contains
very long, non-trivial and little related comments.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In asymmetric scenarios, i.e., when some bfq_queue or bfq_group needs to
be guaranteed a different bandwidth than other bfq_queues or bfq_groups,
these service guaranteed can be provided only by plugging I/O dispatch,
completely or partially, when the queue in service remains temporarily
empty. A case where asymmetry is particularly strong is when some active
bfq_queues belong to a higher-priority class than some other active
bfq_queues. Unfortunately, this important case is not considered at all
in the code for detecting asymmetric scenarios. This commit adds the
missing logic.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Before commit 18e5a57d79 ("block, bfq: postpone rq preparation to
insert or merge"), the destination queue for a request was chosen by a
different hook than the one that then inserted the request. So, between
the execution of the two hooks, the bic of the process generating the
request could happen to be redirected to a different bfq_queue. As a
consequence, the destination bfq_queue stored in the request could be
wrong. Such an event does not need to ba handled any longer.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With some unlucky sequences of events, the function bfq_updated_next_req
updates the current budget of a bfq_queue to a lower value than the
service received by the queue using such a budget. Unfortunately, if
this happens, then the return value of the function bfq_bfqq_budget_left
becomes inconsistent. This commit solves this problem by lower-bounding
the budget computed in bfq_updated_next_req to the service currently
charged to the queue.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
To boost throughput on devices with internal queueing and in scenarios
where device idling is not strictly needed, bfq immediately starts
serving a new bfq_queue if the in-service bfq_queue remains without
pending I/O, even if new I/O may arrive soon for the latter queue. Then,
if such I/O actually arrives soon, bfq preempts the new in-service
bfq_queue so as to give the previous queue a chance to go on being
served (in case the previous queue should actually be the one to be
served, according to its timestamps).
However, the in-service bfq_queue, say Q, may also be without further
budget when it remains also pending I/O. Since bfq changes budgets
dynamically to fit the needs of bfq_queues, this happens more often than
one may expect. If this happens, then there is no point in trying to go
on serving Q when new I/O arrives for it soon: Q would be expired
immediately after being selected for service. This would only cause
useless overhead. This commit avoids such a useless selection.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The speed at which a bfq_queue receives I/O is one of the parameters by
which bfq decides whether the queue is soft real-time (i.e., whether the
queue contains the I/O of a soft real-time application). In particular,
when a bfq_queue remains without outstanding I/O requests, bfq computes
the minimum time instant, named soft_rt_next_start, at which the next
request of the queue may arrive for the queue to be deemed as soft real
time.
Unfortunately this filtering may cause problems with a queue in
interactive weight raising. In fact, such a queue may be conveying the
I/O needed to load a soft real-time application. The latter will
actually exhibit a soft real-time I/O pattern after it finally starts
doing its job. But, if soft_rt_next_start is updated for an interactive
bfq_queue, and the queue has received a lot of service before remaining
with no outstanding request (likely to happen on a fast device), then
soft_rt_next_start is assigned such a high value that, for a very long
time, the queue is prevented from being possibly considered as soft real
time.
This commit removes the updating of soft_rt_next_start for bfq_queues in
interactive weight raising.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
-----BEGIN PGP SIGNATURE-----
iQFSBAABCAA8FiEEq68RxlopcLEwq+PEeb4+QwBBGIYFAlwNpb0eHHRvcnZhbGRz
QGxpbnV4LWZvdW5kYXRpb24ub3JnAAoJEHm+PkMAQRiGwGwH/00UHnXfxww3ixxz
zwTVDzptA6SPm6s84yJOWatM5fXhPiAltZaHSYF9lzRzNU71NCq7Frhq3fQUIXKM
OxqDn9nfSTWcjWTk2q5n2keyRV/KIn67YX7UgqFc1bO/mqtVjEgNWaMyblhI+e9E
giu1ZXayHr43jK1cDOmGExZubXUq7Vsc9TOlrd+d2SwIqeEP7TCMrPhnHDwCNvX2
UU5dtANpVzGtHaBcr37wJj+L8kODCc0f+PQ3g2ar5jTHst5SLlHp2u0AMRnUmgdi
VkGx+mu/uk8mtwUqMIMqhplklVoqK6LTeLqsY5Xt32SKruw9UqyJGdphLjW2QP/g
MkmA1lI=
=7kaD
-----END PGP SIGNATURE-----
Merge tag 'v4.20-rc6' into for-4.21/block
Pull in v4.20-rc6 to resolve the conflict in NVMe, but also to get the
two corruption fixes. We're going to be overhauling the direct dispatch
path, and we need to do that on top of the changes we made for that
in mainline.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The bio_blkcg() function turns out to be inconsistent and consequently
dangerous to use. The first part returns a blkcg where a reference is
owned by the bio meaning it does not need to be rcu protected. However,
the third case, the last line, is problematic:
return css_to_blkcg(task_css(current, io_cgrp_id));
This can race against task migration and the cgroup dying. It is also
semantically different as it must be called rcu protected and is
susceptible to failure when trying to get a reference to it.
This patch adds association ahead of calling bio_blkcg() rather than
after. This makes association a required and explicit step along the
code paths for calling bio_blkcg(). In blk-iolatency, association is
moved above the bio_blkcg() call to ensure it will not return %NULL.
BFQ uses the old bio_blkcg() function, but I do not want to address it
in this series due to the complexity. I have created a private version
documenting the inconsistency and noting not to use it.
Signed-off-by: Dennis Zhou <dennis@kernel.org>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since commit '2d29c9f89fcd ("block, bfq: improve asymmetric scenarios
detection")', if there are process groups with I/O requests waiting for
completion, then BFQ tags the scenario as 'asymmetric'. This detection
is needed for preserving service guarantees (for details, see comments
on the computation * of the variable asymmetric_scenario in the
function bfq_better_to_idle).
Unfortunately, commit '2d29c9f89fcd ("block, bfq: improve asymmetric
scenarios detection")' contains an error exactly in the updating of
the number of groups with I/O requests waiting for completion: if a
group has more than one descendant process, then the above number of
groups, which is renamed from num_active_groups to a more appropriate
num_groups_with_pending_reqs by this commit, may happen to be wrongly
decremented multiple times, namely every time one of the descendant
processes gets all its pending I/O requests completed.
A correct, complete solution should work as follows. Consider a group
that is inactive, i.e., that has no descendant process with pending
I/O inside BFQ queues. Then suppose that num_groups_with_pending_reqs
is still accounting for this group, because the group still has some
descendant process with some I/O request still in
flight. num_groups_with_pending_reqs should be decremented when the
in-flight request of the last descendant process is finally completed
(assuming that nothing else has changed for the group in the meantime,
in terms of composition of the group and active/inactive state of
child groups and processes). To accomplish this, an additional
pending-request counter must be added to entities, and must be
updated correctly.
To avoid this additional field and operations, this commit resorts to
the following tradeoff between simplicity and accuracy: for an
inactive group that is still counted in num_groups_with_pending_reqs,
this commit decrements num_groups_with_pending_reqs when the first
descendant process of the group remains with no request waiting for
completion.
This simplified scheme provides a fix to the unbalanced decrements
introduced by 2d29c9f89f. Since this error was also caused by lack
of comments on this non-trivial issue, this commit also adds related
comments.
Fixes: 2d29c9f89f ("block, bfq: improve asymmetric scenarios detection")
Reported-by: Steven Barrett <steven@liquorix.net>
Tested-by: Steven Barrett <steven@liquorix.net>
Tested-by: Lucjan Lucjanov <lucjan.lucjanov@gmail.com>
Reviewed-by: Federico Motta <federico@willer.it>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With the legacy request path gone there is no good reason to keep
queue_lock as a pointer, we can always use the embedded lock now.
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Fixed floppy and blk-cgroup missing conversions and half done edits.
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>
bfq defines as asymmetric a scenario where an active entity, say E
(representing either a single bfq_queue or a group of other entities),
has a higher weight than some other entities. If the entity E does sync
I/O in such a scenario, then bfq plugs the dispatch of the I/O of the
other entities in the following situation: E is in service but
temporarily has no pending I/O request. In fact, without this plugging,
all the times that E stops being temporarily idle, it may find the
internal queues of the storage device already filled with an
out-of-control number of extra requests, from other entities. So E may
have to wait for the service of these extra requests, before finally
having its own requests served. This may easily break service
guarantees, with E getting less than its fair share of the device
throughput. Usually, the end result is that E gets the same fraction of
the throughput as the other entities, instead of getting more, according
to its higher weight.
Yet there are two other more subtle cases where E, even if its weight is
actually equal to or even lower than the weight of any other active
entities, may get less than its fair share of the throughput in case the
above I/O plugging is not performed:
1. other entities issue larger requests than E;
2. other entities contain more active child entities than E (or in
general tend to have more backlog than E).
In the first case, other entities may get more service than E because
they get larger requests, than those of E, served during the temporary
idle periods of E. In the second case, other entities get more service
because, by having many child entities, they have many requests ready
for dispatching while E is temporarily idle.
This commit addresses this issue by extending the definition of
asymmetric scenario: a scenario is asymmetric when
- active entities representing bfq_queues have differentiated weights,
as in the original definition
or (inclusive)
- one or more entities representing groups of entities are active.
This broader definition makes sure that I/O plugging will be performed
in all the above cases, provided that there is at least one active
group. Of course, this definition is very coarse, so it will trigger
I/O plugging also in cases where it is not needed, such as, e.g.,
multiple active entities with just one child each, and all with the same
I/O-request size. The reason for this coarse definition is just that a
finer-grained definition would be rather heavy to compute.
On the opposite end, even this new definition does not trigger I/O
plugging in all cases where there is no active group, and all bfq_queues
have the same weight. So, in these cases some unfairness may occur if
there are asymmetries in I/O-request sizes. We made this choice because
I/O plugging may lower throughput, and probably a user that has not
created any group cares more about throughput than about perfect
fairness. At any rate, as for possible applications that may care about
service guarantees, bfq already guarantees a high responsiveness and a
low latency to soft real-time applications automatically.
Signed-off-by: Federico Motta <federico@willer.it>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>