On scheduler init, a reference to the root group, and a reference to
its corresponding blkg are taken for the oom queue. Yet these
references are not released on scheduler exit, which prevents these
objects from be freed. This commit adds the missing reference
releases.
Reported-by: Davide Ferrari <davideferrari8@gmail.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Commit a33801e8b4 ("block, bfq: move debug blkio stats behind
CONFIG_DEBUG_BLK_CGROUP") introduced two batches of confusing ifdefs:
one reported in [1], plus a similar one in another function. This
commit removes both batches, in the way suggested in [1].
[1] https://www.spinics.net/lists/linux-block/msg20043.html
Fixes: a33801e8b4 ("block, bfq: move debug blkio stats behind CONFIG_DEBUG_BLK_CGROUP")
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Tested-by: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
BFQ privileges the I/O of soft real-time applications, such as video
players, to guarantee to these application a high bandwidth and a low
latency. In this respect, it is not easy to correctly detect when an
application is soft real-time. A particularly nasty false positive is
that of an I/O-bound application that occasionally happens to meet all
requirements to be deemed as soft real-time. After being detected as
soft real-time, such an application monopolizes the device. Fortunately,
BFQ will realize soon that the application is actually not soft
real-time and suspend every privilege. Yet, the application may happen
again to be wrongly detected as soft real-time, and so on.
As highlighted by our tests, this problem causes BFQ to occasionally
fail to guarantee a high responsiveness, in the presence of heavy
background I/O workloads. The reason is that the background workload
happens to be detected as soft real-time, more or less frequently,
during the execution of the interactive task under test. To give an
idea, because of this problem, Libreoffice Writer occasionally takes 8
seconds, instead of 3, to start up, if there are sequential reads and
writes in the background, on a Kingston SSDNow V300.
This commit addresses this issue by leveraging the following facts.
The reason why some applications are detected as soft real-time despite
all BFQ checks to avoid false positives, is simply that, during high
CPU or storage-device load, I/O-bound applications may happen to do
I/O slowly enough to meet all soft real-time requirements, and pass
all BFQ extra checks. Yet, this happens only for limited time periods:
slow-speed time intervals are usually interspersed between other time
intervals during which these applications do I/O at a very high speed.
To exploit these facts, this commit introduces a little change, in the
detection of soft real-time behavior, to systematically consider also
the recent past: the higher the speed was in the recent past, the
later next I/O should arrive for the application to be considered as
soft real-time. At the beginning of a slow-speed interval, the minimum
arrival time allowed for the next I/O usually happens to still be so
high, to fall *after* the end of the slow-speed period itself. As a
consequence, the application does not risk to be deemed as soft
real-time during the slow-speed interval. Then, during the next
high-speed interval, the application cannot, evidently, be deemed as
soft real-time (exactly because of its speed), and so on.
This extra filtering proved to be rather effective: in the above test,
the frequency of false positives became so low that the start-up time
was 3 seconds in all iterations (apart from occasional outliers,
caused by page-cache-management issues, which are out of the scope of
this commit, and cannot be solved by an I/O scheduler).
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When two or more processes do I/O in a way that the their requests are
sequential in respect to one another, BFQ merges the bfq_queues associated
with the processes. This way the overall I/O pattern becomes sequential,
and thus there is a boost in througput.
These cooperating processes usually start or restart to do I/O shortly
after each other. So, in order to avoid merging non-cooperating processes,
BFQ ensures that none of these queues has been in weight raising for too
long.
In this respect, from commit "block, bfq-sq, bfq-mq: let a queue be merged
only shortly after being created", BFQ checks whether any queue (and not
only weight-raised ones) is doing I/O continuously from too long to be
merged.
This new additional check makes the first one useless: a queue doing
I/O from long enough, if being weight-raised, is also a queue in
weight raising for too long to be merged. Accordingly, this commit
removes the first check.
Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In BFQ and CFQ, two processes are said to be cooperating if they do
I/O in such a way that the union of their I/O requests yields a
sequential I/O pattern. To get such a sequential I/O pattern out of
the non-sequential pattern of each cooperating process, BFQ and CFQ
merge the queues associated with these processes. In more detail,
cooperating processes, and thus their associated queues, usually
start, or restart, to do I/O shortly after each other. This is the
case, e.g., for the I/O threads of KVM/QEMU and of the dump
utility. Basing on this assumption, this commit allows a bfq_queue to
be merged only during a short time interval (100ms) after it starts,
or re-starts, to do I/O. This filtering provides two important
benefits.
First, it greatly reduces the probability that two non-cooperating
processes have their queues merged by mistake, if they just happen to
do I/O close to each other for a short time interval. These spurious
merges cause loss of service guarantees. A low-weight bfq_queue may
unjustly get more than its expected share of the throughput: if such a
low-weight queue is merged with a high-weight queue, then the I/O for
the low-weight queue is served as if the queue had a high weight. This
may damage other high-weight queues unexpectedly. For instance,
because of this issue, lxterminal occasionally took 7.5 seconds to
start, instead of 6.5 seconds, when some sequential readers and
writers did I/O in the background on a FUJITSU MHX2300BT HDD. The
reason is that the bfq_queues associated with some of the readers or
the writers were merged with the high-weight queues of some processes
that had to do some urgent but little I/O. The readers then exploited
the inherited high weight for all or most of their I/O, during the
start-up of terminal. The filtering introduced by this commit
eliminated any outlier caused by spurious queue merges in our start-up
time tests.
This filtering also provides a little boost of the throughput
sustainable by BFQ: 3-4%, depending on the CPU. The reason is that,
once a bfq_queue cannot be merged any longer, this commit makes BFQ
stop updating the data needed to handle merging for the queue.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A just-created bfq_queue will certainly be deemed as interactive on
the arrival of its first I/O request, if the low_latency flag is
set. Yet, if the queue is merged with another queue on the arrival of
its first I/O request, it will not have the chance to be flagged as
interactive. Nevertheless, if the queue is then split soon enough, it
has to be flagged as interactive after the split.
To handle this early-merge scenario correctly, BFQ saves the state of
the queue, on the merge, as if the latter had already been deemed
interactive. So, if the queue is split soon, it will get
weight-raised, because the previous state of the queue is resumed on
the split.
Unfortunately, in the act of saving the state of the newly-created
queue, BFQ doesn't check whether the low_latency flag is set, and this
causes early-merged queues to be then weight-raised, on queue splits,
even if low_latency is off. This commit addresses this problem by
adding the missing check.
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>
If two processes do I/O close to each other, then BFQ merges the
bfq_queues associated with these processes, to get a more sequential
I/O, and thus a higher throughput. In this respect, to detect whether
two processes are doing I/O close to each other, BFQ keeps a list of
the head-of-line I/O requests of all active bfq_queues. The list is
ordered by initial sectors, and implemented through a red-black tree
(rq_pos_tree).
Unfortunately, the update of the rq_pos_tree was incomplete, because
the tree was not updated on the removal of the head-of-line I/O
request of a bfq_queue, in case the queue did not remain empty. This
commit adds the missing update.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If two processes do I/O close to each other, i.e., are cooperating
processes in BFQ (and CFQ'S) nomenclature, then BFQ merges their
associated bfq_queues, so as to get sequential I/O from the union of
the I/O requests of the processes, and thus reach a higher
throughput. A merged queue is then split if its I/O stops being
sequential. In this respect, BFQ deems the I/O of a bfq_queue as
(mostly) sequential only if less than 4 I/O requests are random, out
of the last 32 requests inserted into the queue.
Unfortunately, extensive testing (with the interleaved_io benchmark of
the S suite [1], and with real applications spawning cooperating
processes) has clearly shown that, with such a low threshold, only a
rather low I/O throughput may be reached when several cooperating
processes do I/O. In particular, the outcome of each test run was
bimodal: if queue merging occurred and was stable during the test,
then the throughput was close to the peak rate of the storage device,
otherwise the throughput was arbitrarily low (usually around 1/10 of
the peak rate with a rotational device). The probability to get the
unlucky outcomes grew with the number of cooperating processes: it was
already significant with 5 processes, and close to one with 7 or more
processes.
The cause of the low throughput in the unlucky runs was that the
merged queues containing the I/O of these cooperating processes were
soon split, because they contained more random I/O requests than those
tolerated by the 4/32 threshold, but
- that I/O would have however allowed the storage device to reach
peak throughput or almost peak throughput;
- in contrast, the I/O of these processes, if served individually
(from separate queues) yielded a rather low throughput.
So we repeated our tests with increasing values of the threshold,
until we found the minimum value (19) for which we obtained maximum
throughput, reliably, with at least up to 9 cooperating
processes. Then we checked that the use of that higher threshold value
did not cause any regression for any other benchmark in the suite [1].
This commit raises the threshold to such a higher value.
[1] https://github.com/Algodev-github/S
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 currently creates, and updates, its own instance of the whole
set of blkio statistics that cfq creates. Yet, from the comments
of Tejun Heo in [1], it turned out that most of these statistics
are meant/useful only for debugging. This commit makes BFQ create
the latter, debugging statistics only if the option
CONFIG_DEBUG_BLK_CGROUP is set.
By doing so, this commit also enables BFQ to enjoy a high perfomance
boost. The reason is that, if CONFIG_DEBUG_BLK_CGROUP is not set, then
BFQ has to update far fewer statistics, and, in particular, not the
heaviest to update. To give an idea of the benefits, if
CONFIG_DEBUG_BLK_CGROUP is not set, then, on an Intel i7-4850HQ, and
with 8 threads doing random I/O in parallel on null_blk (configured
with 0 latency), the throughput of BFQ grows from 310 to 400 KIOPS
(+30%). We have measured similar or even much higher boosts with other
CPUs: e.g., +45% with an ARM CortexTM-A53 Octa-core. Our results have
been obtained and can be reproduced very easily with the script in [1].
[1] https://www.spinics.net/lists/linux-block/msg18943.html
Suggested-by: Tejun Heo <tj@kernel.org>
Suggested-by: Ulf Hansson <ulf.hansson@linaro.org>
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
bfq invokes various blkg_*stats_* functions to update the statistics
contained in the special files blkio.bfq.* in the blkio controller
groups, i.e., the I/O accounting related to the proportional-share
policy provided by bfq. The execution of these functions takes a
considerable percentage, about 40%, of the total per-request execution
time of bfq (i.e., of the sum of the execution time of all the bfq
functions that have to be executed to process an I/O request from its
creation to its destruction). This reduces the request-processing
rate sustainable by bfq noticeably, even on a multicore CPU. In fact,
the bfq functions that invoke blkg_*stats_* functions cannot be
executed in parallel with the rest of the code of bfq, because both
are executed under the same same per-device scheduler lock.
To reduce this slowdown, this commit moves, wherever possible, the
invocation of these functions (more precisely, of the bfq functions
that invoke blkg_*stats_* functions) outside the critical sections
protected by the scheduler lock.
With this change, and with all blkio.bfq.* statistics enabled, the
throughput grows, e.g., from 250 to 310 KIOPS (+25%) on an Intel
i7-4850HQ, in case of 8 threads doing random I/O in parallel on
null_blk, with the latter configured with 0 latency. We obtained the
same or higher throughput boosts, up to +30%, with other processors
(some figures are reported in the documentation). For our tests, we
used the script [1], with which our results can be easily reproduced.
NOTE. This commit still protects the invocation of blkg_*stats_*
functions with the request_queue lock, because the group these
functions are invoked on may otherwise disappear before or while these
functions are executed. Fortunately, tests without even this lock
show, by difference, that the serialization caused by this lock has a
little impact (at most ~5% of throughput reduction).
[1] https://github.com/Algodev-github/IOSpeed
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
bfqg_stats_update_io_add and bfqg_stats_update_io_remove are to be
invoked, respectively, when an I/O request enters and when an I/O
request exits the scheduler. Unfortunately, bfq does not fully comply
with this scheme, because it does not invoke these functions for
requests that are inserted into or extracted from its priority
dispatch list. This commit fixes this mistake.
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The commit "block, bfq: decrease burst size when queues in burst
exit" introduced the decrement of burst_size on the removal of a
bfq_queue from the burst list. Unfortunately, this decrement can
happen to be performed even when burst size is already equal to 0,
because of unbalanced decrements. A description follows of the cause
of these unbalanced decrements, namely a wrong assumption, and of the
way how this wrong assumption leads to unbalanced decrements.
The wrong assumption is that a bfq_queue can exit only if the process
associated with the bfq_queue has exited. This is false, because a
bfq_queue, say Q, may exit also as a consequence of a merge with
another bfq_queue. In this case, Q exits because the I/O of its
associated process has been redirected to another bfq_queue.
The decrement unbalance occurs because Q may then be re-created after
a split, and added back to the current burst list, *without*
incrementing burst_size. burst_size is not incremented because Q is
not a new bfq_queue added to the burst list, but a bfq_queue only
temporarily removed from the list, and, before the commit "bfq-sq,
bfq-mq: decrease burst size when queues in burst exit", burst_size was
not decremented when Q was removed.
This commit addresses this issue by just checking whether the exiting
bfq_queue is a merged bfq_queue, and, in that case, not decrementing
burst_size. Unfortunately, this still leaves room for unbalanced
decrements, in the following rarer case: on a split, the bfq_queue
happens to be inserted into a different burst list than that it was
removed from when merged. If this happens, the number of elements in
the new burst list becomes higher than burst_size (by one). When the
bfq_queue then exits, it is of course not in a merged state any
longer, thus burst_size is decremented, which results in an unbalanced
decrement. To handle this sporadic, unlucky case in a simple way,
this commit also checks that burst_size is larger than 0 before
decrementing it.
Finally, this commit removes an useless, extra check: the check that
the bfq_queue is sync, performed before checking whether the bfq_queue
is in the burst list. This extra check is redundant, because only sync
bfq_queues can be inserted into the burst list.
Fixes: 7cb04004fa ("block, bfq: decrease burst size when queues in burst exit")
Reported-by: Philip Müller <philm@manjaro.org>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com>
Tested-by: Philip Müller <philm@manjaro.org>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Similarly to CFQ, BFQ has its write-throttling heuristics, and it
is better not to combine them with further write-throttling
heuristics of a different nature.
So this commit disables write-back throttling for a device if BFQ
is used as I/O scheduler for that device.
Signed-off-by: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If many queues belonging to the same group happen to be created
shortly after each other, then the concurrent processes associated
with these queues have typically a common goal, and they get it done
as soon as possible if not hampered by device idling. Examples are
processes spawned by git grep, or by systemd during boot. As for
device idling, this mechanism is currently necessary for weight
raising to succeed in its goal: privileging I/O. In view of these
facts, BFQ does not provide the above queues with either weight
raising or device idling.
On the other hand, a burst of queue creations may be caused also by
the start-up of a complex application. In this case, these queues need
usually to be served one after the other, and as quickly as possible,
to maximise responsiveness. Therefore, in this case the best strategy
is to weight-raise all the queues created during the burst, i.e., the
exact opposite of the strategy for the above case.
To distinguish between the two cases, BFQ uses an empirical burst-size
threshold, found through extensive tests and monitoring of daily
usage. Only large bursts, i.e., burst with a size above this
threshold, are considered as generated by a high number of parallel
processes. In this respect, upstart-based boot proved to be rather
hard to detect as generating a large burst of queue creations, because
with upstart most of the queues created in a burst exit *before* the
next queues in the same burst are created. To address this issue, I
changed the burst-detection mechanism so as to not decrease the size
of the current burst even if one of the queues in the burst is
eliminated.
Unfortunately, this missing decrease causes false positives on very
fast systems: on the start-up of a complex application, such as
libreoffice writer, so many queues are created, served and exited
shortly after each other, that a large burst of queue creations is
wrongly detected as occurring. These false positives just disappear if
the size of a burst is decreased when one of the queues in the burst
exits. This commit restores the missing burst-size decrease, relying
of the fact that upstart is apparently unlikely to be used on systems
running this and future versions of the kernel.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Mauro Andreolini <mauro.andreolini@unimore.it>
Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com>
Tested-by: Mirko Montanari <mirkomontanari91@gmail.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A just-created bfq_queue, say Q, may happen to be merged with another
bfq_queue on the very first invocation of the function
__bfq_insert_request. In such a case, even if Q would clearly deserve
interactive weight raising (as it has just been created), the function
bfq_add_request does not make it to be invoked for Q, and thus to
activate weight raising for Q. As a consequence, when the state of Q
is saved for a possible future restore, after a split of Q from the
other bfq_queue(s), such a state happens to be (unjustly)
non-weight-raised. Then the bfq_queue will not enjoy any weight
raising on the split, even if should still be in an interactive
weight-raising period when the split occurs.
This commit solves this problem as follows, for a just-created
bfq_queue that is being early-merged: it stores directly, in the saved
state of the bfq_queue, the weight-raising state that would have been
assigned to the bfq_queue if not early-merged.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Angelo Ruocco <angeloruocco90@gmail.com>
Tested-by: Mirko Montanari <mirkomontanari91@gmail.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
As already explained in the message of commit "block, bfq: fix
wrong init of saved start time for weight raising", if a soft
real-time weight-raising period happens to be nested in a larger
interactive weight-raising period, then BFQ restores the interactive
weight raising at the end of the soft real-time weight raising. In
particular, BFQ checks whether the latter has ended only on request
dispatches.
Unfortunately, the above scheme fails to restore interactive weight
raising in the following corner case: if a bfq_queue, say Q,
1) Is merged with another bfq_queue while it is in a nested soft
real-time weight-raising period. The weight-raising state of Q is
then saved, and not considered any longer until a split occurs.
2) Is split from the other bfq_queue(s) at a time instant when its
soft real-time weight raising is already finished.
On the split, while resuming the previous, soft real-time
weight-raised state of the bfq_queue Q, BFQ checks whether the
current soft real-time weight-raising period is actually over. If so,
BFQ switches weight raising off for Q, *without* checking whether the
soft real-time period was actually nested in a non-yet-finished
interactive weight-raising period.
This commit addresses this issue by adding the above missing check in
bfq_queue splits, and restoring interactive weight raising if needed.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Angelo Ruocco <angeloruocco90@gmail.com>
Tested-by: Mirko Montanari <mirkomontanari91@gmail.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This commit fixes a bug that causes bfq to fail to guarantee a high
responsiveness on some drives, if there is heavy random read+write I/O
in the background. More precisely, such a failure allowed this bug to
be found [1], but the bug may well cause other yet unreported
anomalies.
BFQ raises the weight of the bfq_queues associated with soft real-time
applications, to privilege the I/O, and thus reduce latency, for these
applications. This mechanism is named soft-real-time weight raising in
BFQ. A soft real-time period may happen to be nested into an
interactive weight raising period, i.e., it may happen that, when a
bfq_queue switches to a soft real-time weight-raised state, the
bfq_queue is already being weight-raised because deemed interactive
too. In this case, BFQ saves in a special variable
wr_start_at_switch_to_srt, the time instant when the interactive
weight-raising period started for the bfq_queue, i.e., the time
instant when BFQ started to deem the bfq_queue interactive. This value
is then used to check whether the interactive weight-raising period
would still be in progress when the soft real-time weight-raising
period ends. If so, interactive weight raising is restored for the
bfq_queue. This restore is useful, in particular, because it prevents
bfq_queues from losing their interactive weight raising prematurely,
as a consequence of spurious, short-lived soft real-time
weight-raising periods caused by wrong detections as soft real-time.
If, instead, a bfq_queue switches to soft-real-time weight raising
while it *is not* already in an interactive weight-raising period,
then the variable wr_start_at_switch_to_srt has no meaning during the
following soft real-time weight-raising period. Unfortunately the
handling of this case is wrong in BFQ: not only the variable is not
flagged somehow as meaningless, but it is also set to the time when
the switch to soft real-time weight-raising occurs. This may cause an
interactive weight-raising period to be considered mistakenly as still
in progress, and thus a spurious interactive weight-raising period to
start for the bfq_queue, at the end of the soft-real-time
weight-raising period. In particular the spurious interactive
weight-raising period will be considered as still in progress, if the
soft-real-time weight-raising period does not last very long. The
bfq_queue will then be wrongly privileged and, if I/O bound, will
unjustly steal bandwidth to truly interactive or soft real-time
bfq_queues, harming responsiveness and low latency.
This commit fixes this issue by just setting wr_start_at_switch_to_srt
to minus infinity (farthest past time instant according to jiffies
macros): when the soft-real-time weight-raising period ends, certainly
no interactive weight-raising period will be considered as still in
progress.
[1] Background I/O Type: Random - Background I/O mix: Reads and writes
- Application to start: LibreOffice Writer in
http://www.phoronix.com/scan.php?page=news_item&px=Linux-4.13-IO-Laptop
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Tested-by: Mirko Montanari <mirkomontanari91@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull followup block layer updates from Jens Axboe:
"I ended up splitting the main pull request for this series into two,
mainly because of clashes between NVMe fixes that went into 4.13 after
the for-4.14 branches were split off. This pull request is mostly
NVMe, but not exclusively. In detail, it contains:
- Two pull request for NVMe changes from Christoph. Nothing new on
the feature front, basically just fixes all over the map for the
core bits, transport, rdma, etc.
- Series from Bart, cleaning up various bits in the BFQ scheduler.
- Series of bcache fixes, which has been lingering for a release or
two. Coly sent this in, but patches from various people in this
area.
- Set of patches for BFQ from Paolo himself, updating both
documentation and fixing some corner cases in performance.
- Series from Omar, attempting to now get the 4k loop support
correct. Our confidence level is higher this time.
- Series from Shaohua for loop as well, improving O_DIRECT
performance and fixing a use-after-free"
* 'for-4.14/block-postmerge' of git://git.kernel.dk/linux-block: (74 commits)
bcache: initialize dirty stripes in flash_dev_run()
loop: set physical block size to logical block size
bcache: fix bch_hprint crash and improve output
bcache: Update continue_at() documentation
bcache: silence static checker warning
bcache: fix for gc and write-back race
bcache: increase the number of open buckets
bcache: Correct return value for sysfs attach errors
bcache: correct cache_dirty_target in __update_writeback_rate()
bcache: gc does not work when triggering by manual command
bcache: Don't reinvent the wheel but use existing llist API
bcache: do not subtract sectors_to_gc for bypassed IO
bcache: fix sequential large write IO bypass
bcache: Fix leak of bdev reference
block/loop: remove unused field
block/loop: fix use after free
bfq: Use icq_to_bic() consistently
bfq: Suppress compiler warnings about comparisons
bfq: Check kstrtoul() return value
bfq: Declare local functions static
...
Some code uses icq_to_bic() to convert an io_cq pointer to a
bfq_io_cq pointer while other code uses a direct cast. Convert
the code that uses a direct cast such that it uses icq_to_bic().
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This patch avoids that the following warnings are reported when
building with W=1:
block/bfq-iosched.c: In function 'bfq_back_seek_max_store':
block/bfq-iosched.c:4860:13: warning: comparison of unsigned expression < 0 is always false [-Wtype-limits]
if (__data < (MIN)) \
^
block/bfq-iosched.c:4876:1: note: in expansion of macro 'STORE_FUNCTION'
STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
^~~~~~~~~~~~~~
block/bfq-iosched.c: In function 'bfq_slice_idle_store':
block/bfq-iosched.c:4860:13: warning: comparison of unsigned expression < 0 is always false [-Wtype-limits]
if (__data < (MIN)) \
^
block/bfq-iosched.c:4879:1: note: in expansion of macro 'STORE_FUNCTION'
STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
^~~~~~~~~~~~~~
block/bfq-iosched.c: In function 'bfq_slice_idle_us_store':
block/bfq-iosched.c:4892:13: warning: comparison of unsigned expression < 0 is always false [-Wtype-limits]
if (__data < (MIN)) \
^
block/bfq-iosched.c:4899:1: note: in expansion of macro 'USEC_STORE_FUNCTION'
USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
^~~~~~~~~~~~~~~~~~~
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Make sysfs writes fail for invalid numbers instead of storing
uninitialized data copied from the stack. This patch removes
all uninitialized_var() occurrences from the BFQ source code.
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This patch avoids that gcc 7 issues a warning about fall-through
when building with W=1.
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
To provide a very smooth service, bfq starts to serve a bfq_queue
only if the queue is 'eligible', i.e., if the same queue would
have started to be served in the ideal, perfectly fair system that
bfq simulates internally. This is obtained by associating each
queue with a virtual start time, and by computing a special system
virtual time quantity: a queue is eligible only if the system
virtual time has reached the virtual start time of the
queue. Finally, bfq guarantees that, when a new queue must be set
in service, there is always at least one eligible entity for each
active parent entity in the scheduler. To provide this guarantee,
the function __bfq_lookup_next_entity pushes up, for each parent
entity on which it is invoked, the system virtual time to the
minimum among the virtual start times of the entities in the
active tree for the parent entity (more precisely, the push up
occurs if the system virtual time happens to be lower than all
such virtual start times).
There is however a circumstance in which __bfq_lookup_next_entity
cannot push up the system virtual time for a parent entity, even
if the system virtual time is lower than the virtual start times
of all the child entities in the active tree. It happens if one of
the child entities is in service. In fact, in such a case, there
is already an eligible entity, the in-service one, even if it may
not be not present in the active tree (because in-service entities
may be removed from the active tree).
Unfortunately, in the last re-design of the
hierarchical-scheduling engine, the reset of the pointer to the
in-service entity for a given parent entity--reset to be done as a
consequence of the expiration of the in-service entity--always
happens after the function __bfq_lookup_next_entity has been
invoked. This causes the function to think that there is still an
entity in service for the parent entity, and then that the system
virtual time cannot be pushed up, even if actually such a
no-more-in-service entity has already been properly reinserted
into the active tree (or in some other tree if no more
active). Yet, the system virtual time *had* to be pushed up, to be
ready to correctly choose the next queue to serve. Because of the
lack of this push up, bfq may wrongly set in service a queue that
had been speculatively pre-computed as the possible
next-in-service queue, but that would no more be the one to serve
after the expiration and the reinsertion into the active trees of
the previously in-service entities.
This commit addresses this issue by making
__bfq_lookup_next_entity properly push up the system virtual time
if an expiration is occurring.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The block core requests modules with the "-iosched" name suffix, but
bfq no longer has that suffix. Add an alias.
Fixes: ea25da4808 ("block, bfq: split bfq-iosched.c into multiple ...")
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The last parameter "count" never be used in xxx_var_store,
convert these functions to void.
Signed-off-by: weiping zhang <zhangweiping@didichuxing.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When a queue associated with a process remains empty, there are cases
where throughput gets boosted if the device is idled to await the
arrival of a new I/O request for that queue. Currently, BFQ assumes
that one of these cases is when the device has no internal queueing
(regardless of the properties of the I/O being served). Unfortunately,
this condition has proved to be too general. So, this commit refines it
as "the device has no internal queueing and is rotational".
This refinement provides a significant throughput boost with random
I/O, on flash-based storage without internal queueing. For example, on
a HiKey board, throughput increases by up to 125%, growing, e.g., from
6.9MB/s to 15.6MB/s with two or three random readers in parallel.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The logic that decides whether to idle the device is scattered across
three functions. Almost all of the logic is in the function
bfq_bfqq_may_idle, but (1) part of the decision is made in
bfq_update_idle_window, and (2) the function bfq_bfqq_must_idle may
switch off idling regardless of the output of bfq_bfqq_may_idle. In
addition, both bfq_update_idle_window and bfq_bfqq_must_idle make
their decisions as a function of parameters that are used, for similar
purposes, also in bfq_bfqq_may_idle. This commit addresses these
issues by moving all the logic into bfq_bfqq_may_idle.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There are mq devices (eg., virtio-blk, nbd and loopback) which don't
invoke blk_mq_run_hw_queues() after the completion of a request.
If bfq is enabled on these devices and the slice_idle attribute or
strict_guarantees attribute is set as zero, it is possible that
after a request completion the remaining requests of busy bfq queue
will stalled in the bfq schedule until a new request arrives.
To fix the scheduler latency problem, we need to check whether or not
all issued requests have completed and dispatch more requests to driver
if there is no request in driver.
The problem can be reproduced by running the following script
on a virtio-blk device with nr_hw_queues as 1:
#!/bin/sh
dev=vdb
# mount point for dev
mp=/tmp/mnt
cd $mp
job=strict.job
cat <<EOF > $job
[global]
direct=1
bs=4k
size=256M
rw=write
ioengine=libaio
iodepth=128
runtime=5
time_based
[1]
filename=1.data
[2]
new_group
filename=2.data
EOF
echo bfq > /sys/block/$dev/queue/scheduler
echo 1 > /sys/block/$dev/queue/iosched/strict_guarantees
fio $job
Signed-off-by: Hou Tao <houtao1@huawei.com>
Reviewed-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
On each deactivation or re-scheduling (after being served) of a
bfq_queue, BFQ invokes the function __bfq_entity_update_weight_prio(),
to perform pending updates of ioprio, weight and ioprio class for the
bfq_queue. BFQ also invokes this function on I/O-request dispatches,
to raise or lower weights more quickly when needed, thereby improving
latency. However, the entity representing the bfq_queue may be on the
active (sub)tree of a service tree when this happens, and, although
with a very low probability, the bfq_queue may happen to also have a
pending change of its ioprio class. If both conditions hold when
__bfq_entity_update_weight_prio() is invoked, then the entity moves to
a sort of hybrid state: the new service tree for the entity, as
returned by bfq_entity_service_tree(), differs from service tree on
which the entity still is. The functions that handle activations and
deactivations of entities do not cope with such a hybrid state (and
would need to become more complex to cope).
This commit addresses this issue by just making
__bfq_entity_update_weight_prio() not perform also a possible pending
change of ioprio class, when invoked on an I/O-request dispatch for a
bfq_queue. Such a change is thus postponed to when
__bfq_entity_update_weight_prio() is invoked on deactivation or
re-scheduling of the bfq_queue.
Reported-by: Marco Piazza <mpiazza@gmail.com>
Reported-by: Laurentiu Nicola <lnicola@dend.ro>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Marco Piazza <mpiazza@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This commit fixes a bug triggered by a non-trivial sequence of
events. These events are briefly described in the next two
paragraphs. The impatiens, or those who are familiar with queue
merging and splitting, can jump directly to the last paragraph.
On each I/O-request arrival for a shared bfq_queue, i.e., for a
bfq_queue that is the result of the merge of two or more bfq_queues,
BFQ checks whether the shared bfq_queue has become seeky (i.e., if too
many random I/O requests have arrived for the bfq_queue; if the device
is non rotational, then random requests must be also small for the
bfq_queue to be tagged as seeky). If the shared bfq_queue is actually
detected as seeky, then a split occurs: the bfq I/O context of the
process that has issued the request is redirected from the shared
bfq_queue to a new non-shared bfq_queue. As a degenerate case, if the
shared bfq_queue actually happens to be shared only by one process
(because of previous splits), then no new bfq_queue is created: the
state of the shared bfq_queue is just changed from shared to non
shared.
Regardless of whether a brand new non-shared bfq_queue is created, or
the pre-existing shared bfq_queue is just turned into a non-shared
bfq_queue, several parameters of the non-shared bfq_queue are set
(restored) to the original values they had when the bfq_queue
associated with the bfq I/O context of the process (that has just
issued an I/O request) was merged with the shared bfq_queue. One of
these parameters is the weight-raising state.
If, on the split of a shared bfq_queue,
1) a pre-existing shared bfq_queue is turned into a non-shared
bfq_queue;
2) the previously shared bfq_queue happens to be busy;
3) the weight-raising state of the previously shared bfq_queue happens
to change;
the number of weight-raised busy queues changes. The field
wr_busy_queues must then be updated accordingly, but such an update
was missing. This commit adds the missing update.
Reported-by: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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>
icq_to_bic is a container_of operation, so we need to check for NULL
before it. Also move the check outside the spinlock while we're at
it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In blk-cgroup, operations on blkg objects are protected with the
request_queue lock. This is no more the lock that protects
I/O-scheduler operations in blk-mq. In fact, the latter are now
protected with a finer-grained per-scheduler-instance lock. As a
consequence, although blkg lookups are also rcu-protected, blk-mq I/O
schedulers may see inconsistent data when they access blkg and
blkg-related objects. BFQ does access these objects, and does incur
this problem, in the following case.
The blkg_lookup performed in bfq_get_queue, being protected (only)
through rcu, may happen to return the address of a copy of the
original blkg. If this is the case, then the blkg_get performed in
bfq_get_queue, to pin down the blkg, is useless: it does not prevent
blk-cgroup code from destroying both the original blkg and all objects
directly or indirectly referred by the copy of the blkg. BFQ accesses
these objects, which typically causes a crash for NULL-pointer
dereference of memory-protection violation.
Some additional protection mechanism should be added to blk-cgroup to
address this issue. In the meantime, this commit provides a quick
temporary fix for BFQ: cache (when safe) blkg data that might
disappear right after a blkg_lookup.
In particular, this commit exploits the following facts to achieve its
goal without introducing further locks. Destroy operations on a blkg
invoke, as a first step, hooks of the scheduler associated with the
blkg. And these hooks are executed with bfqd->lock held for BFQ. As a
consequence, for any blkg associated with the request queue an
instance of BFQ is attached to, we are guaranteed that such a blkg is
not destroyed, and that all the pointers it contains are consistent,
while that instance is holding its bfqd->lock. A blkg_lookup performed
with bfqd->lock held then returns a fully consistent blkg, which
remains consistent until this lock is held. In more detail, this holds
even if the returned blkg is a copy of the original one.
Finally, also the object describing a group inside BFQ needs to be
protected from destruction on the blkg_free of the original blkg
(which invokes bfq_pd_free). This commit adds private refcounting for
this object, to let it disappear only after no bfq_queue refers to it
any longer.
This commit also removes or updates some stale comments on locking
issues related to blk-cgroup operations.
Reported-by: Tomas Konir <tomas.konir@gmail.com>
Reported-by: Lee Tibbert <lee.tibbert@gmail.com>
Reported-by: Marco Piazza <mpiazza@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Tomas Konir <tomas.konir@gmail.com>
Tested-by: Lee Tibbert <lee.tibbert@gmail.com>
Tested-by: Marco Piazza <mpiazza@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The introduction of the BFQ and Kyber I/O schedulers has triggered a
new wave of I/O benchmarks. Unfortunately, comments and discussions on
these benchmarks confirm that there is still little awareness that it
is very hard to achieve, at the same time, a low latency and a high
throughput. In particular, virtually all benchmarks measure
throughput, or throughput-related figures of merit, but, for BFQ, they
use the scheduler in its default configuration. This configuration is
geared, instead, toward a low latency. This is evidently a sign that
BFQ documentation is still too unclear on this important aspect. This
commit addresses this issue by stressing how BFQ configuration must be
(easily) changed if the only goal is maximum throughput.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
The call to bfq_check_ioprio_change will dereference bic, however,
the null check for bic is after this call. Move the the null
check on bic to before the call to avoid any potential null
pointer dereference issues.
Detected by CoverityScan, CID#1430138 ("Dereference before null check")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The BFQ I/O scheduler features an optimal fair-queuing
(proportional-share) scheduling algorithm, enriched with several
mechanisms to boost throughput and reduce latency for interactive and
real-time applications. This makes BFQ a large and complex piece of
code. This commit addresses this issue by splitting BFQ into three
main, independent components, and by moving each component into a
separate source file:
1. Main algorithm: handles the interaction with the kernel, and
decides which requests to dispatch; it uses the following two further
components to achieve its goals.
2. Scheduling engine (Hierarchical B-WF2Q+ scheduling algorithm):
computes the schedule, using weights and budgets provided by the above
component.
3. cgroups support: handles group operations (creation, destruction,
move, ...).
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
When a bfq queue is set in service and when it is merged, a reference
to the I/O context associated with the queue is taken. This reference
is then released when the queue is deselected from service or
split. More precisely, the release of the reference is postponed to
when the scheduler lock is released, to avoid nesting between the
scheduler and the I/O-context lock. In fact, such nesting would lead
to deadlocks, because of other code paths that take the same locks in
the opposite order. This postponing of I/O-context releases does
complicate code.
This commit addresses these issue by modifying involved operations in
such a way to not need to get the above I/O-context references any
more. Then it also removes any get and release of these references.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
Many popular I/O-intensive services or applications spawn or
reactivate many parallel threads/processes during short time
intervals. Examples are systemd during boot or git grep. These
services or applications benefit mostly from a high throughput: the
quicker the I/O generated by their processes is cumulatively served,
the sooner the target job of these services or applications gets
completed. As a consequence, it is almost always counterproductive to
weight-raise any of the queues associated to the processes of these
services or applications: in most cases it would just lower the
throughput, mainly because weight-raising also implies device idling.
To address this issue, an I/O scheduler needs, first, to detect which
queues are associated with these services or applications. In this
respect, we have that, from the I/O-scheduler standpoint, these
services or applications cause bursts of activations, i.e.,
activations of different queues occurring shortly after each
other. However, a shorter burst of activations may be caused also by
the start of an application that does not consist in a lot of parallel
I/O-bound threads (see the comments on the function bfq_handle_burst
for details).
In view of these facts, this commit introduces:
1) an heuristic to detect (only) bursts of queue activations caused by
services or applications consisting in many parallel I/O-bound
threads;
2) the prevention of device idling and weight-raising for the queues
belonging to these bursts.
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
This patch is basically the counterpart, for NCQ-capable rotational
devices, of the previous patch. Exactly as the previous patch does on
flash-based devices and for any workload, this patch disables device
idling on rotational devices, but only for random I/O. In fact, only
with these queues disabling idling boosts the throughput on
NCQ-capable rotational devices. To not break service guarantees,
idling is disabled for NCQ-enabled rotational devices only when the
same symmetry conditions considered in the previous patches hold.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This patch boosts the throughput on NCQ-capable flash-based devices,
while still preserving latency guarantees for interactive and soft
real-time applications. The throughput is boosted by just not idling
the device when the in-service queue remains empty, even if the queue
is sync and has a non-null idle window. This helps to keep the drive's
internal queue full, which is necessary to achieve maximum
performance. This solution to boost the throughput is a port of
commits a68bbdd and f7d7b7a for CFQ.
As already highlighted in a previous patch, allowing the device to
prefetch and internally reorder requests trivially causes loss of
control on the request service order, and hence on service guarantees.
Fortunately, as discussed in detail in the comments on the function
bfq_bfqq_may_idle(), if every process has to receive the same
fraction of the throughput, then the service order enforced by the
internal scheduler of a flash-based device is relatively close to that
enforced by BFQ. In particular, it is close enough to let service
guarantees be substantially preserved.
Things change in an asymmetric scenario, i.e., if not every process
has to receive the same fraction of the throughput. In this case, to
guarantee the desired throughput distribution, the device must be
prevented from prefetching requests. This is exactly what this patch
does in asymmetric scenarios.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
A seeky queue (i..e, a queue containing random requests) is assigned a
very small device-idling slice, for throughput issues. Unfortunately,
given the process associated with a seeky queue, this behavior causes
the following problem: if the process, say P, performs sync I/O and
has a higher weight than some other processes doing I/O and associated
with non-seeky queues, then BFQ may fail to guarantee to P its
reserved share of the throughput. The reason is that idling is key
for providing service guarantees to processes doing sync I/O [1].
This commit addresses this issue by allowing the device-idling slice
to be reduced for a seeky queue only if the scenario happens to be
symmetric, i.e., if all the queues are to receive the same share of
the throughput.
[1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
Scheduler", Proceedings of the First Workshop on Mobile System
Technologies (MST-2015), May 2015.
http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Riccardo Pizzetti <riccardo.pizzetti@gmail.com>
Signed-off-by: Samuele Zecchini <samuele.zecchini92@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
A set of processes may happen to perform interleaved reads, i.e.,
read requests whose union would give rise to a sequential read pattern.
There are two typical cases: first, processes reading fixed-size chunks
of data at a fixed distance from each other; second, processes reading
variable-size chunks at variable distances. The latter case occurs for
example with QEMU, which splits the I/O generated by a guest into
multiple chunks, and lets these chunks be served by a pool of I/O
threads, iteratively assigning the next chunk of I/O to the first
available thread. CFQ denotes as 'cooperating' a set of processes that
are doing interleaved I/O, and when it detects cooperating processes,
it merges their queues to obtain a sequential I/O pattern from the union
of their I/O requests, and hence boost the throughput.
Unfortunately, in the following frequent case, the mechanism
implemented in CFQ for detecting cooperating processes and merging
their queues is not responsive enough to handle also the fluctuating
I/O pattern of the second type of processes. Suppose that one process
of the second type issues a request close to the next request to serve
of another process of the same type. At that time the two processes
would be considered as cooperating. But, if the request issued by the
first process is to be merged with some other already-queued request,
then, from the moment at which this request arrives, to the moment
when CFQ controls whether the two processes are cooperating, the two
processes are likely to be already doing I/O in distant zones of the
disk surface or device memory.
CFQ uses however preemption to get a sequential read pattern out of
the read requests performed by the second type of processes too. As a
consequence, CFQ uses two different mechanisms to achieve the same
goal: boosting the throughput with interleaved I/O.
This patch introduces Early Queue Merge (EQM), a unified mechanism to
get a sequential read pattern with both types of processes. The main
idea is to immediately check whether a newly-arrived request lets some
pair of processes become cooperating, both in the case of actual
request insertion and, to be responsive with the second type of
processes, in the case of request merge. Both types of processes are
then handled by just merging their queues.
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Mauro Andreolini <mauro.andreolini@unimore.it>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
This patch introduces an heuristic that reduces latency when the
I/O-request pool is saturated. This goal is achieved by disabling
device idling, for non-weight-raised queues, when there are weight-
raised queues with pending or in-flight requests. In fact, as
explained in more detail in the comment on the function
bfq_bfqq_may_idle(), this reduces the rate at which processes
associated with non-weight-raised queues grab requests from the pool,
thereby increasing the probability that processes associated with
weight-raised queues get a request immediately (or at least soon) when
they need one. Along the same line, if there are weight-raised queues,
then this patch halves the service rate of async (write) requests for
non-weight-raised queues.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
I/O schedulers typically allow NCQ-capable drives to prefetch I/O
requests, as NCQ boosts the throughput exactly by prefetching and
internally reordering requests.
Unfortunately, as discussed in detail and shown experimentally in [1],
this may cause fairness and latency guarantees to be violated. The
main problem is that the internal scheduler of an NCQ-capable drive
may postpone the service of some unlucky (prefetched) requests as long
as it deems serving other requests more appropriate to boost the
throughput.
This patch addresses this issue by not disabling device idling for
weight-raised queues, even if the device supports NCQ. This allows BFQ
to start serving a new queue, and therefore allows the drive to
prefetch new requests, only after the idling timeout expires. At that
time, all the outstanding requests of the expired queue have been most
certainly served.
[1] P. Valente and M. Andreolini, "Improving Application
Responsiveness with the BFQ Disk I/O Scheduler", Proceedings of
the 5th Annual International Systems and Storage Conference
(SYSTOR '12), June 2012.
Slightly extended version:
http://algogroup.unimore.it/people/paolo/disk_sched/bfq-v1-suite-
results.pdf
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
To guarantee a low latency also to the I/O requests issued by soft
real-time applications, this patch introduces a further heuristic,
which weight-raises (in the sense explained in the previous patch)
also the queues associated to applications deemed as soft real-time.
To be deemed as soft real-time, an application must meet two
requirements. First, the application must not require an average
bandwidth higher than the approximate bandwidth required to playback
or record a compressed high-definition video. Second, the request
pattern of the application must be isochronous, i.e., after issuing a
request or a batch of requests, the application must stop issuing new
requests until all its pending requests have been completed. After
that, the application may issue a new batch, and so on.
As for the second requirement, it is critical to require also that,
after all the pending requests of the application have been completed,
an adequate minimum amount of time elapses before the application
starts issuing new requests. This prevents also greedy (i.e.,
I/O-bound) applications from being incorrectly deemed, occasionally,
as soft real-time. In fact, if *any amount of time* is fine, then even
a greedy application may, paradoxically, meet both the above
requirements, if: (1) the application performs random I/O and/or the
device is slow, and (2) the CPU load is high. The reason is the
following. First, if condition (1) is true, then, during the service
of the application, the throughput may be low enough to let the
application meet the bandwidth requirement. Second, if condition (2)
is true as well, then the application may occasionally behave in an
apparently isochronous way, because it may simply stop issuing
requests while the CPUs are busy serving other processes.
To address this issue, the heuristic leverages the simple fact that
greedy applications issue *all* their requests as quickly as they can,
whereas soft real-time applications spend some time processing data
after each batch of requests is completed. In particular, the
heuristic works as follows. First, according to the above isochrony
requirement, the heuristic checks whether an application may be soft
real-time, thereby giving to the application the opportunity to be
deemed as such, only when both the following two conditions happen to
hold: 1) the queue associated with the application has expired and is
empty, 2) there is no outstanding request of the application.
Suppose that both conditions hold at time, say, t_c and that the
application issues its next request at time, say, t_i. At time t_c the
heuristic computes the next time instant, called soft_rt_next_start in
the code, such that, only if t_i >= soft_rt_next_start, then both the
next conditions will hold when the application issues its next
request: 1) the application will meet the above bandwidth requirement,
2) a given minimum time interval, say Delta, will have elapsed from
time t_c (so as to filter out greedy application).
The current value of Delta is a little bit higher than the value that
we have found, experimentally, to be adequate on a real,
general-purpose machine. In particular we had to increase Delta to
make the filter quite precise also in slower, embedded systems, and in
KVM/QEMU virtual machines (details in the comments on the code).
If the application actually issues its next request after time
soft_rt_next_start, then its associated queue will be weight-raised
for a relatively short time interval. If, during this time interval,
the application proves again to meet the bandwidth and isochrony
requirements, then the end of the weight-raising period for the queue
is moved forward, and so on. Note that an application whose associated
queue never happens to be empty when it expires will never have the
opportunity to be deemed as soft real-time.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This patch introduces a simple heuristic to load applications quickly,
and to perform the I/O requested by interactive applications just as
quickly. To this purpose, both a newly-created queue and a queue
associated with an interactive application (we explain in a moment how
BFQ decides whether the associated application is interactive),
receive the following two special treatments:
1) The weight of the queue is raised.
2) The queue unconditionally enjoys device idling when it empties; in
fact, if the requests of a queue are sync, then performing device
idling for the queue is a necessary condition to guarantee that the
queue receives a fraction of the throughput proportional to its weight
(see [1] for details).
For brevity, we call just weight-raising the combination of these
two preferential treatments. For a newly-created queue,
weight-raising starts immediately and lasts for a time interval that:
1) depends on the device speed and type (rotational or
non-rotational), and 2) is equal to the time needed to load (start up)
a large-size application on that device, with cold caches and with no
additional workload.
Finally, as for guaranteeing a fast execution to interactive,
I/O-related tasks (such as opening a file), consider that any
interactive application blocks and waits for user input both after
starting up and after executing some task. After a while, the user may
trigger new operations, after which the application stops again, and
so on. Accordingly, the low-latency heuristic weight-raises again a
queue in case it becomes backlogged after being idle for a
sufficiently long (configurable) time. The weight-raising then lasts
for the same time as for a just-created queue.
According to our experiments, the combination of this low-latency
heuristic and of the improvements described in the previous patch
allows BFQ to guarantee a high application responsiveness.
[1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
Scheduler", Proceedings of the First Workshop on Mobile System
Technologies (MST-2015), May 2015.
http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This patch deals with two sources of unfairness, which can also cause
high latencies and throughput loss. The first source is related to
write requests. Write requests tend to starve read requests, basically
because, on one side, writes are slower than reads, whereas, on the
other side, storage devices confuse schedulers by deceptively
signaling the completion of write requests immediately after receiving
them. This patch addresses this issue by just throttling writes. In
particular, after a write request is dispatched for a queue, the
budget of the queue is decremented by the number of sectors to write,
multiplied by an (over)charge coefficient. The value of the
coefficient is the result of our tuning with different devices.
The second source of unfairness has to do with slowness detection:
when the in-service queue is expired, BFQ also controls whether the
queue has been "too slow", i.e., has consumed its last-assigned budget
at such a low rate that it would have been impossible to consume all
of this budget within the maximum time slice T_max (Subsec. 3.5 in
[1]). In this case, the queue is always (over)charged the whole
budget, to reduce its utilization of the device. Both this overcharge
and the slowness-detection criterion may cause unfairness.
First, always charging a full budget to a slow queue is too coarse. It
is much more accurate, and this patch lets BFQ do so, to charge an
amount of service 'equivalent' to the amount of time during which the
queue has been in service. As explained in more detail in the comments
on the code, this enables BFQ to provide time fairness among slow
queues.
Secondly, because of ZBR, a queue may be deemed as slow when its
associated process is performing I/O on the slowest zones of a
disk. However, unless the process is truly too slow, not reducing the
disk utilization of the queue is more profitable in terms of disk
throughput than the opposite. A similar problem is caused by logical
block mapping on non-rotational devices. For this reason, this patch
lets a queue be charged time, and not budget, only if the queue has
consumed less than 2/3 of its assigned budget. As an additional,
important benefit, this tolerance allows BFQ to preserve enough
elasticity to still perform bandwidth, and not time, distribution with
little unlucky or quasi-sequential processes.
Finally, for the same reasons as above, this patch makes slowness
detection itself much less harsh: a queue is deemed slow only if it
has consumed its budget at less than half of the peak rate.
[1] P. Valente and M. Andreolini, "Improving Application
Responsiveness with the BFQ Disk I/O Scheduler", Proceedings of
the 5th Annual International Systems and Storage Conference
(SYSTOR '12), June 2012.
Slightly extended version:
http://algogroup.unimore.it/people/paolo/disk_sched/bfq-v1-suite-
results.pdf
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Unless the maximum budget B_max that BFQ can assign to a queue is set
explicitly by the user, BFQ automatically updates B_max. In
particular, BFQ dynamically sets B_max to the number of sectors that
can be read, at the current estimated peak rate, during the maximum
time, T_max, allowed before a budget timeout occurs. In formulas, if
we denote as R_est the estimated peak rate, then B_max = T_max ∗
R_est. Hence, the higher R_est is with respect to the actual device
peak rate, the higher the probability that processes incur budget
timeouts unjustly is. Besides, a too high value of B_max unnecessarily
increases the deviation from an ideal, smooth service.
Unfortunately, it is not trivial to estimate the peak rate correctly:
because of the presence of sw and hw queues between the scheduler and
the device components that finally serve I/O requests, it is hard to
say exactly when a given dispatched request is served inside the
device, and for how long. As a consequence, it is hard to know
precisely at what rate a given set of requests is actually served by
the device.
On the opposite end, the dispatch time of any request is trivially
available, and, from this piece of information, the "dispatch rate"
of requests can be immediately computed. So, the idea in the next
function is to use what is known, namely request dispatch times
(plus, when useful, request completion times), to estimate what is
unknown, namely in-device request service rate.
The main issue is that, because of the above facts, the rate at
which a certain set of requests is dispatched over a certain time
interval can vary greatly with respect to the rate at which the
same requests are then served. But, since the size of any
intermediate queue is limited, and the service scheme is lossless
(no request is silently dropped), the following obvious convergence
property holds: the number of requests dispatched MUST become
closer and closer to the number of requests completed as the
observation interval grows. This is the key property used in
this new version of the peak-rate estimator.
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
