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>
The feedback-loop algorithm used by BFQ to compute queue (process)
budgets is basically a set of three update rules, one for each of the
main reasons why a queue may be expired. If many processes suddenly
switch from sporadic I/O to greedy and sequential I/O, then these
rules are quite slow to assign large budgets to these processes, and
hence to achieve a high throughput. On the opposite side, BFQ assigns
the maximum possible budget B_max to a just-created queue. This allows
a high throughput to be achieved immediately if the associated process
is I/O-bound and performs sequential I/O from the beginning. But it
also increases the worst-case latency experienced by the first
requests issued by the process, because the larger the budget of a
queue waiting for service is, the later the queue will be served by
B-WF2Q+ (Subsec 3.3 in [1]). This is detrimental for an interactive or
soft real-time application.
To tackle these throughput and latency problems, on one hand this
patch changes the initial budget value to B_max/2. On the other hand,
it re-tunes the three rules, adopting a more aggressive,
multiplicative increase/linear decrease scheme. This scheme trades
latency for throughput more than before, and tends to assign large
budgets quickly to processes that are or become I/O-bound. For two of
the expiration reasons, the new version of the rules also contains
some more little improvements, briefly described below.
*No more backlog.* In this case, the budget was larger than the number
of sectors actually read/written by the process before it stopped
doing I/O. Hence, to reduce latency for the possible future I/O
requests of the process, the old rule simply set the next budget to
the number of sectors actually consumed by the process. However, if
there are still outstanding requests, then the process may have not
yet issued its next request just because it is still waiting for the
completion of some of the still outstanding ones. If this sub-case
holds true, then the new rule, instead of decreasing the budget,
doubles it, proactively, in the hope that: 1) a larger budget will fit
the actual needs of the process, and 2) the process is sequential and
hence a higher throughput will be achieved by serving the process
longer after granting it access to the device.
*Budget timeout*. The original rule set the new budget to the maximum
value B_max, to maximize throughput and let all processes experiencing
budget timeouts receive the same share of the device time. In our
experiments we verified that this sudden jump to B_max did not provide
sensible benefits; rather it increased the latency of processes
performing sporadic and short I/O. The new rule only doubles the
budget.
[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>
Add complete support for full hierarchical scheduling, with a cgroups
interface. Full hierarchical scheduling is implemented through the
'entity' abstraction: both bfq_queues, i.e., the internal BFQ queues
associated with processes, and groups are represented in general by
entities. Given the bfq_queues associated with the processes belonging
to a given group, the entities representing these queues are sons of
the entity representing the group. At higher levels, if a group, say
G, contains other groups, then the entity representing G is the parent
entity of the entities representing the groups in G.
Hierarchical scheduling is performed as follows: if the timestamps of
a leaf entity (i.e., of a bfq_queue) change, and such a change lets
the entity become the next-to-serve entity for its parent entity, then
the timestamps of the parent entity are recomputed as a function of
the budget of its new next-to-serve leaf entity. If the parent entity
belongs, in its turn, to a group, and its new timestamps let it become
the next-to-serve for its parent entity, then the timestamps of the
latter parent entity are recomputed as well, and so on. When a new
bfq_queue must be set in service, the reverse path is followed: the
next-to-serve highest-level entity is chosen, then its next-to-serve
child entity, and so on, until the next-to-serve leaf entity is
reached, and the bfq_queue that this entity represents is set in
service.
Writeback is accounted for on a per-group basis, i.e., for each group,
the async I/O requests of the processes of the group are enqueued in a
distinct bfq_queue, and the entity associated with this queue is a
child of the entity associated with the group.
Weights can be assigned explicitly to groups and processes through the
cgroups interface, differently from what happens, for single
processes, if the cgroups interface is not used (as explained in the
description of the previous patch). In particular, since each node has
a full scheduler, each group can be assigned its own weight.
Signed-off-by: Fabio Checconi <fchecconi@gmail.com>
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>
We tag as v0 the version of BFQ containing only BFQ's engine plus
hierarchical support. BFQ's engine is introduced by this commit, while
hierarchical support is added by next commit. We use the v0 tag to
distinguish this minimal version of BFQ from the versions containing
also the features and the improvements added by next commits. BFQ-v0
coincides with the version of BFQ submitted a few years ago [1], apart
from the introduction of preemption, described below.
BFQ is a proportional-share I/O scheduler, whose general structure,
plus a lot of code, are borrowed from CFQ.
- Each process doing I/O on a device is associated with a weight and a
(bfq_)queue.
- BFQ grants exclusive access to the device, for a while, to one queue
(process) at a time, and implements this service model by
associating every queue with a budget, measured in number of
sectors.
- After a queue is granted access to the device, the budget of the
queue is decremented, on each request dispatch, by the size of the
request.
- The in-service queue is expired, i.e., its service is suspended,
only if one of the following events occurs: 1) the queue finishes
its budget, 2) the queue empties, 3) a "budget timeout" fires.
- The budget timeout prevents processes doing random I/O from
holding the device for too long and dramatically reducing
throughput.
- Actually, as in CFQ, a queue associated with a process issuing
sync requests may not be expired immediately when it empties. In
contrast, BFQ may idle the device for a short time interval,
giving the process the chance to go on being served if it issues
a new request in time. Device idling typically boosts the
throughput on rotational devices, if processes do synchronous
and sequential I/O. In addition, under BFQ, device idling is
also instrumental in guaranteeing the desired throughput
fraction to processes issuing sync requests (see [2] for
details).
- With respect to idling for service guarantees, if several
processes are competing for the device at the same time, but
all processes (and groups, after the following commit) have
the same weight, then BFQ guarantees the expected throughput
distribution without ever idling the device. Throughput is
thus as high as possible in this common scenario.
- Queues are scheduled according to a variant of WF2Q+, named
B-WF2Q+, and implemented using an augmented rb-tree to preserve an
O(log N) overall complexity. See [2] for more details. B-WF2Q+ is
also ready for hierarchical scheduling. However, for a cleaner
logical breakdown, the code that enables and completes
hierarchical support is provided in the next commit, which focuses
exactly on this feature.
- B-WF2Q+ guarantees a tight deviation with respect to an ideal,
perfectly fair, and smooth service. In particular, B-WF2Q+
guarantees that each queue receives a fraction of the device
throughput proportional to its weight, even if the throughput
fluctuates, and regardless of: the device parameters, the current
workload and the budgets assigned to the queue.
- The last, budget-independence, property (although probably
counterintuitive in the first place) is definitely beneficial, for
the following reasons:
- First, with any proportional-share scheduler, the maximum
deviation with respect to an ideal service is proportional to
the maximum budget (slice) assigned to queues. As a consequence,
BFQ can keep this deviation tight not only because of the
accurate service of B-WF2Q+, but also because BFQ *does not*
need to assign a larger budget to a queue to let the queue
receive a higher fraction of the device throughput.
- Second, BFQ is free to choose, for every process (queue), the
budget that best fits the needs of the process, or best
leverages the I/O pattern of the process. In particular, BFQ
updates queue budgets with a simple feedback-loop algorithm that
allows a high throughput to be achieved, while still providing
tight latency guarantees to time-sensitive applications. When
the in-service queue expires, this algorithm computes the next
budget of the queue so as to:
- Let large budgets be eventually assigned to the queues
associated with I/O-bound applications performing sequential
I/O: in fact, the longer these applications are served once
got access to the device, the higher the throughput is.
- Let small budgets be eventually assigned to the queues
associated with time-sensitive applications (which typically
perform sporadic and short I/O), because, the smaller the
budget assigned to a queue waiting for service is, the sooner
B-WF2Q+ will serve that queue (Subsec 3.3 in [2]).
- Weights can be assigned to processes only indirectly, through I/O
priorities, and according to the relation:
weight = 10 * (IOPRIO_BE_NR - ioprio).
The next patch provides, instead, a cgroups interface through which
weights can be assigned explicitly.
- If several processes are competing for the device at the same time,
but all processes and groups have the same weight, then BFQ
guarantees the expected throughput distribution without ever idling
the device. It uses preemption instead. Throughput is then much
higher in this common scenario.
- ioprio classes are served in strict priority order, i.e.,
lower-priority queues are not served as long as there are
higher-priority queues. Among queues in the same class, the
bandwidth is distributed in proportion to the weight of each
queue. A very thin extra bandwidth is however guaranteed to the Idle
class, to prevent it from starving.
- If the strict_guarantees parameter is set (default: unset), then BFQ
- always performs idling when the in-service queue becomes empty;
- forces the device to serve one I/O request at a time, by
dispatching a new request only if there is no outstanding
request.
In the presence of differentiated weights or I/O-request sizes,
both the above conditions are needed to guarantee that every
queue receives its allotted share of the bandwidth (see
Documentation/block/bfq-iosched.txt for more details). Setting
strict_guarantees may evidently affect throughput.
[1] https://lkml.org/lkml/2008/4/1/234https://lkml.org/lkml/2008/11/11/148
[2] 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: Fabio Checconi <fchecconi@gmail.com>
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>
The Kyber I/O scheduler is an I/O scheduler for fast devices designed to
scale to multiple queues. Users configure only two knobs, the target
read and synchronous write latencies, and the scheduler tunes itself to
achieve that latency goal.
The implementation is based on "tokens", built on top of the scalable
bitmap library. Tokens serve as a mechanism for limiting requests. There
are two tiers of tokens: queueing tokens and dispatch tokens.
A queueing token is required to allocate a request. In fact, these
tokens are actually the blk-mq internal scheduler tags, but the
scheduler manages the allocation directly in order to implement its
policy.
Dispatch tokens are device-wide and split up into two scheduling
domains: reads vs. writes. Each hardware queue dispatches batches
round-robin between the scheduling domains as long as tokens are
available for that domain.
These tokens can be used as the mechanism to enable various policies.
The policy Kyber uses is inspired by active queue management techniques
for network routing, similar to blk-wbt. The scheduler monitors
latencies and scales the number of dispatch tokens accordingly. Queueing
tokens are used to prevent starvation of synchronous requests by
asynchronous requests.
Various extensions are possible, including better heuristics and ionice
support. The new scheduler isn't set as the default yet.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently, this callback is called right after put_request() and has no
distinguishable purpose. Instead, let's call it before put_request() as
soon as I/O has completed on the request, before we account it in
blk-stat. With this, Kyber can enable stats when it sees a latency
outlier and make sure the outlier gets accounted.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq_finish_request() is required for schedulers that define their own
put_request(). blk_mq_run_hw_queue() is required for schedulers that
hold back requests to be run later.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Wire up the sbitmap_get_shallow() operation to the tag code so that a
caller can limit the number of tags available to it.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
When CFQ calls wbt_disable_default(), it will call
blk_stat_remove_callback() to stop gathering IO statistics for the
purposes of writeback throttling. Later, when request_queue is
unregistered, wbt_exit() will call blk_stat_remove_callback() again
which will try to delete callback from the list again and possibly cause
list corruption.
Fix the problem by making wbt_disable_default() called wbt_exit() which
is properly guarded against being called multiple times.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
Instead of showing the hctx state and flags as numbers, show the
names of the flags.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Omar Sandoval <osandov@fb.com>
Cc: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Make it possible to check whether or not a block layer queue has
been stopped. Make it possible to start and to run a blk-mq queue
from user space.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Omar Sandoval <osandov@fb.com>
Cc: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Now that we use the proper REQ_OP_WRITE_ZEROES operation everywhere we can
kill this hack.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Now that we have REQ_OP_WRITE_ZEROES implemented for all devices that
support efficient zeroing, we can remove the call to blkdev_issue_discard.
This means we only have two ways of zeroing left and can simplify the
code.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
This avoids fallbacks to explicit zeroing in (__)blkdev_issue_zeroout if
the caller doesn't want them.
Also clean up the convoluted check for the return condition that this
new flag is added to.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
If this flag is set logical provisioning capable device should
release space for the zeroed blocks if possible, if it is not set
devices should keep the blocks anchored.
Also remove an out of sync kerneldoc comment for a static function
that would have become even more out of data with this change.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Turn the existing discard flag into a new BLKDEV_ZERO_UNMAP flag with
similar semantics, but without referring to diѕcard.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
We'll always use the WRITE ZEROES code for zeroing now.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Copy and past the REQ_OP_WRITE_SAME code to prepare to implementations
that limit the write zeroes size.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Lets not flood the kernel log with messages unless
the user requests so.
Signed-off-by: Scott Bauer <scott.bauer@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
The blk_mq_dispatch_rq_list() implementation got modified several
times but the comments in that function were not updated every
time. Since it is nontrivial what is going on, update the comments
in blk_mq_dispatch_rq_list().
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Omar Sandoval <osandov@fb.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Since the next patch in this series will use RCU to iterate over
tag_list, make this safe. Add lockdep_assert_held() statements
in functions that iterate over tag_list to make clear that using
list_for_each_entry() instead of list_for_each_entry_rcu() is
fine in these functions.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Minor cleanup that makes it easier to figure out what's going on in the
driver tag allocation failure path of blk_mq_dispatch_rq_list().
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Schedulers need to be informed when a hardware queue is added or removed
at runtime so they can allocate/free per-hardware queue data. So,
replace the blk_mq_sched_init_hctx_data() helper, which only makes sense
at init time, with .init_hctx() and .exit_hctx() hooks.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
We've added a considerable amount of fixes for stalls and issues
with the blk-mq scheduling in the 4.11 series since forking
off the for-4.12/block branch. We need to do improvements on
top of that for 4.12, so pull in the previous fixes to make
our lives easier going forward.
Signed-off-by: Jens Axboe <axboe@fb.com>
To improve scalability, if hardware queues are shared, restart
a single hardware queue in round-robin fashion. Rename
blk_mq_sched_restart_queues() to reflect the new semantics.
Remove blk_mq_sched_mark_restart_queue() because this function
has no callers. Remove flag QUEUE_FLAG_RESTART because this
patch removes the code that uses this flag.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Introduce a function that runs a hardware queue unconditionally
after a delay. Note: there is already a function that stops and
restarts a hardware queue after a delay, namely blk_mq_delay_queue().
This function will be used in the next patch in this series.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Hannes Reinecke <hare@suse.de>
Cc: Long Li <longli@microsoft.com>
Cc: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently only dm and md/raid5 bios trigger
trace_block_bio_complete(). Now that we have bio_chain() and
bio_inc_remaining(), it is not possible, in general, for a driver to
know when the bio is really complete. Only bio_endio() knows that.
So move the trace_block_bio_complete() call to bio_endio().
Now trace_block_bio_complete() pairs with trace_block_bio_queue().
Any bio for which a 'queue' event is traced, will subsequently
generate a 'complete' event.
There are a few cases where completion tracing is not wanted.
1/ If blk_update_request() has already generated a completion
trace event at the 'request' level, there is no point generating
one at the bio level too. In this case the bi_sector and bi_size
will have changed, so the bio level event would be wrong
2/ If the bio hasn't actually been queued yet, but is being aborted
early, then a trace event could be confusing. Some filesystems
call bio_endio() but do not want tracing.
3/ The bio_integrity code interposes itself by replacing bi_end_io,
then restoring it and calling bio_endio() again. This would produce
two identical trace events if left like that.
To handle these, we introduce a flag BIO_TRACE_COMPLETION and only
produce the trace event when this is set.
We address point 1 above by clearing the flag in blk_update_request().
We address point 2 above by only setting the flag when
generic_make_request() is called.
We address point 3 above by clearing the flag after generating a
completion event.
When bio_split() is used on a bio, particularly in blk_queue_split(),
there is an extra complication. A new bio is split off the front, and
may be handle directly without going through generic_make_request().
The old bio, which has been advanced, is passed to
generic_make_request(), so it will trigger a trace event a second
time.
Probably the best result when a split happens is to see a single
'queue' event for the whole bio, then multiple 'complete' events - one
for each component. To achieve this was can:
- copy the BIO_TRACE_COMPLETION flag to the new bio in bio_split()
- avoid generating a 'queue' event if BIO_TRACE_COMPLETION is already set.
This way, the split-off bio won't create a queue event, the original
won't either even if it re-submitted to generic_make_request(),
but both will produce completion events, each for their own range.
So if generic_make_request() is called (which generates a QUEUED
event), then bi_endio() will create a single COMPLETE event for each
range that the bio is split into, unless the driver has explicitly
requested it not to.
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
blk_mq_update_nr_hw_queues() used to remap hardware queues, which is the
behavior that drivers expect. However, commit 4e68a01142 changed
blk_mq_queue_reinit() to not remap queues for the case of CPU
hotplugging, inadvertently making blk_mq_update_nr_hw_queues() not remap
queues as well. This breaks, for example, NBD's multi-connection mode,
leaving the added hardware queues unused. Fix it by making
blk_mq_update_nr_hw_queues() explicitly remap the queues.
Fixes: 4e68a01142 ("blk-mq: don't redistribute hardware queues on a CPU hotplug event")
Reviewed-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
In elevator_switch(), if blk_mq_init_sched() fails, we attempt to fall
back to the original scheduler. However, at this point, we've already
torn down the original scheduler's tags, so this causes a crash. Doing
the fallback like the legacy elevator path is much harder for mq, so fix
it by just falling back to none, instead.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
If a new hardware queue is added at runtime, we don't allocate scheduler
tags for it, leading to a crash. This hooks up the scheduler framework
to blk_mq_{init,exit}_hctx() to make sure everything gets properly
initialized/freed.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Preparation cleanup for the next couple of fixes, push
blk_mq_sched_setup() and e->ops.mq.init_sched() into a helper.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
While dispatching requests, if we fail to get a driver tag, we mark the
hardware queue as waiting for a tag and put the requests on a
hctx->dispatch list to be run later when a driver tag is freed. However,
blk_mq_dispatch_rq_list() may dispatch requests from multiple hardware
queues if using a single-queue scheduler with a multiqueue device. If
blk_mq_get_driver_tag() fails, it doesn't update the hardware queue we
are processing. This means we end up using the hardware queue of the
previous request, which may or may not be the same as that of the
current request. If it isn't, the wrong hardware queue will end up
waiting for a tag, and the requests will be on the wrong dispatch list,
leading to a hang.
The fix is twofold:
1. Make sure we save which hardware queue we were trying to get a
request for in blk_mq_get_driver_tag() regardless of whether it
succeeds or not.
2. Make blk_mq_dispatch_rq_list() take a request_queue instead of a
blk_mq_hw_queue to make it clear that it must handle multiple
hardware queues, since I've already messed this up on a couple of
occasions.
This didn't appear in testing with nvme and mq-deadline because nvme has
more driver tags than the default number of scheduler tags. However,
with the blk_mq_update_nr_hw_queues() fix, it showed up with nbd.
Signed-off-by: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Instead of bloating the generic struct request with it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
The block layer core sets blk_mq_queue_data.list but no block
drivers read that member. Hence remove it and also the code that
is used to set this member.
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
Writeback throttling does not play well with CFQ since that also tries
to throttle async writes. As a result async writeback can get starved in
presence of readers. As an example take a benchmark simulating
postgreSQL database running over a standard rotating SATA drive. There
are 16 processes doing random reads from a huge file (2*machine memory),
1 process doing random writes to the huge file and calling fsync once
per 50000 writes and 1 process doing sequential 8k writes to a
relatively small file wrapping around at the end of the file and calling
fsync every 5 writes. Under this load read latency easily exceeds the
target latency of 75 ms (just because there are so many reads happening
against a relatively slow disk) and thus writeback is throttled to a
point where only 1 write request is allowed at a time. Blktrace data
then looks like:
8,0 1 0 8.347751764 0 m N cfq workload slice:40000000
8,0 1 0 8.347755256 0 m N cfq293A / set_active wl_class: 0 wl_type:0
8,0 1 0 8.347784100 0 m N cfq293A / Not idling. st->count:1
8,0 1 3814 8.347763916 5839 UT N [kworker/u9:2] 1
8,0 0 0 8.347777605 0 m N cfq293A / Not idling. st->count:1
8,0 1 0 8.347784100 0 m N cfq293A / Not idling. st->count:1
8,0 3 1596 8.354364057 0 C R 156109528 + 8 (6906954) [0]
8,0 3 0 8.354383193 0 m N cfq6196SN / complete rqnoidle 0
8,0 3 0 8.354386476 0 m N cfq schedule dispatch
8,0 3 0 8.354399397 0 m N cfq293A / Not idling. st->count:1
8,0 3 0 8.354404705 0 m N cfq293A / dispatch_insert
8,0 3 0 8.354409454 0 m N cfq293A / dispatched a request
8,0 3 0 8.354412527 0 m N cfq293A / activate rq, drv=1
8,0 3 1597 8.354414692 0 D W 145961400 + 24 (6718452) [swapper/0]
8,0 3 0 8.354484184 0 m N cfq293A / Not idling. st->count:1
8,0 3 0 8.354487536 0 m N cfq293A / slice expired t=0
8,0 3 0 8.354498013 0 m N / served: vt=5888102466265088 min_vt=5888074869387264
8,0 3 0 8.354502692 0 m N cfq293A / sl_used=6737519 disp=1 charge=6737519 iops=0 sect=24
8,0 3 0 8.354505695 0 m N cfq293A / del_from_rr
...
8,0 0 1810 8.354728768 0 C W 145961400 + 24 (314076) [0]
8,0 0 0 8.354746927 0 m N cfq293A / complete rqnoidle 0
...
8,0 1 3829 8.389886102 5839 G W 145962968 + 24 [kworker/u9:2]
8,0 1 3830 8.389888127 5839 P N [kworker/u9:2]
8,0 1 3831 8.389908102 5839 A W 145978336 + 24 <- (8,4) 44000
8,0 1 3832 8.389910477 5839 Q W 145978336 + 24 [kworker/u9:2]
8,0 1 3833 8.389914248 5839 I W 145962968 + 24 (28146) [kworker/u9:2]
8,0 1 0 8.389919137 0 m N cfq293A / insert_request
8,0 1 0 8.389924305 0 m N cfq293A / add_to_rr
8,0 1 3834 8.389933175 5839 UT N [kworker/u9:2] 1
...
8,0 0 0 9.455290997 0 m N cfq workload slice:40000000
8,0 0 0 9.455294769 0 m N cfq293A / set_active wl_class:0 wl_type:0
8,0 0 0 9.455303499 0 m N cfq293A / fifo=ffff880003166090
8,0 0 0 9.455306851 0 m N cfq293A / dispatch_insert
8,0 0 0 9.455311251 0 m N cfq293A / dispatched a request
8,0 0 0 9.455314324 0 m N cfq293A / activate rq, drv=1
8,0 0 2043 9.455316210 6204 D W 145962968 + 24 (1065401962) [pgioperf]
8,0 0 0 9.455392407 0 m N cfq293A / Not idling. st->count:1
8,0 0 0 9.455395969 0 m N cfq293A / slice expired t=0
8,0 0 0 9.455404210 0 m N / served: vt=5888958194597888 min_vt=5888941810597888
8,0 0 0 9.455410077 0 m N cfq293A / sl_used=4000000 disp=1 charge=4000000 iops=0 sect=24
8,0 0 0 9.455416851 0 m N cfq293A / del_from_rr
...
8,0 0 2045 9.455648515 0 C W 145962968 + 24 (332305) [0]
8,0 0 0 9.455668350 0 m N cfq293A / complete rqnoidle 0
...
8,0 1 4371 9.455710115 5839 G W 145978336 + 24 [kworker/u9:2]
8,0 1 4372 9.455712350 5839 P N [kworker/u9:2]
8,0 1 4373 9.455730159 5839 A W 145986616 + 24 <- (8,4) 52280
8,0 1 4374 9.455732674 5839 Q W 145986616 + 24 [kworker/u9:2]
8,0 1 4375 9.455737563 5839 I W 145978336 + 24 (27448) [kworker/u9:2]
8,0 1 0 9.455742871 0 m N cfq293A / insert_request
8,0 1 0 9.455747550 0 m N cfq293A / add_to_rr
8,0 1 4376 9.455756629 5839 UT N [kworker/u9:2] 1
So we can see a Q event for a write request, then IO is blocked by
writeback throttling and G and I events for the request happen only once
other writeback IO is completed. Thus CFQ always sees only one write
request. When it sees it, it queues the async queue behind all the read
queues and the async queue gets scheduled after about one second. When
it is scheduled, that one request gets dispatched and async queue is
expired as it has no more requests to submit. Overall we submit about
one write request per second.
Although this scheduling is beneficial for read latency, writes are
heavily starved and this causes large delays all over the system (due to
processes blocking on page lock, transaction starts, etc.). When
writeback throttling is disabled, write throughput is about one fifth of
a read throughput which roughly matches readers/writers ratio and
overall the system stalls are much shorter.
Mixing writeback throttling logic with CFQ throttling logic is always a
recipe for surprises as CFQ assumes it sees the big part of the picture
which is not necessarily true when writeback throttling is blocking
requests. So disable writeback throttling logic by default when CFQ is
used as an IO scheduler.
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
In 4.10 I introduced a patch that associates the ioc priority with
each request in the block layer. This work was done in the single queue
block layer code. This patch unifies ioc priority to request mapping across
the single/multi queue block layers.
I have tested this patch with the null block device driver with the following
parameters.
null_blk queue_mode=2 irqmode=0 use_per_node_hctx=1 nr_devices=1
I have not seen a performance regression with this patch and I would appreciate
any feedback or additional testing.
I have also verified that io priorities are passed to the device when using
the SQ and MQ path to a SATA HDD that supports io priorities.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Adam Manzanares <adam.manzanares@wdc.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Commit a4d907b6a3 unified the single and multi queue request handlers,
but in the process, it also screwed up the locking balance and calls
blk_mq_try_issue_directly() with the ctx preempt lock held. This is a
problem for drivers that have set BLK_MQ_F_BLOCKING, since now they
can't reliably sleep.
While in there, protect against similar issues in the future, by adding
a might_sleep() trigger in the BLOCKING path for direct issue or queue
run.
Reported-by: Josef Bacik <josef@toxicpanda.com>
Tested-by: Josef Bacik <josef@toxicpanda.com>
Fixes: a4d907b6a3 ("blk-mq: streamline blk_mq_make_request")
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
