Commit Graph

36 Commits

Author SHA1 Message Date
Paolo Valente
2ec5a5c483 block, bfq: always inject I/O of queues blocked by wakers
Suppose that I/O dispatch is plugged, to wait for new I/O for the
in-service bfq-queue, say bfqq.  Suppose then that there is a further
bfq_queue woken by bfqq, and that this woken queue has pending I/O. A
woken queue does not steal bandwidth from bfqq, because it remains
soon without I/O if bfqq is not served. So there is virtually no risk
of loss of bandwidth for bfqq if this woken queue has I/O dispatched
while bfqq is waiting for new I/O. In contrast, this extra I/O
injection boosts throughput. This commit performs this extra
injection.

Tested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Link: https://lore.kernel.org/r/20210304174627.161-2-paolo.valente@linaro.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-03-25 10:50:07 -06:00
huhai
1a23e06cda bfq: don't duplicate code for different paths
As we can see, returns parent_sched_may_change whether
sd->next_in_service changes or not, so remove this judgment.

Signed-off-by: huhai <huhai@tj.kylinos.cn>
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2021-01-24 18:20:47 -07:00
Dmitry Monakhov
2de791ab49 bfq: fix blkio cgroup leakage v4
Changes from v1:
    - update commit description with proper ref-accounting justification

commit db37a34c56 ("block, bfq: get a ref to a group when adding it to a service tree")
introduce leak forbfq_group and blkcg_gq objects because of get/put
imbalance.
In fact whole idea of original commit is wrong because bfq_group entity
can not dissapear under us because it is referenced by child bfq_queue's
entities from here:
 -> bfq_init_entity()
    ->bfqg_and_blkg_get(bfqg);
    ->entity->parent = bfqg->my_entity

 -> bfq_put_queue(bfqq)
    FINAL_PUT
    ->bfqg_and_blkg_put(bfqq_group(bfqq))
    ->kmem_cache_free(bfq_pool, bfqq);

So parent entity can not disappear while child entity is in tree,
and child entities already has proper protection.
This patch revert commit db37a34c56 ("block, bfq: get a ref to a group when adding it to a service tree")

bfq_group leak trace caused by bad commit:
-> blkg_alloc
   -> bfq_pq_alloc
     -> bfqg_get (+1)
->bfq_activate_bfqq
  ->bfq_activate_requeue_entity
    -> __bfq_activate_entity
       ->bfq_get_entity
         ->bfqg_and_blkg_get (+1)  <==== : Note1
->bfq_del_bfqq_busy
  ->bfq_deactivate_entity+0x53/0xc0 [bfq]
    ->__bfq_deactivate_entity+0x1b8/0x210 [bfq]
      -> bfq_forget_entity(is_in_service = true)
	 entity->on_st_or_in_serv = false   <=== :Note2
	 if (is_in_service)
	     return;  ==> do not touch reference
-> blkcg_css_offline
 -> blkcg_destroy_blkgs
  -> blkg_destroy
   -> bfq_pd_offline
    -> __bfq_deactivate_entity
         if (!entity->on_st_or_in_serv) /* true, because (Note2)
		return false;
 -> bfq_pd_free
    -> bfqg_put() (-1, byt bfqg->ref == 2) because of (Note2)
So bfq_group and blkcg_gq  will leak forever, see test-case below.

##TESTCASE_BEGIN:
#!/bin/bash

max_iters=${1:-100}
#prep cgroup mounts
mount -t tmpfs cgroup_root /sys/fs/cgroup
mkdir /sys/fs/cgroup/blkio
mount -t cgroup -o blkio none /sys/fs/cgroup/blkio

# Prepare blkdev
grep blkio /proc/cgroups
truncate -s 1M img
losetup /dev/loop0 img
echo bfq > /sys/block/loop0/queue/scheduler

grep blkio /proc/cgroups
for ((i=0;i<max_iters;i++))
do
    mkdir -p /sys/fs/cgroup/blkio/a
    echo 0 > /sys/fs/cgroup/blkio/a/cgroup.procs
    dd if=/dev/loop0 bs=4k count=1 of=/dev/null iflag=direct 2> /dev/null
    echo 0 > /sys/fs/cgroup/blkio/cgroup.procs
    rmdir /sys/fs/cgroup/blkio/a
    grep blkio /proc/cgroups
done
##TESTCASE_END:

Fixes: db37a34c56 ("block, bfq: get a ref to a group when adding it to a service tree")
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Dmitry Monakhov <dmtrmonakhov@yandex-team.ru>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-08-18 07:48:08 -07:00
Paolo Valente
db37a34c56 block, bfq: get a ref to a group when adding it to a service tree
BFQ schedules generic entities, which may represent either bfq_queues
or groups of bfq_queues. When an entity is inserted into a service
tree, a reference must be taken, to make sure that the entity does not
disappear while still referred in the tree. Unfortunately, such a
reference is mistakenly taken only if the entity represents a
bfq_queue. This commit takes a reference also in case the entity
represents a group.

Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Chris Evich <cevich@redhat.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-02-03 06:58:15 -07:00
Paolo Valente
33a16a9804 block, bfq: extend incomplete name of field on_st
The flag on_st in the bfq_entity data structure is true if the entity
is on a service tree or is in service. Yet the name of the field,
confusingly, does not mention the second, very important case. Extend
the name to mention the second case too.

Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-02-03 06:58:15 -07:00
Wen Yang
554d21efb0 block, bfq: improve arithmetic division in bfq_delta()
do_div() does a 64-by-32 division. Use div64_ul() instead of it
if the divisor is unsigned long, to avoid truncation to 32-bit.
And as a nice side effect also cleans up the function a bit.

Signed-off-by: Wen Yang <wenyang@linux.alibaba.com>
Cc: Paolo Valente <paolo.valente@linaro.org>
Cc: Jens Axboe <axboe@fb.com>
Cc: linux-block@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2020-01-22 10:34:11 -07:00
Fam Zheng
e9d3c866bf bfq: Fix the missing barrier in __bfq_entity_update_weight_prio
The comment of bfq_group_set_weight says the reading of prio_changed
should happen before the reading of weight, but a memory barrier is
missing here. Add it now, to match the smp_wmb() there.

Signed-off-by: Fam Zheng <zhengfeiran@bytedance.com>
Reviewed-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-09-06 14:33:48 -06:00
Christoph Hellwig
a497ee34a4 block: switch all files cleared marked as GPLv2 or later to SPDX tags
All these files have some form of the usual GPLv2 or later boilerplate.
Switch them to use SPDX tags instead.

Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-30 16:11:59 -06:00
Jens Axboe
5c61ee2cd5 Linux 5.1-rc6
-----BEGIN PGP SIGNATURE-----
 
 iQFSBAABCAA8FiEEq68RxlopcLEwq+PEeb4+QwBBGIYFAly8rGYeHHRvcnZhbGRz
 QGxpbnV4LWZvdW5kYXRpb24ub3JnAAoJEHm+PkMAQRiGmZMH/1IRB0E1Qmzz8yzw
 wj79UuRGYPqxDDSWW+wNc8sU4Ic7iYirn9APHAztCdQqsjmzU/OVLfSa3JhdBe5w
 THo7pbGKBqEDcWnKfNk/21jXFNLZ1vr9BoQv2DGU2MMhHAyo/NZbalo2YVtpQPmM
 OCRth5n+LzvH7rGrX7RYgWu24G9l3NMfgtaDAXBNXesCGFAjVRrdkU5CBAaabvtU
 4GWh/nnutndOOLdByL3x+VZ3H3fIBnbNjcIGCglvvqzk7h3hrfGEl4UCULldTxcM
 IFsfMUhSw1ENy7F6DHGbKIG90cdCJcrQ8J/ziEzjj/KLGALluutfFhVvr6YCM2J6
 2RgU8CY=
 =CfY1
 -----END PGP SIGNATURE-----

Merge tag 'v5.1-rc6' into for-5.2/block

Pull in v5.1-rc6 to resolve two conflicts. One is in BFQ, in just a
comment, and is trivial. The other one is a conflict due to a later fix
in the bio multi-page work, and needs a bit more care.

* tag 'v5.1-rc6': (770 commits)
  Linux 5.1-rc6
  block: make sure that bvec length can't be overflow
  block: kill all_q_node in request_queue
  x86/cpu/intel: Lower the "ENERGY_PERF_BIAS: Set to normal" message's log priority
  coredump: fix race condition between mmget_not_zero()/get_task_mm() and core dumping
  mm/kmemleak.c: fix unused-function warning
  init: initialize jump labels before command line option parsing
  kernel/watchdog_hld.c: hard lockup message should end with a newline
  kcov: improve CONFIG_ARCH_HAS_KCOV help text
  mm: fix inactive list balancing between NUMA nodes and cgroups
  mm/hotplug: treat CMA pages as unmovable
  proc: fixup proc-pid-vm test
  proc: fix map_files test on F29
  mm/vmstat.c: fix /proc/vmstat format for CONFIG_DEBUG_TLBFLUSH=y CONFIG_SMP=n
  mm/memory_hotplug: do not unlock after failing to take the device_hotplug_lock
  mm: swapoff: shmem_unuse() stop eviction without igrab()
  mm: swapoff: take notice of completion sooner
  mm: swapoff: remove too limiting SWAP_UNUSE_MAX_TRIES
  mm: swapoff: shmem_find_swap_entries() filter out other types
  slab: store tagged freelist for off-slab slabmgmt
  ...

Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-22 09:47:36 -06:00
Paolo Valente
eed47d19d9 block, bfq: fix use after free in bfq_bfqq_expire
The function bfq_bfqq_expire() invokes the function
__bfq_bfqq_expire(), and the latter may free the in-service bfq-queue.
If this happens, then no other instruction of bfq_bfqq_expire() must
be executed, or a use-after-free will occur.

Basing on the assumption that __bfq_bfqq_expire() invokes
bfq_put_queue() on the in-service bfq-queue exactly once, the queue is
assumed to be freed if its refcounter is equal to one right before
invoking __bfq_bfqq_expire().

But, since commit 9dee8b3b05 ("block, bfq: fix queue removal from
weights tree") this assumption is false. __bfq_bfqq_expire() may also
invoke bfq_weights_tree_remove() and, since commit 9dee8b3b05
("block, bfq: fix queue removal from weights tree"), also
the latter function may invoke bfq_put_queue(). So __bfq_bfqq_expire()
may invoke bfq_put_queue() twice, and this is the actual case where
the in-service queue may happen to be freed.

To address this issue, this commit moves the check on the refcounter
of the queue right around the last bfq_put_queue() that may be invoked
on the queue.

Fixes: 9dee8b3b05 ("block, bfq: fix queue removal from weights tree")
Reported-by: Dmitrii Tcvetkov <demfloro@demfloro.ru>
Reported-by: Douglas Anderson <dianders@chromium.org>
Tested-by: Dmitrii Tcvetkov <demfloro@demfloro.ru>
Tested-by: Douglas Anderson <dianders@chromium.org>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-10 07:54:38 -06:00
Angelo Ruocco
636b8fe86b block, bfq: fix some typos in comments
Some of the comments in the bfq files had typos. This patch fixes them.

Signed-off-by: Angelo Ruocco <angeloruocco90@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-08 10:05:43 -06:00
Paolo Valente
fb53ac6cd0 block, bfq: do not idle for lowest-weight queues
In most cases, it is detrimental for throughput to plug I/O dispatch
when the in-service bfq_queue becomes temporarily empty (plugging is
performed to wait for the possible arrival, soon, of new I/O from the
in-service queue). There is however a case where plugging is needed
for service guarantees. If a bfq_queue, say Q, has a higher weight
than some other active bfq_queue, and is sync, i.e., contains sync
I/O, then, to guarantee that Q does receive a higher share of the
throughput than other lower-weight queues, it is necessary to plug I/O
dispatch when Q remains temporarily empty while being served.

For this reason, BFQ performs I/O plugging when some active bfq_queue
has a higher weight than some other active bfq_queue. But this is
unnecessarily overkill. In fact, if the in-service bfq_queue actually
has a weight lower than or equal to the other queues, then the queue
*must not* be guaranteed a higher share of the throughput than the
other queues. So, not plugging I/O cannot cause any harm to the
queue. And can boost throughput.

Taking advantage of this fact, this commit does not plug I/O for sync
bfq_queues with a weight lower than or equal to the weights of the
other queues. Here is an example of the resulting throughput boost
with the dbench workload, which is particularly nasty for BFQ. With
the dbench test in the Phoronix suite, BFQ reaches its lowest total
throughput with 6 clients on a filesystem with journaling, in case the
journaling daemon has a higher weight than normal processes. Before
this commit, the total throughput was ~80 MB/sec on a PLEXTOR PX-256M5,
after this commit it is ~100 MB/sec.

Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01 08:15:39 -06:00
Konstantin Khlebnikov
42b1bd33dc block/bfq: fix ifdef for CONFIG_BFQ_GROUP_IOSCHED=y
Replace BFQ_GROUP_IOSCHED_ENABLED with CONFIG_BFQ_GROUP_IOSCHED.
Code under these ifdefs never worked, something might be broken.

Fixes: 0471559c2f ("block, bfq: add/remove entity weights correctly")
Fixes: 73d5811849 ("block, bfq: consider also ioprio classes in symmetry detection")
Reviewed-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Signed-off-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-04-01 06:56:15 -06:00
Paolo Valente
9dee8b3b05 block, bfq: fix queue removal from weights tree
bfq maintains an ordered list, through a red-black tree, of unique
weights of active bfq_queues. This list is used to detect whether there
are active queues with differentiated weights. The weight of a queue is
removed from the list when both the following two conditions become
true:

(1) the bfq_queue is flagged as inactive
(2) the has no in-flight request any longer;

Unfortunately, in the rare cases where condition (2) becomes true before
condition (1), the removal fails, because the function to remove the
weight of the queue (bfq_weights_tree_remove) is rightly invoked in the
path that deactivates the bfq_queue, but mistakenly invoked *before* the
function that actually performs the deactivation (bfq_deactivate_bfqq).

This commits moves the invocation of bfq_weights_tree_remove for
condition (1) to after bfq_deactivate_bfqq. As a consequence of this
move, it is necessary to add a further reference to the queue when the
weight of a queue is added, because the queue might otherwise be freed
before bfq_weights_tree_remove is invoked. This commit adds this
reference and makes all related modifications.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31 12:50:24 -07:00
Paolo Valente
73d5811849 block, bfq: consider also ioprio classes in symmetry detection
In asymmetric scenarios, i.e., when some bfq_queue or bfq_group needs to
be guaranteed a different bandwidth than other bfq_queues or bfq_groups,
these service guaranteed can be provided only by plugging I/O dispatch,
completely or partially, when the queue in service remains temporarily
empty. A case where asymmetry is particularly strong is when some active
bfq_queues belong to a higher-priority class than some other active
bfq_queues. Unfortunately, this important case is not considered at all
in the code for detecting asymmetric scenarios. This commit adds the
missing logic.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-31 12:50:23 -07:00
Paolo Valente
5bf859081f block, bfq: fix comments on __bfq_deactivate_entity
Comments on function __bfq_deactivate_entity contains two imprecise or
wrong statements:
1) The function performs the deactivation of the entity.
2) The function must be invoked only if the entity is on a service tree.

This commits replaces both statements with the correct ones:
1) The functions updates sched_data and service trees for the entity,
so as to represent entity as inactive (which is only part of the steps
needed for the deactivation of the entity).
2) The function must be invoked on every entity being deactivated.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2019-01-14 06:25:44 -07:00
Paolo Valente
ba7aeae553 block, bfq: fix decrement of num_active_groups
Since commit '2d29c9f89fcd ("block, bfq: improve asymmetric scenarios
detection")', if there are process groups with I/O requests waiting for
completion, then BFQ tags the scenario as 'asymmetric'. This detection
is needed for preserving service guarantees (for details, see comments
on the computation * of the variable asymmetric_scenario in the
function bfq_better_to_idle).

Unfortunately, commit '2d29c9f89fcd ("block, bfq: improve asymmetric
scenarios detection")' contains an error exactly in the updating of
the number of groups with I/O requests waiting for completion: if a
group has more than one descendant process, then the above number of
groups, which is renamed from num_active_groups to a more appropriate
num_groups_with_pending_reqs by this commit, may happen to be wrongly
decremented multiple times, namely every time one of the descendant
processes gets all its pending I/O requests completed.

A correct, complete solution should work as follows. Consider a group
that is inactive, i.e., that has no descendant process with pending
I/O inside BFQ queues. Then suppose that num_groups_with_pending_reqs
is still accounting for this group, because the group still has some
descendant process with some I/O request still in
flight. num_groups_with_pending_reqs should be decremented when the
in-flight request of the last descendant process is finally completed
(assuming that nothing else has changed for the group in the meantime,
in terms of composition of the group and active/inactive state of
child groups and processes). To accomplish this, an additional
pending-request counter must be added to entities, and must be
updated correctly.

To avoid this additional field and operations, this commit resorts to
the following tradeoff between simplicity and accuracy: for an
inactive group that is still counted in num_groups_with_pending_reqs,
this commit decrements num_groups_with_pending_reqs when the first
descendant process of the group remains with no request waiting for
completion.

This simplified scheme provides a fix to the unbalanced decrements
introduced by 2d29c9f89f. Since this error was also caused by lack
of comments on this non-trivial issue, this commit also adds related
comments.

Fixes: 2d29c9f89f ("block, bfq: improve asymmetric scenarios detection")
Reported-by: Steven Barrett <steven@liquorix.net>
Tested-by: Steven Barrett <steven@liquorix.net>
Tested-by: Lucjan Lucjanov <lucjan.lucjanov@gmail.com>
Reviewed-by: Federico Motta <federico@willer.it>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-12-07 07:40:07 -07:00
Federico Motta
98fa7a3e00 block, bfq: fix asymmetric scenarios detection
Since commit 2d29c9f89f ("block, bfq: improve asymmetric scenarios
detection"), a scenario is defined asymmetric when one of the
following conditions holds:
- active bfq_queues have different weights
- one or more group of entities (bfq_queue or other groups of entities)
  are active
bfq grants fairness and low latency also in such asymmetric scenarios,
by plugging the dispatching of I/O if the bfq_queue in service happens
to be temporarily idle. This plugging may lower throughput, so it is
important to do it only when strictly needed.

By mistake, in commit '2d29c9f89fcd' ("block, bfq: improve asymmetric
scenarios detection") the num_active_groups counter was firstly
incremented and subsequently decremented at any entity (group or
bfq_queue) weight change.

This is useless, because only transitions from active to inactive and
vice versa matter for that counter. Unfortunately this is also
incorrect in the following case: the entity at issue is a bfq_queue
and it is under weight raising. In fact in this case there is a
spurious increment of the num_active_groups counter.

This spurious increment may cause scenarios to be wrongly detected as
asymmetric, thus causing useless plugging and loss of throughput.

This commit fixes this issue by simply removing the above useless and
wrong increments and decrements.

Fixes: 2d29c9f89f ("block, bfq: improve asymmetric scenarios detection")
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Federico Motta <federico@willer.it>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-10-25 11:17:40 -06:00
Federico Motta
2d29c9f89f block, bfq: improve asymmetric scenarios detection
bfq defines as asymmetric a scenario where an active entity, say E
(representing either a single bfq_queue or a group of other entities),
has a higher weight than some other entities.  If the entity E does sync
I/O in such a scenario, then bfq plugs the dispatch of the I/O of the
other entities in the following situation: E is in service but
temporarily has no pending I/O request.  In fact, without this plugging,
all the times that E stops being temporarily idle, it may find the
internal queues of the storage device already filled with an
out-of-control number of extra requests, from other entities. So E may
have to wait for the service of these extra requests, before finally
having its own requests served. This may easily break service
guarantees, with E getting less than its fair share of the device
throughput.  Usually, the end result is that E gets the same fraction of
the throughput as the other entities, instead of getting more, according
to its higher weight.

Yet there are two other more subtle cases where E, even if its weight is
actually equal to or even lower than the weight of any other active
entities, may get less than its fair share of the throughput in case the
above I/O plugging is not performed:
1. other entities issue larger requests than E;
2. other entities contain more active child entities than E (or in
   general tend to have more backlog than E).

In the first case, other entities may get more service than E because
they get larger requests, than those of E, served during the temporary
idle periods of E.  In the second case, other entities get more service
because, by having many child entities, they have many requests ready
for dispatching while E is temporarily idle.

This commit addresses this issue by extending the definition of
asymmetric scenario: a scenario is asymmetric when
- active entities representing bfq_queues have differentiated weights,
  as in the original definition
or (inclusive)
- one or more entities representing groups of entities are active.

This broader definition makes sure that I/O plugging will be performed
in all the above cases, provided that there is at least one active
group.  Of course, this definition is very coarse, so it will trigger
I/O plugging also in cases where it is not needed, such as, e.g.,
multiple active entities with just one child each, and all with the same
I/O-request size.  The reason for this coarse definition is just that a
finer-grained definition would be rather heavy to compute.

On the opposite end, even this new definition does not trigger I/O
plugging in all cases where there is no active group, and all bfq_queues
have the same weight.  So, in these cases some unfairness may occur if
there are asymmetries in I/O-request sizes.  We made this choice because
I/O plugging may lower throughput, and probably a user that has not
created any group cares more about throughput than about perfect
fairness.  At any rate, as for possible applications that may care about
service guarantees, bfq already guarantees a high responsiveness and a
low latency to soft real-time applications automatically.

Signed-off-by: Federico Motta <federico@willer.it>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-10-13 15:40:00 -06:00
Paolo Valente
cbeb869a3d block, bfq: correctly charge and reset entity service in all cases
BFQ schedules entities (which represent either per-process queues or
groups of queues) as a function of their timestamps. In particular, as
a function of their (virtual) finish times. The finish time of an
entity is computed as a function of the budget assigned to the entity,
assuming, tentatively, that the entity, once in service, will receive
an amount of service equal to its budget. Then, when the entity is
expired because it finishes to be served, this finish time is updated
as a function of the actual service received by the entity. This
allows the entity to be correctly charged with only the service
received, and then to be correctly re-scheduled.

Yet an entity may receive service also while not being the entity in
service (in the scheduling environment of its parent entity), for
several reasons. If the entity remains with no backlog while receiving
this 'unofficial' service, then it is expired. Also on such an
expiration, the finish time of the entity should be updated to account
for only the service actually received by the entity. Unfortunately,
such an update is not performed for an entity expiring without being
the entity in service.

In a similar vein, the service counter of the entity in service is
reset when the entity is expired, to be ready to be used for next
service cycle. This reset too should be performed also in case an
entity is expired because it remains empty after receiving service
while not being the entity in service. But in this case the reset is
not performed.

This commit performs the above update of the finish time and reset of
the service received, also for an entity expiring while not being the
entity in service.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-14 13:06:02 -06:00
Paolo Valente
f812164869 block, bfq: improve code of bfq_bfqq_charge_time
bfq_bfqq_charge_time contains some lengthy and redundant code. This
commit trims and condenses that code.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-08-16 13:08:15 -06:00
Paolo Valente
e02a0aa26b block, bfq: always update the budget of an entity when needed
When the next child entity to serve changes for a given parent entity,
the budget of that parent entity must be updated accordingly.
Unfortunately, this update is not performed, by mistake, for the
entities that happen to switch from having no child entity to serve,
to having one child entity to serve.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-08-16 13:08:12 -06:00
Paolo Valente
9fae8dd59f block, bfq: fix service being wrongly set to zero in case of preemption
If
- a bfq_queue Q preempts another queue, because one request of Q
arrives in time,
- but, after this preemption, Q is not the queue that is set in service,
then Q->entity.service is set to 0 when Q is eventually set in
service. But Q should have continued receiving service with its old
budget (which is why preemption has occurred) and its old service.

This commit addresses this issue by resetting service on queue real
expiration.

Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-09 09:07:52 -06:00
Paolo Valente
0471559c2f block, bfq: add/remove entity weights correctly
To keep I/O throughput high as often as possible, BFQ performs
I/O-dispatch plugging (aka device idling) only when beneficial exactly
for throughput, or when needed for service guarantees (low latency,
fairness). An important case where the latter condition holds is when
the scenario is 'asymmetric' in terms of weights: i.e., when some
bfq_queue or whole group of queues has a higher weight, and thus has
to receive more service, than other queues or groups. Without dispatch
plugging, lower-weight queues/groups may unjustly steal bandwidth to
higher-weight queues/groups.

To detect asymmetric scenarios, BFQ checks some sufficient
conditions. One of these conditions is that active groups have
different weights. BFQ controls this condition by maintaining a
special set of unique weights of active groups
(group_weights_tree). To this purpose, in the function
bfq_active_insert/bfq_active_extract BFQ adds/removes the weight of a
group to/from this set.

Unfortunately, the function bfq_active_extract may happen to be
invoked also for a group that is still active (to preserve the correct
update of the next queue to serve, see comments in function
bfq_no_longer_next_in_service() for details). In this case, removing
the weight of the group makes the set group_weights_tree
inconsistent. Service-guarantee violations follow.

This commit addresses this issue by moving group_weights_tree
insertions from their previous location (in bfq_active_insert) into
the function __bfq_activate_entity, and by moving group_weights_tree
extractions from bfq_active_extract to when the entity that represents
a group remains throughly idle, i.e., with no request either enqueued
or dispatched.

Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-09 09:07:52 -06:00
Paolo Valente
8a8747dc01 block, bfq: limit sectors served with interactive weight raising
To maximise responsiveness, BFQ raises the weight, and performs device
idling, for bfq_queues associated with processes deemed as
interactive. In particular, weight raising has a maximum duration,
equal to the time needed to start a large application. If a
weight-raised process goes on doing I/O beyond this maximum duration,
it loses weight-raising.

This mechanism is evidently vulnerable to the following false
positives: I/O-bound applications that will go on doing I/O for much
longer than the duration of weight-raising. These applications have
basically no benefit from being weight-raised at the beginning of
their I/O. On the opposite end, while being weight-raised, these
applications
a) unjustly steal throughput to applications that may truly need
low latency;
b) make BFQ uselessly perform device idling; device idling results
in loss of device throughput with most flash-based storage, and may
increase latencies when used purposelessly.

This commit adds a countermeasure to reduce both the above
problems. To introduce this countermeasure, we provide the following
extra piece of information (full details in the comments added by this
commit). During the start-up of the large application used as a
reference to set the duration of weight-raising, involved processes
transfer at most ~110K sectors each. Accordingly, a process initially
deemed as interactive has no right to be weight-raised any longer,
once transferred 110K sectors or more.

Basing on this consideration, this commit early-ends weight-raising
for a bfq_queue if the latter happens to have received an amount of
service at least equal to 110K sectors (actually, a little bit more,
to keep a safety margin). I/O-bound applications that reach a high
throughput, such as file copy, get to this threshold much before the
allowed weight-raising period finishes. Thus this early ending of
weight-raising reduces the amount of time during which these
applications cause the problems described above.

Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
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>
2018-01-18 08:21:37 -07:00
Paolo Valente
7b8fa3b900 block, bfq: let a queue be merged only shortly after starting I/O
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>
2018-01-05 09:26:09 -07:00
Paolo Valente
24bfd19bb7 block, bfq: update blkio stats outside the scheduler lock
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>
2017-11-14 20:13:33 -07:00
Paolo Valente
24d90bb2f3 block, bfq: guarantee update_next_in_service always returns an eligible entity
If the function bfq_update_next_in_service is invoked as a consequence
of the activation or requeueing of an entity, say E, then it doesn't
invoke bfq_lookup_next_entity to get the next-in-service entity. In
contrast, it follows a shorter path: if E happens to be eligible (see
commit "bfq-sq-mq: make lookup_next_entity push up vtime on
expirations" for details on eligibility) and to have a lower virtual
finish time than the current candidate as next-in-service entity, then
E directly becomes the next-in-service entity. Unfortunately, there is
a corner case for which this shorter path makes
bfq_update_next_in_service choose a non eligible entity: it occurs if
both E and the current next-in-service entity happen to be non
eligible when bfq_update_next_in_service is invoked. In this case, E
is not set as next-in-service, and, since bfq_lookup_next_entity is
not invoked, the state of the parent entity is not updated so as to
end up with an eligible entity as the proper next-in-service entity.

In this respect, next-in-service is actually allowed to be non
eligible while some queue is in service: since no system-virtual-time
push-up can be performed in that case (see again commit "bfq-sq-mq:
make lookup_next_entity push up vtime on expirations" for details),
next-in-service is chosen, speculatively, as a function of the
possible value that the system virtual time may get after a push
up. But the correctness of the schedule breaks if next-in-service is
still a non eligible entity when it is time to set in service the next
entity. Unfortunately, this may happen in the above corner case.

This commit fixes this problem by making bfq_update_next_in_service
invoke bfq_lookup_next_entity not only if the above shorter path
cannot be taken, but also if the shorter path is taken but fails to
yield an eligible next-in-service entity.

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>
2017-08-31 08:20:31 -06:00
Paolo Valente
a02195ce86 block, bfq: remove direct switch to an entity in higher class
If the function bfq_update_next_in_service is invoked as a consequence
of the activation or requeueing of an entity, say E, and finds out
that E belongs to a higher-priority class than that of the current
next-in-service entity, then it sets next_in_service directly to
E. But this may lead to anomalous schedules, because E may happen not
be eligible for service, because its virtual start time is higher than
the system virtual time for its service tree.

This commit addresses this issue by simply removing this direct
switch.

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>
2017-08-31 08:20:29 -06:00
Paolo Valente
80294c3bbf block, bfq: make lookup_next_entity push up vtime on expirations
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>
2017-08-31 08:20:28 -06:00
Paolo Valente
46d556e6aa block, bfq: consider also in_service_entity to state whether an entity is active
Groups of BFQ queues are represented by generic entities in BFQ. When
a queue belonging to a parent entity is deactivated, the parent entity
may need to be deactivated too, in case the deactivated queue was the
only active queue for the parent entity. This deactivation may need to
be propagated upwards if the entity belongs, in its turn, to a further
higher-level entity, and so on. In particular, the upward propagation
of deactivation stops at the first parent entity that remains active
even if one of its child entities has been deactivated.

To decide whether the last non-deactivation condition holds for a
parent entity, BFQ checks whether the field next_in_service is still
not NULL for the parent entity, after the deactivation of one of its
child entity. If it is not NULL, then there are certainly other active
entities in the parent entity, and deactivations can stop.

Unfortunately, this check misses a corner case: if in_service_entity
is not NULL, then next_in_service may happen to be NULL, although the
parent entity is evidently active. This happens if: 1) the entity
pointed by in_service_entity is the only active entity in the parent
entity, and 2) according to the definition of next_in_service, the
in_service_entity cannot be considered as next_in_service. See the
comments on the definition of next_in_service for details on this
second point.

Hitting the above corner case causes crashes.

To address this issue, this commit:
1) Extends the above check on only next_in_service to controlling both
next_in_service and in_service_entity (if any of them is not NULL,
then no further deactivation is performed)
2) Improves the (important) comments on how next_in_service is defined
and updated; in particular it fixes a few rather obscure paragraphs

Reported-by: Eric Wheeler <bfq-sched@lists.ewheeler.net>
Reported-by: Rick Yiu <rick_yiu@htc.com>
Reported-by: Tom X Nguyen <tom81094@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Eric Wheeler <bfq-sched@lists.ewheeler.net>
Tested-by: Rick Yiu <rick_yiu@htc.com>
Tested-by: Laurentiu Nicola <lnicola@dend.ro>
Tested-by: Tom X Nguyen <tom81094@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2017-07-29 15:32:49 -06:00
Paolo Valente
6ab1d8da97 block, bfq: reset in_service_entity if it becomes idle
BFQ implements hierarchical scheduling by representing each group of
queues with a generic parent entity. For each parent entity, BFQ
maintains an in_service_entity pointer: if one of the child entities
happens to be in service, in_service_entity points to it.  The
resetting of these pointers happens only on queue expirations: when
the in-service queue is expired, i.e., stops to be the queue in
service, BFQ resets all in_service_entity pointers along the
parent-entity path from this queue to the root entity.

Functions handling the scheduling of entities assume, naturally, that
in-service entities are active, i.e., have pending I/O requests (or,
as a special case, even if they have no pending requests, they are
expected to receive a new request very soon, with the scheduler idling
the storage device while waiting for such an event). Unfortunately,
the above resetting scheme of the in_service_entity pointers may cause
this assumption to be violated.  For example, the in-service queue may
happen to remain without requests because of a request merge. In this
case the queue does become idle, and all related data structures are
updated accordingly. But in_service_entity still points to the queue
in the parent entity. This inconsistency may even propagate to
higher-level parent entities, if they happen to become idle as well,
as a consequence of the leaf queue becoming idle. For this queue and
parent entities, scheduling functions have an undefined behaviour,
and, as reported, may easily lead to kernel crashes or hangs.

This commit addresses this issue by simply resetting the
in_service_entity field also when it is detected to point to an entity
becoming idle (regardless of why the entity becomes idle).

Reported-by: Laurentiu Nicola <lnicola@dend.ro>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Tested-by: Laurentiu Nicola <lnicola@dend.ro>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2017-07-29 15:32:47 -06:00
Hou Tao
38c9140740 bfq: fix typos in comments about B-WF2Q+ algorithm
The start time of eligible entity should be less than or equal to
the current virtual time, and the entity in idle tree has a finish
time being greater than the current virtual time.

Signed-off-by: Hou Tao <houtao1@huawei.com>
Reviewed-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
2017-07-12 08:32:02 -06:00
Paolo Valente
431b17f9d5 block, bfq: don't change ioprio class for a bfq_queue on a service tree
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>
2017-07-03 16:50:00 -06:00
Paolo Valente
a66c38a171 block, bfq: use pointer entity->sched_data only if set
In the function __bfq_deactivate_entity, the pointer
entity->sched_data could happen to be used before being properly
initialized. This led to a NULL pointer dereference. This commit fixes
this bug by just using this pointer only where it is safe to do so.

Reported-by: Tom Harrison <l12436.tw@gmail.com>
Tested-by: Tom Harrison <l12436.tw@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
2017-05-10 07:39:43 -06:00
Paolo Valente
ea25da4808 block, bfq: split bfq-iosched.c into multiple source files
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>
2017-04-19 08:48:24 -06:00