The tg_may_dispatch() will call tg_with_in_bps_limit() and
tg_with_in_iops_limit() to check if we can dispatch a bio or
not, which will calculate bps/iops limitation multiple times.
But tg_may_dispatch() is always called under queue lock, which
means the bps/iops limitation will not change in tg_may_dispatch().
So we can calculate the bps/iops limitation only once, and pass
them to tg_with_in_bps_limit() and tg_with_in_iops_limit() to
avoid calculating bps/iops limitation repeatedly.
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The 'throtl_grp_quantum' and 'throtl_quantum' are both read-only
variables, thus better to use readable macros instead of static
variables, which can also save some spaces for .bss area.
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
adjust_inuse_and_calc_cost() is responsible for reducing the amount of
donated weights dynamically in period as the budget runs low. Because we
don't want to do full donation calculation in period, we keep latching up
inuse by INUSE_ADJ_STEP_PCT of the active weight of the cgroup until the
resulting hweight_inuse is satisfactory.
Unfortunately, the adj_step calculation was reading the active weight before
acquiring ioc->lock. Because the current thread could have lost race to
activate the iocg to another thread before entering this function, it may
read the active weight as zero before acquiring ioc->lock. When this
happens, the adj_step is calculated as zero and the incremental adjustment
loop becomes an infinite one.
Fix it by fetching the active weight after acquiring ioc->lock.
Fixes: b0853ab4a2 ("blk-iocost: revamp in-period donation snapbacks")
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Conceptually, root_iocg->hweight_donating must be less than WEIGHT_ONE but
all hweight calculations round up and thus it may end up >= WEIGHT_ONE
triggering divide-by-zero and other issues. Bound the value to avoid
surprises.
Fixes: e08d02aa5f ("blk-iocost: implement Andy's method for donation weight updates")
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
These functions can be used to enable iostat for partitions on devices
like md, bcache.
Signed-off-by: Song Liu <songliubraving@fb.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
NVMe shares tagset between fabric queue and admin queue or between
connect_q and NS queue, so hctx_may_queue() can be called to allocate
request for these queues.
Tags can be reserved in these tagset. Before error recovery, there is
often lots of in-flight requests which can't be completed, and new
reserved request may be needed in error recovery path. However,
hctx_may_queue() can always return false because there is too many
in-flight requests which can't be completed during error handling.
Finally, nothing can proceed.
Fix this issue by always allowing reserved tag allocation in
hctx_may_queue(). This is reasonable because reserved tags are supposed
to always be available.
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Cc: David Milburn <dmilburn@redhat.com>
Cc: Ewan D. Milne <emilne@redhat.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
scsi/sg.h is included more than once, Remove the one that isn't
necessary.
Signed-off-by: Tian Tao <tiantao6@hisilicon.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The test and the explaination of the patch as bellow.
Before test we added more debug code in blkg_async_bio_workfn():
int count = 0
if (bios.head && bios.head->bi_next) {
need_plug = true;
blk_start_plug(&plug);
}
while ((bio = bio_list_pop(&bios))) {
/*io_punt is a sysctl user interface to control the print*/
if(io_punt) {
printk("[%s:%d] bio start,size:%llu,%d count=%d plug?%d\n",
current->comm, current->pid, bio->bi_iter.bi_sector,
(bio->bi_iter.bi_size)>>9, count++, need_plug);
}
submit_bio(bio);
}
if (need_plug)
blk_finish_plug(&plug);
Steps that need to be set to trigger *PUNT* io before testing:
mount -t btrfs -o compress=lzo /dev/sda6 /btrfs
mount -t cgroup2 nodev /cgroup2
mkdir /cgroup2/cg3
echo "+io" > /cgroup2/cgroup.subtree_control
echo "8:0 wbps=1048576000" > /cgroup2/cg3/io.max #1000M/s
echo $$ > /cgroup2/cg3/cgroup.procs
Then use dd command to test btrfs PUNT io in current shell:
dd if=/dev/zero of=/btrfs/file bs=64K count=100000
Test hardware environment as below:
[root@localhost btrfs]# lscpu
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 32
On-line CPU(s) list: 0-31
Thread(s) per core: 2
Core(s) per socket: 8
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
With above debug code, test command and test environment, I did the
tests under 3 different system loads, which are triggered by stress:
1, Run 64 threads by command "stress -c 64 &"
[53615.975974] [kworker/u66:18:1490] bio start,size:45583056,8 count=0 plug?1
[53615.975980] [kworker/u66:18:1490] bio start,size:45583064,8 count=1 plug?1
[53615.975984] [kworker/u66:18:1490] bio start,size:45583072,8 count=2 plug?1
[53615.975987] [kworker/u66:18:1490] bio start,size:45583080,8 count=3 plug?1
[53615.975990] [kworker/u66:18:1490] bio start,size:45583088,8 count=4 plug?1
[53615.975993] [kworker/u66:18:1490] bio start,size:45583096,8 count=5 plug?1
... ...
[53615.977041] [kworker/u66:18:1490] bio start,size:45585480,8 count=303 plug?1
[53615.977044] [kworker/u66:18:1490] bio start,size:45585488,8 count=304 plug?1
[53615.977047] [kworker/u66:18:1490] bio start,size:45585496,8 count=305 plug?1
[53615.977050] [kworker/u66:18:1490] bio start,size:45585504,8 count=306 plug?1
[53615.977053] [kworker/u66:18:1490] bio start,size:45585512,8 count=307 plug?1
[53615.977056] [kworker/u66:18:1490] bio start,size:45585520,8 count=308 plug?1
[53615.977058] [kworker/u66:18:1490] bio start,size:45585528,8 count=309 plug?1
2, Run 32 threads by command "stress -c 32 &"
[50586.290521] [kworker/u66:6:32351] bio start,size:45806496,8 count=0 plug?1
[50586.290526] [kworker/u66:6:32351] bio start,size:45806504,8 count=1 plug?1
[50586.290529] [kworker/u66:6:32351] bio start,size:45806512,8 count=2 plug?1
[50586.290531] [kworker/u66:6:32351] bio start,size:45806520,8 count=3 plug?1
[50586.290533] [kworker/u66:6:32351] bio start,size:45806528,8 count=4 plug?1
[50586.290535] [kworker/u66:6:32351] bio start,size:45806536,8 count=5 plug?1
... ...
[50586.299640] [kworker/u66:5:32350] bio start,size:45808576,8 count=252 plug?1
[50586.299643] [kworker/u66:5:32350] bio start,size:45808584,8 count=253 plug?1
[50586.299646] [kworker/u66:5:32350] bio start,size:45808592,8 count=254 plug?1
[50586.299649] [kworker/u66:5:32350] bio start,size:45808600,8 count=255 plug?1
[50586.299652] [kworker/u66:5:32350] bio start,size:45808608,8 count=256 plug?1
[50586.299663] [kworker/u66:5:32350] bio start,size:45808616,8 count=257 plug?1
[50586.299665] [kworker/u66:5:32350] bio start,size:45808624,8 count=258 plug?1
[50586.299668] [kworker/u66:5:32350] bio start,size:45808632,8 count=259 plug?1
3, Don't run thread by stress
[50861.355246] [kworker/u66:19:32376] bio start,size:13544504,8 count=0 plug?0
[50861.355288] [kworker/u66:19:32376] bio start,size:13544512,8 count=0 plug?0
[50861.355322] [kworker/u66:19:32376] bio start,size:13544520,8 count=0 plug?0
[50861.355353] [kworker/u66:19:32376] bio start,size:13544528,8 count=0 plug?0
[50861.355392] [kworker/u66:19:32376] bio start,size:13544536,8 count=0 plug?0
[50861.355431] [kworker/u66:19:32376] bio start,size:13544544,8 count=0 plug?0
[50861.355468] [kworker/u66:19:32376] bio start,size:13544552,8 count=0 plug?0
[50861.355499] [kworker/u66:19:32376] bio start,size:13544560,8 count=0 plug?0
[50861.355532] [kworker/u66:19:32376] bio start,size:13544568,8 count=0 plug?0
[50861.355575] [kworker/u66:19:32376] bio start,size:13544576,8 count=0 plug?0
[50861.355618] [kworker/u66:19:32376] bio start,size:13544584,8 count=0 plug?0
[50861.355659] [kworker/u66:19:32376] bio start,size:13544592,8 count=0 plug?0
[50861.355740] [kworker/u66:0:32346] bio start,size:13544600,8 count=0 plug?1
[50861.355748] [kworker/u66:0:32346] bio start,size:13544608,8 count=1 plug?1
[50861.355962] [kworker/u66:2:32347] bio start,size:13544616,8 count=0 plug?0
[50861.356272] [kworker/u66:7:31962] bio start,size:13544624,8 count=0 plug?0
[50861.356446] [kworker/u66:7:31962] bio start,size:13544632,8 count=0 plug?0
[50861.356567] [kworker/u66:7:31962] bio start,size:13544640,8 count=0 plug?0
[50861.356707] [kworker/u66:19:32376] bio start,size:13544648,8 count=0 plug?0
[50861.356748] [kworker/u66:15:32355] bio start,size:13544656,8 count=0 plug?0
[50861.356825] [kworker/u66:17:31970] bio start,size:13544664,8 count=0 plug?0
Analysis of above 3 test results with different system load:
>From above test, we can see more and more continuous bios can be plugged
with system load increasing. When run "stress -c 64 &", 310 continuous
bios are plugged; When run "stress -c 32 &", 260 continuous bios are
plugged; When don't run stress, at most only 2 continuous bios are
plugged, in most cases, bio_list only contains one single bio.
How to explain above phenomenon:
We know, in submit_bio(), if the bio is a REQ_CGROUP_PUNT io, it will
queue a work to workqueue blkcg_punt_bio_wq. But when the workqueue is
scheduled, it depends on the system load. When system load is low, the
workqueue will be quickly scheduled, and the bio in bio_list will be
quickly processed in blkg_async_bio_workfn(), so there is less chance
that the same io submit thread can add multiple continuous bios to
bio_list before workqueue is scheduled to run. The analysis aligned with
above test "3".
When system load is high, there is some delay before the workqueue can
be scheduled to run, the higher the system load the greater the delay.
So there is more chance that the same io submit thread can add multiple
continuous bios to bio_list. Then when the workqueue is scheduled to run,
there are more continuous bios in bio_list, which will be processed in
blkg_async_bio_workfn(). The analysis aligned with above test "1" and "2".
According to test, we can get io performance improved with the patch,
especially when system load is higher. Another optimazition is to use
the plug only when bio_list contains at least 2 bios.
Signed-off-by: Xianting Tian <tian.xianting@h3c.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Like check_disk_changed, except that it does not call ->revalidate_disk
but leaves that to the caller.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Switch to the naming used by the other entries so that we can use the
QUEUE_RW_ENTRY helper.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Add two helpers macros to avoid boilerplate code for the queue sysfs
entries.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Now we usually free the hctx->sched_data by e->type->ops.exit_hctx(),
and no users will use blk_mq_sched_free_hctx_data() function.
Remove it.
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Discarding blocks and buffers under a mounted filesystem is hardly
anything admin wants to do. Usually it will confuse the filesystem and
sometimes the loss of buffer_head state (including b_private field) can
even cause crashes like:
BUG: unable to handle kernel NULL pointer dereference at 0000000000000008
PGD 0 P4D 0
Oops: 0002 [#1] SMP PTI
CPU: 4 PID: 203778 Comm: jbd2/dm-3-8 Kdump: loaded Tainted: G O --------- - - 4.18.0-147.5.0.5.h126.eulerosv2r9.x86_64 #1
Hardware name: Huawei RH2288H V3/BC11HGSA0, BIOS 1.57 08/11/2015
RIP: 0010:jbd2_journal_grab_journal_head+0x1b/0x40 [jbd2]
...
Call Trace:
__jbd2_journal_insert_checkpoint+0x23/0x70 [jbd2]
jbd2_journal_commit_transaction+0x155f/0x1b60 [jbd2]
kjournald2+0xbd/0x270 [jbd2]
So if we don't have block device open with O_EXCL already, claim the
block device while we truncate buffer cache. This makes sure any
exclusive block device user (such as filesystem) cannot operate on the
device while we are discarding buffer cache.
Reported-by: Ye Bin <yebin10@huawei.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Christoph Hellwig <hch@lst.de>
[axboe: fix !CONFIG_BLOCK error in truncate_bdev_range()]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
High CPU utilization on "native_queued_spin_lock_slowpath" due to lock
contention is possible for mq-deadline and bfq IO schedulers
when nr_hw_queues is more than one.
It is because kblockd work queue can submit IO from all online CPUs
(through blk_mq_run_hw_queues()) even though only one hctx has pending
commands.
The elevator callback .has_work for mq-deadline and bfq scheduler considers
pending work if there are any IOs on request queue but it does not account
hctx context.
Add a per-hctx 'elevator_queued' count to the hctx to avoid triggering
the elevator even though there are no requests queued.
[jpg: Relocated atomic_dec() in dd_dispatch_request(), update commit message per Kashyap]
Signed-off-by: Kashyap Desai <kashyap.desai@broadcom.com>
Signed-off-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For when using a shared sbitmap, no longer should the number of active
request queues per hctx be relied on for when judging how to share the tag
bitmap.
Instead maintain the number of active request queues per tag_set, and make
the judgement based on that.
Originally-from: Kashyap Desai <kashyap.desai@broadcom.com>
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Don Brace<don.brace@microsemi.com> #SCSI resv cmds patches used
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The per-hctx nr_active value can no longer be used to fairly assign a share
of tag depth per request queue for when using a shared sbitmap, as it does
not consider that the tags are shared tags over all hctx's.
For this case, record the nr_active_requests per request_queue, and make
the judgement based on that value.
Co-developed-with: Kashyap Desai <kashyap.desai@broadcom.com>
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Don Brace<don.brace@microsemi.com> #SCSI resv cmds patches used
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk-mq.h and blk-mq-tag.h include on each other, which is less than ideal.
Locate hctx_may_queue() to blk-mq.h, as it is not really tag specific code.
In this way, we can drop the blk-mq-tag.h include of blk-mq.h
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Some SCSI HBAs (such as HPSA, megaraid, mpt3sas, hisi_sas_v3 ..) support
multiple reply queues with single hostwide tags.
In addition, these drivers want to use interrupt assignment in
pci_alloc_irq_vectors(PCI_IRQ_AFFINITY). However, as discussed in [0],
CPU hotplug may cause in-flight IO completion to not be serviced when an
interrupt is shutdown. That problem is solved in commit bf0beec060
("blk-mq: drain I/O when all CPUs in a hctx are offline").
However, to take advantage of that blk-mq feature, the HBA HW queuess are
required to be mapped to that of the blk-mq hctx's; to do that, the HBA HW
queues need to be exposed to the upper layer.
In making that transition, the per-SCSI command request tags are no
longer unique per Scsi host - they are just unique per hctx. As such, the
HBA LLDD would have to generate this tag internally, which has a certain
performance overhead.
However another problem is that blk-mq assumes the host may accept
(Scsi_host.can_queue * #hw queue) commands. In commit 6eb045e092 ("scsi:
core: avoid host-wide host_busy counter for scsi_mq"), the Scsi host busy
counter was removed, which would stop the LLDD being sent more than
.can_queue commands; however, it should still be ensured that the block
layer does not issue more than .can_queue commands to the Scsi host.
To solve this problem, introduce a shared sbitmap per blk_mq_tag_set,
which may be requested at init time.
New flag BLK_MQ_F_TAG_HCTX_SHARED should be set when requesting the
tagset to indicate whether the shared sbitmap should be used.
Even when BLK_MQ_F_TAG_HCTX_SHARED is set, a full set of tags and requests
are still allocated per hctx; the reason for this is that if tags and
requests were only allocated for a single hctx - like hctx0 - it may break
block drivers which expect a request be associated with a specific hctx,
i.e. not always hctx0. This will introduce extra memory usage.
This change is based on work originally from Ming Lei in [1] and from
Bart's suggestion in [2].
[0] https://lore.kernel.org/linux-block/alpine.DEB.2.21.1904051331270.1802@nanos.tec.linutronix.de/
[1] https://lore.kernel.org/linux-block/20190531022801.10003-1-ming.lei@redhat.com/
[2] https://lore.kernel.org/linux-block/ff77beff-5fd9-9f05-12b6-826922bace1f@huawei.com/T/#m3db0a602f095cbcbff27e9c884d6b4ae826144be
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Don Brace<don.brace@microsemi.com> #SCSI resv cmds patches used
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Introduce pointers for the blk_mq_tags regular and reserved bitmap tags,
with the goal of later being able to use a common shared tag bitmap across
all HW contexts in a set.
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Don Brace<don.brace@microsemi.com> #SCSI resv cmds patches used
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pass hctx/tagset flags argument down to blk_mq_init_tags() and
blk_mq_free_tags() for selective init/free.
For now, make it include the alloc policy flag, which can be evaluated
when needed (in blk_mq_init_tags()).
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since the tags are allocated in blk_mq_init_tags(), it's better practice
to free in that same function upon error, rather than a callee which is to
init the bitmap tags (blk_mq_init_tags()).
[jpg: Split from an earlier patch with a new commit message]
Signed-off-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The function does not set the depth, but rather transitions from
shared to non-shared queues and vice versa.
So rename it to blk_mq_update_tag_set_shared() to better reflect
its purpose.
[jpg: take out some unrelated changes in blk_mq_init_bitmap_tags()]
Signed-off-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
BLK_MQ_F_TAG_SHARED actually means that tags is shared among request
queues, all of which should belong to LUNs attached to same HBA.
So rename it to make the point explicitly.
[jpg: rebase a few times, add rnbd-clt.c change]
Suggested-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: John Garry <john.garry@huawei.com>
Tested-by: Douglas Gilbert <dgilbert@interlog.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Only virtio_blk and xen-blkfront set the revalidate argument to true,
and both do not implement the ->revalidate_disk method. So switch
to the helper that just updates the size instead.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Replace bd_invalidate with a new BDEV_NEED_PART_SCAN flag in a bd_flags
variable to better describe the condition.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Reviewed-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Remove a duplicative condition to remove below cppcheck warnings:
"warning: Redundant condition: sched_allow_merge. '!A || (A && B)' is
equivalent to '!A || B' [redundantCondition]"
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If WRITE_ZERO/WRITE_SAME operation is not supported by the storage,
blk_cloned_rq_check_limits() will return IO error which will cause
device-mapper to fail the paths.
Instead, if the queue limit is set to 0, return BLK_STS_NOTSUPP.
BLK_STS_NOTSUPP will be ignored by device-mapper and will not fail the
paths.
Suggested-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Ritika Srivastava <ritika.srivastava@oracle.com>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
These are really cheap to collect and can be useful in debugging iocost
behavior. Add them as debug stats for now.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When an iocg accumulates too much vtime or gets deactivated, we throw away
some vtime, which lowers the overall device utilization. As the exact amount
which is being thrown away is known, we can compensate by accelerating the
vrate accordingly so that the extra vtime generated in the current period
matches what got lost.
This significantly improves work conservation when involving high weight
cgroups with intermittent and bursty IO patterns.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A low weight iocg can amass a large amount of debt, for example, when
anonymous memory gets reclaimed aggressively. If the system has a lot of
memory paired with a slow IO device, the debt can span multiple seconds or
more. If there are no other subsequent IO issuers, the in-debt iocg may end
up blocked paying its debt while the IO device is idle.
This patch implements a mechanism to protect against such pathological
cases. If the device has been sufficiently idle for a substantial amount of
time, the debts are halved. The criteria are on the conservative side as we
want to resolve the rare extreme cases without impacting regular operation
by forgiving debts too readily.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Curently, iocost syncs the delay duration to the outstanding debt amount,
which seemed enough to protect the system from anon memory hogs. However,
that was mostly because the delay calcuation was using hweight_inuse which
quickly converges towards zero under debt for delay duration calculation,
often pusnishing debtors overly harshly for longer than deserved.
The previous patch fixed the delay calcuation and now the protection against
anonymous memory hogs isn't enough because the effect of delay is indirect
and non-linear and a huge amount of future debt can accumulate abruptly
while unthrottled.
This patch implements delay hysteresis so that delay is decayed
exponentially over time instead of getting cleared immediately as debt is
paid off. While the overall behavior is similar to the blk-cgroup
implementation used by blk-iolatency, a lot of the details are different and
due to the empirical nature of the mechanism, it's challenging to adapt the
mechanism for one controller without negatively impacting the other.
As the delay is gradually decayed now, there's no point in running it from
its own hrtimer. Periodic updates are now performed from ioc_timer_fn() and
the dedicated hrtimer is removed.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Debt handling had several issues.
* How much inuse a debtor carries wasn't clearly defined. inuse would be
driven down over time from not issuing IOs but it'd be better to clamp it
to minimum immediately once in debt.
* How much can be paid off was determined by hweight_inuse. As inuse was
driven down, the payment amount would fall together regardless of the
debtor's active weight. This means that the debtors were punished harshly.
* ioc_rqos_merge() wasn't calling blkcg_schedule_throttle() after
iocg_kick_delay().
This patch revamps debt handling so that
* Debt handling owns inuse for iocgs in debt and keeps them at zero.
* Payment amount is determined by hweight_active. This is more deterministic
and safer than hweight_inuse but still far from ideal in that it doesn't
factor in possible donations from other iocgs for debt payments. This
likely needs further improvements in the future.
* iocg_rqos_merge() now calls blkcg_schedule_throttle() as necessary.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Andy Newell <newella@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When the margin drops below the minimum on a donating iocg, donation is
immediately canceled in full. There are a couple shortcomings with the
current behavior.
* It's abrupt. A small temporary budget deficit can lead to a wide swing in
weight allocation and a large surplus.
* It's open coded in the issue path but not implemented for the merge path.
A series of merges at a low inuse can make the iocg incur debts and stall
incorrectly.
This patch reimplements in-period donation snapbacks so that
* inuse adjustment and cost calculations are factored into
adjust_inuse_and_calc_cost() which is called from both the issue and merge
paths.
* Snapbacks are more gradual. It occurs in quarter steps.
* A snapback triggers if the margin goes below the low threshold and is
lower than the budget at the time of the last adjustment.
* For the above, __propagate_weights() stores the margin in
iocg->saved_margin. Move iocg->last_inuse storing together into
__propagate_weights() for consistency.
* Full snapback is guaranteed when there are waiters.
* With precise donation and gradual snapbacks, inuse adjustments are now a
lot more effective and the value of scaling inuse on weight changes isn't
clear. Removed inuse scaling from weight_update().
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
iocost has various safety nets to combat inuse adjustment calculation
inaccuracies. With Andy's method implemented in transfer_surpluses(), inuse
adjustment calculations are now accurate and we can make donation amount
determinations accurate too.
* Stop keeping track of past usage history and using the maximum. Act on the
immediate usage information.
* Remove donation constraints defined by SURPLUS_* constants. Donate
whatever isn't used.
* Determine the donation amount so that the iocg will end up with
MARGIN_TARGET_PCT budget at the end of the coming period assuming the same
usage as the previous period. TARGET is set at 50% of period, which is the
previous maximum. This provides smooth convergence for most repetitive IO
patterns.
* Apply donation logic early at 20% budget. There's no risk in doing so as
the calculation is based on the delta between the current budget and the
target budget at the end of the coming period.
* Remove preemptive iocg activation for zero cost IOs. As donation can reach
near zero now, the mere activation doesn't provide any protection anymore.
In the unlikely case that this becomes a problem, the right solution is
assigning appropriate costs for such IOs.
This significantly improves the donation determination logic while also
simplifying it. Now all donations are immediate, exact and smooth.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Andy Newell <newella@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
iocost implements work conservation by reducing iocg->inuse and propagating
the adjustment upwards proportionally. However, while I knew the target
absolute hierarchical proportion - adjusted hweight_inuse, I couldn't figure
out how to determine the iocg->inuse adjustment to achieve that and
approximated the adjustment by scaling iocg->inuse using the proportion of
the needed hweight_inuse changes.
When nested, these scalings aren't accurate even when adjusting a single
node as the donating node also receives the benefit of the donated portion.
When multiple nodes are donating as they often do, they can be wildly wrong.
iocost employed various safety nets to combat the inaccuracies. There are
ample buffers in determining how much to donate, the adjustments are
conservative and gradual. While it can achieve a reasonable level of work
conservation in simple scenarios, the inaccuracies can easily add up leading
to significant loss of total work. This in turn makes it difficult to
closely cap vrate as vrate adjustment is needed to compensate for the loss
of work. The combination of inaccurate donation calculations and vrate
adjustments can lead to wide fluctuations and clunky overall behaviors.
Andy Newell devised a method to calculate the needed ->inuse updates to
achieve the target hweight_inuse's. The method is compatible with the
proportional inuse adjustment propagation which allows all hot path
operations to be local to each iocg.
To roughly summarize, Andy's method divides the tree into donating and
non-donating parts, calculates global donation rate which is used to
determine the target hweight_inuse for each node, and then derives per-level
proportions. There's non-trivial amount of math involved. Please refer to
the following pdfs for detailed descriptions.
https://drive.google.com/file/d/1PsJwxPFtjUnwOY1QJ5AeICCcsL7BM3bohttps://drive.google.com/file/d/1vONz1-fzVO7oY5DXXsLjSxEtYYQbOvsEhttps://drive.google.com/file/d/1WcrltBOSPN0qXVdBgnKm4mdp9FhuEFQN
This patch implements Andy's method in transfer_surpluses(). This makes the
donation calculations accurate per cycle and enables further improvements in
other parts of the donation logic.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Andy Newell <newella@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The way the surplus donation logic is structured isn't great. There are two
separate paths for starting/increasing donations and decreasing them making
the logic harder to follow and is prone to unnecessary behavior differences.
In preparation for improved donation handling, this patch restructures the
code so that
* All donors - new, increasing and decreasing - are funneled through the
same code path.
* The target donation calculation is factored into hweight_after_donation()
which is called once from the same spot for all possible donors.
* Actual inuse adjustment is factored into trasnfer_surpluses().
This change introduces a few behavior differences - e.g. donation amount
reduction now uses the max usage of the recent three periods just like new
and increasing donations, and inuse now gets adjusted upwards the same way
it gets downwards. These differences are unlikely to have severely negative
implications and the whole logic will be revamped soon.
This patch also removes two tracepoints. The existing TPs don't quite fit
the new implementation. A later patch will update and reinstate them.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Budget donations are inaccurate and could take multiple periods to converge.
To prevent triggering vrate adjustments while surplus transfers were
catching up, vrate adjustment was suppressed if donations were increasing,
which was indicated by non-zero nr_surpluses.
This entangling won't be necessary with the scheduled rewrite of donation
mechanism which will make it precise and immediate. Let's decouple the two
in preparation.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Instead of marking iocgs with surplus with a flag and filtering for them
while walking all active iocgs, build a surpluses list. This doesn't make
much difference now but will help implementing improved donation logic which
will iterate iocgs with surplus multiple times.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, iocg->usages[] which are used to guide inuse adjustments are
calculated from vtime deltas. This, however, assumes that the hierarchical
inuse weight at the time of calculation held for the entire period, which
often isn't true and can lead to significant errors.
Now that we have absolute usage information collected, we can derive
iocg->usages[] from iocg->local_stat.usage_us so that inuse adjustment
decisions are made based on actual absolute usage. The calculated usage is
clamped between 1 and WEIGHT_ONE and WEIGHT_ONE is also used to signal
saturation regardless of the current hierarchical inuse weight.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, iocost doesn't collect or expose any statistics punting off all
monitoring duties to drgn based iocost_monitor.py. While it works for some
scenarios, there are some usability and data availability challenges. For
example, accurate per-cgroup usage information can't be tracked by vtime
progression at all and the number available in iocg->usages[] are really
short-term snapshots used for control heuristics with possibly significant
errors.
This patch implements per-cgroup absolute usage stat counter and exposes it
through io.stat along with the current vrate. Usage stat collection and
flushing employ the same method as cgroup rstat on the active iocg's and the
only hot path overhead is preemption toggling and adding to a percpu
counter.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, debt handling requires only iocg->waitq.lock. In the future, we
want to adjust and propagate inuse changes depending on debt status. Let's
grab ioc->lock in debt handling paths in preparation.
* Because ioc->lock nests outside iocg->waitq.lock, the decision to grab
ioc->lock needs to be made before entering the critical sections.
* Add and use iocg_[un]lock() which handles the conditional double locking.
* Add @pay_debt to iocg_kick_waitq() so that debt payment happens only when
the caller grabbed both locks.
This patch is prepatory and the comments contain references to future
changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The margin handling was pretty inconsistent.
* ioc->margin_us and ioc->inuse_margin_vtime were used as vtime margin
thresholds. However, the two are in different units with the former
requiring conversion to vtime on use.
* iocg_kick_waitq() was using a quarter of WAITQ_TIMER_MARGIN_PCT of
period_us as the timer slack - ~1.2%. While iocg_kick_delay() was using a
quarter of ioc->margin_us - ~12.5%. There aren't strong reasons to use
different values for the two.
This patch cleans up margin and timer slack handling:
* vtime margins are now recorded in ioc->margins.{min, max} on period
duration changes and used consistently.
* Timer slack is now 1% of period_us and recorded in ioc->timer_slack_ns and
used consistently for iocg_kick_waitq() and iocg_kick_delay().
The only functional change is shortening of timer slack. No meaningful
visible change is expected.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
They are in microseconds and wrap in around 1.2 hours with u32. While
unlikely, confusions from wraparounds are still possible. We aren't saving
anything meaningful by keeping these u32. Let's make them u64.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
To improve weight donations, we want to able to scale inuse with a greater
accuracy and down below 1. Let's make non-hierarchical weights to use
WEIGHT_ONE based fixed point numbers too like hierarchical ones.
This doesn't cause any behavior changes yet.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We're gonna use HWEIGHT_WHOLE for regular weights too. Let's rename it to
WEIGHT_ONE.
Pure rename.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
