platform/kernel/linux-rpi.git
5 years agoblock: bypass blk_set_runtime_active for uninitialized q->dev
Stanley Chu [Thu, 12 Sep 2019 08:35:27 +0000 (16:35 +0800)]
block: bypass blk_set_runtime_active for uninitialized q->dev

Some devices may skip blk_pm_runtime_init() and have null pointer
in its request_queue->dev. For example, SCSI devices of UFS Well-Known
LUNs.

Currently the null pointer is checked by the user of
blk_set_runtime_active(), i.e., scsi_dev_type_resume(). It is better to
check it by blk_set_runtime_active() itself instead of by its users.

Signed-off-by: Stanley Chu <stanley.chu@mediatek.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agonull_blk: validate the number of devices
André Almeida [Wed, 11 Sep 2019 14:46:36 +0000 (11:46 -0300)]
null_blk: validate the number of devices

A negative number of devices is nonsensical, so change the type to
unsigned. If the number of devices is 0, it is impossible for userspace
to interact with the module, so refuse loading the driver for that case.

Signed-off-by: André Almeida <andrealmeid@collabora.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agonull_blk: fix module name at log message
André Almeida [Wed, 11 Sep 2019 14:46:35 +0000 (11:46 -0300)]
null_blk: fix module name at log message

The name of the module is "null_blk", not "null". Make `pr_info()` follow
the pattern of `pr_err()` log messages.

Signed-off-by: André Almeida <andrealmeid@collabora.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agodocs: block: null_blk: enhance document style
André Almeida [Wed, 11 Sep 2019 14:46:34 +0000 (11:46 -0300)]
docs: block: null_blk: enhance document style

Use proper ReST syntax for chapters. Add more information to enhance
standardization in the file and to make the rendering more homogeneous.
Add a SPDX identifier. Mark single-queue mode as deprecated.

Signed-off-by: André Almeida <andrealmeid@collabora.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoiocost_monitor: Report debt
Tejun Heo [Wed, 4 Sep 2019 19:45:56 +0000 (12:45 -0700)]
iocost_monitor: Report debt

Report debt and rename del_ms row to delay for consistency.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoiocost_monitor: Report more info with higher accuracy
Tejun Heo [Wed, 4 Sep 2019 19:45:55 +0000 (12:45 -0700)]
iocost_monitor: Report more info with higher accuracy

When outputting json:

* Don't truncate numbers.

* Report address of iocg to ease drilling down further.

When outputting table:

* Use math.ceil() for delay_ms so that small delays don't read as 0.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoiocost_monitor: Always use strings for json values
Tejun Heo [Wed, 4 Sep 2019 19:45:54 +0000 (12:45 -0700)]
iocost_monitor: Always use strings for json values

Json has limited accuracy for numbers and can silently truncate 64bit
values, which can be extremely confusing.  Let's consistently use
string encapsulated values for json output.

While at it, convert an unnecesary f-string to str().

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblk-iocost: Don't let merges push vtime into the future
Tejun Heo [Wed, 4 Sep 2019 19:45:53 +0000 (12:45 -0700)]
blk-iocost: Don't let merges push vtime into the future

Merges have the same problem that forced-bios had which is fixed by
the previous patch.  The cost of a merge is calculated at the time of
issue and force-advances vtime into the future.  Until global vtime
catches up, how the cgroup's hweight changes in the meantime doesn't
matter and it often leads to situations where the cost is calculated
at one hweight and paid at a very different one.  See the previous
patch for more details.

Fix it by never advancing vtime into the future for merges.  If budget
is available, vtime is advanced.  Otherwise, the cost is charged as
debt.

This brings merge cost handling in line with issue cost handling in
ioc_rqos_throttle().

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblk-iocost: Account force-charged overage in absolute vtime
Tejun Heo [Wed, 4 Sep 2019 19:45:52 +0000 (12:45 -0700)]
blk-iocost: Account force-charged overage in absolute vtime

Currently, when a bio needs to be force-charged and there isn't enough
budget, vtime is simply pushed into the future.  This means that the
cost of the whole bio is scaled using the current hweight and then
charged immediately.  Until the global vtime advances beyond this
future vtime, the cgroup won't be allowed to issue normal IOs.

This is incorrect and can lead to, for example, exploding vrate or
extended stalls if vrate range is constrained.  Consider the following
scenario.

1. A cgroup with a very low hweight runs out of budget.

2. A storm of swap-out happens on it.  All of them are scaled
   according to the current low hweight and charged to vtime pushing
   it to a far future.

3. All other cgroups go idle and now the above cgroup has access to
   the whole device.  However, because vtime is already wound using
   the past low hweight, what its current hweight is doesn't matter
   until global vtime catches up to the local vtime.

4. As a result, either vrate gets ramped up extremely or the IOs stall
   while the underlying device is idle.

This is because the hweight the overage is calculated at is different
from the hweight that it's being paid at.

Fix it by remembering the overage in absoulte vtime and continuously
paying with the actual budget according to the current hweight at each
period.

Note that non-forced bios which wait already remembers the cost in
absolute vtime.  This brings forced-bio accounting in line.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblk-iocost: Fix incorrect operation order during iocg free
Tejun Heo [Tue, 10 Sep 2019 16:15:25 +0000 (09:15 -0700)]
blk-iocost: Fix incorrect operation order during iocg free

ioc_pd_free() first cancels the hrtimers and then deactivates the
iocg.  However, the iocg timer can run inbetween and reschedule the
hrtimers which will end up running after the iocg is freed leading to
crashes like the following.

  general protection fault: 0000 [#1] SMP
  ...
  RIP: 0010:iocg_kick_delay+0xbe/0x1b0
  RSP: 0018:ffffc90003598ea0 EFLAGS: 00010046
  RAX: 1cee00fd69512b54 RBX: ffff8881bba48400 RCX: 00000000000003e8
  RDX: 0000000000000000 RSI: 0000000000000001 RDI: ffff8881bba48400
  RBP: 0000000000004e20 R08: 0000000000000002 R09: 00000000000003e8
  R10: 0000000000000000 R11: 0000000000000000 R12: ffffc90003598ef0
  R13: 00979f3810ad461f R14: ffff8881bba4b400 R15: 25439f950d26e1d1
  FS:  0000000000000000(0000) GS:ffff88885f800000(0000) knlGS:0000000000000000
  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00007f64328c7e40 CR3: 0000000002409005 CR4: 00000000003606e0
  DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
  DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
  Call Trace:
   <IRQ>
   iocg_delay_timer_fn+0x3d/0x60
   __hrtimer_run_queues+0xfe/0x270
   hrtimer_interrupt+0xf4/0x210
   smp_apic_timer_interrupt+0x5e/0x120
   apic_timer_interrupt+0xf/0x20
   </IRQ>

Fix it by canceling hrtimers after deactivating the iocg.

Fixes: 7caa47151ab2 ("blkcg: implement blk-iocost")
Reported-by: Dave Jones <davej@codemonkey.org.uk>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobfq: Add per-device weight
Fam Zheng [Wed, 28 Aug 2019 03:54:53 +0000 (11:54 +0800)]
bfq: Add per-device weight

This adds to BFQ the missing per-device weight interfaces:
blkio.bfq.weight_device on legacy and io.bfq.weight on unified. The
implementation pretty closely resembles what we had in CFQ and the parsing code
is basically reused.

Tests
=====

Using two cgroups and three block devices, having weights setup as:

Cgroup          test1           test2
============================================
default         100             500
sda             500             100
sdb             default         default
sdc             200             200

cgroup v1 runs
--------------

    sda.test1.out:   READ: bw=913MiB/s
    sda.test2.out:   READ: bw=183MiB/s

    sdb.test1.out:   READ: bw=213MiB/s
    sdb.test2.out:   READ: bw=1054MiB/s

    sdc.test1.out:   READ: bw=650MiB/s
    sdc.test2.out:   READ: bw=650MiB/s

cgroup v2 runs
--------------

    sda.test1.out:   READ: bw=915MiB/s
    sda.test2.out:   READ: bw=184MiB/s

    sdb.test1.out:   READ: bw=216MiB/s
    sdb.test2.out:   READ: bw=1069MiB/s

    sdc.test1.out:   READ: bw=621MiB/s
    sdc.test2.out:   READ: bw=622MiB/s

Signed-off-by: Fam Zheng <zhengfeiran@bytedance.com>
Acked-by: Tejun Heo <tj@kernel.org>
Reviewed-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobfq: Extract bfq_group_set_weight from bfq_io_set_weight_legacy
Fam Zheng [Wed, 28 Aug 2019 03:54:52 +0000 (11:54 +0800)]
bfq: Extract bfq_group_set_weight from bfq_io_set_weight_legacy

This function will be useful when we update weight from the soon-coming
per-device interface.

Signed-off-by: Fam Zheng <zhengfeiran@bytedance.com>
Reviewed-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobfq: Fix the missing barrier in __bfq_entity_update_weight_prio
Fam Zheng [Wed, 28 Aug 2019 03:54:51 +0000 (11:54 +0800)]
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>
5 years agoblock: fix elevator_get_by_features()
Jens Axboe [Fri, 6 Sep 2019 13:02:31 +0000 (07:02 -0600)]
block: fix elevator_get_by_features()

The lookup logic is broken - 'e' will never be NULL, even if the
list is empty. Maintain lookup hit in a separate variable instead.

Fixes: a0958ba7fcdc ("block: Improve default elevator selection")
Reported-by: Julia Lawall <julia.lawall@lip6.fr>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agosd: Set ELEVATOR_F_ZBD_SEQ_WRITE for ZBC disks
Damien Le Moal [Thu, 5 Sep 2019 09:51:35 +0000 (18:51 +0900)]
sd: Set ELEVATOR_F_ZBD_SEQ_WRITE for ZBC disks

Using the helper blk_queue_required_elevator_features(), set the
elevator feature ELEVATOR_F_ZBD_SEQ_WRITE as required for the request
queue of SCSI ZBC disks.

This feature requirement can always be satisfied as the mq-deadline
elevator is always selected for in-kernel compilation when
CONFIG_BLK_DEV_ZONED (zoned block device support) is enabled.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: Set ELEVATOR_F_ZBD_SEQ_WRITE for nullblk zoned disks
Damien Le Moal [Thu, 5 Sep 2019 09:51:34 +0000 (18:51 +0900)]
block: Set ELEVATOR_F_ZBD_SEQ_WRITE for nullblk zoned disks

Using the helper blk_queue_required_elevator_features(), set the
elevator feature ELEVATOR_F_ZBD_SEQ_WRITE as required for the request
queue of null_blk devices created with zoned mode enabled.

This feature requirement can always be satisfied as the mq-deadline
elevator is always selected for in-kernel compilation when
CONFIG_BLK_DEV_ZONED (zoned block device support) is enabled.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: Delay default elevator initialization
Damien Le Moal [Thu, 5 Sep 2019 09:51:33 +0000 (18:51 +0900)]
block: Delay default elevator initialization

When elevator_init_mq() is called from blk_mq_init_allocated_queue(),
the only information known about the device is the number of hardware
queues as the block device scan by the device driver is not completed
yet for most drivers. The device type and elevator required features
are not set yet, preventing to correctly select the default elevator
most suitable for the device.

This currently affects all multi-queue zoned block devices which default
to the "none" elevator instead of the required "mq-deadline" elevator.
These drives currently include host-managed SMR disks connected to a
smartpqi HBA and null_blk block devices with zoned mode enabled.
Upcoming NVMe Zoned Namespace devices will also be affected.

Fix this by adding the boolean elevator_init argument to
blk_mq_init_allocated_queue() to control the execution of
elevator_init_mq(). Two cases exist:
1) elevator_init = false is used for calls to
   blk_mq_init_allocated_queue() within blk_mq_init_queue(). In this
   case, a call to elevator_init_mq() is added to __device_add_disk(),
   resulting in the delayed initialization of the queue elevator
   after the device driver finished probing the device information. This
   effectively allows elevator_init_mq() access to more information
   about the device.
2) elevator_init = true preserves the current behavior of initializing
   the elevator directly from blk_mq_init_allocated_queue(). This case
   is used for the special request based DM devices where the device
   gendisk is created before the queue initialization and device
   information (e.g. queue limits) is already known when the queue
   initialization is executed.

Additionally, to make sure that the elevator initialization is never
done while requests are in-flight (there should be none when the device
driver calls device_add_disk()), freeze and quiesce the device request
queue before calling blk_mq_init_sched() in elevator_init_mq().

Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: Improve default elevator selection
Damien Le Moal [Thu, 5 Sep 2019 09:51:32 +0000 (18:51 +0900)]
block: Improve default elevator selection

For block devices that do not specify required features, preserve the
current default elevator selection (mq-deadline for single queue
devices, none for multi-queue devices). However, for devices specifying
required features (e.g. zoned block devices ELEVATOR_F_ZBD_SEQ_WRITE
feature), select the first available elevator providing the required
features.

In all cases, default to "none" if no elevator is available or if the
initialization of the default elevator fails.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: Introduce elevator features
Damien Le Moal [Thu, 5 Sep 2019 09:51:31 +0000 (18:51 +0900)]
block: Introduce elevator features

Introduce the definition of elevator features through the
elevator_features flags in the elevator_type structure. Each flag can
represent a feature supported by an elevator. The first feature defined
by this patch is support for zoned block device sequential write
constraint with the flag ELEVATOR_F_ZBD_SEQ_WRITE, which is implemented
by the mq-deadline elevator using zone write locking.

Other possible features are IO priorities, write hints, latency targets
or single-LUN dual-actuator disks (for which the elevator could maintain
one LBA ordered list per actuator).

The required_elevator_features field is also added to the request_queue
structure to allow a device driver to specify elevator feature flags
that an elevator must support for the correct operation of the device
(e.g. device drivers for zoned block devices can have the
ELEVATOR_F_ZBD_SEQ_WRITE flag as a required feature).
The helper function blk_queue_required_elevator_features() is
defined for setting this new field.

With these two new fields in place, the elevator functions
elevator_match() and elevator_find() are modified to allow a user to set
only an elevator with a set of features that satisfies the device
required features. Elevators not matching the device requirements are
not shown in the device sysfs queue/scheduler file to prevent their use.

The "none" elevator can always be selected as before.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: Change elevator_init_mq() to always succeed
Damien Le Moal [Thu, 5 Sep 2019 09:51:30 +0000 (18:51 +0900)]
block: Change elevator_init_mq() to always succeed

If the default elevator chosen is mq-deadline, elevator_init_mq() may
return an error if mq-deadline initialization fails, leading to
blk_mq_init_allocated_queue() returning an error, which in turn will
cause the block device initialization to fail and the device not being
exposed.

Instead of taking such extreme measure, handle mq-deadline
initialization failures in the same manner as when mq-deadline is not
available (no module to load), that is, default to the "none" scheduler.
With this change, elevator_init_mq() return type can be changed to void.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: Cleanup elevator_init_mq() use
Damien Le Moal [Thu, 5 Sep 2019 09:51:29 +0000 (18:51 +0900)]
block: Cleanup elevator_init_mq() use

Instead of checking a queue tag_set BLK_MQ_F_NO_SCHED flag before
calling elevator_init_mq() to make sure that the queue supports IO
scheduling, use the elevator.c function elv_support_iosched() in
elevator_init_mq(). This does not introduce any functional change but
ensure that elevator_init_mq() does the right thing based on the queue
settings.

Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agolightnvm: print error when target is not found
Minwoo Im [Thu, 5 Sep 2019 19:04:33 +0000 (21:04 +0200)]
lightnvm: print error when target is not found

If userspace requests target to be removed, nvm_remove_tgt() will
iterate the nvm_devices to find out the given target, but if not
found, then it should print out an error.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Updated output string and patch description.
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agolightnvm: introduce pr_fmt for the prefix nvm
Minwoo Im [Thu, 5 Sep 2019 19:04:32 +0000 (21:04 +0200)]
lightnvm: introduce pr_fmt for the prefix nvm

all the pr_() family can have this prefix by pr_fmt.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Javier González <javier@javigon.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoparide/pcd: need to check if cd->disk is null in pcd_detect
zhengbin [Tue, 13 Aug 2019 11:23:12 +0000 (19:23 +0800)]
paride/pcd: need to check if cd->disk is null in pcd_detect

If alloc_disk fails in pcd_init_units, cd->disk & pi are empty, we need
to check if cd->disk is null in pcd_detect.

Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: zhengbin <zhengbin13@huawei.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoparide/pcd: need to set queue to NULL before put_disk
zhengbin [Tue, 13 Aug 2019 11:27:40 +0000 (19:27 +0800)]
paride/pcd: need to set queue to NULL before put_disk

In pcd_init_units, if blk_mq_init_sq_queue fails, need to set queue to
NULL before put_disk, otherwise null-ptr-deref Read will occur.

put_disk
  kobject_put
    disk_release
      blk_put_queue(disk->queue)

Fixes: f0d176255401 ("paride/pcd: Fix potential NULL pointer dereference and mem leak")
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: zhengbin <zhengbin13@huawei.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoparide/pf: need to set queue to NULL before put_disk
zhengbin [Tue, 13 Aug 2019 08:59:44 +0000 (16:59 +0800)]
paride/pf: need to set queue to NULL before put_disk

In pf_init_units, if blk_mq_init_sq_queue fails, need to set queue to
NULL before put_disk, otherwise null-ptr-deref Read will occur.

put_disk
  kobject_put
    disk_release
      blk_put_queue(disk->queue)

Fixes: 77218ddf46d8 ("paride: convert pf to blk-mq")
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: zhengbin <zhengbin13@huawei.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoMerge branch 'md-next' of git://git.kernel.org/pub/scm/linux/kernel/git/song/md into...
Jens Axboe [Wed, 4 Sep 2019 00:00:22 +0000 (18:00 -0600)]
Merge branch 'md-next' of git://git./linux/kernel/git/song/md into for-5.4/block

Pull MD fixes from Song.

* 'md-next' of git://git.kernel.org/pub/scm/linux/kernel/git/song/md:
  md/raid5: use bio_end_sector to calculate last_sector
  md/raid1: fail run raid1 array when active disk less than one
  md raid0/linear: Mark array as 'broken' and fail BIOs if a member is gone

5 years agomd/raid5: use bio_end_sector to calculate last_sector
Guoqing Jiang [Tue, 3 Sep 2019 09:41:03 +0000 (11:41 +0200)]
md/raid5: use bio_end_sector to calculate last_sector

Use the common way to get last_sector.

Signed-off-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agomd/raid1: fail run raid1 array when active disk less than one
Yufen Yu [Tue, 3 Sep 2019 13:12:41 +0000 (21:12 +0800)]
md/raid1: fail run raid1 array when active disk less than one

When run test case:
  mdadm -CR /dev/md1 -l 1 -n 4 /dev/sd[a-d] --assume-clean --bitmap=internal
  mdadm -S /dev/md1
  mdadm -A /dev/md1 /dev/sd[b-c] --run --force

  mdadm --zero /dev/sda
  mdadm /dev/md1 -a /dev/sda

  echo offline > /sys/block/sdc/device/state
  echo offline > /sys/block/sdb/device/state
  sleep 5
  mdadm -S /dev/md1

  echo running > /sys/block/sdb/device/state
  echo running > /sys/block/sdc/device/state
  mdadm -A /dev/md1 /dev/sd[a-c] --run --force

mdadm run fail with kernel message as follow:
[  172.986064] md: kicking non-fresh sdb from array!
[  173.004210] md: kicking non-fresh sdc from array!
[  173.022383] md/raid1:md1: active with 0 out of 4 mirrors
[  173.022406] md1: failed to create bitmap (-5)

In fact, when active disk in raid1 array less than one, we
need to return fail in raid1_run().

Reviewed-by: NeilBrown <neilb@suse.de>
Signed-off-by: Yufen Yu <yuyufen@huawei.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agomd raid0/linear: Mark array as 'broken' and fail BIOs if a member is gone
Guilherme G. Piccoli [Tue, 3 Sep 2019 19:49:00 +0000 (16:49 -0300)]
md raid0/linear: Mark array as 'broken' and fail BIOs if a member is gone

Currently md raid0/linear are not provided with any mechanism to validate
if an array member got removed or failed. The driver keeps sending BIOs
regardless of the state of array members, and kernel shows state 'clean'
in the 'array_state' sysfs attribute. This leads to the following
situation: if a raid0/linear array member is removed and the array is
mounted, some user writing to this array won't realize that errors are
happening unless they check dmesg or perform one fsync per written file.
Despite udev signaling the member device is gone, 'mdadm' cannot issue the
STOP_ARRAY ioctl successfully, given the array is mounted.

In other words, no -EIO is returned and writes (except direct ones) appear
normal. Meaning the user might think the wrote data is correctly stored in
the array, but instead garbage was written given that raid0 does stripping
(and so, it requires all its members to be working in order to not corrupt
data). For md/linear, writes to the available members will work fine, but
if the writes go to the missing member(s), it'll cause a file corruption
situation, whereas the portion of the writes to the missing devices aren't
written effectively.

This patch changes this behavior: we check if the block device's gendisk
is UP when submitting the BIO to the array member, and if it isn't, we flag
the md device as MD_BROKEN and fail subsequent I/Os to that device; a read
request to the array requiring data from a valid member is still completed.
While flagging the device as MD_BROKEN, we also show a rate-limited warning
in the kernel log.

A new array state 'broken' was added too: it mimics the state 'clean' in
every aspect, being useful only to distinguish if the array has some member
missing. We rely on the MD_BROKEN flag to put the array in the 'broken'
state. This state cannot be written in 'array_state' as it just shows
one or more members of the array are missing but acts like 'clean', it
wouldn't make sense to write it.

With this patch, the filesystem reacts much faster to the event of missing
array member: after some I/O errors, ext4 for instance aborts the journal
and prevents corruption. Without this change, we're able to keep writing
in the disk and after a machine reboot, e2fsck shows some severe fs errors
that demand fixing. This patch was tested in ext4 and xfs filesystems, and
requires a 'mdadm' counterpart to handle the 'broken' state.

Cc: Song Liu <songliubraving@fb.com>
Reviewed-by: NeilBrown <neilb@suse.de>
Signed-off-by: Guilherme G. Piccoli <gpiccoli@canonical.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agoclosures: fix a race on wakeup from closure_sync
Kent Overstreet [Tue, 3 Sep 2019 13:25:45 +0000 (21:25 +0800)]
closures: fix a race on wakeup from closure_sync

The race was when a thread using closure_sync() notices cl->s->done == 1
before the thread calling closure_put() calls wake_up_process(). Then,
it's possible for that thread to return and exit just before
wake_up_process() is called - so we're trying to wake up a process that
no longer exists.

rcu_read_lock() is sufficient to protect against this, as there's an rcu
barrier somewhere in the process teardown path.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Acked-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: Fix an error code in bch_dump_read()
Dan Carpenter [Tue, 3 Sep 2019 13:25:44 +0000 (21:25 +0800)]
bcache: Fix an error code in bch_dump_read()

The copy_to_user() function returns the number of bytes remaining to be
copied, but the intention here was to return -EFAULT if the copy fails.

Fixes: cafe56359144 ("bcache: A block layer cache")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: add cond_resched() in __bch_cache_cmp()
Shile Zhang [Tue, 3 Sep 2019 13:25:43 +0000 (21:25 +0800)]
bcache: add cond_resched() in __bch_cache_cmp()

Read /sys/fs/bcache/<uuid>/cacheN/priority_stats can take very long
time with huge cache after long run.

Signed-off-by: Shile Zhang <shile.zhang@linux.alibaba.com>
Tested-by: Heitor Alves de Siqueira <halves@canonical.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoDocumentation:kernel-per-CPU-kthreads.txt: Remove reference to elevator=
Marcos Paulo de Souza [Tue, 3 Sep 2019 14:05:37 +0000 (08:05 -0600)]
Documentation:kernel-per-CPU-kthreads.txt: Remove reference to elevator=

This argument was not being considered since blk-mq was set by default,
so removed this documentation to avoid confusion.

Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Marcos Paulo de Souza <marcos.souza.org@gmail.com>
.txt file is now .rst

Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoDocumenation: switching-sched: Remove notes about elevator argument
Marcos Paulo de Souza [Tue, 3 Sep 2019 14:04:02 +0000 (08:04 -0600)]
Documenation: switching-sched: Remove notes about elevator argument

This argument was ignored since blk-mq was set as default, so remove it
from documentation.

Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Marcos Paulo de Souza <marcos.souza.org@gmail.com>
.txt file is now .rst

Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: elevator.c: Remove now unused elevator= argument
Marcos Paulo de Souza [Wed, 28 Aug 2019 01:19:27 +0000 (22:19 -0300)]
block: elevator.c: Remove now unused elevator= argument

Since the inclusion of blk-mq, elevator argument was not being
considered anymore, and it's utility died long with the legacy IO path,
now removed too.

Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Bob Liu <bob.liu@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Marcos Paulo de Souza <marcos.souza.org@gmail.com>
Fold with doc removal patch.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: mq-deadline: Fix queue restart handling
Damien Le Moal [Wed, 28 Aug 2019 04:40:20 +0000 (13:40 +0900)]
block: mq-deadline: Fix queue restart handling

Commit 7211aef86f79 ("block: mq-deadline: Fix write completion
handling") added a call to blk_mq_sched_mark_restart_hctx() in
dd_dispatch_request() to make sure that write request dispatching does
not stall when all target zones are locked. This fix left a subtle race
when a write completion happens during a dispatch execution on another
CPU:

CPU 0: Dispatch CPU1: write completion

dd_dispatch_request()
    lock(&dd->lock);
    ...
    lock(&dd->zone_lock); dd_finish_request()
    rq = find request lock(&dd->zone_lock);
    unlock(&dd->zone_lock);
     zone write unlock
unlock(&dd->zone_lock);
...
__blk_mq_free_request
                                      check restart flag (not set)
      -> queue not run
    ...
    if (!rq && have writes)
        blk_mq_sched_mark_restart_hctx()
    unlock(&dd->lock)

Since the dispatch context finishes after the write request completion
handling, marking the queue as needing a restart is not seen from
__blk_mq_free_request() and blk_mq_sched_restart() not executed leading
to the dispatch stall under 100% write workloads.

Fix this by moving the call to blk_mq_sched_mark_restart_hctx() from
dd_dispatch_request() into dd_finish_request() under the zone lock to
ensure full mutual exclusion between write request dispatch selection
and zone unlock on write request completion.

Fixes: 7211aef86f79 ("block: mq-deadline: Fix write completion handling")
Cc: stable@vger.kernel.org
Reported-by: Hans Holmberg <Hans.Holmberg@wdc.com>
Reviewed-by: Hans Holmberg <hans.holmberg@wdc.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agowriteback: don't access page->mapping directly in track_foreign_dirty TP
Tejun Heo [Fri, 30 Aug 2019 23:39:54 +0000 (16:39 -0700)]
writeback: don't access page->mapping directly in track_foreign_dirty TP

page->mapping may encode different values in it and page_mapping()
should always be used to access the mapping pointer.
track_foreign_dirty tracepoint was incorrectly accessing page->mapping
directly.  Use page_mapping() instead.  Also, add NULL checks while at
it.

Fixes: 3a8e9ac89e6a ("writeback: add tracepoints for cgroup foreign writebacks")
Reported-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoMerge branch 'nvme-5.4' of git://git.infradead.org/nvme into for-5.4/block
Jens Axboe [Fri, 30 Aug 2019 20:21:27 +0000 (14:21 -0600)]
Merge branch 'nvme-5.4' of git://git.infradead.org/nvme into for-5.4/block

Pull NVMe changes from Sagi:

"The nvme updates include:
 - ana log parse fix from Anton
 - nvme quirks support for Apple devices from Ben
 - fix missing bio completion tracing for multipath stack devices from
   Hannes and Mikhail
 - IP TOS settings for nvme rdma and tcp transports from Israel
 - rq_dma_dir cleanups from Israel
 - tracing for Get LBA Status command from Minwoo
 - Some nvme-tcp cleanups from Minwoo, Potnuri and Myself
 - Some consolidation between the fabrics transports for handling the CAP
   register
 - reset race with ns scanning fix for fabrics (move fabrics commands to
   a dedicated request queue with a different lifetime from the admin
   request queue)."

* 'nvme-5.4' of git://git.infradead.org/nvme: (30 commits)
  nvme-rdma: Use rq_dma_dir macro
  nvme-fc: Use rq_dma_dir macro
  nvme-pci: Tidy up nvme_unmap_data
  nvme: make fabrics command run on a separate request queue
  nvme-pci: Support shared tags across queues for Apple 2018 controllers
  nvme-pci: Add support for Apple 2018+ models
  nvme-pci: Add support for variable IO SQ element size
  nvme-pci: Pass the queue to SQ_SIZE/CQ_SIZE macros
  nvme: trace bio completion
  nvme-multipath: fix ana log nsid lookup when nsid is not found
  nvmet-tcp: Add TOS for tcp transport
  nvme-tcp: Add TOS for tcp transport
  nvme-tcp: Use struct nvme_ctrl directly
  nvme-rdma: Add TOS for rdma transport
  nvme-fabrics: Add type of service (TOS) configuration
  nvmet-tcp: fix possible memory leak
  nvmet-tcp: fix possible NULL deref
  nvmet: trace: parse Get LBA Status command in detail
  nvme: trace: parse Get LBA Status command in detail
  nvme: trace: support for Get LBA Status opcode parsed
  ...

5 years agowriteback: add tracepoints for cgroup foreign writebacks
Tejun Heo [Thu, 29 Aug 2019 22:47:19 +0000 (15:47 -0700)]
writeback: add tracepoints for cgroup foreign writebacks

cgroup foreign inode handling has quite a bit of heuristics and
internal states which sometimes makes it difficult to understand
what's going on.  Add tracepoints to improve visibility.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblkcg: add missing NULL check in ioc_cpd_alloc()
Tejun Heo [Fri, 30 Aug 2019 13:10:58 +0000 (06:10 -0700)]
blkcg: add missing NULL check in ioc_cpd_alloc()

ioc_cpd_alloc() forgot to check NULL return from kzalloc().  Add it.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: kbuild test robot <lkp@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agonvme-rdma: Use rq_dma_dir macro
Israel Rukshin [Wed, 28 Aug 2019 11:11:50 +0000 (14:11 +0300)]
nvme-rdma: Use rq_dma_dir macro

Remove code duplication.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-fc: Use rq_dma_dir macro
Israel Rukshin [Wed, 28 Aug 2019 11:11:49 +0000 (14:11 +0300)]
nvme-fc: Use rq_dma_dir macro

Remove code duplication.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-pci: Tidy up nvme_unmap_data
Israel Rukshin [Wed, 28 Aug 2019 11:11:48 +0000 (14:11 +0300)]
nvme-pci: Tidy up nvme_unmap_data

Remove pointless local variable and use rq_dma_dir macro.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme: make fabrics command run on a separate request queue
Sagi Grimberg [Sat, 3 Aug 2019 02:33:59 +0000 (19:33 -0700)]
nvme: make fabrics command run on a separate request queue

We have a fundamental issue that fabric commands use the admin_q.
The reason is, that admin-connect, register reads and writes and
admin commands cannot be guaranteed ordering while we are running
controller resets.

For example, when we reset a controller we perform:
1. disable the controller
2. teardown the admin queue
3. re-establish the admin queue
4. enable the controller

In order to perform (3), we need to unquiesce the admin queue, however
we may have some admin commands that are already pending on the
quiesced admin_q and will immediate execute when we unquiesce it before
we execute (4). The host must not send admin commands to the controller
before enabling the controller.

To fix this, we have the fabric commands (admin connect and property
get/set, but not I/O queue connect) use a separate fabrics_q and make
sure to quiesce the admin_q before we disable the controller, and
unquiesce it only after we enable the controller.

This fixes the error prints from nvmet in a controller reset storm test:
kernel: nvmet: got cmd 6 while CC.EN == 0 on qid = 0
Which indicate that the host is sending an admin command when the
controller is not enabled.

Reviewed-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-pci: Support shared tags across queues for Apple 2018 controllers
Benjamin Herrenschmidt [Wed, 7 Aug 2019 07:51:22 +0000 (17:51 +1000)]
nvme-pci: Support shared tags across queues for Apple 2018 controllers

Another issue with the Apple T2 based 2018 controllers seem to be
that they blow up (and shut the machine down) if there's a tag
collision between the IO queue and the Admin queue.

My suspicion is that they use our tags for their internal tracking
and don't mix them with the queue id. They also seem to not like
when tags go beyond the IO queue depth, ie 128 tags.

This adds a quirk that marks tags 0..31 of the IO queue reserved

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Acked-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-pci: Add support for Apple 2018+ models
Benjamin Herrenschmidt [Wed, 7 Aug 2019 07:51:21 +0000 (17:51 +1000)]
nvme-pci: Add support for Apple 2018+ models

Based on reverse engineering and original patch by

Paul Pawlowski <paul@mrarm.io>

This adds support for Apple weird implementation of NVME in their
2018 or later machines. It accounts for the twice-as-big SQ entries
for the IO queues, and the fact that only interrupt vector 0 appears
to function properly.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-pci: Add support for variable IO SQ element size
Benjamin Herrenschmidt [Wed, 7 Aug 2019 07:51:20 +0000 (17:51 +1000)]
nvme-pci: Add support for variable IO SQ element size

The size of a submission queue element should always be 6 (64 bytes)
by spec.

However some controllers such as Apple's are not properly implementing
the standard and require a different size.

This provides the ground work for the subsequent quirks for these
controllers.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-pci: Pass the queue to SQ_SIZE/CQ_SIZE macros
Benjamin Herrenschmidt [Wed, 7 Aug 2019 07:51:19 +0000 (17:51 +1000)]
nvme-pci: Pass the queue to SQ_SIZE/CQ_SIZE macros

This will make it easier to handle variable queue entry sizes
later. No functional change.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme: trace bio completion
Hannes Reinecke [Wed, 24 Jul 2019 13:47:55 +0000 (15:47 +0200)]
nvme: trace bio completion

When native multipathing is enabled we cannot enable blktrace for
the underlying paths, so any completion is never traced.

Signed-off-by: Hannes Reinecke <hare@suse.com>
[fixed-up by Mikhail for non-multipath-build]
Signed-off-by: Mikhail Skorzhinskii <mskorzhinskiy@solarflare.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-multipath: fix ana log nsid lookup when nsid is not found
Anton Eidelman [Fri, 16 Aug 2019 20:00:10 +0000 (13:00 -0700)]
nvme-multipath: fix ana log nsid lookup when nsid is not found

ANA log parsing invokes nvme_update_ana_state() per ANA group desc.
This updates the state of namespaces with nsids in desc->nsids[].

Both ctrl->namespaces list and desc->nsids[] array are sorted by nsid.
Hence nvme_update_ana_state() performs a single walk over ctrl->namespaces:
- if current namespace matches the current desc->nsids[n],
  this namespace is updated, and n is incremented.
- the process stops when it encounters the end of either
  ctrl->namespaces end or desc->nsids[]

In case desc->nsids[n] does not match any of ctrl->namespaces,
the remaining nsids following desc->nsids[n] will not be updated.
Such situation was considered abnormal and generated WARN_ON_ONCE.

However ANA log MAY contain nsids not (yet) found in ctrl->namespaces.
For example, lets consider the following scenario:
- nvme0 exposes namespaces with nsids = [2, 3] to the host
- a new namespace nsid = 1 is added dynamically
- also, a ANA topology change is triggered
- NS_CHANGED aen is generated and triggers scan_work
- before scan_work discovers nsid=1 and creates a namespace, a NOTICE_ANA
  aen was issues and ana_work receives ANA log with nsids=[1, 2, 3]

Result: ana_work fails to update ANA state on existing namespaces [2, 3]

Solution:
Change the way nvme_update_ana_state() namespace list walk
checks the current namespace against desc->nsids[n] as follows:
a) ns->head->ns_id < desc->nsids[n]: keep walking ctrl->namespaces.
b) ns->head->ns_id == desc->nsids[n]: match, update the namespace
c) ns->head->ns_id >= desc->nsids[n]: skip to desc->nsids[n+1]

This enables correct operation in the scenario described above.
This also allows ANA log to contain nsids currently invisible
to the host, i.e. inactive nsids.

Signed-off-by: Anton Eidelman <anton@lightbitslabs.com>
Reviewed-by: James Smart <james.smart@broadcom.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvmet-tcp: Add TOS for tcp transport
Israel Rukshin [Sun, 18 Aug 2019 09:08:55 +0000 (12:08 +0300)]
nvmet-tcp: Add TOS for tcp transport

Set the outgoing packets type of service (TOS) according to the
receiving TOS.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Suggested-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-tcp: Add TOS for tcp transport
Israel Rukshin [Sun, 18 Aug 2019 09:08:54 +0000 (12:08 +0300)]
nvme-tcp: Add TOS for tcp transport

TOS provide clients the ability to segregate traffic flows for
different type of data.
One of the TOS usage is bandwidth management which allows setting bandwidth
limits for QoS classes, e.g. 80% bandwidth to controllers at QoS class A
and 20% to controllers at QoS class B.

usage examples:
nvme connect --tos=0 --transport=tcp --traddr=10.0.1.1 --nqn=test-nvme

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-tcp: Use struct nvme_ctrl directly
Israel Rukshin [Sun, 18 Aug 2019 09:08:53 +0000 (12:08 +0300)]
nvme-tcp: Use struct nvme_ctrl directly

This patch doesn't change any functionality.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-rdma: Add TOS for rdma transport
Israel Rukshin [Sun, 18 Aug 2019 09:08:52 +0000 (12:08 +0300)]
nvme-rdma: Add TOS for rdma transport

For RDMA transports, TOS is an extension of IB QoS to provide clients
the ability to segregate traffic flows for different type of data.
RDMA CM abstract it for ULPs using rdma_set_service_type().
Internally, each traffic flow is represented by a connection with all of
its independent resources like that of a normal connection, and is
differentiated by service type. In other words, there can be multiple qp
connections between an IP pair and each supports a unique service type.

One of the TOS usage is bandwidth management which allows setting bandwidth
limits for QoS classes, e.g. 80% bandwidth to controllers at QoS class A
and 20% to controllers at QoS class B.

Note: In addition to the TOS configuration, QOS must be configured on the
relevant HCA on the target (send RDMA commands) and initiator to effect
the traffic.

usage examples:
nvme connect --tos=0 --transport=rdma --traddr=10.0.1.1 --nqn=test-nvme

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-fabrics: Add type of service (TOS) configuration
Israel Rukshin [Sun, 18 Aug 2019 09:08:51 +0000 (12:08 +0300)]
nvme-fabrics: Add type of service (TOS) configuration

TOS is user-defined and needs to be configured via nvme-cli.
It must be set before initiating any traffic and once set the TOS
cannot be changed.

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvmet-tcp: fix possible memory leak
Sagi Grimberg [Sat, 3 Aug 2019 03:29:11 +0000 (20:29 -0700)]
nvmet-tcp: fix possible memory leak

when we uninit a command in error flow we also need to
free an iovec if it was allocated.

Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvmet-tcp: fix possible NULL deref
Sagi Grimberg [Sat, 3 Aug 2019 03:23:38 +0000 (20:23 -0700)]
nvmet-tcp: fix possible NULL deref

We must only call sgl_free for sgl that we actually
allocated.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvmet: trace: parse Get LBA Status command in detail
Minwoo Im [Sun, 4 Aug 2019 07:50:50 +0000 (16:50 +0900)]
nvmet: trace: parse Get LBA Status command in detail

Four different fields are in CDWs of Get LBA Status command which means
it would be great if we can see in detail when tracing in target side
also.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme: trace: parse Get LBA Status command in detail
Minwoo Im [Sun, 4 Aug 2019 07:50:49 +0000 (16:50 +0900)]
nvme: trace: parse Get LBA Status command in detail

Four different fields are in CDWs of Get LBA Status command which means
it would be great if we can see in detail when tracing.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme: trace: support for Get LBA Status opcode parsed
Minwoo Im [Sun, 4 Aug 2019 07:50:48 +0000 (16:50 +0900)]
nvme: trace: support for Get LBA Status opcode parsed

This patch adds Get LBA Status command's opcode to the macro that is
used by the trace feature.  Now we can see "get_lba_status" instead of
the opcode value itself.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme: add Get LBA Status command opcode
Minwoo Im [Sun, 4 Aug 2019 07:50:47 +0000 (16:50 +0900)]
nvme: add Get LBA Status command opcode

NVMe 1.4 added Get LBA Status command with opcode 0x86.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvmet: fix data units read and written counters in SMART log
Tom Wu [Thu, 8 Aug 2019 02:22:36 +0000 (02:22 +0000)]
nvmet: fix data units read and written counters in SMART log

In nvme spec 1.3 there is a definition for data write/read counters
from SMART log, (See section 5.14.1.2):
This value is reported in thousands (i.e., a value of 1
corresponds to 1000 units of 512 bytes read) and is rounded up.

However, in nvme target where value is reported with actual units,
but not thousands of units as the spec requires.

Signed-off-by: Tom Wu <tomwu@mellanox.com>
Reviewed-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-tcp: support simple polling
Sagi Grimberg [Wed, 3 Jul 2019 21:08:04 +0000 (14:08 -0700)]
nvme-tcp: support simple polling

Simple polling support via socket busy_poll interface.
Although we do not shutdown interrupts but simply hammer
the socket poll, we can sometimes find completions faster
than the normal interrupt driven RX path.

We add per queue nr_cqe counter that resets every time
RX path is invoked such that .poll callback can return it
to stay consistent with the semantics.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme: tcp: selects CRYPTO_CRC32C for nvme-tcp
Minwoo Im [Sun, 14 Jul 2019 08:18:42 +0000 (17:18 +0900)]
nvme: tcp: selects CRYPTO_CRC32C for nvme-tcp

The tcp host module is now taking those APIs from crypto ahash:
(1) crypto_ahash_final()
(2) crypto_ahash_digest()
(3) crypto_alloc_ahash()

nvme-tcp should depends on CRYPTO_CRC32C.

Cc: Christoph Hellwig <hch@lst.de>
Cc: Keith Busch <kbusch@kernel.org>
Cc: Jens Axboe <axboe@fb.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme: don't pass cap to nvme_disable_ctrl
Sagi Grimberg [Tue, 23 Jul 2019 00:06:54 +0000 (17:06 -0700)]
nvme: don't pass cap to nvme_disable_ctrl

All seem to call it with ctrl->cap so no need to pass it
at all.

Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme: move sqsize setting to the core
Sagi Grimberg [Tue, 23 Jul 2019 00:06:53 +0000 (17:06 -0700)]
nvme: move sqsize setting to the core

nvme_enable_ctrl reads the cap register right after, so
no need to do that locally in the transport driver. Have
sqsize setting in nvme_init_identify.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-pci: set ctrl sqsize to the device q_depth
Sagi Grimberg [Thu, 22 Aug 2019 17:51:17 +0000 (10:51 -0700)]
nvme-pci: set ctrl sqsize to the device q_depth

Align with what the rest of the transports are doing.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme: have nvme_init_identify set ctrl->cap
Sagi Grimberg [Tue, 23 Jul 2019 00:06:52 +0000 (17:06 -0700)]
nvme: have nvme_init_identify set ctrl->cap

No need to use a stack cap variable.

Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-tcp: Use protocol specific operations while reading socket
Potnuri Bharat Teja [Mon, 8 Jul 2019 09:52:00 +0000 (15:22 +0530)]
nvme-tcp: Use protocol specific operations while reading socket

Using socket specific read_sock() calls instead of directly calling
tcp_read_sock() helps lld module registered handlers if any, to be called
from nvme-tcp host.
This patch therefore replaces the tcp_read_sock() with socket specific
prot_ops.

Signed-off-by: Potnuri Bharat Teja <bharat@chelsio.com>
Acked-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agonvme-tcp: cleanup nvme_tcp_recv_pdu
Sagi Grimberg [Fri, 19 Jul 2019 19:46:46 +0000 (12:46 -0700)]
nvme-tcp: cleanup nvme_tcp_recv_pdu

Can return directly in the switch statement

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
5 years agoblkcg: fix missing free on error path of blk_iocost_init()
Tejun Heo [Thu, 29 Aug 2019 15:53:06 +0000 (08:53 -0700)]
blkcg: fix missing free on error path of blk_iocost_init()

blk_iocost_init() forgot to free its percpu stat on the error path.
Fix it.

Fixes: 7caa47151ab2 ("blkcg: implement blk-iocost")
Reported-by: Hillf Danton <hdanton@sina.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblkcg: blk-iocost: predeclare used structs
Stephen Rothwell [Thu, 29 Aug 2019 15:43:34 +0000 (09:43 -0600)]
blkcg: blk-iocost: predeclare used structs

Fixes: 7caa47151ab2 ("blkcg: implement blk-iocost")
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblkcg: add tools/cgroup/iocost_coef_gen.py
Tejun Heo [Wed, 28 Aug 2019 22:06:00 +0000 (15:06 -0700)]
blkcg: add tools/cgroup/iocost_coef_gen.py

Add a script which can be used to generate device-specific iocost
linear model coefficients.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblkcg: add tools/cgroup/iocost_monitor.py
Tejun Heo [Wed, 28 Aug 2019 22:05:59 +0000 (15:05 -0700)]
blkcg: add tools/cgroup/iocost_monitor.py

Instead of mucking with debugfs and ->pd_stat(), add drgn based
monitoring script.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Omar Sandoval <osandov@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblkcg: implement blk-iocost
Tejun Heo [Wed, 28 Aug 2019 22:05:58 +0000 (15:05 -0700)]
blkcg: implement blk-iocost

This patchset implements IO cost model based work-conserving
proportional controller.

While io.latency provides the capability to comprehensively prioritize
and protect IOs depending on the cgroups, its protection is binary -
the lowest latency target cgroup which is suffering is protected at
the cost of all others.  In many use cases including stacking multiple
workload containers in a single system, it's necessary to distribute
IO capacity with better granularity.

One challenge of controlling IO resources is the lack of trivially
observable cost metric.  The most common metrics - bandwidth and iops
- can be off by orders of magnitude depending on the device type and
IO pattern.  However, the cost isn't a complete mystery.  Given
several key attributes, we can make fairly reliable predictions on how
expensive a given stream of IOs would be, at least compared to other
IO patterns.

The function which determines the cost of a given IO is the IO cost
model for the device.  This controller distributes IO capacity based
on the costs estimated by such model.  The more accurate the cost
model the better but the controller adapts based on IO completion
latency and as long as the relative costs across differents IO
patterns are consistent and sensible, it'll adapt to the actual
performance of the device.

Currently, the only implemented cost model is a simple linear one with
a few sets of default parameters for different classes of device.
This covers most common devices reasonably well.  All the
infrastructure to tune and add different cost models is already in
place and a later patch will also allow using bpf progs for cost
models.

Please see the top comment in blk-iocost.c and documentation for
more details.

v2: Rebased on top of RQ_ALLOC_TIME changes and folded in Rik's fix
    for a divide-by-zero bug in current_hweight() triggered by zero
    inuse_sum.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Andy Newell <newella@fb.com>
Cc: Josef Bacik <jbacik@fb.com>
Cc: Rik van Riel <riel@surriel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblk-mq: add optional request->alloc_time_ns
Tejun Heo [Wed, 28 Aug 2019 22:05:57 +0000 (15:05 -0700)]
blk-mq: add optional request->alloc_time_ns

There are currently two start time timestamps - start_time_ns and
io_start_time_ns.  The former marks the request allocation and and the
second issue-to-device time.  The planned io.weight controller needs
to measure the total time bios take to execute after it leaves rq_qos
including the time spent waiting for request to become available,
which can easily dominate on saturated devices.

This patch adds request->alloc_time_ns which records when the request
allocation attempt started.  As it isn't used for the usual stats,
make it optional behind CONFIG_BLK_RQ_ALLOC_TIME and
QUEUE_FLAG_RQ_ALLOC_TIME so that it can be compiled out when there are
no users and it's active only on queues which need it even when
compiled in.

v2: s/pre_start_time/alloc_time/ and add CONFIG_BLK_RQ_ALLOC_TIME
    gating as suggested by Jens.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblkcg: s/RQ_QOS_CGROUP/RQ_QOS_LATENCY/
Tejun Heo [Wed, 28 Aug 2019 22:05:56 +0000 (15:05 -0700)]
blkcg: s/RQ_QOS_CGROUP/RQ_QOS_LATENCY/

io.weight is gonna be another rq_qos cgroup mechanism.  Let's rename
RQ_QOS_CGROUP which is being used by io.latency to RQ_QOS_LATENCY in
preparation.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock/rq_qos: implement rq_qos_ops->queue_depth_changed()
Tejun Heo [Wed, 28 Aug 2019 22:05:55 +0000 (15:05 -0700)]
block/rq_qos: implement rq_qos_ops->queue_depth_changed()

wbt already gets queue depth changed notification through
wbt_set_queue_depth().  Generalize it into
rq_qos_ops->queue_depth_changed() so that other rq_qos policies can
easily hook into the events too.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock/rq_qos: add rq_qos_merge()
Tejun Heo [Wed, 28 Aug 2019 22:05:54 +0000 (15:05 -0700)]
block/rq_qos: add rq_qos_merge()

Add a merge hook for rq_qos.  This will be used by io.weight.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblkcg: separate blkcg_conf_get_disk() out of blkg_conf_prep()
Tejun Heo [Wed, 28 Aug 2019 22:05:53 +0000 (15:05 -0700)]
blkcg: separate blkcg_conf_get_disk() out of blkg_conf_prep()

Separate out blkcg_conf_get_disk() so that it can be used by blkcg
policy interface file input parsers before the policy is actually
enabled.  This doesn't introduce any functional changes.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblkcg: make ->cpd_init_fn() optional
Tejun Heo [Wed, 28 Aug 2019 22:05:52 +0000 (15:05 -0700)]
blkcg: make ->cpd_init_fn() optional

For policies which can do enough initialization from ->cpd_alloc_fn(),
make ->cpd_init_fn() optional.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblkcg: pass @q and @blkcg into blkcg_pol_alloc_pd_fn()
Tejun Heo [Wed, 28 Aug 2019 22:05:51 +0000 (15:05 -0700)]
blkcg: pass @q and @blkcg into blkcg_pol_alloc_pd_fn()

Instead of @node, pass in @q and @blkcg so that the alloc function has
more context.  This doesn't cause any behavior change and will be used
by io.weight implementation.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoMerge branch 'md-next' of git://git.kernel.org/pub/scm/linux/kernel/git/song/md into...
Jens Axboe [Tue, 27 Aug 2019 19:56:59 +0000 (13:56 -0600)]
Merge branch 'md-next' of git://git./linux/kernel/git/song/md into for-5.4/block

Pull MD fixes from Song.

* 'md-next' of git://git.kernel.org/pub/scm/linux/kernel/git/song/md:
  raid5 improve too many read errors msg by adding limits
  md: don't report active array_state until after revalidate_disk() completes.
  md: only call set_in_sync() when it is expected to succeed.

5 years agoraid5 improve too many read errors msg by adding limits
Nigel Croxon [Wed, 21 Aug 2019 13:27:08 +0000 (09:27 -0400)]
raid5 improve too many read errors msg by adding limits

Often limits can be changed by admin. When discussing such things
it helps if you can provide "self-sustained" facts. Also
sometimes the admin thinks he changed a limit, but it did not
take effect for some reason or he changed the wrong thing.

V3: Only pr_warn when Faulty is 0.
V2: Add read_errors value to pr_warn.

Signed-off-by: Nigel Croxon <ncroxon@redhat.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agomd: don't report active array_state until after revalidate_disk() completes.
NeilBrown [Tue, 20 Aug 2019 00:21:09 +0000 (10:21 +1000)]
md: don't report active array_state until after revalidate_disk() completes.

Until revalidate_disk() has completed, the size of a new md array will
appear to be zero.
So we shouldn't report, through array_state, that the array is active
until that time.
udev rules check array_state to see if the array is ready.  As soon as
it appear to be zero, fsck can be run.  If it find the size to be
zero, it will fail.

So add a new flag to provide an interlock between do_md_run() and
array_state_show().  This flag is set while do_md_run() is active and
it prevents array_state_show() from reporting that the array is
active.

Before do_md_run() is called, ->pers will be NULL so array is
definitely not active.
After do_md_run() is called, revalidate_disk() will have run and the
array will be completely ready.

We also move various sysfs_notify*() calls out of md_run() into
do_md_run() after MD_NOT_READY is cleared.  This ensure the
information is ready before the notification is sent.

Prior to v4.12, array_state_show() was called with the
mddev->reconfig_mutex held, which provided exclusion with do_md_run().

Note that MD_NOT_READY cleared twice.  This is deliberate to cover
both success and error paths with minimal noise.

Fixes: b7b17c9b67e5 ("md: remove mddev_lock() from md_attr_show()")
Cc: stable@vger.kernel.org (v4.12++)
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agomd: only call set_in_sync() when it is expected to succeed.
NeilBrown [Tue, 20 Aug 2019 00:21:09 +0000 (10:21 +1000)]
md: only call set_in_sync() when it is expected to succeed.

Since commit 4ad23a976413 ("MD: use per-cpu counter for
writes_pending"), set_in_sync() is substantially more expensive: it
can wait for a full RCU grace period which can be 10s of milliseconds.

So we should only call it when the cost is justified.

md_check_recovery() currently calls set_in_sync() every time it finds
anything to do (on non-external active arrays).  For an array
performing resync or recovery, this will be quite often.
Each call will introduce a delay to the md thread, which can noticeable
affect IO submission latency.

In md_check_recovery() we only need to call set_in_sync() if
'safemode' was non-zero at entry, meaning that there has been not
recent IO.  So we save this "safemode was nonzero" state, and only
call set_in_sync() if it was non-zero.

This measurably reduces mean and maximum IO submission latency during
resync/recovery.

Reported-and-tested-by: Jack Wang <jinpu.wang@cloud.ionos.com>
Fixes: 4ad23a976413 ("MD: use per-cpu counter for writes_pending")
Cc: stable@vger.kernel.org (v4.12+)
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agoblock: split .sysfs_lock into two locks
Ming Lei [Tue, 27 Aug 2019 11:01:48 +0000 (19:01 +0800)]
block: split .sysfs_lock into two locks

The kernfs built-in lock of 'kn->count' is held in sysfs .show/.store
path. Meantime, inside block's .show/.store callback, q->sysfs_lock is
required.

However, when mq & iosched kobjects are removed via
blk_mq_unregister_dev() & elv_unregister_queue(), q->sysfs_lock is held
too. This way causes AB-BA lock because the kernfs built-in lock of
'kn-count' is required inside kobject_del() too, see the lockdep warning[1].

On the other hand, it isn't necessary to acquire q->sysfs_lock for
both blk_mq_unregister_dev() & elv_unregister_queue() because
clearing REGISTERED flag prevents storing to 'queue/scheduler'
from being happened. Also sysfs write(store) is exclusive, so no
necessary to hold the lock for elv_unregister_queue() when it is
called in switching elevator path.

So split .sysfs_lock into two: one is still named as .sysfs_lock for
covering sync .store, the other one is named as .sysfs_dir_lock
for covering kobjects and related status change.

sysfs itself can handle the race between add/remove kobjects and
showing/storing attributes under kobjects. For switching scheduler
via storing to 'queue/scheduler', we use the queue flag of
QUEUE_FLAG_REGISTERED with .sysfs_lock for avoiding the race, then
we can avoid to hold .sysfs_lock during removing/adding kobjects.

[1]  lockdep warning
    ======================================================
    WARNING: possible circular locking dependency detected
    5.3.0-rc3-00044-g73277fc75ea0 #1380 Not tainted
    ------------------------------------------------------
    rmmod/777 is trying to acquire lock:
    00000000ac50e981 (kn->count#202){++++}, at: kernfs_remove_by_name_ns+0x59/0x72

    but task is already holding lock:
    00000000fb16ae21 (&q->sysfs_lock){+.+.}, at: blk_unregister_queue+0x78/0x10b

    which lock already depends on the new lock.

    the existing dependency chain (in reverse order) is:

    -> #1 (&q->sysfs_lock){+.+.}:
           __lock_acquire+0x95f/0xa2f
           lock_acquire+0x1b4/0x1e8
           __mutex_lock+0x14a/0xa9b
           blk_mq_hw_sysfs_show+0x63/0xb6
           sysfs_kf_seq_show+0x11f/0x196
           seq_read+0x2cd/0x5f2
           vfs_read+0xc7/0x18c
           ksys_read+0xc4/0x13e
           do_syscall_64+0xa7/0x295
           entry_SYSCALL_64_after_hwframe+0x49/0xbe

    -> #0 (kn->count#202){++++}:
           check_prev_add+0x5d2/0xc45
           validate_chain+0xed3/0xf94
           __lock_acquire+0x95f/0xa2f
           lock_acquire+0x1b4/0x1e8
           __kernfs_remove+0x237/0x40b
           kernfs_remove_by_name_ns+0x59/0x72
           remove_files+0x61/0x96
           sysfs_remove_group+0x81/0xa4
           sysfs_remove_groups+0x3b/0x44
           kobject_del+0x44/0x94
           blk_mq_unregister_dev+0x83/0xdd
           blk_unregister_queue+0xa0/0x10b
           del_gendisk+0x259/0x3fa
           null_del_dev+0x8b/0x1c3 [null_blk]
           null_exit+0x5c/0x95 [null_blk]
           __se_sys_delete_module+0x204/0x337
           do_syscall_64+0xa7/0x295
           entry_SYSCALL_64_after_hwframe+0x49/0xbe

    other info that might help us debug this:

     Possible unsafe locking scenario:

           CPU0                    CPU1
           ----                    ----
      lock(&q->sysfs_lock);
                                   lock(kn->count#202);
                                   lock(&q->sysfs_lock);
      lock(kn->count#202);

     *** DEADLOCK ***

    2 locks held by rmmod/777:
     #0: 00000000e69bd9de (&lock){+.+.}, at: null_exit+0x2e/0x95 [null_blk]
     #1: 00000000fb16ae21 (&q->sysfs_lock){+.+.}, at: blk_unregister_queue+0x78/0x10b

    stack backtrace:
    CPU: 0 PID: 777 Comm: rmmod Not tainted 5.3.0-rc3-00044-g73277fc75ea0 #1380
    Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS ?-20180724_192412-buildhw-07.phx4
    Call Trace:
     dump_stack+0x9a/0xe6
     check_noncircular+0x207/0x251
     ? print_circular_bug+0x32a/0x32a
     ? find_usage_backwards+0x84/0xb0
     check_prev_add+0x5d2/0xc45
     validate_chain+0xed3/0xf94
     ? check_prev_add+0xc45/0xc45
     ? mark_lock+0x11b/0x804
     ? check_usage_forwards+0x1ca/0x1ca
     __lock_acquire+0x95f/0xa2f
     lock_acquire+0x1b4/0x1e8
     ? kernfs_remove_by_name_ns+0x59/0x72
     __kernfs_remove+0x237/0x40b
     ? kernfs_remove_by_name_ns+0x59/0x72
     ? kernfs_next_descendant_post+0x7d/0x7d
     ? strlen+0x10/0x23
     ? strcmp+0x22/0x44
     kernfs_remove_by_name_ns+0x59/0x72
     remove_files+0x61/0x96
     sysfs_remove_group+0x81/0xa4
     sysfs_remove_groups+0x3b/0x44
     kobject_del+0x44/0x94
     blk_mq_unregister_dev+0x83/0xdd
     blk_unregister_queue+0xa0/0x10b
     del_gendisk+0x259/0x3fa
     ? disk_events_poll_msecs_store+0x12b/0x12b
     ? check_flags+0x1ea/0x204
     ? mark_held_locks+0x1f/0x7a
     null_del_dev+0x8b/0x1c3 [null_blk]
     null_exit+0x5c/0x95 [null_blk]
     __se_sys_delete_module+0x204/0x337
     ? free_module+0x39f/0x39f
     ? blkcg_maybe_throttle_current+0x8a/0x718
     ? rwlock_bug+0x62/0x62
     ? __blkcg_punt_bio_submit+0xd0/0xd0
     ? trace_hardirqs_on_thunk+0x1a/0x20
     ? mark_held_locks+0x1f/0x7a
     ? do_syscall_64+0x4c/0x295
     do_syscall_64+0xa7/0x295
     entry_SYSCALL_64_after_hwframe+0x49/0xbe
    RIP: 0033:0x7fb696cdbe6b
    Code: 73 01 c3 48 8b 0d 1d 20 0c 00 f7 d8 64 89 01 48 83 c8 ff c3 66 2e 0f 1f 84 00 00 008
    RSP: 002b:00007ffec9588788 EFLAGS: 00000206 ORIG_RAX: 00000000000000b0
    RAX: ffffffffffffffda RBX: 0000559e589137c0 RCX: 00007fb696cdbe6b
    RDX: 000000000000000a RSI: 0000000000000800 RDI: 0000559e58913828
    RBP: 0000000000000000 R08: 00007ffec9587701 R09: 0000000000000000
    R10: 00007fb696d4eae0 R11: 0000000000000206 R12: 00007ffec95889b0
    R13: 00007ffec95896b3 R14: 0000559e58913260 R15: 0000559e589137c0

Cc: Christoph Hellwig <hch@infradead.org>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Greg KH <gregkh@linuxfoundation.org>
Cc: Mike Snitzer <snitzer@redhat.com>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: add helper for checking if queue is registered
Ming Lei [Tue, 27 Aug 2019 11:01:47 +0000 (19:01 +0800)]
block: add helper for checking if queue is registered

There are 4 users which check if queue is registered, so add one helper
to check it.

Cc: Christoph Hellwig <hch@infradead.org>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Greg KH <gregkh@linuxfoundation.org>
Cc: Mike Snitzer <snitzer@redhat.com>
Cc: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblk-mq: don't hold q->sysfs_lock in blk_mq_map_swqueue
Ming Lei [Tue, 27 Aug 2019 11:01:46 +0000 (19:01 +0800)]
blk-mq: don't hold q->sysfs_lock in blk_mq_map_swqueue

blk_mq_map_swqueue() is called from blk_mq_init_allocated_queue()
and blk_mq_update_nr_hw_queues(). For the former caller, the kobject
isn't exposed to userspace yet. For the latter caller, hctx sysfs entries
and debugfs are un-registered before updating nr_hw_queues.

On the other hand, commit 2f8f1336a48b ("blk-mq: always free hctx after
request queue is freed") moves freeing hctx into queue's release
handler, so there won't be race with queue release path too.

So don't hold q->sysfs_lock in blk_mq_map_swqueue().

Cc: Christoph Hellwig <hch@infradead.org>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Greg KH <gregkh@linuxfoundation.org>
Cc: Mike Snitzer <snitzer@redhat.com>
Cc: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: don't hold q->sysfs_lock in elevator_init_mq
Ming Lei [Tue, 27 Aug 2019 11:01:45 +0000 (19:01 +0800)]
block: don't hold q->sysfs_lock in elevator_init_mq

The original comment says:

q->sysfs_lock must be held to provide mutual exclusion between
elevator_switch() and here.

Which is simply wrong. elevator_init_mq() is only called from
blk_mq_init_allocated_queue, which is always called before the request
queue is registered via blk_register_queue(), for dm-rq or normal rq
based driver. However, queue's kobject is only exposed and added to sysfs
in blk_register_queue(). So there isn't such race between elevator_switch()
and elevator_init_mq().

So avoid to hold q->sysfs_lock in elevator_init_mq().

Cc: Christoph Hellwig <hch@infradead.org>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Greg KH <gregkh@linuxfoundation.org>
Cc: Mike Snitzer <snitzer@redhat.com>
Cc: Bart Van Assche <bvanassche@acm.org>
Cc: Damien Le Moal <damien.lemoal@wdc.com>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: Remove blk_mq_register_dev()
Bart Van Assche [Tue, 27 Aug 2019 11:01:44 +0000 (19:01 +0800)]
block: Remove blk_mq_register_dev()

This function has no callers. Hence remove it.

Cc: Christoph Hellwig <hch@infradead.org>
Cc: Ming Lei <ming.lei@redhat.com>
Cc: Hannes Reinecke <hare@suse.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agowriteback, memcg: Implement foreign dirty flushing
Tejun Heo [Mon, 26 Aug 2019 16:06:56 +0000 (09:06 -0700)]
writeback, memcg: Implement foreign dirty flushing

There's an inherent mismatch between memcg and writeback.  The former
trackes ownership per-page while the latter per-inode.  This was a
deliberate design decision because honoring per-page ownership in the
writeback path is complicated, may lead to higher CPU and IO overheads
and deemed unnecessary given that write-sharing an inode across
different cgroups isn't a common use-case.

Combined with inode majority-writer ownership switching, this works
well enough in most cases but there are some pathological cases.  For
example, let's say there are two cgroups A and B which keep writing to
different but confined parts of the same inode.  B owns the inode and
A's memory is limited far below B's.  A's dirty ratio can rise enough
to trigger balance_dirty_pages() sleeps but B's can be low enough to
avoid triggering background writeback.  A will be slowed down without
a way to make writeback of the dirty pages happen.

This patch implements foreign dirty recording and foreign mechanism so
that when a memcg encounters a condition as above it can trigger
flushes on bdi_writebacks which can clean its pages.  Please see the
comment on top of mem_cgroup_track_foreign_dirty_slowpath() for
details.

A reproducer follows.

write-range.c::

  #include <stdio.h>
  #include <stdlib.h>
  #include <unistd.h>
  #include <fcntl.h>
  #include <sys/types.h>

  static const char *usage = "write-range FILE START SIZE\n";

  int main(int argc, char **argv)
  {
  int fd;
  unsigned long start, size, end, pos;
  char *endp;
  char buf[4096];

  if (argc < 4) {
  fprintf(stderr, usage);
  return 1;
  }

  fd = open(argv[1], O_WRONLY);
  if (fd < 0) {
  perror("open");
  return 1;
  }

  start = strtoul(argv[2], &endp, 0);
  if (*endp != '\0') {
  fprintf(stderr, usage);
  return 1;
  }

  size = strtoul(argv[3], &endp, 0);
  if (*endp != '\0') {
  fprintf(stderr, usage);
  return 1;
  }

  end = start + size;

  while (1) {
  for (pos = start; pos < end; ) {
  long bread, bwritten = 0;

  if (lseek(fd, pos, SEEK_SET) < 0) {
  perror("lseek");
  return 1;
  }

  bread = read(0, buf, sizeof(buf) < end - pos ?
       sizeof(buf) : end - pos);
  if (bread < 0) {
  perror("read");
  return 1;
  }
  if (bread == 0)
  return 0;

  while (bwritten < bread) {
  long this;

  this = write(fd, buf + bwritten,
       bread - bwritten);
  if (this < 0) {
  perror("write");
  return 1;
  }

  bwritten += this;
  pos += bwritten;
  }
  }
  }
  }

repro.sh::

  #!/bin/bash

  set -e
  set -x

  sysctl -w vm.dirty_expire_centisecs=300000
  sysctl -w vm.dirty_writeback_centisecs=300000
  sysctl -w vm.dirtytime_expire_seconds=300000
  echo 3 > /proc/sys/vm/drop_caches

  TEST=/sys/fs/cgroup/test
  A=$TEST/A
  B=$TEST/B

  mkdir -p $A $B
  echo "+memory +io" > $TEST/cgroup.subtree_control
  echo $((1<<30)) > $A/memory.high
  echo $((32<<30)) > $B/memory.high

  rm -f testfile
  touch testfile
  fallocate -l 4G testfile

  echo "Starting B"

  (echo $BASHPID > $B/cgroup.procs
   pv -q --rate-limit 70M < /dev/urandom | ./write-range testfile $((2<<30)) $((2<<30))) &

  echo "Waiting 10s to ensure B claims the testfile inode"
  sleep 5
  sync
  sleep 5
  sync
  echo "Starting A"

  (echo $BASHPID > $A/cgroup.procs
   pv < /dev/urandom | ./write-range testfile 0 $((2<<30)))

v2: Added comments explaining why the specific intervals are being used.

v3: Use 0 @nr when calling cgroup_writeback_by_id() to use best-effort
    flushing while avoding possible livelocks.

v4: Use get_jiffies_64() and time_before/after64() instead of raw
    jiffies_64 and arthimetic comparisons as suggested by Jan.

Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agowriteback, memcg: Implement cgroup_writeback_by_id()
Tejun Heo [Mon, 26 Aug 2019 16:06:55 +0000 (09:06 -0700)]
writeback, memcg: Implement cgroup_writeback_by_id()

Implement cgroup_writeback_by_id() which initiates cgroup writeback
from bdi and memcg IDs.  This will be used by memcg foreign inode
flushing.

v2: Use wb_get_lookup() instead of wb_get_create() to avoid creating
    spurious wbs.

v3: Interpret 0 @nr as 1.25 * nr_dirty to implement best-effort
    flushing while avoding possible livelocks.

Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agowriteback: Separate out wb_get_lookup() from wb_get_create()
Tejun Heo [Mon, 26 Aug 2019 16:06:54 +0000 (09:06 -0700)]
writeback: Separate out wb_get_lookup() from wb_get_create()

Separate out wb_get_lookup() which doesn't try to create one if there
isn't already one from wb_get_create().  This will be used by later
patches.

Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobdi: Add bdi->id
Tejun Heo [Mon, 26 Aug 2019 16:06:53 +0000 (09:06 -0700)]
bdi: Add bdi->id

There currently is no way to universally identify and lookup a bdi
without holding a reference and pointer to it.  This patch adds an
non-recycling bdi->id and implements bdi_get_by_id() which looks up
bdis by their ids.  This will be used by memcg foreign inode flushing.

I left bdi_list alone for simplicity and because while rb_tree does
support rcu assignment it doesn't seem to guarantee lossless walk when
walk is racing aginst tree rebalance operations.

Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agowriteback: Generalize and expose wb_completion
Tejun Heo [Mon, 26 Aug 2019 16:06:52 +0000 (09:06 -0700)]
writeback: Generalize and expose wb_completion

wb_completion is used to track writeback completions.  We want to use
it from memcg side for foreign inode flushes.  This patch updates it
to remember the target waitq instead of assuming bdi->wb_waitq and
expose it outside of fs-writeback.c.

Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agonull_blk: fix inline misuse
Jens Axboe [Fri, 23 Aug 2019 18:49:00 +0000 (12:49 -0600)]
null_blk: fix inline misuse

You can't magically mark a function inline and expect that to work.

Fixes: fceb5d1b19cb ("null_blk: create a helper for zoned devices")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agonull_blk: create a helper for req completion
Chaitanya Kulkarni [Fri, 23 Aug 2019 04:45:19 +0000 (21:45 -0700)]
null_blk: create a helper for req completion

This patch creates a helper function for handling the request
completion in the null_handle_cmd().

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agonull_blk: create a helper for zoned devices
Chaitanya Kulkarni [Fri, 23 Aug 2019 04:45:18 +0000 (21:45 -0700)]
null_blk: create a helper for zoned devices

This patch creates a helper function for handling zoned block device
operations.

This patch also restructured the code for null_blk_zoned.c and uses the
pattern to return blk_status_t and catch the error in the function
null_handle_cmd() into cmd->error variable instead of setting it up in
the deeper layer just like the way it is done for flush, badblocks and
memory backed case in the null_handle_cmd(). We also move
null_handle_zoned() to the null_blk_zoned.c to keep the zoned code
separate.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>