Merge tag 'md-6.9-20240301' of https://git.kernel.org/pub/scm/linux/kernel/git/song/md into for-6.9/block

Pull MD updates from Song:

"The major changes are:

 1. Refactor raid1 read_balance, by Yu Kuai and Paul Luse.
 2. Clean up and fix for md_ioctl, by Li Nan.
 3. Other small fixes, by Gui-Dong Han and Heming Zhao."

* tag 'md-6.9-20240301' of https://git.kernel.org/pub/scm/linux/kernel/git/song/md: (22 commits)
  md/raid1: factor out helpers to choose the best rdev from read_balance()
  md/raid1: factor out the code to manage sequential IO
  md/raid1: factor out choose_bb_rdev() from read_balance()
  md/raid1: factor out choose_slow_rdev() from read_balance()
  md/raid1: factor out read_first_rdev() from read_balance()
  md/raid1-10: factor out a new helper raid1_should_read_first()
  md/raid1-10: add a helper raid1_check_read_range()
  md/raid1: fix choose next idle in read_balance()
  md/raid1: record nonrot rdevs while adding/removing rdevs to conf
  md/raid1: factor out helpers to add rdev to conf
  md: add a new helper rdev_has_badblock()
  md/raid5: fix atomicity violation in raid5_cache_count
  md/md-bitmap: fix incorrect usage for sb_index
  md: check mddev->pers before calling md_set_readonly()
  md: clean up openers check in do_md_stop() and md_set_readonly()
  md: sync blockdev before stopping raid or setting readonly
  md: factor out a helper to sync mddev
  md: Don't clear MD_CLOSING when the raid is about to stop
  md: return directly before setting did_set_md_closing
  md: clean up invalid BUG_ON in md_ioctl
  ...

From: Yu Kuai <yukuai3@huawei.com>
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>

The original idea is that Paul want to optimize raid1 read
performance([1]), however, we think that the original code for
read_balance() is quite complex, and we don't want to add more
complexity. Hence we decide to refactor read_balance() first, to make
code cleaner and easier for follow up.

Before this patchset, read_balance() has many local variables and many
branches, it want to consider all the scenarios in one iteration. The
idea of this patch is to divide them into 4 different steps:

1) If resync is in progress, find the first usable disk, patch 5;
Otherwise:
2) Loop through all disks and skipping slow disks and disks with bad
blocks, choose the best disk, patch 10. If no disk is found:
3) Look for disks with bad blocks and choose the one with most number of
sectors, patch 8. If no disk is found:
4) Choose first found slow disk with no bad blocks, or slow disk with
most number of sectors, patch 7.

Note that step 3) and step 4) are super code path, and performance
should not be considered.

And after this patchset, we'll continue to optimize read_balance for
step 2), specifically how to choose the best rdev to read.

[1] https://lore.kernel.org/all/20240102125115.129261-1-paul.e.luse@linux.intel.com/

Yu Kuai (11):
  md: add a new helper rdev_has_badblock()
  md/raid1: factor out helpers to add rdev to conf
  md/raid1: record nonrot rdevs while adding/removing rdevs to conf
  md/raid1: fix choose next idle in read_balance()
  md/raid1-10: add a helper raid1_check_read_range()
  md/raid1-10: factor out a new helper raid1_should_read_first()
  md/raid1: factor out read_first_rdev() from read_balance()
  md/raid1: factor out choose_slow_rdev() from read_balance()
  md/raid1: factor out choose_bb_rdev() from read_balance()
  md/raid1: factor out the code to manage sequential IO
  md/raid1: factor out helpers to choose the best rdev from
    read_balance()

The way that best rdev is chosen:

1) If the read is sequential from one rdev:
 - if rdev is rotational, use this rdev;
 - if rdev is non-rotational, use this rdev until total read length
   exceed disk opt io size;

2) If the read is not sequential:
 - if there is idle disk, use it, otherwise:
 - if the array has non-rotational disk, choose the rdev with minimal
   inflight IO;
 - if all the underlaying disks are rotational disk, choose the rdev
   with closest IO;

There are no functional changes, just to make code cleaner and prepare
for following refactor.
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-12-yukuai1@huaweicloud.com

There is no functional change for now, make read_balance() cleaner and
prepare to fix problems and refactor the handler of sequential IO.
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-11-yukuai1@huaweicloud.com

read_balance() is hard to understand because there are too many status
and branches, and it's overlong.

This patch factor out the case to read the rdev with bad blocks from
read_balance(), there are no functional changes.
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-10-yukuai1@huaweicloud.com

read_balance() is hard to understand because there are too many status
and branches, and it's overlong.

This patch factor out the case to read the slow rdev from
read_balance(), there are no functional changes.
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-9-yukuai1@huaweicloud.com

read_balance() is hard to understand because there are too many status
and branches, and it's overlong.

This patch factor out the case to read the first rdev from
read_balance(), there are no functional changes.
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-8-yukuai1@huaweicloud.com

If resync is in progress, read_balance() should find the first usable
disk, otherwise, data could be inconsistent after resync is done. raid1
and raid10 implement the same checking, hence factor out the checking
to make code cleaner.

Noted that raid1 is using 'mddev->recovery_cp', which is updated after
all resync IO is done, while raid10 is using 'conf->next_resync', which
is inaccurate because raid10 update it before submitting resync IO.
Fortunately, raid10 read IO can't concurrent with resync IO, hence there
is no problem. And this patch also switch raid10 to use
'mddev->recovery_cp'.
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-7-yukuai1@huaweicloud.com

The checking and handler of bad blocks appear many timers during
read_balance() in raid1 and raid10. This helper will be used in later
patches to simplify read_balance() a lot.
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-6-yukuai1@huaweicloud.com

Commit 12cee5a8 ("md/raid1: prevent merging too large request") add
the case choose next idle in read_balance():

read_balance:
 for_each_rdev
  if(next_seq_sect == this_sector || dist == 0)
  -> sequential reads
   best_disk = disk;
   if (...)
    choose_next_idle = 1
    continue;

 for_each_rdev
 -> iterate next rdev
  if (pending == 0)
   best_disk = disk;
   -> choose the next idle disk
   break;

  if (choose_next_idle)
   -> keep using this rdev if there are no other idle disk
   contine

However, commit 2e52d449 ("md/raid1: add failfast handling for reads.")
remove the code:

-               /* If device is idle, use it */
-               if (pending == 0) {
-                       best_disk = disk;
-                       break;
-               }

Hence choose next idle will never work now, fix this problem by
following:

1) don't set best_disk in this case, read_balance() will choose the best
   disk after iterating all the disks;
2) add 'pending' so that other idle disk will be chosen;
3) add a new local variable 'sequential_disk' to record the disk, and if
   there is no other idle disk, 'sequential_disk' will be chosen;

Fixes: 2e52d449 ("md/raid1: add failfast handling for reads.")
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-5-yukuai1@huaweicloud.com

For raid1, each read will iterate all the rdevs from conf and check if
any rdev is non-rotational, then choose rdev with minimal IO inflight
if so, or rdev with closest distance otherwise.

Disk nonrot info can be changed through sysfs entry:

/sys/block/[disk_name]/queue/rotational

However, consider that this should only be used for testing, and user
really shouldn't do this in real life. Record the number of non-rotational
disks in conf, to avoid checking each rdev in IO fast path and simplify
read_balance() a little bit.
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-4-yukuai1@huaweicloud.com

There are no functional changes, just make code cleaner and prepare to
record disk non-rotational information while adding and removing rdev to
conf
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-3-yukuai1@huaweicloud.com

The current api is_badblock() must pass in 'first_bad' and
'bad_sectors', however, many caller just want to know if there are
badblocks or not, and these caller must define two local variable that
will never be used.

Add a new helper rdev_has_badblock() that will only return if there are
badblocks or not, remove unnecessary local variables and replace
is_badblock() with the new helper in many places.

There are no functional changes, and the new helper will also be used
later to refactor read_balance().
Co-developed-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Paul Luse <paul.e.luse@linux.intel.com>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Xiao Ni <xni@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240229095714.926789-2-yukuai1@huaweicloud.com

The current command UBLK_CMD_DEL_DEV won't return until the device is
released, this way looks more reliable, but makes userspace more
difficult to implement, especially about orders: unmap command
buffer(which holds one ublkc reference), ublkc close,
io_uring_file_unregister, ublkb close.

Add UBLK_CMD_DEL_DEV_ASYNC so that device deletion won't wait release,
then userspace needn't worry about the above order. Actually both loop
and nbd is deleted in this async way.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/20240223075539.89945-3-ming.lei@redhat.comSigned-off-by: Jens Axboe <axboe@kernel.dk>

Firstly convert get_device() and put_device() into ublk_get_device()
and ublk_put_device().

Secondly annotate ublk_get_device() & ublk_put_device() as noinline
for trace, especially it is often to trigger device deletion hang
when incorrect order is used on ublkc mmap, ublkc close,
io_uring_sqe_unregister_file, ublkb close.
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Link: https://lore.kernel.org/r/20240223075539.89945-2-ming.lei@redhat.comSigned-off-by: Jens Axboe <axboe@kernel.dk>

For most of ARCHs, 'nr_cpus=1' is passed for kdump kernel, so
nr_hw_queues for each mapping is supposed to be 1 already.

More importantly, this way may cause trouble for driver, because blk-mq and
driver see different queue mapping since driver should setup hardware
queue setting before calling into allocating blk-mq tagset.

So not overriding nr_hw_queues and nr_maps for kdump kernel.

Cc: Wen Xiong <wenxiong@us.ibm.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20240228040857.306483-1-ming.lei@redhat.comSigned-off-by: Jens Axboe <axboe@kernel.dk>

In raid5_cache_count():
    if (conf->max_nr_stripes < conf->min_nr_stripes)
        return 0;
    return conf->max_nr_stripes - conf->min_nr_stripes;
The current check is ineffective, as the values could change immediately
after being checked.

In raid5_set_cache_size():
    ...
    conf->min_nr_stripes = size;
    ...
    while (size > conf->max_nr_stripes)
        conf->min_nr_stripes = conf->max_nr_stripes;
    ...

Due to intermediate value updates in raid5_set_cache_size(), concurrent
execution of raid5_cache_count() and raid5_set_cache_size() may lead to
inconsistent reads of conf->max_nr_stripes and conf->min_nr_stripes.
The current checks are ineffective as values could change immediately
after being checked, raising the risk of conf->min_nr_stripes exceeding
conf->max_nr_stripes and potentially causing an integer overflow.

This possible bug is found by an experimental static analysis tool
developed by our team. This tool analyzes the locking APIs to extract
function pairs that can be concurrently executed, and then analyzes the
instructions in the paired functions to identify possible concurrency bugs
including data races and atomicity violations. The above possible bug is
reported when our tool analyzes the source code of Linux 6.2.

To resolve this issue, it is suggested to introduce local variables
'min_stripes' and 'max_stripes' in raid5_cache_count() to ensure the
values remain stable throughout the check. Adding locks in
raid5_cache_count() fails to resolve atomicity violations, as
raid5_set_cache_size() may hold intermediate values of
conf->min_nr_stripes while unlocked. With this patch applied, our tool no
longer reports the bug, with the kernel configuration allyesconfig for
x86_64. Due to the lack of associated hardware, we cannot test the patch
in runtime testing, and just verify it according to the code logic.

Fixes: edbe83ab ("md/raid5: allow the stripe_cache to grow and shrink.")
Cc: stable@vger.kernel.org
Signed-off-by: Gui-Dong Han <2045gemini@gmail.com>
Reviewed-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240112071017.16313-1-2045gemini@gmail.comSigned-off-by: Song Liu <song@kernel.org>

Opening the backing device only when the block device is opened is
a bit weird as no one configures block devices to not use them.
Opend them at add time, close them at remove time and remove the
now superflous opened counter as remove can simply check for
disk_openers.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Richard Weinberger <richard@nod.at>
Link: https://lore.kernel.org/r/20240222072417.3773131-8-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

No need for it now, everything goes through the gendisk.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Richard Weinberger <richard@nod.at>
Link: https://lore.kernel.org/r/20240222072417.3773131-7-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

No need to delay this until open time.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Richard Weinberger <richard@nod.at>
Link: https://lore.kernel.org/r/20240222072417.3773131-6-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

No reason to delay this until open time.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Richard Weinberger <richard@nod.at>
Link: https://lore.kernel.org/r/20240222072417.3773131-5-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

No reason to delay this until open time.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Richard Weinberger <richard@nod.at>
Link: https://lore.kernel.org/r/20240222072417.3773131-4-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

Fold it into the only caller to remove lots of references to the
global ubd_devs array.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Richard Weinberger <richard@nod.at>
Link: https://lore.kernel.org/r/20240222072417.3773131-3-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

And add a disk pointer to the ubd structure instead to keep all
the per-device information together.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Richard Weinberger <richard@nod.at>
Link: https://lore.kernel.org/r/20240222072417.3773131-2-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

Pass the initial queue limits to blk_mq_alloc_disk and use the
blkif_set_queue_limits API to update the limits on reconnect.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Link: https://lore.kernel.org/r/20240221125845.3610668-5-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

blkif_set_queue_limits already sets the max_sements limits, so don't do
it a second time.  Also remove a comment about a long fixe bug in
blk_mq_update_nr_hw_queues.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Link: https://lore.kernel.org/r/20240221125845.3610668-4-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

The block layer now sets the discard granularity to the physical
block size default.  Take advantage of that in xen-blkfront and only
set the discard granularity if explicitly specified.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Link: https://lore.kernel.org/r/20240221125845.3610668-3-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

Currently xen-blkfront set the max discard limit to the capacity of
the device, which is suboptimal when the capacity changes.  Just set
it to UINT_MAX, which has the same effect and is simpler.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Link: https://lore.kernel.org/r/20240221125845.3610668-2-hch@lst.deSigned-off-by: Jens Axboe <axboe@kernel.dk>

Commit d7038f95 ("md-bitmap: don't use ->index for pages backing the
bitmap file") removed page->index from bitmap code, but left wrong code
logic for clustered-md. current code never set slot offset for cluster
nodes, will sometimes cause crash in clustered env.

Call trace (partly):
 md_bitmap_file_set_bit+0x110/0x1d8 [md_mod]
 md_bitmap_startwrite+0x13c/0x240 [md_mod]
 raid1_make_request+0x6b0/0x1c08 [raid1]
 md_handle_request+0x1dc/0x368 [md_mod]
 md_submit_bio+0x80/0xf8 [md_mod]
 __submit_bio+0x178/0x300
 submit_bio_noacct_nocheck+0x11c/0x338
 submit_bio_noacct+0x134/0x614
 submit_bio+0x28/0xdc
 submit_bh_wbc+0x130/0x1cc
 submit_bh+0x1c/0x28

Fixes: d7038f95 ("md-bitmap: don't use ->index for pages backing the bitmap file")
Cc: stable@vger.kernel.org # v6.6+
Signed-off-by: Heming Zhao <heming.zhao@suse.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240223121128.28985-1-heming.zhao@suse.com

If 'mddev->pers' is NULL, there is nothing to do in md_set_readonly().
Except for md_ioctl(), the other two callers of md_set_readonly() have
already checked 'mddev->pers'. To simplify the code, move the check of
'mddev->pers' to the caller.
Signed-off-by: Li Nan <linan122@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240226031444.3606764-10-linan666@huaweicloud.com

Before stopping or setting readonly, mddev_set_closing_and_sync_blockdev()
is always called to check the openers. So no longer need to check it again
in do_md_stop() and md_set_readonly(). Clean it up.
Signed-off-by: Li Nan <linan122@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240226031444.3606764-9-linan666@huaweicloud.com

Commit a05b7ea0 ("md: avoid crash when stopping md array races
with closing other open fds.") added sync_block before stopping raid and
setting readonly. Later in commit 260fa034 ("md: avoid deadlock when
dirty buffers during md_stop.") it is moved to ioctl. array_state_store()
was ignored. Add sync blockdev to array_state_store() now.
Signed-off-by: Li Nan <linan122@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240226031444.3606764-8-linan666@huaweicloud.com

There are no functional changes, prepare to sync mddev in
array_state_store().
Signed-off-by: Li Nan <linan122@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240226031444.3606764-7-linan666@huaweicloud.com

The raid should not be opened anymore when it is about to be stopped.
However, other processes can open it again if the flag MD_CLOSING is
cleared before exiting. From now on, this flag will not be cleared when
the raid will be stopped.

Fixes: 065e519e ("md: MD_CLOSING needs to be cleared after called md_set_readonly or do_md_stop")
Signed-off-by: Li Nan <linan122@huawei.com>
Reviewed-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240226031444.3606764-6-linan666@huaweicloud.com

There is nothing to do at 'out' before setting 'did_set_md_closing'
in md_ioctl(). Return directly, and it will help us to remove
'did_set_md_closing' later.
Signed-off-by: Li Nan <linan122@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240226031444.3606764-5-linan666@huaweicloud.com

'disk->private_data' is set to mddev in md_alloc() and never set to NULL,
and users need to open mddev before submitting ioctl. So mddev must not
have been freed during ioctl, and there is no need to check mddev here.
Clean up it.
Signed-off-by: Li Nan <linan122@huawei.com>
Reviewed-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240226031444.3606764-4-linan666@huaweicloud.com

There is only one case of this 'switch'. Change it to 'if'.
Signed-off-by: Li Nan <linan122@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240226031444.3606764-3-linan666@huaweicloud.com

There is no functional change. Just to make code cleaner.
Signed-off-by: Li Nan <linan122@huawei.com>
Reviewed-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>
Link: https://lore.kernel.org/r/20240226031444.3606764-2-linan666@huaweicloud.com

The SLAB_MEM_SPREAD flag is already a no-op as of 6.8-rc1, remove
its usage so we can delete it from slab. No functional change.
Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
Link: https://lore.kernel.org/r/20240224134646.829105-1-chengming.zhou@linux.devSigned-off-by: Jens Axboe <axboe@kernel.dk>

The logic in blk_mq_complete_need_ipi() assumes SMP systems where all
CPUs have equal compute capacities and only LLC cache can make
a different on perceived performance. But this assumption falls apart on
HMP systems where LLC is shared, but the CPUs have different capacities.
Staying local then can have a big performance impact if the IO request
was done from a CPU with higher capacity but the interrupt is serviced
on a lower capacity CPU.

Use the new cpus_equal_capacity() function to check if we need to send
an IPI.

Without the patch I see the BLOCK softirq always running on little cores
(where the hardirq is serviced). With it I can see it running on all
cores.

This was noticed after the topology change [1] where now on a big.LITTLE
we truly get that the LLC is shared between all cores where as in the
past it was being misrepresented for historical reasons. The logic
exposed a missing dependency on capacities for such systems where there
can be a big performance difference between the CPUs.

This of course introduced a noticeable change in behavior depending on
how the topology is presented. Leading to regressions in some workloads
as the performance of the BLOCK softirq on littles can be noticeably
worse on some platforms.

Worth noting that we could have checked for capacities being greater
than or equal instead for equality. This will lead to favouring higher
performance always. But opted for equality instead to match the
performance of the requester without making an assumption that can lead
to power trade-offs which these systems tend to be sensitive about. If
the requester would like to run faster, it's better to rely on the
scheduler to give the IO requester via some facility to run on a faster
core; and then if the interrupt triggered on a CPU with different
capacity we'll make sure to match the performance the requester is
supposed to run at.

[1] https://lpc.events/event/16/contributions/1342/attachments/962/1883/LPC-2022-Android-MC-Phantom-Domains.pdfSigned-off-by: Qais Yousef <qyousef@layalina.io>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Link: https://lore.kernel.org/r/20240223155749.2958009-3-qyousef@layalina.ioSigned-off-by: Jens Axboe <axboe@kernel.dk>