nvme-pci: factor the iod mempool creation into a helper

Add a helper to create the iod mempool.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Tested-by Gerd Bayer <gbayer@linxu.ibm.com>

nvme-pci: move more teardown work to nvme_remove

nvme_dbbuf_dma_free frees dma coherent memory, so it must not be called
after ->remove has returned.  Fortunately there is no way to use it
after shutdown as no more I/O is possible so it can be moved.  Similarly
the iod_mempool can't be used for a device kept alive after shutdown, so
move it next to freeing the PRP pools.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Tested-by Gerd Bayer <gbayer@linxu.ibm.com>

nvme-pci: put the admin queue in nvme_dev_remove_admin

Once the controller is shutdown no one can access the admin queue.  Tear
it down in nvme_dev_remove_admin, which matches the flow in the other
drivers.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Tested-by Gerd Bayer <gbayer@linxu.ibm.com>

nvme: simplify transport specific device attribute handling

Allow the transport driver to override the attribute groups for the
control device, so that the PCIe driver doesn't manually have to add a
group after device creation and keep track of it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Tested-by Gerd Bayer <gbayer@linxu.ibm.com>

nvme: move OPAL setup from PCIe to core

Nothing about the TCG Opal support is PCIe transport specific, so move it
to the core code.  For this nvme_init_ctrl_finish grows a new
was_suspended argument that allows the transport driver to tell the OPAL
code if the controller came out of a suspend cycle.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: James Smart <jsmart2021@gmail.com>
Tested-by Gerd Bayer <gbayer@linxu.ibm.com>

nvme: don't call nvme_init_ctrl_finish from nvme_passthru_end

nvme_passthrough_end can race with a reset, which can lead to
racing stores to the cels xarray as well as further shengians
with upcoming more complicated initialization.

So drop the call and just log that the controller capabilities
might have changed and a reset could be required to use the new
controller capabilities.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Tested-by Gerd Bayer <gbayer@linxu.ibm.com>

nvme: implement the DEAC bit for the Write Zeroes command

While the specification allows devices to either deallocate data
or to actually write zeroes on any Write Zeroes command, many SSDs
only do the sensible thing and deallocate data when the DEAC bit
is specific.  Set it when it is supported and the caller doesn't
explicitly opt out of deallocation.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>

nvme: identify-namespace without CAP_SYS_ADMIN

Allow all identify-namespace variants (CNS 00h, 05h and 08h) without
requiring CAP_SYS_ADMIN. The information (retrieved using id-ns) is
needed to form IO commands for passthrough interface.

Signed-off-by: Kanchan Joshi <joshi.k@samsung.com>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme: fine-granular CAP_SYS_ADMIN for nvme io commands

Currently both io and admin commands are kept under a
coarse-granular CAP_SYS_ADMIN check, disregarding file mode completely.

$ ls -l /dev/ng*
crw-rw-rw- 1 root root 242, 0 Sep  9 19:20 /dev/ng0n1
crw------- 1 root root 242, 1 Sep  9 19:20 /dev/ng0n2

In the example above, ng0n1 appears as if it may allow unprivileged
read/write operation but it does not and behaves same as ng0n2.

This patch implements a shift from CAP_SYS_ADMIN to more fine-granular
control for io-commands.
If CAP_SYS_ADMIN is present, nothing else is checked as before.
Otherwise, following rules are in place
- any admin-cmd is not allowed
- vendor-specific and fabric commmand are not allowed
- io-commands that can write are allowed if matching FMODE_WRITE
permission is present
- io-commands that read are allowed

Add a helper nvme_cmd_allowed that implements above policy.
Change all the callers of CAP_SYS_ADMIN to go through nvme_cmd_allowed
for any decision making.
Since file open mode is counted for any approval/denial, change at
various places to keep file-mode information handy.

Signed-off-by: Kanchan Joshi <joshi.k@samsung.com>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvme-fc: improve memory usage in nvme_fc_rcv_ls_req()

sizeof( struct nvmefc_ls_rcv_op ) = 64
sizeof( union nvmefc_ls_requests ) = 1024
sizeof( union nvmefc_ls_responses ) = 128

So, in nvme_fc_rcv_ls_req(), 1216 bytes of memory are requested when
kzalloc() is called.

Because of the way memory allocations are performed, 2048 bytes are
allocated. So about 800 bytes are wasted for each request.

Switch to 3 distinct memory allocations, in order to:
   - save these 800 bytes
   - avoid zeroing this extra memory
   - make sure that memory is properly aligned in case of DMA access
    ("fc_dma_map_single(lsop->rspbuf)" just a few lines below)

Signed-off-by: Christophe JAILLET <christophe.jaillet@wanadoo.fr>
Reviewed-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvmet: only allocate a single slab for bvecs

There is no need to have a separate slab cache for each namespace,
and having separate ones creates duplicate debugs file names as well.

Fixes: d5eff33ee6f8 ("nvmet: add simple file backed ns support")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>

nvmet: force reconnect when number of queue changes

In order to test queue number changes we need to make sure that the
host reconnects. Because only when the host disconnects from the
target the number of queues are allowed to change according the spec.

The initial idea was to disable and re-enable the ports and have the
host wait until the KATO timer expires, triggering error
recovery. Though the host would see a DNR reply when trying to
reconnect. Because of the DNR bit the connection is dropped
completely. There is no point in trying to reconnect with the same
parameters according the spec.

We can force to reconnect the host is by deleting all controllers. The
host will observe any newly posted request to fail and thus starts the
error recovery but this time without the DNR bit set.

Signed-off-by: Daniel Wagner <dwagner@suse.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Acked-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

nvmet: use try_cmpxchg in nvmet_update_sq_head

Use try_cmpxchg instead of cmpxchg (*ptr, old, new) == old in
nvmet_update_sq_head.  x86 CMPXCHG instruction returns success in ZF flag, so
this change saves a compare after cmpxchg (and related move instruction in
front of cmpxchg).

Also, try_cmpxchg implicitly assigns old *ptr value to "old" when cmpxchg
fails. There is no need to re-read the value in the loop.

Note that the value from *ptr should be read using READ_ONCE to prevent
the compiler from merging, refetching or reordering the read.

No functional change intended.

Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>

Merge branch 'md-next' of https://git./linux/kernel/git/song/md into for-6.2/block

Pull MD fixes from Song.

* 'md-next' of https://git.kernel.org/pub/scm/linux/kernel/git/song/md:
  md/raid1: stop mdx_raid1 thread when raid1 array run failed
  md/raid5: use bdev_write_cache instead of open coding it
  md: fix a crash in mempool_free
  md/raid0, raid10: Don't set discard sectors for request queue
  md/bitmap: Fix bitmap chunk size overflow issues
  md: introduce md_ro_state
  md: factor out __md_set_array_info()
  lib/raid6: drop RAID6_USE_EMPTY_ZERO_PAGE
  raid5-cache: use try_cmpxchg in r5l_wake_reclaim
  drivers/md/md-bitmap: check the return value of md_bitmap_get_counter()

md/raid1: stop mdx_raid1 thread when raid1 array run failed

fail run raid1 array when we assemble array with the inactive disk only,
but the mdx_raid1 thread were not stop, Even if the associated resources
have been released. it will caused a NULL dereference when we do poweroff.

This causes the following Oops:
    [  287.587787] BUG: kernel NULL pointer dereference, address: 0000000000000070
    [  287.594762] #PF: supervisor read access in kernel mode
    [  287.599912] #PF: error_code(0x0000) - not-present page
    [  287.605061] PGD 0 P4D 0
    [  287.607612] Oops: 0000 [#1] SMP NOPTI
    [  287.611287] CPU: 3 PID: 5265 Comm: md0_raid1 Tainted: G     U            5.10.146 #0
    [  287.619029] Hardware name: xxxxxxx/To be filled by O.E.M, BIOS 5.19 06/16/2022
    [  287.626775] RIP: 0010:md_check_recovery+0x57/0x500 [md_mod]
    [  287.632357] Code: fe 01 00 00 48 83 bb 10 03 00 00 00 74 08 48 89 ......
    [  287.651118] RSP: 0018:ffffc90000433d78 EFLAGS: 00010202
    [  287.656347] RAX: 0000000000000000 RBX: ffff888105986800 RCX: 0000000000000000
    [  287.663491] RDX: ffffc90000433bb0 RSI: 00000000ffffefff RDI: ffff888105986800
    [  287.670634] RBP: ffffc90000433da0 R08: 0000000000000000 R09: c0000000ffffefff
    [  287.677771] R10: 0000000000000001 R11: ffffc90000433ba8 R12: ffff888105986800
    [  287.684907] R13: 0000000000000000 R14: fffffffffffffe00 R15: ffff888100b6b500
    [  287.692052] FS:  0000000000000000(0000) GS:ffff888277f80000(0000) knlGS:0000000000000000
    [  287.700149] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
    [  287.705897] CR2: 0000000000000070 CR3: 000000000320a000 CR4: 0000000000350ee0
    [  287.713033] Call Trace:
    [  287.715498]  raid1d+0x6c/0xbbb [raid1]
    [  287.719256]  ? __schedule+0x1ff/0x760
    [  287.722930]  ? schedule+0x3b/0xb0
    [  287.726260]  ? schedule_timeout+0x1ed/0x290
    [  287.730456]  ? __switch_to+0x11f/0x400
    [  287.734219]  md_thread+0xe9/0x140 [md_mod]
    [  287.738328]  ? md_thread+0xe9/0x140 [md_mod]
    [  287.742601]  ? wait_woken+0x80/0x80
    [  287.746097]  ? md_register_thread+0xe0/0xe0 [md_mod]
    [  287.751064]  kthread+0x11a/0x140
    [  287.754300]  ? kthread_park+0x90/0x90
    [  287.757974]  ret_from_fork+0x1f/0x30

In fact, when raid1 array run fail, we need to do
md_unregister_thread() before raid1_free().

Signed-off-by: Jiang Li <jiang.li@ugreen.com>
Signed-off-by: Song Liu <song@kernel.org>

md/raid5: use bdev_write_cache instead of open coding it

Use the bdev_write_cache instead of two equivalent open coded checks.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Song Liu <song@kernel.org>

md: fix a crash in mempool_free

There's a crash in mempool_free when running the lvm test
shell/lvchange-rebuild-raid.sh.

The reason for the crash is this:
* super_written calls atomic_dec_and_test(&mddev->pending_writes) and
  wake_up(&mddev->sb_wait). Then it calls rdev_dec_pending(rdev, mddev)
  and bio_put(bio).
* so, the process that waited on sb_wait and that is woken up is racing
  with bio_put(bio).
* if the process wins the race, it calls bioset_exit before bio_put(bio)
  is executed.
* bio_put(bio) attempts to free a bio into a destroyed bio set - causing
  a crash in mempool_free.

We fix this bug by moving bio_put before atomic_dec_and_test.

We also move rdev_dec_pending before atomic_dec_and_test as suggested by
Neil Brown.

The function md_end_flush has a similar bug - we must call bio_put before
we decrement the number of in-progress bios.

 BUG: kernel NULL pointer dereference, address: 0000000000000000
 #PF: supervisor write access in kernel mode
 #PF: error_code(0x0002) - not-present page
 PGD 11557f0067 P4D 11557f0067 PUD 0
 Oops: 0002 [#1] PREEMPT SMP
 CPU: 0 PID: 73 Comm: kworker/0:1 Not tainted 6.1.0-rc3 #5
 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-2 04/01/2014
 Workqueue: kdelayd flush_expired_bios [dm_delay]
 RIP: 0010:mempool_free+0x47/0x80
 Code: 48 89 ef 5b 5d ff e0 f3 c3 48 89 f7 e8 32 45 3f 00 48 63 53 08 48 89 c6 3b 53 04 7d 2d 48 8b 43 10 8d 4a 01 48 89 df 89 4b 08 <48> 89 2c d0 e8 b0 45 3f 00 48 8d 7b 30 5b 5d 31 c9 ba 01 00 00 00
 RSP: 0018:ffff88910036bda8 EFLAGS: 00010093
 RAX: 0000000000000000 RBX: ffff8891037b65d8 RCX: 0000000000000001
 RDX: 0000000000000000 RSI: 0000000000000202 RDI: ffff8891037b65d8
 RBP: ffff8891447ba240 R08: 0000000000012908 R09: 00000000003d0900
 R10: 0000000000000000 R11: 0000000000173544 R12: ffff889101a14000
 R13: ffff8891562ac300 R14: ffff889102b41440 R15: ffffe8ffffa00d05
 FS:  0000000000000000(0000) GS:ffff88942fa00000(0000) knlGS:0000000000000000
 CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
 CR2: 0000000000000000 CR3: 0000001102e99000 CR4: 00000000000006b0
 Call Trace:
  <TASK>
  clone_endio+0xf4/0x1c0 [dm_mod]
  clone_endio+0xf4/0x1c0 [dm_mod]
  __submit_bio+0x76/0x120
  submit_bio_noacct_nocheck+0xb6/0x2a0
  flush_expired_bios+0x28/0x2f [dm_delay]
  process_one_work+0x1b4/0x300
  worker_thread+0x45/0x3e0
  ? rescuer_thread+0x380/0x380
  kthread+0xc2/0x100
  ? kthread_complete_and_exit+0x20/0x20
  ret_from_fork+0x1f/0x30
  </TASK>
 Modules linked in: brd dm_delay dm_raid dm_mod af_packet uvesafb cfbfillrect cfbimgblt cn cfbcopyarea fb font fbdev tun autofs4 binfmt_misc configfs ipv6 virtio_rng virtio_balloon rng_core virtio_net pcspkr net_failover failover qemu_fw_cfg button mousedev raid10 raid456 libcrc32c async_raid6_recov async_memcpy async_pq raid6_pq async_xor xor async_tx raid1 raid0 md_mod sd_mod t10_pi crc64_rocksoft crc64 virtio_scsi scsi_mod evdev psmouse bsg scsi_common [last unloaded: brd]
 CR2: 0000000000000000
 ---[ end trace 0000000000000000 ]---

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Song Liu <song@kernel.org>

md/raid0, raid10: Don't set discard sectors for request queue

It should use disk_stack_limits to get a proper max_discard_sectors
rather than setting a value by stack drivers.

And there is a bug. If all member disks are rotational devices,
raid0/raid10 set max_discard_sectors. So the member devices are
not ssd/nvme, but raid0/raid10 export the wrong value. It reports
warning messages in function __blkdev_issue_discard when mkfs.xfs
like this:

[ 4616.022599] ------------[ cut here ]------------
[ 4616.027779] WARNING: CPU: 4 PID: 99634 at block/blk-lib.c:50 __blkdev_issue_discard+0x16a/0x1a0
[ 4616.140663] RIP: 0010:__blkdev_issue_discard+0x16a/0x1a0
[ 4616.146601] Code: 24 4c 89 20 31 c0 e9 fe fe ff ff c1 e8 09 8d 48 ff 4c 89 f0 4c 09 e8 48 85 c1 0f 84 55 ff ff ff b8 ea ff ff ff e9 df fe ff ff <0f> 0b 48 8d 74 24 08 e8 ea d6 00 00 48 c7 c6 20 1e 89 ab 48 c7 c7
[ 4616.167567] RSP: 0018:ffffaab88cbffca8 EFLAGS: 00010246
[ 4616.173406] RAX: ffff9ba1f9e44678 RBX: 0000000000000000 RCX: ffff9ba1c9792080
[ 4616.181376] RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffff9ba1c9792080
[ 4616.189345] RBP: 0000000000000cc0 R08: ffffaab88cbffd10 R09: 0000000000000000
[ 4616.197317] R10: 0000000000000012 R11: 0000000000000000 R12: 0000000000000000
[ 4616.205288] R13: 0000000000400000 R14: 0000000000000cc0 R15: ffff9ba1c9792080
[ 4616.213259] FS:  00007f9a5534e980(0000) GS:ffff9ba1b7c80000(0000) knlGS:0000000000000000
[ 4616.222298] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 4616.228719] CR2: 000055a390a4c518 CR3: 0000000123e40006 CR4: 00000000001706e0
[ 4616.236689] Call Trace:
[ 4616.239428]  blkdev_issue_discard+0x52/0xb0
[ 4616.244108]  blkdev_common_ioctl+0x43c/0xa00
[ 4616.248883]  blkdev_ioctl+0x116/0x280
[ 4616.252977]  __x64_sys_ioctl+0x8a/0xc0
[ 4616.257163]  do_syscall_64+0x5c/0x90
[ 4616.261164]  ? handle_mm_fault+0xc5/0x2a0
[ 4616.265652]  ? do_user_addr_fault+0x1d8/0x690
[ 4616.270527]  ? do_syscall_64+0x69/0x90
[ 4616.274717]  ? exc_page_fault+0x62/0x150
[ 4616.279097]  entry_SYSCALL_64_after_hwframe+0x63/0xcd
[ 4616.284748] RIP: 0033:0x7f9a55398c6b

Signed-off-by: Xiao Ni <xni@redhat.com>
Reported-by: Yi Zhang <yi.zhang@redhat.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Song Liu <song@kernel.org>

md/bitmap: Fix bitmap chunk size overflow issues

- limit bitmap chunk size internal u64 variable to values not overflowing
  the u32 bitmap superblock structure variable stored on persistent media
- assign bitmap chunk size internal u64 variable from unsigned values to
  avoid possible sign extension artifacts when assigning from a s32 value

The bug has been there since at least kernel 4.0.
Steps to reproduce it:
1: mdadm -C /dev/mdx -l 1 --bitmap=internal --bitmap-chunk=256M -e 1.2
-n2 /dev/rnbd1 /dev/rnbd2
2 resize member device rnbd1 and rnbd2 to 8 TB
3 mdadm --grow /dev/mdx --size=max

The bitmap_chunksize will overflow without patch.

Cc: stable@vger.kernel.org
Signed-off-by: Florian-Ewald Mueller <florian-ewald.mueller@ionos.com>
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Signed-off-by: Song Liu <song@kernel.org>

md: introduce md_ro_state

Introduce md_ro_state for mddev->ro, so it is easy to understand.

Signed-off-by: Ye Bin <yebin10@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>

md: factor out __md_set_array_info()

Factor out __md_set_array_info(). No functional change.

Signed-off-by: Ye Bin <yebin10@huawei.com>
Signed-off-by: Song Liu <song@kernel.org>

lib/raid6: drop RAID6_USE_EMPTY_ZERO_PAGE

RAID6_USE_EMPTY_ZERO_PAGE is unused and hardcoded to 0, so let's drop it.

Signed-off-by: Giulio Benetti <giulio.benetti@benettiengineering.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Song Liu <song@kernel.org>

raid5-cache: use try_cmpxchg in r5l_wake_reclaim

Use try_cmpxchg instead of cmpxchg (*ptr, old, new) == old in
r5l_wake_reclaim. 86 CMPXCHG instruction returns success in ZF flag, so
this change saves a compare after cmpxchg (and related move instruction in
front of cmpxchg).

Also, try_cmpxchg implicitly assigns old *ptr value to "old" when cmpxchg
fails. There is no need to re-read the value in the loop.

Note that the value from *ptr should be read using READ_ONCE to prevent
the compiler from merging, refetching or reordering the read.

No functional change intended.

Cc: Song Liu <song@kernel.org>
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Signed-off-by: Song Liu <song@kernel.org>

drivers/md/md-bitmap: check the return value of md_bitmap_get_counter()

Check the return value of md_bitmap_get_counter() in case it returns
NULL pointer, which will result in a null pointer dereference.

v2: update the check to include other dereference

Signed-off-by: Li Zhong <floridsleeves@gmail.com>
Signed-off-by: Song Liu <song@kernel.org>

sbitmap: Use single per-bitmap counting to wake up queued tags

sbitmap suffers from code complexity, as demonstrated by recent fixes,
and eventual lost wake ups on nested I/O completion.  The later happens,
from what I understand, due to the non-atomic nature of the updates to
wait_cnt, which needs to be subtracted and eventually reset when equal
to zero.  This two step process can eventually miss an update when a
nested completion happens to interrupt the CPU in between the wait_cnt
updates.  This is very hard to fix, as shown by the recent changes to
this code.

The code complexity arises mostly from the corner cases to avoid missed
wakes in this scenario.  In addition, the handling of wake_batch
recalculation plus the synchronization with sbq_queue_wake_up is
non-trivial.

This patchset implements the idea originally proposed by Jan [1], which
removes the need for the two-step updates of wait_cnt.  This is done by
tracking the number of completions and wakeups in always increasing,
per-bitmap counters.  Instead of having to reset the wait_cnt when it
reaches zero, we simply keep counting, and attempt to wake up N threads
in a single wait queue whenever there is enough space for a batch.
Waking up less than batch_wake shouldn't be a problem, because we
haven't changed the conditions for wake up, and the existing batch
calculation guarantees at least enough remaining completions to wake up
a batch for each queue at any time.

Performance-wise, one should expect very similar performance to the
original algorithm for the case where there is no queueing.  In both the
old algorithm and this implementation, the first thing is to check
ws_active, which bails out if there is no queueing to be managed. In the
new code, we took care to avoid accounting completions and wakeups when
there is no queueing, to not pay the cost of atomic operations
unnecessarily, since it doesn't skew the numbers.

For more interesting cases, where there is queueing, we need to take
into account the cross-communication of the atomic operations.  I've
been benchmarking by running parallel fio jobs against a single hctx
nullb in different hardware queue depth scenarios, and verifying both
IOPS and queueing.

Each experiment was repeated 5 times on a 20-CPU box, with 20 parallel
jobs. fio was issuing fixed-size randwrites with qd=64 against nullb,
varying only the hardware queue length per test.

queue size 2                 4                 8                 16                 32                 64
6.1-rc2    1681.1K (1.6K)    2633.0K (12.7K)   6940.8K (16.3K)   8172.3K (617.5K)   8391.7K (367.1K)   8606.1K (351.2K)
patched    1721.8K (15.1K)   3016.7K (3.8K)    7543.0K (89.4K)   8132.5K (303.4K)   8324.2K (230.6K)   8401.8K (284.7K)

The following is a similar experiment, ran against a nullb with a single
bitmap shared by 20 hctx spread across 2 NUMA nodes. This has 40
parallel fio jobs operating on the same device

queue size 2               4                 8               16                  32        64
6.1-rc2    1081.0K (2.3K)    957.2K (1.5K)     1699.1K (5.7K)  6178.2K (124.6K)    12227.9K (37.7K)   13286.6K (92.9K)
patched    1081.8K (2.8K)    1316.5K (5.4K)    2364.4K (1.8K)  6151.4K  (20.0K)    11893.6K (17.5K)   12385.6K (18.4K)

It has also survived blktests and a 12h-stress run against nullb. I also
ran the code against nvme and a scsi SSD, and I didn't observe
performance regression in those. If there are other tests you think I
should run, please let me know and I will follow up with results.

[1] https://lore.kernel.org/all/aef9de29-e9f5-259a-f8be-12d1b734e72@google.com/

Cc: Hugh Dickins <hughd@google.com>
Cc: Keith Busch <kbusch@kernel.org>
Cc: Liu Song <liusong@linux.alibaba.com>
Suggested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Gabriel Krisman Bertazi <krisman@suse.de>
Link: https://lore.kernel.org/r/20221105231055.25953-1-krisman@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: simplify blk_mq_realloc_tag_set_tags

Use set->nr_hw_queues for the current number of tags, and remove the
duplicate set->nr_hw_queues update in the caller.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221109100811.2413423-2-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: remove blk_mq_alloc_tag_set_tags

There is no point in trying to share any code with the realloc case when
all that is needed by the initial tagset allocation is a simple
kcalloc_node.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221109100811.2413423-1-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bfq: ignore oom_bfqq in bfq_check_waker

oom_bfqq is just a fallback bfqq, so shouldn't be used with waker
detection.

Suggested-by: Jan Kara <jack@suse.cz>
Signed-off-by: Khazhismel Kumykov <khazhy@google.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20221108181030.1611703-2-khazhy@google.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bfq: fix waker_bfqq inconsistency crash

This fixes crashes in bfq_add_bfqq_busy due to waker_bfqq being NULL,
but woken_list_node still being hashed. This would happen when
bfq_init_rq() expects a brand new allocated queue to be returned from
bfq_get_bfqq_handle_split() and unconditionally updates waker_bfqq
without resetting woken_list_node. Since we can always return oom_bfqq
when attempting to allocate, we cannot assume waker_bfqq starts as NULL.

Avoid setting woken_bfqq for oom_bfqq entirely, as it's not useful.

Crashes would have a stacktrace like:
[160595.656560]  bfq_add_bfqq_busy+0x110/0x1ec
[160595.661142]  bfq_add_request+0x6bc/0x980
[160595.666602]  bfq_insert_request+0x8ec/0x1240
[160595.671762]  bfq_insert_requests+0x58/0x9c
[160595.676420]  blk_mq_sched_insert_request+0x11c/0x198
[160595.682107]  blk_mq_submit_bio+0x270/0x62c
[160595.686759]  __submit_bio_noacct_mq+0xec/0x178
[160595.691926]  submit_bio+0x120/0x184
[160595.695990]  ext4_mpage_readpages+0x77c/0x7c8
[160595.701026]  ext4_readpage+0x60/0xb0
[160595.705158]  filemap_read_page+0x54/0x114
[160595.711961]  filemap_fault+0x228/0x5f4
[160595.716272]  do_read_fault+0xe0/0x1f0
[160595.720487]  do_fault+0x40/0x1c8

Tested by injecting random failures into bfq_get_queue, crashes go away
completely.

Fixes: 8ef3fc3a043c ("block, bfq: make shared queues inherit wakers")
Signed-off-by: Khazhismel Kumykov <khazhy@google.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/r/20221108181030.1611703-1-khazhy@google.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

drbd: Store op in drbd_peer_request

(Sort of) cherry-picked from the out-of-tree drbd9 branch. Original
commit message by Joel Colledge:

    This simplifies drbd_submit_peer_request by removing most of the
    arguments. It also makes the treatment of the op better aligned with
    that in struct bio.

    Determine fault_type dynamically using information which is already
    available instead of passing it in as a parameter.

Note: The opf in receive_rs_deallocated was changed from
REQ_OP_WRITE_ZEROES to REQ_OP_DISCARD. This was required in the
out-of-tree module, and does not matter in-tree. The opf is ignored
anyway in drbd_submit_peer_request, since the discard/zero-out is
decided by the EE_TRIM flag.

Signed-off-by: Joel Colledge <joel.colledge@linbit.com>
Signed-off-by: Christoph Böhmwalder <christoph.boehmwalder@linbit.com>
Link: https://lore.kernel.org/r/20221109133453.51652-4-christoph.boehmwalder@linbit.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

drbd: disable discard support if granularity > max

The discard_granularity describes the minimum unit of a discard.
If that is larger than the maximal discard size, we need to disable
discards completely.

Reviewed-by: Joel Colledge <joel.colledge@linbit.com>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Christoph Böhmwalder <christoph.boehmwalder@linbit.com>
Link: https://lore.kernel.org/r/20221109133453.51652-3-christoph.boehmwalder@linbit.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

drbd: use blk_queue_max_discard_sectors helper

We currently only set q->limits.max_discard_sectors, but that is not
enough. Another field, max_hw_discard_sectors, was introduced in
commit 0034af036554 ("block: make /sys/block/<dev>/queue/discard_max_bytes
writeable").

The difference is that max_discard_sectors can be changed from user
space via sysfs, while max_hw_discard_sectors is the "hardware" upper
limit.

So use this helper, which sets both.

This is also a fixup for commit 998e9cbcd615 ("drbd: cleanup
decide_on_discard_support"): if discards are not supported, that does
not necessarily mean we also want to disable write_zeroes.

Fixes: 998e9cbcd615 ("drbd: cleanup decide_on_discard_support")
Reviewed-by: Joel Colledge <joel.colledge@linbit.com>
Signed-off-by: Christoph Böhmwalder <christoph.boehmwalder@linbit.com>
Link: https://lore.kernel.org/r/20221109133453.51652-2-christoph.boehmwalder@linbit.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

ABI: sysfs-bus-pci: add documentation for p2pmem allocate

Add documentation for the p2pmem/allocate binary file which allows
for allocating p2pmem buffers in userspace for passing to drivers
that support them. (Currently only O_DIRECT to NVMe devices.)

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221021174116.7200-10-logang@deltatee.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

PCI/P2PDMA: Allow userspace VMA allocations through sysfs

Create a sysfs bin attribute called "allocate" under the existing
"p2pmem" group. The only allowable operation on this file is the mmap()
call.

When mmap() is called on this attribute, the kernel allocates a chunk of
memory from the genalloc and inserts the pages into the VMA. The
dev_pagemap .page_free callback will indicate when these pages are no
longer used and they will be put back into the genalloc.

On device unbind, remove the sysfs file before the memremap_pages are
cleaned up. This ensures unmap_mapping_range() is called on the files
inode and no new mappings can be created.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Acked-by: Bjorn Helgaas <bhelgaas@google.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221021174116.7200-9-logang@deltatee.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: set FOLL_PCI_P2PDMA in bio_map_user_iov()

When a bio's queue supports PCI P2PDMA, set FOLL_PCI_P2PDMA for
iov_iter_get_pages_flags(). This allows PCI P2PDMA pages to be
passed from userspace and enables the NVMe passthru requests to
use P2PDMA pages.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221021174116.7200-8-logang@deltatee.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: set FOLL_PCI_P2PDMA in __bio_iov_iter_get_pages()

When a bio's queue supports PCI P2PDMA, set FOLL_PCI_P2PDMA for
iov_iter_get_pages_flags(). This allows PCI P2PDMA pages to be passed
from userspace and enables the O_DIRECT path in iomap based filesystems
and direct to block devices.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221021174116.7200-7-logang@deltatee.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lib/scatterlist: add check when merging zone device pages

Consecutive zone device pages should not be merged into the same sgl
or bvec segment with other types of pages or if they belong to different
pgmaps. Otherwise getting the pgmap of a given segment is not possible
without scanning the entire segment. This helper returns true either if
both pages are not zone device pages or both pages are zone device
pages with the same pgmap.

Factor out the check for page mergability into a pages_are_mergable()
helper and add a check with zone_device_pages_are_mergeable().

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221021174116.7200-6-logang@deltatee.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: add check when merging zone device pages

Consecutive zone device pages should not be merged into the same sgl
or bvec segment with other types of pages or if they belong to different
pgmaps. Otherwise getting the pgmap of a given segment is not possible
without scanning the entire segment. This helper returns true either if
both pages are not zone device pages or both pages are zone device
pages with the same pgmap.

Add a helper to determine if zone device pages are mergeable and use
this helper in page_is_mergeable().

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221021174116.7200-5-logang@deltatee.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

iov_iter: introduce iov_iter_get_pages_[alloc_]flags()

Add iov_iter_get_pages_flags() and iov_iter_get_pages_alloc_flags()
which take a flags argument that is passed to get_user_pages_fast().

This is so that FOLL_PCI_P2PDMA can be passed when appropriate.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221021174116.7200-4-logang@deltatee.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

mm: introduce FOLL_PCI_P2PDMA to gate getting PCI P2PDMA pages

GUP Callers that expect PCI P2PDMA pages can now set FOLL_PCI_P2PDMA to
allow obtaining P2PDMA pages. If GUP is called without the flag and a
P2PDMA page is found, it will return an error in try_grab_page() or
try_grab_folio().

The check is safe to do before taking the reference to the page in both
cases seeing the page should be protected by either the appropriate
ptl or mmap_lock; or the gup fast guarantees preventing TLB flushes.

try_grab_folio() has one call site that WARNs on failure and cannot
actually deal with the failure of this function (it seems it will
get into an infinite loop). Expand the comment there to document a
couple more conditions on why it will not fail.

FOLL_PCI_P2PDMA cannot be set if FOLL_LONGTERM is set. This is to copy
fsdax until pgmap refcounts are fixed (see the link below for more
information).

Link: https://lkml.kernel.org/r/Yy4Ot5MoOhsgYLTQ@ziepe.ca
Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221021174116.7200-3-logang@deltatee.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

mm: allow multiple error returns in try_grab_page()

In order to add checks for P2PDMA memory into try_grab_page(), expand
the error return from a bool to an int/error code. Update all the
callsites handle change in usage.

Also remove the WARN_ON_ONCE() call at the callsites seeing there
already is a WARN_ON_ONCE() inside the function if it fails.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221021174116.7200-2-logang@deltatee.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: Fix some kernel-doc comments

Remove the description of @required_features in elevator_match()
to clear the below warning:

block/elevator.c:103: warning: Excess function parameter 'required_features' description in 'elevator_match'

Link: https://bugzilla.openanolis.cn/show_bug.cgi?id=2734
Fixes: ffb86425ee2c ("block: don't check for required features in elevator_match")
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Signed-off-by: Yang Li <yang.lee@linux.alibaba.com>
Link: https://lore.kernel.org/r/20221107062255.2685-1-yang.lee@linux.alibaba.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: use if-else instead of goto in blk_mq_alloc_cached_request()

if-else is more readable than goto here.

Signed-off-by: Jinlong Chen <nickyc975@zju.edu.cn>
Link: https://lore.kernel.org/r/d3306fa4e92dc9cc614edc8f1802686096bafef2.1667356813.git.nickyc975@zju.edu.cn
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: improve error handling in blk_mq_alloc_rq_map()

Use goto-style error handling like we do elsewhere in the kernel.

Signed-off-by: Jinlong Chen <nickyc975@zju.edu.cn>
Link: https://lore.kernel.org/r/bbbc2d9b17b137798c7fb92042141ca4cbbc58cc.1667356813.git.nickyc975@zju.edu.cn
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: use blk_mq_[un]quiesce_tagset

All controller namespaces share the same tagset, so we can use this
interface which does the optimal operation for parallel quiesce based on
the tagset type(e.g. blocking tagsets and non-blocking tagsets).

nvme connect_q should not be quiesced when quiesce tagset, so set the
QUEUE_FLAG_SKIP_TAGSET_QUIESCE to skip it when init connect_q.

Currently we use NVME_NS_STOPPED to ensure pairing quiescing and
unquiescing. If use blk_mq_[un]quiesce_tagset, NVME_NS_STOPPED will be
invalided, so introduce NVME_CTRL_STOPPED to replace NVME_NS_STOPPED.
In addition, we never really quiesce a single namespace. It is a better
choice to move the flag from ns to ctrl.

Signed-off-by: Chao Leng <lengchao@huawei.com>
[hch: rebased on top of prep patches]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Chao Leng <lengchao@huawei.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-15-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: add tagset quiesce interface

Drivers that have shared tagsets may need to quiesce potentially a lot
of request queues that all share a single tagset (e.g. nvme). Add an
interface to quiesce all the queues on a given tagset. This interface is
useful because it can speedup the quiesce by doing it in parallel.

Because some queues should not need to be quiesced (e.g. the nvme
connect_q) when quiescing the tagset, introduce a
QUEUE_FLAG_SKIP_TAGSET_QUIESCE flag to allow this new interface to
ski quiescing a particular queue.

Signed-off-by: Chao Leng <lengchao@huawei.com>
[hch: simplify for the per-tag_set srcu_struct]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Chao Leng <lengchao@huawei.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-14-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: pass a tagset to blk_mq_wait_quiesce_done

Nothing in blk_mq_wait_quiesce_done needs the request_queue now, so just
pass the tagset, and move the non-mq check into the only caller that
needs it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chao Leng <lengchao@huawei.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-13-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: move the srcu_struct used for quiescing to the tagset

All I/O submissions have fairly similar latencies, and a tagset-wide
quiesce is a fairly common operation.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Chao Leng <lengchao@huawei.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-12-hch@lst.de
[axboe: fix whitespace]
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: skip non-mq queues in blk_mq_quiesce_queue

For submit_bio based queues there is no (S)RCU critical section during
I/O submission and thus nothing to wait for in blk_mq_wait_quiesce_done,
so skip doing any synchronization.  No non-mq driver should be calling
this, but for now we have core callers that unconditionally call into it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-11-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme-apple: don't unquiesce the I/O queues in apple_nvme_reset_work

apple_nvme_reset_work schedules apple_nvme_remove, to be called, which
will call apple_nvme_disable and unquiesce the I/O queues.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-10-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme-pci: don't unquiesce the I/O queues in nvme_remove_dead_ctrl

nvme_remove_dead_ctrl schedules nvme_remove to be called, which will
call nvme_dev_disable and unquiesce the I/O queues.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-9-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: split nvme_kill_queues

nvme_kill_queues does two things:

 1) mark the gendisk of all namespaces dead
 2) unquiesce all I/O queues

These used to be be intertwined due to block layer issues, but aren't
any more.  So move the unquiscing of the I/O queues into the callers,
and rename the rest of the function to the now more descriptive
nvme_mark_namespaces_dead.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-8-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: don't unquiesce the admin queue in nvme_kill_queues

None of the callers of nvme_kill_queues needs it to unquiesce the
admin queues, as all of them already do it themselves:

 1) nvme_reset_work explicit call nvme_start_admin_queue toward the
    beginning of the function.  The extra call to nvme_start_admin_queue
    in nvme_reset_work this won't do anything as
    NVME_CTRL_ADMIN_Q_STOPPED will already be cleared.
 2) nvme_remove calls nvme_dev_disable with shutdown flag set to true at
    the very beginning of the function if the PCIe device was not present,
    which is the precondition for the call to nvme_kill_queues.
    nvme_dev_disable already calls nvme_start_admin_queue toward the
    end of the function when the shutdown flag is set to true, so the
    admin queue is already enabled at this point.
 3) nvme_remove_dead_ctrl schedules a workqueue to unbind the driver,
    which will end up in nvme_remove, which calls nvme_dev_disable with
    the shutdown flag.  This case will call nvme_start_admin_queue a bit
    later than before.
 4) apple_nvme_remove uses the same sequence as nvme_remove_dead_ctrl
    above.
 5) nvme_remove_namespaces only calls nvme_kill_queues when the
    controller is in the DEAD state.  That can only happen in the PCIe
    driver, and only from nvme_remove. See item 2) above for the
    conditions there.

So it is safe to just remove the call to nvme_start_admin_queue in
nvme_kill_queues without replacement.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-7-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: remove the NVME_NS_DEAD check in nvme_validate_ns

At the point where namespaces are marked dead, the controller is in a
non-live state and we won't get pass the identify commands.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-6-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: remove the NVME_NS_DEAD check in nvme_remove_invalid_namespaces

The NVME_NS_DEAD check only made sense when we revalidated namespaces
in nvme_passthrough_end for commands that affected the namespace inventory.
These days NVME_NS_DEAD is only set during reset or when tearing down
namespaces, and we always remove all namespaces right after that.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-5-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: don't remove namespaces in nvme_passthru_end

The call to nvme_remove_invalid_namespaces made sense when
nvme_passthru_end revalidated all namespaces and had to remove those that
didn't exist any more.  Since we don't revalidate from nvme_passthru_end
now, this call is entirely spurious.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-4-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme-pci: refactor the tagset handling in nvme_reset_work

The code to create, update or delete a tagset and namespaces in
nvme_reset_work is a bit convoluted.  Refactor it with a two high-level
conditionals for first probe vs reset and I/O queues vs no I/O queues
to make the code flow more clear.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-3-hch@lst.de
[axboe: fix whitespace issue]
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: set the disk capacity to 0 in blk_mark_disk_dead

nvme and xen-blkfront are already doing this to stop buffered writes from
creating dirty pages that can't be written out later.  Move it to the
common code.

This also removes the comment about the ordering from nvme, as bd_mutex
not only is gone entirely, but also hasn't been used for locking updates
to the disk size long before that, and thus the ordering requirement
documented there doesn't apply any more.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Chao Leng <lengchao@huawei.com>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20221101150050.3510-2-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: don't declare 'bfqd' as type 'void *' in bfq_group

Prevent unnecessary format conversion for bfqg->bfqd in multiple
places.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Paolo Valente <paolo.valente@unimore.it>
Link: https://lore.kernel.org/r/20221102022542.3621219-6-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: remove dead code for updating 'rq_in_driver'

Such code are not even compiled since they are inside marco "#if 0".

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Paolo Valente <paolo.valente@unimore.it>
Link: https://lore.kernel.org/r/20221102022542.3621219-5-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: cleanup bfq_activate_requeue_entity()

Just make the code a litter cleaner by removing the unnecessary
variable 'sd'.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Paolo Valente <paolo.valente@unimore.it>
Link: https://lore.kernel.org/r/20221102022542.3621219-4-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: factor out code to update 'active_entities'

Current code is a bit ugly and hard to read.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Paolo Valente <paolo.valente@unimore.it>
Link: https://lore.kernel.org/r/20221102022542.3621219-3-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: remove set but not used variable in __bfq_entity_update_weight_prio

After the patch "block, bfq: cleanup bfq_weights_tree add/remove apis"),
the local variable 'bfqd' is not used anymore, thus remove it.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Link: https://lore.kernel.org/r/20221102022542.3621219-2-yukuai1@huaweicloud.com
Fixes: afdba1461262 ("block, bfq: cleanup bfq_weights_tree add/remove apis")
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: Replace struct rq_depth with unsigned int in struct iolatency_grp

We only need a max queue depth for every iolatency to limit the inflight io
number. Replace struct rq_depth with unsigned int to simplfy "struct
iolatency_grp" and save memory.

Signed-off-by: Kemeng Shi <shikemeng@huawei.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Link: https://lore.kernel.org/r/20221018111240.22612-4-shikemeng@huawei.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: Correct comment for scale_cookie_change

Default queue depth of iolatency_grp is unlimited, so we scale down
quickly(once by half) in scale_cookie_change. Remove the "subtract
1/16th" part which is not the truth and add the actual way we
scale down.

Signed-off-by: Kemeng Shi <shikemeng@huawei.com>
Link: https://lore.kernel.org/r/20221018111240.22612-3-shikemeng@huawei.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: Remove redundant parent blkcg_gp check in check_scale_change

Function blkcg_iolatency_throttle will make sure blkg->parent is not
NULL before calls check_scale_change. And function check_scale_change
is only called in blkcg_iolatency_throttle.

Signed-off-by: Kemeng Shi <shikemeng@huawei.com>
Reviewed-by: Josef Bacik <josef@toxicpanda.com>
Link: https://lore.kernel.org/r/20221018111240.22612-2-shikemeng@huawei.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: split elevator_switch

Split an elevator_disable helper from elevator_switch for the case where
we want to switch to no scheduler at all.  This includes removing the
pointless elevator_switch_mq helper and removing the switch to no
schedule logic from blk_mq_init_sched.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221030100714.876891-8-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: don't check for required features in elevator_match

Checking for the required features in the callers simplifies the code
quite a bit, so do that.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221030100714.876891-7-hch@lst.de
[axboe: adjust for dropping patch 1, use __elevator_find()]
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: simplify the check for the current elevator in elv_iosched_show

Just compare the pointers instead of using the string based
elevator_match.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221030100714.876891-6-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: cleanup the variable naming in elv_iosched_store

Use eq for the elevator_queue as done elsewhere.  This frees e to be used
for the loop iterator instead of the odd __ prefix.  In addition rename
elv to cur to make it more clear it is the currently selected elevator.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221030100714.876891-5-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: exit elv_iosched_show early when I/O schedulers are not supported

If the tag_set has BLK_MQ_F_NO_SCHED flag set we will never show any
scheduler, so exit early.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221030100714.876891-4-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: cleanup elevator_get

Do the request_module and repeated lookup in the only caller that cares,
pick a saner name that explains where are actually doing a lookup and
use a sane calling conventions that passes the queue first.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221030100714.876891-3-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: cleanup __bfq_weights_tree_remove()

It's the same with bfq_weights_tree_remove() now.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Link: https://lore.kernel.org/r/20220916071942.214222-7-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: cleanup bfq_weights_tree add/remove apis

The 'bfq_data' and 'rb_root_cached' can both be accessed through
'bfq_queue', thus only pass 'bfq_queue' as parameter.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Link: https://lore.kernel.org/r/20220916071942.214222-6-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: do not idle if only one group is activated

Now that root group is counted into 'num_groups_with_pending_reqs',
'num_groups_with_pending_reqs > 0' is always true in
bfq_asymmetric_scenario(). Thus change the condition to '> 1'.

On the other hand, this change can enable concurrent sync io if only
one group is activated.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Link: https://lore.kernel.org/r/20220916071942.214222-5-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: refactor the counting of 'num_groups_with_pending_reqs'

Currently, bfq can't handle sync io concurrently as long as they
are not issued from root group. This is because
'bfqd->num_groups_with_pending_reqs > 0' is always true in
bfq_asymmetric_scenario().

The way that bfqg is counted into 'num_groups_with_pending_reqs':

Before this patch:
 1) root group will never be counted.
 2) Count if bfqg or it's child bfqgs have pending requests.
 3) Don't count if bfqg and it's child bfqgs complete all the requests.

After this patch:
 1) root group is counted.
 2) Count if bfqg have pending requests.
 3) Don't count if bfqg complete all the requests.

With this change, the occasion that only one group is activated can be
detected, and next patch will support concurrent sync io in the
occasion.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Link: https://lore.kernel.org/r/20220916071942.214222-4-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: record how many queues have pending requests

Prepare to refactor the counting of 'num_groups_with_pending_reqs'.

Add a counter in bfq_group, update it while tracking if bfqq have pending
requests and when bfq_bfqq_move() is called.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Link: https://lore.kernel.org/r/20220916071942.214222-3-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block, bfq: support to track if bfqq has pending requests

If entity belongs to bfqq, then entity->in_groups_with_pending_reqs
is not used currently. This patch use it to track if bfqq has pending
requests through callers of weights_tree insertion and removal.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Acked-by: Paolo Valente <paolo.valente@linaro.org>
Link: https://lore.kernel.org/r/20220916071942.214222-2-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: remove redundant call to blk_freeze_queue_start in blk_mq_destroy_queue

The calling relationship in blk_mq_destroy_queue() is as follows:

blk_mq_destroy_queue()
    ...
    -> blk_queue_start_drain()
        -> blk_freeze_queue_start()  <- called
        ...
    -> blk_freeze_queue()
        -> blk_freeze_queue_start()  <- called again
        -> blk_mq_freeze_queue_wait()
    ...

So there is a redundant call to blk_freeze_queue_start().

Replace blk_freeze_queue() with blk_mq_freeze_queue_wait() to avoid the
redundant call.

Signed-off-by: Jinlong Chen <nickyc975@zju.edu.cn>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221030083212.1251255-1-nickyc975@zju.edu.cn
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: move queue_is_mq out of blk_mq_cancel_work_sync

The only caller that needs queue_is_mq check is del_gendisk, so move the
check into it.

Signed-off-by: Jinlong Chen <nickyc975@zju.edu.cn>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221030094730.1275463-1-nickyc975@zju.edu.cn
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: simplify blksize_bits() implementation

Convert current looping-based implementation into bit operation,
which can bring improvement for:

1) bitops is more efficient for its arch-level optimization.

2) Given that blksize_bits() is inline, _if_ @size is compile-time
constant, it's possible that order_base_2() _may_ make output
compile-time evaluated, depending on code context and compiler behavior.

Signed-off-by: Dawei Li <set_pte_at@outlook.com>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/TYCP286MB23238842958D7C083D6B67CECA349@TYCP286MB2323.JPNP286.PROD.OUTLOOK.COM
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: avoid double ->queue_rq() because of early timeout

David Jeffery found one double ->queue_rq() issue, so far it can
be triggered in VM use case because of long vmexit latency or preempt
latency of vCPU pthread or long page fault in vCPU pthread, then block
IO req could be timed out before queuing the request to hardware but after
calling blk_mq_start_request() during ->queue_rq(), then timeout handler
may handle it by requeue, then double ->queue_rq() is caused, and kernel
panic.

So far, it is driver's responsibility to cover the race between timeout
and completion, so it seems supposed to be solved in driver in theory,
given driver has enough knowledge.

But it is really one common problem, lots of driver could have similar
issue, and could be hard to fix all affected drivers, even it isn't easy
for driver to handle the race. So David suggests this patch by draining
in-progress ->queue_rq() for solving this issue.

Cc: Stefan Hajnoczi <stefanha@redhat.com>
Cc: Keith Busch <kbusch@kernel.org>
Cc: virtualization@lists.linux-foundation.org
Cc: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: David Jeffery <djeffery@redhat.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Link: https://lore.kernel.org/r/20221026051957.358818-1-ming.lei@redhat.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: Micro-optimize get_max_segment_size()

This patch removes a conditional jump from get_max_segment_size(). The
x86-64 assembler code for this function without this patch is as follows:

206             return min_not_zero(mask - offset + 1,
   0x0000000000000118 <+72>:    not    %rax
   0x000000000000011b <+75>:    and    0x8(%r10),%rax
   0x000000000000011f <+79>:    add    $0x1,%rax
   0x0000000000000123 <+83>:    je     0x138 <bvec_split_segs+104>
   0x0000000000000125 <+85>:    cmp    %rdx,%rax
   0x0000000000000128 <+88>:    mov    %rdx,%r12
   0x000000000000012b <+91>:    cmovbe %rax,%r12
   0x000000000000012f <+95>:    test   %rdx,%rdx
   0x0000000000000132 <+98>:    mov    %eax,%edx
   0x0000000000000134 <+100>:   cmovne %r12d,%edx

With this patch applied:

206             return min(mask - offset, (unsigned long)lim->max_segment_size - 1) + 1;
   0x000000000000003f <+63>:    mov    0x28(%rdi),%ebp
   0x0000000000000042 <+66>:    not    %rax
   0x0000000000000045 <+69>:    and    0x8(%rdi),%rax
   0x0000000000000049 <+73>:    sub    $0x1,%rbp
   0x000000000000004d <+77>:    cmp    %rbp,%rax
   0x0000000000000050 <+80>:    cmova  %rbp,%rax
   0x0000000000000054 <+84>:    add    $0x1,%eax

Reviewed-by: Ming Lei <ming.lei@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Keith Busch <kbusch@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Link: https://lore.kernel.org/r/20221025191755.1711437-4-bvanassche@acm.org
Reviewed-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: Constify most queue limits pointers

Document which functions do not modify the queue limits.

Reviewed-by: Ming Lei <ming.lei@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Keith Busch <kbusch@kernel.org>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Link: https://lore.kernel.org/r/20221025191755.1711437-3-bvanassche@acm.org
Reviewed-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: Remove request.write_hint

Commit c75e707fe1aa ("block: remove the per-bio/request write hint")
removed all code that uses the struct request write_hint member. Hence
also remove 'write_hint' itself.

Reviewed-by: Ming Lei <ming.lei@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Link: https://lore.kernel.org/r/20221025191755.1711437-2-bvanassche@acm.org
Reviewed-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: remove bio_start_io_acct_time

bio_start_io_acct_time is not actually used anywhere, so remove it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221025155916.270303-1-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme-apple: remove an extra queue reference

Now that blk_mq_destroy_queue does not release the queue reference, there
is no need for a second admin queue reference to be held by the
apple_nvme structure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Sven Peter <sven@svenpeter.dev>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Link: https://lore.kernel.org/r/20221018135720.670094-5-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme-pci: remove an extra queue reference

Now that blk_mq_destroy_queue does not release the queue reference, there
is no need for a second admin queue reference to be held by the nvme_dev.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Link: https://lore.kernel.org/r/20221018135720.670094-4-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

scsi: remove an extra queue reference

Now that blk_mq_destroy_queue does not release the queue reference, there
is no need for a second queue reference to be held by the scsi_device.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Link: https://lore.kernel.org/r/20221018135720.670094-3-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: move the call to blk_put_queue out of blk_mq_destroy_queue

The fact that blk_mq_destroy_queue also drops a queue reference leads
to various places having to grab an extra reference.  Move the call to
blk_put_queue into the callers to allow removing the extra references.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Link: https://lore.kernel.org/r/20221018135720.670094-2-hch@lst.de
[axboe: fix fabrics_q vs admin_q conflict in nvme core.c]
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: fix up elevator_type refcounting

The current reference management logic of io scheduler modules contains
refcnt problems. For example, blk_mq_init_sched may fail before or after
the calling of e->ops.init_sched. If it fails before the calling, it does
nothing to the reference to the io scheduler module. But if it fails after
the calling, it releases the reference by calling kobject_put(&eq->kobj).

As the callers of blk_mq_init_sched can't know exactly where the failure
happens, they can't handle the reference to the io scheduler module
properly: releasing the reference on failure results in double-release if
blk_mq_init_sched has released it, and not releasing the reference results
in ghost reference if blk_mq_init_sched did not release it either.

The same problem also exists in io schedulers' init_sched implementations.

We can address the problem by adding releasing statements to the error
handling procedures of blk_mq_init_sched and init_sched implementations.
But that is counterintuitive and requires modifications to existing io
schedulers.

Instead, We make elevator_alloc get the io scheduler module references
that will be released by elevator_release. And then, we match each
elevator_get with an elevator_put. Therefore, each reference to an io
scheduler module explicitly has its own getter and releaser, and we no
longer need to worry about the refcnt problems.

The bugs and the patch can be validated with tools here:
https://github.com/nickyc975/linux_elv_refcnt_bug.git

[hch: split out a few bits into separate patches, use a non-try
      module_get in elevator_alloc]

Signed-off-by: Jinlong Chen <nickyc975@zju.edu.cn>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221020064819.1469928-5-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: check for an unchanged elevator earlier in __elevator_change

No need to find the actual elevator_type struct for this comparism,
the name is all that is needed.

Signed-off-by: Jinlong Chen <nickyc975@zju.edu.cn>
[hch: split from a larger patch]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221020064819.1469928-4-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: sanitize the elevator name before passing it to __elevator_change

The stripped name should also be used for the none check.  To do so
strip it in the caller and pass in the sanitized name.  Drop the pointless
__ prefix in the function name while we're at it.

Based on a patch from Jinlong Chen <nickyc975@zju.edu.cn>.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221020064819.1469928-3-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: add proper helpers for elevator_type module refcount management

Make sure we have helpers for all relevant module refcount operations on
the elevator_type in elevator.h, and use them.  Move the call to the get
helper in blk_mq_elv_switch_none a bit so that it is obvious with a less
verbose comment.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221020064819.1469928-2-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-wbt: don't enable throttling if default elevator is bfq

Commit b5dc5d4d1f4f ("block,bfq: Disable writeback throttling") tries to
disable wbt for bfq, it's done by calling wbt_disable_default() in
bfq_init_queue(). However, wbt is still enabled if default elevator is
bfq:

device_add_disk
 elevator_init_mq
  bfq_init_queue
   wbt_disable_default -> done nothing

 blk_register_queue
  wbt_enable_default -> wbt is enabled

Fix the problem by adding a new flag ELEVATOR_FLAG_DISBALE_WBT, bfq
will set the flag in bfq_init_queue, and following wbt_enable_default()
won't enable wbt while the flag is set.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221019121518.3865235-7-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

elevator: add new field flags in struct elevator_queue

There are only one flag to indicate that elevator is registered currently,
prepare to add a flag to disable wbt if default elevator is bfq.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221019121518.3865235-6-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-wbt: don't show valid wbt_lat_usec in sysfs while wbt is disabled

Currently, if wbt is initialized and then disabled by
wbt_disable_default(), sysfs will still show valid wbt_lat_usec, which
will confuse users that wbt is still enabled.

This patch shows wbt_lat_usec as zero if it's disabled.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reported-and-tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221019121518.3865235-5-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-wbt: make enable_state more accurate

Currently, if user disable wbt through sysfs, 'enable_state' will be
'WBT_STATE_ON_MANUAL', which will be confusing. Add a new state
'WBT_STATE_OFF_MANUAL' to cover that case.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221019121518.3865235-4-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-wbt: remove unnecessary check in wbt_enable_default()

If CONFIG_BLK_WBT_MQ is disabled, wbt_init() won't do anything.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20221019121518.3865235-3-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>

elevator: remove redundant code in elv_unregister_queue()

"elevator_queue *e" is already declared and initialized in the beginning
of elv_unregister_queue().

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Link: https://lore.kernel.org/r/20221019121518.3865235-2-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>