bcache: option to automatically run gc thread after writeback

The option gc_after_writeback is disabled by default, because garbage
collection will discard SSD data which drops cached data.

Echo 1 into /sys/fs/bcache/<UUID>/internal/gc_after_writeback will
enable this option, which wakes up gc thread when writeback accomplished
and all cached data is clean.

This option is helpful for people who cares writing performance more. In
heavy writing workload, all cached data can be clean only happens when
writeback thread cleans all cached data in I/O idle time. In such
situation a following gc running may help to shrink bcache B+ tree and
discard more clean data, which may be helpful for future writing
requests.

If you are not sure whether this is helpful for your own workload,
please leave it as disabled by default.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: introduce force_wake_up_gc()

Garbage collection thread starts to work when c->sectors_to_gc is
negative value, otherwise nothing will happen even the gc thread is
woken up by wake_up_gc().

force_wake_up_gc() sets c->sectors_to_gc to -1 before calling
wake_up_gc(), then gc thread may have chance to run if no one else sets
c->sectors_to_gc to a positive value before gc_should_run().

This routine can be called where the gc thread is woken up and required
to run in force.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: cannot set writeback_running via sysfs if no writeback kthread created

"echo 1 > writeback_running" marks writeback_running even if no
writeback kthread created as "d_strtoul(writeback_running)" will simply
set dc-> writeback_running without checking the existence of
dc->writeback_thread.

Add check for setting writeback_running via sysfs: if no writeback
kthread available, reject setting to 1.

v2 -> v3:
  * Make message on wrong assignment more clear.
  * Print name of bcache device instead of name of backing device.

Signed-off-by: Shenghui Wang <shhuiw@foxmail.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: do not mark writeback_running too early

A fresh backing device is not attached to any cache_set, and
has no writeback kthread created until first attached to some
cache_set.

But bch_cached_dev_writeback_init run
"
dc->writeback_running = true;
WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING,
&dc->disk.flags));
"
for any newly formatted backing devices.

For a fresh standalone backing device, we can get something like
following even if no writeback kthread created:
------------------------
/sys/block/bcache0/bcache# cat writeback_running
1
/sys/block/bcache0/bcache# cat writeback_rate_debug
rate: 512.0k/sec
dirty: 0.0k
target: 0.0k
proportional: 0.0k
integral: 0.0k
change: 0.0k/sec
next io: -15427384ms

The none ZERO fields are misleading as no alive writeback kthread yet.

Set dc->writeback_running false as no writeback thread created in
bch_cached_dev_writeback_init().

We have writeback thread created and woken up in bch_cached_dev_writeback
_start(). Set dc->writeback_running true before bch_writeback_queue()
called, as a writeback thread will check if dc->writeback_running is true
before writing back dirty data, and hung if false detected.

After the change, we can get the following output for a fresh standalone
backing device:
-----------------------
/sys/block/bcache0/bcache$ cat writeback_running
0
/sys/block/bcache0/bcache# cat writeback_rate_debug
rate: 0.0k/sec
dirty: 0.0k
target: 0.0k
proportional: 0.0k
integral: 0.0k
change: 0.0k/sec
next io: 0ms

v1 -> v2:
  Set dc->writeback_running before bch_writeback_queue() called,

Signed-off-by: Shenghui Wang <shhuiw@foxmail.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: update comment in sysfs.c

We have struct cached_dev allocated by kzalloc in register_bcache(),
which initializes all the fields of cached_dev with 0s. And commit
ce4c3e19e520 ("bcache: Replace bch_read_string_list() by
__sysfs_match_string()") has remove the string "default".

Update the comment.

Signed-off-by: Shenghui Wang <shhuiw@foxmail.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: update comment for bch_data_insert

commit 220bb38c21b8 ("bcache: Break up struct search") introduced
changes to struct search and s->iop. bypass/bio are fields of struct
data_insert_op now. Update the comment.

Signed-off-by: Shenghui Wang <shhuiw@foxmail.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: do not check if debug dentry is ERR or NULL explicitly on remove

debugfs_remove and debugfs_remove_recursive will check if the dentry
pointer is NULL or ERR, and will do nothing in that case.

Remove the check in cache_set_free and bch_debug_init.

Signed-off-by: Shenghui Wang <shhuiw@foxmail.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

bcache: add comment for cache_set->fill_iter

We have the following define for btree iterator:
struct btree_iter {
size_t size, used;
#ifdef CONFIG_BCACHE_DEBUG
struct btree_keys *b;
#endif
struct btree_iter_set {
struct bkey *k, *end;
} data[MAX_BSETS];
};

We can see that the length of data[] field is static MAX_BSETS, which is
defined as 4 currently.

But a btree node on disk could have too many bsets for an iterator to fit
on the stack - maybe far more that MAX_BSETS. Have to dynamically allocate
space to host more btree_iter_sets.

bch_cache_set_alloc() will make sure the pool cache_set->fill_iter can
allocate an iterator equipped with enough room that can host
(sb.bucket_size / sb.block_size)
btree_iter_sets, which is more than static MAX_BSETS.

bch_btree_node_read_done() will use that pool to allocate one iterator, to
host many bsets in one btree node.

Add more comment around cache_set->fill_iter to make code less confusing.

Signed-off-by: Shenghui Wang <shhuiw@foxmail.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blkcg: handle dying request_queue when associating a blkg

Between v3 [1] and v4 [2] of the blkg association series, the
association point moved from generic_make_request_checks(), which is
called after the request enters the queue, to bio_set_dev(), which is when
the bio is formed before submit_bio(). When the request_queue goes away,
the blkgs supporting the request_queue are destroyed and then the
q->root_blkg is set to %NULL.

This patch adds a %NULL check to blkg_tryget_closest() to prevent the
NPE caused by the above. It also adds a guard to see if the
request_queue is dying when creating a blkg to prevent creating a blkg
for a dead request_queue.

[1] https://lore.kernel.org/lkml/20180911184137.35897-1-dennisszhou@gmail.com/
[2] https://lore.kernel.org/lkml/20181126211946.77067-1-dennis@kernel.org/

Fixes: 5cdf2e3fea5e ("blkcg: associate blkg when associating a device")
Reported-and-tested-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Dennis Zhou <dennis@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: deactivate blk_stat timer in wbt_disable_default()

rwb_enabled() can't be changed when there is any inflight IO.

wbt_disable_default() may set rwb->wb_normal as zero, however the
blk_stat timer may still be pending, and the timer function will update
wrb->wb_normal again.

This patch introduces blk_stat_deactivate() and applies it in
wbt_disable_default(), then the following IO hang triggered when running
parted & switching io scheduler can be fixed:

[  369.937806] INFO: task parted:3645 blocked for more than 120 seconds.
[  369.938941]       Not tainted 4.20.0-rc6-00284-g906c801e5248 #498
[  369.939797] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[  369.940768] parted          D    0  3645   3239 0x00000000
[  369.941500] Call Trace:
[  369.941874]  ? __schedule+0x6d9/0x74c
[  369.942392]  ? wbt_done+0x5e/0x5e
[  369.942864]  ? wbt_cleanup_cb+0x16/0x16
[  369.943404]  ? wbt_done+0x5e/0x5e
[  369.943874]  schedule+0x67/0x78
[  369.944298]  io_schedule+0x12/0x33
[  369.944771]  rq_qos_wait+0xb5/0x119
[  369.945193]  ? karma_partition+0x1c2/0x1c2
[  369.945691]  ? wbt_cleanup_cb+0x16/0x16
[  369.946151]  wbt_wait+0x85/0xb6
[  369.946540]  __rq_qos_throttle+0x23/0x2f
[  369.947014]  blk_mq_make_request+0xe6/0x40a
[  369.947518]  generic_make_request+0x192/0x2fe
[  369.948042]  ? submit_bio+0x103/0x11f
[  369.948486]  ? __radix_tree_lookup+0x35/0xb5
[  369.949011]  submit_bio+0x103/0x11f
[  369.949436]  ? blkg_lookup_slowpath+0x25/0x44
[  369.949962]  submit_bio_wait+0x53/0x7f
[  369.950469]  blkdev_issue_flush+0x8a/0xae
[  369.951032]  blkdev_fsync+0x2f/0x3a
[  369.951502]  do_fsync+0x2e/0x47
[  369.951887]  __x64_sys_fsync+0x10/0x13
[  369.952374]  do_syscall_64+0x89/0x149
[  369.952819]  entry_SYSCALL_64_after_hwframe+0x49/0xbe
[  369.953492] RIP: 0033:0x7f95a1e729d4
[  369.953996] Code: Bad RIP value.
[  369.954456] RSP: 002b:00007ffdb570dd48 EFLAGS: 00000246 ORIG_RAX: 000000000000004a
[  369.955506] RAX: ffffffffffffffda RBX: 000055c2139c6be0 RCX: 00007f95a1e729d4
[  369.956389] RDX: 0000000000000001 RSI: 0000000000001261 RDI: 0000000000000004
[  369.957325] RBP: 0000000000000002 R08: 0000000000000000 R09: 000055c2139c6ce0
[  369.958199] R10: 0000000000000000 R11: 0000000000000246 R12: 000055c2139c0380
[  369.959143] R13: 0000000000000004 R14: 0000000000000100 R15: 0000000000000008

Cc: stable@vger.kernel.org
Cc: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

sbitmap: flush deferred clears for resize and shallow gets

We're missing a deferred clear off the shallow get, which can cause
a hang. Additionally, when we resize the sbitmap, we should also
flush deferred clears for good measure.

Ensure we have full coverage on batch clears, even for paths where
we would not be doing deferred clear. This makes it less error
prone for future additions.

Reported-by: Bart Van Assche <bvanassche@acm.org>
Tested-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: do not overwrite ppa list with meta list

Ehen using pblk with 0 sized metadata both ppa list and meta list
points to the same memory since pblk_dma_meta_size() returns 0 in
that case.

This patch fix that issue by ensuring that pblk_dma_meta_size()
always returns space equal to sizeof(struct pblk_sec_meta) and thus
ppa list and meta list points to different memory address.

Even that in that case drive does not really care about meta_list
pointer, this is the easiest way to fix that issue without introducing
changes in many places in the code just for 0 sized metadata case.

The same approach needs to be also done for pblk_get_sec_meta()
since we also cannot point to the same memory address in meta buffer
when we are using it for pblk recovery process

Reported-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Tested-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Igor Konopko <igor.j.konopko@intel.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: support packed metadata

pblk performs recovery of open lines by storing the LBA in the per LBA
metadata field. Recovery therefore only works for drives that has this
field.

This patch adds support for packed metadata, which store l2p mapping
for open lines in last sector of every write unit and enables drives
without per IO metadata to recover open lines.

After this patch, drives with OOB size <16B will use packed metadata
and metadata size larger than16B will continue to use the device per
IO metadata.

Reviewed-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Igor Konopko <igor.j.konopko@intel.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: disable interleaved metadata

Currently pblk only check the size of I/O metadata and does not take
into account if this metadata is in a separate buffer or interleaved
in a single metadata buffer.

In reality only the first scenario is supported, where second mode will
break pblk functionality during any IO operation.

This patch prevents pblk to be instantiated in case device only
supports interleaved metadata.

Reviewed-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Igor Konopko <igor.j.konopko@intel.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: dynamic DMA pool entry size

Currently lightnvm and pblk uses single DMA pool, for which the entry
size always is equal to PAGE_SIZE. The contents of each entry allocated
from the DMA pool consists of a PPA list (8bytes * 64), leaving
56bytes * 64 space for metadata. Since the metadata field can be bigger,
such as 128 bytes, the static size does not cover this use-case.

This patch adds support for I/O metadata above 56 bytes by changing DMA
pool size based on device meta size and allows pblk to use OOB metadata
>=16B.

Reviewed-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Igor Konopko <igor.j.konopko@intel.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: add helpers for OOB metadata

pblk currently assumes that size of OOB metadata on drive is always
equal to size of pblk_sec_meta struct. This commit add helpers which will
allow to handle different sizes of OOB metadata on drive in the future.

After this patch only OOB metadata equal to 16 bytes is supported.

Reviewed-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Igor Konopko <igor.j.konopko@intel.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: move lba list to partial read context

Currently DMA allocated memory is reused on partial read
for lba_list_mem and lba_list_media arrays. In preparation
for dynamic DMA pool sizes we need to move this arrays
into pblk_pr_ctx structures.

Reviewed-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Igor Konopko <igor.j.konopko@intel.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: avoid ref warning on cache creation

The current kref implementation around pblk global caches triggers a
false positive on refcount_inc_checked() (when called) as the kref is
initialized to 0. Instead of usint kref_inc() on a 0 reference, which is
in principle correct, use kref_init() to avoid the check. This is also
more explicit about what actually happens on cache creation.

In the process, do a small refactoring to use kref helpers.

Fixes: 1864de94ec9d6 "lightnvm: pblk: stop recreating global caches"
Signed-off-by: Javier González <javier@cnexlabs.com>
Reviewed-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: simplify geometry enumeration

Currently the geometry of an OCSSD is enumerated using a two step
approach:

First, nvm_register is called, the OCSSD identify command is issued,
and second the geometry sos and csecs values are read either from the
OCSSD identify if it is a 1.2 drive, or from the NVMe namespace data
structure if it is a 2.0 device.

This patch recombines it into a single step, such that nvm_register can
use the csecs and sos fields independent of which version is used. This
enables one to dynamically size the lightnvm subsystem dma pool.

Reviewed-by: Igor Konopko <igor.j.konopko@intel.com>
Reviewed-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: add comments wrt locking in recovery path

pblk's recovery path is single threaded and therefore a number of
assumptions regarding concurrency can be made. To avoid confusion, make
this explicit with a couple of comments in the code.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: add lock protection to list operations

Protect the list_add on the pblk_line_init_bb() error
path in case this code is used for some other purpose
in the future.

Signed-off-by: Hua Su <suhua.tanke@gmail.com>
Reviewed-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: fix spelling in comment

Signed-off-by: Hua Su <suhua.tanke@gmail.com>
Updated description.
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: remove dead code in pblk_recov_l2p

Remove the call to pblk_line_replace_data as it returns
directly because we have not set l_mg->data_next yet.

Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Reviewed-by: Javier González <javier@javigon.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: fix pblk_lines_init error handling path

The chunk metadata is allocated with vmalloc, so we need to use
vfree to free it.

Fixes: 090ee26fd512 ("lightnvm: use internal allocation for chunk log page")
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Reviewed-by: Javier González <javier@javigon.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: remove unused macro

ADDR_POOL_SIZE is not used anymore, so remove the macro.

Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Reviewed-by: Javier González <javier@javigon.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: set conservative threshold for user writes

In a worst-case scenario (random writes), OP% of sectors
in each line will be invalid, and we will then need
to move data out of 100/OP% lines to free a single line.

So, to prevent the possibility of running out of lines,
temporarily block user writes when there is less than
100/OP% free lines.

Also ensure that pblk creation does not produce instances
with insufficient over provisioning.

Insufficient over-provising is not a problem on real hardware,
but often an issue when running QEMU simulations (with few lines).
100 lines is enough to create a sane instance with the standard
(11%) over provisioning.

Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Reviewed-by: Javier González <javier@javigon.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: stop writes gracefully when running out of lines

If mapping fails (i.e. when running out of lines), handle the error
and stop writing.

Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Reviewed-by: Javier González <javier@javigon.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: account for write error sectors in emeta

Lines inflicted with write errors lines might be recovered
if they have not been recycled after write error garbage collection.

Ensure that the emeta accounting of valid lbas is correct
for such lines to avoid recovery inconsistencies.

Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Reviewed-by: Javier González <javier@javigon.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: fix resubmission of overwritten write err lbas

Make sure we only look up valid lba addresses on the resubmission path.

If an lba is invalidated in the write buffer, that sector will be
submitted to disk (as it is already mapped to a ppa), and that write
might fail, resulting in a crash when trying to look up the lba in the
mapping table (as the lba is marked as invalid).

Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Reviewed-by: Javier González <javier@javigon.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: fix chunk close trace event check

The check for chunk closes suffers from an off-by-one issue, leading
to chunk close events not being traced.

Fixes: 4c44abf43d00 ("lightnvm: pblk: add trace events for chunk states")
Signed-off-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: Fix uninitialized return value in nvm_get_chunk_meta()

With gcc 4.1:

    drivers/lightnvm/core.c: In function ‘nvm_get_bb_meta’:
    drivers/lightnvm/core.c:977: warning: ‘ret’ may be used uninitialized in this function

and

    drivers/nvme/host/lightnvm.c: In function ‘nvme_nvm_get_chk_meta’:
    drivers/nvme/host/lightnvm.c:580: warning: ‘ret’ may be used uninitialized in this function

Indeed, if (for the former) the number of channels or LUNs is zero, or
(for both) the passed number of chunks is zero, ret will be returned
uninitialized.

Fix this by preinitializing ret to zero.

Fixes: aff3fb18f957de93 ("lightnvm: move bad block and chunk state logic to core")
Fixes: a294c199455187d1 ("lightnvm: implement get log report chunk helpers")
Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

lightnvm: pblk: ignore the smeta oob area scan

The smeta area l2p mapping is empty, and actually the
recovery procedure only need to restore data sector's l2p
mapping. So ignore the smeta oob scan.

Signed-off-by: Zhoujie Wu <zjwu@marvell.com>
Reviewed-by: Javier González <javier@javigon.com>
Reviewed-by: Hans Holmberg <hans.holmberg@cnexlabs.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

dm: fix request-based dm's use of dm_wait_for_completion

The md->wait waitqueue is used by both bio-based and request-based DM.
Commit dbd3bbd291 ("dm rq: leverage blk_mq_queue_busy() to check for
outstanding IO") lost sight of the requirement that
dm_wait_for_completion() must work with all types of DM devices.

Fix md_in_flight() to call the blk-mq or bio-based method accordingly.

Fixes: dbd3bbd291 ("dm rq: leverage blk_mq_queue_busy() to check for outstanding IO")
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: fix irq vs io_queue calculations

Guenter reported an boot hang issue on HPPA after we default to 0 poll
queues. We have two issues in the queue count calculations:

1) We don't separate the poll queues from the read/write queues. This is
   important, since the former doesn't need interrupts.
2) The adjust logic is broken.

Adjust the poll queue count before doing nvme_calc_io_queues(). The poll
queue count is only limited by the IO queue count we were able to get
from the controller, not failures in the IRQ allocation loop. This
leaves nvme_calc_io_queues() just adjusting the read/write queue map.

Reported-by: Reported-by: Guenter Roeck <linux@roeck-us.net>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

dm: fix inflight IO check

After switching to percpu inflight counters, the inflight check
is totally buggy. It's perfectly valid for some counters to be
non-zero while having a total inflight IO count of 0, that's how
these kinds of counters work (inc on one CPU, dec on another).
Fix the md_in_flight() check to sum all counters before returning
a false positive, potentially.

While at it, remove the inflight read for IO completion. We don't
need it, just wake anyone that's waiting for the IO count to drop
to zero. The caller needs to re-check that value anyway when woken,
which it does.

Fixes: 6f75723190d8 ("dm: remove the pending IO accounting")
Acked-by: Mike Snitzer <snitzer@redhat.com>
Reported-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

mtip32xx: use BLK_STS_DEV_RESOURCE for device resources

For cases where we can only fail with IO in-flight, we should be using
BLK_STS_DEV_RESOURCE instead of BLK_STS_RESOURCE. The latter refers to
system wide resource constraints.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

mtip32xx: avoid using semaphores

The "cmd_slot_unal" semaphore is never used in a blocking way
but only as an atomic counter. Change the code to using
atomic_dec_if_positive() as a better API.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

dm: remove the pending IO accounting

Remove the "pending" atomic counters, that duplicate block-core's
in_flight counters, and update md_in_flight() to look at percpu
in_flight counters.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: return just one value from part_in_flight

The previous patches deleted all the code that needed the second value
returned from part_in_flight - now the kernel only uses the first value.

Consequently, part_in_flight (and blk_mq_in_flight) may be changed so that
it only returns one value.

This patch just refactors the code, there's no functional change.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: switch to per-cpu in-flight counters

Now when part_round_stats is gone, we can switch to per-cpu in-flight
counters.

We use the local-atomic type local_t, so that if part_inc_in_flight or
part_dec_in_flight is reentrantly called from an interrupt, the value will
be correct.

The other counters could be corrupted due to reentrant interrupt, but the
corruption only results in slight counter skew - the in_flight counter
must be exact, so it needs local_t.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: delete part_round_stats and switch to less precise counting

We want to convert to per-cpu in_flight counters.

The function part_round_stats needs the in_flight counter every jiffy, it
would be too costly to sum all the percpu variables every jiffy, so it
must be deleted. part_round_stats is used to calculate two counters -
time_in_queue and io_ticks.

time_in_queue can be calculated without part_round_stats, by adding the
duration of the I/O when the I/O ends (the value is almost as exact as the
previously calculated value, except that time for in-progress I/Os is not
counted).

io_ticks can be approximated by increasing the value when I/O is started
or ended and the jiffies value has changed. If the I/Os take less than a
jiffy, the value is as exact as the previously calculated value. If the
I/Os take more than a jiffy, io_ticks can drift behind the previously
calculated value.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

block: stop passing 'cpu' to all percpu stats methods

All of part_stat_* and related methods are used with preempt disabled,
so there is no need to pass cpu around to allow of them.  Just call
smp_processor_id() as needed.

Suggested-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

dm rq: leverage blk_mq_queue_busy() to check for outstanding IO

Now that request-based dm-multipath only supports blk-mq, make use of
the newly introduced blk_mq_queue_busy() to check for outstanding IO --
rather than (ab)using the block core's in_flight counters.

Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

dm: dont rewrite dm_disk(md)->part0.in_flight

generic_start_io_acct and generic_end_io_acct already update the variable
in_flight using atomic operations, so we don't have to overwrite them
again.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Merge tag 'v4.20-rc6' into for-4.21/block

Pull in v4.20-rc6 to resolve the conflict in NVMe, but also to get the
two corruption fixes. We're going to be overhauling the direct dispatch
path, and we need to do that on top of the changes we made for that
in mainline.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

sbitmap: silence bogus lockdep IRQ warning

Ming reports that lockdep spews the following trace. What this
essentially says is that the sbitmap swap_lock was used inconsistently
in IRQ enabled and disabled context, and that is usually indicative of a
bug that will cause a deadlock.

For this case, it's a false positive. The swap_lock is used from process
context only, when we swap the bits in the word and cleared mask. We
also end up doing that when we are getting a driver tag, from the
blk_mq_mark_tag_wait(), and from there we hold the waitqueue lock with
IRQs disabled. However, this isn't from an actual IRQ, it's still
process context.

In lieu of a better way to fix this, simply always disable interrupts
when grabbing the swap_lock if lockdep is enabled.

[  100.967642] ================start test sanity/001================
[  101.238280] null: module loaded
[  106.093735]
[  106.094012] =====================================================
[  106.094854] WARNING: SOFTIRQ-safe -> SOFTIRQ-unsafe lock order detected
[  106.095759] 4.20.0-rc3_5d2ee7122c73_for-next+ #1 Not tainted
[  106.096551] -----------------------------------------------------
[  106.097386] fio/1043 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire:
[  106.098231] 000000004c43fa71
(&(&sb->map[i].swap_lock)->rlock){+.+.}, at: sbitmap_get+0xd5/0x22c
[  106.099431]
[  106.099431] and this task is already holding:
[  106.100229] 000000007eec8b2f
(&(&hctx->dispatch_wait_lock)->rlock){....}, at:
blk_mq_dispatch_rq_list+0x4c1/0xd7c
[  106.101630] which would create a new lock dependency:
[  106.102326]  (&(&hctx->dispatch_wait_lock)->rlock){....} ->
(&(&sb->map[i].swap_lock)->rlock){+.+.}
[  106.103553]
[  106.103553] but this new dependency connects a SOFTIRQ-irq-safe lock:
[  106.104580]  (&sbq->ws[i].wait){..-.}
[  106.104582]
[  106.104582] ... which became SOFTIRQ-irq-safe at:
[  106.105751]   _raw_spin_lock_irqsave+0x4b/0x82
[  106.106284]   __wake_up_common_lock+0x119/0x1b9
[  106.106825]   sbitmap_queue_wake_up+0x33f/0x383
[  106.107456]   sbitmap_queue_clear+0x4c/0x9a
[  106.108046]   __blk_mq_free_request+0x188/0x1d3
[  106.108581]   blk_mq_free_request+0x23b/0x26b
[  106.109102]   scsi_end_request+0x345/0x5d7
[  106.109587]   scsi_io_completion+0x4b5/0x8f0
[  106.110099]   scsi_finish_command+0x412/0x456
[  106.110615]   scsi_softirq_done+0x23f/0x29b
[  106.111115]   blk_done_softirq+0x2a7/0x2e6
[  106.111608]   __do_softirq+0x360/0x6ad
[  106.112062]   run_ksoftirqd+0x2f/0x5b
[  106.112499]   smpboot_thread_fn+0x3a5/0x3db
[  106.113000]   kthread+0x1d4/0x1e4
[  106.113457]   ret_from_fork+0x3a/0x50
[  106.113969]
[  106.113969] to a SOFTIRQ-irq-unsafe lock:
[  106.114672]  (&(&sb->map[i].swap_lock)->rlock){+.+.}
[  106.114674]
[  106.114674] ... which became SOFTIRQ-irq-unsafe at:
[  106.116000] ...
[  106.116003]   _raw_spin_lock+0x33/0x64
[  106.116676]   sbitmap_get+0xd5/0x22c
[  106.117134]   __sbitmap_queue_get+0xe8/0x177
[  106.117731]   __blk_mq_get_tag+0x1e6/0x22d
[  106.118286]   blk_mq_get_tag+0x1db/0x6e4
[  106.118756]   blk_mq_get_driver_tag+0x161/0x258
[  106.119383]   blk_mq_dispatch_rq_list+0x28e/0xd7c
[  106.120043]   blk_mq_do_dispatch_sched+0x23a/0x287
[  106.120607]   blk_mq_sched_dispatch_requests+0x379/0x3fc
[  106.121234]   __blk_mq_run_hw_queue+0x137/0x17e
[  106.121781]   __blk_mq_delay_run_hw_queue+0x80/0x25f
[  106.122366]   blk_mq_run_hw_queue+0x151/0x187
[  106.122887]   blk_mq_sched_insert_requests+0x13f/0x175
[  106.123492]   blk_mq_flush_plug_list+0x7d6/0x81b
[  106.124042]   blk_flush_plug_list+0x392/0x3d7
[  106.124557]   blk_finish_plug+0x37/0x4f
[  106.125019]   read_pages+0x3ef/0x430
[  106.125446]   __do_page_cache_readahead+0x18e/0x2fc
[  106.126027]   force_page_cache_readahead+0x121/0x133
[  106.126621]   page_cache_sync_readahead+0x35f/0x3bb
[  106.127229]   generic_file_buffered_read+0x410/0x1860
[  106.127932]   __vfs_read+0x319/0x38f
[  106.128415]   vfs_read+0xd2/0x19a
[  106.128817]   ksys_read+0xb9/0x135
[  106.129225]   do_syscall_64+0x140/0x385
[  106.129684]   entry_SYSCALL_64_after_hwframe+0x49/0xbe
[  106.130292]
[  106.130292] other info that might help us debug this:
[  106.130292]
[  106.131226] Chain exists of:
[  106.131226]   &sbq->ws[i].wait -->
&(&hctx->dispatch_wait_lock)->rlock -->
&(&sb->map[i].swap_lock)->rlock
[  106.131226]
[  106.132865]  Possible interrupt unsafe locking scenario:
[  106.132865]
[  106.133659]        CPU0                    CPU1
[  106.134194]        ----                    ----
[  106.134733]   lock(&(&sb->map[i].swap_lock)->rlock);
[  106.135318]                                local_irq_disable();
[  106.136014]                                lock(&sbq->ws[i].wait);
[  106.136747]
lock(&(&hctx->dispatch_wait_lock)->rlock);
[  106.137742]   <Interrupt>
[  106.138110]     lock(&sbq->ws[i].wait);
[  106.138625]
[  106.138625]  *** DEADLOCK ***
[  106.138625]
[  106.139430] 3 locks held by fio/1043:
[  106.139947]  #0: 0000000076ff0fd9 (rcu_read_lock){....}, at:
hctx_lock+0x29/0xe8
[  106.140813]  #1: 000000002feb1016 (&sbq->ws[i].wait){..-.}, at:
blk_mq_dispatch_rq_list+0x4ad/0xd7c
[  106.141877]  #2: 000000007eec8b2f
(&(&hctx->dispatch_wait_lock)->rlock){....}, at:
blk_mq_dispatch_rq_list+0x4c1/0xd7c
[  106.143267]
[  106.143267] the dependencies between SOFTIRQ-irq-safe lock and the
holding lock:
[  106.144351]  -> (&sbq->ws[i].wait){..-.} ops: 82 {
[  106.144926]     IN-SOFTIRQ-W at:
[  106.145314]                       _raw_spin_lock_irqsave+0x4b/0x82
[  106.146042]                       __wake_up_common_lock+0x119/0x1b9
[  106.146785]                       sbitmap_queue_wake_up+0x33f/0x383
[  106.147567]                       sbitmap_queue_clear+0x4c/0x9a
[  106.148379]                       __blk_mq_free_request+0x188/0x1d3
[  106.149148]                       blk_mq_free_request+0x23b/0x26b
[  106.149864]                       scsi_end_request+0x345/0x5d7
[  106.150546]                       scsi_io_completion+0x4b5/0x8f0
[  106.151367]                       scsi_finish_command+0x412/0x456
[  106.152157]                       scsi_softirq_done+0x23f/0x29b
[  106.152855]                       blk_done_softirq+0x2a7/0x2e6
[  106.153537]                       __do_softirq+0x360/0x6ad
[  106.154280]                       run_ksoftirqd+0x2f/0x5b
[  106.155020]                       smpboot_thread_fn+0x3a5/0x3db
[  106.155828]                       kthread+0x1d4/0x1e4
[  106.156526]                       ret_from_fork+0x3a/0x50
[  106.157267]     INITIAL USE at:
[  106.157713]                      _raw_spin_lock_irqsave+0x4b/0x82
[  106.158542]                      prepare_to_wait_exclusive+0xa8/0x215
[  106.159421]                      blk_mq_get_tag+0x34f/0x6e4
[  106.160186]                      blk_mq_get_request+0x48e/0xaef
[  106.160997]                      blk_mq_make_request+0x27e/0xbd2
[  106.161828]                      generic_make_request+0x4d1/0x873
[  106.162661]                      submit_bio+0x20c/0x253
[  106.163379]                      mpage_bio_submit+0x44/0x4b
[  106.164142]                      mpage_readpages+0x3c2/0x407
[  106.164919]                      read_pages+0x13a/0x430
[  106.165633]                      __do_page_cache_readahead+0x18e/0x2fc
[  106.166530]                      force_page_cache_readahead+0x121/0x133
[  106.167439]                      page_cache_sync_readahead+0x35f/0x3bb
[  106.168337]                      generic_file_buffered_read+0x410/0x1860
[  106.169255]                      __vfs_read+0x319/0x38f
[  106.169977]                      vfs_read+0xd2/0x19a
[  106.170662]                      ksys_read+0xb9/0x135
[  106.171356]                      do_syscall_64+0x140/0x385
[  106.172120]                      entry_SYSCALL_64_after_hwframe+0x49/0xbe
[  106.173051]   }
[  106.173308]   ... key      at: [<ffffffff85094600>] __key.26481+0x0/0x40
[  106.174219]   ... acquired at:
[  106.174646]    _raw_spin_lock+0x33/0x64
[  106.175183]    blk_mq_dispatch_rq_list+0x4c1/0xd7c
[  106.175843]    blk_mq_do_dispatch_sched+0x23a/0x287
[  106.176518]    blk_mq_sched_dispatch_requests+0x379/0x3fc
[  106.177262]    __blk_mq_run_hw_queue+0x137/0x17e
[  106.177900]    __blk_mq_delay_run_hw_queue+0x80/0x25f
[  106.178591]    blk_mq_run_hw_queue+0x151/0x187
[  106.179207]    blk_mq_sched_insert_requests+0x13f/0x175
[  106.179926]    blk_mq_flush_plug_list+0x7d6/0x81b
[  106.180571]    blk_flush_plug_list+0x392/0x3d7
[  106.181187]    blk_finish_plug+0x37/0x4f
[  106.181737]    __se_sys_io_submit+0x171/0x304
[  106.182346]    do_syscall_64+0x140/0x385
[  106.182895]    entry_SYSCALL_64_after_hwframe+0x49/0xbe
[  106.183607]
[  106.183830] -> (&(&hctx->dispatch_wait_lock)->rlock){....} ops: 1 {
[  106.184691]    INITIAL USE at:
[  106.185119]                    _raw_spin_lock+0x33/0x64
[  106.185838]                    blk_mq_dispatch_rq_list+0x4c1/0xd7c
[  106.186697]                    blk_mq_do_dispatch_sched+0x23a/0x287
[  106.187551]                    blk_mq_sched_dispatch_requests+0x379/0x3fc
[  106.188481]                    __blk_mq_run_hw_queue+0x137/0x17e
[  106.189307]                    __blk_mq_delay_run_hw_queue+0x80/0x25f
[  106.190189]                    blk_mq_run_hw_queue+0x151/0x187
[  106.190989]                    blk_mq_sched_insert_requests+0x13f/0x175
[  106.191902]                    blk_mq_flush_plug_list+0x7d6/0x81b
[  106.192739]                    blk_flush_plug_list+0x392/0x3d7
[  106.193535]                    blk_finish_plug+0x37/0x4f
[  106.194269]                    __se_sys_io_submit+0x171/0x304
[  106.195059]                    do_syscall_64+0x140/0x385
[  106.195794]                    entry_SYSCALL_64_after_hwframe+0x49/0xbe
[  106.196705]  }
[  106.196950]  ... key      at: [<ffffffff84880620>] __key.51231+0x0/0x40
[  106.197853]  ... acquired at:
[  106.198270]    lock_acquire+0x280/0x2f3
[  106.198806]    _raw_spin_lock+0x33/0x64
[  106.199337]    sbitmap_get+0xd5/0x22c
[  106.199850]    __sbitmap_queue_get+0xe8/0x177
[  106.200450]    __blk_mq_get_tag+0x1e6/0x22d
[  106.201035]    blk_mq_get_tag+0x1db/0x6e4
[  106.201589]    blk_mq_get_driver_tag+0x161/0x258
[  106.202237]    blk_mq_dispatch_rq_list+0x5b9/0xd7c
[  106.202902]    blk_mq_do_dispatch_sched+0x23a/0x287
[  106.203572]    blk_mq_sched_dispatch_requests+0x379/0x3fc
[  106.204316]    __blk_mq_run_hw_queue+0x137/0x17e
[  106.204956]    __blk_mq_delay_run_hw_queue+0x80/0x25f
[  106.205649]    blk_mq_run_hw_queue+0x151/0x187
[  106.206269]    blk_mq_sched_insert_requests+0x13f/0x175
[  106.206997]    blk_mq_flush_plug_list+0x7d6/0x81b
[  106.207644]    blk_flush_plug_list+0x392/0x3d7
[  106.208264]    blk_finish_plug+0x37/0x4f
[  106.208814]    __se_sys_io_submit+0x171/0x304
[  106.209415]    do_syscall_64+0x140/0x385
[  106.209965]    entry_SYSCALL_64_after_hwframe+0x49/0xbe
[  106.210684]
[  106.210904]
[  106.210904] the dependencies between the lock to be acquired
[  106.210905]  and SOFTIRQ-irq-unsafe lock:
[  106.212541] -> (&(&sb->map[i].swap_lock)->rlock){+.+.} ops: 1969 {
[  106.213393]    HARDIRQ-ON-W at:
[  106.213840]                     _raw_spin_lock+0x33/0x64
[  106.214570]                     sbitmap_get+0xd5/0x22c
[  106.215282]                     __sbitmap_queue_get+0xe8/0x177
[  106.216086]                     __blk_mq_get_tag+0x1e6/0x22d
[  106.216876]                     blk_mq_get_tag+0x1db/0x6e4
[  106.217627]                     blk_mq_get_driver_tag+0x161/0x258
[  106.218465]                     blk_mq_dispatch_rq_list+0x28e/0xd7c
[  106.219326]                     blk_mq_do_dispatch_sched+0x23a/0x287
[  106.220198]                     blk_mq_sched_dispatch_requests+0x379/0x3fc
[  106.221138]                     __blk_mq_run_hw_queue+0x137/0x17e
[  106.221975]                     __blk_mq_delay_run_hw_queue+0x80/0x25f
[  106.222874]                     blk_mq_run_hw_queue+0x151/0x187
[  106.223686]                     blk_mq_sched_insert_requests+0x13f/0x175
[  106.224597]                     blk_mq_flush_plug_list+0x7d6/0x81b
[  106.225444]                     blk_flush_plug_list+0x392/0x3d7
[  106.226255]                     blk_finish_plug+0x37/0x4f
[  106.227006]                     read_pages+0x3ef/0x430
[  106.227717]                     __do_page_cache_readahead+0x18e/0x2fc
[  106.228595]                     force_page_cache_readahead+0x121/0x133
[  106.229491]                     page_cache_sync_readahead+0x35f/0x3bb
[  106.230373]                     generic_file_buffered_read+0x410/0x1860
[  106.231277]                     __vfs_read+0x319/0x38f
[  106.231986]                     vfs_read+0xd2/0x19a
[  106.232666]                     ksys_read+0xb9/0x135
[  106.233350]                     do_syscall_64+0x140/0x385
[  106.234097]                     entry_SYSCALL_64_after_hwframe+0x49/0xbe
[  106.235012]    SOFTIRQ-ON-W at:
[  106.235460]                     _raw_spin_lock+0x33/0x64
[  106.236195]                     sbitmap_get+0xd5/0x22c
[  106.236913]                     __sbitmap_queue_get+0xe8/0x177
[  106.237715]                     __blk_mq_get_tag+0x1e6/0x22d
[  106.238488]                     blk_mq_get_tag+0x1db/0x6e4
[  106.239244]                     blk_mq_get_driver_tag+0x161/0x258
[  106.240079]                     blk_mq_dispatch_rq_list+0x28e/0xd7c
[  106.240937]                     blk_mq_do_dispatch_sched+0x23a/0x287
[  106.241806]                     blk_mq_sched_dispatch_requests+0x379/0x3fc
[  106.242751]                     __blk_mq_run_hw_queue+0x137/0x17e
[  106.243579]                     __blk_mq_delay_run_hw_queue+0x80/0x25f
[  106.244469]                     blk_mq_run_hw_queue+0x151/0x187
[  106.245277]                     blk_mq_sched_insert_requests+0x13f/0x175
[  106.246191]                     blk_mq_flush_plug_list+0x7d6/0x81b
[  106.247044]                     blk_flush_plug_list+0x392/0x3d7
[  106.247859]                     blk_finish_plug+0x37/0x4f
[  106.248749]                     read_pages+0x3ef/0x430
[  106.249463]                     __do_page_cache_readahead+0x18e/0x2fc
[  106.250357]                     force_page_cache_readahead+0x121/0x133
[  106.251263]                     page_cache_sync_readahead+0x35f/0x3bb
[  106.252157]                     generic_file_buffered_read+0x410/0x1860
[  106.253084]                     __vfs_read+0x319/0x38f
[  106.253808]                     vfs_read+0xd2/0x19a
[  106.254488]                     ksys_read+0xb9/0x135
[  106.255186]                     do_syscall_64+0x140/0x385
[  106.255943]                     entry_SYSCALL_64_after_hwframe+0x49/0xbe
[  106.256867]    INITIAL USE at:
[  106.257300]                    _raw_spin_lock+0x33/0x64
[  106.258033]                    sbitmap_get+0xd5/0x22c
[  106.258747]                    __sbitmap_queue_get+0xe8/0x177
[  106.259542]                    __blk_mq_get_tag+0x1e6/0x22d
[  106.260320]                    blk_mq_get_tag+0x1db/0x6e4
[  106.261072]                    blk_mq_get_driver_tag+0x161/0x258
[  106.261902]                    blk_mq_dispatch_rq_list+0x28e/0xd7c
[  106.262762]                    blk_mq_do_dispatch_sched+0x23a/0x287
[  106.263626]                    blk_mq_sched_dispatch_requests+0x379/0x3fc
[  106.264571]                    __blk_mq_run_hw_queue+0x137/0x17e
[  106.265409]                    __blk_mq_delay_run_hw_queue+0x80/0x25f
[  106.266302]                    blk_mq_run_hw_queue+0x151/0x187
[  106.267111]                    blk_mq_sched_insert_requests+0x13f/0x175
[  106.268028]                    blk_mq_flush_plug_list+0x7d6/0x81b
[  106.268878]                    blk_flush_plug_list+0x392/0x3d7
[  106.269694]                    blk_finish_plug+0x37/0x4f
[  106.270432]                    read_pages+0x3ef/0x430
[  106.271139]                    __do_page_cache_readahead+0x18e/0x2fc
[  106.272040]                    force_page_cache_readahead+0x121/0x133
[  106.272932]                    page_cache_sync_readahead+0x35f/0x3bb
[  106.273811]                    generic_file_buffered_read+0x410/0x1860
[  106.274709]                    __vfs_read+0x319/0x38f
[  106.275407]                    vfs_read+0xd2/0x19a
[  106.276074]                    ksys_read+0xb9/0x135
[  106.276764]                    do_syscall_64+0x140/0x385
[  106.277500]                    entry_SYSCALL_64_after_hwframe+0x49/0xbe
[  106.278417]  }
[  106.278676]  ... key      at: [<ffffffff85094640>] __key.26212+0x0/0x40
[  106.279586]  ... acquired at:
[  106.280026]    lock_acquire+0x280/0x2f3
[  106.280559]    _raw_spin_lock+0x33/0x64
[  106.281101]    sbitmap_get+0xd5/0x22c
[  106.281610]    __sbitmap_queue_get+0xe8/0x177
[  106.282221]    __blk_mq_get_tag+0x1e6/0x22d
[  106.282809]    blk_mq_get_tag+0x1db/0x6e4
[  106.283368]    blk_mq_get_driver_tag+0x161/0x258
[  106.284018]    blk_mq_dispatch_rq_list+0x5b9/0xd7c
[  106.284685]    blk_mq_do_dispatch_sched+0x23a/0x287
[  106.285371]    blk_mq_sched_dispatch_requests+0x379/0x3fc
[  106.286135]    __blk_mq_run_hw_queue+0x137/0x17e
[  106.286806]    __blk_mq_delay_run_hw_queue+0x80/0x25f
[  106.287515]    blk_mq_run_hw_queue+0x151/0x187
[  106.288149]    blk_mq_sched_insert_requests+0x13f/0x175
[  106.289041]    blk_mq_flush_plug_list+0x7d6/0x81b
[  106.289912]    blk_flush_plug_list+0x392/0x3d7
[  106.290590]    blk_finish_plug+0x37/0x4f
[  106.291238]    __se_sys_io_submit+0x171/0x304
[  106.291864]    do_syscall_64+0x140/0x385
[  106.292534]    entry_SYSCALL_64_after_hwframe+0x49/0xbe

Reported-by: Ming Lei <ming.lei@redhat.com>
Tested-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Linux 4.20-rc6

Merge git://git./linux/kernel/git/davem/net

Pull networking fixes from David Miller:
 "A decent batch of fixes here. I'd say about half are for problems that
  have existed for a while, and half are for new regressions added in
  the 4.20 merge window.

   1) Fix 10G SFP phy module detection in mvpp2, from Baruch Siach.

   2) Revert bogus emac driver change, from Benjamin Herrenschmidt.

   3) Handle BPF exported data structure with pointers when building
      32-bit userland, from Daniel Borkmann.

   4) Memory leak fix in act_police, from Davide Caratti.

   5) Check RX checksum offload in RX descriptors properly in aquantia
      driver, from Dmitry Bogdanov.

   6) SKB unlink fix in various spots, from Edward Cree.

   7) ndo_dflt_fdb_dump() only works with ethernet, enforce this, from
      Eric Dumazet.

   8) Fix FID leak in mlxsw driver, from Ido Schimmel.

   9) IOTLB locking fix in vhost, from Jean-Philippe Brucker.

  10) Fix SKB truesize accounting in ipv4/ipv6/netfilter frag memory
      limits otherwise namespace exit can hang. From Jiri Wiesner.

  11) Address block parsing length fixes in x25 from Martin Schiller.

  12) IRQ and ring accounting fixes in bnxt_en, from Michael Chan.

  13) For tun interfaces, only iface delete works with rtnl ops, enforce
      this by disallowing add. From Nicolas Dichtel.

  14) Use after free in liquidio, from Pan Bian.

  15) Fix SKB use after passing to netif_receive_skb(), from Prashant
      Bhole.

  16) Static key accounting and other fixes in XPS from Sabrina Dubroca.

  17) Partially initialized flow key passed to ip6_route_output(), from
      Shmulik Ladkani.

  18) Fix RTNL deadlock during reset in ibmvnic driver, from Thomas
      Falcon.

  19) Several small TCP fixes (off-by-one on window probe abort, NULL
      deref in tail loss probe, SNMP mis-estimations) from Yuchung
      Cheng"

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (93 commits)
  net/sched: cls_flower: Reject duplicated rules also under skip_sw
  bnxt_en: Fix _bnxt_get_max_rings() for 57500 chips.
  bnxt_en: Fix NQ/CP rings accounting on the new 57500 chips.
  bnxt_en: Keep track of reserved IRQs.
  bnxt_en: Fix CNP CoS queue regression.
  net/mlx4_core: Correctly set PFC param if global pause is turned off.
  Revert "net/ibm/emac: wrong bit is used for STA control"
  neighbour: Avoid writing before skb->head in neigh_hh_output()
  ipv6: Check available headroom in ip6_xmit() even without options
  tcp: lack of available data can also cause TSO defer
  ipv6: sr: properly initialize flowi6 prior passing to ip6_route_output
  mlxsw: spectrum_switchdev: Fix VLAN device deletion via ioctl
  mlxsw: spectrum_router: Relax GRE decap matching check
  mlxsw: spectrum_switchdev: Avoid leaking FID's reference count
  mlxsw: spectrum_nve: Remove easily triggerable warnings
  ipv4: ipv6: netfilter: Adjust the frag mem limit when truesize changes
  sctp: frag_point sanity check
  tcp: fix NULL ref in tail loss probe
  tcp: Do not underestimate rwnd_limited
  net: use skb_list_del_init() to remove from RX sublists
  ...

Merge branch 'x86-urgent-for-linus' of git://git./linux/kernel/git/tip/tip

Pull x86 fixes from Ingo Molnar:
 "Three fixes: a boot parameter re-(re-)fix, a retpoline build artifact
  fix and an LLVM workaround"

* 'x86-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  x86/vdso: Drop implicit common-page-size linker flag
  x86/build: Fix compiler support check for CONFIG_RETPOLINE
  x86/boot: Clear RSDP address in boot_params for broken loaders

Merge branch 'perf-urgent-for-linus' of git://git./linux/kernel/git/tip/tip

Pull kprobes fixes from Ingo Molnar:
 "Two kprobes fixes: a blacklist fix and an instruction patching related
  corruption fix"

* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  kprobes/x86: Blacklist non-attachable interrupt functions
  kprobes/x86: Fix instruction patching corruption when copying more than one RIP-relative instruction

Merge branch 'efi-urgent-for-linus' of git://git./linux/kernel/git/tip/tip

Pull EFI fixes from Ingo Molnar:
 "Two fixes: a large-system fix and an earlyprintk fix with certain
  resolutions"

* 'efi-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
  x86/earlyprintk/efi: Fix infinite loop on some screen widths
  x86/efi: Allocate e820 buffer before calling efi_exit_boot_service

net/sched: cls_flower: Reject duplicated rules also under skip_sw

Currently, duplicated rules are rejected only for skip_hw or "none",
hence allowing users to push duplicates into HW for no reason.

Use the flower tables to protect for that.

Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Paul Blakey <paulb@mellanox.com>
Reported-by: Chris Mi <chrism@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Merge branch 'bnxt_en-Bug-fixes'

Michael Chan says:

====================
bnxt_en: Bug fixes.

The first patch fixes a regression on CoS queue setup, introduced
recently by the 57500 new chip support patches.  The rest are
fixes related to ring and resource accounting on the new 57500 chips.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

bnxt_en: Fix _bnxt_get_max_rings() for 57500 chips.

The CP rings are accounted differently on the new 57500 chips.  There
must be enough CP rings for the sum of RX and TX rings on the new
chips.  The current logic may be over-estimating the RX and TX rings.

The output parameter max_cp should be the maximum NQs capped by
MSIX vectors available for networking in the context of 57500 chips.
The existing code which uses CMPL rings capped by the MSIX vectors
works most of the time but is not always correct.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

bnxt_en: Fix NQ/CP rings accounting on the new 57500 chips.

The new 57500 chips have introduced the NQ structure in addition to
the existing CP rings in all chips.  We need to introduce a new
bnxt_nq_rings_in_use().  On legacy chips, the 2 functions are the
same and one will just call the other.  On the new chips, they
refer to the 2 separate ring structures.  The new function is now
called to determine the resource (NQ or CP rings) associated with
MSIX that are in use.

On 57500 chips, the RDMA driver does not use the CP rings so
we don't need to do the subtraction adjustment.

Fixes: 41e8d7983752 ("bnxt_en: Modify the ring reservation functions for 57500 series chips.")
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

bnxt_en: Keep track of reserved IRQs.

The new 57500 chips use 1 NQ per MSIX vector, whereas legacy chips use
1 CP ring per MSIX vector.  To better unify this, add a resv_irqs
field to struct bnxt_hw_resc.  On legacy chips, we initialize resv_irqs
with resv_cp_rings.  On new chips, we initialize it with the allocated
MSIX resources.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

bnxt_en: Fix CNP CoS queue regression.

Recent changes to support the 57500 devices have created this
regression.  The bnxt_hwrm_queue_qportcfg() call was moved to be
called earlier before the RDMA support was determined, causing
the CoS queues configuration to be set before knowing whether RDMA
was supported or not.  Fix it by moving it to the right place right
after RDMA support is determined.

Fixes: 98f04cf0f1fc ("bnxt_en: Check context memory requirements from firmware.")
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Merge tag 'char-misc-4.20-rc6' of git://git./linux/kernel/git/gregkh/char-misc

Pull char/misc driver fixes from Greg KH:
 "Here are some small driver fixes for 4.20-rc6.

  There is a hyperv fix that for some reaon took forever to get into a
  shape that could be applied to the tree properly, but resolves a much
  reported issue. The others are some gnss patches, one a bugfix and the
  two others updates to the MAINTAINERS file to properly match the gnss
  files in the tree.

  All have been in linux-next for a while with no reported issues"

* tag 'char-misc-4.20-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/char-misc:
  MAINTAINERS: exclude gnss from SIRFPRIMA2 regex matching
  MAINTAINERS: add gnss scm tree
  gnss: sirf: fix activation retry handling
  Drivers: hv: vmbus: Offload the handling of channels to two workqueues

Merge tag 'staging-4.20-rc6' of git://git./linux/kernel/git/gregkh/staging

Pull staging fixes from Greg KH:
 "Here are two staging driver bugfixes for 4.20-rc6.

  One is a revert of a previously incorrect patch that was merged a
  while ago, and the other resolves a possible buffer overrun that was
  found by code inspection.

  Both of these have been in the linux-next tree with no reported
  issues"

* tag 'staging-4.20-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/staging:
  Revert commit ef9209b642f "staging: rtl8723bs: Fix indenting errors and an off-by-one mistake in core/rtw_mlme_ext.c"
  staging: rtl8712: Fix possible buffer overrun

Merge tag 'tty-4.20-rc6' of git://git./linux/kernel/git/gregkh/tty

Pull tty driver fixes from Greg KH:
 "Here are three small tty driver fixes for 4.20-rc6

  Nothing major, just some bug fixes for reported issues. Full details
  are in the shortlog.

  All of these have been in linux-next for a while with no reported
  issues"

* tag 'tty-4.20-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/tty:
  kgdboc: fix KASAN global-out-of-bounds bug in param_set_kgdboc_var()
  tty: serial: 8250_mtk: always resume the device in probe.
  tty: do not set TTY_IO_ERROR flag if console port

Merge tag 'usb-4.20-rc6' of git://git./linux/kernel/git/gregkh/usb

Pull USB fixes from Greg KH:
 "Here are some small USB fixes for 4.20-rc6

  The "largest" here are some xhci fixes for reported issues. Also here
  is a USB core fix, some quirk additions, and a usb-serial fix which
  required the export of one of the tty layer's functions to prevent
  code duplication. The tty maintainer agreed with this change.

  All of these have been in linux-next with no reported issues"

* tag 'usb-4.20-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/usb:
  xhci: Prevent U1/U2 link pm states if exit latency is too long
  xhci: workaround CSS timeout on AMD SNPS 3.0 xHC
  USB: check usb_get_extra_descriptor for proper size
  USB: serial: console: fix reported terminal settings
  usb: quirk: add no-LPM quirk on SanDisk Ultra Flair device
  USB: Fix invalid-free bug in port_over_current_notify()
  usb: appledisplay: Add 27" Apple Cinema Display

Merge tag '4.20-rc5-smb3-fixes' of git://git.samba.org/sfrench/cifs-2.6

Pull cifs fixes from Steve French:
 "Three small fixes: a fix for smb3 direct i/o, a fix for CIFS DFS for
  stable and a minor cifs Kconfig fix"

* tag '4.20-rc5-smb3-fixes' of git://git.samba.org/sfrench/cifs-2.6:
  CIFS: Avoid returning EBUSY to upper layer VFS
  cifs: Fix separator when building path from dentry
  cifs: In Kconfig CONFIG_CIFS_POSIX needs depends on legacy (insecure cifs)

Merge tag 'dax-fixes-4.20-rc6' of git://git./linux/kernel/git/nvdimm/nvdimm

Pull dax fixes from Dan Williams:
 "The last of the known regression fixes and fallout from the Xarray
  conversion of the filesystem-dax implementation.

  On the path to debugging why the dax memory-failure injection test
  started failing after the Xarray conversion a couple more fixes for
  the dax_lock_mapping_entry(), now called dax_lock_page(), surfaced.
  Those plus the bug that started the hunt are now addressed. These
  patches have appeared in a -next release with no issues reported.

  Note the touches to mm/memory-failure.c are just the conversion to the
  new function signature for dax_lock_page().

  Summary:

   - Fix the Xarray conversion of fsdax to properly handle
     dax_lock_mapping_entry() in the presense of pmd entries

   - Fix inode destruction racing a new lock request"

* tag 'dax-fixes-4.20-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm:
  dax: Fix unlock mismatch with updated API
  dax: Don't access a freed inode
  dax: Check page->mapping isn't NULL

Merge tag 'libnvdimm-fixes-4.20-rc6' of git://git./linux/kernel/git/nvdimm/nvdimm

Pull libnvdimm fixes from Dan Williams:
 "A regression fix for the Address Range Scrub implementation, yes
  another one, and support for platforms that misalign persistent memory
  relative to the Linux memory hotplug section constraint. Longer term,
  support for sub-section memory hotplug would alleviate alignment
  waste, but until then this hack allows a 'struct page' memmap to be
  established for these misaligned memory regions.

  These have all appeared in a -next release, and thanks to Patrick for
  reporting and testing the alignment padding fix.

  Summary:

   - Unless and until the core mm handles memory hotplug units smaller
     than a section (128M), persistent memory namespaces must be padded
     to section alignment.

     The libnvdimm core already handled section collision with "System
     RAM", but some configurations overlap independent "Persistent
     Memory" ranges within a section, so additional padding injection is
     added for that case.

   - The recent reworks of the ARS (address range scrub) state machine
     to reduce the number of state flags inadvertantly missed a
     conversion of acpi_nfit_ars_rescan() call sites. Fix the regression
     whereby user-requested ARS results in a "short" scrub rather than a
     "long" scrub.

   - Fixup the unit tests to handle / test the 128M section alignment of
     mocked test resources.

* tag 'libnvdimm-fixes-4.20-rc6' of git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm:
  acpi/nfit: Fix user-initiated ARS to be "ARS-long" rather than "ARS-short"
  libnvdimm, pfn: Pad pfn namespaces relative to other regions
  tools/testing/nvdimm: Align test resources to 128M

net/mlx4_core: Correctly set PFC param if global pause is turned off.

rx_ppp and tx_ppp can be set between 0 and 255, so don't clamp to 1.

Fixes: 6e8814ceb7e8 ("net/mlx4_en: Fix mixed PFC and Global pause user control requests")
Signed-off-by: Tarick Bedeir <tarick@google.com>
Reviewed-by: Eran Ben Elisha <eranbe@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Merge branch 'fixes' of git://git./linux/kernel/git/evalenti/linux-soc-thermal

Pull thermal SoC fixes from Eduardo Valentin:
 "Fixes for armada and broadcom thermal drivers"

* 'fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/evalenti/linux-soc-thermal:
  thermal: broadcom: constify thermal_zone_of_device_ops structure
  thermal: armada: constify thermal_zone_of_device_ops structure
  thermal: bcm2835: Switch to SPDX identifier
  thermal: armada: fix legacy resource fixup
  thermal: armada: fix legacy validity test sense

Merge tag 'asm-generic-4.20' of git://git./linux/kernel/git/arnd/asm-generic

Pull asm-generic fix from Arnd Bergmann:
 "Multiple people reported a bug I introduced in asm-generic/unistd.h in
  4.20, this is the obvious bugfix to get glibc and others to correctly
  build again on new architectures that no longer provide the old
  fstatat64() family of system calls"

* tag 'asm-generic-4.20' of git://git.kernel.org/pub/scm/linux/kernel/git/arnd/asm-generic:
  asm-generic: unistd.h: fixup broken macro include.

Merge tag 'clk-fixes-for-linus' of git://git./linux/kernel/git/clk/linux

Pull clk fixes from Stephen Boyd:
 "A few clk driver fixes this time:

   - Introduce protected-clock DT binding to fix breakage on qcom
     sdm845-mtp boards where the qspi clks introduced this merge window
     cause the firmware on those boards to take down the system if we
     try to read the clk registers

   - Fix a couple off-by-one errors found by Dan Carpenter

   - Handle failure in zynq fixed factor clk driver to avoid using
     uninitialized data"

* tag 'clk-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/clk/linux:
  clk: zynqmp: Off by one in zynqmp_is_valid_clock()
  clk: mmp: Off by one in mmp_clk_add()
  clk: mvebu: Off by one bugs in cp110_of_clk_get()
  arm64: dts: qcom: sdm845-mtp: Mark protected gcc clocks
  clk: qcom: Support 'protected-clocks' property
  dt-bindings: clk: Introduce 'protected-clocks' property
  clk: zynqmp: handle fixed factor param query error

Merge tag 'xfs-4.20-fixes-3' of git://git./fs/xfs/xfs-linux

Pull xfs fixes from Darrick Wong:
 "Here are hopefully the last set of fixes for 4.20.

  There's a fix for a longstanding statfs reporting problem with project
  quotas, a correction for page cache invalidation behaviors when
  fallocating near EOF, and a fix for a broken metadata verifier return
  code.

  Finally, the most important fix is to the pipe splicing code (aka the
  generic copy_file_range fallback) to avoid pointless short directio
  reads by only asking the filesystem for as much data as there are
  available pages in the pipe buffer. Our previous fix (simulated short
  directio reads because the number of pages didn't match the length of
  the read requested) caused subtle problems on overlayfs, so that part
  is reverted.

  Anyhow, this series passes fstests -g all on xfs and overlay+xfs, and
  has passed 17 billion fsx operations problem-free since I started
  testing

  Summary:

   - Fix broken project quota inode counts

   - Fix incorrect PAGE_MASK/PAGE_SIZE usage

   - Fix incorrect return value in btree verifier

   - Fix WARN_ON remap flags false positive

   - Fix splice read overflows"

* tag 'xfs-4.20-fixes-3' of git://git.kernel.org/pub/scm/fs/xfs/xfs-linux:
  iomap: partially revert 4721a601099 (simulated directio short read on EFAULT)
  splice: don't read more than available pipe space
  vfs: allow some remap flags to be passed to vfs_clone_file_range
  xfs: fix inverted return from xfs_btree_sblock_verify_crc
  xfs: fix PAGE_MASK usage in xfs_free_file_space
  fs/xfs: fix f_ffree value for statfs when project quota is set

Revert "mm, thp: consolidate THP gfp handling into alloc_hugepage_direct_gfpmask"

This reverts commit 89c83fb539f95491be80cdd5158e6f0ce329e317.

This should have been done as part of 2f0799a0ffc0 ("mm, thp: restore
node-local hugepage allocations").  The movement of the thp allocation
policy from alloc_pages_vma() to alloc_hugepage_direct_gfpmask() was
intended to only set __GFP_THISNODE for mempolicies that are not
MPOL_BIND whereas the revert could set this regardless of mempolicy.

While the check for MPOL_BIND between alloc_hugepage_direct_gfpmask()
and alloc_pages_vma() was racy, that has since been removed since the
revert.  What is left is the possibility to use __GFP_THISNODE in
policy_node() when it is unexpected because the special handling for
hugepages in alloc_pages_vma()  was removed as part of the consolidation.

Secondly, prior to 89c83fb539f9, alloc_pages_vma() implemented a somewhat
different policy for hugepage allocations, which were allocated through
alloc_hugepage_vma().  For hugepage allocations, if the allocating
process's node is in the set of allowed nodes, allocate with
__GFP_THISNODE for that node (for MPOL_PREFERRED, use that node with
__GFP_THISNODE instead).  This was changed for shmem_alloc_hugepage() to
allow fallback to other nodes in 89c83fb539f9 as it did for new_page() in
mm/mempolicy.c which is functionally different behavior and removes the
requirement to only allocate hugepages locally.

So this commit does a full revert of 89c83fb539f9 instead of the partial
revert that was done in 2f0799a0ffc0.  The result is the same thp
allocation policy for 4.20 that was in 4.19.

Fixes: 89c83fb539f9 ("mm, thp: consolidate THP gfp handling into alloc_hugepage_direct_gfpmask")
Fixes: 2f0799a0ffc0 ("mm, thp: restore node-local hugepage allocations")
Signed-off-by: David Rientjes <rientjes@google.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Revert "net/ibm/emac: wrong bit is used for STA control"

This reverts commit 624ca9c33c8a853a4a589836e310d776620f4ab9.

This commit is completely bogus. The STACR register has two formats, old
and new, depending on the version of the IP block used. There's a pair of
device-tree properties that can be used to specify the format used:

has-inverted-stacr-oc
has-new-stacr-staopc

What this commit did was to change the bit definition used with the old
parts to match the new parts. This of course breaks the driver on all
the old ones.

Instead, the author should have set the appropriate properties in the
device-tree for the variant used on his board.

Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

scsi: Fix a harmless double shift bug

Smatch generates a warning:

    drivers/scsi/scsi_lib.c:1656 scsi_mq_done() warn: test_bit() takes a bit number

The problem is that SCMD_STATE_COMPLETE is supposed to be bit number 0
and not a mask like "(1 << 0)".  It is used like this:

        if (test_and_set_bit(SCMD_STATE_COMPLETE, &scmd->state))

The test_and_set_bit() has a shift built in so it's a double left shift
and uses bit number 1 instead of number 0.  This bug is harmless because
it's done consistently and it doesn't clash with any other flags.

Fixes: f1342709d18a ("scsi: Do not rely on blk-mq for double completions")
Reviewed-by: Keith Busch <keith.busch@intel.com>
Acked-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: remove unused function nvme_ctrl_ready

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: implement Enhanced Command Retry

A controller may have an internal state that is not able to successfully
process commands for a short duration. In such states, an immediate
command requeue is expected to fail. The driver may exceed its max
retry count, which permanently ends the command in failure when the same
command would succeed after waiting for the controller to be ready.

NVMe ratified TP 4033 provides a delay hint in the completion status
code for failed commands. Implement the retry delay based on the command
completion status and the controller's requested delay.

Note that requeued commands are handled per request_queue, not per
individual request. If multiple commands fail, the controller should
consistently report the desired delay time for retryable commands in
all CQEs, otherwise the requeue list may be kicked too soon.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: fix the structure member indentation

This is a cleanup patch which fixes the structure member indentation
introduced by the p2p:

commit c6925093d0b2 ("nvmet: Optionally use PCI P2P memory").
We don't change any functionality in this patch.

This is needed so that any future members will also follow the uniform
indentation.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Acked-by: Logan Gunthorpe <logang@deltatee.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: use unlikely for req status check

This patch adds unlikely in the nvmet request completion path for the
status check in the low level function __nvmet_req_complete.
This is helpful in the scenario where host and target connection is
working smoothly.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet-rdma: Add unlikely for response allocated check

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: Remove unused forward declaration

Signed-off-by: Israel Rukshin <israelr@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: disable fabrics SQ flow control when asked by the user

As for now, we don't care about sq_head pointer updates anyway, so
at least allow the controller to micro-optimize by omiting this update.

Note that we will probably need to support it when a controller
that requires this comes along.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: expose support for fabrics SQ flow control disable in treq

Technical Proposal introduces an indication for SQ flow control
disable support. Expose it since we are able to operate in this mode.

Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: don't override treq upon modification.

Only override the allowed parts of it.

Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
[hch: slight tweak to the NVME_TREQ_SECURE_CHANNEL_MASK definition]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: support fabrics sq flow control

Technical proposal 8005 "fabrics SQ flow control" introduces a mode
where a host and controller agree to omit sq_head pointer updates
when sending nvme completions.

In case the host indicated desire to operate in this mode (connect attribute)
the controller will return back a connect completion with sq_head value
of 0xffff as indication that it will omit sq_head pointer updates.

This mode saves us an atomic update in the I/O path.

Reviewed-by: Hannes Reinecke <hare@suse.com>
[hch: suggested better implementation]
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet-fc: remove the IN_ISR deferred scheduling options

All target lldd's call the cmd receive and op completions in non-isr
thread contexts. As such the IN_ISR options are not necessary.
Remove the functionality and flags, which also removes cpu assignments
to queues.

Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: mark nvmet_genctr static

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: enable Discovery Controller AENs

Add functions to find connections requesting Discovery Change events
and send a notification to hosts that maintain an explicit persistent
connection and have and active Asynchronous Event Request pending.
Only Hosts that have access to the Subsystem effected by the change
will receive notifications of Discovery Change event.

Call these functions each time there is a configfs change that effects
the Discover Log Pages.

Set the OAES field in the Identify Controller response to advertise the
support for Asynchronous Event Notifications.

Signed-off-by: Jay Sternberg <jay.e.sternberg@intel.com>
Reviewed-by: Phil Cayton <phil.cayton@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: allow host connect even if no allowed subsystems are exported

It is perfectly valid that a host connects to a discovery subsystem
and gets an empty discovery log page since no subsystems are
provisioned to it. No reason to disallow connecting to the discovery
subsystem all together.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Jay Sternberg <jay.e.sternberg@intel.com>
Reviewed-by: Phil Cayton <phil.cayton@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: add support to Discovery controllers for commands

Add custom get/set features to commands allowed by Discovery controllers.

Signed-off-by: Jay Sternberg <jay.e.sternberg@intel.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: add defines for discovery change async events

Add AEN/AER values as defined by the specification

Signed-off-by: Jay Sternberg <jay.e.sternberg@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: make kato and AEN processing for use by other controllers

Make common process of get/set features available to other controllers by
making simple functions static inline and others not static and prototypes
in nvmet.h file

Also remove static from nvmet_execute_async_event and add prototype to
nvmet.h to allow used by other controllers

Signed-off-by: Jay Sternberg <jay.e.sternberg@intel.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: allow Keep Alive for Discovery controller

Per change to specification allowing Discovery controllers to have
explicit persistent connections, remove restriction on Discovery
controllers allowing kato on connect.

Signed-off-by: Jay Sternberg <jay.e.sternberg@intel.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: change aen mask functions to use bit numbers

Functions nvmet_aen_disabled and nvmet_clear_aen were using
values not bit numbers ie 1 << 9 not 9 for bit function clear_bit
and test_and_set_bit.

Signed-off-by: Jay Sternberg <jay.e.sternberg@intel.com>
Reviewed-by: Phil Cayton <phil.cayton@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: provide aen bit functions for multiple controller types

Move nvmet_aen_disabled and nvmet_clear_aen in preparation for other types
of controllers to use, initially the discovery controller.

Signed-off-by: Jay Sternberg <jay.e.sternberg@intel.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvmet: use IOCB_NOWAIT for file-ns buffered I/O

This patch optimizes read command behavior when file-ns configured
with buffered I/O. Instead of offloading the buffered I/O read operations
to the worker threads, we first issue the read operation with IOCB_NOWAIT
and try and access the data from the cache. Here we only offload the
request to the worker thread and complete the request in the worker
thread context when IOCB_NOWAIT request fails.

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

nvmet: support for traffic based keep-alive

A controller that supports traffic based keep-alive can restart the keep
alive timer even when no keep-alive was not received in the kato period
as long as other admin or I/O commands were received.  For each command
set ctrl->cmd_seen to true, and when keep-alive timer expires, if any
commands were seen, resched ka_work instead of escalating to a fatal
error.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: support traffic based keep-alive

If the controller supports traffic based keep alive, we restart the keep
alive timer if any admin or io commands was completed during the kato
period.  This prevents a possible starvation of keep alive commands in
the presence of heavy traffic as in such case, we already have a health
indication from the host perspective.

Only set a comp_seen indicator in case the controller supports keep
alive to minimize the overhead for pci controllers.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: cache controller attributes

We get the controller attributes in identify, cache them as we'll need
them for traffic based keep alive support.

Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: introduce ctrl attributes enumeration

We are growing more controller attributes, so use a proper enumeration
for it.  For now just add the 128-bit hostid which we support.

Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: add a numa_node field to struct nvme_ctrl

Instead of directly poking into the struct device add a new numa_node
field to struct nvme_ctrl.  This allows fabrics drivers where ctrl->dev
is a virtual device to support NUMA affinity as well.

Also expose the field as a sysfs attribute, and populate it for the
RDMA and FC transports.

Signed-off-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

nvme: consolidate memset calls in the nvme_setup_cmd path

In function nvme_setup_cmd() we call command specific setup function
for flush, rw, and discard. Instead of calling memset in each function
lets call it once in the parent function.

This is purely code cleanup patch and it does not change any existing
functionality.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

blk-mq: re-build queue map in case of kdump kernel

Now almost all .map_queues() implementation based on managed irq
affinity doesn't update queue mapping and it just retrieves the
old built mapping, so if nr_hw_queues is changed, the mapping talbe
includes stale mapping. And only blk_mq_map_queues() may rebuild
the mapping talbe.

One case is that we limit .nr_hw_queues as 1 in case of kdump kernel.
However, drivers often builds queue mapping before allocating tagset
via pci_alloc_irq_vectors_affinity(), but set->nr_hw_queues can be set
as 1 in case of kdump kernel, so wrong queue mapping is used, and
kernel panic[1] is observed during booting.

This patch fixes the kernel panic triggerd on nvme by rebulding the
mapping table via blk_mq_map_queues().

[1] kernel panic log
[    4.438371] nvme nvme0: 16/0/0 default/read/poll queues
[    4.443277] BUG: unable to handle kernel NULL pointer dereference at 0000000000000098
[    4.444681] PGD 0 P4D 0
[    4.445367] Oops: 0000 [#1] SMP NOPTI
[    4.446342] CPU: 3 PID: 201 Comm: kworker/u33:10 Not tainted 4.20.0-rc5-00664-g5eb02f7ee1eb-dirty #459
[    4.447630] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.10.2-2.fc27 04/01/2014
[    4.448689] Workqueue: nvme-wq nvme_scan_work [nvme_core]
[    4.449368] RIP: 0010:blk_mq_map_swqueue+0xfb/0x222
[    4.450596] Code: 04 f5 20 28 ef 81 48 89 c6 39 55 30 76 93 89 d0 48 c1 e0 04 48 03 83 f8 05 00 00 48 8b 00 42 8b 3c 28 48 8b 43 58 48 8b 04 f8 <48> 8b b8 98 00 00 00 4c 0f a3 37 72 42 f0 4c 0f ab 37 66 8b b8 f6
[    4.453132] RSP: 0018:ffffc900023b3cd8 EFLAGS: 00010286
[    4.454061] RAX: 0000000000000000 RBX: ffff888174448000 RCX: 0000000000000001
[    4.456480] RDX: 0000000000000001 RSI: ffffe8feffc506c0 RDI: 0000000000000001
[    4.458750] RBP: ffff88810722d008 R08: ffff88817647a880 R09: 0000000000000002
[    4.464580] R10: ffffc900023b3c10 R11: 0000000000000004 R12: ffff888174448538
[    4.467803] R13: 0000000000000004 R14: 0000000000000001 R15: 0000000000000001
[    4.469220] FS:  0000000000000000(0000) GS:ffff88817bac0000(0000) knlGS:0000000000000000
[    4.471554] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[    4.472464] CR2: 0000000000000098 CR3: 0000000174e4e001 CR4: 0000000000760ee0
[    4.474264] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[    4.476007] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[    4.477061] PKRU: 55555554
[    4.477464] Call Trace:
[    4.478731]  blk_mq_init_allocated_queue+0x36a/0x3ad
[    4.479595]  blk_mq_init_queue+0x32/0x4e
[    4.480178]  nvme_validate_ns+0x98/0x623 [nvme_core]
[    4.480963]  ? nvme_submit_sync_cmd+0x1b/0x20 [nvme_core]
[    4.481685]  ? nvme_identify_ctrl.isra.8+0x70/0xa0 [nvme_core]
[    4.482601]  nvme_scan_work+0x23a/0x29b [nvme_core]
[    4.483269]  ? _raw_spin_unlock_irqrestore+0x25/0x38
[    4.483930]  ? try_to_wake_up+0x38d/0x3b3
[    4.484478]  ? process_one_work+0x179/0x2fc
[    4.485118]  process_one_work+0x1d3/0x2fc
[    4.485655]  ? rescuer_thread+0x2ae/0x2ae
[    4.486196]  worker_thread+0x1e9/0x2be
[    4.486841]  kthread+0x115/0x11d
[    4.487294]  ? kthread_park+0x76/0x76
[    4.487784]  ret_from_fork+0x3a/0x50
[    4.488322] Modules linked in: nvme nvme_core qemu_fw_cfg virtio_scsi ip_tables
[    4.489428] Dumping ftrace buffer:
[    4.489939]    (ftrace buffer empty)
[    4.490492] CR2: 0000000000000098
[    4.491052] ---[ end trace 03cd268ad5a86ff7 ]---

Cc: Christoph Hellwig <hch@lst.de>
Cc: linux-nvme@lists.infradead.org
Cc: David Milburn <dmilburn@redhat.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>