openwrt/staging/blogic.git
5 years agoiomap: use bio_release_pages in iomap_dio_bio_end_io
Christoph Hellwig [Wed, 26 Jun 2019 13:49:24 +0000 (15:49 +0200)]
iomap: use bio_release_pages in iomap_dio_bio_end_io

Use bio_release_pages instead of duplicating it.

Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: use bio_release_pages in bio_map_user_iov
Christoph Hellwig [Wed, 26 Jun 2019 13:49:23 +0000 (15:49 +0200)]
block: use bio_release_pages in bio_map_user_iov

Use bio_release_pages instead of open coding it.

Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: use bio_release_pages in bio_unmap_user
Christoph Hellwig [Wed, 26 Jun 2019 13:49:22 +0000 (15:49 +0200)]
block: use bio_release_pages in bio_unmap_user

Use bio_release_pages instead of open coding it.

Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: optionally mark pages dirty in bio_release_pages
Christoph Hellwig [Wed, 26 Jun 2019 13:49:21 +0000 (15:49 +0200)]
block: optionally mark pages dirty in bio_release_pages

A lot of callers of bio_release_pages also want to mark the released
pages as dirty.  Add a mark_dirty parameter to avoid a second
relatively expensive bio_for_each_segment_all loop.

Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: move the BIO_NO_PAGE_REF check into bio_release_pages
Christoph Hellwig [Wed, 26 Jun 2019 13:49:20 +0000 (15:49 +0200)]
block: move the BIO_NO_PAGE_REF check into bio_release_pages

Move the BIO_NO_PAGE_REF check into bio_release_pages instead of
duplicating it in both callers.

Also make the function available outside of bio.c so that we can
reuse it in other direct I/O implementations.

Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: skd_main.c: Remove call to memset after dma_alloc_coherent
Fuqian Huang [Thu, 27 Jun 2019 17:35:16 +0000 (01:35 +0800)]
block: skd_main.c: Remove call to memset after dma_alloc_coherent

In commit af7ddd8a627c
("Merge tag 'dma-mapping-4.21' of git://git.infradead.org/users/hch/dma-mapping"),
dma_alloc_coherent has already zeroed the memory.
So memset is not needed.

Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Fuqian Huang <huangfq.daxian@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: mtip32xx: Remove call to memset after dma_alloc_coherent
Fuqian Huang [Thu, 27 Jun 2019 17:35:04 +0000 (01:35 +0800)]
block: mtip32xx: Remove call to memset after dma_alloc_coherent

In commit af7ddd8a627c
("Merge tag 'dma-mapping-4.21' of git://git.infradead.org/users/hch/dma-mapping"),
dma_alloc_coherent has already zeroed the memory.
So memset is not needed.

Signed-off-by: Fuqian Huang <huangfq.daxian@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: sed-opal: "Never True" conditions
Revanth Rajashekar [Thu, 27 Jun 2019 22:31:09 +0000 (16:31 -0600)]
block: sed-opal: "Never True" conditions

'who' an unsigned variable in stucture opal_session_info
can never be lesser than zero. Hence, the condition
"who < OPAL_ADMIN1" can never be true.

Signed-off-by: Revanth Rajashekar <revanth.rajashekar@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: sed-opal: PSID reverttper capability
Revanth Rajashekar [Thu, 27 Jun 2019 22:30:02 +0000 (16:30 -0600)]
block: sed-opal: PSID reverttper capability

PSID is a 32 character password printed on the drive label,
to prove its physical access. This PSID reverttper function
is very useful to regain the control over the drive when it
is locked and the user can no longer access it because of some
failures. However, *all the data on the drive is completely
erased*. This method is advisable only when the user is exhausted
of all other recovery methods.

PSID capabilities are described in:
https://trustedcomputinggroup.org/wp-content/uploads/TCG_Storage-Opal_Feature_Set_PSID_v1.00_r1.00.pdf

Signed-off-by: Revanth Rajashekar <revanth.rajashekar@intel.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, documentation: Document discard_zeroes_data, fua, max_discard_segments and...
Bart Van Assche [Fri, 28 Jun 2019 20:07:45 +0000 (13:07 -0700)]
block, documentation: Document discard_zeroes_data, fua, max_discard_segments and write_zeroes_max_bytes

Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, documentation: Explain the word 'segments'
Bart Van Assche [Fri, 28 Jun 2019 20:07:44 +0000 (13:07 -0700)]
block, documentation: Explain the word 'segments'

Several block layer users who are not kernel developers do not know that
the word 'segment' refers to an element in a DMA scatter/gather list. Make
the block layer documentation easier to understand by stating explicitly
what the word 'segment' stands for.

Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, documentation: Sort queue sysfs attribute names alphabetically
Bart Van Assche [Fri, 28 Jun 2019 20:07:43 +0000 (13:07 -0700)]
block, documentation: Sort queue sysfs attribute names alphabetically

Commit f9824952ee1c ("block: update sysfs documentation") # v5.0 broke the
alphabetical order of the sysfs attribute names. List queue sysfs attribute
names alphabetically.

Cc: Damien Le Moal <damien.lemoal@wdc.com>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, documentation: Fix wbt_lat_usec documentation
Bart Van Assche [Fri, 28 Jun 2019 20:07:42 +0000 (13:07 -0700)]
block, documentation: Fix wbt_lat_usec documentation

Fix the spelling of the wbt_lat_usec sysfs attribute.

Fixes: 87760e5eef35 ("block: hook up writeback throttling") # v4.10.
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agonull_blk: fix type mismatch null_handle_cmd()
Chaitanya Kulkarni [Fri, 28 Jun 2019 23:29:04 +0000 (16:29 -0700)]
null_blk: fix type mismatch null_handle_cmd()

In null_handle_cmd() when device is configured as zoned, variable op is
decalred as an int, where it is used to hold values of type
REQ_OP_XXX which is of type enum req_opf. Change the type from
int to enum req_opf.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, bfq: NULL out the bic when it's no longer valid
Douglas Anderson [Fri, 28 Jun 2019 04:44:09 +0000 (21:44 -0700)]
block, bfq: NULL out the bic when it's no longer valid

In reboot tests on several devices we were seeing a "use after free"
when slub_debug or KASAN was enabled.  The kernel complained about:

  Unable to handle kernel paging request at virtual address 6b6b6c2b

...which is a classic sign of use after free under slub_debug.  The
stack crawl in kgdb looked like:

 0  test_bit (addr=<optimized out>, nr=<optimized out>)
 1  bfq_bfqq_busy (bfqq=<optimized out>)
 2  bfq_select_queue (bfqd=<optimized out>)
 3  __bfq_dispatch_request (hctx=<optimized out>)
 4  bfq_dispatch_request (hctx=<optimized out>)
 5  0xc056ef00 in blk_mq_do_dispatch_sched (hctx=0xed249440)
 6  0xc056f728 in blk_mq_sched_dispatch_requests (hctx=0xed249440)
 7  0xc0568d24 in __blk_mq_run_hw_queue (hctx=0xed249440)
 8  0xc0568d94 in blk_mq_run_work_fn (work=<optimized out>)
 9  0xc024c5c4 in process_one_work (worker=0xec6d4640, work=0xed249480)
 10 0xc024cff4 in worker_thread (__worker=0xec6d4640)

Digging in kgdb, it could be found that, though bfqq looked fine,
bfqq->bic had been freed.

Through further digging, I postulated that perhaps it is illegal to
access a "bic" (AKA an "icq") after bfq_exit_icq() had been called
because the "bic" can be freed at some point in time after this call
is made.  I confirmed that there certainly were cases where the exact
crashing code path would access the "bic" after bfq_exit_icq() had
been called.  Sspecifically I set the "bfqq->bic" to (void *)0x7 and
saw that the bic was 0x7 at the time of the crash.

To understand a bit more about why this crash was fairly uncommon (I
saw it only once in a few hundred reboots), you can see that much of
the time bfq_exit_icq_fbqq() fully frees the bfqq and thus it can't
access the ->bic anymore.  The only case it doesn't is if
bfq_put_queue() sees a reference still held.

However, even in the case when bfqq isn't freed, the crash is still
rare.  Why?  I tracked what happened to the "bic" after the exit
routine.  It doesn't get freed right away.  Rather,
put_io_context_active() eventually called put_io_context() which
queued up freeing on a workqueue.  The freeing then actually happened
later than that through call_rcu().  Despite all these delays, some
extra debugging showed that all the hoops could be jumped through in
time and the memory could be freed causing the original crash.  Phew!

To make a long story short, assuming it truly is illegal to access an
icq after the "exit_icq" callback is finished, this patch is needed.

Cc: stable@vger.kernel.org
Reviewed-by: Paolo Valente <paolo.valente@unimore.it>
Signed-off-by: Douglas Anderson <dianders@chromium.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: add reclaimed_journal_buckets to struct cache_set
Coly Li [Fri, 28 Jun 2019 12:00:00 +0000 (20:00 +0800)]
bcache: add reclaimed_journal_buckets to struct cache_set

Now we have counters for how many times jouranl is reclaimed, how many
times cached dirty btree nodes are flushed, but we don't know how many
jouranl buckets are really reclaimed.

This patch adds reclaimed_journal_buckets into struct cache_set, this
is an increasing only counter, to tell how many journal buckets are
reclaimed since cache set runs. From all these three counters (reclaim,
reclaimed_journal_buckets, flush_write), we can have idea how well
current journal space reclaim code works.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: performance improvement for btree_flush_write()
Coly Li [Fri, 28 Jun 2019 11:59:59 +0000 (19:59 +0800)]
bcache: performance improvement for btree_flush_write()

This patch improves performance for btree_flush_write() in following
ways,
- Use another spinlock journal.flush_write_lock to replace the very
  hot journal.lock. We don't have to use journal.lock here, selecting
  candidate btree nodes takes a lot of time, hold journal.lock here will
  block other jouranling threads and drop the overall I/O performance.
- Only select flushing btree node from c->btree_cache list. When the
  machine has a large system memory, mca cache may have a huge number of
  cached btree nodes. Iterating all the cached nodes will take a lot
  of CPU time, and most of the nodes on c->btree_cache_freeable and
  c->btree_cache_freed lists are cleared and have need to flush. So only
  travel mca list c->btree_cache to select flushing btree node should be
  enough for most of the cases.
- Don't iterate whole c->btree_cache list, only reversely select first
  BTREE_FLUSH_NR btree nodes to flush. Iterate all btree nodes from
  c->btree_cache and select the oldest journal pin btree nodes consumes
  huge number of CPU cycles if the list is huge (push and pop a node
  into/out of a heap is expensive). The last several dirty btree nodes
  on the tail of c->btree_cache list are earlest allocated and cached
  btree nodes, they are relative to the oldest journal pin btree nodes.
  Therefore only flushing BTREE_FLUSH_NR btree nodes from tail of
  c->btree_cache probably includes the oldest journal pin btree nodes.

In my testing, the above change decreases 50%+ CPU consumption when
journal space is full. Some times IOPS drops to 0 for 5-8 seconds,
comparing blocking I/O for 120+ seconds in previous code, this is much
better. Maybe there is room to improve in future, but at this momment
the fix looks fine and performs well in my testing.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: fix race in btree_flush_write()
Coly Li [Fri, 28 Jun 2019 11:59:58 +0000 (19:59 +0800)]
bcache: fix race in btree_flush_write()

There is a race between mca_reap(), btree_node_free() and journal code
btree_flush_write(), which results very rare and strange deadlock or
panic and are very hard to reproduce.

Let me explain how the race happens. In btree_flush_write() one btree
node with oldest journal pin is selected, then it is flushed to cache
device, the select-and-flush is a two steps operation. Between these two
steps, there are something may happen inside the race window,
- The selected btree node was reaped by mca_reap() and allocated to
  other requesters for other btree node.
- The slected btree node was selected, flushed and released by mca
  shrink callback bch_mca_scan().
When btree_flush_write() tries to flush the selected btree node, firstly
b->write_lock is held by mutex_lock(). If the race happens and the
memory of selected btree node is allocated to other btree node, if that
btree node's write_lock is held already, a deadlock very probably
happens here. A worse case is the memory of the selected btree node is
released, then all references to this btree node (e.g. b->write_lock)
will trigger NULL pointer deference panic.

This race was introduced in commit cafe56359144 ("bcache: A block layer
cache"), and enlarged by commit c4dc2497d50d ("bcache: fix high CPU
occupancy during journal"), which selected 128 btree nodes and flushed
them one-by-one in a quite long time period.

Such race is not easy to reproduce before. On a Lenovo SR650 server with
48 Xeon cores, and configure 1 NVMe SSD as cache device, a MD raid0
device assembled by 3 NVMe SSDs as backing device, this race can be
observed around every 10,000 times btree_flush_write() gets called. Both
deadlock and kernel panic all happened as aftermath of the race.

The idea of the fix is to add a btree flag BTREE_NODE_journal_flush. It
is set when selecting btree nodes, and cleared after btree nodes
flushed. Then when mca_reap() selects a btree node with this bit set,
this btree node will be skipped. Since mca_reap() only reaps btree node
without BTREE_NODE_journal_flush flag, such race is avoided.

Once corner case should be noticed, that is btree_node_free(). It might
be called in some error handling code path. For example the following
code piece from btree_split(),
        2149 err_free2:
        2150         bkey_put(b->c, &n2->key);
        2151         btree_node_free(n2);
        2152         rw_unlock(true, n2);
        2153 err_free1:
        2154         bkey_put(b->c, &n1->key);
        2155         btree_node_free(n1);
        2156         rw_unlock(true, n1);
At line 2151 and 2155, the btree node n2 and n1 are released without
mac_reap(), so BTREE_NODE_journal_flush also needs to be checked here.
If btree_node_free() is called directly in such error handling path,
and the selected btree node has BTREE_NODE_journal_flush bit set, just
delay for 1 us and retry again. In this case this btree node won't
be skipped, just retry until the BTREE_NODE_journal_flush bit cleared,
and free the btree node memory.

Fixes: cafe56359144 ("bcache: A block layer cache")
Signed-off-by: Coly Li <colyli@suse.de>
Reported-and-tested-by: kbuild test robot <lkp@intel.com>
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: remove retry_flush_write from struct cache_set
Coly Li [Fri, 28 Jun 2019 11:59:57 +0000 (19:59 +0800)]
bcache: remove retry_flush_write from struct cache_set

In struct cache_set, retry_flush_write is added for commit c4dc2497d50d
("bcache: fix high CPU occupancy during journal") which is reverted in
previous patch.

Now it is useless anymore, and this patch removes it from bcache code.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: add comments for mutex_lock(&b->write_lock)
Coly Li [Fri, 28 Jun 2019 11:59:56 +0000 (19:59 +0800)]
bcache: add comments for mutex_lock(&b->write_lock)

When accessing or modifying BTREE_NODE_dirty bit, it is not always
necessary to acquire b->write_lock. In bch_btree_cache_free() and
mca_reap() acquiring b->write_lock is necessary, and this patch adds
comments to explain why mutex_lock(&b->write_lock) is necessary for
checking or clearing BTREE_NODE_dirty bit there.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: only clear BTREE_NODE_dirty bit when it is set
Coly Li [Fri, 28 Jun 2019 11:59:55 +0000 (19:59 +0800)]
bcache: only clear BTREE_NODE_dirty bit when it is set

In bch_btree_cache_free() and btree_node_free(), BTREE_NODE_dirty is
always set no matter btree node is dirty or not. The code looks like
this,
if (btree_node_dirty(b))
btree_complete_write(b, btree_current_write(b));
clear_bit(BTREE_NODE_dirty, &b->flags);

Indeed if btree_node_dirty(b) returns false, it means BTREE_NODE_dirty
bit is cleared, then it is unnecessary to clear the bit again.

This patch only clears BTREE_NODE_dirty when btree_node_dirty(b) is
true (the bit is set), to save a few CPU cycles.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: Revert "bcache: fix high CPU occupancy during journal"
Coly Li [Fri, 28 Jun 2019 11:59:54 +0000 (19:59 +0800)]
bcache: Revert "bcache: fix high CPU occupancy during journal"

This reverts commit c4dc2497d50d9c6fb16aa0d07b6a14f3b2adb1e0.

This patch enlarges a race between normal btree flush code path and
flush_btree_write(), which causes deadlock when journal space is
exhausted. Reverts this patch makes the race window from 128 btree
nodes to only 1 btree nodes.

Fixes: c4dc2497d50d ("bcache: fix high CPU occupancy during journal")
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Cc: Tang Junhui <tang.junhui.linux@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: Revert "bcache: free heap cache_set->flush_btree in bch_journal_free"
Coly Li [Fri, 28 Jun 2019 11:59:53 +0000 (19:59 +0800)]
bcache: Revert "bcache: free heap cache_set->flush_btree in bch_journal_free"

This reverts commit 6268dc2c4703aabfb0b35681be709acf4c2826c6.

This patch depends on commit c4dc2497d50d ("bcache: fix high CPU
occupancy during journal") which is reverted in previous patch. So
revert this one too.

Fixes: 6268dc2c4703 ("bcache: free heap cache_set->flush_btree in bch_journal_free")
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Cc: Shenghui Wang <shhuiw@foxmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: shrink btree node cache after bch_btree_check()
Coly Li [Fri, 28 Jun 2019 11:59:52 +0000 (19:59 +0800)]
bcache: shrink btree node cache after bch_btree_check()

When cache set starts, bch_btree_check() will check all bkeys on cache
device by calculating the checksum. This operation will consume a huge
number of system memory if there are a lot of data cached. Since bcache
uses its own mca cache to maintain all its read-in btree nodes, and only
releases the cache space when system memory manage code starts to shrink
caches. Then before memory manager code to call the mca cache shrinker
callback, bcache mca cache will compete memory resource with user space
application, which may have nagive effect to performance of user space
workloads (e.g. data base, or I/O service of distributed storage node).

This patch tries to call bcache mca shrinker routine to proactively
release mca cache memory, to decrease the memory pressure of system and
avoid negative effort of the overall system I/O performance.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: set largest seq to ja->seq[bucket_index] in journal_read_bucket()
Coly Li [Fri, 28 Jun 2019 11:59:51 +0000 (19:59 +0800)]
bcache: set largest seq to ja->seq[bucket_index] in journal_read_bucket()

In journal_read_bucket() when setting ja->seq[bucket_index], there might
be potential case that a later non-maximum overwrites a better sequence
number to ja->seq[bucket_index]. This patch adds a check to make sure
that ja->seq[bucket_index] will be only set a new value if it is bigger
then current value.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: add code comments for journal_read_bucket()
Coly Li [Fri, 28 Jun 2019 11:59:50 +0000 (19:59 +0800)]
bcache: add code comments for journal_read_bucket()

This patch adds more code comments in journal_read_bucket(), this is an
effort to make the code to be more understandable.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: fix potential deadlock in cached_def_free()
Coly Li [Fri, 28 Jun 2019 11:59:49 +0000 (19:59 +0800)]
bcache: fix potential deadlock in cached_def_free()

When enable lockdep and reboot system with a writeback mode bcache
device, the following potential deadlock warning is reported by lockdep
engine.

[  101.536569][  T401] kworker/2:2/401 is trying to acquire lock:
[  101.538575][  T401] 00000000bbf6e6c7 ((wq_completion)bcache_writeback_wq){+.+.}, at: flush_workqueue+0x87/0x4c0
[  101.542054][  T401]
[  101.542054][  T401] but task is already holding lock:
[  101.544587][  T401] 00000000f5f305b3 ((work_completion)(&cl->work)#2){+.+.}, at: process_one_work+0x21e/0x640
[  101.548386][  T401]
[  101.548386][  T401] which lock already depends on the new lock.
[  101.548386][  T401]
[  101.551874][  T401]
[  101.551874][  T401] the existing dependency chain (in reverse order) is:
[  101.555000][  T401]
[  101.555000][  T401] -> #1 ((work_completion)(&cl->work)#2){+.+.}:
[  101.557860][  T401]        process_one_work+0x277/0x640
[  101.559661][  T401]        worker_thread+0x39/0x3f0
[  101.561340][  T401]        kthread+0x125/0x140
[  101.562963][  T401]        ret_from_fork+0x3a/0x50
[  101.564718][  T401]
[  101.564718][  T401] -> #0 ((wq_completion)bcache_writeback_wq){+.+.}:
[  101.567701][  T401]        lock_acquire+0xb4/0x1c0
[  101.569651][  T401]        flush_workqueue+0xae/0x4c0
[  101.571494][  T401]        drain_workqueue+0xa9/0x180
[  101.573234][  T401]        destroy_workqueue+0x17/0x250
[  101.575109][  T401]        cached_dev_free+0x44/0x120 [bcache]
[  101.577304][  T401]        process_one_work+0x2a4/0x640
[  101.579357][  T401]        worker_thread+0x39/0x3f0
[  101.581055][  T401]        kthread+0x125/0x140
[  101.582709][  T401]        ret_from_fork+0x3a/0x50
[  101.584592][  T401]
[  101.584592][  T401] other info that might help us debug this:
[  101.584592][  T401]
[  101.588355][  T401]  Possible unsafe locking scenario:
[  101.588355][  T401]
[  101.590974][  T401]        CPU0                    CPU1
[  101.592889][  T401]        ----                    ----
[  101.594743][  T401]   lock((work_completion)(&cl->work)#2);
[  101.596785][  T401]                                lock((wq_completion)bcache_writeback_wq);
[  101.600072][  T401]                                lock((work_completion)(&cl->work)#2);
[  101.602971][  T401]   lock((wq_completion)bcache_writeback_wq);
[  101.605255][  T401]
[  101.605255][  T401]  *** DEADLOCK ***
[  101.605255][  T401]
[  101.608310][  T401] 2 locks held by kworker/2:2/401:
[  101.610208][  T401]  #0: 00000000cf2c7d17 ((wq_completion)events){+.+.}, at: process_one_work+0x21e/0x640
[  101.613709][  T401]  #1: 00000000f5f305b3 ((work_completion)(&cl->work)#2){+.+.}, at: process_one_work+0x21e/0x640
[  101.617480][  T401]
[  101.617480][  T401] stack backtrace:
[  101.619539][  T401] CPU: 2 PID: 401 Comm: kworker/2:2 Tainted: G        W         5.2.0-rc4-lp151.20-default+ #1
[  101.623225][  T401] Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00 04/13/2018
[  101.627210][  T401] Workqueue: events cached_dev_free [bcache]
[  101.629239][  T401] Call Trace:
[  101.630360][  T401]  dump_stack+0x85/0xcb
[  101.631777][  T401]  print_circular_bug+0x19a/0x1f0
[  101.633485][  T401]  __lock_acquire+0x16cd/0x1850
[  101.635184][  T401]  ? __lock_acquire+0x6a8/0x1850
[  101.636863][  T401]  ? lock_acquire+0xb4/0x1c0
[  101.638421][  T401]  ? find_held_lock+0x34/0xa0
[  101.640015][  T401]  lock_acquire+0xb4/0x1c0
[  101.641513][  T401]  ? flush_workqueue+0x87/0x4c0
[  101.643248][  T401]  flush_workqueue+0xae/0x4c0
[  101.644832][  T401]  ? flush_workqueue+0x87/0x4c0
[  101.646476][  T401]  ? drain_workqueue+0xa9/0x180
[  101.648303][  T401]  drain_workqueue+0xa9/0x180
[  101.649867][  T401]  destroy_workqueue+0x17/0x250
[  101.651503][  T401]  cached_dev_free+0x44/0x120 [bcache]
[  101.653328][  T401]  process_one_work+0x2a4/0x640
[  101.655029][  T401]  worker_thread+0x39/0x3f0
[  101.656693][  T401]  ? process_one_work+0x640/0x640
[  101.658501][  T401]  kthread+0x125/0x140
[  101.660012][  T401]  ? kthread_create_worker_on_cpu+0x70/0x70
[  101.661985][  T401]  ret_from_fork+0x3a/0x50
[  101.691318][  T401] bcache: bcache_device_free() bcache0 stopped

Here is how the above potential deadlock may happen in reboot/shutdown
code path,
1) bcache_reboot() is called firstly in the reboot/shutdown code path,
   then in bcache_reboot(), bcache_device_stop() is called.
2) bcache_device_stop() sets BCACHE_DEV_CLOSING on d->falgs, then call
   closure_queue(&d->cl) to invoke cached_dev_flush(). And in turn
   cached_dev_flush() calls cached_dev_free() via closure_at()
3) In cached_dev_free(), after stopped writebach kthread
   dc->writeback_thread, the kwork dc->writeback_write_wq is stopping by
   destroy_workqueue().
4) Inside destroy_workqueue(), drain_workqueue() is called. Inside
   drain_workqueue(), flush_workqueue() is called. Then wq->lockdep_map
   is acquired by lock_map_acquire() in flush_workqueue(). After the
   lock acquired the rest part of flush_workqueue() just wait for the
   workqueue to complete.
5) Now we look back at writeback thread routine bch_writeback_thread(),
   in the main while-loop, write_dirty() is called via continue_at() in
   read_dirty_submit(), which is called via continue_at() in while-loop
   level called function read_dirty(). Inside write_dirty() it may be
   re-called on workqueeu dc->writeback_write_wq via continue_at().
   It means when the writeback kthread is stopped in cached_dev_free()
   there might be still one kworker queued on dc->writeback_write_wq
   to execute write_dirty() again.
6) Now this kworker is scheduled on dc->writeback_write_wq to run by
   process_one_work() (which is called by worker_thread()). Before
   calling the kwork routine, wq->lockdep_map is acquired.
7) But wq->lockdep_map is acquired already in step 4), so a A-A lock
   (lockdep terminology) scenario happens.

Indeed on multiple cores syatem, the above deadlock is very rare to
happen, just as the code comments in process_one_work() says,
2263     * AFAICT there is no possible deadlock scenario between the
2264     * flush_work() and complete() primitives (except for
   single-threaded
2265     * workqueues), so hiding them isn't a problem.

But it is still good to fix such lockdep warning, even no one running
bcache on single core system.

The fix is simple. This patch solves the above potential deadlock by,
- Do not destroy workqueue dc->writeback_write_wq in cached_dev_free().
- Flush and destroy dc->writeback_write_wq in writebach kthread routine
  bch_writeback_thread(), where after quit the thread main while-loop
  and before cached_dev_put() is called.

By this fix, dc->writeback_write_wq will be stopped and destroy before
the writeback kthread stopped, so the chance for a A-A locking on
wq->lockdep_map is disappeared, such A-A deadlock won't happen
any more.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: acquire bch_register_lock later in cached_dev_free()
Coly Li [Fri, 28 Jun 2019 11:59:48 +0000 (19:59 +0800)]
bcache: acquire bch_register_lock later in cached_dev_free()

When enable lockdep engine, a lockdep warning can be observed when
reboot or shutdown system,

[ 3142.764557][    T1] bcache: bcache_reboot() Stopping all devices:
[ 3142.776265][ T2649]
[ 3142.777159][ T2649] ======================================================
[ 3142.780039][ T2649] WARNING: possible circular locking dependency detected
[ 3142.782869][ T2649] 5.2.0-rc4-lp151.20-default+ #1 Tainted: G        W
[ 3142.785684][ T2649] ------------------------------------------------------
[ 3142.788479][ T2649] kworker/3:67/2649 is trying to acquire lock:
[ 3142.790738][ T2649] 00000000aaf02291 ((wq_completion)bcache_writeback_wq){+.+.}, at: flush_workqueue+0x87/0x4c0
[ 3142.794678][ T2649]
[ 3142.794678][ T2649] but task is already holding lock:
[ 3142.797402][ T2649] 000000004fcf89c5 (&bch_register_lock){+.+.}, at: cached_dev_free+0x17/0x120 [bcache]
[ 3142.801462][ T2649]
[ 3142.801462][ T2649] which lock already depends on the new lock.
[ 3142.801462][ T2649]
[ 3142.805277][ T2649]
[ 3142.805277][ T2649] the existing dependency chain (in reverse order) is:
[ 3142.808902][ T2649]
[ 3142.808902][ T2649] -> #2 (&bch_register_lock){+.+.}:
[ 3142.812396][ T2649]        __mutex_lock+0x7a/0x9d0
[ 3142.814184][ T2649]        cached_dev_free+0x17/0x120 [bcache]
[ 3142.816415][ T2649]        process_one_work+0x2a4/0x640
[ 3142.818413][ T2649]        worker_thread+0x39/0x3f0
[ 3142.820276][ T2649]        kthread+0x125/0x140
[ 3142.822061][ T2649]        ret_from_fork+0x3a/0x50
[ 3142.823965][ T2649]
[ 3142.823965][ T2649] -> #1 ((work_completion)(&cl->work)#2){+.+.}:
[ 3142.827244][ T2649]        process_one_work+0x277/0x640
[ 3142.829160][ T2649]        worker_thread+0x39/0x3f0
[ 3142.830958][ T2649]        kthread+0x125/0x140
[ 3142.832674][ T2649]        ret_from_fork+0x3a/0x50
[ 3142.834915][ T2649]
[ 3142.834915][ T2649] -> #0 ((wq_completion)bcache_writeback_wq){+.+.}:
[ 3142.838121][ T2649]        lock_acquire+0xb4/0x1c0
[ 3142.840025][ T2649]        flush_workqueue+0xae/0x4c0
[ 3142.842035][ T2649]        drain_workqueue+0xa9/0x180
[ 3142.844042][ T2649]        destroy_workqueue+0x17/0x250
[ 3142.846142][ T2649]        cached_dev_free+0x52/0x120 [bcache]
[ 3142.848530][ T2649]        process_one_work+0x2a4/0x640
[ 3142.850663][ T2649]        worker_thread+0x39/0x3f0
[ 3142.852464][ T2649]        kthread+0x125/0x140
[ 3142.854106][ T2649]        ret_from_fork+0x3a/0x50
[ 3142.855880][ T2649]
[ 3142.855880][ T2649] other info that might help us debug this:
[ 3142.855880][ T2649]
[ 3142.859663][ T2649] Chain exists of:
[ 3142.859663][ T2649]   (wq_completion)bcache_writeback_wq --> (work_completion)(&cl->work)#2 --> &bch_register_lock
[ 3142.859663][ T2649]
[ 3142.865424][ T2649]  Possible unsafe locking scenario:
[ 3142.865424][ T2649]
[ 3142.868022][ T2649]        CPU0                    CPU1
[ 3142.869885][ T2649]        ----                    ----
[ 3142.871751][ T2649]   lock(&bch_register_lock);
[ 3142.873379][ T2649]                                lock((work_completion)(&cl->work)#2);
[ 3142.876399][ T2649]                                lock(&bch_register_lock);
[ 3142.879727][ T2649]   lock((wq_completion)bcache_writeback_wq);
[ 3142.882064][ T2649]
[ 3142.882064][ T2649]  *** DEADLOCK ***
[ 3142.882064][ T2649]
[ 3142.885060][ T2649] 3 locks held by kworker/3:67/2649:
[ 3142.887245][ T2649]  #0: 00000000e774cdd0 ((wq_completion)events){+.+.}, at: process_one_work+0x21e/0x640
[ 3142.890815][ T2649]  #1: 00000000f7df89da ((work_completion)(&cl->work)#2){+.+.}, at: process_one_work+0x21e/0x640
[ 3142.894884][ T2649]  #2: 000000004fcf89c5 (&bch_register_lock){+.+.}, at: cached_dev_free+0x17/0x120 [bcache]
[ 3142.898797][ T2649]
[ 3142.898797][ T2649] stack backtrace:
[ 3142.900961][ T2649] CPU: 3 PID: 2649 Comm: kworker/3:67 Tainted: G        W         5.2.0-rc4-lp151.20-default+ #1
[ 3142.904789][ T2649] Hardware name: VMware, Inc. VMware Virtual Platform/440BX Desktop Reference Platform, BIOS 6.00 04/13/2018
[ 3142.909168][ T2649] Workqueue: events cached_dev_free [bcache]
[ 3142.911422][ T2649] Call Trace:
[ 3142.912656][ T2649]  dump_stack+0x85/0xcb
[ 3142.914181][ T2649]  print_circular_bug+0x19a/0x1f0
[ 3142.916193][ T2649]  __lock_acquire+0x16cd/0x1850
[ 3142.917936][ T2649]  ? __lock_acquire+0x6a8/0x1850
[ 3142.919704][ T2649]  ? lock_acquire+0xb4/0x1c0
[ 3142.921335][ T2649]  ? find_held_lock+0x34/0xa0
[ 3142.923052][ T2649]  lock_acquire+0xb4/0x1c0
[ 3142.924635][ T2649]  ? flush_workqueue+0x87/0x4c0
[ 3142.926375][ T2649]  flush_workqueue+0xae/0x4c0
[ 3142.928047][ T2649]  ? flush_workqueue+0x87/0x4c0
[ 3142.929824][ T2649]  ? drain_workqueue+0xa9/0x180
[ 3142.931686][ T2649]  drain_workqueue+0xa9/0x180
[ 3142.933534][ T2649]  destroy_workqueue+0x17/0x250
[ 3142.935787][ T2649]  cached_dev_free+0x52/0x120 [bcache]
[ 3142.937795][ T2649]  process_one_work+0x2a4/0x640
[ 3142.939803][ T2649]  worker_thread+0x39/0x3f0
[ 3142.941487][ T2649]  ? process_one_work+0x640/0x640
[ 3142.943389][ T2649]  kthread+0x125/0x140
[ 3142.944894][ T2649]  ? kthread_create_worker_on_cpu+0x70/0x70
[ 3142.947744][ T2649]  ret_from_fork+0x3a/0x50
[ 3142.970358][ T2649] bcache: bcache_device_free() bcache0 stopped

Here is how the deadlock happens.
1) bcache_reboot() calls bcache_device_stop(), then inside
   bcache_device_stop() BCACHE_DEV_CLOSING bit is set on d->flags.
   Then closure_queue(&d->cl) is called to invoke cached_dev_flush().
2) In cached_dev_flush(), cached_dev_free() is called by continu_at().
3) In cached_dev_free(), when stopping the writeback kthread of the
   cached device by kthread_stop(), dc->writeback_thread will be waken
   up to quite the kthread while-loop, then cached_dev_put() is called
   in bch_writeback_thread().
4) Calling cached_dev_put() in writeback kthread may drop dc->count to
   0, then dc->detach kworker is scheduled, which is initialized as
   cached_dev_detach_finish().
5) Inside cached_dev_detach_finish(), the last line of code is to call
   closure_put(&dc->disk.cl), which drops the last reference counter of
   closrure dc->disk.cl, then the callback cached_dev_flush() gets
   called.
Now cached_dev_flush() is called for second time in the code path, the
first time is in step 2). And again bch_register_lock will be acquired
again, and a A-A lock (lockdep terminology) is happening.

The root cause of the above A-A lock is in cached_dev_free(), mutex
bch_register_lock is held before stopping writeback kthread and other
kworkers. Fortunately now we have variable 'bcache_is_reboot', which may
prevent device registration or unregistration during reboot/shutdown
time, so it is unncessary to hold bch_register_lock such early now.

This is how this patch fixes the reboot/shutdown time A-A lock issue:
After moving mutex_lock(&bch_register_lock) to a later location where
before atomic_read(&dc->running) in cached_dev_free(), such A-A lock
problem can be solved without any reboot time registration race.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: acquire bch_register_lock later in cached_dev_detach_finish()
Coly Li [Fri, 28 Jun 2019 11:59:47 +0000 (19:59 +0800)]
bcache: acquire bch_register_lock later in cached_dev_detach_finish()

Now there is variable bcache_is_reboot to prevent device register or
unregister during reboot, it is unncessary to still hold mutex lock
bch_register_lock before stopping writeback_rate_update kworker and
writeback kthread. And if the stopping kworker or kthread holding
bch_register_lock inside their routine (we used to have such problem
in writeback thread, thanks to Junhui Wang fixed it), it is very easy
to introduce deadlock during reboot/shutdown procedure.

Therefore in this patch, the location to acquire bch_register_lock is
moved to the location before calling calc_cached_dev_sectors(). Which
is later then original location in cached_dev_detach_finish().

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: avoid a deadlock in bcache_reboot()
Coly Li [Fri, 28 Jun 2019 11:59:46 +0000 (19:59 +0800)]
bcache: avoid a deadlock in bcache_reboot()

It is quite frequently to observe deadlock in bcache_reboot() happens
and hang the system reboot process. The reason is, in bcache_reboot()
when calling bch_cache_set_stop() and bcache_device_stop() the mutex
bch_register_lock is held. But in the process to stop cache set and
bcache device, bch_register_lock will be acquired again. If this mutex
is held here, deadlock will happen inside the stopping process. The
aftermath of the deadlock is, whole system reboot gets hung.

The fix is to avoid holding bch_register_lock for the following loops
in bcache_reboot(),
       list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
                bch_cache_set_stop(c);

        list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
                bcache_device_stop(&dc->disk);

A module range variable 'bcache_is_reboot' is added, it sets to true
in bcache_reboot(). In register_bcache(), if bcache_is_reboot is checked
to be true, reject the registration by returning -EBUSY immediately.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: stop writeback kthread and kworker when bch_cached_dev_run() failed
Coly Li [Fri, 28 Jun 2019 11:59:45 +0000 (19:59 +0800)]
bcache: stop writeback kthread and kworker when bch_cached_dev_run() failed

In bch_cached_dev_attach() after bch_cached_dev_writeback_start()
called, the wrireback kthread and writeback rate update kworker of the
cached device are created, if the following bch_cached_dev_run()
failed, bch_cached_dev_attach() will return with -ENOMEM without
stopping the writeback related kthread and kworker.

This patch stops writeback kthread and writeback rate update kworker
before returning -ENOMEM if bch_cached_dev_run() returns error.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: destroy dc->writeback_write_wq if failed to create dc->writeback_thread
Coly Li [Fri, 28 Jun 2019 11:59:44 +0000 (19:59 +0800)]
bcache: destroy dc->writeback_write_wq if failed to create dc->writeback_thread

Commit 9baf30972b55 ("bcache: fix for gc and write-back race") added a
new work queue dc->writeback_write_wq, but forgot to destroy it in the
error condition when creating dc->writeback_thread failed.

This patch destroys dc->writeback_write_wq if kthread_create() returns
error pointer to dc->writeback_thread, then a memory leak is avoided.

Fixes: 9baf30972b55 ("bcache: fix for gc and write-back race")
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: fix mistaken sysfs entry for io_error counter
Coly Li [Fri, 28 Jun 2019 11:59:43 +0000 (19:59 +0800)]
bcache: fix mistaken sysfs entry for io_error counter

In bch_cached_dev_files[] from driver/md/bcache/sysfs.c, sysfs_errors is
incorrectly inserted in. The correct entry should be sysfs_io_errors.

This patch fixes the problem and now I/O errors of cached device can be
read from /sys/block/bcache<N>/bcache/io_errors.

Fixes: c7b7bd07404c5 ("bcache: add io_disable to struct cached_dev")
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: add pendings_cleanup to stop pending bcache device
Coly Li [Fri, 28 Jun 2019 11:59:42 +0000 (19:59 +0800)]
bcache: add pendings_cleanup to stop pending bcache device

If a bcache device is in dirty state and its cache set is not
registered, this bcache device will not appear in /dev/bcache<N>,
and there is no way to stop it or remove the bcache kernel module.

This is an as-designed behavior, but sometimes people has to reboot
whole system to release or stop the pending backing device.

This sysfs interface may remove such pending bcache devices when
write anything into the sysfs file manually.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: make bset_search_tree() be more understandable
Coly Li [Fri, 28 Jun 2019 11:59:41 +0000 (19:59 +0800)]
bcache: make bset_search_tree() be more understandable

The purpose of following code in bset_search_tree() is to avoid a branch
instruction,
 994         if (likely(f->exponent != 127))
 995                 n = j * 2 + (((unsigned int)
 996                               (f->mantissa -
 997                                bfloat_mantissa(search, f))) >> 31);
 998         else
 999                 n = (bkey_cmp(tree_to_bkey(t, j), search) > 0)
1000                         ? j * 2
1001                         : j * 2 + 1;

This piece of code is not very clear to understand, even when I tried to
add code comment for it, I made mistake. This patch removes the implict
bit operation and uses explicit branch to calculate next location in
binary tree search.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: remove "XXX:" comment line from run_cache_set()
Coly Li [Fri, 28 Jun 2019 11:59:40 +0000 (19:59 +0800)]
bcache: remove "XXX:" comment line from run_cache_set()

In previous bcache patches for Linux v5.2, the failure code path of
run_cache_set() is tested and fixed. So now the following comment
line can be removed from run_cache_set(),
/* XXX: test this, it's broken */

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: improve error message in bch_cached_dev_run()
Coly Li [Fri, 28 Jun 2019 11:59:39 +0000 (19:59 +0800)]
bcache: improve error message in bch_cached_dev_run()

This patch adds more error message in bch_cached_dev_run() to indicate
the exact reason why an error value is returned. Please notice when
printing out the "is running already" message, pr_info() is used here,
because in this case also -EBUSY is returned, the bcache device can
continue to attach to the cache devince and run, so it won't be an
error level message in kernel message.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: add more error message in bch_cached_dev_attach()
Coly Li [Fri, 28 Jun 2019 11:59:38 +0000 (19:59 +0800)]
bcache: add more error message in bch_cached_dev_attach()

This patch adds more error message for attaching cached device, this is
helpful to debug code failure during bache device start up.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: more detailed error message to bcache_device_link()
Coly Li [Fri, 28 Jun 2019 11:59:37 +0000 (19:59 +0800)]
bcache: more detailed error message to bcache_device_link()

This patch adds more accurate error message for specific
ssyfs_create_link() call, to help debugging failure during
bcache device start tup.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: check CACHE_SET_IO_DISABLE bit in bch_journal()
Coly Li [Fri, 28 Jun 2019 11:59:36 +0000 (19:59 +0800)]
bcache: check CACHE_SET_IO_DISABLE bit in bch_journal()

When too many I/O errors happen on cache set and CACHE_SET_IO_DISABLE
bit is set, bch_journal() may continue to work because the journaling
bkey might be still in write set yet. The caller of bch_journal() may
believe the journal still work but the truth is in-memory journal write
set won't be written into cache device any more. This behavior may
introduce potential inconsistent metadata status.

This patch checks CACHE_SET_IO_DISABLE bit at the head of bch_journal(),
if the bit is set, bch_journal() returns NULL immediately to notice
caller to know journal does not work.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: check CACHE_SET_IO_DISABLE in allocator code
Coly Li [Fri, 28 Jun 2019 11:59:35 +0000 (19:59 +0800)]
bcache: check CACHE_SET_IO_DISABLE in allocator code

If CACHE_SET_IO_DISABLE of a cache set flag is set by too many I/O
errors, currently allocator routines can still continue allocate
space which may introduce inconsistent metadata state.

This patch checkes CACHE_SET_IO_DISABLE bit in following allocator
routines,
- bch_bucket_alloc()
- __bch_bucket_alloc_set()
Once CACHE_SET_IO_DISABLE is set on cache set, the allocator routines
may reject allocation request earlier to avoid potential inconsistent
metadata.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: remove unncessary code in bch_btree_keys_init()
Coly Li [Fri, 28 Jun 2019 11:59:34 +0000 (19:59 +0800)]
bcache: remove unncessary code in bch_btree_keys_init()

Function bch_btree_keys_init() initializes b->set[].size and
b->set[].data to zero. As the code comments indicates, these code indeed
is unncessary, because both struct btree_keys and struct bset_tree are
nested embedded into struct btree, when struct btree is filled with 0
bits by kzalloc() in mca_bucket_alloc(), b->set[].size and
b->set[].data are initialized to 0 (a.k.a NULL) already.

This patch removes the redundant code, and add comments in
bch_btree_keys_init() and mca_bucket_alloc() to explain why it's safe.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: add return value check to bch_cached_dev_run()
Coly Li [Fri, 28 Jun 2019 11:59:33 +0000 (19:59 +0800)]
bcache: add return value check to bch_cached_dev_run()

This patch adds return value check to bch_cached_dev_run(), now if there
is error happens inside bch_cached_dev_run(), it can be catched.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: use sysfs_match_string() instead of __sysfs_match_string()
Alexandru Ardelean [Fri, 28 Jun 2019 11:59:32 +0000 (19:59 +0800)]
bcache: use sysfs_match_string() instead of __sysfs_match_string()

The arrays (of strings) that are passed to __sysfs_match_string() are
static, so use sysfs_match_string() which does an implicit ARRAY_SIZE()
over these arrays.

Functionally, this doesn't change anything.
The change is more cosmetic.

It only shrinks the static arrays by 1 byte each.

Signed-off-by: Alexandru Ardelean <alexandru.ardelean@analog.com>
Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: remove unnecessary prefetch() in bset_search_tree()
Coly Li [Fri, 28 Jun 2019 11:59:31 +0000 (19:59 +0800)]
bcache: remove unnecessary prefetch() in bset_search_tree()

In function bset_search_tree(), when p >= t->size, t->tree[0] will be
prefetched by the following code piece,
 974                 unsigned int p = n << 4;
 975
 976                 p &= ((int) (p - t->size)) >> 31;
 977
 978                 prefetch(&t->tree[p]);

The purpose of the above code is to avoid a branch instruction, but
when p >= t->size, prefetch(&t->tree[0]) has no positive performance
contribution at all. This patch avoids the unncessary prefetch by only
calling prefetch() when p < t->size.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: add io error counting in write_bdev_super_endio()
Coly Li [Fri, 28 Jun 2019 11:59:30 +0000 (19:59 +0800)]
bcache: add io error counting in write_bdev_super_endio()

When backing device super block is written by bch_write_bdev_super(),
the bio complete callback write_bdev_super_endio() simply ignores I/O
status. Indeed such write request also contribute to backing device
health status if the request failed.

This patch checkes bio->bi_status in write_bdev_super_endio(), if there
is error, bch_count_backing_io_errors() will be called to count an I/O
error to dc->io_errors.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: ignore read-ahead request failure on backing device
Coly Li [Fri, 28 Jun 2019 11:59:29 +0000 (19:59 +0800)]
bcache: ignore read-ahead request failure on backing device

When md raid device (e.g. raid456) is used as backing device, read-ahead
requests on a degrading and recovering md raid device might be failured
immediately by md raid code, but indeed this md raid array can still be
read or write for normal I/O requests. Therefore such failed read-ahead
request are not real hardware failure. Further more, after degrading and
recovering accomplished, read-ahead requests will be handled by md raid
array again.

For such condition, I/O failures of read-ahead requests don't indicate
real health status (because normal I/O still be served), they should not
be counted into I/O error counter dc->io_errors.

Since there is no simple way to detect whether the backing divice is a
md raid device, this patch simply ignores I/O failures for read-ahead
bios on backing device, to avoid bogus backing device failure on a
degrading md raid array.

Suggested-and-tested-by: Thorsten Knabe <linux@thorsten-knabe.de>
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: avoid flushing btree node in cache_set_flush() if io disabled
Coly Li [Fri, 28 Jun 2019 11:59:28 +0000 (19:59 +0800)]
bcache: avoid flushing btree node in cache_set_flush() if io disabled

When cache_set_flush() is called for too many I/O errors detected on
cache device and the cache set is retiring, inside the function it
doesn't make sense to flushing cached btree nodes from c->btree_cache
because CACHE_SET_IO_DISABLE is set on c->flags already and all I/Os
onto cache device will be rejected.

This patch checks in cache_set_flush() that whether CACHE_SET_IO_DISABLE
is set. If yes, then avoids to flush the cached btree nodes to reduce
more time and make cache set retiring more faster.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoRevert "bcache: set CACHE_SET_IO_DISABLE in bch_cached_dev_error()"
Coly Li [Fri, 28 Jun 2019 11:59:27 +0000 (19:59 +0800)]
Revert "bcache: set CACHE_SET_IO_DISABLE in bch_cached_dev_error()"

This reverts commit 6147305c73e4511ca1a975b766b97a779d442567.

Although this patch helps the failed bcache device to stop faster when
too many I/O errors detected on corresponding cached device, setting
CACHE_SET_IO_DISABLE bit to cache set c->flags was not a good idea. This
operation will disable all I/Os on cache set, which means other attached
bcache devices won't work neither.

Without this patch, the failed bcache device can also be stopped
eventually if internal I/O accomplished (e.g. writeback). Therefore here
I revert it.

Fixes: 6147305c73e4 ("bcache: set CACHE_SET_IO_DISABLE in bch_cached_dev_error()")
Reported-by: Yong Li <mr.liyong@qq.com>
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: fix return value error in bch_journal_read()
Coly Li [Fri, 28 Jun 2019 11:59:26 +0000 (19:59 +0800)]
bcache: fix return value error in bch_journal_read()

When everything is OK in bch_journal_read(), finally the return value
is returned by,
return ret;
which assumes ret will be 0 here. This assumption is wrong when all
journal buckets as are full and filled with valid journal entries. In
such cache the last location referencess read_bucket() sets 'ret' to
1, which means new jset added into jset list. The jset list is list
'journal' in caller run_cache_set().

Return 1 to run_cache_set() means something wrong and the cache set
won't start, but indeed everything is OK.

This patch changes the line at end of bch_journal_read() to directly
return 0 since everything if verything is good. Then a bogus error
is fixed.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: check c->gc_thread by IS_ERR_OR_NULL in cache_set_flush()
Coly Li [Fri, 28 Jun 2019 11:59:25 +0000 (19:59 +0800)]
bcache: check c->gc_thread by IS_ERR_OR_NULL in cache_set_flush()

When system memory is in heavy pressure, bch_gc_thread_start() from
run_cache_set() may fail due to out of memory. In such condition,
c->gc_thread is assigned to -ENOMEM, not NULL pointer. Then in following
failure code path bch_cache_set_error(), when cache_set_flush() gets
called, the code piece to stop c->gc_thread is broken,
         if (!IS_ERR_OR_NULL(c->gc_thread))
                 kthread_stop(c->gc_thread);

And KASAN catches such NULL pointer deference problem, with the warning
information:

[  561.207881] ==================================================================
[  561.207900] BUG: KASAN: null-ptr-deref in kthread_stop+0x3b/0x440
[  561.207904] Write of size 4 at addr 000000000000001c by task kworker/15:1/313

[  561.207913] CPU: 15 PID: 313 Comm: kworker/15:1 Tainted: G        W         5.0.0-vanilla+ #3
[  561.207916] Hardware name: Lenovo ThinkSystem SR650 -[7X05CTO1WW]-/-[7X05CTO1WW]-, BIOS -[IVE136T-2.10]- 03/22/2019
[  561.207935] Workqueue: events cache_set_flush [bcache]
[  561.207940] Call Trace:
[  561.207948]  dump_stack+0x9a/0xeb
[  561.207955]  ? kthread_stop+0x3b/0x440
[  561.207960]  ? kthread_stop+0x3b/0x440
[  561.207965]  kasan_report+0x176/0x192
[  561.207973]  ? kthread_stop+0x3b/0x440
[  561.207981]  kthread_stop+0x3b/0x440
[  561.207995]  cache_set_flush+0xd4/0x6d0 [bcache]
[  561.208008]  process_one_work+0x856/0x1620
[  561.208015]  ? find_held_lock+0x39/0x1d0
[  561.208028]  ? drain_workqueue+0x380/0x380
[  561.208048]  worker_thread+0x87/0xb80
[  561.208058]  ? __kthread_parkme+0xb6/0x180
[  561.208067]  ? process_one_work+0x1620/0x1620
[  561.208072]  kthread+0x326/0x3e0
[  561.208079]  ? kthread_create_worker_on_cpu+0xc0/0xc0
[  561.208090]  ret_from_fork+0x3a/0x50
[  561.208110] ==================================================================
[  561.208113] Disabling lock debugging due to kernel taint
[  561.208115] irq event stamp: 11800231
[  561.208126] hardirqs last  enabled at (11800231): [<ffffffff83008538>] do_syscall_64+0x18/0x410
[  561.208127] BUG: unable to handle kernel NULL pointer dereference at 000000000000001c
[  561.208129] #PF error: [WRITE]
[  561.312253] hardirqs last disabled at (11800230): [<ffffffff830052ff>] trace_hardirqs_off_thunk+0x1a/0x1c
[  561.312259] softirqs last  enabled at (11799832): [<ffffffff850005c7>] __do_softirq+0x5c7/0x8c3
[  561.405975] PGD 0 P4D 0
[  561.442494] softirqs last disabled at (11799821): [<ffffffff831add2c>] irq_exit+0x1ac/0x1e0
[  561.791359] Oops: 0002 [#1] SMP KASAN NOPTI
[  561.791362] CPU: 15 PID: 313 Comm: kworker/15:1 Tainted: G    B   W         5.0.0-vanilla+ #3
[  561.791363] Hardware name: Lenovo ThinkSystem SR650 -[7X05CTO1WW]-/-[7X05CTO1WW]-, BIOS -[IVE136T-2.10]- 03/22/2019
[  561.791371] Workqueue: events cache_set_flush [bcache]
[  561.791374] RIP: 0010:kthread_stop+0x3b/0x440
[  561.791376] Code: 00 00 65 8b 05 26 d5 e0 7c 89 c0 48 0f a3 05 ec aa df 02 0f 82 dc 02 00 00 4c 8d 63 20 be 04 00 00 00 4c 89 e7 e8 65 c5 53 00 <f0> ff 43 20 48 8d 7b 24 48 b8 00 00 00 00 00 fc ff df 48 89 fa 48
[  561.791377] RSP: 0018:ffff88872fc8fd10 EFLAGS: 00010286
[  561.838895] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  561.838916] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  561.838934] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  561.838948] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  561.838966] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  561.838979] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  561.838996] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  563.067028] RAX: 0000000000000000 RBX: fffffffffffffffc RCX: ffffffff832dd314
[  563.067030] RDX: 0000000000000000 RSI: 0000000000000004 RDI: 0000000000000297
[  563.067032] RBP: ffff88872fc8fe88 R08: fffffbfff0b8213d R09: fffffbfff0b8213d
[  563.067034] R10: 0000000000000001 R11: fffffbfff0b8213c R12: 000000000000001c
[  563.408618] R13: ffff88dc61cc0f68 R14: ffff888102b94900 R15: ffff88dc61cc0f68
[  563.408620] FS:  0000000000000000(0000) GS:ffff888f7dc00000(0000) knlGS:0000000000000000
[  563.408622] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  563.408623] CR2: 000000000000001c CR3: 0000000f48a1a004 CR4: 00000000007606e0
[  563.408625] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[  563.408627] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[  563.904795] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  563.915796] PKRU: 55555554
[  563.915797] Call Trace:
[  563.915807]  cache_set_flush+0xd4/0x6d0 [bcache]
[  563.915812]  process_one_work+0x856/0x1620
[  564.001226] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  564.033563]  ? find_held_lock+0x39/0x1d0
[  564.033567]  ? drain_workqueue+0x380/0x380
[  564.033574]  worker_thread+0x87/0xb80
[  564.062823] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  564.118042]  ? __kthread_parkme+0xb6/0x180
[  564.118046]  ? process_one_work+0x1620/0x1620
[  564.118048]  kthread+0x326/0x3e0
[  564.118050]  ? kthread_create_worker_on_cpu+0xc0/0xc0
[  564.167066] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  564.252441]  ret_from_fork+0x3a/0x50
[  564.252447] Modules linked in: msr rpcrdma sunrpc rdma_ucm ib_iser ib_umad rdma_cm ib_ipoib i40iw configfs iw_cm ib_cm libiscsi scsi_transport_iscsi mlx4_ib ib_uverbs mlx4_en ib_core nls_iso8859_1 nls_cp437 vfat fat intel_rapl skx_edac x86_pkg_temp_thermal coretemp iTCO_wdt iTCO_vendor_support crct10dif_pclmul crc32_pclmul crc32c_intel ghash_clmulni_intel ses raid0 aesni_intel cdc_ether enclosure usbnet ipmi_ssif joydev aes_x86_64 i40e scsi_transport_sas mii bcache md_mod crypto_simd mei_me ioatdma crc64 ptp cryptd pcspkr i2c_i801 mlx4_core glue_helper pps_core mei lpc_ich dca wmi ipmi_si ipmi_devintf nd_pmem dax_pmem nd_btt ipmi_msghandler device_dax pcc_cpufreq button hid_generic usbhid mgag200 i2c_algo_bit drm_kms_helper syscopyarea sysfillrect xhci_pci sysimgblt fb_sys_fops xhci_hcd ttm megaraid_sas drm usbcore nfit libnvdimm sg dm_multipath dm_mod scsi_dh_rdac scsi_dh_emc scsi_dh_alua efivarfs
[  564.299390] bcache: bch_count_io_errors() nvme0n1: IO error on writing btree.
[  564.348360] CR2: 000000000000001c
[  564.348362] ---[ end trace b7f0e5cc7b2103b0 ]---

Therefore, it is not enough to only check whether c->gc_thread is NULL,
we should use IS_ERR_OR_NULL() to check both NULL pointer and error
value.

This patch changes the above buggy code piece in this way,
         if (!IS_ERR_OR_NULL(c->gc_thread))
                 kthread_stop(c->gc_thread);

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agobcache: don't set max writeback rate if gc is running
Coly Li [Fri, 28 Jun 2019 11:59:24 +0000 (19:59 +0800)]
bcache: don't set max writeback rate if gc is running

When gc is running, user space I/O processes may wait inside
bcache code, so no new I/O coming. Indeed this is not a real idle
time, maximum writeback rate should not be set in such situation.
Otherwise a faster writeback thread may compete locks with gc thread
and makes garbage collection slower, which results a longer I/O
freeze period.

This patch checks c->gc_mark_valid in set_at_max_writeback_rate(). If
c->gc_mark_valid is 0 (gc running), set_at_max_writeback_rate() returns
false, then update_writeback_rate() will not set writeback rate to
maximum value even c->idle_counter reaches an idle threshold.

Now writeback thread won't interfere gc thread performance.

Signed-off-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock: Remove unused code
Damien Le Moal [Thu, 27 Jun 2019 02:59:41 +0000 (11:59 +0900)]
block: Remove unused code

bio_flush_dcache_pages() is unused. Remove it.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Damien Le Moal <damien.lemoal@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, bfq: Init saved_wr_start_at_switch_to_srt in unlikely case
Douglas Anderson [Wed, 26 Jun 2019 19:59:19 +0000 (12:59 -0700)]
block, bfq: Init saved_wr_start_at_switch_to_srt in unlikely case

Some debug code suggested by Paolo was tripping when I did reboot
stress tests.  Specifically in bfq_bfqq_resume_state()
"bic->saved_wr_start_at_switch_to_srt" was later than the current
value of "jiffies".  A bit of debugging showed that
"bic->saved_wr_start_at_switch_to_srt" was actually 0 and a bit more
debugging showed that was because we had run through the "unlikely"
case in the bfq_bfqq_save_state() function.

Let's init "saved_wr_start_at_switch_to_srt" in the unlikely case to
something sane.

NOTE: this fixes no known real-world errors.

Reviewed-by: Paolo Valente <paolo.valente@linaro.org>
Reviewed-by: Guenter Roeck <groeck@chromium.org>
Signed-off-by: Douglas Anderson <dianders@chromium.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoMerge branch 'md-next' of https://github.com/liu-song-6/linux into for-5.3/block
Jens Axboe [Wed, 26 Jun 2019 19:49:01 +0000 (13:49 -0600)]
Merge branch 'md-next' of https://github.com/liu-song-6/linux into for-5.3/block

Pull single MD warning fix from Song.

* 'md-next' of https://github.com/liu-song-6/linux:
  md/raid1: Fix a warning message in remove_wb()

5 years agomd/raid1: Fix a warning message in remove_wb()
Dan Carpenter [Wed, 26 Jun 2019 09:42:51 +0000 (12:42 +0300)]
md/raid1: Fix a warning message in remove_wb()

The WARN_ON() macro doesn't take an error message, it just takes a
condition.  I've changed this to use WARN(1, "...") instead.

Fixes: 3e148a320979 ("md/raid1: fix potential data inconsistency issue with write behind device")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agoblock, bfq: re-schedule empty queues if they deserve I/O plugging
Paolo Valente [Tue, 25 Jun 2019 05:12:49 +0000 (07:12 +0200)]
block, bfq: re-schedule empty queues if they deserve I/O plugging

Consider, on one side, a bfq_queue Q that remains empty while in
service, and, on the other side, the pending I/O of bfq_queues that,
according to their timestamps, have to be served after Q.  If an
uncontrolled amount of I/O from the latter bfq_queues were dispatched
while Q is waiting for its new I/O to arrive, then Q's bandwidth
guarantees would be violated. To prevent this, I/O dispatch is plugged
until Q receives new I/O (except for a properly controlled amount of
injected I/O). Unfortunately, preemption breaks I/O-dispatch plugging,
for the following reason.

Preemption is performed in two steps. First, Q is expired and
re-scheduled. Second, the new bfq_queue to serve is chosen. The first
step is needed by the second, as the second can be performed only
after Q's timestamps have been properly updated (done in the
expiration step), and Q has been re-queued for service. This
dependency is a consequence of the way how BFQ's scheduling algorithm
is currently implemented.

But Q is not re-scheduled at all in the first step, because Q is
empty. As a consequence, an uncontrolled amount of I/O may be
dispatched until Q becomes non empty again. This breaks Q's service
guarantees.

This commit addresses this issue by re-scheduling Q even if it is
empty. This in turn breaks the assumption that all scheduled queues
are non empty. Then a few extra checks are now needed.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, bfq: preempt lower-weight or lower-priority queues
Paolo Valente [Tue, 25 Jun 2019 05:12:48 +0000 (07:12 +0200)]
block, bfq: preempt lower-weight or lower-priority queues

BFQ enqueues the I/O coming from each process into a separate
bfq_queue, and serves bfq_queues one at a time. Each bfq_queue may be
served for at most timeout_sync milliseconds (default: 125 ms). This
service scheme is prone to the following inaccuracy.

While a bfq_queue Q1 is in service, some empty bfq_queue Q2 may
receive I/O, and, according to BFQ's scheduling policy, may become the
right bfq_queue to serve, in place of the currently in-service
bfq_queue. In this respect, postponing the service of Q2 to after the
service of Q1 finishes may delay the completion of Q2's I/O, compared
with an ideal service in which all non-empty bfq_queues are served in
parallel, and every non-empty bfq_queue is served at a rate
proportional to the bfq_queue's weight. This additional delay is equal
at most to the time Q1 may unjustly remain in service before switching
to Q2.

If Q1 and Q2 have the same weight, then this time is most likely
negligible compared with the completion time to be guaranteed to Q2's
I/O. In addition, first, one of the reasons why BFQ may want to serve
Q1 for a while is that this boosts throughput and, second, serving Q1
longer reduces BFQ's overhead. As a conclusion, it is usually better
not to preempt Q1 if both Q1 and Q2 have the same weight.

In contrast, as Q2's weight or priority becomes higher and higher
compared with that of Q1, the above delay becomes larger and larger,
compared with the I/O completion times that have to be guaranteed to
Q2 according to Q2's weight. So reducing this delay may be more
important than avoiding the costs of preempting Q1.

Accordingly, this commit preempts Q1 if Q2 has a higher weight or a
higher priority than Q1. Preemption causes Q1 to be re-scheduled, and
triggers a new choice of the next bfq_queue to serve. If Q2 really is
the next bfq_queue to serve, then Q2 will be set in service
immediately.

This change reduces the component of the I/O latency caused by the
above delay by about 80%. For example, on an (old) PLEXTOR PX-256M5
SSD, the maximum latency reported by fio drops from 15.1 to 3.2 ms for
a process doing sporadic random reads while another process is doing
continuous sequential reads.

Signed-off-by: Nicola Bottura <bottura.nicola95@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, bfq: detect wakers and unconditionally inject their I/O
Paolo Valente [Tue, 25 Jun 2019 05:12:47 +0000 (07:12 +0200)]
block, bfq: detect wakers and unconditionally inject their I/O

A bfq_queue Q may happen to be synchronized with another
bfq_queue Q2, i.e., the I/O of Q2 may need to be completed for Q to
receive new I/O. We call Q2 "waker queue".

If I/O plugging is being performed for Q, and Q is not receiving any
more I/O because of the above synchronization, then, thanks to BFQ's
injection mechanism, the waker queue is likely to get served before
the I/O-plugging timeout fires.

Unfortunately, this fact may not be sufficient to guarantee a high
throughput during the I/O plugging, because the inject limit for Q may
be too low to guarantee a lot of injected I/O. In addition, the
duration of the plugging, i.e., the time before Q finally receives new
I/O, may not be minimized, because the waker queue may happen to be
served only after other queues.

To address these issues, this commit introduces the explicit detection
of the waker queue, and the unconditional injection of a pending I/O
request of the waker queue on each invocation of
bfq_dispatch_request().

One may be concerned that this systematic injection of I/O from the
waker queue delays the service of Q's I/O. Fortunately, it doesn't. On
the contrary, next Q's I/O is brought forward dramatically, for it is
not blocked for milliseconds.

Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, bfq: bring forward seek&think time update
Paolo Valente [Tue, 25 Jun 2019 05:12:46 +0000 (07:12 +0200)]
block, bfq: bring forward seek&think time update

Until the base value for request service times gets finally computed
for a bfq_queue, the inject limit for that queue does depend on the
think-time state (short|long) of the queue. A timely update of the
think time then guarantees a quicker activation or deactivation of the
injection. Fortunately, the think time of a bfq_queue is updated in
the same code path as the inject limit; but after the inject limit.

This commits moves the update of the think time before the update of
the inject limit. For coherence, it moves the update of the seek time
too.

Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, bfq: update base request service times when possible
Paolo Valente [Tue, 25 Jun 2019 05:12:45 +0000 (07:12 +0200)]
block, bfq: update base request service times when possible

I/O injection gets reduced if it increases the request service times
of the victim queue beyond a certain threshold.  The threshold, in its
turn, is computed as a function of the base service time enjoyed by
the queue when it undergoes no injection.

As a consequence, for injection to work properly, the above base value
has to be accurate. In this respect, such a value may vary over
time. For example, it varies if the size or the spatial locality of
the I/O requests in the queue change. It is then important to update
this value whenever possible. This commit performs this update.

Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, bfq: fix rq_in_driver check in bfq_update_inject_limit
Paolo Valente [Tue, 25 Jun 2019 05:12:44 +0000 (07:12 +0200)]
block, bfq: fix rq_in_driver check in bfq_update_inject_limit

One of the cases where the parameters for injection may be updated is
when there are no more in-flight I/O requests. The number of in-flight
requests is stored in the field bfqd->rq_in_driver of the descriptor
bfqd of the device. So, the controlled condition is
bfqd->rq_in_driver == 0.

Unfortunately, this is wrong because, the instruction that checks this
condition is in the code path that handles the completion of a
request, and, in particular, the instruction is executed before
bfqd->rq_in_driver is decremented in such a code path.

This commit fixes this issue by just replacing 0 with 1 in the
comparison.

Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoblock, bfq: reset inject limit when think-time state changes
Paolo Valente [Tue, 25 Jun 2019 05:12:43 +0000 (07:12 +0200)]
block, bfq: reset inject limit when think-time state changes

Until the base value of the request service times gets finally
computed for a bfq_queue, the inject limit does depend on the
think-time state (short|long). The limit must be 0 or 1 if the think
time is deemed, respectively, as short or long. However, such a check
and possible limit update is performed only periodically, once per
second. So, to make the injection mechanism much more reactive, this
commit performs the update also every time the think-time state
changes.

In addition, in the following special case, this commit lets the
inject limit of a bfq_queue bfqq remain equal to 1 even if bfqq's
think time is short: bfqq's I/O is synchronized with that of some
other queue, i.e., bfqq may receive new I/O only after the I/O of the
other queue is completed. Keeping the inject limit to 1 allows the
blocking I/O to be served while bfqq is in service. And this is very
convenient both for bfqq and for the total throughput, as explained
in detail in the comments in bfq_update_has_short_ttime().

Reported-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Tested-by: Srivatsa S. Bhat (VMware) <srivatsa@csail.mit.edu>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoMerge branch 'nvme-5.3' of git://git.infradead.org/nvme into for-5.3/block
Jens Axboe [Mon, 24 Jun 2019 16:10:35 +0000 (10:10 -0600)]
Merge branch 'nvme-5.3' of git://git.infradead.org/nvme into for-5.3/block

Pull NVMe updates from Christoph:

"A large chunk of NVMe updates for 5.3.  Highlights:

 - improved PCIe suspent support (Keith Busch)
 - error injection support for the admin queue (Akinobu Mita)
 - Fibre Channel discovery improvements (James Smart)
 - tracing improvements including nvmetc tracing support (Minwoo Im)
 - misc fixes and cleanups (Anton Eidelman, Minwoo Im, Chaitanya
   Kulkarni)"

* 'nvme-5.3' of git://git.infradead.org/nvme: (26 commits)
  Documentation: nvme: add an example for nvme fault injection
  nvme: enable to inject errors into admin commands
  nvme: prepare for fault injection into admin commands
  nvmet: introduce target-side trace
  nvme-trace: print result and status in hex format
  nvme-trace: support for fabrics commands in host-side
  nvme-trace: move opcode symbol print to nvme.h
  nvme-trace: do not export nvme_trace_disk_name
  nvme-pci: clean up nvme_remove_dead_ctrl a bit
  nvme-pci: properly report state change failure in nvme_reset_work
  nvme-pci: set the errno on ctrl state change error
  nvme-pci: adjust irq max_vector using num_possible_cpus()
  nvme-pci: remove queue_count_ops for write_queues and poll_queues
  nvme-pci: remove unnecessary zero for static var
  nvme-pci: use host managed power state for suspend
  nvme: introduce nvme_is_fabrics to check fabrics cmd
  nvme: export get and set features
  nvme: fix possible io failures when removing multipathed ns
  nvme-fc: add message when creating new association
  lpfc: add sysfs interface to post NVME RSCN
  ...

5 years agoDocumentation: nvme: add an example for nvme fault injection
Akinobu Mita [Sun, 9 Jun 2019 14:17:02 +0000 (23:17 +0900)]
Documentation: nvme: add an example for nvme fault injection

This adds an example of how to inject errors into admin commands.

Suggested-by: Thomas Tai <thomas.tai@oracle.com>
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Minwoo Im <minwoo.im@samsung.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme: enable to inject errors into admin commands
Akinobu Mita [Sun, 9 Jun 2019 14:17:01 +0000 (23:17 +0900)]
nvme: enable to inject errors into admin commands

This enables to inject errors into the commands submitted to the admin
queue.

It is useful to test error handling in the controller initialization.

# echo 100 > /sys/kernel/debug/nvme0/fault_inject/probability
# echo 1 > /sys/kernel/debug/nvme0/fault_inject/times
# echo 10 > /sys/kernel/debug/nvme0/fault_inject/space
# nvme reset /dev/nvme0
# dmesg
...
nvme nvme0: Could not set queue count (16385)
nvme nvme0: IO queues not created

Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme: prepare for fault injection into admin commands
Akinobu Mita [Thu, 20 Jun 2019 06:49:02 +0000 (08:49 +0200)]
nvme: prepare for fault injection into admin commands

Currenlty fault injection support for nvme only enables to inject errors
into the commands submitted to I/O queues.

In preparation for fault injection into the admin commands, this makes
the helper functions independent of struct nvme_ns.

Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvmet: introduce target-side trace
Minwoo Im [Wed, 12 Jun 2019 12:45:33 +0000 (21:45 +0900)]
nvmet: introduce target-side trace

This patch introduces target-side request tracing.  As Christoph
suggested, the trace would not be in a core or module to avoid
disadvantages like cache miss:
  http://lists.infradead.org/pipermail/linux-nvme/2019-June/024721.html

The target-side trace code is entirely based on the Johannes's trace code
from the host side.  It has lots of codes duplicated, but it would be
better than having advantages mentioned above.

It also traces not only fabrics commands, but also nvme normal commands.
Once the codes to be shared gets bigger, then we can make it common as
suggsted.

This also removed the create_sq and create_cq trace parsing functions
because it will be done by the connect fabrics command.

Example:
  echo 1 > /sys/kernel/debug/tracing/event/nvmet/nvmet_req_init/enable
  echo 1 > /sys/kernel/debug/tracing/event/nvmet/nvmet_req_complete/enable
  cat /sys/kernel/debug/tracing/trace

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
[hch: fixed the symbol namespace and a an endianess conversion]
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agolightnvm: fix uninitialized pointer in nvm_remove_tgt()
Geert Uytterhoeven [Fri, 21 Jun 2019 09:12:00 +0000 (11:12 +0200)]
lightnvm: fix uninitialized pointer in nvm_remove_tgt()

With gcc 4.1:

    drivers/lightnvm/core.c: In function â€˜nvm_remove_tgt’:
    drivers/lightnvm/core.c:510: warning: â€˜t’ is used uninitialized in this function

Indeed, if no NVM devices have been registered, t will be an
uninitialized pointer, and may be dereferenced later.  A call to
nvm_remove_tgt() can be triggered from userspace by issuing the
NVM_DEV_REMOVE ioctl on the lightnvm control device.

Fix this by preinitializing t to NULL.

Fixes: 843f2edbdde085b4 ("lightnvm: do not remove instance under global lock")
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>
5 years agolightnvm: pblk: fix freeing of merged pages
Heiner Litz [Fri, 21 Jun 2019 09:11:59 +0000 (11:11 +0200)]
lightnvm: pblk: fix freeing of merged pages

bio_add_pc_page() may merge pages when a bio is padded due to a flush.
Fix iteration over the bio to free the correct pages in case of a merge.

Signed-off-by: Heiner Litz <hlitz@ucsc.edu>
Reviewed-by: Javier González <javier@javigon.com>
Signed-off-by: Matias Bjørling <mb@lightnvm.io>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agonvme-trace: print result and status in hex format
Minwoo Im [Wed, 12 Jun 2019 12:45:32 +0000 (21:45 +0900)]
nvme-trace: print result and status in hex format

The "result" field is in 64bit to be printed out which means it could be
like:
  nvme_complete_rq: nvme0: qid=0, cmdid=0, res=18446612684158962624, etries=0, flags=0x0, status=0

Switch both the result and status field to be printed in hexadecimal
format to be easier to read.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-trace: support for fabrics commands in host-side
Minwoo Im [Wed, 12 Jun 2019 12:45:31 +0000 (21:45 +0900)]
nvme-trace: support for fabrics commands in host-side

This patch introduces fabrics commands tracing feature from host-side.
This patch does not include any changes for the previous host-side
tracing, but just add fabrics commands parsing in cmd=() format.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
[hch: fixed some whitespace damage]
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-trace: move opcode symbol print to nvme.h
Minwoo Im [Wed, 12 Jun 2019 12:45:30 +0000 (21:45 +0900)]
nvme-trace: move opcode symbol print to nvme.h

The following patches are going to provide the target-side trace which
might need these kind of macros.  It would be great if it can be shared
between host and target side both.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-trace: do not export nvme_trace_disk_name
Minwoo Im [Wed, 12 Jun 2019 12:45:29 +0000 (21:45 +0900)]
nvme-trace: do not export nvme_trace_disk_name

nvme_trace_disk_name() is now already being invoked with the function
prototype in trace.h.  We don't need to export this symbol at all.

The following patches are going to provide target-side trace feature
with the exactly same function with this so that this patch removes the
EXPORT_SYMBOL() for this function.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-pci: clean up nvme_remove_dead_ctrl a bit
Chaitanya Kulkarni [Sat, 8 Jun 2019 20:16:32 +0000 (13:16 -0700)]
nvme-pci: clean up nvme_remove_dead_ctrl a bit

Remove the status parameter o nvme_remove_dead_ctrl(), which is only
used for printing it.

We move the print message to the same function where actual error is
occurring.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-pci: properly report state change failure in nvme_reset_work
Minwoo Im [Sat, 8 Jun 2019 18:35:20 +0000 (03:35 +0900)]
nvme-pci: properly report state change failure in nvme_reset_work

If the state change to NVME_CTRL_CONNECTING fails, the dmesg is going to
be like:

  [  293.689160] nvme nvme0: failed to mark controller CONNECTING
  [  293.689160] nvme nvme0: Removing after probe failure status: 0

Even it prints the first line to indicate the situation, the second line
is not proper because the status is 0 which means normally success of
the previous operation.

This patch makes it indicate the proper error value when it fails.
  [   25.932367] nvme nvme0: failed to mark controller CONNECTING
  [   25.932369] nvme nvme0: Removing after probe failure status: -16

This situation is able to be easily reproduced by:
  root@target:~# rmmod nvme && modprobe nvme && rmmod nvme

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-pci: set the errno on ctrl state change error
Chaitanya Kulkarni [Sat, 8 Jun 2019 20:01:02 +0000 (13:01 -0700)]
nvme-pci: set the errno on ctrl state change error

This patch removes the confusing assignment of the variable result at
the time of declaration and sets the value in error cases next to the
places where the actual error is happening.

Here we also set the result value to -ENODEV when we fail at the final
ctrl state transition in nvme_reset_work(). Without this assignment
result will hold 0 from nvme_setup_io_queue() and on failure 0 will be
passed to he nvme_remove_dead_ctrl() from final state transition.

Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-pci: adjust irq max_vector using num_possible_cpus()
Minwoo Im [Sat, 8 Jun 2019 18:02:19 +0000 (03:02 +0900)]
nvme-pci: adjust irq max_vector using num_possible_cpus()

If the "irq_queues" are greater than num_possible_cpus(),
nvme_calc_irq_sets() can have irq set_size for HCTX_TYPE_DEFAULT greater
than it can be afforded.
2039         affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;

It might cause a WARN() from the irq_build_affinity_masks() like [1]:
220         if (nr_present < numvecs)
221                 WARN_ON(nr_present + nr_others < numvecs);

This patch prevents it from the WARN() by adjusting the max_vector value
from the nvme_setup_irqs().

[1] WARN messages when modprobe nvme write_queues=32 poll_queues=0:
root@target:~/nvme# nproc
8
root@target:~/nvme# modprobe nvme write_queues=32 poll_queues=0
[   17.925326] nvme nvme0: pci function 0000:00:04.0
[   17.940601] WARNING: CPU: 3 PID: 1030 at kernel/irq/affinity.c:221 irq_create_affinity_masks+0x222/0x330
[   17.940602] Modules linked in: nvme nvme_core [last unloaded: nvme]
[   17.940605] CPU: 3 PID: 1030 Comm: kworker/u17:4 Tainted: G        W         5.1.0+ #156
[   17.940605] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.1-0-ga5cab58e9a3f-prebuilt.qemu.org 04/01/2014
[   17.940608] Workqueue: nvme-reset-wq nvme_reset_work [nvme]
[   17.940609] RIP: 0010:irq_create_affinity_masks+0x222/0x330
[   17.940611] Code: 4c 8d 4c 24 28 4c 8d 44 24 30 e8 c9 fa ff ff 89 44 24 18 e8 c0 38 fa ff 8b 44 24 18 44 8b 54 24 1c 5a 44 01 d0 41 39 c4 76 02 <0f> 0b 48 89 df 44 01 e5 e8 f1 ce 10 00 48 8b 34 24 44 89 f0 44 01
[   17.940611] RSP: 0018:ffffc90002277c50 EFLAGS: 00010216
[   17.940612] RAX: 0000000000000008 RBX: ffff88807ca48860 RCX: 0000000000000000
[   17.940612] RDX: ffff88807bc03800 RSI: 0000000000000020 RDI: 0000000000000000
[   17.940613] RBP: 0000000000000001 R08: ffffc90002277c78 R09: ffffc90002277c70
[   17.940613] R10: 0000000000000008 R11: 0000000000000001 R12: 0000000000000020
[   17.940614] R13: 0000000000025d08 R14: 0000000000000001 R15: ffff88807bc03800
[   17.940614] FS:  0000000000000000(0000) GS:ffff88807db80000(0000) knlGS:0000000000000000
[   17.940616] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[   17.940617] CR2: 00005635e583f790 CR3: 000000000240a000 CR4: 00000000000006e0
[   17.940617] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[   17.940618] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[   17.940618] Call Trace:
[   17.940622]  __pci_enable_msix_range+0x215/0x540
[   17.940623]  ? kernfs_put+0x117/0x160
[   17.940625]  pci_alloc_irq_vectors_affinity+0x74/0x110
[   17.940626]  nvme_reset_work+0xc30/0x1397 [nvme]
[   17.940628]  ? __switch_to_asm+0x34/0x70
[   17.940628]  ? __switch_to_asm+0x40/0x70
[   17.940629]  ? __switch_to_asm+0x34/0x70
[   17.940630]  ? __switch_to_asm+0x40/0x70
[   17.940630]  ? __switch_to_asm+0x34/0x70
[   17.940631]  ? __switch_to_asm+0x40/0x70
[   17.940632]  ? nvme_irq_check+0x30/0x30 [nvme]
[   17.940633]  process_one_work+0x20b/0x3e0
[   17.940634]  worker_thread+0x1f9/0x3d0
[   17.940635]  ? cancel_delayed_work+0xa0/0xa0
[   17.940636]  kthread+0x117/0x120
[   17.940637]  ? kthread_stop+0xf0/0xf0
[   17.940638]  ret_from_fork+0x3a/0x50
[   17.940639] ---[ end trace aca8a131361cd42a ]---
[   17.942124] nvme nvme0: 7/1/0 default/read/poll queues

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-pci: remove queue_count_ops for write_queues and poll_queues
Minwoo Im [Sat, 8 Jun 2019 18:02:18 +0000 (03:02 +0900)]
nvme-pci: remove queue_count_ops for write_queues and poll_queues

queue_count_set() seems like that it has been provided to limit the
number of queue entries for write/poll queues.  But, the
queue_count_set() has been doing nothing but a parameter check even it
has num_possible_cpus() which is nop.

This patch removes entire queue_count_ops from the write_queues and
poll_queues.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-pci: remove unnecessary zero for static var
Minwoo Im [Sat, 8 Jun 2019 18:02:17 +0000 (03:02 +0900)]
nvme-pci: remove unnecessary zero for static var

poll_queues will be zero even without zero initialization here.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-pci: use host managed power state for suspend
Keith Busch [Thu, 23 May 2019 15:27:35 +0000 (09:27 -0600)]
nvme-pci: use host managed power state for suspend

The nvme pci driver prepares its devices for power loss during suspend
by shutting down the controllers. The power setting is deferred to
pci driver's power management before the platform removes power. The
suspend-to-idle mode, however, does not remove power.

NVMe devices that implement host managed power settings can achieve
lower power and better transition latencies than using generic PCI power
settings. Try to use this feature if the platform is not involved with
the suspend. If successful, restore the previous power state on resume.

Tested-by: Kai-Heng Feng <kai.heng.feng@canonical.com>
Tested-by: Mario Limonciello <mario.limonciello@dell.com>
Reviewed-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
[hch: fixed the compilation for the !CONFIG_PM_SLEEP case]
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme: introduce nvme_is_fabrics to check fabrics cmd
Minwoo Im [Thu, 6 Jun 2019 05:30:14 +0000 (14:30 +0900)]
nvme: introduce nvme_is_fabrics to check fabrics cmd

This patch introduces a nvme_is_fabrics() inline function to check
whether or not the given command structure is for fabrics.

Signed-off-by: Minwoo Im <minwoo.im.dev@gmail.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme: export get and set features
Keith Busch [Sun, 26 May 2019 16:29:01 +0000 (01:29 +0900)]
nvme: export get and set features

Future use intends to make use of both, so export these functions. And
since their implementation is identical except for the opcode, provide a
new function that implement both.

[akinobu.mita@gmail.com>: fix line over 80 characters]
Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Minwoo Im <minwoo.im.dev@gmail.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme: fix possible io failures when removing multipathed ns
Anton Eidelman [Thu, 20 Jun 2019 06:48:10 +0000 (08:48 +0200)]
nvme: fix possible io failures when removing multipathed ns

When a shared namespace is removed, we call blk_cleanup_queue()
when the device can still be accessed as the current path and this can
result in submission to a dying queue. Hence, direct_make_request()
called by our mpath device may fail (propagating the failure to userspace).
Instead, we want to failover this I/O to a different path if one exists.
Thus, before we cleanup the request queue, we make sure that the device is
cleared from the current path nor it can be selected again as such.

Fix this by:
- clear the ns from the head->list and synchronize rcu to make sure there is
  no concurrent path search that restores it as the current path
- clear the mpath current path in order to trigger a subsequent path search
  and sync srcu to wait for any ongoing request submissions
- safely continue to namespace removal and blk_cleanup_queue

Signed-off-by: Anton Eidelman <anton@lightbitslabs.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-fc: add message when creating new association
James Smart [Wed, 29 May 2019 22:25:26 +0000 (15:25 -0700)]
nvme-fc: add message when creating new association

When looking at console messages to troubleshoot, there are one
maybe two messages before creation of the controller is complete.
However, a lot of io takes place to reach that point. It's unclear
when things have started.

Add a message when the controller is attempting to create a new
association. Thus we know what controller, between what host and
remote port, and what NQN is being put into place for any
subsequent success or failure messages.

Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Reviewed-by: Giridhar Malavali <gmalavali@marvell.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agolpfc: add sysfs interface to post NVME RSCN
James Smart [Tue, 14 May 2019 21:58:08 +0000 (14:58 -0700)]
lpfc: add sysfs interface to post NVME RSCN

To support scenarios which aren't bound to nvmetcli add port scenarios,
which is currently where the nvmet_fc transport invokes the discovery
event callbacks, a syfs attribute is added to lpfc which can be written
to cause an RSCN to be generated for the nport.

Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Arun Easi <aeasi@marvell.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agolpfc: add support for translating an RSCN rcv into a discovery rescan
James Smart [Tue, 14 May 2019 21:58:07 +0000 (14:58 -0700)]
lpfc: add support for translating an RSCN rcv into a discovery rescan

This patch updates RSCN receive processing to check for the remote
port being an NVME port, and if so, invoke the nvme_fc callback to
rescan the remote port.  The rescan will generate a discovery udev
event.

Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Arun Easi <aeasi@marvell.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agolpfc: add nvmet discovery_event op support
James Smart [Tue, 14 May 2019 21:58:06 +0000 (14:58 -0700)]
lpfc: add nvmet discovery_event op support

This patch adds support for the nvmet discovery op. When the callback
routine is called, the driver will call the routine to generate an RSCN
to the port on the other end of the link.

Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Arun Easi <aeasi@marvell.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agolpfc: add support to generate RSCN events for nport
James Smart [Tue, 14 May 2019 21:58:05 +0000 (14:58 -0700)]
lpfc: add support to generate RSCN events for nport

This patch adds general RSCN support:

 - The ability to transmit an RSCN to the port on the other end of
   the link (regular port if pt2pt, or fabric controller if fabric).
 - And general recognition of an RSCN ELS when an ELS is received.

Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Arun Easi <aeasi@marvell.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvme-fcloop: add support for nvmet discovery_event op
James Smart [Tue, 14 May 2019 21:58:04 +0000 (14:58 -0700)]
nvme-fcloop: add support for nvmet discovery_event op

Update fcloop to support the discovery_event operation and
invoke a nvme rescan. In a real fc adapter, this would generate an
RSCN, which the host would receive and convert into a nvme rescan
on the remote port specified in the rscn payload.

Signed-off-by: James Smart <jsmart2021@gmail.com>
[kbuild-bot: fcloop_tgt_discovery_evt can be static]
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Arun Easi <aeasi@marvell.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvmet-fc: add transport discovery change event callback support
James Smart [Tue, 14 May 2019 21:58:03 +0000 (14:58 -0700)]
nvmet-fc: add transport discovery change event callback support

This patch adds support for the nvmet discovery_change transport op.
In turn, the transport adds it's own LLDD api callback discovery_event
op to request the LLDD to generate an RSCN for the discovery change.

Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Arun Easi <aeasi@marvell.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agonvmet: add transport discovery change op
James Smart [Tue, 14 May 2019 21:58:02 +0000 (14:58 -0700)]
nvmet: add transport discovery change op

Some transports, such as FC-NVME, support discovery controller change
events without the use of a persistent discovery controller. FC receives
events via RSCN from the FC Fabric Controller or subsystem FC port.

This patch adds a nvmet transport op that is called whenever a
discovery change event occurs in the nvmet layer.

To facilitate the callback without adding another layer to cross into
core.c to reference the transport ops, the port structure snapshots
the transport ops when the port is enabled and clears them when disabled.

Signed-off-by: James Smart <jsmart2021@gmail.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Reviewed-by: Arun Easi <aeasi@marvell.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
5 years agocgroup: export css_next_descendant_pre for bfq
Christoph Hellwig [Fri, 21 Jun 2019 08:22:48 +0000 (10:22 +0200)]
cgroup: export css_next_descendant_pre for bfq

The bfq schedule now uses css_next_descendant_pre directly after
the stats functionality depending on it has been from the core
blk-cgroup code to bfq.  Export the symbol so that bfq can still
be build modular.

Fixes: d6258980daf2 ("bfq-iosched: move bfq_stat_recursive_sum into the only caller")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
5 years agoMerge branch 'md-next' of https://github.com/liu-song-6/linux into for-5.3/block
Jens Axboe [Fri, 21 Jun 2019 05:37:20 +0000 (23:37 -0600)]
Merge branch 'md-next' of https://github.com/liu-song-6/linux into for-5.3/block

Pull MD changes from Song.

* 'md-next' of https://github.com/liu-song-6/linux:
  md: add bitmap_abort label in md_run
  md-bitmap: create and destroy wb_info_pool with the change of bitmap
  md-bitmap: create and destroy wb_info_pool with the change of backlog
  md: introduce mddev_create/destroy_wb_pool for the change of member device
  md/raid1: fix potential data inconsistency issue with write behind device

5 years agomd: add bitmap_abort label in md_run
Guoqing Jiang [Fri, 14 Jun 2019 09:10:39 +0000 (17:10 +0800)]
md: add bitmap_abort label in md_run

Now, there are two places need to consider about
the failure of destroy bitmap, so move the common
part between bitmap_abort and abort label.

Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agomd-bitmap: create and destroy wb_info_pool with the change of bitmap
Guoqing Jiang [Fri, 14 Jun 2019 09:10:38 +0000 (17:10 +0800)]
md-bitmap: create and destroy wb_info_pool with the change of bitmap

The write-behind attribute is part of bitmap, since bitmap
can be added/removed dynamically with the following.

1. mdadm --grow /dev/md0 --bitmap=none
2. mdadm --grow /dev/md0 --bitmap=internal --write-behind

So we need to destroy wb_info_pool in md_bitmap_destroy,
and create the pool before load bitmap.

Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agomd-bitmap: create and destroy wb_info_pool with the change of backlog
Guoqing Jiang [Fri, 14 Jun 2019 09:10:37 +0000 (17:10 +0800)]
md-bitmap: create and destroy wb_info_pool with the change of backlog

Since we can enable write-behind mode by write backlog node,
so create wb_info_pool if the mode is just enabled, also call
call md_bitmap_update_sb to make user aware the write-behind
mode is enabled. Conversely, wb_info_pool should be destroyed
when write-behind mode is disabled.

Beside above, it is better to update bitmap sb if we change
the number of max_write_behind.

Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agomd: introduce mddev_create/destroy_wb_pool for the change of member device
Guoqing Jiang [Fri, 14 Jun 2019 09:10:36 +0000 (17:10 +0800)]
md: introduce mddev_create/destroy_wb_pool for the change of member device

Previously, we called rdev_init_wb to avoid potential data
inconsistency when array is created.

Now, we need to call the function and create mempool if a
device is added or just be flaged as "writemostly". So
mddev_create_wb_pool is introduced and called accordingly.
And for safety reason, we mark implicit GFP_NOIO allocation
scope for create mempool during mddev_suspend/mddev_resume.

And mempool should be removed conversely after remove a
member device or its's "writemostly" flag, which is done
by call mddev_destroy_wb_pool.

Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agomd/raid1: fix potential data inconsistency issue with write behind device
Guoqing Jiang [Wed, 19 Jun 2019 09:30:46 +0000 (17:30 +0800)]
md/raid1: fix potential data inconsistency issue with write behind device

For write-behind mode, we think write IO is complete once it has
reached all the non-writemostly devices. It works fine for single
queue devices.

But for multiqueue device, if there are lots of IOs come from upper
layer, then the write-behind device could issue those IOs to different
queues, depends on the each queue's delay, so there is no guarantee
that those IOs can arrive in order.

To address the issue, we need to check the collision among write
behind IOs, we can only continue without collision, otherwise wait
for the completion of previous collisioned IO.

And WBCollision is introduced for multiqueue device which is worked
under write-behind mode.

But this patch doesn't handle below cases which could have the data
inconsistency issue as well, these cases will be handled in later
patches.

1. modify max_write_behind by write backlog node.
2. add or remove array's bitmap dynamically.
3. the change of member disk.

Reviewed-by: NeilBrown <neilb@suse.com>
Signed-off-by: Guoqing Jiang <gqjiang@suse.com>
Signed-off-by: Song Liu <songliubraving@fb.com>
5 years agof2fs: use block layer helper for show_bio_op macro
Chaitanya Kulkarni [Thu, 20 Jun 2019 17:59:19 +0000 (10:59 -0700)]
f2fs: use block layer helper for show_bio_op macro

Adjust the f2fs tracing code to use newly introduced block layer
function blk_op_str() which converts the REQ_OP_XXX into the string
XXX.

Reviewed-by: Chao Yu <yuchao0@huawei.com>
Signed-off-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>