1. 09 Jan, 2018 15 commits
    • Michael Lyle's avatar
      bcache: closures: move control bits one bit right · 3609c471
      Michael Lyle authored
      Otherwise, architectures that do negated adds of atomics (e.g. s390)
      to do atomic_sub fail in closure_set_stopped.
      Signed-off-by: default avatarMichael Lyle <mlyle@lyle.org>
      Cc: Kent Overstreet <kent.overstreet@gmail.com>
      Reported-by: default avatarkbuild test robot <lkp@intel.com>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      3609c471
    • Bart Van Assche's avatar
      block: Fix kernel-doc warnings reported when building with W=1 · aa98192d
      Bart Van Assche authored
      Commit 3a025e1d ("Add optional check for bad kernel-doc comments")
      causes W=1 the kernel-doc script to be run and thereby causes several
      new warnings to appear when building the kernel with W=1. Fix the
      block layer kernel-doc headers such that the block layer again builds
      cleanly with W=1.
      Signed-off-by: default avatarBart Van Assche <bart.vanassche@wdc.com>
      Cc: Martin K. Petersen <martin.petersen@oracle.com>
      Cc: Christoph Hellwig <hch@lst.de>
      Cc: Hannes Reinecke <hare@suse.de>
      Cc: Johannes Thumshirn <jthumshirn@suse.de>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      aa98192d
    • Bart Van Assche's avatar
      blk-mq: Fix spelling in a source code comment · ee3e4de5
      Bart Van Assche authored
      Change "nedeing" into "needing" and "caes" into "cases".
      
      Fixes: commit f906a6a0 ("blk-mq: improve tag waiting setup for non-shared tags")
      Signed-off-by: default avatarBart Van Assche <bart.vanassche@wdc.com>
      Cc: Christoph Hellwig <hch@lst.de>
      Cc: Omar Sandoval <osandov@fb.com>
      Cc: Hannes Reinecke <hare@suse.de>
      Cc: Johannes Thumshirn <jthumshirn@suse.de>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      ee3e4de5
    • Jens Axboe's avatar
      blk-mq: silence false positive warnings in hctx_unlock() · 08b5a6e2
      Jens Axboe authored
      In some stupider versions of gcc, it complains:
      
      block/blk-mq.c: In function ‘blk_mq_complete_request’:
      ./include/linux/srcu.h:175:2: warning: ‘srcu_idx’ may be used uninitialized in this function [-Wmaybe-uninitialized]
        __srcu_read_unlock(sp, idx);
        ^
      block/blk-mq.c:620:6: note: ‘srcu_idx’ was declared here
        int srcu_idx;
            ^
      
      which is completely bogus, since we only use srcu_idx when
      hctx->flags & BLK_MQ_F_BLOCKING is set, and that's the case where
      hctx_lock() has initialized it.
      
      Just set it to '0' in the normal path in hctx_lock() to silence
      this annoying warning.
      
      Fixes: 04ced159 ("blk-mq: move hctx lock/unlock into a helper")
      Fixes: 5197c05e ("blk-mq: protect completion path with RCU")
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      08b5a6e2
    • Tejun Heo's avatar
      blk-mq: rename blk_mq_hw_ctx->queue_rq_srcu to ->srcu · 05707b64
      Tejun Heo authored
      The RCU protection has been expanded to cover both queueing and
      completion paths making ->queue_rq_srcu a misnomer.  Rename it to
      ->srcu as suggested by Bart.
      Signed-off-by: default avatarTejun Heo <tj@kernel.org>
      Cc: Bart Van Assche <Bart.VanAssche@wdc.com>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      05707b64
    • Tejun Heo's avatar
      blk-mq: remove REQ_ATOM_STARTED · 5a61c363
      Tejun Heo authored
      After the recent updates to use generation number and state based
      synchronization, we can easily replace REQ_ATOM_STARTED usages by
      adding an extra state to distinguish completed but not yet freed
      state.
      
      Add MQ_RQ_COMPLETE and replace REQ_ATOM_STARTED usages with
      blk_mq_rq_state() tests.  REQ_ATOM_STARTED no longer has any users
      left and is removed.
      Signed-off-by: default avatarTejun Heo <tj@kernel.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      5a61c363
    • Tejun Heo's avatar
      blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq · 634f9e46
      Tejun Heo authored
      After the recent updates to use generation number and state based
      synchronization, blk-mq no longer depends on REQ_ATOM_COMPLETE except
      to avoid firing the same timeout multiple times.
      
      Remove all REQ_ATOM_COMPLETE usages and use a new rq_flags flag
      RQF_MQ_TIMEOUT_EXPIRED to avoid firing the same timeout multiple
      times.  This removes atomic bitops from hot paths too.
      
      v2: Removed blk_clear_rq_complete() from blk_mq_rq_timed_out().
      
      v3: Added RQF_MQ_TIMEOUT_EXPIRED flag.
      Signed-off-by: default avatarTejun Heo <tj@kernel.org>
      Cc: "jianchao.wang" <jianchao.w.wang@oracle.com>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      634f9e46
    • Tejun Heo's avatar
      blk-mq: make blk_abort_request() trigger timeout path · 358f70da
      Tejun Heo authored
      With issue/complete and timeout paths now using the generation number
      and state based synchronization, blk_abort_request() is the only one
      which depends on REQ_ATOM_COMPLETE for arbitrating completion.
      
      There's no reason for blk_abort_request() to be a completely separate
      path.  This patch makes blk_abort_request() piggyback on the timeout
      path instead of trying to terminate the request directly.
      
      This removes the last dependency on REQ_ATOM_COMPLETE in blk-mq.
      
      Note that this makes blk_abort_request() asynchronous - it initiates
      abortion but the actual termination will happen after a short while,
      even when the caller owns the request.  AFAICS, SCSI and ATA should be
      fine with that and I think mtip32xx and dasd should be safe but not
      completely sure.  It'd be great if people who know the drivers take a
      look.
      
      v2: - Add comment explaining the lack of synchronization around
            ->deadline update as requested by Bart.
      Signed-off-by: default avatarTejun Heo <tj@kernel.org>
      Cc: Asai Thambi SP <asamymuthupa@micron.com>
      Cc: Stefan Haberland <sth@linux.vnet.ibm.com>
      Cc: Jan Hoeppner <hoeppner@linux.vnet.ibm.com>
      Cc: Bart Van Assche <Bart.VanAssche@wdc.com>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      358f70da
    • Tejun Heo's avatar
      blk-mq: use blk_mq_rq_state() instead of testing REQ_ATOM_COMPLETE · 67818d25
      Tejun Heo authored
      blk_mq_check_inflight() and blk_mq_poll_hybrid_sleep() test
      REQ_ATOM_COMPLETE to determine the request state.  Both uses are
      speculative and we can test REQ_ATOM_STARTED and blk_mq_rq_state() for
      equivalent results.  Replace the tests.  This will allow removing
      REQ_ATOM_COMPLETE usages from blk-mq.
      Signed-off-by: default avatarTejun Heo <tj@kernel.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      67818d25
    • Tejun Heo's avatar
      blk-mq: replace timeout synchronization with a RCU and generation based scheme · 1d9bd516
      Tejun Heo authored
      Currently, blk-mq timeout path synchronizes against the usual
      issue/completion path using a complex scheme involving atomic
      bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence
      rules.  Unfortunately, it contains quite a few holes.
      
      There's a complex dancing around REQ_ATOM_STARTED and
      REQ_ATOM_COMPLETE between issue/completion and timeout paths; however,
      they don't have a synchronization point across request recycle
      instances and it isn't clear what the barriers add.
      blk_mq_check_expired() can easily read STARTED from N-2'th iteration,
      deadline from N-1'th, blk_mark_rq_complete() against Nth instance.
      
      In fact, it's pretty easy to make blk_mq_check_expired() terminate a
      later instance of a request.  If we induce 5 sec delay before
      time_after_eq() test in blk_mq_check_expired(), shorten the timeout to
      2s, and issue back-to-back large IOs, blk-mq starts timing out
      requests spuriously pretty quickly.  Nothing actually timed out.  It
      just made the call on a recycle instance of a request and then
      terminated a later instance long after the original instance finished.
      The scenario isn't theoretical either.
      
      This patch replaces the broken synchronization mechanism with a RCU
      and generation number based one.
      
      1. Each request has a u64 generation + state value, which can be
         updated only by the request owner.  Whenever a request becomes
         in-flight, the generation number gets bumped up too.  This provides
         the basis for the timeout path to distinguish different recycle
         instances of the request.
      
         Also, marking a request in-flight and setting its deadline are
         protected with a seqcount so that the timeout path can fetch both
         values coherently.
      
      2. The timeout path fetches the generation, state and deadline.  If
         the verdict is timeout, it records the generation into a dedicated
         request abortion field and does RCU wait.
      
      3. The completion path is also protected by RCU (from the previous
         patch) and checks whether the current generation number and state
         match the abortion field.  If so, it skips completion.
      
      4. The timeout path, after RCU wait, scans requests again and
         terminates the ones whose generation and state still match the ones
         requested for abortion.
      
         By now, the timeout path knows that either the generation number
         and state changed if it lost the race or the completion will yield
         to it and can safely timeout the request.
      
      While it's more lines of code, it's conceptually simpler, doesn't
      depend on direct use of subtle memory ordering or coherence, and
      hopefully doesn't terminate the wrong instance.
      
      While this change makes REQ_ATOM_COMPLETE synchronization unnecessary
      between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't
      removed yet as it's still used in other places.  Future patches will
      move all state tracking to the new mechanism and remove all bitops in
      the hot paths.
      
      Note that this patch adds a comment explaining a race condition in
      BLK_EH_RESET_TIMER path.  The race has always been there and this
      patch doesn't change it.  It's just documenting the existing race.
      
      v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao.
          - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter.
          - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter.
      
      v3: - Fixed possible extended seqcount / u64_stats_sync read looping
            spotted by Peter.
          - MQ_RQ_IDLE was incorrectly being set in complete_request instead
            of free_request.  Fixed.
      
      v4: - Rebased on top of hctx_lock() refactoring patch.
          - Added comment explaining the use of hctx_lock() in completion path.
      
      v5: - Added comments requested by Bart.
          - Note the addition of BLK_EH_RESET_TIMER race condition in the
            commit message.
      Signed-off-by: default avatarTejun Heo <tj@kernel.org>
      Cc: "jianchao.wang" <jianchao.w.wang@oracle.com>
      Cc: Peter Zijlstra <peterz@infradead.org>
      Cc: Christoph Hellwig <hch@lst.de>
      Cc: Bart Van Assche <Bart.VanAssche@wdc.com>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      1d9bd516
    • Tejun Heo's avatar
      blk-mq: protect completion path with RCU · 5197c05e
      Tejun Heo authored
      Currently, blk-mq protects only the issue path with RCU.  This patch
      puts the completion path under the same RCU protection.  This will be
      used to synchronize issue/completion against timeout by later patches,
      which will also add the comments.
      Signed-off-by: default avatarTejun Heo <tj@kernel.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      5197c05e
    • Jens Axboe's avatar
      blk-mq: move hctx lock/unlock into a helper · 04ced159
      Jens Axboe authored
      Move the RCU vs SRCU logic into lock/unlock helpers, which makes
      the actual functional bits within the locked region much easier
      to read.
      
      tj: Reordered in front of timeout revamp patches and added the missing
          blk_mq_run_hw_queue() conversion.
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      Signed-off-by: default avatarTejun Heo <tj@kernel.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      04ced159
    • Paolo Valente's avatar
      block, bfq: release oom-queue ref to root group on exit · 0d52af59
      Paolo Valente authored
      On scheduler init, a reference to the root group, and a reference to
      its corresponding blkg are taken for the oom queue. Yet these
      references are not released on scheduler exit, which prevents these
      objects from be freed. This commit adds the missing reference
      releases.
      Reported-by: default avatarDavide Ferrari <davideferrari8@gmail.com>
      Tested-by: default avatarHolger Hoffstätte <holger@applied-asynchrony.com>
      Signed-off-by: default avatarPaolo Valente <paolo.valente@linaro.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      0d52af59
    • Paolo Valente's avatar
      block, bfq: put async queues for root bfq groups too · 52257ffb
      Paolo Valente authored
      For each pair [device for which bfq is selected as I/O scheduler,
      group in blkio/io], bfq maintains a corresponding bfq group. Each such
      bfq group contains a set of async queues, with each async queue
      created on demand, i.e., when some I/O request arrives for it.  On
      creation, an async queue gets an extra reference, to make sure that
      the queue is not freed as long as its bfq group exists.  Accordingly,
      to allow the queue to be freed after the group exited, this extra
      reference must released on group exit.
      
      The above holds also for a bfq root group, i.e., for the bfq group
      corresponding to the root blkio/io root for a given device. Yet, by
      mistake, the references to the existing async queues of a root group
      are not released when the latter exits. This causes a memory leak when
      the instance of bfq for a given device exits. In a similar vein,
      bfqg_stats_xfer_dead is not executed for a root group.
      
      This commit fixes bfq_pd_offline so that the latter executes the above
      missing operations for a root group too.
      Reported-by: default avatarHolger Hoffstätte <holger@applied-asynchrony.com>
      Reported-by: default avatarGuoqing Jiang <gqjiang@suse.com>
      Tested-by: default avatarHolger Hoffstätte <holger@applied-asynchrony.com>
      Signed-off-by: default avatarDavide Ferrari <davideferrari8@gmail.com>
      Signed-off-by: default avatarPaolo Valente <paolo.valente@linaro.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      52257ffb
    • Ming Lei's avatar
      blk-mq: fix kernel oops in blk_mq_tag_idle() · 8ab0b7dc
      Ming Lei authored
      HW queues may be unmapped in some cases, such as blk_mq_update_nr_hw_queues(),
      then we need to check it before calling blk_mq_tag_idle(), otherwise
      the following kernel oops can be triggered, so fix it by checking if
      the hw queue is unmapped since it doesn't make sense to idle the tags
      any more after hw queues are unmapped.
      
      [  440.771298] Workqueue: nvme-wq nvme_rdma_del_ctrl_work [nvme_rdma]
      [  440.779104] task: ffff894bae755ee0 ti: ffff893bf9bc8000 task.ti: ffff893bf9bc8000
      [  440.788359] RIP: 0010:[<ffffffffb730e2b4>]  [<ffffffffb730e2b4>] __blk_mq_tag_idle+0x24/0x40
      [  440.798697] RSP: 0018:ffff893bf9bcbd10  EFLAGS: 00010286
      [  440.805538] RAX: 0000000000000000 RBX: ffff895bb131dc00 RCX: 000000000000011f
      [  440.814426] RDX: 00000000ffffffff RSI: 0000000000000120 RDI: ffff895bb131dc00
      [  440.823301] RBP: ffff893bf9bcbd10 R08: 000000000001b860 R09: 4a51d361c00c0000
      [  440.832193] R10: b5907f32b4cc7003 R11: ffffd6cabfb57000 R12: ffff894bafd1e008
      [  440.841091] R13: 0000000000000001 R14: ffff895baf770000 R15: 0000000000000080
      [  440.849988] FS:  0000000000000000(0000) GS:ffff894bbdcc0000(0000) knlGS:0000000000000000
      [  440.859955] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
      [  440.867274] CR2: 0000000000000008 CR3: 000000103d098000 CR4: 00000000001407e0
      [  440.876169] Call Trace:
      [  440.879818]  [<ffffffffb7309d68>] blk_mq_exit_hctx+0xd8/0xe0
      [  440.887051]  [<ffffffffb730dc40>] blk_mq_free_queue+0xf0/0x160
      [  440.894465]  [<ffffffffb72ff679>] blk_cleanup_queue+0xd9/0x150
      [  440.901881]  [<ffffffffc08a802b>] nvme_ns_remove+0x5b/0xb0 [nvme_core]
      [  440.910068]  [<ffffffffc08a811b>] nvme_remove_namespaces+0x3b/0x60 [nvme_core]
      [  440.919026]  [<ffffffffc08b817b>] __nvme_rdma_remove_ctrl+0x2b/0xb0 [nvme_rdma]
      [  440.928079]  [<ffffffffc08b8237>] nvme_rdma_del_ctrl_work+0x17/0x20 [nvme_rdma]
      [  440.937126]  [<ffffffffb70ab58a>] process_one_work+0x17a/0x440
      [  440.944517]  [<ffffffffb70ac3a8>] worker_thread+0x278/0x3c0
      [  440.951607]  [<ffffffffb70ac130>] ? manage_workers.isra.24+0x2a0/0x2a0
      [  440.959760]  [<ffffffffb70b352f>] kthread+0xcf/0xe0
      [  440.966055]  [<ffffffffb70b3460>] ? insert_kthread_work+0x40/0x40
      [  440.973715]  [<ffffffffb76d8658>] ret_from_fork+0x58/0x90
      [  440.980586]  [<ffffffffb70b3460>] ? insert_kthread_work+0x40/0x40
      [  440.988229] Code: 5b 41 5c 5d c3 66 90 0f 1f 44 00 00 48 8b 87 20 01 00 00 f0 0f ba 77 40 01 19 d2 85 d2 75 08 c3 0f 1f 80 00 00 00 00 55 48 89 e5 <f0> ff 48 08 48 8d 78 10 e8 7f 0f 05 00 5d c3 0f 1f 00 66 2e 0f
      [  441.011620] RIP  [<ffffffffb730e2b4>] __blk_mq_tag_idle+0x24/0x40
      [  441.019301]  RSP <ffff893bf9bcbd10>
      [  441.024052] CR2: 0000000000000008
      Reported-by: default avatarZhang Yi <yizhan@redhat.com>
      Tested-by: default avatarZhang Yi <yizhan@redhat.com>
      Signed-off-by: default avatarMing Lei <ming.lei@redhat.com>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      8ab0b7dc
  2. 08 Jan, 2018 25 commits
    • Michael Lyle's avatar
      bcache: fix writeback target calc on large devices · 616486ab
      Michael Lyle authored
      Bcache needs to scale the dirty data in the cache over the multiple
      backing disks in order to calculate writeback rates for each.
      The previous code did this by multiplying the target number of dirty
      sectors by the backing device size, and expected it to fit into a
      uint64_t; this blows up on relatively small backing devices.
      
      The new approach figures out the bdev's share in 16384ths of the overall
      cached data.  This is chosen to cope well when bdevs drastically vary in
      size and to ensure that bcache can cross the petabyte boundary for each
      backing device.
      
      This has been improved based on Tang Junhui's feedback to ensure that
      every device gets a share of dirty data, no matter how small it is
      compared to the total backing pool.
      
      The existing mechanism is very limited; this is purely a bug fix to
      remove limits on volume size.  However, there still needs to be change
      to make this "fair" over many volumes where some are idle.
      Reported-by: default avatarJack Douglas <jack@douglastechnology.co.uk>
      Signed-off-by: default avatarMichael Lyle <mlyle@lyle.org>
      Reviewed-by: default avatarTang Junhui <tang.junhui@zte.com.cn>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      616486ab
    • Coly Li's avatar
      bcache: fix misleading error message in bch_count_io_errors() · 5138ac67
      Coly Li authored
      Bcache only does recoverable I/O for read operations by calling
      cached_dev_read_error(). For write opertions there is no I/O recovery for
      failed requests.
      
      But in bch_count_io_errors() no matter read or write I/Os, before errors
      counter reaches io error limit, pr_err() always prints "IO error on %,
      recoverying". For write requests this information is misleading, because
      there is no I/O recovery at all.
      
      This patch adds a parameter 'is_read' to bch_count_io_errors(), and only
      prints "recovering" by pr_err() when the bio direction is READ.
      Signed-off-by: default avatarColy Li <colyli@suse.de>
      Reviewed-by: default avatarMichael Lyle <mlyle@lyle.org>
      Reviewed-by: default avatarTang Junhui <tang.junhui@zte.com.cn>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      5138ac67
    • Coly Li's avatar
      bcache: reduce cache_set devices iteration by devices_max_used · 2831231d
      Coly Li authored
      Member devices of struct cache_set is used to reference all attached
      bcache devices to this cache set. If it is treated as array of pointers,
      size of devices[] is indicated by member nr_uuids of struct cache_set.
      
      nr_uuids is calculated in drivers/md/super.c:bch_cache_set_alloc(),
      	bucket_bytes(c) / sizeof(struct uuid_entry)
      Bucket size is determined by user space tool "make-bcache", by default it
      is 1024 sectors (defined in bcache-tools/make-bcache.c:main()). So default
      nr_uuids value is 4096 from the above calculation.
      
      Every time when bcache code iterates bcache devices of a cache set, all
      the 4096 pointers are checked even only 1 bcache device is attached to the
      cache set, that's a wast of time and unncessary.
      
      This patch adds a member devices_max_used to struct cache_set. Its value
      is 1 + the maximum used index of devices[] in a cache set. When iterating
      all valid bcache devices of a cache set, use c->devices_max_used in
      for-loop may reduce a lot of useless checking.
      
      Personally, my motivation of this patch is not for performance, I use it
      in bcache debugging, which helps me to narrow down the scape to check
      valid bcached devices of a cache set.
      Signed-off-by: default avatarColy Li <colyli@suse.de>
      Reviewed-by: default avatarMichael Lyle <mlyle@lyle.org>
      Reviewed-by: default avatarTang Junhui <tang.junhui@zte.com.cn>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      2831231d
    • Zhai Zhaoxuan's avatar
      bcache: fix unmatched generic_end_io_acct() & generic_start_io_acct() · b40503ea
      Zhai Zhaoxuan authored
      The function cached_dev_make_request() and flash_dev_make_request() call
      generic_start_io_acct() with (struct bcache_device)->disk when they start a
      closure. Then the function bio_complete() calls generic_end_io_acct() with
      (struct search)->orig_bio->bi_disk when the closure has done.
      Since the `bi_disk` is not the bcache device, the generic_end_io_acct() is
      called with a wrong device queue.
      
      It causes the "inflight" (in struct hd_struct) counter keep increasing
      without decreasing.
      
      This patch fix the problem by calling generic_end_io_acct() with
      (struct bcache_device)->disk.
      Signed-off-by: default avatarZhai Zhaoxuan <kxuanobj@gmail.com>
      Reviewed-by: default avatarMichael Lyle <mlyle@lyle.org>
      Reviewed-by: default avatarColy Li <colyli@suse.de>
      Reviewed-by: default avatarTang Junhui <tang.junhui@zte.com.cn>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      b40503ea
    • Kent Overstreet's avatar
      bcache: mark closure_sync() __sched · ce439bf7
      Kent Overstreet authored
      [edit by mlyle: include sched/debug.h to get __sched]
      Signed-off-by: default avatarKent Overstreet <kent.overstreet@gmail.com>
      Signed-off-by: default avatarMichael Lyle <mlyle@lyle.org>
      Reviewed-by: default avatarMichael Lyle <mlyle@lyle.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      ce439bf7
    • Kent Overstreet's avatar
      bcache: Fix, improve efficiency of closure_sync() · e4bf7919
      Kent Overstreet authored
      Eliminates cases where sync can race and fail to complete / get stuck.
      Removes many status flags and simplifies entering-and-exiting closure
      sleeping behaviors.
      
      [mlyle: fixed conflicts due to changed return behavior in mainline.
      extended commit comment, and squashed down two commits that were mostly
      contradictory to get to this state.  Changed __set_current_state to
      set_current_state per Jens review comment]
      Signed-off-by: default avatarKent Overstreet <kent.overstreet@gmail.com>
      Signed-off-by: default avatarMichael Lyle <mlyle@lyle.org>
      Reviewed-by: default avatarMichael Lyle <mlyle@lyle.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      e4bf7919
    • Michael Lyle's avatar
      bcache: allow quick writeback when backing idle · b1092c9a
      Michael Lyle authored
      If the control system would wait for at least half a second, and there's
      been no reqs hitting the backing disk for awhile: use an alternate mode
      where we have at most one contiguous set of writebacks in flight at a
      time. (But don't otherwise delay).  If front-end IO appears, it will
      still be quick, as it will only have to contend with one real operation
      in flight.  But otherwise, we'll be sending data to the backing disk as
      quickly as it can accept it (with one op at a time).
      Signed-off-by: default avatarMichael Lyle <mlyle@lyle.org>
      Reviewed-by: default avatarTang Junhui <tang.junhui@zte.com.cn>
      Acked-by: default avatarColy Li <colyli@suse.de>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      b1092c9a
    • Michael Lyle's avatar
      bcache: writeback: properly order backing device IO · 6e6ccc67
      Michael Lyle authored
      Writeback keys are presently iterated and dispatched for writeback in
      order of the logical block address on the backing device.  Multiple may
      be, in parallel, read from the cache device and then written back
      (especially when there are contiguous I/O).
      
      However-- there was no guarantee with the existing code that the writes
      would be issued in LBA order, as the reads from the cache device are
      often re-ordered.  In turn, when writing back quickly, the backing disk
      often has to seek backwards-- this slows writeback and increases
      utilization.
      
      This patch introduces an ordering mechanism that guarantees that the
      original order of issue is maintained for the write portion of the I/O.
      Performance for writeback is significantly improved when there are
      multiple contiguous keys or high writeback rates.
      Signed-off-by: default avatarMichael Lyle <mlyle@lyle.org>
      Reviewed-by: default avatarTang Junhui <tang.junhui@zte.com.cn>
      Tested-by: default avatarTang Junhui <tang.junhui@zte.com.cn>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      6e6ccc67
    • Tang Junhui's avatar
      bcache: fix wrong return value in bch_debug_init() · 539d39eb
      Tang Junhui authored
      in bch_debug_init(), ret is always 0, and the return value is useless,
      change it to return 0 if be success after calling debugfs_create_dir(),
      else return a non-zero value.
      Signed-off-by: default avatarTang Junhui <tang.junhui@zte.com.cn>
      Reviewed-by: default avatarMichael Lyle <mlyle@lyle.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      539d39eb
    • Tang Junhui's avatar
      bcache: segregate flash only volume write streams · 4eca1cb2
      Tang Junhui authored
      In such scenario that there are some flash only volumes
      , and some cached devices, when many tasks request these devices in
      writeback mode, the write IOs may fall to the same bucket as bellow:
      | cached data | flash data | cached data | cached data| flash data|
      then after writeback of these cached devices, the bucket would
      be like bellow bucket:
      | free | flash data | free | free | flash data |
      
      So, there are many free space in this bucket, but since data of flash
      only volumes still exists, so this bucket cannot be reclaimable,
      which would cause waste of bucket space.
      
      In this patch, we segregate flash only volume write streams from
      cached devices, so data from flash only volumes and cached devices
      can store in different buckets.
      
      Compare to v1 patch, this patch do not add a additionally open bucket
      list, and it is try best to segregate flash only volume write streams
      from cached devices, sectors of flash only volumes may still be mixed
      with dirty sectors of cached device, but the number is very small.
      
      [mlyle: fixed commit log formatting, permissions, line endings]
      Signed-off-by: default avatarTang Junhui <tang.junhui@zte.com.cn>
      Reviewed-by: default avatarMichael Lyle <mlyle@lyle.org>
      Signed-off-by: default avatarMichael Lyle <mlyle@lyle.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      4eca1cb2
    • Vasyl Gomonovych's avatar
      bcache: Use PTR_ERR_OR_ZERO() · 9d134117
      Vasyl Gomonovych authored
      Fix ptr_ret.cocci warnings:
      drivers/md/bcache/btree.c:1800:1-3: WARNING: PTR_ERR_OR_ZERO can be used
      
      Use PTR_ERR_OR_ZERO rather than if(IS_ERR(...)) + PTR_ERR
      
      Generated by: scripts/coccinelle/api/ptr_ret.cocci
      Signed-off-by: default avatarVasyl Gomonovych <gomonovych@gmail.com>
      Reviewed-by: default avatarMichael Lyle <mlyle@lyle.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      9d134117
    • Tang Junhui's avatar
      bcache: stop writeback thread after detaching · 8d29c442
      Tang Junhui authored
      Currently, when a cached device detaching from cache, writeback thread is
      not stopped, and writeback_rate_update work is not canceled. For example,
      after the following command:
      echo 1 >/sys/block/sdb/bcache/detach
      you can still see the writeback thread. Then you attach the device to the
      cache again, bcache will create another writeback thread, for example,
      after below command:
      echo  ba0fb5cd-658a-4533-9806-6ce166d883b9 > /sys/block/sdb/bcache/attach
      then you will see 2 writeback threads.
      This patch stops writeback thread and cancels writeback_rate_update work
      when cached device detaching from cache.
      
      Compare with patch v1, this v2 patch moves code down into the register
      lock for safety in case of any future changes as Coly and Mike suggested.
      
      [edit by mlyle: commit log spelling/formatting]
      Signed-off-by: default avatarTang Junhui <tang.junhui@zte.com.cn>
      Reviewed-by: default avatarMichael Lyle <mlyle@lyle.org>
      Signed-off-by: default avatarMichael Lyle <mlyle@lyle.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      8d29c442
    • Rui Hua's avatar
      bcache: ret IOERR when read meets metadata error · b221fc13
      Rui Hua authored
      The read request might meet error when searching the btree, but the error
      was not handled in cache_lookup(), and this kind of metadata failure will
      not go into cached_dev_read_error(), finally, the upper layer will receive
      bi_status=0.  In this patch we judge the metadata error by the return
      value of bch_btree_map_keys(), there are two potential paths give rise to
      the error:
      
      1. Because the btree is not totally cached in memery, we maybe get error
         when read btree node from cache device (see bch_btree_node_get()), the
         likely errno is -EIO, -ENOMEM
      
      2. When read miss happens, bch_btree_insert_check_key() will be called to
         insert a "replace_key" to btree(see cached_dev_cache_miss(), just for
         doing preparatory work before insert the missed data to cache device),
         a failure can also happen in this situation, the likely errno is
         -ENOMEM
      
      bch_btree_map_keys() will return MAP_DONE in normal scenario, but we will
      get either -EIO or -ENOMEM in above two cases. if this happened, we should
      NOT recover data from backing device (when cache device is dirty) because
      we don't know whether bkeys the read request covered are all clean.  And
      after that happened, s->iop.status is still its initially value(0) before
      we submit s->bio.bio, we set it to BLK_STS_IOERR, so it can go into
      cached_dev_read_error(), and finally it can be passed to upper layer, or
      recovered by reread from backing device.
      
      [edit by mlyle: patch formatting, word-wrap, comment spelling,
      commit log format]
      Signed-off-by: default avatarHua Rui <huarui.dev@gmail.com>
      Reviewed-by: default avatarMichael Lyle <mlyle@lyle.org>
      Signed-off-by: default avatarMichael Lyle <mlyle@lyle.org>
      Signed-off-by: default avatarJens Axboe <axboe@kernel.dk>
      b221fc13
    • Jens Axboe's avatar
      Merge branch 'nvme-4.16' of git://git.infradead.org/nvme into for-4.16/block · 550203e6
      Jens Axboe authored
      Pull NVMe fixes from Christoph:
      
      "Below are the pending nvme updates for Linux 4.16. Just fixes and
       cleanups from various contributors this time around."
      550203e6
    • Israel Rukshin's avatar
      nvme: fix subsystem multiple controllers support check · b837b283
      Israel Rukshin authored
      There is a problem when another module (e.g. nvmet) takes a reference on
      the nvme block device and the physical nvme drive is removed.  In that
      case nvme_free_ctrl() will not be called and the controller state will be
      "deleting" or "dead" unless nvmet module releases the block device.
      Later on, the same nvme drive probes back and nvme_init_subsystem() will
      be called and fail due to duplicate subnqn (if the nvme device doesn't
      support subsystem with multiple controllers). This will cause a probe
      failure.  This commit changes the check of multiple controllers support
      at nvme_init_subsystem() by not counting all the controllers at "dead" or
      "deleting" state (this is safe because controllers at this state will
      never be active again).
      
      Fixes: ab9e00cc ("nvme: track subsystems")
      Reviewed-by: default avatarMax Gurtovoy <maxg@mellanox.com>
      Signed-off-by: default avatarIsrael Rukshin <israelr@mellanox.com>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      b837b283
    • Nitzan Carmi's avatar
      nvme: take refcount on transport module · 85088c4a
      Nitzan Carmi authored
      The block device is backed by the transport so we must ensure that the
      transport driver will not be removed until all references are released.
      Otherwise, we might end up referencing freed memory.
      Reviewed-by: default avatarMax Gurtovoy <maxg@mellanox.com>
      Signed-off-by: default avatarNitzan Carmi <nitzanc@mellanox.com>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      85088c4a
    • Jianchao Wang's avatar
      nvme-pci: fix NULL pointer reference in nvme_alloc_ns · 2b1b7e78
      Jianchao Wang authored
      When the io queues setup or tagset allocation failed, ctrl.tagset is
      NULL.  But the scan work will still be queued and executed, then panic
      comes up due to NULL pointer reference of ctrl.tagset.
      
      To fix this, add a new ctrl state NVME_CTRL_ADMIN_ONLY to inidcate only
      admin queue is live. When non io queues or tagset allocation failed, ctrl
      enters into this state, scan work will not be started.  But async event
      work and nvme dev ioctl will be still available.  This will be helpful to
      do further investigation and recovery.
      Suggested-by: default avatarSagi Grimberg <sagi@grimberg.me>
      Signed-off-by: default avatarJianchao Wang <jianchao.w.wang@oracle.com>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      2b1b7e78
    • Max Gurtovoy's avatar
      nvme: modify the debug level for setting shutdown timeout · 1a3838d7
      Max Gurtovoy authored
      When an NVMe controller reports RTD3 Entry Latency larger than the value
      of shutdown_timeout module parameter, we update the shutdown_timeout
      accordingly to honor RTD3 Entry Latency. Use an informational debug level
      instead of a warning level for it.
      Signed-off-by: default avatarMax Gurtovoy <maxg@mellanox.com>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      1a3838d7
    • Sagi Grimberg's avatar
      nvme-pci: don't open-code nvme_reset_ctrl · 4caff8fc
      Sagi Grimberg authored
      Signed-off-by: default avatarSagi Grimberg <sagi@grimberg.me>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      4caff8fc
    • Israel Rukshin's avatar
      nvmet: rearrange nvmet_ctrl_free() · 6b1943af
      Israel Rukshin authored
      Make it symmetric to nvmet_alloc_ctrl().
      Signed-off-by: default avatarIsrael Rukshin <israelr@mellanox.com>
      Reviewed-by: default avatarMax Gurtovoy <maxg@mellanox.com>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      6b1943af
    • Israel Rukshin's avatar
      nvmet: fix error flow in nvmet_alloc_ctrl() · eca19dc1
      Israel Rukshin authored
      Remove the allocated id on error.
      Signed-off-by: default avatarIsrael Rukshin <israelr@mellanox.com>
      Reviewed-by: default avatarMax Gurtovoy <maxg@mellanox.com>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      eca19dc1
    • Minwoo Im's avatar
      nvme-pci: remove an unnecessary initialization in HMB code · 6fbcde66
      Minwoo Im authored
      The local variable __size__ will be set a bit later in a for-loop.
      Remove the explicit initialization at the beginning of this function.
      Signed-off-by: default avatarMinwoo Im <minwoo.im.dev@gmail.com>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      6fbcde66
    • Roy Shterman's avatar
      nvme-fabrics: protect against module unload during create_ctrl · 0de5cd36
      Roy Shterman authored
      NVMe transport driver module unload may (and usually does) trigger
      iteration over the active controllers and delete them all (sometimes
      under a mutex).  However, a controller can be created concurrently with
      module unload which can lead to leakage of resources (most important char
      device node leakage) in case the controller creation occured after the
      unload delete and drain sequence.  To protect against this, we take a
      module reference to guarantee that the nvme transport driver is not
      unloaded while creating a controller.
      Signed-off-by: default avatarRoy Shterman <roys@lightbitslabs.com>
      Signed-off-by: default avatarSagi Grimberg <sagi@grimberg.me>
      Reviewed-by: default avatarMax Gurtovoy <maxg@mellanox.com>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      0de5cd36
    • James Smart's avatar
      nvmet-fc: cleanup nvmet add_port/remove_port · 9ce1f2e1
      James Smart authored
      The current fc transport add_port routine validates that there is a
      matching port to the target port config. It then takes a reference
      on the targetport. The del_port removes the reference.
      
      Unfortunately, if the LLDD undergoes a hw reset or driver unload and
      wants to unreg the targetport, due to the reference, the targetport
      effectively can't be removed. It requires the admin to remove the
      port from the nvmet config first, which calls the del_port.
      Note: it appears nvmetcli clear skips over the del_port call (I'm
      not attempting to change that).
      
      There's no real reason to take the reference. With FC, there is nothing
      to enable or disable as the presence of the FC targetport implicitly
      means its enabled, and removal of the targtport means its disabled.
      
      Change add_port to simply validate and change remove_port to a noop.
      No references are taken on the targetport.
      Signed-off-by: default avatarJames Smart <james.smart@broadcom.com>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      9ce1f2e1
    • James Smart's avatar
      nvme_fcloop: refactor host/target io job access · b6f80773
      James Smart authored
      The split between what the host accesses on its flows vs what the
      target side accesses was flawed. Abort handling didn't properly
      clear initiator vs target structure cross-reference and locks
      weren't used for synchronization. Thus, there were issues of
      freeing structures too soon and access after free.
      
      A couple of these existed pre the IN_ISR mods, but when the
      target upcalls were converted to work items, thus adding delays
      between the 2 sides of accesses, the problems became pronounced.
      
      Resolve by:
      - tracking io state mainly in the tgt-side io structure.
      - make the tgt-side io structure released by reference not by
        code flow.
      - when changing initiator structures, use locks for
        synchronization
      - aborts are clearly tracked for which side saw the abort, and
        after seeing the abort, cross-references are cleared under lock.
      Signed-off-by: default avatarJames Smart <james.smart@broadcom.com>
      Signed-off-by: default avatarChristoph Hellwig <hch@lst.de>
      b6f80773