scsi_dh_alua: Fix a recently introduced deadlock

While retesting the SRP initiator I ran the command "rmmod mlx4_ib"
while I/O was in progress. That command triggers SCSI device removal
indirectly. Avoid that this action triggers the following deadlock:

=================================
[ INFO: inconsistent lock state ]
4.6.0-rc0-dbg+ #2 Tainted: G           O
---------------------------------
inconsistent {IN-SOFTIRQ-W} -> {SOFTIRQ-ON-W} usage.
multipathd/484 [HC0[0]:SC0[0]:HE1:SE1] takes:
 (&(&pg->lock)->rlock){+.?...}, at: [<ffffffffa04f50a2>] alua_bus_detach+0x52/0xa0 [scsi_dh_alua]
{IN-SOFTIRQ-W} state was registered at:
  [<ffffffff810a64a9>] __lock_acquire+0x7e9/0x1ad0
  [<ffffffff810a7fd0>] lock_acquire+0x60/0x80
  [<ffffffff8159910e>] _raw_spin_lock_irqsave+0x3e/0x60
  [<ffffffffa04f5131>] alua_rtpg_queue+0x41/0x1d0 [scsi_dh_alua]
  [<ffffffffa04f5531>] alua_check+0xe1/0x220 [scsi_dh_alua]
  [<ffffffffa04f5709>] alua_check_sense+0x99/0xb0 [scsi_dh_alua]
  [<ffffffff813f0d01>] scsi_check_sense+0x71/0x3f0
  [<ffffffff813f2f8b>] scsi_decide_disposition+0x18b/0x1d0
  [<ffffffff813f6e52>] scsi_softirq_done+0x52/0x140
  [<ffffffff812a26f2>] blk_done_softirq+0x52/0x90
  [<ffffffff8105bc1f>] __do_softirq+0x10f/0x230
  [<ffffffff8105bec8>] irq_exit+0xa8/0xb0
  [<ffffffff8101a675>] do_IRQ+0x65/0x110
  [<ffffffff8159a2c9>] ret_from_intr+0x0/0x19
  [<ffffffff811732f1>] kmem_cache_alloc+0x151/0x190
  [<ffffffff8118e534>] create_object+0x34/0x2d0
  [<ffffffff8158eaa6>] kmemleak_alloc_percpu+0x56/0xd0
  [<ffffffff8113ab0d>] pcpu_alloc+0x38d/0x660
  [<ffffffff8113aded>] __alloc_percpu_gfp+0xd/0x10
  [<ffffffff812e56a5>] __percpu_counter_init+0x55/0xb0
  [<ffffffff812b4989>] blkg_alloc+0x79/0x230
  [<ffffffff812b6756>] blkcg_init_queue+0x26/0x1d0
  [<ffffffff81297eed>] blk_alloc_queue_node+0x27d/0x2e0
  [<ffffffffa017766c>] dm_create+0x20c/0x570 [dm_mod]
  [<ffffffffa017e356>] dev_create+0x56/0x2c0 [dm_mod]
  [<ffffffffa017dcae>] ctl_ioctl+0x26e/0x520 [dm_mod]
  [<ffffffffa017df6e>] dm_ctl_ioctl+0xe/0x20 [dm_mod]
  [<ffffffff811aa8ee>] do_vfs_ioctl+0x8e/0x660
  [<ffffffff811aaefc>] SyS_ioctl+0x3c/0x70
  [<ffffffff81599929>] entry_SYSCALL_64_fastpath+0x1c/0xac
irq event stamp: 4290931
hardirqs last  enabled at (4290931): [ 1662.892772]
[<ffffffff81599341>] _raw_spin_unlock_irqrestore+0x31/0x50
hardirqs last disabled at (4290930): [<ffffffff815990e7>] _raw_spin_lock_irqsave+0x17/0x60
softirqs last  enabled at (4290774): [<ffffffff8105bcdb>] __do_softirq+0x1cb/0x230
softirqs last disabled at (4289831): [<ffffffff8105bec8>] irq_exit+0xa8/0xb0

other info that might help us debug this:
 Possible unsafe locking scenario:

       CPU0
       ----
  lock(&(&pg->lock)->rlock);
  <Interrupt>
    lock(&(&pg->lock)->rlock);

 *** DEADLOCK ***

2 locks held by multipathd/484:
 #0:  (&bdev->bd_mutex){+.+.+.}, at: [<ffffffff811d1cc3>] __blkdev_put+0x33/0x360
 #1:  (sd_ref_mutex){+.+...}, at: [<ffffffff81400afc>] scsi_disk_put+0x1c/0x40

stack backtrace:
CPU: 6 PID: 484 Comm: multipathd Tainted: G           O    4.6.0-rc0-dbg+ #2
Call Trace:
 [<ffffffff812bd115>] dump_stack+0x67/0x92
 [<ffffffff810a5175>] print_usage_bug+0x215/0x240
 [<ffffffff810a56ea>] mark_lock+0x54a/0x610
 [<ffffffff810a6505>] __lock_acquire+0x845/0x1ad0
 [<ffffffff810a7fd0>] lock_acquire+0x60/0x80
 [<ffffffff81598f23>] _raw_spin_lock+0x33/0x50
 [<ffffffffa04f50a2>] alua_bus_detach+0x52/0xa0 [scsi_dh_alua]
 [<ffffffff813ff6f7>] scsi_dh_release_device+0x17/0x50
 [<ffffffff813fb8da>] scsi_device_dev_release_usercontext+0x2a/0x120
 [<ffffffff810701f0>] execute_in_process_context+0x80/0x90
 [<ffffffff813fb8a7>] scsi_device_dev_release+0x17/0x20
 [<ffffffff813c8cfd>] device_release+0x2d/0x90
 [<ffffffff812bfa8a>] kobject_release+0x7a/0x190
 [<ffffffff812bf946>] kobject_put+0x26/0x50
 [<ffffffff813c8ee2>] put_device+0x12/0x20
 [<ffffffff813edc86>] scsi_device_put+0x26/0x30
 [<ffffffff81400b0d>] scsi_disk_put+0x2d/0x40
 [<ffffffff81400b68>] sd_release+0x48/0xb0
 [<ffffffff811d1f2e>] __blkdev_put+0x29e/0x360
 [<ffffffff811d24b9>] blkdev_put+0x49/0x170
 [<ffffffff811d2600>] blkdev_close+0x20/0x30
 [<ffffffff81198f48>] __fput+0xe8/0x1f0
 [<ffffffff81199089>] ____fput+0x9/0x10
 [<ffffffff81075d9e>] task_work_run+0x6e/0xa0
 [<ffffffff81001119>] exit_to_usermode_loop+0xa9/0xb0
 [<ffffffff81001590>] syscall_return_slowpath+0xb0/0xc0
 [<ffffffff815999b7>] entry_SYSCALL_64_fastpath+0xaa/0xac

Fixes: cb0a168cb6b8 (scsi_dh_alua: update 'access_state' field)
Cc: Hannes Reinecke <hare@suse.de>
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Laurence Oberman <loberman@redhat.com>
Reviewed-by: Hannes Reinicke <hare@suse.de>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ewan Milne <emilne@redhat.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c b/drivers/scsi/device_handler/scsi_dh_alua.c
index a404a41e871c23b18aa4ea10d83af322f9107b40..8eaed0522aa36e83ddbe259a4fe2293c5ac61579 100644 (file)
--- a/drivers/scsi/device_handler/scsi_dh_alua.c
+++ b/drivers/scsi/device_handler/scsi_dh_alua.c
@@ -1112,9 +1112,9 @@ static void alua_bus_detach(struct scsi_device *sdev)
        h->sdev = NULL;
        spin_unlock(&h->pg_lock);
        if (pg) {
-               spin_lock(&pg->lock);
+               spin_lock_irq(&pg->lock);
                list_del_rcu(&h->node);
-               spin_unlock(&pg->lock);
+               spin_unlock_irq(&pg->lock);
                kref_put(&pg->kref, release_port_group);
        }
        sdev->handler_data = NULL;