Revert "scsi: make 'state' device attribute pollable"
authorLinus Torvalds <torvalds@linux-foundation.org>
Tue, 7 Nov 2017 17:04:32 +0000 (09:04 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 7 Nov 2017 17:04:32 +0000 (09:04 -0800)
commita817e73fe693f0718b6210f4b959478877fb2e2f
tree377b6fd9c622f4d67448bf1c716522ebc8601140
parente4880bc5dfb1f02b152e62a894b5c6f3e995b3cf
Revert "scsi: make 'state' device attribute pollable"

This reverts commit 8a97712e5314aefe16b3ffb4583a34deaa49de04.

This commit added a call to sysfs_notify() from within
scsi_device_set_state(), which in turn turns out to make libata very
unhappy, because ata_eh_detach_dev() does

        spin_lock_irqsave(ap->lock, flags);
        ..
        if (ata_scsi_offline_dev(dev)) {
                dev->flags |= ATA_DFLAG_DETACHED;
                ap->pflags |= ATA_PFLAG_SCSI_HOTPLUG;
        }

and ata_scsi_offline_dev() then does that scsi_device_set_state() to set
it offline.

So now we called sysfs_notify() from within a spinlocked region, which
really doesn't work.  The 0day robot reported this as:

   BUG: sleeping function called from invalid context at kernel/locking/mutex.c:238

because sysfs_notify() ends up calling kernfs_find_and_get_ns() which
then does mutex_lock(&kernfs_mutex)..

The pollability of the device state isn't critical, so revert this all
for now, and maybe we'll do it differently in the future.

Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Acked-by: Tejun Heo <tj@kernel.org>
Acked-by: Martin K. Petersen <martin.petersen@oracle.com>
Acked-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/scsi/scsi_lib.c
drivers/scsi/scsi_transport_srp.c