+0x3/0x20 [fat]
   [<f8b3a962>] fat_sync_inode+0x12/0x20 [fat]
   [<f8b37c7e>] fat_remove_entries+0xbe/0x120 [fat]
   [<f8b422ef>] vfat_unlink+0x5f/0x90 [vfat]
   [<f8b42290>] ? vfat_unlink+0x0/0x90 [vfat]
   [<c01b0968>] vfs_unlink+0x98/0x110
   [<c01b2400>] do_unlinkat+0x130/0x140
   [<c016a8f5>] ? audit_syscall_entry+0x105/0x150
   [<c01b253b>] sys_unlinkat+0x3b/0x40
   [<c01040d3>] sysenter_do_call+0x12/0x3f
   =======================

where the deadlock is due to the nesting of lock_super from vfat_unlink
to fat_write_inode:

 - do_unlinkat
   - vfs_unlink
     - vfat_unlink
       * lock_super
       - fat_remove_entries
         - fat_sync_inode
           - fat_write_inode
             * lock_super

and the fix is to simply remove the use of lock_super() in fat_write_inode.

The lock_super() there had been just an automatic conversion of the
kernel lock to the superblock lock, but no locking was actually needed
there, since the code in fat_write_inode already protected all relevant
accesses with a spinlock (sbi->inode_hash_lock to be exact).  The only
code inside the BKL (and thus the superblock lock) was accesses tp local
variables or calls to functions that have long been SMP-safe (i.e.
sb_bread, mark_buffe_dirty and brlese).

Bart reports:
 "Looks good.  I ran 10 parallel processes creating 1M files truncating
  them, writing to them again and then deleting them.  This patch fixes
  the issue I ran into.

  Signed-off-by: Bart Trojanowski <bart@jukie.net>"

Reported-and-tested-by: Bart Trojanowski <bart@jukie.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Tç’