summaryrefslogtreecommitdiffstats
path: root/drivers/mtd
diff options
context:
space:
mode:
authorJosh Cartwright <joshc@linux.com>2012-03-29 19:34:53 -0400
committerDavid Woodhouse <David.Woodhouse@intel.com>2012-05-07 20:30:14 +0100
commit226bb7df3d22bcf4a1c0fe8206c80cc427498eae (patch)
tree9e876ce5c9a8699ecacccf9dc3ee4a9c436015c3 /drivers/mtd
parent7a84477c4acebf6299b6a8bd6a1d5894eb838ffa (diff)
downloadlinux-226bb7df3d22bcf4a1c0fe8206c80cc427498eae.tar.bz2
jffs2: Fix lock acquisition order bug in gc path
The locking policy is such that the erase_complete_block spinlock is nested within the alloc_sem mutex. This fixes a case in which the acquisition order was erroneously reversed. This issue was caught by the following lockdep splat: ======================================================= [ INFO: possible circular locking dependency detected ] 3.0.5 #1 ------------------------------------------------------- jffs2_gcd_mtd6/299 is trying to acquire lock: (&c->alloc_sem){+.+.+.}, at: [<c01f7714>] jffs2_garbage_collect_pass+0x314/0x890 but task is already holding lock: (&(&c->erase_completion_lock)->rlock){+.+...}, at: [<c01f7708>] jffs2_garbage_collect_pass+0x308/0x890 which lock already depends on the new lock. the existing dependency chain (in reverse order) is: -> #1 (&(&c->erase_completion_lock)->rlock){+.+...}: [<c008bec4>] validate_chain+0xe6c/0x10bc [<c008c660>] __lock_acquire+0x54c/0xba4 [<c008d240>] lock_acquire+0xa4/0x114 [<c046780c>] _raw_spin_lock+0x3c/0x4c [<c01f744c>] jffs2_garbage_collect_pass+0x4c/0x890 [<c01f937c>] jffs2_garbage_collect_thread+0x1b4/0x1cc [<c0071a68>] kthread+0x98/0xa0 [<c000f264>] kernel_thread_exit+0x0/0x8 -> #0 (&c->alloc_sem){+.+.+.}: [<c008ad2c>] print_circular_bug+0x70/0x2c4 [<c008c08c>] validate_chain+0x1034/0x10bc [<c008c660>] __lock_acquire+0x54c/0xba4 [<c008d240>] lock_acquire+0xa4/0x114 [<c0466628>] mutex_lock_nested+0x74/0x33c [<c01f7714>] jffs2_garbage_collect_pass+0x314/0x890 [<c01f937c>] jffs2_garbage_collect_thread+0x1b4/0x1cc [<c0071a68>] kthread+0x98/0xa0 [<c000f264>] kernel_thread_exit+0x0/0x8 other info that might help us debug this: Possible unsafe locking scenario: CPU0 CPU1 ---- ---- lock(&(&c->erase_completion_lock)->rlock); lock(&c->alloc_sem); lock(&(&c->erase_completion_lock)->rlock); lock(&c->alloc_sem); *** DEADLOCK *** 1 lock held by jffs2_gcd_mtd6/299: #0: (&(&c->erase_completion_lock)->rlock){+.+...}, at: [<c01f7708>] jffs2_garbage_collect_pass+0x308/0x890 stack backtrace: [<c00155dc>] (unwind_backtrace+0x0/0x100) from [<c0463dc0>] (dump_stack+0x20/0x24) [<c0463dc0>] (dump_stack+0x20/0x24) from [<c008ae84>] (print_circular_bug+0x1c8/0x2c4) [<c008ae84>] (print_circular_bug+0x1c8/0x2c4) from [<c008c08c>] (validate_chain+0x1034/0x10bc) [<c008c08c>] (validate_chain+0x1034/0x10bc) from [<c008c660>] (__lock_acquire+0x54c/0xba4) [<c008c660>] (__lock_acquire+0x54c/0xba4) from [<c008d240>] (lock_acquire+0xa4/0x114) [<c008d240>] (lock_acquire+0xa4/0x114) from [<c0466628>] (mutex_lock_nested+0x74/0x33c) [<c0466628>] (mutex_lock_nested+0x74/0x33c) from [<c01f7714>] (jffs2_garbage_collect_pass+0x314/0x890) [<c01f7714>] (jffs2_garbage_collect_pass+0x314/0x890) from [<c01f937c>] (jffs2_garbage_collect_thread+0x1b4/0x1cc) [<c01f937c>] (jffs2_garbage_collect_thread+0x1b4/0x1cc) from [<c0071a68>] (kthread+0x98/0xa0) [<c0071a68>] (kthread+0x98/0xa0) from [<c000f264>] (kernel_thread_exit+0x0/0x8) This was introduce in '81cfc9f jffs2: Fix serious write stall due to erase'. Cc: stable@kernel.org [2.6.37+] Signed-off-by: Josh Cartwright <joshc@linux.com> Signed-off-by: Artem Bityutskiy <artem.bityutskiy@linux.intel.com> Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Diffstat (limited to 'drivers/mtd')
0 files changed, 0 insertions, 0 deletions