Btrfs: find_free_extent: Do not erroneously skip LOOP_CACHING_WAIT state
authorchandan <chandan@linux.vnet.ibm.com>
Mon, 2 Nov 2015 08:29:46 +0000 (13:59 +0530)
committerChris Mason <clm@fb.com>
Tue, 3 Nov 2015 15:44:20 +0000 (07:44 -0800)
When executing generic/001 in a loop on a ppc64 machine (with both sectorsize
and nodesize set to 64k), the following call trace is observed,

WARNING: at /root/repos/linux/fs/btrfs/locking.c:253
Modules linked in:
CPU: 2 PID: 8353 Comm: umount Not tainted 4.3.0-rc5-13676-ga5e681d #54
task: c0000000f2b1f560 ti: c0000000f6008000 task.ti: c0000000f6008000
NIP: c000000000520c88 LR: c0000000004a3b34 CTR: 0000000000000000
REGS: c0000000f600a820 TRAP: 0700   Not tainted  (4.3.0-rc5-13676-ga5e681d)
MSR: 8000000102029032 <SF,VEC,EE,ME,IR,DR,RI>  CR: 24444884  XER: 00000000
CFAR: c0000000004a3b30 SOFTE: 1
GPR00: c0000000004a3b34 c0000000f600aaa0 c00000000108ac00 c0000000f5a808c0
GPR04: 0000000000000000 c0000000f600ae60 0000000000000000 0000000000000005
GPR08: 00000000000020a1 0000000000000001 c0000000f2b1f560 0000000000000030
GPR12: 0000000084842882 c00000000fdc0900 c0000000f600ae60 c0000000f070b800
GPR16: 0000000000000000 c0000000f3c8a000 0000000000000000 0000000000000049
GPR20: 0000000000000001 0000000000000001 c0000000f5aa01f8 0000000000000000
GPR24: 0f83e0f83e0f83e1 c0000000f5a808c0 c0000000f3c8d000 c000000000000000
GPR28: c0000000f600ae74 0000000000000001 c0000000f3c8d000 c0000000f5a808c0
NIP [c000000000520c88] .btrfs_tree_lock+0x48/0x2a0
LR [c0000000004a3b34] .btrfs_lock_root_node+0x44/0x80
Call Trace:
[c0000000f600aaa0] [c0000000f600ab80] 0xc0000000f600ab80 (unreliable)
[c0000000f600ab80] [c0000000004a3b34] .btrfs_lock_root_node+0x44/0x80
[c0000000f600ac00] [c0000000004a99dc] .btrfs_search_slot+0xa8c/0xc00
[c0000000f600ad40] [c0000000004ab878] .btrfs_insert_empty_items+0x98/0x120
[c0000000f600adf0] [c00000000050da44] .btrfs_finish_chunk_alloc+0x1d4/0x620
[c0000000f600af20] [c0000000004be854] .btrfs_create_pending_block_groups+0x1d4/0x2c0
[c0000000f600b020] [c0000000004bf188] .do_chunk_alloc+0x3c8/0x420
[c0000000f600b100] [c0000000004c27cc] .find_free_extent+0xbfc/0x1030
[c0000000f600b260] [c0000000004c2ce8] .btrfs_reserve_extent+0xe8/0x250
[c0000000f600b330] [c0000000004c2f90] .btrfs_alloc_tree_block+0x140/0x590
[c0000000f600b440] [c0000000004a47b4] .__btrfs_cow_block+0x124/0x780
[c0000000f600b530] [c0000000004a4fc0] .btrfs_cow_block+0xf0/0x250
[c0000000f600b5e0] [c0000000004a917c] .btrfs_search_slot+0x22c/0xc00
[c0000000f600b720] [c00000000050aa40] .btrfs_remove_chunk+0x1b0/0x9f0
[c0000000f600b850] [c0000000004c4e04] .btrfs_delete_unused_bgs+0x434/0x570
[c0000000f600b950] [c0000000004d3cb8] .close_ctree+0x2e8/0x3b0
[c0000000f600ba20] [c00000000049d178] .btrfs_put_super+0x18/0x30
[c0000000f600ba90] [c000000000243cd4] .generic_shutdown_super+0xa4/0x1a0
[c0000000f600bb10] [c0000000002441d8] .kill_anon_super+0x18/0x30
[c0000000f600bb90] [c00000000049c898] .btrfs_kill_super+0x18/0xc0
[c0000000f600bc10] [c0000000002444f8] .deactivate_locked_super+0x98/0xe0
[c0000000f600bc90] [c000000000269f94] .cleanup_mnt+0x54/0xa0
[c0000000f600bd10] [c0000000000bd744] .task_work_run+0xc4/0x100
[c0000000f600bdb0] [c000000000016334] .do_notify_resume+0x74/0x80
[c0000000f600be30] [c0000000000098b8] .ret_from_except_lite+0x64/0x68
Instruction dump:
fba1ffe8 fbc1fff0 fbe1fff8 7c791b78 f8010010 f821ff21 e94d0290 81030040
812a04e8 7d094a78 7d290034 5529d97e <0b0900003b400000 3be30050 3bc3004c

The above call trace is seen even on x86_64; albeit very rarely and that too
with nodesize set to 64k and with nospace_cache mount option being used.

The reason for the above call trace is,
btrfs_remove_chunk
  check_system_chunk
    Allocate chunk if required
  For each physical stripe on underlying device,
    btrfs_free_dev_extent
      ...
      Take lock on Device tree's root node
      btrfs_cow_block("dev tree's root node");
        btrfs_reserve_extent
          find_free_extent
    index = BTRFS_RAID_DUP;
    have_caching_bg = false;

            When in LOOP_CACHING_NOWAIT state, Assume we find a block group
    which is being cached; Hence have_caching_bg is set to true

            When repeating the search for the next RAID index, we set
    have_caching_bg to false.

Hence right after completing the LOOP_CACHING_NOWAIT state, we incorrectly
skip LOOP_CACHING_WAIT state and move to LOOP_ALLOC_CHUNK state where we
allocate a chunk and try to add entries corresponding to the chunk's physical
stripe into the device tree. When doing so the task deadlocks itself waiting
for the blocking lock on the root node of the device tree.

This commit fixes the issue by introducing a new local variable to help
indicate as to whether a block group of any RAID type is being cached.

Signed-off-by: Chandan Rajendra <chandan@linux.vnet.ibm.com>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/extent-tree.c

index f50c7c2e2b59d2fb7d5110b5248f8fa01ee1eabf..99a8e57da8a11fbbdcbc0d833a687d79629200cc 100644 (file)
@@ -7029,6 +7029,7 @@ static noinline int find_free_extent(struct btrfs_root *orig_root,
        bool failed_alloc = false;
        bool use_cluster = true;
        bool have_caching_bg = false;
+       bool orig_have_caching_bg = false;
        bool full_search = false;
 
        WARN_ON(num_bytes < root->sectorsize);
@@ -7378,6 +7379,10 @@ loop:
        }
        up_read(&space_info->groups_sem);
 
+       if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
+               && !orig_have_caching_bg)
+               orig_have_caching_bg = true;
+
        if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
                goto search;
 
@@ -7400,7 +7405,7 @@ loop:
                         * don't have any unached bgs and we've alrelady done a
                         * full search through.
                         */
-                       if (have_caching_bg || !full_search)
+                       if (orig_have_caching_bg || !full_search)
                                loop = LOOP_CACHING_WAIT;
                        else
                                loop = LOOP_ALLOC_CHUNK;