btrfs BUG() in __set_extent_bit on GFP_ATOMIC failure.
While playing with fault injection, I hit this quite easily.
kernel BUG at fs/btrfs/extent_io.c:990! invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
CPU: 1 PID: 1270 Comm: fsx Not tainted 3.16.0+ #41 task: ffff88023fe46d60 ti: ffff8802405a8000 task.ti: ffff8802405a8000 RIP: 0010:[<ffffffffc0319af4>] [<ffffffffc0319af4>] __set_extent_bit+0x574/0x660 [btrfs] ... [<ffffffff851be7fc>] ? set_track+0x9c/0x140 [<ffffffffc031aa94>] lock_extent_bits+0x94/0x310 [btrfs] [<ffffffff85166454>] ? pagecache_get_page+0xb4/0x210 [<ffffffffc030d4de>] lock_and_cleanup_extent_if_need+0xee/0x1f0 [btrfs] [<ffffffffc030e6e1>] __btrfs_buffered_write+0x1b1/0x680 [btrfs] [<ffffffff850a258b>] ? preempt_count_sub+0xab/0x100 [<ffffffffc030ed2e>] btrfs_file_write_iter+0x17e/0x570 [btrfs] [<ffffffff851d78ce>] new_sync_write+0x8e/0xd0 [<ffffffff851d8127>] vfs_write+0xb7/0x1f0 [<ffffffff851d8d68>] SyS_write+0x58/0xd0 [<ffffffff8576371f>] tracesys+0xdd/0xe2
989 prealloc = alloc_extent_state_atomic(prealloc);
990 BUG_ON(!prealloc);
541 static struct extent_state *
542 alloc_extent_state_atomic(struct extent_state *prealloc) 543 {
544 if (!prealloc)
545 prealloc = alloc_extent_state(GFP_ATOMIC);
546
547 return prealloc;
548 }
Going BUG() on a GFP_ATOMIC allocation failure seems a bit excessive. Surely there's something better we can do here ?
Dave
Quoted 34 lines · Dave Jones
While playing with fault injection, I hit this quite easily.
kernel BUG at
fs/btrfs/extent_io.c:990! invalid opcode: 0000 [#1] PREEMPT SMPDEBUG_PAGEALLOCCPU: 1 PID: 1270 Comm: fsx Not tainted 3.16.0+ #41 task: ffff88023fe46d60 ti: ffff8802405a8000 task.ti: ffff8802405a8000 RIP: 0010:[<ffffffffc0319af4>] [<ffffffffc0319af4>] __set_extent_bit+0x574/0x660 [btrfs] ... [<ffffffff851be7fc>] ? set_track+0x9c/0x140 [<ffffffffc031aa94>] lock_extent_bits+0x94/0x310 [btrfs] [<ffffffff85166454>] ? pagecache_get_page+0xb4/0x210 [<ffffffffc030d4de>] lock_and_cleanup_extent_if_need+0xee/0x1f0 [btrfs] [<ffffffffc030e6e1>] __btrfs_buffered_write+0x1b1/0x680 [btrfs] [<ffffffff850a258b>] ? preempt_count_sub+0xab/0x100 [<ffffffffc030ed2e>] btrfs_file_write_iter+0x17e/0x570 [btrfs] [<ffffffff851d78ce>] new_sync_write+0x8e/0xd0 [<ffffffff851d8127>] vfs_write+0xb7/0x1f0 [<ffffffff851d8d68>] SyS_write+0x58/0xd0 [<ffffffff8576371f>] tracesys+0xdd/0xe2989 prealloc =
alloc_extent_state_atomic(prealloc);
990BUG_ON(!prealloc);541 static struct
extent_state*
542alloc_extent_state_atomic(struct extent_state *prealloc)543 {
544 if (!prealloc)
545 prealloc =alloc_extent_state(GFP_ATOMIC);
546
547 return prealloc;
548 }
Quoted 3 lines
Going
BUG()on aGFP_ATOMICallocation failure seems a bit excessive. Surely there's something better we can do here ?
Ugh, yes we can. It should jump back to outside the lock and get the prealloc there.
-chris