[PATCH AUTOSEL 5.4 156/350] md/bitmap: avoid race window between md_bitmap_resize and bitmap_file_clear_bit

From: Sasha Levin
Date: Tue Dec 10 2019 - 16:29:30 EST


From: Guoqing Jiang <guoqing.jiang@xxxxxxxxxxxxxxx>

[ Upstream commit fadcbd2901a0f7c8721f3bdb69eac95c272dc8ed ]

We need to move "spin_lock_irq(&bitmap->counts.lock)" before unmap previous
storage, otherwise panic like belows could happen as follows.

[ 902.353802] sdl: detected capacity change from 1077936128 to 3221225472
[ 902.616948] general protection fault: 0000 [#1] SMP
[snip]
[ 902.618588] CPU: 12 PID: 33698 Comm: md0_raid1 Tainted: G O 4.14.144-1-pserver #4.14.144-1.1~deb10
[ 902.618870] Hardware name: Supermicro SBA-7142G-T4/BHQGE, BIOS 3.00 10/24/2012
[ 902.619120] task: ffff9ae1860fc600 task.stack: ffffb52e4c704000
[ 902.619301] RIP: 0010:bitmap_file_clear_bit+0x90/0xd0 [md_mod]
[ 902.619464] RSP: 0018:ffffb52e4c707d28 EFLAGS: 00010087
[ 902.619626] RAX: ffe8008b0d061000 RBX: ffff9ad078c87300 RCX: 0000000000000000
[ 902.619792] RDX: ffff9ad986341868 RSI: 0000000000000803 RDI: ffff9ad078c87300
[ 902.619986] RBP: ffff9ad0ed7a8000 R08: 0000000000000000 R09: 0000000000000000
[ 902.620154] R10: ffffb52e4c707ec0 R11: ffff9ad987d1ed44 R12: ffff9ad0ed7a8360
[ 902.620320] R13: 0000000000000003 R14: 0000000000060000 R15: 0000000000000800
[ 902.620487] FS: 0000000000000000(0000) GS:ffff9ad987d00000(0000) knlGS:0000000000000000
[ 902.620738] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 902.620901] CR2: 000055ff12aecec0 CR3: 0000001005207000 CR4: 00000000000406e0
[ 902.621068] Call Trace:
[ 902.621256] bitmap_daemon_work+0x2dd/0x360 [md_mod]
[ 902.621429] ? find_pers+0x70/0x70 [md_mod]
[ 902.621597] md_check_recovery+0x51/0x540 [md_mod]
[ 902.621762] raid1d+0x5c/0xeb0 [raid1]
[ 902.621939] ? try_to_del_timer_sync+0x4d/0x80
[ 902.622102] ? del_timer_sync+0x35/0x40
[ 902.622265] ? schedule_timeout+0x177/0x360
[ 902.622453] ? call_timer_fn+0x130/0x130
[ 902.622623] ? find_pers+0x70/0x70 [md_mod]
[ 902.622794] ? md_thread+0x94/0x150 [md_mod]
[ 902.622959] md_thread+0x94/0x150 [md_mod]
[ 902.623121] ? wait_woken+0x80/0x80
[ 902.623280] kthread+0x119/0x130
[ 902.623437] ? kthread_create_on_node+0x60/0x60
[ 902.623600] ret_from_fork+0x22/0x40
[ 902.624225] RIP: bitmap_file_clear_bit+0x90/0xd0 [md_mod] RSP: ffffb52e4c707d28

Because mdadm was running on another cpu to do resize, so bitmap_resize was
called to replace bitmap as below shows.

PID: 38801 TASK: ffff9ad074a90e00 CPU: 0 COMMAND: "mdadm"
[exception RIP: queued_spin_lock_slowpath+56]
[snip]
-- <NMI exception stack> --
#5 [ffffb52e60f17c58] queued_spin_lock_slowpath at ffffffff9c0b27b8
#6 [ffffb52e60f17c58] bitmap_resize at ffffffffc0399877 [md_mod]
#7 [ffffb52e60f17d30] raid1_resize at ffffffffc0285bf9 [raid1]
#8 [ffffb52e60f17d50] update_size at ffffffffc038a31a [md_mod]
#9 [ffffb52e60f17d70] md_ioctl at ffffffffc0395ca4 [md_mod]

And the procedure to keep resize bitmap safe is allocate new storage
space, then quiesce, copy bits, replace bitmap, and re-start.

However the daemon (bitmap_daemon_work) could happen even the array is
quiesced, which means when bitmap_file_clear_bit is triggered by raid1d,
then it thinks it should be fine to access store->filemap since
counts->lock is held, but resize could change the storage without the
protection of the lock.

Cc: Jack Wang <jinpu.wang@xxxxxxxxxxxxxxx>
Cc: NeilBrown <neilb@xxxxxxxx>
Signed-off-by: Guoqing Jiang <guoqing.jiang@xxxxxxxxxxxxxxx>
Signed-off-by: Song Liu <songliubraving@xxxxxx>
Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx>
---
drivers/md/md-bitmap.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/md/md-bitmap.c b/drivers/md/md-bitmap.c
index b092c7b5282f9..3ad18246fcb3c 100644
--- a/drivers/md/md-bitmap.c
+++ b/drivers/md/md-bitmap.c
@@ -2139,6 +2139,7 @@ int md_bitmap_resize(struct bitmap *bitmap, sector_t blocks,
memcpy(page_address(store.sb_page),
page_address(bitmap->storage.sb_page),
sizeof(bitmap_super_t));
+ spin_lock_irq(&bitmap->counts.lock);
md_bitmap_file_unmap(&bitmap->storage);
bitmap->storage = store;

@@ -2154,7 +2155,6 @@ int md_bitmap_resize(struct bitmap *bitmap, sector_t blocks,
blocks = min(old_counts.chunks << old_counts.chunkshift,
chunks << chunkshift);

- spin_lock_irq(&bitmap->counts.lock);
/* For cluster raid, need to pre-allocate bitmap */
if (mddev_is_clustered(bitmap->mddev)) {
unsigned long page;
--
2.20.1