Hi folks, fstests running shared/032 on XFS with a default mkfs and mount config causes a panic in the block layer when userspace is operating directly on the block device like this: SECTION -- xfs FSTYP -- xfs (debug) PLATFORM -- Linux/x86_64 test3 6.3.0-dgc+ #1792 SMP PREEMPT_DYNAMIC Wed May 3 15:20:20 AEST 2023 MKFS_OPTIONS -- -f -m rmapbt=1 /dev/pmem1 MOUNT_OPTIONS -- -o dax=never -o context=system_u:object_r:root_t:s0 /dev/pmem1 /mnt/scratch .... [ 56.070695] run fstests shared/032 at 2023-05-03 15:21:55 [ 56.768890] BTRFS: device fsid 355df15c-7bc5-49b0-9b5d-dc25ce855a9d devid 1 transid 6 /dev/pmem1 scanned by mkfs.btrfs (5836) [ 57.285879] pmem1: p1 [ 57.301845] BUG: kernel NULL pointer dereference, address: 00000000000000a8 [ 57.304562] #PF: supervisor read access in kernel mode [ 57.306499] #PF: error_code(0x0000) - not-present page [ 57.308414] PGD 0 P4D 0 [ 57.309401] Oops: 0000 [#1] PREEMPT SMP [ 57.310876] CPU: 3 PID: 4478 Comm: (udev-worker) Not tainted 6.3.0-dgc+ #1792 [ 57.313517] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.2-debian-1.16.2-1 04/01/2014 [ 57.317089] RIP: 0010:blk_mq_sched_bio_merge+0x7b/0x100 [ 57.319059] Code: c0 49 8b 5c 24 38 48 03 1c c5 c0 69 71 82 b8 02 00 00 00 f7 c2 00 00 40 00 75 07 31 c0 84 d2 0f 94 c0 48 8b 94 c3 90 00 00 00 <f6> 82 a8 0d [ 57.325898] RSP: 0018:ffffc900042ab880 EFLAGS: 00010246 [ 57.327835] RAX: 0000000000000001 RBX: ffff888237d80000 RCX: 0000000000000000 [ 57.330492] RDX: 0000000000000000 RSI: ffff88810135e000 RDI: ffff8885c1928000 [ 57.333118] RBP: ffffc900042ab8b0 R08: 0000000000001000 R09: 0000000000000001 [ 57.335791] R10: ffff8885c1928000 R11: 0000000000000008 R12: ffff8885c1928000 [ 57.338298] R13: ffff88810135e000 R14: 0000000000000001 R15: ffff88810135e000 [ 57.340092] FS: 00007f4adf5438c0(0000) GS:ffff888237d80000(0000) knlGS:0000000000000000 [ 57.342132] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 57.343536] CR2: 00000000000000a8 CR3: 0000000102e4a005 CR4: 0000000000060ee0 [ 57.345150] Call Trace: [ 57.345732] <TASK> [ 57.346265] ? mempool_alloc_slab+0x15/0x20 [ 57.347242] blk_mq_attempt_bio_merge+0x4e/0x50 [ 57.348288] blk_mq_submit_bio+0x232/0x580 [ 57.349230] __submit_bio+0x1e/0x110 [ 57.350066] submit_bio_noacct_nocheck+0x24a/0x330 [ 57.351161] submit_bio_noacct+0x196/0x490 [ 57.352104] submit_bio+0x43/0x60 [ 57.352879] mpage_readahead+0xf4/0x130 [ 57.353762] ? blkdev_write_begin+0x30/0x30 [ 57.354741] blkdev_readahead+0x15/0x20 [ 57.355719] read_pages+0x5c/0x230 [ 57.356625] page_cache_ra_unbounded+0x148/0x190 [ 57.357764] force_page_cache_ra+0x9a/0xc0 [ 57.358845] page_cache_sync_ra+0x2e/0x50 [ 57.359839] filemap_get_pages+0x10f/0x670 [ 57.360870] ? walk_component+0xc7/0x170 [ 57.361859] filemap_read+0xed/0x380 [ 57.362819] ? __ia32_compat_sys_lookup_dcookie+0x410/0xe80 [ 57.364086] ? __might_fault+0x22/0x30 [ 57.364953] blkdev_read_iter+0xe3/0x1e0 [ 57.365871] vfs_read+0x213/0x2e0 [ 57.366646] ksys_read+0x71/0xf0 [ 57.367396] __x64_sys_read+0x19/0x20 [ 57.368244] do_syscall_64+0x34/0x80 [ 57.369064] entry_SYSCALL_64_after_hwframe+0x63/0xcd [ 57.370225] RIP: 0033:0x7f4adf7c503d [ 57.371054] Code: 31 c0 e9 c6 fe ff ff 50 48 8d 3d a6 55 0a 00 e8 39 fe 01 00 66 0f 1f 84 00 00 00 00 00 80 3d a1 25 0e 00 00 74 17 31 c0 0f 05 <48> 3d 00 fc [ 57.375229] RSP: 002b:00007ffd91350748 EFLAGS: 00000246 ORIG_RAX: 0000000000000000 [ 57.376933] RAX: ffffffffffffffda RBX: 000055c11070df70 RCX: 00007f4adf7c503d [ 57.378565] RDX: 0000000000000040 RSI: 000055c1105c48f8 RDI: 000000000000000f [ 57.380177] RBP: 00000001fffe0000 R08: 000055c1106f41c0 R09: 00007f4adf8a02e0 [ 57.381790] R10: 0000000000000000 R11: 0000000000000246 R12: 000055c1105c48d0 [ 57.383406] R13: 0000000000000040 R14: 000055c11070dfc8 R15: 000055c1105c48e8 [ 57.385026] </TASK> shared/032 is entirely a mkfs test - it is checking that mkfs for the filesystem under test recognises other filesystem signatures on the block device and does not overwrite them by accident. This test does not involve kernel filesystem code at all. Somewhere in amongst the running of various mkfs operations during the test, a udev-worker is triggered and is doing something with the block device (probing it?) and that results in the above oops occurring. This is only happening on the machine I have configured with (fake) PMEM devices - virtio, sd and nvme devices do not appear to be triggering this. That kinda implies a timing issue - pmem completes IO synchronously, all the the others use async completion... -Dave. -- Dave Chinner david@xxxxxxxxxxxxx