| Message ID | 20210112155502.426331-1-fengli@smartx.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v2] blk: avoid divide-by-zero with zero granularity | expand |
Li,
> If the physical_block_size and io_min is less than a sector,
That's not supposed to happen. What device/driver is this?
Hi Martin, I use the nvme-tcp as the host, the target is spdk nvme-tcp target, and set a wrong block size(i.g. bs=8), then the host prints this oops: [ 63.153018] nvme nvme0: creating 3 I/O queues. [ 63.181644] nvme nvme0: mapped 3/0/0 default/read/poll queues. [ 63.185568] nvme nvme0: new ctrl: NQN "nqn.2018-11.io.spdk:nvmf-test", addr 192.168.64.217:4421 [ 63.189440] divide error: 0000 [#1] SMP NOPTI [ 63.190963] CPU: 0 PID: 122 Comm: kworker/u8:2 Not tainted 5.9.9-200.fc33.x86_64 #1 [ 63.193426] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-1.fc33 04/01/2014 [ 63.196263] Workqueue: nvme-wq nvme_scan_work [nvme_core] [ 63.198015] RIP: 0010:blk_stack_limits+0x189/0x440 [ 63.199549] Code: 43 28 0f b6 45 62 08 43 62 8b 4d 2c 39 4d 38 0f 43 4d 38 31 d2 45 31 f6 48 8b 04 24 8b 75 34 89 cf 44 8b 43 34 c1 ef 09 09 [ 63.203169] RSP: 0018:ffffb312001ffc98 EFLAGS: 00010202 [ 63.204034] RAX: 0000000000000000 RBX: ffff9463b270ecc8 RCX: 0000000000000008 [ 63.205205] RDX: 0000000000000000 RSI: 0000000000000008 RDI: 0000000000000000 [ 63.206377] RBP: ffff9463b27083f8 R08: 0000000000000000 R09: 0000000000000008 [ 63.207553] R10: ffff9463b81d6420 R11: 0000035f00000000 R12: ffff9463b2bea000 [ 63.208725] R13: ffff9463b6d31258 R14: 0000000000000000 R15: ffff9463b6c22800 [ 63.209914] FS: 0000000000000000(0000) GS:ffff9463bec00000(0000) knlGS:0000000000000000 [ 63.211239] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 63.212186] CR2: 000055d73270e690 CR3: 0000000036b32000 CR4: 00000000000006f0 [ 63.213383] Call Trace: [ 63.213826] ? blk_mq_freeze_queue_wait+0x13/0x80 [ 63.214609] __nvme_revalidate_disk+0x191/0x2d0 [nvme_core] [ 63.215528] nvme_validate_ns+0x3fa/0x960 [nvme_core] [ 63.216374] nvme_scan_work+0x165/0x3c0 [nvme_core] [ 63.217183] process_one_work+0x1b4/0x370 [ 63.217852] worker_thread+0x53/0x3e0 [ 63.218458] ? process_one_work+0x370/0x370 [ 63.219156] kthread+0x11b/0x140 [ 63.219700] ? __kthread_bind_mask+0x60/0x60 [ 63.220403] ret_from_fork+0x22/0x30 Martin K. Petersen <martin.petersen@oracle.com> 于2021年1月13日周三 上午1:10写道： > > > Li, > > > If the physical_block_size and io_min is less than a sector, > > That's not supposed to happen. What device/driver is this? > > -- > Martin K. Petersen Oracle Linux Engineering
On 12/01/2021 18:29, Feng Li wrote: > I use the nvme-tcp as the host, the target is spdk nvme-tcp target, > and set a wrong block size(i.g. bs=8), then the host prints this oops: I think the better fix here is to reject devices which report a block size small than a sector.
Johannes, >> I use the nvme-tcp as the host, the target is spdk nvme-tcp target, >> and set a wrong block size(i.g. bs=8), then the host prints this oops: > > I think the better fix here is to reject devices which report a block size > small than a sector. Yep, Linux doesn't support logical block sizes < 512 bytes. Also, the NVMe spec states: "A value smaller than 9 (i.e., 512 bytes) is not supported."
Yes, Reject the device is the right fix. I will try to send another fix. By the way, I think this fix is good protection, maybe some other devices violate this block size constraint. Divide zero is unacceptable. Thanks, Feng Li Martin K. Petersen <martin.petersen@oracle.com> 于2021年1月13日周三 上午1:48写道： > > > Johannes, > > >> I use the nvme-tcp as the host, the target is spdk nvme-tcp target, > >> and set a wrong block size(i.g. bs=8), then the host prints this oops: > > > > I think the better fix here is to reject devices which report a block size > > small than a sector. > > Yep, Linux doesn't support logical block sizes < 512 bytes. > > Also, the NVMe spec states: > > "A value smaller than 9 (i.e., 512 bytes) is not supported." > > -- > Martin K. Petersen Oracle Linux Engineering
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h index f94ee3089e01..ffffb04ad113 100644 --- a/include/linux/blkdev.h +++ b/include/linux/blkdev.h @@ -1485,7 +1485,11 @@ static inline int queue_alignment_offset(const struct request_queue *q) static inline int queue_limit_alignment_offset(struct queue_limits *lim, sector_t sector) { unsigned int granularity = max(lim->physical_block_size, lim->io_min); - unsigned int alignment = sector_div(sector, granularity >> SECTOR_SHIFT) + unsigned int alignment; + if (granularity >> SECTOR_SHIFT == 0) + return 0; + + alignment = sector_div(sector, granularity >> SECTOR_SHIFT) << SECTOR_SHIFT; return (granularity + lim->alignment_offset - alignment) % granularity;
If the physical_block_size and io_min is less than a sector, the 'granularity >> SECTOR_SHIFT' will be zero. Signed-off-by: Li Feng <fengli@smartx.com> --- include/linux/blkdev.h | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)