Re: [RFC PATCH v2 2/2] blk-mq: Lockout tagset iter when freeing rqs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 23/12/2020 15:47, Bart Van Assche wrote:
On 12/23/20 3:40 AM, John Garry wrote:
Sorry, I got the 2x iter functions mixed up.

So if we use mutex to solve blk_mq_queue_tag_busy_iter() problem, then we
still have this issue in blk_mq_tagset_busy_iter() which I report previously
[0]:

[  319.771745] BUG: KASAN: use-after-free in bt_tags_iter+0xe0/0x128
[  319.777832] Read of size 4 at addr ffff0010b6bd27cc by task more/1866
[  319.784262]
[  319.785753] CPU: 61 PID: 1866 Comm: more Tainted: G        W
5.10.0-rc4-18118-gaa7b9c30d8ff #1070
[  319.795312] Hardware name: Huawei Taishan 2280 /D05, BIOS Hisilicon
D05 IT21 Nemo 2.0 RC0 04/18/2018
[  319.804437] Call trace:
[  319.806892]  dump_backtrace+0x0/0x2d0
[  319.810552]  show_stack+0x18/0x68
[  319.813865]  dump_stack+0x100/0x16c
[  319.817348]  print_address_description.constprop.12+0x6c/0x4e8
[  319.823176]  kasan_report+0x130/0x200
[  319.826831]  __asan_load4+0x9c/0xd8
[  319.830315]  bt_tags_iter+0xe0/0x128
[  319.833884]  __blk_mq_all_tag_iter+0x320/0x3a8
[  319.838320]  blk_mq_tagset_busy_iter+0x8c/0xd8
[  319.842760]  scsi_host_busy+0x88/0xb8
[  319.846418]  show_host_busy+0x1c/0x48
[  319.850079]  dev_attr_show+0x44/0x90
[  319.853655]  sysfs_kf_seq_show+0x128/0x1c8
[  319.857744]  kernfs_seq_show+0xa0/0xb8
[  319.861489]  seq_read_iter+0x1ec/0x6a0
[  319.865230]  seq_read+0x1d0/0x250
[  319.868539]  kernfs_fop_read+0x70/0x330
[  319.872369]  vfs_read+0xe4/0x250
[  319.875590]  ksys_read+0xc8/0x178
[  319.878898]  __arm64_sys_read+0x44/0x58
[  319.882730]  el0_svc_common.constprop.2+0xc4/0x1e8
[  319.887515]  do_el0_svc+0x90/0xa0
[  319.890824]  el0_sync_handler+0x128/0x178
[  319.894825]  el0_sync+0x158/0x180
[  319.898131]
[  319.899614] The buggy address belongs to the page:
[  319.904403] page:000000004e9e6864 refcount:0 mapcount:0
mapping:0000000000000000 index:0x0 pfn:0x10b6bd2
[  319.913876] flags: 0xbfffc0000000000()
[  319.917626] raw: 0bfffc0000000000 0000000000000000 fffffe0000000000
0000000000000000
[  319.925363] raw: 0000000000000000 0000000000000000 00000000ffffffff
0000000000000000
[  319.933096] page dumped because: kasan: bad access detected
[  319.938658]
[  319.940141] Memory state around the buggy address:
[  319.944925]  ffff0010b6bd2680: ff ff ff ff ff ff ff ff ff ff ff ff ff
ff ff ff
[  319.952139]  ffff0010b6bd2700: ff ff ff ff ff ff ff ff ff ff ff ff ff
ff ff ff
[  319.959354] >ffff0010b6bd2780: ff ff ff ff ff ff ff ff ff ff ff ff ff
ff ff ff
[  319.966566] ^
[  319.972131]  ffff0010b6bd2800: ff ff ff ff ff ff ff ff ff ff ff ff ff
ff ff ff
[  319.979344]  ffff0010b6bd2880: ff ff ff ff ff ff ff ff ff ff ff ff ff
ff ff ff
[  319.986557]
==================================================================
[  319.993770] Disabling lock debugging due to kernel taint

So to trigger this, I start fio on a disk, and then have one script
which constantly enables and disables an IO scheduler for that disk, and
another script which constantly reads /sys/class/scsi_host/host0/host_busy .

And in this problem, the driver tag we iterate may point to a stale IO sched
request.

Hi John,

Hi Bart,

Sorry for the slow reply, but I was on vacation since before you sent this mail.


I propose to change the order in which blk_mq_sched_free_requests(q) and
blk_mq_debugfs_unregister(q) are called. Today blk_mq_sched_free_requests(q)
is called by blk_cleanup_queue() before blk_put_queue() is called.
blk_put_queue() calls blk_release_queue() if the last reference is dropped.
blk_release_queue() calls blk_mq_debugfs_unregister(). I prefer removing the
debugfs attributes earlier over modifying the tag iteration functions
because I think removing the debugfs attributes earlier is less risky.

But don't we already have this following path to remove the per-hctx debugfs dir earlier than blk_mq_sched_free_requests() or blk_release_queue():

blk_cleanup_queue() -> blk_mq_exit_queue() -> blk_mq_exit_hw_queues() -> blk_mq_debugfs_unregister_hctx() -> blk_mq_debugfs_unregister_hctx(hctx->debugfs_dir)

Having said that, I am not sure how this is related directly to the problem I mentioned. In that problem, above, we trigger the blk_mq_tagset_busy_iter() from the SCSI host sysfs file, and the use-after-free comes about from disabling the elevator (and freeing the sched requests) in parallel.

Although this will make it harder to debug lockups that happen while
removing a request queue, kernel developers who are analyzing such an issue
can undo this change in their development kernel tree.


Thanks,
John




[Index of Archives]     [Linux RAID]     [Linux SCSI]     [Linux ATA RAID]     [IDE]     [Linux Wireless]     [Linux Kernel]     [ATH6KL]     [Linux Bluetooth]     [Linux Netdev]     [Kernel Newbies]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Device Mapper]

  Powered by Linux