On Tue, Jun 5, 2018 at 8:41 AM, Ming Lei <ming.lei@xxxxxxxxxx> wrote: > On Tue, Jun 05, 2018 at 09:27:41AM +0900, Tetsuo Handa wrote: >> Jens Axboe wrote: >> > On 6/1/18 4:10 AM, Tetsuo Handa wrote: >> > > Tetsuo Handa wrote: >> > >> Since sum of percpu_count did not change after percpu_ref_kill(), this is >> > >> not a race condition while folding percpu counter values into atomic counter >> > >> value. That is, for some reason, someone who is responsible for calling >> > >> percpu_ref_put(&q->q_usage_counter) (presumably via blk_queue_exit()) is >> > >> unable to call percpu_ref_put(). >> > >> But I don't know how to find someone who is failing to call percpu_ref_put()... >> > > >> > > I found the someone. It was already there in the backtrace... >> > > >> > >> > Ahh, nicely spotted! One idea would be the one below. For this case, >> > we're recursing, so we can either do a non-block queue enter, or we >> > can just do a live enter. >> > >> >> While "block: don't use blocking queue entered for recursive bio submits" was >> already applied, syzbot is still reporting a hung task with same signature but >> different trace. >> >> https://syzkaller.appspot.com/text?tag=CrashLog&x=1432cedf800000 >> ---------------------------------------- >> [ 492.512243] INFO: task syz-executor1:20263 blocked for more than 120 seconds. >> [ 492.519604] Not tainted 4.17.0+ #83 >> [ 492.523793] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message. >> [ 492.531787] syz-executor1 D23384 20263 4574 0x00000004 >> [ 492.537443] Call Trace: >> [ 492.540041] __schedule+0x801/0x1e30 >> [ 492.580958] schedule+0xef/0x430 >> [ 492.610154] blk_queue_enter+0x8da/0xdf0 >> [ 492.716327] generic_make_request+0x651/0x1790 >> [ 492.765680] submit_bio+0xba/0x460 >> [ 492.793198] submit_bio_wait+0x134/0x1e0 >> [ 492.801891] blkdev_issue_flush+0x204/0x300 >> [ 492.806236] blkdev_fsync+0x93/0xd0 >> [ 492.813620] vfs_fsync_range+0x140/0x220 >> [ 492.817702] vfs_fsync+0x29/0x30 >> [ 492.821081] __loop_update_dio+0x4de/0x6a0 >> [ 492.825341] lo_ioctl+0xd28/0x2190 >> [ 492.833442] blkdev_ioctl+0x9b6/0x2020 >> [ 492.872146] block_ioctl+0xee/0x130 >> [ 492.880139] do_vfs_ioctl+0x1cf/0x16a0 >> [ 492.927550] ksys_ioctl+0xa9/0xd0 >> [ 492.931036] __x64_sys_ioctl+0x73/0xb0 >> [ 492.934952] do_syscall_64+0x1b1/0x800 >> [ 492.963624] entry_SYSCALL_64_after_hwframe+0x49/0xbe >> [ 493.212768] 1 lock held by syz-executor1/20263: >> [ 493.217448] #0: 00000000956bf5a3 (&lo->lo_ctl_mutex/1){+.+.}, at: lo_ioctl+0x8d/0x2190 >> ---------------------------------------- >> >> Is it OK to call [__]loop_update_dio() between blk_mq_freeze_queue() and >> blk_mq_unfreeze_queue(), for vfs_fsync() from __loop_update_dio() is calling >> blk_queue_enter() after blk_mq_freeze_queue() started blocking blk_queue_enter() >> by caling atomic_inc_return() and percpu_ref_kill() ? >> > > The vfs_fsync() isn't necessary in loop_update_dio() since both > generic_file_write_iter() and generic_file_read_iter() can handle > buffered io vs dio well. > > I will send one patch to remove the vfs_sync() later. Hi Tetsuo, The issue might be fixed by removing this vfs_sync(), but I'd like to understand the idea behind since vfs_sync() shouldn't have caused any IO to this loop queue. I also tried to do the test via the following c syzbot, but can't reproduce it yet after running it for several hours. https://syzkaller.appspot.com/x/repro.c?id=4727023951937536 Could you share us how you reproduce it? Thanks, Ming Lei
