On 2018/06/19 20:53, Dmitry Vyukov wrote: > On Tue, Jun 19, 2018 at 1:44 PM, Tetsuo Handa > <penguin-kernel@xxxxxxxxxxxxxxxxxxx> wrote: >> This bug report is getting no feedback, but I guess that this bug is in >> block or mm or locking layer rather than fs layer. >> >> NMI backtrace for this bug tends to report that sb_bread() from fill_super() >> from mount_bdev() is stalling is the cause of keep holding s_umount_key for >> more than 120 seconds. What is strange is that NMI backtrace for this bug tends >> to point at rcu_read_lock()/pagecache_get_page()/radix_tree_deref_slot()/ >> rcu_read_unlock() which is expected not to stall. >> >> Since CONFIG_RCU_CPU_STALL_TIMEOUT is set to 120 (and actually +5 due to >> CONFIG_PROVE_RCU=y) which is longer than CONFIG_DEFAULT_HUNG_TASK_TIMEOUT, >> maybe setting CONFIG_RCU_CPU_STALL_TIMEOUT to smaller values (e.g. 25) can >> give us some hints... > > If an rcu stall is the true root cause of this, then I guess would see > "rcu stall" bug too. Rcu stall is detected after 120 seconds, but task > hang after 120-240 seconds. So rcu stall has much higher chances to be > detected. Do you see the corresponding "rcu stall" bug? RCU stall is detected after 125 seconds due to CONFIG_PROVE_RCU=y (e.g. https://syzkaller.appspot.com/bug?id=1fac0fd91219f3f2a03d6fa7deafc95fbed79cc2 ). I didn't find the corresponding "rcu stall" bug. But it is not required that one RCU stall takes longer than 120 seconds. down(); // Will take 120 seconds due to multiple RCU stalls rcu_read_lock(): do_something(); rcu_read_unlock(): // Took 30 seconds for unknown reason. rcu_read_lock(): do_something(); rcu_read_unlock(): // Took 30 seconds for unknown reason. rcu_read_lock(): do_something(); rcu_read_unlock(): // Took 30 seconds for unknown reason. rcu_read_lock(): do_something(); rcu_read_unlock(): // Took 30 seconds for unknown reason. up();
