Re: Hung tasks with multiple partitions

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

 



On 1/30/20 11:34 AM, Salman Qazi wrote:
I am writing on behalf of the Chromium OS team at Google.  We found
the root cause for some hung tasks we were experiencing and we would
like to get your opinion on potential solutions.  The bugs were
encountered on 4.19 kernel.
However my reading of the code suggests that the relevant portions of the
code have not changed since then.

We have an eMMC flash drive that has been carved into partitions on an
8 CPU system.  The repro case that we came up with, is to use 8
threaded fio write-mostly workload against one partition, let the
system use the other partition as the read-write filesystem (i.e. just
background activity) and then run the following loop:

while true; do sync; sleep 1 ; done

The hung task stack traces look like the following:

[  128.994891] jbd2/dm-1-8     D    0   367      2 0x00000028
last_sleep: 96340206998.  last_runnable: 96340140151
[  128.994898] Call trace:
[  128.994903]  __switch_to+0x120/0x13c
[  128.994909]  __schedule+0x60c/0x7dc
[  128.994914]  schedule+0x74/0x94
[  128.994919]  io_schedule+0x1c/0x40
[  128.994925]  bit_wait_io+0x18/0x58
[  128.994930]  __wait_on_bit+0x78/0xdc
[  128.994935]  out_of_line_wait_on_bit+0xa0/0xcc
[  128.994943]  __wait_on_buffer+0x48/0x54
[  128.994948]  jbd2_journal_commit_transaction+0x1198/0x1a4c
[  128.994956]  kjournald2+0x19c/0x268
[  128.994961]  kthread+0x120/0x130
[  128.994967]  ret_from_fork+0x10/0x18

I added some more information to trace points to understand what was
going on.  It turns out that blk_mq_sched_dispatch_requests had
checked hctx->dispatch, found it empty, and then began consuming
requests from the io scheduler (in blk_mq_do_dispatch_sched).
Unfortunately, the deluge from the I/O scheduler (BFQ in our case)
doesn't stop for 30 seconds and there is no mechanism present in
blk_mq_do_dispatch_sched to terminate early or reconsider
hctx->dispatch contents.  In the meantime, a flush command arrives in
hctx->dispatch (via insertion in  blk_mq_sched_bypass_insert) and
languishes there.  Eventually the thread waiting on the flush triggers
the hung task watchdog.

The solution that comes to mind is to periodically check
hctx->dispatch in blk_mq_do_dispatch_sched and exit early if it is
non-empty.  However, not being an expert in this subsystem, I am not
sure if there would be other consequences.

The call stack shown in your e-mail usually means that an I/O request got stuck. How about determining first whether this is caused by the BFQ scheduler or by the eMMC driver? I think the developers of these software components need that information anyway before they can step in.

The attached script may help to identify which requests got stuck.

Bart.
#!/bin/bash

show_state() {
    local a dev=$1

    for a in device/state queue/scheduler; do
	[ -e "$dev/$a" ] && grep -aH . "$dev/$a"
    done
}

if [ -e /sys/kernel/debug/block ]; then
    devs=($(cd /sys/kernel/debug/block && echo ./*))
else
    devs=($(cd /sys/class/block && echo ./*))
fi

cd /sys/class/block || exit $?
for dev in "${devs[@]}"; do
    dev="${dev#./}"
    echo "$dev"
    pending=0
    if [ -e "$dev/mq" ]; then
	for f in "$dev"/mq/*/{pending,*/rq_list}; do
	    [ -e "$f" ] || continue
	    if { read -r line1 && read -r line2; } <"$f"; then
		echo "$f"
		echo "$line1 $line2" >/dev/null
		head -n 9 "$f"
		((pending++))
	    fi
	done
    fi
    (
	busy=0
	cd /sys/kernel/debug/block >&/dev/null &&
	    { grep -aH . $dev/requeue_list; true; } &&
	    for d in "$dev"/mq/hctx* "$dev"/hctx*; do
		[ ! -d "$d" ] && continue
		{ [ ! -e "$d/tags" ] ||
		      grep -q '^busy=0$' "$d/tags"; } &&
		    { [ ! -e "$d/sched_tags" ] ||
			  [ "$(<"$d/sched_tags")" = "" ] ||
			  grep -q '^busy=0$' "$d/sched_tags"; } && continue
		((busy++))
	        for f in "$d"/{active,busy,dispatch,flags,requeue_list,sched_tags,state,tags*,cpu*/rq_list,sched/*rqs}; do
		    [ -e "$f" ] && grep -aH . "$f"
		done
	    done
	exit $busy
    )
    pending=$((pending+$?))
    if [ "$pending" -gt 0 ]; then
	(
	    cd /sys/kernel/debug/block >&/dev/null &&
		if [ -e "$dev/mq/state" ]; then
		    grep -aH . "$dev/mq/state"
		else
		    grep -aH . "$dev/state"
		fi
	)
	show_state "$dev"
    fi
done

[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