On Tue, Apr 04 2017, Ming Lei wrote:

> On Mon, Apr 3, 2017 at 9:18 AM, NeilBrown <neilb@suse.com> wrote:
>>
>> When a filesystem is mounted from a loop device, writes are
>> throttled by balance_dirty_pages() twice: once when writing
>> to the filesystem and once when the loop_handle_cmd() writes
>> to the backing file.  This double-throttling can trigger
>> positive feedback loops that create significant delays.  The
>> throttling at the lower level is seen by the upper level as
>> a slow device, so it throttles extra hard.
>>
>> The PF_LESS_THROTTLE flag was created to handle exactly this
>> circumstance, though with an NFS filesystem mounted from a
>> local NFS server.  It reduces the throttling on the lower
>> layer so that it can proceed largely unthrottled.
>>
>> To demonstrate this, create a filesystem on a loop device
>> and write (e.g. with dd) several large files which combine
>> to consume significantly more than the limit set by
>> /proc/sys/vm/dirty_ratio or dirty_bytes.  Measure the total
>> time taken.
>>
>> When I do this directly on a device (no loop device) the
>> total time for several runs (mkfs, mount, write 200 files,
>> umount) is fairly stable: 28-35 seconds.
>> When I do this over a loop device the times are much worse
>> and less stable.  52-460 seconds.  Half below 100seconds,
>> half above.
>> When I apply this patch, the times become stable again,
>> though not as fast as the no-loop-back case: 53-72 seconds.
>>
>> There may be room for further improvement as the total overhead still
>> seems too high, but this is a big improvement.
>>
>> Signed-off-by: NeilBrown <neilb@suse.com>
>> ---
>>  drivers/block/loop.c | 3 +++
>>  1 file changed, 3 insertions(+)
>>
>> diff --git a/drivers/block/loop.c b/drivers/block/loop.c
>> index 0ecb6461ed81..a7e1dd215fc2 100644
>> --- a/drivers/block/loop.c
>> +++ b/drivers/block/loop.c
>> @@ -1694,8 +1694,11 @@ static void loop_queue_work(struct kthread_work *work)
>>  {
>>         struct loop_cmd *cmd =
>>                 container_of(work, struct loop_cmd, work);
>> +       int oldflags = current->flags & PF_LESS_THROTTLE;
>>
>> +       current->flags |= PF_LESS_THROTTLE;
>>         loop_handle_cmd(cmd);
>> +       current->flags = (current->flags & ~PF_LESS_THROTTLE) | oldflags;
>>  }
>
> You can do it against 'lo->worker_task' instead of doing it in each
> loop_queue_work(),
> and this flag needn't to be restored because the kernel thread is loop
> specialized.
>

good point.  I'll do that.  Thanks,
NeilBrown