Re: loop_set_block_size: loop0 () has still dirty pages (nrpages=2)

[Ingo Franzki]

On 10.06.2021 16:45, Ming Lei wrote:
> On Tue, Jun 08, 2021 at 02:01:29PM +0200, Ingo Franzki wrote:
>> Hi all,
>>
>> we occasionally encounter a problem when setting up a loop device in one of our automated testcases.
>>
>> We set up a loop device as follows:
>>
>>     # dd if=/dev/zero of=/var/tmp/loopbackfile1.img bs=1M count=2500 status=none
>>     # losetup --sector-size 4096 -fP --show /var/tmp/loopbackfile1.img
>>
>> This works fine most of the times, but in the seldom case of the error, we get 'losetup: /var/tmp/loopbackfile1.img: failed to set up loop device: Resource temporarily unavailable'.
>>
>> I am sure that no other loop device is currently defined, so we don't run out of loop devices.
>>
>> We also see the following message in the syslog when the error occurs:
>>
>>      loop_set_block_size: loop0 () has still dirty pages (nrpages=2)
>>
>> The nrpages number varies from time to time. 
>>
>> "Resource temporarily unavailable" is EAGAIN, and function loop_set_block_size() in drivers/block/loop.c returns this after printing the syslog message via pr_warn:
>>
>> static int loop_set_block_size(struct loop_device *lo, unsigned long arg)
>> {
>> 	int err = 0;
>>
>> 	if (lo->lo_state != Lo_bound)
>> 		return -ENXIO;
>>
>> 	err = loop_validate_block_size(arg);
>> 	if (err)
>> 		return err;
>>
>> 	if (lo->lo_queue->limits.logical_block_size == arg)
>> 		return 0;
>>
>> 	sync_blockdev(lo->lo_device);
>> 	invalidate_bdev(lo->lo_device);
>>
>> 	blk_mq_freeze_queue(lo->lo_queue);
>>
>> 	/* invalidate_bdev should have truncated all the pages */
>> 	if (lo->lo_device->bd_inode->i_mapping->nrpages) {
>> 		err = -EAGAIN;
>> 		pr_warn("%s: loop%d (%s) has still dirty pages (nrpages=%lu)\n",
>> 			__func__, lo->lo_number, lo->lo_file_name,
>> 			lo->lo_device->bd_inode->i_mapping->nrpages);
>> 		goto out_unfreeze;
>> 	}
>>
>> 	blk_queue_logical_block_size(lo->lo_queue, arg);
>> 	blk_queue_physical_block_size(lo->lo_queue, arg);
>> 	blk_queue_io_min(lo->lo_queue, arg);
>> 	loop_update_dio(lo);
>> out_unfreeze:
>> 	blk_mq_unfreeze_queue(lo->lo_queue);
>>
>> 	return err;
>> }
>>
>> So looks like invalidate_bdev() did actually not truncate all the pages under some circumstances....
>>
>> The problem only happens when '--sector-size 4096' is specified, with the default sector size is always works. It does not call loop_set_block_size() in the default case I guess.
>>
>> The loop0 device has certainly be used by other testcases before, most likely with the default block size. But at the time of this run, no loop device is currently active (losetup shows nothing). 
>>
>> Anyone have an idea what goes wrong here? 
> 
> It returns '-EAGAIN' to ask userspace to try again.
> 
> I understand loop_set_block_size() doesn't prevent page cache of this
> loop disk from being dirtied, so it isn't strange to
> see lo_device->bd_inode->i_mapping->nrpages isn't zero after sync_blockdev()
> & invalidate_bdev() on loop.
> 

OK, that makes sense from the kernel perspective. 

However, shouldn't then the losetup userspace utility implement some kind of retry logic in case of -EAGAIN ?
I don't see that in the source of losetup.c nor in loopdev.c in the util-linux package. There is a retry loop in create_loop() in losetup.c retrying loopcxt_setup_device() in case of EBUSY, but not in case of EAGAIN.

And losetup also hides the original error code and just returns EXIT_FAILURE in case of a failure. So no real good chance for the script that uses losetup to catch that error situation and perform a retry itself.

Adding Karel Zak (the maintainer of util-linux).

@Karel Zak: How about adding EAGAIN to the condition for performing a retry? 

Something like this:

-		if (errno == EBUSY && !hasdev && ntries < 64) {
+		if ((errno == EBUSY || errno == EAGAIN) && !hasdev && ntries < 64) {
			xusleep(200000);
			ntries++;
			continue;
		}

>>
>> This happens on upstream kernels on the s390x platform, but I can't tell if is related to the platform or a specific kernel version. 
> 
> It can be reproduced easily when you run buffered write on loop disk,
> meantime keeping to change block size from one to another on this loop.

Yes it is very likely that this is happening in our use case. 
> 
> 
> Thanks,
> Ming
> 


-- 
Ingo Franzki
eMail: ifranzki@xxxxxxxxxxxxx  
Linux on IBM Z Development, Schoenaicher Str. 220, 71032 Boeblingen, Germany

IBM Deutschland Research & Development GmbH / Vorsitzender des Aufsichtsrats: Matthias Hartmann
Geschäftsführung: Dirk Wittkopp
Sitz der Gesellschaft: Böblingen / Registergericht: Amtsgericht Stuttgart, HRB 243294
IBM DATA Privacy Statement: https://www.ibm.com/privacy/us/en/