Re: Overagressive failing of disk reads, both LIBATA and IDE

[Mark Lord <liml@rtr.ca>]

Andrew Morton wrote:
> On Thu, 19 Mar 2009 23:32:55 -0400 Mark Lord <liml@rtr.ca> wrote:
> 
>> Norman Diamond wrote:
>>> For months I was wondering how a disk could do this:
>>> dd if=/dev/hda of=/dev/null bs=512 skip=551540 count=4  # succeeds
>>> dd if=/dev/hda of=/dev/null bs=512 skip=551544 count=4  # succeeds
>>> dd if=/dev/hda of=/dev/null bs=512 skip=551540 count=8  # fails
>>>
>>> It turns out the disk isn't doing that.  Linux is.  The old IDE drivers did
>>> it, but with LIBATA the same thing happens to /dev/sda.  In later examples
>>> also, the same happens to /dev/sda as /dev/hda.
>> ..
>>
>> You can blame me for the IDE driver not doing that properly.
>> But for libata, it's the SCSI layer.
>>
>> I've been patching this for years for my clients,
>> and will be updating the patch soon-ish and trying
>> again to get it into upstream kernels.
>>
>> Here's the (now ancient) 2.6.20 version for SLES10:
>>
>> * * *
>>
>> Allow SCSI to continue with the remaining blocks of a request
>> after encountering a media error.  Otherwise, it may just fail
>> the entire request, even though some blocks were fine and needed
>> by a completely different process than the one that wanted the bad block(s).
>>
>> Signed-off-by: Mark Lord <mlord@pobox.com>
>>
>> --- linux-2.6.16.60-0.6/drivers/scsi/scsi_lib.c	2008-03-10 13:46:03.000000000 -0400
>> +++ linux/drivers/scsi/scsi_lib.c	2008-03-21 11:54:09.000000000 -0400
>> @@ -888,6 +888,12 @@
>>  	 */
>>  	if (sense_valid && !sense_deferred) {
>>  		switch (sshdr.sense_key) {
>> +		case MEDIUM_ERROR:
>> +		/* Bad sector.  Fail it, and then continue the rest of the request. */
>> +		if (scsi_end_request(cmd, 0, cmd->device->sector_size, 1) == NULL) {
>> +			cmd->retries = 0;       // go around again..
>> +			return;
>> +		}
>>  		case UNIT_ATTENTION:
>>  			if (cmd->device->removable) {
>>  				/* Detected disc change.  Set a bit
> 
> Once upon a time the VFS would fall back to single page reads when a large
> readahead request failed.  That's probably still the case.
> 
> It was more by accident than by design, but it had (has) the desired effect?
..

Ahh.. but the block layer efficiently merges adjacent sectors
from multiple processes into single requests.  Generally a good thing,
that, but it does mean that one bad sector can currently trigger I/O
failures for several processes which aren't even interested in the
bad sector.

-ml