[PATCH 2 of 4] Handle transient secondary mirror leg failures

[Takahiro Yasui <tyasui@redhat.com>]

On 12/18/09 13:49, malahal at us.ibm.com wrote:
> Takahiro Yasui [tyasui at redhat.com] wrote:
>> On 12/18/09 12:10, Jonathan Brassow wrote:
>>> 2) If you don't get a new table loaded, it will behave as a suspend/ 
>>> resume only.  Recent code changes in dm-raid1.c are causing  
>>> 'log_failure' and 'leg_failure' to not be reset in those cases.  IOW,  
>>> all these steps could be for nothing.  :(
>>
>> I would like to know how effective the retry is. As Jon explained
>> above, recent upstream kernel blocks all write I/Os on NOSYNC regions.
>> This means that those write I/Os are kept blocked for a long time.
>> For example, mirror retry interval in your patch #4 is 30 seconds and
>> application or  filesystem will be waited for 30 seconds (330 seconds
>> if retry count is 10). Can your application wait for more than 5 minutes?
>>
>> This behaviour will not been solved even if kernel is fixed so that
>> log_failure and leg_failure are reset. The write I/Os blocked will
>> be re-queued in the kernel when suspend/resume are done, but they
>> will be put in the hold queue again if the device failure is not
>> transient but permanent.
>>
>> I would like to know the use case of this patch set.
> 
> I have tested with RHEL5.4 kernel that doesn't block on device failure.

I see. Your retry approach looks good for the kernels on RHEL5.4 kernel
and 2.6.32 kernel since write I/Os aren't blocked even after the error
is detected on a mirror leg.

> IMHO, suspend/resume is doing two things if the kernel code blocks on
> failure -- 1) letting the kernel module unblock 2) start resync. We
> should separate those two actions. Maybe, we should do something here???
> Have an ioctl to start resync or have a message to unblock....

I'm sorry I don't get your point to separate unblock and resync.
Currently "unblock" is done by "suspend" and resync is done by "resume"
We need to reset log_failure/leg_failre, but anything else?

> Also, blocking the kernel module introduced a regression -- the
> mirror will block forever on a transient failure!

You are right. Refreshing block I/Os is necessary when dmeventd noticed
an error event and lvconvert command found nothing to be done for repair.
dmeventd should kick an action to refresh the mirror device.

Thanks,
Taka