Re: Current qc_defer implementation may lead to infinite recursion

[Elias Oltmanns <eo@nebensachen.de>]

Tejun Heo <htejun@gmail.com> wrote:
> Elias Oltmanns wrote:
>> Tejun Heo <htejun@gmail.com> wrote:
>>> Elias Oltmanns wrote:
>>>> This proves that piix_qc_defer() has declined the same command 100
>>>> times in succession. However, this will only happen if the status of
>>>> all the commands enqueued for one port hasn't changed in the
>>>> meantime. This suggests to me that the threads scheduled for command
>>>> execution and completion aren't served for some reason. Any ideas?
>>> Blocked counts of 1 will cause busy looping because when blk_run_queue()
>>> returns because it's recursing too deep, it schedules unplug work right
>>> away, so it will easily loop 100 times.  Max blocked counts should be
>>> adjusted to two (needs some testing before actually submitting the
>>> change).  But that still shouldn't cause any lock up.  What happens if
>>> you remove the 100 times limit?  Does the machine hang on IO?
>> 
>> Yes, it does. In fact, I had already verified that before sending the
>> previous email.
>
> Hmmm.... it's supposed not to lock up although it can cause busy wait.

The same problem still exitst in 2.6.25-rc9. As I understand, not all
configurations are affected. So, perhaps I should bring this to the
attention of those who are working on the scheduler. What do you think?

Regards,

Elias