I think it is time to make two changes to get the project more focused:
1.) remove all non pagebuf meta data code. linux/xfs_fs_bio.c
2.) remove the delwri data write path.
These were both temporary and since their "real" counterparts are working
most of the time, let's focus on making the real thing work.
Russell, can you remove the non pagebuf meta data code and remove the config
option.
Ananth, can you remove the delwri code path and have delay_alloc always on?
This shouldn't be done 'till the test I sent you works (mmap_l). OK?
Thanks,
Jim
>
>Daniel Moore wrote:
>>
>> The fixes over the last couple of days have made the stuff I've been
>> running here behave a whole heap better. However there's still
>> corruption occuring.
>>
>> I have a simple test program that seems to reproducably tickle the
>> problem I'm seeing - if someone's interested I can pass it on.
>>
>> As I suspected from trying to get the kernel build test working,
>> previously written pages are getting passed back as newly read pages
>> (incorrectly). I'm pretty sure this only happens when the written
>> pages get flushed to disk and it can happen down to single 4k blocks
>> at a time.
>>
>> The test case is single threaded and only starts to fail (reliably) when
>> it is set to read and write big enough files such that it can't operate
>> entirely out of buffered pages.
>>
>> I've never seen meta data/junk returned - only file data that's probably
>> still buffered.
>
>My experience with all the latest changes:
>
>(1) Corruption in kernel makes same as before
>(2) doio-2 threads corrruption same as before
>
>On Chait's system (booted with 64M) & his kernel:
>
>(3) doio-2 threads corruption same as (2)
>
>Daniel, I'd be interested in your tests. Please send
>all the relevant details.
>
>thanks,
>
>--------------------------------------------------------------------------
>Rajagopal Ananthanarayanan ("ananth")
>Member Technical Staff, SGI.
>--------------------------------------------------------------------------
>
|