Lachlan McIlroy wrote:
> Russell Cattelan wrote:
>>
>> Reference:
>> http://oss.sgi.com/archives/xfs/2008-06/msg00209.html
>>
>>
>> It turns out that the code in question is still broken.
>>
>> xfs_iext_irec_compact_full will still corrupt the incore extent list 
>> if it does
>> the the partial copy of extents from one page to the next.
>> I haven't quit figured out where things get out of sync but somehow
>> if_real_bytes which tracks the total number of bytes available in
>> the extent buffers and if_bytes (which tracks the total bytes used
>> for extents.
>>
>> So at some point the inode thinks is has more extents than allocated 
>> pages allow.
>> So what happens is xfs_iext_idx_to_irec which uses idxp to pass IN the
>> absolute extent index is suppose to change  idxp on the way OUT and 
>> erp_idxp
>> to be a buffer index pair. What happens is that it doesn't find the 
>> extent so idxp
>> is left holding the same value as was passed in and erp_idx is 0.
>> This causes the extent code to then index way past the end of extent 
>> buffer 0
>> into garbage mem.
>>
>> with 4k ext buffers max extent count per buffer is 256.
>> example being:
>> IN:
>> idxp = 400
>> should become:
>> idexp = 144
>> erp_idxp = 1
>>
>> but we end up not finding the extent so
>> we have
>> idxp = 400
>> erp_idxp =0
>>
>> so we now index 6400 bytes into a 4k buffer.
>>
>> Which often times  is a pages of mostly 0 so we end up with access to 
>> block 0 errors.
>>
>> The more I looked at this code the more it didn't make sense to do 
>> partial moves.
>> Since the list of extent buffers is only scanned once vs restarting 
>> the list whenever a partial move is done,
>> it is very unlikely to actually free an extent buffer. (granted it's 
>> possible but unlikely)
>>
>> xfs_iext_irec_compact_pages does the same thing as 
>> xfs_iext_irec_compact_full except that doesn't handle partial moves.
>>
>> xfs_iext_irec_compact is written such that ratio of current extents 
>> has to be significantly smaller than the current allocated space
>> xfs_inode: 4513
>> nextents < ( nlists  * XFS_LINEAR_EXT) >> 3
>>
>> As it turns out 99% of the time it calls xfs_iext_irec_compact_pages
>> (which is why it has been so hard to track this bug down).
>>
>> If you change the 3 to a 1 so the code alway calls compact_full vs 
>> compact_pages the extent list will corrupt amost
>> immediately.
> Awesome work Russell, we'll give it a go.
>
>>
>> Since the code is broken, almost never used, and provides only micro 
>> optimization of incore space I propose we just
>> remove it all together.
> Are you sure the bug is in xfs_iext_irec_compact_full?
>
> Is it possible that we are still indexing beyond the page when using
> xfs_iext_irec_compact_pages but the pages just happen to be sequential
> so the indexing gets the extent anyway?
I added a bunch of printk to track this down,  the compact_pages path is hit
a lot in fact as far as I can tell all running file systems that shrink 
extents and don't crash :-)

I should have done this originally my I'm including the modified 
makeextents that
I used to tickle this problem.
It reserves a bunch of space to create a contiguous extents then in 
unreserves space to
poke a bunch of holes creating a big extent list, it then goes back and 
writes the whole
file hopefully collapsing extents as it goes.

i.e.
makeextents -v -c 512 foo ; xfs_bmap -v foo
should give you 1024 extents
makeextents -v -f -c 512 foo ; xfs_bmap -v foo
will do the same thing but fill in the file with writes.
The number of resulting extents vary, but sometimes you
end up with one extent. sometimes more.

If you change the 3 to a 1 in the current code so compact_full is used 
vs compact_pages
and run the test it will hit some problem every time.
xexlist in kdb will show the corruption in the incore list.

This will run the code through all 3 formats so if you are lucky you end 
up hitting all
the cases indirect > 256, direct <= 256, and inline <= 2

note:  xfs_iext_indirect_to_direct does call compact_full but in that 
case we are already down
to under 256 extents (at least we should be ) and at that point 
compact_full will behave just like compact_pages.


>
>>
>> I'm also including an xfsidbg patch that for xexlist that prints out 
>> buffer offset and index.
>>
>> -Russell Cattelan
>>
>>
>