lkml.org 
[lkml]   [2009]   [Mar]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: ftruncate-mmap: pages are lost after writing to mmaped file.
Date
On Friday 20 March 2009 03:46:39 Jan Kara wrote:
> On Fri 20-03-09 02:48:21, Nick Piggin wrote:

> > Holding mapping->private_lock over the __set_page_dirty should
> > fix it, although I guess you'd want to release it before calling
> > __mark_inode_dirty so as not to put inode_lock under there. I
> > have a patch for this if it sounds reasonable.
>
> Yes, that seems to be a bug - the function actually looked suspitious to
> me yesterday but I somehow convinced myself that it's fine. Probably
> because fsx-linux is single-threaded.


After a whole lot of chasing my own tail in the VM and buffer layers,
I think it is a problem in ext2 (and I haven't been able to reproduce
with ext3 yet, which might lend weight to that, although as we have
seen, it is very timing dependent).

That would be slightly unfortunate because we still have Jan's ext3
problem, and also another reported problem of corruption on ext3 (on
brd driver).

Anyway, when I have reproduced the problem with the test case, the
"lost" writes are all reported to be holes. Unfortunately, that doesn't
point straight to the filesystem, because ext2 allocates blocks in this
case at writeout time, so if dirty bits are getting lost, then it would
be normal to see holes.

I then put in a whole lot of extra infrastructure to track metadata about
each struct page (when it was last written out, when it last had the number
of writable ptes reach 0, when the dirty bits were last cleared etc). And
none of the normal asertions were triggering: eg. when any page is removed
from pagecache (except truncates), it has always had all its buffers
written out *after* all ptes were made readonly or unmapped. Lots of other
tests and crap like that.

So I tried what I should have done to start with and did an e2fsck after
seeing corruption. Yes, it comes up with errors. Now that is unusual
because that should be largely insulated from the vm: if a dirty bit gets
lost, then the filesystem image should be quite happy and error-free with
a hole or unwritten data there.

I don't know ext? locking very well, except that it looks pretty overly
complex and crufty.

Usually, blocks are instantiated by write(2), under i_mutex, serialising
the allocator somewhat. mmap-write blocks are instantiated at writeout
time, unserialised. I moved truncate_mutex to cover the entire get_blocks
function, and can no longer trigger the problem. Might be a timing issue
though -- Ying, can you try this and see if you can still reproduce?

I close my eyes and pick something out of a hat. a686cd89. Search for XXX.
Nice. Whether or not this cased the problem, can someone please tell me
why it got merged in that state?

I'm leaving ext3 running for now. It looks like a nasty task to bisect
ext2 down to that commit :( and I would be more interested in trying to
reproduce Jan's ext3 problem, however, because I'm not too interested in
diving into ext2 locking to work out exactly what is racing and how to
fix it properly. I suspect it would be most productive to wire up some
ioctls right into the block allocator/lookup and code up a userspace
tester for it that could probably stress it a lot harder than kernel
writeout can.



\
 
 \ /
  Last update: 2009-03-24 08:47    [W:0.800 / U:0.864 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site