[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: FW: [dm-devel] [PATCH 0/9] dm snapshot: shared exception store




On Fri, 31 Oct 2008, FUJITA Tomonori wrote:

> On Fri, 31 Oct 2008 03:12:11 -0400 (EDT)
> Mikulas Patocka <mpatocka redhat com> wrote:
> 
> > I looked at your implementation and I have these comments on it.
> 
> Thanks for the reviewing!
> 
> btw, can we discuss this on the mailing list? Some developers might be
> interested in the progress.

I see. I CC'd mailing list.

> > You have alloc_chunk_buffer in alloc_chunk_buffer. vmalloc can trigger 
> > writeouts and wait for them, resulting in a deadlock. Try
> > __vmalloc(size, GFP_NOFS, PAGE_KERNEL)
> > instead. Or you can use bufio instead of this, see below.

Oh, now I realized that it should be GFP_NOIO, not GFP_NOFS (both here and 
in kzalloc). GFP_NOFS causes that the allocator won't reenter filesystem, 
but it can still wait on IO on pages or swap. GFP_NOIO supresses any IO 
requests from the allocator.

Please, also rename brelse and set_buffer_dirty to something else, because 
these functions already exists in Linux --- they're static in your 
implementation, but if structure of include files change, you could easily 
end up including that one that contains brelse and set_buffer_dirty and 
get an error from conflicting definitions.

> If dm provides something that can work for the
> shared-exception-snapshot target, I'd love to use it and dump the
> home-grown caching code.

The core dm doesn't provide it now, but my patch (into core dm) does. The 
patch would be merged one day into core dm.

> The shared-exception-snapshot target wants to cache chunks and update
> multiple chunks atomically.
> 
> 
> > How do you handle the situation when the user removes origin with dmsetup, 
> > but leaves snapshots there? Don't you access already closed devices if 
> > that happens? This is kind of tricky to solve, because:
> > - there is no way to prevent removal of a target in device mapper 
> > architecture. So you can't say "I refuse to remove origin if I have open 
> > snapshots".
> > - open devices are bound to dm_target structure, so you couldn't open 
> > something in origin constructor, hand it over to snapshot and close it in 
> > snapshot destructor.
> > The only solution I see for this problem is: when destructing the origin, 
> > mark snapshots as unusable (with some flag, so that they always return 
> > -EIO), wait for all IOs on the snapshots to drain and then procceed with 
> > closing the origin.
> > If you have some other ideas, how to solve this, say them.
> 
> Yeah, this issue is on my todo list. I've not investigated this yet.
> 
> Your solution is fine by me. If an admin removes the origin, all the
> snapshots of it has to go. It sounds reasonable. Well, if we add a new
> mechanism for the first option to dm, that might be better.

I thought about it, but preventing removal of target would affect some 
other code (for example, dmsetup remove_all command would have to remove 
targets according to their dependencies), so it'd be better just solve 
this problem in a target and do not try to change device mapper 
architecture.

> Anyway, I'd like to go with the simplest option. As long as the kernel
> doesn't crash, I have no complaint.
> 
> > You can look at bufio layer, I wrote it for my implementation: 
> > http://people.redhat.com/mpatocka/patches/kernel/2.6.27/dm-bufio.patch
> > Bufio does exactly the same thing as your caching, but it is implemented 
> > in such a way that multiple targets can reuse it. So it could be good to 
> > use it to reduce code duplication.
> > 
> > With bufio, you can hold at most one buffer per thread, more threads can 
> > hold buffers concurrently. The buffer is accessed with dm_bufio_read 
> > (reads from disk) or dm_bufio_new (doesn't read from disk, it is expected 
> > that the caller immediatelly initializes the buffer). After access, it is 
> > released with dm_bufio_release. dm_bufio_mark_buffer_dirty marks the 
> > buffer dirty after you modified it and dm_bufio_write_dirty_buffers writes 
> > dirty buffers. Dirty buffers can be written automatically prior to 
> > dm_bufio_write_dirty_buffers if there is memory pressure. If you want to 
> > use it and have some comments on it (or you need some extensions), write 
> > to me.
> 
> As I wrote above, I'm pretty happy to try bufio if it will be merged
> into mainline. I guess that bufio doesn't update multiple chunks
> atomically but it's fine. I can work on it later on.

Well, nothing can update multiple chunks atomically :) The disks only 
guarantee you that they update 512-byte sector atomically, not even the 
whole chunk. You have to use journaling or some other method (phase trees, 
crash counts...) to get atomic updates --- and you can use it with 
dm-bufio.

If you want to use dm-bufio, you have to either change the btree code so 
that it holds at most one buffer (i.e. no more brelse_path). Another 
possiblity would be that I change dm-bufio for you to have more reserved 
buffers --- but then, you must guarantee that only one thread can take 
buffers. If you had multiple threads, each taking multiple buffers, there 
could be an out-of-memory deadlock.

Mikulas

> Thanks!
> 


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]