[dm-devel] [PATCH 12/13] Make generic_make_request handle arbitrarily large bios

[Kent Overstreet]

On Mon, May 21, 2012 at 01:17:06PM -0400, Vivek Goyal wrote:
> May be I am missing something, hence I will ask.  Is punting to workqueue
> will really solve the issue raised by Neil. 
> 
> Due to spliting required, we will be holding some bios in the stack and
> these bios can't be submitted till further allocation from pool happens. So
> will it matter whether we are waiting for allocation in submitting process
> context or in worker thread context.

Punting to workqueue allows all the bio splits that have already been
allocated to be submitted.

> IOW, say you have a pool of 2 bios. We allocate 1 bio (say bio A),  and submit
> it for IO (now 1 bio left in pool). Now, bio A needs to be split up, so we
> allocate bio B and submit it (pool is empty now). Now we try to submit bio
> B and this also needs to be split. There are no more free bios so we will
> wait for some to get free but none of the bios (A and B) have actually
> been submitted for IO so nothing will get freed and we have a deadlock
> (This is assuming that memory is tight enough that we are not able to do
> any allocations from the slab backing the mempool).

You're talking about a different issue. You can't safely split a split
from the same bio pool - but this code doesn't do that since it adds a
separate bio pool for each request_queue that's only for splits.