Re: iomap: writeback ioend/bio allocation deadlock risk

[Christoph Hellwig <hch@lst.de>]

On Fri, May 21, 2021 at 04:54:45PM +0800, Ming Lei wrote:
> On Fri, May 21, 2021 at 10:36:35AM +0200, Christoph Hellwig wrote:
> > On Fri, May 21, 2021 at 04:35:03PM +0800, Ming Lei wrote:
> > > Just wondering why the ioend isn't submitted out after it becomes full?
> > 
> > block layer plugging?  Although failing bio allocations will kick that,
> > so that is not a deadlock risk.
> 
> These ioends are just added to one list stored on local stack variable(submit_list),

Yes.  But only until the code finished iterating over a page.  The
worst case number of bios for a page is PAGE_SIZE / SECTOR_SIZE, and
the bio_set is side to handle that comfortably.

> how can block layer plugging observe & submit them out?

It can't.  I was talking about the high-level plug that is held
over multiple pages.  At that point the bios are submitted to the
block layer already, but they might be held in a plug.  And looking at
the bio_alloc_bioset code we don't actually flush a plug at the moment
in that case.  Something like the patch below would do that:

diff --git a/block/bio.c b/block/bio.c
index 44205dfb6b60..5b9d2f4d7c08 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -432,6 +432,7 @@ struct bio *bio_alloc_bioset(gfp_t gfp_mask, unsigned short nr_iovecs,
 
 	p = mempool_alloc(&bs->bio_pool, gfp_mask);
 	if (!p && gfp_mask != saved_gfp) {
+		blk_schedule_flush_plug(current);
 		punt_bios_to_rescuer(bs);
 		gfp_mask = saved_gfp;
 		p = mempool_alloc(&bs->bio_pool, gfp_mask);