On Fri 10-06-16 16:24:59, Sergey Senozhatsky wrote:

that was fast!

On (06/10/16 08:34), Michal Hocko wrote:
[..]

quoted

OK, so this is flags & GFP_SLAB_BUG_MASK BUG_ON because gfp is
___GFP_HIGHMEM. It is my [1] patch which has introduced it.
I think we need the following. Andrew could you fold it into
mm-memcg-use-consistent-gfp-flags-during-readahead.patch or maybe keep
it as a separate patch?

[1] http://lkml.kernel.org/r/1465301556-26431-1-git-send-email-mhocko@kernel.org

Thanks for the report Sergey!

after quick tests -- works for me. please see below.

[...]

so the first bio_alloc() is ok now. what about the second bio_alloc()
in mpage_alloc()? it'll still see the ___GFP_HIGHMEM?

Sure, early morning for me... Thanks for catching that.
---

From a2712312c0a36506ba003747c593dfbdf8eaa8be Mon Sep 17 00:00:00 2001

From: Michal Hocko <mhocko@suse.com>
Date: Fri, 10 Jun 2016 08:27:33 +0200
Subject: [PATCH] mm: restrict gfp mask in mpage_alloc

Sergey has reported that we might hit BUG_ON in new_slab() because
unrestricted gfp mask used for the readahead purposes contains
incompatible flags (__GFP_HIGHMEM in his case):
[  429.191962] gfp: 2
[  429.192634] ------------[ cut here ]------------
[  429.193281] kernel BUG at mm/slub.c:1616!
[...]
[  429.217369]  [<ffffffff811ca221>] bio_alloc_bioset+0xbd/0x1b1
[  429.218013]  [<ffffffff81148078>] mpage_alloc+0x28/0x7b
[  429.218650]  [<ffffffff8114856a>] do_mpage_readpage+0x43d/0x545
[  429.219282]  [<ffffffff81148767>] mpage_readpages+0xf5/0x152

Make sure that mpage_alloc always restricts the mask to GFP_KERNEL subset.
This is what was done before "mm, memcg: use consistent gfp flags during
readahead" explicitly by mapping_gfp_constraint(mapping, GFP_KERNEL) in
mpage_readpages.

Reported-by: Sergey Senozhatsky <redacted>
Signed-off-by: Michal Hocko <mhocko@suse.com>
---
 fs/mpage.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/fs/mpage.c b/fs/mpage.c
index 9c11255b0797..c8a05901a37b 100644
--- a/fs/mpage.c
+++ b/fs/mpage.c

@@ -71,6 +71,8 @@ mpage_alloc(struct block_device *bdev,
 {
 	struct bio *bio;
 
+	/* Restrict the given (page cache) mask for slab allocations */
+	gfp_flags &= GFP_KERNEL;
 	bio = bio_alloc(gfp_flags, nr_vecs);
 
 	if (bio == NULL && (current->flags & PF_MEMALLOC)) {

-- 
2.8.1

-- 
Michal Hocko
SUSE Labs