Re: [PATCH] mm: page_alloc: document kmemleak's non-blockable __GFP_NOFAIL case

From: Yang Shi
Date: Thu Jul 25 2019 - 14:08:48 EST

On 7/24/19 7:48 PM, Andrew Morton wrote:
On Sat, 13 Jul 2019 04:49:04 +0800 Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx> wrote:

When running ltp's oom test with kmemleak enabled, the below warning was
triggerred since kernel detects __GFP_NOFAIL & ~__GFP_DIRECT_RECLAIM is
passed in:


The mempool_alloc_slab() clears __GFP_DIRECT_RECLAIM, kmemleak has
__GFP_NOFAIL set all the time due to commit
d9570ee3bd1d4f20ce63485f5ef05663866fe6c0 ("kmemleak: allow to coexist
with fault injection").

The fault-injection would not try to fail slab or page allocation if
__GFP_NOFAIL is used and that commit tries to turn off fault injection
for kmemleak allocation. Although __GFP_NOFAIL doesn't guarantee no
failure for all the cases (i.e. non-blockable allocation may fail), it
still makes sense to the most cases. Kmemleak is also a debugging tool,
so it sounds not worth changing the behavior.

It also meaks sense to keep the warning, so just document the special
case in the comment.


--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -4531,8 +4531,14 @@ bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
if (gfp_mask & __GFP_NOFAIL) {
- * All existing users of the __GFP_NOFAIL are blockable, so warn
- * of any new users that actually require GFP_NOWAIT
+ * The users of the __GFP_NOFAIL are expected be blockable,
+ * and this is true for the most cases except for kmemleak.
+ * The kmemleak pass in __GFP_NOFAIL to skip fault injection,
+ * however kmemleak may allocate object at some non-blockable
+ * context to trigger this warning.
+ *
+ * Keep this warning since it is still useful for the most
+ * normal cases.
Comment has rather a lot of typos. I'd normally fix them but I think
I'll duck this patch until the kmemleak situation is addressed, so we
can add a kmemleakless long-term comment, if desired.

Actually, this has been replaced by reverting the problematic commit. And, the patch has been in -mm tree. Please see: revert-kmemleak-allow-to-coexist-with-fault-injection.patch

I think we would like to have this merged in 5.3-rc1 or rc2?