slob: Fix list_head bug during allocation

Hi Andrew,

This patch is in response to an email from the 0day kernel test robot subject:

340d3d6178 ("mm/slob.c: respect list_head abstraction layer"): kernel BUG at lib/list_debug.c:31!

This patch applies on top of linux-next tag: next-20190401

It fixes a patch that was merged recently into mm:

The patch titled
Subject: mm/slob.c: respect list_head abstraction layer
has been added to the -mm tree. Its filename is
slob-respect-list_head-abstraction-layer.patch

This patch should soon appear at
http://ozlabs.org/~akpm/mmots/broken-out/slob-respect-list_head-abstraction-layer.patch
and later at
http://ozlabs.org/~akpm/mmotm/broken-out/slob-respect-list_head-abstraction-layer.patch

If reverting is easier than patching I can re-work this into another version of the original (buggy) patch set which was the series:

[PATCH 0/4] mm: Use slab_list list_head instead of lru

Please don't be afraid to give a firm response. I'm new to mm and I'd like to not be a nuisance if I can manage it ;) I'd also like to fix this in a way that makes your day as easy as possible.

The 0day kernel test robot found a bug in the slob allocator caused by a patch from me recently merged into the mm tree. This is the first time the 0day has found a bug in already merged code of mine so I do not know the exact protocol in regards to linking the fix with the report, patching, reverting etc.

I was unable to reproduce the crash, I tried building with the config attached to the email above but the kernel booted fine for me in Qemu.

So I re-worked the module originally used for testing, it can be found here:

https://github.com/tcharding/ktest/tree/master/list_head

ok.

Next I tried to find the bug just using my eyes. This patch is the result. Unfortunately I can not understand why this bug was not triggered _before_ I originally patched it. Perhaps I'm not juggling all the state perfectly in my head. Anyways, this patch stops and code calling list manipulation functions if the slab_list page member has been modified during allocation.

The code in question revolves around an optimisation aimed at preventing fragmentation at the start of a slab due to the first fit nature of the allocation algorithm.

Full explanation is in the commit log for the patch, the short version is; skip optimisation if page list is modified, this only occurs when an allocation completely fills the slab and in this case the optimisation is unnecessary since we have not fragmented the slab by this allocation.

This is more than just a bug fix, it significantly reduces the complexity of the function while still fixing the reason for originally touching this code (violation of list_head abstraction).

The only testing I've done is to build and boot a kernel in Qemu (with CONFIG_LIST_DEBUG and CONFIG_SLOB) enabled). However, as mentioned, this method of testing did _not_ reproduce the 0day crash so if there are better suggestions on how I should test these I'm happy to do so.

thanks,
Tobin.

Tobin C. Harding (1):
slob: Only use list functions when safe to do so

 mm/slob.c | 50 ++++++++++++++++++++++++++++++--------------------
 1 file changed, 30 insertions(+), 20 deletions(-)

2.21.0