Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <8097505c-66a6-f60d-d8aa-2c36062813ff@suse.cz>
Date:   Tue, 3 Jan 2023 15:36:56 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.6.1
Subject: Re: [linus:master] [mm, slub] 0af8489b02:
 kernel_BUG_at_include/linux/mm.h
To:     Oliver Sang <oliver.sang@intel.com>
Cc:     Hyeonggon Yoo <42.hyeyoo@gmail.com>, oe-lkp@lists.linux.dev,
        lkp@intel.com, Mike Rapoport <rppt@linux.ibm.com>,
        Christoph Lameter <cl@linux.com>, linux-kernel@vger.kernel.org,
        linux-mm@kvack.org, Matthew Wilcox <willy@infradead.org>
References: <202212312021.bc1efe86-oliver.sang@intel.com>
 <41276905-b8a5-76ae-8a17-a8ec6558e988@suse.cz>
 <Y7Qxucg5le7WOzr7@xsang-OptiPlex-9020>
Content-Language: en-US
From:   Vlastimil Babka <vbabka@suse.cz>
In-Reply-To: <Y7Qxucg5le7WOzr7@xsang-OptiPlex-9020>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 1/3/23 14:46, Oliver Sang wrote:
> hi,
> 
>> 
>> Yet the page owner info tells us, 0x2daee was most recently used as order-1,
>> and now it seems to be (based on the raw struct page dump) sitting on a pcplist.
>> 
>> So the events leading up to this could be something like:
>> 
>> - 0x2daee is order-1 slab folio of the inode cache, sitting on the partial list
>> - despite being on partial list, it's freed ???
>> - somebody else allocates order-2 page 0x2daec and uses it for whatever,
>> then frees it
>> - 0x2daec is reallocated as order-1 slab from names_cache, then freed
>> - we try to allocate from the slab page 0x2daee and trip on the PageTail
>> 
>> Except, the freeing of order-2 page would have reset the PageTail and
>> compound_head in 0x2daec, so this is even more complicated or involves some
>> extra race?
> 
> FYI, we ran tests more up to 500 times, then saw different issues but rate is
> actually low
> 
> 56d5a2b9ba85a390 0af8489b0216fa1dd83e264bef8
> ---------------- ---------------------------
>        fail:runs  %reproduction    fail:runs
>            |             |             |
>            :500         12%          61:500   dmesg.invalid_opcode:#[##]
>            :500          3%          14:500   dmesg.kernel_BUG_at_include/linux/mm.h
>            :500          3%          17:500   dmesg.kernel_BUG_at_include/linux/page-flags.h
>            :500          5%          26:500   dmesg.kernel_BUG_at_lib/list_debug.c
>            :500          0%           2:500   dmesg.kernel_BUG_at_mm/page_alloc.c
>            :500          0%           2:500   dmesg.kernel_BUG_at_mm/usercopy.c

Hm even if rate is low, the different kinds of reports could be useful to
see, if all of that is caused by the commit.

>> 
>> In any case, this is something a debug_pagealloc kernel could have a chance
>> of catching earlier. Would it be possible to enable CONFIG_DEBUG_PAGEALLOC
>> and DEBUG_PAGEALLOC_ENABLE_DEFAULT additionally to the rest of the
>> configuration, and repeat the test?
> 
> ok, we are starting to test by these 2 additional configs now.

Great, thanks!