Re: [fuse-devel] Kernel panic under load

[Miklos Szeredi]

On Fri, Sep 9, 2016 at 9:42 PM, Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> Hi Miklos,
>
> On Fri, Sep 09, 2016 at 04:32:49PM +0200, Miklos Szeredi wrote:
>> On Fri, Sep 9, 2016 at 3:34 PM, Antonio SJ Musumeci <trapexit@xxxxxxxxxx> wrote:
>> > https://gist.github.com/bauruine/3bc00075c4d0b5b3353071d208ded30f
>> > https://github.com/trapexit/mergerfs/issues/295
>> >
>> > I've some users which are having issues with my filesystem where the
>> > system's load increases and then the kernel panics.
>> >
>> > Has anyone seen this before?
>>
>> Quite possibly this is caused by fuse, but the BUG is deep in mm
>> territory and I have zero clue about what it means.
>>
>> Hannes,  can you please look a the above crash in mm/workingset.c?
>
> The MM maintains a reclaimable list of page cache tree nodes that have
> gone empty (all pages evicted) except for the shadow entries reclaimed
> pages leave behind. When faulting a regular page back into such a node
> the code in page_cache_tree_insert() removes it from the list again:
>
>                 workingset_node_pages_inc(node);
>                 /*
>                  * Don't track node that contains actual pages.
>                  *
>                  * Avoid acquiring the list_lru lock if already
>                  * untracked.  The list_empty() test is safe as
>                  * node->private_list is protected by
>                  * mapping->tree_lock.
>                  */
>                 if (!list_empty(&node->private_list))
>                         list_lru_del(&workingset_shadow_nodes,
>                                      &node->private_list);
>
> The BUG_ON() triggers when we later walk the reclaimable list and find
> a radix tree node that has actual pages in it. This could happen when
> pages are inserted into a mapping without using add_to_page_cache and
> related functions. Does that maybe ring a bell?

Fuse allows pages to be spliced into the page cache when reading the
file.  It does this with replace_page_cache_page(), which is an atomic
version of delete_from_page_cache()+add_to_page_cache().

Fuse is the only user of replace_page_cache_page(), so I imagine bugs
can more easily escape notice than the more commonly used variants.

Could you please take a look at this function.  "git blame" shows that
it's older than the add/remove variants, but I haven't gone into the
details.

Thanks,
Miklos