Re: [syzbot] general protection fault in list_lru_add

[Muchun Song]

On Thu, Mar 24, 2022 at 7:11 AM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> Muchun,
>  mind taking a look at this asap? This seems like a pretty core thing,
> and the fact that it oopses on that

Sorry for this. I am looking at this now.

>
> > RIP: 0010:list_add_tail include/linux/list.h:102 [inline]
> > RIP: 0010:list_lru_add+0x277/0x510 mm/list_lru.c:129
> >  d_lru_add fs/dcache.c:431 [inline]
> >  retain_dentry fs/dcache.c:685 [inline]
> >  dput+0x7a7/0xdb0 fs/dcache.c:908
>
> just worries me a lot.
>
> The dentry lru list rules are odd but not outrageously so. The main
> oddity is that the DCACHE_LRU_LIST bit in the dentry flags indicate
> whether the dentry is on a LRU list or not.
>
> And it's not one single list - it can be *either* the usual
> sb->lists_dentry_lru list, or the special "shrink list".
>
> But this oops is for the regular d_lru_add() path that adds the dentry
> to the sb->s_dentry_lru list as the dentry count goes down to zero
> (and it's not one of the dentries that get insta-free'd).

After this commit, the rules of dentry allocations changed.
The dentry should be allocated by kmem_cache_alloc_lru()
to set up the dentry reclaim context correctly (e.g. allocating
its list_lru_one). This issue seems that list_lru_one wasn't
allocated, then NULL pointer reference.

I'm trying to reproduce this and looking for the root cause.

Thanks.