Re: WARNING: CPU: 1 PID: 14735 at fs/dcache.c:365 dentry_free+0x100/0x128

[Al Viro <viro@zeniv.linux.org.uk>]

On Wed, Jul 20, 2022 at 07:00:32PM +0800, Hillf Danton wrote:

> To help debug it, de-union d_in_lookup_hash with d_alias and add debug
> info after dentry is killed. If any warning hits, we know where to add
> something like
> 
> 	WARN_ON(dentry->d_flags & DCACHE_DENTRY_KILLED);
> 
> before hlist_bl_add or hlist_add.

IDGI.  That clearly has nothing to do with in-lookup stuff - no
DCACHE_PAR_LOOKUP in reported flags, so it either never had it set,
or it went through __d_lookup_done() already.

If anything, it might have already been through d_free(), with
d_rcu being confused for d_alias.

I'd do something like
	WARN_ON(dentry->d_flags & (1U<<31));
	dentry->d_flags |= 1U << 31;
in the begining of d_free() (possibly with dumping dentry state if we
hit that, not that there would be much to report; d_name.name might
be informative, though).

Again, in-lookup looks like a red herring - DCACHE_PAR_LOOKUP is set
only in d_alloc_parallel(), right next to the insertion into the list
and removed only in __d_lookup_free(), right next to the removal from
the same.  No DCACHE_PAR_LOOKUP in ->d_flags (it's 0x8008 in reported
cases, i.e. DCACHE_OP_REVALIDATE | DCACHE_DENTRY_KILLED).

What's more, take a look at retain_dentry(); WARN_ON(d_in_lookup(dentry))
right at the top and it had not triggered in any of the reports I've
seen in that thread.  Granted, it's not called on each path to
__dentry_kill(), but it is on the call chains I've seen reported...

Another thing that might be interesting to know is ->d_sb, along with
->d_sb->s_type->name and ->d_sb->s_id.  That should tell which fs it's
on...

I wonder if anyone had seen anything similar outside of parisc...
I don't know if I have any chance to reproduce it here - the only
parisc box I've got is a 715/100 (assuming the disk is still alive)
and it's 32bit, unlike the reported setups and, er, not fast.
qemu seems to have some parisc support, but it's 32bit-only at the
moment...

PS: please, Cc fsdevel on anything VFS-related.  Very few people are
still subscribed to l-k these days - I am, but it's impossible to read
through and postings can easily get missed.