Re: [PATCH] fs: Kill DCACHE_DONTCACHE dentry even if DCACHE_REFERENCED is set

[Dave Chinner <david@fromorbit.com>]

On Thu, Aug 27, 2020 at 05:58:07PM +0800, Li, Hao wrote:
> On 2020/8/27 14:37, Dave Chinner wrote:
> > On Fri, Aug 21, 2020 at 09:59:53AM +0800, Hao Li wrote:
> >> Currently, DCACHE_REFERENCED prevents the dentry with DCACHE_DONTCACHE
> >> set from being killed, so the corresponding inode can't be evicted. If
> >> the DAX policy of an inode is changed, we can't make policy changing
> >> take effects unless dropping caches manually.
> >>
> >> This patch fixes this problem and flushes the inode to disk to prepare
> >> for evicting it.
> >>
> >> Signed-off-by: Hao Li <lihao2018.fnst@cn.fujitsu.com>
> >> ---
> >>  fs/dcache.c | 3 ++-
> >>  fs/inode.c  | 2 +-
> >>  2 files changed, 3 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/fs/dcache.c b/fs/dcache.c
> >> index ea0485861d93..486c7409dc82 100644
> >> --- a/fs/dcache.c
> >> +++ b/fs/dcache.c
> >> @@ -796,7 +796,8 @@ static inline bool fast_dput(struct dentry *dentry)
> >>  	 */
> >>  	smp_rmb();
> >>  	d_flags = READ_ONCE(dentry->d_flags);
> >> -	d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED;
> >> +	d_flags &= DCACHE_REFERENCED | DCACHE_LRU_LIST | DCACHE_DISCONNECTED
> >> +			| DCACHE_DONTCACHE;
> > Seems reasonable, but you need to update the comment above as to
> > how this flag fits into this code....
> 
> Yes. I will change it. Thanks.
> 
> >
> >>  	/* Nothing to do? Dropping the reference was all we needed? */
> >>  	if (d_flags == (DCACHE_REFERENCED | DCACHE_LRU_LIST) && !d_unhashed(dentry))
> >> diff --git a/fs/inode.c b/fs/inode.c
> >> index 72c4c347afb7..5218a8aebd7f 100644
> >> --- a/fs/inode.c
> >> +++ b/fs/inode.c
> >> @@ -1632,7 +1632,7 @@ static void iput_final(struct inode *inode)
> >>  	}
> >>  
> >>  	state = inode->i_state;
> >> -	if (!drop) {
> >> +	if (!drop || (drop && (inode->i_state & I_DONTCACHE))) {
> >>  		WRITE_ONCE(inode->i_state, state | I_WILL_FREE);
> >>  		spin_unlock(&inode->i_lock);
> > What's this supposed to do? We'll only get here with drop set if the
> > filesystem is mounting or unmounting.
> 
> The variable drop will also be set to True if I_DONTCACHE is set on
> inode->i_state.
> Although mounting/unmounting will set the drop variable, it won't set
> I_DONTCACHE if I understand correctly. As a result,
> drop && (inode->i_state & I_DONTCACHE) will filter out mounting/unmounting.

So what does this have to do with changing the way the dcache
treats DCACHE_DONTCACHE?

Also, if I_DONTCACHE is set, but the inode has also been unlinked or
is unhashed, should we be writing it back? i.e. it might have been
dropped for a different reason to I_DONTCACHE....

IOWs, if there is a problem with how I_DONTCACHE is being handled,
then the problem must already exists regardless of the
DCACHE_DONTCACHE behaviour, right? So shouldn't this be a separate
bug fix with it's own explanation of the problem and the fix?

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com