On Mon, Jun 24, 2024 at 09:41:34AM +0200, Mateusz Guzik wrote:
> On kernels compiled without this option (which is currently the default
> state) filemap_nr_thps expands to 0.
>
> do_dentry_open has a big chunk dependent on it, most of which gets
> optimized away, except for a branch and a full fence:
>
> if (f->f_mode & FMODE_WRITE) {
> [snip]
> smp_mb();
> if (filemap_nr_thps(inode->i_mapping)) {
> [snip]
> }
> }
>
> While the branch is pretty minor the fence really does not need to be
> there.
>
[the To: recipient bounces, thus got dropped]
Now that I sent this I remembered that some of the atomic ops in Linux
provide a full fence regardless of an architecture.
get_write_access uses atomic_inc_unless_negative which qualifies?
1551 static __always_inline bool
1552 atomic_inc_unless_negative(atomic_t *v)
1553 {
1554 │ kcsan_mb();
1555 │ instrument_atomic_read_write(v, sizeof(*v));
1556 │ return raw_atomic_inc_unless_negative(v);
1557 }
If so, the ifdefed-out smp_mb can be eliminated altogether if I got my
fences right.
On top of that mnt_get_write_access injects another smp_mb() anyway.
> This is a bare-minimum patch which takes care of it until someone(tm)
> cleans this up. Notably it does not conditionally compile other spots
> which issue the matching fence.
>
> I did not bother benchmarking it, not issuing a spurious full fence in
> the fast path does not warrant justification from perf standpoint.
>
> Signed-off-by: Mateusz Guzik <mjguzik@xxxxxxxxx>
> ---
>
> I am not particularly familiar with any of this, the smp_mb in the open
> for write path was sticking out like a sore thumb on code read so I
> figured there may be One Weird Trick to whack it.
>
> If the stock code is correct as is, then the ifdef as above is fine.
>
> The ifdefed chunk is big enough that it should probably be its own
> routine. I don't want to bikeshed so I did not go for it.
>
> For a moment I considered adding filemap_nr_thps_mb which would expand
> to 0 or issue the fence + do the read, but then I figured a routine
> claiming to post a fence and only conditionally do it is misleading at
> best.
>
> As per the commit message fences in collapse_file remain compiled in.
> It is unclear to me if the code following them is doing anything useful
> on kernels !CONFIG_READ_ONLY_THP_FOR_FS.
>
> All that said, if there is cosmetic touch ups you want done here, I can
> do them.
>
> However, a nice full patch would take care of all of the above and I
> have neither the information needed to do it nor the interest to get it,
> so should someone insinst on a full version I'm going to suggest they
> write it themselves. I repeat this is merely a damage control until
> someone sorts thigs out.
>
> fs/open.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/fs/open.c b/fs/open.c
> index 28f2fcbebb1b..654c300b3c33 100644
> --- a/fs/open.c
> +++ b/fs/open.c
> @@ -980,6 +980,7 @@ static int do_dentry_open(struct file *f,
> if ((f->f_flags & O_DIRECT) && !(f->f_mode & FMODE_CAN_ODIRECT))
> return -EINVAL;
>
> +#ifdef CONFIG_READ_ONLY_THP_FOR_FS
> /*
> * XXX: Huge page cache doesn't support writing yet. Drop all page
> * cache for this file before processing writes.
> @@ -1007,6 +1008,7 @@ static int do_dentry_open(struct file *f,
> filemap_invalidate_unlock(inode->i_mapping);
> }
> }
> +#endif
>
> return 0;
>
> --
> 2.43.0
>
