Re: [PATCH 2/2] fuse: Adjust readdir() buffer to requesting buffer size.

[Jaco Kroon]

Hi,

On 2025/04/01 16:40, Miklos Szeredi wrote:
> On Tue, 1 Apr 2025 at 16:29, Jaco Kroon <jaco@xxxxxxxxx> wrote:
> > After:
> >
> > getdents64(3, 0x7ffae8eed040 /* 276 entries */, 131072) = 6696
> > getdents64(3, 0x7ffae8eed040 /* 0 entries */, 131072) = 0
> This looks great.  But see below.
>
> > Signed-off-by: Jaco Kroon <jaco@xxxxxxxxx>
> > ---
> >   fs/fuse/readdir.c | 19 +++++++++++++++----
> >   1 file changed, 15 insertions(+), 4 deletions(-)
> >
> > diff --git a/fs/fuse/readdir.c b/fs/fuse/readdir.c
> > index 17ce9636a2b1..a13534f411b4 100644
> > --- a/fs/fuse/readdir.c
> > +++ b/fs/fuse/readdir.c
> > @@ -12,6 +12,7 @@
> >   #include <linux/posix_acl.h>
> >   #include <linux/pagemap.h>
> >   #include <linux/highmem.h>
> > +#include <linux/minmax.h>
> >
> >   static bool fuse_use_readdirplus(struct inode *dir, struct dir_context *ctx)
> >   {
> > @@ -337,11 +338,21 @@ static int fuse_readdir_uncached(struct file *file, struct dir_context *ctx)
> >          struct fuse_mount *fm = get_fuse_mount(inode);
> >          struct fuse_io_args ia = {};
> >          struct fuse_args_pages *ap = &ia.ap;
> > -       struct fuse_folio_desc desc = { .length = PAGE_SIZE };
> > +       struct fuse_folio_desc desc = { .length = ctx->count };
> >          u64 attr_version = 0, evict_ctr = 0;
> >          bool locked;
> > +       int order;
> >
> > -       folio = folio_alloc(GFP_KERNEL, 0);
> > +       desc.length = clamp(desc.length, PAGE_SIZE, fm->fc->max_pages << PAGE_SHIFT);
> > +       order = get_count_order(desc.length >> CONFIG_PAGE_SHIFT);
> > +
> > +       do {
> > +               folio = folio_alloc(GFP_KERNEL, order);
> > +               if (folio)
> > +                       break;
> > +               --order;
> > +               desc.length = PAGE_SIZE << order;
> > +       } while (order >= 0);
> >          if (!folio)
> >                  return -ENOMEM;
> Why not use kvmalloc instead?
Because fuse_simple_request via fuse_args_pages (ap) via fuse_io_args 
(ia) expects folios and changing that is more than what I'm capable off, 
and has larger overall impact.
> We could also implement allocation based on size of result in dev.c to
> optimize this, as most directories will be small, but that can be done
> later.

This indeed sounds interesting and would be great, but again, beyond 
what I'm capable of doing at this stage.

Great insights.

Thank you.

Kind regards,
Jaco