Re: [PATCH v2 08/19] xen/shbuf: switch xen-front-pgdir-shbuf to use INVALID_GRANT_REF

[Oleksandr <olekstysh@xxxxxxxxx>]

On 02.05.22 16:31, Juergen Gross wrote:


Hello Juergen


> On 29.04.22 17:28, Oleksandr wrote:
> > Hello Juergen
> >
> >
> > On 28.04.22 21:03, Oleksandr Tyshchenko wrote:
> > > On Thu, Apr 28, 2022 at 11:28 AM Juergen Gross <jgross@xxxxxxxx 
> > > <mailto:jgross@xxxxxxxx>> wrote:
> > > Hello Juergen
> > >
> > > [sorry for the possible format issue]
> > >
> > >     Instead of using a private macro for an invalid grant reference use
> > >     the common one.
> > >
> > >     Signed-off-by: Juergen Gross <jgross@xxxxxxxx
> > >     <mailto:jgross@xxxxxxxx>>
> > >     ---
> > >      drivers/xen/xen-front-pgdir-shbuf.c | 17 ++++-------------
> > >      1 file changed, 4 insertions(+), 13 deletions(-)
> > >
> > >     diff --git a/drivers/xen/xen-front-pgdir-shbuf.c
> > >     b/drivers/xen/xen-front-pgdir-shbuf.c
> > >     index a959dee21134..fa2921d4fbfc 100644
> > >     --- a/drivers/xen/xen-front-pgdir-shbuf.c
> > >     +++ b/drivers/xen/xen-front-pgdir-shbuf.c
> > >     @@ -21,15 +21,6 @@
> > >
> > >      #include <xen/xen-front-pgdir-shbuf.h>
> > >
> > >     -#ifndef GRANT_INVALID_REF
> > >     -/*
> > >     - * FIXME: usage of grant reference 0 as invalid grant reference:
> > >     - * grant reference 0 is valid, but never exposed to a PV driver,
> > >     - * because of the fact it is already in use/reserved by the PV
> > >     console.
> > >     - */
> > >     -#define GRANT_INVALID_REF      0
> > >     -#endif
> > >     -
> > >      /**
> > >       * This structure represents the structure of a shared page
> > >       * that contains grant references to the pages of the shared
> > >     @@ -83,7 +74,7 @@ grant_ref_t
> > >      xen_front_pgdir_shbuf_get_dir_start(struct xen_front_pgdir_shbuf
> > >     *buf)
> > >      {
> > >             if (!buf->grefs)
> > >     -               return GRANT_INVALID_REF;
> > >     +               return INVALID_GRANT_REF;
> > >
> > >             return buf->grefs[0];
> > >      }
> > >     @@ -142,7 +133,7 @@ void xen_front_pgdir_shbuf_free(struct
> > >     xen_front_pgdir_shbuf *buf)
> > >                     int i;
> > >
> > >                     for (i = 0; i < buf->num_grefs; i++)
> > >     -                       if (buf->grefs[i] != GRANT_INVALID_REF)
> > >     +                       if (buf->grefs[i] != INVALID_GRANT_REF)
> > >     gnttab_end_foreign_access(buf->grefs[i], 0UL);
> > >             }
> > >             kfree(buf->grefs);
> > >     @@ -355,7 +346,7 @@ static void backend_fill_page_dir(struct
> > >     xen_front_pgdir_shbuf *buf)
> > >             }
> > >             /* Last page must say there is no more pages. */
> > >             page_dir = (struct xen_page_directory *)ptr;
> > >     -       page_dir->gref_dir_next_page = GRANT_INVALID_REF;
> > >     +       page_dir->gref_dir_next_page = INVALID_GRANT_REF;
> > >      }
> > >
> > >      /**
> > >     @@ -384,7 +375,7 @@ static void guest_fill_page_dir(struct
> > >     xen_front_pgdir_shbuf *buf)
> > >
> > >                     if (grefs_left <= XEN_NUM_GREFS_PER_PAGE) {
> > >                             to_copy = grefs_left;
> > >     -                       page_dir->gref_dir_next_page =
> > >     GRANT_INVALID_REF;
> > >     +                       page_dir->gref_dir_next_page =
> > >     INVALID_GRANT_REF;
> > >
> > >
> > > I faced an issue with testing PV Sound with the current series.
> > >
> > > root@salvator-x-h3-4x2g-xt-domu:~# aplay /media/MoodyLoop.wav
> > > Playing WAVE '/media/MoodyLoop.wav' : Signed 16 bit Little Endian, 
> > > Rate 44100 Hz, Stereo
> > > (XEN) common/grant_table.c:1053:d1v2 Bad ref 0xffffffff for d6
> > >
> > > Here we have an interesting situation. PV Sound frontend uses this 
> > > xen-front-pgdir-shbuf framework. Technically, this patch changes 
> > > page_dir->gref_dir_next_page (reference to the next page describing 
> > > page directory) from 0 to 0xffffffff here.
> > > #define INVALID_GRANT_REF  ((grant_ref_t)-1)
> > >
> > > But according to the protocol (sndif.h), "0" means that there are no 
> > > more pages in the list and the user space backend expects only that 
> > > value. So receiving 0xffffffff it assumes there are pages in the 
> > > list and trying to process...
> > > https://elixir.bootlin.com/linux/v5.18-rc4/source/include/xen/interface/io/sndif.h#L650 
> > >
> > >
> > > I think, the same is relevant to backend_fill_page_dir() as well.
> >
> > In addition to what I said yesterday:
> >
> > PV Display also uses this xen-front-pgdir-shbuf framework. It's 
> > protocol (displif.h) also mentions the same as sndif.h if the context 
> > of gref_dir_next_page:
> >   * gref_dir_next_page - grant_ref_t, reference to the next page 
> > describing
> >   *   page directory. Must be 0 if there are no more pages in the list.
> >
> >
> > With that local change both PV devices work in my environment.
> >
> > diff --git a/drivers/xen/xen-front-pgdir-shbuf.c 
> > b/drivers/xen/xen-front-pgdir-shbuf.c
> > index fa2921d..ad4a88e 100644
> > --- a/drivers/xen/xen-front-pgdir-shbuf.c
> > +++ b/drivers/xen/xen-front-pgdir-shbuf.c
> > @@ -346,7 +346,7 @@ static void backend_fill_page_dir(struct 
> > xen_front_pgdir_shbuf *buf)
> >          }
> >          /* Last page must say there is no more pages. */
> >          page_dir = (struct xen_page_directory *)ptr;
> > -       page_dir->gref_dir_next_page = INVALID_GRANT_REF;
> > +       page_dir->gref_dir_next_page = 0;
> >   }
> >
> >   /**
> > @@ -375,7 +375,7 @@ static void guest_fill_page_dir(struct 
> > xen_front_pgdir_shbuf *buf)
> >                  if (grefs_left <= XEN_NUM_GREFS_PER_PAGE) {
> >                          to_copy = grefs_left;
> > -                       page_dir->gref_dir_next_page = 
> > INVALID_GRANT_REF;
> > +                       page_dir->gref_dir_next_page = 0;
> >                  } else {
> >                          to_copy = XEN_NUM_GREFS_PER_PAGE;
> >                          page_dir->gref_dir_next_page = buf->grefs[i 
> > + 1];
> I think I'll introduce a proper define for that purpose.

I think it would be the best option.



> Juergen
--
Regards,

Oleksandr Tyshchenko