Re: [Intel-gfx] [PATCH 3/4] drm/ttm: Don't leak a resource on eviction error

[Thomas Hellström]

On 6/22/23 16:48, Christian König wrote:
> Am 22.06.23 um 16:08 schrieb Thomas Hellström:
> > On 6/22/23 15:55, Andi Shyti wrote:
> > > Hi Thomas,
> > >
> > > On Thu, Jun 22, 2023 at 12:14:11PM +0200, Thomas Hellström wrote:
> > > > On eviction errors other than -EMULTIHOP we were leaking a resource.
> > > > Fix.
> > > >
> > > > Fixes: 403797925768 ("drm/ttm: Fix multihop assert on eviction.")
> > > > Cc: Andrey Grodzovsky <andrey.grodzovsky@xxxxxxx>
> > > > Cc: Christian König <christian.koenig@xxxxxxx>
> > > > Cc: Christian Koenig <christian.koenig@xxxxxxx>
> > > > Cc: Huang Rui <ray.huang@xxxxxxx>
> > > > Cc: dri-devel@xxxxxxxxxxxxxxxxxxxxx
> > > > Cc: <stable@xxxxxxxxxxxxxxx> # v5.15+
> > > > Signed-off-by: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
> > > > ---
> > > >   drivers/gpu/drm/ttm/ttm_bo.c | 16 ++++++++--------
> > > >   1 file changed, 8 insertions(+), 8 deletions(-)
> > > >
> > > > diff --git a/drivers/gpu/drm/ttm/ttm_bo.c 
> > > > b/drivers/gpu/drm/ttm/ttm_bo.c
> > > > index 615d30c4262d..89530f2a027f 100644
> > > > --- a/drivers/gpu/drm/ttm/ttm_bo.c
> > > > +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> > > > @@ -462,14 +462,14 @@ static int ttm_bo_evict(struct 
> > > > ttm_buffer_object *bo,
> > > >       ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
> > > >       if (ret == -EMULTIHOP) {
> > > >           ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem, ctx, &hop);
> > > > -        if (ret) {
> > > > -            if (ret != -ERESTARTSYS && ret != -EINTR)
> > > > -                pr_err("Buffer eviction failed\n");
> > > > -            ttm_resource_free(bo, &evict_mem);
> > > > -            goto out;
> > > > -        }
> > > > -        /* try and move to final place now. */
> > > > -        goto bounce;
> > > > +        if (!ret)
> > > > +            /* try and move to final place now. */
> > > > +            goto bounce;
> > > As we are at this, can't we replace this with a while()? Goto's
> > > used instead of a while loop are a fist in the eye...
> >
> > I'm completely OK with that. this patch already did away with one of 
> > them. Let's hear Christian's opinion first, though.
>
> I'm not a fan of that goto either, but could we somehow avoid the 
> while(1) ? E.g. something like do { } while (!ret) after handling the 
> multihop?

I think the construct that makes it most obvious what's happening, 
although it needs two tests for -EMULTIHOP is something like

do {
....
   if (ret != -EMULTIHOP)
      break;
   ....
} while (ret ==-EMULTIHOP);

Will be out tomorrow, though, so I don't have time to respin before Monday.

/Thomas


> Christian.
>
> > Thanks,
> >
> > Thomas
> >
> >
> >
> >
> >
> > > It looks even better:
> > >
> > >     while (1) {
> > >         ret = ttm_bo_handle_move_mem(bo, evict_mem, true, ctx, &hop);
> > >         if (!ret)
> > >             break;
> > >
> > >         if (ret == -EMULTIHOP)
> > >             ret = ttm_bo_bounce_temp_buffer(bo, &evict_mem,
> > >                             ctx, &hop);
> > >
> > >         /* try again */
> > >         if (!ret)
> > >             continue;
> > >
> > >         ttm_resource_free(bo, &evict_mem);
> > >         if (ret != -ERESTARTSYS && ret != -EINTR)
> > >             pr_err("Buffer eviction failed\n");
> > >
> > >         break;
> > >     }
> > >
> > > Andi
> > >
> > > > +    }
> > > > +    if (ret) {
> > > > +        ttm_resource_free(bo, &evict_mem);
> > > > +        if (ret != -ERESTARTSYS && ret != -EINTR)
> > > > +            pr_err("Buffer eviction failed\n");
> > > >       }
> > > >   out:
> > > >       return ret;
> > > > --
> > > > 2.40.1