On Sat, 2 Sep 2023 21:15:39 +0300
Dmitry Osipenko <dmitry.osipenko@xxxxxxxxxxxxx> wrote:
> On 8/28/23 14:16, Boris Brezillon wrote:
> > On Sun, 27 Aug 2023 20:54:27 +0300
> > Dmitry Osipenko <dmitry.osipenko@xxxxxxxxxxxxx> wrote:
> >
> >> Freeing drm-shmem GEM right after creating it using
> >> drm_gem_shmem_prime_import_sg_table() frees SGT of the imported dma-buf
> >> and then dma-buf frees this SGT second time.
> >>
> >> The v3d_prime_import_sg_table() is example of a error code path where
> >> dma-buf's SGT is freed by drm-shmem and then it's freed second time by
> >> dma_buf_unmap_attachment() in drm_gem_prime_import_dev().
> >>
> >> Add drm-shmem GEM flag telling that this is imported SGT shall not be
> >> treated as own SGT, fixing the use-after-free bug.
> >>
> >> Cc: stable@xxxxxxxxxxxxxxx
> >> Fixes: 2194a63a818d ("drm: Add library for shmem backed GEM objects")
> >> Signed-off-by: Dmitry Osipenko <dmitry.osipenko@xxxxxxxxxxxxx>
> >> ---
> >> drivers/gpu/drm/drm_gem_shmem_helper.c | 3 ++-
> >> include/drm/drm_gem_shmem_helper.h | 7 +++++++
> >> 2 files changed, 9 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/drivers/gpu/drm/drm_gem_shmem_helper.c b/drivers/gpu/drm/drm_gem_shmem_helper.c
> >> index a783d2245599..78d9cf2355a5 100644
> >> --- a/drivers/gpu/drm/drm_gem_shmem_helper.c
> >> +++ b/drivers/gpu/drm/drm_gem_shmem_helper.c
> >> @@ -141,7 +141,7 @@ void drm_gem_shmem_free(struct drm_gem_shmem_object *shmem)
> >>
> >> if (obj->import_attach) {
> >> drm_prime_gem_destroy(obj, shmem->sgt);
> >> - } else {
> >> + } else if (!shmem->imported_sgt) {
> >> dma_resv_lock(shmem->base.resv, NULL);
> >>
> >> drm_WARN_ON(obj->dev, shmem->vmap_use_count);
> >> @@ -758,6 +758,7 @@ drm_gem_shmem_prime_import_sg_table(struct drm_device *dev,
> >> return ERR_CAST(shmem);
> >>
> >> shmem->sgt = sgt;
> >> + shmem->imported_sgt = true;
> >
> >
> > I feel like adding more fields that can be used to do the is_imported()
> > check is going to be even more confusing. Can we instead have
> >
> > /* drm_gem_shmem_prime_import_sg_table() can be called from a
> > * driver specific ->import_sg_table() implementations that
> > * have extra failable initialization steps. Assign
> > * drm_gem_object::import_attach here (even though it's
> > * assigned in drm_gem_prime_import_dev()), so we don't end up
> > * with driver error paths calling drm_gem_shmem_free() with an
> > * imported sg_table assigned to drm_gem_shmem_object::sgt and
> > * drm_gem_object::import_attach left uninitialized.
> > */
> > shmem->base.import_attach = attach;
> >
> > here?
>
> AFAICT, this is not going to work because obj->import_attach will be
> released by drm_prime core by the time drm_gem_shmem_free() is invoked
> and drm_gem_shmem_free() uses obj->import_attach as well.
How can this happen? If something wrong happens in the driver-specific
->gem_prime_import_sg_table() implementation, drm_gem_shmem_free() will
be called before ->gem_prime_import_sg_table() returns, and the
attachment will only be released after that [1].
> I'll keep this
> patch around unless there will be other suggestions. To me the flag is
> good enough, I'll add a clarifying comment to the code in v16.
I really think this is a bad idea, for the same reasons I gave in my
reply to patch 2 (adding fields that need to be maintained when the
state can be inferred from other fields is error prone).
[1]https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/drm_prime.c#L958
