| Message ID | 20180227115000.4105-1-christian.koenig@amd.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Christian, Am Dienstag, den 27.02.2018, 12:49 +0100 schrieb Christian König: > Unpin the GEM object only after freeing the sg table. What is the race that is being fixed here? The SG table is private to the importer and the importer should hopefully only call map_detach if it is done with all operations using the SG table. Thus it shouldn't matter that the SG table might point to moved pages during execution of this function. Regards, Lucas > Signed-off-by: Christian König <christian.koenig@amd.com> > --- > drivers/gpu/drm/drm_prime.c | 32 ++++++++++++++++---------------- > 1 file changed, 16 insertions(+), 16 deletions(-) > > diff --git a/drivers/gpu/drm/drm_prime.c > b/drivers/gpu/drm/drm_prime.c > index e82a976f0fba..c38dacda6119 100644 > --- a/drivers/gpu/drm/drm_prime.c > +++ b/drivers/gpu/drm/drm_prime.c > @@ -230,26 +230,26 @@ void drm_gem_map_detach(struct dma_buf > *dma_buf, > struct drm_prime_attachment *prime_attach = attach->priv; > struct drm_gem_object *obj = dma_buf->priv; > struct drm_device *dev = obj->dev; > - struct sg_table *sgt; > > - if (dev->driver->gem_prime_unpin) > - dev->driver->gem_prime_unpin(obj); > + if (prime_attach) { > + struct sg_table *sgt = prime_attach->sgt; > > - if (!prime_attach) > - return; > - > - sgt = prime_attach->sgt; > - if (sgt) { > - if (prime_attach->dir != DMA_NONE) > - dma_unmap_sg_attrs(attach->dev, sgt->sgl, > sgt->nents, > - prime_attach->dir, > - DMA_ATTR_SKIP_CPU_SYNC); > - sg_free_table(sgt); > + if (sgt) { > + if (prime_attach->dir != DMA_NONE) > + dma_unmap_sg_attrs(attach->dev, sgt- > >sgl, > + sgt->nents, > + prime_attach- > >dir, > + DMA_ATTR_SKIP_CPU > _SYNC); > + sg_free_table(sgt); > + } > + > + kfree(sgt); > + kfree(prime_attach); > + attach->priv = NULL; > } > > - kfree(sgt); > - kfree(prime_attach); > - attach->priv = NULL; > + if (dev->driver->gem_prime_unpin) > + dev->driver->gem_prime_unpin(obj); > } > EXPORT_SYMBOL(drm_gem_map_detach); >
Am 28.02.2018 um 10:48 schrieb Lucas Stach: > Hi Christian, > > Am Dienstag, den 27.02.2018, 12:49 +0100 schrieb Christian König: >> Unpin the GEM object only after freeing the sg table. > What is the race that is being fixed here? The SG table is private to > the importer and the importer should hopefully only call map_detach if > it is done with all operations using the SG table. Thus it shouldn't > matter that the SG table might point to moved pages during execution of > this function. Exactly, it shouldn't matter. This is just a precaution. When the device driver is buggy I want proper error messages from IOMMU and not accessing pages which might already be reused for something else. Regards, Christian. > > Regards, > Lucas > >> Signed-off-by: Christian König <christian.koenig@amd.com> >> --- >> drivers/gpu/drm/drm_prime.c | 32 ++++++++++++++++---------------- >> 1 file changed, 16 insertions(+), 16 deletions(-) >> >> diff --git a/drivers/gpu/drm/drm_prime.c >> b/drivers/gpu/drm/drm_prime.c >> index e82a976f0fba..c38dacda6119 100644 >> --- a/drivers/gpu/drm/drm_prime.c >> +++ b/drivers/gpu/drm/drm_prime.c >> @@ -230,26 +230,26 @@ void drm_gem_map_detach(struct dma_buf >> *dma_buf, >> struct drm_prime_attachment *prime_attach = attach->priv; >> struct drm_gem_object *obj = dma_buf->priv; >> struct drm_device *dev = obj->dev; >> - struct sg_table *sgt; >> >> - if (dev->driver->gem_prime_unpin) >> - dev->driver->gem_prime_unpin(obj); >> + if (prime_attach) { >> + struct sg_table *sgt = prime_attach->sgt; >> >> - if (!prime_attach) >> - return; >> - >> - sgt = prime_attach->sgt; >> - if (sgt) { >> - if (prime_attach->dir != DMA_NONE) >> - dma_unmap_sg_attrs(attach->dev, sgt->sgl, >> sgt->nents, >> - prime_attach->dir, >> - DMA_ATTR_SKIP_CPU_SYNC); >> - sg_free_table(sgt); >> + if (sgt) { >> + if (prime_attach->dir != DMA_NONE) >> + dma_unmap_sg_attrs(attach->dev, sgt- >>> sgl, >> + sgt->nents, >> + prime_attach- >>> dir, >> + DMA_ATTR_SKIP_CPU >> _SYNC); >> + sg_free_table(sgt); >> + } >> + >> + kfree(sgt); >> + kfree(prime_attach); >> + attach->priv = NULL; >> } >> >> - kfree(sgt); >> - kfree(prime_attach); >> - attach->priv = NULL; >> + if (dev->driver->gem_prime_unpin) >> + dev->driver->gem_prime_unpin(obj); >> } >> EXPORT_SYMBOL(drm_gem_map_detach); >>
Ping? Anybody who could give me an review on those changes? The first one is just nice to have, the rest is a nice memory usage reduction in some cases. Christian. Am 27.02.2018 um 12:49 schrieb Christian König: > Unpin the GEM object only after freeing the sg table. > > Signed-off-by: Christian König <christian.koenig@amd.com> > --- > drivers/gpu/drm/drm_prime.c | 32 ++++++++++++++++---------------- > 1 file changed, 16 insertions(+), 16 deletions(-) > > diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c > index e82a976f0fba..c38dacda6119 100644 > --- a/drivers/gpu/drm/drm_prime.c > +++ b/drivers/gpu/drm/drm_prime.c > @@ -230,26 +230,26 @@ void drm_gem_map_detach(struct dma_buf *dma_buf, > struct drm_prime_attachment *prime_attach = attach->priv; > struct drm_gem_object *obj = dma_buf->priv; > struct drm_device *dev = obj->dev; > - struct sg_table *sgt; > > - if (dev->driver->gem_prime_unpin) > - dev->driver->gem_prime_unpin(obj); > + if (prime_attach) { > + struct sg_table *sgt = prime_attach->sgt; > > - if (!prime_attach) > - return; > - > - sgt = prime_attach->sgt; > - if (sgt) { > - if (prime_attach->dir != DMA_NONE) > - dma_unmap_sg_attrs(attach->dev, sgt->sgl, sgt->nents, > - prime_attach->dir, > - DMA_ATTR_SKIP_CPU_SYNC); > - sg_free_table(sgt); > + if (sgt) { > + if (prime_attach->dir != DMA_NONE) > + dma_unmap_sg_attrs(attach->dev, sgt->sgl, > + sgt->nents, > + prime_attach->dir, > + DMA_ATTR_SKIP_CPU_SYNC); > + sg_free_table(sgt); > + } > + > + kfree(sgt); > + kfree(prime_attach); > + attach->priv = NULL; > } > > - kfree(sgt); > - kfree(prime_attach); > - attach->priv = NULL; > + if (dev->driver->gem_prime_unpin) > + dev->driver->gem_prime_unpin(obj); > } > EXPORT_SYMBOL(drm_gem_map_detach); >
On Wed, Feb 28, 2018 at 11:25:59AM +0100, Christian König wrote: > Am 28.02.2018 um 10:48 schrieb Lucas Stach: > > Hi Christian, > > > > Am Dienstag, den 27.02.2018, 12:49 +0100 schrieb Christian König: > > > Unpin the GEM object only after freeing the sg table. > > What is the race that is being fixed here? The SG table is private to > > the importer and the importer should hopefully only call map_detach if > > it is done with all operations using the SG table. Thus it shouldn't > > matter that the SG table might point to moved pages during execution of > > this function. > > Exactly, it shouldn't matter. This is just a precaution. > > When the device driver is buggy I want proper error messages from IOMMU and > not accessing pages which might already be reused for something else. Please add this to the commit message, rather crucial to understand the motivation. With that fixed you can have my Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> And pls push to drm-misc. -Daniel > > Regards, > Christian. > > > > > Regards, > > Lucas > > > > > Signed-off-by: Christian König <christian.koenig@amd.com> > > > --- > > > drivers/gpu/drm/drm_prime.c | 32 ++++++++++++++++---------------- > > > 1 file changed, 16 insertions(+), 16 deletions(-) > > > > > > diff --git a/drivers/gpu/drm/drm_prime.c > > > b/drivers/gpu/drm/drm_prime.c > > > index e82a976f0fba..c38dacda6119 100644 > > > --- a/drivers/gpu/drm/drm_prime.c > > > +++ b/drivers/gpu/drm/drm_prime.c > > > @@ -230,26 +230,26 @@ void drm_gem_map_detach(struct dma_buf > > > *dma_buf, > > > struct drm_prime_attachment *prime_attach = attach->priv; > > > struct drm_gem_object *obj = dma_buf->priv; > > > struct drm_device *dev = obj->dev; > > > - struct sg_table *sgt; > > > - if (dev->driver->gem_prime_unpin) > > > - dev->driver->gem_prime_unpin(obj); > > > + if (prime_attach) { > > > + struct sg_table *sgt = prime_attach->sgt; > > > - if (!prime_attach) > > > - return; > > > - > > > - sgt = prime_attach->sgt; > > > - if (sgt) { > > > - if (prime_attach->dir != DMA_NONE) > > > - dma_unmap_sg_attrs(attach->dev, sgt->sgl, > > > sgt->nents, > > > - prime_attach->dir, > > > - DMA_ATTR_SKIP_CPU_SYNC); > > > - sg_free_table(sgt); > > > + if (sgt) { > > > + if (prime_attach->dir != DMA_NONE) > > > + dma_unmap_sg_attrs(attach->dev, sgt- > > > > sgl, > > > + sgt->nents, > > > + prime_attach- > > > > dir, > > > + DMA_ATTR_SKIP_CPU > > > _SYNC); > > > + sg_free_table(sgt); > > > + } > > > + > > > + kfree(sgt); > > > + kfree(prime_attach); > > > + attach->priv = NULL; > > > } > > > - kfree(sgt); > > > - kfree(prime_attach); > > > - attach->priv = NULL; > > > + if (dev->driver->gem_prime_unpin) > > > + dev->driver->gem_prime_unpin(obj); > > > } > > > EXPORT_SYMBOL(drm_gem_map_detach); > > _______________________________________________ > dri-devel mailing list > dri-devel@lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/dri-devel
Am 06.03.2018 um 10:15 schrieb Daniel Vetter: > On Wed, Feb 28, 2018 at 11:25:59AM +0100, Christian König wrote: >> Am 28.02.2018 um 10:48 schrieb Lucas Stach: >>> Hi Christian, >>> >>> Am Dienstag, den 27.02.2018, 12:49 +0100 schrieb Christian König: >>>> Unpin the GEM object only after freeing the sg table. >>> What is the race that is being fixed here? The SG table is private to >>> the importer and the importer should hopefully only call map_detach if >>> it is done with all operations using the SG table. Thus it shouldn't >>> matter that the SG table might point to moved pages during execution of >>> this function. >> Exactly, it shouldn't matter. This is just a precaution. >> >> When the device driver is buggy I want proper error messages from IOMMU and >> not accessing pages which might already be reused for something else. > Please add this to the commit message, rather crucial to understand the > motivation. With that fixed you can have my > > Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> > > And pls push to drm-misc. Can I use standard git for that now? I really don't want to mess with dim in my environment. Christian.
On Tue, Mar 06, 2018 at 10:30:56AM +0100, Christian König wrote: > Am 06.03.2018 um 10:15 schrieb Daniel Vetter: > > On Wed, Feb 28, 2018 at 11:25:59AM +0100, Christian König wrote: > > > Am 28.02.2018 um 10:48 schrieb Lucas Stach: > > > > Hi Christian, > > > > > > > > Am Dienstag, den 27.02.2018, 12:49 +0100 schrieb Christian König: > > > > > Unpin the GEM object only after freeing the sg table. > > > > What is the race that is being fixed here? The SG table is private to > > > > the importer and the importer should hopefully only call map_detach if > > > > it is done with all operations using the SG table. Thus it shouldn't > > > > matter that the SG table might point to moved pages during execution of > > > > this function. > > > Exactly, it shouldn't matter. This is just a precaution. > > > > > > When the device driver is buggy I want proper error messages from IOMMU and > > > not accessing pages which might already be reused for something else. > > Please add this to the commit message, rather crucial to understand the > > motivation. With that fixed you can have my > > > > Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> > > > > And pls push to drm-misc. > > Can I use standard git for that now? I really don't want to mess with dim in > my environment. Ping Alex to run it for you please. In an ideal world we'd run all that stuff server-side, but that's not happening anytime soon. Also if you have any specific issues about dim stomping over your setup, we'll be happy to fix it. If you set the (relative) paths correctly you can hide the various additional checkouts it needs rather well. -Daniel
diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c index e82a976f0fba..c38dacda6119 100644 --- a/drivers/gpu/drm/drm_prime.c +++ b/drivers/gpu/drm/drm_prime.c @@ -230,26 +230,26 @@ void drm_gem_map_detach(struct dma_buf *dma_buf, struct drm_prime_attachment *prime_attach = attach->priv; struct drm_gem_object *obj = dma_buf->priv; struct drm_device *dev = obj->dev; - struct sg_table *sgt; - if (dev->driver->gem_prime_unpin) - dev->driver->gem_prime_unpin(obj); + if (prime_attach) { + struct sg_table *sgt = prime_attach->sgt; - if (!prime_attach) - return; - - sgt = prime_attach->sgt; - if (sgt) { - if (prime_attach->dir != DMA_NONE) - dma_unmap_sg_attrs(attach->dev, sgt->sgl, sgt->nents, - prime_attach->dir, - DMA_ATTR_SKIP_CPU_SYNC); - sg_free_table(sgt); + if (sgt) { + if (prime_attach->dir != DMA_NONE) + dma_unmap_sg_attrs(attach->dev, sgt->sgl, + sgt->nents, + prime_attach->dir, + DMA_ATTR_SKIP_CPU_SYNC); + sg_free_table(sgt); + } + + kfree(sgt); + kfree(prime_attach); + attach->priv = NULL; } - kfree(sgt); - kfree(prime_attach); - attach->priv = NULL; + if (dev->driver->gem_prime_unpin) + dev->driver->gem_prime_unpin(obj); } EXPORT_SYMBOL(drm_gem_map_detach);
Unpin the GEM object only after freeing the sg table. Signed-off-by: Christian König <christian.koenig@amd.com> --- drivers/gpu/drm/drm_prime.c | 32 ++++++++++++++++---------------- 1 file changed, 16 insertions(+), 16 deletions(-)