virtio-gpu uses cached mappings. shmem helpers use writecombine though. So roll our own mmap function, wrapping drm_gem_shmem_mmap(), to tweak vm_page_prot accordingly. Reported-by: Gurchetan Singh <gurchetansingh at chromium.org> Signed-off-by: Gerd Hoffmann <kraxel at redhat.com> --- drivers/gpu/drm/virtio/virtgpu_object.c | 18 +++++++++++++++++- 1 file changed, 17 insertions(+), 1 deletion(-) diff --git a/drivers/gpu/drm/virtio/virtgpu_object.c b/drivers/gpu/drm/virtio/virtgpu_object.c index 017a9e0fc3bb..158610902054 100644 --- a/drivers/gpu/drm/virtio/virtgpu_object.c +++ b/drivers/gpu/drm/virtio/virtgpu_object.c @@ -75,6 +75,22 @@ static void virtio_gpu_free_object(struct drm_gem_object *obj) drm_gem_shmem_free_object(obj); } +static int virtio_gpu_gem_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma) +{ + pgprot_t prot; + int ret; + + ret = drm_gem_shmem_mmap(obj, vma); + if (ret < 0) + return ret; + + /* virtio-gpu needs normal caching, so clear writecombine */ + prot = vm_get_page_prot(vma->vm_flags); + prot = pgprot_decrypted(prot); + vma->vm_page_prot = prot; + return 0; +} + static const struct drm_gem_object_funcs virtio_gpu_gem_funcs = { .free = virtio_gpu_free_object, .open = virtio_gpu_gem_object_open, @@ -86,7 +102,7 @@ static const struct drm_gem_object_funcs virtio_gpu_gem_funcs = { .get_sg_table = drm_gem_shmem_get_sg_table, .vmap = drm_gem_shmem_vmap, .vunmap = drm_gem_shmem_vunmap, - .mmap = &drm_gem_shmem_mmap, + .mmap = &virtio_gpu_gem_mmap, }; struct drm_gem_object *virtio_gpu_create_object(struct drm_device *dev, -- 2.18.1
Hi Am 10.12.19 um 09:57 schrieb Gerd Hoffmann:> virtio-gpu uses cached mappings. shmem helpers use writecombine though. > So roll our own mmap function, wrapping drm_gem_shmem_mmap(), to tweak > vm_page_prot accordingly. > > Reported-by: Gurchetan Singh <gurchetansingh at chromium.org> > Signed-off-by: Gerd Hoffmann <kraxel at redhat.com> > --- > drivers/gpu/drm/virtio/virtgpu_object.c | 18 +++++++++++++++++- > 1 file changed, 17 insertions(+), 1 deletion(-) > > diff --git a/drivers/gpu/drm/virtio/virtgpu_object.c b/drivers/gpu/drm/virtio/virtgpu_object.c > index 017a9e0fc3bb..158610902054 100644 > --- a/drivers/gpu/drm/virtio/virtgpu_object.c > +++ b/drivers/gpu/drm/virtio/virtgpu_object.c > @@ -75,6 +75,22 @@ static void virtio_gpu_free_object(struct drm_gem_object *obj) > drm_gem_shmem_free_object(obj); > } > > +static int virtio_gpu_gem_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma) > +{ > + pgprot_t prot; > + int ret; > + > + ret = drm_gem_shmem_mmap(obj, vma); > + if (ret < 0) > + return ret; > + > + /* virtio-gpu needs normal caching, so clear writecombine */ > + prot = vm_get_page_prot(vma->vm_flags); > + prot = pgprot_decrypted(prot); > + vma->vm_page_prot = prot; > + return 0; > +}There's similar code in udl, [1] which still uses writecombine for imported buffers. Virtio does not need this? Aside from this, do you think we could handle all special cases within shmem? Best regards Thomas [1] https://cgit.freedesktop.org/drm/drm-tip/tree/drivers/gpu/drm/udl/udl_gem.c?id=28ecf94a6f1072fc4744c06f5b3d267297125b37#n20> + > static const struct drm_gem_object_funcs virtio_gpu_gem_funcs = { > .free = virtio_gpu_free_object, > .open = virtio_gpu_gem_object_open, > @@ -86,7 +102,7 @@ static const struct drm_gem_object_funcs virtio_gpu_gem_funcs = { > .get_sg_table = drm_gem_shmem_get_sg_table, > .vmap = drm_gem_shmem_vmap, > .vunmap = drm_gem_shmem_vunmap, > - .mmap = &drm_gem_shmem_mmap, > + .mmap = &virtio_gpu_gem_mmap, > }; > > struct drm_gem_object *virtio_gpu_create_object(struct drm_device *dev, >-- Thomas Zimmermann Graphics Driver Developer SUSE Software Solutions Germany GmbH Maxfeldstr. 5, 90409 N?rnberg, Germany (HRB 36809, AG N?rnberg) Gesch?ftsf?hrer: Felix Imend?rffer -------------- next part -------------- A non-text attachment was scrubbed... Name: signature.asc Type: application/pgp-signature Size: 488 bytes Desc: OpenPGP digital signature URL: <http://lists.linuxfoundation.org/pipermail/virtualization/attachments/20191210/b42b6074/attachment-0001.sig>
On Tue, Dec 10, 2019 at 12:58 AM Gerd Hoffmann <kraxel at redhat.com> wrote:> > virtio-gpu uses cached mappings. shmem helpers use writecombine though. > So roll our own mmap function, wrapping drm_gem_shmem_mmap(), to tweak > vm_page_prot accordingly. > > Reported-by: Gurchetan Singh <gurchetansingh at chromium.org> > Signed-off-by: Gerd Hoffmann <kraxel at redhat.com> > --- > drivers/gpu/drm/virtio/virtgpu_object.c | 18 +++++++++++++++++- > 1 file changed, 17 insertions(+), 1 deletion(-) > > diff --git a/drivers/gpu/drm/virtio/virtgpu_object.c b/drivers/gpu/drm/virtio/virtgpu_object.c > index 017a9e0fc3bb..158610902054 100644 > --- a/drivers/gpu/drm/virtio/virtgpu_object.c > +++ b/drivers/gpu/drm/virtio/virtgpu_object.c > @@ -75,6 +75,22 @@ static void virtio_gpu_free_object(struct drm_gem_object *obj) > drm_gem_shmem_free_object(obj); > } > > +static int virtio_gpu_gem_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma) > +{ > + pgprot_t prot; > + int ret; > + > + ret = drm_gem_shmem_mmap(obj, vma); > + if (ret < 0) > + return ret; > + > + /* virtio-gpu needs normal caching, so clear writecombine */ > + prot = vm_get_page_prot(vma->vm_flags); > + prot = pgprot_decrypted(prot); > + vma->vm_page_prot = prot; > + return 0; > +} > + > static const struct drm_gem_object_funcs virtio_gpu_gem_funcs = { > .free = virtio_gpu_free_object, > .open = virtio_gpu_gem_object_open, > @@ -86,7 +102,7 @@ static const struct drm_gem_object_funcs virtio_gpu_gem_funcs = { > .get_sg_table = drm_gem_shmem_get_sg_table, > .vmap = drm_gem_shmem_vmap,Do we need vmap/vmunap? It seems optionable and also uses non-cacheable memory?> .vunmap = drm_gem_shmem_vunmap, > - .mmap = &drm_gem_shmem_mmap, > + .mmap = &virtio_gpu_gem_mmap,Why the &virtio_gpu_gem_mmap? Shouldn't just virtio_gpu_gem_mmap work?> }; > > struct drm_gem_object *virtio_gpu_create_object(struct drm_device *dev, > -- > 2.18.1 >
Hi,> There's similar code in udl, [1] which still uses writecombine for > imported buffers. Virtio does not need this?virtio doesn't support dma-buf imports (yet). So no worries for now. Why pick writecombine for the imported buffer btw? It'll probably be correct for the majority of imports, but it still looks like a educated guess to me. What if you import from vgem? I guess we should either ... (1) Ask the exporting driver to handle things, simliar to how it is done for vmaps already, probably by calling dma_buf_mmap(), or (2) Refuse to mmap imported objects via drm api.> Aside from this, do you think we could handle all special cases within > shmem?Probably makes sense to teach shmem helpers about caching. cheers, Gerd
Apparently Analagous Threads
- [PATCH] drm/virtio: fix mmap page attributes
- [PATCH] drm/virtio: fix mmap page attributes
- [PATCH 2/8] drm/shmem: switch shmem helper to &drm_gem_object_funcs.mmap
- [PATCH v2 02/11] drm/shmem: switch shmem helper to &drm_gem_object_funcs.mmap
- [PATCH v3 02/11] drm/shmem: switch shmem helper to &drm_gem_object_funcs.mmap