Re: [Intel-gfx] [PATCH] Revert "drm/vgem: fix cache synchronization on arm/arm64"
On Fri, Aug 02, 2019 at 10:18:10AM -0400, Sean Paul wrote:
> On Thu, Aug 01, 2019 at 01:44:58PM +0100, Chris Wilson wrote:
> > commit 7e9e5ead55be ("drm/vgem: fix cache synchronization on arm/arm64")
> > broke all of the !llc i915-vgem coherency tests in CI, and left the HW
> > very, very unhappy (which is even more scary).
> >
> > Fixes: 7e9e5ead55be ("drm/vgem: fix cache synchronization on arm/arm64")
> > Signed-off-by: Chris Wilson
>
> Acked-by: Sean Paul
Applied to drm-fixes directly.
Thanks, Daniel
>
> > Cc: Daniel Vetter
> > Cc: Rob Clark
> > Cc: Sean Paul
> > ---
> > drivers/gpu/drm/vgem/vgem_drv.c | 130
> > 1 file changed, 47 insertions(+), 83 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/vgem/vgem_drv.c
> > b/drivers/gpu/drm/vgem/vgem_drv.c
> > index b98689fb0d5d..5bd60ded3d81 100644
> > --- a/drivers/gpu/drm/vgem/vgem_drv.c
> > +++ b/drivers/gpu/drm/vgem/vgem_drv.c
> > @@ -54,16 +54,10 @@ static struct vgem_device {
> > struct platform_device *platform;
> > } *vgem_device;
> >
> > -static void sync_and_unpin(struct drm_vgem_gem_object *bo);
> > -static struct page **pin_and_sync(struct drm_vgem_gem_object *bo);
> > -
> > static void vgem_gem_free_object(struct drm_gem_object *obj)
> > {
> > struct drm_vgem_gem_object *vgem_obj = to_vgem_bo(obj);
> >
> > - if (!obj->import_attach)
> > - sync_and_unpin(vgem_obj);
> > -
> > kvfree(vgem_obj->pages);
> > mutex_destroy(&vgem_obj->pages_lock);
> >
> > @@ -91,15 +85,40 @@ static vm_fault_t vgem_gem_fault(struct vm_fault *vmf)
> > return VM_FAULT_SIGBUS;
> >
> > mutex_lock(&obj->pages_lock);
> > - if (!obj->pages)
> > - pin_and_sync(obj);
> > if (obj->pages) {
> > get_page(obj->pages[page_offset]);
> > vmf->page = obj->pages[page_offset];
> > ret = 0;
> > }
> > mutex_unlock(&obj->pages_lock);
> > + if (ret) {
> > + struct page *page;
> > +
> > + page = shmem_read_mapping_page(
> > + file_inode(obj->base.filp)->i_mapping,
> > + page_offset);
> > + if (!IS_ERR(page)) {
> > + vmf->page = page;
> > + ret = 0;
> > + } else switch (PTR_ERR(page)) {
> > + case -ENOSPC:
> > + case -ENOMEM:
> > + ret = VM_FAULT_OOM;
> > + break;
> > + case -EBUSY:
> > + ret = VM_FAULT_RETRY;
> > + break;
> > + case -EFAULT:
> > + case -EINVAL:
> > + ret = VM_FAULT_SIGBUS;
> > + break;
> > + default:
> > + WARN_ON(PTR_ERR(page));
> > + ret = VM_FAULT_SIGBUS;
> > + break;
> > + }
> >
> > + }
> > return ret;
> > }
> >
> > @@ -265,93 +284,32 @@ static const struct file_operations vgem_driver_fops
> > = {
> > .release= drm_release,
> > };
> >
> > -/* Called under pages_lock, except in free path (where it can't race): */
> > -static void sync_and_unpin(struct drm_vgem_gem_object *bo)
> > -{
> > - struct drm_device *dev = bo->base.dev;
> > -
> > - if (bo->table) {
> > - dma_sync_sg_for_cpu(dev->dev, bo->table->sgl,
> > - bo->table->nents, DMA_BIDIRECTIONAL);
> > - sg_free_table(bo->table);
> > - kfree(bo->table);
> > - bo->table = NULL;
> > - }
> > -
> > - if (bo->pages) {
> > - drm_gem_put_pages(&bo->base, bo->pages, true, true);
> > - bo->pages = NULL;
> > - }
> > -}
> > -
> > -static struct page **pin_and_sync(struct drm_vgem_gem_object *bo)
> > -{
> > - struct drm_device *dev = bo->base.dev;
> > - int npages = bo->base.size >> PAGE_SHIFT;
> > - struct page **pages;
> > - struct sg_table *sgt;
> > -
> > - WARN_ON(!mutex_is_locked(&bo->pages_lock));
> > -
> > - pages = drm_gem_get_pages(&bo->base);
> > - if (IS_ERR(pages)) {
> > - bo->pages_pin_count--;
> > - mutex_unlock(&bo->pages_lock);
> > - return pages;
> > - }
> > -
> > - sgt = drm_prime_pages_to_sg(pages, npages);
> > - if (IS_ERR(sgt)) {
> > - dev_err(dev->dev,
> > - "failed to allocate sgt: %ld\n",
> > - PTR_ERR(bo->table));
> > - drm_gem_put_pages(&bo->base, pages, false, false);
> > - mutex_unlock(&bo->pages_lock);
> > - return ERR_CAST(bo->table);
> > - }
> > -
> > - /*
> > -* Flush the object from the CPU cache so that importers
> > -* can rely on coherent indirect access via the exported
> > -* dma-address.
> > -*/
> > - dma_sync_sg_for_device(dev->dev, sgt->sgl,
> > - sgt->nents, DMA_
Re: [Intel-gfx] [PATCH] Revert "drm/vgem: fix cache synchronization on arm/arm64"
On Thu, Aug 01, 2019 at 01:44:58PM +0100, Chris Wilson wrote:
> commit 7e9e5ead55be ("drm/vgem: fix cache synchronization on arm/arm64")
> broke all of the !llc i915-vgem coherency tests in CI, and left the HW
> very, very unhappy (which is even more scary).
>
> Fixes: 7e9e5ead55be ("drm/vgem: fix cache synchronization on arm/arm64")
> Signed-off-by: Chris Wilson
Acked-by: Sean Paul
> Cc: Daniel Vetter
> Cc: Rob Clark
> Cc: Sean Paul
> ---
> drivers/gpu/drm/vgem/vgem_drv.c | 130
> 1 file changed, 47 insertions(+), 83 deletions(-)
>
> diff --git a/drivers/gpu/drm/vgem/vgem_drv.c b/drivers/gpu/drm/vgem/vgem_drv.c
> index b98689fb0d5d..5bd60ded3d81 100644
> --- a/drivers/gpu/drm/vgem/vgem_drv.c
> +++ b/drivers/gpu/drm/vgem/vgem_drv.c
> @@ -54,16 +54,10 @@ static struct vgem_device {
> struct platform_device *platform;
> } *vgem_device;
>
> -static void sync_and_unpin(struct drm_vgem_gem_object *bo);
> -static struct page **pin_and_sync(struct drm_vgem_gem_object *bo);
> -
> static void vgem_gem_free_object(struct drm_gem_object *obj)
> {
> struct drm_vgem_gem_object *vgem_obj = to_vgem_bo(obj);
>
> - if (!obj->import_attach)
> - sync_and_unpin(vgem_obj);
> -
> kvfree(vgem_obj->pages);
> mutex_destroy(&vgem_obj->pages_lock);
>
> @@ -91,15 +85,40 @@ static vm_fault_t vgem_gem_fault(struct vm_fault *vmf)
> return VM_FAULT_SIGBUS;
>
> mutex_lock(&obj->pages_lock);
> - if (!obj->pages)
> - pin_and_sync(obj);
> if (obj->pages) {
> get_page(obj->pages[page_offset]);
> vmf->page = obj->pages[page_offset];
> ret = 0;
> }
> mutex_unlock(&obj->pages_lock);
> + if (ret) {
> + struct page *page;
> +
> + page = shmem_read_mapping_page(
> + file_inode(obj->base.filp)->i_mapping,
> + page_offset);
> + if (!IS_ERR(page)) {
> + vmf->page = page;
> + ret = 0;
> + } else switch (PTR_ERR(page)) {
> + case -ENOSPC:
> + case -ENOMEM:
> + ret = VM_FAULT_OOM;
> + break;
> + case -EBUSY:
> + ret = VM_FAULT_RETRY;
> + break;
> + case -EFAULT:
> + case -EINVAL:
> + ret = VM_FAULT_SIGBUS;
> + break;
> + default:
> + WARN_ON(PTR_ERR(page));
> + ret = VM_FAULT_SIGBUS;
> + break;
> + }
>
> + }
> return ret;
> }
>
> @@ -265,93 +284,32 @@ static const struct file_operations vgem_driver_fops = {
> .release= drm_release,
> };
>
> -/* Called under pages_lock, except in free path (where it can't race): */
> -static void sync_and_unpin(struct drm_vgem_gem_object *bo)
> -{
> - struct drm_device *dev = bo->base.dev;
> -
> - if (bo->table) {
> - dma_sync_sg_for_cpu(dev->dev, bo->table->sgl,
> - bo->table->nents, DMA_BIDIRECTIONAL);
> - sg_free_table(bo->table);
> - kfree(bo->table);
> - bo->table = NULL;
> - }
> -
> - if (bo->pages) {
> - drm_gem_put_pages(&bo->base, bo->pages, true, true);
> - bo->pages = NULL;
> - }
> -}
> -
> -static struct page **pin_and_sync(struct drm_vgem_gem_object *bo)
> -{
> - struct drm_device *dev = bo->base.dev;
> - int npages = bo->base.size >> PAGE_SHIFT;
> - struct page **pages;
> - struct sg_table *sgt;
> -
> - WARN_ON(!mutex_is_locked(&bo->pages_lock));
> -
> - pages = drm_gem_get_pages(&bo->base);
> - if (IS_ERR(pages)) {
> - bo->pages_pin_count--;
> - mutex_unlock(&bo->pages_lock);
> - return pages;
> - }
> -
> - sgt = drm_prime_pages_to_sg(pages, npages);
> - if (IS_ERR(sgt)) {
> - dev_err(dev->dev,
> - "failed to allocate sgt: %ld\n",
> - PTR_ERR(bo->table));
> - drm_gem_put_pages(&bo->base, pages, false, false);
> - mutex_unlock(&bo->pages_lock);
> - return ERR_CAST(bo->table);
> - }
> -
> - /*
> - * Flush the object from the CPU cache so that importers
> - * can rely on coherent indirect access via the exported
> - * dma-address.
> - */
> - dma_sync_sg_for_device(dev->dev, sgt->sgl,
> - sgt->nents, DMA_BIDIRECTIONAL);
> -
> - bo->pages = pages;
> - bo->table = sgt;
> -
> - return pages;
> -}
> -
> static struct page **vgem_pin_pages(struct drm_vgem_gem_object *bo)
> {
> - struct page **pages;
> -
> mu
Re: [Intel-gfx] [PATCH] Revert "drm/vgem: fix cache synchronization on arm/arm64"
On Thu, Aug 01, 2019 at 01:44:58PM +0100, Chris Wilson wrote:
> commit 7e9e5ead55be ("drm/vgem: fix cache synchronization on arm/arm64")
> broke all of the !llc i915-vgem coherency tests in CI, and left the HW
> very, very unhappy (which is even more scary).
>
> Fixes: 7e9e5ead55be ("drm/vgem: fix cache synchronization on arm/arm64")
> Signed-off-by: Chris Wilson
> Cc: Daniel Vetter
> Cc: Rob Clark
> Cc: Sean Paul
Acked-by: Daniel Vetter
Not that this counts a hole lot unfortunately :-/
Trouble here is that x86 assumes that everything is coherent (it's not)
and dma-api assumes that coherentness is a property of the device and
opaque to the device driver (it's not). So maybe we should do a mix of v1
and v2 of Rob's patches, i.e. not inflict the cache flushing into
drm_clflush.c (to avoid the pointless struct device nonsense), but then
also do this in vgem here only with #ifdef ARM64 and leave the clflush for
everything else.
-Daniel
> ---
> drivers/gpu/drm/vgem/vgem_drv.c | 130
> 1 file changed, 47 insertions(+), 83 deletions(-)
>
> diff --git a/drivers/gpu/drm/vgem/vgem_drv.c b/drivers/gpu/drm/vgem/vgem_drv.c
> index b98689fb0d5d..5bd60ded3d81 100644
> --- a/drivers/gpu/drm/vgem/vgem_drv.c
> +++ b/drivers/gpu/drm/vgem/vgem_drv.c
> @@ -54,16 +54,10 @@ static struct vgem_device {
> struct platform_device *platform;
> } *vgem_device;
>
> -static void sync_and_unpin(struct drm_vgem_gem_object *bo);
> -static struct page **pin_and_sync(struct drm_vgem_gem_object *bo);
> -
> static void vgem_gem_free_object(struct drm_gem_object *obj)
> {
> struct drm_vgem_gem_object *vgem_obj = to_vgem_bo(obj);
>
> - if (!obj->import_attach)
> - sync_and_unpin(vgem_obj);
> -
> kvfree(vgem_obj->pages);
> mutex_destroy(&vgem_obj->pages_lock);
>
> @@ -91,15 +85,40 @@ static vm_fault_t vgem_gem_fault(struct vm_fault *vmf)
> return VM_FAULT_SIGBUS;
>
> mutex_lock(&obj->pages_lock);
> - if (!obj->pages)
> - pin_and_sync(obj);
> if (obj->pages) {
> get_page(obj->pages[page_offset]);
> vmf->page = obj->pages[page_offset];
> ret = 0;
> }
> mutex_unlock(&obj->pages_lock);
> + if (ret) {
> + struct page *page;
> +
> + page = shmem_read_mapping_page(
> + file_inode(obj->base.filp)->i_mapping,
> + page_offset);
> + if (!IS_ERR(page)) {
> + vmf->page = page;
> + ret = 0;
> + } else switch (PTR_ERR(page)) {
> + case -ENOSPC:
> + case -ENOMEM:
> + ret = VM_FAULT_OOM;
> + break;
> + case -EBUSY:
> + ret = VM_FAULT_RETRY;
> + break;
> + case -EFAULT:
> + case -EINVAL:
> + ret = VM_FAULT_SIGBUS;
> + break;
> + default:
> + WARN_ON(PTR_ERR(page));
> + ret = VM_FAULT_SIGBUS;
> + break;
> + }
>
> + }
> return ret;
> }
>
> @@ -265,93 +284,32 @@ static const struct file_operations vgem_driver_fops = {
> .release= drm_release,
> };
>
> -/* Called under pages_lock, except in free path (where it can't race): */
> -static void sync_and_unpin(struct drm_vgem_gem_object *bo)
> -{
> - struct drm_device *dev = bo->base.dev;
> -
> - if (bo->table) {
> - dma_sync_sg_for_cpu(dev->dev, bo->table->sgl,
> - bo->table->nents, DMA_BIDIRECTIONAL);
> - sg_free_table(bo->table);
> - kfree(bo->table);
> - bo->table = NULL;
> - }
> -
> - if (bo->pages) {
> - drm_gem_put_pages(&bo->base, bo->pages, true, true);
> - bo->pages = NULL;
> - }
> -}
> -
> -static struct page **pin_and_sync(struct drm_vgem_gem_object *bo)
> -{
> - struct drm_device *dev = bo->base.dev;
> - int npages = bo->base.size >> PAGE_SHIFT;
> - struct page **pages;
> - struct sg_table *sgt;
> -
> - WARN_ON(!mutex_is_locked(&bo->pages_lock));
> -
> - pages = drm_gem_get_pages(&bo->base);
> - if (IS_ERR(pages)) {
> - bo->pages_pin_count--;
> - mutex_unlock(&bo->pages_lock);
> - return pages;
> - }
> -
> - sgt = drm_prime_pages_to_sg(pages, npages);
> - if (IS_ERR(sgt)) {
> - dev_err(dev->dev,
> - "failed to allocate sgt: %ld\n",
> - PTR_ERR(bo->table));
> - drm_gem_put_pages(&bo->base, pages, false, false);
> - mutex_unlock(&bo->pages_lock);
> - return ERR_CAST(bo->table)
[Intel-gfx] [PATCH] Revert "drm/vgem: fix cache synchronization on arm/arm64"
commit 7e9e5ead55be ("drm/vgem: fix cache synchronization on arm/arm64")
broke all of the !llc i915-vgem coherency tests in CI, and left the HW
very, very unhappy (which is even more scary).
Fixes: 7e9e5ead55be ("drm/vgem: fix cache synchronization on arm/arm64")
Signed-off-by: Chris Wilson
Cc: Daniel Vetter
Cc: Rob Clark
Cc: Sean Paul
---
drivers/gpu/drm/vgem/vgem_drv.c | 130
1 file changed, 47 insertions(+), 83 deletions(-)
diff --git a/drivers/gpu/drm/vgem/vgem_drv.c b/drivers/gpu/drm/vgem/vgem_drv.c
index b98689fb0d5d..5bd60ded3d81 100644
--- a/drivers/gpu/drm/vgem/vgem_drv.c
+++ b/drivers/gpu/drm/vgem/vgem_drv.c
@@ -54,16 +54,10 @@ static struct vgem_device {
struct platform_device *platform;
} *vgem_device;
-static void sync_and_unpin(struct drm_vgem_gem_object *bo);
-static struct page **pin_and_sync(struct drm_vgem_gem_object *bo);
-
static void vgem_gem_free_object(struct drm_gem_object *obj)
{
struct drm_vgem_gem_object *vgem_obj = to_vgem_bo(obj);
- if (!obj->import_attach)
- sync_and_unpin(vgem_obj);
-
kvfree(vgem_obj->pages);
mutex_destroy(&vgem_obj->pages_lock);
@@ -91,15 +85,40 @@ static vm_fault_t vgem_gem_fault(struct vm_fault *vmf)
return VM_FAULT_SIGBUS;
mutex_lock(&obj->pages_lock);
- if (!obj->pages)
- pin_and_sync(obj);
if (obj->pages) {
get_page(obj->pages[page_offset]);
vmf->page = obj->pages[page_offset];
ret = 0;
}
mutex_unlock(&obj->pages_lock);
+ if (ret) {
+ struct page *page;
+
+ page = shmem_read_mapping_page(
+ file_inode(obj->base.filp)->i_mapping,
+ page_offset);
+ if (!IS_ERR(page)) {
+ vmf->page = page;
+ ret = 0;
+ } else switch (PTR_ERR(page)) {
+ case -ENOSPC:
+ case -ENOMEM:
+ ret = VM_FAULT_OOM;
+ break;
+ case -EBUSY:
+ ret = VM_FAULT_RETRY;
+ break;
+ case -EFAULT:
+ case -EINVAL:
+ ret = VM_FAULT_SIGBUS;
+ break;
+ default:
+ WARN_ON(PTR_ERR(page));
+ ret = VM_FAULT_SIGBUS;
+ break;
+ }
+ }
return ret;
}
@@ -265,93 +284,32 @@ static const struct file_operations vgem_driver_fops = {
.release= drm_release,
};
-/* Called under pages_lock, except in free path (where it can't race): */
-static void sync_and_unpin(struct drm_vgem_gem_object *bo)
-{
- struct drm_device *dev = bo->base.dev;
-
- if (bo->table) {
- dma_sync_sg_for_cpu(dev->dev, bo->table->sgl,
- bo->table->nents, DMA_BIDIRECTIONAL);
- sg_free_table(bo->table);
- kfree(bo->table);
- bo->table = NULL;
- }
-
- if (bo->pages) {
- drm_gem_put_pages(&bo->base, bo->pages, true, true);
- bo->pages = NULL;
- }
-}
-
-static struct page **pin_and_sync(struct drm_vgem_gem_object *bo)
-{
- struct drm_device *dev = bo->base.dev;
- int npages = bo->base.size >> PAGE_SHIFT;
- struct page **pages;
- struct sg_table *sgt;
-
- WARN_ON(!mutex_is_locked(&bo->pages_lock));
-
- pages = drm_gem_get_pages(&bo->base);
- if (IS_ERR(pages)) {
- bo->pages_pin_count--;
- mutex_unlock(&bo->pages_lock);
- return pages;
- }
-
- sgt = drm_prime_pages_to_sg(pages, npages);
- if (IS_ERR(sgt)) {
- dev_err(dev->dev,
- "failed to allocate sgt: %ld\n",
- PTR_ERR(bo->table));
- drm_gem_put_pages(&bo->base, pages, false, false);
- mutex_unlock(&bo->pages_lock);
- return ERR_CAST(bo->table);
- }
-
- /*
-* Flush the object from the CPU cache so that importers
-* can rely on coherent indirect access via the exported
-* dma-address.
-*/
- dma_sync_sg_for_device(dev->dev, sgt->sgl,
- sgt->nents, DMA_BIDIRECTIONAL);
-
- bo->pages = pages;
- bo->table = sgt;
-
- return pages;
-}
-
static struct page **vgem_pin_pages(struct drm_vgem_gem_object *bo)
{
- struct page **pages;
-
mutex_lock(&bo->pages_lock);
- if (bo->pages_pin_count++ == 0 && !bo->pages) {
- pages = pin_and_sync(bo);
- } else {
- WARN_ON(!bo->pages);
- page
