Re: [PATCH v3 7/7] dma-buf: system_heap: Add a system-uncached heap re-using the system heap
On Fri, Oct 16, 2020 at 12:04 PM John Stultz wrote: > > On Thu, Oct 8, 2020 at 4:51 AM Brian Starkey wrote: > > On Sat, Oct 03, 2020 at 04:02:57AM +, John Stultz wrote: > > > @@ -393,6 +424,16 @@ static int system_heap_allocate(struct dma_heap > > > *heap, > > > /* just return, as put will call release and that will free > > > */ > > > return ret; > > > } > > > + > > > + /* > > > + * For uncached buffers, we need to initially flush cpu cache, since > > > + * the __GFP_ZERO on the allocation means the zeroing was done by > > > the > > > + * cpu and thus it is likely cached. Map (and implicitly flush) it > > > out > > > + * now so we don't get corruption later on. > > > + */ > > > + if (buffer->uncached) > > > + dma_map_sgtable(dma_heap_get_dev(heap), table, > > > DMA_BIDIRECTIONAL, 0); > > > > Do we have to keep this mapping around for the entire lifetime of the > > buffer? > > Yea, I guess we can just map and unmap it right there. It will look a > little absurd, but that sort of aligns with your next point. > > > Also, this problem (and solution) keeps lingering around. It really > > feels like there should be a better way to solve "clean the linear > > mapping all the way to DRAM", but I don't know what that should be. > > Yea, something better here would be nice... In ION, we had a little helper function named ion_buffer_prep_noncached that called arch_dma_prep_coherent() on all sg entries like so for_each_sg(table->sgl, sg, table->orig_nents, i) arch_dma_prep_coherent(sg_page(sg), sg->length); Would that help? ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [PATCH v3 7/7] dma-buf: system_heap: Add a system-uncached heap re-using the system heap
On Fri, Oct 16, 2020 at 12:03 PM John Stultz wrote: > On Thu, Oct 8, 2020 at 4:51 AM Brian Starkey wrote: > > On Sat, Oct 03, 2020 at 04:02:57AM +, John Stultz wrote: > > > @@ -426,6 +487,16 @@ static int system_heap_create(void) > > > if (IS_ERR(sys_heap)) > > > return PTR_ERR(sys_heap); > > > > > > + exp_info.name = "system-uncached"; > > > + exp_info.ops = _uncached_heap_ops; > > > + exp_info.priv = NULL; > > > + > > > + sys_uncached_heap = dma_heap_add(_info); > > > + if (IS_ERR(sys_uncached_heap)) > > > + return PTR_ERR(sys_heap); > > > + > > > > In principle, there's a race here between the heap getting registered > > to sysfs and the dma_mask getting updated. > > > > I don't think it would cause a problem in practice, but with the API > > as it is, there's no way to avoid it - so I wonder if the > > dma_heap_get_dev() accessor isn't the right API design. > > Hrm. I guess to address your concern we would need split > dma_heap_add() into something like: > dma_heap_create() > dma_heap_add() > > Which breaks the creation of the heap with the registering it to the > subsystem, so some attributes can be tweaked inbetween? Looking at this some more, this approach isn't going to work. We create the device and then we call dma_coerce_mask_and_coherent() on it, but as soon as the device is created it seems possible for userland to directly access it. Again, though, as you mentioned this isn't terribly likely in practice. The best thing I can think of for now is to have the uncached heap's allocate pointer initially point to a dummy function that returns EBUSY and then after we update the dma mask then we can set it to the real alloc. I'll go with that for now, but let me know if you have other suggestions. thanks -john ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [PATCH v3 7/7] dma-buf: system_heap: Add a system-uncached heap re-using the system heap
On Thu, Oct 8, 2020 at 4:51 AM Brian Starkey wrote: > On Sat, Oct 03, 2020 at 04:02:57AM +, John Stultz wrote: > > @@ -215,8 +236,12 @@ static void *system_heap_do_vmap(struct > > system_heap_buffer *buffer) > > struct page **pages = vmalloc(sizeof(struct page *) * npages); > > struct page **tmp = pages; > > struct sg_page_iter piter; > > + pgprot_t pgprot = PAGE_KERNEL; > > void *vaddr; > > > > + if (buffer->uncached) > > + pgprot = pgprot_writecombine(PAGE_KERNEL); > > I think this should go after the 'if (!pages)' check. Best to get the > allocation failure check as close to the allocation as possible IMO. Sounds good. Changed in my tree. > > @@ -393,6 +424,16 @@ static int system_heap_allocate(struct dma_heap *heap, > > /* just return, as put will call release and that will free */ > > return ret; > > } > > + > > + /* > > + * For uncached buffers, we need to initially flush cpu cache, since > > + * the __GFP_ZERO on the allocation means the zeroing was done by the > > + * cpu and thus it is likely cached. Map (and implicitly flush) it out > > + * now so we don't get corruption later on. > > + */ > > + if (buffer->uncached) > > + dma_map_sgtable(dma_heap_get_dev(heap), table, > > DMA_BIDIRECTIONAL, 0); > > Do we have to keep this mapping around for the entire lifetime of the > buffer? Yea, I guess we can just map and unmap it right there. It will look a little absurd, but that sort of aligns with your next point. > Also, this problem (and solution) keeps lingering around. It really > feels like there should be a better way to solve "clean the linear > mapping all the way to DRAM", but I don't know what that should be. Yea, something better here would be nice... > > @@ -426,6 +487,16 @@ static int system_heap_create(void) > > if (IS_ERR(sys_heap)) > > return PTR_ERR(sys_heap); > > > > + exp_info.name = "system-uncached"; > > + exp_info.ops = _uncached_heap_ops; > > + exp_info.priv = NULL; > > + > > + sys_uncached_heap = dma_heap_add(_info); > > + if (IS_ERR(sys_uncached_heap)) > > + return PTR_ERR(sys_heap); > > + > > In principle, there's a race here between the heap getting registered > to sysfs and the dma_mask getting updated. > > I don't think it would cause a problem in practice, but with the API > as it is, there's no way to avoid it - so I wonder if the > dma_heap_get_dev() accessor isn't the right API design. Hrm. I guess to address your concern we would need split dma_heap_add() into something like: dma_heap_create() dma_heap_add() Which breaks the creation of the heap with the registering it to the subsystem, so some attributes can be tweaked inbetween? I'll see about taking a stab at this, but I'll probably submit my updated series sooner with this un-addressed so I can get some further review. thanks -john ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [PATCH v3 7/7] dma-buf: system_heap: Add a system-uncached heap re-using the system heap
On Sat, Oct 03, 2020 at 04:02:57AM +, John Stultz wrote:
> This adds a heap that allocates non-contiguous buffers that are
> marked as writecombined, so they are not cached by the CPU.
>
> This is useful, as most graphics buffers are usually not touched
> by the CPU or only written into once by the CPU. So when mapping
> the buffer over and over between devices, we can skip the CPU
> syncing, which saves a lot of cache management overhead, greatly
> improving performance.
>
> For folk using ION, there was a ION_FLAG_CACHED flag, which
> signaled if the returned buffer should be CPU cacheable or not.
> With DMA-BUF heaps, we do not yet have such a flag, and by default
> the current heaps (system and cma) produce CPU cachable buffers.
> So for folks transitioning from ION to DMA-BUF Heaps, this fills
> in some of that missing functionality.
>
> There has been a suggestion to make this functionality a flag
> (DMAHEAP_FLAG_UNCACHED?) on the system heap, similar to how
> ION used the ION_FLAG_CACHED. But I want to make sure an
> _UNCACHED flag would truely be a generic attribute across all
> heaps. So far that has been unclear, so having it as a separate
> heap seemes better for now. (But I'm open to discussion on this
> point!)
>
> This is a rework of earlier efforts to add a uncached system heap,
> done utilizing the exisitng system heap, adding just a bit of
> logic to handle the uncached case.
>
> Feedback would be very welcome!
>
> Many thanks to Liam Mark for his help to get this working.
>
> Pending opensource users of this code include:
> * AOSP HiKey960 gralloc:
> - https://android-review.googlesource.com/c/device/linaro/hikey/+/1399519
> - Visibly improves performance over the system heap
> * AOSP Codec2 (possibly, needs more review):
> -
> https://android-review.googlesource.com/c/platform/frameworks/av/+/1360640/17/media/codec2/vndk/C2DmaBufAllocator.cpp#325
>
> Cc: Sumit Semwal
> Cc: Liam Mark
> Cc: Laura Abbott
> Cc: Brian Starkey
> Cc: Hridya Valsaraju
> Cc: Suren Baghdasaryan
> Cc: Sandeep Patil
> Cc: Daniel Mentz
> Cc: Chris Goldsworthy
> Cc: �rjan Eide
> Cc: Robin Murphy
> Cc: Ezequiel Garcia
> Cc: Simon Ser
> Cc: James Jones
> Cc: linux-me...@vger.kernel.org
> Cc: dri-devel@lists.freedesktop.org
> Signed-off-by: John Stultz
> ---
> drivers/dma-buf/heaps/system_heap.c | 87 ++---
> 1 file changed, 79 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/dma-buf/heaps/system_heap.c
> b/drivers/dma-buf/heaps/system_heap.c
> index 2b8d4b6abacb..952f1fd9dacf 100644
> --- a/drivers/dma-buf/heaps/system_heap.c
> +++ b/drivers/dma-buf/heaps/system_heap.c
> @@ -22,6 +22,7 @@
> #include
>
> struct dma_heap *sys_heap;
> +struct dma_heap *sys_uncached_heap;
>
> struct system_heap_buffer {
> struct dma_heap *heap;
> @@ -31,6 +32,8 @@ struct system_heap_buffer {
> struct sg_table sg_table;
> int vmap_cnt;
> void *vaddr;
> +
> + bool uncached;
> };
>
> struct dma_heap_attachment {
> @@ -38,6 +41,8 @@ struct dma_heap_attachment {
> struct sg_table *table;
> struct list_head list;
> bool mapped;
> +
> + bool uncached;
> };
>
> #define HIGH_ORDER_GFP (((GFP_HIGHUSER | __GFP_ZERO | __GFP_NOWARN \
> @@ -94,7 +99,7 @@ static int system_heap_attach(struct dma_buf *dmabuf,
> a->dev = attachment->dev;
> INIT_LIST_HEAD(>list);
> a->mapped = false;
> -
> + a->uncached = buffer->uncached;
> attachment->priv = a;
>
> mutex_lock(>lock);
> @@ -124,9 +129,13 @@ static struct sg_table *system_heap_map_dma_buf(struct
> dma_buf_attachment *attac
> {
> struct dma_heap_attachment *a = attachment->priv;
> struct sg_table *table = a->table;
> + int attr = 0;
> int ret;
>
> - ret = dma_map_sgtable(attachment->dev, table, direction, 0);
> + if (a->uncached)
> + attr = DMA_ATTR_SKIP_CPU_SYNC;
> +
> + ret = dma_map_sgtable(attachment->dev, table, direction, attr);
> if (ret)
> return ERR_PTR(ret);
>
> @@ -139,9 +148,12 @@ static void system_heap_unmap_dma_buf(struct
> dma_buf_attachment *attachment,
> enum dma_data_direction direction)
> {
> struct dma_heap_attachment *a = attachment->priv;
> + int attr = 0;
>
> + if (a->uncached)
> + attr = DMA_ATTR_SKIP_CPU_SYNC;
> a->mapped = false;
> - dma_unmap_sgtable(attachment->dev, table, direction, 0);
> + dma_unmap_sgtable(attachment->dev, table, direction, attr);
> }
>
> static int system_heap_dma_buf_begin_cpu_access(struct dma_buf *dmabuf,
> @@ -150,6 +162,9 @@ static int system_heap_dma_buf_begin_cpu_access(struct
> dma_buf *dmabuf,
> struct system_heap_buffer *buffer = dmabuf->priv;
> struct dma_heap_attachment *a;
>
> + if (buffer->uncached)
> + return 0;
> +
> mutex_lock(>lock);
>
> if (buffer->vmap_cnt)
> @@ -171,6
Re: [PATCH v3 7/7] dma-buf: system_heap: Add a system-uncached heap re-using the system heap
On Mon, Oct 5, 2020 at 6:45 AM Christoph Hellwig wrote: > > How is this going to deal with VIVT caches? Hrm. That's a good question. I'm not sure I totally have my head around it but, I guess we could make sure to call invalidate_kernel_vmap_range() in begin_cpu_access() and flush_kernel_vmap_range() in end_cpu_access() rather then exiting out early as we do now? Unless you have better guidance? Worse case we could check CONFIG_CPU_CACHE_VIVT and not register the system-uncached heap. thanks -john ___ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
[PATCH v3 7/7] dma-buf: system_heap: Add a system-uncached heap re-using the system heap
This adds a heap that allocates non-contiguous buffers that are
marked as writecombined, so they are not cached by the CPU.
This is useful, as most graphics buffers are usually not touched
by the CPU or only written into once by the CPU. So when mapping
the buffer over and over between devices, we can skip the CPU
syncing, which saves a lot of cache management overhead, greatly
improving performance.
For folk using ION, there was a ION_FLAG_CACHED flag, which
signaled if the returned buffer should be CPU cacheable or not.
With DMA-BUF heaps, we do not yet have such a flag, and by default
the current heaps (system and cma) produce CPU cachable buffers.
So for folks transitioning from ION to DMA-BUF Heaps, this fills
in some of that missing functionality.
There has been a suggestion to make this functionality a flag
(DMAHEAP_FLAG_UNCACHED?) on the system heap, similar to how
ION used the ION_FLAG_CACHED. But I want to make sure an
_UNCACHED flag would truely be a generic attribute across all
heaps. So far that has been unclear, so having it as a separate
heap seemes better for now. (But I'm open to discussion on this
point!)
This is a rework of earlier efforts to add a uncached system heap,
done utilizing the exisitng system heap, adding just a bit of
logic to handle the uncached case.
Feedback would be very welcome!
Many thanks to Liam Mark for his help to get this working.
Pending opensource users of this code include:
* AOSP HiKey960 gralloc:
- https://android-review.googlesource.com/c/device/linaro/hikey/+/1399519
- Visibly improves performance over the system heap
* AOSP Codec2 (possibly, needs more review):
-
https://android-review.googlesource.com/c/platform/frameworks/av/+/1360640/17/media/codec2/vndk/C2DmaBufAllocator.cpp#325
Cc: Sumit Semwal
Cc: Liam Mark
Cc: Laura Abbott
Cc: Brian Starkey
Cc: Hridya Valsaraju
Cc: Suren Baghdasaryan
Cc: Sandeep Patil
Cc: Daniel Mentz
Cc: Chris Goldsworthy
Cc: Ørjan Eide
Cc: Robin Murphy
Cc: Ezequiel Garcia
Cc: Simon Ser
Cc: James Jones
Cc: linux-me...@vger.kernel.org
Cc: dri-devel@lists.freedesktop.org
Signed-off-by: John Stultz
---
drivers/dma-buf/heaps/system_heap.c | 87 ++---
1 file changed, 79 insertions(+), 8 deletions(-)
diff --git a/drivers/dma-buf/heaps/system_heap.c
b/drivers/dma-buf/heaps/system_heap.c
index 2b8d4b6abacb..952f1fd9dacf 100644
--- a/drivers/dma-buf/heaps/system_heap.c
+++ b/drivers/dma-buf/heaps/system_heap.c
@@ -22,6 +22,7 @@
#include
struct dma_heap *sys_heap;
+struct dma_heap *sys_uncached_heap;
struct system_heap_buffer {
struct dma_heap *heap;
@@ -31,6 +32,8 @@ struct system_heap_buffer {
struct sg_table sg_table;
int vmap_cnt;
void *vaddr;
+
+ bool uncached;
};
struct dma_heap_attachment {
@@ -38,6 +41,8 @@ struct dma_heap_attachment {
struct sg_table *table;
struct list_head list;
bool mapped;
+
+ bool uncached;
};
#define HIGH_ORDER_GFP (((GFP_HIGHUSER | __GFP_ZERO | __GFP_NOWARN \
@@ -94,7 +99,7 @@ static int system_heap_attach(struct dma_buf *dmabuf,
a->dev = attachment->dev;
INIT_LIST_HEAD(>list);
a->mapped = false;
-
+ a->uncached = buffer->uncached;
attachment->priv = a;
mutex_lock(>lock);
@@ -124,9 +129,13 @@ static struct sg_table *system_heap_map_dma_buf(struct
dma_buf_attachment *attac
{
struct dma_heap_attachment *a = attachment->priv;
struct sg_table *table = a->table;
+ int attr = 0;
int ret;
- ret = dma_map_sgtable(attachment->dev, table, direction, 0);
+ if (a->uncached)
+ attr = DMA_ATTR_SKIP_CPU_SYNC;
+
+ ret = dma_map_sgtable(attachment->dev, table, direction, attr);
if (ret)
return ERR_PTR(ret);
@@ -139,9 +148,12 @@ static void system_heap_unmap_dma_buf(struct
dma_buf_attachment *attachment,
enum dma_data_direction direction)
{
struct dma_heap_attachment *a = attachment->priv;
+ int attr = 0;
+ if (a->uncached)
+ attr = DMA_ATTR_SKIP_CPU_SYNC;
a->mapped = false;
- dma_unmap_sgtable(attachment->dev, table, direction, 0);
+ dma_unmap_sgtable(attachment->dev, table, direction, attr);
}
static int system_heap_dma_buf_begin_cpu_access(struct dma_buf *dmabuf,
@@ -150,6 +162,9 @@ static int system_heap_dma_buf_begin_cpu_access(struct
dma_buf *dmabuf,
struct system_heap_buffer *buffer = dmabuf->priv;
struct dma_heap_attachment *a;
+ if (buffer->uncached)
+ return 0;
+
mutex_lock(>lock);
if (buffer->vmap_cnt)
@@ -171,6 +186,9 @@ static int system_heap_dma_buf_end_cpu_access(struct
dma_buf *dmabuf,
struct system_heap_buffer *buffer = dmabuf->priv;
struct dma_heap_attachment *a;
+ if (buffer->uncached)
+ return 0;
+
mutex_lock(>lock);
if
