[PATCH] dma-buf: system_heap: Optimize sg_table-to-pages conversion in vmap

[PATCH] dma-buf: system_heap: Optimize sg_table-to-pages conversion in vmap

[Barry Song (Xiaomi)]

From: Xueyuan Chen <Xueyuan.chen21@gmail.com>

Replace the heavy for_each_sgtable_page() iterator in system_heap_do_vmap()
with a more efficient nested loop approach.

Instead of iterating page by page, we now iterate through the scatterlist
entries via for_each_sgtable_sg(). Because pages within a single sg entry
are physically contiguous, we can populate the page array with a in an
inner loop using simple pointer math. This save a lot of time.

The WARN_ON check is also pulled out of the loop to save branch
instructions.

Performance results mapping a 2GB buffer on Radxa O6:
- Before: ~1440000 ns
- After:  ~232000 ns
(~84% reduction in iteration time, or ~6.2x faster)

Cc: Sumit Semwal <sumit.semwal@linaro.org>
Cc: Benjamin Gaignard <benjamin.gaignard@collabora.com>
Cc: Brian Starkey <Brian.Starkey@arm.com>
Cc: John Stultz <jstultz@google.com>
Cc: T.J. Mercier <tjmercier@google.com>
Cc: Christian König <christian.koenig@amd.com>
Signed-off-by: Xueyuan Chen <Xueyuan.chen21@gmail.com>
Signed-off-by: Barry Song (Xiaomi) <baohua@kernel.org>
---
 drivers/dma-buf/heaps/system_heap.c | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/drivers/dma-buf/heaps/system_heap.c b/drivers/dma-buf/heaps/system_heap.c
index b3650d8fd651..769f01f0cc96 100644
--- a/drivers/dma-buf/heaps/system_heap.c
+++ b/drivers/dma-buf/heaps/system_heap.c
@@ -224,16 +224,21 @@ static void *system_heap_do_vmap(struct system_heap_buffer *buffer)
 	int npages = PAGE_ALIGN(buffer->len) / PAGE_SIZE;
 	struct page **pages = vmalloc(sizeof(struct page *) * npages);
 	struct page **tmp = pages;
-	struct sg_page_iter piter;
 	void *vaddr;
+	u32 i, j, count;
+	struct page *base_page;
+	struct scatterlist *sg;
 
 	if (!pages)
 		return ERR_PTR(-ENOMEM);
 
-	for_each_sgtable_page(table, &piter, 0) {
-		WARN_ON(tmp - pages >= npages);
-		*tmp++ = sg_page_iter_page(&piter);
+	for_each_sgtable_sg(table, sg, i) {
+		base_page = sg_page(sg);
+		count = sg->length >> PAGE_SHIFT;
+		for (j = 0; j < count; j++)
+			*tmp++ = base_page + j;
 	}
+	WARN_ON(tmp - pages != npages);
 
 	vaddr = vmap(pages, npages, VM_MAP, PAGE_KERNEL);
 	vfree(pages);
-- 
2.39.3 (Apple Git-146)

Re: [PATCH] dma-buf: system_heap: Optimize sg_table-to-pages conversion in vmap

[Christian König]

On 4/6/26 23:49, Barry Song (Xiaomi) wrote:
> From: Xueyuan Chen <Xueyuan.chen21@gmail.com>
> 
> Replace the heavy for_each_sgtable_page() iterator in system_heap_do_vmap()
> with a more efficient nested loop approach.
> 
> Instead of iterating page by page, we now iterate through the scatterlist
> entries via for_each_sgtable_sg(). Because pages within a single sg entry
> are physically contiguous, we can populate the page array with a in an
> inner loop using simple pointer math. This save a lot of time.
> 
> The WARN_ON check is also pulled out of the loop to save branch
> instructions.
> 
> Performance results mapping a 2GB buffer on Radxa O6:
> - Before: ~1440000 ns
> - After:  ~232000 ns
> (~84% reduction in iteration time, or ~6.2x faster)

Well real question is why do you care about the vmap performance?

That should basically only be used for fbdev emulation (except for VMGFX) and we absolutely don't care about performance there.

Regards,
Christian.

> 
> Cc: Sumit Semwal <sumit.semwal@linaro.org>
> Cc: Benjamin Gaignard <benjamin.gaignard@collabora.com>
> Cc: Brian Starkey <Brian.Starkey@arm.com>
> Cc: John Stultz <jstultz@google.com>
> Cc: T.J. Mercier <tjmercier@google.com>
> Cc: Christian König <christian.koenig@amd.com>
> Signed-off-by: Xueyuan Chen <Xueyuan.chen21@gmail.com>
> Signed-off-by: Barry Song (Xiaomi) <baohua@kernel.org>
> ---
>  drivers/dma-buf/heaps/system_heap.c | 13 +++++++++----
>  1 file changed, 9 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/dma-buf/heaps/system_heap.c b/drivers/dma-buf/heaps/system_heap.c
> index b3650d8fd651..769f01f0cc96 100644
> --- a/drivers/dma-buf/heaps/system_heap.c
> +++ b/drivers/dma-buf/heaps/system_heap.c
> @@ -224,16 +224,21 @@ static void *system_heap_do_vmap(struct system_heap_buffer *buffer)
>  	int npages = PAGE_ALIGN(buffer->len) / PAGE_SIZE;
>  	struct page **pages = vmalloc(sizeof(struct page *) * npages);
>  	struct page **tmp = pages;
> -	struct sg_page_iter piter;
>  	void *vaddr;
> +	u32 i, j, count;
> +	struct page *base_page;
> +	struct scatterlist *sg;
>  
>  	if (!pages)
>  		return ERR_PTR(-ENOMEM);
>  
> -	for_each_sgtable_page(table, &piter, 0) {
> -		WARN_ON(tmp - pages >= npages);
> -		*tmp++ = sg_page_iter_page(&piter);
> +	for_each_sgtable_sg(table, sg, i) {
> +		base_page = sg_page(sg);
> +		count = sg->length >> PAGE_SHIFT;
> +		for (j = 0; j < count; j++)
> +			*tmp++ = base_page + j;
>  	}
> +	WARN_ON(tmp - pages != npages);
>  
>  	vaddr = vmap(pages, npages, VM_MAP, PAGE_KERNEL);
>  	vfree(pages);

Re: [PATCH] dma-buf: system_heap: Optimize sg_table-to-pages conversion in vmap

[Barry Song]

On Tue, Apr 7, 2026 at 3:58 PM Christian König <christian.koenig@amd.com> wrote:
>
> On 4/6/26 23:49, Barry Song (Xiaomi) wrote:
> > From: Xueyuan Chen <Xueyuan.chen21@gmail.com>
> >
> > Replace the heavy for_each_sgtable_page() iterator in system_heap_do_vmap()
> > with a more efficient nested loop approach.
> >
> > Instead of iterating page by page, we now iterate through the scatterlist
> > entries via for_each_sgtable_sg(). Because pages within a single sg entry
> > are physically contiguous, we can populate the page array with a in an
> > inner loop using simple pointer math. This save a lot of time.
> >
> > The WARN_ON check is also pulled out of the loop to save branch
> > instructions.
> >
> > Performance results mapping a 2GB buffer on Radxa O6:
> > - Before: ~1440000 ns
> > - After:  ~232000 ns
> > (~84% reduction in iteration time, or ~6.2x faster)
>
> Well real question is why do you care about the vmap performance?
>
> That should basically only be used for fbdev emulation (except for VMGFX) and we absolutely don't care about performance there.

I agree that in mainline, dma_buf_vmap is not used very often.
Here’s what I was able to find:

  1   1638  drivers/dma-buf/dma-buf.c <<dma_buf_vmap_unlocked>>
             ret = dma_buf_vmap(dmabuf, map);
   2    376  drivers/gpu/drm/drm_gem_shmem_helper.c
<<drm_gem_shmem_vmap_locked>>
             ret = dma_buf_vmap(obj->import_attach->dmabuf, map);
   3     85  drivers/gpu/drm/etnaviv/etnaviv_gem_prime.c
<<etnaviv_gem_prime_vmap_impl>>
             ret = dma_buf_vmap(etnaviv_obj->base.import_attach->dmabuf, &map);
   4    433  drivers/gpu/drm/vmwgfx/vmwgfx_blit.c <<map_external>>
             ret = dma_buf_vmap(bo->tbo.base.dma_buf, map);
   5     88  drivers/gpu/drm/vmwgfx/vmwgfx_gem.c <<vmw_gem_vmap>>
             ret = dma_buf_vmap(obj->import_attach->dmabuf, map);

However, in the Android ecosystem, system_heap and similar heaps
are widely used across camera, NPU, and media drivers. Many of these
drivers are not in mainline but do use vmap() in real code paths.

As I can show you some of them from MTK platforms:

1:
[    6.689849] system_heap_vmap+0x17c/0x254 [system_heap
8d35d4ce35bb30d8a623f0b9863998a2528e4175]
[    6.689859] dma_buf_vmap_unlocked+0xb8/0x130
[    6.689861] aov_core_init+0x310/0x718 [mtk_aov
96e2e5e9457dcdacce3a7629b0600c5dbeca623b]
[    6.689873] mtk_aov_probe+0x434/0x5b4 [mtk_aov
96e2e5e9457dcdacce3a7629b0600c5dbeca623b]

2:
[  116.181643] __vmap_pages_range_noflush+0x7c4/0x814
[  116.181645] vmap+0xb4/0x148
[  116.181647] system_heap_vmap+0x17c/0x254 [system_heap
8d35d4ce35bb30d8a623f0b9863998a2528e4175]
[  116.181651] dma_buf_vmap_unlocked+0xb8/0x130
[  116.181653] mtk_cam_vb2_vaddr+0xa0/0xfc [mtk_cam_isp8s
0cf9be6c773a8f14aab9db9ebf53feacb499846a]
[  116.181682] vb2_plane_vaddr+0x5c/0x78
[  116.181684] mtk_cam_job_fill_ipi_frame+0xa8c/0x128c [mtk_cam_isp8s
0cf9be6c773a8f14aab9db9ebf53feacb499846a]

3:
[  116.306178] __vmap_pages_range_noflush+0x7c4/0x814
[  116.306183] vmap+0xb4/0x148
[  116.306187] system_heap_vmap+0x17c/0x254 [system_heap
8d35d4ce35bb30d8a623f0b9863998a2528e4175]
[  116.306209] dma_buf_vmap_unlocked+0xb8/0x130
[  116.306212] apu_sysmem_alloc+0x168/0x360 [apusys
8fb33cbce3b858d651b9da26fc370090a67cfb70]
[  116.306468] mdw_mem_alloc+0xd8/0x314 [apusys
8fb33cbce3b858d651b9da26fc370090a67cfb70]
[  116.306591] mdw_mem_pool_chunk_add+0x11c/0x400 [apusys
8fb33cbce3b858d651b9da26fc370090a67cfb70]
[  116.306712] mdw_mem_pool_create+0x190/0x2c8 [apusys
8fb33cbce3b858d651b9da26fc370090a67cfb70]
[  116.306833] mdw_drv_open+0x21c/0x47c [apusys
8fb33cbce3b858d651b9da26fc370090a67cfb70]

While we may want to encourage more of these drivers to upstream,
some aspects are beyond our control (different SoC vendors), but we
can at least contribute upstream ourselves.

Best Regards
Barry

Re: [PATCH] dma-buf: system_heap: Optimize sg_table-to-pages conversion in vmap

[Christian König]

On 4/7/26 13:29, Barry Song wrote:
> On Tue, Apr 7, 2026 at 3:58 PM Christian König <christian.koenig@amd.com> wrote:
>>
>> On 4/6/26 23:49, Barry Song (Xiaomi) wrote:
>>> From: Xueyuan Chen <Xueyuan.chen21@gmail.com>
>>>
>>> Replace the heavy for_each_sgtable_page() iterator in system_heap_do_vmap()
>>> with a more efficient nested loop approach.
>>>
>>> Instead of iterating page by page, we now iterate through the scatterlist
>>> entries via for_each_sgtable_sg(). Because pages within a single sg entry
>>> are physically contiguous, we can populate the page array with a in an
>>> inner loop using simple pointer math. This save a lot of time.
>>>
>>> The WARN_ON check is also pulled out of the loop to save branch
>>> instructions.
>>>
>>> Performance results mapping a 2GB buffer on Radxa O6:
>>> - Before: ~1440000 ns
>>> - After:  ~232000 ns
>>> (~84% reduction in iteration time, or ~6.2x faster)
>>
>> Well real question is why do you care about the vmap performance?
>>
>> That should basically only be used for fbdev emulation (except for VMGFX) and we absolutely don't care about performance there.
> 
> I agree that in mainline, dma_buf_vmap is not used very often.
> Here’s what I was able to find:
> 
>   1   1638  drivers/dma-buf/dma-buf.c <<dma_buf_vmap_unlocked>>
>              ret = dma_buf_vmap(dmabuf, map);
>    2    376  drivers/gpu/drm/drm_gem_shmem_helper.c
> <<drm_gem_shmem_vmap_locked>>
>              ret = dma_buf_vmap(obj->import_attach->dmabuf, map);
>    3     85  drivers/gpu/drm/etnaviv/etnaviv_gem_prime.c
> <<etnaviv_gem_prime_vmap_impl>>
>              ret = dma_buf_vmap(etnaviv_obj->base.import_attach->dmabuf, &map);
>    4    433  drivers/gpu/drm/vmwgfx/vmwgfx_blit.c <<map_external>>
>              ret = dma_buf_vmap(bo->tbo.base.dma_buf, map);
>    5     88  drivers/gpu/drm/vmwgfx/vmwgfx_gem.c <<vmw_gem_vmap>>
>              ret = dma_buf_vmap(obj->import_attach->dmabuf, map);
> 
> However, in the Android ecosystem, system_heap and similar heaps
> are widely used across camera, NPU, and media drivers. Many of these
> drivers are not in mainline but do use vmap() in real code paths.

Well out of tree drivers are not a justification to make an upstream changes.

Apart from a handful of workarounds which need to CPU access as fallback DMA-buf vmap is only used to provide fb dev emulation.

The vmap interface has already given us quite a headache in the first place and there are a couple of unresolved problems regarding synchronization and coherency. 

When a driver would be pushed upstream which makes so frequent use of the dma_buf_vmap function that it matters for the performance I think there would be push back on that and the driver developer would require a very good explanation why that is necessary.

So for now I have to reject that patch.

Regards,
Christian.

> 
> As I can show you some of them from MTK platforms:
> 
> 1:
> [    6.689849] system_heap_vmap+0x17c/0x254 [system_heap
> 8d35d4ce35bb30d8a623f0b9863998a2528e4175]
> [    6.689859] dma_buf_vmap_unlocked+0xb8/0x130
> [    6.689861] aov_core_init+0x310/0x718 [mtk_aov
> 96e2e5e9457dcdacce3a7629b0600c5dbeca623b]
> [    6.689873] mtk_aov_probe+0x434/0x5b4 [mtk_aov
> 96e2e5e9457dcdacce3a7629b0600c5dbeca623b]
> 
> 2:
> [  116.181643] __vmap_pages_range_noflush+0x7c4/0x814
> [  116.181645] vmap+0xb4/0x148
> [  116.181647] system_heap_vmap+0x17c/0x254 [system_heap
> 8d35d4ce35bb30d8a623f0b9863998a2528e4175]
> [  116.181651] dma_buf_vmap_unlocked+0xb8/0x130
> [  116.181653] mtk_cam_vb2_vaddr+0xa0/0xfc [mtk_cam_isp8s
> 0cf9be6c773a8f14aab9db9ebf53feacb499846a]
> [  116.181682] vb2_plane_vaddr+0x5c/0x78
> [  116.181684] mtk_cam_job_fill_ipi_frame+0xa8c/0x128c [mtk_cam_isp8s
> 0cf9be6c773a8f14aab9db9ebf53feacb499846a]
> 
> 3:
> [  116.306178] __vmap_pages_range_noflush+0x7c4/0x814
> [  116.306183] vmap+0xb4/0x148
> [  116.306187] system_heap_vmap+0x17c/0x254 [system_heap
> 8d35d4ce35bb30d8a623f0b9863998a2528e4175]
> [  116.306209] dma_buf_vmap_unlocked+0xb8/0x130
> [  116.306212] apu_sysmem_alloc+0x168/0x360 [apusys
> 8fb33cbce3b858d651b9da26fc370090a67cfb70]
> [  116.306468] mdw_mem_alloc+0xd8/0x314 [apusys
> 8fb33cbce3b858d651b9da26fc370090a67cfb70]
> [  116.306591] mdw_mem_pool_chunk_add+0x11c/0x400 [apusys
> 8fb33cbce3b858d651b9da26fc370090a67cfb70]
> [  116.306712] mdw_mem_pool_create+0x190/0x2c8 [apusys
> 8fb33cbce3b858d651b9da26fc370090a67cfb70]
> [  116.306833] mdw_drv_open+0x21c/0x47c [apusys
> 8fb33cbce3b858d651b9da26fc370090a67cfb70]
> 
> While we may want to encourage more of these drivers to upstream,
> some aspects are beyond our control (different SoC vendors), but we
> can at least contribute upstream ourselves.
> 
> Best Regards
> Barry