Skip to contentSkip to conversation

nouveau/hmm: fix migrate zero page to GPU

PATCH In review Ralph Campbell 3 participants 4 messages May 20, 2020 – May 22, 2020

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 nouveau/hmm: fix migrate zero page to GPU

When calling OpenCL clEnqueueSVMMigrateMem() on a region of memory that is backed by pte_none() or zero pages, migrate_vma_setup() will fill the source PFN array with an entry indicating the source page is zero. Use this to optimize migration to device private memory by allocating GPU memory and zero filling it instead of failing to migrate the page.

---

This patch applies cleanly to Jason's Gunthorpe's hmm tree plus two patches I posted earlier. The first is queued in Ben Skegg's nouveau tree and the second is still pending review/not queued.
[1] ("nouveau/hmm: map pages after migration")
https://lore.kernel.org/linux-mm/[email protected]/
[2] ("nouveau/hmm: fix nouveau_dmem_chunk allocations")
https://lore.kernel.org/lkml/[email protected]/

 drivers/gpu/drm/nouveau/nouveau_dmem.c | 75 ++++++++++++++++++++++----
 1 file changed, 66 insertions(+), 9 deletions(-)
drivers/gpu/drm/nouveau/nouveau_dmem.c +66 -9
index cbc71567f9a5..e5c230d9ae24 100644
@@ -56,6 +56,8 @@ enum nouveau_aper {
 typedef int (*nouveau_migrate_copy_t)(struct nouveau_drm *drm, u64 npages,
 				      enum nouveau_aper, u64 dst_addr,
 				      enum nouveau_aper, u64 src_addr);
+typedef int (*nouveau_clear_page_t)(struct nouveau_drm *drm, u32 length,
+				      enum nouveau_aper, u64 dst_addr);
 
 struct nouveau_dmem_chunk {
 	struct list_head list;
@@ -67,6 +69,7 @@ struct nouveau_dmem_chunk {
 
 struct nouveau_dmem_migrate {
 	nouveau_migrate_copy_t copy_func;
+	nouveau_clear_page_t clear_func;
 	struct nouveau_channel *chan;
 };
 
@@ -436,6 +439,52 @@ nvc0b5_migrate_copy(struct nouveau_drm *drm, u64 npages,
 	return 0;
 }
 
+static int
+nvc0b5_migrate_clear(struct nouveau_drm *drm, u32 length,
+		     enum nouveau_aper dst_aper, u64 dst_addr)
+{
+	struct nouveau_channel *chan = drm->dmem->migrate.chan;
+	u32 launch_dma = (1 << 10) /* REMAP_ENABLE_TRUE */ |
+			 (1 << 8) /* DST_MEMORY_LAYOUT_PITCH. */ |
+			 (1 << 7) /* SRC_MEMORY_LAYOUT_PITCH. */ |
+			 (1 << 2) /* FLUSH_ENABLE_TRUE. */ |
+			 (2 << 0) /* DATA_TRANSFER_TYPE_NON_PIPELINED. */;
+	u32 remap = (4 <<  0) /* DST_X_CONST_A */ |
+		    (5 <<  4) /* DST_Y_CONST_B */ |
+		    (3 << 16) /* COMPONENT_SIZE_FOUR */ |
+		    (1 << 24) /* NUM_DST_COMPONENTS_TWO */;
+	int ret;
+
+	ret = RING_SPACE(chan, 12);
+	if (ret)
+		return ret;
+
+	switch (dst_aper) {
+	case NOUVEAU_APER_VRAM:
+		BEGIN_IMC0(chan, NvSubCopy, 0x0264, 0);
+			break;
+	case NOUVEAU_APER_HOST:
+		BEGIN_IMC0(chan, NvSubCopy, 0x0264, 1);
+		break;
+	default:
+		return -EINVAL;
+	}
+	launch_dma |= 0x00002000; /* DST_TYPE_PHYSICAL. */
+
+	BEGIN_NVC0(chan, NvSubCopy, 0x0700, 3);
+	OUT_RING(chan, 0);
+	OUT_RING(chan, 0);
+	OUT_RING(chan, remap);
+	BEGIN_NVC0(chan, NvSubCopy, 0x0408, 2);
+	OUT_RING(chan, upper_32_bits(dst_addr));
+	OUT_RING(chan, lower_32_bits(dst_addr));
+	BEGIN_NVC0(chan, NvSubCopy, 0x0418, 1);
+	OUT_RING(chan, length >> 3);
+	BEGIN_NVC0(chan, NvSubCopy, 0x0300, 1);
+	OUT_RING(chan, launch_dma);
+	return 0;
+}
+
 static int
 nouveau_dmem_migrate_init(struct nouveau_drm *drm)
 {
@@ -445,6 +494,7 @@ nouveau_dmem_migrate_init(struct nouveau_drm *drm)
 	case  VOLTA_DMA_COPY_A:
 	case TURING_DMA_COPY_A:
 		drm->dmem->migrate.copy_func = nvc0b5_migrate_copy;
+		drm->dmem->migrate.clear_func = nvc0b5_migrate_clear;
 		drm->dmem->migrate.chan = drm->ttm.chan;
 		return 0;
 	default:
@@ -487,21 +537,28 @@ static unsigned long nouveau_dmem_migrate_copy_one(struct nouveau_drm *drm,
 	unsigned long paddr;
 
 	spage = migrate_pfn_to_page(src);
-	if (!spage || !(src & MIGRATE_PFN_MIGRATE))
+	if (!(src & MIGRATE_PFN_MIGRATE))
 		goto out;
 
 	dpage = nouveau_dmem_page_alloc_locked(drm);
 	if (!dpage)
 		goto out;
 
-	*dma_addr = dma_map_page(dev, spage, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
-	if (dma_mapping_error(dev, *dma_addr))
-		goto out_free_page;
-
 	paddr = nouveau_dmem_page_addr(dpage);
-	if (drm->dmem->migrate.copy_func(drm, 1, NOUVEAU_APER_VRAM,
-			paddr, NOUVEAU_APER_HOST, *dma_addr))
-		goto out_dma_unmap;
+	if (spage) {
+		*dma_addr = dma_map_page(dev, spage, 0, page_size(spage),
+					 DMA_BIDIRECTIONAL);
+		if (dma_mapping_error(dev, *dma_addr))
+			goto out_free_page;
+		if (drm->dmem->migrate.copy_func(drm, page_size(spage),
+			NOUVEAU_APER_VRAM, paddr, NOUVEAU_APER_HOST, *dma_addr))
+			goto out_dma_unmap;
+	} else {
+		*dma_addr = DMA_MAPPING_ERROR;
+		if (drm->dmem->migrate.clear_func(drm, page_size(dpage),
+			NOUVEAU_APER_VRAM, paddr))
+			goto out_free_page;
+	}
 
 	*pfn = NVIF_VMM_PFNMAP_V0_V | NVIF_VMM_PFNMAP_V0_VRAM |
 		((paddr >> PAGE_SHIFT) << NVIF_VMM_PFNMAP_V0_ADDR_SHIFT);
@@ -528,7 +585,7 @@ static void nouveau_dmem_migrate_chunk(struct nouveau_drm *drm,
 	for (i = 0; addr < args->end; i++) {
 		args->dst[i] = nouveau_dmem_migrate_copy_one(drm, args->src[i],
 				dma_addrs + nr_dma, pfns + i);
-		if (args->dst[i])
+		if (!dma_mapping_error(drm->dev->dev, dma_addrs[nr_dma]))
 			nr_dma++;
 		addr += PAGE_SIZE;
 	}
Signature · 1 line
2.20.1
#1 lore raw
Quoted 8 lines · Ralph Campbell

When calling OpenCL clEnqueueSVMMigrateMem() on a region of memory that is backed by pte_none() or zero pages, migrate_vma_setup() will fill the source PFN array with an entry indicating the source page is zero. Use this to optimize migration to device private memory by allocating GPU memory and zero filling it instead of failing to migrate the page.

Quoted 7 lines

This patch applies cleanly to Jason's Gunthorpe's hmm tree plus two patches I posted earlier. The first is queued in Ben Skegg's nouveau tree and the second is still pending review/not queued.
[1] ("nouveau/hmm: map pages after migration")
https://lore.kernel.org/linux-mm/[email protected]/
[2] ("nouveau/hmm: fix nouveau_dmem_chunk allocations")
https://lore.kernel.org/lkml/[email protected]/

It would be best if it goes through Ben's tree if it doesn't have conflicts with the hunks I have in the hmm tree.. Is it the case?

Jason

#2 lore raw
Quoted 17 lines · Jason Gunthorpe
Quoted 8 lines · Ralph Campbell

When calling OpenCL clEnqueueSVMMigrateMem() on a region of memory that is backed by pte_none() or zero pages, migrate_vma_setup() will fill the source PFN array with an entry indicating the source page is zero. Use this to optimize migration to device private memory by allocating GPU memory and zero filling it instead of failing to migrate the page.

Quoted 7 lines

This patch applies cleanly to Jason's Gunthorpe's hmm tree plus two patches I posted earlier. The first is queued in Ben Skegg's nouveau tree and the second is still pending review/not queued.
[1] ("nouveau/hmm: map pages after migration")
https://lore.kernel.org/linux-mm/[email protected]/
[2] ("nouveau/hmm: fix nouveau_dmem_chunk allocations")
https://lore.kernel.org/lkml/[email protected]/

Quoted 4 lines

It would be best if it goes through Ben's tree if it doesn't have conflicts with the hunks I have in the hmm tree.. Is it the case?

Jason

I think there might be some merge conflicts even though it is semantically independent of the other changes. I guess since we are at 5.7-rc6 and not far from the merge window, I can rebase after 5.8-rc1 and resend. I posted this mostly to get some review and as a "heads up" of the issue.

#3 lore raw
Quoted 23 lines · Ralph Campbell
Quoted 20 lines · Jason Gunthorpe
Quoted 8 lines · Ralph Campbell

When calling OpenCL clEnqueueSVMMigrateMem() on a region of memory that is backed by pte_none() or zero pages, migrate_vma_setup() will fill the source PFN array with an entry indicating the source page is zero. Use this to optimize migration to device private memory by allocating GPU memory and zero filling it instead of failing to migrate the page.

Quoted 7 lines

This patch applies cleanly to Jason's Gunthorpe's hmm tree plus two patches I posted earlier. The first is queued in Ben Skegg's nouveau tree and the second is still pending review/not queued.
[1] ("nouveau/hmm: map pages after migration")
https://lore.kernel.org/linux-mm/20200304001339.8248-5-rcampbell-DDmLM1+adcrQT0dZR+AlfA@public.gmane.org/
[2] ("nouveau/hmm: fix nouveau_dmem_chunk allocations")
https://lore.kernel.org/lkml/20200421231107.30958-1-rcampbell-DDmLM1+adcrQT0dZR+AlfA@public.gmane.org/

It would be best if it goes through Ben's tree if it doesn't have conflicts with the hunks I have in the hmm tree.. Is it the case?

Quoted 6 lines
Quoted 1 line

Jason

I think there might be some merge conflicts even though it is semantically independent of the other changes. I guess since we are at 5.7-rc6 and not far from the merge window, I can rebase after 5.8-rc1 and resend. I posted this mostly to get some review and as a "heads up" of the issue.

Both look alright to me, and apply cleanly on top of my tree already, so I've got them locally.

Ben.

Quoted 4 lines

_______________________________________________
Nouveau mailing list [email protected]
https://lists.freedesktop.org/mailman/listinfo/nouveau