New subject: [PATCH 2/4] drm/radeon/kms: add evergreen/cayman CS parser for async DMA

13 Dec 2012

From: Alex Deucher alexander.deucher@amd.com
Allows us to use the DMA ring from userspace.
DMA doesn't have a good NOP packet in which to embed the
reloc idx, so userspace has to add a reloc for each
buffer used and order them to match the command stream.
Signed-off-by: Alex Deucher alexander.deucher@amd.com
---
 drivers/gpu/drm/radeon/r600_cs.c     |  193 ++++++++++++++++++++++++++++++++++
 drivers/gpu/drm/radeon/radeon.h      |    1 +
 drivers/gpu/drm/radeon/radeon_asic.c |    6 +-
 drivers/gpu/drm/radeon/radeon_asic.h |    1 +
 drivers/gpu/drm/radeon/radeon_cs.c   |    1 +
 5 files changed, 199 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/radeon/r600_cs.c b/drivers/gpu/drm/radeon/r600_cs.c
index 5d6e7f9..f23609ac 100644
--- a/drivers/gpu/drm/radeon/r600_cs.c
+++ b/drivers/gpu/drm/radeon/r600_cs.c
@@ -2568,3 +2568,196 @@ void r600_cs_legacy_init(void)
 {
    r600_cs_packet_next_reloc = &r600_cs_packet_next_reloc_nomm;
 }
+
+/*
+ *  DMA
+ */
+/**
+ * r600_dma_cs_next_reloc() - parse next reloc
+ * @p:		parser structure holding parsing context.
+ * @cs_reloc:		reloc informations
+ *
+ * Return the next reloc, do bo validation and compute
+ * GPU offset using the provided start.
+ **/
+int r600_dma_cs_next_reloc(struct radeon_cs_parser *p,
+			   struct radeon_cs_reloc **cs_reloc)
+{
+	struct radeon_cs_chunk *relocs_chunk;
+	unsigned idx;
+
+	if (p->chunk_relocs_idx == -1) {
+		DRM_ERROR("No relocation chunk !\n");
+		return -EINVAL;
+	}
+	*cs_reloc = NULL;
+	relocs_chunk = &p->chunks[p->chunk_relocs_idx];
+	idx = p->dma_reloc_idx;
+	if (idx >= relocs_chunk->length_dw) {
+		DRM_ERROR("Relocs at %d after relocations chunk end %d !\n",
+			  idx, relocs_chunk->length_dw);
+		return -EINVAL;
+	}
+	/* FIXME: we assume reloc size is 4 dwords */
+	*cs_reloc = p->relocs_ptr[(idx / 4)];
+	p->dma_reloc_idx++;
+	return 0;
+}
+
+#define GET_DMA_CMD(h) (((h) & 0xf0000000) >> 28)
+#define GET_DMA_COUNT(h) ((h) & 0x0000ffff)
+#define GET_DMA_T(h) (((h) & 0x00800000) >> 23)
+
+/**
+ * r600_dma_cs_parse() - parse the DMA IB
+ * @p:		parser structure holding parsing context.
+ *
+ * Parses the DMA IB from the CS ioctl and updates
+ * the GPU addresses based on the reloc information and
+ * checks for errors. (R6xx-R7xx)
+ * Returns 0 for success and an error on failure.
+ **/
+int r600_dma_cs_parse(struct radeon_cs_parser *p)
+{
+	struct radeon_cs_chunk *ib_chunk = &p->chunks[p->chunk_ib_idx];
+	struct radeon_cs_reloc *src_reloc, *dst_reloc;
+	u32 header, cmd, count, tiled;
+	volatile u32 *ib = p->ib.ptr;
+	u32 idx, idx_value;
+	u64 src_offset, dst_offset;
+	int r;
+
+	do {
+		if (p->idx >= ib_chunk->length_dw) {
+			DRM_ERROR("Can not parse packet at %d after CS end %d !\n",
+				  p->idx, ib_chunk->length_dw);
+			return -EINVAL;
+		}
+		idx = p->idx;
+		header = radeon_get_ib_value(p, idx);
+		cmd = GET_DMA_CMD(header);
+		count = GET_DMA_COUNT(header);
+		tiled = GET_DMA_T(header);
+
+		switch (cmd) {
+		case DMA_PACKET_WRITE:
+			r = r600_dma_cs_next_reloc(p, &dst_reloc);
+			if (r) {
+				DRM_ERROR("bad DMA_PACKET_WRITE\n");
+				return -EINVAL;
+			}
+			if (tiled) {
+				ib[idx+1] += (u32)(dst_reloc->lobj.gpu_offset >> 8);
+				dst_offset = ib[idx+1];
+				dst_offset <<= 8;
+				p->idx += count + 5;
+			} else {
+				ib[idx+1] += (u32)(dst_reloc->lobj.gpu_offset & 0xfffffffc);
+				ib[idx+2] += upper_32_bits(dst_reloc->lobj.gpu_offset) & 0xff;
+				dst_offset = ib[idx+1];
+				dst_offset |= ((u64)(ib[idx+2] & 0xff)) << 32;
+				p->idx += count + 3;
+			}
+			if ((dst_offset + (count * 4)) > radeon_bo_size(dst_reloc->robj)) {
+				dev_warn(p->dev, "DMA write buffer too small (%llu %lu)\n",
+					 dst_offset, radeon_bo_size(dst_reloc->robj));
+				return -EINVAL;
+			}
+			break;
+		case DMA_PACKET_COPY:
+			r = r600_dma_cs_next_reloc(p, &src_reloc);
+			if (r) {
+				DRM_ERROR("bad DMA_PACKET_COPY\n");
+				return -EINVAL;
+			}
+			r = r600_dma_cs_next_reloc(p, &dst_reloc);
+			if (r) {
+				DRM_ERROR("bad DMA_PACKET_COPY\n");
+				return -EINVAL;
+			}
+			if (tiled) {
+				idx_value = radeon_get_ib_value(p, idx + 2);
+				/* detile bit */
+				if (idx_value & (1 << 31)) {
+					/* tiled src, linear dst */
+					ib[idx+1] += (u32)(src_reloc->lobj.gpu_offset >> 8);
+					src_offset = ib[idx+1];
+					src_offset <<= 8;
+
+					ib[idx+5] += (u32)(dst_reloc->lobj.gpu_offset & 0xfffffffc);
+					ib[idx+6] += upper_32_bits(dst_reloc->lobj.gpu_offset) & 0xff;
+					dst_offset = ib[idx+5];
+					dst_offset |= ((u64)(ib[idx+6] & 0xff)) << 32;
+				} else {
+					/* linear src, tiled dst */
+					ib[idx+5] += (u32)(src_reloc->lobj.gpu_offset & 0xfffffffc);
+					ib[idx+6] += upper_32_bits(src_reloc->lobj.gpu_offset) & 0xff;
+					src_offset = ib[idx+5];
+					src_offset |= ((u64)(ib[idx+6] & 0xff)) << 32;
+
+					ib[idx+1] += (u32)(dst_reloc->lobj.gpu_offset >> 8);
+					dst_offset = ib[idx+1];
+					dst_offset <<= 8;
+				}
+				p->idx += 7;
+			} else {
+				ib[idx+1] += (u32)(dst_reloc->lobj.gpu_offset & 0xfffffffc);
+				ib[idx+2] += (u32)(src_reloc->lobj.gpu_offset & 0xfffffffc);
+				ib[idx+3] += upper_32_bits(dst_reloc->lobj.gpu_offset) & 0xff;
+				ib[idx+4] += upper_32_bits(src_reloc->lobj.gpu_offset) & 0xff;
+
+				src_offset = ib[idx+2];
+				src_offset |= ((u64)(ib[idx+4] & 0xff)) << 32;
+				dst_offset = ib[idx+1];
+				dst_offset |= ((u64)(ib[idx+3] & 0xff)) << 32;
+
+				p->idx += 5;
+			}
+			if ((src_offset + (count * 4)) > radeon_bo_size(src_reloc->robj)) {
+				dev_warn(p->dev, "DMA copy src buffer too small (%llu %lu)\n",
+					 src_offset, radeon_bo_size(src_reloc->robj));
+				return -EINVAL;
+			}
+			if ((dst_offset + (count * 4)) > radeon_bo_size(dst_reloc->robj)) {
+				dev_warn(p->dev, "DMA write dst buffer too small (%llu %lu)\n",
+					 dst_offset, radeon_bo_size(dst_reloc->robj));
+				return -EINVAL;
+			}
+			break;
+		case DMA_PACKET_CONSTANT_FILL:
+			if (p->family < CHIP_RV770) {
+				DRM_ERROR("Constant Fill is 7xx only !\n");
+				return -EINVAL;
+			}
+			r = r600_dma_cs_next_reloc(p, &dst_reloc);
+			if (r) {
+				DRM_ERROR("bad DMA_PACKET_WRITE\n");
+				return -EINVAL;
+			}
+			ib[idx+1] += (u32)(dst_reloc->lobj.gpu_offset & 0xfffffffc);
+			ib[idx+3] += (upper_32_bits(dst_reloc->lobj.gpu_offset) << 16) & 0x00ff0000;
+			dst_offset = ib[idx+1];
+			dst_offset |= ((u64)(ib[idx+3] & 0x00ff0000)) << 16;
+			if ((dst_offset + (count * 4)) > radeon_bo_size(dst_reloc->robj)) {
+				dev_warn(p->dev, "DMA constant fill buffer too small (%llu %lu)\n",
+					 dst_offset, radeon_bo_size(dst_reloc->robj));
+				return -EINVAL;
+			}
+			p->idx += 4;
+			break;
+		case DMA_PACKET_NOP:
+			p->idx += 1;
+			break;
+		default:
+			DRM_ERROR("Unknown packet type %d at %d !\n", cmd, idx);
+			return -EINVAL;
+		}
+	} while (p->idx < p->chunks[p->chunk_ib_idx].length_dw);
+#if 0
+	for (r = 0; r < p->ib->length_dw; r++) {
+		printk(KERN_INFO "%05d  0x%08X\n", r, p->ib.ptr[r]);
+		mdelay(1);
+	}
+#endif
+	return 0;
+}
diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index 285fb3f..5dc744d 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -839,6 +839,7 @@ struct radeon_cs_parser {
    struct radeon_cs_reloc	*relocs;
    struct radeon_cs_reloc	**relocs_ptr;
    struct list_head	validated;
+	unsigned		dma_reloc_idx;
    /* indices of various chunks */
    int			chunk_ib_idx;
    int			chunk_relocs_idx;
diff --git a/drivers/gpu/drm/radeon/radeon_asic.c b/drivers/gpu/drm/radeon/radeon_asic.c
index 3ea0475..d360341 100644
--- a/drivers/gpu/drm/radeon/radeon_asic.c
+++ b/drivers/gpu/drm/radeon/radeon_asic.c
@@ -952,7 +952,7 @@ static struct radeon_asic r600_asic = {
    		.ib_execute = &r600_dma_ring_ib_execute,
    		.emit_fence = &r600_dma_fence_ring_emit,
    		.emit_semaphore = &r600_dma_semaphore_ring_emit,
-			.cs_parse = NULL,
+			.cs_parse = &r600_dma_cs_parse,
    		.ring_test = &r600_dma_ring_test,
    		.ib_test = &r600_dma_ib_test,
    		.is_lockup = &r600_dma_is_lockup,
@@ -1036,7 +1036,7 @@ static struct radeon_asic rs780_asic = {
    		.ib_execute = &r600_dma_ring_ib_execute,
    		.emit_fence = &r600_dma_fence_ring_emit,
    		.emit_semaphore = &r600_dma_semaphore_ring_emit,
-			.cs_parse = NULL,
+			.cs_parse = &r600_dma_cs_parse,
    		.ring_test = &r600_dma_ring_test,
    		.ib_test = &r600_dma_ib_test,
    		.is_lockup = &r600_dma_is_lockup,
@@ -1120,7 +1120,7 @@ static struct radeon_asic rv770_asic = {
    		.ib_execute = &r600_dma_ring_ib_execute,
    		.emit_fence = &r600_dma_fence_ring_emit,
    		.emit_semaphore = &r600_dma_semaphore_ring_emit,
-			.cs_parse = NULL,
+			.cs_parse = &r600_dma_cs_parse,
    		.ring_test = &r600_dma_ring_test,
    		.ib_test = &r600_dma_ib_test,
    		.is_lockup = &r600_dma_is_lockup,
diff --git a/drivers/gpu/drm/radeon/radeon_asic.h b/drivers/gpu/drm/radeon/radeon_asic.h
index c338931..b311c0a 100644
--- a/drivers/gpu/drm/radeon/radeon_asic.h
+++ b/drivers/gpu/drm/radeon/radeon_asic.h
@@ -304,6 +304,7 @@ void r600_pcie_gart_tlb_flush(struct radeon_device *rdev);
 uint32_t r600_pciep_rreg(struct radeon_device *rdev, uint32_t reg);
 void r600_pciep_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v);
 int r600_cs_parse(struct radeon_cs_parser *p);
+int r600_dma_cs_parse(struct radeon_cs_parser *p);
 void r600_fence_ring_emit(struct radeon_device *rdev,
    		  struct radeon_fence *fence);
 void r600_semaphore_ring_emit(struct radeon_device *rdev,
diff --git a/drivers/gpu/drm/radeon/radeon_cs.c b/drivers/gpu/drm/radeon/radeon_cs.c
index 41672cc..1b32a5a 100644
--- a/drivers/gpu/drm/radeon/radeon_cs.c
+++ b/drivers/gpu/drm/radeon/radeon_cs.c
@@ -43,6 +43,7 @@ static int radeon_cs_parser_relocs(struct radeon_cs_parser *p)
    	return 0;
    }
    chunk = &p->chunks[p->chunk_relocs_idx];
+	p->dma_reloc_idx = 0;
    /* FIXME: we assume that each relocs use 4 dwords */
    p->nrelocs = chunk->length_dw / 4;
    p->relocs_ptr = kcalloc(p->nrelocs, sizeof(void *), GFP_KERNEL);
-- 
1.7.7.5


    

[PATCH 1/4] drm/radeon/kms: add 6xx/7xx CS parser for async DMA