drm/radeon/kms: add 6xx/7xx CS parser for async DMA (v2)
Alex Deucher [Fri, 18 Nov 2011 15:19:47 +0000 (10:19 -0500)]
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.

v2: fix address bounds checking, reloc indexing

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/radeon/r600_cs.c
drivers/gpu/drm/radeon/radeon.h
drivers/gpu/drm/radeon/radeon_asic.c
drivers/gpu/drm/radeon/radeon_asic.h
drivers/gpu/drm/radeon/radeon_cs.c

index 0b4d833..0be768b 100644 (file)
@@ -2514,3 +2514,196 @@ 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;
+       }
+       *cs_reloc = p->relocs_ptr[idx];
+       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) {
+                               dst_offset = ib[idx+1];
+                               dst_offset <<= 8;
+
+                               ib[idx+1] += (u32)(dst_reloc->lobj.gpu_offset >> 8);
+                               p->idx += count + 5;
+                       } else {
+                               dst_offset = ib[idx+1];
+                               dst_offset |= ((u64)(ib[idx+2] & 0xff)) << 32;
+
+                               ib[idx+1] += (u32)(dst_reloc->lobj.gpu_offset & 0xfffffffc);
+                               ib[idx+2] += upper_32_bits(dst_reloc->lobj.gpu_offset) & 0xff;
+                               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 + (count * 4), 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 */
+                                       src_offset = ib[idx+1];
+                                       src_offset <<= 8;
+                                       ib[idx+1] += (u32)(src_reloc->lobj.gpu_offset >> 8);
+
+                                       dst_offset = ib[idx+5];
+                                       dst_offset |= ((u64)(ib[idx+6] & 0xff)) << 32;
+                                       ib[idx+5] += (u32)(dst_reloc->lobj.gpu_offset & 0xfffffffc);
+                                       ib[idx+6] += upper_32_bits(dst_reloc->lobj.gpu_offset) & 0xff;
+                               } else {
+                                       /* linear src, tiled dst */
+                                       src_offset = ib[idx+5];
+                                       src_offset |= ((u64)(ib[idx+6] & 0xff)) << 32;
+                                       ib[idx+5] += (u32)(src_reloc->lobj.gpu_offset & 0xfffffffc);
+                                       ib[idx+6] += upper_32_bits(src_reloc->lobj.gpu_offset) & 0xff;
+
+                                       dst_offset = ib[idx+1];
+                                       dst_offset <<= 8;
+                                       ib[idx+1] += (u32)(dst_reloc->lobj.gpu_offset >> 8);
+                               }
+                               p->idx += 7;
+                       } else {
+                               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;
+
+                               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;
+                               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 + (count * 4), 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 + (count * 4), 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;
+                       }
+                       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 + (count * 4), radeon_bo_size(dst_reloc->robj));
+                               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;
+                       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;
+}
index 285fb3f..5dc744d 100644 (file)
@@ -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;
index 3ea0475..d360341 100644 (file)
@@ -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,
index c338931..b311c0a 100644 (file)
@@ -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,
index 41672cc..1b32a5a 100644 (file)
@@ -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);