drm/i915/bdw: Populate LR contexts (somewhat)
authorOscar Mateo <oscar.mateo@intel.com>
Thu, 24 Jul 2014 16:04:17 +0000 (17:04 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Mon, 11 Aug 2014 14:21:53 +0000 (16:21 +0200)
For the most part, logical ring context objects are similar to hardware
contexts in that the backing object is meant to be opaque. There are
some exceptions where we need to poke certain offsets of the object for
initialization, updating the tail pointer or updating the PDPs.

For our basic execlist implementation we'll only need our PPGTT PDs,
and ringbuffer addresses in order to set up the context. With previous
patches, we have both, so start prepping the context to be load.

Before running a context for the first time you must populate some
fields in the context object. These fields begin 1 PAGE + LRCA, ie. the
first page (in 0 based counting) of the context  image. These same
fields will be read and written to as contexts are saved and restored
once the system is up and running.

Many of these fields are completely reused from previous global
registers: ringbuffer head/tail/control, context control matches some
previous MI_SET_CONTEXT flags, and page directories. There are other
fields which we don't touch which we may want in the future.

v2: CTX_LRI_HEADER_0 is MI_LOAD_REGISTER_IMM(14) for render and (11)
for other engines.

v3: Several rebases and general changes to the code.

v4: Squash with "Extract LR context object populating"
Also, Damien's review comments:
- Set the Force Posted bit on the LRI header, as the BSpec suggest we do.
- Prevent warning when compiling a 32-bits kernel without HIGHMEM64.
- Add a clarifying comment to the context population code.

v5: Damien's review comments:
- The third MI_LOAD_REGISTER_IMM in the context does not set Force Posted.
- Remove dead code.

v6: Add a note about the (presumed) differences between BDW and CHV state
contexts. Also, Brad's review comments:
- Use the _MASKED_BIT_ENABLE, upper_32_bits and lower_32_bits macros.
- Be less magical about how we set the ring size in the context.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net> (v1)
Signed-off-by: Rafael Barbalho <rafael.barbalho@intel.com> (v2)
Signed-off-by: Oscar Mateo <oscar.mateo@intel.com>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_reg.h
drivers/gpu/drm/i915/intel_lrc.c

index 7a6cc69cdc2b6eb6a8e97b95f82d27130e1d51d2..c1d24242a02d282c5eab3581e81c0ff100bd6a19 100644 (file)
  *   address/value pairs. Don't overdue it, though, x <= 2^4 must hold!
  */
 #define MI_LOAD_REGISTER_IMM(x)        MI_INSTR(0x22, 2*(x)-1)
+#define   MI_LRI_FORCE_POSTED          (1<<12)
 #define MI_STORE_REGISTER_MEM(x) MI_INSTR(0x24, 2*(x)-1)
 #define MI_STORE_REGISTER_MEM_GEN8(x) MI_INSTR(0x24, 3*(x)-1)
 #define   MI_SRM_LRM_GLOBAL_GTT                (1<<22)
index 8f2d14da6228cc40078fc1882906ebab0aa3ce2e..a7a08a85edb36b1aa4e28f78851a385ded8ac0d1 100644 (file)
 
 #define GEN8_LR_CONTEXT_ALIGN 4096
 
+#define RING_ELSP(ring)                        ((ring)->mmio_base+0x230)
+#define RING_CONTEXT_CONTROL(ring)     ((ring)->mmio_base+0x244)
+
+#define CTX_LRI_HEADER_0               0x01
+#define CTX_CONTEXT_CONTROL            0x02
+#define CTX_RING_HEAD                  0x04
+#define CTX_RING_TAIL                  0x06
+#define CTX_RING_BUFFER_START          0x08
+#define CTX_RING_BUFFER_CONTROL                0x0a
+#define CTX_BB_HEAD_U                  0x0c
+#define CTX_BB_HEAD_L                  0x0e
+#define CTX_BB_STATE                   0x10
+#define CTX_SECOND_BB_HEAD_U           0x12
+#define CTX_SECOND_BB_HEAD_L           0x14
+#define CTX_SECOND_BB_STATE            0x16
+#define CTX_BB_PER_CTX_PTR             0x18
+#define CTX_RCS_INDIRECT_CTX           0x1a
+#define CTX_RCS_INDIRECT_CTX_OFFSET    0x1c
+#define CTX_LRI_HEADER_1               0x21
+#define CTX_CTX_TIMESTAMP              0x22
+#define CTX_PDP3_UDW                   0x24
+#define CTX_PDP3_LDW                   0x26
+#define CTX_PDP2_UDW                   0x28
+#define CTX_PDP2_LDW                   0x2a
+#define CTX_PDP1_UDW                   0x2c
+#define CTX_PDP1_LDW                   0x2e
+#define CTX_PDP0_UDW                   0x30
+#define CTX_PDP0_LDW                   0x32
+#define CTX_LRI_HEADER_2               0x41
+#define CTX_R_PWR_CLK_STATE            0x42
+#define CTX_GPGPU_CSR_BASE_ADDRESS     0x44
+
 int intel_sanitize_enable_execlists(struct drm_device *dev, int enable_execlists)
 {
        WARN_ON(i915.enable_ppgtt == -1);
@@ -59,6 +91,115 @@ int intel_sanitize_enable_execlists(struct drm_device *dev, int enable_execlists
        return 0;
 }
 
+static int
+populate_lr_context(struct intel_context *ctx, struct drm_i915_gem_object *ctx_obj,
+                   struct intel_engine_cs *ring, struct intel_ringbuffer *ringbuf)
+{
+       struct drm_i915_gem_object *ring_obj = ringbuf->obj;
+       struct i915_hw_ppgtt *ppgtt = ctx_to_ppgtt(ctx);
+       struct page *page;
+       uint32_t *reg_state;
+       int ret;
+
+       ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
+       if (ret) {
+               DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
+               return ret;
+       }
+
+       ret = i915_gem_object_get_pages(ctx_obj);
+       if (ret) {
+               DRM_DEBUG_DRIVER("Could not get object pages\n");
+               return ret;
+       }
+
+       i915_gem_object_pin_pages(ctx_obj);
+
+       /* The second page of the context object contains some fields which must
+        * be set up prior to the first execution. */
+       page = i915_gem_object_get_page(ctx_obj, 1);
+       reg_state = kmap_atomic(page);
+
+       /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
+        * commands followed by (reg, value) pairs. The values we are setting here are
+        * only for the first context restore: on a subsequent save, the GPU will
+        * recreate this batchbuffer with new values (including all the missing
+        * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
+       if (ring->id == RCS)
+               reg_state[CTX_LRI_HEADER_0] = MI_LOAD_REGISTER_IMM(14);
+       else
+               reg_state[CTX_LRI_HEADER_0] = MI_LOAD_REGISTER_IMM(11);
+       reg_state[CTX_LRI_HEADER_0] |= MI_LRI_FORCE_POSTED;
+       reg_state[CTX_CONTEXT_CONTROL] = RING_CONTEXT_CONTROL(ring);
+       reg_state[CTX_CONTEXT_CONTROL+1] =
+                       _MASKED_BIT_ENABLE((1<<3) | MI_RESTORE_INHIBIT);
+       reg_state[CTX_RING_HEAD] = RING_HEAD(ring->mmio_base);
+       reg_state[CTX_RING_HEAD+1] = 0;
+       reg_state[CTX_RING_TAIL] = RING_TAIL(ring->mmio_base);
+       reg_state[CTX_RING_TAIL+1] = 0;
+       reg_state[CTX_RING_BUFFER_START] = RING_START(ring->mmio_base);
+       reg_state[CTX_RING_BUFFER_START+1] = i915_gem_obj_ggtt_offset(ring_obj);
+       reg_state[CTX_RING_BUFFER_CONTROL] = RING_CTL(ring->mmio_base);
+       reg_state[CTX_RING_BUFFER_CONTROL+1] =
+                       ((ringbuf->size - PAGE_SIZE) & RING_NR_PAGES) | RING_VALID;
+       reg_state[CTX_BB_HEAD_U] = ring->mmio_base + 0x168;
+       reg_state[CTX_BB_HEAD_U+1] = 0;
+       reg_state[CTX_BB_HEAD_L] = ring->mmio_base + 0x140;
+       reg_state[CTX_BB_HEAD_L+1] = 0;
+       reg_state[CTX_BB_STATE] = ring->mmio_base + 0x110;
+       reg_state[CTX_BB_STATE+1] = (1<<5);
+       reg_state[CTX_SECOND_BB_HEAD_U] = ring->mmio_base + 0x11c;
+       reg_state[CTX_SECOND_BB_HEAD_U+1] = 0;
+       reg_state[CTX_SECOND_BB_HEAD_L] = ring->mmio_base + 0x114;
+       reg_state[CTX_SECOND_BB_HEAD_L+1] = 0;
+       reg_state[CTX_SECOND_BB_STATE] = ring->mmio_base + 0x118;
+       reg_state[CTX_SECOND_BB_STATE+1] = 0;
+       if (ring->id == RCS) {
+               /* TODO: according to BSpec, the register state context
+                * for CHV does not have these. OTOH, these registers do
+                * exist in CHV. I'm waiting for a clarification */
+               reg_state[CTX_BB_PER_CTX_PTR] = ring->mmio_base + 0x1c0;
+               reg_state[CTX_BB_PER_CTX_PTR+1] = 0;
+               reg_state[CTX_RCS_INDIRECT_CTX] = ring->mmio_base + 0x1c4;
+               reg_state[CTX_RCS_INDIRECT_CTX+1] = 0;
+               reg_state[CTX_RCS_INDIRECT_CTX_OFFSET] = ring->mmio_base + 0x1c8;
+               reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] = 0;
+       }
+       reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9);
+       reg_state[CTX_LRI_HEADER_1] |= MI_LRI_FORCE_POSTED;
+       reg_state[CTX_CTX_TIMESTAMP] = ring->mmio_base + 0x3a8;
+       reg_state[CTX_CTX_TIMESTAMP+1] = 0;
+       reg_state[CTX_PDP3_UDW] = GEN8_RING_PDP_UDW(ring, 3);
+       reg_state[CTX_PDP3_LDW] = GEN8_RING_PDP_LDW(ring, 3);
+       reg_state[CTX_PDP2_UDW] = GEN8_RING_PDP_UDW(ring, 2);
+       reg_state[CTX_PDP2_LDW] = GEN8_RING_PDP_LDW(ring, 2);
+       reg_state[CTX_PDP1_UDW] = GEN8_RING_PDP_UDW(ring, 1);
+       reg_state[CTX_PDP1_LDW] = GEN8_RING_PDP_LDW(ring, 1);
+       reg_state[CTX_PDP0_UDW] = GEN8_RING_PDP_UDW(ring, 0);
+       reg_state[CTX_PDP0_LDW] = GEN8_RING_PDP_LDW(ring, 0);
+       reg_state[CTX_PDP3_UDW+1] = upper_32_bits(ppgtt->pd_dma_addr[3]);
+       reg_state[CTX_PDP3_LDW+1] = lower_32_bits(ppgtt->pd_dma_addr[3]);
+       reg_state[CTX_PDP2_UDW+1] = upper_32_bits(ppgtt->pd_dma_addr[2]);
+       reg_state[CTX_PDP2_LDW+1] = lower_32_bits(ppgtt->pd_dma_addr[2]);
+       reg_state[CTX_PDP1_UDW+1] = upper_32_bits(ppgtt->pd_dma_addr[1]);
+       reg_state[CTX_PDP1_LDW+1] = lower_32_bits(ppgtt->pd_dma_addr[1]);
+       reg_state[CTX_PDP0_UDW+1] = upper_32_bits(ppgtt->pd_dma_addr[0]);
+       reg_state[CTX_PDP0_LDW+1] = lower_32_bits(ppgtt->pd_dma_addr[0]);
+       if (ring->id == RCS) {
+               reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
+               reg_state[CTX_R_PWR_CLK_STATE] = 0x20c8;
+               reg_state[CTX_R_PWR_CLK_STATE+1] = 0;
+       }
+
+       kunmap_atomic(reg_state);
+
+       ctx_obj->dirty = 1;
+       set_page_dirty(page);
+       i915_gem_object_unpin_pages(ctx_obj);
+
+       return 0;
+}
+
 void intel_lr_context_free(struct intel_context *ctx)
 {
        int i;
@@ -151,14 +292,24 @@ int intel_lr_context_deferred_create(struct intel_context *ctx,
        if (ret) {
                DRM_DEBUG_DRIVER("Failed to allocate ringbuffer obj %s: %d\n",
                                ring->name, ret);
-               kfree(ringbuf);
-               i915_gem_object_ggtt_unpin(ctx_obj);
-               drm_gem_object_unreference(&ctx_obj->base);
-               return ret;
+               goto error;
+       }
+
+       ret = populate_lr_context(ctx, ctx_obj, ring, ringbuf);
+       if (ret) {
+               DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
+               intel_destroy_ringbuffer_obj(ringbuf);
+               goto error;
        }
 
        ctx->engine[ring->id].ringbuf = ringbuf;
        ctx->engine[ring->id].state = ctx_obj;
 
        return 0;
+
+error:
+       kfree(ringbuf);
+       i915_gem_object_ggtt_unpin(ctx_obj);
+       drm_gem_object_unreference(&ctx_obj->base);
+       return ret;
 }