#include "intel_batchbuffer.h"
#include "intel_ioctl.h"
#include "intel_decode.h"
-#include "i915_debug.h"
+#include "intel_reg.h"
/* Relocations in kernel space:
* - pass dma buffer seperately
batch->buf = NULL;
}
- batch->buf = dri_bo_alloc(intel->intelScreen->bufmgr, "batchbuffer",
- intel->intelScreen->maxBatchSize, 4096,
- DRM_BO_FLAG_MEM_TT);
- dri_bo_map(batch->buf, GL_TRUE);
- batch->map = batch->buf->virtual;
- batch->size = intel->intelScreen->maxBatchSize;
+ if (!batch->buffer && intel->ttm == GL_TRUE)
+ batch->buffer = malloc (intel->maxBatchSize);
+
+ batch->buf = dri_bo_alloc(intel->bufmgr, "batchbuffer",
+ intel->maxBatchSize, 4096,
+ DRM_BO_FLAG_MEM_LOCAL | DRM_BO_FLAG_CACHED | DRM_BO_FLAG_CACHED_MAPPED);
+ if (batch->buffer)
+ batch->map = batch->buffer;
+ else {
+ dri_bo_map(batch->buf, GL_TRUE);
+ batch->map = batch->buf->virtual;
+ }
+ batch->size = intel->maxBatchSize;
batch->ptr = batch->map;
+ batch->dirty_state = ~0;
+ batch->cliprect_mode = IGNORE_CLIPRECTS;
+
+ /* account batchbuffer in aperture */
+ dri_bufmgr_check_aperture_space(batch->buf);
+
}
struct intel_batchbuffer *
struct intel_batchbuffer *batch = calloc(sizeof(*batch), 1);
batch->intel = intel;
- batch->last_fence = NULL;
intel_batchbuffer_reset(batch);
return batch;
void
intel_batchbuffer_free(struct intel_batchbuffer *batch)
{
- if (batch->last_fence) {
- dri_fence_wait(batch->last_fence);
- dri_fence_unreference(batch->last_fence);
- batch->last_fence = NULL;
- }
- if (batch->map) {
- dri_bo_unmap(batch->buf);
- batch->map = NULL;
+ if (batch->buffer)
+ free (batch->buffer);
+ else {
+ if (batch->map) {
+ dri_bo_unmap(batch->buf);
+ batch->map = NULL;
+ }
}
dri_bo_unreference(batch->buf);
batch->buf = NULL;
free(batch);
}
-static int
-relocation_sort(const void *a_in, const void *b_in) {
- const struct buffer_reloc *a = a_in, *b = b_in;
-
- return (intptr_t)a->buf < (intptr_t)b->buf ? -1 : 1;
-}
/* TODO: Push this whole function into bufmgr.
*/
static void
do_flush_locked(struct intel_batchbuffer *batch,
- GLuint used,
- GLboolean ignore_cliprects, GLboolean allow_unlock)
+ GLuint used, GLboolean allow_unlock)
{
- GLuint *ptr;
- GLuint i;
struct intel_context *intel = batch->intel;
- dri_fence *fo;
- GLboolean performed_rendering = GL_FALSE;
-
- assert(batch->buf->virtual != NULL);
- ptr = batch->buf->virtual;
-
- /* Sort our relocation list in terms of referenced buffer pointer.
- * This lets us uniquely validate the buffers with the sum of all the flags,
- * while avoiding O(n^2) on number of relocations.
- */
- qsort(batch->reloc, batch->nr_relocs, sizeof(batch->reloc[0]),
- relocation_sort);
+ int ret = 0;
- /* Perform the necessary validations of buffers, and enter the relocations
- * in the batchbuffer.
- */
- for (i = 0; i < batch->nr_relocs; i++) {
- struct buffer_reloc *r = &batch->reloc[i];
-
- if (r->validate_flags & DRM_BO_FLAG_WRITE)
- performed_rendering = GL_TRUE;
-
- /* If this is the first time we've seen this buffer in the relocation
- * list, figure out our flags and validate it.
- */
- if (i == 0 || batch->reloc[i - 1].buf != r->buf) {
- uint32_t validate_flags;
- int j, ret;
-
- /* Accumulate the flags we need for validating this buffer. */
- validate_flags = r->validate_flags;
- for (j = i + 1; j < batch->nr_relocs; j++) {
- if (batch->reloc[j].buf != r->buf)
- break;
- validate_flags |= batch->reloc[j].validate_flags;
- }
-
- /* Validate. If we fail, fence to clear the unfenced list and bail
- * out.
- */
- ret = dri_bo_validate(r->buf, validate_flags);
- if (ret != 0) {
- dri_bo_unmap(batch->buf);
- fo = dri_fence_validated(intel->intelScreen->bufmgr,
- "batchbuffer failure fence", GL_TRUE);
- dri_fence_unreference(fo);
- goto done;
- }
- }
- ptr[r->offset / 4] = r->buf->offset + r->delta;
- dri_bo_unreference(r->buf);
- }
+ if (batch->buffer)
+ dri_bo_subdata (batch->buf, 0, used, batch->buffer);
+ else
+ dri_bo_unmap(batch->buf);
- dri_bo_unmap(batch->buf);
batch->map = NULL;
batch->ptr = NULL;
- dri_bo_validate(batch->buf, DRM_BO_FLAG_MEM_TT | DRM_BO_FLAG_EXE);
-
- batch->list_count = 0;
- batch->nr_relocs = 0;
- batch->flags = 0;
-
/* Throw away non-effective packets. Won't work once we have
* hardware contexts which would preserve statechanges beyond a
* single buffer.
*/
- if (!(intel->numClipRects == 0 && !ignore_cliprects)) {
- intel_batch_ioctl(batch->intel,
- batch->buf->offset,
- used, ignore_cliprects, allow_unlock);
+ if (!(intel->numClipRects == 0 &&
+ batch->cliprect_mode == LOOP_CLIPRECTS)) {
+ if (intel->ttm == GL_TRUE) {
+ struct drm_i915_gem_execbuffer *execbuf;
+
+ execbuf = dri_process_relocs(batch->buf);
+ ret = intel_exec_ioctl(batch->intel,
+ used,
+ batch->cliprect_mode != LOOP_CLIPRECTS,
+ allow_unlock,
+ execbuf);
+ } else {
+ dri_process_relocs(batch->buf);
+ ret = intel_batch_ioctl(batch->intel,
+ batch->buf->offset,
+ used,
+ batch->cliprect_mode != LOOP_CLIPRECTS,
+ allow_unlock);
+ }
}
- /* Associate a fence with the validated buffers, and note that we included
- * a flush at the end.
- */
- fo = dri_fence_validated(intel->intelScreen->bufmgr,
- "Batch fence", GL_TRUE);
-
- if (performed_rendering) {
- dri_fence_unreference(batch->last_fence);
- batch->last_fence = fo;
- } else {
- /* If we didn't validate any buffers for writing by the card, we don't
- * need to track the fence for glFinish().
- */
- dri_fence_unreference(fo);
- }
+ dri_post_submit(batch->buf);
- if (intel->numClipRects == 0 && !ignore_cliprects) {
+ if (intel->numClipRects == 0 &&
+ batch->cliprect_mode == LOOP_CLIPRECTS) {
if (allow_unlock) {
/* If we are not doing any actual user-visible rendering,
* do a sched_yield to keep the app from pegging the cpu while
sched_yield();
LOCK_HARDWARE(intel);
}
- intel->vtbl.lost_hardware(intel);
}
-done:
if (INTEL_DEBUG & DEBUG_BATCH) {
dri_bo_map(batch->buf, GL_FALSE);
- intel_decode(ptr, used / 4, batch->buf->offset);
+ intel_decode(batch->buf->virtual, used / 4, batch->buf->offset,
+ intel->intelScreen->deviceID);
dri_bo_unmap(batch->buf);
+
+ if (intel->vtbl.debug_batch != NULL)
+ intel->vtbl.debug_batch(intel);
}
-}
+ if (ret != 0) {
+ UNLOCK_HARDWARE(intel);
+ exit(1);
+ }
+ intel->vtbl.new_batch(intel);
+}
void
-intel_batchbuffer_flush(struct intel_batchbuffer *batch)
+_intel_batchbuffer_flush(struct intel_batchbuffer *batch, const char *file,
+ int line)
{
struct intel_context *intel = batch->intel;
GLuint used = batch->ptr - batch->map;
if (used == 0)
return;
- /* Add the MI_BATCH_BUFFER_END. Always add an MI_FLUSH - this is a
- * performance drain that we would like to avoid.
- */
- if (used & 4) {
- ((int *) batch->ptr)[0] = intel->vtbl.flush_cmd();
- ((int *) batch->ptr)[1] = 0;
- ((int *) batch->ptr)[2] = MI_BATCH_BUFFER_END;
- used += 12;
+ if (INTEL_DEBUG & DEBUG_BATCH)
+ fprintf(stderr, "%s:%d: Batchbuffer flush with %db used\n", file, line,
+ used);
+
+ /* Emit a flush if the bufmgr doesn't do it for us. */
+ if (!intel->ttm) {
+ *(GLuint *) (batch->ptr) = intel->vtbl.flush_cmd();
+ batch->ptr += 4;
+ used = batch->ptr - batch->map;
}
- else {
- ((int *) batch->ptr)[0] = intel->vtbl.flush_cmd();
- ((int *) batch->ptr)[1] = MI_BATCH_BUFFER_END;
- used += 8;
+
+ /* Round batchbuffer usage to 2 DWORDs. */
+
+ if ((used & 4) == 0) {
+ *(GLuint *) (batch->ptr) = 0; /* noop */
+ batch->ptr += 4;
+ used = batch->ptr - batch->map;
}
+ /* Mark the end of the buffer. */
+ *(GLuint *) (batch->ptr) = MI_BATCH_BUFFER_END; /* noop */
+ batch->ptr += 4;
+ used = batch->ptr - batch->map;
+
+ /* Workaround for recursive batchbuffer flushing: If the window is
+ * moved, we can get into a case where we try to flush during a
+ * flush. What happens is that when we try to grab the lock for
+ * the first flush, we detect that the window moved which then
+ * causes another flush (from the intel_draw_buffer() call in
+ * intelUpdatePageFlipping()). To work around this we reset the
+ * batchbuffer tail pointer before trying to get the lock. This
+ * prevent the nested buffer flush, but a better fix would be to
+ * avoid that in the first place. */
+ batch->ptr = batch->map;
+
/* TODO: Just pass the relocation list and dma buffer up to the
* kernel.
*/
if (!was_locked)
LOCK_HARDWARE(intel);
- do_flush_locked(batch, used, !(batch->flags & INTEL_BATCH_CLIPRECTS),
- GL_FALSE);
+ do_flush_locked(batch, used, GL_FALSE);
if (!was_locked)
UNLOCK_HARDWARE(intel);
+ if (INTEL_DEBUG & DEBUG_SYNC) {
+ int irq;
+
+ fprintf(stderr, "waiting for idle\n");
+ LOCK_HARDWARE(intel);
+ irq = intelEmitIrqLocked(intel);
+ UNLOCK_HARDWARE(intel);
+ intelWaitIrq(intel, irq);
+ }
+
/* Reset the buffer:
*/
intel_batchbuffer_reset(batch);
}
-void
-intel_batchbuffer_finish(struct intel_batchbuffer *batch)
-{
- intel_batchbuffer_flush(batch);
- if (batch->last_fence != NULL)
- dri_fence_wait(batch->last_fence);
-}
-
/* This is the only way buffers get added to the validate list.
*/
GLboolean
intel_batchbuffer_emit_reloc(struct intel_batchbuffer *batch,
dri_bo *buffer,
- GLuint flags, GLuint delta)
+ uint32_t read_domains, uint32_t write_domain,
+ uint32_t delta)
{
- struct buffer_reloc *r = &batch->reloc[batch->nr_relocs++];
-
- assert(batch->nr_relocs <= MAX_RELOCS);
-
- dri_bo_reference(buffer);
- r->buf = buffer;
- r->offset = batch->ptr - batch->map;
- r->delta = delta;
- r->validate_flags = flags;
+ int ret;
+
+ if (batch->ptr - batch->map > batch->buf->size)
+ _mesa_printf ("bad relocation ptr %p map %p offset %d size %d\n",
+ batch->ptr, batch->map, batch->ptr - batch->map, batch->buf->size);
+ ret = dri_emit_reloc(batch->buf, read_domains, write_domain,
+ delta, batch->ptr - batch->map, buffer);
+
+ /*
+ * Using the old buffer offset, write in what the right data would be, in case
+ * the buffer doesn't move and we can short-circuit the relocation processing
+ * in the kernel
+ */
+ intel_batchbuffer_emit_dword (batch, buffer->offset + delta);
- batch->ptr += 4;
return GL_TRUE;
}
-
-
void
intel_batchbuffer_data(struct intel_batchbuffer *batch,
- const void *data, GLuint bytes, GLuint flags)
+ const void *data, GLuint bytes,
+ enum cliprect_mode cliprect_mode)
{
assert((bytes & 3) == 0);
- intel_batchbuffer_require_space(batch, bytes, flags);
+ intel_batchbuffer_require_space(batch, bytes, cliprect_mode);
__memcpy(batch->ptr, data, bytes);
batch->ptr += bytes;
}
projects / mesa.git / blobdiff