[igt-dev] [PATCH i-g-t 2/4] lib/intel_batchbuffer: Extend intel_bb
Dominik Grzegorzek
dominik.grzegorzek at intel.com
Mon Jun 29 11:10:02 UTC 2020
From: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
As we're going toward removing libdrm from rendercopy tests some
additional code in intel_bb is required. So add new functions
and fix memory issues in the reset path.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
Cc: Chris Wilson <chris at chris-wilson.co.uk>
---
lib/intel_batchbuffer.c | 189 ++++++++++++++++++++++++++++++++++++----
lib/intel_batchbuffer.h | 12 +++
2 files changed, 185 insertions(+), 16 deletions(-)
diff --git a/lib/intel_batchbuffer.c b/lib/intel_batchbuffer.c
index 2a882627..c041465d 100644
--- a/lib/intel_batchbuffer.c
+++ b/lib/intel_batchbuffer.c
@@ -1250,24 +1250,32 @@ static void __do_nothing(void *node)
(void) node;
}
-static void __intel_bb_destroy_objects(struct intel_bb *ibb)
+static void __intel_bb_destroy_relocations(struct intel_bb *ibb)
{
uint32_t i;
/* Free relocations */
- for (i = 0; i < ibb->num_objects; i++)
+ for (i = 0; i < ibb->num_objects; i++) {
free(from_user_pointer(ibb->objects[i].relocs_ptr));
+ ibb->objects[i].relocs_ptr = to_user_pointer(NULL);
+ ibb->objects[i].relocation_count = 0;
+ }
- free(ibb->objects);
- tdestroy(ibb->root, __do_nothing);
+ ibb->relocs = NULL;
+ ibb->num_relocs = 0;
+ ibb->allocated_relocs = 0;
+}
+static void __intel_bb_destroy_objects(struct intel_bb *ibb)
+{
+ free(ibb->objects);
ibb->objects = NULL;
+
+ tdestroy(ibb->root, __do_nothing);
ibb->root = NULL;
+
ibb->num_objects = 0;
- ibb->num_relocs = 0;
ibb->allocated_objects = 0;
- ibb->allocated_relocs = 0;
- ibb->ptr = ibb->batch;
}
/**
@@ -1280,6 +1288,7 @@ void intel_bb_destroy(struct intel_bb *ibb)
{
igt_assert(ibb);
+ __intel_bb_destroy_relocations(ibb);
__intel_bb_destroy_objects(ibb);
gem_close(ibb->i915, ibb->handle);
@@ -1295,6 +1304,8 @@ void intel_bb_destroy(struct intel_bb *ibb)
*/
void intel_bb_reset(struct intel_bb *ibb, bool purge_objects_cache)
{
+ __intel_bb_destroy_relocations(ibb);
+
if (purge_objects_cache) {
__intel_bb_destroy_objects(ibb);
__reallocate_objects(ibb);
@@ -1305,6 +1316,61 @@ void intel_bb_reset(struct intel_bb *ibb, bool purge_objects_cache)
intel_bb_add_object(ibb, ibb->handle, 0, false);
ibb->ptr = ibb->batch;
+ memset(ibb->batch, 0, ibb->size);
+}
+
+/*
+ * intel_bb_sync:
+ * @ibb: pointer to intel_bb
+ *
+ * Waits for bb completion.
+ */
+void intel_bb_sync(struct intel_bb *ibb)
+{
+ gem_sync(ibb->i915, ibb->handle);
+}
+
+/*
+ * intel_bb_print:
+ * @ibb: pointer to intel_bb
+ *
+ * Prints batch to stdout.
+ */
+void intel_bb_print(struct intel_bb *ibb)
+{
+ igt_info("drm fd: %d, gen: %d, devid: %u, debug: %d\n",
+ ibb->i915, ibb->gen, ibb->devid, ibb->debug);
+ igt_info("handle: %u, size: %u, batch: %p, ptr: %p\n",
+ ibb->handle, ibb->size, ibb->batch, ibb->ptr);
+ igt_info("prng: %u, gtt_size: %" PRIu64 ", supports 48bit: %d\n",
+ ibb->prng, ibb->gtt_size, ibb->supports_48b_address);
+ igt_info("ctx: %u\n", ibb->ctx);
+ igt_info("root: %p\n", ibb->root);
+ igt_info("objects: %p, num_objects: %u, allocated obj: %u\n",
+ ibb->objects, ibb->num_objects, ibb->allocated_objects);
+ igt_info("relocs: %p, num_relocs: %u, allocated_relocs: %u\n----\n",
+ ibb->relocs, ibb->num_relocs, ibb->allocated_relocs);
+}
+
+/*
+ * intel_bb_dump:
+ * @ibb: pointer to intel_bb
+ * @filename: name to which write bb
+ *
+ * Dump batch bo to file.
+ */
+void intel_bb_dump(struct intel_bb *ibb, const char *filename)
+{
+ FILE *out;
+ void *ptr;
+
+ ptr = gem_mmap__device_coherent(ibb->i915, ibb->handle, 0, ibb->size,
+ PROT_READ);
+ out = fopen(filename, "wb");
+ igt_assert(out);
+ fwrite(ptr, ibb->size, 1, out);
+ fclose(out);
+ munmap(ptr, ibb->size);
}
/**
@@ -1382,6 +1448,23 @@ intel_bb_add_object(struct intel_bb *ibb, uint32_t handle,
return object;
}
+static bool intel_bb_object_set_fence(struct intel_bb *ibb, uint32_t handle)
+{
+ struct drm_i915_gem_exec_object2 object = { .handle = handle };
+ struct drm_i915_gem_exec_object2 **found;
+
+ found = tfind((void *) &object, &ibb->root, __compare_objects);
+ if (!found) {
+ igt_warn("Trying to set fence on not found handle: %u\n",
+ handle);
+ return false;
+ }
+
+ (*found)->flags |= EXEC_OBJECT_NEEDS_FENCE;
+
+ return true;
+}
+
/*
* intel_bb_add_reloc:
* @ibb: pointer to intel_bb
@@ -1431,9 +1514,10 @@ static uint64_t intel_bb_add_reloc(struct intel_bb *ibb,
igt_debug("add reloc: handle: %u, r/w: 0x%x/0x%x, "
"delta: 0x%" PRIx64 ", "
"offset: 0x%" PRIx64 ", "
- "poffset: 0x%" PRIx64 "\n",
+ "poffset: %p\n",
handle, read_domains, write_domain,
- delta, offset, presumed_offset);
+ delta, offset,
+ from_user_pointer(relocs[i].presumed_offset));
return object->offset;
}
@@ -1479,6 +1563,38 @@ uint64_t intel_bb_emit_reloc(struct intel_bb *ibb,
return address;
}
+/**
+ * intel_bb_emit_reloc_fenced:
+ * @ibb: pointer to intel_bb
+ * @handle: object handle which address will be taken to patch the bb
+ * @read_domains: gem domain bits for the relocation
+ * @write_domain: gem domain bit for the relocation
+ * @delta: delta value to add to @buffer's gpu address
+ * @presumed_offset: address of the object in address space, important for
+ * I915_EXEC_NO_RELOC flag
+ * @write: does a handle is a render target
+ *
+ * Function works similar to intel_bb_emit_reloc but additionally marks an
+ * object with EXEC_OBJECT_NEEDS_FENCE to ensure appropriate fence will
+ * be configured when it is required on older gens.
+ */
+uint64_t intel_bb_emit_reloc_fenced(struct intel_bb *ibb,
+ uint32_t handle,
+ uint32_t read_domains,
+ uint32_t write_domain,
+ uint64_t delta,
+ uint64_t presumed_offset)
+{
+ uint64_t address;
+
+ address = intel_bb_emit_reloc(ibb, handle, read_domains, write_domain,
+ delta, presumed_offset);
+
+ intel_bb_object_set_fence(ibb, handle);
+
+ return address;
+}
+
/**
* intel_bb_offset_reloc:
* @ibb: pointer to intel_bb
@@ -1509,7 +1625,8 @@ uint64_t intel_bb_offset_reloc(struct intel_bb *ibb,
0, offset, presumed_offset);
}
-static void intel_bb_dump_execbuf(struct drm_i915_gem_execbuffer2 *execbuf)
+static void intel_bb_dump_execbuf(struct intel_bb *ibb,
+ struct drm_i915_gem_execbuffer2 *execbuf)
{
struct drm_i915_gem_exec_object2 *objects;
struct drm_i915_gem_relocation_entry *relocs, *reloc;
@@ -1535,7 +1652,8 @@ static void intel_bb_dump_execbuf(struct drm_i915_gem_execbuffer2 *execbuf)
i, objects->handle, objects->relocation_count,
relocs,
objects->alignment,
- objects->offset, objects->flags,
+ objects->offset & (ibb->gtt_size - 1),
+ objects->flags,
objects->rsvd1, objects->rsvd2);
if (objects->relocation_count) {
igt_info("\texecbuf relocs:\n");
@@ -1548,7 +1666,7 @@ static void intel_bb_dump_execbuf(struct drm_i915_gem_execbuffer2 *execbuf)
"write_domain: 0x%x\n",
j, reloc->target_handle,
reloc->offset, reloc->delta,
- reloc->presumed_offset,
+ reloc->presumed_offset & (ibb->gtt_size - 1),
reloc->read_domains,
reloc->write_domain);
}
@@ -1609,14 +1727,16 @@ int __intel_bb_exec(struct intel_bb *ibb, uint32_t end_offset,
execbuf.flags = flags | I915_EXEC_BATCH_FIRST;
ret = __gem_execbuf(ibb->i915, &execbuf);
- if (ret)
+ if (ret) {
+ intel_bb_dump_execbuf(ibb, &execbuf);
return ret;
+ }
if (sync || ibb->debug)
gem_sync(ibb->i915, ibb->handle);
if (ibb->debug) {
- intel_bb_dump_execbuf(&execbuf);
+ intel_bb_dump_execbuf(ibb, &execbuf);
if (intel_bb_debug_tree) {
igt_info("\nTree:\n");
twalk(ibb->root, print_node);
@@ -1679,7 +1799,7 @@ uint64_t intel_bb_get_object_offset(struct intel_bb *ibb, uint32_t handle)
if (!found)
return 0;
- return (*found)->offset;
+ return (*found)->offset & (ibb->gtt_size - 1);
}
/**
@@ -1706,8 +1826,45 @@ bool intel_bb_object_offset_to_buf(struct intel_bb *ibb, struct intel_buf *buf)
return false;
}
- buf->addr.offset = (*found)->offset;
+ buf->addr.offset = (*found)->offset & (ibb->gtt_size - 1);
buf->addr.ctx = ibb->ctx;
return true;
}
+
+/*
+ * intel_bb_out_flush:
+ * @ibb: batchbuffer
+ *
+ * Outputs flushing code to the batch.
+ */
+#define CMD_POLY_STIPPLE_OFFSET 0x7906
+uint32_t intel_bb_out_flush(struct intel_bb *ibb)
+{
+ if (intel_bb_offset(ibb) == 0)
+ return 0;
+
+ if (ibb->gen == 5) {
+ intel_bb_out(ibb, CMD_POLY_STIPPLE_OFFSET << 16);
+ intel_bb_out(ibb, 0);
+ }
+
+ /* Round batchbuffer usage to 2 DWORDs. */
+ intel_bb_ptr_align(ibb, 8);
+
+ if (ibb->gen <= 4) {
+ intel_bb_out(ibb, MI_FLUSH);
+ } else {
+ intel_bb_out(ibb, MI_FLUSH_DW | 2);
+ intel_bb_out(ibb, 0);
+ intel_bb_out(ibb, 0);
+ intel_bb_out(ibb, 0);
+ }
+
+ /* Mark the end of the buffer. */
+ intel_bb_out(ibb, MI_BATCH_BUFFER_END); /* noop */
+ intel_bb_out(ibb, 0);
+
+ return intel_bb_offset(ibb);
+}
+
diff --git a/lib/intel_batchbuffer.h b/lib/intel_batchbuffer.h
index 0649fc22..2ced3ea6 100644
--- a/lib/intel_batchbuffer.h
+++ b/lib/intel_batchbuffer.h
@@ -460,6 +460,9 @@ struct intel_bb *intel_bb_create(int i915, uint32_t size);
void intel_bb_destroy(struct intel_bb *ibb);
void intel_bb_reset(struct intel_bb *ibb, bool purge_objects_cache);
+void intel_bb_sync(struct intel_bb *ibb);
+void intel_bb_print(struct intel_bb *ibb);
+void intel_bb_dump(struct intel_bb *ibb, const char *filename);
void intel_bb_set_debug(struct intel_bb *ibb, bool debug);
static inline uint32_t intel_bb_offset(struct intel_bb *ibb)
@@ -510,6 +513,13 @@ uint64_t intel_bb_emit_reloc(struct intel_bb *ibb,
uint64_t delta,
uint64_t presumed_offset);
+uint64_t intel_bb_emit_reloc_fenced(struct intel_bb *ibb,
+ uint32_t handle,
+ uint32_t read_domains,
+ uint32_t write_domain,
+ uint64_t delta,
+ uint64_t presumed_offset);
+
uint64_t intel_bb_offset_reloc(struct intel_bb *ibb,
uint32_t handle,
uint32_t read_domains,
@@ -529,4 +539,6 @@ void intel_bb_exec_with_context(struct intel_bb *ibb, uint32_t end_offset,
uint64_t intel_bb_get_object_offset(struct intel_bb *ibb, uint32_t handle);
bool intel_bb_object_offset_to_buf(struct intel_bb *ibb, struct intel_buf *buf);
+uint32_t intel_bb_out_flush(struct intel_bb *ibb);
+
#endif
--
2.20.1
More information about the igt-dev
mailing list