From: Mika Kuoppala <mika.kuoppala@xxxxxxxxxxxxxxx>
To simplify emitting the recursive batch, make batch
always the first object on the execbuf list.
This will require kernel v4.13 or greater.
v2: set handles early, poll_ptr indecency (Chris)
v3: allow dep with poll
v4: fix gem_exec_schedule
v5: rebase
v6: rebase
v6: gem_ctx_shared
v7: conditional close of poll handle
v8: bw compat note (Chris)
Cc: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Signed-off-by: Mika Kuoppala <mika.kuoppala@xxxxxxxxxxxxxxx>
---
lib/igt_dummyload.c | 115 +++++++++++++++++----------------
lib/igt_dummyload.h | 8 ++-
tests/i915/gem_exec_balancer.c | 8 +--
tests/i915/gem_spin_batch.c | 13 ++--
4 files changed, 75 insertions(+), 69 deletions(-)
diff --git a/lib/igt_dummyload.c b/lib/igt_dummyload.c
index b9e239db3..c079bd045 100644
--- a/lib/igt_dummyload.c
+++ b/lib/igt_dummyload.c
@@ -63,6 +63,7 @@
#define MI_ARB_CHK (0x5 << 23)
static const int BATCH_SIZE = 4096;
+static const int POLL_SIZE = 4096;
static const int LOOP_START_OFFSET = 64;
static IGT_LIST_HEAD(spin_list);
@@ -72,16 +73,23 @@ static int
emit_recursive_batch(igt_spin_t *spin,
int fd, const struct igt_spin_factory *opts)
{
-#define SCRATCH 0
+
#define BATCH IGT_SPIN_BATCH
+#define POLL 1
+#define DEP 2
const int gen = intel_gen(intel_get_drm_devid(fd));
- struct drm_i915_gem_relocation_entry relocs[2], *r;
+ struct drm_i915_gem_exec_object2 * const batch =
+ &spin->obj[BATCH];
+ struct drm_i915_gem_exec_object2 * const poll =
+ &spin->obj[POLL];
+ struct drm_i915_gem_exec_object2 * const dep =
+ &spin->obj[DEP];
+ struct drm_i915_gem_relocation_entry relocs[4], *r;
struct drm_i915_gem_execbuffer2 *execbuf;
- struct drm_i915_gem_exec_object2 *obj;
unsigned int flags[GEM_MAX_ENGINES];
unsigned int nengine;
int fence_fd = -1;
- uint32_t *cs, *batch;
+ uint32_t *cs, *batch_start;
int i;
nengine = 0;
@@ -103,64 +111,48 @@ emit_recursive_batch(igt_spin_t *spin,
memset(&spin->execbuf, 0, sizeof(spin->execbuf));
execbuf = &spin->execbuf;
memset(spin->obj, 0, sizeof(spin->obj));
- obj = spin->obj;
memset(relocs, 0, sizeof(relocs));
- obj[BATCH].handle = gem_create(fd, BATCH_SIZE);
- batch = __gem_mmap__wc(fd, obj[BATCH].handle,
- 0, BATCH_SIZE, PROT_WRITE);
- if (!batch)
- batch = gem_mmap__gtt(fd, obj[BATCH].handle,
- BATCH_SIZE, PROT_WRITE);
+ batch->handle = gem_create(fd, BATCH_SIZE);
+ spin->handle = batch->handle;
- gem_set_domain(fd, obj[BATCH].handle,
+ batch_start = __gem_mmap__wc(fd, batch->handle,
+ 0, BATCH_SIZE, PROT_WRITE);
+ if (!batch_start)
+ batch_start = gem_mmap__gtt(fd, batch->handle,
+ BATCH_SIZE, PROT_WRITE);
+ gem_set_domain(fd, batch->handle,
I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
execbuf->buffer_count++;
- cs = batch;
-
- if (opts->dependency) {
- igt_assert(!(opts->flags & IGT_SPIN_POLL_RUN));
-
- r = &relocs[obj[BATCH].relocation_count++];
+ cs = batch_start;
- /* dummy write to dependency */
- obj[SCRATCH].handle = opts->dependency;
- r->presumed_offset = 0;
- r->target_handle = obj[SCRATCH].handle;
- r->offset = sizeof(uint32_t) * 1020;
- r->delta = 0;
- r->read_domains = I915_GEM_DOMAIN_RENDER;
- r->write_domain = I915_GEM_DOMAIN_RENDER;
-
- execbuf->buffer_count++;
- } else if (opts->flags & IGT_SPIN_POLL_RUN) {
- r = &relocs[obj[BATCH].relocation_count++];
+ poll->handle = gem_create(fd, POLL_SIZE);
+ spin->poll_handle = poll->handle;
+ execbuf->buffer_count++;
- igt_assert(!opts->dependency);
+ if (opts->flags & IGT_SPIN_POLL_RUN) {
+ r = &relocs[batch->relocation_count++];
if (gen == 4 || gen == 5) {
execbuf->flags |= I915_EXEC_SECURE;
igt_require(__igt_device_set_master(fd) == 0);
}
- spin->poll_handle = gem_create(fd, 4096);
- obj[SCRATCH].handle = spin->poll_handle;
-
- if (__gem_set_caching(fd, spin->poll_handle,
+ if (__gem_set_caching(fd, poll->handle,
I915_CACHING_CACHED) == 0)
- spin->poll = gem_mmap__cpu(fd, spin->poll_handle,
- 0, 4096,
+ spin->poll = gem_mmap__cpu(fd, poll->handle,
+ 0, POLL_SIZE,
PROT_READ | PROT_WRITE);
else
- spin->poll = gem_mmap__wc(fd, spin->poll_handle,
- 0, 4096,
+ spin->poll = gem_mmap__wc(fd, poll->handle,
+ 0, POLL_SIZE,
PROT_READ | PROT_WRITE);
igt_assert_eq(spin->poll[SPIN_POLL_START_IDX], 0);
/* batch is first */
- r->presumed_offset = 4096;
- r->target_handle = obj[SCRATCH].handle;
+ r->presumed_offset = BATCH_SIZE;
+ r->target_handle = poll->handle;
r->offset = sizeof(uint32_t) * 1;
r->delta = sizeof(uint32_t) * SPIN_POLL_START_IDX;
@@ -179,14 +171,25 @@ emit_recursive_batch(igt_spin_t *spin,
}
*cs++ = 1;
+ }
+
+ if (opts->dependency) {
+ r = &relocs[batch->relocation_count++];
+
+ /* dummy write to dependency */
+ dep->handle = opts->dependency;
+ r->presumed_offset = BATCH_SIZE + POLL_SIZE;
+ r->target_handle = dep->handle;
+ r->offset = sizeof(uint32_t) * 1020;
+ r->delta = 0;
+ r->read_domains = I915_GEM_DOMAIN_RENDER;
+ r->write_domain = I915_GEM_DOMAIN_RENDER;
execbuf->buffer_count++;
}
- spin->handle = obj[BATCH].handle;
-
- igt_assert_lt(cs - batch, LOOP_START_OFFSET / sizeof(*cs));
- spin->condition = batch + LOOP_START_OFFSET / sizeof(*cs);
+ igt_assert_lt(cs - batch_start, LOOP_START_OFFSET / sizeof(*cs));
+ spin->condition = batch_start + LOOP_START_OFFSET / sizeof(*cs);
cs = spin->condition;
/* Allow ourselves to be preempted */
@@ -208,9 +211,9 @@ emit_recursive_batch(igt_spin_t *spin,
cs += 1000;
/* recurse */
- r = &relocs[obj[BATCH].relocation_count++];
- r->target_handle = obj[BATCH].handle;
- r->offset = (cs + 1 - batch) * sizeof(*cs);
+ r = &relocs[batch->relocation_count++];
+ r->target_handle = batch->handle;
+ r->offset = (cs + 1 - batch_start) * sizeof(*cs);
r->read_domains = I915_GEM_DOMAIN_COMMAND;
r->delta = LOOP_START_OFFSET;
if (gen >= 8) {
@@ -227,10 +230,10 @@ emit_recursive_batch(igt_spin_t *spin,
*cs = r->delta;
cs++;
}
- obj[BATCH].relocs_ptr = to_user_pointer(relocs);
+ batch->relocs_ptr = to_user_pointer(relocs);
- execbuf->buffers_ptr = to_user_pointer(obj +
- (2 - execbuf->buffer_count));
+ execbuf->buffers_ptr = to_user_pointer(spin->obj);
+ execbuf->flags |= I915_EXEC_BATCH_FIRST;
execbuf->rsvd1 = opts->ctx;
if (opts->flags & IGT_SPIN_FENCE_OUT)
@@ -264,10 +267,9 @@ emit_recursive_batch(igt_spin_t *spin,
}
}
- igt_assert_lt(cs - batch, BATCH_SIZE / sizeof(*cs));
+ igt_assert_lt(cs - batch_start, BATCH_SIZE / sizeof(*cs));
- /* Make it easier for callers to resubmit. */
- for (i = 0; i < ARRAY_SIZE(spin->obj); i++) {
+ for (i = 0; i < execbuf->buffer_count; i++) {
spin->obj[i].relocation_count = 0;
spin->obj[i].relocs_ptr = 0;
spin->obj[i].flags = EXEC_OBJECT_PINNED;
@@ -445,10 +447,11 @@ void igt_spin_free(int fd, igt_spin_t *spin)
gem_munmap((void *)((unsigned long)spin->condition & (~4095UL)),
BATCH_SIZE);
- if (spin->poll) {
+ if (spin->poll)
gem_munmap(spin->poll, 4096);
+
+ if (spin->poll_handle)
gem_close(fd, spin->poll_handle);
- }
if (spin->handle)
gem_close(fd, spin->handle);
diff --git a/lib/igt_dummyload.h b/lib/igt_dummyload.h
index 421ca183b..de5781d7d 100644
--- a/lib/igt_dummyload.h
+++ b/lib/igt_dummyload.h
@@ -34,6 +34,7 @@
typedef struct igt_spin {
unsigned int handle;
+
timer_t timer;
struct igt_list_head link;
@@ -41,9 +42,12 @@ typedef struct igt_spin {
uint32_t cmd_precondition;
int out_fence;
- struct drm_i915_gem_exec_object2 obj[2];
-#define IGT_SPIN_BATCH 1
+ struct drm_i915_gem_exec_object2 obj[3];
+#define IGT_SPIN_BATCH 0
+#define IGT_SPIN_POLL 1
+
struct drm_i915_gem_execbuffer2 execbuf;
+
uint32_t poll_handle;
uint32_t *poll;
#define SPIN_POLL_START_IDX 0
diff --git a/tests/i915/gem_exec_balancer.c b/tests/i915/gem_exec_balancer.c
index 70c4529b4..19df406cf 100644
--- a/tests/i915/gem_exec_balancer.c
+++ b/tests/i915/gem_exec_balancer.c
@@ -718,7 +718,7 @@ static uint32_t create_semaphore_to_spinner(int i915, igt_spin_t *spin)
cs = map = gem_mmap__cpu(i915, handle, 0, 4096, PROT_WRITE);
/* Wait until the spinner is running */
- addr = spin->obj[0].offset + 4 * SPIN_POLL_START_IDX;
+ addr = spin->obj[IGT_SPIN_POLL].offset + 4 * SPIN_POLL_START_IDX;
*cs++ = MI_SEMAPHORE_WAIT |
MI_SEMAPHORE_POLL |
MI_SEMAPHORE_SAD_NEQ_SDD |
@@ -797,9 +797,9 @@ static void bonded_slice(int i915)
igt_spin_reset(spin);
/* igt_spin_t poll and batch obj must be laid out as we expect */
- igt_assert_eq(IGT_SPIN_BATCH, 1);
- obj[0] = spin->obj[0];
- obj[1] = spin->obj[1];
+ igt_assert_eq(IGT_SPIN_BATCH, 0);
+ obj[0] = spin->obj[IGT_SPIN_POLL];
+ obj[1] = spin->obj[IGT_SPIN_BATCH];
obj[2].handle = create_semaphore_to_spinner(i915, spin);
eb.buffers_ptr = to_user_pointer(obj);
diff --git a/tests/i915/gem_spin_batch.c b/tests/i915/gem_spin_batch.c
index c67f015ff..04707fdc2 100644
--- a/tests/i915/gem_spin_batch.c
+++ b/tests/i915/gem_spin_batch.c
@@ -78,12 +78,9 @@ static void spin_resubmit(int fd, const struct intel_execution_engine2 *e2,
gem_context_create(fd) : ctx0;
igt_spin_t *spin = __igt_spin_new(fd, .ctx = ctx0, .engine = e2->flags);
const struct intel_execution_engine2 *other;
+ struct drm_i915_gem_execbuffer2 eb = spin->execbuf;
- struct drm_i915_gem_execbuffer2 eb = {
- .buffer_count = 1,
- .buffers_ptr = to_user_pointer(&spin->obj[IGT_SPIN_BATCH]),
- .rsvd1 = ctx1,
- };
+ eb.rsvd1 = ctx1;
igt_assert(gem_context_has_engine_map(fd, 0) ||
!(flags & RESUBMIT_ALL_ENGINES));
@@ -97,11 +94,13 @@ static void spin_resubmit(int fd, const struct intel_execution_engine2 *e2,
if (gem_engine_is_equal(other, e2))
continue;
- eb.flags = other->flags;
+ eb.flags &= ~(I915_EXEC_RING_MASK | I915_EXEC_BSD_MASK);
+ eb.flags |= other->flags;
gem_execbuf(fd, &eb);
}
} else {
- eb.flags = e2->flags;
+ eb.flags &= ~(I915_EXEC_RING_MASK | I915_EXEC_BSD_MASK);
+ eb.flags |= e2->flags;
gem_execbuf(fd, &eb);
}
--
2.24.0
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
