@@ -3081,15 +3081,60 @@ static int bond_virtual_engine(struct intel_gt *gt,
struct i915_gem_context *ctx;
struct i915_request *rq[16];
enum intel_engine_id id;
+ struct igt_spinner spin;
unsigned long n;
int err;
+ /*
+ * A set of bonded requests is intended to be run concurrently
+ * across a number of engines. We use one request per-engine
+ * and a magic fence to schedule each of the bonded requests
+ * at the same time. A consequence of our currently scheduler
+ * is that we only move requests to the HW ready queue when
+ * the request becomes ready, that is when all of its prerequisite
+ * fences have been signaled. As one of those fences is the master
+ * submit fence, there is a delay on all secondary fences as the
+ * HW may be currently busy. Equally, as all the requests are
+ * independent, they may have other fences that delay individual
+ * request submission to HW. Ergo, we do not guarantee that
+ * all requests are immediately submitted to HW at the same time,
+ * just that if the rules are abided by, they are ready at the
+ * same time as the first is submitted. Userspace can embed semaphores
+ * in its batch to ensure parallel execution of phases as it requires.
+ * Though naturally it gets requested that perhaps the scheduler should
+ * take care of parallel execution, even across preemption events
+ * on different HW. (The proper answer is of course "lalalala".)
+ *
+ * With the submit-fence, we have identified three possible phases
+ * of synchronisation depending on the master fence: queued (not
+ * ready), ready or executing, signaled. The first two are quite
+ * simple and checked below. However, the signaled master fence
+ * handling is contentious. Currently we do not distinguish between
+ * a signaled fence and an expired fence, as once signaled it does
+ * not convey any information about the previous execution, it may
+ * be freed and hence checking later it may not exist at all. Ergo
+ * we currently do not apply the bonding constraint for an already
+ * signaled fence, as our expectation is that it should not constrain
+ * the secondaries and is outside of the scope of the bonded request
+ * API (i.e. all requests are meant to be running in parallel). As
+ * it imposes no constraint, and is effectively a no-op, we do not
+ * check below as normal execution flows are checked extensively above.
+ *
+ * XXX Is the degenerate handling of signaled submit fences the
+ * expected behaviour for userpace?
+ */
+
GEM_BUG_ON(nsibling >= ARRAY_SIZE(rq) - 1);
- ctx = kernel_context(gt->i915);
- if (!ctx)
+ if (igt_spinner_init(&spin, gt))
return -ENOMEM;
+ ctx = kernel_context(gt->i915);
+ if (!ctx) {
+ err = -ENOMEM;
+ goto err_spin;
+ }
+
err = 0;
rq[0] = ERR_PTR(-ENOMEM);
for_each_engine(master, gt, id) {
@@ -3100,7 +3145,7 @@ static int bond_virtual_engine(struct intel_gt *gt,
memset_p((void *)rq, ERR_PTR(-EINVAL), ARRAY_SIZE(rq));
- rq[0] = igt_request_alloc(ctx, master);
+ rq[0] = spinner_create_request(&spin, ctx, master, MI_NOOP);
if (IS_ERR(rq[0])) {
err = PTR_ERR(rq[0]);
goto out;
@@ -3113,10 +3158,17 @@ static int bond_virtual_engine(struct intel_gt *gt,
&fence,
GFP_KERNEL);
}
+
i915_request_add(rq[0]);
if (err < 0)
goto out;
+ if (!(flags & BOND_SCHEDULE) &&
+ !igt_wait_for_spinner(&spin, rq[0])) {
+ err = -EIO;
+ goto out;
+ }
+
for (n = 0; n < nsibling; n++) {
struct intel_context *ve;
@@ -3164,6 +3216,8 @@ static int bond_virtual_engine(struct intel_gt *gt,
}
}
onstack_fence_fini(&fence);
+ intel_engine_flush_submission(master);
+ igt_spinner_end(&spin);
if (i915_request_wait(rq[0], 0, HZ / 10) < 0) {
pr_err("Master request did not execute (on %s)!\n",
@@ -3201,6 +3255,8 @@ static int bond_virtual_engine(struct intel_gt *gt,
err = -EIO;
kernel_context_close(ctx);
+err_spin:
+ igt_spinner_fini(&spin);
return err;
}
Bonded request submission is designed to allow requests to execute in parallel as laid out by the user. If the master request is already finished before its bonded pair is submitted, the pair were not destined to run in parallel and we lose the information about the master engine to dictate selection of the secondary. If the second request was required to be run on a particular engine in a virtual set, that should have been specified, rather than left to the whims of a random unconnected requests! In the selftest, I made the mistake of not ensuring the master would overlap with its bonded pairs, meaning that it could indeed complete before we submitted the bonds. Those bonds were then free to select any available engine in their virtual set, and not the one expected by the test. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> --- drivers/gpu/drm/i915/gt/selftest_lrc.c | 62 ++++++++++++++++++++++++-- 1 file changed, 59 insertions(+), 3 deletions(-)