drm/i915/execlists: Virtual engine bonding
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 21 May 2019 21:11:33 +0000 (22:11 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 22 May 2019 07:40:46 +0000 (08:40 +0100)
Some users require that when a master batch is executed on one particular
engine, a companion batch is run simultaneously on a specific slave
engine. For this purpose, we introduce virtual engine bonding, allowing
maps of master:slaves to be constructed to constrain which physical
engines a virtual engine may select given a fence on a master engine.

For the moment, we continue to ignore the issue of preemption deferring
the master request for later. Ideally, we would like to then also remove
the slave and run something else rather than have it stall the pipeline.
With load balancing, we should be able to move workload around it, but
there is a similar stall on the master pipeline while it may wait for
the slave to be executed. At the cost of more latency for the bonded
request, it may be interesting to launch both on their engines in
lockstep. (Bubbles abound.)

Opens: Also what about bonding an engine as its own master? It doesn't
break anything internally, so allow the silliness.

v2: Emancipate the bonds
v3: Couple in delayed scheduling for the selftests
v4: Handle invalid mutually exclusive bonding
v5: Mention what the uapi does
v6: s/nbond/num_bonds/

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190521211134.16117-9-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/gt/intel_engine_types.h
drivers/gpu/drm/i915/gt/intel_lrc.c
drivers/gpu/drm/i915/gt/intel_lrc.h
drivers/gpu/drm/i915/gt/selftest_lrc.c
drivers/gpu/drm/i915/i915_gem_context.c
drivers/gpu/drm/i915/selftests/lib_sw_fence.c
include/uapi/drm/i915_drm.h

index 7b47e00..f3fc2e8 100644 (file)
@@ -405,6 +405,13 @@ struct intel_engine_cs {
         */
        void            (*submit_request)(struct i915_request *rq);
 
+       /*
+        * Called on signaling of a SUBMIT_FENCE, passing along the signaling
+        * request down to the bonded pairs.
+        */
+       void            (*bond_execute)(struct i915_request *rq,
+                                       struct dma_fence *signal);
+
        /*
         * Call when the priority on a request has changed and it and its
         * dependencies may need rescheduling. Note the request itself may
index 0b4d29d..1f7bee0 100644 (file)
@@ -189,6 +189,18 @@ struct virtual_engine {
                int prio;
        } nodes[I915_NUM_ENGINES];
 
+       /*
+        * Keep track of bonded pairs -- restrictions upon on our selection
+        * of physical engines any particular request may be submitted to.
+        * If we receive a submit-fence from a master engine, we will only
+        * use one of sibling_mask physical engines.
+        */
+       struct ve_bond {
+               const struct intel_engine_cs *master;
+               intel_engine_mask_t sibling_mask;
+       } *bonds;
+       unsigned int num_bonds;
+
        /* And finally, which physical engines this virtual engine maps onto. */
        unsigned int num_siblings;
        struct intel_engine_cs *siblings[0];
@@ -960,6 +972,7 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
                        rb_erase_cached(rb, &execlists->virtual);
                        RB_CLEAR_NODE(rb);
 
+                       GEM_BUG_ON(!(rq->execution_mask & engine->mask));
                        rq->engine = engine;
 
                        if (engine != ve->siblings[0]) {
@@ -3069,6 +3082,8 @@ static void virtual_context_destroy(struct kref *kref)
        if (ve->context.state)
                __execlists_context_fini(&ve->context);
 
+       kfree(ve->bonds);
+
        i915_timeline_fini(&ve->base.timeline);
        kfree(ve);
 }
@@ -3265,6 +3280,38 @@ static void virtual_submit_request(struct i915_request *rq)
        tasklet_schedule(&ve->base.execlists.tasklet);
 }
 
+static struct ve_bond *
+virtual_find_bond(struct virtual_engine *ve,
+                 const struct intel_engine_cs *master)
+{
+       int i;
+
+       for (i = 0; i < ve->num_bonds; i++) {
+               if (ve->bonds[i].master == master)
+                       return &ve->bonds[i];
+       }
+
+       return NULL;
+}
+
+static void
+virtual_bond_execute(struct i915_request *rq, struct dma_fence *signal)
+{
+       struct virtual_engine *ve = to_virtual_engine(rq->engine);
+       struct ve_bond *bond;
+
+       bond = virtual_find_bond(ve, to_request(signal)->engine);
+       if (bond) {
+               intel_engine_mask_t old, new, cmp;
+
+               cmp = READ_ONCE(rq->execution_mask);
+               do {
+                       old = cmp;
+                       new = cmp & bond->sibling_mask;
+               } while ((cmp = cmpxchg(&rq->execution_mask, old, new)) != old);
+       }
+}
+
 struct intel_context *
 intel_execlists_create_virtual(struct i915_gem_context *ctx,
                               struct intel_engine_cs **siblings,
@@ -3305,6 +3352,7 @@ intel_execlists_create_virtual(struct i915_gem_context *ctx,
 
        ve->base.schedule = i915_schedule;
        ve->base.submit_request = virtual_submit_request;
+       ve->base.bond_execute = virtual_bond_execute;
 
        ve->base.execlists.queue_priority_hint = INT_MIN;
        tasklet_init(&ve->base.execlists.tasklet,
@@ -3394,9 +3442,59 @@ intel_execlists_clone_virtual(struct i915_gem_context *ctx,
        if (IS_ERR(dst))
                return dst;
 
+       if (se->num_bonds) {
+               struct virtual_engine *de = to_virtual_engine(dst->engine);
+
+               de->bonds = kmemdup(se->bonds,
+                                   sizeof(*se->bonds) * se->num_bonds,
+                                   GFP_KERNEL);
+               if (!de->bonds) {
+                       intel_context_put(dst);
+                       return ERR_PTR(-ENOMEM);
+               }
+
+               de->num_bonds = se->num_bonds;
+       }
+
        return dst;
 }
 
+int intel_virtual_engine_attach_bond(struct intel_engine_cs *engine,
+                                    const struct intel_engine_cs *master,
+                                    const struct intel_engine_cs *sibling)
+{
+       struct virtual_engine *ve = to_virtual_engine(engine);
+       struct ve_bond *bond;
+       int n;
+
+       /* Sanity check the sibling is part of the virtual engine */
+       for (n = 0; n < ve->num_siblings; n++)
+               if (sibling == ve->siblings[n])
+                       break;
+       if (n == ve->num_siblings)
+               return -EINVAL;
+
+       bond = virtual_find_bond(ve, master);
+       if (bond) {
+               bond->sibling_mask |= sibling->mask;
+               return 0;
+       }
+
+       bond = krealloc(ve->bonds,
+                       sizeof(*bond) * (ve->num_bonds + 1),
+                       GFP_KERNEL);
+       if (!bond)
+               return -ENOMEM;
+
+       bond[ve->num_bonds].master = master;
+       bond[ve->num_bonds].sibling_mask = sibling->mask;
+
+       ve->bonds = bond;
+       ve->num_bonds++;
+
+       return 0;
+}
+
 void intel_execlists_show_requests(struct intel_engine_cs *engine,
                                   struct drm_printer *m,
                                   void (*show_request)(struct drm_printer *m,
index 5530606..e029aee 100644 (file)
@@ -123,4 +123,8 @@ struct intel_context *
 intel_execlists_clone_virtual(struct i915_gem_context *ctx,
                              struct intel_engine_cs *src);
 
+int intel_virtual_engine_attach_bond(struct intel_engine_cs *engine,
+                                    const struct intel_engine_cs *master,
+                                    const struct intel_engine_cs *sibling);
+
 #endif /* _INTEL_LRC_H_ */
index 61637f5..a8c5090 100644 (file)
@@ -13,6 +13,7 @@
 #include "selftests/igt_gem_utils.h"
 #include "selftests/igt_live_test.h"
 #include "selftests/igt_spinner.h"
+#include "selftests/lib_sw_fence.h"
 #include "selftests/mock_context.h"
 
 static int live_sanitycheck(void *arg)
@@ -1619,6 +1620,195 @@ out_unlock:
        return err;
 }
 
+static int bond_virtual_engine(struct drm_i915_private *i915,
+                              unsigned int class,
+                              struct intel_engine_cs **siblings,
+                              unsigned int nsibling,
+                              unsigned int flags)
+#define BOND_SCHEDULE BIT(0)
+{
+       struct intel_engine_cs *master;
+       struct i915_gem_context *ctx;
+       struct i915_request *rq[16];
+       enum intel_engine_id id;
+       unsigned long n;
+       int err;
+
+       GEM_BUG_ON(nsibling >= ARRAY_SIZE(rq) - 1);
+
+       ctx = kernel_context(i915);
+       if (!ctx)
+               return -ENOMEM;
+
+       err = 0;
+       rq[0] = ERR_PTR(-ENOMEM);
+       for_each_engine(master, i915, id) {
+               struct i915_sw_fence fence = {};
+
+               if (master->class == class)
+                       continue;
+
+               memset_p((void *)rq, ERR_PTR(-EINVAL), ARRAY_SIZE(rq));
+
+               rq[0] = igt_request_alloc(ctx, master);
+               if (IS_ERR(rq[0])) {
+                       err = PTR_ERR(rq[0]);
+                       goto out;
+               }
+               i915_request_get(rq[0]);
+
+               if (flags & BOND_SCHEDULE) {
+                       onstack_fence_init(&fence);
+                       err = i915_sw_fence_await_sw_fence_gfp(&rq[0]->submit,
+                                                              &fence,
+                                                              GFP_KERNEL);
+               }
+               i915_request_add(rq[0]);
+               if (err < 0)
+                       goto out;
+
+               for (n = 0; n < nsibling; n++) {
+                       struct intel_context *ve;
+
+                       ve = intel_execlists_create_virtual(ctx,
+                                                           siblings,
+                                                           nsibling);
+                       if (IS_ERR(ve)) {
+                               err = PTR_ERR(ve);
+                               onstack_fence_fini(&fence);
+                               goto out;
+                       }
+
+                       err = intel_virtual_engine_attach_bond(ve->engine,
+                                                              master,
+                                                              siblings[n]);
+                       if (err) {
+                               intel_context_put(ve);
+                               onstack_fence_fini(&fence);
+                               goto out;
+                       }
+
+                       err = intel_context_pin(ve);
+                       intel_context_put(ve);
+                       if (err) {
+                               onstack_fence_fini(&fence);
+                               goto out;
+                       }
+
+                       rq[n + 1] = i915_request_create(ve);
+                       intel_context_unpin(ve);
+                       if (IS_ERR(rq[n + 1])) {
+                               err = PTR_ERR(rq[n + 1]);
+                               onstack_fence_fini(&fence);
+                               goto out;
+                       }
+                       i915_request_get(rq[n + 1]);
+
+                       err = i915_request_await_execution(rq[n + 1],
+                                                          &rq[0]->fence,
+                                                          ve->engine->bond_execute);
+                       i915_request_add(rq[n + 1]);
+                       if (err < 0) {
+                               onstack_fence_fini(&fence);
+                               goto out;
+                       }
+               }
+               onstack_fence_fini(&fence);
+
+               if (i915_request_wait(rq[0],
+                                     I915_WAIT_LOCKED,
+                                     HZ / 10) < 0) {
+                       pr_err("Master request did not execute (on %s)!\n",
+                              rq[0]->engine->name);
+                       err = -EIO;
+                       goto out;
+               }
+
+               for (n = 0; n < nsibling; n++) {
+                       if (i915_request_wait(rq[n + 1],
+                                             I915_WAIT_LOCKED,
+                                             MAX_SCHEDULE_TIMEOUT) < 0) {
+                               err = -EIO;
+                               goto out;
+                       }
+
+                       if (rq[n + 1]->engine != siblings[n]) {
+                               pr_err("Bonded request did not execute on target engine: expected %s, used %s; master was %s\n",
+                                      siblings[n]->name,
+                                      rq[n + 1]->engine->name,
+                                      rq[0]->engine->name);
+                               err = -EINVAL;
+                               goto out;
+                       }
+               }
+
+               for (n = 0; !IS_ERR(rq[n]); n++)
+                       i915_request_put(rq[n]);
+               rq[0] = ERR_PTR(-ENOMEM);
+       }
+
+out:
+       for (n = 0; !IS_ERR(rq[n]); n++)
+               i915_request_put(rq[n]);
+       if (igt_flush_test(i915, I915_WAIT_LOCKED))
+               err = -EIO;
+
+       kernel_context_close(ctx);
+       return err;
+}
+
+static int live_virtual_bond(void *arg)
+{
+       static const struct phase {
+               const char *name;
+               unsigned int flags;
+       } phases[] = {
+               { "", 0 },
+               { "schedule", BOND_SCHEDULE },
+               { },
+       };
+       struct drm_i915_private *i915 = arg;
+       struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
+       unsigned int class, inst;
+       int err = 0;
+
+       if (USES_GUC_SUBMISSION(i915))
+               return 0;
+
+       mutex_lock(&i915->drm.struct_mutex);
+
+       for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
+               const struct phase *p;
+               int nsibling;
+
+               nsibling = 0;
+               for (inst = 0; inst <= MAX_ENGINE_INSTANCE; inst++) {
+                       if (!i915->engine_class[class][inst])
+                               break;
+
+                       GEM_BUG_ON(nsibling == ARRAY_SIZE(siblings));
+                       siblings[nsibling++] = i915->engine_class[class][inst];
+               }
+               if (nsibling < 2)
+                       continue;
+
+               for (p = phases; p->name; p++) {
+                       err = bond_virtual_engine(i915,
+                                                 class, siblings, nsibling,
+                                                 p->flags);
+                       if (err) {
+                               pr_err("%s(%s): failed class=%d, nsibling=%d, err=%d\n",
+                                      __func__, p->name, class, nsibling, err);
+                               goto out_unlock;
+                       }
+               }
+       }
+
+out_unlock:
+       mutex_unlock(&i915->drm.struct_mutex);
+       return err;
+}
+
 int intel_execlists_live_selftests(struct drm_i915_private *i915)
 {
        static const struct i915_subtest tests[] = {
@@ -1633,6 +1823,7 @@ int intel_execlists_live_selftests(struct drm_i915_private *i915)
                SUBTEST(live_preempt_smoke),
                SUBTEST(live_virtual_engine),
                SUBTEST(live_virtual_mask),
+               SUBTEST(live_virtual_bond),
        };
 
        if (!HAS_EXECLISTS(i915))
index 68db6b2..5d2f8ba 100644 (file)
@@ -1504,8 +1504,92 @@ out_siblings:
        return err;
 }
 
+static int
+set_engines__bond(struct i915_user_extension __user *base, void *data)
+{
+       struct i915_context_engines_bond __user *ext =
+               container_of_user(base, typeof(*ext), base);
+       const struct set_engines *set = data;
+       struct i915_engine_class_instance ci;
+       struct intel_engine_cs *virtual;
+       struct intel_engine_cs *master;
+       u16 idx, num_bonds;
+       int err, n;
+
+       if (get_user(idx, &ext->virtual_index))
+               return -EFAULT;
+
+       if (idx >= set->engines->num_engines) {
+               DRM_DEBUG("Invalid index for virtual engine: %d >= %d\n",
+                         idx, set->engines->num_engines);
+               return -EINVAL;
+       }
+
+       idx = array_index_nospec(idx, set->engines->num_engines);
+       if (!set->engines->engines[idx]) {
+               DRM_DEBUG("Invalid engine at %d\n", idx);
+               return -EINVAL;
+       }
+       virtual = set->engines->engines[idx]->engine;
+
+       err = check_user_mbz(&ext->flags);
+       if (err)
+               return err;
+
+       for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) {
+               err = check_user_mbz(&ext->mbz64[n]);
+               if (err)
+                       return err;
+       }
+
+       if (copy_from_user(&ci, &ext->master, sizeof(ci)))
+               return -EFAULT;
+
+       master = intel_engine_lookup_user(set->ctx->i915,
+                                         ci.engine_class, ci.engine_instance);
+       if (!master) {
+               DRM_DEBUG("Unrecognised master engine: { class:%u, instance:%u }\n",
+                         ci.engine_class, ci.engine_instance);
+               return -EINVAL;
+       }
+
+       if (get_user(num_bonds, &ext->num_bonds))
+               return -EFAULT;
+
+       for (n = 0; n < num_bonds; n++) {
+               struct intel_engine_cs *bond;
+
+               if (copy_from_user(&ci, &ext->engines[n], sizeof(ci)))
+                       return -EFAULT;
+
+               bond = intel_engine_lookup_user(set->ctx->i915,
+                                               ci.engine_class,
+                                               ci.engine_instance);
+               if (!bond) {
+                       DRM_DEBUG("Unrecognised engine[%d] for bonding: { class:%d, instance: %d }\n",
+                                 n, ci.engine_class, ci.engine_instance);
+                       return -EINVAL;
+               }
+
+               /*
+                * A non-virtual engine has no siblings to choose between; and
+                * a submit fence will always be directed to the one engine.
+                */
+               if (intel_engine_is_virtual(virtual)) {
+                       err = intel_virtual_engine_attach_bond(virtual,
+                                                              master,
+                                                              bond);
+                       if (err)
+                               return err;
+               }
+       }
+
+       return 0;
+}
+
 static const i915_user_extension_fn set_engines__extensions[] = {
        [I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE] = set_engines__load_balance,
+       [I915_CONTEXT_ENGINES_EXT_BOND] = set_engines__bond,
 };
 
 static int
index 2bfa72c..b976c12 100644 (file)
@@ -45,6 +45,9 @@ void __onstack_fence_init(struct i915_sw_fence *fence,
 
 void onstack_fence_fini(struct i915_sw_fence *fence)
 {
+       if (!fence->flags)
+               return;
+
        i915_sw_fence_commit(fence);
        i915_sw_fence_fini(fence);
 }
index f977094..e2da902 100644 (file)
@@ -1543,6 +1543,10 @@ struct drm_i915_gem_context_param {
  * sized argument, will revert back to default settings.
  *
  * See struct i915_context_param_engines.
+ *
+ * Extensions:
+ *   i915_context_engines_load_balance (I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE)
+ *   i915_context_engines_bond (I915_CONTEXT_ENGINES_EXT_BOND)
  */
 #define I915_CONTEXT_PARAM_ENGINES     0xa
 /* Must be kept compact -- no holes and well documented */
@@ -1646,9 +1650,49 @@ struct i915_context_engines_load_balance {
        struct i915_engine_class_instance engines[N__]; \
 } __attribute__((packed)) name__
 
+/*
+ * i915_context_engines_bond:
+ *
+ * Constructed bonded pairs for execution within a virtual engine.
+ *
+ * All engines are equal, but some are more equal than others. Given
+ * the distribution of resources in the HW, it may be preferable to run
+ * a request on a given subset of engines in parallel to a request on a
+ * specific engine. We enable this selection of engines within a virtual
+ * engine by specifying bonding pairs, for any given master engine we will
+ * only execute on one of the corresponding siblings within the virtual engine.
+ *
+ * To execute a request in parallel on the master engine and a sibling requires
+ * coordination with a I915_EXEC_FENCE_SUBMIT.
+ */
+struct i915_context_engines_bond {
+       struct i915_user_extension base;
+
+       struct i915_engine_class_instance master;
+
+       __u16 virtual_index; /* index of virtual engine in ctx->engines[] */
+       __u16 num_bonds;
+
+       __u64 flags; /* all undefined flags must be zero */
+       __u64 mbz64[4]; /* reserved for future use; must be zero */
+
+       struct i915_engine_class_instance engines[0];
+} __attribute__((packed));
+
+#define I915_DEFINE_CONTEXT_ENGINES_BOND(name__, N__) struct { \
+       struct i915_user_extension base; \
+       struct i915_engine_class_instance master; \
+       __u16 virtual_index; \
+       __u16 num_bonds; \
+       __u64 flags; \
+       __u64 mbz64[4]; \
+       struct i915_engine_class_instance engines[N__]; \
+} __attribute__((packed)) name__
+
 struct i915_context_param_engines {
        __u64 extensions; /* linked chain of extension blocks, 0 terminates */
 #define I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE 0 /* see i915_context_engines_load_balance */
+#define I915_CONTEXT_ENGINES_EXT_BOND 1 /* see i915_context_engines_bond */
        struct i915_engine_class_instance engines[0];
 } __attribute__((packed));