drm/i915: Unwind incomplete legacy context switches
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 23 Nov 2017 15:26:30 +0000 (15:26 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 23 Nov 2017 16:12:04 +0000 (16:12 +0000)
The legacy context switch for ringbuffer submission is multistaged,
where each of those stages may fail. However, we were updating global
state after some stages, and so we had to force the incomplete request
to be submitted because we could not unwind. Save the global state
before performing the switches, and so enable us to unwind back to the
previous global state should any phase fail. We then must cancel the
request instead of submitting it should the construction fail.

v2: s/saved_ctx/from_ctx/; s/ctx/to_ctx/ etc.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171123152631.31385-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/i915_gem_context.c
drivers/gpu/drm/i915/i915_gem_request.c
drivers/gpu/drm/i915/intel_ringbuffer.c
drivers/gpu/drm/i915/intel_ringbuffer.h

index 6ca56e482d793cfcd1f6e9bf0713debf3c1e91f3..da4a64b3484ca22413bf5a7691e457c3bd97ff33 100644 (file)
@@ -507,6 +507,7 @@ void i915_gem_contexts_lost(struct drm_i915_private *dev_priv)
 
        for_each_engine(engine, dev_priv, id) {
                engine->legacy_active_context = NULL;
+               engine->legacy_active_ppgtt = NULL;
 
                if (!engine->last_retired_context)
                        continue;
@@ -681,68 +682,49 @@ static int remap_l3(struct drm_i915_gem_request *req, int slice)
        return 0;
 }
 
-static inline bool skip_rcs_switch(struct i915_hw_ppgtt *ppgtt,
-                                  struct intel_engine_cs *engine,
-                                  struct i915_gem_context *to)
-{
-       if (to->remap_slice)
-               return false;
-
-       if (ppgtt && (intel_engine_flag(engine) & ppgtt->pd_dirty_rings))
-               return false;
-
-       return to == engine->legacy_active_context;
-}
-
-static bool
-needs_pd_load_pre(struct i915_hw_ppgtt *ppgtt, struct intel_engine_cs *engine)
-{
-       struct i915_gem_context *from = engine->legacy_active_context;
-
-       if (!ppgtt)
-               return false;
-
-       /* Always load the ppgtt on first use */
-       if (!from)
-               return true;
-
-       /* Same context without new entries, skip */
-       if ((!from->ppgtt || from->ppgtt == ppgtt) &&
-           !(intel_engine_flag(engine) & ppgtt->pd_dirty_rings))
-               return false;
-
-       if (engine->id != RCS)
-               return true;
-
-       return true;
-}
-
-static int do_rcs_switch(struct drm_i915_gem_request *req)
+/**
+ * i915_switch_context() - perform a GPU context switch.
+ * @rq: request for which we'll execute the context switch
+ *
+ * The context life cycle is simple. The context refcount is incremented and
+ * decremented by 1 on create and destroy. If the context is in use by the GPU,
+ * it will have a refcount > 1. This allows us to destroy the context abstract
+ * object while letting the normal object tracking destroy the backing BO.
+ *
+ * This function should not be used in execlists mode.  Instead the context is
+ * switched by writing to the ELSP and requests keep a reference to their
+ * context.
+ */
+int i915_switch_context(struct drm_i915_gem_request *rq)
 {
-       struct i915_gem_context *to = req->ctx;
-       struct intel_engine_cs *engine = req->engine;
-       struct i915_hw_ppgtt *ppgtt = to->ppgtt ?: req->i915->mm.aliasing_ppgtt;
-       struct i915_gem_context *from = engine->legacy_active_context;
-       u32 hw_flags;
+       struct intel_engine_cs *engine = rq->engine;
+       struct i915_gem_context *to_ctx = rq->ctx;
+       struct i915_hw_ppgtt *to_mm =
+               to_ctx->ppgtt ?: rq->i915->mm.aliasing_ppgtt;
+       struct i915_gem_context *from_ctx = engine->legacy_active_context;
+       struct i915_hw_ppgtt *from_mm = engine->legacy_active_ppgtt;
+       u32 hw_flags = 0;
        int ret, i;
 
-       GEM_BUG_ON(engine->id != RCS);
+       lockdep_assert_held(&rq->i915->drm.struct_mutex);
+       GEM_BUG_ON(HAS_EXECLISTS(rq->i915));
 
-       if (skip_rcs_switch(ppgtt, engine, to))
-               return 0;
-
-       if (needs_pd_load_pre(ppgtt, engine)) {
-               /* Older GENs and non render rings still want the load first,
-                * "PP_DCLV followed by PP_DIR_BASE register through Load
-                * Register Immediate commands in Ring Buffer before submitting
-                * a context."*/
-               trace_switch_mm(engine, to);
-               ret = ppgtt->switch_mm(ppgtt, req);
+       if (to_mm != from_mm ||
+           (to_mm && intel_engine_flag(engine) & to_mm->pd_dirty_rings)) {
+               trace_switch_mm(engine, to_ctx);
+               ret = to_mm->switch_mm(to_mm, rq);
                if (ret)
-                       return ret;
+                       goto err;
+
+               to_mm->pd_dirty_rings &= ~intel_engine_flag(engine);
+               engine->legacy_active_ppgtt = to_mm;
+               hw_flags = MI_FORCE_RESTORE;
        }
 
-       if (i915_gem_context_is_kernel(to))
+       if (to_ctx->engine[engine->id].state &&
+           (to_ctx != from_ctx || hw_flags & MI_FORCE_RESTORE)) {
+               GEM_BUG_ON(engine->id != RCS);
+
                /*
                 * The kernel context(s) is treated as pure scratch and is not
                 * expected to retain any state (as we sacrifice it during
@@ -750,78 +732,37 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
                 * as nothing actually executes using the kernel context; it
                 * is purely used for flushing user contexts.
                 */
-               hw_flags = MI_RESTORE_INHIBIT;
-       else if (ppgtt && intel_engine_flag(engine) & ppgtt->pd_dirty_rings)
-               hw_flags = MI_FORCE_RESTORE;
-       else
-               hw_flags = 0;
+               if (i915_gem_context_is_kernel(to_ctx))
+                       hw_flags = MI_RESTORE_INHIBIT;
 
-       if (to != from || (hw_flags & MI_FORCE_RESTORE)) {
-               ret = mi_set_context(req, hw_flags);
+               ret = mi_set_context(rq, hw_flags);
                if (ret)
-                       return ret;
+                       goto err_mm;
 
-               engine->legacy_active_context = to;
+               engine->legacy_active_context = to_ctx;
        }
 
-       if (ppgtt)
-               ppgtt->pd_dirty_rings &= ~intel_engine_flag(engine);
+       if (to_ctx->remap_slice) {
+               for (i = 0; i < MAX_L3_SLICES; i++) {
+                       if (!(to_ctx->remap_slice & BIT(i)))
+                               continue;
 
-       for (i = 0; i < MAX_L3_SLICES; i++) {
-               if (!(to->remap_slice & (1<<i)))
-                       continue;
-
-               ret = remap_l3(req, i);
-               if (ret)
-                       return ret;
-
-               to->remap_slice &= ~(1<<i);
-       }
-
-       return 0;
-}
-
-/**
- * i915_switch_context() - perform a GPU context switch.
- * @req: request for which we'll execute the context switch
- *
- * The context life cycle is simple. The context refcount is incremented and
- * decremented by 1 and create and destroy. If the context is in use by the GPU,
- * it will have a refcount > 1. This allows us to destroy the context abstract
- * object while letting the normal object tracking destroy the backing BO.
- *
- * This function should not be used in execlists mode.  Instead the context is
- * switched by writing to the ELSP and requests keep a reference to their
- * context.
- */
-int i915_switch_context(struct drm_i915_gem_request *req)
-{
-       struct intel_engine_cs *engine = req->engine;
-
-       lockdep_assert_held(&req->i915->drm.struct_mutex);
-       GEM_BUG_ON(HAS_EXECLISTS(req->i915));
-
-       if (!req->ctx->engine[engine->id].state) {
-               struct i915_gem_context *to = req->ctx;
-               struct i915_hw_ppgtt *ppgtt =
-                       to->ppgtt ?: req->i915->mm.aliasing_ppgtt;
-
-               if (needs_pd_load_pre(ppgtt, engine)) {
-                       int ret;
-
-                       trace_switch_mm(engine, to);
-                       ret = ppgtt->switch_mm(ppgtt, req);
+                       ret = remap_l3(rq, i);
                        if (ret)
-                               return ret;
-
-                       ppgtt->pd_dirty_rings &= ~intel_engine_flag(engine);
+                               goto err_ctx;
                }
 
-               engine->legacy_active_context = to;
-               return 0;
+               to_ctx->remap_slice = 0;
        }
 
-       return do_rcs_switch(req);
+       return 0;
+
+err_ctx:
+       engine->legacy_active_context = from_ctx;
+err_mm:
+       engine->legacy_active_ppgtt = from_mm;
+err:
+       return ret;
 }
 
 static bool engine_has_idle_kernel_context(struct intel_engine_cs *engine)
index 7325469ce754ac044e13f860d08fd03480f08837..a90bdd26571f58cffe8300d1a1b74eebf3c6005b 100644 (file)
@@ -719,25 +719,19 @@ i915_gem_request_alloc(struct intel_engine_cs *engine,
        /* Unconditionally invalidate GPU caches and TLBs. */
        ret = engine->emit_flush(req, EMIT_INVALIDATE);
        if (ret)
-               goto err_ctx;
+               goto err_unwind;
 
        ret = engine->request_alloc(req);
-       if (ret) {
-               /*
-                * Past the point-of-no-return. Since we may have updated
-                * global state after partially completing the request alloc,
-                * we need to commit any commands so far emitted in the
-                * request to the HW.
-                */
-               __i915_add_request(req, false);
-               return ERR_PTR(ret);
-       }
+       if (ret)
+               goto err_unwind;
 
        /* Check that we didn't interrupt ourselves with a new request */
        GEM_BUG_ON(req->timeline->seqno != req->fence.seqno);
        return req;
 
-err_ctx:
+err_unwind:
+       req->ring->emit = req->head;
+
        /* Make sure we didn't add ourselves to external state before freeing */
        GEM_BUG_ON(!list_empty(&req->active_list));
        GEM_BUG_ON(!list_empty(&req->priotree.signalers_list));
index bfa11a84e476e51fd01e0b766dd84c062e36f706..a904b0353bec73275a7c52d3552713817a0051fe 100644 (file)
@@ -591,6 +591,7 @@ static void reset_ring_common(struct intel_engine_cs *engine,
                        request->ring->head = request->postfix;
        } else {
                engine->legacy_active_context = NULL;
+               engine->legacy_active_ppgtt = NULL;
        }
 }
 
index d38d059285dcac309097e4a73daac0946d3f5025..ef304c99954c8573412757d9fece2d7c533d5a8d 100644 (file)
@@ -531,6 +531,7 @@ struct intel_engine_cs {
         * stream (ring).
         */
        struct i915_gem_context *legacy_active_context;
+       struct i915_hw_ppgtt *legacy_active_ppgtt;
 
        /* status_notifier: list of callbacks for context-switch changes */
        struct atomic_notifier_head context_status_notifier;