drm/i915/execlists: Always reset the context's RING registers
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 11 Apr 2019 13:05:15 +0000 (14:05 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 11 Apr 2019 19:48:52 +0000 (20:48 +0100)
During reset, we try and stop the active ring. This has the consequence
that we often clobber the RING registers within the context image. When
we find an active request, we update the context image to rerun that
request (if it was guilty, we replace the hanging user payload with
NOPs). However, we were ignoring an active context if the request had
completed, with the consequence that the next submission on that request
would start with RING_HEAD==0 and not the tail of the previous request,
causing all requests still in the ring to be rerun. Rare, but
occasionally seen within CI where we would spot that the context seqno
would reverse and complain that we were retiring an incomplete request.

    <0> [412.390350]   <idle>-0       3d.s2 408373352us : __i915_request_submit: rcs0 fence 1e95b:3640 -> current 3638
    <0> [412.390350]   <idle>-0       3d.s2 408373353us : __i915_request_submit: rcs0 fence 1e95b:3642 -> current 3638
    <0> [412.390350]   <idle>-0       3d.s2 408373354us : __i915_request_submit: rcs0 fence 1e95b:3644 -> current 3638
    <0> [412.390350]   <idle>-0       3d.s2 408373354us : __i915_request_submit: rcs0 fence 1e95b:3646 -> current 3638
    <0> [412.390350]   <idle>-0       3d.s2 408373356us : __execlists_submission_tasklet: rcs0 in[0]:  ctx=2.1, fence 1e95b:3646 (current 3638), prio=4
    <0> [412.390350] i915_sel-4613    0.... 408373374us : __i915_request_commit: rcs0 fence 1e95b:3648
    <0> [412.390350] i915_sel-4613    0d..1 408373377us : process_csb: rcs0 cs-irq head=2, tail=3
    <0> [412.390350] i915_sel-4613    0d..1 408373377us : process_csb: rcs0 csb[3]: status=0x00000001:0x00000000, active=0x1
    <0> [412.390350] i915_sel-4613    0d..1 408373378us : __i915_request_submit: rcs0 fence 1e95b:3648 -> current 3638
    <0> [412.390350]   <idle>-0       3..s1 408373378us : execlists_submission_tasklet: rcs0 awake?=1, active=5
    <0> [412.390350] i915_sel-4613    0d..1 408373379us : __execlists_submission_tasklet: rcs0 in[0]:  ctx=2.2, fence 1e95b:3648 (current 3638), prio=4
    <0> [412.390350] i915_sel-4613    0.... 408373381us : i915_reset_engine: rcs0 flags=4
    <0> [412.390350] i915_sel-4613    0.... 408373382us : execlists_reset_prepare: rcs0: depth<-0
    <0> [412.390350]   <idle>-0       3d.s2 408373390us : process_csb: rcs0 cs-irq head=3, tail=4
    <0> [412.390350]   <idle>-0       3d.s2 408373390us : process_csb: rcs0 csb[4]: status=0x00008002:0x00000002, active=0x1
    <0> [412.390350]   <idle>-0       3d.s2 408373390us : process_csb: rcs0 out[0]: ctx=2.2, fence 1e95b:3648 (current 3640), prio=4
    <0> [412.390350] i915_sel-4613    0.... 408373401us : intel_engine_stop_cs: rcs0
    <0> [412.390350] i915_sel-4613    0d..1 408373402us : process_csb: rcs0 cs-irq head=4, tail=4
    <0> [412.390350] i915_sel-4613    0.... 408373403us : intel_gpu_reset: engine_mask=1
    <0> [412.390350] i915_sel-4613    0d..1 408373408us : execlists_cancel_port_requests: rcs0:port0 fence 1e95b:3648, (current 3648)
    <0> [412.390350] i915_sel-4613    0.... 408373442us : intel_engine_cancel_stop_cs: rcs0
    <0> [412.390350] i915_sel-4613    0.... 408373442us : execlists_reset_finish: rcs0: depth->0
    <0> [412.390350] ksoftirq-26      3..s. 408373442us : execlists_submission_tasklet: rcs0 awake?=1, active=0
    <0> [412.390350] ksoftirq-26      3d.s1 408373443us : process_csb: rcs0 cs-irq head=5, tail=5
    <0> [412.390350] i915_sel-4613    0.... 408373475us : i915_request_retire: rcs0 fence 1e95b:3640, current 3648
    <0> [412.390350] i915_sel-4613    0.... 408373476us : i915_request_retire: __retire_engine_request(rcs0) fence 1e95b:3640, current 3648
    <0> [412.390350] i915_sel-4613    0.... 408373494us : __i915_request_commit: rcs0 fence 1e95b:3650
    <0> [412.390350] i915_sel-4613    0d..1 408373496us : process_csb: rcs0 cs-irq head=5, tail=5
    <0> [412.390350] i915_sel-4613    0d..1 408373496us : __i915_request_submit: rcs0 fence 1e95b:3650 -> current 3648
    <0> [412.390350] i915_sel-4613    0d..1 408373498us : __execlists_submission_tasklet: rcs0 in[0]:  ctx=2.1, fence 1e95b:3650 (current 3648), prio=6
    <0> [412.390350] i915_sel-4613    0.... 408373500us : i915_request_retire_upto: rcs0 fence 1e95b:3648, current 3648
    <0> [412.390350] i915_sel-4613    0.... 408373500us : i915_request_retire: rcs0 fence 1e95b:3642, current 3648
    <0> [412.390350] i915_sel-4613    0.... 408373501us : i915_request_retire: __retire_engine_request(rcs0) fence 1e95b:3642, current 3648
    <0> [412.390350] i915_sel-4613    0.... 408373514us : i915_request_retire: rcs0 fence 1e95b:3644, current 3648
    <0> [412.390350] i915_sel-4613    0.... 408373515us : i915_request_retire: __retire_engine_request(rcs0) fence 1e95b:3644, current 3648
    <0> [412.390350] i915_sel-4613    0.... 408373527us : i915_request_retire: rcs0 fence 1e95b:3646, current 3640
    <0> [412.390350]   <idle>-0       3..s1 408373569us : execlists_submission_tasklet: rcs0 awake?=1, active=1
    <0> [412.390350]   <idle>-0       3d.s2 408373569us : process_csb: rcs0 cs-irq head=5, tail=1
    <0> [412.390350]   <idle>-0       3d.s2 408373570us : process_csb: rcs0 csb[0]: status=0x00000001:0x00000000, active=0x1
    <0> [412.390350]   <idle>-0       3d.s2 408373570us : process_csb: rcs0 csb[1]: status=0x00000018:0x00000002, active=0x5
    <0> [412.390350]   <idle>-0       3d.s2 408373570us : process_csb: rcs0 out[0]: ctx=2.1, fence 1e95b:3650 (current 3650), prio=6
    <0> [412.390350]   <idle>-0       3d.s2 408373571us : process_csb: rcs0 completed ctx=2
    <0> [412.390350] i915_sel-4613    0.... 408373621us : i915_request_retire: i915_request_retire:253 GEM_BUG_ON(!i915_request_completed(request))

v2: Fixup the cancellation path to drain the CSB and reset the pointers.

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/20190411130515.20716-2-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/intel_lrc.c

index 95ec4f7ad0838164f146d61a9e9390534088df3b..3cb788a223ef07bb418461b51bec34a6ad33f013 100644 (file)
@@ -893,96 +893,6 @@ invalidate_csb_entries(const u32 *first, const u32 *last)
        clflush((void *)last);
 }
 
-static void reset_csb_pointers(struct intel_engine_execlists *execlists)
-{
-       const unsigned int reset_value = execlists->csb_size - 1;
-
-       /*
-        * After a reset, the HW starts writing into CSB entry [0]. We
-        * therefore have to set our HEAD pointer back one entry so that
-        * the *first* entry we check is entry 0. To complicate this further,
-        * as we don't wait for the first interrupt after reset, we have to
-        * fake the HW write to point back to the last entry so that our
-        * inline comparison of our cached head position against the last HW
-        * write works even before the first interrupt.
-        */
-       execlists->csb_head = reset_value;
-       WRITE_ONCE(*execlists->csb_write, reset_value);
-
-       invalidate_csb_entries(&execlists->csb_status[0],
-                              &execlists->csb_status[GEN8_CSB_ENTRIES - 1]);
-}
-
-static void nop_submission_tasklet(unsigned long data)
-{
-       /* The driver is wedged; don't process any more events. */
-}
-
-static void execlists_cancel_requests(struct intel_engine_cs *engine)
-{
-       struct intel_engine_execlists * const execlists = &engine->execlists;
-       struct i915_request *rq, *rn;
-       struct rb_node *rb;
-       unsigned long flags;
-
-       GEM_TRACE("%s\n", engine->name);
-
-       /*
-        * Before we call engine->cancel_requests(), we should have exclusive
-        * access to the submission state. This is arranged for us by the
-        * caller disabling the interrupt generation, the tasklet and other
-        * threads that may then access the same state, giving us a free hand
-        * to reset state. However, we still need to let lockdep be aware that
-        * we know this state may be accessed in hardirq context, so we
-        * disable the irq around this manipulation and we want to keep
-        * the spinlock focused on its duties and not accidentally conflate
-        * coverage to the submission's irq state. (Similarly, although we
-        * shouldn't need to disable irq around the manipulation of the
-        * submission's irq state, we also wish to remind ourselves that
-        * it is irq state.)
-        */
-       spin_lock_irqsave(&engine->timeline.lock, flags);
-
-       /* Cancel the requests on the HW and clear the ELSP tracker. */
-       execlists_cancel_port_requests(execlists);
-       execlists_user_end(execlists);
-
-       /* Mark all executing requests as skipped. */
-       list_for_each_entry(rq, &engine->timeline.requests, link) {
-               if (!i915_request_signaled(rq))
-                       dma_fence_set_error(&rq->fence, -EIO);
-
-               i915_request_mark_complete(rq);
-       }
-
-       /* Flush the queued requests to the timeline list (for retiring). */
-       while ((rb = rb_first_cached(&execlists->queue))) {
-               struct i915_priolist *p = to_priolist(rb);
-               int i;
-
-               priolist_for_each_request_consume(rq, rn, p, i) {
-                       list_del_init(&rq->sched.link);
-                       __i915_request_submit(rq);
-                       dma_fence_set_error(&rq->fence, -EIO);
-                       i915_request_mark_complete(rq);
-               }
-
-               rb_erase_cached(&p->node, &execlists->queue);
-               i915_priolist_free(p);
-       }
-
-       /* Remaining _unready_ requests will be nop'ed when submitted */
-
-       execlists->queue_priority_hint = INT_MIN;
-       execlists->queue = RB_ROOT_CACHED;
-       GEM_BUG_ON(port_isset(execlists->port));
-
-       GEM_BUG_ON(__tasklet_is_enabled(&execlists->tasklet));
-       execlists->tasklet.func = nop_submission_tasklet;
-
-       spin_unlock_irqrestore(&engine->timeline.lock, flags);
-}
-
 static inline bool
 reset_in_progress(const struct intel_engine_execlists *execlists)
 {
@@ -1152,7 +1062,7 @@ static void process_csb(struct intel_engine_cs *engine)
         * the wash as hardware, working or not, will need to do the
         * invalidation before.
         */
-       invalidate_csb_entries(&buf[0], &buf[GEN8_CSB_ENTRIES - 1]);
+       invalidate_csb_entries(&buf[0], &buf[num_entries - 1]);
 }
 
 static void __execlists_submission_tasklet(struct intel_engine_cs *const engine)
@@ -1921,7 +1831,6 @@ static void execlists_reset_prepare(struct intel_engine_cs *engine)
 
        /* And flush any current direct submission. */
        spin_lock_irqsave(&engine->timeline.lock, flags);
-       process_csb(engine); /* drain preemption events */
        spin_unlock_irqrestore(&engine->timeline.lock, flags);
 }
 
@@ -1942,14 +1851,47 @@ static bool lrc_regs_ok(const struct i915_request *rq)
        return true;
 }
 
-static void execlists_reset(struct intel_engine_cs *engine, bool stalled)
+static void reset_csb_pointers(struct intel_engine_execlists *execlists)
+{
+       const unsigned int reset_value = execlists->csb_size - 1;
+
+       /*
+        * After a reset, the HW starts writing into CSB entry [0]. We
+        * therefore have to set our HEAD pointer back one entry so that
+        * the *first* entry we check is entry 0. To complicate this further,
+        * as we don't wait for the first interrupt after reset, we have to
+        * fake the HW write to point back to the last entry so that our
+        * inline comparison of our cached head position against the last HW
+        * write works even before the first interrupt.
+        */
+       execlists->csb_head = reset_value;
+       WRITE_ONCE(*execlists->csb_write, reset_value);
+
+       invalidate_csb_entries(&execlists->csb_status[0],
+                              &execlists->csb_status[reset_value]);
+}
+
+static void __execlists_reset(struct intel_engine_cs *engine, bool stalled)
 {
        struct intel_engine_execlists * const execlists = &engine->execlists;
+       struct intel_context *ce;
        struct i915_request *rq;
-       unsigned long flags;
        u32 *regs;
 
-       spin_lock_irqsave(&engine->timeline.lock, flags);
+       process_csb(engine); /* drain preemption events */
+
+       /* Following the reset, we need to reload the CSB read/write pointers */
+       reset_csb_pointers(&engine->execlists);
+
+       /*
+        * Save the currently executing context, even if we completed
+        * its request, it was still running at the time of the
+        * reset and will have been clobbered.
+        */
+       if (!port_isset(execlists->port))
+               goto out_clear;
+
+       ce = port_request(execlists->port)->hw_context;
 
        /*
         * Catch up with any missed context-switch interrupts.
@@ -1964,12 +1906,13 @@ static void execlists_reset(struct intel_engine_cs *engine, bool stalled)
 
        /* Push back any incomplete requests for replay after the reset. */
        rq = __unwind_incomplete_requests(engine);
-
-       /* Following the reset, we need to reload the CSB read/write pointers */
-       reset_csb_pointers(&engine->execlists);
-
        if (!rq)
-               goto out_unlock;
+               goto out_replay;
+
+       if (rq->hw_context != ce) { /* caught just before a CS event */
+               rq = NULL;
+               goto out_replay;
+       }
 
        /*
         * If this request hasn't started yet, e.g. it is waiting on a
@@ -1984,7 +1927,7 @@ static void execlists_reset(struct intel_engine_cs *engine, bool stalled)
         * perfectly and we do not need to flag the result as being erroneous.
         */
        if (!i915_request_started(rq) && lrc_regs_ok(rq))
-               goto out_unlock;
+               goto out_replay;
 
        /*
         * If the request was innocent, we leave the request in the ELSP
@@ -1999,7 +1942,7 @@ static void execlists_reset(struct intel_engine_cs *engine, bool stalled)
         */
        i915_reset_request(rq, stalled);
        if (!stalled && lrc_regs_ok(rq))
-               goto out_unlock;
+               goto out_replay;
 
        /*
         * We want a simple context + ring to execute the breadcrumb update.
@@ -2009,21 +1952,103 @@ static void execlists_reset(struct intel_engine_cs *engine, bool stalled)
         * future request will be after userspace has had the opportunity
         * to recreate its own state.
         */
-       regs = rq->hw_context->lrc_reg_state;
+       regs = ce->lrc_reg_state;
        if (engine->pinned_default_state) {
                memcpy(regs, /* skip restoring the vanilla PPHWSP */
                       engine->pinned_default_state + LRC_STATE_PN * PAGE_SIZE,
                       engine->context_size - PAGE_SIZE);
        }
+       execlists_init_reg_state(regs, ce, engine, ce->ring);
 
        /* Rerun the request; its payload has been neutered (if guilty). */
-       rq->ring->head = intel_ring_wrap(rq->ring, rq->head);
-       intel_ring_update_space(rq->ring);
+out_replay:
+       ce->ring->head =
+               rq ? intel_ring_wrap(ce->ring, rq->head) : ce->ring->tail;
+       intel_ring_update_space(ce->ring);
+       __execlists_update_reg_state(ce, engine);
+
+out_clear:
+       execlists_clear_all_active(execlists);
+}
+
+static void execlists_reset(struct intel_engine_cs *engine, bool stalled)
+{
+       unsigned long flags;
+
+       GEM_TRACE("%s\n", engine->name);
+
+       spin_lock_irqsave(&engine->timeline.lock, flags);
+
+       __execlists_reset(engine, stalled);
+
+       spin_unlock_irqrestore(&engine->timeline.lock, flags);
+}
+
+static void nop_submission_tasklet(unsigned long data)
+{
+       /* The driver is wedged; don't process any more events. */
+}
+
+static void execlists_cancel_requests(struct intel_engine_cs *engine)
+{
+       struct intel_engine_execlists * const execlists = &engine->execlists;
+       struct i915_request *rq, *rn;
+       struct rb_node *rb;
+       unsigned long flags;
 
-       execlists_init_reg_state(regs, rq->hw_context, engine, rq->ring);
-       __execlists_update_reg_state(rq->hw_context, engine);
+       GEM_TRACE("%s\n", engine->name);
+
+       /*
+        * Before we call engine->cancel_requests(), we should have exclusive
+        * access to the submission state. This is arranged for us by the
+        * caller disabling the interrupt generation, the tasklet and other
+        * threads that may then access the same state, giving us a free hand
+        * to reset state. However, we still need to let lockdep be aware that
+        * we know this state may be accessed in hardirq context, so we
+        * disable the irq around this manipulation and we want to keep
+        * the spinlock focused on its duties and not accidentally conflate
+        * coverage to the submission's irq state. (Similarly, although we
+        * shouldn't need to disable irq around the manipulation of the
+        * submission's irq state, we also wish to remind ourselves that
+        * it is irq state.)
+        */
+       spin_lock_irqsave(&engine->timeline.lock, flags);
+
+       __execlists_reset(engine, true);
+
+       /* Mark all executing requests as skipped. */
+       list_for_each_entry(rq, &engine->timeline.requests, link) {
+               if (!i915_request_signaled(rq))
+                       dma_fence_set_error(&rq->fence, -EIO);
+
+               i915_request_mark_complete(rq);
+       }
+
+       /* Flush the queued requests to the timeline list (for retiring). */
+       while ((rb = rb_first_cached(&execlists->queue))) {
+               struct i915_priolist *p = to_priolist(rb);
+               int i;
+
+               priolist_for_each_request_consume(rq, rn, p, i) {
+                       list_del_init(&rq->sched.link);
+                       __i915_request_submit(rq);
+                       dma_fence_set_error(&rq->fence, -EIO);
+                       i915_request_mark_complete(rq);
+               }
+
+               rb_erase_cached(&p->node, &execlists->queue);
+               i915_priolist_free(p);
+       }
+
+       /* Remaining _unready_ requests will be nop'ed when submitted */
+
+       execlists->queue_priority_hint = INT_MIN;
+       execlists->queue = RB_ROOT_CACHED;
+       GEM_BUG_ON(port_isset(execlists->port));
+
+       GEM_BUG_ON(__tasklet_is_enabled(&execlists->tasklet));
+       execlists->tasklet.func = nop_submission_tasklet;
 
-out_unlock:
        spin_unlock_irqrestore(&engine->timeline.lock, flags);
 }