]> git.itanic.dy.fi Git - linux-stable/commitdiff
drm/i915/gt: Incrementally check for rewinding
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 9 Jun 2020 15:17:23 +0000 (16:17 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 24 Jun 2020 15:49:22 +0000 (17:49 +0200)
commit 8ab3a3812aa90e488813e719308ffd807b865624 upstream.

In commit 5ba32c7be81e ("drm/i915/execlists: Always force a context
reload when rewinding RING_TAIL"), we placed the check for rewinding a
context on actually submitting the next request in that context. This
was so that we only had to check once, and could do so with precision
avoiding as many forced restores as possible. For example, to ensure
that we can resubmit the same request a couple of times, we include a
small wa_tail such that on the next submission, the ring->tail will
appear to move forwards when resubmitting the same request. This is very
common as it will happen for every lite-restore to fill the second port
after a context switch.

However, intel_ring_direction() is limited in precision to movements of
upto half the ring size. The consequence being that if we tried to
unwind many requests, we could exceed half the ring and flip the sense
of the direction, so missing a force restore. As no request can be
greater than half the ring (i.e. 2048 bytes in the smallest case), we
can check for rollback incrementally. As we check against the tail that
would be submitted, we do not lose any sensitivity and allow lite
restores for the simple case. We still need to double check upon
submitting the context, to allow for multiple preemptions and
resubmissions.

Fixes: 5ba32c7be81e ("drm/i915/execlists: Always force a context reload when rewinding RING_TAIL")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: <stable@vger.kernel.org> # v5.4+
Reviewed-by: Bruce Chang <yu.bruce.chang@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200609151723.12971-1-chris@chris-wilson.co.uk
(cherry picked from commit e36ba817fa966f81fb1c8d16f3721b5a644b2fa9)
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/gpu/drm/i915/gt/intel_engine_cs.c
drivers/gpu/drm/i915/gt/intel_lrc.c
drivers/gpu/drm/i915/gt/intel_ring.c
drivers/gpu/drm/i915/gt/selftest_mocs.c
drivers/gpu/drm/i915/gt/selftest_ring.c [new file with mode: 0644]
drivers/gpu/drm/i915/selftests/i915_mock_selftests.h

index 883a9b7fe88d5a6a0b4398d3a3793b727b015343..55b9165e7533be29a1466401f6dfb9efee8f75ad 100644 (file)
@@ -639,7 +639,7 @@ static int engine_setup_common(struct intel_engine_cs *engine)
 struct measure_breadcrumb {
        struct i915_request rq;
        struct intel_ring ring;
-       u32 cs[1024];
+       u32 cs[2048];
 };
 
 static int measure_breadcrumb_dw(struct intel_context *ce)
@@ -661,6 +661,8 @@ static int measure_breadcrumb_dw(struct intel_context *ce)
 
        frame->ring.vaddr = frame->cs;
        frame->ring.size = sizeof(frame->cs);
+       frame->ring.wrap =
+               BITS_PER_TYPE(frame->ring.size) - ilog2(frame->ring.size);
        frame->ring.effective_size = frame->ring.size;
        intel_ring_update_space(&frame->ring);
        frame->rq.ring = &frame->ring;
index 2dfaddb8811edf863a1cec02bdefcebfd8941dd0..ba82193b4e319be4172f86f303d188402fa52891 100644 (file)
@@ -972,6 +972,13 @@ __unwind_incomplete_requests(struct intel_engine_cs *engine)
                        list_move(&rq->sched.link, pl);
                        set_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);
 
+                       /* Check in case we rollback so far we wrap [size/2] */
+                       if (intel_ring_direction(rq->ring,
+                                                intel_ring_wrap(rq->ring,
+                                                                rq->tail),
+                                                rq->ring->tail) > 0)
+                               rq->context->lrc.desc |= CTX_DESC_FORCE_RESTORE;
+
                        active = rq;
                } else {
                        struct intel_engine_cs *owner = rq->context->engine;
@@ -1383,8 +1390,9 @@ static u64 execlists_update_context(struct i915_request *rq)
         * HW has a tendency to ignore us rewinding the TAIL to the end of
         * an earlier request.
         */
+       GEM_BUG_ON(ce->lrc_reg_state[CTX_RING_TAIL] != rq->ring->tail);
+       prev = rq->ring->tail;
        tail = intel_ring_set_tail(rq->ring, rq->tail);
-       prev = ce->lrc_reg_state[CTX_RING_TAIL];
        if (unlikely(intel_ring_direction(rq->ring, tail, prev) <= 0))
                desc |= CTX_DESC_FORCE_RESTORE;
        ce->lrc_reg_state[CTX_RING_TAIL] = tail;
@@ -4213,6 +4221,14 @@ static int gen12_emit_flush_render(struct i915_request *request,
        return 0;
 }
 
+static void assert_request_valid(struct i915_request *rq)
+{
+       struct intel_ring *ring __maybe_unused = rq->ring;
+
+       /* Can we unwind this request without appearing to go forwards? */
+       GEM_BUG_ON(intel_ring_direction(ring, rq->wa_tail, rq->head) <= 0);
+}
+
 /*
  * Reserve space for 2 NOOPs at the end of each request to be
  * used as a workaround for not being allowed to do lite
@@ -4225,6 +4241,9 @@ static u32 *gen8_emit_wa_tail(struct i915_request *request, u32 *cs)
        *cs++ = MI_NOOP;
        request->wa_tail = intel_ring_offset(request, cs);
 
+       /* Check that entire request is less than half the ring */
+       assert_request_valid(request);
+
        return cs;
 }
 
index 8cda1b7e17ba374d3d892bccfd689df3814fdb18..bdb324167ef3356728297edad1d1f512ba0496f9 100644 (file)
@@ -315,3 +315,7 @@ int intel_ring_cacheline_align(struct i915_request *rq)
        GEM_BUG_ON(rq->ring->emit & (CACHELINE_BYTES - 1));
        return 0;
 }
+
+#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
+#include "selftest_ring.c"
+#endif
index 8831ffee2061e9efc56ccd8491a9d81d8edd4b93..63f87d8608c3f860c995b610c2ed5dace1de688c 100644 (file)
@@ -18,6 +18,20 @@ struct live_mocs {
        void *vaddr;
 };
 
+static struct intel_context *mocs_context_create(struct intel_engine_cs *engine)
+{
+       struct intel_context *ce;
+
+       ce = intel_context_create(engine);
+       if (IS_ERR(ce))
+               return ce;
+
+       /* We build large requests to read the registers from the ring */
+       ce->ring = __intel_context_ring_size(SZ_16K);
+
+       return ce;
+}
+
 static int request_add_sync(struct i915_request *rq, int err)
 {
        i915_request_get(rq);
@@ -301,7 +315,7 @@ static int live_mocs_clean(void *arg)
        for_each_engine(engine, gt, id) {
                struct intel_context *ce;
 
-               ce = intel_context_create(engine);
+               ce = mocs_context_create(engine);
                if (IS_ERR(ce)) {
                        err = PTR_ERR(ce);
                        break;
@@ -395,7 +409,7 @@ static int live_mocs_reset(void *arg)
        for_each_engine(engine, gt, id) {
                struct intel_context *ce;
 
-               ce = intel_context_create(engine);
+               ce = mocs_context_create(engine);
                if (IS_ERR(ce)) {
                        err = PTR_ERR(ce);
                        break;
diff --git a/drivers/gpu/drm/i915/gt/selftest_ring.c b/drivers/gpu/drm/i915/gt/selftest_ring.c
new file mode 100644 (file)
index 0000000..2a8c534
--- /dev/null
@@ -0,0 +1,110 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright © 2020 Intel Corporation
+ */
+
+static struct intel_ring *mock_ring(unsigned long sz)
+{
+       struct intel_ring *ring;
+
+       ring = kzalloc(sizeof(*ring) + sz, GFP_KERNEL);
+       if (!ring)
+               return NULL;
+
+       kref_init(&ring->ref);
+       ring->size = sz;
+       ring->wrap = BITS_PER_TYPE(ring->size) - ilog2(sz);
+       ring->effective_size = sz;
+       ring->vaddr = (void *)(ring + 1);
+       atomic_set(&ring->pin_count, 1);
+
+       intel_ring_update_space(ring);
+
+       return ring;
+}
+
+static void mock_ring_free(struct intel_ring *ring)
+{
+       kfree(ring);
+}
+
+static int check_ring_direction(struct intel_ring *ring,
+                               u32 next, u32 prev,
+                               int expected)
+{
+       int result;
+
+       result = intel_ring_direction(ring, next, prev);
+       if (result < 0)
+               result = -1;
+       else if (result > 0)
+               result = 1;
+
+       if (result != expected) {
+               pr_err("intel_ring_direction(%u, %u):%d != %d\n",
+                      next, prev, result, expected);
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static int check_ring_step(struct intel_ring *ring, u32 x, u32 step)
+{
+       u32 prev = x, next = intel_ring_wrap(ring, x + step);
+       int err = 0;
+
+       err |= check_ring_direction(ring, next, next,  0);
+       err |= check_ring_direction(ring, prev, prev,  0);
+       err |= check_ring_direction(ring, next, prev,  1);
+       err |= check_ring_direction(ring, prev, next, -1);
+
+       return err;
+}
+
+static int check_ring_offset(struct intel_ring *ring, u32 x, u32 step)
+{
+       int err = 0;
+
+       err |= check_ring_step(ring, x, step);
+       err |= check_ring_step(ring, intel_ring_wrap(ring, x + 1), step);
+       err |= check_ring_step(ring, intel_ring_wrap(ring, x - 1), step);
+
+       return err;
+}
+
+static int igt_ring_direction(void *dummy)
+{
+       struct intel_ring *ring;
+       unsigned int half = 2048;
+       int step, err = 0;
+
+       ring = mock_ring(2 * half);
+       if (!ring)
+               return -ENOMEM;
+
+       GEM_BUG_ON(ring->size != 2 * half);
+
+       /* Precision of wrap detection is limited to ring->size / 2 */
+       for (step = 1; step < half; step <<= 1) {
+               err |= check_ring_offset(ring, 0, step);
+               err |= check_ring_offset(ring, half, step);
+       }
+       err |= check_ring_step(ring, 0, half - 64);
+
+       /* And check unwrapped handling for good measure */
+       err |= check_ring_offset(ring, 0, 2 * half + 64);
+       err |= check_ring_offset(ring, 3 * half, 1);
+
+       mock_ring_free(ring);
+       return err;
+}
+
+int intel_ring_mock_selftests(void)
+{
+       static const struct i915_subtest tests[] = {
+               SUBTEST(igt_ring_direction),
+       };
+
+       return i915_subtests(tests, NULL);
+}
index 5b39bab4da1d0cb3246392dd64a19c9837889bfb..86baed226b53dd06485c3004a8b09f4d5036ae13 100644 (file)
@@ -20,6 +20,7 @@ selftest(fence, i915_sw_fence_mock_selftests)
 selftest(scatterlist, scatterlist_mock_selftests)
 selftest(syncmap, i915_syncmap_mock_selftests)
 selftest(uncore, intel_uncore_mock_selftests)
+selftest(ring, intel_ring_mock_selftests)
 selftest(engine, intel_engine_cs_mock_selftests)
 selftest(timelines, intel_timeline_mock_selftests)
 selftest(requests, i915_request_mock_selftests)