drm/i915/selftests: Exercise intel_timeline_read_hwsp()

intel_timeline_read_hwsp() is used to support semaphore waits between
engines, that may themselves be deferred for arbitrary periods -- that
is the read of the target request's HWSP is at an indeterminant point in
the future. To support this, we need to prevent overwriting a HWSP that
is being watched across a seqno wrap (otherwise the next request will
write its value into the old HWSP preventing the watcher from making
progress, ad infinitum.) To simulate the observer across a wrap, let's
create a request that reads from the HWSP and dispatch it at different
points around a wrap to see if the value is lost.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20201021220411.5777-2-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/gt/selftest_timeline.c

@@ -17,8 +17,9 @@
 #include "../selftests/i915_random.h"
 #include "../i915_selftest.h"
 
-#include "../selftests/igt_flush_test.h"
-#include "../selftests/mock_gem_device.h"
+#include "selftests/igt_flush_test.h"
+#include "selftests/lib_sw_fence.h"
+#include "selftests/mock_gem_device.h"
 #include "selftests/mock_timeline.h"
 
 static struct page *hwsp_page(struct intel_timeline *tl)
@@ -755,6 +756,378 @@ static int live_hwsp_wrap(void *arg)
 	return err;
 }
 
+static int emit_read_hwsp(struct i915_request *rq,
+			  u32 seqno, u32 hwsp,
+			  u32 *addr)
+{
+	const u32 gpr = i915_mmio_reg_offset(GEN8_RING_CS_GPR(rq->engine->mmio_base, 0));
+	u32 *cs;
+
+	cs = intel_ring_begin(rq, 12);
+	if (IS_ERR(cs))
+		return PTR_ERR(cs);
+
+	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
+	*cs++ = *addr;
+	*cs++ = 0;
+	*cs++ = seqno;
+	*addr += 4;
+
+	*cs++ = MI_LOAD_REGISTER_MEM_GEN8 | MI_USE_GGTT;
+	*cs++ = gpr;
+	*cs++ = hwsp;
+	*cs++ = 0;
+
+	*cs++ = MI_STORE_REGISTER_MEM_GEN8 | MI_USE_GGTT;
+	*cs++ = gpr;
+	*cs++ = *addr;
+	*cs++ = 0;
+	*addr += 4;
+
+	intel_ring_advance(rq, cs);
+
+	return 0;
+}
+
+struct hwsp_watcher {
+	struct i915_vma *vma;
+	struct i915_request *rq;
+	u32 addr;
+	u32 *map;
+};
+
+static bool cmp_lt(u32 a, u32 b)
+{
+	return a < b;
+}
+
+static bool cmp_gte(u32 a, u32 b)
+{
+	return a >= b;
+}
+
+static int setup_watcher(struct hwsp_watcher *w, struct intel_gt *gt)
+{
+	struct drm_i915_gem_object *obj;
+	struct i915_vma *vma;
+
+	obj = i915_gem_object_create_internal(gt->i915, SZ_2M);
+	if (IS_ERR(obj))
+		return PTR_ERR(obj);
+
+	w->map = i915_gem_object_pin_map(obj, I915_MAP_WB);
+	if (IS_ERR(w->map)) {
+		i915_gem_object_put(obj);
+		return PTR_ERR(w->map);
+	}
+
+	vma = i915_gem_object_ggtt_pin_ww(obj, NULL, NULL, 0, 0, 0);
+	if (IS_ERR(vma)) {
+		i915_gem_object_put(obj);
+		return PTR_ERR(vma);
+	}
+
+	w->vma = vma;
+	w->addr = i915_ggtt_offset(vma);
+	return 0;
+}
+
+static int create_watcher(struct hwsp_watcher *w,
+			  struct intel_engine_cs *engine,
+			  int ringsz)
+{
+	struct intel_context *ce;
+	struct intel_timeline *tl;
+
+	ce = intel_context_create(engine);
+	if (IS_ERR(ce))
+		return PTR_ERR(ce);
+
+	ce->ring = __intel_context_ring_size(ringsz);
+	w->rq = intel_context_create_request(ce);
+	intel_context_put(ce);
+	if (IS_ERR(w->rq))
+		return PTR_ERR(w->rq);
+
+	w->addr = i915_ggtt_offset(w->vma);
+	tl = w->rq->context->timeline;
+
+	/* some light mutex juggling required; think co-routines */
+	lockdep_unpin_lock(&tl->mutex, w->rq->cookie);
+	mutex_unlock(&tl->mutex);
+
+	return 0;
+}
+
+static int check_watcher(struct hwsp_watcher *w, const char *name,
+			 bool (*op)(u32 hwsp, u32 seqno))
+{
+	struct i915_request *rq = fetch_and_zero(&w->rq);
+	struct intel_timeline *tl = rq->context->timeline;
+	u32 offset, end;
+	int err;
+
+	GEM_BUG_ON(w->addr - i915_ggtt_offset(w->vma) > w->vma->size);
+
+	i915_request_get(rq);
+	mutex_lock(&tl->mutex);
+	rq->cookie = lockdep_pin_lock(&tl->mutex);
+	i915_request_add(rq);
+
+	if (i915_request_wait(rq, 0, HZ) < 0) {
+		err = -ETIME;
+		goto out;
+	}
+
+	err = 0;
+	offset = 0;
+	end = (w->addr - i915_ggtt_offset(w->vma)) / sizeof(*w->map);
+	while (offset < end) {
+		if (!op(w->map[offset + 1], w->map[offset])) {
+			pr_err("Watcher '%s' found HWSP value %x for seqno %x\n",
+			       name, w->map[offset + 1], w->map[offset]);
+			err = -EINVAL;
+		}
+
+		offset += 2;
+	}
+
+out:
+	i915_request_put(rq);
+	return err;
+}
+
+static void cleanup_watcher(struct hwsp_watcher *w)
+{
+	if (w->rq) {
+		struct intel_timeline *tl = w->rq->context->timeline;
+
+		mutex_lock(&tl->mutex);
+		w->rq->cookie = lockdep_pin_lock(&tl->mutex);
+
+		i915_request_add(w->rq);
+	}
+
+	i915_vma_unpin_and_release(&w->vma, I915_VMA_RELEASE_MAP);
+}
+
+static bool retire_requests(struct intel_timeline *tl)
+{
+	struct i915_request *rq, *rn;
+
+	mutex_lock(&tl->mutex);
+	list_for_each_entry_safe(rq, rn, &tl->requests, link)
+		if (!i915_request_retire(rq))
+			break;
+	mutex_unlock(&tl->mutex);
+
+	return !i915_active_fence_isset(&tl->last_request);
+}
+
+static struct i915_request *wrap_timeline(struct i915_request *rq)
+{
+	struct intel_context *ce = rq->context;
+	struct intel_timeline *tl = ce->timeline;
+	u32 seqno = rq->fence.seqno;
+
+	while (tl->seqno >= seqno) { /* Cause a wrap */
+		i915_request_put(rq);
+		rq = intel_context_create_request(ce);
+		if (IS_ERR(rq))
+			return rq;
+
+		i915_request_get(rq);
+		i915_request_add(rq);
+	}
+
+	i915_request_put(rq);
+	rq = intel_context_create_request(ce);
+	if (IS_ERR(rq))
+		return rq;
+
+	i915_request_get(rq);
+	i915_request_add(rq);
+
+	return rq;
+}
+
+static int live_hwsp_read(void *arg)
+{
+	struct intel_gt *gt = arg;
+	struct hwsp_watcher watcher[2] = {};
+	struct intel_engine_cs *engine;
+	struct intel_timeline *tl;
+	enum intel_engine_id id;
+	int err = 0;
+	int i;
+
+	/*
+	 * If we take a reference to the HWSP for reading on the GPU, that
+	 * read may be arbitrarily delayed (either by foreign fence or
+	 * priority saturation) and a wrap can happen within 30 minutes.
+	 * When the GPU read is finally submitted it should be correct,
+	 * even across multiple wraps.
+	 */
+
+	if (INTEL_GEN(gt->i915) < 8) /* CS convenience [SRM/LRM] */
+		return 0;
+
+	tl = intel_timeline_create(gt);
+	if (IS_ERR(tl))
+		return PTR_ERR(tl);
+
+	if (!tl->hwsp_cacheline)
+		goto out_free;
+
+	for (i = 0; i < ARRAY_SIZE(watcher); i++) {
+		err = setup_watcher(&watcher[i], gt);
+		if (err)
+			goto out;
+	}
+
+	for_each_engine(engine, gt, id) {
+		struct intel_context *ce;
+		unsigned long count = 0;
+		IGT_TIMEOUT(end_time);
+
+		/* Create a request we can use for remote reading of the HWSP */
+		err = create_watcher(&watcher[1], engine, SZ_512K);
+		if (err)
+			goto out;
+
+		do {
+			struct i915_sw_fence *submit;
+			struct i915_request *rq;
+			u32 hwsp;
+
+			submit = heap_fence_create(GFP_KERNEL);
+			if (!submit) {
+				err = -ENOMEM;
+				goto out;
+			}
+
+			err = create_watcher(&watcher[0], engine, SZ_4K);
+			if (err)
+				goto out;
+
+			ce = intel_context_create(engine);
+			if (IS_ERR(ce)) {
+				err = PTR_ERR(ce);
+				goto out;
+			}
+
+			/* Skip to the end, saving 30 minutes of nops */
+			tl->seqno = -10u + 2 * (count & 3);
+			WRITE_ONCE(*(u32 *)tl->hwsp_seqno, tl->seqno);
+			ce->timeline = intel_timeline_get(tl);
+
+			rq = intel_context_create_request(ce);
+			if (IS_ERR(rq)) {
+				err = PTR_ERR(rq);
+				intel_context_put(ce);
+				goto out;
+			}
+
+			err = i915_sw_fence_await_dma_fence(&rq->submit,
+							    &watcher[0].rq->fence, 0,
+							    GFP_KERNEL);
+			if (err < 0) {
+				i915_request_add(rq);
+				intel_context_put(ce);
+				goto out;
+			}
+
+			mutex_lock(&watcher[0].rq->context->timeline->mutex);
+			err = intel_timeline_read_hwsp(rq, watcher[0].rq, &hwsp);
+			if (err == 0)
+				err = emit_read_hwsp(watcher[0].rq, /* before */
+						     rq->fence.seqno, hwsp,
+						     &watcher[0].addr);
+			mutex_unlock(&watcher[0].rq->context->timeline->mutex);
+			if (err) {
+				i915_request_add(rq);
+				intel_context_put(ce);
+				goto out;
+			}
+
+			mutex_lock(&watcher[1].rq->context->timeline->mutex);
+			err = intel_timeline_read_hwsp(rq, watcher[1].rq, &hwsp);
+			if (err == 0)
+				err = emit_read_hwsp(watcher[1].rq, /* after */
+						     rq->fence.seqno, hwsp,
+						     &watcher[1].addr);
+			mutex_unlock(&watcher[1].rq->context->timeline->mutex);
+			if (err) {
+				i915_request_add(rq);
+				intel_context_put(ce);
+				goto out;
+			}
+
+			i915_request_get(rq);
+			i915_request_add(rq);
+
+			rq = wrap_timeline(rq);
+			intel_context_put(ce);
+			if (IS_ERR(rq)) {
+				err = PTR_ERR(rq);
+				goto out;
+			}
+
+			err = i915_sw_fence_await_dma_fence(&watcher[1].rq->submit,
+							    &rq->fence, 0,
+							    GFP_KERNEL);
+			if (err < 0) {
+				i915_request_put(rq);
+				goto out;
+			}
+
+			err = check_watcher(&watcher[0], "before", cmp_lt);
+			i915_sw_fence_commit(submit);
+			heap_fence_put(submit);
+			if (err) {
+				i915_request_put(rq);
+				goto out;
+			}
+			count++;
+
+			if (8 * watcher[1].rq->ring->emit >
+			    3 * watcher[1].rq->ring->size) {
+				i915_request_put(rq);
+				break;
+			}
+
+			/* Flush the timeline before manually wrapping again */
+			if (i915_request_wait(rq,
+					      I915_WAIT_INTERRUPTIBLE,
+					      HZ) < 0) {
+				err = -ETIME;
+				i915_request_put(rq);
+				goto out;
+			}
+
+			retire_requests(tl);
+			i915_request_put(rq);
+		} while (!__igt_timeout(end_time, NULL));
+		WRITE_ONCE(*(u32 *)tl->hwsp_seqno, 0xdeadbeef);
+
+		pr_info("%s: simulated %lu wraps\n", engine->name, count);
+		err = check_watcher(&watcher[1], "after", cmp_gte);
+		if (err)
+			goto out;
+	}
+
+out:
+	for (i = 0; i < ARRAY_SIZE(watcher); i++)
+		cleanup_watcher(&watcher[i]);
+
+	if (igt_flush_test(gt->i915))
+		err = -EIO;
+
+out_free:
+	intel_timeline_put(tl);
+	return err;
+}
+
 static int live_hwsp_rollover_kernel(void *arg)
 {
 	struct intel_gt *gt = arg;
@@ -998,6 +1371,7 @@ int intel_timeline_live_selftests(struct drm_i915_private *i915)
 		SUBTEST(live_hwsp_engine),
 		SUBTEST(live_hwsp_alternate),
 		SUBTEST(live_hwsp_wrap),
+		SUBTEST(live_hwsp_read),
 		SUBTEST(live_hwsp_rollover_kernel),
 		SUBTEST(live_hwsp_rollover_user),
 	};