drm/i915/selftests: Exercise intel_timeline_read_hwsp()
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 21 Oct 2020 22:04:11 +0000 (23:04 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 23 Oct 2020 12:35:39 +0000 (13:35 +0100)
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

index 96d164a..d5ea6a9 100644 (file)
@@ -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 @@ out_free:
        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),
        };