@@ -352,6 +352,8 @@ static int max_pte_pkt_size(struct i915_request *rq, int pkt)
return pkt;
}
+#define I915_EMIT_PTE_NUM_DWORDS 6
+
static int emit_pte(struct i915_request *rq,
struct sgt_dma *it,
enum i915_cache_level cache_level,
@@ -393,7 +395,7 @@ static int emit_pte(struct i915_request *rq,
offset += (u64)rq->engine->instance << 32;
- cs = intel_ring_begin(rq, 6);
+ cs = intel_ring_begin(rq, I915_EMIT_PTE_NUM_DWORDS);
if (IS_ERR(cs))
return PTR_ERR(cs);
@@ -416,7 +418,7 @@ static int emit_pte(struct i915_request *rq,
intel_ring_advance(rq, cs);
intel_ring_update_space(ring);
- cs = intel_ring_begin(rq, 6);
+ cs = intel_ring_begin(rq, I915_EMIT_PTE_NUM_DWORDS);
if (IS_ERR(cs))
return PTR_ERR(cs);
@@ -8,6 +8,7 @@
#include "gem/i915_gem_internal.h"
#include "gem/i915_gem_lmem.h"
+#include "selftests/igt_spinner.h"
#include "selftests/i915_random.h"
static const unsigned int sizes[] = {
@@ -529,6 +530,149 @@ static int live_migrate_clear(void *arg)
return 0;
}
+struct spinner_timer {
+ struct timer_list timer;
+ struct igt_spinner spin;
+};
+
+static void spinner_kill(struct timer_list *timer)
+{
+ struct spinner_timer *st = from_timer(st, timer, timer);
+
+ igt_spinner_end(&st->spin);
+ pr_info("%s\n", __func__);
+}
+
+static int live_emit_pte_full_ring(void *arg)
+{
+ struct intel_gt *gt = arg;
+ struct intel_migrate *migrate = >->migrate;
+ struct drm_i915_private *i915 = migrate->context->engine->i915;
+ struct drm_i915_gem_object *obj;
+ struct intel_context *ce;
+ struct i915_request *rq, *prev;
+ struct spinner_timer st;
+ struct sgt_dma it;
+ int len, sz, err;
+ u32 *cs;
+
+ /*
+ * Simple regression test to check that we don't trample the
+ * rq->reserved_space when returning from emit_pte(), if the ring is
+ * nearly full.
+ */
+
+ if (igt_spinner_init(&st.spin, to_gt(i915)))
+ return -ENOMEM;
+
+ obj = i915_gem_object_create_internal(i915, 2 * PAGE_SIZE);
+ if (IS_ERR(obj)) {
+ err = PTR_ERR(obj);
+ goto out_spinner;
+ }
+
+ err = i915_gem_object_pin_pages_unlocked(obj);
+ if (err)
+ goto out_obj;
+
+ ce = intel_migrate_create_context(migrate);
+ if (IS_ERR(ce)) {
+ err = PTR_ERR(ce);
+ goto out_obj;
+ }
+
+ ce->ring_size = SZ_4K; /* Not too big */
+
+ err = intel_context_pin(ce);
+ if (err)
+ goto out_put;
+
+ rq = igt_spinner_create_request(&st.spin, ce, MI_ARB_CHECK);
+ if (IS_ERR(rq)) {
+ err = PTR_ERR(rq);
+ goto out_unpin;
+ }
+
+ i915_request_add(rq);
+ if (!igt_wait_for_spinner(&st.spin, rq)) {
+ err = -EIO;
+ goto out_unpin;
+ }
+
+ /*
+ * Fill the rest of the ring leaving I915_EMIT_PTE_NUM_DWORDS +
+ * ring->reserved_space at the end. To actually emit the PTEs we require
+ * slightly more than I915_EMIT_PTE_NUM_DWORDS, since our object size is
+ * greater than PAGE_SIZE. The correct behaviour is to wait for more
+ * ring space in emit_pte(), otherwise we trample on the reserved_space
+ * resulting in crashes when later submitting the rq.
+ */
+
+ prev = NULL;
+ do {
+ if (prev)
+ i915_request_add(rq);
+
+ rq = i915_request_create(ce);
+ if (IS_ERR(rq)) {
+ err = PTR_ERR(rq);
+ goto out_unpin;
+ }
+
+ sz = (rq->ring->space - rq->reserved_space) / sizeof(u32) -
+ I915_EMIT_PTE_NUM_DWORDS;
+ sz = min_t(u32, sz, (SZ_1K - rq->reserved_space) / sizeof(u32) -
+ I915_EMIT_PTE_NUM_DWORDS);
+ cs = intel_ring_begin(rq, sz);
+ if (IS_ERR(cs)) {
+ err = PTR_ERR(cs);
+ goto out_rq;
+ }
+
+ memset32(cs, MI_NOOP, sz);
+ cs += sz;
+ intel_ring_advance(rq, cs);
+
+ pr_info("%s emit=%u sz=%d\n", __func__, rq->ring->emit, sz);
+
+ prev = rq;
+ } while (rq->ring->space > (rq->reserved_space +
+ I915_EMIT_PTE_NUM_DWORDS * sizeof(u32)));
+
+ timer_setup_on_stack(&st.timer, spinner_kill, 0);
+ mod_timer(&st.timer, jiffies + 2 * HZ);
+
+ /*
+ * This should wait for the spinner to be killed, otherwise we should go
+ * down in flames when doing i915_request_add().
+ */
+ pr_info("%s emite_pte ring space=%u\n", __func__, rq->ring->space);
+ it = sg_sgt(obj->mm.pages->sgl);
+ len = emit_pte(rq, &it, obj->cache_level, false, 0, CHUNK_SZ);
+ if (!len) {
+ err = -EINVAL;
+ goto out_rq;
+ }
+ if (len < 0) {
+ err = len;
+ goto out_rq;
+ }
+
+out_rq:
+ i915_request_add(rq); /* GEM_BUG_ON(rq->reserved_space > ring->space)? */
+ del_timer_sync(&st.timer);
+ destroy_timer_on_stack(&st.timer);
+out_unpin:
+ intel_context_unpin(ce);
+out_put:
+ intel_context_put(ce);
+out_obj:
+ i915_gem_object_put(obj);
+out_spinner:
+ igt_spinner_fini(&st.spin);
+ return err;
+}
+
struct threaded_migrate {
struct intel_migrate *migrate;
struct task_struct *tsk;
@@ -651,6 +795,7 @@ int intel_migrate_live_selftests(struct drm_i915_private *i915)
static const struct i915_subtest tests[] = {
SUBTEST(live_migrate_copy),
SUBTEST(live_migrate_clear),
+ SUBTEST(live_emit_pte_full_ring),
SUBTEST(thread_migrate_copy),
SUBTEST(thread_migrate_clear),
SUBTEST(thread_global_copy),
Simple regression test to check that we don't trample the rq->reserved_space when returning from emit_pte(), if the ring is nearly full. v2: Make spinner_kill() static v3: Reduce the ring size further, which should mean we need to execute less noops; hopefully this appeases bsw. Also add some debug logging. v4: Fix the min request construction to account for reserved_space + I915_EMIT_PTE_NUM_DWORDS v5: Use a simple on-stack timer to kill the spinner instead of kthread (Chris) References: https://gitlab.freedesktop.org/drm/intel/-/issues/7535 References: https://gitlab.freedesktop.org/drm/intel/-/issues/6889 Signed-off-by: Matthew Auld <matthew.auld@intel.com> Cc: Chris Wilson <chris.p.wilson@intel.com> Cc: Andi Shyti <andi.shyti@linux.intel.com> Cc: Andrzej Hajda <andrzej.hajda@intel.com> Cc: Nirmoy Das <nirmoy.das@intel.com> --- drivers/gpu/drm/i915/gt/intel_migrate.c | 6 +- drivers/gpu/drm/i915/gt/selftest_migrate.c | 145 +++++++++++++++++++++ 2 files changed, 149 insertions(+), 2 deletions(-)