On 12/20/2022 12:46 PM, Matthew Auld wrote:
We should still be able to GTT evict objects during execbuf (old
bindings can linger around), even if there is object lock contention. In
the worst case the execbuf should just wait on the contented locks.
Returning -ENOSPC smells like a regression from past behaviour, and
seems to break userspace.
v2:
- Add coverage for explicit softpin
- Add timeout for the spinner
v3:
- Improve the test description
References: https://gitlab.freedesktop.org/drm/intel/-/issues/7570
Signed-off-by: Matthew Auld <matthew.auld@xxxxxxxxx>
Cc: Andrzej Hajda <andrzej.hajda@xxxxxxxxx>
Cc: Nirmoy Das <nirmoy.das@xxxxxxxxx>
Cc: Mani Milani <mani@xxxxxxxxxxxx>
---
tests/i915/gem_ppgtt.c | 133 +++++++++++++++++++++++++++++++++++++++++
1 file changed, 133 insertions(+)
diff --git a/tests/i915/gem_ppgtt.c b/tests/i915/gem_ppgtt.c
index 9673ce22..024e8d47 100644
--- a/tests/i915/gem_ppgtt.c
+++ b/tests/i915/gem_ppgtt.c
@@ -255,6 +255,131 @@ static void flink_and_close(void)
close(fd2);
}
+#define PAGE_SIZE 4096
+
+static uint32_t batch_create_size(int fd, uint64_t size)
+{
+ const uint32_t bbe = MI_BATCH_BUFFER_END;
+ uint32_t handle;
+
+ handle = gem_create(fd, size);
+ gem_write(fd, handle, 0, &bbe, sizeof(bbe));
+
+ return handle;
+}
+
+#define IGT_USE_ANY 0x1
+#define IGT_USE_PINNED 0x2
+static void upload(int fd, uint32_t handle, uint32_t in_fence, uint32_t ctx_id,
+ unsigned int flags)
+{
+ struct drm_i915_gem_exec_object2 exec[2] = {};
+ struct drm_i915_gem_execbuffer2 execbuf = {
+ .buffers_ptr = to_user_pointer(&exec),
+ .buffer_count = 1,
+ .rsvd1 = ctx_id,
+ };
+
+ if (in_fence) {
+ execbuf.rsvd2 = in_fence;
+ execbuf.flags = I915_EXEC_FENCE_IN;
+ }
+
+ exec[0].handle = handle;
+ exec[0].flags = EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+
+ if (flags & IGT_USE_PINNED)
+ exec[0].flags |= EXEC_OBJECT_PINNED; /* offset = 0 */
+
+ if (flags & IGT_USE_ANY) {
+ exec[0].flags |= EXEC_OBJECT_PAD_TO_SIZE;
+ exec[0].pad_to_size = gem_aperture_size(fd);
+ }
+
+ gem_execbuf(fd, &execbuf);
+}
+
+static void shrink_vs_evict(unsigned int flags)
+{
+ const unsigned int nproc = sysconf(_SC_NPROCESSORS_ONLN) + 1;
+ const uint64_t timeout_5s = 5000000000LL;
5*NSEC_PER_SEC would be nice.
+ int fd = drm_open_driver(DRIVER_INTEL);
+ uint64_t ahnd = get_reloc_ahnd(fd, 0);
+ const intel_ctx_t *ctx_arr[nproc];
+ igt_spin_t *spinner;
+ uint32_t handle1;
+ int i;
+
+ /*
+ * Try to simulate some nasty object lock contention during GTT
+ * eviction. Create a BO and bind across several different VMs. Invoke
+ * the shrinker on that shared BO, followed by triggering GTT eviction
+ * across all VMs. Both require the object lock to make forward
+ * progress when trying to unbind the BO, but the shrinker will be
+ * blocked by the spinner (until killed). Once the spinner is killed
+ * the shrinker should be able to unbind the object and drop the object
+ * lock, and GTT eviction should eventually succeed. At no point should
+ * we see -ENOSPC from the execbuf, even if we can't currently grab the
+ * object lock.
+ */
+
+ igt_require(gem_uses_full_ppgtt(fd));
+
+ igt_drop_caches_set(fd, DROP_ALL);
+
+ handle1 = gem_create(fd, PAGE_SIZE);
+
+ spinner = igt_spin_new(fd,
+ .ahnd = ahnd,
+ .flags = IGT_SPIN_FENCE_OUT);
+ igt_spin_set_timeout(spinner, timeout_5s);
+
+ /*
+ * Create several VMs to ensure we don't block on the same vm lock. The
+ * goal of the test is to ensure that object lock contention doesn't
+ * somehow result in -ENOSPC from execbuf, if we need to trigger GTT
+ * eviction.
+ */
+ for (i = 0; i < nproc; i++) {
+ ctx_arr[i] = intel_ctx_create(fd, NULL);
+
+ upload(fd, handle1, spinner->execbuf.rsvd2 >> 32,
+ ctx_arr[i]->id, flags);
+ }
+
+ igt_fork(child, 1)
+ igt_drop_caches_set(fd, DROP_ALL);
+
+ sleep(2); /* Give the shrinker time to find handle1 */
+
+ igt_fork(child, nproc) {
+ uint32_t handle2 = gem_create(fd, PAGE_SIZE);
+
+ /*
+ * One of these forks will be stuck on the vm mutex, since the
+ * shrinker is holding it (along with the object lock) while
+ * trying to unbind the chosen vma, but is blocked by the
+ * spinner. The rest should only block waiting to grab the
+ * object lock for handle1, before then trying to GTT evict it
+ * from their respective vm. In either case the contention of
+ * the vm->mutex or object lock should never result in -ENOSPC
+ * or some other error.
+ */
+ handle2 = batch_create_size(fd, PAGE_SIZE);
This can be
uint32_t handle2 = batch_create_size(fd, PAGE_SIZE);
Above gem_create seems unnecessary. With those
Reviewed-by: Nirmoy Das <nirmoy.das@xxxxxxxxx>
+
+ upload(fd, handle2, 0, ctx_arr[child]->id, flags);
+ gem_close(fd, handle2);
+ }
+
+ igt_waitchildren();
+ igt_spin_free(fd, spinner);
+
+ for (i = 0; i < nproc; i++)
+ intel_ctx_destroy(fd, ctx_arr[i]);
+
+ gem_close(fd, handle1);
+}
+
static bool has_contexts(void)
{
bool result;
@@ -331,4 +456,12 @@ igt_main
igt_subtest("flink-and-close-vma-leak")
flink_and_close();
+
+ igt_describe("Regression test to verify GTT eviction can't randomly fail due to object lock contention");
+ igt_subtest_group {
+ igt_subtest("shrink-vs-evict-any")
+ shrink_vs_evict(IGT_USE_ANY);
+ igt_subtest("shrink-vs-evict-pinned")
+ shrink_vs_evict(IGT_USE_PINNED);
+ }
}
