@@ -43,6 +43,8 @@
#include <getopt.h>
#include <pthread.h>
#include <stdatomic.h>
+#include <setjmp.h>
+#include <signal.h>
#include "drm.h"
#include "drmtest.h"
@@ -481,6 +483,327 @@ static void create_ext_placement_each(int fd)
free(regions);
}
+static bool supports_needs_cpu_access(int fd)
+{
+ struct drm_i915_gem_memory_class_instance regions[] = {
+ { I915_MEMORY_CLASS_DEVICE, },
+ { I915_MEMORY_CLASS_SYSTEM, },
+ };
+ struct drm_i915_gem_create_ext_memory_regions setparam_region = {
+ .base = { .name = I915_GEM_CREATE_EXT_MEMORY_REGIONS },
+ .regions = to_user_pointer(®ions),
+ .num_regions = ARRAY_SIZE(regions),
+ };
+ uint64_t size = PAGE_SIZE;
+ uint32_t handle;
+ int ret;
+
+ ret = __gem_create_ext(fd, &size,
+ I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+ &handle, &setparam_region.base);
+ if (!ret) {
+ gem_close(fd, handle);
+ igt_assert(gem_has_lmem(fd)); /* Should be dgpu only */
+ }
+
+ return ret == 0;
+}
+
+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;
+}
+
+static int upload(int fd, uint32_t handle)
+{
+ struct drm_i915_gem_exec_object2 exec[2] = {};
+ struct drm_i915_gem_execbuffer2 execbuf = {
+ .buffers_ptr = to_user_pointer(&exec),
+ .buffer_count = 2,
+ };
+
+ /*
+ * To be reasonably sure that we are not being swindled, let's make sure
+ * to 'touch' the pages from the GPU first to ensure the object is for
+ * sure placed in one of requested regions.
+ */
+ exec[0].handle = handle;
+ exec[1].handle = batch_create_size(fd, PAGE_SIZE);
+
+ return __gem_execbuf(fd, &execbuf);
+}
+
+static int alloc_lmem(int fd, uint32_t *handle,
+ struct drm_i915_gem_memory_class_instance ci,
+ uint64_t size, bool cpu_access, bool do_upload)
+{
+ struct drm_i915_gem_memory_class_instance regions[] = {
+ ci, { I915_MEMORY_CLASS_SYSTEM, },
+ };
+ struct drm_i915_gem_create_ext_memory_regions setparam_region = {
+ .base = { .name = I915_GEM_CREATE_EXT_MEMORY_REGIONS },
+ .regions = to_user_pointer(®ions),
+ };
+ uint32_t flags;
+
+ igt_assert_eq(ci.memory_class, I915_MEMORY_CLASS_DEVICE);
+
+ flags = 0;
+ setparam_region.num_regions = 1;
+ if (cpu_access) {
+ flags = I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+ setparam_region.num_regions = 2;
+ }
+
+ *handle = gem_create_ext(fd, size, flags, &setparam_region.base);
+
+ if (do_upload)
+ return upload(fd, *handle);
+
+ return 0;
+}
+
+static void create_ext_cpu_access_sanity_check(int fd)
+{
+ struct drm_i915_gem_create_ext_memory_regions setparam_region = {
+ .base = { .name = I915_GEM_CREATE_EXT_MEMORY_REGIONS },
+ };
+ struct drm_i915_query_memory_regions *regions;
+ uint64_t size = PAGE_SIZE;
+ uint32_t handle;
+ int i;
+
+ /*
+ * The ABI is that FLAG_NEEDS_CPU_ACCESS can only be applied to LMEM +
+ * SMEM objects. Make sure the kernel follows that. Let's check if we
+ * can indeed fault the object.
+ */
+
+ /* Implicit placement; should fail */
+ igt_assert_eq(__gem_create_ext(fd, &size,
+ I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+ &handle, NULL), -EINVAL);
+
+ regions = gem_get_query_memory_regions(fd);
+ igt_assert(regions);
+ igt_assert(regions->num_regions);
+
+ for (i = 0; i < regions->num_regions; i++) {
+ struct drm_i915_gem_memory_class_instance ci_regions[2] = {
+ regions->regions[i].region,
+ { I915_MEMORY_CLASS_SYSTEM, },
+ };
+ uint32_t *ptr;
+
+ setparam_region.regions = to_user_pointer(ci_regions);
+ setparam_region.num_regions = 1;
+
+ /* Single explicit placement; should fail */
+ igt_assert_eq(__gem_create_ext(fd, &size,
+ I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+ &handle, &setparam_region.base),
+ -EINVAL);
+
+ if (ci_regions[0].memory_class == I915_MEMORY_CLASS_SYSTEM)
+ continue;
+
+ /*
+ * Now combine with system memory; should pass. We should also
+ * be able to fault it.
+ */
+ setparam_region.num_regions = 2;
+ igt_assert_eq(__gem_create_ext(fd, &size,
+ I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS,
+ &handle, &setparam_region.base),
+ 0);
+ upload(fd, handle);
+ ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+ PROT_READ | PROT_WRITE);
+ ptr[0] = 0xdeadbeaf;
+ gem_close(fd, handle);
+
+ /*
+ * It should also work just fine without the flag, where in the
+ * worst case we need to migrate it when faulting it.
+ */
+ igt_assert_eq(__gem_create_ext(fd, &size,
+ 0,
+ &handle, &setparam_region.base),
+ 0);
+ upload(fd, handle);
+ ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+ PROT_READ | PROT_WRITE);
+ ptr[0] = 0xdeadbeaf;
+ gem_close(fd, handle);
+ }
+
+ free(regions);
+}
+
+static uint64_t get_visible_size(int device, uint16_t instance)
+{
+ FILE *file;
+ char *line = NULL;
+ size_t line_size;
+ int fd, i;
+
+ fd = igt_debugfs_open(device, "i915_gem_objects", O_RDONLY);
+ file = fdopen(fd, "r");
+ igt_require(getline(&line, &line_size, file) > 0);
+
+ i = 0;
+ while (getline(&line, &line_size, file) > 0) {
+ const char needle[] = "visible_size: ";
+ const char *ptr = strstr(line, needle);
+ long int size_mb;
+ char *end;
+
+ if (!ptr)
+ continue;
+
+ if (i++ != instance)
+ continue;
+
+ size_mb = strtol(ptr + ARRAY_SIZE(needle)-1, &end, 10);
+
+ free(line);
+ fclose(file);
+ close(fd);
+
+ return size_mb << 20ULL;
+ }
+
+ igt_assert(!"reached");
+}
+
+static jmp_buf jmp;
+
+__noreturn static void sigtrap(int sig)
+{
+ siglongjmp(jmp, sig);
+}
+
+static void trap_sigbus(uint32_t *ptr)
+{
+ sighandler_t old_sigbus;
+
+ old_sigbus = signal(SIGBUS, sigtrap);
+ switch (sigsetjmp(jmp, SIGBUS)) {
+ case SIGBUS:
+ break;
+ case 0:
+ *ptr = 0xdeadbeaf;
+ default:
+ igt_assert(!"reached");
+ break;
+ }
+ signal(SIGBUS, old_sigbus);
+}
+
+static void create_ext_cpu_access_big(int fd)
+{
+ struct drm_i915_query_memory_regions *regions;
+ int i;
+
+ /*
+ * Sanity check that we can still CPU map an overly large object, even
+ * if it happens to be larger the CPU visible portion of LMEM. Also
+ * check that an overly large allocation, which can't be spilled into
+ * system memory does indeed fail.
+ */
+
+ regions = gem_get_query_memory_regions(fd);
+ igt_assert(regions);
+ igt_assert(regions->num_regions);
+
+ for (i = 0; i < regions->num_regions; i++) {
+ struct drm_i915_memory_region_info qmr = regions->regions[i];
+ struct drm_i915_gem_memory_class_instance ci = qmr.region;
+ uint64_t size, visible_size, lmem_size;
+ uint32_t handle;
+ uint32_t *ptr;
+
+ if (ci.memory_class == I915_MEMORY_CLASS_SYSTEM)
+ continue;
+
+ lmem_size = qmr.probed_size;
+ visible_size = get_visible_size(fd, ci.memory_instance);
+ if (!visible_size)
+ continue;
+
+ if (visible_size <= (0.70 * lmem_size)) {
+ /*
+ * Too big. We should still be able to allocate it just
+ * fine, but faulting should result in tears.
+ */
+ size = visible_size;
+ igt_assert_eq(alloc_lmem(fd, &handle, ci, size, false, true), 0);
+ ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+ PROT_READ | PROT_WRITE);
+ trap_sigbus(ptr);
+ gem_close(fd, handle);
+
+ /*
+ * Too big again, but this time we can spill to system
+ * memory when faulting the object.
+ */
+ size = visible_size;
+ igt_assert_eq(alloc_lmem(fd, &handle, ci, size, true, true), 0);
+ ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+ PROT_READ | PROT_WRITE);
+ ptr[0] = 0xdeadbeaf;
+ gem_close(fd, handle);
+
+ /*
+ * Let's also move the upload to after faulting the
+ * pages. The current behaviour is that the pages are
+ * only allocated in device memory when initially
+ * touched by the GPU. With this in mind we should also
+ * make sure that the pages are indeed migrated, as
+ * expected.
+ */
+ size = visible_size;
+ igt_assert_eq(alloc_lmem(fd, &handle, ci, size, false, false), 0);
+ ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+ PROT_READ | PROT_WRITE);
+ ptr[0] = 0xdeadbeaf; /* temp system memory */
+ igt_assert_eq(upload(fd, handle), 0);
+ trap_sigbus(ptr); /* non-mappable device memory */
+ gem_close(fd, handle);
+ }
+
+ /*
+ * Should fit. We likely need to migrate to the mappable portion
+ * on fault though, if this device has a small BAR, given how
+ * large the initial allocation is.
+ */
+ size = visible_size >> 1;
+ igt_assert_eq(alloc_lmem(fd, &handle, ci, size, false, true), 0);
+ ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+ PROT_READ | PROT_WRITE);
+ ptr[0] = 0xdeadbeaf;
+ gem_close(fd, handle);
+
+ /*
+ * And then with the CPU_ACCESS flag enabled; should also be no
+ * surprises here.
+ */
+ igt_assert_eq(alloc_lmem(fd, &handle, ci, size, true, true), 0);
+ ptr = gem_mmap_offset__fixed(fd, handle, 0, size,
+ PROT_READ | PROT_WRITE);
+ ptr[0] = 0xdeadbeaf;
+ gem_close(fd, handle);
+ }
+
+ free(regions);
+}
+
igt_main
{
int fd = -1;
@@ -516,4 +839,15 @@ igt_main
igt_subtest("create-ext-placement-all")
create_ext_placement_all(fd);
+ igt_describe("Verify the basic functionally and expected ABI contract around I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS");
+ igt_subtest("create-ext-cpu-access-sanity-check") {
+ igt_require(supports_needs_cpu_access(fd));
+ create_ext_cpu_access_sanity_check(fd);
+ }
+
+ igt_describe("Verify the extreme cases with very large objects and I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS");
+ igt_subtest("create-ext-cpu-access-big") {
+ igt_require(supports_needs_cpu_access(fd));
+ create_ext_cpu_access_big(fd);
+ }
}
Add some basic tests for this new flag. Signed-off-by: Matthew Auld <matthew.auld@intel.com> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> --- tests/i915/gem_create.c | 334 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 334 insertions(+)