[PATCH i-g-t 4/5] tests/intel/xe_exec_system_allocator: Add preferred_loc_smem test
Matthew Brost
matthew.brost at intel.com
Fri Aug 29 19:55:05 UTC 2025
On Thu, Aug 28, 2025 at 04:58:16PM +0000, nishit.sharma at intel.com wrote:
> From: Nishit Sharma <nishit.sharma at intel.com>
>
> Added preferred-loc-smem test which is called in combination with other
> tests as well. In this test the buffer object preferred location is
> system memory. MADVISE ioctl is called with preferred_loc attribute and
> default_system system memory as preferred location.
>
> Signed-off-by: Nishit Sharma <nishit.sharma at intel.com>
> ---
> tests/intel/xe_exec_system_allocator.c | 225 ++++++++++++++++++++-----
> 1 file changed, 187 insertions(+), 38 deletions(-)
>
> diff --git a/tests/intel/xe_exec_system_allocator.c b/tests/intel/xe_exec_system_allocator.c
> index 007d9bdc0..70ca5fc2e 100644
> --- a/tests/intel/xe_exec_system_allocator.c
> +++ b/tests/intel/xe_exec_system_allocator.c
> @@ -138,7 +138,6 @@ static void signal_pdata(struct process_data *pdata)
> #define CPU_FAULT_THREADS (0x1 << 2)
> #define CPU_FAULT_PROCESS (0x1 << 3)
> #define CPU_FAULT_SAME_PAGE (0x1 << 4)
> -
> static void process_check(void *ptr, uint64_t alloc_size, uint64_t stride,
> unsigned int flags)
> {
> @@ -406,6 +405,39 @@ static void __aligned_partial_free(struct aligned_alloc_type *aligned_alloc_typ
> aligned_alloc_type->__size - aligned_alloc_type->size - begin_size);
> }
>
> +#define MAX_N_EXEC_QUEUES 16
> +
> +#define MMAP (0x1 << 0)
> +#define NEW (0x1 << 1)
> +#define BO_UNMAP (0x1 << 2)
> +#define FREE (0x1 << 3)
> +#define BUSY (0x1 << 4)
> +#define BO_MAP (0x1 << 5)
> +#define RACE (0x1 << 6)
> +#define SKIP_MEMSET (0x1 << 7)
> +#define FAULT (0x1 << 8)
> +#define FILE_BACKED (0x1 << 9)
> +#define LOCK (0x1 << 10)
> +#define MMAP_SHARED (0x1 << 11)
> +#define HUGE_PAGE (0x1 << 12)
> +#define SHARED_ALLOC (0x1 << 13)
> +#define FORK_READ (0x1 << 14)
> +#define FORK_READ_AFTER (0x1 << 15)
> +#define MREMAP (0x1 << 16)
> +#define DONTUNMAP (0x1 << 17)
> +#define READ_ONLY_REMAP (0x1 << 18)
> +#define SYNC_EXEC (0x1 << 19)
> +#define EVERY_OTHER_CHECK (0x1 << 20)
> +#define MULTI_FAULT (0x1 << 21)
> +#define PREFETCH (0x1 << 22)
> +#define THREADS (0x1 << 23)
> +#define PROCESSES (0x1 << 24)
> +#define PREFETCH_BENCHMARK (0x1 << 25)
> +#define PREFETCH_SYS_BENCHMARK (0x1 << 26)
> +#define PREFERRED_LOC_SMEM (0x1 << 27)
> +
These flags are for sections that call test_exec() only.
> +#define N_MULTI_FAULT 4
> +
> /**
> * SUBTEST: unaligned-alloc
> * Description: allocate unaligned sizes of memory
> @@ -460,7 +492,7 @@ many_allocs(int fd, struct drm_xe_engine_class_instance *eci,
> uint32_t *bos = NULL;
> struct timespec tv = {};
> uint64_t submit, read, elapsed;
> - int i;
> + int i, err;
>
> vm = xe_vm_create(fd, DRM_XE_VM_CREATE_FLAG_LR_MODE |
> DRM_XE_VM_CREATE_FLAG_FAULT_MODE, 0);
> @@ -500,6 +532,15 @@ many_allocs(int fd, struct drm_xe_engine_class_instance *eci,
> alloc.ptr = aligned_alloc(SZ_2M, alloc_size);
> igt_assert(alloc.ptr);
> }
> +
> + if (flags & PREFERRED_LOC_SMEM) {
> + err = xe_vm_madvise(fd, vm, to_user_pointer(alloc.ptr), alloc_size, 0,
> + DRM_XE_MEM_RANGE_ATTR_PREFERRED_LOC,
> + DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM, 0);
> + if (err)
> + igt_warn("MADVISE_FAILURE err = %s, vm =%u data=%"PRIu64" alloc_size =%"PRIu64"\n",
> + strerror(errno), vm, to_user_pointer(alloc.ptr), alloc_size);
> + }
I agree with Pravalika, this is never set at the caller.
Also see above PREFERRED_LOC_SMEM flags shouldn't be used here, the
flags for many_allocs are defined here.
136 /* many_alloc flags */
137 #define MIX_BO_ALLOC (0x1 << 0)
138 #define BENCHMARK (0x1 << 1)
139 #define CPU_FAULT_THREADS (0x1 << 2)
140 #define CPU_FAULT_PROCESS (0x1 << 3)
141 #define CPU_FAULT_SAME_PAGE (0x1 << 4)
If you want call madvise in many_alloc(), define a flag there. I don't
think it actually makes sense set DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM in
many_allocs() though as these sections are focused on migration perf /
eviction.
> allocs[i] = alloc;
>
> touch_all_pages(fd, exec_queue, allocs[i].ptr, alloc_size, stride,
> @@ -662,7 +703,7 @@ partial(int fd, struct drm_xe_engine_class_instance *eci, unsigned int flags)
> size_t bo_size = SZ_2M, unmap_offset = 0;
> uint32_t vm, exec_queue;
> u64 *exec_ufence = NULL;
> - int i;
> + int i, err;
> void *old, *new = NULL;
> struct aligned_alloc_type alloc;
>
> @@ -688,6 +729,15 @@ partial(int fd, struct drm_xe_engine_class_instance *eci, unsigned int flags)
> xe_wait_ufence(fd, &data[0].vm_sync, USER_FENCE_VALUE, 0, FIVE_SEC);
> data[0].vm_sync = 0;
>
> + if (flags & PREFERRED_LOC_SMEM) {
> + err = xe_vm_madvise(fd, vm, to_user_pointer(data), bo_size, 0,
> + DRM_XE_MEM_RANGE_ATTR_PREFERRED_LOC,
> + DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM, 0);
> + if (err)
> + igt_warn("MADVISE_FAILURE err = %s, vm =%u data=%"PRIu64" alloc_size = %zu\n",
> + strerror(errno), vm, to_user_pointer(data), bo_size);
> + }
> +
Same as above, partial flags are defined here;
605 #define CPU_FAULT (0x1 << 0)
606 #define REMAP (0x1 << 1)
607 #define MIDDLE (0x1 << 2)
Also this section really is only interesting if migrations happens, so
DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM doesn't make a ton of sense here.
> exec_ufence = mmap(NULL, SZ_4K, PROT_READ |
> PROT_WRITE, MAP_SHARED |
> MAP_ANONYMOUS, -1, 0);
> @@ -747,38 +797,6 @@ partial(int fd, struct drm_xe_engine_class_instance *eci, unsigned int flags)
> xe_vm_destroy(fd, vm);
> }
>
> -#define MAX_N_EXEC_QUEUES 16
> -
> -#define MMAP (0x1 << 0)
> -#define NEW (0x1 << 1)
> -#define BO_UNMAP (0x1 << 2)
> -#define FREE (0x1 << 3)
> -#define BUSY (0x1 << 4)
> -#define BO_MAP (0x1 << 5)
> -#define RACE (0x1 << 6)
> -#define SKIP_MEMSET (0x1 << 7)
> -#define FAULT (0x1 << 8)
> -#define FILE_BACKED (0x1 << 9)
> -#define LOCK (0x1 << 10)
> -#define MMAP_SHARED (0x1 << 11)
> -#define HUGE_PAGE (0x1 << 12)
> -#define SHARED_ALLOC (0x1 << 13)
> -#define FORK_READ (0x1 << 14)
> -#define FORK_READ_AFTER (0x1 << 15)
> -#define MREMAP (0x1 << 16)
> -#define DONTUNMAP (0x1 << 17)
> -#define READ_ONLY_REMAP (0x1 << 18)
> -#define SYNC_EXEC (0x1 << 19)
> -#define EVERY_OTHER_CHECK (0x1 << 20)
> -#define MULTI_FAULT (0x1 << 21)
> -#define PREFETCH (0x1 << 22)
> -#define THREADS (0x1 << 23)
> -#define PROCESSES (0x1 << 24)
> -#define PREFETCH_BENCHMARK (0x1 << 25)
> -#define PREFETCH_SYS_BENCHMARK (0x1 << 26)
> -
> -#define N_MULTI_FAULT 4
> -
As stated above, please don't move these flags.
> /**
> * SUBTEST: once-%s
> * Description: Run %arg[1] system allocator test only once
> @@ -951,6 +969,80 @@ partial(int fd, struct drm_xe_engine_class_instance *eci, unsigned int flags)
> * @mmap-new-nomemset: mmap a new buffer for each exec, skip memset of buffers
> * @mmap-new-huge-nomemset: mmap huge page new buffer for each exec, skip memset of buffers
> * @mmap-new-race-nomemset: mmap a new buffer for each exec with race between cpu and gpu access, skip memset of buffers
> + * @free-nomemset-preferred-loc-smem: malloc and free buffer for each exec and perform preferred loc madvise operation
> + * @free-preferred-loc-smem: free buffer for each exec and perform preferred loc madvise operation
> + * @free-race-nomemset-preferred-loc-smem: free buffer for each exec with race between cpu and gpu access and perform madvise operation skipping memset
> + * @free-race-preferred-loc-smem: free buffer for each exec with race between cpu and gpu access and perform madvise operation
> + * @malloc-bo-unmap-nomemset-preferred-loc-smem: malloc single buffer for all execs, bind and unbind a BO to same address, skip memset and perform madvise operation
> + * @malloc-busy-nomemset-preferred-loc-smem: malloc single buffer for all execs, try to unbind while buffer valid, skip memset of buffers and perform madvise operation
> + * @malloc-busy-preferred-loc-smem: malloc single buffer for all execs, try to unbind while buffer valid and perform madvise operation
> + * @malloc-fork-read-after-preferred-loc-smem: malloc single buffer for all execs, fork a process to read test output, perform madvise operation
> + * @malloc-fork-read-preferred-loc-smem: malloc single buffer for all execs, fork a process to read test output, perform madvise operation
> + * @malloc-mlock-nomemset-preferred-loc-smem: malloc and mlock single buffer for all execs, skip memset of buffers, perform madvise operation
> + * @malloc-mlock-preferred-loc-smem: malloc and mlock single buffer for all execs, perform madvise operation
> + * @malloc-multi-fault-preferred-loc-smem: malloc single buffer for all execs and perform madvise operation
> + * @malloc-nomemset-preferred-loc-smem: malloc single buffer for all execs, skip memset of buffers and perform madvise operation
> + * @malloc-preferred-loc-smem: malloc single buffer for all execs, issue a command which will trigger multiple faults, perform madvise operation
> + * @malloc-prefetch-preferred-loc-smem: malloc single buffer for all execs, prefetch buffer before each exec, perform madvise operation
> + * @malloc-prefetch-race-preferred-loc-smem: malloc single buffer for all execs, prefetch buffer before each exec, perform madvise operation
> + * @malloc-race-nomemset-preferred-loc-smem: malloc single buffer for all execs with race between cpu and gpu access, perform madvise operation
> + * @malloc-race-preferred-loc-smem: malloc single buffer for all execs with race between cpu and gpu access, perform madvise operation
> + * @free-race-nomemset-preferred-loc-smem: malloc and free buffer for each exec with race between cpu and gpu access, perform madvise operation
> + * @free-race-preferred-loc-smem: malloc and free buffer for each exec with race between cpu and gpu access, perform madvise operation
> + * @malloc-bo-unmap-nomemset-preferred-loc-smem: malloc single buffer for all execs, bind and unbind a BO to same address before execs, perform madvise operation
> + * @malloc-bo-unmap-preferred-loc-smem: malloc single buffer for all execs and perform madvise operation
> + * @malloc-busy-nomemset-preferred-loc-smem: malloc single buffer for all execs and perform madvise operation
> + * @malloc-busy-preferred-loc-smem: malloc single buffer for all execs and perform madvise
> + * @mmap-file-mlock-nomemset-preferred-loc-smem: mmap and mlock single buffer, with file backing, perform madvise
> + * @mmap-file-mlock-preferred-loc-smem: mmap and mlock single buffer, with file backing, perform madvise
> + * @mmap-file-nomemset-preferred-loc-smem: mmap single buffer, with file backing and perform madvise
> + * @mmap-file-preferred-loc-smem: mmap single buffer, with file backing and perform madvise
> + * @mmap-free-huge-nomemset-preferred-loc-smem: mmap huge page and free buffer for each exec and perform madvise
> + * @mmap-free-huge-preferred-loc-smem: mmap huge page and free buffer for each exec and perform madvise
> + * @mmap-free-nomemset-preferred-loc-smem: mmap and free buffer for each exec and perform madvise
> + * @mmap-free-preferred-loc-smem: mmap and free buffer for each exec and perform madvise
> + * @mmap-free-race-nomemset-preferred-loc-smem:
> + * @mmap-free-race-preferred-loc-smem:
> + * @mmap-huge-nomemset-preferred-loc-smem:
> + * @mmap-huge-preferred-loc-smem:
> + * @mmap-mlock-nomemset-preferred-loc-smem:
> + * @mmap-mlock-preferred-loc-smem:
> + * @mmap-new-huge-nomemset-preferred-loc-smem:
> + * @mmap-new-huge-preferred-loc-smem:
> + * @mmap-new-nomemset-preferred-loc-smem:
> + * @mmap-new-preferred-loc-smem:
> + * @mmap-new-race-nomemset-preferred-loc-smem:
> + * @mmap-new-race-preferred-loc-smem:
> + * @mmap-nomemset-preferred-loc-smem:
> + * @mmap-preferred-loc-smem:
> + * @mmap-prefetch-preferred-loc-smem:
> + * @mmap-prefetch-shared-preferred-loc-smem:
> + * @mmap-race-nomemset-preferred-loc-smem:
> + * @mmap-race-preferred-loc-smem:
> + * @mmap-remap-dontunmap-eocheck-preferred-loc-smem:
> + * @mmap-remap-dontunmap-preferred-loc-smem:
> + * @mmap-remap-eocheck-preferred-loc-smem:
> + * @mmap-remap-preferred-loc-smem:
> + * @mmap-remap-ro-dontunmap-eocheck-preferred-loc-smem:
> + * @mmap-remap-ro-dontunmap-preferred-loc-smem:
> + * @mmap-remap-ro-eocheck-preferred-loc-smem:
> + * @mmap-remap-ro-preferred-loc-smem:
> + * @mmap-shared-nomemset-preferred-loc-smem:
> + * @mmap-shared-preferred-loc-smem:
> + * @mmap-shared-remap-dontunmap-eocheck-preferred-loc-smem:
> + * @mmap-shared-remap-dontunmap-preferred-loc-smem:
> + * @mmap-shared-remap-eocheck-preferred-loc-smem:
> + * @mmap-shared-remap-preferred-loc-smem:
> + * @new-bo-map-nomemset-preferred-loc-smem:
> + * @new-bo-map-preferred-loc-smem:
> + * @new-busy-nomemset-preferred-loc-smem:
> + * @new-busy-preferred-loc-smem:
> + * @new-nomemset-preferred-loc-smem:
> + * @new-preferred-loc-smem:
> + * @new-prefetch-preferred-loc-smem:
> + * @new-race-nomemset-preferred-loc-smem:
> + * @new-race-preferred-loc-smem:
> + * @prefetch-benchmark:
> *
> * SUBTEST: prefetch-benchmark
> * Description: Prefetch a 64M buffer 128 times, measure bandwidth of prefetch
> @@ -1020,7 +1112,7 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
> uint32_t bo = 0, bind_sync = 0;
> void **pending_free;
> u64 *exec_ufence = NULL, *bind_ufence = NULL;
> - int i, j, b, file_fd = -1, prev_idx, pf_count;
> + int i, j, b, file_fd = -1, prev_idx, pf_count, err;
> bool free_vm = false;
> size_t aligned_size = bo_size ?: xe_get_default_alignment(fd);
> size_t orig_size = bo_size;
> @@ -1133,6 +1225,15 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
>
> addr = to_user_pointer(data);
>
> + if (flags & PREFERRED_LOC_SMEM) {
> + err = xe_vm_madvise(fd, vm, to_user_pointer(data), bo_size, 0,
> + DRM_XE_MEM_RANGE_ATTR_PREFERRED_LOC,
> + DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM, 0);
> + if (err)
> + igt_warn("MADVISE_FAILURE err = %s, vm =%u data=%"PRIu64" alloc_size = %zu\n",
> + strerror(errno), vm, to_user_pointer(data), bo_size);
> + }
What I would here, rather than have PREFERRED_LOC_SMEM, maybe flag like
MADVISE_SWIZZLE. On each pass of the main loop, switch between
DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM and
DRM_XE_PREFERRED_LOC_DEFAULT_DEVICE. I think that should create pretty
good coverage.
> +
> if (flags & BO_UNMAP) {
> bo_flags = DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM;
> bo = xe_bo_create(fd, vm, bo_size,
> @@ -1202,7 +1303,7 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
> uint64_t batch_addr = addr + batch_offset;
> uint64_t sdi_offset = (char *)&data[idx].data - (char *)data;
> uint64_t sdi_addr = addr + sdi_offset;
> - int e = i % n_exec_queues, err;
> + int e = i % n_exec_queues;
> bool fault_inject = (FAULT & flags) && i == n_execs / 2;
> bool fault_injected = (FAULT & flags) && i > n_execs;
>
> @@ -1232,6 +1333,16 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
> aligned_alloc_type = __aligned_alloc(aligned_size, bo_size);
> next_data = aligned_alloc_type.ptr;
> igt_assert(next_data);
> +
> + if (flags & PREFERRED_LOC_SMEM) {
> + err = xe_vm_madvise(fd, vm, to_user_pointer(next_data), bo_size, 0,
> + DRM_XE_MEM_RANGE_ATTR_PREFERRED_LOC,
> + DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM, 0);
> + if (err)
> + igt_warn("MADVISE_FAILURE err = %s, vm =%u data=%"PRIu64" alloc_size = %zu\n",
> + strerror(errno), vm, to_user_pointer(next_data), bo_size);
> + }
> +
> __aligned_partial_free(&aligned_alloc_type);
>
> b = 0;
> @@ -1253,6 +1364,7 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
> sync[0].addr = to_user_pointer(bind_ufence);
>
> start = igt_nsec_elapsed(&tv);
> +
> xe_vm_prefetch_async(fd, vm, 0, 0, addr, bo_size, sync,
> 1, region);
> end = igt_nsec_elapsed(&tv);
> @@ -1355,6 +1467,7 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
> } else {
> igt_assert_eq(data[idx].data,
> READ_VALUE(&data[idx]));
> +
> if (flags & PREFETCH_SYS_BENCHMARK) {
> struct timespec tv = {};
> u64 start, end;
> @@ -1429,6 +1542,7 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
> aligned_alloc_type = __aligned_alloc(aligned_size, bo_size);
> data = aligned_alloc_type.ptr;
> igt_assert(data);
> +
> __aligned_partial_free(&aligned_alloc_type);
>
> bo_flags = DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM;
> @@ -1450,6 +1564,7 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
> }
> bo = 0;
> data = aligned_alloc(aligned_size, bo_size);
> +
> igt_assert(data);
> }
> addr = to_user_pointer(data);
> @@ -1460,6 +1575,15 @@ test_exec(int fd, struct drm_xe_engine_class_instance *eci,
> prev_idx = idx;
> }
>
> + if (flags & PREFERRED_LOC_SMEM) {
> + err = xe_vm_madvise(fd, vm, to_user_pointer(data), bo_size, 0,
> + DRM_XE_MEM_RANGE_ATTR_PREFERRED_LOC,
> + DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM, 0);
> + if (err)
> + igt_warn("MADVISE_FAILURE err = %s, vm =%u data=%"PRIu64" alloc_size = %zu\n",
> + strerror(errno), vm, to_user_pointer(data), bo_size);
> + }
> +
> if (flags & PREFETCH_BENCHMARK) {
> igt_info("Prefetch VRAM execution took %.3fms, %.1f5 GB/s\n",
> 1e-6 * prefetch_ns,
> @@ -1587,6 +1711,7 @@ threads(int fd, int n_exec_queues, int n_execs, size_t bo_size,
> uint32_t vm = 0;
> bool go = false;
> void *alloc = NULL;
> + int err;
>
> if ((FILE_BACKED | FORK_READ) & flags)
> return;
> @@ -1614,6 +1739,15 @@ threads(int fd, int n_exec_queues, int n_execs, size_t bo_size,
> alloc = aligned_alloc(SZ_2M, alloc_size);
> igt_assert(alloc);
>
> + if (flags & PREFERRED_LOC_SMEM) {
> + err = xe_vm_madvise(fd, vm, to_user_pointer(alloc), alloc_size, 0,
> + DRM_XE_MEM_RANGE_ATTR_PREFERRED_LOC,
> + DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM, 0);
> + if (err)
> + igt_warn("MADVISE_FAILURE err = %s, vm =%u data=%"PRIu64" alloc_size = %zu\n",
> + strerror(errno), vm, to_user_pointer(alloc), alloc_size);
> + }
> +
> memset(alloc, 5, alloc_size);
> flags &= ~SHARED_ALLOC;
> }
> @@ -1831,6 +1965,7 @@ igt_main
> { NULL },
> };
> int fd;
> + int num_sections;
>
> igt_fixture {
> struct xe_device *xe;
> @@ -1843,7 +1978,21 @@ igt_main
> open_sync_file();
> }
>
> - for (const struct section *s = sections; s->name; s++) {
> +
> + num_sections = 0;
> + for (const struct section *s = sections; s[num_sections].name; num_sections++)
> + ;
> +
> + for (int i = 0; i < num_sections * 2; i++) {
> + struct section *s = §ions[i % num_sections];
> +
> + if (i/num_sections == 0) {
You are adding too many sections. I'd pick a interesting combination of
flags and add table entries for them.
I'd suggest:
malloc-madvise
malloc-prefetch-madvise
mmap-remap-madvise
new-madvise
free-madvise
mmap-free-madvise
If isn't clear copy the existing flags in those sections and add the
suggested MADVISE_SWIZZLE.
Matt
> + static char modified_name[256];
> + snprintf(modified_name, sizeof(modified_name), "%s-preferred-loc-smem", s->name);
> + s->name = modified_name;
> + s->flags |= PREFERRED_LOC_SMEM;
> + }
> +
> igt_subtest_f("once-%s", s->name)
> xe_for_each_engine(fd, hwe)
> test_exec(fd, hwe, 1, 1, 0, 0, 0, NULL,
> --
> 2.43.0
>
More information about the igt-dev
mailing list