From: "Sharma, Nishit" <nishit.sharma@intel.com>
To: Sobin Thomas <sobin.thomas@intel.com>,
<igt-dev@lists.freedesktop.org>,
"Hellstrom, Thomas" <thomas.hellstrom@intel.com>
Subject: Re: [PATCH i-g-t 1/1] tests/intel/xe_evict: overcommit tests for fault-mode and non-fault-mode VMs
Date: Mon, 19 Jan 2026 10:36:35 +0530 [thread overview]
Message-ID: <d4b4a604-02d6-42e6-910a-2b7c7792fec2@intel.com> (raw)
In-Reply-To: <20260113070806.189718-2-sobin.thomas@intel.com>
[-- Attachment #1: Type: text/plain, Size: 14720 bytes --]
On 1/13/2026 12:38 PM, Sobin Thomas wrote:
> The existing tests in xe_evict focuses on system-wide memory allocation
> across multiple processes. However, OOM error handling in different VM
> modes was not being tested, and the previous test_svm_overcommit() had
> a critical bug that prevented proper overcommit scenarios.
>
> Add three new tests to verify graceful OOM failure handling:
>
> - test_evict_oom(): Allocates BOs aggressively in a loop until
> OOM occurs. Tests error handling in LR mode and expects
> -ENOSPC or -ENOMEM.
>
> - test_vm_nonfault_mode_overcommit(): Verifies that non-fault mode VMs
> properly reject overcommit attempts with -ENOSPC or -ENOMEM as
> expected.
>
> - test_vm_fault_mode_overcommit(): Validates that fault-mode VMs can
> handle memory pressure gracefully by touching pages to trigger page
> faults.
>
> Signed-off-by: Sobin Thomas<sobin.thomas@intel.com>
> ---
> tests/intel/xe_evict.c | 330 +++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 330 insertions(+)
>
> diff --git a/tests/intel/xe_evict.c b/tests/intel/xe_evict.c
> index 2350487c8..0f739ec7b 100644
> --- a/tests/intel/xe_evict.c
> +++ b/tests/intel/xe_evict.c
> @@ -210,6 +210,274 @@ test_evict(int fd, struct drm_xe_engine_class_instance *eci,
> drm_close_driver(fd);
> }
>
> +static int
> +test_evict_oom(int fd, struct drm_xe_engine_class_instance *eci,
> + int n_exec_queues, int n_execs, size_t bo_size, unsigned long flags)
> +{
> + uint32_t vm;
> + uint32_t bind_exec_queues[1] = { 0 };
> + uint64_t addr = 0x100000000;
> + uint64_t available_ram;
> + uint64_t total_alloc_size;
> + int bind_err = 0;
> + uint32_t *bo;
> + int i;
> + const size_t min_map_size = 4096; /* Add constant with proper type */
> +
> +#define USER_FENCE_VALUE 0xdeadbeefdeadbeefull
> + struct drm_xe_sync sync[1] = {
> + { .type = DRM_XE_SYNC_TYPE_USER_FENCE, .flags = DRM_XE_SYNC_FLAG_SIGNAL,
> + .timeline_value = USER_FENCE_VALUE },
> + };
> +
> + /* Check if we have enough system memory to proceed */
> + available_ram = igt_get_avail_ram_mb() << 20;
> + /* Calculate total allocation size - now for full n_execs */
> + total_alloc_size = (uint64_t)n_execs * bo_size;
> +
> + if (available_ram < (total_alloc_size / 4)) {
> + igt_info("Insufficient memory to run OOM test safely\n");
> + return -ENOMEM;
> + }
> + if (total_alloc_size > available_ram) {
> + int safe_n_execs = available_ram / bo_size;
> +
> + safe_n_execs = ALIGN_DOWN(safe_n_execs, 4);
> + if (safe_n_execs < 4) {
> + igt_info("Not enough memory to run OOM test\n");
> + return -ENOMEM;
> + }
> + igt_info("Reducing n_execs from %d to %d to fit in available memory\n",
> + n_execs, safe_n_execs);
> + n_execs = safe_n_execs;
> + total_alloc_size = (uint64_t)n_execs * bo_size;
> + }
> + igt_info("OOM test: n_execs=%d, bo_size=%llu MB, total_alloc=%llu MB, available=%llu MB\n",
> + n_execs, (unsigned long long)(bo_size >> 20),
> + (unsigned long long)(total_alloc_size >> 20),
> + (unsigned long long)(available_ram >> 20));
> +
> + /* Allocate array for n_execs BOs */
> + bo = calloc(n_execs, sizeof(*bo));
> + igt_assert(bo);
> +
> + fd = drm_reopen_driver(fd);
> +
> + vm = xe_vm_create(fd, DRM_XE_VM_CREATE_FLAG_LR_MODE, 0);
> + bind_exec_queues[0] = xe_bind_exec_queue_create(fd, vm, 0);
> +
> + igt_info("\n n_execs = %d, bo_size = %zu\n", n_execs, bo_size);
> +
> + /* Try to allocate and bind more than available memory */
> + for (i = 0; i < n_execs; i++) {
> + uint32_t __bo;
> + struct {
> + uint64_t vm_sync;
> + } *data;
> + int create_ret;
> + size_t map_size; /* Use size_t for map size */
> +
> + /* Use __xe_bo_create to handle allocation failures gracefully */
> + create_ret = __xe_bo_create(fd, 0, bo_size, vram_memory(fd, eci->gt_id),
> + DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM,
> + NULL, &__bo);
> + if (create_ret) {
> + bind_err = -errno;
> + igt_info("BO create failed at iteration %d with error %d (%s)\n",
> + i, -bind_err, strerror(-bind_err));
> + break;
> + }
> +
> + bo[i] = __bo;
> +
> + /* Calculate map size with same types */
> + map_size = sizeof(*data) > min_map_size ? sizeof(*data) : min_map_size;
> + data = xe_bo_map(fd, __bo, map_size);
> + sync[0].addr = to_user_pointer(&data->vm_sync);
> + bind_err = __xe_vm_bind(fd, vm, bind_exec_queues[0], __bo, 0, addr,
> + bo_size, DRM_XE_VM_BIND_OP_MAP, 0, sync,
> + 1, 0, 0, 0);
> + /* Attempt bind - should eventually fail with -ENOSPC or -ENOMEM */
> + if (bind_err) {
> + bind_err = -errno;
> + igt_info("Bind failed at iteration %d with error %d (%s)\n",
> + i, -bind_err, strerror(-bind_err));
> + munmap(data, map_size);
> + gem_close(fd, __bo);
> + bo[i] = 0;
> + break;
> + }
> +
> + xe_wait_ufence(fd, &data->vm_sync, USER_FENCE_VALUE,
> + bind_exec_queues[0], 20 * NSEC_PER_SEC);
> +
> + /* Unmap with the same size we used for mapping */
> + munmap(data, map_size);
> + addr += bo_size;
> + }
> +
> + /* Cleanup allocated BOs - iterate through all n_execs */
> + for (int j = 0; j < n_execs; j++) {
> + if (bo[j])
> + gem_close(fd, bo[j]);
> + }
> +
> + xe_exec_queue_destroy(fd, bind_exec_queues[0]);
> + xe_vm_destroy(fd, vm);
> + drm_close_driver(fd);
> + free(bo);
> +
> + return bind_err;
> +}
> +
> +static unsigned int oom_working_set(uint64_t vram_size, uint64_t system_size,
> + uint64_t bo_size)
> +{
> + uint64_t target_allocation;
> + unsigned int set_size;
> +
> + /*
> + * For VRAM eviction testing, we want to allocate MORE than VRAM size
> + * to force eviction to system memory. Target 150-200% of VRAM.
> + * The BOs are created with VRAM placement, so they'll initially go to VRAM
> + * and then get evicted to system when VRAM fills up.
> + */
> + target_allocation = (vram_size * 150) / 100; /* 150% of VRAM */
> +
> + /* But ensure we don't exceed available system memory */
> + if (target_allocation > (system_size * 80) / 100) {
> + target_allocation = (system_size * 80) / 100;
> + igt_info("Limited by system memory: reducing target from %llu MB to %llu MB\n",
> + (unsigned long long)((vram_size * 150 / 100) >> 20),
> + (unsigned long long)(target_allocation >> 20));
> + }
> +
> + set_size = target_allocation / bo_size;
> +
> + igt_info("VRAM stress calculation: vram_size=%" PRIu64 " MB, system=%" PRIu64 " MB, "
> + "target_alloc=%" PRIu64 " MB (%.1f%% of VRAM), bo_size=%" PRIu64 " MB, set_size=%u\n",
> + (uint64_t)(vram_size >> 20), (uint64_t)(system_size >> 20),
> + (uint64_t)(target_allocation >> 20),
> + (double)(target_allocation * 100) / vram_size,
> + (uint64_t)(bo_size >> 20), set_size);
> +
> + return ALIGN_DOWN(set_size, 4);
> +}
> +
> +static void
> +test_vm_nonfault_mode_overcommit(int fd, struct drm_xe_engine_class_instance *eci,
> + uint64_t vram_size, uint64_t overcommit_multiplier)
> +{
> + uint32_t bo;
> + uint64_t phys_mem;
> + uint64_t overcommit_size;
> + uint32_t vm;
> + int ret, cret;
> +
> + fd = drm_open_driver(DRIVER_XE);
fd already passed from calling function. So no need to call
drm_open_driver() again
> + phys_mem = (uint64_t)igt_get_total_ram_mb() << 20;
system_size defined in igt_fixture() which can be passed as argument in
test_vm_nonfault_mode_overcommit() and can be used.
> + igt_info("Total system memory = %llu MB\n", (unsigned long long)phys_mem);
> + overcommit_size = ALIGN(phys_mem * 2, 4096);
> + vm = xe_vm_create(fd, 0, 0);
> + cret = __xe_bo_create(fd, vm, overcommit_size,
> + vram_memory(fd, eci->gt_id) | system_memory(fd),
> + 0, NULL, &bo);
> + if (cret) {
> + igt_assert_f(errno == E2BIG || errno == ENOMEM || errno == ENOSPC,
> + "xe_bo_create failed with unexpected errno=%d (%s)\n",
> + errno, strerror(errno));
> + igt_info("Non-fault overcommit: BO create failed as expected: %d (%s)\n",
> + errno, strerror(errno));
is it expected to get error code apart from E2BIG, ENOMEM and ENOSPC
which is being printed in igt_info? If it's not then igt_info can be
removed as error code printing and test halt done in igt_assert().
> + xe_vm_destroy(fd, vm);
> + drm_close_driver(fd);
> + return;
> + }
> +
> + ret = __xe_vm_bind(fd, vm, 0, bo, 0, 0x100000000,
> + overcommit_size, DRM_XE_VM_BIND_OP_MAP, 0,
> + NULL, 0, 0, 0, 0);
We can use 6th argument which is offset as 0 so that it can take address
to bind overcommit_size automatically
> + igt_assert_f(ret == -ENOMEM || ret == -ENOSPC,
> + "Expected bind to fail with -ENOMEM/-ENOSPC, got %d\n", ret);
> +
> + gem_close(fd, bo);
> + xe_vm_destroy(fd, vm);
> + drm_close_driver(fd);
> +}
> +
> +static void test_vm_fault_mode_overcommit(int fd, struct drm_xe_engine_class_instance *eci,
> + uint64_t vram_size, uint64_t overcommit_multiplier)
> +{
> + uint64_t phys_mem = igt_get_total_ram_mb() << 20;
> + uint64_t available_mem = igt_get_avail_ram_mb() << 20;
> + uint64_t overcommit_size;
> + uint32_t vm;
> + uint32_t bo;
> + uint64_t *ptr;
> + int create_ret;
> +
> + /* Limit overcommit to available memory to avoid OOM killer */
> + overcommit_size = ALIGN(phys_mem * overcommit_multiplier, 4096);
Same as above. physical memory size defined in igt_fixture() which can
be passed as argument in test_vm_fault_mode_overcommit()
> + if (overcommit_size > available_mem) {
> + igt_info("Limiting overcommit size from %llu MB to %llu MB (available)\n",
> + (unsigned long long)(overcommit_size >> 20),
> + (unsigned long long)(available_mem >> 20));
> + overcommit_size = ALIGN(available_mem, 4096);
> + }
> +
> + /* Need at least some reasonable size to test */
> + if (overcommit_size < (vram_size * 2)) {
> + igt_skip("Insufficient memory for fault mode overcommit test\n");
> + return;
> + }
> +
> + igt_info("Fault mode overcommit test: size=%llu MB, vram=%llu MB\n",
> + (unsigned long long)(overcommit_size >> 20),
> + (unsigned long long)(vram_size >> 20));
> +
> + fd = drm_reopen_driver(fd);
fd already passed in function so why drm_reopen_driver()?
> +
> + vm = xe_vm_create(fd, DRM_XE_VM_CREATE_FLAG_LR_MODE |
> + DRM_XE_VM_CREATE_FLAG_FAULT_MODE, 0);
> + igt_assert(vm);
> +
> + create_ret = __xe_bo_create(fd, 0, overcommit_size,
> + vram_memory(fd, eci->gt_id) | system_memory(fd),
> + DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM,
> + NULL, &bo);
> + if (create_ret) {
> + igt_info("BO creation failed (expected for overcommit): %s\n",
> + strerror(-create_ret));
why not igt_assert() if bo_create() failed with error?
> + xe_vm_destroy(fd, vm);
> + drm_close_driver(fd);
> + return;
> + }
> +
> + ptr = xe_bo_map(fd, bo, overcommit_size);
> + if (ptr == MAP_FAILED) {
> + igt_info("BO mapping failed (expected for overcommit): %s\n",
> + strerror(errno));
Can't we assert here?
> + gem_close(fd, bo);
> + xe_vm_destroy(fd, vm);
> + drm_close_driver(fd);
> + return;
> + }
> +
> + /* Touch first page */
> + memset(ptr, 0xAA, 4096);
> +
> + /* Touch sparse pages to test fault handling - limit to avoid OOM */
> + for (uint64_t off = 0; off < overcommit_size && off < (available_mem / 2);
This should be for (off = 0; off < overcommit_size; off += 4096) as page
alignment above set is 4096 and we are checking whether pages in
overcommit_size is accessible
> + off += (4096 * 256))
> + ((char *)ptr)[off] = 0xBB;
> +
> + igt_info("Fault mode overcommit test completed successfully\n");
> +
> + munmap(ptr, overcommit_size);
> + gem_close(fd, bo);
> + xe_vm_destroy(fd, vm);
> + drm_close_driver(fd);
> +}
> +
> static void
> test_evict_cm(int fd, struct drm_xe_engine_class_instance *eci,
> int n_exec_queues, int n_execs, size_t bo_size, unsigned long flags,
> @@ -665,7 +933,29 @@ static unsigned int working_set(uint64_t vram_size, uint64_t system_size,
> * @beng-threads-large: bind exec_queue threads large
> *
> */
> +/**
> + * SUBTEST: evict-vm-nonfault-overcommit
> + * Description: VM non-fault mode overcommit test - expects bind failure
> + * Test category: functionality test
> + * Feature: VM bind
> + */
>
> +/**
> + * SUBTEST: evict-vm-fault-overcommit
> + * Description: VM fault mode overcommit test - touch pages to trigger faults
> + * Test category: functionality test
> + * Feature: VM bind, fault mode
> + */
> +/**
> + * SUBTEST: evict-%s
> + * Description: %arg[1] out-of-memory evict test - expects graceful failure
> + * Test category: functionality test
> + *
> + * arg[1]:
> + *
> + * @oom-graceful: OOM graceful failure with small BOs
> + * @oom-graceful-large: OOM graceful failure with large BOs
> + */
> /*
> * Table driven test that attempts to cover all possible scenarios of eviction
> * (small / large objects, compute mode vs non-compute VMs, external BO or BOs
> @@ -730,6 +1020,17 @@ int igt_main()
> MULTI_VM },
> { NULL },
> };
> + const struct section_oom {
> + const char *name;
> + int n_exec_queues;
> + int mul;
> + int div;
> + unsigned int flags;
> +} sections_oom[] = {
> + { "oom-graceful", 1, 1, 128, BIND_EXEC_QUEUE },
> + { "oom-graceful-large", 1, 1, 16, BIND_EXEC_QUEUE },
> + { NULL },
> +};
> const struct section_threads {
> const char *name;
> int n_threads;
> @@ -836,6 +1137,14 @@ int igt_main()
> }
> }
>
> + igt_subtest("evict-vm-nonfault-overcommit") {
> + test_vm_nonfault_mode_overcommit(fd, hwe, vram_size, 2);
> + }
> +
> + igt_subtest("evict-vm-fault-overcommit") {
> + test_vm_fault_mode_overcommit(fd, hwe, vram_size, 2);
> + }
> +
> for (const struct section_cm *s = sections_cm; s->name; s++) {
> igt_subtest_f("evict-%s", s->name) {
> uint64_t bo_size = calc_bo_size(vram_size, s->mul, s->div);
> @@ -862,6 +1171,27 @@ int igt_main()
> min(ws, s->n_execs), bo_size, s->flags);
> }
> }
> + for (const struct section_oom *s = sections_oom; s->name; s++) {
> + igt_subtest_f("evict-%s", s->name) {
> + uint64_t bo_size = calc_bo_size(vram_size, s->mul, s->div);
> + int n_execs = oom_working_set(vram_size, system_size, bo_size);
> + int ret;
> +
> + igt_info("OOM test: n_execs %d, bo_size %" PRIu64 " MiB\n",
> + n_execs, bo_size >> 20);
> +
> + ret = test_evict_oom(fd, hwe, s->n_exec_queues, n_execs,
> + bo_size, s->flags);
> +
> + /* Accept success or graceful OOM errors */
> + igt_assert(ret == 0 || ret == -ENOSPC || ret == -ENOMEM);
> + if (ret != 0)
> + igt_info("Test passed: Got expected error %d (%s)\n",
> + ret, strerror(-ret));
> + else
> + igt_info("Test passed: All allocations and binds succeeded\n");
> + }
> +}
>
> igt_fixture()
> drm_close_driver(fd);
Let @Hellstrom, Thomas have a look on this.
Regards
Nishit
[-- Attachment #2: Type: text/html, Size: 51954 bytes --]
next prev parent reply other threads:[~2026-01-19 5:06 UTC|newest]
Thread overview: 8+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-01-13 7:08 [PATCH i-g-t 0/1] tests/intel/xe_evict: overcommit tests for fault-mode and non-fault-mode VMs Sobin Thomas
2026-01-13 7:08 ` [PATCH i-g-t 1/1] " Sobin Thomas
2026-01-19 5:06 ` Sharma, Nishit [this message]
2026-01-13 8:31 ` ✓ Xe.CI.BAT: success for tests/intel/xe_evict: overcommit tests for fault-mode and non-fault-mode VMs (rev2) Patchwork
2026-01-13 8:47 ` ✓ i915.CI.BAT: " Patchwork
2026-01-13 12:14 ` ✗ i915.CI.Full: failure " Patchwork
2026-01-13 14:02 ` ✗ Xe.CI.Full: " Patchwork
-- strict thread matches above, loose matches on Subject: below --
2026-01-13 3:58 [PATCH i-g-t 0/1] tests/intel/xe_evict: overcommit tests for fault-mode and non-fault-mode VMs Sobin Thomas
2026-01-13 3:58 ` [PATCH i-g-t 1/1] " Sobin Thomas
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=d4b4a604-02d6-42e6-910a-2b7c7792fec2@intel.com \
--to=nishit.sharma@intel.com \
--cc=igt-dev@lists.freedesktop.org \
--cc=sobin.thomas@intel.com \
--cc=thomas.hellstrom@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox