From: Sobin Thomas <sobin.thomas@intel.com>
To: igt-dev@lists.freedesktop.org, thomas.hellstrom@intel.com
Cc: nishit.sharma@intel.com, kamil.konieczny@intel.com,
"Sobin Thomas" <sobin.thomas@intel.com>,
"Thomas Hellström" <thomas.hellstrom@linux.intel.com>
Subject: [PATCH i-g-t v13 2/2] tests/intel/xe_vm: Add support for overcommit tests
Date: Tue, 28 Apr 2026 03:25:51 +0000 [thread overview]
Message-ID: <20260428032551.752707-3-sobin.thomas@intel.com> (raw)
In-Reply-To: <20260428032551.752707-1-sobin.thomas@intel.com>
Current tests focus on VM creation with basic mode selection and do not
support overcommit scenarios.
This change adds tests to verify overcommit behavior across different VM
modes.
Non-fault mode tests:
- vram-lr-defer: DEFER_BACKING rejects overcommit at bind time
- vram-lr-external-nodefer: Long-running mode with external BO and
no defer backing
- vram-no-lr: Non-LR mode
Fault mode tests:
- vram-lr-fault: Fault handling allows graceful overcommit via page
faults
- vram-lr-fault-no-overcommit: Verifies NO_VM_OVERCOMMIT blocks same-VM
BO eviction during VM_BIND while still allowing eviction during
pagefault OOM
These tests validate that VMs handle memory pressure appropriately based
on their configuration—rejecting at bind, failing at exec, or handling
it gracefully via page faults.
v2 - Added Additional test cases for LR mode and No Overcommit.
v3 - Refactored into single api call based on the VM / BO Flags.
v5 - Addressed review comments (reset sync objects and nits).
Added check in cleanup
v6 - Replaced __xe_vm_bind with xe_vm_bind_lr_sync and refactored.
v7 - Add failable xe_vm_bind_lr_sync to handle the failure in the
vm bind in case over commit happens.
v9 - Replaced xe_vm_bind_lr_sync_failable with __xe_vm_bind_lr_sync
v10 - Add ENOSPC error, moved BO map after bind is completed.
Removed special casing LR Mode.
v11 - Add stage checks for the over commits for different stages
(bind / exec).
v12 - Removed bind user ptr exclusive for fault mode.
Replaced igt_assert with igt_require for xe_visible_vram_size
Signed-off-by: Sobin Thomas <sobin.thomas@intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
---
tests/intel/xe_vm.c | 391 +++++++++++++++++++++++++++++++++++++++++++-
1 file changed, 390 insertions(+), 1 deletion(-)
diff --git a/tests/intel/xe_vm.c b/tests/intel/xe_vm.c
index d75b0730d..408bfdb71 100644
--- a/tests/intel/xe_vm.c
+++ b/tests/intel/xe_vm.c
@@ -20,6 +20,14 @@
#include "xe/xe_query.h"
#include "xe/xe_spin.h"
#include <string.h>
+#define USER_FENCE_VALUE 0xdeadbeefdeadbeefull
+
+enum overcommit_stage {
+ EXPECT_NONE,
+ EXPECT_CREATE,
+ EXPECT_BIND,
+ EXPECT_EXEC,
+};
static uint32_t
addr_low(uint64_t addr)
@@ -2376,6 +2384,376 @@ static void invalid_vm_id(int fd)
do_ioctl_err(fd, DRM_IOCTL_XE_VM_DESTROY, &destroy, ENOENT);
}
+static enum overcommit_stage create_data_bos(int fd, uint32_t vm, uint32_t *bos, int num_bos,
+ uint64_t nf_bo_size, bool use_vram, uint64_t data_addr,
+ uint32_t bo_flags, int gt_id, int *num_bound_out)
+{
+ uint32_t placement = use_vram ? vram_memory(fd, gt_id) : system_memory(fd);
+
+ *num_bound_out = 0;
+
+ for (int i = 0; i < num_bos; i++) {
+ int bind_err;
+ int create_ret = 0;
+
+ /* Create BO using the case's create function */
+ create_ret = __xe_bo_create(fd, vm, nf_bo_size, placement,
+ bo_flags, NULL, &bos[i]);
+
+ if (create_ret) {
+ if (errno == ENOMEM || errno == ENOSPC) {
+ igt_debug("BO create failed at %d/%d with error %d (%s)\n",
+ i, num_bos, errno, strerror(errno));
+ return EXPECT_CREATE;
+ }
+ igt_assert_f(0, "Unexpected BO create error %d (%s)\n", errno,
+ strerror(errno));
+ }
+
+ bind_err = __xe_vm_bind_lr_sync(fd, vm, bos[i], 0, data_addr +
+ ((uint64_t)i * nf_bo_size), nf_bo_size, 0);
+ if (bind_err) {
+ if (errno == ENOMEM || errno == ENOSPC) {
+ igt_debug("BO bind failed at %d/%d - error %d (%s), %d BOs bound\n",
+ i, num_bos, errno, strerror(errno), i);
+ return EXPECT_BIND;
+ }
+ igt_assert_f(0, "Unexpected BO bind error %d (%s)\n", errno,
+ strerror(errno));
+ }
+
+ *num_bound_out = i + 1;
+ igt_debug("Created and bound BO %d/%d at 0x%llx\n",
+ i + 1, num_bos,
+ (unsigned long long)(data_addr + ((uint64_t)i * nf_bo_size)));
+ }
+ return EXPECT_NONE;
+}
+
+static void verify_bo(int fd, uint32_t *bos, int num_bos, uint64_t nf_bo_size, uint64_t stride)
+{
+ for (int i = 0; i < num_bos; i++) {
+ uint32_t *verify_data;
+ int errors = 0;
+
+ verify_data = xe_bo_map(fd, bos[i], nf_bo_size);
+ igt_assert(verify_data);
+
+ for (int off = 0; off < nf_bo_size; off += stride) {
+ uint32_t expected = 0xBB;
+ uint32_t actual = *(uint32_t *)((char *)verify_data + off);
+
+ if (actual != expected) {
+ if (errors < 5)
+ igt_debug("Mismatch at BO %d offset 0x%llx",
+ i, (unsigned long long)off);
+ errors++;
+ }
+ }
+
+ munmap(verify_data, nf_bo_size);
+ igt_assert_f(errors == 0, "Data verification failed for BO %d with %d errors\n",
+ i, errors);
+ }
+}
+
+static void bind_userptr_sync(int fd, uint32_t vm, uint32_t bind_exec_queue, void *userptr,
+ uint64_t addr, uint64_t size, struct drm_xe_sync *sync,
+ uint64_t *sync_mem)
+{
+ *sync_mem = 0;
+
+ sync->addr = to_user_pointer(sync_mem);
+ xe_vm_bind_userptr_async(fd, vm, bind_exec_queue, to_user_pointer(userptr), addr, size,
+ sync, 1);
+ xe_wait_ufence(fd, sync_mem, USER_FENCE_VALUE, bind_exec_queue, 20 * NSEC_PER_SEC);
+}
+
+/**
+ * SUBTEST: overcommit-fault-%s
+ * Description: Test VM overcommit behavior in fault mode with %arg[1] configuration
+ * Functionality: overcommit
+ * Test category: functionality test
+ *
+ * arg[1]:
+ *
+ * @vram-lr:VRAM with LR and fault mode, expects exec to pass
+ * @vram-lr-no-overcommit:VRAM with LR, fault and NO_VM_OVERCOMMIT; exec succeeds via migration
+ */
+
+/**
+ * SUBTEST: overcommit-nonfault-%s
+ * Description: Test VM overcommit behavior in nonfault mode with %arg[1] configuration
+ * Functionality: overcommit
+ * Test category: functionality test
+ *
+ * arg[1]:
+ *
+ * @vram-lr-defer:VRAM with LR and defer backing, expects bind rejection
+ * @vram-lr-external-nodefer:VRAM with LR and external BO without defer, expects bind fail
+ * @vram-no-lr:VRAM without LR mode, expects exec to fail
+ */
+struct vm_overcommit_case {
+ const char *name;
+ uint32_t vm_flags;
+ uint32_t bo_flags;
+ bool use_vram;
+ uint64_t data_addr;
+ int overcommit_mult;
+ enum overcommit_stage expected_stage;
+};
+
+static const struct vm_overcommit_case overcommit_cases[] = {
+ /* DEFER_BACKING */
+ {
+ .name = "vram-lr-defer",
+ .vm_flags = DRM_XE_VM_CREATE_FLAG_LR_MODE,
+ .bo_flags = DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING |
+ DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM,
+ .use_vram = true,
+ .data_addr = 0x1a0000,
+ .overcommit_mult = 2,
+ .expected_stage = EXPECT_BIND,
+ },
+ /* External BO without defer backing */
+ {
+ .name = "vram-lr-external-nodefer",
+ .vm_flags = DRM_XE_VM_CREATE_FLAG_LR_MODE,
+ .bo_flags = DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM,
+ .use_vram = true,
+ .data_addr = 0x1a0000,
+ .overcommit_mult = 2,
+ .expected_stage = EXPECT_BIND,
+ },
+ /* LR + FAULT - should not fail on exec */
+ {
+ .name = "vram-lr",
+ .vm_flags = DRM_XE_VM_CREATE_FLAG_LR_MODE |
+ DRM_XE_VM_CREATE_FLAG_FAULT_MODE,
+ .bo_flags = DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM,
+ .use_vram = true,
+ .data_addr = 0x300000000,
+ .overcommit_mult = 2,
+ .expected_stage = EXPECT_NONE,
+ },
+ /* !LR - overcommit should fail on exec */
+ {
+ .name = "vram-no-lr",
+ .vm_flags = 0,
+ .bo_flags = DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM,
+ .use_vram = true,
+ .data_addr = 0x300000000,
+ .overcommit_mult = 2,
+ .expected_stage = EXPECT_EXEC,
+ },
+ /* LR + FAULT + NO_VM_OVERCOMMIT */
+ {
+ .name = "vram-lr-no-overcommit",
+ .vm_flags = DRM_XE_VM_CREATE_FLAG_NO_VM_OVERCOMMIT | DRM_XE_VM_CREATE_FLAG_LR_MODE |
+ DRM_XE_VM_CREATE_FLAG_FAULT_MODE,
+ .bo_flags = DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING |
+ DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM,
+ .use_vram = true,
+ .data_addr = 0x300000000,
+ .overcommit_mult = 2,
+ /*
+ * FAULT_MODE handles VRAM pressure via migration even with
+ * NO_VM_OVERCOMMIT; DEFER_BACKING defers physical allocation
+ * to fault time so bind-time rejection does not occur.
+ */
+ .expected_stage = EXPECT_NONE,
+ },
+ { }
+};
+
+static void
+test_vm_overcommit(int fd, struct drm_xe_engine_class_instance *eci,
+ const struct vm_overcommit_case *c,
+ uint64_t system_size, uint64_t vram_size)
+{
+ uint32_t vm = 0, *bos, batch_bo = 0, exec_queue = 0, bind_exec_queue = 0;
+ uint64_t sync_addr = 0x1000000000, batch_addr = 0x200000000;
+ int i, num_bos, num_bound = 0, bind_err, create_ret, wait_ret;
+ size_t sync_size, nf_bo_size = 64 * 1024 * 1024;
+ enum overcommit_stage actual_stage = EXPECT_NONE;
+ uint64_t stride = 1024 * 1024, base_size;
+ uint64_t overcommit_size, off, data_addr;
+ int64_t timeout = 20 * NSEC_PER_SEC;
+ struct drm_xe_sync bind_sync[1] = {
+ {
+ .type = DRM_XE_SYNC_TYPE_USER_FENCE,
+ .flags = DRM_XE_SYNC_FLAG_SIGNAL,
+ .timeline_value = USER_FENCE_VALUE
+ },
+ };
+ struct drm_xe_sync exec_sync[1] = {
+ {
+ .type = DRM_XE_SYNC_TYPE_USER_FENCE,
+ .flags = DRM_XE_SYNC_FLAG_SIGNAL,
+ .timeline_value = USER_FENCE_VALUE,
+ },
+ };
+ struct drm_xe_exec exec = {
+ .num_batch_buffer = 1,
+ .num_syncs = 1,
+ .syncs = to_user_pointer(exec_sync),
+ };
+ struct {
+ uint32_t batch[16];
+ uint64_t pad;
+ uint32_t data;
+ uint64_t vm_sync;
+ } *batch_data = NULL;
+ uint64_t *user_fence_sync = NULL;
+
+ data_addr = c->data_addr;
+ base_size = c->use_vram ? vram_size : system_size;
+ overcommit_size = ALIGN((uint64_t)(base_size * c->overcommit_mult), 4096);
+
+ num_bos = (overcommit_size / nf_bo_size) + 1;
+ bos = calloc(num_bos, sizeof(*bos));
+ igt_assert(bos);
+
+ igt_debug("Overcommit test: allocating %d BOs of %llu MB each",
+ num_bos, (unsigned long long)(nf_bo_size >> 20));
+ igt_debug("total=%llu MB, vram=%llu MB\n",
+ (unsigned long long)(num_bos * nf_bo_size >> 20),
+ (unsigned long long)(vram_size >> 20));
+ /* Create VM with appropriate flags */
+ vm = xe_vm_create(fd, c->vm_flags, 0);
+ igt_assert(vm);
+
+ bind_exec_queue = xe_bind_exec_queue_create(fd, vm, 0);
+ sync_size = xe_bb_size(fd, sizeof(uint64_t) * num_bos);
+ user_fence_sync = mmap(NULL, sync_size, PROT_READ | PROT_WRITE,
+ MAP_SHARED | MAP_ANONYMOUS, -1, 0);
+ igt_assert(user_fence_sync != MAP_FAILED);
+ memset(user_fence_sync, 0, sync_size);
+ exec_sync->addr = to_user_pointer(&user_fence_sync[0]);
+
+ /* Create and bind data BOs */
+ actual_stage = create_data_bos(fd, vm, bos, num_bos, nf_bo_size, c->use_vram, data_addr,
+ c->bo_flags, eci->gt_id, &num_bound);
+ /*
+ * On EXPECT_CREATE nothing was bound so bail out entirely.
+ * On EXPECT_BIND with no BOs bound there is nothing to execute either.
+ * On EXPECT_BIND with some BOs bound, continue executing so that the
+ * already-bound BOs can still be executed, verifying they are usable
+ * after a partial bind failure.
+ */
+ if (actual_stage == EXPECT_CREATE || (actual_stage == EXPECT_BIND && num_bound == 0))
+ goto check_and_cleanup;
+
+ /*
+ * Create batch buffer first in SRAM as focus is to
+ * check overcommit in VRAM
+ */
+ create_ret = __xe_bo_create(fd, vm, 0x1000, system_memory(fd), 0, NULL, &batch_bo);
+ igt_assert_f(create_ret == 0, "Unexpected batch BO create error %d (%s)\n",
+ create_ret, strerror(errno));
+
+ igt_debug("Mapping the created BO\n");
+ batch_data = xe_bo_map(fd, batch_bo, 0x1000);
+ igt_assert(batch_data);
+ memset(batch_data, 0, 0x1000);
+
+ bind_userptr_sync(fd, vm, bind_exec_queue, user_fence_sync, sync_addr, sync_size, bind_sync,
+ &batch_data->vm_sync);
+ exec_sync->addr = sync_addr;
+ batch_data->vm_sync = 0;
+ bind_err = __xe_vm_bind_lr_sync(fd, vm, batch_bo, 0, batch_addr, 0x1000, 0);
+ if (bind_err) {
+ if (errno == ENOMEM || errno == ENOSPC) {
+ actual_stage = EXPECT_BIND;
+ goto check_and_cleanup;
+ } else { /* Assert on any unexpected bind error */
+ igt_assert_f(0, "Unexpected bind error %d (%s)\n", bind_err,
+ strerror(errno));
+ }
+ }
+
+ igt_debug("VM binds done - batch_bo at 0x%llx\n", (unsigned long long)batch_addr);
+ exec_queue = xe_exec_queue_create(fd, vm, eci, 0);
+
+ /* Use GPU to write to each successfully bound BO */
+ for (i = 0; i < num_bound; i++) {
+ igt_debug("Writing to BO %d/%d via GPU\n", i + 1, num_bos);
+ timeout = 20 * NSEC_PER_SEC;
+
+ for (off = 0; off < nf_bo_size; off += stride) {
+ uint64_t target_addr = data_addr + ((uint64_t)i * nf_bo_size) + off;
+ int b_idx = 0;
+
+ batch_data->batch[b_idx++] = MI_STORE_DWORD_IMM_GEN4;
+ batch_data->batch[b_idx++] = target_addr & 0xFFFFFFFF;
+ batch_data->batch[b_idx++] = (target_addr >> 32) & 0xFFFFFFFF;
+ batch_data->batch[b_idx++] = 0xBB;
+ batch_data->batch[b_idx++] = MI_BATCH_BUFFER_END;
+
+ /* Submit batch */
+ exec.exec_queue_id = exec_queue;
+ exec.address = batch_addr;
+
+ if (igt_ioctl(fd, DRM_IOCTL_XE_EXEC, &exec)) {
+ if (errno == ENOMEM || errno == ENOSPC) {
+ igt_debug("Expected fault/error: %d (%s)\n",
+ errno, strerror(errno));
+ actual_stage = EXPECT_EXEC;
+ goto check_and_cleanup;
+ }
+ igt_assert_f(0, "Unexpected exec error: %d\n", errno);
+ }
+ wait_ret = __xe_wait_ufence(fd, &user_fence_sync[0],
+ USER_FENCE_VALUE, exec_queue, &timeout);
+ /*
+ * EIO means the exec queue was banned due to VRAM
+ * exhaustion in non-fault mode after partial bind.
+ */
+ if (wait_ret == -EIO) {
+ igt_assert_f(c->expected_stage == EXPECT_BIND ||
+ c->expected_stage == EXPECT_EXEC,
+ "Unexpected queue reset\n");
+ actual_stage = EXPECT_EXEC;
+ goto check_and_cleanup;
+ }
+ igt_assert_eq(wait_ret, 0);
+ user_fence_sync[0] = 0;
+ }
+ igt_debug("Accessed BO %d/%d via GPU\n", i + 1, num_bos);
+ }
+ igt_debug("All batches submitted - waiting for GPU completion\n");
+
+ /* Verify GPU writes for bound BOs */
+ if (actual_stage == EXPECT_NONE || (actual_stage == EXPECT_BIND && num_bound > 0))
+ verify_bo(fd, bos, num_bound, nf_bo_size, stride);
+
+check_and_cleanup:
+ igt_assert_f(actual_stage == c->expected_stage, "Expected overcommit at stage %d, got %d\n",
+ c->expected_stage, actual_stage);
+ /* Cleanup */
+ if (exec_queue)
+ xe_exec_queue_destroy(fd, exec_queue);
+ if (bind_exec_queue)
+ xe_exec_queue_destroy(fd, bind_exec_queue);
+ if (batch_data)
+ munmap(batch_data, 0x1000);
+ if (batch_bo)
+ gem_close(fd, batch_bo);
+
+ if (user_fence_sync)
+ munmap(user_fence_sync, sync_size);
+
+ if (bos) {
+ for (i = 0; i < num_bos; i++) {
+ if (bos[i])
+ gem_close(fd, bos[i]);
+ }
+ free(bos);
+ }
+ if (vm > 0)
+ xe_vm_destroy(fd, vm);
+}
+
/**
* SUBTEST: out-of-memory
* Description: Test if vm_bind ioctl results in oom
@@ -2385,7 +2763,6 @@ static void invalid_vm_id(int fd)
*/
static void test_oom(int fd)
{
-#define USER_FENCE_VALUE 0xdeadbeefdeadbeefull
#define BO_SIZE xe_bb_size(fd, SZ_512M)
#define MAX_BUFS ((int)(xe_visible_vram_size(fd, 0) / BO_SIZE))
uint64_t addr = 0x1a0000;
@@ -3115,6 +3492,18 @@ int igt_main()
test_get_property(fd, f->test);
}
+ for (int i = 0; overcommit_cases[i].name; i++) {
+ const struct vm_overcommit_case *c = &overcommit_cases[i];
+ const char *mode = (c->vm_flags & DRM_XE_VM_CREATE_FLAG_FAULT_MODE) ?
+ "fault" : "nonfault";
+ igt_subtest_f("overcommit-%s-%s", mode, c->name) {
+ igt_require(xe_has_vram(fd));
+ igt_require(xe_visible_vram_size(fd, 0));
+ test_vm_overcommit(fd, hwe_non_copy, c, (igt_get_avail_ram_mb() << 20),
+ xe_visible_vram_size(fd, 0));
+ }
+ }
+
igt_fixture()
drm_close_driver(fd);
}
--
2.52.0
next prev parent reply other threads:[~2026-04-28 3:26 UTC|newest]
Thread overview: 9+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-04-28 3:25 [PATCH i-g-t v13 0/2] tests/intel/xe_vm: Add support for overcommit tests Sobin Thomas
2026-04-28 3:25 ` [PATCH i-g-t v13 1/2] lib/xe: Add failable variant of xe_vm_bind_lr_sync() Sobin Thomas
2026-04-28 3:25 ` Sobin Thomas [this message]
2026-04-28 4:27 ` ✓ i915.CI.BAT: success for tests/intel/xe_vm: Add support for overcommit tests (rev9) Patchwork
2026-04-28 4:37 ` ✓ Xe.CI.BAT: " Patchwork
2026-04-28 10:16 ` ✓ i915.CI.Full: " Patchwork
2026-04-28 11:11 ` ✗ Xe.CI.FULL: failure " Patchwork
2026-04-28 12:17 ` Thomas, Sobin
2026-04-28 13:36 ` Thomas, Sobin
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=20260428032551.752707-3-sobin.thomas@intel.com \
--to=sobin.thomas@intel.com \
--cc=igt-dev@lists.freedesktop.org \
--cc=kamil.konieczny@intel.com \
--cc=nishit.sharma@intel.com \
--cc=thomas.hellstrom@intel.com \
--cc=thomas.hellstrom@linux.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