From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org Received: from gabe.freedesktop.org (gabe.freedesktop.org [131.252.210.177]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by smtp.lore.kernel.org (Postfix) with ESMTPS id F0EF9E93813 for ; Mon, 13 Apr 2026 04:33:46 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id 8C96810E300; Mon, 13 Apr 2026 04:33:46 +0000 (UTC) Authentication-Results: gabe.freedesktop.org; dkim=pass (2048-bit key; unprotected) header.d=intel.com header.i=@intel.com header.b="Or9AkLEb"; dkim-atps=neutral Received: from mgamail.intel.com (mgamail.intel.com [192.198.163.15]) by gabe.freedesktop.org (Postfix) with ESMTPS id EC64C10E300 for ; Mon, 13 Apr 2026 04:33:27 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=intel.com; i=@intel.com; q=dns/txt; s=Intel; t=1776054808; x=1807590808; h=from:to:cc:subject:date:message-id:mime-version: content-transfer-encoding; bh=0toSuDHDO6I14hNnI9/eQtw7E0TCaknT4kuEVz9M030=; b=Or9AkLEbr+QW45yGNk9v/ID+p3n2XnFGCFrhw71HLva2hQHKlPUQQ1a3 UPo41rqb3wd0X1YVqau0teZsa0F9MuJzzOPUChyHkafgG2KBKihyiqt4m WDO9L/NaUnhOMjqDeqrBiy+1J/YmHt5s63A0vfpW3SPM2uCRow/4FR3LU 6/YaAr2W3srqBLgAhRALGDNHPYmViXMa7hNQ+nhKdT2bnjEVxMWVsaNhC LhRqCxGRO6j2INmOV1ePjfWS2opFs30D/9eksrCs4c/aKE1I8VFBWzwmK SgUgKksuf+vq/QIxAVamsq5kYZ/H4bnSYiQNwbj3En5p3xUjIS1dZwexT g==; X-CSE-ConnectionGUID: XVyaFBPBTHW4yTwQ0PyZgA== X-CSE-MsgGUID: jbitXYxTRzmuKo2rcTNt/Q== X-IronPort-AV: E=McAfee;i="6800,10657,11757"; a="77097290" X-IronPort-AV: E=Sophos;i="6.23,176,1770624000"; d="scan'208";a="77097290" Received: from orviesa008.jf.intel.com ([10.64.159.148]) by fmvoesa109.fm.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 12 Apr 2026 21:33:27 -0700 X-CSE-ConnectionGUID: kC6N2sawQo+3JHp1kdLtyA== X-CSE-MsgGUID: IOgCLGqTTvKZEu4z+5A3Dw== X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="6.23,176,1770624000"; d="scan'208";a="229585122" Received: from dut6245dg2frd.fm.intel.com ([10.80.54.109]) by orviesa008-auth.jf.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 12 Apr 2026 21:33:27 -0700 From: Sobin Thomas To: igt-dev@lists.freedesktop.org, thomas.hellstrom@intel.com Cc: nishit.sharma@intel.com, Sobin Thomas Subject: [PATCH i-g-t v3 1/1] test/intel/xe_vm:Add oversubscribe concurrent bind stress subtest Date: Mon, 13 Apr 2026 04:33:18 +0000 Message-ID: <20260413043318.17931-1-sobin.thomas@intel.com> X-Mailer: git-send-email 2.52.0 MIME-Version: 1.0 Content-Transfer-Encoding: 8bit X-BeenThere: igt-dev@lists.freedesktop.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Development mailing list for IGT GPU Tools List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: igt-dev-bounces@lists.freedesktop.org Sender: "igt-dev" Add test for oversubscribing VRAM in multi process environment that creates VM, bind large BOs and submit workloads nearly simultaneously. Previous coverage lacked a scenario combining multi-process bind with VRAM oversubscription. This generates memory pressure with multi-process VM Bind activity and concurrent submission, exercising the bind pipeline under eviction pressure. v2: Removed helper APIs usage clock_nanosleep and commented code.(Nishit) v3: Refactored code to smaller functions. Added check for available SRAM usage and keep the max process to 20. v4: Remove explicit macros definition Replace Bind ioctl with library calls.(Thomas) Signed-off-by: Sobin Thomas --- tests/intel/xe_vm.c | 433 ++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 433 insertions(+) diff --git a/tests/intel/xe_vm.c b/tests/intel/xe_vm.c index d75b0730d..42effbd48 100644 --- a/tests/intel/xe_vm.c +++ b/tests/intel/xe_vm.c @@ -19,8 +19,100 @@ #include "xe/xe_ioctl.h" #include "xe/xe_query.h" #include "xe/xe_spin.h" +#include #include +#define GB(x) (1024ULL * 1024ULL * 1024ULL * (x)) +#define MAX_THREADS 20 +struct gem_bo { + uint32_t handle; + uint64_t size; + int *ptr; + uint64_t addr; +}; + +struct xe_test_ctx { + int fd; + uint32_t vm_id; + uint32_t exec_queue_id; + uint16_t sram_instance; + uint16_t vram_instance; + bool has_vram; +}; + +struct mem_bind_sync { + struct gem_bo *bufs; + int n_bufs; + uint64_t *binds_ufence; +}; + +static uint64_t align_to_page_size(uint64_t size) +{ + return (size + 4095UL) & ~4095UL; +} + +static void create_exec_queue(int fd, struct xe_test_ctx *ctx) +{ + struct drm_xe_engine_class_instance *hwe; + struct drm_xe_engine_class_instance eci = { 0 }; + + /* Use first available engine */ + xe_for_each_engine(fd, hwe) { + eci = *hwe; + break; + } + ctx->exec_queue_id = xe_exec_queue_create(fd, ctx->vm_id, &eci, 0); +} + +static uint64_t * +vm_bind_bo_batch(int fd, struct xe_test_ctx *ctx, struct gem_bo *bos, int size) +{ + uint64_t *ufence; + struct drm_xe_sync bind_sync; + struct drm_xe_vm_bind_op binds[size]; + int i; + + ufence = malloc(sizeof(uint64_t)); + *ufence = 0; + bind_sync = (struct drm_xe_sync) { + .type = DRM_XE_SYNC_TYPE_USER_FENCE, + .flags = DRM_XE_SYNC_FLAG_SIGNAL, + .addr = to_user_pointer(ufence), + .timeline_value = 1, + }; + + for (i = 0; i < size; i++) { + binds[i] = (struct drm_xe_vm_bind_op) { + .obj = bos[i].handle, + .obj_offset = 0, + .range = bos[i].size, + .addr = bos[i].addr, + .op = DRM_XE_VM_BIND_OP_MAP, + .flags = 0, + }; + } + xe_vm_bind_array(fd, ctx->vm_id, 0, binds, size, &bind_sync, 1); + return ufence; +} + +static void query_mem_info(int fd, struct xe_test_ctx *ctx) +{ + uint64_t vram_reg, sys_reg; + struct drm_xe_mem_region *region; + + ctx->has_vram = xe_has_vram(fd); + if (ctx->has_vram) { + vram_reg = vram_memory(fd, 0); + region = xe_mem_region(fd, vram_reg); + ctx->vram_instance = region->instance; + } + + sys_reg = system_memory(fd); + region = xe_mem_region(fd, sys_reg); + ctx->sram_instance = region->instance; + igt_debug("has_vram: %d\n", ctx->has_vram); +} + static uint32_t addr_low(uint64_t addr) { @@ -2696,6 +2788,342 @@ static void test_get_property(int fd, void (*func)(int fd, uint32_t vm)) xe_vm_destroy(fd, vm); } +static int build_add_batch(struct gem_bo *batch_bo, struct gem_bo *integers_bo, + struct gem_bo *result_bo, int ints_to_add) +{ + int pos = 0; + uint64_t tmp_addr; + #define GPR_RX_ADDR(x) (0x600 + (x) * 8) + + batch_bo->ptr[pos++] = MI_LOAD_REGISTER_MEM_CMD | MI_LRI_LRM_CS_MMIO | 2; + batch_bo->ptr[pos++] = GPR_RX_ADDR(0); + tmp_addr = integers_bo->addr + 0 * sizeof(uint32_t); + batch_bo->ptr[pos++] = tmp_addr & 0xFFFFFFFF; + batch_bo->ptr[pos++] = (tmp_addr >> 32) & 0xFFFFFFFF; + for (int i = 1; i < ints_to_add; i++) { + /* r1 = integers_bo[i] */ + batch_bo->ptr[pos++] = MI_LOAD_REGISTER_MEM_CMD | MI_LRI_LRM_CS_MMIO | 2; + batch_bo->ptr[pos++] = GPR_RX_ADDR(1); + tmp_addr = integers_bo->addr + i * sizeof(uint32_t); + batch_bo->ptr[pos++] = tmp_addr & 0xFFFFFFFF; + batch_bo->ptr[pos++] = (tmp_addr >> 32) & 0xFFFFFFFF; + /* r0 = r0 + r1 */ + batch_bo->ptr[pos++] = MI_MATH(4); + batch_bo->ptr[pos++] = MI_MATH_LOAD(MI_MATH_REG_SRCA, MI_MATH_REG(0)); + batch_bo->ptr[pos++] = MI_MATH_LOAD(MI_MATH_REG_SRCB, MI_MATH_REG(1)); + batch_bo->ptr[pos++] = MI_MATH_ADD; + batch_bo->ptr[pos++] = MI_MATH_STORE(MI_MATH_REG(0), MI_MATH_REG_ACCU); + } + /* result_bo[0] = r0 */ + batch_bo->ptr[pos++] = MI_STORE_REGISTER_MEM_GEN8 | MI_LRI_LRM_CS_MMIO; + batch_bo->ptr[pos++] = GPR_RX_ADDR(0); + tmp_addr = result_bo->addr + 0 * sizeof(uint32_t); + batch_bo->ptr[pos++] = tmp_addr & 0xFFFFFFFF; + batch_bo->ptr[pos++] = (tmp_addr >> 32) & 0xFFFFFFFF; + + batch_bo->ptr[pos++] = MI_BATCH_BUFFER_END; + while (pos % 4 != 0) + batch_bo->ptr[pos++] = MI_NOOP; + return pos; +} + +static void create_test_bos(int fd, struct xe_test_ctx *ctx, struct mem_bind_sync *bind, + uint32_t placement, uint64_t *addr) +{ + const char *mem_type = (placement & vram_memory(fd, 0)) ? "VRAM" : "SRAM"; + + for (int i = 0; i < bind->n_bufs; i++) { + struct gem_bo *bo = &bind->bufs[i]; + int ret; + + bo->size = GB(1); + ret = __xe_bo_create_caching(fd, ctx->vm_id, bo->size, + placement, 0, + DRM_XE_GEM_CPU_CACHING_WC, + &bo->handle); + if (ret == -ENOMEM || ret == -ENOSPC) { + bind->n_bufs = i;/* stop creating more */ + igt_debug("%s allocation failed at buffer %d\n", mem_type, i); + break; + } + bo->ptr = NULL; + bo->addr = *addr; + *addr += bo->size; + igt_debug("%s buffer %d created at 0x%016lx\n", mem_type, i, bo->addr); + } +} + +static int fill_random_integers(struct gem_bo *int_bo, int ints_to_add) +{ + uint32_t expected_result = 0; + + for (int i = 0; i < ints_to_add; i++) { + int random_int = rand() % 8; + + int_bo->ptr[i] = random_int; + expected_result += random_int; + + igt_debug("%d", random_int); + if (i + 1 != ints_to_add) + igt_debug(" + "); + else + igt_debug(" = "); + } + igt_debug("%d\n", expected_result); + return expected_result; +} + +static uint32_t setup_batch_syncs(int fd, struct drm_xe_sync *batch_syncs, int *n_batch_syncs, + uint64_t timeline_val, struct mem_bind_sync *vram_bind, + struct mem_bind_sync *sram_bind) +{ + uint32_t batch_syncobj = syncobj_create(fd, 0); + + batch_syncs[*n_batch_syncs] = (struct drm_xe_sync) { + .extensions = 0, + .type = DRM_XE_SYNC_TYPE_TIMELINE_SYNCOBJ, + .flags = DRM_XE_SYNC_FLAG_SIGNAL, + .handle = batch_syncobj, + .timeline_value = timeline_val, + }; + (*n_batch_syncs)++; + + return batch_syncobj; +} + +/* + * In concurrent VM bind stress tests, multiple threads simultaneously bind + * buffers to GPU virtual address space and submit batch operations. This + * creates significant GPU memory pressure where the kernel may transiently + * fail batch submission with ENOMEM when: + * - GPU page tables are being updated across multiple bindings + * - GPU memory is fragmented across many concurrent buffer mappings + * - Multiple processes compete for finite GPU resources + * + * Without retries, transient ENOMEM failures cause false test failures. + * Retrying ensures we distinguish temporary resource exhaustion from actual + * driver bugs. Non-ENOMEM errors still fail immediately and are properly + * reported with full errno context for debugging. + */ +static int xe_exec_with_retry(int fd, struct drm_xe_exec *exec, int max_retries) +{ + int rc = 0, retries = 0; + + for (retries = 0; retries < max_retries; retries++) { + rc = igt_ioctl(fd, DRM_IOCTL_XE_EXEC, exec); + + if (!(rc && errno == ENOMEM)) + break; + + usleep(100 * retries); + if (retries == 0) + igt_warn("got ENOMEM\n"); + } + + if (retries == max_retries) + igt_warn("gave up after %d retries\n", retries); + + if (rc) + igt_warn("errno: %d (%s)\n", errno, strerror(errno)); + + return rc; +} + +static void cleanup_bo_resources(int fd, struct gem_bo *bo) +{ + if (bo->ptr) { + igt_assert_eq(munmap(bo->ptr, bo->size), 0); + bo->ptr = NULL; + } + if (bo->handle) + gem_close(fd, bo->handle); +} + +static void cleanup_sram_vram_objs(int fd, struct mem_bind_sync *vram_bind, + struct mem_bind_sync *sram_bind) +{ + for (int i = 0; i < vram_bind->n_bufs; i++) + gem_close(fd, vram_bind->bufs[i].handle); + for (int i = 0; i < sram_bind->n_bufs; i++) + gem_close(fd, sram_bind->bufs[i].handle); + free(vram_bind->bufs); + free(sram_bind->bufs); + if (vram_bind->n_bufs) + free(vram_bind->binds_ufence); + if (vram_bind->n_bufs) + free(sram_bind->binds_ufence); +} + +/** + * SUBTEST: oversubscribe-concurrent-bind + * Description: Test for oversubscribing the VM with multiple processes + * doing binds at the same time, and ensure they all complete successfully. + * Functionality: This check is for a specific bug where if multiple processes + * oversubscribe the VM, some of the binds may fail with ENOMEM due to + * deadlock in the bind code. + * Test category: stress test + */ +static void test_vm_oversubscribe_concurrent_bind(int fd) +{ + #define MIN_BUFS_PER_PROC 2 + int n_proc = 0, n_vram_bufs = 0, n_sram_bufs = 0; + uint32_t max_by_mem; + uint64_t total_vram_demand = 0; + uint64_t vram_size = xe_visible_available_vram_size(fd, 0); + uint64_t sram_avail = (uint64_t)igt_get_avail_ram_mb() << 20; + uint64_t target_vram = vram_size * 2; /* 2 of VRAM */ + uint64_t target_sram = sram_avail * 50 / 100; /* 50% system RAM */ + + int total_vram_bufs = target_vram / GB(1); + int total_sram_bufs = target_sram / GB(1); + + /* determine concurrency from memory pressure */ + + pthread_barrier_t *barrier; + pthread_barrierattr_t attr; + + max_by_mem = min(total_vram_bufs / MIN_BUFS_PER_PROC, + total_sram_bufs / MIN_BUFS_PER_PROC); + n_proc = min_t(uint32_t, max_by_mem, MIN_BUFS_PER_PROC); + igt_assert(n_proc > 0); + + n_vram_bufs = max(2, total_vram_bufs / n_proc); + n_sram_bufs = max(2, total_sram_bufs / n_proc); + total_vram_demand = (uint64_t)n_proc * n_vram_bufs * GB(1); + + igt_debug("VRAM size: %" PRIu64 "MB, System RAM available: %" PRIu64 "MB\n", + vram_size >> 20, sram_avail >> 20); + + igt_debug(" n_proc = %d\n", n_proc); + igt_debug("VRAM: %" PRIu64 "GB\n", vram_size >> 30); + igt_debug("VRAM demand: %" PRIu64 "MB (%.2fx oversubscription)\n", + total_vram_demand >> 20, (double)total_vram_demand / vram_size); + igt_debug("Processes=%d VRAM_bufs=%d SRAM_bufs=%d\n", n_proc, + n_vram_bufs, n_sram_bufs); + + barrier = mmap(NULL, sizeof(pthread_barrier_t), + PROT_READ | PROT_WRITE, + MAP_SHARED | MAP_ANONYMOUS, -1, 0); + igt_assert(barrier != MAP_FAILED); + pthread_barrierattr_init(&attr); + pthread_barrierattr_setpshared(&attr, PTHREAD_PROCESS_SHARED); + pthread_barrier_init(barrier, &attr, n_proc); + + igt_fork(child, n_proc) { + struct xe_test_ctx ctx = {0}; + int rc; + uint64_t addr = 0x40000000; + int expected_result = 0, ints_to_add = 4; + int max_retries = 1024; + uint32_t batch_syncobj; + struct gem_bo integers_bo, result_bo, batch_bo, *vram_bufs, *sram_bufs; + struct gem_bo bo_arr[] = {result_bo, batch_bo, integers_bo}; + struct drm_xe_sync batch_syncs[3]; + int n_batch_syncs = 0; + int pos = 0; + uint64_t timeline_val = 1; + struct drm_xe_exec exec; + struct mem_bind_sync vram_bind = {0}; + struct mem_bind_sync sram_bind = {0}; + + vram_bufs = (struct gem_bo *)calloc(n_vram_bufs, sizeof(struct gem_bo)); + sram_bufs = (struct gem_bo *)calloc(n_sram_bufs, sizeof(struct gem_bo)); + srand(child); + + igt_assert(vram_bufs && sram_bufs); + + ctx.vm_id = xe_vm_create(fd, DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE, 0); + query_mem_info(fd, &ctx); + create_exec_queue(fd, &ctx); + vram_bind.bufs = vram_bufs; + vram_bind.n_bufs = n_vram_bufs; + sram_bind.bufs = sram_bufs; + sram_bind.n_bufs = n_sram_bufs; + + create_test_bos(fd, &ctx, &vram_bind, vram_memory(fd, 0), &addr); + create_test_bos(fd, &ctx, &sram_bind, system_memory(fd), &addr); + + pthread_barrier_wait(barrier); + + if (n_vram_bufs) + vram_bind.binds_ufence = vm_bind_bo_batch(fd, &ctx, vram_bufs, + n_vram_bufs); + + if (n_sram_bufs) + sram_bind.binds_ufence = vm_bind_bo_batch(fd, &ctx, sram_bufs, + n_sram_bufs); + + integers_bo.size = align_to_page_size(sizeof(int) * ints_to_add); + integers_bo.handle = xe_bo_create_caching(fd, ctx.vm_id, integers_bo.size, + system_memory(fd), 0, + DRM_XE_GEM_CPU_CACHING_WC); + integers_bo.ptr = (int *)xe_bo_map(fd, integers_bo.handle, integers_bo.size); + integers_bo.addr = 0x100000; + + expected_result = fill_random_integers(&integers_bo, ints_to_add); + igt_debug("%d\n", expected_result); + + result_bo.size = align_to_page_size(sizeof(int)); + result_bo.handle = xe_bo_create_caching(fd, ctx.vm_id, result_bo.size, + system_memory(fd), 0, + DRM_XE_GEM_CPU_CACHING_WC); + result_bo.ptr = NULL; + result_bo.addr = 0x200000; + + batch_bo.size = 4096; + batch_bo.handle = xe_bo_create_caching(fd, ctx.vm_id, batch_bo.size, + system_memory(fd), 0, + DRM_XE_GEM_CPU_CACHING_WC); + + batch_bo.ptr = (int *)xe_bo_map(fd, batch_bo.handle, batch_bo.size); + batch_bo.addr = 0x300000; + + pos = build_add_batch(&batch_bo, &integers_bo, &result_bo, ints_to_add); + + igt_assert(pos * sizeof(int) <= batch_bo.size); + xe_vm_bind_lr_sync(fd, ctx.vm_id, integers_bo.handle, 0, integers_bo.addr, + integers_bo.size, 0); + xe_vm_bind_lr_sync(fd, ctx.vm_id, result_bo.handle, 0, result_bo.addr, + result_bo.size, 0); + xe_vm_bind_lr_sync(fd, ctx.vm_id, batch_bo.handle, 0, batch_bo.addr, + batch_bo.size, 0); + batch_syncobj = setup_batch_syncs(fd, batch_syncs, &n_batch_syncs, timeline_val, + &vram_bind, &sram_bind); + + /* Wait for bind operations to complete on CPU */ + if (vram_bind.n_bufs) + xe_wait_ufence(fd, vram_bind.binds_ufence, 1, 0, INT64_MAX); + if (sram_bind.n_bufs) + xe_wait_ufence(fd, sram_bind.binds_ufence, 1, 0, INT64_MAX); + + exec = (struct drm_xe_exec) { + .exec_queue_id = ctx.exec_queue_id, + .num_syncs = n_batch_syncs, + .syncs = (uintptr_t)batch_syncs, + .address = batch_bo.addr, + .num_batch_buffer = 1, + }; + + rc = xe_exec_with_retry(fd, &exec, max_retries); + igt_assert_eq(rc, 0); + + igt_assert(syncobj_timeline_wait(fd, &batch_syncobj, + &timeline_val, 1, INT64_MAX, 0, NULL)); + result_bo.ptr = (int *)xe_bo_map(fd, result_bo.handle, result_bo.size); + igt_assert_eq(result_bo.ptr[0], expected_result); + cleanup_bo_resources(fd, bo_arr); + cleanup_sram_vram_objs(fd, &vram_bind, &sram_bind); + syncobj_destroy(fd, batch_syncobj); + xe_exec_queue_destroy(fd, ctx.exec_queue_id); + xe_vm_destroy(fd, ctx.vm_id); + close(fd); + } + igt_waitchildren(); + pthread_barrier_destroy(barrier); + pthread_barrierattr_destroy(&attr); + igt_assert_eq(munmap(barrier, sizeof(pthread_barrier_t)), 0); +} + int igt_main() { struct drm_xe_engine_class_instance *hwe, *hwe_non_copy = NULL; @@ -3088,6 +3516,11 @@ int igt_main() igt_subtest_f("invalid-flag-%s", s->name) invalid_flag(fd, s->flags); } + igt_subtest("oversubscribe-concurrent-bind") + { + igt_require(xe_has_vram(fd)); + test_vm_oversubscribe_concurrent_bind(fd); + } igt_subtest("invalid-extensions") invalid_extensions(fd); -- 2.52.0