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 0A1C7CA6019 for ; Fri, 9 Oct 2026 10:32:38 +0000 (UTC) Received: from gabe.freedesktop.org (localhost [127.0.0.1]) by gabe.freedesktop.org (Postfix) with ESMTP id AC7F2112047; Fri, 9 Oct 2026 10:32:37 +0000 (UTC) Authentication-Results: gabe.freedesktop.org; dkim=pass (2048-bit key; unprotected) header.d=intel.com header.i=@intel.com header.b="UfCnI2Vd"; dkim-atps=neutral Received: from mgamail.intel.com (mgamail.intel.com [198.175.65.14]) by gabe.freedesktop.org (Postfix) with ESMTPS id B8E09112048 for ; Fri, 9 Oct 2026 10:28:34 +0000 (UTC) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=intel.com; i=@intel.com; q=dns/txt; s=Intel; t=1791541715; x=1823077715; h=from:to:cc:subject:date:message-id:in-reply-to: references:mime-version:content-transfer-encoding; bh=1CgZlbeweeqYjUfkyMKTpTvJzwj/wlJ2QZX3LRKO8RM=; b=UfCnI2VdcfJR9A/qphgI06QzGLX66JL489QXlwv/ItulEESVHiRcao8P KMR55/U5eF1f27eE7P7aPbRG/TwakqpxStVXC47Ypy8WWPZecIGVUOaoM Wz1iyTRovEe6NSk4paq+ge8ddT8YBCH2V/cLul27vU0SfXWZi+PiAXlpG Gz4njwoP/MmquzstRSRomSBfF5tCcpzJDDKRR6YbytTNbUgM3n7cz9ffw /K9PK6XrtFwuV9ytyJFTNacdq58cooxWfA8Xfo9t4qXpsL96Etdtc28hv 9cnmNn+o7wyfmzOp5OaH6fxnVEuNWDNTtBzXbybush3StlZWiyjUxR740 w==; X-CSE-ConnectionGUID: rnhh0RcWQFWiUAXV+yLW2A== X-CSE-MsgGUID: jCf0mQ1NQzSk5Ps66XTEbg== X-IronPort-AV: E=McAfee;i="6800,10657,11929"; a="218782" X-IronPort-AV: E=Sophos;i="6.27,148,1787036400"; d="scan'208";a="218782" Received: from fmviesa006.fm.intel.com ([10.60.135.146]) by orvoesa106.jf.intel.com with ESMTP/TLS/ECDHE-RSA-AES256-GCM-SHA384; 09 Oct 2026 03:28:35 -0700 X-CSE-ConnectionGUID: Wp7NB8z5SR+aBcBOHG+s9g== X-CSE-MsgGUID: o/LUAjFARCOY3Am19/Kagw== X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="6.27,148,1787036400"; d="scan'208";a="621494" Received: from dut6245dg2frd.fm.intel.com ([10.36.24.131]) by fmviesa006.fm.intel.com with ESMTP; 09 Oct 2026 03:28:34 -0700 From: Sobin Thomas To: igt-dev@lists.freedesktop.org, matthew.brost@intel.com Cc: nishit.sharma@intel.com, priyanka.dandamudi@intel.com Subject: [PATCH i-g-t v6 5/5] tests/intel: Add oversubscribe concurrent bind stress subtest Date: Fri, 9 Oct 2026 10:28:32 +0000 Message-ID: <20261009102832.2839902-6-sobin.thomas@intel.com> X-Mailer: git-send-email 2.55.0 In-Reply-To: <20261009102832.2839902-1-sobin.thomas@intel.com> References: <20261009102832.2839902-1-sobin.thomas@intel.com> 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. v3: Fixed minor nits v4: Refactoring of the create_bo ( Nishit) Fix child process exit deadlock in release_ready_processes() (Nishit) v6: - Abort all child threads and fail with a reaon on timeout (Nishit) - Track the setup fault , OOM Faults and completed child process.(Nishit) Signed-off-by: Sobin Thomas --- tests/intel/xe_bo_alloc.c | 616 +++++++++++++++++++++++++++++++++++++- 1 file changed, 609 insertions(+), 7 deletions(-) diff --git a/tests/intel/xe_bo_alloc.c b/tests/intel/xe_bo_alloc.c index 4b62be299..c820421bd 100644 --- a/tests/intel/xe_bo_alloc.c +++ b/tests/intel/xe_bo_alloc.c @@ -31,6 +31,18 @@ #define N_BOS_PER_BIND 10 #define THREAD_VA_STRIDE GB(1) +#define INT_ADD_CNT 4 +#define TIMEOUT_NS (30ULL * 1000000000ULL) +#define MAX_SRAM_TEST_SIZE GB(32) +#define GPR_RX_ADDR(x) (0x600 + (x) * 8) +#define MAX_PROCS 20 + +#define EXEC_DATA_ADDR 0x100000ULL +#define EXEC_RESULT_ADDR 0x200000ULL +#define EXEC_BATCH_ADDR 0x300000ULL +#define EXEC_UFENCE_ADDR 0x400000ULL +#define STRESS_BIND_ADDR 0x40000000ULL + /** * SUBTEST: all-sizes-once * Description: Test all BO allocations sizes in test table @@ -85,12 +97,51 @@ * Description: Test 50 random BO allocation sizes in test table with a thread per engine, * leak the binding, unaligned bind addresses * Test category: stress test + * + * SUBTEST: test_vm_oversubscribe_concurrent_bind + * Description: Test enough random BO allocation sizes, bound as arrays of binds, to trigger + * evictions with 2 processes per engine, leak the BO munmap / gem close, unaligned bind addresses + * Test category: stress test */ + #define GB(x) (1024ULL * 1024ULL * 1024ULL * (x)) +#define MIN_BUFS_PER_PROC 2 #define N_ALLOC_SIZES 256 +#define USER_FENCE_VALUE 0xdeadbeefdeadbeefull + static uint64_t *alloc_sizes; +struct gem_bo { + uint32_t handle; + uint64_t size; + uint32_t *ptr; + uint64_t addr; +}; + +struct xe_oversubscribe_ctx { + uint32_t vm_id; + uint32_t exec_queue_id; +}; + +struct mem_bind_sync { + struct gem_bo *bufs; + int n_bufs; + uint64_t *binds_ufence; +}; + +struct process_data { + pthread_mutex_t mutex; + pthread_cond_t cond; + int ready; + int failed; + int fault_count; + int child_completed; + int bind_oom; + bool go; + bool abort; +}; + /* * Data-driven subtest matrix. * @@ -111,6 +162,7 @@ enum test_type { TYPE_RANDOM_SIZE, TYPE_ARRAY_BIND, TYPE_THREAD, + TYPE_OVERSUBSCRIBE, }; struct test_case { @@ -126,6 +178,14 @@ struct bo_group { int n_bos; }; +struct oversubscribe_params { + int n_proc; + int n_vram_bufs; + int n_sram_bufs; + uint64_t vram_size; + uint64_t sram_size; + uint64_t total_vram_demand; +}; #define LEAK_BINDING (0x1 << 0) #define LEAK_BO (0x1 << 1) #define EVICT (0x1 << 2) @@ -251,6 +311,544 @@ static int __xe_vm_bind_array(int fd, uint32_t vm, return 0; } +static void init_pdata(struct process_data *pdata) +{ + pthread_mutexattr_t mattr; + pthread_condattr_t cattr; + + pthread_mutexattr_init(&mattr); + igt_assert_eq(pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED), 0); + igt_assert_eq(pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST), 0); + igt_assert_eq(pthread_mutex_init(&pdata->mutex, &mattr), 0); + pthread_mutexattr_destroy(&mattr); + + pthread_condattr_init(&cattr); + igt_assert_eq(pthread_condattr_setpshared(&cattr, PTHREAD_PROCESS_SHARED), 0); + igt_assert_eq(pthread_cond_init(&pdata->cond, &cattr), 0); + pthread_condattr_destroy(&cattr); + pdata->ready = 0; + pdata->failed = 0; + pdata->child_completed = 0; + pdata->fault_count = 0; + pdata->bind_oom = 0; + pdata->go = false; + pdata->abort = false; +} + +static bool process_ready_and_wait(struct process_data *pdata) +{ + struct timespec timeout; + bool success; + int ret = pthread_mutex_lock(&pdata->mutex); + + if (ret == EOWNERDEAD) + igt_assert_eq(pthread_mutex_consistent(&pdata->mutex), 0); + else + igt_assert_eq(ret, 0); + pdata->ready++; + pthread_cond_broadcast(&pdata->cond); + clock_gettime(CLOCK_REALTIME, &timeout); + timeout.tv_sec += 30; + /* We are handling timeout and exit the while if child wait > 30s */ + while (!pdata->go && !pdata->abort) { + ret = pthread_cond_timedwait(&pdata->cond, &pdata->mutex, &timeout); + if (ret == ETIMEDOUT) { + pdata->abort = true; + pdata->go = true; + pthread_cond_broadcast(&pdata->cond); + break; + } + igt_assert_eq(ret, 0); + } + success = !pdata->abort; + pthread_mutex_unlock(&pdata->mutex); + return success; +} + +static void process_setup_failed(struct process_data *pdata, bool fault) +{ + int ret = pthread_mutex_lock(&pdata->mutex); + + if (ret == EOWNERDEAD) + igt_assert_eq(pthread_mutex_consistent(&pdata->mutex), 0); + else + igt_assert_eq(ret, 0); + /* This will be incremented for all real error / OOM */ + pdata->failed++; + if (fault) + pdata->fault_count++; /*Will be incremented for fault case */ + pthread_cond_broadcast(&pdata->cond); + pthread_mutex_unlock(&pdata->mutex); +} + +static void release_ready_processes(struct process_data *pdata, int n_proc) +{ + struct timespec timeout; + int ret; + + ret = pthread_mutex_lock(&pdata->mutex); + if (ret == EOWNERDEAD) + igt_assert_eq(pthread_mutex_consistent(&pdata->mutex), 0); + else + igt_assert_eq(ret, 0); + + clock_gettime(CLOCK_REALTIME, &timeout); + timeout.tv_sec += 30; + /* + * Wait until all children have either reached VM Bind rendevouz + * Or, failed setup and exited the test path. + */ + while (pdata->ready + pdata->failed < n_proc) { + ret = pthread_cond_timedwait(&pdata->cond, &pdata->mutex, &timeout); + if (ret == EOWNERDEAD) { + igt_assert_eq(pthread_mutex_consistent(&pdata->mutex), 0); + continue; /* re-check loop condition */ + } else if (ret == ETIMEDOUT) { + igt_warn("Timeout waiting for the child process: ready=%d," + " failed=%d, expected=%d\n", + pdata->ready, pdata->failed, n_proc); + pdata->failed++; + break; + } + igt_assert_eq(ret, 0); + } + + igt_debug("Process rendezvous: ready=%d failed=%d total = %d\n", + pdata->ready, pdata->failed, n_proc); + + /* Release every successful child into VM_Bind at same time*/ + pdata->go = true; + pthread_cond_broadcast(&pdata->cond); + pthread_mutex_unlock(&pdata->mutex); +} + +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; + int i; + uint64_t tmp_addr; + + 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 (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_exec_queue(int fd, struct xe_oversubscribe_ctx *ctx) +{ + ctx->exec_queue_id = xe_exec_queue_create(fd, ctx->vm_id, + &xe_engine(fd, 0)->instance, 0); +} + +static uint64_t * +vm_bind_bo_batch(int fd, struct xe_oversubscribe_ctx *ctx, struct gem_bo *bos, int size, + int *out_err) +{ + uint64_t *ufence; + struct drm_xe_sync bind_sync; + struct drm_xe_vm_bind_op *binds; + int i; + + binds = calloc(size, sizeof(*binds)); + igt_assert(binds); + + ufence = calloc(1, sizeof(*ufence)); + igt_assert(ufence); + 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, + }; + } + *out_err = __xe_vm_bind_array(fd, ctx->vm_id, binds, size, &bind_sync, 1); + free(binds); + return ufence; +} + +/* Returns 1 (caller should skip exec) on an allowed OOM error, 0 otherwise. */ +static int bind_bos(int fd, struct xe_oversubscribe_ctx *ctx, struct gem_bo *bufs, + struct mem_bind_sync *bind, int *bind_err, bool allow_oom) +{ + if (!bind->n_bufs) + return 0; + + bind->binds_ufence = vm_bind_bo_batch(fd, ctx, bufs, bind->n_bufs, bind_err); + + if (*bind_err) { + if (allow_oom) { + igt_assert_f(*bind_err == -ENOMEM || *bind_err == -ENOSPC, + "Unexpected bind error: %d (%s)\n", + *bind_err, strerror(-*bind_err)); + igt_debug("Bind failed with expected OOM (%s), skipping exec\n", + strerror(-*bind_err)); + return 1; + } + + igt_assert_f(false, "Unexpected bind error: %d", *bind_err); + } + + xe_wait_ufence(fd, bind->binds_ufence, 1, 0, TIMEOUT_NS); + return 0; +} + +static int fill_random_integers(struct gem_bo *int_bo, int ints_to_add) +{ + uint32_t expected_result = 0; + char expr[256]; + int len = 0; + + for (int i = 0; i < ints_to_add; i++) { + uint32_t random_int = rand() % 8; + + int_bo->ptr[i] = random_int; + expected_result += random_int; + + len += snprintf(expr + len, sizeof(expr) - len, "%s%u", + i ? " + " : "", random_int); + } + igt_debug("%s = %u\n", expr, expected_result); + return expected_result; +} + +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 int create_test_bos(int fd, struct xe_oversubscribe_ctx *ctx, + struct mem_bind_sync *bind, uint32_t placement, + uint64_t *addr) +{ + const char *mem_type = (placement & vram_memory(fd, 0)) ? "VRAM" : "SRAM"; + int ret; + + for (int i = 0; i < bind->n_bufs; i++) { + struct gem_bo *bo = &bind->bufs[i]; + + 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) { + int saved_errno = errno; /* capture before anything can clobber it */ + + bind->n_bufs = i; + if (saved_errno == ENOMEM || saved_errno == ENOSPC) { + /* Continue on OOM, expected when oversubscribing the VM */ + igt_debug("%s allocation failed at buffer %d (OOM)\n", mem_type, i); + break; + } + /* We are returning as this is a fail scenario */ + igt_warn("%s allocation failed at buffer %d: %s\n", + mem_type, i, strerror(saved_errno)); + return -saved_errno; + } + bo->ptr = NULL; + bo->addr = *addr; + *addr += bo->size; + igt_debug("%s buffer %d created at 0x%016lx\n", mem_type, i, bo->addr); + } + return 0; +} + +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->binds_ufence) + free(vram_bind->binds_ufence); + if (sram_bind->binds_ufence) + free(sram_bind->binds_ufence); +} + +static void create_bo(int fd, uint32_t vm, struct gem_bo *bo, uint64_t size, + uint64_t addr, bool map_now) +{ + bo->size = ALIGN(size, 4096); + bo->handle = xe_bo_create_caching(fd, vm, bo->size, system_memory(fd), 0, + DRM_XE_GEM_CPU_CACHING_WC); + igt_assert(bo->handle); + bo->addr = addr; + if (map_now) { + bo->ptr = xe_bo_map(fd, bo->handle, bo->size); + igt_assert(bo->ptr != MAP_FAILED); + } else { + bo->ptr = NULL; + } +} + +static bool calc_oversubscribe_params(int fd, struct oversubscribe_params *p) +{ + uint64_t target_vram, target_sram; + uint64_t total_vram_bufs, total_sram_bufs, max_by_mem; + + p->vram_size = xe_visible_available_vram_size(fd, 0); + p->sram_size = (uint64_t)igt_get_avail_ram_mb() << 20; + /* + * Dynamically cap VRAM oversubscription so the overflow into system + * RAM stays within 25% of available RAM. On small-VRAM platforms + * the 2x target fits within the cap and behavior is + * unchanged; on large-VRAM platforms (e.g. PVC) this prevents OOM. + */ + + target_vram = min(p->vram_size * 2, p->vram_size + p->sram_size / 4); + target_sram = min_t(uint64_t, p->sram_size * 50 / 100, + MAX_SRAM_TEST_SIZE); + total_vram_bufs = target_vram / GB(1); + total_sram_bufs = target_sram / GB(1); + max_by_mem = min(total_vram_bufs / MIN_BUFS_PER_PROC, + total_sram_bufs / MIN_BUFS_PER_PROC); + + p->n_proc = min_t(int, max_by_mem, MAX_PROCS); + /* Ensure we have minimum of 2 process */ + if (p->n_proc < 2) { + igt_warn("At least 2 processes are required for oversubscription.\n"); + return false; + } + p->n_vram_bufs = max_t(int, 2, total_vram_bufs / p->n_proc); + p->n_sram_bufs = max_t(int, 2, total_sram_bufs / p->n_proc); + p->total_vram_demand = (uint64_t)p->n_proc * p->n_vram_bufs * GB(1); + return true; +} + +static void test_vm_oversubscribe_concurrent_bind(int fd) +{ + struct oversubscribe_params p; + struct process_data *pdata; + + igt_require_f(calc_oversubscribe_params(fd, &p), + "Not enough VRAM/RAM for oversubscription test\n"); + igt_debug("VRAM size: %" PRIu64 "MB, System RAM available: %" PRIu64 "MB\n", + p.vram_size >> 20, p.sram_size >> 20); + + igt_debug("n_proc = %d\n", p.n_proc); + igt_debug("VRAM: %" PRIu64 "GB\n", p.vram_size >> 30); + igt_debug("VRAM demand: %" PRIu64 "MB (%.2fx oversubscription)\n", + p.total_vram_demand >> 20, (double)p.total_vram_demand / p.vram_size); + igt_debug("Processes=%d VRAM_bufs=%d SRAM_bufs=%d\n", p.n_proc, + p.n_vram_bufs, p.n_sram_bufs); + + pdata = mmap(NULL, sizeof(*pdata), PROT_READ | PROT_WRITE, + MAP_SHARED | MAP_ANONYMOUS, -1, 0); + igt_assert(pdata != MAP_FAILED); + init_pdata(pdata); + + igt_fork(child, p.n_proc) { + struct xe_oversubscribe_ctx ctx = {0}; + int rc, ret, fd_c; + uint64_t addr = STRESS_BIND_ADDR; + uint32_t expected_result = 0; + struct gem_bo integers_bo = {0}, result_bo = {0}, batch_bo = {0}; + struct gem_bo *vram_bufs, *sram_bufs; + int pos = 0; + struct mem_bind_sync vram_bind = {0}; + struct mem_bind_sync sram_bind = {0}; + struct drm_xe_sync batch_syncs[1]; + struct drm_xe_exec exec; + struct gem_bo ufence_bo = {0}; + int vram_bind_err = 0, sram_bind_err = 0; + + vram_bufs = calloc(p.n_vram_bufs, sizeof(*vram_bufs)); + sram_bufs = calloc(p.n_sram_bufs, sizeof(*sram_bufs)); + srand(child); + + fd_c = drm_open_driver(DRIVER_XE); + + igt_assert(vram_bufs && sram_bufs); + + ctx.vm_id = xe_vm_create(fd_c, DRM_XE_VM_CREATE_FLAG_LR_MODE, 0); + create_exec_queue(fd_c, &ctx); + vram_bind.bufs = vram_bufs; + vram_bind.n_bufs = p.n_vram_bufs; + sram_bind.bufs = sram_bufs; + sram_bind.n_bufs = p.n_sram_bufs; + + ret = create_test_bos(fd_c, &ctx, &vram_bind, vram_memory(fd_c, 0), &addr); + if (ret) { + igt_debug("Child process failed while bo creation in VRAM"); + process_setup_failed(pdata, true); + goto cleanup; + } + + ret = create_test_bos(fd_c, &ctx, &sram_bind, system_memory(fd_c), &addr); + if (ret) { + igt_debug("Child process failed while bo creation in SRAM"); + process_setup_failed(pdata, true); + goto cleanup; + } + + if (!vram_bind.n_bufs || !sram_bind.n_bufs) { + /* Skipping test as we are running out of memory + * We will keep track in the pdata->failed count + */ + igt_debug("Skipping No BOs allocated; VRAM/SRAM unavailable\n"); + process_setup_failed(pdata, false); + goto cleanup; + } + + /* + * All allocations are complete. Report Ready and wait until every + * child has either reached this point or reported a setup failure. + */ + + if (!process_ready_and_wait(pdata)) { + igt_debug("Rendezvous aborted/timed out, skipping bind + exec\n"); + process_setup_failed(pdata, true); + goto cleanup; + } + /* + * VM_Bind starts only after the parent signal all ready + * childen. + */ + + if (bind_bos(fd_c, &ctx, vram_bufs, &vram_bind, &vram_bind_err, true)) { + ret = pthread_mutex_lock(&pdata->mutex); + if (ret == EOWNERDEAD) + igt_assert_eq(pthread_mutex_consistent(&pdata->mutex), 0); + else + igt_assert_eq(ret, 0); + pdata->bind_oom++; + pthread_mutex_unlock(&pdata->mutex); + goto cleanup; + } + + if (bind_bos(fd_c, &ctx, sram_bufs, &sram_bind, &sram_bind_err, false)) { + ret = pthread_mutex_lock(&pdata->mutex); + if (ret == EOWNERDEAD) + igt_assert_eq(pthread_mutex_consistent(&pdata->mutex), 0); + else + igt_assert_eq(ret, 0); + pdata->bind_oom++; + pthread_mutex_unlock(&pdata->mutex); + goto cleanup; + } + + create_bo(fd_c, ctx.vm_id, &integers_bo, 4096, EXEC_DATA_ADDR, true); + expected_result = fill_random_integers(&integers_bo, INT_ADD_CNT); + igt_debug("%d\n", expected_result); + create_bo(fd_c, ctx.vm_id, &result_bo, 4096, EXEC_RESULT_ADDR, false); + create_bo(fd_c, ctx.vm_id, &batch_bo, 4096, EXEC_BATCH_ADDR, true); + pos = build_add_batch(&batch_bo, &integers_bo, &result_bo, INT_ADD_CNT); + + igt_assert(pos * sizeof(int) <= batch_bo.size); + + xe_vm_bind_lr_sync(fd_c, ctx.vm_id, integers_bo.handle, 0, integers_bo.addr, + integers_bo.size, 0); + xe_vm_bind_lr_sync(fd_c, ctx.vm_id, result_bo.handle, 0, result_bo.addr, + result_bo.size, 0); + xe_vm_bind_lr_sync(fd_c, ctx.vm_id, batch_bo.handle, 0, batch_bo.addr, + batch_bo.size, 0); + create_bo(fd_c, ctx.vm_id, &ufence_bo, 4096, EXEC_UFENCE_ADDR, true); + + memset(ufence_bo.ptr, 0, ufence_bo.size); + xe_vm_bind_lr_sync(fd_c, ctx.vm_id, ufence_bo.handle, 0, ufence_bo.addr, + ufence_bo.size, 0); + + batch_syncs[0] = (struct drm_xe_sync){ + .type = DRM_XE_SYNC_TYPE_USER_FENCE, + .flags = DRM_XE_SYNC_FLAG_SIGNAL, + .addr = ufence_bo.addr, + .timeline_value = USER_FENCE_VALUE, + }; + + exec = (struct drm_xe_exec) { + .exec_queue_id = ctx.exec_queue_id, + .num_syncs = 1, + .syncs = (uintptr_t)batch_syncs, + .address = batch_bo.addr, + .num_batch_buffer = 1, + }; + + rc = igt_ioctl(fd_c, DRM_IOCTL_XE_EXEC, &exec); + igt_assert_f(rc == 0, "xe_exec failed unexpectedly: %s (%d)\n", + strerror(errno), errno); + xe_wait_ufence(fd_c, (uint64_t *)ufence_bo.ptr, USER_FENCE_VALUE, ctx.exec_queue_id, + TIMEOUT_NS); + result_bo.ptr = xe_bo_map(fd_c, result_bo.handle, result_bo.size); + igt_assert(result_bo.ptr != MAP_FAILED); + igt_assert_eq(result_bo.ptr[0], expected_result); + ret = pthread_mutex_lock(&pdata->mutex); + if (ret == EOWNERDEAD) + igt_assert_eq(pthread_mutex_consistent(&pdata->mutex), 0); + else + igt_assert_eq(ret, 0); + pdata->child_completed++; + pthread_mutex_unlock(&pdata->mutex); +cleanup: + cleanup_bo_resources(fd_c, &ufence_bo); + cleanup_bo_resources(fd_c, &result_bo); + cleanup_bo_resources(fd_c, &batch_bo); + cleanup_bo_resources(fd_c, &integers_bo); + cleanup_sram_vram_objs(fd_c, &vram_bind, &sram_bind); + xe_exec_queue_destroy(fd_c, ctx.exec_queue_id); + xe_vm_destroy(fd_c, ctx.vm_id); + drm_close_driver(fd_c); + } + + release_ready_processes(pdata, p.n_proc); + /* Wait for 60 sec to kill the hanging child process */ + igt_waitchildren_timeout(60, "Over subscription child failed"); + igt_assert_f(pdata->child_completed >= 2, + "Oversubscription test requites atlread 2 processes"); + igt_reset_timeout(); + igt_info("Oversubscribe summary : %d processes\n", p.n_proc); + igt_info("Setup Faulted (unexpected error): %d\n", pdata->fault_count); + igt_info("Setup skipped (OOM cases) %d\n", pdata->failed - pdata->fault_count); + igt_info("Bind failed (ENOMEM/ENOSPC) %d\n", pdata->bind_oom); + igt_info("Successfully completd oversubscription process (exec + verify) %d\n", + pdata->child_completed); + pthread_cond_destroy(&pdata->cond); + pthread_mutex_destroy(&pdata->mutex); + igt_assert_eq(munmap(pdata, sizeof(*pdata)), 0); +} + static void alloc_sizes_init(void) { int i; @@ -978,13 +1576,12 @@ static void run_gt_dynamic_subtests(int fd, const struct test_case *t) * engine per invocation, so they're wired up as a dynamic child subtest * per engine below (igt_subtest_with_dynamic_f() + igt_dynamic_f()). * - * TYPE_THREAD / TYPE_PROCESS are deliberately NOT split per engine here: - * threads() and processes() already do their own xe_for_each_engine() to - * spawn one thread/process per engine internally. Wrapping those in a - * second, outer per-engine loop would run the full all-engine fan-out - * once for every engine on the machine, multiplying the runtime by the - * engine count and re-exercising every engine N times over. So these - * stay flat, single subtests. + * TYPE_THREAD are deliberately NOT split per engine here: + * threads() already do their own xe_for_each_engine() to spawn one thread per + * engine internally. Wrapping those in a second, outer per-engine loop would + * run the full all-engine fan-out once for every engine on the machine, + * multiplying the runtime by the engine count and re-exercising every engine + * N times over. So these stay flat, single subtests. */ static const struct test_case test_matrix[] = { @@ -1002,6 +1599,7 @@ static const struct test_case test_matrix[] = { { "threads-leak-binding-rand-sizes-50", 50, LEAK_BINDING, TYPE_THREAD, false }, { "threads-leak-binding-rand-sizes-50-unaligned", 50, LEAK_BINDING | UNALIGNED, TYPE_THREAD, false }, + { "test_vm_oversubscribe_concurrent_bind", 0, 0, TYPE_OVERSUBSCRIBE, false}, }; int igt_main() @@ -1027,6 +1625,10 @@ int igt_main() run_threaded_bo_alloc_test(fd, t->count, t->flags); } break; + case TYPE_OVERSUBSCRIBE: + igt_subtest_f("%s", t->name) + test_vm_oversubscribe_concurrent_bind(fd); + break; case TYPE_ALL_SIZES: case TYPE_RANDOM_SIZE: case TYPE_ARRAY_BIND: -- 2.55.0