[igt-dev] [PATCH i-g-t v2 3/4] tests/xe_evict_ccs: Add evict ccs test
Zbigniew Kempczyński
zbigniew.kempczynski at intel.com
Wed Oct 18 17:37:45 UTC 2023
On Tue, Oct 17, 2023 at 04:28:12PM +0100, Matthew Auld wrote:
> On Fri, 13 Oct 2023 at 12:31, Zbigniew Kempczyński
> <zbigniew.kempczynski at intel.com> wrote:
> >
> > Exercise is flat-ccs eviction working fine in the kernel driver
> > when buffers takes more than available vram. Differentiate with
> > standalone/parallel execution, same or separate drm fd and buffer
> > freeing time. Tests are divided to two groups - first which won't
> > exceed vram memory size (thus don't trigger eviction, but it is
> > good for the reference logic is properly compress/decompress
> > buffers) and second which exceeds.
> >
> > v2:
> > - Add command line switches to exercise kernel with different
> > sizes, number of objects and vram overcommitment
> > - Add -simple test which creates single big object which enforces
> > eviction (Matt)
> >
> > Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski at intel.com>
> > Cc: Matthew Auld <matthew.auld at intel.com>
> > ---
> > tests/intel/xe_evict_ccs.c | 522 +++++++++++++++++++++++++++++++++++++
> > tests/meson.build | 1 +
> > 2 files changed, 523 insertions(+)
> > create mode 100644 tests/intel/xe_evict_ccs.c
> >
> > diff --git a/tests/intel/xe_evict_ccs.c b/tests/intel/xe_evict_ccs.c
> > new file mode 100644
> > index 0000000000..72872f952b
> > --- /dev/null
> > +++ b/tests/intel/xe_evict_ccs.c
> > @@ -0,0 +1,522 @@
> > +// SPDX-License-Identifier: MIT
> > +/*
> > + * Copyright © 2023 Intel Corporation
> > + */
> > +
> > +/**
> > + * TEST: Check flat-ccs eviction
> > + * Category: Software building block
> > + * Sub-category: Flat-CCS
> > + * Functionality: evict
> > + * GPU requirements: GPU needs to have dedicated VRAM
> > + */
> > +
> > +#include "igt.h"
> > +#include "igt_list.h"
> > +#include "intel_blt.h"
> > +#include "intel_mocs.h"
> > +#include "lib/igt_syncobj.h"
> > +#include "lib/intel_reg.h"
> > +#include "xe_drm.h"
> > +
> > +#include "xe/xe_ioctl.h"
> > +#include "xe/xe_query.h"
> > +#include <math.h>
> > +#include <string.h>
> > +
> > +#define OVERCOMMIT_VRAM_PERCENT 110
> > +#define MIN_OBJ_KB 64
> > +#define MAX_OBJ_KB (256 * 1024)
> > +#define DUMP_FILENAME "/tmp/object.data"
> > +#define DUMP_EXPFILENAME "/tmp/object.expected"
> > +
> > +static struct param {
> > + bool print_bb;
> > + bool disable_compression;
> > + bool dump_corrupted_surface;
> > + int num_objs;
> > + int vram_percent;
> > + int min_size_kb;
> > + int max_size_kb;
> > + bool verify;
> > +} params = {
> > + .num_objs = 0,
> > + .vram_percent = OVERCOMMIT_VRAM_PERCENT,
> > + .min_size_kb = MIN_OBJ_KB,
> > + .max_size_kb = MAX_OBJ_KB,
> > +};
> > +
> > +struct object {
> > + uint64_t size;
> > + uint32_t start_value;
> > + struct blt_copy_object *blt_obj;
> > + struct igt_list_head link;
> > +};
> > +
> > +#define TEST_PARALLEL (1 << 0)
> > +#define TEST_INSTANTFREE (1 << 1)
> > +#define TEST_REOPEN (1 << 2)
> > +#define TEST_SIMPLE (1 << 3)
> > +
> > +#define MAX_NPROC 8
> > +struct config {
> > + uint32_t flags;
> > + int nproc;
> > + int free_mb, total_mb;
> > + int test_mb, mb_per_proc;
> > + const struct param *param;
> > +};
> > +
> > +static void copy_obj(struct blt_copy_data *blt,
> > + struct blt_copy_object *src_obj,
> > + struct blt_copy_object *dst_obj,
> > + intel_ctx_t *ctx,
> > + uint64_t ahnd)
> > +{
> > + struct blt_block_copy_data_ext ext = {};
> > + int fd = blt->fd;
> > + uint64_t bb_size = xe_get_default_alignment(fd);
> > + uint32_t bb;
> > + uint32_t w, h;
> > +
> > + w = src_obj->x2;
> > + h = src_obj->y2;
> > +
> > + bb = xe_bo_create_flags(fd, 0, bb_size,
> > + vram_memory(fd, 0) | XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM);
>
> Nit: we can use visible_vram_memory() here.
Previously I tried to limit the test to visible vram and
it passes whereas this one which works with full vram size
failed.
I missed in last uapi change series that vram_memory() now
may implicitly switch to system memory and this is likely
not we want.
>
> > +
> > + blt->color_depth = CD_32bit;
> > + blt->print_bb = params.print_bb;
> > + blt_set_copy_object(&blt->src, src_obj);
> > + blt_set_copy_object(&blt->dst, dst_obj);
> > + blt_set_object_ext(&ext.src, 0, w, h, SURFACE_TYPE_2D);
> > + blt_set_object_ext(&ext.dst, 0, w, h, SURFACE_TYPE_2D);
> > + blt_set_batch(&blt->bb, bb, bb_size, vram_if_possible(fd, 0));
> > + blt_block_copy(fd, ctx, NULL, ahnd, blt, &ext);
> > + intel_ctx_xe_sync(ctx, true);
> > +
> > + gem_close(fd, bb);
> > + put_offset(ahnd, bb);
> > + put_offset(ahnd, blt->src.handle);
> > + put_offset(ahnd, blt->dst.handle);
> > + intel_allocator_bind(ahnd, 0, 0);
> > +}
> > +
> > +static uint32_t rand_and_update(uint32_t *left, uint32_t min, uint32_t max)
> > +{
> > + int left_bit, min_bit, max_bit, rand_id, rand_kb;
> > +
> > + left_bit = igt_fls(*left) - 1;
> > + min_bit = igt_fls(min) - 1;
> > + max_bit = max_t(int, min_t(int, igt_fls(max) - 1, left_bit), igt_fls(max));
> > + rand_id = rand() % (max_bit - min_bit);
> > + rand_kb = 1 << (rand_id + min_bit);
> > +
> > + if (*left >= rand_kb)
> > + *left -= rand_kb;
> > + else
> > + *left = 0;
> > +
> > + return rand_kb;
> > +}
> > +
> > +static struct object *create_obj(struct blt_copy_data *blt,
> > + intel_ctx_t *ctx, uint64_t ahnd,
> > + uint64_t size, int start_value,
> > + bool disable_compression)
> > +{
> > + int fd = blt->fd;
> > + struct object *obj;
> > + uint32_t w, h;
> > + uint8_t uc_mocs = intel_get_uc_mocs_index(fd);
> > + int i;
> > + struct blt_copy_object *src;
> > +
> > + obj = calloc(1, sizeof(*obj));
> > + igt_assert(obj);
> > + obj->size = size;
> > + obj->start_value = start_value;
> > +
> > + w = max_t(int, 1024, roundup_power_of_two(sqrt(size/4)));
> > + h = size / w / 4; /* /4 - 32bpp */
> > +
> > + igt_debug("Obj size: %ldKiB (%ldMiB) <w: %d, h: %d>\n",
> > + size / SZ_1K, size / SZ_1M, w, h);
> > +
> > + src = blt_create_object(blt,
> > + system_memory(fd),
> > + w, h, 32, uc_mocs,
> > + T_LINEAR, COMPRESSION_DISABLED,
> > + COMPRESSION_TYPE_3D, true);
> > +
> > + obj->blt_obj = blt_create_object(blt,
> > + vram_memory(fd, 0) | XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM,
>
> Same here. I assume blt_create_object() knows to ignore the non-region
> bits when setting obj->region, or doesn't it really matter?
Eh, you're right. This bits should be get rid off before storing
on to obj->region. Thanks for spotting this, I'm going to fix this
and respin in v3.
>
> > + w, h, 32, uc_mocs,
> > + T_LINEAR,
> > + disable_compression ? COMPRESSION_DISABLED :
> > + COMPRESSION_ENABLED,
> > + COMPRESSION_TYPE_3D, true);
> > +
> > + for (i = 0; i < size / sizeof(uint32_t); i++)
> > + src->ptr[i] = start_value++;
> > +
> > + copy_obj(blt, src, obj->blt_obj, ctx, ahnd);
> > +
> > + blt_destroy_object_and_alloc_free(fd, ahnd, src);
> > + intel_allocator_bind(ahnd, 0, 0);
> > +
> > + return obj;
> > +}
> > +
> > +static void dump_obj(const struct blt_copy_object *obj, int start_value)
> > +{
> > + FILE *out;
> > +
> > + if (!params.dump_corrupted_surface)
> > + return;
> > +
> > + out = fopen(DUMP_FILENAME, "wb");
> > + fwrite(obj->ptr, obj->size, 1, out);
> > + fclose(out);
> > +
> > + out = fopen(DUMP_EXPFILENAME, "wb");
> > + for (int i = 0; i < obj->size / 4; i++) {
> > + int v = start_value + i;
> > +
> > + fwrite(&v, sizeof(int), 1, out);
> > + }
> > + fclose(out);
> > +}
> > +
> > +static void check_obj(const char *check_mode,
> > + const struct blt_copy_object *obj, uint64_t size,
> > + int start_value, int num_obj)
> > +{
> > + int i, idx;
> > +
> > + if (obj->ptr[0] != start_value ||
> > + (obj->ptr[size/4 - 1] != start_value + size/4 - 1)) {
> > + igt_info("[%s] Failed object w: %d, h: %d, size: %ldKiB (%ldMiB)\n",
> > + check_mode, obj->x2, obj->y2, obj->size / SZ_1K, obj->size / SZ_1M);
> > + dump_obj(obj, start_value);
> > + }
> > +
> > + igt_assert_eq(obj->ptr[0], start_value);
> > + igt_assert_eq(obj->ptr[size/4 - 1], start_value + size/4 - 1);
> > +
> > + /* Couple of checks of random indices */
> > + for (i = 0; i < 128; i++) {
> > + idx = rand() % (size/4);
> > +
> > + if (obj->ptr[idx] != start_value + idx) {
> > + igt_info("[%s] Failed object w: %d, h: %d, size: %ldKiB (%ldMiB)\n",
> > + check_mode, obj->x2, obj->y2,
> > + obj->size / SZ_1K, obj->size / SZ_1M);
> > + dump_obj(obj, start_value);
> > + }
> > +
> > + igt_assert_f(obj->ptr[idx] == start_value + idx,
> > + "[%s] Object number %d doesn't contain valid data",
> > + check_mode, num_obj);
> > + }
> > +}
> > +
> > +static void evict_single(int fd, int child, const struct config *config)
> > +{
> > + struct blt_copy_data blt = {};
> > + struct blt_copy_object *orig_obj;
> > + uint32_t kb_left = config->mb_per_proc * SZ_1K;
> > + uint32_t min_alloc_kb = config->param->min_size_kb;
> > + uint32_t max_alloc_kb = config->param->max_size_kb;
> > + uint32_t vm = xe_vm_create(fd, DRM_XE_VM_CREATE_ASYNC_BIND_OPS, 0);
> > + uint64_t ahnd = intel_allocator_open(fd, vm, INTEL_ALLOCATOR_RELOC);
> > + uint8_t uc_mocs = intel_get_uc_mocs_index(fd);
> > + struct object *obj, *tmp;
> > + struct igt_list_head list;
> > + struct drm_xe_engine_class_instance inst = {
> > + .engine_class = DRM_XE_ENGINE_CLASS_COPY,
> > + };
> > + intel_ctx_t *ctx;
> > + uint32_t exec_queue, big_obj;
> > + int num_obj = 0;
> > +
> > + srandom(time(NULL));
> > + IGT_INIT_LIST_HEAD(&list);
> > + igt_debug("[%2d] child : to allocate: %uMiB\n", child, kb_left/SZ_1K);
> > +
> > + blt_copy_init(fd, &blt);
> > +
> > + exec_queue = xe_exec_queue_create(fd, vm, &inst, 0);
> > + ctx = intel_ctx_xe(fd, vm, exec_queue, 0, 0, 0);
> > +
> > + while (kb_left) {
> > + struct blt_copy_object *verify_obj;
> > + uint64_t obj_size = rand_and_update(&kb_left, min_alloc_kb, max_alloc_kb) * SZ_1K;
> > + int start_value = rand();
> > +
> > + if (config->flags & TEST_SIMPLE)
> > + obj_size = max_alloc_kb * SZ_1K;
> > +
> > + obj = create_obj(&blt, ctx, ahnd, obj_size, start_value,
> > + config->param->disable_compression);
> > + igt_list_add(&obj->link, &list);
> > +
> > + if (config->param->verify) {
> > + verify_obj = blt_create_object(&blt, system_memory(fd),
> > + obj->blt_obj->x2,
> > + obj->blt_obj->y2,
> > + 32, uc_mocs,
> > + T_LINEAR, COMPRESSION_DISABLED,
> > + 0, true);
> > + copy_obj(&blt, obj->blt_obj, verify_obj, ctx, ahnd);
> > + check_obj("Verify", verify_obj, obj->blt_obj->size,
> > + obj->start_value, num_obj++);
> > + blt_destroy_object_and_alloc_free(fd, ahnd, verify_obj);
> > + intel_allocator_bind(ahnd, 0, 0);
> > + }
> > +
> > + if (config->flags & TEST_SIMPLE) {
> > + big_obj = xe_bo_create_flags(fd, vm, kb_left * SZ_1K,
> > + vram_memory(fd, 0));
> > + break;
> > + }
> > +
> > + if (config->param->num_objs && ++num_obj == config->param->num_objs)
> > + break;
> > + }
> > +
> > + if (config->param->verify)
> > + igt_info("[%8d] Verify ok\n", getpid());
> > +
> > + num_obj = 0;
> > + igt_list_for_each_entry_safe(obj, tmp, &list, link) {
> > + orig_obj = blt_create_object(&blt, system_memory(fd),
> > + obj->blt_obj->x2,
> > + obj->blt_obj->y2,
> > + 32, uc_mocs,
> > + T_LINEAR, COMPRESSION_DISABLED,
> > + 0, true);
> > + copy_obj(&blt, obj->blt_obj, orig_obj, ctx, ahnd);
> > + check_obj("Check", orig_obj, obj->blt_obj->size, obj->start_value, num_obj++);
> > + blt_destroy_object_and_alloc_free(fd, ahnd, orig_obj);
> > +
> > + if (config->flags & TEST_INSTANTFREE) {
> > + igt_list_del(&obj->link);
> > + blt_destroy_object_and_alloc_free(fd, ahnd, obj->blt_obj);
> > + free(obj);
> > + }
> > + intel_allocator_bind(ahnd, 0, 0);
> > + }
> > +
> > + if (!(config->flags & TEST_INSTANTFREE))
> > + igt_list_for_each_entry_safe(obj, tmp, &list, link) {
> > + igt_list_del(&obj->link);
> > + blt_destroy_object_and_alloc_free(fd, ahnd, obj->blt_obj);
> > + free(obj);
> > + }
> > +
> > + if (config->flags & TEST_SIMPLE)
> > + gem_close(fd, big_obj);
> > +}
> > +
> > +static void set_config(int fd, uint32_t flags, const struct param *param,
> > + struct config *config)
> > +{
> > + int nproc = 1;
> > +
> > + config->param = param;
> > + config->flags = flags;
> > + config->free_mb = xe_vram_available(fd, 0) / SZ_1M;
>
> Should this be visible avail?
That was something weird here, I'm going to recheck this.
Thanks for the review.
--
Zbigniew
>
> > + config->total_mb = xe_visible_vram_size(fd, 0) / SZ_1M;
> > + config->test_mb = min_t(int, config->free_mb * config->param->vram_percent / 100,
> > + config->total_mb * config->param->vram_percent / 100);
> > +
> > + igt_debug("VRAM memory size: %dMB/%dMB (use %dMB), overcommit perc: %d\n",
> > + config->free_mb, config->total_mb,
> > + config->test_mb, config->param->vram_percent);
> > +
> > + if (flags & TEST_PARALLEL)
> > + nproc = min_t(int, sysconf(_SC_NPROCESSORS_ONLN), MAX_NPROC);
> > + config->nproc = nproc;
> > + config->mb_per_proc = config->test_mb / nproc;
> > +
> > + igt_debug("nproc: %d, mem per proc: %dMB\n", nproc, config->mb_per_proc);
> > +}
> > +
> > +static void evict_ccs(int fd, uint32_t flags, const struct param *param)
> > +{
> > + struct config config;
> > + char numstr[32];
> > +
> > + igt_info("Test mode <parallel: %d, instant free: %d, reopen: %d, simple: %d>\n",
> > + !!(flags & TEST_PARALLEL),
> > + !!(flags & TEST_INSTANTFREE),
> > + !!(flags & TEST_REOPEN),
> > + !!(flags & TEST_SIMPLE));
> > + if (param->num_objs)
> > + snprintf(numstr, sizeof(numstr), "%d", param->num_objs);
> > + else
> > + strncpy(numstr, "limited to vram", sizeof(numstr));
> > + igt_info("Params: compression: %s, num objects: %s, vram percent: %d, kb <min: %d, max: %d>\n",
> > + param->disable_compression ? "disabled" : "enabled",
> > + numstr, param->vram_percent,
> > + param->min_size_kb, param->max_size_kb);
> > +
> > + set_config(fd, flags, param, &config);
> > +
> > + if (flags & TEST_PARALLEL) {
> > + igt_fork(n, config.nproc) {
> > + if (flags & TEST_REOPEN) {
> > + fd = drm_reopen_driver(fd);
> > + intel_allocator_init();
> > + }
> > + evict_single(fd, n, &config);
> > + }
> > + igt_waitchildren();
> > + } else {
> > + if (flags & TEST_REOPEN)
> > + fd = drm_reopen_driver(fd);
> > + evict_single(fd, 0, &config);
> > + }
> > +}
> > +
> > +/**
> > + *
> > + * SUBTEST: evict-ccs-overcommit-simple
> > + * Description: FlatCCS eviction test.
> > + * Feature: flatccs
> > + * Test category: stress test
> > + */
> > +/**
> > + *
> > + * SUBTEST: evict-ccs-overcommit-%s-%s-%s
> > + * Description: FlatCCS eviction test.
> > + * Feature: flatccs
> > + * Test category: stress test
> > + *
> > + * arg[1]:
> > + *
> > + * @standalone: single process
> > + * @parallel: multiple processes
> > + *
> > + * arg[2]:
> > + *
> > + * @nofree: keep objects till the end of the test
> > + * @instantfree: free object after it was verified and it won't
> > + * be used anymore
> > + *
> > + * arg[3]:
> > + *
> > + * @samefd: operate on same opened drm fd
> > + * @reopen: use separately opened drm fds
> > + *
> > + */
> > +static int opt_handler(int opt, int opt_index, void *data)
> > +{
> > + switch (opt) {
> > + case 'b':
> > + params.print_bb = true;
> > + igt_debug("Print bb: %d\n", params.print_bb);
> > + break;
> > + case 'd':
> > + params.disable_compression = true;
> > + igt_debug("Print bb: %d\n", params.disable_compression);
> > + break;
> > + case 'D':
> > + params.dump_corrupted_surface = true;
> > + igt_debug("Print bb: %d\n", params.dump_corrupted_surface);
> > + break;
> > + case 'n':
> > + params.num_objs = atoi(optarg);
> > + igt_debug("Number objects: %d\n", params.num_objs);
> > + break;
> > + case 'p':
> > + params.vram_percent = atoi(optarg);
> > + igt_debug("Percent vram: %d\n", params.vram_percent);
> > + break;
> > + case 's':
> > + params.min_size_kb = atoi(optarg);
> > + igt_debug("Min size kb: %d\n", params.min_size_kb);
> > + break;
> > + case 'S':
> > + params.max_size_kb = atoi(optarg);
> > + igt_debug("Max size kb: %d\n", params.max_size_kb);
> > + break;
> > + case 'V':
> > + params.verify = true;
> > + igt_debug("Verify: %d\n", params.verify);
> > + break;
> > + default:
> > + return IGT_OPT_HANDLER_ERROR;
> > + }
> > +
> > + return IGT_OPT_HANDLER_SUCCESS;
> > +}
> > +
> > +const char *help_str =
> > + " -b\tPrint bb\n"
> > + " -d\tDisable compression (don't use flatccs area)\n"
> > + " -D\tDump surface which doesn't match\n"
> > + " -e\tAdd temporary object which enforce eviction\n"
> > + " -n\tNumber of objects to create (0 - 31)\n"
> > + " -p\tPercent of VRAM to alloc\n"
> > + " -s\tMinimum size of object in kb\n"
> > + " -S\tMaximum size of object in kb\n"
> > + " -V\tVerify object after compressing\n"
> > + ;
> > +
> > +igt_main_args("bdDn:p:s:S:V", NULL, help_str, opt_handler, NULL)
> > +{
> > + struct drm_xe_engine_class_instance *hwe;
> > +
> > + const struct ccs {
> > + const char *name;
> > + uint32_t flags;
> > + } ccs[] = {
> > + { "simple",
> > + TEST_SIMPLE },
> > + { "standalone-nofree-samefd",
> > + 0 },
> > + { "standalone-nofree-reopen",
> > + TEST_REOPEN },
> > + { "standalone-instantfree-samefd",
> > + TEST_INSTANTFREE },
> > + { "standalone-instantfree-reopen",
> > + TEST_INSTANTFREE | TEST_REOPEN },
> > + { "parallel-nofree-samefd",
> > + TEST_PARALLEL },
> > + { "parallel-nofree-reopen",
> > + TEST_PARALLEL | TEST_REOPEN },
> > + { "parallel-instantfree-samefd",
> > + TEST_PARALLEL | TEST_INSTANTFREE },
> > + { "parallel-instantfree-reopen",
> > + TEST_PARALLEL | TEST_INSTANTFREE | TEST_REOPEN },
> > + { },
> > + };
> > + uint64_t vram_size;
> > + int fd;
> > +
> > + igt_fixture {
> > + fd = drm_open_driver(DRIVER_XE);
> > + igt_require(xe_has_vram(fd));
> > + vram_size = xe_visible_vram_size(fd, 0);
> > + igt_assert(vram_size);
> > +
> > + xe_for_each_hw_engine(fd, hwe)
> > + if (hwe->engine_class != DRM_XE_ENGINE_CLASS_COPY)
> > + break;
> > + }
> > +
> > + igt_fixture
> > + intel_allocator_multiprocess_start();
> > +
> > + for (const struct ccs *s = ccs; s->name; s++) {
> > + igt_subtest_f("evict-ccs-overcommit-%s", s->name)
> > + evict_ccs(fd, s->flags, ¶ms);
> > + }
> > +
> > + igt_fixture {
> > + intel_allocator_multiprocess_stop();
> > + drm_close_driver(fd);
> > + }
> > +}
> > diff --git a/tests/meson.build b/tests/meson.build
> > index 2c2e1ca9ac..831947567f 100644
> > --- a/tests/meson.build
> > +++ b/tests/meson.build
> > @@ -279,6 +279,7 @@ intel_xe_progs = [
> > 'xe_debugfs',
> > 'xe_drm_fdinfo',
> > 'xe_evict',
> > + 'xe_evict_ccs',
> > 'xe_exec_balancer',
> > 'xe_exec_basic',
> > 'xe_exec_compute_mode',
> > --
> > 2.34.1
> >
More information about the igt-dev
mailing list