rooty/open-gpu-kernel-modules
1
0
Fork 0
mirror of https://github.com/NVIDIA/open-gpu-kernel-modules.git synced 2024-12-12 21:08:56 +01:00
Code Issues Projects Releases Wiki Activity
c700e8f91c
open-gpu-kernel-modules/kernel-open/nvidia-uvm/uvm_pmm_test.c
Andy Ritger 1739a20efc
515.43.04
2022-05-09 13:18:59 -07:00

1450 lines
51 KiB
C
Raw Blame History

/*******************************************************************************
Copyright (c) 2015-2022 NVIDIA Corporation
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to
deal in the Software without restriction, including without limitation the
rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
sell copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be
included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
DEALINGS IN THE SOFTWARE.
*******************************************************************************/
#include "uvm_common.h"
#include "uvm_pmm_gpu.h"
#include "uvm_global.h"
#include "uvm_gpu.h"
#include "uvm_hal.h"
#include "uvm_va_block.h"
#include "uvm_va_range.h"
#include "uvm_va_space.h"
#include "uvm_tracker.h"
#include "uvm_push.h"
#include "uvm_mem.h"
#include "uvm_kvmalloc.h"
#include "uvm_test.h"
#include "uvm_test_ioctl.h"
#include "uvm_test_rng.h"
#define CHUNKS_PER_BUCKET 128
typedef struct
{
struct list_head entry;
uvm_gpu_chunk_t *chunks[CHUNKS_PER_BUCKET];
} pmm_leak_bucket_t;
typedef struct
{
uvm_gpu_chunk_t *chunk;
uvm_tracker_t tracker;
NvU32 pattern;
struct list_head node;
} test_chunk_t;
// When the basic_test free_pattern is BASIC_TEST_FREE_PATTERN_EVERY_N, this
// controls how many allocs to do before a free.
#define BASIC_TEST_FREE_EVERY_N 3
// Number of allocations to make in part of basic_test. This is 33 because
// that's a decent balance between the widest gap between chunk levels (causing
// us to fill up at least one root chunk), and 33*UVM_CHUNK_SIZE_MAX isn't too
// big.
#define BASIC_TEST_STATIC_ALLOCATIONS 33
typedef enum
{
BASIC_TEST_FREE_PATTERN_IMMEDIATE,
BASIC_TEST_FREE_PATTERN_ALL_FORWARD,
BASIC_TEST_FREE_PATTERN_ALL_REVERSE,
BASIC_TEST_FREE_PATTERN_EVERY_N,
BASIC_TEST_FREE_PATTERN_COUNT
} basic_test_free_pattern_t;
typedef struct
{
// List of all allocated test_chunk_t's
struct list_head list;
// Total number of chunks allocated in this test
size_t num_chunks_total;
uvm_va_space_t *va_space;
uvm_pmm_gpu_t *pmm;
uvm_mem_t *verif_mem;
uvm_pmm_gpu_memory_type_t type;
basic_test_free_pattern_t free_pattern;
} basic_test_state_t;
typedef enum
{
SPLIT_TEST_MODE_NORMAL,
SPLIT_TEST_MODE_MERGE,
SPLIT_TEST_MODE_INJECT_ERROR,
SPLIT_TEST_MODE_COUNT
} split_test_mode_t;
// This helper needs to stay in sync with the one in uvm_pmm_gpu.c
// It is duplicated because we do not want to expose it as an API.
static uvm_pmm_gpu_memory_type_t pmm_squash_memory_type(uvm_parent_gpu_t *parent_gpu, uvm_pmm_gpu_memory_type_t type)
{
return type;
}
// Verify that the input chunks are in the correct state following alloc
static NV_STATUS check_chunks(uvm_pmm_gpu_t *pmm,
uvm_gpu_chunk_t **chunks,
size_t num_chunks,
uvm_chunk_size_t chunk_size,
uvm_pmm_gpu_memory_type_t mem_type)
{
uvm_gpu_t *gpu = uvm_pmm_to_gpu(pmm);
size_t i;
mem_type = pmm_squash_memory_type(gpu->parent, mem_type);
for (i = 0; i < num_chunks; i++) {
TEST_CHECK_RET(chunks[i]);
TEST_CHECK_RET(chunks[i]->suballoc == NULL);
TEST_CHECK_RET(chunks[i]->type == mem_type);
TEST_CHECK_RET(chunks[i]->state == UVM_PMM_GPU_CHUNK_STATE_TEMP_PINNED);
TEST_CHECK_RET(uvm_gpu_chunk_get_size(chunks[i]) == chunk_size);
TEST_CHECK_RET(IS_ALIGNED(chunks[i]->address, chunk_size));
}
return NV_OK;
}
static NV_STATUS check_alloc_tracker(uvm_pmm_gpu_t *pmm, uvm_tracker_t *tracker)
{
uvm_gpu_t *gpu = uvm_pmm_to_gpu(pmm);
uvm_tracker_entry_t *tracker_entry;
// The tracker entries returned from an alloc are not allowed to contain
// entries for any GPU other than the owner. This is to prevent leaking
// tracker entries from other GPUs into VA spaces which never registered
// those GPUs.
for_each_tracker_entry(tracker_entry, tracker)
TEST_CHECK_RET(uvm_tracker_entry_gpu(tracker_entry) == gpu);
return NV_OK;
}
static NV_STATUS chunk_alloc_check_common(uvm_pmm_gpu_t *pmm,
size_t num_chunks,
uvm_chunk_size_t chunk_size,
uvm_pmm_gpu_memory_type_t mem_type,
uvm_pmm_alloc_flags_t flags,
uvm_gpu_chunk_t **chunks,
uvm_tracker_t *local_tracker,
uvm_tracker_t *tracker)
{
NV_STATUS status;
NV_STATUS check_status;
check_status = check_alloc_tracker(pmm, local_tracker);
if (tracker) {
status = uvm_tracker_add_tracker_safe(tracker, local_tracker);
uvm_tracker_clear(local_tracker);
}
else {
status = uvm_tracker_wait(local_tracker);
}
uvm_tracker_deinit(local_tracker);
if (check_status == NV_OK)
check_status = status;
if (check_status != NV_OK)
return check_status;
return check_chunks(pmm, chunks, num_chunks, chunk_size, mem_type);
}
static NV_STATUS chunk_alloc_check(uvm_pmm_gpu_t *pmm,
size_t num_chunks,
uvm_chunk_size_t chunk_size,
uvm_pmm_gpu_memory_type_t mem_type,
uvm_pmm_alloc_flags_t flags,
uvm_gpu_chunk_t **chunks,
uvm_tracker_t *tracker)
{
uvm_gpu_t *gpu = uvm_pmm_to_gpu(pmm);
NV_STATUS status;
uvm_tracker_t local_tracker = UVM_TRACKER_INIT();
// If the GPU has no memory then PMA will fail with NV_ERR_INVALID_ARGUMENT.
// Callers assert that only NV_OK or NV_ERR_NO_MEMORY are returned here, so
// replace the error to avoid noise.
if (gpu->mem_info.size == 0)
return NV_ERR_NO_MEMORY;
status = uvm_pmm_gpu_alloc(pmm, num_chunks, chunk_size, mem_type, flags, chunks, &local_tracker);
if (status != NV_OK)
return status;
return chunk_alloc_check_common(pmm, num_chunks, chunk_size, mem_type, flags, chunks, &local_tracker, tracker);
}
static NV_STATUS chunk_alloc_user_check(uvm_pmm_gpu_t *pmm,
size_t num_chunks,
uvm_chunk_size_t chunk_size,
uvm_pmm_gpu_memory_type_t mem_type,
uvm_pmm_alloc_flags_t flags,
uvm_gpu_chunk_t **chunks,
uvm_tracker_t *tracker)
{
NV_STATUS status;
uvm_tracker_t local_tracker = UVM_TRACKER_INIT();
status = uvm_pmm_gpu_alloc(pmm, num_chunks, chunk_size, mem_type, flags, chunks, &local_tracker);
if (status != NV_OK)
return status;
return chunk_alloc_check_common(pmm, num_chunks, chunk_size, mem_type, flags, chunks, &local_tracker, tracker);
}
static NV_STATUS check_leak(uvm_gpu_t *gpu, uvm_chunk_size_t chunk_size, uvm_pmm_gpu_memory_type_t type, NvS64 limit, NvU64 *chunks)
{
NV_STATUS status = NV_OK;
pmm_leak_bucket_t *bucket, *next;
LIST_HEAD(allocations);
*chunks = 0;
while (limit != *chunks) {
int k;
pmm_leak_bucket_t *allocated;
allocated = kzalloc(sizeof(pmm_leak_bucket_t), GFP_KERNEL);
if (allocated == NULL) {
status = NV_ERR_NO_MEMORY;
goto cleanup;
}
list_add(&allocated->entry, &allocations);
for (k = 0; k < CHUNKS_PER_BUCKET && limit != *chunks; k++) {
status = chunk_alloc_check(&gpu->pmm,
1,
chunk_size,
type,
UVM_PMM_ALLOC_FLAGS_NONE,
&allocated->chunks[k],
NULL);
UVM_ASSERT(status == NV_OK || status == NV_ERR_NO_MEMORY);
if (status != NV_OK) {
if (limit == -1 && status == NV_ERR_NO_MEMORY)
status = NV_OK;
goto cleanup;
}
(*chunks)++;
if (fatal_signal_pending(current)) {
status = NV_ERR_SIGNAL_PENDING;
goto cleanup;
}
}
}
cleanup:
list_for_each_entry_safe(bucket, next, &allocations, entry) {
int k;
for (k = 0; k < CHUNKS_PER_BUCKET; k++) {
if (!bucket->chunks[k])
break;
uvm_pmm_gpu_free(&gpu->pmm, bucket->chunks[k], NULL);
}
list_del(&bucket->entry);
kfree(bucket);
}
return status;
}
// Tracker is an in/out dependency
static NV_STATUS do_memset_4(uvm_gpu_t *gpu, uvm_gpu_address_t dst, NvU32 val, size_t size, uvm_tracker_t *tracker)
{
NV_STATUS status;
uvm_push_t push;
status = uvm_push_begin_acquire(gpu->channel_manager,
UVM_CHANNEL_TYPE_GPU_INTERNAL,
tracker,
&push,
"memset {%s, 0x%llx} %zu bytes to 0x%08x",
uvm_gpu_address_aperture_string(dst),
dst.address,
size,
val);
if (status != NV_OK)
return status;
gpu->parent->ce_hal->memset_4(&push, dst, val, size);
uvm_push_end(&push);
uvm_tracker_overwrite_with_push(tracker, &push);
return NV_OK;
}
// Tracker is an in/out dependency
static NV_STATUS gpu_mem_check(uvm_gpu_t *gpu,
uvm_mem_t *verif_mem,
uvm_gpu_address_t src,
size_t size,
NvU32 expected,
uvm_tracker_t *tracker)
{
NV_STATUS status;
uvm_push_t push;
uvm_gpu_address_t verif_gpu_addr;
NvU32 *verif_cpu_addr = uvm_mem_get_cpu_addr_kernel(verif_mem);
size_t i;
UVM_ASSERT(verif_mem->size >= size);
memset(verif_cpu_addr, 0, size);
status = uvm_push_begin_acquire(gpu->channel_manager,
UVM_CHANNEL_TYPE_GPU_TO_CPU,
tracker,
&push,
"GPU -> CPU {%s, 0x%llx} %zu bytes expecting 0x%08x",
uvm_gpu_address_aperture_string(src),
src.address,
size,
expected);
if (status != NV_OK)
return status;
verif_gpu_addr = uvm_mem_gpu_address_virtual_kernel(verif_mem, gpu);
gpu->parent->ce_hal->memcopy(&push, verif_gpu_addr, src, size);
TEST_NV_CHECK_RET(uvm_push_end_and_wait(&push));
for (i = 0; i < size / sizeof(verif_cpu_addr[0]); i++) {
if (verif_cpu_addr[i] != expected) {
UVM_TEST_PRINT("GPU read of {%s, 0x%llx} %zu bytes expected pattern 0x%08x, but offset %zu is 0x%08x\n",
uvm_gpu_address_aperture_string(src),
src.address,
size,
expected,
i * sizeof(verif_cpu_addr[0]),
verif_cpu_addr[i]);
return NV_ERR_INVALID_STATE;
}
}
return NV_OK;
}
static NV_STATUS init_test_chunk(uvm_va_space_t *va_space,
uvm_pmm_gpu_t *pmm,
test_chunk_t *test_chunk,
uvm_pmm_gpu_memory_type_t type,
uvm_chunk_size_t size,
NvU32 pattern)
{
uvm_gpu_t *gpu = uvm_pmm_to_gpu(pmm);
NV_STATUS status = NV_OK;
uvm_push_t push;
uvm_gpu_address_t chunk_addr;
uvm_gpu_t *other_gpu;
INIT_LIST_HEAD(&test_chunk->node);
uvm_tracker_init(&test_chunk->tracker);
test_chunk->pattern = pattern;
MEM_NV_CHECK_RET(chunk_alloc_check(pmm, 1, size, type, UVM_PMM_ALLOC_FLAGS_EVICT, &test_chunk->chunk, &test_chunk->tracker), NV_OK);
TEST_NV_CHECK_GOTO(uvm_mmu_chunk_map(test_chunk->chunk), chunk_free);
if (uvm_mmu_gpu_needs_static_vidmem_mapping(gpu) || uvm_mmu_gpu_needs_dynamic_vidmem_mapping(gpu))
chunk_addr = uvm_gpu_address_virtual_from_vidmem_phys(gpu, test_chunk->chunk->address);
else
chunk_addr = uvm_gpu_address_physical(UVM_APERTURE_VID, test_chunk->chunk->address);
// Fill the chunk
TEST_NV_CHECK_GOTO(do_memset_4(gpu, chunk_addr, pattern, size, &test_chunk->tracker), chunk_unmap);
// Launch dummy pushes on all other GPUs. This will increase the chances of
// a subsequent re-alloc of this chunk needing to synchronize the tracker.
// See the tracker comment in check_alloc_tracker.
for_each_va_space_gpu(other_gpu, va_space) {
if (other_gpu == gpu)
continue;
status = uvm_push_begin(other_gpu->channel_manager,
UVM_CHANNEL_TYPE_MEMOPS,
&push,
"dummy push for chunk {%s, %u} with 0x%08x",
uvm_pmm_gpu_memory_type_string(type),
size,
pattern);
TEST_NV_CHECK_GOTO(status, chunk_unmap);
other_gpu->parent->host_hal->noop(&push, 4);
uvm_push_end(&push);
TEST_NV_CHECK_GOTO(uvm_tracker_add_push_safe(&test_chunk->tracker, &push), chunk_unmap);
}
return NV_OK;
chunk_unmap:
uvm_mmu_chunk_unmap(test_chunk->chunk, &test_chunk->tracker);
chunk_free:
uvm_pmm_gpu_free(pmm, test_chunk->chunk, &test_chunk->tracker);
uvm_tracker_deinit(&test_chunk->tracker);
return status;
}
static NV_STATUS destroy_test_chunk(uvm_pmm_gpu_t *pmm, test_chunk_t *test_chunk, uvm_mem_t *verif_mem)
{
uvm_gpu_t *gpu = uvm_pmm_to_gpu(pmm);
NV_STATUS status;
uvm_gpu_address_t chunk_addr;
uvm_gpu_chunk_t *chunk = test_chunk->chunk;
uvm_chunk_size_t size = uvm_gpu_chunk_get_size(chunk);
if (uvm_mmu_gpu_needs_static_vidmem_mapping(gpu) || uvm_mmu_gpu_needs_dynamic_vidmem_mapping(gpu))
chunk_addr = uvm_gpu_address_virtual_from_vidmem_phys(gpu, chunk->address);
else
chunk_addr = uvm_gpu_address_physical(UVM_APERTURE_VID, chunk->address);
status = gpu_mem_check(gpu, verif_mem, chunk_addr, size, test_chunk->pattern, &test_chunk->tracker);
list_del(&test_chunk->node);
uvm_mmu_chunk_unmap(chunk, &test_chunk->tracker);
uvm_pmm_gpu_free(pmm, chunk, &test_chunk->tracker);
uvm_tracker_deinit(&test_chunk->tracker);
return status;
}
static bool basic_test_should_free(basic_test_state_t *test_state)
{
if (test_state->free_pattern == BASIC_TEST_FREE_PATTERN_IMMEDIATE)
return true;
return test_state->free_pattern == BASIC_TEST_FREE_PATTERN_EVERY_N &&
(test_state->num_chunks_total % BASIC_TEST_FREE_EVERY_N) == 0;
}
static NV_STATUS basic_test_alloc(basic_test_state_t *test_state, uvm_chunk_size_t size)
{
test_chunk_t *test_chunk;
NvU32 pattern;
NV_STATUS status = NV_OK;
test_chunk = uvm_kvmalloc_zero(sizeof(*test_chunk));
if (!test_chunk) {
UVM_TEST_PRINT("Failed to allocate test_chunk\n");
return NV_ERR_NO_MEMORY;
}
pattern = current->pid | (test_state->num_chunks_total << 16);
status = init_test_chunk(test_state->va_space, test_state->pmm, test_chunk, test_state->type, size, pattern);
if (status != NV_OK) {
uvm_kvfree(test_chunk);
return status;
}
list_add_tail(&test_chunk->node, &test_state->list);
++test_state->num_chunks_total;
if (basic_test_should_free(test_state)) {
test_chunk = list_first_entry(&test_state->list, test_chunk_t, node);
status = destroy_test_chunk(test_state->pmm, test_chunk, test_state->verif_mem);
uvm_kvfree(test_chunk);
}
return status;
}
static NV_STATUS basic_test_free_all(basic_test_state_t *test_state)
{
test_chunk_t *test_chunk;
NV_STATUS temp_status, status = NV_OK;
while (!list_empty(&test_state->list)) {
if (test_state->free_pattern == BASIC_TEST_FREE_PATTERN_ALL_REVERSE)
test_chunk = list_last_entry(&test_state->list, test_chunk_t, node);
else // Handles cleanup and BASIC_TEST_FREE_PATTERN_ALL_FORWARD
test_chunk = list_first_entry(&test_state->list, test_chunk_t, node);
temp_status = destroy_test_chunk(test_state->pmm, test_chunk, test_state->verif_mem);
if (status == NV_OK)
status = temp_status;
uvm_kvfree(test_chunk);
}
return status;
}
// Try to allocate enough smaller chunks to fully fill the largest chunk, plus
// a little extra.
static size_t basic_test_num_allocations(uvm_chunk_size_t size)
{
return (UVM_CHUNK_SIZE_MAX / size) + 1;
}
// - Allocate multiple chunks of all possible sizes and types using various
// patterns
// - Write a unique value to each chunk
// - Free those chunks in various patterns, verifying the unique value
static NV_STATUS basic_test(uvm_va_space_t *va_space, uvm_gpu_t *gpu,
UvmTestPmmSanityMode mode)
{
uvm_chunk_size_t size;
uvm_chunk_sizes_mask_t chunk_sizes;
basic_test_state_t test_state;
NV_STATUS status = NV_OK; // Implicitly modified by TEST_NV_CHECK_GOTO
size_t i;
int first_memory_type, last_memory_type;
int first_free_pattern, last_free_pattern;
if (mode == UvmTestPmmSanityModeBasic) {
first_memory_type = UVM_PMM_GPU_MEMORY_TYPE_USER;
last_memory_type = UVM_PMM_GPU_MEMORY_TYPE_USER;
first_free_pattern = BASIC_TEST_FREE_PATTERN_EVERY_N;
last_free_pattern = BASIC_TEST_FREE_PATTERN_EVERY_N;
}
else {
first_memory_type = 0;
last_memory_type = UVM_PMM_GPU_MEMORY_TYPE_COUNT - 1;
first_free_pattern = 0;
last_free_pattern = BASIC_TEST_FREE_PATTERN_COUNT - 1;
}
// Note that we can't really test PMM in isolation, since even pushing work
// to the GPU requires using PMM to create GPU page tables for the
// pushbuffers. We could handle that in theory by forcing sysmem page
// tables, but that would require re-allocating the entire GPU address
// space.
memset(&test_state, 0, sizeof(test_state));
INIT_LIST_HEAD(&test_state.list);
test_state.va_space = va_space;
test_state.pmm = &gpu->pmm;
MEM_NV_CHECK_RET(uvm_mem_alloc_sysmem_and_map_cpu_kernel(UVM_CHUNK_SIZE_MAX, current->mm, &test_state.verif_mem),
NV_OK);
TEST_NV_CHECK_GOTO(uvm_mem_map_gpu_kernel(test_state.verif_mem, gpu), out);
for (test_state.type = first_memory_type; test_state.type <= last_memory_type; test_state.type++) {
// In SR-IOV heavy, virtual mappings will be created for each chunk
// this test allocates before accessing it on the GPU. But currently
// it is not allowed to map a kernel chunk, so skip those.
if (uvm_gpu_is_virt_mode_sriov_heavy(gpu) && uvm_pmm_gpu_memory_type_is_kernel(test_state.type))
continue;
chunk_sizes = gpu->pmm.chunk_sizes[test_state.type];
for (test_state.free_pattern = first_free_pattern;
test_state.free_pattern <= last_free_pattern;
test_state.free_pattern++) {
// Outer loop over size, increasing
size = uvm_chunk_find_first_size(chunk_sizes);
for_each_chunk_size_from(size, chunk_sizes) {
for (i = 0; i < basic_test_num_allocations(size); i++)
TEST_NV_CHECK_GOTO(basic_test_alloc(&test_state, size), out);
}
TEST_NV_CHECK_GOTO(basic_test_free_all(&test_state), out);
// Outer loop over size, decreasing
size = uvm_chunk_find_last_size(chunk_sizes);
for_each_chunk_size_rev_from(size, chunk_sizes) {
for (i = 0; i < basic_test_num_allocations(size); i++)
TEST_NV_CHECK_GOTO(basic_test_alloc(&test_state, size), out);
}
TEST_NV_CHECK_GOTO(basic_test_free_all(&test_state), out);
// Inner loop over size, increasing
for (i = 0; i < BASIC_TEST_STATIC_ALLOCATIONS; i++) {
size = uvm_chunk_find_first_size(chunk_sizes);
for_each_chunk_size_from(size, chunk_sizes)
TEST_NV_CHECK_GOTO(basic_test_alloc(&test_state, size), out);
}
TEST_NV_CHECK_GOTO(basic_test_free_all(&test_state), out);
// Inner loop over size, decreasing
for (i = 0; i < BASIC_TEST_STATIC_ALLOCATIONS; i++) {
size = uvm_chunk_find_last_size(chunk_sizes);
for_each_chunk_size_rev_from(size, chunk_sizes)
TEST_NV_CHECK_GOTO(basic_test_alloc(&test_state, size), out);
}
TEST_NV_CHECK_GOTO(basic_test_free_all(&test_state), out);
}
}
out:
if (status != NV_OK)
basic_test_free_all(&test_state);
UVM_ASSERT(list_empty(&test_state.list));
uvm_mem_free(test_state.verif_mem);
return status;
}
static NV_STATUS get_subchunks_test(uvm_pmm_gpu_t *pmm,
uvm_gpu_chunk_t *parent,
uvm_gpu_chunk_t **expected_children,
size_t num_children)
{
uvm_gpu_chunk_t **subchunks = NULL;
NV_STATUS status = NV_OK;
size_t count, start_index, size = num_children * sizeof(subchunks[0]);
subchunks = uvm_kvmalloc(size);
if (!subchunks) {
UVM_TEST_PRINT("Failed to allocate subchunks\n");
return NV_ERR_NO_MEMORY;
}
// Verify all
memset(subchunks, 0, size);
TEST_CHECK_GOTO(uvm_pmm_gpu_get_subchunks(pmm, parent, 0, num_children, subchunks) == num_children, out);
TEST_CHECK_GOTO(memcmp(expected_children, subchunks, num_children * sizeof(subchunks[0])) == 0, out);
// Get first half
count = num_children / 2;
memset(subchunks, 0, size);
TEST_CHECK_GOTO(uvm_pmm_gpu_get_subchunks(pmm, parent, 0, count, subchunks) == count, out);
TEST_CHECK_GOTO(memcmp(expected_children, subchunks, count * sizeof(subchunks[0])) == 0, out);
// Get second half, intentionally requesting more subchunks than available
start_index = num_children / 2;
count = num_children - start_index;
memset(subchunks, 0, size);
TEST_CHECK_GOTO(uvm_pmm_gpu_get_subchunks(pmm, parent, start_index, num_children, subchunks) == count, out);
TEST_CHECK_GOTO(memcmp(&expected_children[start_index], subchunks, count * sizeof(subchunks[0])) == 0, out);
// Larger-than-possible start_index
TEST_CHECK_GOTO(uvm_pmm_gpu_get_subchunks(pmm, parent, num_children, 1, subchunks) == 0, out);
out:
uvm_kvfree(subchunks);
return status;
}
// Always frees parent chunk, even on error return
static NV_STATUS split_test_single(uvm_pmm_gpu_t *pmm,
test_chunk_t *parent,
uvm_chunk_size_t child_size,
split_test_mode_t mode,
uvm_mem_t *verif_mem)
{
uvm_pmm_gpu_memory_type_t parent_type = parent->chunk->type;
uvm_chunk_size_t parent_size = uvm_gpu_chunk_get_size(parent->chunk);
NvU64 parent_addr = parent->chunk->address;
size_t i, num_children = (size_t)(parent_size / child_size);
uvm_gpu_chunk_t **split_chunks = NULL;
uvm_gpu_chunk_t *temp_chunk;
test_chunk_t child_wrapper;
NV_STATUS temp_status, status = NV_OK;
// Verify that we can get "subchunks" of a non-split chunk
TEST_CHECK_RET(uvm_pmm_gpu_get_subchunks(pmm, parent->chunk, 0, 2, &temp_chunk) == 1);
TEST_CHECK_RET(temp_chunk == parent->chunk);
split_chunks = uvm_kvmalloc(num_children * sizeof(split_chunks[0]));
if (!split_chunks) {
UVM_TEST_PRINT("Failed to allocate split_chunks\n");
status = NV_ERR_NO_MEMORY;
goto error;
}
if (mode == SPLIT_TEST_MODE_INJECT_ERROR)
parent->chunk->inject_split_error = true;
status = uvm_pmm_gpu_split_chunk(pmm, parent->chunk, child_size, split_chunks);
if (mode == SPLIT_TEST_MODE_INJECT_ERROR) {
// This case verifies that a split failure will leave the chunk in its
// original state.
if (status != NV_ERR_NO_MEMORY) {
UVM_TEST_PRINT("Injecting split error failed, returned %s\n", nvstatusToString(status));
status = NV_ERR_INVALID_STATE;
// Let the error label clean up the split children
parent->chunk = NULL;
goto error;
}
status = destroy_test_chunk(pmm, parent, verif_mem);
}
else {
TEST_NV_CHECK_GOTO(status, error);
temp_chunk = parent->chunk;
parent->chunk = NULL;
// Sanity check split
for (i = 0; i < num_children; i++) {
TEST_CHECK_GOTO(split_chunks[i], error);
TEST_CHECK_GOTO(split_chunks[i]->address == parent_addr + i * child_size, error);
TEST_CHECK_GOTO(split_chunks[i]->suballoc == NULL, error);
TEST_CHECK_GOTO(split_chunks[i]->type == parent_type, error);
TEST_CHECK_GOTO(split_chunks[i]->state == UVM_PMM_GPU_CHUNK_STATE_TEMP_PINNED, error);
TEST_CHECK_GOTO(uvm_gpu_chunk_get_size(split_chunks[i]) == child_size, error);
}
status = get_subchunks_test(pmm, temp_chunk, split_chunks, num_children);
if (status != NV_OK)
goto error;
if (mode == SPLIT_TEST_MODE_MERGE) {
parent->chunk = temp_chunk;
uvm_pmm_gpu_merge_chunk(pmm, parent->chunk);
TEST_CHECK_GOTO(parent->chunk->address == parent_addr, error);
TEST_CHECK_GOTO(parent->chunk->suballoc == NULL, error);
TEST_CHECK_GOTO(parent->chunk->state == UVM_PMM_GPU_CHUNK_STATE_TEMP_PINNED, error);
status = destroy_test_chunk(pmm, parent, verif_mem);
}
else {
// Destroy split chunks, verifying the original pattern
for (i = 0; i < num_children; i++) {
child_wrapper.chunk = split_chunks[i];
child_wrapper.pattern = parent->pattern;
temp_status = uvm_tracker_init_from(&child_wrapper.tracker, &parent->tracker);
if (status == NV_OK)
status = temp_status;
// destroy_test_chunk does list_del
INIT_LIST_HEAD(&child_wrapper.node);
temp_status = destroy_test_chunk(pmm, &child_wrapper, verif_mem);
if (status == NV_OK)
status = temp_status;
}
uvm_tracker_deinit(&parent->tracker);
}
}
uvm_kvfree(split_chunks);
return status;
error:
if (parent->chunk) {
uvm_mmu_chunk_unmap(parent->chunk, &parent->tracker);
uvm_pmm_gpu_free(pmm, parent->chunk, &parent->tracker);
}
else {
for (i = 0; i < num_children; i++) {
uvm_mmu_chunk_unmap(split_chunks[i], &parent->tracker);
uvm_pmm_gpu_free(pmm, split_chunks[i], &parent->tracker);
}
}
uvm_kvfree(split_chunks);
return status;
}
// Splits each possible non-leaf chunk size into all possible sizes below that
// size, and verifies that the data in the chunk remains intact.
static NV_STATUS split_test(uvm_va_space_t *va_space, uvm_gpu_t *gpu)
{
uvm_pmm_gpu_memory_type_t type;
uvm_chunk_size_t parent_size, child_size;
NvU32 pattern;
NvU32 count = 0;
test_chunk_t parent_test_chunk;
NV_STATUS status = NV_OK;
uvm_mem_t *verif_mem = NULL;
split_test_mode_t mode;
// Check the num_subchunks == 0 case
TEST_CHECK_RET(uvm_pmm_gpu_get_subchunks(&gpu->pmm, NULL, 0, 0, NULL) == 0);
MEM_NV_CHECK_RET(uvm_mem_alloc_sysmem_and_map_cpu_kernel(UVM_CHUNK_SIZE_MAX, current->mm, &verif_mem), NV_OK);
TEST_NV_CHECK_GOTO(uvm_mem_map_gpu_kernel(verif_mem, gpu), out);
for (type = 0; type < UVM_PMM_GPU_MEMORY_TYPE_COUNT; type++) {
// In SR-IOV heavy, virtual mappings will be created for each chunk
// this test allocates before accessing it on the GPU. But currently
// it is not allowed to map a kernel chunk, so skip those.
if (uvm_gpu_is_virt_mode_sriov_heavy(gpu) && uvm_pmm_gpu_memory_type_is_kernel(type))
continue;
// Test every available parent size except the smallest, which obviously
// can't be split.
parent_size = uvm_chunk_find_next_size(gpu->pmm.chunk_sizes[type],
uvm_chunk_find_first_size(gpu->pmm.chunk_sizes[type]));
for_each_chunk_size_from(parent_size, gpu->pmm.chunk_sizes[type]) {
// Split from parent_size to every smaller supported size
child_size = uvm_chunk_find_prev_size(gpu->pmm.chunk_sizes[type], parent_size);
for_each_chunk_size_rev_from(child_size, gpu->pmm.chunk_sizes[type]) {
for (mode = 0; mode < SPLIT_TEST_MODE_COUNT; mode++) {
pattern = current->pid | (count << 16);
++count;
status = init_test_chunk(va_space, &gpu->pmm, &parent_test_chunk, type, parent_size, pattern);
if (status != NV_OK)
goto out;
status = split_test_single(&gpu->pmm, &parent_test_chunk, child_size, mode, verif_mem);
if (status != NV_OK)
goto out;
}
}
}
}
out:
uvm_mem_free(verif_mem);
return status;
}
NV_STATUS uvm_test_pmm_query(UVM_TEST_PMM_QUERY_PARAMS *params, struct file *filp)
{
uvm_va_space_t *va_space = uvm_va_space_get(filp);
NV_STATUS status = NV_OK;
uvm_gpu_t *gpu;
gpu = uvm_va_space_retain_gpu_by_uuid(va_space, &params->gpu_uuid);
if (!gpu)
return NV_ERR_INVALID_DEVICE;
switch (params->key) {
case UVM_TEST_CHUNK_SIZE_GET_USER_SIZE:
params->value = gpu->pmm.chunk_sizes[UVM_PMM_GPU_MEMORY_TYPE_USER];
status = NV_OK;
break;
default:
status = NV_ERR_INVALID_ARGUMENT;
break;
}
uvm_gpu_release(gpu);
return status;
}
NV_STATUS uvm_test_pmm_sanity(UVM_TEST_PMM_SANITY_PARAMS *params, struct file *filp)
{
NV_STATUS status = NV_OK;
uvm_va_space_t *va_space = uvm_va_space_get(filp);
uvm_gpu_t *gpu;
if (params->mode != UvmTestPmmSanityModeBasic &&
params->mode != UvmTestPmmSanityModeFull) {
return NV_ERR_INVALID_ARGUMENT;
}
uvm_va_space_down_read(va_space);
for_each_va_space_gpu(gpu, va_space) {
status = basic_test(va_space, gpu, params->mode);
if (status != NV_OK)
goto out;
status = split_test(va_space, gpu);
if (status != NV_OK)
goto out;
}
out:
uvm_va_space_up_read(va_space);
return status;
}
NV_STATUS uvm_test_pmm_check_leak(UVM_TEST_PMM_CHECK_LEAK_PARAMS *params, struct file *filp)
{
uvm_va_space_t *va_space = uvm_va_space_get(filp);
NV_STATUS status = NV_OK;
uvm_gpu_t *gpu;
uvm_pmm_gpu_memory_type_t first_user_mode = UVM_PMM_GPU_MEMORY_TYPE_USER;
uvm_pmm_gpu_memory_type_t last_user_mode = UVM_PMM_GPU_MEMORY_TYPE_USER;
uvm_pmm_gpu_memory_type_t current_user_mode = first_user_mode;
if (params->alloc_limit < -1)
return NV_ERR_INVALID_ARGUMENT;
gpu = uvm_va_space_retain_gpu_by_uuid(va_space, &params->gpu_uuid);
if (!gpu)
return NV_ERR_INVALID_DEVICE;
for (; current_user_mode <= last_user_mode; current_user_mode++) {
status = check_leak(gpu, params->chunk_size, current_user_mode, params->alloc_limit, &params->allocated);
if (status != NV_OK)
break;
}
uvm_gpu_release(gpu);
return status;
}
NV_STATUS __test_pmm_async_alloc_type(uvm_va_space_t *va_space,
uvm_gpu_t *gpu,
size_t num_chunks,
uvm_pmm_gpu_memory_type_t mem_type,
size_t work_iterations)
{
NV_STATUS status;
NV_STATUS tracker_status = NV_OK;
uvm_gpu_chunk_t **chunks;
uvm_chunk_size_t chunk_size = PAGE_SIZE;
uvm_gpu_t *work_gpu;
uvm_mem_t *dummy_buffer = NULL;
uvm_mem_alloc_params_t mem_params;
uvm_tracker_t tracker = UVM_TRACKER_INIT();
uvm_push_t push;
NvU32 i;
uvm_global_processor_mask_t global_mask;
chunks = uvm_kvmalloc_zero(num_chunks * sizeof(chunks[0]));
if (!chunks)
return NV_ERR_NO_MEMORY;
memset(&mem_params, 0, sizeof(mem_params));
mem_params.backing_gpu = NULL;
mem_params.size = 1024*1024;
mem_params.mm = current->mm;
status = uvm_mem_alloc(&mem_params, &dummy_buffer);
if (status != NV_OK)
goto out;
uvm_va_space_global_gpus(va_space, &global_mask);
status = uvm_mem_map_kernel(dummy_buffer, &global_mask);
if (status != NV_OK)
goto out;
// Alloc lots of small chunks to trigger suballocation
status = chunk_alloc_user_check(&gpu->pmm,
num_chunks,
chunk_size,
mem_type,
UVM_PMM_ALLOC_FLAGS_NONE,
chunks,
&tracker);
if (status != NV_OK)
goto out;
// Push long-running work on all GPUs and collect it all in the tracker
for (i = 0; i < work_iterations; i++) {
if (fatal_signal_pending(current)) {
status = NV_ERR_SIGNAL_PENDING;
goto out;
}
for_each_va_space_gpu(work_gpu, va_space) {
// Acquire the prior iteration just to make things even slower
status = uvm_push_begin_acquire(work_gpu->channel_manager,
UVM_CHANNEL_TYPE_GPU_INTERNAL,
&tracker,
&push,
"memset");
if (status != NV_OK)
goto out;
work_gpu->parent->ce_hal->memset_1(&push,
uvm_mem_gpu_address_virtual_kernel(dummy_buffer, work_gpu),
0,
mem_params.size);
uvm_push_end(&push);
TEST_NV_CHECK_GOTO(uvm_tracker_add_push_safe(&tracker, &push), out);
}
}
// Free every other chunk to keep the suballocation around
for (i = 0; i < num_chunks; i += 2) {
uvm_pmm_gpu_free(&gpu->pmm, chunks[i], &tracker);
chunks[i] = NULL;
}
// Re-alloc chunks to verify that the returned trackers don't have work for
// other GPUs (chunk_alloc_user_check() checks that).
for (i = 0; i < num_chunks; i += 2) {
status = chunk_alloc_user_check(&gpu->pmm,
1,
chunk_size,
mem_type,
UVM_PMM_ALLOC_FLAGS_NONE,
&chunks[i],
NULL);
if (status != NV_OK)
goto out;
}
out:
if (chunks) {
for (i = 0; i < num_chunks; i++) {
if (chunks[i])
uvm_pmm_gpu_free(&gpu->pmm, chunks[i], &tracker);
}
}
tracker_status = uvm_tracker_wait_deinit(&tracker);
uvm_mem_free(dummy_buffer);
uvm_kvfree(chunks);
return status == NV_OK ? tracker_status : status;
}
NV_STATUS uvm_test_pmm_async_alloc(UVM_TEST_PMM_ASYNC_ALLOC_PARAMS *params, struct file *filp)
{
NV_STATUS status = NV_OK;
uvm_va_space_t *va_space = uvm_va_space_get(filp);
uvm_gpu_t *gpu;
uvm_pmm_gpu_memory_type_t first_user_mode = UVM_PMM_GPU_MEMORY_TYPE_USER;
uvm_pmm_gpu_memory_type_t last_user_mode = UVM_PMM_GPU_MEMORY_TYPE_USER;
uvm_pmm_gpu_memory_type_t current_user_mode = first_user_mode;
uvm_va_space_down_read(va_space);
gpu = uvm_va_space_get_gpu_by_uuid(va_space, &params->gpu_uuid);
if (!gpu) {
uvm_va_space_up_read(va_space);
return NV_ERR_INVALID_DEVICE;
}
for (; current_user_mode <= last_user_mode; current_user_mode++) {
status = __test_pmm_async_alloc_type(va_space,
gpu,
params->num_chunks,
current_user_mode,
params->num_work_iterations);
if (status != NV_OK)
break;
}
uvm_va_space_up_read(va_space);
return status;
}
static uvm_reverse_map_t g_reverse_map_entries[PAGES_PER_UVM_VA_BLOCK * 4];
static NV_STATUS test_pmm_reverse_map_single(uvm_gpu_t *gpu, uvm_va_space_t *va_space, NvU64 addr)
{
NV_STATUS status = NV_OK;
NvU32 num_translations;
uvm_va_block_t *va_block;
uvm_gpu_phys_address_t phys_addr;
bool is_resident;
status = uvm_va_block_find(va_space, addr, &va_block);
if (status != NV_OK)
return status;
TEST_CHECK_RET(uvm_va_block_size(va_block) == UVM_VA_BLOCK_SIZE);
// Verify that all pages are populated on the GPU
uvm_mutex_lock(&va_block->lock);
is_resident = uvm_processor_mask_test(&va_block->resident, gpu->id) &&
uvm_page_mask_full(uvm_va_block_resident_mask_get(va_block, gpu->id));
if (is_resident)
phys_addr = uvm_va_block_gpu_phys_page_address(va_block, 0, gpu);
uvm_mutex_unlock(&va_block->lock);
TEST_CHECK_RET(is_resident);
// In this test a single VA range covers the whole 2MB physical region. We
// expect a single translation to be returned for a 2MB chunk.
num_translations = uvm_pmm_gpu_phys_to_virt(&gpu->pmm, phys_addr.address, UVM_VA_BLOCK_SIZE, g_reverse_map_entries);
TEST_CHECK_RET(num_translations == 1);
TEST_CHECK_RET(g_reverse_map_entries[0].va_block == va_block);
TEST_CHECK_RET(g_reverse_map_entries[0].region.first == 0);
TEST_CHECK_RET(uvm_va_block_region_num_pages(g_reverse_map_entries[0].region) == uvm_va_block_num_cpu_pages(va_block));
uvm_va_block_release(va_block);
return NV_OK;
}
static NV_STATUS test_pmm_reverse_map_many_blocks(uvm_gpu_t *gpu, uvm_va_space_t *va_space, NvU64 addr, NvU64 size)
{
uvm_va_range_t *va_range;
uvm_va_block_t *va_block = NULL;
NvU32 num_blocks;
NvU32 index = 0;
uvm_gpu_phys_address_t phys_addr = {0};
bool is_resident;
// In this test, the [addr:addr + size) VA region contains
// several VA ranges with different sizes.
// Find the first block to compute the base physical address of the root
// chunk
uvm_for_each_va_range_in(va_range, va_space, addr, addr + size - 1) {
va_block = uvm_va_range_block(va_range, 0);
if (va_block)
break;
}
TEST_CHECK_RET(va_block);
uvm_mutex_lock(&va_block->lock);
is_resident = uvm_id_equal(uvm_va_block_page_get_closest_resident(va_block, 0, gpu->id), gpu->id);
if (is_resident) {
phys_addr = uvm_va_block_gpu_phys_page_address(va_block, 0, gpu);
phys_addr.address = UVM_ALIGN_DOWN(phys_addr.address, UVM_VA_BLOCK_SIZE);
}
uvm_mutex_unlock(&va_block->lock);
TEST_CHECK_RET(is_resident);
// Perform the lookup for the whole root chunk
num_blocks = uvm_pmm_gpu_phys_to_virt(&gpu->pmm, phys_addr.address, size, g_reverse_map_entries);
TEST_CHECK_RET(num_blocks != 0);
// Iterate over all VA ranges and their VA blocks within the 2MB VA region.
// Some blocks are not populated. However, we assume that blocks have been
// populated in order so they have been assigned physical addresses
// incrementally. Therefore, the reverse translations will show them in
// order.
uvm_for_each_va_range_in(va_range, va_space, addr, addr + size - 1) {
uvm_va_block_t *va_block;
for_each_va_block_in_va_range(va_range, va_block) {
NvU32 num_va_block_pages = 0;
// Iterate over all the translations for the current VA block. One
// translation per chunk is returned. We compute the total number of
// pages covered in the translations to check that they match with
// the number of pages in the VA block.
while (g_reverse_map_entries[index].va_block == va_block) {
uvm_reverse_map_t *reverse_mapping;
reverse_mapping = &g_reverse_map_entries[index];
uvm_va_block_release(va_block);
num_va_block_pages += uvm_va_block_region_num_pages(reverse_mapping->region);
UVM_ASSERT(uvm_va_block_contains_address(va_block, uvm_reverse_map_start(reverse_mapping)));
UVM_ASSERT(uvm_va_block_contains_address(va_block, uvm_reverse_map_end(reverse_mapping)));
uvm_mutex_lock(&va_block->lock);
// Verify that all pages are populated on the GPU
is_resident = uvm_page_mask_region_full(uvm_va_block_resident_mask_get(va_block, gpu->id),
reverse_mapping->region);
uvm_mutex_unlock(&va_block->lock);
TEST_CHECK_RET(is_resident);
++index;
}
if (num_va_block_pages)
TEST_CHECK_RET(num_va_block_pages == uvm_va_block_num_cpu_pages(va_block));
}
}
TEST_CHECK_RET(index == num_blocks);
return NV_OK;
}
NV_STATUS uvm_test_pmm_reverse_map(UVM_TEST_PMM_REVERSE_MAP_PARAMS *params, struct file *filp)
{
NV_STATUS status;
uvm_gpu_t *gpu;
uvm_va_space_t *va_space;
va_space = uvm_va_space_get(filp);
// Take the global lock to void interferences from different instances of
// the test, since we use global variables
uvm_mutex_lock(&g_uvm_global.global_lock);
uvm_va_space_down_write(va_space);
gpu = uvm_va_space_get_gpu_by_uuid(va_space, &params->gpu_uuid);
if (!gpu || !uvm_processor_mask_test(&va_space->registered_gpus, gpu->id)) {
status = NV_ERR_INVALID_DEVICE;
goto exit_unlock;
}
status = test_pmm_reverse_map_single(gpu, va_space, params->range_address1);
if (status == NV_OK)
status = test_pmm_reverse_map_many_blocks(gpu, va_space, params->range_address2, params->range_size2);
exit_unlock:
uvm_va_space_up_write(va_space);
uvm_mutex_unlock(&g_uvm_global.global_lock);
return status;
}
static NV_STATUS test_indirect_peers(uvm_gpu_t *owning_gpu, uvm_gpu_t *accessing_gpu)
{
uvm_pmm_gpu_t *pmm = &owning_gpu->pmm;
size_t chunk_size = uvm_chunk_find_first_size(pmm->chunk_sizes[UVM_PMM_GPU_MEMORY_TYPE_USER]);
uvm_gpu_chunk_t *parent_chunk = NULL;
uvm_gpu_chunk_t **chunks = NULL;
size_t i, num_chunks = UVM_CHUNK_SIZE_MAX / chunk_size;
NV_STATUS tracker_status, status = NV_OK;
uvm_mem_t *verif_mem = NULL;
uvm_tracker_t tracker = UVM_TRACKER_INIT();
uvm_gpu_address_t local_addr;
uvm_gpu_address_t peer_addr;
NvU32 init_val = 0x12345678;
NvU32 new_val = 0xabcdc0de;
chunks = uvm_kvmalloc_zero(num_chunks * sizeof(chunks[0]));
if (!chunks)
return NV_ERR_NO_MEMORY;
TEST_NV_CHECK_GOTO(uvm_mem_alloc_sysmem_and_map_cpu_kernel(UVM_CHUNK_SIZE_MAX, current->mm, &verif_mem), out);
TEST_NV_CHECK_GOTO(uvm_mem_map_gpu_kernel(verif_mem, owning_gpu), out);
TEST_NV_CHECK_GOTO(uvm_mem_map_gpu_kernel(verif_mem, accessing_gpu), out);
// Allocate a root chunk then split it to test multiple mappings across
// contiguous chunks under the same root.
TEST_NV_CHECK_GOTO(uvm_pmm_gpu_alloc_user(pmm,
1,
UVM_CHUNK_SIZE_MAX,
UVM_PMM_ALLOC_FLAGS_EVICT,
&parent_chunk,
NULL), out);
TEST_NV_CHECK_GOTO(uvm_pmm_gpu_split_chunk(pmm, parent_chunk, chunk_size, chunks), out);
parent_chunk = NULL;
// Verify contiguity and multiple mappings under a root chunk
for (i = 0; i < num_chunks; i++) {
TEST_NV_CHECK_GOTO(uvm_pmm_gpu_indirect_peer_map(pmm, chunks[i], accessing_gpu), out);
TEST_CHECK_GOTO(uvm_pmm_gpu_indirect_peer_addr(pmm, chunks[i], accessing_gpu) ==
uvm_pmm_gpu_indirect_peer_addr(pmm, chunks[0], accessing_gpu) + i * chunk_size,
out);
}
// Check that accessing_gpu can read and write
local_addr = uvm_gpu_address_physical(UVM_APERTURE_VID, chunks[0]->address);
peer_addr = uvm_pmm_gpu_peer_copy_address(&owning_gpu->pmm, chunks[0], accessing_gpu);
// Init on local GPU
TEST_NV_CHECK_GOTO(do_memset_4(owning_gpu, local_addr, init_val, UVM_CHUNK_SIZE_MAX, &tracker), out);
// Read using indirect peer and verify
TEST_NV_CHECK_GOTO(gpu_mem_check(accessing_gpu,
verif_mem,
peer_addr,
UVM_CHUNK_SIZE_MAX,
init_val,
&tracker), out);
// Write from indirect peer
TEST_NV_CHECK_GOTO(do_memset_4(accessing_gpu, peer_addr, new_val, UVM_CHUNK_SIZE_MAX, &tracker), out);
// Read using local gpu and verify
TEST_NV_CHECK_GOTO(gpu_mem_check(owning_gpu, verif_mem, local_addr, UVM_CHUNK_SIZE_MAX, new_val, &tracker), out);
out:
tracker_status = uvm_tracker_wait_deinit(&tracker);
if (status == NV_OK && tracker_status != NV_OK) {
UVM_TEST_PRINT("Tracker wait failed\n");
status = tracker_status;
}
if (parent_chunk) {
uvm_pmm_gpu_free(pmm, parent_chunk, NULL);
}
else {
for (i = 0; i < num_chunks; i++) {
if (chunks[i])
uvm_pmm_gpu_free(pmm, chunks[i], NULL);
}
}
if (verif_mem)
uvm_mem_free(verif_mem);
uvm_kvfree(chunks);
return status;
}
NV_STATUS uvm_test_pmm_indirect_peers(UVM_TEST_PMM_INDIRECT_PEERS_PARAMS *params, struct file *filp)
{
NV_STATUS status = NV_OK;
uvm_va_space_t *va_space = uvm_va_space_get(filp);
uvm_gpu_t *owning_gpu, *accessing_gpu;
bool ran_test = false;
uvm_va_space_down_read(va_space);
for_each_va_space_gpu(owning_gpu, va_space) {
for_each_va_space_gpu_in_mask(accessing_gpu, va_space, &va_space->indirect_peers[uvm_id_value(owning_gpu->id)]) {
ran_test = true;
status = test_indirect_peers(owning_gpu, accessing_gpu);
if (status != NV_OK)
goto out;
}
}
if (!ran_test)
status = NV_WARN_NOTHING_TO_DO;
out:
uvm_va_space_up_read(va_space);
return status;
}
static NV_STATUS test_chunk_with_elevated_page(uvm_gpu_t *gpu)
{
uvm_pmm_gpu_t *pmm = &gpu->pmm;
size_t chunk_size = uvm_chunk_find_first_size(pmm->chunk_sizes[UVM_PMM_GPU_MEMORY_TYPE_USER]);
uvm_gpu_chunk_t *parent_chunk = NULL, *parent_root = NULL;
uvm_gpu_chunk_t **chunks = NULL;
uvm_gpu_chunk_t *new_chunk = NULL;
size_t i, num_chunks = UVM_CHUNK_SIZE_MAX / chunk_size;
NV_STATUS status = NV_OK;
struct page *page = NULL;
chunks = uvm_kvmalloc_zero(num_chunks * sizeof(chunks[0]));
if (!chunks)
return NV_ERR_NO_MEMORY;
// Allocate a root chunk then split it to test multiple mappings across
// contiguous chunks under the same root.
TEST_NV_CHECK_GOTO(uvm_pmm_gpu_alloc_user(pmm,
1,
UVM_CHUNK_SIZE_MAX,
UVM_PMM_ALLOC_FLAGS_EVICT,
&parent_chunk,
NULL), out);
// Keep an extra reference to just one page within the parent chunk.
// This will make the whole root chunk non-allocatable.
page = uvm_gpu_chunk_to_page(pmm, parent_chunk);
get_page(page);
TEST_NV_CHECK_GOTO(uvm_pmm_gpu_split_chunk(pmm, parent_chunk, chunk_size, chunks), out);
parent_root = parent_chunk;
parent_chunk = NULL;
// Free some of the chunks
for (i = 0; i < num_chunks/2; i++) {
UVM_ASSERT(chunks[i]);
uvm_pmm_gpu_free(pmm, chunks[i], NULL);
chunks[i] = NULL;
}
// Now try alloc a chunk of that size.
// Expect that the allocation will fail or return a chunk with a
// different parent.
status = chunk_alloc_check(pmm,
1,
chunk_size,
UVM_PMM_GPU_MEMORY_TYPE_USER,
UVM_PMM_ALLOC_FLAGS_NONE,
&new_chunk,
NULL);
UVM_ASSERT(status == NV_OK || status == NV_ERR_NO_MEMORY);
if (status == NV_OK)
TEST_CHECK_GOTO(!uvm_gpu_chunk_same_root(new_chunk, parent_root), out);
else if (status == NV_ERR_NO_MEMORY)
status = NV_OK;
for (i = num_chunks/2; i < num_chunks; i++) {
UVM_ASSERT(chunks[i]);
uvm_pmm_gpu_free(pmm, chunks[i], NULL);
chunks[i] = NULL;
}
out:
if (parent_chunk) {
uvm_pmm_gpu_free(pmm, parent_chunk, NULL);
}
else {
for (i = 0; i < num_chunks; i++) {
if (chunks[i])
uvm_pmm_gpu_free(pmm, chunks[i], NULL);
}
}
if (new_chunk)
uvm_pmm_gpu_free(pmm, new_chunk, NULL);
if (page)
put_page(page);
uvm_kvfree(chunks);
return status;
}
NV_STATUS uvm_test_pmm_chunk_with_elevated_page(UVM_TEST_PMM_CHUNK_WITH_ELEVATED_PAGE_PARAMS *params, struct file *filp)
{
NV_STATUS status = NV_OK;
uvm_va_space_t *va_space = uvm_va_space_get(filp);
uvm_gpu_t *gpu;
bool ran_test = false;
uvm_va_space_down_read(va_space);
for_each_va_space_gpu(gpu, va_space) {
if (!gpu->parent->numa_info.enabled)
continue;
ran_test = true;
status = test_chunk_with_elevated_page(gpu);
if (status != NV_OK)
goto out;
}
if (!ran_test)
status = NV_WARN_NOTHING_TO_DO;
out:
uvm_va_space_up_read(va_space);
return status;
}
Reference in New Issue View Git Blame Copy Permalink