| 1 | /* |
| 2 | * Memory helpers |
| 3 | */ |
| 4 | #include <sys/types.h> |
| 5 | #include <sys/stat.h> |
| 6 | #include <fcntl.h> |
| 7 | #include <unistd.h> |
| 8 | #include <sys/mman.h> |
| 9 | |
| 10 | #include "fio.h" |
| 11 | #ifndef FIO_NO_HAVE_SHM_H |
| 12 | #include <sys/shm.h> |
| 13 | #endif |
| 14 | |
| 15 | void fio_unpin_memory(struct thread_data *td) |
| 16 | { |
| 17 | if (td->pinned_mem) { |
| 18 | dprint(FD_MEM, "unpinning %llu bytes\n", td->o.lockmem); |
| 19 | if (munlock(td->pinned_mem, td->o.lockmem) < 0) |
| 20 | perror("munlock"); |
| 21 | munmap(td->pinned_mem, td->o.lockmem); |
| 22 | td->pinned_mem = NULL; |
| 23 | } |
| 24 | } |
| 25 | |
| 26 | int fio_pin_memory(struct thread_data *td) |
| 27 | { |
| 28 | unsigned long long phys_mem; |
| 29 | |
| 30 | if (!td->o.lockmem) |
| 31 | return 0; |
| 32 | |
| 33 | dprint(FD_MEM, "pinning %llu bytes\n", td->o.lockmem); |
| 34 | |
| 35 | /* |
| 36 | * Don't allow mlock of more than real_mem-128MiB |
| 37 | */ |
| 38 | phys_mem = os_phys_mem(); |
| 39 | if (phys_mem) { |
| 40 | if ((td->o.lockmem + 128 * 1024 * 1024) > phys_mem) { |
| 41 | td->o.lockmem = phys_mem - 128 * 1024 * 1024; |
| 42 | log_info("fio: limiting mlocked memory to %lluMiB\n", |
| 43 | td->o.lockmem >> 20); |
| 44 | } |
| 45 | } |
| 46 | |
| 47 | td->pinned_mem = mmap(NULL, td->o.lockmem, PROT_READ | PROT_WRITE, |
| 48 | MAP_PRIVATE | OS_MAP_ANON, -1, 0); |
| 49 | if (td->pinned_mem == MAP_FAILED) { |
| 50 | perror("malloc locked mem"); |
| 51 | td->pinned_mem = NULL; |
| 52 | return 1; |
| 53 | } |
| 54 | if (mlock(td->pinned_mem, td->o.lockmem) < 0) { |
| 55 | perror("mlock"); |
| 56 | munmap(td->pinned_mem, td->o.lockmem); |
| 57 | td->pinned_mem = NULL; |
| 58 | return 1; |
| 59 | } |
| 60 | |
| 61 | return 0; |
| 62 | } |
| 63 | |
| 64 | static int alloc_mem_shm(struct thread_data *td, unsigned int total_mem) |
| 65 | { |
| 66 | #ifndef CONFIG_NO_SHM |
| 67 | int flags = IPC_CREAT | S_IRUSR | S_IWUSR; |
| 68 | |
| 69 | if (td->o.mem_type == MEM_SHMHUGE) { |
| 70 | unsigned long mask = td->o.hugepage_size - 1; |
| 71 | |
| 72 | flags |= SHM_HUGETLB; |
| 73 | total_mem = (total_mem + mask) & ~mask; |
| 74 | } |
| 75 | |
| 76 | td->shm_id = shmget(IPC_PRIVATE, total_mem, flags); |
| 77 | dprint(FD_MEM, "shmget %u, %d\n", total_mem, td->shm_id); |
| 78 | if (td->shm_id < 0) { |
| 79 | td_verror(td, errno, "shmget"); |
| 80 | if (geteuid() != 0 && (errno == ENOMEM || errno == EPERM)) |
| 81 | log_err("fio: you may need to run this job as root\n"); |
| 82 | if (td->o.mem_type == MEM_SHMHUGE) { |
| 83 | if (errno == EINVAL) { |
| 84 | log_err("fio: check that you have free huge" |
| 85 | " pages and that hugepage-size is" |
| 86 | " correct.\n"); |
| 87 | } else if (errno == ENOSYS) { |
| 88 | log_err("fio: your system does not appear to" |
| 89 | " support huge pages.\n"); |
| 90 | } else if (errno == ENOMEM) { |
| 91 | log_err("fio: no huge pages available, do you" |
| 92 | " need to allocate some? See HOWTO.\n"); |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | return 1; |
| 97 | } |
| 98 | |
| 99 | td->orig_buffer = shmat(td->shm_id, NULL, 0); |
| 100 | dprint(FD_MEM, "shmat %d, %p\n", td->shm_id, td->orig_buffer); |
| 101 | if (td->orig_buffer == (void *) -1) { |
| 102 | td_verror(td, errno, "shmat"); |
| 103 | td->orig_buffer = NULL; |
| 104 | return 1; |
| 105 | } |
| 106 | |
| 107 | return 0; |
| 108 | #else |
| 109 | log_err("fio: shm not supported\n"); |
| 110 | return 1; |
| 111 | #endif |
| 112 | } |
| 113 | |
| 114 | static void free_mem_shm(struct thread_data *td) |
| 115 | { |
| 116 | #ifndef CONFIG_NO_SHM |
| 117 | struct shmid_ds sbuf; |
| 118 | |
| 119 | dprint(FD_MEM, "shmdt/ctl %d %p\n", td->shm_id, td->orig_buffer); |
| 120 | shmdt(td->orig_buffer); |
| 121 | shmctl(td->shm_id, IPC_RMID, &sbuf); |
| 122 | #endif |
| 123 | } |
| 124 | |
| 125 | static int alloc_mem_mmap(struct thread_data *td, size_t total_mem) |
| 126 | { |
| 127 | int flags = 0; |
| 128 | |
| 129 | td->mmapfd = -1; |
| 130 | |
| 131 | if (td->o.mem_type == MEM_MMAPHUGE) { |
| 132 | unsigned long mask = td->o.hugepage_size - 1; |
| 133 | |
| 134 | /* TODO: make sure the file is a real hugetlbfs file */ |
| 135 | if (!td->o.mmapfile) |
| 136 | flags |= MAP_HUGETLB; |
| 137 | total_mem = (total_mem + mask) & ~mask; |
| 138 | } |
| 139 | |
| 140 | if (td->o.mmapfile) { |
| 141 | td->mmapfd = open(td->o.mmapfile, O_RDWR|O_CREAT, 0644); |
| 142 | |
| 143 | if (td->mmapfd < 0) { |
| 144 | td_verror(td, errno, "open mmap file"); |
| 145 | td->orig_buffer = NULL; |
| 146 | return 1; |
| 147 | } |
| 148 | if (td->o.mem_type != MEM_MMAPHUGE && |
| 149 | td->o.mem_type != MEM_MMAPSHARED && |
| 150 | ftruncate(td->mmapfd, total_mem) < 0) { |
| 151 | td_verror(td, errno, "truncate mmap file"); |
| 152 | td->orig_buffer = NULL; |
| 153 | return 1; |
| 154 | } |
| 155 | if (td->o.mem_type == MEM_MMAPHUGE || |
| 156 | td->o.mem_type == MEM_MMAPSHARED) |
| 157 | flags |= MAP_SHARED; |
| 158 | else |
| 159 | flags |= MAP_PRIVATE; |
| 160 | } else |
| 161 | flags |= OS_MAP_ANON | MAP_PRIVATE; |
| 162 | |
| 163 | td->orig_buffer = mmap(NULL, total_mem, PROT_READ | PROT_WRITE, flags, |
| 164 | td->mmapfd, 0); |
| 165 | dprint(FD_MEM, "mmap %llu/%d %p\n", (unsigned long long) total_mem, |
| 166 | td->mmapfd, td->orig_buffer); |
| 167 | if (td->orig_buffer == MAP_FAILED) { |
| 168 | td_verror(td, errno, "mmap"); |
| 169 | td->orig_buffer = NULL; |
| 170 | if (td->mmapfd != 1 && td->mmapfd != -1) { |
| 171 | close(td->mmapfd); |
| 172 | if (td->o.mmapfile) |
| 173 | unlink(td->o.mmapfile); |
| 174 | } |
| 175 | |
| 176 | return 1; |
| 177 | } |
| 178 | |
| 179 | return 0; |
| 180 | } |
| 181 | |
| 182 | static void free_mem_mmap(struct thread_data *td, size_t total_mem) |
| 183 | { |
| 184 | dprint(FD_MEM, "munmap %llu %p\n", (unsigned long long) total_mem, |
| 185 | td->orig_buffer); |
| 186 | munmap(td->orig_buffer, td->orig_buffer_size); |
| 187 | if (td->o.mmapfile) { |
| 188 | if (td->mmapfd != -1) |
| 189 | close(td->mmapfd); |
| 190 | unlink(td->o.mmapfile); |
| 191 | free(td->o.mmapfile); |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | static int alloc_mem_malloc(struct thread_data *td, size_t total_mem) |
| 196 | { |
| 197 | td->orig_buffer = malloc(total_mem); |
| 198 | dprint(FD_MEM, "malloc %llu %p\n", (unsigned long long) total_mem, |
| 199 | td->orig_buffer); |
| 200 | |
| 201 | return td->orig_buffer == NULL; |
| 202 | } |
| 203 | |
| 204 | static void free_mem_malloc(struct thread_data *td) |
| 205 | { |
| 206 | dprint(FD_MEM, "free malloc mem %p\n", td->orig_buffer); |
| 207 | free(td->orig_buffer); |
| 208 | } |
| 209 | |
| 210 | static int alloc_mem_cudamalloc(struct thread_data *td, size_t total_mem) |
| 211 | { |
| 212 | #ifdef CONFIG_CUDA |
| 213 | CUresult ret; |
| 214 | char name[128]; |
| 215 | |
| 216 | ret = cuInit(0); |
| 217 | if (ret != CUDA_SUCCESS) { |
| 218 | log_err("fio: failed initialize cuda driver api\n"); |
| 219 | return 1; |
| 220 | } |
| 221 | |
| 222 | ret = cuDeviceGetCount(&td->gpu_dev_cnt); |
| 223 | if (ret != CUDA_SUCCESS) { |
| 224 | log_err("fio: failed get device count\n"); |
| 225 | return 1; |
| 226 | } |
| 227 | dprint(FD_MEM, "found %d GPU devices\n", td->gpu_dev_cnt); |
| 228 | |
| 229 | if (td->gpu_dev_cnt == 0) { |
| 230 | log_err("fio: no GPU device found. " |
| 231 | "Can not perform GPUDirect RDMA.\n"); |
| 232 | return 1; |
| 233 | } |
| 234 | |
| 235 | td->gpu_dev_id = td->o.gpu_dev_id; |
| 236 | ret = cuDeviceGet(&td->cu_dev, td->gpu_dev_id); |
| 237 | if (ret != CUDA_SUCCESS) { |
| 238 | log_err("fio: failed get GPU device\n"); |
| 239 | return 1; |
| 240 | } |
| 241 | |
| 242 | ret = cuDeviceGetName(name, sizeof(name), td->gpu_dev_id); |
| 243 | if (ret != CUDA_SUCCESS) { |
| 244 | log_err("fio: failed get device name\n"); |
| 245 | return 1; |
| 246 | } |
| 247 | dprint(FD_MEM, "dev_id = [%d], device name = [%s]\n", \ |
| 248 | td->gpu_dev_id, name); |
| 249 | |
| 250 | ret = cuCtxCreate(&td->cu_ctx, CU_CTX_MAP_HOST, td->cu_dev); |
| 251 | if (ret != CUDA_SUCCESS) { |
| 252 | log_err("fio: failed to create cuda context: %d\n", ret); |
| 253 | return 1; |
| 254 | } |
| 255 | |
| 256 | ret = cuMemAlloc(&td->dev_mem_ptr, total_mem); |
| 257 | if (ret != CUDA_SUCCESS) { |
| 258 | log_err("fio: cuMemAlloc %zu bytes failed\n", total_mem); |
| 259 | return 1; |
| 260 | } |
| 261 | td->orig_buffer = (void *) td->dev_mem_ptr; |
| 262 | |
| 263 | dprint(FD_MEM, "cudaMalloc %llu %p\n", \ |
| 264 | (unsigned long long) total_mem, td->orig_buffer); |
| 265 | return 0; |
| 266 | #else |
| 267 | return -EINVAL; |
| 268 | #endif |
| 269 | } |
| 270 | |
| 271 | static void free_mem_cudamalloc(struct thread_data *td) |
| 272 | { |
| 273 | #ifdef CONFIG_CUDA |
| 274 | if (td->dev_mem_ptr != NULL) |
| 275 | cuMemFree(td->dev_mem_ptr); |
| 276 | |
| 277 | if (cuCtxDestroy(td->cu_ctx) != CUDA_SUCCESS) |
| 278 | log_err("fio: failed to destroy cuda context\n"); |
| 279 | #endif |
| 280 | } |
| 281 | |
| 282 | /* |
| 283 | * Set up the buffer area we need for io. |
| 284 | */ |
| 285 | int allocate_io_mem(struct thread_data *td) |
| 286 | { |
| 287 | size_t total_mem; |
| 288 | int ret = 0; |
| 289 | |
| 290 | if (td_ioengine_flagged(td, FIO_NOIO)) |
| 291 | return 0; |
| 292 | |
| 293 | total_mem = td->orig_buffer_size; |
| 294 | |
| 295 | if (td->o.odirect || td->o.mem_align || td->o.oatomic || |
| 296 | td_ioengine_flagged(td, FIO_MEMALIGN)) { |
| 297 | total_mem += page_mask; |
| 298 | if (td->o.mem_align && td->o.mem_align > page_size) |
| 299 | total_mem += td->o.mem_align - page_size; |
| 300 | } |
| 301 | |
| 302 | dprint(FD_MEM, "Alloc %llu for buffers\n", (unsigned long long) total_mem); |
| 303 | |
| 304 | /* |
| 305 | * If the IO engine has hooks to allocate/free memory, use those. But |
| 306 | * error out if the user explicitly asked for something else. |
| 307 | */ |
| 308 | if (td->io_ops->iomem_alloc) { |
| 309 | if (fio_option_is_set(&td->o, mem_type)) { |
| 310 | log_err("fio: option 'mem/iomem' conflicts with specified IO engine\n"); |
| 311 | ret = 1; |
| 312 | } else |
| 313 | ret = td->io_ops->iomem_alloc(td, total_mem); |
| 314 | } else if (td->o.mem_type == MEM_MALLOC) |
| 315 | ret = alloc_mem_malloc(td, total_mem); |
| 316 | else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE) |
| 317 | ret = alloc_mem_shm(td, total_mem); |
| 318 | else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE || |
| 319 | td->o.mem_type == MEM_MMAPSHARED) |
| 320 | ret = alloc_mem_mmap(td, total_mem); |
| 321 | else if (td->o.mem_type == MEM_CUDA_MALLOC) |
| 322 | ret = alloc_mem_cudamalloc(td, total_mem); |
| 323 | else { |
| 324 | log_err("fio: bad mem type: %d\n", td->o.mem_type); |
| 325 | ret = 1; |
| 326 | } |
| 327 | |
| 328 | if (ret) |
| 329 | td_verror(td, ENOMEM, "iomem allocation"); |
| 330 | |
| 331 | return ret; |
| 332 | } |
| 333 | |
| 334 | void free_io_mem(struct thread_data *td) |
| 335 | { |
| 336 | unsigned int total_mem; |
| 337 | |
| 338 | total_mem = td->orig_buffer_size; |
| 339 | if (td->o.odirect || td->o.oatomic) |
| 340 | total_mem += page_mask; |
| 341 | |
| 342 | if (td->io_ops->iomem_alloc) { |
| 343 | if (td->io_ops->iomem_free) |
| 344 | td->io_ops->iomem_free(td); |
| 345 | } else if (td->o.mem_type == MEM_MALLOC) |
| 346 | free_mem_malloc(td); |
| 347 | else if (td->o.mem_type == MEM_SHM || td->o.mem_type == MEM_SHMHUGE) |
| 348 | free_mem_shm(td); |
| 349 | else if (td->o.mem_type == MEM_MMAP || td->o.mem_type == MEM_MMAPHUGE || |
| 350 | td->o.mem_type == MEM_MMAPSHARED) |
| 351 | free_mem_mmap(td, total_mem); |
| 352 | else if (td->o.mem_type == MEM_CUDA_MALLOC) |
| 353 | free_mem_cudamalloc(td); |
| 354 | else |
| 355 | log_err("Bad memory type %u\n", td->o.mem_type); |
| 356 | |
| 357 | td->orig_buffer = NULL; |
| 358 | td->orig_buffer_size = 0; |
| 359 | } |