| 1 | #ifndef FIO_OS_LINUX_H |
| 2 | #define FIO_OS_LINUX_H |
| 3 | |
| 4 | #define FIO_OS os_linux |
| 5 | |
| 6 | #include <sys/ioctl.h> |
| 7 | #include <sys/uio.h> |
| 8 | #include <sys/syscall.h> |
| 9 | #include <sys/sysmacros.h> |
| 10 | #include <sys/vfs.h> |
| 11 | #include <sys/mman.h> |
| 12 | #include <unistd.h> |
| 13 | #include <fcntl.h> |
| 14 | #include <errno.h> |
| 15 | #include <sched.h> |
| 16 | #include <linux/unistd.h> |
| 17 | #include <linux/raw.h> |
| 18 | #include <linux/major.h> |
| 19 | #include <linux/fs.h> |
| 20 | #include <scsi/sg.h> |
| 21 | |
| 22 | #include "./os-linux-syscall.h" |
| 23 | #include "binject.h" |
| 24 | #include "../file.h" |
| 25 | |
| 26 | #ifndef __has_builtin // Optional of course. |
| 27 | #define __has_builtin(x) 0 // Compatibility with non-clang compilers. |
| 28 | #endif |
| 29 | |
| 30 | #define FIO_HAVE_CPU_AFFINITY |
| 31 | #define FIO_HAVE_DISK_UTIL |
| 32 | #define FIO_HAVE_SGIO |
| 33 | #define FIO_HAVE_IOPRIO |
| 34 | #define FIO_HAVE_IOPRIO_CLASS |
| 35 | #define FIO_HAVE_IOSCHED_SWITCH |
| 36 | #define FIO_HAVE_ODIRECT |
| 37 | #define FIO_HAVE_HUGETLB |
| 38 | #define FIO_HAVE_RAWBIND |
| 39 | #define FIO_HAVE_BLKTRACE |
| 40 | #define FIO_HAVE_CL_SIZE |
| 41 | #define FIO_HAVE_CGROUPS |
| 42 | #define FIO_HAVE_FS_STAT |
| 43 | #define FIO_HAVE_TRIM |
| 44 | #define FIO_HAVE_BINJECT |
| 45 | #define FIO_HAVE_GETTID |
| 46 | #define FIO_USE_GENERIC_INIT_RANDOM_STATE |
| 47 | #define FIO_HAVE_PWRITEV2 |
| 48 | #define FIO_HAVE_SHM_ATTACH_REMOVED |
| 49 | |
| 50 | #ifdef MAP_HUGETLB |
| 51 | #define FIO_HAVE_MMAP_HUGE |
| 52 | #endif |
| 53 | |
| 54 | #define OS_MAP_ANON MAP_ANONYMOUS |
| 55 | |
| 56 | typedef cpu_set_t os_cpu_mask_t; |
| 57 | |
| 58 | typedef struct drand48_data os_random_state_t; |
| 59 | |
| 60 | #ifdef CONFIG_3ARG_AFFINITY |
| 61 | #define fio_setaffinity(pid, cpumask) \ |
| 62 | sched_setaffinity((pid), sizeof(cpumask), &(cpumask)) |
| 63 | #define fio_getaffinity(pid, ptr) \ |
| 64 | sched_getaffinity((pid), sizeof(cpu_set_t), (ptr)) |
| 65 | #elif defined(CONFIG_2ARG_AFFINITY) |
| 66 | #define fio_setaffinity(pid, cpumask) \ |
| 67 | sched_setaffinity((pid), &(cpumask)) |
| 68 | #define fio_getaffinity(pid, ptr) \ |
| 69 | sched_getaffinity((pid), (ptr)) |
| 70 | #endif |
| 71 | |
| 72 | #define fio_cpu_clear(mask, cpu) (void) CPU_CLR((cpu), (mask)) |
| 73 | #define fio_cpu_set(mask, cpu) (void) CPU_SET((cpu), (mask)) |
| 74 | #define fio_cpu_isset(mask, cpu) CPU_ISSET((cpu), (mask)) |
| 75 | #define fio_cpu_count(mask) CPU_COUNT((mask)) |
| 76 | |
| 77 | static inline int fio_cpuset_init(os_cpu_mask_t *mask) |
| 78 | { |
| 79 | CPU_ZERO(mask); |
| 80 | return 0; |
| 81 | } |
| 82 | |
| 83 | static inline int fio_cpuset_exit(os_cpu_mask_t *mask) |
| 84 | { |
| 85 | return 0; |
| 86 | } |
| 87 | |
| 88 | #define FIO_MAX_CPUS CPU_SETSIZE |
| 89 | |
| 90 | enum { |
| 91 | IOPRIO_CLASS_NONE, |
| 92 | IOPRIO_CLASS_RT, |
| 93 | IOPRIO_CLASS_BE, |
| 94 | IOPRIO_CLASS_IDLE, |
| 95 | }; |
| 96 | |
| 97 | enum { |
| 98 | IOPRIO_WHO_PROCESS = 1, |
| 99 | IOPRIO_WHO_PGRP, |
| 100 | IOPRIO_WHO_USER, |
| 101 | }; |
| 102 | |
| 103 | #define IOPRIO_BITS 16 |
| 104 | #define IOPRIO_CLASS_SHIFT 13 |
| 105 | |
| 106 | #define IOPRIO_MIN_PRIO 0 /* highest priority */ |
| 107 | #define IOPRIO_MAX_PRIO 7 /* lowest priority */ |
| 108 | |
| 109 | #define IOPRIO_MIN_PRIO_CLASS 0 |
| 110 | #define IOPRIO_MAX_PRIO_CLASS 3 |
| 111 | |
| 112 | static inline int ioprio_set(int which, int who, int ioprio_class, int ioprio) |
| 113 | { |
| 114 | /* |
| 115 | * If no class is set, assume BE |
| 116 | */ |
| 117 | if (!ioprio_class) |
| 118 | ioprio_class = IOPRIO_CLASS_BE; |
| 119 | |
| 120 | ioprio |= ioprio_class << IOPRIO_CLASS_SHIFT; |
| 121 | return syscall(__NR_ioprio_set, which, who, ioprio); |
| 122 | } |
| 123 | |
| 124 | static inline int gettid(void) |
| 125 | { |
| 126 | return syscall(__NR_gettid); |
| 127 | } |
| 128 | |
| 129 | #define SPLICE_DEF_SIZE (64*1024) |
| 130 | |
| 131 | #ifndef BLKGETSIZE64 |
| 132 | #define BLKGETSIZE64 _IOR(0x12,114,size_t) |
| 133 | #endif |
| 134 | |
| 135 | #ifndef BLKFLSBUF |
| 136 | #define BLKFLSBUF _IO(0x12,97) |
| 137 | #endif |
| 138 | |
| 139 | #ifndef BLKDISCARD |
| 140 | #define BLKDISCARD _IO(0x12,119) |
| 141 | #endif |
| 142 | |
| 143 | static inline int blockdev_invalidate_cache(struct fio_file *f) |
| 144 | { |
| 145 | return ioctl(f->fd, BLKFLSBUF); |
| 146 | } |
| 147 | |
| 148 | static inline int blockdev_size(struct fio_file *f, unsigned long long *bytes) |
| 149 | { |
| 150 | if (!ioctl(f->fd, BLKGETSIZE64, bytes)) |
| 151 | return 0; |
| 152 | |
| 153 | return errno; |
| 154 | } |
| 155 | |
| 156 | static inline unsigned long long os_phys_mem(void) |
| 157 | { |
| 158 | long pagesize, pages; |
| 159 | |
| 160 | pagesize = sysconf(_SC_PAGESIZE); |
| 161 | pages = sysconf(_SC_PHYS_PAGES); |
| 162 | if (pages == -1 || pagesize == -1) |
| 163 | return 0; |
| 164 | |
| 165 | return (unsigned long long) pages * (unsigned long long) pagesize; |
| 166 | } |
| 167 | |
| 168 | static inline void os_random_seed(unsigned long seed, os_random_state_t *rs) |
| 169 | { |
| 170 | srand48_r(seed, rs); |
| 171 | } |
| 172 | |
| 173 | static inline long os_random_long(os_random_state_t *rs) |
| 174 | { |
| 175 | long val; |
| 176 | |
| 177 | lrand48_r(rs, &val); |
| 178 | return val; |
| 179 | } |
| 180 | |
| 181 | static inline int fio_lookup_raw(dev_t dev, int *majdev, int *mindev) |
| 182 | { |
| 183 | struct raw_config_request rq; |
| 184 | int fd; |
| 185 | |
| 186 | if (major(dev) != RAW_MAJOR) |
| 187 | return 1; |
| 188 | |
| 189 | /* |
| 190 | * we should be able to find /dev/rawctl or /dev/raw/rawctl |
| 191 | */ |
| 192 | fd = open("/dev/rawctl", O_RDONLY); |
| 193 | if (fd < 0) { |
| 194 | fd = open("/dev/raw/rawctl", O_RDONLY); |
| 195 | if (fd < 0) |
| 196 | return 1; |
| 197 | } |
| 198 | |
| 199 | rq.raw_minor = minor(dev); |
| 200 | if (ioctl(fd, RAW_GETBIND, &rq) < 0) { |
| 201 | close(fd); |
| 202 | return 1; |
| 203 | } |
| 204 | |
| 205 | close(fd); |
| 206 | *majdev = rq.block_major; |
| 207 | *mindev = rq.block_minor; |
| 208 | return 0; |
| 209 | } |
| 210 | |
| 211 | #ifdef O_NOATIME |
| 212 | #define FIO_O_NOATIME O_NOATIME |
| 213 | #else |
| 214 | #define FIO_O_NOATIME 0 |
| 215 | #endif |
| 216 | |
| 217 | #ifdef O_ATOMIC |
| 218 | #define OS_O_ATOMIC O_ATOMIC |
| 219 | #else |
| 220 | #define OS_O_ATOMIC 040000000 |
| 221 | #endif |
| 222 | |
| 223 | #ifdef MADV_REMOVE |
| 224 | #define FIO_MADV_FREE MADV_REMOVE |
| 225 | #endif |
| 226 | |
| 227 | /* Check for GCC or Clang byte swap intrinsics */ |
| 228 | #if (__has_builtin(__builtin_bswap16) && __has_builtin(__builtin_bswap32) \ |
| 229 | && __has_builtin(__builtin_bswap64)) || (__GNUC__ > 4 \ |
| 230 | || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)) /* fio_swapN */ |
| 231 | #define fio_swap16(x) __builtin_bswap16(x) |
| 232 | #define fio_swap32(x) __builtin_bswap32(x) |
| 233 | #define fio_swap64(x) __builtin_bswap64(x) |
| 234 | #else |
| 235 | #include <byteswap.h> |
| 236 | #define fio_swap16(x) bswap_16(x) |
| 237 | #define fio_swap32(x) bswap_32(x) |
| 238 | #define fio_swap64(x) bswap_64(x) |
| 239 | #endif /* fio_swapN */ |
| 240 | |
| 241 | #define CACHE_LINE_FILE \ |
| 242 | "/sys/devices/system/cpu/cpu0/cache/index0/coherency_line_size" |
| 243 | |
| 244 | static inline int arch_cache_line_size(void) |
| 245 | { |
| 246 | char size[32]; |
| 247 | int fd, ret; |
| 248 | |
| 249 | fd = open(CACHE_LINE_FILE, O_RDONLY); |
| 250 | if (fd < 0) |
| 251 | return -1; |
| 252 | |
| 253 | ret = read(fd, size, sizeof(size)); |
| 254 | |
| 255 | close(fd); |
| 256 | |
| 257 | if (ret <= 0) |
| 258 | return -1; |
| 259 | else |
| 260 | return atoi(size); |
| 261 | } |
| 262 | |
| 263 | #ifdef __powerpc64__ |
| 264 | #define FIO_HAVE_CPU_ONLINE_SYSCONF |
| 265 | static inline unsigned int cpus_online(void) |
| 266 | { |
| 267 | return sysconf(_SC_NPROCESSORS_CONF); |
| 268 | } |
| 269 | #endif |
| 270 | |
| 271 | static inline unsigned long long get_fs_free_size(const char *path) |
| 272 | { |
| 273 | unsigned long long ret; |
| 274 | struct statfs s; |
| 275 | |
| 276 | if (statfs(path, &s) < 0) |
| 277 | return -1ULL; |
| 278 | |
| 279 | ret = s.f_bsize; |
| 280 | ret *= (unsigned long long) s.f_bfree; |
| 281 | return ret; |
| 282 | } |
| 283 | |
| 284 | static inline int os_trim(int fd, unsigned long long start, |
| 285 | unsigned long long len) |
| 286 | { |
| 287 | uint64_t range[2]; |
| 288 | |
| 289 | range[0] = start; |
| 290 | range[1] = len; |
| 291 | |
| 292 | if (!ioctl(fd, BLKDISCARD, range)) |
| 293 | return 0; |
| 294 | |
| 295 | return errno; |
| 296 | } |
| 297 | |
| 298 | #ifdef CONFIG_SCHED_IDLE |
| 299 | static inline int fio_set_sched_idle(void) |
| 300 | { |
| 301 | struct sched_param p = { .sched_priority = 0, }; |
| 302 | return sched_setscheduler(gettid(), SCHED_IDLE, &p); |
| 303 | } |
| 304 | #endif |
| 305 | |
| 306 | #ifndef F_GET_RW_HINT |
| 307 | #ifndef F_LINUX_SPECIFIC_BASE |
| 308 | #define F_LINUX_SPECIFIC_BASE 1024 |
| 309 | #endif |
| 310 | #define F_GET_RW_HINT (F_LINUX_SPECIFIC_BASE + 11) |
| 311 | #define F_SET_RW_HINT (F_LINUX_SPECIFIC_BASE + 12) |
| 312 | #define F_GET_FILE_RW_HINT (F_LINUX_SPECIFIC_BASE + 13) |
| 313 | #define F_SET_FILE_RW_HINT (F_LINUX_SPECIFIC_BASE + 14) |
| 314 | #endif |
| 315 | |
| 316 | #ifndef RWH_WRITE_LIFE_NONE |
| 317 | #define RWH_WRITE_LIFE_NOT_SET 0 |
| 318 | #define RWH_WRITE_LIFE_NONE 1 |
| 319 | #define RWH_WRITE_LIFE_SHORT 2 |
| 320 | #define RWH_WRITE_LIFE_MEDIUM 3 |
| 321 | #define RWH_WRITE_LIFE_LONG 4 |
| 322 | #define RWH_WRITE_LIFE_EXTREME 5 |
| 323 | #endif |
| 324 | |
| 325 | #define FIO_HAVE_WRITE_HINT |
| 326 | |
| 327 | #ifndef RWF_HIPRI |
| 328 | #define RWF_HIPRI 0x00000001 |
| 329 | #endif |
| 330 | #ifndef RWF_DSYNC |
| 331 | #define RWF_DSYNC 0x00000002 |
| 332 | #endif |
| 333 | #ifndef RWF_SYNC |
| 334 | #define RWF_SYNC 0x00000004 |
| 335 | #endif |
| 336 | |
| 337 | #ifndef RWF_WRITE_LIFE_SHIFT |
| 338 | #define RWF_WRITE_LIFE_SHIFT 4 |
| 339 | #define RWF_WRITE_LIFE_SHORT (1 << RWF_WRITE_LIFE_SHIFT) |
| 340 | #define RWF_WRITE_LIFE_MEDIUM (2 << RWF_WRITE_LIFE_SHIFT) |
| 341 | #define RWF_WRITE_LIFE_LONG (3 << RWF_WRITE_LIFE_SHIFT) |
| 342 | #define RWF_WRITE_LIFE_EXTREME (4 << RWF_WRITE_LIFE_SHIFT) |
| 343 | #endif |
| 344 | |
| 345 | #ifndef CONFIG_PWRITEV2 |
| 346 | #ifdef __NR_preadv2 |
| 347 | static inline void make_pos_h_l(unsigned long *pos_h, unsigned long *pos_l, |
| 348 | off_t offset) |
| 349 | { |
| 350 | #if BITS_PER_LONG == 64 |
| 351 | *pos_l = offset; |
| 352 | *pos_h = 0; |
| 353 | #else |
| 354 | *pos_l = offset & 0xffffffff; |
| 355 | *pos_h = ((uint64_t) offset) >> 32; |
| 356 | #endif |
| 357 | } |
| 358 | static inline ssize_t preadv2(int fd, const struct iovec *iov, int iovcnt, |
| 359 | off_t offset, unsigned int flags) |
| 360 | { |
| 361 | unsigned long pos_l, pos_h; |
| 362 | |
| 363 | make_pos_h_l(&pos_h, &pos_l, offset); |
| 364 | return syscall(__NR_preadv2, fd, iov, iovcnt, pos_l, pos_h, flags); |
| 365 | } |
| 366 | static inline ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt, |
| 367 | off_t offset, unsigned int flags) |
| 368 | { |
| 369 | unsigned long pos_l, pos_h; |
| 370 | |
| 371 | make_pos_h_l(&pos_h, &pos_l, offset); |
| 372 | return syscall(__NR_pwritev2, fd, iov, iovcnt, pos_l, pos_h, flags); |
| 373 | } |
| 374 | #else |
| 375 | static inline ssize_t preadv2(int fd, const struct iovec *iov, int iovcnt, |
| 376 | off_t offset, unsigned int flags) |
| 377 | { |
| 378 | errno = ENOSYS; |
| 379 | return -1; |
| 380 | } |
| 381 | static inline ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt, |
| 382 | off_t offset, unsigned int flags) |
| 383 | { |
| 384 | errno = ENOSYS; |
| 385 | return -1; |
| 386 | } |
| 387 | #endif /* __NR_preadv2 */ |
| 388 | #endif /* CONFIG_PWRITEV2 */ |
| 389 | |
| 390 | static inline int shm_attach_to_open_removed(void) |
| 391 | { |
| 392 | return 1; |
| 393 | } |
| 394 | |
| 395 | #endif |