Commit | Line | Data |
---|---|---|
ebac4655 JA |
1 | #ifndef FIO_OS_LINUX_H |
2 | #define FIO_OS_LINUX_H | |
3 | ||
cca84643 JA |
4 | #define FIO_OS os_linux |
5 | ||
ebac4655 | 6 | #include <sys/ioctl.h> |
8756e4d4 | 7 | #include <sys/uio.h> |
3c39a379 | 8 | #include <sys/syscall.h> |
250e878a | 9 | #include <sys/sysmacros.h> |
2e3bd4c2 | 10 | #include <sys/vfs.h> |
39b93568 | 11 | #include <sys/mman.h> |
3c39a379 | 12 | #include <unistd.h> |
97fbdfa1 | 13 | #include <fcntl.h> |
3c2d93ed | 14 | #include <errno.h> |
d09913c2 | 15 | #include <sched.h> |
ea421790 | 16 | #include <linux/unistd.h> |
07e5b264 JA |
17 | #include <linux/raw.h> |
18 | #include <linux/major.h> | |
1894d171 TK |
19 | #include <linux/fs.h> |
20 | #include <scsi/sg.h> | |
ebac4655 | 21 | |
58828d07 SW |
22 | #ifdef ARCH_HAVE_CRC_CRYPTO |
23 | #include <sys/auxv.h> | |
24 | #ifndef HWCAP_CRC32 | |
25 | #define HWCAP_CRC32 (1 << 7) | |
26 | #endif /* HWCAP_CRC32 */ | |
27 | #endif /* ARCH_HAVE_CRC_CRYPTO */ | |
28 | ||
1c764dbe | 29 | #include "./os-linux-syscall.h" |
79a43187 | 30 | #include "binject.h" |
e2e58886 | 31 | #include "../file.h" |
f3de88a7 | 32 | |
7922a7b7 OM |
33 | #ifndef __has_builtin // Optional of course. |
34 | #define __has_builtin(x) 0 // Compatibility with non-clang compilers. | |
35 | #endif | |
36 | ||
ebac4655 JA |
37 | #define FIO_HAVE_CPU_AFFINITY |
38 | #define FIO_HAVE_DISK_UTIL | |
39 | #define FIO_HAVE_SGIO | |
ba4f8923 | 40 | #define FIO_HAVE_IOPRIO |
32ef447a | 41 | #define FIO_HAVE_IOPRIO_CLASS |
22f78b32 | 42 | #define FIO_HAVE_IOSCHED_SWITCH |
2c0ecd28 | 43 | #define FIO_HAVE_ODIRECT |
74b025b0 | 44 | #define FIO_HAVE_HUGETLB |
07e5b264 | 45 | #define FIO_HAVE_RAWBIND |
5e62c22a | 46 | #define FIO_HAVE_BLKTRACE |
eb7ccf38 | 47 | #define FIO_HAVE_CL_SIZE |
a696fa2a | 48 | #define FIO_HAVE_CGROUPS |
2e3bd4c2 | 49 | #define FIO_HAVE_FS_STAT |
a5f3027c | 50 | #define FIO_HAVE_TRIM |
79a43187 | 51 | #define FIO_HAVE_BINJECT |
47f767c1 | 52 | #define FIO_HAVE_GETTID |
93bcfd20 | 53 | #define FIO_USE_GENERIC_INIT_RANDOM_STATE |
6562685f | 54 | #define FIO_HAVE_PWRITEV2 |
97900ebf | 55 | #define FIO_HAVE_SHM_ATTACH_REMOVED |
b6959b55 | 56 | |
39b93568 JA |
57 | #ifdef MAP_HUGETLB |
58 | #define FIO_HAVE_MMAP_HUGE | |
59 | #endif | |
60 | ||
dc873b6f | 61 | #define OS_MAP_ANON MAP_ANONYMOUS |
ebac4655 JA |
62 | |
63 | typedef cpu_set_t os_cpu_mask_t; | |
4e78e405 | 64 | |
6dfd46b9 | 65 | typedef struct drand48_data os_random_state_t; |
ebac4655 | 66 | |
67bf9823 | 67 | #ifdef CONFIG_3ARG_AFFINITY |
e8462bd8 JA |
68 | #define fio_setaffinity(pid, cpumask) \ |
69 | sched_setaffinity((pid), sizeof(cpumask), &(cpumask)) | |
ebac4655 JA |
70 | #define fio_getaffinity(pid, ptr) \ |
71 | sched_getaffinity((pid), sizeof(cpu_set_t), (ptr)) | |
67bf9823 | 72 | #elif defined(CONFIG_2ARG_AFFINITY) |
e8462bd8 JA |
73 | #define fio_setaffinity(pid, cpumask) \ |
74 | sched_setaffinity((pid), &(cpumask)) | |
c8f025f8 JA |
75 | #define fio_getaffinity(pid, ptr) \ |
76 | sched_getaffinity((pid), (ptr)) | |
77 | #endif | |
be4ecfdf | 78 | |
1f4c0a4f JA |
79 | #define fio_cpu_clear(mask, cpu) (void) CPU_CLR((cpu), (mask)) |
80 | #define fio_cpu_set(mask, cpu) (void) CPU_SET((cpu), (mask)) | |
52fd65f4 | 81 | #define fio_cpu_isset(mask, cpu) (CPU_ISSET((cpu), (mask)) != 0) |
d004a209 | 82 | #define fio_cpu_count(mask) CPU_COUNT((mask)) |
d2ce18b5 JA |
83 | |
84 | static inline int fio_cpuset_init(os_cpu_mask_t *mask) | |
85 | { | |
86 | CPU_ZERO(mask); | |
87 | return 0; | |
88 | } | |
89 | ||
90 | static inline int fio_cpuset_exit(os_cpu_mask_t *mask) | |
91 | { | |
92 | return 0; | |
93 | } | |
6d459ee7 JA |
94 | |
95 | #define FIO_MAX_CPUS CPU_SETSIZE | |
4d8947de | 96 | |
28727df7 JA |
97 | enum { |
98 | IOPRIO_CLASS_NONE, | |
99 | IOPRIO_CLASS_RT, | |
100 | IOPRIO_CLASS_BE, | |
101 | IOPRIO_CLASS_IDLE, | |
102 | }; | |
103 | ||
104 | enum { | |
105 | IOPRIO_WHO_PROCESS = 1, | |
106 | IOPRIO_WHO_PGRP, | |
107 | IOPRIO_WHO_USER, | |
108 | }; | |
109 | ||
110 | #define IOPRIO_BITS 16 | |
111 | #define IOPRIO_CLASS_SHIFT 13 | |
112 | ||
1767bd34 TK |
113 | #define IOPRIO_MIN_PRIO 0 /* highest priority */ |
114 | #define IOPRIO_MAX_PRIO 7 /* lowest priority */ | |
115 | ||
116 | #define IOPRIO_MIN_PRIO_CLASS 0 | |
117 | #define IOPRIO_MAX_PRIO_CLASS 3 | |
118 | ||
28727df7 | 119 | static inline int ioprio_set(int which, int who, int ioprio_class, int ioprio) |
ebac4655 | 120 | { |
28727df7 JA |
121 | /* |
122 | * If no class is set, assume BE | |
123 | */ | |
124 | if (!ioprio_class) | |
125 | ioprio_class = IOPRIO_CLASS_BE; | |
126 | ||
127 | ioprio |= ioprio_class << IOPRIO_CLASS_SHIFT; | |
ebac4655 JA |
128 | return syscall(__NR_ioprio_set, which, who, ioprio); |
129 | } | |
130 | ||
47f767c1 JA |
131 | static inline int gettid(void) |
132 | { | |
133 | return syscall(__NR_gettid); | |
134 | } | |
135 | ||
3feedc60 JA |
136 | #define SPLICE_DEF_SIZE (64*1024) |
137 | ||
ebac4655 JA |
138 | #ifndef BLKGETSIZE64 |
139 | #define BLKGETSIZE64 _IOR(0x12,114,size_t) | |
140 | #endif | |
141 | ||
e5b401d4 JA |
142 | #ifndef BLKFLSBUF |
143 | #define BLKFLSBUF _IO(0x12,97) | |
144 | #endif | |
145 | ||
a5f3027c JA |
146 | #ifndef BLKDISCARD |
147 | #define BLKDISCARD _IO(0x12,119) | |
148 | #endif | |
149 | ||
e2e58886 | 150 | static inline int blockdev_invalidate_cache(struct fio_file *f) |
e5b401d4 | 151 | { |
ecc314ba | 152 | return ioctl(f->fd, BLKFLSBUF); |
e5b401d4 JA |
153 | } |
154 | ||
ecc314ba | 155 | static inline int blockdev_size(struct fio_file *f, unsigned long long *bytes) |
ebac4655 | 156 | { |
ecc314ba | 157 | if (!ioctl(f->fd, BLKGETSIZE64, bytes)) |
ebac4655 JA |
158 | return 0; |
159 | ||
160 | return errno; | |
161 | } | |
162 | ||
32cd46a0 JA |
163 | static inline unsigned long long os_phys_mem(void) |
164 | { | |
165 | long pagesize, pages; | |
166 | ||
167 | pagesize = sysconf(_SC_PAGESIZE); | |
168 | pages = sysconf(_SC_PHYS_PAGES); | |
169 | if (pages == -1 || pagesize == -1) | |
170 | return 0; | |
171 | ||
172 | return (unsigned long long) pages * (unsigned long long) pagesize; | |
173 | } | |
174 | ||
6dfd46b9 JA |
175 | static inline void os_random_seed(unsigned long seed, os_random_state_t *rs) |
176 | { | |
177 | srand48_r(seed, rs); | |
178 | } | |
179 | ||
180 | static inline long os_random_long(os_random_state_t *rs) | |
181 | { | |
182 | long val; | |
183 | ||
184 | lrand48_r(rs, &val); | |
185 | return val; | |
186 | } | |
187 | ||
8cc7afa9 | 188 | static inline int fio_lookup_raw(dev_t dev, int *majdev, int *mindev) |
07e5b264 JA |
189 | { |
190 | struct raw_config_request rq; | |
191 | int fd; | |
192 | ||
193 | if (major(dev) != RAW_MAJOR) | |
8cc7afa9 | 194 | return 1; |
07e5b264 JA |
195 | |
196 | /* | |
197 | * we should be able to find /dev/rawctl or /dev/raw/rawctl | |
198 | */ | |
199 | fd = open("/dev/rawctl", O_RDONLY); | |
200 | if (fd < 0) { | |
201 | fd = open("/dev/raw/rawctl", O_RDONLY); | |
202 | if (fd < 0) | |
8cc7afa9 | 203 | return 1; |
07e5b264 JA |
204 | } |
205 | ||
206 | rq.raw_minor = minor(dev); | |
207 | if (ioctl(fd, RAW_GETBIND, &rq) < 0) { | |
208 | close(fd); | |
8cc7afa9 | 209 | return 1; |
07e5b264 JA |
210 | } |
211 | ||
212 | close(fd); | |
213 | *majdev = rq.block_major; | |
214 | *mindev = rq.block_minor; | |
8cc7afa9 | 215 | return 0; |
07e5b264 JA |
216 | } |
217 | ||
5b6f5c66 | 218 | #ifdef O_NOATIME |
5921e80c | 219 | #define FIO_O_NOATIME O_NOATIME |
5b6f5c66 JA |
220 | #else |
221 | #define FIO_O_NOATIME 0 | |
222 | #endif | |
5921e80c | 223 | |
d01612f3 CM |
224 | #ifdef O_ATOMIC |
225 | #define OS_O_ATOMIC O_ATOMIC | |
226 | #else | |
227 | #define OS_O_ATOMIC 040000000 | |
228 | #endif | |
229 | ||
a1c58075 JA |
230 | #ifdef MADV_REMOVE |
231 | #define FIO_MADV_FREE MADV_REMOVE | |
232 | #endif | |
233 | ||
7922a7b7 OM |
234 | /* Check for GCC or Clang byte swap intrinsics */ |
235 | #if (__has_builtin(__builtin_bswap16) && __has_builtin(__builtin_bswap32) \ | |
236 | && __has_builtin(__builtin_bswap64)) || (__GNUC__ > 4 \ | |
237 | || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)) /* fio_swapN */ | |
df73aab1 | 238 | #define fio_swap16(x) __builtin_bswap16(x) |
df73aab1 | 239 | #define fio_swap32(x) __builtin_bswap32(x) |
df73aab1 JA |
240 | #define fio_swap64(x) __builtin_bswap64(x) |
241 | #else | |
7922a7b7 OM |
242 | #include <byteswap.h> |
243 | #define fio_swap16(x) bswap_16(x) | |
244 | #define fio_swap32(x) bswap_32(x) | |
245 | #define fio_swap64(x) bswap_64(x) | |
246 | #endif /* fio_swapN */ | |
ff245192 | 247 | |
eb7ccf38 JA |
248 | #define CACHE_LINE_FILE \ |
249 | "/sys/devices/system/cpu/cpu0/cache/index0/coherency_line_size" | |
250 | ||
251 | static inline int arch_cache_line_size(void) | |
252 | { | |
253 | char size[32]; | |
254 | int fd, ret; | |
255 | ||
256 | fd = open(CACHE_LINE_FILE, O_RDONLY); | |
257 | if (fd < 0) | |
258 | return -1; | |
259 | ||
260 | ret = read(fd, size, sizeof(size)); | |
ec76f15a JA |
261 | |
262 | close(fd); | |
263 | ||
eb7ccf38 JA |
264 | if (ret <= 0) |
265 | return -1; | |
ec76f15a JA |
266 | else |
267 | return atoi(size); | |
eb7ccf38 JA |
268 | } |
269 | ||
d1419c4f TK |
270 | #ifdef __powerpc64__ |
271 | #define FIO_HAVE_CPU_ONLINE_SYSCONF | |
272 | static inline unsigned int cpus_online(void) | |
273 | { | |
274 | return sysconf(_SC_NPROCESSORS_CONF); | |
275 | } | |
276 | #endif | |
277 | ||
c08ad04c | 278 | static inline unsigned long long get_fs_free_size(const char *path) |
2e3bd4c2 JA |
279 | { |
280 | unsigned long long ret; | |
281 | struct statfs s; | |
282 | ||
283 | if (statfs(path, &s) < 0) | |
284 | return -1ULL; | |
285 | ||
286 | ret = s.f_bsize; | |
287 | ret *= (unsigned long long) s.f_bfree; | |
288 | return ret; | |
289 | } | |
290 | ||
496b1f9e | 291 | static inline int os_trim(struct fio_file *f, unsigned long long start, |
a5f3027c JA |
292 | unsigned long long len) |
293 | { | |
294 | uint64_t range[2]; | |
295 | ||
296 | range[0] = start; | |
297 | range[1] = len; | |
298 | ||
496b1f9e | 299 | if (!ioctl(f->fd, BLKDISCARD, range)) |
a5f3027c JA |
300 | return 0; |
301 | ||
302 | return errno; | |
303 | } | |
304 | ||
7e09a9f1 | 305 | #ifdef CONFIG_SCHED_IDLE |
f2a2ce0e HL |
306 | static inline int fio_set_sched_idle(void) |
307 | { | |
308 | struct sched_param p = { .sched_priority = 0, }; | |
309 | return sched_setscheduler(gettid(), SCHED_IDLE, &p); | |
310 | } | |
7e09a9f1 | 311 | #endif |
f2a2ce0e | 312 | |
ae8e559e JA |
313 | #ifndef F_GET_RW_HINT |
314 | #ifndef F_LINUX_SPECIFIC_BASE | |
315 | #define F_LINUX_SPECIFIC_BASE 1024 | |
316 | #endif | |
317 | #define F_GET_RW_HINT (F_LINUX_SPECIFIC_BASE + 11) | |
318 | #define F_SET_RW_HINT (F_LINUX_SPECIFIC_BASE + 12) | |
bd553af6 JA |
319 | #define F_GET_FILE_RW_HINT (F_LINUX_SPECIFIC_BASE + 13) |
320 | #define F_SET_FILE_RW_HINT (F_LINUX_SPECIFIC_BASE + 14) | |
ae8e559e JA |
321 | #endif |
322 | ||
323 | #ifndef RWH_WRITE_LIFE_NONE | |
bd553af6 JA |
324 | #define RWH_WRITE_LIFE_NOT_SET 0 |
325 | #define RWH_WRITE_LIFE_NONE 1 | |
326 | #define RWH_WRITE_LIFE_SHORT 2 | |
327 | #define RWH_WRITE_LIFE_MEDIUM 3 | |
328 | #define RWH_WRITE_LIFE_LONG 4 | |
329 | #define RWH_WRITE_LIFE_EXTREME 5 | |
37659335 JA |
330 | #endif |
331 | ||
ae8e559e | 332 | #define FIO_HAVE_WRITE_HINT |
37659335 | 333 | |
6562685f JA |
334 | #ifndef RWF_HIPRI |
335 | #define RWF_HIPRI 0x00000001 | |
336 | #endif | |
337 | #ifndef RWF_DSYNC | |
338 | #define RWF_DSYNC 0x00000002 | |
339 | #endif | |
340 | #ifndef RWF_SYNC | |
341 | #define RWF_SYNC 0x00000004 | |
342 | #endif | |
343 | ||
72cd4146 JA |
344 | #ifndef RWF_WRITE_LIFE_SHIFT |
345 | #define RWF_WRITE_LIFE_SHIFT 4 | |
346 | #define RWF_WRITE_LIFE_SHORT (1 << RWF_WRITE_LIFE_SHIFT) | |
347 | #define RWF_WRITE_LIFE_MEDIUM (2 << RWF_WRITE_LIFE_SHIFT) | |
348 | #define RWF_WRITE_LIFE_LONG (3 << RWF_WRITE_LIFE_SHIFT) | |
349 | #define RWF_WRITE_LIFE_EXTREME (4 << RWF_WRITE_LIFE_SHIFT) | |
350 | #endif | |
351 | ||
6562685f JA |
352 | #ifndef CONFIG_PWRITEV2 |
353 | #ifdef __NR_preadv2 | |
354 | static inline void make_pos_h_l(unsigned long *pos_h, unsigned long *pos_l, | |
355 | off_t offset) | |
356 | { | |
7b5c648f | 357 | #if BITS_PER_LONG == 64 |
7b5c648f JA |
358 | *pos_l = offset; |
359 | *pos_h = 0; | |
360 | #else | |
6562685f JA |
361 | *pos_l = offset & 0xffffffff; |
362 | *pos_h = ((uint64_t) offset) >> 32; | |
7b5c648f | 363 | #endif |
6562685f JA |
364 | } |
365 | static inline ssize_t preadv2(int fd, const struct iovec *iov, int iovcnt, | |
366 | off_t offset, unsigned int flags) | |
367 | { | |
368 | unsigned long pos_l, pos_h; | |
369 | ||
370 | make_pos_h_l(&pos_h, &pos_l, offset); | |
371 | return syscall(__NR_preadv2, fd, iov, iovcnt, pos_l, pos_h, flags); | |
372 | } | |
373 | static inline ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt, | |
374 | off_t offset, unsigned int flags) | |
375 | { | |
376 | unsigned long pos_l, pos_h; | |
377 | ||
378 | make_pos_h_l(&pos_h, &pos_l, offset); | |
379 | return syscall(__NR_pwritev2, fd, iov, iovcnt, pos_l, pos_h, flags); | |
380 | } | |
381 | #else | |
382 | static inline ssize_t preadv2(int fd, const struct iovec *iov, int iovcnt, | |
383 | off_t offset, unsigned int flags) | |
384 | { | |
385 | errno = ENOSYS; | |
386 | return -1; | |
387 | } | |
388 | static inline ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt, | |
389 | off_t offset, unsigned int flags) | |
390 | { | |
391 | errno = ENOSYS; | |
392 | return -1; | |
393 | } | |
394 | #endif /* __NR_preadv2 */ | |
395 | #endif /* CONFIG_PWRITEV2 */ | |
396 | ||
0cfe2089 TK |
397 | static inline int shm_attach_to_open_removed(void) |
398 | { | |
399 | return 1; | |
400 | } | |
401 | ||
2c3e17be SW |
402 | #ifdef CONFIG_LINUX_FALLOCATE |
403 | #define FIO_HAVE_NATIVE_FALLOCATE | |
404 | static inline bool fio_fallocate(struct fio_file *f, uint64_t offset, | |
405 | uint64_t len) | |
406 | { | |
407 | int ret; | |
408 | ret = fallocate(f->fd, 0, 0, len); | |
409 | if (ret == 0) | |
410 | return true; | |
411 | ||
412 | /* Work around buggy old glibc versions... */ | |
413 | if (ret > 0) | |
414 | errno = ret; | |
415 | ||
416 | return false; | |
417 | } | |
418 | #endif | |
419 | ||
58828d07 SW |
420 | #define FIO_HAVE_CPU_HAS |
421 | static inline bool os_cpu_has(cpu_features feature) | |
422 | { | |
423 | bool have_feature; | |
424 | unsigned long fio_unused hwcap; | |
425 | ||
426 | switch (feature) { | |
427 | #ifdef ARCH_HAVE_CRC_CRYPTO | |
428 | case CPU_ARM64_CRC32C: | |
429 | hwcap = getauxval(AT_HWCAP); | |
430 | have_feature = (hwcap & HWCAP_CRC32) != 0; | |
431 | break; | |
432 | #endif | |
433 | default: | |
434 | have_feature = false; | |
435 | } | |
436 | ||
437 | return have_feature; | |
438 | } | |
439 | ||
ebac4655 | 440 | #endif |