Merge branch 'fio-c++-engine' of https://github.com/tchaikov/fio
[fio.git] / os / os-linux.h
CommitLineData
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"
e2e58886 30#include "../file.h"
f3de88a7 31
7922a7b7
OM
32#ifndef __has_builtin // Optional of course.
33 #define __has_builtin(x) 0 // Compatibility with non-clang compilers.
34#endif
35
ebac4655
JA
36#define FIO_HAVE_CPU_AFFINITY
37#define FIO_HAVE_DISK_UTIL
38#define FIO_HAVE_SGIO
ba4f8923 39#define FIO_HAVE_IOPRIO
32ef447a 40#define FIO_HAVE_IOPRIO_CLASS
22f78b32 41#define FIO_HAVE_IOSCHED_SWITCH
2c0ecd28 42#define FIO_HAVE_ODIRECT
74b025b0 43#define FIO_HAVE_HUGETLB
07e5b264 44#define FIO_HAVE_RAWBIND
5e62c22a 45#define FIO_HAVE_BLKTRACE
eb7ccf38 46#define FIO_HAVE_CL_SIZE
a696fa2a 47#define FIO_HAVE_CGROUPS
2e3bd4c2 48#define FIO_HAVE_FS_STAT
a5f3027c 49#define FIO_HAVE_TRIM
47f767c1 50#define FIO_HAVE_GETTID
93bcfd20 51#define FIO_USE_GENERIC_INIT_RANDOM_STATE
6562685f 52#define FIO_HAVE_PWRITEV2
97900ebf 53#define FIO_HAVE_SHM_ATTACH_REMOVED
b6959b55 54
39b93568
JA
55#ifdef MAP_HUGETLB
56#define FIO_HAVE_MMAP_HUGE
57#endif
58
dc873b6f 59#define OS_MAP_ANON MAP_ANONYMOUS
ebac4655
JA
60
61typedef cpu_set_t os_cpu_mask_t;
4e78e405 62
6dfd46b9 63typedef struct drand48_data os_random_state_t;
ebac4655 64
67bf9823 65#ifdef CONFIG_3ARG_AFFINITY
e8462bd8
JA
66#define fio_setaffinity(pid, cpumask) \
67 sched_setaffinity((pid), sizeof(cpumask), &(cpumask))
ebac4655
JA
68#define fio_getaffinity(pid, ptr) \
69 sched_getaffinity((pid), sizeof(cpu_set_t), (ptr))
67bf9823 70#elif defined(CONFIG_2ARG_AFFINITY)
e8462bd8
JA
71#define fio_setaffinity(pid, cpumask) \
72 sched_setaffinity((pid), &(cpumask))
c8f025f8
JA
73#define fio_getaffinity(pid, ptr) \
74 sched_getaffinity((pid), (ptr))
75#endif
be4ecfdf 76
1f4c0a4f
JA
77#define fio_cpu_clear(mask, cpu) (void) CPU_CLR((cpu), (mask))
78#define fio_cpu_set(mask, cpu) (void) CPU_SET((cpu), (mask))
52fd65f4 79#define fio_cpu_isset(mask, cpu) (CPU_ISSET((cpu), (mask)) != 0)
d004a209 80#define fio_cpu_count(mask) CPU_COUNT((mask))
d2ce18b5
JA
81
82static inline int fio_cpuset_init(os_cpu_mask_t *mask)
83{
84 CPU_ZERO(mask);
85 return 0;
86}
87
88static inline int fio_cpuset_exit(os_cpu_mask_t *mask)
89{
90 return 0;
91}
6d459ee7
JA
92
93#define FIO_MAX_CPUS CPU_SETSIZE
4d8947de 94
28727df7
JA
95enum {
96 IOPRIO_CLASS_NONE,
97 IOPRIO_CLASS_RT,
98 IOPRIO_CLASS_BE,
99 IOPRIO_CLASS_IDLE,
100};
101
102enum {
103 IOPRIO_WHO_PROCESS = 1,
104 IOPRIO_WHO_PGRP,
105 IOPRIO_WHO_USER,
106};
107
108#define IOPRIO_BITS 16
109#define IOPRIO_CLASS_SHIFT 13
110
1767bd34
TK
111#define IOPRIO_MIN_PRIO 0 /* highest priority */
112#define IOPRIO_MAX_PRIO 7 /* lowest priority */
113
114#define IOPRIO_MIN_PRIO_CLASS 0
115#define IOPRIO_MAX_PRIO_CLASS 3
116
28727df7 117static inline int ioprio_set(int which, int who, int ioprio_class, int ioprio)
ebac4655 118{
28727df7
JA
119 /*
120 * If no class is set, assume BE
121 */
122 if (!ioprio_class)
123 ioprio_class = IOPRIO_CLASS_BE;
124
125 ioprio |= ioprio_class << IOPRIO_CLASS_SHIFT;
ebac4655
JA
126 return syscall(__NR_ioprio_set, which, who, ioprio);
127}
128
47f767c1
JA
129static inline int gettid(void)
130{
131 return syscall(__NR_gettid);
132}
133
3feedc60
JA
134#define SPLICE_DEF_SIZE (64*1024)
135
ebac4655
JA
136#ifndef BLKGETSIZE64
137#define BLKGETSIZE64 _IOR(0x12,114,size_t)
138#endif
139
e5b401d4
JA
140#ifndef BLKFLSBUF
141#define BLKFLSBUF _IO(0x12,97)
142#endif
143
a5f3027c
JA
144#ifndef BLKDISCARD
145#define BLKDISCARD _IO(0x12,119)
146#endif
147
e2e58886 148static inline int blockdev_invalidate_cache(struct fio_file *f)
e5b401d4 149{
ecc314ba 150 return ioctl(f->fd, BLKFLSBUF);
e5b401d4
JA
151}
152
ecc314ba 153static inline int blockdev_size(struct fio_file *f, unsigned long long *bytes)
ebac4655 154{
ecc314ba 155 if (!ioctl(f->fd, BLKGETSIZE64, bytes))
ebac4655
JA
156 return 0;
157
158 return errno;
159}
160
32cd46a0
JA
161static inline unsigned long long os_phys_mem(void)
162{
163 long pagesize, pages;
164
165 pagesize = sysconf(_SC_PAGESIZE);
166 pages = sysconf(_SC_PHYS_PAGES);
167 if (pages == -1 || pagesize == -1)
168 return 0;
169
170 return (unsigned long long) pages * (unsigned long long) pagesize;
171}
172
6dfd46b9
JA
173static inline void os_random_seed(unsigned long seed, os_random_state_t *rs)
174{
175 srand48_r(seed, rs);
176}
177
178static inline long os_random_long(os_random_state_t *rs)
179{
180 long val;
181
182 lrand48_r(rs, &val);
183 return val;
184}
185
8cc7afa9 186static inline int fio_lookup_raw(dev_t dev, int *majdev, int *mindev)
07e5b264
JA
187{
188 struct raw_config_request rq;
189 int fd;
190
191 if (major(dev) != RAW_MAJOR)
8cc7afa9 192 return 1;
07e5b264
JA
193
194 /*
195 * we should be able to find /dev/rawctl or /dev/raw/rawctl
196 */
197 fd = open("/dev/rawctl", O_RDONLY);
198 if (fd < 0) {
199 fd = open("/dev/raw/rawctl", O_RDONLY);
200 if (fd < 0)
8cc7afa9 201 return 1;
07e5b264
JA
202 }
203
204 rq.raw_minor = minor(dev);
205 if (ioctl(fd, RAW_GETBIND, &rq) < 0) {
206 close(fd);
8cc7afa9 207 return 1;
07e5b264
JA
208 }
209
210 close(fd);
211 *majdev = rq.block_major;
212 *mindev = rq.block_minor;
8cc7afa9 213 return 0;
07e5b264
JA
214}
215
5b6f5c66 216#ifdef O_NOATIME
5921e80c 217#define FIO_O_NOATIME O_NOATIME
5b6f5c66
JA
218#else
219#define FIO_O_NOATIME 0
220#endif
5921e80c 221
d01612f3
CM
222#ifdef O_ATOMIC
223#define OS_O_ATOMIC O_ATOMIC
224#else
225#define OS_O_ATOMIC 040000000
226#endif
227
a1c58075
JA
228#ifdef MADV_REMOVE
229#define FIO_MADV_FREE MADV_REMOVE
230#endif
231
7922a7b7
OM
232/* Check for GCC or Clang byte swap intrinsics */
233#if (__has_builtin(__builtin_bswap16) && __has_builtin(__builtin_bswap32) \
234 && __has_builtin(__builtin_bswap64)) || (__GNUC__ > 4 \
235 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)) /* fio_swapN */
df73aab1 236#define fio_swap16(x) __builtin_bswap16(x)
df73aab1 237#define fio_swap32(x) __builtin_bswap32(x)
df73aab1
JA
238#define fio_swap64(x) __builtin_bswap64(x)
239#else
7922a7b7
OM
240#include <byteswap.h>
241#define fio_swap16(x) bswap_16(x)
242#define fio_swap32(x) bswap_32(x)
243#define fio_swap64(x) bswap_64(x)
244#endif /* fio_swapN */
ff245192 245
eb7ccf38
JA
246#define CACHE_LINE_FILE \
247 "/sys/devices/system/cpu/cpu0/cache/index0/coherency_line_size"
248
249static inline int arch_cache_line_size(void)
250{
251 char size[32];
252 int fd, ret;
253
254 fd = open(CACHE_LINE_FILE, O_RDONLY);
255 if (fd < 0)
256 return -1;
257
258 ret = read(fd, size, sizeof(size));
ec76f15a
JA
259
260 close(fd);
261
eb7ccf38
JA
262 if (ret <= 0)
263 return -1;
ec76f15a
JA
264 else
265 return atoi(size);
eb7ccf38
JA
266}
267
d1419c4f
TK
268#ifdef __powerpc64__
269#define FIO_HAVE_CPU_ONLINE_SYSCONF
270static inline unsigned int cpus_online(void)
271{
272 return sysconf(_SC_NPROCESSORS_CONF);
273}
274#endif
275
c08ad04c 276static inline unsigned long long get_fs_free_size(const char *path)
2e3bd4c2
JA
277{
278 unsigned long long ret;
279 struct statfs s;
280
281 if (statfs(path, &s) < 0)
282 return -1ULL;
283
284 ret = s.f_bsize;
285 ret *= (unsigned long long) s.f_bfree;
286 return ret;
287}
288
496b1f9e 289static inline int os_trim(struct fio_file *f, unsigned long long start,
a5f3027c
JA
290 unsigned long long len)
291{
292 uint64_t range[2];
293
294 range[0] = start;
295 range[1] = len;
296
496b1f9e 297 if (!ioctl(f->fd, BLKDISCARD, range))
a5f3027c
JA
298 return 0;
299
300 return errno;
301}
302
7e09a9f1 303#ifdef CONFIG_SCHED_IDLE
f2a2ce0e
HL
304static inline int fio_set_sched_idle(void)
305{
306 struct sched_param p = { .sched_priority = 0, };
307 return sched_setscheduler(gettid(), SCHED_IDLE, &p);
308}
7e09a9f1 309#endif
f2a2ce0e 310
ae8e559e
JA
311#ifndef F_GET_RW_HINT
312#ifndef F_LINUX_SPECIFIC_BASE
313#define F_LINUX_SPECIFIC_BASE 1024
314#endif
315#define F_GET_RW_HINT (F_LINUX_SPECIFIC_BASE + 11)
316#define F_SET_RW_HINT (F_LINUX_SPECIFIC_BASE + 12)
bd553af6
JA
317#define F_GET_FILE_RW_HINT (F_LINUX_SPECIFIC_BASE + 13)
318#define F_SET_FILE_RW_HINT (F_LINUX_SPECIFIC_BASE + 14)
ae8e559e
JA
319#endif
320
321#ifndef RWH_WRITE_LIFE_NONE
bd553af6
JA
322#define RWH_WRITE_LIFE_NOT_SET 0
323#define RWH_WRITE_LIFE_NONE 1
324#define RWH_WRITE_LIFE_SHORT 2
325#define RWH_WRITE_LIFE_MEDIUM 3
326#define RWH_WRITE_LIFE_LONG 4
327#define RWH_WRITE_LIFE_EXTREME 5
37659335
JA
328#endif
329
ae8e559e 330#define FIO_HAVE_WRITE_HINT
37659335 331
6562685f
JA
332#ifndef RWF_HIPRI
333#define RWF_HIPRI 0x00000001
334#endif
335#ifndef RWF_DSYNC
336#define RWF_DSYNC 0x00000002
337#endif
338#ifndef RWF_SYNC
339#define RWF_SYNC 0x00000004
340#endif
341
72cd4146
JA
342#ifndef RWF_WRITE_LIFE_SHIFT
343#define RWF_WRITE_LIFE_SHIFT 4
344#define RWF_WRITE_LIFE_SHORT (1 << RWF_WRITE_LIFE_SHIFT)
345#define RWF_WRITE_LIFE_MEDIUM (2 << RWF_WRITE_LIFE_SHIFT)
346#define RWF_WRITE_LIFE_LONG (3 << RWF_WRITE_LIFE_SHIFT)
347#define RWF_WRITE_LIFE_EXTREME (4 << RWF_WRITE_LIFE_SHIFT)
348#endif
349
6562685f
JA
350#ifndef CONFIG_PWRITEV2
351#ifdef __NR_preadv2
352static inline void make_pos_h_l(unsigned long *pos_h, unsigned long *pos_l,
353 off_t offset)
354{
7b5c648f 355#if BITS_PER_LONG == 64
7b5c648f
JA
356 *pos_l = offset;
357 *pos_h = 0;
358#else
6562685f
JA
359 *pos_l = offset & 0xffffffff;
360 *pos_h = ((uint64_t) offset) >> 32;
7b5c648f 361#endif
6562685f
JA
362}
363static inline ssize_t preadv2(int fd, const struct iovec *iov, int iovcnt,
364 off_t offset, unsigned int flags)
365{
366 unsigned long pos_l, pos_h;
367
368 make_pos_h_l(&pos_h, &pos_l, offset);
369 return syscall(__NR_preadv2, fd, iov, iovcnt, pos_l, pos_h, flags);
370}
371static inline ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt,
372 off_t offset, unsigned int flags)
373{
374 unsigned long pos_l, pos_h;
375
376 make_pos_h_l(&pos_h, &pos_l, offset);
377 return syscall(__NR_pwritev2, fd, iov, iovcnt, pos_l, pos_h, flags);
378}
379#else
380static inline ssize_t preadv2(int fd, const struct iovec *iov, int iovcnt,
381 off_t offset, unsigned int flags)
382{
383 errno = ENOSYS;
384 return -1;
385}
386static inline ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt,
387 off_t offset, unsigned int flags)
388{
389 errno = ENOSYS;
390 return -1;
391}
392#endif /* __NR_preadv2 */
393#endif /* CONFIG_PWRITEV2 */
394
0cfe2089
TK
395static inline int shm_attach_to_open_removed(void)
396{
397 return 1;
398}
399
2c3e17be
SW
400#ifdef CONFIG_LINUX_FALLOCATE
401#define FIO_HAVE_NATIVE_FALLOCATE
402static inline bool fio_fallocate(struct fio_file *f, uint64_t offset,
403 uint64_t len)
404{
405 int ret;
e69dfc2b 406 ret = fallocate(f->fd, 0, offset, len);
2c3e17be
SW
407 if (ret == 0)
408 return true;
409
410 /* Work around buggy old glibc versions... */
411 if (ret > 0)
412 errno = ret;
413
414 return false;
415}
416#endif
417
58828d07
SW
418#define FIO_HAVE_CPU_HAS
419static inline bool os_cpu_has(cpu_features feature)
420{
421 bool have_feature;
422 unsigned long fio_unused hwcap;
423
424 switch (feature) {
425#ifdef ARCH_HAVE_CRC_CRYPTO
426 case CPU_ARM64_CRC32C:
427 hwcap = getauxval(AT_HWCAP);
428 have_feature = (hwcap & HWCAP_CRC32) != 0;
429 break;
430#endif
431 default:
432 have_feature = false;
433 }
434
435 return have_feature;
436}
437
ebac4655 438#endif