docs: update for new data placement options
[fio.git] / os / os-linux.h
CommitLineData
ebac4655
JA
1#ifndef FIO_OS_LINUX_H
2#define FIO_OS_LINUX_H
3
910f378a
BVA
4#ifdef __ANDROID__
5#define FIO_OS os_android
6#else
cca84643 7#define FIO_OS os_linux
910f378a 8#endif
cca84643 9
ebac4655 10#include <sys/ioctl.h>
8756e4d4 11#include <sys/uio.h>
3c39a379 12#include <sys/syscall.h>
250e878a 13#include <sys/sysmacros.h>
2e3bd4c2 14#include <sys/vfs.h>
39b93568 15#include <sys/mman.h>
3c39a379 16#include <unistd.h>
97fbdfa1 17#include <fcntl.h>
3c2d93ed 18#include <errno.h>
d09913c2 19#include <sched.h>
ea421790 20#include <linux/unistd.h>
07e5b264 21#include <linux/major.h>
1894d171
TK
22#include <linux/fs.h>
23#include <scsi/sg.h>
e69d8385 24#include <asm/byteorder.h>
910f378a
BVA
25#ifdef __ANDROID__
26#include "os-ashmem.h"
27#define FIO_NO_HAVE_SHM_H
28#endif
ebac4655 29
58828d07
SW
30#ifdef ARCH_HAVE_CRC_CRYPTO
31#include <sys/auxv.h>
3721c7fe
SW
32#ifndef HWCAP_PMULL
33#define HWCAP_PMULL (1 << 4)
34#endif /* HWCAP_PMULL */
58828d07
SW
35#ifndef HWCAP_CRC32
36#define HWCAP_CRC32 (1 << 7)
37#endif /* HWCAP_CRC32 */
38#endif /* ARCH_HAVE_CRC_CRYPTO */
39
1c764dbe 40#include "./os-linux-syscall.h"
e2e58886 41#include "../file.h"
f3de88a7 42
7922a7b7
OM
43#ifndef __has_builtin // Optional of course.
44 #define __has_builtin(x) 0 // Compatibility with non-clang compilers.
45#endif
46
ebac4655
JA
47#define FIO_HAVE_CPU_AFFINITY
48#define FIO_HAVE_DISK_UTIL
49#define FIO_HAVE_SGIO
ba4f8923 50#define FIO_HAVE_IOPRIO
32ef447a 51#define FIO_HAVE_IOPRIO_CLASS
22f78b32 52#define FIO_HAVE_IOSCHED_SWITCH
2c0ecd28 53#define FIO_HAVE_ODIRECT
74b025b0 54#define FIO_HAVE_HUGETLB
5e62c22a 55#define FIO_HAVE_BLKTRACE
eb7ccf38 56#define FIO_HAVE_CL_SIZE
a696fa2a 57#define FIO_HAVE_CGROUPS
2e3bd4c2 58#define FIO_HAVE_FS_STAT
a5f3027c 59#define FIO_HAVE_TRIM
47f767c1 60#define FIO_HAVE_GETTID
93bcfd20 61#define FIO_USE_GENERIC_INIT_RANDOM_STATE
e69d8385 62#define FIO_HAVE_BYTEORDER_FUNCS
6562685f 63#define FIO_HAVE_PWRITEV2
97900ebf 64#define FIO_HAVE_SHM_ATTACH_REMOVED
b6959b55 65
39b93568
JA
66#ifdef MAP_HUGETLB
67#define FIO_HAVE_MMAP_HUGE
68#endif
69
dc873b6f 70#define OS_MAP_ANON MAP_ANONYMOUS
ebac4655 71
4db9ac71 72#define FIO_EXT_ENG_DIR "/usr/local/lib/fio"
5a8a6a03 73
ebac4655 74typedef cpu_set_t os_cpu_mask_t;
4e78e405 75
67bf9823 76#ifdef CONFIG_3ARG_AFFINITY
e8462bd8
JA
77#define fio_setaffinity(pid, cpumask) \
78 sched_setaffinity((pid), sizeof(cpumask), &(cpumask))
ebac4655
JA
79#define fio_getaffinity(pid, ptr) \
80 sched_getaffinity((pid), sizeof(cpu_set_t), (ptr))
67bf9823 81#elif defined(CONFIG_2ARG_AFFINITY)
e8462bd8
JA
82#define fio_setaffinity(pid, cpumask) \
83 sched_setaffinity((pid), &(cpumask))
c8f025f8
JA
84#define fio_getaffinity(pid, ptr) \
85 sched_getaffinity((pid), (ptr))
86#endif
be4ecfdf 87
e9d2a04d
TK
88#ifdef CONFIG_PTHREAD_GETAFFINITY
89#define FIO_HAVE_GET_THREAD_AFFINITY
874d55e5
JA
90#define fio_get_thread_affinity(mask) \
91 pthread_getaffinity_np(pthread_self(), sizeof(mask), &(mask))
e9d2a04d 92#endif
874d55e5 93
6e0cefdc
BVA
94#define fio_cpu_clear(mask, cpu) CPU_CLR((cpu), (mask))
95#define fio_cpu_set(mask, cpu) CPU_SET((cpu), (mask))
52fd65f4 96#define fio_cpu_isset(mask, cpu) (CPU_ISSET((cpu), (mask)) != 0)
d004a209 97#define fio_cpu_count(mask) CPU_COUNT((mask))
d2ce18b5
JA
98
99static inline int fio_cpuset_init(os_cpu_mask_t *mask)
100{
101 CPU_ZERO(mask);
102 return 0;
103}
104
105static inline int fio_cpuset_exit(os_cpu_mask_t *mask)
106{
107 return 0;
108}
6d459ee7
JA
109
110#define FIO_MAX_CPUS CPU_SETSIZE
4d8947de 111
28727df7
JA
112enum {
113 IOPRIO_CLASS_NONE,
114 IOPRIO_CLASS_RT,
115 IOPRIO_CLASS_BE,
116 IOPRIO_CLASS_IDLE,
117};
118
119enum {
120 IOPRIO_WHO_PROCESS = 1,
121 IOPRIO_WHO_PGRP,
122 IOPRIO_WHO_USER,
123};
124
125#define IOPRIO_BITS 16
126#define IOPRIO_CLASS_SHIFT 13
127
b3048c1b
DLM
128#define IOPRIO_HINT_BITS 10
129#define IOPRIO_HINT_SHIFT 3
130
1767bd34
TK
131#define IOPRIO_MIN_PRIO 0 /* highest priority */
132#define IOPRIO_MAX_PRIO 7 /* lowest priority */
133
134#define IOPRIO_MIN_PRIO_CLASS 0
135#define IOPRIO_MAX_PRIO_CLASS 3
136
b3048c1b
DLM
137#define IOPRIO_MIN_PRIO_HINT 0
138#define IOPRIO_MAX_PRIO_HINT ((1 << IOPRIO_HINT_BITS) - 1)
139
d4d503a5
DLM
140#define ioprio_class(ioprio) ((ioprio) >> IOPRIO_CLASS_SHIFT)
141#define ioprio(ioprio) ((ioprio) & IOPRIO_MAX_PRIO)
b3048c1b
DLM
142#define ioprio_hint(ioprio) \
143 (((ioprio) >> IOPRIO_HINT_SHIFT) & IOPRIO_MAX_PRIO_HINT)
d4d503a5 144
b3048c1b 145static inline int ioprio_value(int ioprio_class, int ioprio, int ioprio_hint)
ebac4655 146{
28727df7
JA
147 /*
148 * If no class is set, assume BE
149 */
8d6e8d99
DLM
150 if (!ioprio_class)
151 ioprio_class = IOPRIO_CLASS_BE;
152
b3048c1b
DLM
153 return (ioprio_class << IOPRIO_CLASS_SHIFT) |
154 (ioprio_hint << IOPRIO_HINT_SHIFT) |
155 ioprio;
8d6e8d99 156}
28727df7 157
03ec570f
DLM
158static inline bool ioprio_value_is_class_rt(unsigned int priority)
159{
d4d503a5 160 return ioprio_class(priority) == IOPRIO_CLASS_RT;
03ec570f
DLM
161}
162
b3048c1b
DLM
163static inline int ioprio_set(int which, int who, int ioprio_class, int ioprio,
164 int ioprio_hint)
8d6e8d99
DLM
165{
166 return syscall(__NR_ioprio_set, which, who,
b3048c1b 167 ioprio_value(ioprio_class, ioprio, ioprio_hint));
ebac4655
JA
168}
169
de5ed0e4 170#ifndef CONFIG_HAVE_GETTID
47f767c1
JA
171static inline int gettid(void)
172{
173 return syscall(__NR_gettid);
174}
de5ed0e4 175#endif
47f767c1 176
3feedc60
JA
177#define SPLICE_DEF_SIZE (64*1024)
178
ebac4655
JA
179#ifndef BLKGETSIZE64
180#define BLKGETSIZE64 _IOR(0x12,114,size_t)
181#endif
182
e5b401d4
JA
183#ifndef BLKFLSBUF
184#define BLKFLSBUF _IO(0x12,97)
185#endif
186
a5f3027c
JA
187#ifndef BLKDISCARD
188#define BLKDISCARD _IO(0x12,119)
189#endif
190
e2e58886 191static inline int blockdev_invalidate_cache(struct fio_file *f)
e5b401d4 192{
ecc314ba 193 return ioctl(f->fd, BLKFLSBUF);
e5b401d4
JA
194}
195
ecc314ba 196static inline int blockdev_size(struct fio_file *f, unsigned long long *bytes)
ebac4655 197{
ecc314ba 198 if (!ioctl(f->fd, BLKGETSIZE64, bytes))
ebac4655
JA
199 return 0;
200
201 return errno;
202}
203
32cd46a0
JA
204static inline unsigned long long os_phys_mem(void)
205{
206 long pagesize, pages;
207
208 pagesize = sysconf(_SC_PAGESIZE);
209 pages = sysconf(_SC_PHYS_PAGES);
210 if (pages == -1 || pagesize == -1)
211 return 0;
212
213 return (unsigned long long) pages * (unsigned long long) pagesize;
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
a1c58075
JA
222#ifdef MADV_REMOVE
223#define FIO_MADV_FREE MADV_REMOVE
224#endif
225
7922a7b7
OM
226/* Check for GCC or Clang byte swap intrinsics */
227#if (__has_builtin(__builtin_bswap16) && __has_builtin(__builtin_bswap32) \
228 && __has_builtin(__builtin_bswap64)) || (__GNUC__ > 4 \
229 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)) /* fio_swapN */
df73aab1 230#define fio_swap16(x) __builtin_bswap16(x)
df73aab1 231#define fio_swap32(x) __builtin_bswap32(x)
df73aab1
JA
232#define fio_swap64(x) __builtin_bswap64(x)
233#else
7922a7b7
OM
234#include <byteswap.h>
235#define fio_swap16(x) bswap_16(x)
236#define fio_swap32(x) bswap_32(x)
237#define fio_swap64(x) bswap_64(x)
238#endif /* fio_swapN */
ff245192 239
eb7ccf38
JA
240#define CACHE_LINE_FILE \
241 "/sys/devices/system/cpu/cpu0/cache/index0/coherency_line_size"
242
243static inline int arch_cache_line_size(void)
244{
245 char size[32];
246 int fd, ret;
247
248 fd = open(CACHE_LINE_FILE, O_RDONLY);
249 if (fd < 0)
250 return -1;
251
252 ret = read(fd, size, sizeof(size));
ec76f15a
JA
253
254 close(fd);
255
eb7ccf38
JA
256 if (ret <= 0)
257 return -1;
ec76f15a
JA
258 else
259 return atoi(size);
eb7ccf38
JA
260}
261
c08ad04c 262static inline unsigned long long get_fs_free_size(const char *path)
2e3bd4c2
JA
263{
264 unsigned long long ret;
265 struct statfs s;
266
267 if (statfs(path, &s) < 0)
268 return -1ULL;
269
270 ret = s.f_bsize;
271 ret *= (unsigned long long) s.f_bfree;
272 return ret;
273}
274
496b1f9e 275static inline int os_trim(struct fio_file *f, unsigned long long start,
a5f3027c
JA
276 unsigned long long len)
277{
278 uint64_t range[2];
279
280 range[0] = start;
281 range[1] = len;
282
496b1f9e 283 if (!ioctl(f->fd, BLKDISCARD, range))
a5f3027c
JA
284 return 0;
285
286 return errno;
287}
288
7e09a9f1 289#ifdef CONFIG_SCHED_IDLE
f2a2ce0e
HL
290static inline int fio_set_sched_idle(void)
291{
292 struct sched_param p = { .sched_priority = 0, };
293 return sched_setscheduler(gettid(), SCHED_IDLE, &p);
294}
7e09a9f1 295#endif
f2a2ce0e 296
ae8e559e
JA
297#ifndef F_GET_RW_HINT
298#ifndef F_LINUX_SPECIFIC_BASE
299#define F_LINUX_SPECIFIC_BASE 1024
300#endif
301#define F_GET_RW_HINT (F_LINUX_SPECIFIC_BASE + 11)
302#define F_SET_RW_HINT (F_LINUX_SPECIFIC_BASE + 12)
bd553af6
JA
303#define F_GET_FILE_RW_HINT (F_LINUX_SPECIFIC_BASE + 13)
304#define F_SET_FILE_RW_HINT (F_LINUX_SPECIFIC_BASE + 14)
ae8e559e
JA
305#endif
306
307#ifndef RWH_WRITE_LIFE_NONE
bd553af6
JA
308#define RWH_WRITE_LIFE_NOT_SET 0
309#define RWH_WRITE_LIFE_NONE 1
310#define RWH_WRITE_LIFE_SHORT 2
311#define RWH_WRITE_LIFE_MEDIUM 3
312#define RWH_WRITE_LIFE_LONG 4
313#define RWH_WRITE_LIFE_EXTREME 5
37659335
JA
314#endif
315
ae8e559e 316#define FIO_HAVE_WRITE_HINT
37659335 317
6562685f
JA
318#ifndef RWF_HIPRI
319#define RWF_HIPRI 0x00000001
320#endif
321#ifndef RWF_DSYNC
322#define RWF_DSYNC 0x00000002
323#endif
324#ifndef RWF_SYNC
325#define RWF_SYNC 0x00000004
326#endif
7d42e66e
KK
327#ifndef RWF_NOWAIT
328#define RWF_NOWAIT 0x00000008
329#endif
6562685f 330
4a87b584 331#ifndef RWF_UNCACHED
8beee62d 332#define RWF_UNCACHED 0x00000040
4a87b584
JA
333#endif
334
72cd4146
JA
335#ifndef RWF_WRITE_LIFE_SHIFT
336#define RWF_WRITE_LIFE_SHIFT 4
337#define RWF_WRITE_LIFE_SHORT (1 << RWF_WRITE_LIFE_SHIFT)
338#define RWF_WRITE_LIFE_MEDIUM (2 << RWF_WRITE_LIFE_SHIFT)
339#define RWF_WRITE_LIFE_LONG (3 << RWF_WRITE_LIFE_SHIFT)
340#define RWF_WRITE_LIFE_EXTREME (4 << RWF_WRITE_LIFE_SHIFT)
341#endif
342
6562685f
JA
343#ifndef CONFIG_PWRITEV2
344#ifdef __NR_preadv2
345static inline void make_pos_h_l(unsigned long *pos_h, unsigned long *pos_l,
346 off_t offset)
347{
7b5c648f 348#if BITS_PER_LONG == 64
7b5c648f
JA
349 *pos_l = offset;
350 *pos_h = 0;
351#else
6562685f
JA
352 *pos_l = offset & 0xffffffff;
353 *pos_h = ((uint64_t) offset) >> 32;
7b5c648f 354#endif
6562685f
JA
355}
356static inline ssize_t preadv2(int fd, const struct iovec *iov, int iovcnt,
357 off_t offset, unsigned int flags)
358{
359 unsigned long pos_l, pos_h;
360
361 make_pos_h_l(&pos_h, &pos_l, offset);
362 return syscall(__NR_preadv2, fd, iov, iovcnt, pos_l, pos_h, flags);
363}
364static inline ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt,
365 off_t offset, unsigned int flags)
366{
367 unsigned long pos_l, pos_h;
368
369 make_pos_h_l(&pos_h, &pos_l, offset);
370 return syscall(__NR_pwritev2, fd, iov, iovcnt, pos_l, pos_h, flags);
371}
372#else
373static inline ssize_t preadv2(int fd, const struct iovec *iov, int iovcnt,
374 off_t offset, unsigned int flags)
375{
376 errno = ENOSYS;
377 return -1;
378}
379static inline ssize_t pwritev2(int fd, const struct iovec *iov, int iovcnt,
380 off_t offset, unsigned int flags)
381{
382 errno = ENOSYS;
383 return -1;
384}
385#endif /* __NR_preadv2 */
386#endif /* CONFIG_PWRITEV2 */
387
0cfe2089
TK
388static inline int shm_attach_to_open_removed(void)
389{
390 return 1;
391}
392
2c3e17be
SW
393#ifdef CONFIG_LINUX_FALLOCATE
394#define FIO_HAVE_NATIVE_FALLOCATE
395static inline bool fio_fallocate(struct fio_file *f, uint64_t offset,
396 uint64_t len)
397{
398 int ret;
e69dfc2b 399 ret = fallocate(f->fd, 0, offset, len);
2c3e17be
SW
400 if (ret == 0)
401 return true;
402
403 /* Work around buggy old glibc versions... */
404 if (ret > 0)
405 errno = ret;
406
407 return false;
408}
409#endif
410
58828d07
SW
411#define FIO_HAVE_CPU_HAS
412static inline bool os_cpu_has(cpu_features feature)
413{
414 bool have_feature;
415 unsigned long fio_unused hwcap;
416
417 switch (feature) {
418#ifdef ARCH_HAVE_CRC_CRYPTO
419 case CPU_ARM64_CRC32C:
420 hwcap = getauxval(AT_HWCAP);
3721c7fe
SW
421 have_feature = (hwcap & (HWCAP_PMULL | HWCAP_CRC32)) ==
422 (HWCAP_PMULL | HWCAP_CRC32);
58828d07
SW
423 break;
424#endif
425 default:
426 have_feature = false;
427 }
428
429 return have_feature;
430}
431
ebac4655 432#endif