t/aio-ring: use syscall defines
[fio.git] / t / aio-ring.c
1 /*
2  * gcc -D_GNU_SOURCE -Wall -O2 -o aio-ring aio-ring.c  -lpthread -laio
3  */
4 #include <stdio.h>
5 #include <errno.h>
6 #include <assert.h>
7 #include <stdlib.h>
8 #include <stddef.h>
9 #include <signal.h>
10 #include <inttypes.h>
11
12 #include <sys/types.h>
13 #include <sys/stat.h>
14 #include <sys/ioctl.h>
15 #include <sys/syscall.h>
16 #include <sys/resource.h>
17 #include <sys/mman.h>
18 #include <linux/fs.h>
19 #include <fcntl.h>
20 #include <unistd.h>
21 #include <libaio.h>
22 #include <string.h>
23 #include <pthread.h>
24 #include <sched.h>
25
26 #include "../arch/arch.h"
27
28 #define IOCTX_FLAG_SCQRING      (1 << 0)        /* Use SQ/CQ rings */
29 #define IOCTX_FLAG_IOPOLL       (1 << 1)
30 #define IOCTX_FLAG_FIXEDBUFS    (1 << 2)
31 #define IOCTX_FLAG_SQTHREAD     (1 << 3)        /* Use SQ thread */
32 #define IOCTX_FLAG_SQWQ         (1 << 4)        /* Use SQ wq */
33 #define IOCTX_FLAG_SQPOLL       (1 << 5)
34
35 #define IOEV_RES2_CACHEHIT      (1 << 0)
36
37 #define barrier()       __asm__ __volatile__("": : :"memory")
38
39 #define min(a, b)               ((a < b) ? (a) : (b))
40
41 typedef uint64_t u64;
42 typedef uint32_t u32;
43 typedef uint16_t u16;
44
45 #define IORING_OFF_SQ_RING      0ULL
46 #define IORING_OFF_CQ_RING      0x8000000ULL
47 #define IORING_OFF_IOCB         0x10000000ULL
48
49 struct aio_uring_offsets {
50         u32 head;
51         u32 tail;
52         u32 ring_mask;
53         u32 ring_entries;
54         u32 flags;
55         u32 elems;
56 };
57
58 struct aio_uring_params {
59         u32 sq_entries;
60         u32 cq_entries;
61         u32 flags;
62         u16 sq_thread_cpu;
63         u16 resv[9];
64         struct aio_uring_offsets sq_off;
65         struct aio_uring_offsets cq_off;
66 };
67
68 struct aio_sq_ring {
69         u32 *head;
70         u32 *tail;
71         u32 *ring_mask;
72         u32 *ring_entries;
73         u32 *array;
74 };
75
76 struct aio_cq_ring {
77         u32 *head;
78         u32 *tail;
79         u32 *ring_mask;
80         u32 *ring_entries;
81         struct io_event *events;
82 };
83
84 #define IORING_ENTER_GETEVENTS  (1 << 0)
85
86 #define DEPTH                   32
87
88 #define BATCH_SUBMIT            8
89 #define BATCH_COMPLETE          8
90
91 #define BS                      4096
92
93 static unsigned sq_ring_mask, cq_ring_mask;
94
95 struct submitter {
96         pthread_t thread;
97         unsigned long max_blocks;
98         int fd;
99         struct drand48_data rand;
100         struct aio_sq_ring sq_ring;
101         struct iocb *iocbs;
102         struct iovec iovecs[DEPTH];
103         struct aio_cq_ring cq_ring;
104         int inflight;
105         unsigned long reaps;
106         unsigned long done;
107         unsigned long calls;
108         unsigned long cachehit, cachemiss;
109         volatile int finish;
110         char filename[128];
111 };
112
113 static struct submitter submitters[1];
114 static volatile int finish;
115
116 static int polled = 0;          /* use IO polling */
117 static int fixedbufs = 0;       /* use fixed user buffers */
118 static int buffered = 1;        /* use buffered IO, not O_DIRECT */
119 static int sq_thread = 0;       /* use kernel submission thread */
120 static int sq_thread_cpu = 0;   /* pin above thread to this CPU */
121
122 static int io_uring_setup(unsigned entries, struct iovec *iovecs,
123                           struct aio_uring_params *p)
124 {
125         return syscall(__NR_sys_io_uring_setup, entries, iovecs, p);
126 }
127
128 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
129                           unsigned int min_complete, unsigned int flags)
130 {
131         return syscall(__NR_sys_io_uring_enter, s->fd, to_submit, min_complete,
132                         flags);
133 }
134
135 static int gettid(void)
136 {
137         return syscall(__NR_gettid);
138 }
139
140 static void init_io(struct submitter *s, int fd, unsigned index)
141 {
142         struct iocb *iocb = &s->iocbs[index];
143         unsigned long offset;
144         long r;
145
146         lrand48_r(&s->rand, &r);
147         offset = (r % (s->max_blocks - 1)) * BS;
148
149         iocb->aio_fildes = fd;
150         iocb->aio_lio_opcode = IO_CMD_PREAD;
151         iocb->u.c.buf = s->iovecs[index].iov_base;
152         iocb->u.c.nbytes = BS;
153         iocb->u.c.offset = offset;
154 }
155
156 static int prep_more_ios(struct submitter *s, int fd, int max_ios)
157 {
158         struct aio_sq_ring *ring = &s->sq_ring;
159         u32 index, tail, next_tail, prepped = 0;
160
161         next_tail = tail = *ring->tail;
162         do {
163                 next_tail++;
164                 barrier();
165                 if (next_tail == *ring->head)
166                         break;
167
168                 index = tail & sq_ring_mask;
169                 init_io(s, fd, index);
170                 ring->array[index] = index;
171                 prepped++;
172                 tail = next_tail;
173         } while (prepped < max_ios);
174
175         if (*ring->tail != tail) {
176                 /* order tail store with writes to iocbs above */
177                 barrier();
178                 *ring->tail = tail;
179                 barrier();
180         }
181         return prepped;
182 }
183
184 static int get_file_size(int fd, unsigned long *blocks)
185 {
186         struct stat st;
187
188         if (fstat(fd, &st) < 0)
189                 return -1;
190         if (S_ISBLK(st.st_mode)) {
191                 unsigned long long bytes;
192
193                 if (ioctl(fd, BLKGETSIZE64, &bytes) != 0)
194                         return -1;
195
196                 *blocks = bytes / BS;
197                 return 0;
198         } else if (S_ISREG(st.st_mode)) {
199                 *blocks = st.st_size / BS;
200                 return 0;
201         }
202
203         return -1;
204 }
205
206 static int reap_events(struct submitter *s)
207 {
208         struct aio_cq_ring *ring = &s->cq_ring;
209         struct io_event *ev;
210         u32 head, reaped = 0;
211
212         head = *ring->head;
213         do {
214                 barrier();
215                 if (head == *ring->tail)
216                         break;
217                 ev = &ring->events[head & cq_ring_mask];
218                 if (ev->res != BS) {
219                         struct iocb *iocb = ev->obj;
220
221                         printf("io: unexpected ret=%ld\n", ev->res);
222                         printf("offset=%lu, size=%lu\n", (unsigned long) iocb->u.c.offset, (unsigned long) iocb->u.c.nbytes);
223                         return -1;
224                 }
225                 if (ev->res2 & IOEV_RES2_CACHEHIT)
226                         s->cachehit++;
227                 else
228                         s->cachemiss++;
229                 reaped++;
230                 head++;
231         } while (1);
232
233         s->inflight -= reaped;
234         *ring->head = head;
235         barrier();
236         return reaped;
237 }
238
239 static void *submitter_fn(void *data)
240 {
241         struct submitter *s = data;
242         int fd, ret, prepped, flags;
243
244         printf("submitter=%d\n", gettid());
245
246         flags = O_RDONLY;
247         if (!buffered)
248                 flags |= O_DIRECT;
249         fd = open(s->filename, flags);
250         if (fd < 0) {
251                 perror("open");
252                 goto done;
253         }
254
255         if (get_file_size(fd, &s->max_blocks)) {
256                 printf("failed getting size of device/file\n");
257                 goto err;
258         }
259         if (!s->max_blocks) {
260                 printf("Zero file/device size?\n");
261                 goto err;
262         }
263
264         s->max_blocks--;
265
266         srand48_r(pthread_self(), &s->rand);
267
268         prepped = 0;
269         do {
270                 int to_wait, to_submit, this_reap;
271
272                 if (!prepped && s->inflight < DEPTH)
273                         prepped = prep_more_ios(s, fd, min(DEPTH - s->inflight, BATCH_SUBMIT));
274                 s->inflight += prepped;
275 submit_more:
276                 to_submit = prepped;
277 submit:
278                 if (s->inflight + BATCH_SUBMIT < DEPTH)
279                         to_wait = 0;
280                 else
281                         to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
282
283                 ret = io_uring_enter(s, to_submit, to_wait, IORING_ENTER_GETEVENTS);
284                 s->calls++;
285
286                 this_reap = reap_events(s);
287                 if (this_reap == -1)
288                         break;
289                 s->reaps += this_reap;
290
291                 if (ret >= 0) {
292                         if (!ret) {
293                                 to_submit = 0;
294                                 if (s->inflight)
295                                         goto submit;
296                                 continue;
297                         } else if (ret < to_submit) {
298                                 int diff = to_submit - ret;
299
300                                 s->done += ret;
301                                 prepped -= diff;
302                                 goto submit_more;
303                         }
304                         s->done += ret;
305                         prepped = 0;
306                         continue;
307                 } else if (ret < 0) {
308                         if ((ret == -1 && errno == EAGAIN) || ret == -EAGAIN) {
309                                 if (s->finish)
310                                         break;
311                                 if (this_reap)
312                                         goto submit;
313                                 to_submit = 0;
314                                 goto submit;
315                         }
316                         if (ret == -1)
317                                 printf("io_submit: %s\n", strerror(errno));
318                         else
319                                 printf("io_submit: %s\n", strerror(-ret));
320                         break;
321                 }
322         } while (!s->finish);
323 err:
324         close(fd);
325 done:
326         finish = 1;
327         return NULL;
328 }
329
330 static void sig_int(int sig)
331 {
332         printf("Exiting on signal %d\n", sig);
333         submitters[0].finish = 1;
334         finish = 1;
335 }
336
337 static void arm_sig_int(void)
338 {
339         struct sigaction act;
340
341         memset(&act, 0, sizeof(act));
342         act.sa_handler = sig_int;
343         act.sa_flags = SA_RESTART;
344         sigaction(SIGINT, &act, NULL);
345 }
346
347 static int setup_ring(struct submitter *s)
348 {
349         struct aio_sq_ring *sring = &s->sq_ring;
350         struct aio_cq_ring *cring = &s->cq_ring;
351         struct aio_uring_params p;
352         void *ptr;
353         int fd;
354
355         memset(&p, 0, sizeof(p));
356
357         p.flags = IOCTX_FLAG_SCQRING;
358         if (polled)
359                 p.flags |= IOCTX_FLAG_IOPOLL;
360         if (fixedbufs)
361                 p.flags |= IOCTX_FLAG_FIXEDBUFS;
362         if (buffered)
363                 p.flags |= IOCTX_FLAG_SQWQ;
364         else if (sq_thread) {
365                 p.flags |= IOCTX_FLAG_SQTHREAD;
366                 p.sq_thread_cpu = sq_thread_cpu;
367         }
368
369         if (fixedbufs)
370                 fd = io_uring_setup(DEPTH, s->iovecs, &p);
371         else
372                 fd = io_uring_setup(DEPTH, NULL, &p);
373         if (fd < 0) {
374                 perror("io_uring_setup");
375                 return 1;
376         }
377
378         s->fd = fd;
379
380         ptr = mmap(0, p.sq_off.elems + p.sq_entries * sizeof(u32),
381                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE,
382                         fd, IORING_OFF_SQ_RING);
383         printf("sq_ring ptr = 0x%p\n", ptr);
384         sring->head = ptr + p.sq_off.head;
385         sring->tail = ptr + p.sq_off.tail;
386         sring->ring_mask = ptr + p.sq_off.ring_mask;
387         sring->ring_entries = ptr + p.sq_off.ring_entries;
388         sring->array = ptr + p.sq_off.elems;
389         sq_ring_mask = *sring->ring_mask;
390
391         s->iocbs = mmap(0, p.sq_entries * sizeof(struct iocb), PROT_READ | PROT_WRITE,
392                         MAP_SHARED | MAP_POPULATE, fd, IORING_OFF_IOCB);
393         printf("iocbs ptr   = 0x%p\n", s->iocbs);
394
395         ptr = mmap(0, p.cq_off.elems + p.cq_entries * sizeof(struct io_event),
396                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE,
397                         fd, IORING_OFF_CQ_RING);
398         printf("cq_ring ptr = 0x%p\n", ptr);
399         cring->head = ptr + p.cq_off.head;
400         cring->tail = ptr + p.cq_off.tail;
401         cring->ring_mask = ptr + p.cq_off.ring_mask;
402         cring->ring_entries = ptr + p.cq_off.ring_entries;
403         cring->events = ptr + p.cq_off.elems;
404         cq_ring_mask = *cring->ring_mask;
405         return 0;
406 }
407
408 int main(int argc, char *argv[])
409 {
410         struct submitter *s = &submitters[0];
411         unsigned long done, calls, reap, cache_hit, cache_miss;
412         int err, i;
413         struct rlimit rlim;
414         void *ret;
415
416         if (argc < 2) {
417                 printf("%s: filename\n", argv[0]);
418                 return 1;
419         }
420
421         rlim.rlim_cur = RLIM_INFINITY;
422         rlim.rlim_max = RLIM_INFINITY;
423         if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
424                 perror("setrlimit");
425                 return 1;
426         }
427
428         arm_sig_int();
429
430         for (i = 0; i < DEPTH; i++) {
431                 void *buf;
432
433                 if (posix_memalign(&buf, BS, BS)) {
434                         printf("failed alloc\n");
435                         return 1;
436                 }
437                 s->iovecs[i].iov_base = buf;
438                 s->iovecs[i].iov_len = BS;
439         }
440
441         err = setup_ring(s);
442         if (err) {
443                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
444                 return 1;
445         }
446         printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
447         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
448         strcpy(s->filename, argv[1]);
449
450         pthread_create(&s->thread, NULL, submitter_fn, s);
451
452         cache_hit = cache_miss = reap = calls = done = 0;
453         do {
454                 unsigned long this_done = 0;
455                 unsigned long this_reap = 0;
456                 unsigned long this_call = 0;
457                 unsigned long this_cache_hit = 0;
458                 unsigned long this_cache_miss = 0;
459                 unsigned long rpc = 0, ipc = 0;
460                 double hit = 0.0;
461
462                 sleep(1);
463                 this_done += s->done;
464                 this_call += s->calls;
465                 this_reap += s->reaps;
466                 this_cache_hit += s->cachehit;
467                 this_cache_miss += s->cachemiss;
468                 if (this_cache_hit && this_cache_miss) {
469                         unsigned long hits, total;
470
471                         hits = this_cache_hit - cache_hit;
472                         total = hits + this_cache_miss - cache_miss;
473                         hit = (double) hits / (double) total;
474                         hit *= 100.0;
475                 }
476                 if (this_call - calls) {
477                         rpc = (this_done - done) / (this_call - calls);
478                         ipc = (this_reap - reap) / (this_call - calls);
479                 }
480                 printf("IOPS=%lu, IOS/call=%lu/%lu, inflight=%u (head=%u tail=%u), Cachehit=%0.2f%%\n",
481                                 this_done - done, rpc, ipc, s->inflight,
482                                 *s->cq_ring.head, *s->cq_ring.tail, hit);
483                 done = this_done;
484                 calls = this_call;
485                 reap = this_reap;
486                 cache_hit = s->cachehit;
487                 cache_miss = s->cachemiss;
488         } while (!finish);
489
490         pthread_join(s->thread, &ret);
491         return 0;
492 }