t/io_uring.c

   1 #include <stdio.h>
   2 #include <errno.h>
   3 #include <assert.h>
   4 #include <stdlib.h>
   5 #include <stddef.h>
   6 #include <signal.h>
   7 #include <inttypes.h>
   8
   9 #include <sys/types.h>
  10 #include <sys/stat.h>
  11 #include <sys/ioctl.h>
  12 #include <sys/syscall.h>
  13 #include <sys/resource.h>
  14 #include <sys/mman.h>
  15 #include <sys/uio.h>
  16 #include <linux/fs.h>
  17 #include <fcntl.h>
  18 #include <unistd.h>
  19 #include <string.h>
  20 #include <pthread.h>
  21 #include <sched.h>
  22
  23 #include "../arch/arch.h"
  24 #include "../lib/types.h"
  25 #include "../os/io_uring.h"
  26
  27 #define barrier()       __asm__ __volatile__("": : :"memory")
  28
  29 #define min(a, b)               ((a < b) ? (a) : (b))
  30
  31 struct io_sq_ring {
  32         unsigned *head;
  33         unsigned *tail;
  34         unsigned *ring_mask;
  35         unsigned *ring_entries;
  36         unsigned *array;
  37 };
  38
  39 struct io_cq_ring {
  40         unsigned *head;
  41         unsigned *tail;
  42         unsigned *ring_mask;
  43         unsigned *ring_entries;
  44         struct io_uring_event *events;
  45 };
  46
  47 #define DEPTH                   32
  48
  49 #define BATCH_SUBMIT            8
  50 #define BATCH_COMPLETE          8
  51
  52 #define BS                      4096
  53
  54 static unsigned sq_ring_mask, cq_ring_mask;
  55
  56 struct submitter {
  57         pthread_t thread;
  58         unsigned long max_blocks;
  59         int ring_fd;
  60         struct drand48_data rand;
  61         struct io_sq_ring sq_ring;
  62         struct io_uring_iocb *iocbs;
  63         struct iovec iovecs[DEPTH];
  64         struct io_cq_ring cq_ring;
  65         int inflight;
  66         unsigned long reaps;
  67         unsigned long done;
  68         unsigned long calls;
  69         unsigned long cachehit, cachemiss;
  70         volatile int finish;
  71         char filename[128];
  72 };
  73
  74 static struct submitter submitters[1];
  75 static volatile int finish;
  76
  77 static int polled = 0;          /* use IO polling */
  78 static int fixedbufs = 0;       /* use fixed user buffers */
  79 static int buffered = 1;        /* use buffered IO, not O_DIRECT */
  80 static int sq_thread = 0;       /* use kernel submission thread */
  81 static int sq_thread_cpu = 0;   /* pin above thread to this CPU */
  82
  83 static int io_uring_setup(unsigned entries, struct iovec *iovecs,
  84                           struct io_uring_params *p)
  85 {
  86         return syscall(__NR_sys_io_uring_setup, entries, iovecs, p);
  87 }
  88
  89 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
  90                           unsigned int min_complete, unsigned int flags)
  91 {
  92         return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit,
  93                         min_complete, flags);
  94 }
  95
  96 static int gettid(void)
  97 {
  98         return syscall(__NR_gettid);
  99 }
 100
 101 static void init_io(struct submitter *s, int fd, unsigned index)
 102 {
 103         struct io_uring_iocb *iocb = &s->iocbs[index];
 104         unsigned long offset;
 105         long r;
 106
 107         lrand48_r(&s->rand, &r);
 108         offset = (r % (s->max_blocks - 1)) * BS;
 109
 110         if (fixedbufs)
 111                 iocb->opcode = IORING_OP_READ_FIXED;
 112         else
 113                 iocb->opcode = IORING_OP_READ;
 114         iocb->flags = 0;
 115         iocb->ioprio = 0;
 116         iocb->fd = fd;
 117         iocb->off = offset;
 118         iocb->addr = s->iovecs[index].iov_base;
 119         iocb->len = BS;
 120 }
 121
 122 static int prep_more_ios(struct submitter *s, int fd, int max_ios)
 123 {
 124         struct io_sq_ring *ring = &s->sq_ring;
 125         unsigned index, tail, next_tail, prepped = 0;
 126
 127         next_tail = tail = *ring->tail;
 128         do {
 129                 next_tail++;
 130                 barrier();
 131                 if (next_tail == *ring->head)
 132                         break;
 133
 134                 index = tail & sq_ring_mask;
 135                 init_io(s, fd, index);
 136                 ring->array[index] = index;
 137                 prepped++;
 138                 tail = next_tail;
 139         } while (prepped < max_ios);
 140
 141         if (*ring->tail != tail) {
 142                 /* order tail store with writes to iocbs above */
 143                 barrier();
 144                 *ring->tail = tail;
 145                 barrier();
 146         }
 147         return prepped;
 148 }
 149
 150 static int get_file_size(int fd, unsigned long *blocks)
 151 {
 152         struct stat st;
 153
 154         if (fstat(fd, &st) < 0)
 155                 return -1;
 156         if (S_ISBLK(st.st_mode)) {
 157                 unsigned long long bytes;
 158
 159                 if (ioctl(fd, BLKGETSIZE64, &bytes) != 0)
 160                         return -1;
 161
 162                 *blocks = bytes / BS;
 163                 return 0;
 164         } else if (S_ISREG(st.st_mode)) {
 165                 *blocks = st.st_size / BS;
 166                 return 0;
 167         }
 168
 169         return -1;
 170 }
 171
 172 static int reap_events(struct submitter *s)
 173 {
 174         struct io_cq_ring *ring = &s->cq_ring;
 175         struct io_uring_event *ev;
 176         unsigned head, reaped = 0;
 177
 178         head = *ring->head;
 179         do {
 180                 barrier();
 181                 if (head == *ring->tail)
 182                         break;
 183                 ev = &ring->events[head & cq_ring_mask];
 184                 if (ev->res != BS) {
 185                         struct io_uring_iocb *iocb = &s->iocbs[ev->index];
 186
 187                         printf("io: unexpected ret=%d\n", ev->res);
 188                         printf("offset=%lu, size=%lu\n",
 189                                         (unsigned long) iocb->off,
 190                                         (unsigned long) iocb->len);
 191                         return -1;
 192                 }
 193                 if (ev->flags & IOEV_FLAG_CACHEHIT)
 194                         s->cachehit++;
 195                 else
 196                         s->cachemiss++;
 197                 reaped++;
 198                 head++;
 199         } while (1);
 200
 201         s->inflight -= reaped;
 202         *ring->head = head;
 203         barrier();
 204         return reaped;
 205 }
 206
 207 static void *submitter_fn(void *data)
 208 {
 209         struct submitter *s = data;
 210         int fd, ret, prepped, flags;
 211
 212         printf("submitter=%d\n", gettid());
 213
 214         flags = O_RDONLY;
 215         if (!buffered)
 216                 flags |= O_DIRECT;
 217         fd = open(s->filename, flags);
 218         if (fd < 0) {
 219                 perror("open");
 220                 goto done;
 221         }
 222
 223         if (get_file_size(fd, &s->max_blocks)) {
 224                 printf("failed getting size of device/file\n");
 225                 goto err;
 226         }
 227         if (s->max_blocks <= 1) {
 228                 printf("Zero file/device size?\n");
 229                 goto err;
 230         }
 231         s->max_blocks--;
 232
 233         srand48_r(pthread_self(), &s->rand);
 234
 235         prepped = 0;
 236         do {
 237                 int to_wait, to_submit, this_reap, to_prep;
 238
 239                 if (!prepped && s->inflight < DEPTH) {
 240                         to_prep = min(DEPTH - s->inflight, BATCH_SUBMIT);
 241                         prepped = prep_more_ios(s, fd, to_prep);
 242                 }
 243                 s->inflight += prepped;
 244 submit_more:
 245                 to_submit = prepped;
 246 submit:
 247                 if (s->inflight + BATCH_SUBMIT < DEPTH)
 248                         to_wait = 0;
 249                 else
 250                         to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
 251
 252                 ret = io_uring_enter(s, to_submit, to_wait,
 253                                         IORING_ENTER_GETEVENTS);
 254                 s->calls++;
 255
 256                 this_reap = reap_events(s);
 257                 if (this_reap == -1)
 258                         break;
 259                 s->reaps += this_reap;
 260
 261                 if (ret >= 0) {
 262                         if (!ret) {
 263                                 to_submit = 0;
 264                                 if (s->inflight)
 265                                         goto submit;
 266                                 continue;
 267                         } else if (ret < to_submit) {
 268                                 int diff = to_submit - ret;
 269
 270                                 s->done += ret;
 271                                 prepped -= diff;
 272                                 goto submit_more;
 273                         }
 274                         s->done += ret;
 275                         prepped = 0;
 276                         continue;
 277                 } else if (ret < 0) {
 278                         if (errno == EAGAIN) {
 279                                 if (s->finish)
 280                                         break;
 281                                 if (this_reap)
 282                                         goto submit;
 283                                 to_submit = 0;
 284                                 goto submit;
 285                         }
 286                         printf("io_submit: %s\n", strerror(errno));
 287                         break;
 288                 }
 289         } while (!s->finish);
 290 err:
 291         close(fd);
 292 done:
 293         finish = 1;
 294         return NULL;
 295 }
 296
 297 static void sig_int(int sig)
 298 {
 299         printf("Exiting on signal %d\n", sig);
 300         submitters[0].finish = 1;
 301         finish = 1;
 302 }
 303
 304 static void arm_sig_int(void)
 305 {
 306         struct sigaction act;
 307
 308         memset(&act, 0, sizeof(act));
 309         act.sa_handler = sig_int;
 310         act.sa_flags = SA_RESTART;
 311         sigaction(SIGINT, &act, NULL);
 312 }
 313
 314 static int setup_ring(struct submitter *s)
 315 {
 316         struct io_sq_ring *sring = &s->sq_ring;
 317         struct io_cq_ring *cring = &s->cq_ring;
 318         struct io_uring_params p;
 319         void *ptr;
 320         int fd;
 321
 322         memset(&p, 0, sizeof(p));
 323
 324         if (polled)
 325                 p.flags |= IORING_SETUP_IOPOLL;
 326         if (fixedbufs)
 327                 p.flags |= IORING_SETUP_FIXEDBUFS;
 328         if (buffered)
 329                 p.flags |= IORING_SETUP_SQWQ;
 330         else if (sq_thread) {
 331                 p.flags |= IORING_SETUP_SQTHREAD;
 332                 p.sq_thread_cpu = sq_thread_cpu;
 333         }
 334
 335         if (fixedbufs)
 336                 fd = io_uring_setup(DEPTH, s->iovecs, &p);
 337         else
 338                 fd = io_uring_setup(DEPTH, NULL, &p);
 339         if (fd < 0) {
 340                 perror("io_uring_setup");
 341                 return 1;
 342         }
 343
 344         s->ring_fd = fd;
 345         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 346                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 347                         IORING_OFF_SQ_RING);
 348         printf("sq_ring ptr = 0x%p\n", ptr);
 349         sring->head = ptr + p.sq_off.head;
 350         sring->tail = ptr + p.sq_off.tail;
 351         sring->ring_mask = ptr + p.sq_off.ring_mask;
 352         sring->ring_entries = ptr + p.sq_off.ring_entries;
 353         sring->array = ptr + p.sq_off.array;
 354         sq_ring_mask = *sring->ring_mask;
 355
 356         s->iocbs = mmap(0, p.sq_entries * sizeof(struct io_uring_iocb),
 357                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 358                         IORING_OFF_IOCB);
 359         printf("iocbs ptr   = 0x%p\n", s->iocbs);
 360
 361         ptr = mmap(0, p.cq_off.events + p.cq_entries * sizeof(struct io_uring_event),
 362                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 363                         IORING_OFF_CQ_RING);
 364         printf("cq_ring ptr = 0x%p\n", ptr);
 365         cring->head = ptr + p.cq_off.head;
 366         cring->tail = ptr + p.cq_off.tail;
 367         cring->ring_mask = ptr + p.cq_off.ring_mask;
 368         cring->ring_entries = ptr + p.cq_off.ring_entries;
 369         cring->events = ptr + p.cq_off.events;
 370         cq_ring_mask = *cring->ring_mask;
 371         return 0;
 372 }
 373
 374 int main(int argc, char *argv[])
 375 {
 376         struct submitter *s = &submitters[0];
 377         unsigned long done, calls, reap, cache_hit, cache_miss;
 378         int err, i;
 379         struct rlimit rlim;
 380         void *ret;
 381
 382         if (argc < 2) {
 383                 printf("%s: filename\n", argv[0]);
 384                 return 1;
 385         }
 386
 387         rlim.rlim_cur = RLIM_INFINITY;
 388         rlim.rlim_max = RLIM_INFINITY;
 389         if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 390                 perror("setrlimit");
 391                 return 1;
 392         }
 393
 394         arm_sig_int();
 395
 396         for (i = 0; i < DEPTH; i++) {
 397                 void *buf;
 398
 399                 if (posix_memalign(&buf, BS, BS)) {
 400                         printf("failed alloc\n");
 401                         return 1;
 402                 }
 403                 s->iovecs[i].iov_base = buf;
 404                 s->iovecs[i].iov_len = BS;
 405         }
 406
 407         err = setup_ring(s);
 408         if (err) {
 409                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
 410                 return 1;
 411         }
 412         printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
 413         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 414         strcpy(s->filename, argv[1]);
 415
 416         pthread_create(&s->thread, NULL, submitter_fn, s);
 417
 418         cache_hit = cache_miss = reap = calls = done = 0;
 419         do {
 420                 unsigned long this_done = 0;
 421                 unsigned long this_reap = 0;
 422                 unsigned long this_call = 0;
 423                 unsigned long this_cache_hit = 0;
 424                 unsigned long this_cache_miss = 0;
 425                 unsigned long rpc = 0, ipc = 0;
 426                 double hit = 0.0;
 427
 428                 sleep(1);
 429                 this_done += s->done;
 430                 this_call += s->calls;
 431                 this_reap += s->reaps;
 432                 this_cache_hit += s->cachehit;
 433                 this_cache_miss += s->cachemiss;
 434                 if (this_cache_hit && this_cache_miss) {
 435                         unsigned long hits, total;
 436
 437                         hits = this_cache_hit - cache_hit;
 438                         total = hits + this_cache_miss - cache_miss;
 439                         hit = (double) hits / (double) total;
 440                         hit *= 100.0;
 441                 }
 442                 if (this_call - calls) {
 443                         rpc = (this_done - done) / (this_call - calls);
 444                         ipc = (this_reap - reap) / (this_call - calls);
 445                 }
 446                 printf("IOPS=%lu, IOS/call=%lu/%lu, inflight=%u (head=%u tail=%u), Cachehit=%0.2f%%\n",
 447                                 this_done - done, rpc, ipc, s->inflight,
 448                                 *s->cq_ring.head, *s->cq_ring.tail, hit);
 449                 done = this_done;
 450                 calls = this_call;
 451                 reap = this_reap;
 452                 cache_hit = s->cachehit;
 453                 cache_miss = s->cachemiss;
 454         } while (!finish);
 455
 456         pthread_join(s->thread, &ret);
 457         close(s->ring_fd);
 458         return 0;
 459 }