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_cqe *cqes;
  45 };
  46
  47 #define DEPTH                   128
  48
  49 #define BATCH_SUBMIT            64
  50 #define BATCH_COMPLETE          64
  51
  52 #define BS                      4096
  53
  54 #define MAX_FDS                 16
  55
  56 static unsigned sq_ring_mask, cq_ring_mask;
  57
  58 struct file {
  59         unsigned long max_blocks;
  60         unsigned pending_ios;
  61         int fd;
  62 };
  63
  64 struct submitter {
  65         pthread_t thread;
  66         int ring_fd;
  67         struct drand48_data rand;
  68         struct io_sq_ring sq_ring;
  69         struct io_uring_sqe *sqes;
  70         struct iovec iovecs[DEPTH];
  71         struct io_cq_ring cq_ring;
  72         int inflight;
  73         unsigned long reaps;
  74         unsigned long done;
  75         unsigned long calls;
  76         unsigned long cachehit, cachemiss;
  77         volatile int finish;
  78
  79         struct file files[MAX_FDS];
  80         unsigned nr_files;
  81         unsigned cur_file;
  82 };
  83
  84 static struct submitter submitters[1];
  85 static volatile int finish;
  86
  87 static int polled = 1;          /* use IO polling */
  88 static int fixedbufs = 1;       /* use fixed user buffers */
  89 static int buffered = 0;        /* use buffered IO, not O_DIRECT */
  90 static int sq_thread_poll = 0;  /* use kernel submission/poller thread */
  91 static int sq_thread_cpu = -1;  /* pin above thread to this CPU */
  92
  93 static int io_uring_register_buffers(struct submitter *s)
  94 {
  95         struct io_uring_register_buffers reg = {
  96                 .iovecs = s->iovecs,
  97                 .nr_iovecs = DEPTH
  98         };
  99
 100         return syscall(__NR_sys_io_uring_register, s->ring_fd,
 101                         IORING_REGISTER_BUFFERS, &reg);
 102 }
 103
 104 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 105 {
 106         return syscall(__NR_sys_io_uring_setup, entries, p);
 107 }
 108
 109 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
 110                           unsigned int min_complete, unsigned int flags)
 111 {
 112         return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit,
 113                         min_complete, flags);
 114 }
 115
 116 static int gettid(void)
 117 {
 118         return syscall(__NR_gettid);
 119 }
 120
 121 static unsigned file_depth(struct submitter *s)
 122 {
 123         return (DEPTH + s->nr_files - 1) / s->nr_files;
 124 }
 125
 126 static void init_io(struct submitter *s, unsigned index)
 127 {
 128         struct io_uring_sqe *sqe = &s->sqes[index];
 129         unsigned long offset;
 130         struct file *f;
 131         long r;
 132
 133         if (s->nr_files == 1) {
 134                 f = &s->files[0];
 135         } else {
 136                 f = &s->files[s->cur_file];
 137                 if (f->pending_ios >= file_depth(s)) {
 138                         s->cur_file++;
 139                         if (s->cur_file == s->nr_files)
 140                                 s->cur_file = 0;
 141                 }
 142         }
 143         f->pending_ios++;
 144
 145         lrand48_r(&s->rand, &r);
 146         offset = (r % (f->max_blocks - 1)) * BS;
 147
 148         sqe->flags = 0;
 149         sqe->opcode = IORING_OP_READV;
 150         if (fixedbufs) {
 151                 sqe->addr = s->iovecs[index].iov_base;
 152                 sqe->len = BS;
 153                 sqe->buf_index = index;
 154                 sqe->flags |= IOSQE_FIXED_BUFFER;
 155         } else {
 156                 sqe->addr = &s->iovecs[index];
 157                 sqe->len = 1;
 158                 sqe->buf_index = 0;
 159         }
 160         sqe->ioprio = 0;
 161         sqe->fd = f->fd;
 162         sqe->off = offset;
 163         sqe->data = (unsigned long) f;
 164 }
 165
 166 static int prep_more_ios(struct submitter *s, int max_ios)
 167 {
 168         struct io_sq_ring *ring = &s->sq_ring;
 169         unsigned index, tail, next_tail, prepped = 0;
 170
 171         next_tail = tail = *ring->tail;
 172         do {
 173                 next_tail++;
 174                 barrier();
 175                 if (next_tail == *ring->head)
 176                         break;
 177
 178                 index = tail & sq_ring_mask;
 179                 init_io(s, index);
 180                 ring->array[index] = index;
 181                 prepped++;
 182                 tail = next_tail;
 183         } while (prepped < max_ios);
 184
 185         if (*ring->tail != tail) {
 186                 /* order tail store with writes to sqes above */
 187                 barrier();
 188                 *ring->tail = tail;
 189                 barrier();
 190         }
 191         return prepped;
 192 }
 193
 194 static int get_file_size(struct file *f)
 195 {
 196         struct stat st;
 197
 198         if (fstat(f->fd, &st) < 0)
 199                 return -1;
 200         if (S_ISBLK(st.st_mode)) {
 201                 unsigned long long bytes;
 202
 203                 if (ioctl(f->fd, BLKGETSIZE64, &bytes) != 0)
 204                         return -1;
 205
 206                 f->max_blocks = bytes / BS;
 207                 return 0;
 208         } else if (S_ISREG(st.st_mode)) {
 209                 f->max_blocks = st.st_size / BS;
 210                 return 0;
 211         }
 212
 213         return -1;
 214 }
 215
 216 static int reap_events(struct submitter *s)
 217 {
 218         struct io_cq_ring *ring = &s->cq_ring;
 219         struct io_uring_cqe *cqe;
 220         unsigned head, reaped = 0;
 221
 222         head = *ring->head;
 223         do {
 224                 struct file *f;
 225
 226                 barrier();
 227                 if (head == *ring->tail)
 228                         break;
 229                 cqe = &ring->cqes[head & cq_ring_mask];
 230                 f = (struct file *) cqe->data;
 231                 f->pending_ios--;
 232                 if (cqe->res != BS) {
 233                         printf("io: unexpected ret=%d\n", cqe->res);
 234                         return -1;
 235                 }
 236                 if (cqe->flags & IOCQE_FLAG_CACHEHIT)
 237                         s->cachehit++;
 238                 else
 239                         s->cachemiss++;
 240                 reaped++;
 241                 head++;
 242         } while (1);
 243
 244         s->inflight -= reaped;
 245         *ring->head = head;
 246         barrier();
 247         return reaped;
 248 }
 249
 250 static void *submitter_fn(void *data)
 251 {
 252         struct submitter *s = data;
 253         int ret, prepped;
 254
 255         printf("submitter=%d\n", gettid());
 256
 257         srand48_r(pthread_self(), &s->rand);
 258
 259         prepped = 0;
 260         do {
 261                 int to_wait, to_submit, this_reap, to_prep;
 262
 263                 if (!prepped && s->inflight < DEPTH) {
 264                         to_prep = min(DEPTH - s->inflight, BATCH_SUBMIT);
 265                         prepped = prep_more_ios(s, to_prep);
 266                 }
 267                 s->inflight += prepped;
 268 submit_more:
 269                 to_submit = prepped;
 270 submit:
 271                 if (s->inflight + BATCH_SUBMIT < DEPTH)
 272                         to_wait = 0;
 273                 else
 274                         to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
 275
 276                 ret = io_uring_enter(s, to_submit, to_wait,
 277                                         IORING_ENTER_GETEVENTS);
 278                 s->calls++;
 279
 280                 this_reap = reap_events(s);
 281                 if (this_reap == -1)
 282                         break;
 283                 s->reaps += this_reap;
 284
 285                 if (ret >= 0) {
 286                         if (!ret) {
 287                                 to_submit = 0;
 288                                 if (s->inflight)
 289                                         goto submit;
 290                                 continue;
 291                         } else if (ret < to_submit) {
 292                                 int diff = to_submit - ret;
 293
 294                                 s->done += ret;
 295                                 prepped -= diff;
 296                                 goto submit_more;
 297                         }
 298                         s->done += ret;
 299                         prepped = 0;
 300                         continue;
 301                 } else if (ret < 0) {
 302                         if (errno == EAGAIN) {
 303                                 if (s->finish)
 304                                         break;
 305                                 if (this_reap)
 306                                         goto submit;
 307                                 to_submit = 0;
 308                                 goto submit;
 309                         }
 310                         printf("io_submit: %s\n", strerror(errno));
 311                         break;
 312                 }
 313         } while (!s->finish);
 314
 315         finish = 1;
 316         return NULL;
 317 }
 318
 319 static void sig_int(int sig)
 320 {
 321         printf("Exiting on signal %d\n", sig);
 322         submitters[0].finish = 1;
 323         finish = 1;
 324 }
 325
 326 static void arm_sig_int(void)
 327 {
 328         struct sigaction act;
 329
 330         memset(&act, 0, sizeof(act));
 331         act.sa_handler = sig_int;
 332         act.sa_flags = SA_RESTART;
 333         sigaction(SIGINT, &act, NULL);
 334 }
 335
 336 static int setup_ring(struct submitter *s)
 337 {
 338         struct io_sq_ring *sring = &s->sq_ring;
 339         struct io_cq_ring *cring = &s->cq_ring;
 340         struct io_uring_params p;
 341         int ret, fd;
 342         void *ptr;
 343
 344         memset(&p, 0, sizeof(p));
 345
 346         if (polled)
 347                 p.flags |= IORING_SETUP_IOPOLL;
 348         if (sq_thread_poll) {
 349                 p.flags |= IORING_SETUP_SQPOLL;
 350                 if (sq_thread_cpu != -1)
 351                         p.flags |= IORING_SETUP_SQ_AFF;
 352         }
 353
 354         fd = io_uring_setup(DEPTH, &p);
 355         if (fd < 0) {
 356                 perror("io_uring_setup");
 357                 return 1;
 358         }
 359         s->ring_fd = fd;
 360
 361         if (fixedbufs) {
 362                 ret = io_uring_register_buffers(s);
 363                 if (ret < 0) {
 364                         perror("io_uring_register");
 365                         return 1;
 366                 }
 367         }
 368
 369         ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
 370                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 371                         IORING_OFF_SQ_RING);
 372         printf("sq_ring ptr = 0x%p\n", ptr);
 373         sring->head = ptr + p.sq_off.head;
 374         sring->tail = ptr + p.sq_off.tail;
 375         sring->ring_mask = ptr + p.sq_off.ring_mask;
 376         sring->ring_entries = ptr + p.sq_off.ring_entries;
 377         sring->array = ptr + p.sq_off.array;
 378         sq_ring_mask = *sring->ring_mask;
 379
 380         s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
 381                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 382                         IORING_OFF_SQES);
 383         printf("sqes ptr    = 0x%p\n", s->sqes);
 384
 385         ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
 386                         PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
 387                         IORING_OFF_CQ_RING);
 388         printf("cq_ring ptr = 0x%p\n", ptr);
 389         cring->head = ptr + p.cq_off.head;
 390         cring->tail = ptr + p.cq_off.tail;
 391         cring->ring_mask = ptr + p.cq_off.ring_mask;
 392         cring->ring_entries = ptr + p.cq_off.ring_entries;
 393         cring->cqes = ptr + p.cq_off.cqes;
 394         cq_ring_mask = *cring->ring_mask;
 395         return 0;
 396 }
 397
 398 int main(int argc, char *argv[])
 399 {
 400         struct submitter *s = &submitters[0];
 401         unsigned long done, calls, reap, cache_hit, cache_miss;
 402         int err, i, flags, fd;
 403         struct rlimit rlim;
 404         void *ret;
 405
 406         if (argc < 2) {
 407                 printf("%s: filename\n", argv[0]);
 408                 return 1;
 409         }
 410
 411         flags = O_RDONLY | O_NOATIME;
 412         if (!buffered)
 413                 flags |= O_DIRECT;
 414
 415         i = 1;
 416         while (i < argc) {
 417                 struct file *f = &s->files[s->nr_files];
 418
 419                 fd = open(argv[i], flags);
 420                 if (fd < 0) {
 421                         perror("open");
 422                         return 1;
 423                 }
 424                 f->fd = fd;
 425                 if (get_file_size(f)) {
 426                         printf("failed getting size of device/file\n");
 427                         return 1;
 428                 }
 429                 if (f->max_blocks <= 1) {
 430                         printf("Zero file/device size?\n");
 431                         return 1;
 432                 }
 433                 f->max_blocks--;
 434
 435                 printf("Added file %s\n", argv[i]);
 436                 s->nr_files++;
 437                 i++;
 438         }
 439
 440         rlim.rlim_cur = RLIM_INFINITY;
 441         rlim.rlim_max = RLIM_INFINITY;
 442         if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
 443                 perror("setrlimit");
 444                 return 1;
 445         }
 446
 447         arm_sig_int();
 448
 449         for (i = 0; i < DEPTH; i++) {
 450                 void *buf;
 451
 452                 if (posix_memalign(&buf, BS, BS)) {
 453                         printf("failed alloc\n");
 454                         return 1;
 455                 }
 456                 s->iovecs[i].iov_base = buf;
 457                 s->iovecs[i].iov_len = BS;
 458         }
 459
 460         err = setup_ring(s);
 461         if (err) {
 462                 printf("ring setup failed: %s, %d\n", strerror(errno), err);
 463                 return 1;
 464         }
 465         printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
 466         printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 467
 468         pthread_create(&s->thread, NULL, submitter_fn, s);
 469
 470         cache_hit = cache_miss = reap = calls = done = 0;
 471         do {
 472                 unsigned long this_done = 0;
 473                 unsigned long this_reap = 0;
 474                 unsigned long this_call = 0;
 475                 unsigned long this_cache_hit = 0;
 476                 unsigned long this_cache_miss = 0;
 477                 unsigned long rpc = 0, ipc = 0;
 478                 double hit = 0.0;
 479
 480                 sleep(1);
 481                 this_done += s->done;
 482                 this_call += s->calls;
 483                 this_reap += s->reaps;
 484                 this_cache_hit += s->cachehit;
 485                 this_cache_miss += s->cachemiss;
 486                 if (this_cache_hit && this_cache_miss) {
 487                         unsigned long hits, total;
 488
 489                         hits = this_cache_hit - cache_hit;
 490                         total = hits + this_cache_miss - cache_miss;
 491                         hit = (double) hits / (double) total;
 492                         hit *= 100.0;
 493                 }
 494                 if (this_call - calls) {
 495                         rpc = (this_done - done) / (this_call - calls);
 496                         ipc = (this_reap - reap) / (this_call - calls);
 497                 }
 498                 printf("IOPS=%lu, IOS/call=%lu/%lu, inflight=%u (head=%u tail=%u), Cachehit=%0.2f%%\n",
 499                                 this_done - done, rpc, ipc, s->inflight,
 500                                 *s->cq_ring.head, *s->cq_ring.tail, hit);
 501                 done = this_done;
 502                 calls = this_call;
 503                 reap = this_reap;
 504                 cache_hit = s->cachehit;
 505                 cache_miss = s->cachemiss;
 506         } while (!finish);
 507
 508         pthread_join(s->thread, &ret);
 509         close(s->ring_fd);
 510         return 0;
 511 }