t/io_uring: make bs and polled IO configurable at runtime
[fio.git] / t / io_uring.c
index 3bcb19b647ccc2c13c91098bf9a4231a12b05bfb..7fa84f99075bb8fb2f7c9ad0120edba911889a72 100644 (file)
@@ -44,10 +44,8 @@ struct io_cq_ring {
 };
 
 #define DEPTH                  128
-
 #define BATCH_SUBMIT           32
 #define BATCH_COMPLETE         32
-
 #define BS                     4096
 
 #define MAX_FDS                        16
@@ -64,16 +62,13 @@ struct file {
 struct submitter {
        pthread_t thread;
        int ring_fd;
-       struct drand48_data rand;
        struct io_sq_ring sq_ring;
        struct io_uring_sqe *sqes;
-       struct iovec iovecs[DEPTH];
        struct io_cq_ring cq_ring;
        int inflight;
        unsigned long reaps;
        unsigned long done;
        unsigned long calls;
-       unsigned long cachehit, cachemiss;
        volatile int finish;
 
        __s32 *fds;
@@ -81,11 +76,16 @@ struct submitter {
        struct file files[MAX_FDS];
        unsigned nr_files;
        unsigned cur_file;
+       struct iovec iovecs[];
 };
 
-static struct submitter submitters[1];
+static struct submitter *submitter;
 static volatile int finish;
 
+static int depth = DEPTH;
+static int batch_submit = BATCH_SUBMIT;
+static int batch_complete = BATCH_COMPLETE;
+static int bs = BS;
 static int polled = 1;         /* use IO polling */
 static int fixedbufs = 1;      /* use fixed user buffers */
 static int register_files = 1; /* use fixed files */
@@ -99,8 +99,8 @@ static int io_uring_register_buffers(struct submitter *s)
        if (do_nop)
                return 0;
 
-       return syscall(__NR_sys_io_uring_register, s->ring_fd,
-                       IORING_REGISTER_BUFFERS, s->iovecs, DEPTH);
+       return syscall(__NR_io_uring_register, s->ring_fd,
+                       IORING_REGISTER_BUFFERS, s->iovecs, depth);
 }
 
 static int io_uring_register_files(struct submitter *s)
@@ -116,30 +116,32 @@ static int io_uring_register_files(struct submitter *s)
                s->files[i].fixed_fd = i;
        }
 
-       return syscall(__NR_sys_io_uring_register, s->ring_fd,
+       return syscall(__NR_io_uring_register, s->ring_fd,
                        IORING_REGISTER_FILES, s->fds, s->nr_files);
 }
 
 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 {
-       return syscall(__NR_sys_io_uring_setup, entries, p);
+       return syscall(__NR_io_uring_setup, entries, p);
 }
 
 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
                          unsigned int min_complete, unsigned int flags)
 {
-       return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit,
-                       min_complete, flags);
+       return syscall(__NR_io_uring_enter, s->ring_fd, to_submit, min_complete,
+                       flags, NULL, 0);
 }
 
+#ifndef CONFIG_HAVE_GETTID
 static int gettid(void)
 {
        return syscall(__NR_gettid);
 }
+#endif
 
 static unsigned file_depth(struct submitter *s)
 {
-       return (DEPTH + s->nr_files - 1) / s->nr_files;
+       return (depth + s->nr_files - 1) / s->nr_files;
 }
 
 static void init_io(struct submitter *s, unsigned index)
@@ -167,8 +169,8 @@ static void init_io(struct submitter *s, unsigned index)
        }
        f->pending_ios++;
 
-       lrand48_r(&s->rand, &r);
-       offset = (r % (f->max_blocks - 1)) * BS;
+       r = lrand48();
+       offset = (r % (f->max_blocks - 1)) * bs;
 
        if (register_files) {
                sqe->flags = IOSQE_FIXED_FILE;
@@ -180,7 +182,7 @@ static void init_io(struct submitter *s, unsigned index)
        if (fixedbufs) {
                sqe->opcode = IORING_OP_READ_FIXED;
                sqe->addr = (unsigned long) s->iovecs[index].iov_base;
-               sqe->len = BS;
+               sqe->len = bs;
                sqe->buf_index = index;
        } else {
                sqe->opcode = IORING_OP_READV;
@@ -213,8 +215,6 @@ static int prep_more_ios(struct submitter *s, int max_ios)
        } while (prepped < max_ios);
 
        if (*ring->tail != tail) {
-               /* order tail store with writes to sqes above */
-               write_barrier();
                *ring->tail = tail;
                write_barrier();
        }
@@ -233,10 +233,10 @@ static int get_file_size(struct file *f)
                if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
                        return -1;
 
-               f->max_blocks = bytes / BS;
+               f->max_blocks = bytes / bs;
                return 0;
        } else if (S_ISREG(st.st_mode)) {
-               f->max_blocks = st.st_size / BS;
+               f->max_blocks = st.st_size / bs;
                return 0;
        }
 
@@ -260,15 +260,13 @@ static int reap_events(struct submitter *s)
                if (!do_nop) {
                        f = (struct file *) (uintptr_t) cqe->user_data;
                        f->pending_ios--;
-                       if (cqe->res != BS) {
+                       if (cqe->res != bs) {
                                printf("io: unexpected ret=%d\n", cqe->res);
+                               if (polled && cqe->res == -EOPNOTSUPP)
+                                       printf("Your filesystem/driver/kernel doesn't support polled IO\n");
                                return -1;
                        }
                }
-               if (cqe->flags & IOCQE_FLAG_CACHEHIT)
-                       s->cachehit++;
-               else
-                       s->cachemiss++;
                reaped++;
                head++;
        } while (1);
@@ -287,24 +285,24 @@ static void *submitter_fn(void *data)
 
        printf("submitter=%d\n", gettid());
 
-       srand48_r(pthread_self(), &s->rand);
+       srand48(pthread_self());
 
        prepped = 0;
        do {
                int to_wait, to_submit, this_reap, to_prep;
 
-               if (!prepped && s->inflight < DEPTH) {
-                       to_prep = min(DEPTH - s->inflight, BATCH_SUBMIT);
+               if (!prepped && s->inflight < depth) {
+                       to_prep = min(depth - s->inflight, batch_submit);
                        prepped = prep_more_ios(s, to_prep);
                }
                s->inflight += prepped;
 submit_more:
                to_submit = prepped;
 submit:
-               if (to_submit && (s->inflight + to_submit <= DEPTH))
+               if (to_submit && (s->inflight + to_submit <= depth))
                        to_wait = 0;
                else
-                       to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
+                       to_wait = min(s->inflight + to_submit, batch_complete);
 
                /*
                 * Only need to call io_uring_enter if we're not using SQ thread
@@ -375,7 +373,7 @@ submit:
 static void sig_int(int sig)
 {
        printf("Exiting on signal %d\n", sig);
-       submitters[0].finish = 1;
+       submitter->finish = 1;
        finish = 1;
 }
 
@@ -409,7 +407,7 @@ static int setup_ring(struct submitter *s)
                }
        }
 
-       fd = io_uring_setup(DEPTH, &p);
+       fd = io_uring_setup(depth, &p);
        if (fd < 0) {
                perror("io_uring_setup");
                return 1;
@@ -464,7 +462,7 @@ static int setup_ring(struct submitter *s)
 
 static void file_depths(char *buf)
 {
-       struct submitter *s = &submitters[0];
+       struct submitter *s = submitter;
        char *p;
        int i;
 
@@ -480,32 +478,79 @@ static void file_depths(char *buf)
        }
 }
 
+static void usage(char *argv)
+{
+       printf("%s [options] -- [filenames]\n"
+               " -d <int> : IO Depth, default %d\n"
+               " -s <int> : Batch submit, default %d\n"
+               " -c <int> : Batch complete, default %d\n"
+               " -b <int> : Block size, default %d\n"
+               " -p <bool> : Polled IO, default %d\n",
+               argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled);
+       exit(0);
+}
+
 int main(int argc, char *argv[])
 {
-       struct submitter *s = &submitters[0];
-       unsigned long done, calls, reap, cache_hit, cache_miss;
-       int err, i, flags, fd;
+       struct submitter *s;
+       unsigned long done, calls, reap;
+       int err, i, flags, fd, opt;
        char *fdepths;
        void *ret;
 
        if (!do_nop && argc < 2) {
-               printf("%s: filename\n", argv[0]);
+               printf("%s: filename [options]\n", argv[0]);
                return 1;
        }
 
+       while ((opt = getopt(argc, argv, "d:s:c:b:p:h?")) != -1) {
+               switch (opt) {
+               case 'd':
+                       depth = atoi(optarg);
+                       break;
+               case 's':
+                       batch_submit = atoi(optarg);
+                       break;
+               case 'c':
+                       batch_complete = atoi(optarg);
+                       break;
+               case 'b':
+                       bs = atoi(optarg);
+                       break;
+               case 'p':
+                       polled = !!atoi(optarg);
+                       break;
+               case 'h':
+               case '?':
+               default:
+                       usage(argv[0]);
+                       break;
+               }
+       }
+
+       submitter = malloc(sizeof(*submitter) + depth * sizeof(struct iovec));
+       memset(submitter, 0, sizeof(*submitter) + depth * sizeof(struct iovec));
+       s = submitter;
+
        flags = O_RDONLY | O_NOATIME;
        if (!buffered)
                flags |= O_DIRECT;
 
-       i = 1;
+       i = optind;
        while (!do_nop && i < argc) {
-               struct file *f = &s->files[s->nr_files];
+               struct file *f;
 
+               if (s->nr_files == MAX_FDS) {
+                       printf("Max number of files (%d) reached\n", MAX_FDS);
+                       break;
+               }
                fd = open(argv[i], flags);
                if (fd < 0) {
                        perror("open");
                        return 1;
                }
+
+               f = &s->files[s->nr_files];
                f->real_fd = fd;
                if (get_file_size(f)) {
                        printf("failed getting size of device/file\n");
@@ -535,15 +580,15 @@ int main(int argc, char *argv[])
 
        arm_sig_int();
 
-       for (i = 0; i < DEPTH; i++) {
+       for (i = 0; i < depth; i++) {
                void *buf;
 
-               if (posix_memalign(&buf, BS, BS)) {
+               if (posix_memalign(&buf, bs, bs)) {
                        printf("failed alloc\n");
                        return 1;
                }
                s->iovecs[i].iov_base = buf;
-               s->iovecs[i].iov_len = BS;
+               s->iovecs[i].iov_len = bs;
        }
 
        err = setup_ring(s);
@@ -552,49 +597,34 @@ int main(int argc, char *argv[])
                return 1;
        }
        printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
-       printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
+       printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
 
        pthread_create(&s->thread, NULL, submitter_fn, s);
 
        fdepths = malloc(8 * s->nr_files);
-       cache_hit = cache_miss = reap = calls = done = 0;
+       reap = calls = done = 0;
        do {
                unsigned long this_done = 0;
                unsigned long this_reap = 0;
                unsigned long this_call = 0;
-               unsigned long this_cache_hit = 0;
-               unsigned long this_cache_miss = 0;
                unsigned long rpc = 0, ipc = 0;
-               double hit = 0.0;
 
                sleep(1);
                this_done += s->done;
                this_call += s->calls;
                this_reap += s->reaps;
-               this_cache_hit += s->cachehit;
-               this_cache_miss += s->cachemiss;
-               if (this_cache_hit && this_cache_miss) {
-                       unsigned long hits, total;
-
-                       hits = this_cache_hit - cache_hit;
-                       total = hits + this_cache_miss - cache_miss;
-                       hit = (double) hits / (double) total;
-                       hit *= 100.0;
-               }
                if (this_call - calls) {
                        rpc = (this_done - done) / (this_call - calls);
                        ipc = (this_reap - reap) / (this_call - calls);
                } else
                        rpc = ipc = -1;
                file_depths(fdepths);
-               printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s), Cachehit=%0.2f%%\n",
+               printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s)\n",
                                this_done - done, rpc, ipc, s->inflight,
-                               fdepths, hit);
+                               fdepths);
                done = this_done;
                calls = this_call;
                reap = this_reap;
-               cache_hit = s->cachehit;
-               cache_miss = s->cachemiss;
        } while (!finish);
 
        pthread_join(s->thread, &ret);