t/io_uring: stop when max number of files is reached
[fio.git] / t / io_uring.c
index 595e6f2059d71721bf1b0ed83c18b9560e52a264..c7139f8728249c4dae24831da1890299d439750e 100644 (file)
@@ -22,9 +22,7 @@
 
 #include "../arch/arch.h"
 #include "../lib/types.h"
-#include "../os/io_uring.h"
-
-#define barrier()      __asm__ __volatile__("": : :"memory")
+#include "../os/linux/io_uring.h"
 
 #define min(a, b)              ((a < b) ? (a) : (b))
 
@@ -47,8 +45,8 @@ struct io_cq_ring {
 
 #define DEPTH                  128
 
-#define BATCH_SUBMIT           64
-#define BATCH_COMPLETE         64
+#define BATCH_SUBMIT           32
+#define BATCH_COMPLETE         32
 
 #define BS                     4096
 
@@ -90,36 +88,36 @@ static volatile int finish;
 
 static int polled = 1;         /* use IO polling */
 static int fixedbufs = 1;      /* use fixed user buffers */
+static int register_files = 1; /* use fixed files */
 static int buffered = 0;       /* use buffered IO, not O_DIRECT */
 static int sq_thread_poll = 0; /* use kernel submission/poller thread */
 static int sq_thread_cpu = -1; /* pin above thread to this CPU */
+static int do_nop = 0;         /* no-op SQ ring commands */
 
 static int io_uring_register_buffers(struct submitter *s)
 {
-       struct io_uring_register_buffers reg = {
-               .iovecs = s->iovecs,
-               .nr_iovecs = DEPTH
-       };
+       if (do_nop)
+               return 0;
 
        return syscall(__NR_sys_io_uring_register, s->ring_fd,
-                       IORING_REGISTER_BUFFERS, &reg);
+                       IORING_REGISTER_BUFFERS, s->iovecs, DEPTH);
 }
 
 static int io_uring_register_files(struct submitter *s)
 {
-       struct io_uring_register_files reg;
        int i;
 
+       if (do_nop)
+               return 0;
+
        s->fds = calloc(s->nr_files, sizeof(__s32));
        for (i = 0; i < s->nr_files; i++) {
                s->fds[i] = s->files[i].real_fd;
                s->files[i].fixed_fd = i;
        }
-       reg.fds = s->fds;
-       reg.nr_fds = s->nr_files;
 
        return syscall(__NR_sys_io_uring_register, s->ring_fd,
-                       IORING_REGISTER_FILES, &reg);
+                       IORING_REGISTER_FILES, s->fds, s->nr_files);
 }
 
 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
@@ -131,7 +129,7 @@ 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);
+                       min_complete, flags, NULL, 0);
 }
 
 static int gettid(void)
@@ -151,6 +149,11 @@ static void init_io(struct submitter *s, unsigned index)
        struct file *f;
        long r;
 
+       if (do_nop) {
+               sqe->opcode = IORING_OP_NOP;
+               return;
+       }
+
        if (s->nr_files == 1) {
                f = &s->files[0];
        } else {
@@ -159,6 +162,7 @@ static void init_io(struct submitter *s, unsigned index)
                        s->cur_file++;
                        if (s->cur_file == s->nr_files)
                                s->cur_file = 0;
+                       f = &s->files[s->cur_file];
                }
        }
        f->pending_ios++;
@@ -166,20 +170,25 @@ static void init_io(struct submitter *s, unsigned index)
        lrand48_r(&s->rand, &r);
        offset = (r % (f->max_blocks - 1)) * BS;
 
-       sqe->flags = IOSQE_FIXED_FILE;
+       if (register_files) {
+               sqe->flags = IOSQE_FIXED_FILE;
+               sqe->fd = f->fixed_fd;
+       } else {
+               sqe->flags = 0;
+               sqe->fd = f->real_fd;
+       }
        if (fixedbufs) {
                sqe->opcode = IORING_OP_READ_FIXED;
-               sqe->addr = s->iovecs[index].iov_base;
+               sqe->addr = (unsigned long) s->iovecs[index].iov_base;
                sqe->len = BS;
                sqe->buf_index = index;
        } else {
                sqe->opcode = IORING_OP_READV;
-               sqe->addr = &s->iovecs[index];
+               sqe->addr = (unsigned long) &s->iovecs[index];
                sqe->len = 1;
                sqe->buf_index = 0;
        }
        sqe->ioprio = 0;
-       sqe->fd = f->fixed_fd;
        sqe->off = offset;
        sqe->user_data = (unsigned long) f;
 }
@@ -192,7 +201,7 @@ static int prep_more_ios(struct submitter *s, int max_ios)
        next_tail = tail = *ring->tail;
        do {
                next_tail++;
-               barrier();
+               read_barrier();
                if (next_tail == *ring->head)
                        break;
 
@@ -205,9 +214,9 @@ static int prep_more_ios(struct submitter *s, int max_ios)
 
        if (*ring->tail != tail) {
                /* order tail store with writes to sqes above */
-               barrier();
+               write_barrier();
                *ring->tail = tail;
-               barrier();
+               write_barrier();
        }
        return prepped;
 }
@@ -244,15 +253,19 @@ static int reap_events(struct submitter *s)
        do {
                struct file *f;
 
-               barrier();
+               read_barrier();
                if (head == *ring->tail)
                        break;
                cqe = &ring->cqes[head & cq_ring_mask];
-               f = (struct file *) cqe->user_data;
-               f->pending_ios--;
-               if (cqe->res != BS) {
-                       printf("io: unexpected ret=%d\n", cqe->res);
-                       return -1;
+               if (!do_nop) {
+                       f = (struct file *) (uintptr_t) cqe->user_data;
+                       f->pending_ios--;
+                       if (cqe->res != BS) {
+                               printf("io: unexpected ret=%d\n", cqe->res);
+                               if (polled && cqe->res == -EOPNOTSUPP)
+                                       printf("Your filesystem doesn't support poll\n");
+                               return -1;
+                       }
                }
                if (cqe->flags & IOCQE_FLAG_CACHEHIT)
                        s->cachehit++;
@@ -264,7 +277,7 @@ static int reap_events(struct submitter *s)
 
        s->inflight -= reaped;
        *ring->head = head;
-       barrier();
+       write_barrier();
        return reaped;
 }
 
@@ -290,7 +303,7 @@ static void *submitter_fn(void *data)
 submit_more:
                to_submit = prepped;
 submit:
-               if (s->inflight + BATCH_SUBMIT < DEPTH)
+               if (to_submit && (s->inflight + to_submit <= DEPTH))
                        to_wait = 0;
                else
                        to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
@@ -300,8 +313,13 @@ submit:
                 * poll, or if IORING_SQ_NEED_WAKEUP is set.
                 */
                if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
-                       ret = io_uring_enter(s, to_submit, to_wait,
-                                               IORING_ENTER_GETEVENTS);
+                       unsigned flags = 0;
+
+                       if (to_wait)
+                               flags = IORING_ENTER_GETEVENTS;
+                       if ((*ring->flags & IORING_SQ_NEED_WAKEUP))
+                               flags |= IORING_ENTER_SQ_WAKEUP;
+                       ret = io_uring_enter(s, to_submit, to_wait, flags);
                        s->calls++;
                }
 
@@ -314,9 +332,10 @@ submit:
                do {
                        int r;
                        r = reap_events(s);
-                       if (r == -1)
+                       if (r == -1) {
+                               s->finish = 1;
                                break;
-                       else if (r > 0)
+                       else if (r > 0)
                                this_reap += r;
                } while (sq_thread_poll && this_reap < to_wait);
                s->reaps += this_reap;
@@ -382,7 +401,7 @@ static int setup_ring(struct submitter *s)
 
        memset(&p, 0, sizeof(p));
 
-       if (polled)
+       if (polled && !do_nop)
                p.flags |= IORING_SETUP_IOPOLL;
        if (sq_thread_poll) {
                p.flags |= IORING_SETUP_SQPOLL;
@@ -407,10 +426,12 @@ static int setup_ring(struct submitter *s)
                }
        }
 
-       ret = io_uring_register_files(s);
-       if (ret < 0) {
-               perror("io_uring_register_files");
-               return 1;
+       if (register_files) {
+               ret = io_uring_register_files(s);
+               if (ret < 0) {
+                       perror("io_uring_register_files");
+                       return 1;
+               }
        }
 
        ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
@@ -443,15 +464,33 @@ static int setup_ring(struct submitter *s)
        return 0;
 }
 
+static void file_depths(char *buf)
+{
+       struct submitter *s = &submitters[0];
+       char *p;
+       int i;
+
+       buf[0] = '\0';
+       p = buf;
+       for (i = 0; i < s->nr_files; i++) {
+               struct file *f = &s->files[i];
+
+               if (i + 1 == s->nr_files)
+                       p += sprintf(p, "%d", f->pending_ios);
+               else
+                       p += sprintf(p, "%d, ", f->pending_ios);
+       }
+}
+
 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 rlimit rlim;
+       char *fdepths;
        void *ret;
 
-       if (argc < 2) {
+       if (!do_nop && argc < 2) {
                printf("%s: filename\n", argv[0]);
                return 1;
        }
@@ -461,14 +500,20 @@ int main(int argc, char *argv[])
                flags |= O_DIRECT;
 
        i = 1;
-       while (i < argc) {
-               struct file *f = &s->files[s->nr_files];
+       while (!do_nop && i < argc) {
+               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");
@@ -485,11 +530,15 @@ int main(int argc, char *argv[])
                i++;
        }
 
-       rlim.rlim_cur = RLIM_INFINITY;
-       rlim.rlim_max = RLIM_INFINITY;
-       if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
-               perror("setrlimit");
-               return 1;
+       if (fixedbufs) {
+               struct rlimit rlim;
+
+               rlim.rlim_cur = RLIM_INFINITY;
+               rlim.rlim_max = RLIM_INFINITY;
+               if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
+                       perror("setrlimit");
+                       return 1;
+               }
        }
 
        arm_sig_int();
@@ -515,6 +564,7 @@ int main(int argc, char *argv[])
 
        pthread_create(&s->thread, NULL, submitter_fn, s);
 
+       fdepths = malloc(8 * s->nr_files);
        cache_hit = cache_miss = reap = calls = done = 0;
        do {
                unsigned long this_done = 0;
@@ -544,9 +594,10 @@ int main(int argc, char *argv[])
                        ipc = (this_reap - reap) / (this_call - calls);
                } else
                        rpc = ipc = -1;
-               printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (head=%u tail=%u), Cachehit=%0.2f%%\n",
+               file_depths(fdepths);
+               printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s), Cachehit=%0.2f%%\n",
                                this_done - done, rpc, ipc, s->inflight,
-                               *s->cq_ring.head, *s->cq_ring.tail, hit);
+                               fdepths, hit);
                done = this_done;
                calls = this_call;
                reap = this_reap;
@@ -556,5 +607,6 @@ int main(int argc, char *argv[])
 
        pthread_join(s->thread, &ret);
        close(s->ring_fd);
+       free(fdepths);
        return 0;
 }