t/io_uring: add support for registered files
[fio.git] / t / io_uring.c
index 76da6b292f03e2d21623b1d2f48f4bb88dfdb149..0461329b64d2267ae240fca1545d605338b73a0e 100644 (file)
@@ -33,6 +33,7 @@ struct io_sq_ring {
        unsigned *tail;
        unsigned *ring_mask;
        unsigned *ring_entries;
+       unsigned *flags;
        unsigned *array;
 };
 
@@ -44,10 +45,10 @@ struct io_cq_ring {
        struct io_uring_cqe *cqes;
 };
 
-#define DEPTH                  32
+#define DEPTH                  128
 
-#define BATCH_SUBMIT           8
-#define BATCH_COMPLETE         8
+#define BATCH_SUBMIT           64
+#define BATCH_COMPLETE         64
 
 #define BS                     4096
 
@@ -57,6 +58,7 @@ static unsigned sq_ring_mask, cq_ring_mask;
 
 struct file {
        unsigned long max_blocks;
+       unsigned pending_ios;
        int fd;
 };
 
@@ -74,6 +76,7 @@ struct submitter {
        unsigned long calls;
        unsigned long cachehit, cachemiss;
        volatile int finish;
+
        struct file files[MAX_FDS];
        unsigned nr_files;
        unsigned cur_file;
@@ -83,15 +86,41 @@ static struct submitter submitters[1];
 static volatile int finish;
 
 static int polled = 1;         /* use IO polling */
-static int fixedbufs = 0;      /* use fixed user buffers */
+static int fixedbufs = 1;      /* use fixed user buffers */
 static int buffered = 0;       /* use buffered IO, not O_DIRECT */
-static int sq_thread = 0;      /* use kernel submission thread */
-static int sq_thread_cpu = 0;  /* pin above thread to this CPU */
+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 io_uring_register_buffers(struct submitter *s)
+{
+       struct io_uring_register_buffers reg = {
+               .iovecs = s->iovecs,
+               .nr_iovecs = DEPTH
+       };
+
+       return syscall(__NR_sys_io_uring_register, s->ring_fd,
+                       IORING_REGISTER_BUFFERS, &reg);
+}
+
+static int io_uring_register_files(struct submitter *s)
+{
+       struct io_uring_register_files reg;
+       int i, ret;
+
+       reg.fds = calloc(s->nr_files, sizeof(int));
+       for (i = 0; i < s->nr_files; i++)
+               reg.fds[i] = s->files[i].fd;
+       reg.nr_fds = s->nr_files;
+
+       ret = syscall(__NR_sys_io_uring_register, s->ring_fd,
+                       IORING_REGISTER_FILES, &reg);
+       free(reg.fds);
+       return ret;
+}
 
-static int io_uring_setup(unsigned entries, struct iovec *iovecs,
-                         unsigned nr_iovecs, struct io_uring_params *p)
+static int io_uring_setup(unsigned entries, struct io_uring_params *p)
 {
-       return syscall(__NR_sys_io_uring_setup, entries, iovecs, nr_iovecs, p);
+       return syscall(__NR_sys_io_uring_setup, entries, p);
 }
 
 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
@@ -106,6 +135,11 @@ static int gettid(void)
        return syscall(__NR_gettid);
 }
 
+static unsigned file_depth(struct submitter *s)
+{
+       return (DEPTH + s->nr_files - 1) / s->nr_files;
+}
+
 static void init_io(struct submitter *s, unsigned index)
 {
        struct io_uring_sqe *sqe = &s->sqes[index];
@@ -113,28 +147,37 @@ static void init_io(struct submitter *s, unsigned index)
        struct file *f;
        long r;
 
-       f = &s->files[s->cur_file];
-       s->cur_file++;
-       if (s->cur_file == s->nr_files)
-               s->cur_file = 0;
+       if (s->nr_files == 1) {
+               f = &s->files[0];
+       } else {
+               f = &s->files[s->cur_file];
+               if (f->pending_ios >= file_depth(s)) {
+                       s->cur_file++;
+                       if (s->cur_file == s->nr_files)
+                               s->cur_file = 0;
+               }
+       }
+       f->pending_ios++;
 
        lrand48_r(&s->rand, &r);
        offset = (r % (f->max_blocks - 1)) * BS;
 
+       sqe->flags = IOSQE_FIXED_FILE;
+       sqe->opcode = IORING_OP_READV;
        if (fixedbufs) {
-               sqe->opcode = IORING_OP_READ_FIXED;
                sqe->addr = s->iovecs[index].iov_base;
                sqe->len = BS;
-               sqe->index = index;
+               sqe->buf_index = index;
+               sqe->flags |= IOSQE_FIXED_BUFFER;
        } else {
-               sqe->opcode = IORING_OP_READV;
                sqe->addr = &s->iovecs[index];
                sqe->len = 1;
+               sqe->buf_index = 0;
        }
-       sqe->flags = 0;
        sqe->ioprio = 0;
        sqe->fd = f->fd;
        sqe->off = offset;
+       sqe->data = (unsigned long) f;
 }
 
 static int prep_more_ios(struct submitter *s, int max_ios)
@@ -195,10 +238,14 @@ static int reap_events(struct submitter *s)
 
        head = *ring->head;
        do {
+               struct file *f;
+
                barrier();
                if (head == *ring->tail)
                        break;
                cqe = &ring->cqes[head & cq_ring_mask];
+               f = (struct file *) cqe->data;
+               f->pending_ios--;
                if (cqe->res != BS) {
                        printf("io: unexpected ret=%d\n", cqe->res);
                        return -1;
@@ -220,6 +267,7 @@ static int reap_events(struct submitter *s)
 static void *submitter_fn(void *data)
 {
        struct submitter *s = data;
+       struct io_sq_ring *ring = &s->sq_ring;
        int ret, prepped;
 
        printf("submitter=%d\n", gettid());
@@ -243,13 +291,30 @@ submit:
                else
                        to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
 
-               ret = io_uring_enter(s, to_submit, to_wait,
-                                       IORING_ENTER_GETEVENTS);
-               s->calls++;
+               /*
+                * Only need to call io_uring_enter if we're not using SQ thread
+                * 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);
+                       s->calls++;
+               }
 
-               this_reap = reap_events(s);
-               if (this_reap == -1)
-                       break;
+               /*
+                * For non SQ thread poll, we already got the events we needed
+                * through the io_uring_enter() above. For SQ thread poll, we
+                * need to loop here until we find enough events.
+                */
+               this_reap = 0;
+               do {
+                       int r;
+                       r = reap_events(s);
+                       if (r == -1)
+                               break;
+                       else if (r > 0)
+                               this_reap += r;
+               } while (sq_thread_poll && this_reap < to_wait);
                s->reaps += this_reap;
 
                if (ret >= 0) {
@@ -308,30 +373,40 @@ static int setup_ring(struct submitter *s)
        struct io_sq_ring *sring = &s->sq_ring;
        struct io_cq_ring *cring = &s->cq_ring;
        struct io_uring_params p;
+       int ret, fd;
        void *ptr;
-       int fd;
 
        memset(&p, 0, sizeof(p));
 
        if (polled)
                p.flags |= IORING_SETUP_IOPOLL;
-       if (buffered)
-               p.flags |= IORING_SETUP_SQWQ;
-       else if (sq_thread) {
-               p.flags |= IORING_SETUP_SQTHREAD;
-               p.sq_thread_cpu = sq_thread_cpu;
+       if (sq_thread_poll) {
+               p.flags |= IORING_SETUP_SQPOLL;
+               if (sq_thread_cpu != -1)
+                       p.flags |= IORING_SETUP_SQ_AFF;
        }
 
-       if (fixedbufs)
-               fd = io_uring_setup(DEPTH, s->iovecs, DEPTH, &p);
-       else
-               fd = io_uring_setup(DEPTH, NULL, 0, &p);
+       fd = io_uring_setup(DEPTH, &p);
        if (fd < 0) {
                perror("io_uring_setup");
                return 1;
        }
-
        s->ring_fd = fd;
+
+       if (fixedbufs) {
+               ret = io_uring_register_buffers(s);
+               if (ret < 0) {
+                       perror("io_uring_register_buffers");
+                       return 1;
+               }
+       }
+
+       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),
                        PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
                        IORING_OFF_SQ_RING);
@@ -340,6 +415,7 @@ static int setup_ring(struct submitter *s)
        sring->tail = ptr + p.sq_off.tail;
        sring->ring_mask = ptr + p.sq_off.ring_mask;
        sring->ring_entries = ptr + p.sq_off.ring_entries;
+       sring->flags = ptr + p.sq_off.flags;
        sring->array = ptr + p.sq_off.array;
        sq_ring_mask = *sring->ring_mask;
 
@@ -374,7 +450,7 @@ int main(int argc, char *argv[])
                return 1;
        }
 
-       flags = O_RDONLY;
+       flags = O_RDONLY | O_NOATIME;
        if (!buffered)
                flags |= O_DIRECT;
 
@@ -460,8 +536,9 @@ int main(int argc, char *argv[])
                if (this_call - calls) {
                        rpc = (this_done - done) / (this_call - calls);
                        ipc = (this_reap - reap) / (this_call - calls);
-               }
-               printf("IOPS=%lu, IOS/call=%lu/%lu, inflight=%u (head=%u tail=%u), Cachehit=%0.2f%%\n",
+               } else
+                       rpc = ipc = -1;
+               printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (head=%u tail=%u), Cachehit=%0.2f%%\n",
                                this_done - done, rpc, ipc, s->inflight,
                                *s->cq_ring.head, *s->cq_ring.tail, hit);
                done = this_done;