io_uring: io_uring_setup(2) takes a 'nr_iovecs' field now
[fio.git] / t / io_uring.c
index 83d723f938e88d547bcab93482441fdecd73caa4..76da6b292f03e2d21623b1d2f48f4bb88dfdb149 100644 (file)
 #include <sched.h>
 
 #include "../arch/arch.h"
-
-typedef uint64_t u64;
-typedef uint32_t u32;
-typedef int32_t s32;
-typedef uint16_t u16;
-typedef uint8_t u8;
-
+#include "../lib/types.h"
 #include "../os/io_uring.h"
 
 #define barrier()      __asm__ __volatile__("": : :"memory")
@@ -35,19 +29,19 @@ typedef uint8_t u8;
 #define min(a, b)              ((a < b) ? (a) : (b))
 
 struct io_sq_ring {
-       u32 *head;
-       u32 *tail;
-       u32 *ring_mask;
-       u32 *ring_entries;
-       u32 *array;
+       unsigned *head;
+       unsigned *tail;
+       unsigned *ring_mask;
+       unsigned *ring_entries;
+       unsigned *array;
 };
 
 struct io_cq_ring {
-       u32 *head;
-       u32 *tail;
-       u32 *ring_mask;
-       u32 *ring_entries;
-       struct io_uring_event *events;
+       unsigned *head;
+       unsigned *tail;
+       unsigned *ring_mask;
+       unsigned *ring_entries;
+       struct io_uring_cqe *cqes;
 };
 
 #define DEPTH                  32
@@ -57,15 +51,21 @@ struct io_cq_ring {
 
 #define BS                     4096
 
+#define MAX_FDS                        16
+
 static unsigned sq_ring_mask, cq_ring_mask;
 
+struct file {
+       unsigned long max_blocks;
+       int fd;
+};
+
 struct submitter {
        pthread_t thread;
-       unsigned long max_blocks;
        int ring_fd;
        struct drand48_data rand;
        struct io_sq_ring sq_ring;
-       struct io_uring_iocb *iocbs;
+       struct io_uring_sqe *sqes;
        struct iovec iovecs[DEPTH];
        struct io_cq_ring cq_ring;
        int inflight;
@@ -74,22 +74,24 @@ struct submitter {
        unsigned long calls;
        unsigned long cachehit, cachemiss;
        volatile int finish;
-       char filename[128];
+       struct file files[MAX_FDS];
+       unsigned nr_files;
+       unsigned cur_file;
 };
 
 static struct submitter submitters[1];
 static volatile int finish;
 
-static int polled = 0;         /* use IO polling */
+static int polled = 1;         /* use IO polling */
 static int fixedbufs = 0;      /* use fixed user buffers */
-static int buffered = 1;       /* use buffered IO, not O_DIRECT */
+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 io_uring_setup(unsigned entries, struct iovec *iovecs,
-                         struct io_uring_params *p)
+                         unsigned nr_iovecs, struct io_uring_params *p)
 {
-       return syscall(__NR_sys_io_uring_setup, entries, iovecs, p);
+       return syscall(__NR_sys_io_uring_setup, entries, iovecs, nr_iovecs, p);
 }
 
 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
@@ -104,28 +106,41 @@ static int gettid(void)
        return syscall(__NR_gettid);
 }
 
-static void init_io(struct submitter *s, int fd, unsigned index)
+static void init_io(struct submitter *s, unsigned index)
 {
-       struct io_uring_iocb *iocb = &s->iocbs[index];
+       struct io_uring_sqe *sqe = &s->sqes[index];
        unsigned long offset;
+       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;
+
        lrand48_r(&s->rand, &r);
-       offset = (r % (s->max_blocks - 1)) * BS;
-
-       iocb->opcode = IORING_OP_READ;
-       iocb->flags = 0;
-       iocb->ioprio = 0;
-       iocb->fd = fd;
-       iocb->off = offset;
-       iocb->addr = s->iovecs[index].iov_base;
-       iocb->len = BS;
+       offset = (r % (f->max_blocks - 1)) * BS;
+
+       if (fixedbufs) {
+               sqe->opcode = IORING_OP_READ_FIXED;
+               sqe->addr = s->iovecs[index].iov_base;
+               sqe->len = BS;
+               sqe->index = index;
+       } else {
+               sqe->opcode = IORING_OP_READV;
+               sqe->addr = &s->iovecs[index];
+               sqe->len = 1;
+       }
+       sqe->flags = 0;
+       sqe->ioprio = 0;
+       sqe->fd = f->fd;
+       sqe->off = offset;
 }
 
-static int prep_more_ios(struct submitter *s, int fd, int max_ios)
+static int prep_more_ios(struct submitter *s, int max_ios)
 {
        struct io_sq_ring *ring = &s->sq_ring;
-       u32 index, tail, next_tail, prepped = 0;
+       unsigned index, tail, next_tail, prepped = 0;
 
        next_tail = tail = *ring->tail;
        do {
@@ -135,14 +150,14 @@ static int prep_more_ios(struct submitter *s, int fd, int max_ios)
                        break;
 
                index = tail & sq_ring_mask;
-               init_io(s, fd, index);
+               init_io(s, index);
                ring->array[index] = index;
                prepped++;
                tail = next_tail;
        } while (prepped < max_ios);
 
        if (*ring->tail != tail) {
-               /* order tail store with writes to iocbs above */
+               /* order tail store with writes to sqes above */
                barrier();
                *ring->tail = tail;
                barrier();
@@ -150,22 +165,22 @@ static int prep_more_ios(struct submitter *s, int fd, int max_ios)
        return prepped;
 }
 
-static int get_file_size(int fd, unsigned long *blocks)
+static int get_file_size(struct file *f)
 {
        struct stat st;
 
-       if (fstat(fd, &st) < 0)
+       if (fstat(f->fd, &st) < 0)
                return -1;
        if (S_ISBLK(st.st_mode)) {
                unsigned long long bytes;
 
-               if (ioctl(fd, BLKGETSIZE64, &bytes) != 0)
+               if (ioctl(f->fd, BLKGETSIZE64, &bytes) != 0)
                        return -1;
 
-               *blocks = bytes / BS;
+               f->max_blocks = bytes / BS;
                return 0;
        } else if (S_ISREG(st.st_mode)) {
-               *blocks = st.st_size / BS;
+               f->max_blocks = st.st_size / BS;
                return 0;
        }
 
@@ -175,25 +190,20 @@ static int get_file_size(int fd, unsigned long *blocks)
 static int reap_events(struct submitter *s)
 {
        struct io_cq_ring *ring = &s->cq_ring;
-       struct io_uring_event *ev;
-       u32 head, reaped = 0;
+       struct io_uring_cqe *cqe;
+       unsigned head, reaped = 0;
 
        head = *ring->head;
        do {
                barrier();
                if (head == *ring->tail)
                        break;
-               ev = &ring->events[head & cq_ring_mask];
-               if (ev->res != BS) {
-                       struct io_uring_iocb *iocb = &s->iocbs[ev->index];
-
-                       printf("io: unexpected ret=%d\n", ev->res);
-                       printf("offset=%lu, size=%lu\n",
-                                       (unsigned long) iocb->off,
-                                       (unsigned long) iocb->len);
+               cqe = &ring->cqes[head & cq_ring_mask];
+               if (cqe->res != BS) {
+                       printf("io: unexpected ret=%d\n", cqe->res);
                        return -1;
                }
-               if (ev->flags & IOEV_FLAG_CACHEHIT)
+               if (cqe->flags & IOCQE_FLAG_CACHEHIT)
                        s->cachehit++;
                else
                        s->cachemiss++;
@@ -210,29 +220,10 @@ static int reap_events(struct submitter *s)
 static void *submitter_fn(void *data)
 {
        struct submitter *s = data;
-       int fd, ret, prepped, flags;
+       int ret, prepped;
 
        printf("submitter=%d\n", gettid());
 
-       flags = O_RDONLY;
-       if (!buffered)
-               flags |= O_DIRECT;
-       fd = open(s->filename, flags);
-       if (fd < 0) {
-               perror("open");
-               goto done;
-       }
-
-       if (get_file_size(fd, &s->max_blocks)) {
-               printf("failed getting size of device/file\n");
-               goto err;
-       }
-       if (s->max_blocks <= 1) {
-               printf("Zero file/device size?\n");
-               goto err;
-       }
-       s->max_blocks--;
-
        srand48_r(pthread_self(), &s->rand);
 
        prepped = 0;
@@ -241,7 +232,7 @@ static void *submitter_fn(void *data)
 
                if (!prepped && s->inflight < DEPTH) {
                        to_prep = min(DEPTH - s->inflight, BATCH_SUBMIT);
-                       prepped = prep_more_ios(s, fd, to_prep);
+                       prepped = prep_more_ios(s, to_prep);
                }
                s->inflight += prepped;
 submit_more:
@@ -290,9 +281,7 @@ submit:
                        break;
                }
        } while (!s->finish);
-err:
-       close(fd);
-done:
+
        finish = 1;
        return NULL;
 }
@@ -326,8 +315,6 @@ static int setup_ring(struct submitter *s)
 
        if (polled)
                p.flags |= IORING_SETUP_IOPOLL;
-       if (fixedbufs)
-               p.flags |= IORING_SETUP_FIXEDBUFS;
        if (buffered)
                p.flags |= IORING_SETUP_SQWQ;
        else if (sq_thread) {
@@ -336,16 +323,16 @@ static int setup_ring(struct submitter *s)
        }
 
        if (fixedbufs)
-               fd = io_uring_setup(DEPTH, s->iovecs, &p);
+               fd = io_uring_setup(DEPTH, s->iovecs, DEPTH, &p);
        else
-               fd = io_uring_setup(DEPTH, NULL, &p);
+               fd = io_uring_setup(DEPTH, NULL, 0, &p);
        if (fd < 0) {
                perror("io_uring_setup");
                return 1;
        }
 
        s->ring_fd = fd;
-       ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(u32),
+       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);
        printf("sq_ring ptr = 0x%p\n", ptr);
@@ -356,12 +343,12 @@ static int setup_ring(struct submitter *s)
        sring->array = ptr + p.sq_off.array;
        sq_ring_mask = *sring->ring_mask;
 
-       s->iocbs = mmap(0, p.sq_entries * sizeof(struct io_uring_iocb),
+       s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
                        PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
-                       IORING_OFF_IOCB);
-       printf("iocbs ptr   = 0x%p\n", s->iocbs);
+                       IORING_OFF_SQES);
+       printf("sqes ptr    = 0x%p\n", s->sqes);
 
-       ptr = mmap(0, p.cq_off.events + p.cq_entries * sizeof(struct io_uring_event),
+       ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
                        PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
                        IORING_OFF_CQ_RING);
        printf("cq_ring ptr = 0x%p\n", ptr);
@@ -369,7 +356,7 @@ static int setup_ring(struct submitter *s)
        cring->tail = ptr + p.cq_off.tail;
        cring->ring_mask = ptr + p.cq_off.ring_mask;
        cring->ring_entries = ptr + p.cq_off.ring_entries;
-       cring->events = ptr + p.cq_off.events;
+       cring->cqes = ptr + p.cq_off.cqes;
        cq_ring_mask = *cring->ring_mask;
        return 0;
 }
@@ -378,7 +365,7 @@ int main(int argc, char *argv[])
 {
        struct submitter *s = &submitters[0];
        unsigned long done, calls, reap, cache_hit, cache_miss;
-       int err, i;
+       int err, i, flags, fd;
        struct rlimit rlim;
        void *ret;
 
@@ -387,6 +374,35 @@ int main(int argc, char *argv[])
                return 1;
        }
 
+       flags = O_RDONLY;
+       if (!buffered)
+               flags |= O_DIRECT;
+
+       i = 1;
+       while (i < argc) {
+               struct file *f = &s->files[s->nr_files];
+
+               fd = open(argv[i], flags);
+               if (fd < 0) {
+                       perror("open");
+                       return 1;
+               }
+               f->fd = fd;
+               if (get_file_size(f)) {
+                       printf("failed getting size of device/file\n");
+                       return 1;
+               }
+               if (f->max_blocks <= 1) {
+                       printf("Zero file/device size?\n");
+                       return 1;
+               }
+               f->max_blocks--;
+
+               printf("Added file %s\n", argv[i]);
+               s->nr_files++;
+               i++;
+       }
+
        rlim.rlim_cur = RLIM_INFINITY;
        rlim.rlim_max = RLIM_INFINITY;
        if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
@@ -414,7 +430,6 @@ int main(int argc, char *argv[])
        }
        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);
-       strcpy(s->filename, argv[1]);
 
        pthread_create(&s->thread, NULL, submitter_fn, s);