#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
-#include <errno.h>
-#include <assert.h>
-#include <time.h>
#include <string.h>
-#include <sys/mman.h>
-#include <sys/poll.h>
+#include <dlfcn.h>
+#include <assert.h>
+
#include "fio.h"
#include "os.h"
-static int fill_timespec(struct timespec *ts)
-{
-#ifdef _POSIX_TIMERS
- if (!clock_gettime(CLOCK_MONOTONIC, ts))
- return 0;
-
- perror("clock_gettime");
-#endif
- return 1;
-}
-
-static unsigned long long ts_utime_since_now(struct timespec *t)
-{
- long long sec, nsec;
- struct timespec now;
-
- if (fill_timespec(&now))
- return 0;
-
- sec = now.tv_sec - t->tv_sec;
- nsec = now.tv_nsec - t->tv_nsec;
- if (sec > 0 && nsec < 0) {
- sec--;
- nsec += 1000000000;
- }
-
- sec *= 1000000;
- nsec /= 1000;
- return sec + nsec;
-}
-
-static int fio_io_sync(struct thread_data *td)
-{
- return fsync(td->fd);
-}
-
-#ifdef FIO_HAVE_LIBAIO
-
-#define ev_to_iou(ev) (struct io_u *) ((unsigned long) (ev)->obj)
-
-struct libaio_data {
- io_context_t aio_ctx;
- struct io_event *aio_events;
-};
-
-static int fio_libaio_io_prep(struct thread_data *td, struct io_u *io_u)
-{
- if (io_u->ddir == DDIR_READ)
- io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
- else
- io_prep_pwrite(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
-
- return 0;
-}
-
-static struct io_u *fio_libaio_event(struct thread_data *td, int event)
-{
- struct libaio_data *ld = td->io_data;
-
- return ev_to_iou(ld->aio_events + event);
-}
-
-static int fio_libaio_getevents(struct thread_data *td, int min, int max,
- struct timespec *t)
-{
- struct libaio_data *ld = td->io_data;
- int r;
-
- do {
- r = io_getevents(ld->aio_ctx, min, max, ld->aio_events, t);
- if (r == -EAGAIN) {
- usleep(100);
- continue;
- } else if (r == -EINTR)
- continue;
- else
- break;
- } while (1);
-
- return r;
-}
-
-static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
-{
- struct libaio_data *ld = td->io_data;
- struct iocb *iocb = &io_u->iocb;
- int ret;
-
- do {
- ret = io_submit(ld->aio_ctx, 1, &iocb);
- if (ret == 1)
- return 0;
- else if (ret == -EAGAIN)
- usleep(100);
- else if (ret == -EINTR)
- continue;
- else
- break;
- } while (1);
-
- return ret;
-
-}
+static LIST_HEAD(engine_list);
-static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
+static int check_engine_ops(struct ioengine_ops *ops)
{
- struct libaio_data *ld = td->io_data;
-
- return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
-}
-
-static void fio_libaio_cleanup(struct thread_data *td)
-{
- struct libaio_data *ld = td->io_data;
-
- if (ld) {
- io_destroy(ld->aio_ctx);
- if (ld->aio_events)
- free(ld->aio_events);
-
- free(ld);
- td->io_data = NULL;
- }
-}
-
-int fio_libaio_init(struct thread_data *td)
-{
- struct libaio_data *ld = malloc(sizeof(*ld));
-
- memset(ld, 0, sizeof(*ld));
- if (io_queue_init(td->iodepth, &ld->aio_ctx)) {
- td_verror(td, errno);
+ if (ops->version != FIO_IOOPS_VERSION) {
+ log_err("bad ioops version %d (want %d)\n", ops->version, FIO_IOOPS_VERSION);
return 1;
}
- td->io_prep = fio_libaio_io_prep;
- td->io_queue = fio_libaio_queue;
- td->io_getevents = fio_libaio_getevents;
- td->io_event = fio_libaio_event;
- td->io_cancel = fio_libaio_cancel;
- td->io_cleanup = fio_libaio_cleanup;
- td->io_sync = fio_io_sync;
-
- ld->aio_events = malloc(td->iodepth * sizeof(struct io_event));
- td->io_data = ld;
- return 0;
-}
-
-#else /* FIO_HAVE_LIBAIO */
-
-int fio_libaio_init(struct thread_data *td)
-{
- return EINVAL;
-}
-
-#endif /* FIO_HAVE_LIBAIO */
-
-#ifdef FIO_HAVE_POSIXAIO
-
-struct posixaio_data {
- struct io_u **aio_events;
-};
-
-static int fio_posixaio_cancel(struct thread_data *td, struct io_u *io_u)
-{
- int r = aio_cancel(td->fd, &io_u->aiocb);
-
- if (r == 1 || r == AIO_CANCELED)
- return 0;
-
- return 1;
-}
-
-static int fio_posixaio_prep(struct thread_data *td, struct io_u *io_u)
-{
- struct aiocb *aiocb = &io_u->aiocb;
-
- aiocb->aio_fildes = td->fd;
- aiocb->aio_buf = io_u->buf;
- aiocb->aio_nbytes = io_u->buflen;
- aiocb->aio_offset = io_u->offset;
-
- io_u->seen = 0;
- return 0;
-}
-
-static int fio_posixaio_getevents(struct thread_data *td, int min, int max,
- struct timespec *t)
-{
- struct posixaio_data *pd = td->io_data;
- struct list_head *entry;
- struct timespec start;
- int r, have_timeout = 0;
-
- if (t && !fill_timespec(&start))
- have_timeout = 1;
-
- r = 0;
-restart:
- list_for_each(entry, &td->io_u_busylist) {
- struct io_u *io_u = list_entry(entry, struct io_u, list);
- int err;
-
- if (io_u->seen)
- continue;
-
- err = aio_error(&io_u->aiocb);
- switch (err) {
- default:
- io_u->error = err;
- case ECANCELED:
- case 0:
- pd->aio_events[r++] = io_u;
- io_u->seen = 1;
- break;
- case EINPROGRESS:
- break;
- }
-
- if (r >= max)
- break;
- }
-
- if (r >= min)
- return r;
-
- if (have_timeout) {
- unsigned long long usec;
-
- usec = (t->tv_sec * 1000000) + (t->tv_nsec / 1000);
- if (ts_utime_since_now(&start) > usec)
- return r;
- }
-
/*
- * hrmpf, we need to wait for more. we should use aio_suspend, for
- * now just sleep a little and recheck status of busy-and-not-seen
+ * cpu thread doesn't need to provide anything
*/
- usleep(1000);
- goto restart;
-}
-
-static struct io_u *fio_posixaio_event(struct thread_data *td, int event)
-{
- struct posixaio_data *pd = td->io_data;
-
- return pd->aio_events[event];
-}
-
-static int fio_posixaio_queue(struct thread_data *td, struct io_u *io_u)
-{
- struct aiocb *aiocb = &io_u->aiocb;
- int ret;
-
- if (io_u->ddir == DDIR_READ)
- ret = aio_read(aiocb);
- else
- ret = aio_write(aiocb);
-
- if (ret)
- io_u->error = errno;
-
- return io_u->error;
-}
-
-static void fio_posixaio_cleanup(struct thread_data *td)
-{
- struct posixaio_data *pd = td->io_data;
+ if (ops->flags & FIO_CPUIO)
+ return 0;
- if (pd) {
- free(pd->aio_events);
- free(pd);
- td->io_data = NULL;
+ if (!ops->queue) {
+ log_err("%s: no queue handler\n", ops->name);
+ return 1;
}
-}
-
-int fio_posixaio_init(struct thread_data *td)
-{
- struct posixaio_data *pd = malloc(sizeof(*pd));
-
- pd->aio_events = malloc(td->iodepth * sizeof(struct io_u *));
-
- td->io_prep = fio_posixaio_prep;
- td->io_queue = fio_posixaio_queue;
- td->io_getevents = fio_posixaio_getevents;
- td->io_event = fio_posixaio_event;
- td->io_cancel = fio_posixaio_cancel;
- td->io_cleanup = fio_posixaio_cleanup;
- td->io_sync = fio_io_sync;
-
- td->io_data = pd;
- return 0;
-}
-
-#else /* FIO_HAVE_POSIXAIO */
-
-int fio_posixaio_init(struct thread_data *td)
-{
- return EINVAL;
-}
-
-#endif /* FIO_HAVE_POSIXAIO */
-
-struct syncio_data {
- struct io_u *last_io_u;
-};
-
-static int fio_syncio_getevents(struct thread_data *td, int min, int max,
- struct timespec *t)
-{
- assert(max <= 1);
/*
- * we can only have one finished io_u for sync io, since the depth
- * is always 1
+ * sync engines only need a ->queue()
*/
- if (list_empty(&td->io_u_busylist))
+ if (ops->flags & FIO_SYNCIO)
return 0;
-
- return 1;
-}
-
-static struct io_u *fio_syncio_event(struct thread_data *td, int event)
-{
- struct syncio_data *sd = td->io_data;
-
- assert(event == 0);
-
- return sd->last_io_u;
-}
-
-static int fio_syncio_prep(struct thread_data *td, struct io_u *io_u)
-{
- if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
- td_verror(td, errno);
+
+ if (!ops->event) {
+ log_err("%s: no event handler\n", ops->name);
return 1;
}
-
+ if (!ops->getevents) {
+ log_err("%s: no getevents handler\n", ops->name);
+ return 1;
+ }
+ if (!ops->queue) {
+ log_err("%s: no queue handler\n", ops->name);
+ return 1;
+ }
+
return 0;
}
-static int fio_syncio_queue(struct thread_data *td, struct io_u *io_u)
+void unregister_ioengine(struct ioengine_ops *ops)
{
- struct syncio_data *sd = td->io_data;
- int ret;
-
- if (io_u->ddir == DDIR_READ)
- ret = read(td->fd, io_u->buf, io_u->buflen);
- else
- ret = write(td->fd, io_u->buf, io_u->buflen);
-
- if ((unsigned int) ret != io_u->buflen) {
- if (ret > 0) {
- io_u->resid = io_u->buflen - ret;
- io_u->error = EIO;
- } else
- io_u->error = errno;
- }
-
- if (!io_u->error)
- sd->last_io_u = io_u;
-
- return io_u->error;
+ list_del(&ops->list);
+ INIT_LIST_HEAD(&ops->list);
}
-static void fio_syncio_cleanup(struct thread_data *td)
+void register_ioengine(struct ioengine_ops *ops)
{
- if (td->io_data) {
- free(td->io_data);
- td->io_data = NULL;
- }
+ INIT_LIST_HEAD(&ops->list);
+ list_add_tail(&ops->list, &engine_list);
}
-int fio_syncio_init(struct thread_data *td)
+static struct ioengine_ops *find_ioengine(const char *name)
{
- struct syncio_data *sd = malloc(sizeof(*sd));
-
- td->io_prep = fio_syncio_prep;
- td->io_queue = fio_syncio_queue;
- td->io_getevents = fio_syncio_getevents;
- td->io_event = fio_syncio_event;
- td->io_cancel = NULL;
- td->io_cleanup = fio_syncio_cleanup;
- td->io_sync = fio_io_sync;
-
- sd->last_io_u = NULL;
- td->io_data = sd;
- return 0;
-}
+ struct ioengine_ops *ops;
+ struct list_head *entry;
+ char engine[16];
-static int fio_mmapio_queue(struct thread_data *td, struct io_u *io_u)
-{
- unsigned long long real_off = io_u->offset - td->file_offset;
- struct syncio_data *sd = td->io_data;
+ strncpy(engine, name, sizeof(engine) - 1);
- if (io_u->ddir == DDIR_READ)
- memcpy(io_u->buf, td->mmap + real_off, io_u->buflen);
- else
- memcpy(td->mmap + real_off, io_u->buf, io_u->buflen);
+ if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
+ strcpy(engine, "libaio");
- /*
- * not really direct, but should drop the pages from the cache
- */
- if (td->odirect) {
- if (msync(td->mmap + real_off, io_u->buflen, MS_SYNC) < 0)
- io_u->error = errno;
- if (madvise(td->mmap + real_off, io_u->buflen, MADV_DONTNEED) < 0)
- io_u->error = errno;
+ list_for_each(entry, &engine_list) {
+ ops = list_entry(entry, struct ioengine_ops, list);
+ if (!strcmp(engine, ops->name))
+ return ops;
}
- if (!io_u->error)
- sd->last_io_u = io_u;
-
- return io_u->error;
+ return NULL;
}
-static int fio_mmapio_sync(struct thread_data *td)
+static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
+ const char *engine_lib)
{
- return msync(td->mmap, td->file_size, MS_SYNC);
-}
+ struct ioengine_ops *ops;
+ void *dlhandle;
-int fio_mmapio_init(struct thread_data *td)
-{
- struct syncio_data *sd = malloc(sizeof(*sd));
-
- td->io_prep = NULL;
- td->io_queue = fio_mmapio_queue;
- td->io_getevents = fio_syncio_getevents;
- td->io_event = fio_syncio_event;
- td->io_cancel = NULL;
- td->io_cleanup = fio_syncio_cleanup;
- td->io_sync = fio_mmapio_sync;
-
- sd->last_io_u = NULL;
- td->io_data = sd;
- return 0;
-}
-
-#ifdef FIO_HAVE_SGIO
-
-struct sgio_cmd {
- unsigned char cdb[10];
- int nr;
-};
-
-struct sgio_data {
- struct sgio_cmd *cmds;
- struct io_u **events;
- unsigned int bs;
-};
-
-static void sgio_hdr_init(struct sgio_data *sd, struct sg_io_hdr *hdr,
- struct io_u *io_u, int fs)
-{
- struct sgio_cmd *sc = &sd->cmds[io_u->index];
-
- memset(hdr, 0, sizeof(*hdr));
- memset(sc->cdb, 0, sizeof(sc->cdb));
-
- hdr->interface_id = 'S';
- hdr->cmdp = sc->cdb;
- hdr->cmd_len = sizeof(sc->cdb);
- hdr->pack_id = io_u->index;
- hdr->usr_ptr = io_u;
-
- if (fs) {
- hdr->dxferp = io_u->buf;
- hdr->dxfer_len = io_u->buflen;
+ dlerror();
+ dlhandle = dlopen(engine_lib, RTLD_LAZY);
+ if (!dlhandle) {
+ td_vmsg(td, -1, dlerror(), "dlopen");
+ return NULL;
}
-}
-
-static int fio_sgio_getevents(struct thread_data *td, int min, int max,
- struct timespec *t)
-{
- struct sgio_data *sd = td->io_data;
- struct pollfd pfd = { .fd = td->fd, .events = POLLIN };
- void *buf = malloc(max * sizeof(struct sg_io_hdr));
- int left = max, ret, events, i, r = 0, fl = 0;
/*
- * don't block for !events
+ * Unlike the included modules, external engines should have a
+ * non-static ioengine structure that we can reference.
*/
- if (!min) {
- fl = fcntl(td->fd, F_GETFL);
- fcntl(td->fd, F_SETFL, fl | O_NONBLOCK);
- }
-
- while (left) {
- do {
- if (!min)
- break;
- poll(&pfd, 1, -1);
- if (pfd.revents & POLLIN)
- break;
- } while (1);
-
- ret = read(td->fd, buf, left * sizeof(struct sg_io_hdr));
- if (ret < 0) {
- if (errno == EAGAIN)
- break;
- td_verror(td, errno);
- r = -1;
- break;
- } else if (!ret)
- break;
-
- events = ret / sizeof(struct sg_io_hdr);
- left -= events;
- r += events;
-
- for (i = 0; i < events; i++) {
- struct sg_io_hdr *hdr = (struct sg_io_hdr *) buf + i;
-
- sd->events[i] = hdr->usr_ptr;
- }
+ ops = dlsym(dlhandle, "ioengine");
+ if (!ops) {
+ td_vmsg(td, -1, dlerror(), "dlsym");
+ dlclose(dlhandle);
+ return NULL;
}
- if (!min)
- fcntl(td->fd, F_SETFL, fl);
-
- free(buf);
- return r;
+ ops->dlhandle = dlhandle;
+ return ops;
}
-static int fio_sgio_ioctl_doio(struct thread_data *td, struct io_u *io_u)
+struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
{
- struct sgio_data *sd = td->io_data;
- struct sg_io_hdr *hdr = &io_u->hdr;
+ struct ioengine_ops *ops, *ret;
+ char engine[16];
- sd->events[0] = io_u;
+ strncpy(engine, name, sizeof(engine) - 1);
- return ioctl(td->fd, SG_IO, hdr);
-}
-
-static int fio_sgio_rw_doio(struct thread_data *td, struct io_u *io_u, int sync)
-{
- struct sg_io_hdr *hdr = &io_u->hdr;
- int ret;
+ /*
+ * linux libaio has alias names, so convert to what we want
+ */
+ if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
+ strcpy(engine, "libaio");
- ret = write(td->fd, hdr, sizeof(*hdr));
- if (ret < 0)
- return errno;
+ ops = find_ioengine(engine);
+ if (!ops)
+ ops = dlopen_ioengine(td, name);
- if (sync) {
- ret = read(td->fd, hdr, sizeof(*hdr));
- if (ret < 0)
- return errno;
+ if (!ops) {
+ log_err("fio: engine %s not loadable\n", name);
+ return NULL;
}
- return 0;
-}
-
-static int fio_sgio_doio(struct thread_data *td, struct io_u *io_u, int sync)
-{
- if (td->filetype == FIO_TYPE_BD)
- return fio_sgio_ioctl_doio(td, io_u);
-
- return fio_sgio_rw_doio(td, io_u, sync);
-}
-
-static int fio_sgio_sync(struct thread_data *td)
-{
- struct sgio_data *sd = td->io_data;
- struct sg_io_hdr *hdr;
- struct io_u *io_u;
- int ret;
-
- io_u = __get_io_u(td);
- if (!io_u)
- return ENOMEM;
-
- hdr = &io_u->hdr;
- sgio_hdr_init(sd, hdr, io_u, 0);
- hdr->dxfer_direction = SG_DXFER_NONE;
+ /*
+ * Check that the required methods are there.
+ */
+ if (check_engine_ops(ops))
+ return NULL;
- hdr->cmdp[0] = 0x35;
+ ret = malloc(sizeof(*ret));
+ memcpy(ret, ops, sizeof(*ret));
+ ret->data = NULL;
- ret = fio_sgio_doio(td, io_u, 1);
- put_io_u(td, io_u);
return ret;
}
-static int fio_sgio_prep(struct thread_data *td, struct io_u *io_u)
-{
- struct sg_io_hdr *hdr = &io_u->hdr;
- struct sgio_data *sd = td->io_data;
- int nr_blocks, lba;
-
- if (io_u->buflen & (sd->bs - 1)) {
- fprintf(stderr, "read/write not sector aligned\n");
- return EINVAL;
- }
-
- sgio_hdr_init(sd, hdr, io_u, 1);
-
- if (io_u->ddir == DDIR_READ) {
- hdr->dxfer_direction = SG_DXFER_FROM_DEV;
- hdr->cmdp[0] = 0x28;
- } else {
- hdr->dxfer_direction = SG_DXFER_TO_DEV;
- hdr->cmdp[0] = 0x2a;
- }
-
- nr_blocks = io_u->buflen / sd->bs;
- lba = io_u->offset / sd->bs;
- hdr->cmdp[2] = (lba >> 24) & 0xff;
- hdr->cmdp[3] = (lba >> 16) & 0xff;
- hdr->cmdp[4] = (lba >> 8) & 0xff;
- hdr->cmdp[5] = lba & 0xff;
- hdr->cmdp[7] = (nr_blocks >> 8) & 0xff;
- hdr->cmdp[8] = nr_blocks & 0xff;
- return 0;
-}
-
-static int fio_sgio_queue(struct thread_data *td, struct io_u *io_u)
+void close_ioengine(struct thread_data *td)
{
- struct sg_io_hdr *hdr = &io_u->hdr;
- int ret;
+ if (td->io_ops->cleanup)
+ td->io_ops->cleanup(td);
- ret = fio_sgio_doio(td, io_u, 0);
+ if (td->io_ops->dlhandle)
+ dlclose(td->io_ops->dlhandle);
- if (ret < 0)
- io_u->error = errno;
- else if (hdr->status) {
- io_u->resid = hdr->resid;
- io_u->error = EIO;
- }
-
- return io_u->error;
+#if 0
+ /* we can't do this for threads, so just leak it, it's exiting */
+ free(td->io_ops);
+#endif
+ td->io_ops = NULL;
}
-static struct io_u *fio_sgio_event(struct thread_data *td, int event)
+int td_io_prep(struct thread_data *td, struct io_u *io_u)
{
- struct sgio_data *sd = td->io_data;
+ if (td->io_ops->prep)
+ return td->io_ops->prep(td, io_u);
- return sd->events[event];
+ return 0;
}
-static int fio_sgio_get_bs(struct thread_data *td, unsigned int *bs)
+int td_io_getevents(struct thread_data *td, int min, int max,
+ struct timespec *t)
{
- struct sgio_data *sd = td->io_data;
- struct io_u *io_u;
- struct sg_io_hdr *hdr;
- unsigned char buf[8];
- int ret;
-
- io_u = __get_io_u(td);
- assert(io_u);
+ if (min > 0 && td->io_ops->commit) {
+ int r = td->io_ops->commit(td);
- hdr = &io_u->hdr;
- sgio_hdr_init(sd, hdr, io_u, 0);
- memset(buf, 0, sizeof(buf));
-
- hdr->cmdp[0] = 0x25;
- hdr->dxfer_direction = SG_DXFER_FROM_DEV;
- hdr->dxferp = buf;
- hdr->dxfer_len = sizeof(buf);
-
- ret = fio_sgio_doio(td, io_u, 1);
- if (ret) {
- put_io_u(td, io_u);
- return ret;
+ if (r < 0)
+ return r;
}
+ if (td->io_ops->getevents)
+ return td->io_ops->getevents(td, min, max, t);
- *bs = (buf[4] << 24) | (buf[5] << 16) | (buf[6] << 8) | buf[7];
- put_io_u(td, io_u);
return 0;
}
-int fio_sgio_init(struct thread_data *td)
+int td_io_queue(struct thread_data *td, struct io_u *io_u)
{
- struct sgio_data *sd;
- unsigned int bs;
int ret;
- sd = malloc(sizeof(*sd));
- sd->cmds = malloc(td->iodepth * sizeof(struct sgio_cmd));
- sd->events = malloc(td->iodepth * sizeof(struct io_u *));
- td->io_data = sd;
+ assert((io_u->flags & IO_U_F_FLIGHT) == 0);
+ io_u->flags |= IO_U_F_FLIGHT;
- if (td->filetype == FIO_TYPE_BD) {
- if (ioctl(td->fd, BLKSSZGET, &bs) < 0) {
- td_verror(td, errno);
- return 1;
- }
- } else if (td->filetype == FIO_TYPE_CHAR) {
- int version;
+ if (td->io_ops->flags & FIO_SYNCIO) {
+ fio_gettime(&io_u->issue_time, NULL);
- if (ioctl(td->fd, SG_GET_VERSION_NUM, &version) < 0) {
- td_verror(td, errno);
- return 1;
- }
-
- ret = fio_sgio_get_bs(td, &bs);
- if (ret)
- return ret;
- } else {
- fprintf(stderr, "ioengine sgio only works on block devices\n");
- return 1;
+ /*
+ * for a sync engine, set the timeout upfront
+ */
+ if (mtime_since(&td->timeout_end, &io_u->issue_time) < IO_U_TIMEOUT)
+ io_u_set_timeout(td);
}
- sd->bs = bs;
-
- td->io_prep = fio_sgio_prep;
- td->io_queue = fio_sgio_queue;
+ if (io_u->ddir != DDIR_SYNC)
+ td->io_issues[io_u->ddir]++;
- if (td->filetype == FIO_TYPE_BD)
- td->io_getevents = fio_syncio_getevents;
- else
- td->io_getevents = fio_sgio_getevents;
+ ret = td->io_ops->queue(td, io_u);
- td->io_event = fio_sgio_event;
- td->io_cancel = NULL;
- td->io_cleanup = fio_syncio_cleanup;
- td->io_sync = fio_sgio_sync;
+ if (ret == FIO_Q_QUEUED) {
+ int r;
- /*
- * we want to do it, regardless of whether odirect is set or not
- */
- td->override_sync = 1;
- return 0;
-}
-
-#else /* FIO_HAVE_SGIO */
-
-int fio_sgio_init(struct thread_data *td)
-{
- return EINVAL;
-}
-
-#endif /* FIO_HAVE_SGIO */
-
-#ifdef FIO_HAVE_SPLICE
-struct spliceio_data {
- struct io_u *last_io_u;
- int pipe[2];
-};
-
-static struct io_u *fio_spliceio_event(struct thread_data *td, int event)
-{
- struct spliceio_data *sd = td->io_data;
-
- assert(event == 0);
-
- return sd->last_io_u;
-}
-
-/*
- * For splice reading, we unfortunately cannot (yet) vmsplice the other way.
- * So just splice the data from the file into the pipe, and use regular
- * read to fill the buffer. Doesn't make a lot of sense, but...
- */
-static int fio_splice_read(struct thread_data *td, struct io_u *io_u)
-{
- struct spliceio_data *sd = td->io_data;
- int ret, ret2, buflen;
- off_t offset;
- void *p;
-
- offset = io_u->offset;
- buflen = io_u->buflen;
- p = io_u->buf;
- while (buflen) {
- int this_len = buflen;
-
- if (this_len > SPLICE_DEF_SIZE)
- this_len = SPLICE_DEF_SIZE;
-
- ret = splice(td->fd, &offset, sd->pipe[1], NULL, this_len, SPLICE_F_MORE);
- if (ret < 0) {
- if (errno == ENODATA || errno == EAGAIN)
- continue;
-
- return errno;
- }
-
- buflen -= ret;
-
- while (ret) {
- ret2 = read(sd->pipe[0], p, ret);
- if (ret2 < 0)
- return errno;
-
- ret -= ret2;
- p += ret2;
+ td->io_u_queued++;
+ if (td->io_u_queued > td->iodepth_batch) {
+ r = td_io_commit(td);
+ if (r < 0)
+ return r;
}
}
- return io_u->buflen;
-}
-
-/*
- * For splice writing, we can vmsplice our data buffer directly into a
- * pipe and then splice that to a file.
- */
-static int fio_splice_write(struct thread_data *td, struct io_u *io_u)
-{
- struct spliceio_data *sd = td->io_data;
- struct iovec iov[1] = {
- {
- .iov_base = io_u->buf,
- .iov_len = io_u->buflen,
- }
- };
- struct pollfd pfd = { .fd = sd->pipe[1], .events = POLLOUT, };
- off_t off = io_u->offset;
- int ret, ret2;
-
- while (iov[0].iov_len) {
- if (poll(&pfd, 1, -1) < 0)
- return errno;
-
- ret = vmsplice(sd->pipe[1], iov, 1, SPLICE_F_NONBLOCK);
- if (ret < 0)
- return errno;
-
- iov[0].iov_len -= ret;
- iov[0].iov_base += ret;
+ if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
+ fio_gettime(&io_u->issue_time, NULL);
- while (ret) {
- ret2 = splice(sd->pipe[0], NULL, td->fd, &off, ret, 0);
- if (ret2 < 0)
- return errno;
-
- ret -= ret2;
- }
+ /*
+ * async engine, set the timeout here
+ */
+ if (ret == FIO_Q_QUEUED &&
+ mtime_since(&td->timeout_end, &io_u->issue_time) < IO_U_TIMEOUT)
+ io_u_set_timeout(td);
}
- return io_u->buflen;
+ return ret;
}
-static int fio_spliceio_queue(struct thread_data *td, struct io_u *io_u)
+int td_io_init(struct thread_data *td)
{
- struct spliceio_data *sd = td->io_data;
- int ret;
+ if (td->io_ops->init)
+ return td->io_ops->init(td);
- if (io_u->ddir == DDIR_READ)
- ret = fio_splice_read(td, io_u);
- else
- ret = fio_splice_write(td, io_u);
-
- if ((unsigned int) ret != io_u->buflen) {
- if (ret > 0) {
- io_u->resid = io_u->buflen - ret;
- io_u->error = ENODATA;
- } else
- io_u->error = errno;
- }
-
- if (!io_u->error)
- sd->last_io_u = io_u;
-
- return io_u->error;
+ return 0;
}
-static void fio_spliceio_cleanup(struct thread_data *td)
+int td_io_commit(struct thread_data *td)
{
- struct spliceio_data *sd = td->io_data;
-
- if (sd) {
- close(sd->pipe[0]);
- close(sd->pipe[1]);
- free(sd);
- td->io_data = NULL;
- }
-}
+ if (!td->cur_depth)
+ return 0;
-int fio_spliceio_init(struct thread_data *td)
-{
- struct spliceio_data *sd = malloc(sizeof(*sd));
-
- td->io_queue = fio_spliceio_queue;
- td->io_getevents = fio_syncio_getevents;
- td->io_event = fio_spliceio_event;
- td->io_cancel = NULL;
- td->io_cleanup = fio_spliceio_cleanup;
- td->io_sync = fio_io_sync;
-
- sd->last_io_u = NULL;
- if (pipe(sd->pipe) < 0) {
- td_verror(td, errno);
- free(sd);
- return 1;
- }
+ td->io_u_queued = 0;
+ if (td->io_ops->commit)
+ return td->io_ops->commit(td);
- td->io_data = sd;
return 0;
}
-
-#else /* FIO_HAVE_SPLICE */
-
-int fio_spliceio_init(struct thread_data *td)
-{
- return EINVAL;
-}
-
-#endif /* FIO_HAVE_SPLICE */
projects / fio.git / blobdiff