X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=c78865636b1f52f159bf8ef7d6846deb5fec5ef1;hp=8141991c27ca2619fe630aa903a7c4a959cb1e80;hb=b486f76ad31b593caded74f42468ea9adbdc21a7;hpb=e1161c325f7866bae879e686d1c673ca32ab09ae diff --git a/io_u.c b/io_u.c index 8141991c..c7886563 100644 --- a/io_u.c +++ b/io_u.c @@ -8,9 +8,13 @@ #include "fio.h" #include "os.h" +/* + * Change this define to play with the timeout handling + */ +#undef FIO_USE_TIMEOUT + struct io_completion_data { int nr; /* input */ - endio_handler *handler; /* input */ int error; /* output */ unsigned long bytes_done[2]; /* output */ @@ -33,10 +37,10 @@ static int random_map_free(struct thread_data *td, struct fio_file *f, /* * Mark a given offset as used in the map. */ -static void mark_random_map(struct thread_data *td, struct fio_file *f, - struct io_u *io_u) +static void mark_random_map(struct thread_data *td, struct io_u *io_u) { unsigned int min_bs = td->rw_min_bs; + struct fio_file *f = io_u->file; unsigned long long block; unsigned int blocks; unsigned int nr_blocks; @@ -94,20 +98,23 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f, * until we find a free one. For sequential io, just return the end of * the last io issued. */ -static int get_next_offset(struct thread_data *td, struct fio_file *f, - struct io_u *io_u) +static int get_next_offset(struct thread_data *td, struct io_u *io_u) { + struct fio_file *f = io_u->file; const int ddir = io_u->ddir; unsigned long long b, rb; long r; - if (!td->sequential) { + if (td_random(td)) { unsigned long long max_blocks = f->file_size / td->min_bs[ddir]; int loops = 5; do { r = os_random_long(&td->random_state); - b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); + if (!max_blocks) + b = 0; + else + b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); if (td->norandommap) break; rb = b + (f->file_offset / td->min_bs[ddir]); @@ -130,9 +137,9 @@ static int get_next_offset(struct thread_data *td, struct fio_file *f, return 0; } -static unsigned int get_next_buflen(struct thread_data *td, struct fio_file *f, - struct io_u *io_u) +static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) { + struct fio_file *f = io_u->file; const int ddir = io_u->ddir; unsigned int buflen; long r; @@ -147,8 +154,14 @@ static unsigned int get_next_buflen(struct thread_data *td, struct fio_file *f, } while (buflen + io_u->offset > f->real_file_size) { - if (buflen == td->min_bs[ddir]) + if (buflen == td->min_bs[ddir]) { + if (!td->odirect) { + assert(io_u->offset <= f->real_file_size); + buflen = f->real_file_size - io_u->offset; + return buflen; + } return 0; + } buflen = td->min_bs[ddir]; } @@ -207,14 +220,16 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u) { struct io_u *__io_u = *io_u; + __io_u->flags |= IO_U_F_FREE; + __io_u->flags &= ~IO_U_F_FLIGHT; + list_del(&__io_u->list); list_add_tail(&__io_u->list, &td->io_u_requeues); td->cur_depth--; *io_u = NULL; } -static int fill_io_u(struct thread_data *td, struct fio_file *f, - struct io_u *io_u) +static int fill_io_u(struct thread_data *td, struct io_u *io_u) { /* * If using an iolog, grab next piece if any available. @@ -228,7 +243,6 @@ static int fill_io_u(struct thread_data *td, struct fio_file *f, if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks) && td->io_issues[DDIR_WRITE] && should_fsync(td)) { io_u->ddir = DDIR_SYNC; - io_u->file = f; return 0; } @@ -238,18 +252,18 @@ static int fill_io_u(struct thread_data *td, struct fio_file *f, * No log, let the seq/rand engine retrieve the next buflen and * position. */ - if (get_next_offset(td, f, io_u)) + if (get_next_offset(td, io_u)) return 1; - io_u->buflen = get_next_buflen(td, f, io_u); + io_u->buflen = get_next_buflen(td, io_u); if (!io_u->buflen) return 1; /* * mark entry before potentially trimming io_u */ - if (!td->read_iolog && !td->sequential && !td->norandommap) - mark_random_map(td, f, io_u); + if (!td->read_iolog && td_random(td) && !td->norandommap) + mark_random_map(td, io_u); /* * If using a write iolog, store this entry. @@ -257,14 +271,16 @@ static int fill_io_u(struct thread_data *td, struct fio_file *f, if (td->write_iolog_file) write_iolog_put(td, io_u); - io_u->file = f; return 0; } -static void io_u_mark_depth(struct thread_data *td) +void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) { int index = 0; + if (io_u->ddir == DDIR_SYNC) + return; + switch (td->cur_depth) { default: index++; @@ -282,8 +298,8 @@ static void io_u_mark_depth(struct thread_data *td) break; } - td->io_u_map[index]++; - td->total_io_u++; + td->ts.io_u_map[index]++; + td->ts.total_io_u[io_u->ddir]++; } static void io_u_mark_latency(struct thread_data *td, unsigned long msec) @@ -317,10 +333,34 @@ static void io_u_mark_latency(struct thread_data *td, unsigned long msec) break; } - td->io_u_lat[index]++; + td->ts.io_u_lat[index]++; } -static struct fio_file *get_next_file(struct thread_data *td) +/* + * Get next file to service by choosing one at random + */ +static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, + int badf) +{ + struct fio_file *f; + int fno; + + do { + long r = os_random_long(&td->next_file_state); + + fno = (unsigned int) ((double) td->nr_files * (r / (RAND_MAX + 1.0))); + f = &td->files[fno]; + + if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) + return f; + } while (1); +} + +/* + * Get next file to service by doing round robin between all available ones + */ +static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, + int badf) { unsigned int old_next_file = td->next_file; struct fio_file *f; @@ -332,7 +372,7 @@ static struct fio_file *get_next_file(struct thread_data *td) if (td->next_file >= td->nr_files) td->next_file = 0; - if (f->fd != -1) + if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) break; f = NULL; @@ -341,6 +381,41 @@ static struct fio_file *get_next_file(struct thread_data *td) return f; } +static struct fio_file *get_next_file(struct thread_data *td) +{ + struct fio_file *f; + + assert(td->nr_files <= td->files_index); + + if (!td->nr_open_files) + return NULL; + + f = td->file_service_file; + if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) + return f; + + if (td->file_service_type == FIO_FSERVICE_RR) + f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); + else + f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); + + td->file_service_file = f; + td->file_service_left = td->file_service_nr - 1; + return f; +} + +static struct fio_file *find_next_new_file(struct thread_data *td) +{ + struct fio_file *f; + + if (td->file_service_type == FIO_FSERVICE_RR) + f = get_next_file_rr(td, 0, FIO_FILE_OPEN); + else + f = get_next_file_rand(td, 0, FIO_FILE_OPEN); + + return f; +} + struct io_u *__get_io_u(struct thread_data *td) { struct io_u *io_u = NULL; @@ -353,6 +428,7 @@ struct io_u *__get_io_u(struct thread_data *td) io_u->buflen = 0; io_u->resid = 0; io_u->file = NULL; + io_u->end_io = NULL; } if (io_u) { @@ -363,7 +439,6 @@ struct io_u *__get_io_u(struct thread_data *td) list_del(&io_u->list); list_add(&io_u->list, &td->io_u_busylist); td->cur_depth++; - io_u_mark_depth(td); } return io_u; @@ -377,6 +452,7 @@ struct io_u *get_io_u(struct thread_data *td) { struct fio_file *f; struct io_u *io_u; + int ret; io_u = __get_io_u(td); if (!io_u) @@ -388,24 +464,46 @@ struct io_u *get_io_u(struct thread_data *td) if (io_u->file) goto out; - f = get_next_file(td); - if (!f) { - put_io_u(td, io_u); - return NULL; - } + do { + f = get_next_file(td); + if (!f) { + put_io_u(td, io_u); + return NULL; + } + +set_file: + io_u->file = f; + + if (!fill_io_u(td, io_u)) + break; - io_u->file = f; + /* + * No more to do for this file, close it + */ + io_u->file = NULL; + td_io_close_file(td, f); + + /* + * probably not the right place to do this, but see + * if we need to open a new file + */ + if (td->nr_open_files < td->open_files && + td->open_files != td->nr_files) { + f = find_next_new_file(td); + + if (!f || (ret = td_io_open_file(td, f))) { + put_io_u(td, io_u); + return NULL; + } + goto set_file; + } + } while (1); if (td->zone_bytes >= td->zone_size) { td->zone_bytes = 0; f->last_pos += td->zone_skip; } - if (fill_io_u(td, f, io_u)) { - put_io_u(td, io_u); - return NULL; - } - if (io_u->buflen + io_u->offset > f->real_file_size) { if (td->io_ops->flags & FIO_RAWIO) { put_io_u(td, io_u); @@ -443,6 +541,23 @@ out: return io_u; } +void io_u_log_error(struct thread_data *td, struct io_u *io_u) +{ + const char *msg[] = { "read", "write", "sync" }; + + log_err("fio: io_u error"); + + if (io_u->file) + log_err(" on file %s", io_u->file->file_name); + + log_err(": %s\n", strerror(io_u->error)); + + log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen); + + if (!td->error) + td_verror(td, io_u->error, "io_u error"); +} + static void io_completed(struct thread_data *td, struct io_u *io_u, struct io_completion_data *icd) { @@ -451,6 +566,8 @@ static void io_completed(struct thread_data *td, struct io_u *io_u, assert(io_u->flags & IO_U_F_FLIGHT); io_u->flags &= ~IO_U_F_FLIGHT; + put_file(td, io_u->file); + if (io_u->ddir == DDIR_SYNC) { td->last_was_sync = 1; return; @@ -481,21 +598,21 @@ static void io_completed(struct thread_data *td, struct io_u *io_u, icd->bytes_done[idx] += bytes; - if (icd->handler) { - ret = icd->handler(io_u); + if (io_u->end_io) { + ret = io_u->end_io(io_u); if (ret && !icd->error) icd->error = ret; } - } else + } else { icd->error = io_u->error; + io_u_log_error(td, io_u); + } } -static void init_icd(struct io_completion_data *icd, endio_handler *handler, - int nr) +static void init_icd(struct io_completion_data *icd, int nr) { fio_gettime(&icd->time, NULL); - icd->handler = handler; icd->nr = nr; icd->error = 0; @@ -519,39 +636,31 @@ static void ios_completed(struct thread_data *td, /* * Complete a single io_u for the sync engines. */ -long io_u_sync_complete(struct thread_data *td, struct io_u *io_u, - endio_handler *handler) +long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) { struct io_completion_data icd; - init_icd(&icd, handler, 1); + init_icd(&icd, 1); io_completed(td, io_u, &icd); put_io_u(td, io_u); if (!icd.error) return icd.bytes_done[0] + icd.bytes_done[1]; + td_verror(td, icd.error, "io_u_sync_complete"); return -1; } /* * Called to complete min_events number of io for the async engines. */ -long io_u_queued_complete(struct thread_data *td, int min_events, - endio_handler *handler) - +long io_u_queued_complete(struct thread_data *td, int min_events) { struct io_completion_data icd; struct timespec *tvp = NULL; int ret; - if (min_events > 0) { - ret = td_io_commit(td); - if (ret < 0) { - td_verror(td, -ret, "td_io_commit"); - return ret; - } - } else { + if (!min_events) { struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; tvp = &ts; @@ -564,11 +673,12 @@ long io_u_queued_complete(struct thread_data *td, int min_events, } else if (!ret) return ret; - init_icd(&icd, handler, ret); + init_icd(&icd, ret); ios_completed(td, &icd); if (!icd.error) return icd.bytes_done[0] + icd.bytes_done[1]; + td_verror(td, icd.error, "io_u_queued_complete"); return -1; } @@ -595,6 +705,16 @@ void io_u_set_timeout(struct thread_data *td) setitimer(ITIMER_REAL, &td->timer, NULL); fio_gettime(&td->timeout_end, NULL); } + +static void io_u_dump(struct io_u *io_u) +{ + unsigned long t_start = mtime_since_now(&io_u->start_time); + unsigned long t_issue = mtime_since_now(&io_u->issue_time); + + log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); + log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset); + log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); +} #else void io_u_set_timeout(struct thread_data fio_unused *td) { @@ -606,6 +726,8 @@ static void io_u_timeout_handler(int fio_unused sig) { struct thread_data *td, *__td; pid_t pid = getpid(); + struct list_head *entry; + struct io_u *io_u; int i; log_err("fio: io_u timeout\n"); @@ -632,7 +754,14 @@ static void io_u_timeout_handler(int fio_unused sig) } log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid); - td->error = ETIMEDOUT; + + list_for_each(entry, &td->io_u_busylist) { + io_u = list_entry(entry, struct io_u, list); + + io_u_dump(io_u); + } + + td_verror(td, ETIMEDOUT, "io_u timeout"); exit(1); } #endif