X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=0487857e74851b3752d38747b5307fcbebbf4416;hp=7e5bfcc190585362a075d47b1134fbb049d5f1d2;hb=d460eb318f784876812bedfed382afb00e57c62a;hpb=afab5e46cf0a984a87426329b3157e38d9a419a4 diff --git a/io_u.c b/io_u.c index 7e5bfcc1..0487857e 100644 --- a/io_u.c +++ b/io_u.c @@ -3,13 +3,17 @@ #include #include #include +#include #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 */ @@ -32,10 +36,10 @@ 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; + unsigned int min_bs = td->o.rw_min_bs; + struct fio_file *f = io_u->file; unsigned long long block; unsigned int blocks; unsigned int nr_blocks; @@ -47,7 +51,11 @@ static void mark_random_map(struct thread_data *td, struct fio_file *f, while (blocks < nr_blocks) { unsigned int idx, bit; - if (!random_map_free(td, f, block)) + /* + * If we have a mixed random workload, we may + * encounter blocks we already did IO to. + */ + if (!td->o.ddir_nr && !random_map_free(td, f, block)) break; idx = RAND_MAP_IDX(td, f, block); @@ -74,7 +82,7 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f, i = f->last_free_lookup; *b = (i * BLOCKS_PER_MAP); - while ((*b) * td->rw_min_bs < f->real_file_size) { + while ((*b) * td->o.rw_min_bs < f->real_file_size) { if (f->file_map[i] != -1UL) { *b += ffz(f->file_map[i]); f->last_free_lookup = i; @@ -88,71 +96,110 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f, return 1; } +static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, + int ddir, unsigned long long *b) +{ + unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir]; + unsigned long long r, rb; + int loops = 5; + + do { + r = os_random_long(&td->random_state); + if (!max_blocks) + *b = 0; + else + *b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); + if (td->o.norandommap) + break; + rb = *b + (f->file_offset / td->o.min_bs[ddir]); + loops--; + } while (!random_map_free(td, f, rb) && loops); + + /* + * if we failed to retrieve a truly random offset within + * the loops assigned, see if there are free ones left at all + */ + if (!loops && get_next_free_block(td, f, b)) + return 1; + + return 0; +} + /* * For random io, generate a random new block and see if it's used. Repeat * 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) { - unsigned long long max_blocks = f->file_size / td->min_bs[ddir]; - int loops = 5; + unsigned long long b; - do { - r = os_random_long(&td->random_state); - 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]); - loops--; - } while (!random_map_free(td, f, rb) && loops); + if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { + td->ddir_nr = td->o.ddir_nr; - /* - * if we failed to retrieve a truly random offset within - * the loops assigned, see if there are free ones left at all - */ - if (!loops && get_next_free_block(td, f, &b)) + if (get_next_rand_offset(td, f, ddir, &b)) + return 1; + } else { + if (f->last_pos >= f->real_file_size) return 1; - } else - b = f->last_pos / td->min_bs[ddir]; - io_u->offset = (b * td->min_bs[ddir]) + f->file_offset; + b = f->last_pos / td->o.min_bs[ddir]; + } + + io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset; if (io_u->offset >= f->real_file_size) return 1; 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) { const int ddir = io_u->ddir; unsigned int buflen; long r; - if (td->min_bs[ddir] == td->max_bs[ddir]) - buflen = td->min_bs[ddir]; + if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) + buflen = td->o.min_bs[ddir]; else { r = os_random_long(&td->bsrange_state); - buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0)); - if (!td->bs_unaligned) - buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1); + buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0)); + if (!td->o.bs_unaligned) + buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1); } - while (buflen + io_u->offset > f->real_file_size) { - if (buflen == td->min_bs[ddir]) - return 0; + return buflen; +} - buflen = td->min_bs[ddir]; - } +static void set_rwmix_bytes(struct thread_data *td) +{ + unsigned long long rbytes; + unsigned int diff; - return buflen; + /* + * we do time or byte based switch. this is needed because + * buffered writes may issue a lot quicker than they complete, + * whereas reads do not. + */ + rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes; + diff = td->o.rwmix[td->rwmix_ddir ^ 1]; + + td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff); +} + +static inline enum fio_ddir get_rand_ddir(struct thread_data *td) +{ + unsigned int v; + long r; + + r = os_random_long(&td->rwmix_state); + v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); + if (v < td->o.rwmix[DDIR_READ]) + return DDIR_READ; + + return DDIR_WRITE; } /* @@ -165,23 +212,45 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td) if (td_rw(td)) { struct timeval now; unsigned long elapsed; + unsigned int cycle; fio_gettime(&now, NULL); elapsed = mtime_since_now(&td->rwmix_switch); + /* + * if this is the first cycle, make it shorter + */ + cycle = td->o.rwmixcycle; + if (!td->rwmix_bytes) + cycle /= 10; + /* * Check if it's time to seed a new data direction. */ - if (elapsed >= td->rwmixcycle) { - unsigned int v; - long r; - - r = os_random_long(&td->rwmix_state); - v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); - if (v < td->rwmixread) - td->rwmix_ddir = DDIR_READ; - else - td->rwmix_ddir = DDIR_WRITE; + if (elapsed >= cycle || + td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) { + unsigned long long max_bytes; + enum fio_ddir ddir; + + /* + * Put a top limit on how many bytes we do for + * one data direction, to avoid overflowing the + * ranges too much + */ + ddir = get_rand_ddir(td); + max_bytes = td->this_io_bytes[ddir]; + if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) { + if (!td->rw_end_set[ddir]) { + td->rw_end_set[ddir] = 1; + memcpy(&td->rw_end[ddir], &now, sizeof(now)); + } + ddir ^= 1; + } + + if (ddir != td->rwmix_ddir) + set_rwmix_bytes(td); + + td->rwmix_ddir = ddir; memcpy(&td->rwmix_switch, &now, sizeof(now)); } return td->rwmix_ddir; @@ -193,6 +262,9 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td) void put_io_u(struct thread_data *td, struct io_u *io_u) { + assert((io_u->flags & IO_U_F_FREE) == 0); + io_u->flags |= IO_U_F_FREE; + io_u->file = NULL; list_del(&io_u->list); list_add(&io_u->list, &td->io_u_freelist); @@ -203,29 +275,31 @@ 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. */ - if (td->read_iolog) + if (td->o.read_iolog) return read_iolog_get(td, io_u); /* * see if it's time to sync */ - if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks) - && td->io_issues[DDIR_WRITE] && should_fsync(td)) { + if (td->o.fsync_blocks && + !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && + td->io_issues[DDIR_WRITE] && should_fsync(td)) { io_u->ddir = DDIR_SYNC; - io_u->file = f; - return 0; + goto out; } io_u->ddir = get_rw_ddir(td); @@ -234,33 +308,36 @@ 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_random(td) && !td->o.norandommap) + mark_random_map(td, io_u); /* * If using a write iolog, store this entry. */ - if (td->write_iolog_file) +out: + if (td->o.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++; @@ -278,8 +355,8 @@ 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) @@ -289,23 +366,23 @@ static void io_u_mark_latency(struct thread_data *td, unsigned long msec) switch (msec) { default: index++; - case 1024 ... 2047: + case 1000 ... 1999: index++; - case 512 ... 1023: + case 750 ... 999: index++; - case 256 ... 511: + case 500 ... 749: index++; - case 128 ... 255: + case 250 ... 499: index++; - case 64 ... 127: + case 100 ... 249: index++; - case 32 ... 63: + case 50 ... 99: index++; - case 16 ... 31: + case 20 ... 49: index++; - case 8 ... 15: + case 10 ... 19: index++; - case 4 ... 7: + case 4 ... 9: index++; case 2 ... 3: index++; @@ -313,10 +390,36 @@ 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->o.nr_files * (r / (RAND_MAX + 1.0))); + f = &td->files[fno]; + if (f->flags & FIO_FILE_DONE) + continue; + + 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; @@ -325,10 +428,15 @@ static struct fio_file *get_next_file(struct thread_data *td) f = &td->files[td->next_file]; td->next_file++; - if (td->next_file >= td->nr_files) + if (td->next_file >= td->o.nr_files) td->next_file = 0; - if (f->fd != -1) + if (f->flags & FIO_FILE_DONE) { + f = NULL; + continue; + } + + if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) break; f = NULL; @@ -337,6 +445,44 @@ 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->o.nr_files <= td->files_index); + + if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) + return NULL; + + f = td->file_service_file; + if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) + return f; + + if (td->o.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->nr_open_files || td->nr_done_files >= td->o.nr_files) + return NULL; + + if (td->o.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; @@ -349,14 +495,17 @@ 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) { + assert(io_u->flags & IO_U_F_FREE); + io_u->flags &= ~IO_U_F_FREE; + io_u->error = 0; 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; @@ -370,6 +519,7 @@ 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) @@ -379,33 +529,48 @@ struct io_u *get_io_u(struct thread_data *td) * from a requeue, io_u already setup */ if (io_u->file) - return io_u; + 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; + } - io_u->file = f; +set_file: + io_u->file = f; - if (td->zone_bytes >= td->zone_size) { - td->zone_bytes = 0; - f->last_pos += td->zone_skip; - } + if (!fill_io_u(td, io_u)) + break; - if (fill_io_u(td, f, io_u)) { - put_io_u(td, io_u); - return NULL; - } + /* + * No more to do for this file, close it + */ + io_u->file = NULL; + td_io_close_file(td, f); + f->flags |= FIO_FILE_DONE; + td->nr_done_files++; - if (io_u->buflen + io_u->offset > f->real_file_size) { - if (td->io_ops->flags & FIO_RAWIO) { - put_io_u(td, io_u); - return NULL; + /* + * probably not the right place to do this, but see + * if we need to open a new file + */ + if (td->nr_open_files < td->o.open_files && + td->o.open_files != td->o.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); - io_u->buflen = f->real_file_size - io_u->offset; + if (td->zone_bytes >= td->o.zone_size) { + td->zone_bytes = 0; + f->last_pos += td->o.zone_skip; } if (io_u->ddir != DDIR_SYNC) { @@ -416,13 +581,15 @@ struct io_u *get_io_u(struct thread_data *td) f->last_pos = io_u->offset + io_u->buflen; - if (td->verify != VERIFY_NONE) + if (td->o.verify != VERIFY_NONE) populate_verify_io_u(td, io_u); } /* * Set io data pointers. */ + io_u->endpos = io_u->offset + io_u->buflen; +out: io_u->xfer_buf = io_u->buf; io_u->xfer_buflen = io_u->buflen; @@ -435,11 +602,33 @@ struct io_u *get_io_u(struct thread_data *td) 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) { unsigned long msec; + 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; @@ -457,7 +646,7 @@ static void io_completed(struct thread_data *td, struct io_u *io_u, td->zone_bytes += bytes; td->this_io_bytes[idx] += bytes; - io_u->file->last_completed_pos = io_u->offset + io_u->buflen; + io_u->file->last_completed_pos = io_u->endpos; msec = mtime_since(&io_u->issue_time, &icd->time); @@ -465,26 +654,27 @@ static void io_completed(struct thread_data *td, struct io_u *io_u, add_bw_sample(td, idx, &icd->time); io_u_mark_latency(td, msec); - if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE) + if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE && + td->o.verify != VERIFY_NONE) log_io_piece(td, 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(td, 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; @@ -505,47 +695,140 @@ static void ios_completed(struct thread_data *td, } } -long io_u_sync_complete(struct thread_data *td, struct io_u *io_u, - endio_handler *handler) +/* + * Complete a single io_u for the sync engines. + */ +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; } -long io_u_queued_complete(struct thread_data *td, int min_events, - endio_handler *handler) - +/* + * Called to complete min_events number of io for the async engines. + */ +long io_u_queued_complete(struct thread_data *td, int min_events) { struct io_completion_data icd; + struct timespec *tvp = NULL; int ret; + struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; - if (min_events > 0) { - ret = td_io_commit(td); - if (ret < 0) { - td_verror(td, -ret); - return ret; - } - } + if (!min_events) + tvp = &ts; - ret = td_io_getevents(td, min_events, td->cur_depth, NULL); + ret = td_io_getevents(td, min_events, td->cur_depth, tvp); if (ret < 0) { - td_verror(td, -ret); + td_verror(td, -ret, "td_io_getevents"); return ret; } 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; } + +/* + * Call when io_u is really queued, to update the submission latency. + */ +void io_u_queued(struct thread_data *td, struct io_u *io_u) +{ + unsigned long slat_time; + + slat_time = mtime_since(&io_u->start_time, &io_u->issue_time); + add_slat_sample(td, io_u->ddir, slat_time); +} + +#ifdef FIO_USE_TIMEOUT +void io_u_set_timeout(struct thread_data *td) +{ + assert(td->cur_depth); + + td->timer.it_interval.tv_sec = 0; + td->timer.it_interval.tv_usec = 0; + td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; + td->timer.it_value.tv_usec = 0; + 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) +{ +} +#endif + +#ifdef FIO_USE_TIMEOUT +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"); + + /* + * TLS would be nice... + */ + td = NULL; + for_each_td(__td, i) { + if (__td->pid == pid) { + td = __td; + break; + } + } + + if (!td) { + log_err("fio: io_u timeout, can't find job\n"); + exit(1); + } + + if (!td->cur_depth) { + log_err("fio: timeout without pending work?\n"); + return; + } + + log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); + + 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 + +void io_u_init_timeout(void) +{ +#ifdef FIO_USE_TIMEOUT + signal(SIGALRM, io_u_timeout_handler); +#endif +}