X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=4750ac600a4272142bc85ccf9333382fcb1a4d28;hp=69f2f19d426053d8fc487affc991534a55c3f4f9;hb=e6203e9e73039537d41810e91e3dfe6abeb919d9;hpb=b3605062146bce0136918763bb8eb584478ae042 diff --git a/io_u.c b/io_u.c index 69f2f19d..4750ac60 100644 --- a/io_u.c +++ b/io_u.c @@ -6,7 +6,6 @@ #include #include "fio.h" -#include "os.h" /* * Change this define to play with the timeout handling @@ -39,7 +38,7 @@ static int random_map_free(struct thread_data *td, struct fio_file *f, */ 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; @@ -52,7 +51,11 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u) 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 == 1 && !random_map_free(td, f, block)) break; idx = RAND_MAP_IDX(td, f, block); @@ -79,9 +82,9 @@ 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]); + *b += fio_ffz(f->file_map[i]); f->last_free_lookup = i; return 0; } @@ -93,6 +96,55 @@ 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 we are not maintaining a random map, we are done. + */ + if (td->o.norandommap) + return 0; + + /* + * calculate map offset and chec if it's free + */ + rb = *b + (f->file_offset / td->o.min_bs[ddir]); + if (random_map_free(td, f, rb)) + return 0; + + } while (--loops); + + /* + * we get here, if we didn't suceed in looking up a block. generate + * a random start offset into the filemap, and find the first free + * block from there. + */ + loops = 10; + do { + f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0)); + if (!get_next_free_block(td, f, b)) + return 0; + + r = os_random_long(&td->random_state); + } while (--loops); + + /* + * that didn't work either, try exhaustive search from the start + */ + f->last_free_lookup = 0; + return get_next_free_block(td, f, b); +} + /* * 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 @@ -102,32 +154,22 @@ 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; + unsigned long long b; - if (td_random(td)) { - unsigned long long max_blocks = f->file_size / td->min_bs[ddir]; - int loops = 5; + if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { + td->ddir_nr = td->o.ddir_nr; - 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 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 - b = f->last_pos / td->min_bs[ddir]; + } else { + if (f->last_pos >= f->real_file_size) { + if (!td_random(td) || get_next_rand_offset(td, f, ddir, &b)) + return 1; + } else + b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; + } - io_u->offset = (b * td->min_bs[ddir]) + f->file_offset; + io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset; if (io_u->offset >= f->real_file_size) return 1; @@ -136,30 +178,54 @@ static int get_next_offset(struct thread_data *td, 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; - 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; - - buflen = td->min_bs[ddir]; - } + if (io_u->offset + buflen > io_u->file->real_file_size) + buflen = td->o.min_bs[ddir]; return buflen; } +static void set_rwmix_bytes(struct thread_data *td) +{ + unsigned long long rbytes; + unsigned int diff; + + /* + * 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; +} + /* * Return the data direction for the next io_u. If the job is a * mixed read/write workload, check the rwmix cycle and switch if @@ -170,23 +236,45 @@ 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; @@ -201,6 +289,9 @@ 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; + if (io_u->file) + put_file(td, io_u->file); + io_u->file = NULL; list_del(&io_u->list); list_add(&io_u->list, &td->io_u_freelist); @@ -222,23 +313,27 @@ void requeue_io_u(struct thread_data *td, 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) - 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; - return 0; + goto out; } io_u->ddir = get_rw_ddir(td); + /* + * See if it's time to switch to a new zone + */ + if (td->zone_bytes >= td->o.zone_size) { + td->zone_bytes = 0; + io_u->file->last_pos += td->o.zone_skip; + td->io_skip_bytes += td->o.zone_skip; + } + /* * No log, let the seq/rand engine retrieve the next buflen and * position. @@ -250,18 +345,20 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u) if (!io_u->buflen) return 1; + if (io_u->offset + io_u->buflen > io_u->file->real_file_size) + return 1; + /* * mark entry before potentially trimming io_u */ - if (!td->read_iolog && td_random(td) && !td->norandommap) + 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) - write_iolog_put(td, io_u); - +out: + log_io_u(td, io_u); return 0; } @@ -274,17 +371,22 @@ void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) switch (td->cur_depth) { default: - index++; + index = 6; + break; case 32 ... 63: - index++; + index = 5; + break; case 16 ... 31: - index++; + index = 4; + break; case 8 ... 15: - index++; + index = 3; + break; case 4 ... 7: - index++; + index = 2; + break; case 2 ... 3: - index++; + index = 1; case 1: break; } @@ -293,54 +395,118 @@ void io_u_mark_depth(struct thread_data *td, struct io_u *io_u) td->ts.total_io_u[io_u->ddir]++; } -static void io_u_mark_latency(struct thread_data *td, unsigned long msec) +static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) +{ + int index = 0; + + assert(usec < 1000); + + switch (usec) { + case 750 ... 999: + index = 9; + break; + case 500 ... 749: + index = 8; + break; + case 250 ... 499: + index = 7; + break; + case 100 ... 249: + index = 6; + break; + case 50 ... 99: + index = 5; + break; + case 20 ... 49: + index = 4; + break; + case 10 ... 19: + index = 3; + break; + case 4 ... 9: + index = 2; + break; + case 2 ... 3: + index = 1; + case 0 ... 1: + break; + } + + assert(index < FIO_IO_U_LAT_U_NR); + td->ts.io_u_lat_u[index]++; +} + +static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) { int index = 0; switch (msec) { default: - index++; + index = 11; + break; case 1000 ... 1999: - index++; + index = 10; + break; case 750 ... 999: - index++; + index = 9; + break; case 500 ... 749: - index++; + index = 8; + break; case 250 ... 499: - index++; + index = 7; + break; case 100 ... 249: - index++; + index = 6; + break; case 50 ... 99: - index++; + index = 5; + break; case 20 ... 49: - index++; + index = 4; + break; case 10 ... 19: - index++; + index = 3; + break; case 4 ... 9: - index++; + index = 2; + break; case 2 ... 3: - index++; + index = 1; case 0 ... 1: break; } - td->ts.io_u_lat[index]++; + assert(index < FIO_IO_U_LAT_M_NR); + td->ts.io_u_lat_m[index]++; +} + +static void io_u_mark_latency(struct thread_data *td, unsigned long usec) +{ + if (usec < 1000) + io_u_mark_lat_usec(td, usec); + else + io_u_mark_lat_msec(td, usec / 1000); } /* * Get next file to service by choosing one at random */ -static struct fio_file *get_next_file_rand(struct thread_data *td) +static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, + int badf) { - unsigned int fileno; struct fio_file *f; + int fno; do { long r = os_random_long(&td->next_file_state); - fileno = (unsigned int) ((double) (td->open_files * r) / (RAND_MAX + 1.0)); - f = &td->files[fileno]; - if (f->open) + 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); } @@ -348,7 +514,8 @@ static struct fio_file *get_next_file_rand(struct thread_data *td) /* * 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) +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; @@ -357,10 +524,15 @@ static struct fio_file *get_next_file_rr(struct thread_data *td) f = &td->files[td->next_file]; td->next_file++; - if (td->next_file >= td->open_files) + if (td->next_file >= td->o.nr_files) td->next_file = 0; - if (f->open) + if (f->flags & FIO_FILE_DONE) { + f = NULL; + continue; + } + + if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) break; f = NULL; @@ -373,23 +545,84 @@ static struct fio_file *get_next_file(struct thread_data *td) { struct fio_file *f; - if (!td->nr_open_files) + 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->open && td->file_service_left--) + 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); + 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); + 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; +} + +static int set_io_u_file(struct thread_data *td, struct io_u *io_u) +{ + struct fio_file *f; + + do { + f = get_next_file(td); + if (!f) + return 1; + +set_file: + io_u->file = f; + get_file(f); + + if (!fill_io_u(td, io_u)) + break; + + /* + * td_io_close() does a put_file() as well, so no need to + * do that here. + */ + io_u->file = NULL; + td_io_close_file(td, f); + f->flags |= FIO_FILE_DONE; + td->nr_done_files++; + + /* + * 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 || td_io_open_file(td, f)) + return 1; + + goto set_file; + } + } while (1); + + return 0; +} + + struct io_u *__get_io_u(struct thread_data *td) { struct io_u *io_u = NULL; @@ -437,79 +670,43 @@ struct io_u *get_io_u(struct thread_data *td) if (io_u->file) goto out; - do { - f = get_next_file(td); - if (!f) { - put_io_u(td, io_u); - return NULL; - } - - io_u->file = f; - - if (!fill_io_u(td, io_u)) - break; - - /* - * 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->nr_files && - td->open_files != td->nr_files) { - int ret = td_io_open_file(td, f); - - if (ret) { - put_io_u(td, io_u); - return NULL; - } - } - } while (1); - - if (td->zone_bytes >= td->zone_size) { - td->zone_bytes = 0; - f->last_pos += td->zone_skip; - } - - 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; - } - - io_u->buflen = f->real_file_size - io_u->offset; - } + /* + * If using an iolog, grab next piece if any available. + */ + if (td->o.read_iolog_file) { + if (read_iolog_get(td, io_u)) + goto err_put; + } else if (set_io_u_file(td, io_u)) + goto err_put; + + f = io_u->file; + assert(f->flags & FIO_FILE_OPEN); if (io_u->ddir != DDIR_SYNC) { - if (!io_u->buflen) { - put_io_u(td, io_u); - return NULL; - } + if (!io_u->buflen) + goto err_put; 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; - if (td_io_prep(td, io_u)) { - put_io_u(td, io_u); - return NULL; + if (!td_io_prep(td, io_u)) { + fio_gettime(&io_u->start_time, NULL); + return io_u; } - - fio_gettime(&io_u->start_time, NULL); - return io_u; +err_put: + put_io_u(td, io_u); + return NULL; } void io_u_log_error(struct thread_data *td, struct io_u *io_u) @@ -532,7 +729,7 @@ void io_u_log_error(struct thread_data *td, struct io_u *io_u) static void io_completed(struct thread_data *td, struct io_u *io_u, struct io_completion_data *icd) { - unsigned long msec; + unsigned long usec; assert(io_u->flags & IO_U_F_FLIGHT); io_u->flags &= ~IO_U_F_FLIGHT; @@ -554,21 +751,22 @@ 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); + usec = utime_since(&io_u->issue_time, &icd->time); - add_clat_sample(td, idx, msec); + add_clat_sample(td, idx, usec); add_bw_sample(td, idx, &icd->time); - io_u_mark_latency(td, msec); + io_u_mark_latency(td, usec); - if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE) + if (td_write(td) && idx == DDIR_WRITE && + td->o.verify != VERIFY_NONE) log_io_piece(td, io_u); icd->bytes_done[idx] += bytes; if (io_u->end_io) { - ret = io_u->end_io(io_u); + ret = io_u->end_io(td, io_u); if (ret && !icd->error) icd->error = ret; } @@ -628,12 +826,10 @@ 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) { - struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; - + if (!min_events) tvp = &ts; - } ret = td_io_getevents(td, min_events, td->cur_depth, tvp); if (ret < 0) { @@ -658,7 +854,7 @@ 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); + slat_time = utime_since(&io_u->start_time, &io_u->issue_time); add_slat_sample(td, io_u->ddir, slat_time); } @@ -722,7 +918,7 @@ static void io_u_timeout_handler(int fio_unused sig) return; } - log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid); + 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);