X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=8cc348a835fed67a2310c04f84b4608281db1315;hp=f451d1a756141e3a29eaad8f3f2fb74d113ae82e;hb=5e0baa7f423995c4f75070d6a2ee56619e195a36;hpb=da3758bfa36ee05e25e73f33760f17750bd0687c diff --git a/io_u.c b/io_u.c index f451d1a7..8cc348a8 100644 --- a/io_u.c +++ b/io_u.c @@ -8,6 +8,7 @@ #include "fio.h" #include "hash.h" #include "verify.h" +#include "trim.h" #include "lib/rand.h" struct io_completion_data { @@ -41,10 +42,12 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u) struct fio_file *f = io_u->file; unsigned long long block; unsigned int blocks, nr_blocks; + int busy_check; block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs; nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; blocks = 0; + busy_check = !(io_u->flags & IO_U_F_BUSY_OK); while (nr_blocks) { unsigned int this_blocks, mask; @@ -54,7 +57,11 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u) * 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(f, block)) + if (!busy_check) { + blocks = nr_blocks; + break; + } + if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block)) break; idx = RAND_MAP_IDX(f, block); @@ -97,6 +104,8 @@ static unsigned long long last_block(struct thread_data *td, struct fio_file *f, unsigned long long max_blocks; unsigned long long max_size; + assert(ddir_rw(ddir)); + /* * Hmm, should we make sure that ->io_size <= ->real_file_size? */ @@ -117,16 +126,20 @@ static unsigned long long last_block(struct thread_data *td, struct fio_file *f, static int get_next_free_block(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir, unsigned long long *b) { - unsigned long long min_bs = td->o.rw_min_bs; + unsigned long long min_bs = td->o.rw_min_bs, lastb; int i; + lastb = last_block(td, f, ddir); + if (!lastb) + return 1; + i = f->last_free_lookup; *b = (i * BLOCKS_PER_MAP); while ((*b) * min_bs < f->real_file_size && (*b) * min_bs < f->io_size) { if (f->file_map[i] != (unsigned int) -1) { *b += ffz(f->file_map[i]); - if (*b > last_block(td, f, ddir)) + if (*b > lastb) break; f->last_free_lookup = i; return 0; @@ -143,14 +156,17 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f, static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir, unsigned long long *b) { - unsigned long long r; + unsigned long long r, lastb; int loops = 5; + lastb = last_block(td, f, ddir); + if (!lastb) + return 1; + do { r = os_random_long(&td->random_state); dprint(FD_RANDOM, "off rand %llu\n", r); - *b = (last_block(td, f, ddir) - 1) - * (r / ((unsigned long long) OS_RAND_MAX + 1.0)); + *b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0)); /* * if we are not maintaining a random map, we are done. @@ -190,6 +206,67 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, return get_next_free_block(td, f, ddir, b); } +static int get_next_rand_block(struct thread_data *td, struct fio_file *f, + enum fio_ddir ddir, unsigned long long *b) +{ + if (get_next_rand_offset(td, f, ddir, b)) { + dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", + f->file_name, f->last_pos, f->real_file_size); + return 1; + } + + return 0; +} + +static int get_next_seq_block(struct thread_data *td, struct fio_file *f, + enum fio_ddir ddir, unsigned long long *b) +{ + assert(ddir_rw(ddir)); + + if (f->last_pos < f->real_file_size) { + *b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; + return 0; + } + + return 1; +} + +static int get_next_block(struct thread_data *td, struct io_u *io_u, + enum fio_ddir ddir, int rw_seq, unsigned long long *b) +{ + struct fio_file *f = io_u->file; + int ret; + + assert(ddir_rw(ddir)); + + if (rw_seq) { + if (td_random(td)) + ret = get_next_rand_block(td, f, ddir, b); + else + ret = get_next_seq_block(td, f, ddir, b); + } else { + io_u->flags |= IO_U_F_BUSY_OK; + + if (td->o.rw_seq == RW_SEQ_SEQ) { + ret = get_next_seq_block(td, f, ddir, b); + if (ret) + ret = get_next_rand_block(td, f, ddir, b); + } else if (td->o.rw_seq == RW_SEQ_IDENT) { + if (f->last_start != -1ULL) + *b = (f->last_start - f->file_offset) + / td->o.min_bs[ddir]; + else + *b = 0; + ret = 0; + } else { + log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); + ret = 1; + } + } + + return ret; +} + /* * 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 @@ -200,28 +277,18 @@ static int __get_next_offset(struct thread_data *td, struct io_u *io_u) struct fio_file *f = io_u->file; unsigned long long b; enum fio_ddir ddir = io_u->ddir; + int rw_seq_hit = 0; - if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { - td->ddir_nr = td->o.ddir_nr; + assert(ddir_rw(ddir)); - if (get_next_rand_offset(td, f, ddir, &b)) { - dprint(FD_IO, "%s: getting rand offset failed\n", - f->file_name); - return 1; - } - } else { - if (f->last_pos >= f->real_file_size) { - if (!td_random(td) || - get_next_rand_offset(td, f, ddir, &b)) { - dprint(FD_IO, "%s: pos %llu > size %llu\n", - f->file_name, f->last_pos, - f->real_file_size); - return 1; - } - } else - b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; + if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { + rw_seq_hit = 1; + td->ddir_seq_nr = td->o.ddir_seq_nr; } + if (get_next_block(td, io_u, ddir, rw_seq_hit, &b)) + return 1; + io_u->offset = b * td->o.ba[ddir]; if (io_u->offset >= f->io_size) { dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", @@ -256,6 +323,8 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u) unsigned int minbs, maxbs; long r; + assert(ddir_rw(ddir)); + minbs = td->o.min_bs[ddir]; maxbs = td->o.max_bs[ddir]; @@ -336,6 +405,8 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) struct timeval t; long usec; + assert(ddir_rw(ddir)); + if (td->rate_pending_usleep[ddir] <= 0) return ddir; @@ -437,6 +508,17 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td) return td->rwmix_ddir; } +static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) +{ + io_u->ddir = get_rw_ddir(td); + + if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) && + td->o.barrier_blocks && + !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && + td->io_issues[DDIR_WRITE]) + io_u->flags |= IO_U_F_BARRIER; +} + void put_file_log(struct thread_data *td, struct fio_file *f) { int ret = put_file(td, f); @@ -479,7 +561,7 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u) td_io_u_lock(td); __io_u->flags |= IO_U_F_FREE; - if ((__io_u->flags & IO_U_F_FLIGHT) && !ddir_sync(__io_u->ddir)) + if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir)) td->io_issues[__io_u->ddir]--; __io_u->flags &= ~IO_U_F_FLIGHT; @@ -496,12 +578,12 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u) if (td->io_ops->flags & FIO_NOIO) goto out; - io_u->ddir = get_rw_ddir(td); + set_rw_ddir(td, io_u); /* - * fsync() or fdatasync(), we are done + * fsync() or fdatasync() or trim etc, we are done */ - if (ddir_sync(io_u->ddir)) + if (!ddir_rw(io_u->ddir)) goto out; /* @@ -899,6 +981,7 @@ again: if (io_u) { assert(io_u->flags & IO_U_F_FREE); io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF); + io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER); io_u->error = 0; flist_del(&io_u->list); @@ -918,21 +1001,31 @@ again: return io_u; } -/* - * Return an io_u to be processed. Gets a buflen and offset, sets direction, - * etc. The returned io_u is fully ready to be prepped and submitted. - */ -struct io_u *get_io_u(struct thread_data *td) +static int check_get_trim(struct thread_data *td, struct io_u *io_u) { - struct fio_file *f; - struct io_u *io_u; + if (td->o.trim_backlog && td->trim_entries) { + int get_trim = 0; - io_u = __get_io_u(td); - if (!io_u) { - dprint(FD_IO, "__get_io_u failed\n"); - return NULL; + if (td->trim_batch) { + td->trim_batch--; + get_trim = 1; + } else if (!(td->io_hist_len % td->o.trim_backlog) && + td->last_ddir != DDIR_READ) { + td->trim_batch = td->o.trim_batch; + if (!td->trim_batch) + td->trim_batch = td->o.trim_backlog; + get_trim = 1; + } + + if (get_trim && !get_next_trim(td, io_u)) + return 1; } + return 0; +} + +static int check_get_verify(struct thread_data *td, struct io_u *io_u) +{ if (td->o.verify_backlog && td->io_hist_len) { int get_verify = 0; @@ -948,9 +1041,32 @@ struct io_u *get_io_u(struct thread_data *td) } if (get_verify && !get_next_verify(td, io_u)) - goto out; + return 1; } + return 0; +} + +/* + * Return an io_u to be processed. Gets a buflen and offset, sets direction, + * etc. The returned io_u is fully ready to be prepped and submitted. + */ +struct io_u *get_io_u(struct thread_data *td) +{ + struct fio_file *f; + struct io_u *io_u; + + io_u = __get_io_u(td); + if (!io_u) { + dprint(FD_IO, "__get_io_u failed\n"); + return NULL; + } + + if (check_get_verify(td, io_u)) + goto out; + if (check_get_trim(td, io_u)) + goto out; + /* * from a requeue, io_u already setup */ @@ -971,18 +1087,26 @@ struct io_u *get_io_u(struct thread_data *td) f = io_u->file; assert(fio_file_open(f)); - if (!ddir_sync(io_u->ddir)) { + if (ddir_rw(io_u->ddir)) { if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); goto err_put; } + f->last_start = io_u->offset; f->last_pos = io_u->offset + io_u->buflen; if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE) populate_verify_io_u(td, io_u); else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE) io_u_fill_buffer(td, io_u, io_u->xfer_buflen); + else if (io_u->ddir == DDIR_READ) { + /* + * Reset the buf_filled parameters so next time if the + * buffer is used for writes it is refilled. + */ + io_u->buf_filled_len = 0; + } } /* @@ -992,6 +1116,7 @@ struct io_u *get_io_u(struct thread_data *td) io_u->xfer_buflen = io_u->buflen; out: + assert(io_u->file); if (!td_io_prep(td, io_u)) { if (!td->o.disable_slat) fio_gettime(&io_u->start_time, NULL); @@ -1005,7 +1130,10 @@ err_put: void io_u_log_error(struct thread_data *td, struct io_u *io_u) { - const char *msg[] = { "read", "write", "sync" }; + const char *msg[] = { "read", "write", "sync", "datasync", + "sync_file_range", "wait", "trim" }; + + log_err("fio: io_u error"); @@ -1035,7 +1163,7 @@ static void io_completed(struct thread_data *td, struct io_u *io_u, td_io_u_lock(td); assert(io_u->flags & IO_U_F_FLIGHT); - io_u->flags &= ~IO_U_F_FLIGHT; + io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK); td_io_u_unlock(td); if (ddir_sync(io_u->ddir)) { @@ -1051,7 +1179,7 @@ static void io_completed(struct thread_data *td, struct io_u *io_u, td->last_was_sync = 0; td->last_ddir = io_u->ddir; - if (!io_u->error) { + if (!io_u->error && ddir_rw(io_u->ddir)) { unsigned int bytes = io_u->buflen - io_u->resid; const enum fio_ddir idx = io_u->ddir; const enum fio_ddir odx = io_u->ddir ^ 1; @@ -1079,7 +1207,13 @@ static void io_completed(struct thread_data *td, struct io_u *io_u, if (!td->o.disable_clat || !td->o.disable_bw) lusec = utime_since(&io_u->issue_time, &icd->time); + if (!td->o.disable_lat) { + unsigned long tusec; + tusec = utime_since(&io_u->start_time, + &icd->time); + add_lat_sample(td, idx, tusec, bytes); + } if (!td->o.disable_clat) { add_clat_sample(td, idx, lusec, bytes); io_u_mark_latency(td, lusec); @@ -1111,7 +1245,7 @@ static void io_completed(struct thread_data *td, struct io_u *io_u, if (ret && !icd->error) icd->error = ret; } - } else { + } else if (io_u->error) { icd->error = io_u->error; io_u_log_error(td, io_u); } @@ -1240,20 +1374,8 @@ void io_u_queued(struct thread_data *td, struct io_u *io_u) void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, unsigned int max_bs) { - long *ptr = io_u->buf; - - if (!td->o.zero_buffers) { - unsigned long r = __rand(&__fio_rand_state); - - if (sizeof(int) != sizeof(*ptr)) - r *= (unsigned long) __rand(&__fio_rand_state); - - while ((void *) ptr - io_u->buf < max_bs) { - *ptr = r; - ptr++; - r *= GOLDEN_RATIO_PRIME; - r >>= 3; - } - } else - memset(ptr, 0, max_bs); + if (!td->o.zero_buffers) + fill_random_buf(io_u->buf, max_bs); + else + memset(io_u->buf, 0, max_bs); }