X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=blktrace.c;h=bb0bcbe60ac52e2f7f6c1fda34983f704e94ada0;hp=29eed50c391c422da566eb063cb2bc4863bd19a1;hb=76a00ec6684317902e536365b89531431e7f8867;hpb=25f488581e510c1b440f2e76842ec23f3dad7b57 diff --git a/blktrace.c b/blktrace.c index 29eed50c..bb0bcbe6 100644 --- a/blktrace.c +++ b/blktrace.c @@ -9,6 +9,7 @@ #include "flist.h" #include "fio.h" #include "blktrace_api.h" +#include "lib/linux-dev-lookup.h" #define TRACE_FIFO_SIZE 8192 @@ -108,67 +109,6 @@ int is_blktrace(const char *filename, int *need_swap) return 0; } -static int lookup_device(struct thread_data *td, char *path, unsigned int maj, - unsigned int min) -{ - struct dirent *dir; - struct stat st; - int found = 0; - DIR *D; - - D = opendir(path); - if (!D) - return 0; - - while ((dir = readdir(D)) != NULL) { - char full_path[256]; - - if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, "..")) - continue; - - sprintf(full_path, "%s%s%s", path, FIO_OS_PATH_SEPARATOR, dir->d_name); - if (lstat(full_path, &st) == -1) { - perror("lstat"); - break; - } - - if (S_ISDIR(st.st_mode)) { - found = lookup_device(td, full_path, maj, min); - if (found) { - strcpy(path, full_path); - break; - } - } - - if (!S_ISBLK(st.st_mode)) - continue; - - /* - * If replay_redirect is set then always return this device - * upon lookup which overrides the device lookup based on - * major minor in the actual blktrace - */ - if (td->o.replay_redirect) { - dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden" - " with: %s\n", maj, min, - td->o.replay_redirect); - strcpy(path, td->o.replay_redirect); - found = 1; - break; - } - - if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) { - dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min); - strcpy(path, full_path); - found = 1; - break; - } - } - - closedir(D); - return found; -} - #define FMINORBITS 20 #define FMINORMASK ((1U << FMINORBITS) - 1) #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS)) @@ -212,9 +152,16 @@ static int trace_add_file(struct thread_data *td, __u32 device) } strcpy(dev, "/dev"); - if (lookup_device(td, dev, maj, min)) { + if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) { int fileno; + if (td->o.replay_redirect) + dprint(FD_BLKTRACE, "device lookup: %d/%d\n overridden" + " with: %s\n", maj, min, + td->o.replay_redirect); + else + dprint(FD_BLKTRACE, "device lookup: %d/%d\n", maj, min); + dprint(FD_BLKTRACE, "add devices %s\n", dev); fileno = add_file_exclusive(td, dev); td->o.open_files++; @@ -260,11 +207,11 @@ static void handle_trace_notify(struct blk_io_trace *t) { switch (t->action) { case BLK_TN_PROCESS: - log_info("blktrace: got process notify: %x, %d\n", + dprint(FD_BLKTRACE, "got process notify: %x, %d\n", t->action, t->pid); break; case BLK_TN_TIMESTAMP: - log_info("blktrace: got timestamp notify: %x, %d\n", + dprint(FD_BLKTRACE, "got timestamp notify: %x, %d\n", t->action, t->pid); break; case BLK_TN_MESSAGE: @@ -275,8 +222,10 @@ static void handle_trace_notify(struct blk_io_trace *t) } } -static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t, - unsigned long long ttime, unsigned long *ios) +static void handle_trace_discard(struct thread_data *td, + struct blk_io_trace *t, + unsigned long long ttime, + unsigned long *ios, unsigned int *bs) { struct io_piece *ipo = malloc(sizeof(*ipo)); int fileno; @@ -284,7 +233,10 @@ static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t, init_ipo(ipo); fileno = trace_add_file(td, t->device); - ios[DDIR_WRITE]++; + ios[DDIR_TRIM]++; + if (t->bytes > bs[DDIR_TRIM]) + bs[DDIR_TRIM] = t->bytes; + td->o.size += t->bytes; memset(ipo, 0, sizeof(*ipo)); @@ -329,20 +281,30 @@ static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t, * due to internal workings of the block layer. */ static void handle_trace(struct thread_data *td, struct blk_io_trace *t, - unsigned long long ttime, unsigned long *ios, - unsigned int *bs) + unsigned long *ios, unsigned int *bs) { + static unsigned long long last_ttime; + unsigned long long delay; + if ((t->action & 0xffff) != __BLK_TA_QUEUE) return; - if (t->action & BLK_TC_ACT(BLK_TC_PC)) - return; + + if (!(t->action & BLK_TC_ACT(BLK_TC_NOTIFY))) { + if (!last_ttime || td->o.no_stall) { + last_ttime = t->time; + delay = 0; + } else { + delay = t->time - last_ttime; + last_ttime = t->time; + } + } if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY)) handle_trace_notify(t); else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD)) - handle_trace_discard(td, t, ttime, ios); + handle_trace_discard(td, t, delay, ios, bs); else - handle_trace_fs(td, t, ttime, ios, bs); + handle_trace_fs(td, t, delay, ios, bs); } static void byteswap_trace(struct blk_io_trace *t) @@ -360,21 +322,65 @@ static void byteswap_trace(struct blk_io_trace *t) t->pdu_len = fio_swap16(t->pdu_len); } +static int t_is_write(struct blk_io_trace *t) +{ + return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0; +} + +static enum fio_ddir t_get_ddir(struct blk_io_trace *t) +{ + if (t->action & BLK_TC_ACT(BLK_TC_READ)) + return DDIR_READ; + else if (t->action & BLK_TC_ACT(BLK_TC_WRITE)) + return DDIR_WRITE; + else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD)) + return DDIR_TRIM; + + return DDIR_INVAL; +} + +static void depth_inc(struct blk_io_trace *t, int *depth) +{ + enum fio_ddir ddir; + + ddir = t_get_ddir(t); + if (ddir != DDIR_INVAL) + depth[ddir]++; +} + +static void depth_dec(struct blk_io_trace *t, int *depth) +{ + enum fio_ddir ddir; + + ddir = t_get_ddir(t); + if (ddir != DDIR_INVAL) + depth[ddir]--; +} + +static void depth_end(struct blk_io_trace *t, int *this_depth, int *depth) +{ + enum fio_ddir ddir = DDIR_INVAL; + + ddir = t_get_ddir(t); + if (ddir != DDIR_INVAL) { + depth[ddir] = max(depth[ddir], this_depth[ddir]); + this_depth[ddir] = 0; + } +} + /* * Load a blktrace file by reading all the blk_io_trace entries, and storing * them as io_pieces like the fio text version would do. */ int load_blktrace(struct thread_data *td, const char *filename, int need_swap) { - unsigned long long ttime, delay; struct blk_io_trace t; - unsigned long ios[2], skipped_writes; - unsigned int cpu; - unsigned int rw_bs[2]; + unsigned long ios[DDIR_RWDIR_CNT], skipped_writes; + unsigned int rw_bs[DDIR_RWDIR_CNT]; struct fifo *fifo; int fd, i, old_state; struct fio_file *f; - int this_depth, depth; + int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth; fd = open(filename, O_RDONLY); if (fd < 0) { @@ -388,12 +394,14 @@ int load_blktrace(struct thread_data *td, const char *filename, int need_swap) td->o.size = 0; - cpu = 0; - ttime = 0; - ios[0] = ios[1] = 0; - rw_bs[0] = rw_bs[1] = 0; + for (i = 0; i < DDIR_RWDIR_CNT; i++) { + ios[i] = 0; + rw_bs[i] = 0; + this_depth[i] = 0; + depth[i] = 0; + } + skipped_writes = 0; - this_depth = depth = 0; do { int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t)); @@ -429,38 +437,20 @@ int load_blktrace(struct thread_data *td, const char *filename, int need_swap) } if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) { if ((t.action & 0xffff) == __BLK_TA_QUEUE) - this_depth++; - else if ((t.action & 0xffff) == __BLK_TA_COMPLETE) { - depth = max(depth, this_depth); - this_depth = 0; - } - if (!ttime) { - ttime = t.time; - cpu = t.cpu; - } - - delay = 0; - if (cpu == t.cpu) - delay = t.time - ttime; - if ((t.action & BLK_TC_ACT(BLK_TC_WRITE)) && read_only) + depth_inc(&t, this_depth); + else if (((t.action & 0xffff) == __BLK_TA_BACKMERGE) || + ((t.action & 0xffff) == __BLK_TA_FRONTMERGE)) + depth_dec(&t, this_depth); + else if ((t.action & 0xffff) == __BLK_TA_COMPLETE) + depth_end(&t, this_depth, depth); + + if (t_is_write(&t) && read_only) { skipped_writes++; - else { - /* - * set delay to zero if no_stall enabled for - * fast replay - */ - if (td->o.no_stall) - delay = 0; - - handle_trace(td, &t, delay, ios, rw_bs); + continue; } - - ttime = t.time; - cpu = t.cpu; - } else { - delay = 0; - handle_trace(td, &t, delay, ios, rw_bs); } + + handle_trace(td, &t, ios, rw_bs); } while (1); for (i = 0; i < td->files_index; i++) { @@ -482,8 +472,14 @@ int load_blktrace(struct thread_data *td, const char *filename, int need_swap) * For stacked devices, we don't always get a COMPLETE event so * the depth grows to insane values. Limit it to something sane(r). */ - if (!depth || depth > 1024) - depth = 1024; + max_depth = 0; + for (i = 0; i < DDIR_RWDIR_CNT; i++) { + if (depth[i] > 1024) + depth[i] = 1024; + else if (!depth[i] && ios[i]) + depth[i] = 1; + max_depth = max(depth[i], max_depth); + } if (skipped_writes) log_err("fio: %s skips replay of %lu writes due to read-only\n", @@ -492,7 +488,7 @@ int load_blktrace(struct thread_data *td, const char *filename, int need_swap) if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) { log_err("fio: found no ios in blktrace data\n"); return 1; - } else if (ios[DDIR_READ] && !ios[DDIR_READ]) { + } else if (ios[DDIR_READ] && !ios[DDIR_WRITE]) { td->o.td_ddir = TD_DDIR_READ; td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; } else if (!ios[DDIR_READ] && ios[DDIR_WRITE]) { @@ -502,20 +498,22 @@ int load_blktrace(struct thread_data *td, const char *filename, int need_swap) td->o.td_ddir = TD_DDIR_RW; td->o.max_bs[DDIR_READ] = rw_bs[DDIR_READ]; td->o.max_bs[DDIR_WRITE] = rw_bs[DDIR_WRITE]; + td->o.max_bs[DDIR_TRIM] = rw_bs[DDIR_TRIM]; } /* * We need to do direct/raw ios to the device, to avoid getting - * read-ahead in our way. + * read-ahead in our way. But only do so if the minimum block size + * is a multiple of 4k, otherwise we don't know if it's safe to do so. */ - td->o.odirect = 1; + if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095)) + td->o.odirect = 1; /* - * we don't know if this option was set or not. it defaults to 1, - * so we'll just guess that we should override it if it's still 1 + * If depth wasn't manually set, use probed depth */ - if (td->o.iodepth != 1) - td->o.iodepth = depth; + if (!fio_option_is_set(&td->o, iodepth)) + td->o.iodepth = td->o.iodepth_low = max_depth; return 0; err: