X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=blktrace.c;h=9ce4ae29c8cbe358c4ce0913e84d407721d13c17;hp=64e49f5e4b552d3d4e67dea8ca7eed5dcf6e589e;hb=d1c46c049cfba2028abc45246e2609bcee52d0f3;hpb=4241ea8fb0606f9bd5b7ed1278d3b7825883acf6 diff --git a/blktrace.c b/blktrace.c index 64e49f5e..9ce4ae29 100644 --- a/blktrace.c +++ b/blktrace.c @@ -6,11 +6,11 @@ #include #include -#include "list.h" +#include "flist.h" #include "fio.h" #include "blktrace_api.h" -#define TRACE_FIFO_SIZE 65536 +#define TRACE_FIFO_SIZE 8192 /* * fifo refill frontend, to avoid reading data in trace sized bites @@ -34,6 +34,7 @@ static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd) if (ret > 0) ret = fifo_put(fifo, buf, ret); + dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret); return ret; } @@ -62,6 +63,7 @@ static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd, if (t->pdu_len == 0) return 0; + dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len); return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len); } @@ -95,7 +97,8 @@ int is_blktrace(const char *filename) return 0; } -static int lookup_device(char *path, unsigned int maj, unsigned int min) +static int lookup_device(struct thread_data *td, char *path, unsigned int maj, + unsigned int min) { struct dirent *dir; struct stat st; @@ -119,7 +122,7 @@ static int lookup_device(char *path, unsigned int maj, unsigned int min) } if (S_ISDIR(st.st_mode)) { - found = lookup_device(full_path, maj, min); + found = lookup_device(td, full_path, maj, min); if (found) { strcpy(path, full_path); break; @@ -129,7 +132,22 @@ static int lookup_device(char *path, unsigned int maj, unsigned int min) 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", 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; @@ -145,6 +163,18 @@ static int lookup_device(char *path, unsigned int maj, unsigned int min) #define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS)) #define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK)) +static void trace_add_open_event(struct thread_data *td, int fileno) +{ + struct io_piece *ipo; + + ipo = calloc(1, sizeof(*ipo)); + + ipo->ddir = DDIR_INVAL; + ipo->fileno = fileno; + ipo->file_action = FIO_LOG_OPEN_FILE; + flist_add_tail(&ipo->list, &td->io_log_list); +} + static void trace_add_file(struct thread_data *td, __u32 device) { static unsigned int last_maj, last_min; @@ -168,8 +198,13 @@ static void trace_add_file(struct thread_data *td, __u32 device) return; strcpy(dev, "/dev"); - if (lookup_device(dev, maj, min)) - add_file(td, dev); + if (lookup_device(td, dev, maj, min)) { + int fileno; + + dprint(FD_BLKTRACE, "add devices %s\n", dev); + fileno = add_file_exclusive(td, dev); + trace_add_open_event(td, fileno); + } } /* @@ -181,7 +216,7 @@ static void store_ipo(struct thread_data *td, unsigned long long offset, struct io_piece *ipo = malloc(sizeof(*ipo)); memset(ipo, 0, sizeof(*ipo)); - INIT_LIST_HEAD(&ipo->list); + INIT_FLIST_HEAD(&ipo->list); /* * the 512 is wrong here, it should be the hardware sector size... */ @@ -193,25 +228,61 @@ static void store_ipo(struct thread_data *td, unsigned long long offset, else ipo->ddir = DDIR_READ; - list_add_tail(&ipo->list, &td->io_log_list); + dprint(FD_BLKTRACE, "store ddir=%d, off=%llu, len=%lu, delay=%lu\n", + ipo->ddir, ipo->offset, + ipo->len, ipo->delay); + queue_io_piece(td, ipo); } -/* - * We only care for queue traces, most of the others are side effects - * 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) +static void handle_trace_notify(struct blk_io_trace *t) { - int rw; + switch (t->action) { + case BLK_TN_PROCESS: + printf("got process notify: %x, %d\n", t->action, t->pid); + break; + case BLK_TN_TIMESTAMP: + printf("got timestamp notify: %x, %d\n", t->action, t->pid); + break; + case BLK_TN_MESSAGE: + break; + default: + dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action); + break; + } +} - 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)) - return; +static void handle_trace_discard(struct thread_data *td, struct blk_io_trace *t, + unsigned long long ttime, unsigned long *ios) +{ + struct io_piece *ipo = malloc(sizeof(*ipo)); + + trace_add_file(td, t->device); + + ios[DDIR_WRITE]++; + td->o.size += t->bytes; + + memset(ipo, 0, sizeof(*ipo)); + INIT_FLIST_HEAD(&ipo->list); + + /* + * the 512 is wrong here, it should be the hardware sector size... + */ + ipo->offset = t->sector * 512; + ipo->len = t->bytes; + ipo->delay = ttime / 1000; + ipo->ddir = DDIR_TRIM; + + dprint(FD_BLKTRACE, "store discard, off=%llu, len=%lu, delay=%lu\n", + ipo->offset, ipo->len, + ipo->delay); + queue_io_piece(td, ipo); +} + +static void handle_trace_fs(struct thread_data *td, struct blk_io_trace *t, + unsigned long long ttime, unsigned long *ios, + unsigned int *bs) +{ + int rw; trace_add_file(td, t->device); @@ -225,6 +296,27 @@ static void handle_trace(struct thread_data *td, struct blk_io_trace *t, store_ipo(td, t->sector, t->bytes, rw, ttime); } +/* + * We only care for queue traces, most of the others are side effects + * 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) +{ + 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)) + handle_trace_notify(t); + else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD)) + handle_trace_discard(td, t, ttime, ios); + else + handle_trace_fs(td, t, ttime, ios, bs); +} + /* * 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. @@ -267,11 +359,13 @@ int load_blktrace(struct thread_data *td, const char *filename) } if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) { - log_err("fio: bad magic in blktrace data: %x\n", t.magic); + log_err("fio: bad magic in blktrace data: %x\n", + t.magic); goto err; } if ((t.magic & 0xff) != BLK_IO_TRACE_VERSION) { - log_err("fio: bad blktrace version %d\n", t.magic & 0xff); + log_err("fio: bad blktrace version %d\n", + t.magic & 0xff); goto err; } ret = discard_pdu(td, fifo, fd, &t); @@ -282,28 +376,42 @@ int load_blktrace(struct thread_data *td, const char *filename) log_err("fio: discarded %d of %d\n", ret, t.pdu_len); goto err; } - if (t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) - continue; - if (!ttime) { + if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 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) + 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); + } + 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) - skipped_writes++; - else + } else { + delay = 0; handle_trace(td, &t, delay, ios, rw_bs); - ttime = t.time; - cpu = t.cpu; + } } while (1); fifo_free(fifo); close(fd); if (skipped_writes) - log_err("fio: <%s> skips replay of %lu writes due to read-only\n", td->o.name, skipped_writes); + log_err("fio: %s skips replay of %lu writes due to read-only\n", + td->o.name, skipped_writes); if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) { log_err("fio: found no ios in blktrace data\n");