#include <stdio.h>
#include <stdlib.h>
#include <sys/stat.h>
+#include <sys/ioctl.h>
+#include <linux/fs.h>
#include <dirent.h>
-#include "list.h"
+#include "flist.h"
#include "fio.h"
+#include "blktrace.h"
#include "blktrace_api.h"
+#include "oslib/linux-dev-lookup.h"
-#define TRACE_FIFO_SIZE 65536
+#define TRACE_FIFO_SIZE 8192
/*
* fifo refill frontend, to avoid reading data in trace sized bites
if (ret > 0)
ret = fifo_put(fifo, buf, ret);
+ dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
return ret;
}
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);
}
* Check if this is a blktrace binary data file. We read a single trace
* into memory and check for the magic signature.
*/
-int is_blktrace(const char *filename)
+int is_blktrace(const char *filename, int *need_swap)
{
struct blk_io_trace t;
int fd, ret;
fd = open(filename, O_RDONLY);
- if (fd < 0) {
- perror("open blktrace");
+ if (fd < 0)
return 0;
- }
ret = read(fd, &t, sizeof(t));
close(fd);
return 0;
}
- if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC)
+ if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
+ *need_swap = 0;
return 1;
+ }
+
+ /*
+ * Maybe it needs to be endian swapped...
+ */
+ t.magic = fio_swap32(t.magic);
+ if ((t.magic & 0xffffff00) == BLK_IO_TRACE_MAGIC) {
+ *need_swap = 1;
+ return 1;
+ }
return 0;
}
-static int lookup_device(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;
+#define FMINORBITS 20
+#define FMINORMASK ((1U << FMINORBITS) - 1)
+#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
+#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
- while ((dir = readdir(D)) != NULL) {
- char full_path[256];
+static void trace_add_open_close_event(struct thread_data *td, int fileno, enum file_log_act action)
+{
+ struct io_piece *ipo;
- if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
- continue;
+ ipo = calloc(1, sizeof(*ipo));
+ init_ipo(ipo);
- sprintf(full_path, "%s/%s", path, dir->d_name);
- if (lstat(full_path, &st) == -1) {
- perror("lstat");
- break;
- }
+ ipo->ddir = DDIR_INVAL;
+ ipo->fileno = fileno;
+ ipo->file_action = action;
+ flist_add_tail(&ipo->list, &td->io_log_list);
+}
- if (S_ISDIR(st.st_mode)) {
- found = lookup_device(full_path, maj, min);
- if (found) {
- strcpy(path, full_path);
- break;
- }
- }
+static int get_dev_blocksize(const char *dev, unsigned int *bs)
+{
+ int fd;
- if (!S_ISBLK(st.st_mode))
- continue;
+ fd = open(dev, O_RDONLY);
+ if (fd < 0)
+ return 1;
- if (maj == major(st.st_rdev) && min == minor(st.st_rdev)) {
- strcpy(path, full_path);
- found = 1;
- break;
- }
+ if (ioctl(fd, BLKSSZGET, bs) < 0) {
+ close(fd);
+ return 1;
}
- closedir(D);
- return found;
+ close(fd);
+ return 0;
}
-#define FMINORBITS 20
-#define FMINORMASK ((1U << FMINORBITS) - 1)
-#define FMAJOR(dev) ((unsigned int) ((dev) >> FMINORBITS))
-#define FMINOR(dev) ((unsigned int) ((dev) & FMINORMASK))
-
-static void trace_add_file(struct thread_data *td, __u32 device)
+static int trace_add_file(struct thread_data *td, __u32 device,
+ unsigned int *bs)
{
- static unsigned int last_maj, last_min;
+ static unsigned int last_maj, last_min, last_fileno, last_bs;
unsigned int maj = FMAJOR(device);
unsigned int min = FMINOR(device);
struct fio_file *f;
- char dev[256];
unsigned int i;
+ char dev[256];
- if (last_maj == maj && last_min == min)
- return;
+ if (last_maj == maj && last_min == min) {
+ *bs = last_bs;
+ return last_fileno;
+ }
last_maj = maj;
last_min = min;
/*
* check for this file in our list
*/
- for_each_file(td, f, i)
- if (f->major == maj && f->minor == min)
- return;
+ for_each_file(td, f, i) {
+ if (f->major == maj && f->minor == min) {
+ last_fileno = f->fileno;
+ last_bs = f->bs;
+ goto out;
+ }
+ }
strcpy(dev, "/dev");
- if (lookup_device(dev, maj, min))
- add_file(td, dev);
+ if (blktrace_lookup_device(td->o.replay_redirect, dev, maj, min)) {
+ unsigned int this_bs;
+ 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);
+
+ if (get_dev_blocksize(dev, &this_bs))
+ this_bs = 512;
+
+ td->o.open_files++;
+ td->files[fileno]->major = maj;
+ td->files[fileno]->minor = min;
+ td->files[fileno]->bs = this_bs;
+ trace_add_open_close_event(td, fileno, FIO_LOG_OPEN_FILE);
+
+ last_fileno = fileno;
+ last_bs = this_bs;
+ }
+
+out:
+ *bs = last_bs;
+ return last_fileno;
+}
+
+static void t_bytes_align(struct thread_options *o, struct blk_io_trace *t)
+{
+ if (!o->replay_align)
+ return;
+
+ t->bytes = (t->bytes + o->replay_align - 1) & ~(o->replay_align - 1);
}
/*
* Store blk_io_trace data in an ipo for later retrieval.
*/
static void store_ipo(struct thread_data *td, unsigned long long offset,
- unsigned int bytes, int rw, unsigned long long ttime)
+ unsigned int bytes, int rw, unsigned long long ttime,
+ int fileno, unsigned int bs)
{
struct io_piece *ipo = malloc(sizeof(*ipo));
- memset(ipo, 0, sizeof(*ipo));
- INIT_LIST_HEAD(&ipo->list);
- /*
- * the 512 is wrong here, it should be the hardware sector size...
- */
- ipo->offset = offset * 512;
+ init_ipo(ipo);
+
+ ipo->offset = offset * bs;
+ if (td->o.replay_scale)
+ ipo->offset = ipo->offset / td->o.replay_scale;
+ ipo_bytes_align(td->o.replay_align, ipo);
ipo->len = bytes;
ipo->delay = ttime / 1000;
if (rw)
ipo->ddir = DDIR_WRITE;
else
ipo->ddir = DDIR_READ;
+ ipo->fileno = fileno;
+
+ 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);
+}
+
+static void handle_trace_notify(struct blk_io_trace *t)
+{
+ switch (t->action) {
+ case BLK_TN_PROCESS:
+ dprint(FD_BLKTRACE, "got process notify: %x, %d\n",
+ t->action, t->pid);
+ break;
+ case BLK_TN_TIMESTAMP:
+ dprint(FD_BLKTRACE, "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;
+ }
+}
+
+static void handle_trace_discard(struct thread_data *td,
+ struct blk_io_trace *t,
+ unsigned long long ttime,
+ unsigned long *ios, unsigned int *rw_bs)
+{
+ struct io_piece *ipo = malloc(sizeof(*ipo));
+ unsigned int bs;
+ int fileno;
+
+ init_ipo(ipo);
+ fileno = trace_add_file(td, t->device, &bs);
+
+ ios[DDIR_TRIM]++;
+ if (t->bytes > rw_bs[DDIR_TRIM])
+ rw_bs[DDIR_TRIM] = t->bytes;
+
+ td->o.size += t->bytes;
+
+ memset(ipo, 0, sizeof(*ipo));
+ INIT_FLIST_HEAD(&ipo->list);
+
+ ipo->offset = t->sector * bs;
+ if (td->o.replay_scale)
+ ipo->offset = ipo->offset / td->o.replay_scale;
+ ipo_bytes_align(td->o.replay_align, ipo);
+ ipo->len = t->bytes;
+ ipo->delay = ttime / 1000;
+ ipo->ddir = DDIR_TRIM;
+ ipo->fileno = fileno;
- list_add_tail(&ipo->list, &td->io_log_list);
+ 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 *rw_bs)
+{
+ unsigned int bs;
+ int rw;
+ int fileno;
+
+ fileno = trace_add_file(td, t->device, &bs);
+
+ rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
+
+ if (t->bytes > rw_bs[rw])
+ rw_bs[rw] = t->bytes;
+
+ ios[rw]++;
+ td->o.size += t->bytes;
+ store_ipo(td, t->sector, t->bytes, rw, ttime, fileno, bs);
}
/*
* 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)
{
- int rw;
+ static unsigned long long last_ttime;
+ unsigned long long delay = 0;
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;
+ }
+ }
+
+ t_bytes_align(&td->o, t);
+
if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
- return;
+ handle_trace_notify(t);
+ else if (t->action & BLK_TC_ACT(BLK_TC_DISCARD))
+ handle_trace_discard(td, t, delay, ios, bs);
+ else
+ handle_trace_fs(td, t, delay, ios, bs);
+}
- trace_add_file(td, t->device);
+static void byteswap_trace(struct blk_io_trace *t)
+{
+ t->magic = fio_swap32(t->magic);
+ t->sequence = fio_swap32(t->sequence);
+ t->time = fio_swap64(t->time);
+ t->sector = fio_swap64(t->sector);
+ t->bytes = fio_swap32(t->bytes);
+ t->action = fio_swap32(t->action);
+ t->pid = fio_swap32(t->pid);
+ t->device = fio_swap32(t->device);
+ t->cpu = fio_swap32(t->cpu);
+ t->error = fio_swap16(t->error);
+ t->pdu_len = fio_swap16(t->pdu_len);
+}
- rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
+static int t_is_write(struct blk_io_trace *t)
+{
+ return (t->action & BLK_TC_ACT(BLK_TC_WRITE | BLK_TC_DISCARD)) != 0;
+}
- if (t->bytes > bs[rw])
- bs[rw] = t->bytes;
+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;
+}
- ios[rw]++;
- td->o.size += t->bytes;
- store_ipo(td, t->sector, t->bytes, rw, ttime);
+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 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];
- 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;
+ int fd, i, old_state;
+ struct fio_file *f;
+ int this_depth[DDIR_RWDIR_CNT], depth[DDIR_RWDIR_CNT], max_depth;
fd = open(filename, O_RDONLY);
if (fd < 0) {
fifo = fifo_alloc(TRACE_FIFO_SIZE);
+ old_state = td_bump_runstate(td, TD_SETTING_UP);
+
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;
do {
int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
break;
}
+ if (need_swap)
+ byteswap_trace(&t);
+
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);
log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
goto err;
}
- if (!ttime) {
- ttime = t.time;
- cpu = t.cpu;
+ if ((t.action & BLK_TC_ACT(BLK_TC_NOTIFY)) == 0) {
+ if ((t.action & 0xffff) == __BLK_TA_QUEUE)
+ 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++;
+ continue;
+ }
}
- delay = 0;
- if (cpu == t.cpu)
- delay = t.time - ttime;
- handle_trace(td, &t, delay, ios, rw_bs);
- ttime = t.time;
- cpu = t.cpu;
+
+ handle_trace(td, &t, ios, rw_bs);
} while (1);
+ for_each_file(td, f, i)
+ trace_add_open_close_event(td, f->fileno, FIO_LOG_CLOSE_FILE);
+
fifo_free(fifo);
close(fd);
+ td_restore_runstate(td, old_state);
+
+ if (!td->files_index) {
+ log_err("fio: did not find replay device(s)\n");
+ return 1;
+ }
+
+ /*
+ * For stacked devices, we don't always get a COMPLETE event so
+ * the depth grows to insane values. Limit it to something sane(r).
+ */
+ 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",
+ td->o.name, skipped_writes);
+
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]) {
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.
+ */
+ if (!fio_option_is_set(&td->o, odirect) && !(td_min_bs(td) & 4095))
+ td->o.odirect = 1;
+
+ /*
+ * If depth wasn't manually set, use probed depth
*/
- td->o.odirect = 1;
+ if (!fio_option_is_set(&td->o, iodepth))
+ td->o.iodepth = td->o.iodepth_low = max_depth;
return 0;
err:
projects / fio.git / blobdiff