*/
#include <stdio.h>
#include <stdlib.h>
+#include <sys/stat.h>
+#include <dirent.h>
-#include "list.h"
+#include "flist.h"
#include "fio.h"
#include "blktrace_api.h"
-#define TRACE_FIFO_SIZE (sizeof(struct blk_io_trace) * 1000)
+#define TRACE_FIFO_SIZE 65536
/*
* fifo refill frontend, to avoid reading data in trace sized bites
static int refill_fifo(struct thread_data *td, struct fifo *fifo, int fd)
{
char buf[TRACE_FIFO_SIZE];
- unsigned int total, left;
- void *ptr;
+ unsigned int total;
int ret;
- total = 0;
- ptr = buf;
- while (total < TRACE_FIFO_SIZE) {
- left = TRACE_FIFO_SIZE - total;
+ total = sizeof(buf);
+ if (total > fifo_room(fifo))
+ total = fifo_room(fifo);
- ret = read(fd, ptr, left);
- if (ret < 0) {
- td_verror(td, errno, "read blktrace file");
- return -1;
- } else if (!ret)
- break;
-
- fifo_put(fifo, ptr, ret);
- ptr += ret;
- total += ret;
+ ret = read(fd, buf, total);
+ if (ret < 0) {
+ td_verror(td, errno, "read blktrace file");
+ return -1;
}
- return 0;
+ if (ret > 0)
+ ret = fifo_put(fifo, buf, ret);
+
+ dprint(FD_BLKTRACE, "refill: filled %d bytes\n", ret);
+ return ret;
}
/*
static int trace_fifo_get(struct thread_data *td, struct fifo *fifo, int fd,
void *buf, unsigned int len)
{
- int ret;
-
- if (fifo_len(fifo) >= len)
- return fifo_get(fifo, buf, len);
-
- ret = refill_fifo(td, fifo, fd);
- if (ret < 0)
- return ret;
+ if (fifo_len(fifo) < len) {
+ int ret = refill_fifo(td, fifo, fd);
- if (fifo_len(fifo) < len)
- return 0;
+ if (ret < 0)
+ return ret;
+ }
return fifo_get(fifo, buf, len);
}
/*
* Just discard the pdu by seeking past it.
*/
-static int discard_pdu(int fd, struct blk_io_trace *t)
+static int discard_pdu(struct thread_data *td, struct fifo *fifo, int fd,
+ struct blk_io_trace *t)
{
if (t->pdu_len == 0)
return 0;
- if (lseek(fd, t->pdu_len, SEEK_CUR) < 0)
- return errno;
-
- return 0;
+ dprint(FD_BLKTRACE, "discard pdu len %u\n", t->pdu_len);
+ return trace_fifo_get(td, fifo, fd, NULL, t->pdu_len);
}
/*
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;
}
+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;
+
+ while ((dir = readdir(D)) != NULL) {
+ char full_path[256];
+
+ if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
+ continue;
+
+ sprintf(full_path, "%s/%s", path, dir->d_name);
+ if (lstat(full_path, &st) == -1) {
+ perror("lstat");
+ break;
+ }
+
+ if (S_ISDIR(st.st_mode)) {
+ found = lookup_device(full_path, maj, min);
+ if (found) {
+ strcpy(path, full_path);
+ break;
+ }
+ }
+
+ if (!S_ISBLK(st.st_mode))
+ continue;
+
+ 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))
+#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;
+ unsigned int maj = FMAJOR(device);
+ unsigned int min = FMINOR(device);
+ struct fio_file *f;
+ char dev[256];
+ unsigned int i;
+
+ if (last_maj == maj && last_min == min)
+ return;
+
+ 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;
+
+ strcpy(dev, "/dev");
+ if (lookup_device(dev, maj, min)) {
+ int fileno;
+
+ dprint(FD_BLKTRACE, "add devices %s\n", dev);
+ fileno = add_file(td, dev);
+ trace_add_open_event(td, fileno);
+ }
+}
+
/*
* Store blk_io_trace data in an ipo for later retrieval.
*/
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...
*/
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);
+}
+
+static void handle_trace_notify( struct blk_io_trace *t)
+{
+ 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;
+ default:
+ dprint(FD_BLKTRACE, "unknown trace act %x\n", t->action);
+ break;
+ }
+}
+
+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);
+
+ rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
+
+ if (t->bytes > bs[rw])
+ bs[rw] = t->bytes;
+
+ ios[rw]++;
+ td->o.size += t->bytes;
+ store_ipo(td, t->sector, t->bytes, rw, ttime);
}
/*
unsigned long long ttime, unsigned long *ios,
unsigned int *bs)
{
- int rw;
-
if ((t->action & 0xffff) != __BLK_TA_QUEUE)
return;
if (t->action & BLK_TC_ACT(BLK_TC_PC))
return;
- /*
- * should not happen, need to look into that...
- */
- if (!t->bytes)
- return;
-
- rw = (t->action & BLK_TC_ACT(BLK_TC_WRITE)) != 0;
-
- if (t->bytes > bs[rw])
- bs[rw] = t->bytes;
-
- ios[rw]++;
- td->o.size += t->bytes;
- store_ipo(td, t->sector, t->bytes, rw, ttime);
+ if (t->action & BLK_TC_ACT(BLK_TC_NOTIFY))
+ handle_trace_notify(t);
+ else
+ handle_trace_fs(td, t, ttime, ios, bs);
}
/*
{
unsigned long long ttime, delay;
struct blk_io_trace t;
- unsigned long ios[2];
+ unsigned long ios[2], skipped_writes;
unsigned int cpu;
unsigned int rw_bs[2];
struct fifo *fifo;
ttime = 0;
ios[0] = ios[1] = 0;
rw_bs[0] = rw_bs[1] = 0;
+ skipped_writes = 0;
do {
int ret = trace_fifo_get(td, fifo, fd, &t, sizeof(t));
}
if ((t.magic & 0xffffff00) != BLK_IO_TRACE_MAGIC) {
- log_err("fio: bad magic in blktrace data\n");
+ 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(fd, &t);
- if (ret) {
+ ret = discard_pdu(td, fifo, fd, &t);
+ if (ret < 0) {
td_verror(td, ret, "blktrace lseek");
goto err;
+ } else if (t.pdu_len != ret) {
+ log_err("fio: discarded %d of %d\n", ret, t.pdu_len);
+ goto err;
}
- 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
+ handle_trace(td, &t, delay, ios, rw_bs);
+
ttime = t.time;
cpu = t.cpu;
+ } else {
+ delay = 0;
+ handle_trace(td, &t, delay, ios, rw_bs);
}
- delay = 0;
- if (cpu == t.cpu)
- delay = t.time - ttime;
- 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);
+
if (!ios[DDIR_READ] && !ios[DDIR_WRITE]) {
log_err("fio: found no ios in blktrace data\n");
return 1;