Make experimental_verify=1 handle all cases properly
- Don't track written bytes, just replay the workload by resetting
all the random generators. This should work for any mixture of IO.
- Handle trims for verify.
- Ensure that rwmix is replayed properly for verify.
- Fixup logging for replay.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
diff --git a/backend.c b/backend.c
index 507faa9..902414e 100644
--- a/backend.c
+++ b/backend.c
@@ -393,12 +393,11 @@
* The main verify engine. Runs over the writes we previously submitted,
* reads the blocks back in, and checks the crc/md5 of the data.
*/
-static void do_verify(struct thread_data *td, uint64_t verify_bytes)
+static void do_verify(struct thread_data *td)
{
struct fio_file *f;
struct io_u *io_u;
int ret, min_events;
- uint64_t io_bytes;
unsigned int i;
dprint(FD_VERIFY, "starting loop\n");
@@ -422,7 +421,6 @@
td_set_runstate(td, TD_VERIFYING);
io_u = NULL;
- io_bytes = 0;
while (!td->terminate) {
enum fio_ddir ddir;
int ret2, full;
@@ -455,11 +453,34 @@
break;
}
} else {
- io_u = get_io_u(td);
- if (!io_u)
- break;
+ while ((io_u = get_io_u(td)) != NULL) {
+ /*
+ * We are only interested in the places where
+ * we wrote or trimmed IOs. Turn those into
+ * reads for verification purposes.
+ */
+ if (io_u->ddir == DDIR_READ) {
+ /*
+ * Pretend we issued it for rwmix
+ * accounting
+ */
+ td->io_issues[DDIR_READ]++;
+ put_io_u(td, io_u);
+ continue;
+ } else if (io_u->ddir == DDIR_TRIM) {
+ io_u->ddir = DDIR_READ;
+ io_u->flags |= IO_U_F_TRIMMED;
+ break;
+ } else if (io_u->ddir == DDIR_WRITE) {
+ io_u->ddir = DDIR_READ;
+ break;
+ } else {
+ put_io_u(td, io_u);
+ continue;
+ }
+ }
- if (io_u->buflen + io_bytes > verify_bytes)
+ if (!io_u)
break;
}
@@ -491,7 +512,6 @@
io_u->xfer_buflen = io_u->resid;
io_u->xfer_buf += bytes;
io_u->offset += bytes;
- io_bytes += bytes;
if (ddir_rw(io_u->ddir))
td->ts.short_io_u[io_u->ddir]++;
@@ -507,7 +527,6 @@
if (ret < 0)
break;
}
- io_bytes += io_u->xfer_buflen;
continue;
case FIO_Q_QUEUED:
break;
@@ -542,18 +561,15 @@
min_events = 1;
do {
- unsigned long bytes = 0;
-
/*
* Reap required number of io units, if any,
* and do the verification on them through
* the callback handler
*/
- if (io_u_queued_complete(td, min_events, &bytes) < 0) {
+ if (io_u_queued_complete(td, min_events, NULL) < 0) {
ret = -1;
break;
}
- io_bytes += bytes;
} while (full && (td->cur_depth > td->o.iodepth_low));
}
if (ret < 0)
@@ -1190,8 +1206,6 @@
clear_state = 0;
while (keep_running(td)) {
- uint64_t write_bytes;
-
fio_gettime(&td->start, NULL);
memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
@@ -1212,9 +1226,7 @@
prune_io_piece_log(td);
- write_bytes = td->io_bytes[DDIR_WRITE];
do_io(td);
- write_bytes = td->io_bytes[DDIR_WRITE] - write_bytes;
clear_state = 1;
@@ -1243,7 +1255,7 @@
fio_gettime(&td->start, NULL);
- do_verify(td, write_bytes);
+ do_verify(td);
td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
diff --git a/io_u.c b/io_u.c
index f020cac..8567e11 100644
--- a/io_u.c
+++ b/io_u.c
@@ -172,19 +172,31 @@
return 1;
}
+/*
+ * Sort the reads for a verify phase in batches of verifysort_nr, if
+ * specified.
+ */
+static inline int should_sort_io(struct thread_data *td)
+{
+ if (!td->o.verifysort_nr || !td->o.do_verify)
+ return 0;
+ if (!td_random(td))
+ return 0;
+ if (td->runstate != TD_VERIFYING)
+ return 0;
+ if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
+ return 0;
+
+ return 1;
+}
+
static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
enum fio_ddir ddir, uint64_t *b)
{
struct rand_off *r;
int i, ret = 1;
- /*
- * If sort not enabled, or not a pure random read workload without
- * any stored write metadata, just return a random offset
- */
- if (!td->o.verifysort_nr || !(ddir == DDIR_READ && td->o.do_verify &&
- td->o.verify != VERIFY_NONE && td_random(td)) ||
- td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
+ if (!should_sort_io(td))
return get_off_from_method(td, f, ddir, b);
if (!flist_empty(&td->next_rand_list)) {
@@ -546,12 +558,6 @@
enum fio_ddir ddir;
/*
- * If verify phase started, it's always a READ
- */
- if (td->runstate == TD_VERIFYING)
- return DDIR_READ;
-
- /*
* see if it's time to fsync
*/
if (td->o.fsync_blocks &&
@@ -606,7 +612,7 @@
static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
{
- io_u->ddir = get_rw_ddir(td);
+ io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
td->o.barrier_blocks &&
@@ -650,14 +656,15 @@
void requeue_io_u(struct thread_data *td, struct io_u **io_u)
{
struct io_u *__io_u = *io_u;
+ enum fio_ddir ddir = acct_ddir(__io_u);
dprint(FD_IO, "requeue %p\n", __io_u);
td_io_u_lock(td);
__io_u->flags |= IO_U_F_FREE;
- if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
- td->io_issues[__io_u->ddir]--;
+ if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
+ td->io_issues[ddir]--;
__io_u->flags &= ~IO_U_F_FLIGHT;
if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
@@ -719,13 +726,9 @@
if (td_random(td) && file_randommap(td, io_u->file))
mark_random_map(td, io_u);
- /*
- * If using a write iolog, store this entry.
- */
out:
dprint_io_u(io_u, "fill_io_u");
td->zone_bytes += io_u->buflen;
- log_io_u(td, io_u);
return 0;
}
@@ -1091,6 +1094,7 @@
io_u->flags &= ~IO_U_F_VER_LIST;
io_u->error = 0;
+ io_u->acct_ddir = -1;
flist_del(&io_u->list);
flist_add_tail(&io_u->list, &td->io_u_busylist);
td->cur_depth++;
diff --git a/ioengine.h b/ioengine.h
index 6809501..d5a0dc9 100644
--- a/ioengine.h
+++ b/ioengine.h
@@ -8,7 +8,7 @@
#include <guasi.h>
#endif
-#define FIO_IOOPS_VERSION 14
+#define FIO_IOOPS_VERSION 15
enum {
IO_U_F_FREE = 1 << 0,
@@ -57,6 +57,12 @@
enum fio_ddir ddir;
/*
+ * For replay workloads, we may want to account as a different
+ * IO type than what is being submitted.
+ */
+ enum fio_ddir acct_ddir;
+
+ /*
* Allocated/set buffer and length
*/
unsigned long buflen;
@@ -210,4 +216,12 @@
#define dprint_io_u(io_u, p)
#endif
+static inline enum fio_ddir acct_ddir(struct io_u *io_u)
+{
+ if (io_u->acct_ddir != -1)
+ return io_u->acct_ddir;
+
+ return io_u->ddir;
+}
+
#endif
diff --git a/ioengines.c b/ioengines.c
index f81c46f..234f8ed 100644
--- a/ioengines.c
+++ b/ioengines.c
@@ -260,6 +260,11 @@
assert(fio_file_open(io_u->file));
+ /*
+ * If using a write iolog, store this entry.
+ */
+ log_io_u(td, io_u);
+
io_u->error = 0;
io_u->resid = 0;
@@ -275,8 +280,8 @@
sizeof(struct timeval));
}
- if (ddir_rw(io_u->ddir))
- td->io_issues[io_u->ddir]++;
+ if (ddir_rw(acct_ddir(io_u)))
+ td->io_issues[acct_ddir(io_u)]++;
ret = td->io_ops->queue(td, io_u);
diff --git a/libfio.c b/libfio.c
index 8255072..ac629dc 100644
--- a/libfio.c
+++ b/libfio.c
@@ -67,6 +67,7 @@
static void reset_io_counters(struct thread_data *td)
{
int ddir;
+
for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) {
td->stat_io_bytes[ddir] = 0;
td->this_io_bytes[ddir] = 0;
@@ -74,10 +75,12 @@
td->this_io_blocks[ddir] = 0;
td->rate_bytes[ddir] = 0;
td->rate_blocks[ddir] = 0;
+ td->io_issues[ddir] = 0;
}
td->zone_bytes = 0;
td->last_was_sync = 0;
+ td->rwmix_issues = 0;
/*
* reset file done count if we are to start over