search for: stats_trim

..., st->usecs)); + + fprintf (fp, "\n"); + } } static inline void @@ -117,6 +124,7 @@ print_stats (int64_t usecs) print_stat (&zero_st); print_stat (&extents_st); print_stat (&cache_st); + print_stat (&flush_st); fflush (fp); } @@ -263,6 +271,21 @@ stats_trim (struct nbdkit_next_ops *next_ops, void *nxdata, return r; } +/* Flush. */ +static int +stats_flush (struct nbdkit_next_ops *next_ops, void *nxdata, + void *handle, uint32_t flags, + int *err) +{ + struct timeval start; + int r; + + gettimeofday (&start, NULL);...

- Use more friendly output with GiB and MiB/s - Measure time per operation, providing finer grain stats - Add total stats for understanding system throughput - Add missing stats for flush I hope that these changes will help to understand and improve virt-v2v performance. Changes since v1: - Keep bytes values - Increase precision to 0.001 GiB and 0.001 MiB/s - Add total stats - Show time before

- Use more friendly output with GiB and MiB/s. - Measure time per operation, providing finer grain stats - Add missing stats for flush I hope that these changes will help to understand and imporve virt-v2v performance. Nir Soffer (3): filters: stats: Show size in GiB, rate in MiB/s filters: stats: Measure time per operation filters: stats: Add flush stats filters/stats/stats.c | 117

...{ + struct timeval start; int r; + gettimeofday (&start, NULL); r = next_ops->pwrite (nxdata, buf, count, offset, flags, err); - if (r == 0) record_stat (&pwrite_st, count); + if (r == 0) record_stat (&pwrite_st, count, &start); return r; } @@ -228,10 +254,12 @@ stats_trim (struct nbdkit_next_ops *next_ops, void *nxdata, uint32_t count, uint64_t offset, uint32_t flags, int *err) { + struct timeval start; int r; + gettimeofday (&start, NULL); r = next_ops->trim (nxdata, count, offset, flags, err); - if (r == 0) record_stat...

..., count, offset, flags, err); if (r == 0) { + gettimeofday (&end, NULL); + usecs = tvdiff_usec(&start, &end); + ACQUIRE_LOCK_FOR_CURRENT_SCOPE (&lock); pwrite_ops++; pwrite_bytes += count; + pwrite_usecs += usecs; } return r; } @@ -222,13 +236,20 @@ stats_trim (struct nbdkit_next_ops *next_ops, void *nxdata, uint32_t count, uint64_t offset, uint32_t flags, int *err) { + struct timeval start, end; + int64_t usecs; int r; + gettimeofday (&start, NULL); r = next_ops->trim (nxdata, count, offset, flags, err); i...

...r = next_ops->pwrite (nxdata, buf, count, offset, flags, err); if (r == 0) { - pthread_mutex_lock (&lock); + ACQUIRE_LOCK_FOR_CURRENT_SCOPE (&lock); pwrite_ops++; pwrite_bytes += count; - pthread_mutex_unlock (&lock); } return r; } @@ -201,10 +198,9 @@ stats_trim (struct nbdkit_next_ops *next_ops, void *nxdata, r = next_ops->trim (nxdata, count, offset, flags, err); if (r == 0) { - pthread_mutex_lock (&lock); + ACQUIRE_LOCK_FOR_CURRENT_SCOPE (&lock); trim_ops++; trim_bytes += count; - pthread_mutex_unlock (&lock);...

I do have a question about whether patch 2 is right, or whether I've exposed a bigger problem in the truncate (and possibly other) filter, but the rest seem fairly straightforward. Eric Blake (4): server: Check for pthread lock failures truncate: Factor out reading real_size under mutex plugins: Check for mutex failures filters: Check for mutex failures filters/cache/cache.c

I like this change. I think we were overloading the config_complete method before to do two different things (complete configuration; do any allocation/housekeeping necessary before we can start serving). The only questions in my mind are whether we want this before 1.18, and whether the name ("get_ready") is a good one. Rich.

Since v1: - rework .can_cache to be tri-state, with default of no advertisement (ripple effect through other patches) - add a lot more patches in order to round out filter support And in the meantime, Rich pushed NBD_CMD_CACHE support into libnbd, so in theory we now have a way to test cache commands through the entire stack. Eric Blake (24): server: Internal hooks for implementing