| Message ID | 20180119205847.7141-12-jsnow@redhat.com |
|---|---|
| State | New |
| Headers | show |
| Series | blockjob: refactor mirror_throttle | expand |
On 2018-01-19 21:58, John Snow wrote: > We're attempting to slacken the mirror loop in three different places, > but we can combine these three attempts. Combine the early loop call to > block_job_pause_point with the two late-loop calls to block_job_sleep_ns. > > When delay_ns is 0 and it has not been SLICE_TIME since the last yield, > block_job_relax is merely a call to block_job_pause_point, so this should > be equivalent with the exception that if we have managed to not yield at > all in the last SLICE_TIME ns, we will now do so. > > I am not sure that condition was possible, > so this loop should be equivalent. Well, to me it even sounds like a positive change if it was a change. We want the job to yield after SLICE_TIME ns, after all, and I don't think it matters where that happens, exactly. > > Signed-off-by: John Snow <jsnow@redhat.com> > --- > block/mirror.c | 22 +++++++++++----------- > block/trace-events | 2 +- > 2 files changed, 12 insertions(+), 12 deletions(-) > > diff --git a/block/mirror.c b/block/mirror.c > index a0e0044de2..192e03694f 100644 > --- a/block/mirror.c > +++ b/block/mirror.c > @@ -761,7 +761,7 @@ static void coroutine_fn mirror_run(void *opaque) > assert(!s->dbi); > s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap); > for (;;) { > - uint64_t delay_ns = 0; > + static uint64_t delay_ns = 0; Errr. Are you sure about that? Now every mirror job in the qeny process will share this single variable. Was that your intention? > int64_t cnt, delta; > bool should_complete; > > @@ -770,9 +770,16 @@ static void coroutine_fn mirror_run(void *opaque) > goto immediate_exit; > } > > - block_job_pause_point(&s->common); > - > cnt = bdrv_get_dirty_count(s->dirty_bitmap); > + > + trace_mirror_before_relax(s, cnt, s->synced, delay_ns); > + if (block_job_relax(&s->common, delay_ns)) { See the reply to <a href="HEAD^^">that patch</a>. > + if (!s->synced) { > + goto immediate_exit; > + } > + } > + delay_ns = 0; > + > /* s->common.offset contains the number of bytes already processed so > * far, cnt is the number of dirty bytes remaining and > * s->bytes_in_flight is the number of bytes currently being > @@ -849,15 +856,8 @@ static void coroutine_fn mirror_run(void *opaque) > } > > ret = 0; > - trace_mirror_before_sleep(s, cnt, s->synced, delay_ns); > - if (!s->synced) { > - block_job_sleep_ns(&s->common, delay_ns); > - if (block_job_is_cancelled(&s->common)) { > - break; > - } > - } else if (!should_complete) { > + if (s->synced && !should_complete) { > delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0); > - block_job_sleep_ns(&s->common, delay_ns); > } > } Basic idea looks good to me (apart from the static delay_ns), but, well, block-job-cancel on a busy job still breaks. Max
On 2018-02-07 23:46, Max Reitz wrote: > On 2018-01-19 21:58, John Snow wrote: >> We're attempting to slacken the mirror loop in three different places, >> but we can combine these three attempts. Combine the early loop call to >> block_job_pause_point with the two late-loop calls to block_job_sleep_ns. >> >> When delay_ns is 0 and it has not been SLICE_TIME since the last yield, >> block_job_relax is merely a call to block_job_pause_point, so this should >> be equivalent with the exception that if we have managed to not yield at >> all in the last SLICE_TIME ns, we will now do so. >> >> I am not sure that condition was possible, >> so this loop should be equivalent. > > Well, to me it even sounds like a positive change if it was a change. > We want the job to yield after SLICE_TIME ns, after all, and I don't > think it matters where that happens, exactly. > >> >> Signed-off-by: John Snow <jsnow@redhat.com> >> --- >> block/mirror.c | 22 +++++++++++----------- >> block/trace-events | 2 +- >> 2 files changed, 12 insertions(+), 12 deletions(-) >> >> diff --git a/block/mirror.c b/block/mirror.c >> index a0e0044de2..192e03694f 100644 >> --- a/block/mirror.c >> +++ b/block/mirror.c >> @@ -761,7 +761,7 @@ static void coroutine_fn mirror_run(void *opaque) >> assert(!s->dbi); >> s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap); >> for (;;) { >> - uint64_t delay_ns = 0; >> + static uint64_t delay_ns = 0; > > Errr. Are you sure about that? > > Now every mirror job in the qeny process will share this single > variable. Was that your intention? ("Errr" @myself for "qeny")
On 02/07/2018 05:46 PM, Max Reitz wrote: > Errr. Are you sure about that? > > Now every mirror job in the qeny process will share this single > variable. Was that your intention? > Duh, no. Thanks for the review, also. I'm currently sidetracked on something else, so I'll get back to this soon.
diff --git a/block/mirror.c b/block/mirror.c index a0e0044de2..192e03694f 100644 --- a/block/mirror.c +++ b/block/mirror.c @@ -761,7 +761,7 @@ static void coroutine_fn mirror_run(void *opaque) assert(!s->dbi); s->dbi = bdrv_dirty_iter_new(s->dirty_bitmap); for (;;) { - uint64_t delay_ns = 0; + static uint64_t delay_ns = 0; int64_t cnt, delta; bool should_complete; @@ -770,9 +770,16 @@ static void coroutine_fn mirror_run(void *opaque) goto immediate_exit; } - block_job_pause_point(&s->common); - cnt = bdrv_get_dirty_count(s->dirty_bitmap); + + trace_mirror_before_relax(s, cnt, s->synced, delay_ns); + if (block_job_relax(&s->common, delay_ns)) { + if (!s->synced) { + goto immediate_exit; + } + } + delay_ns = 0; + /* s->common.offset contains the number of bytes already processed so * far, cnt is the number of dirty bytes remaining and * s->bytes_in_flight is the number of bytes currently being @@ -849,15 +856,8 @@ static void coroutine_fn mirror_run(void *opaque) } ret = 0; - trace_mirror_before_sleep(s, cnt, s->synced, delay_ns); - if (!s->synced) { - block_job_sleep_ns(&s->common, delay_ns); - if (block_job_is_cancelled(&s->common)) { - break; - } - } else if (!should_complete) { + if (s->synced && !should_complete) { delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0); - block_job_sleep_ns(&s->common, delay_ns); } } diff --git a/block/trace-events b/block/trace-events index 11c8d5f590..60f36f929a 100644 --- a/block/trace-events +++ b/block/trace-events @@ -27,7 +27,7 @@ mirror_start(void *bs, void *s, void *opaque) "bs %p s %p opaque %p" mirror_restart_iter(void *s, int64_t cnt) "s %p dirty count %"PRId64 mirror_before_flush(void *s) "s %p" mirror_before_drain(void *s, int64_t cnt) "s %p dirty count %"PRId64 -mirror_before_sleep(void *s, int64_t cnt, int synced, uint64_t delay_ns) "s %p dirty count %"PRId64" synced %d delay %"PRIu64"ns" +mirror_before_relax(void *s, int64_t cnt, int synced, uint64_t delay_ns) "s %p dirty count %"PRId64" synced %d delay %"PRIu64"ns" mirror_one_iteration(void *s, int64_t offset, uint64_t bytes) "s %p offset %" PRId64 " bytes %" PRIu64 mirror_iteration_done(void *s, int64_t offset, uint64_t bytes, int ret) "s %p offset %" PRId64 " bytes %" PRIu64 " ret %d" mirror_yield(void *s, int64_t cnt, int buf_free_count, int in_flight) "s %p dirty count %"PRId64" free buffers %d in_flight %d"
We're attempting to slacken the mirror loop in three different places, but we can combine these three attempts. Combine the early loop call to block_job_pause_point with the two late-loop calls to block_job_sleep_ns. When delay_ns is 0 and it has not been SLICE_TIME since the last yield, block_job_relax is merely a call to block_job_pause_point, so this should be equivalent with the exception that if we have managed to not yield at all in the last SLICE_TIME ns, we will now do so. I am not sure that condition was possible, so this loop should be equivalent. Signed-off-by: John Snow <jsnow@redhat.com> --- block/mirror.c | 22 +++++++++++----------- block/trace-events | 2 +- 2 files changed, 12 insertions(+), 12 deletions(-)