The mirror job makes a semi-inaccurate record of the last time we yielded by recording the last time we left a "pause", but this doesn't always correlate to the time we actually last successfully ceded control.
Record the time we last *exited* a yield centrally. In other words, record the time we began execution of this job to know how long we have been selfish for. Signed-off-by: John Snow <js...@redhat.com> --- block/mirror.c | 8 ++------ blockjob.c | 2 ++ include/block/blockjob.h | 5 +++++ 3 files changed, 9 insertions(+), 6 deletions(-) diff --git a/block/mirror.c b/block/mirror.c index c9badc1203..88f4e8964d 100644 --- a/block/mirror.c +++ b/block/mirror.c @@ -63,7 +63,6 @@ typedef struct MirrorBlockJob { QSIMPLEQ_HEAD(, MirrorBuffer) buf_free; int buf_free_count; - uint64_t last_pause_ns; unsigned long *in_flight_bitmap; int in_flight; int64_t bytes_in_flight; @@ -596,8 +595,7 @@ static void mirror_throttle(MirrorBlockJob *s) { int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); - if (now - s->last_pause_ns > SLICE_TIME) { - s->last_pause_ns = now; + if (now - s->common.last_enter_ns > SLICE_TIME) { block_job_sleep_ns(&s->common, 0); } else { block_job_pause_point(&s->common); @@ -769,7 +767,6 @@ static void coroutine_fn mirror_run(void *opaque) mirror_free_init(s); - s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); if (!s->is_none_mode) { ret = mirror_dirty_init(s); if (ret < 0 || block_job_is_cancelled(&s->common)) { @@ -803,7 +800,7 @@ static void coroutine_fn mirror_run(void *opaque) * We do so every SLICE_TIME nanoseconds, or when there is an error, * or when the source is clean, whichever comes first. */ - delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->last_pause_ns; + delta = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - s->common.last_enter_ns; if (delta < SLICE_TIME && s->common.iostatus == BLOCK_DEVICE_IO_STATUS_OK) { if (s->in_flight >= MAX_IN_FLIGHT || s->buf_free_count == 0 || @@ -878,7 +875,6 @@ static void coroutine_fn mirror_run(void *opaque) delay_ns = (s->in_flight == 0 && cnt == 0 ? SLICE_TIME : 0); block_job_sleep_ns(&s->common, delay_ns); } - s->last_pause_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); } immediate_exit: diff --git a/blockjob.c b/blockjob.c index f5cea84e73..2a9ff66b95 100644 --- a/blockjob.c +++ b/blockjob.c @@ -321,6 +321,7 @@ void block_job_start(BlockJob *job) job->pause_count--; job->busy = true; job->paused = false; + job->last_enter_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); bdrv_coroutine_enter(blk_bs(job->blk), job->co); } @@ -786,6 +787,7 @@ static void block_job_do_yield(BlockJob *job, uint64_t ns) job->busy = false; block_job_unlock(); qemu_coroutine_yield(); + job->last_enter_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); /* Set by block_job_enter before re-entering the coroutine. */ assert(job->busy); diff --git a/include/block/blockjob.h b/include/block/blockjob.h index 00403d9482..e965845c94 100644 --- a/include/block/blockjob.h +++ b/include/block/blockjob.h @@ -141,6 +141,11 @@ typedef struct BlockJob { */ QEMUTimer sleep_timer; + /** + * Timestamp of the last yield + */ + uint64_t last_enter_ns; + /** Non-NULL if this job is part of a transaction */ BlockJobTxn *txn; QLIST_ENTRY(BlockJob) txn_list; -- 2.14.3