mirror_throttle attempts to make sure we yield every so often when we're
doing operations that may not otherwise be as cooperative as we'd like.
This pattern is useful to other jobs like commit as well; if commit
continuously decides not to copy data (and calculates delay_ns to be 0),
we'll frequently and aggressively yield, only to be rescheduled immediately.
This causes those "WARNING: I/O thread spun for 1000 iterations"
warnings that everyone loves so much.
Split out the mirror_throttle function to become a new internal
BlockJob API function, block_job_throttle; then use it for all
the other jobs in their runtime loops.
I decided to use it in stream and backup as well, so that regardless of if the
loop structure has the chance to be greedy or not (I did not audit this), the
BlockJob has some kind of failsafe that will occasionally perform a 0ns sleep
every SLICE_TIME.
________________________________________________________________________________
For convenience, this branch is available at:
https://github.com/jnsnow/qemu.git branch block-job-yield
https://github.com/jnsnow/qemu/tree/block-job-yield
This version is tagged block-job-yield-v1:
https://github.com/jnsnow/qemu/releases/tag/block-job-yield-v1
John Snow (7):
blockjob: record time of last yield
blockjob: consolidate SLICE_TIME definition
blockjob: create block_job_throttle
blockjob: allow block_job_throttle to take delay_ns
block/commit: use block_job_throttle
block/stream: use block_job_throttle
block/backup: use block_job_throttle
block/backup.c | 12 ++++++------
block/commit.c | 5 ++---
block/mirror.c | 22 +++-------------------
block/stream.c | 4 +---
blockjob.c | 12 ++++++++++++
include/block/blockjob.h | 5 +++++
include/block/blockjob_int.h | 15 +++++++++++++++
7 files changed, 44 insertions(+), 31 deletions(-)
--
2.14.3