hi,
Thanks for reviewing.
On Mon, Sep 02, 2019 at 10:54:41PM +0800, Xiaoguang Wang wrote:
Sometimes process will be stalled in "wait_on_bit_io(&bh->b_state,
BH_Shadow, TASK_UNINTERRUPTIBLE)" for a while, and in order to analyse
app's latency thoroughly, add a new tracepoint to track this delay.
Trace info likes below:
fsstress-5068 [008] .... 11007.757543: jbd2_sleep_on_shadow: dev 254,17 sleep 1
fsstress-5070 [007] .... 11007.757544: jbd2_sleep_on_shadow: dev 254,17 sleep 2
fsstress-5069 [009] .... 11007.757548: jbd2_sleep_on_shadow: dev 254,17 sleep 2
fsstress-5067 [011] .... 11007.757569: jbd2_sleep_on_shadow: dev 254,17 sleep 1
fsstress-5063 [007] .... 11007.757651: jbd2_sleep_on_shadow: dev 254,17 sleep 2
fsstress-5070 [007] .... 11007.757792: jbd2_sleep_on_shadow: dev 254,17 sleep 0
fsstress-5071 [011] .... 11007.763493: jbd2_sleep_on_shadow: dev 254,17 sleep 1
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@xxxxxxxxxxxxxxxxx>
I think maybe it might be better to use units of microseconds and then
change sleep to usleep so the units are clear? This is a spinlock, so
it should be quick.
Sorry, I may not quite understand you, do you mean that milliseconds is not precise, so
should use microseconds? For these two patches, they do not use usleep or msleep to do
real sleep work, they just record the duration which process takes to wait bh_shadow flag
to be cleared or transaction to be unlocked.
Regards,
Xiaougang Wang
For the other patch in this series, milliseconds seems fine, but if we
change the trace info to use "msleep" instead that would be clearer
--- or you could change it to use microseconds as well just for
consistency; I think either would be fine.
What do you think?
Cheers,
- Ted
