bpo-21302: Add clock_nanosleep() implementation for time.sleep() #28111
Conversation
clock_nanosleep() is available in Linux which has POSIX 2001.12 or newer
|
Oh nice, it's much better! IMO it's worth to mention that time.sleep() has a resolution of 1 nanosecond if the clock_nanosleep() function is available. Do you know if clock_nanosleep() is available on other platforms than Linux? Is it available on macOS or FreeBSD for example? |
I do not think that it is available in macOS and FreeBSD according the gnulib manual, clock_nanosleep() is available sure in Linux from POSIX version 2001.12. So any modern and new Linux based OS has it, moreover it seems to me QNX has also the same clock_nanosleep(). Is there any right place to mention this improved resolution of time.sleep()? |
PEP 7 rules appled for _PyTime_AsTimespec and _PyTime_AsTimeval error checking in pysleep().
In calling clock_nanosleep() EINTR is not stored in errno. Need to use return value of clock_nanosleep()/select() for checking it.
Unix operating systems eg: Linux, macOS, FreeBSD etc. time.sleep() has a resolution of nanoseconds with using clock_nanosleep() or nanosleep() function.
|
In all Unix systems eg: macOS, FreeBSD, Linux etc, nanosleep() is available and as i see there are some codes where select() is used for sleeping microsecs resolution, semaphore.c, _tkinter.c. Sleeping via select() was the state of the art solution in about 1998 in Linux but it is already an outdated solution nowadays. It would be better to use nanosleep() in these codes also. I will improve it in the rest of codes, too. |
In all Unix systems eg: macOS, FreeBSD, Linux etc, nanosleep() is available.
Livius90
changed the title
Waitable timer is 100 nsec resolution. Now, seconds to nanosec conversion is limited in usec, soon will come the next developing part to improve it in next commit.
Livius90
changed the title
Waitable timer resolution is 100 nsec but it is limited to 1 usec by round ceiling, moreover sleep for lower then 1 milisec is not possible in Win32 API.
|
Improved time.sleep() is ready for final review. All platforms have now the state of the art solution for sleeping. In Windows it uses Waitable timer object, in Linux it is clock_nanosleep() and in other Unix systems eg: macOS, FreeBSD it is nanosleep() function. Resolution of Waitable timer object is 100 nsec but it is limited to 1 usec by round ceiling. However, sleeping for lower than 1 millisec is not possible in Win32 API because the overhead of the the event waiting call is more than nanoseconds. Resolution of clock_nanosleep() and nanosleep() is 1 nsec, clock_nanosleep() is suitable for blocking/waiting until an absolute time date, nanosleep() is suitable only for interval sleeping. @abalkin, @pganssle and @vstinner please make a review for my final state. |
|
PR was changed to add clock_nanosleep() only. Can you review it? |
Misc/NEWS.d/next/Library/2021-09-10-00-24-08.bpo-21302.XS_frc.rst
Outdated
Show resolved
Hide resolved
suggested description Co-authored-by: Victor Stinner <[email protected]>
Can not simplify HAVE_CLOCK_NANOSLEEP blocks because do {} while loop need for EINTR error handling
vstinner
merged commit 85a4748
into
python:main
|
Merged, thanks! I created PR #28311 follow-up to update the documentation. |
|
Thanks the merging. What are the formal rules to create a new PR to improve Windows implementation, too? Is it good to create a new PR without any bpo issue number? Or can i make more PR (windows implementation and nanosleep()) for this bpo-21302? |
|
In this case, you can reuse bpo-21302. |
Requested changes from @vstinner are done. It was runt-time tested in Oracle Linux 8.4 and Windows 10.
Some previous pull request history can be found here: #28077
https://bugs.python.org/issue21302