[POSIX] psutil.users() loses precision for "started" attribute

[djast]

In _psutil_linux.c, psutil_users() casts ut->ut_tv.tv_sec to type float (single-precision); this is insufficient to represent contemporary timestamp values with reasonable accuracy.

Currently, psutil.users()[i].started can be inaccurate by more than a full minute.

Note that while Python float values are represented as 64-bit double-precision values, the float type in C is typically only 32-bit. The code should probably use "000d0" _Py_PARSE_PID and (double)ut->ut_tv.tv_sec rather than "000f0" _Py_PARSE_PID and (float)ut->ut_tv.tv_sec in the Py_BuildValue() call.

[giampaolo]

Please create a PR or a diff to attach to this ticket.

[djast]

I don't have a lot of CPython dev experience, but I think this is what's needed for the Linux case:

psutil.txt

However, a similar fix is almost certainly needed for (at least) _psutil_aix.c, _psutil_bsd.c, _psutil_osx.c, and _psutil_sunos.c; some of these may also need comparable changes in their psutil_boot_time() functions (which appear to use utmp rather than /proc/stat as Linux does).

In a nutshell, anywhere the code says "(float)ut->ut_tv.tv_sec" or "(float)utx->ut_tv.tv_sec" requires attention.

[giampaolo]

Yes, I added a unit test and I could verify that the time is off by 1 minute compared to who command.

[giampaolo]

PR: #2226.

[djast]

"boot_time = (float)ut->ut_tv.tv_sec" in _psutil_aix.c and _psutil_sunos.c probably also need similar fixes (the declaration of the boot_time variable, the cast, and the Py_BuildValue call).

[giampaolo]

Yes, I have changed it also on those platforms.

[djast]

I see that you fixed psutil_users() for those platforms; I'm referring to psutil_boot_time(), which appears to have a nearly identical bug:

psutil/psutil/_psutil_aix.c

Line 724 in 90b35e3

boot_time = (float)ut->ut_tv.tv_sec;

psutil/psutil/_psutil_sunos.c

Line 1425 in 90b35e3

boot_time = (float)ut->ut_tv.tv_sec;

Technically these bugs aren't related to this psutils.users() issue, but assuming the code is used, it appears to me that psutil.boot_time() would have a similar loss of precision. (I'd file this as a separate issue, but I don't have those platforms available to confirm the bug.)

[giampaolo]

Thank you. I have created #2228 to keep track of the boot_time() precision issue.