Support some POSIX-RT APIs to run rt-tests (cyclictest) on Osv

[hw-claudio]

I have tried building and running cyclictest on OSv:

https://rt.wiki.kernel.org/index.php/Cyclictest

git clone git://git.kernel.org/pub/scm/linux/kernel/git/clrkwllms/rt-tests.git
or tarballs:

https://www.kernel.org/pub/linux/kernel/people/clrkwllms/

there seem to be these missing functions required to be able to do a minimal cyclictest run:

sched_getscheduler
sched_setscheduler

timer_create
timer_settime

do you think it's possible to implement these POSIX rt functions?

Ciao,

Claudio

[penberg]

AFAICT, it should be possible to implement these but you really should ask @nyh, @tgrabiec, and/or @glommer. 😉

[bamboojsr]

@nyh @hw-claudio , Today I met same problem, But I had add pthread into OSv lib.

Adding /libenviron.so...
Adding /zpool.so...
Adding /libzfs.so...
Adding /libuutil.so...
Adding /zfs.so...
Adding /tools/mkfs.so...
Adding /tools/cpiod.so...
Adding /tools/mount-nfs.so...
Adding /tools/umount.so...
Adding /usr/lib/libgcc_s.so.1...
Adding /usr/lib/libpthread.so.0...
Adding /usr/lib/libboost_filesystem.so.1.55.0...
Adding /usr/lib/libboost_regex.so.1.55.0...
Adding /etc/hosts...
Link /etc/mnttab to /proc/mounts ...
Adding /etc/fstab...
Adding /dev...
Adding /proc...
Adding /tmp...
Adding /tools/cyclictest.so...

[nyh]

libpthread.so.0 is on the list of objects supplied by the OSv kernel itself (see the full list in program::program in elf.cc), so adding it to the image should not make a difference (did it?), and even if it did, it is quite unlikely that anything from libpthread.so (which is part of glibc) will actually work unmodified on OSv.

If you're interested in adding real-time scheduling to OSv (can you please say something about your use case?) you might be happy to know that we actually have working patches for many (but not all) of these features: I have a patch (which I continue to maintain off-tree) adding OSv-specific methods set_realtime_priority(), and set_realtime_slice(), and Claudio had patches to use those to implement some of the Posix API for scheduling classes.

My patch involves tinkering inside the OSv scheduler - it is not something as simple as strapping on an additional library like libthread.so.

[bamboojsr]

Yes, Wireless workload 's scheduling in 1ms, So real-time is very important, we don't want to waste more time at OS schedule. In current time ,we find guest's real-time more than native os(aboute 3doubles).
So I want to know the OSv real-time performance by cyclictest .

[hw-claudio]

Hi all,

I will look at the state of the current RT scheduler patchset from Nadav and my RT API branch when I get back next week.

I got cyclictest working at the time using them.

Ciao!

Claudio

From: bamboojsr [mailto:notifications@github.com]
Sent: Freitag, 5. Februar 2016 09:56
To: cloudius-systems/osv
Cc: Claudio Fontana
Subject: Re: [osv] Support some POSIX-RT APIs to run rt-tests (cyclictest) on Osv (#386)

Yes, Wireless workload 's scheduling in 1ms, So real-time is very important, we don't want to waste more time at OS schedule. In current time ,we find guest's real-time more than native os(aboute 3doubles).
So I want to know the OSv real-time performance by cyclictest .

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/386#issuecomment-180153087.

[nyh]

On Tue, Feb 9, 2016 at 5:58 AM, Claudio Fontana notifications@github.com
wrote:

Hi all,

I will look at the state of the current RT scheduler patchset from Nadav
and my RT API branch when I get back next week.

I got cyclictest working at the time using them.

Thanks Claudio. I'll post my part of the patch again (rebased to the latest
code) in a few days.

[nyh]

sched_getscheduler() is needed for Python 3.6, which is apparently (?) compiled with "-z now" so needs a lot of symbols to be available, but doesn't quite care what they do. See #992.

[nyh]

We should start by trivially implementing sched_getscheduler() and friends for the normal, already implemented, SCHED_OTHER. Only later we can worry about the patches I sent to the mailing list, for supporting SCHED_FIFO and incomplete support for SCHED_RR.

[nyh]

If we want to revive the realtime scheduler feature in OSv, here is my old scheduler patch:
https://groups.google.com/d/msg/osv-dev/cdf5jYSsCuc/bENVrfuX6EkJ
this patch adds a new API for controlling the realtime scheduler, and here is @hw-claudio 's patches for implementing the relevant Posix API (sched_getscheduler() et al.):
https://groups.google.com/d/msg/osv-dev/3QN0uZkW9jA/685VZQS9WFYJ

[nyh]

I rebased by low-level scheduler patch as a pull-request, here: #1223

We also need @hw-claudio's patches (see above links, or last few patches in https://github.com/hw-claudio/osv_aarch64/commits/rt_wip https://github.com/hw-claudio/osv_aarch64/commits/rt) to provide the POSIX API compatibility on top of my low-level support. His first patch (hw-claudio@fcb3e13) should work today except the call to set_realtime() should be split to calling set_realtime_priority() and set_realtime_time_slice() (this aspect of the low-level API changed in the 3rd version of my patch). The second patch (hw-claudio@cd9b24f) should be reconsidered because set_realtime() is gone in the last version of my API, so improving it is pointless - we just need the sched_setscheduler to call the right combination of set_realtime_priority() and set_realtime_time_slice().

And remember that set_realtime_time_slice() doesn't actually work in this version (it generates a warning message). So if SCHED_RR is interesting to you, we need to implement it. It's not very complicated, but I never did implement it.