Re: [PATCH v3 1/2] eal: add thread yield functions

[Thomas Monjalon <thomas@monjalon.net>]

25/10/2023 18:40, Bruce Richardson:
> On Wed, Oct 25, 2023 at 06:31:10PM +0200, Thomas Monjalon wrote:
> > When running real-time threads, we may need to force scheduling
> > kernel threads or other real-time threads.
> > New functions are added to address these cases.
> > 
> > The yield functions should not have any interest for normal threads.
> > Note: other purposes may be addressed with rte_pause() or rte_delay_*().
> > 
> > Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
> > ---
> >  lib/eal/include/rte_thread.h | 22 ++++++++++++++++++++++
> >  lib/eal/unix/rte_thread.c    | 16 ++++++++++++++++
> >  lib/eal/version.map          |  4 ++++
> >  lib/eal/windows/rte_thread.c | 15 +++++++++++++++
> >  4 files changed, 57 insertions(+)
> > 
> > diff --git a/lib/eal/include/rte_thread.h b/lib/eal/include/rte_thread.h
> > index 8da9d4d3fb..139cafac96 100644
> > --- a/lib/eal/include/rte_thread.h
> > +++ b/lib/eal/include/rte_thread.h
> > @@ -183,6 +183,28 @@ int rte_thread_join(rte_thread_t thread_id, uint32_t *value_ptr);
> >   */
> >  int rte_thread_detach(rte_thread_t thread_id);
> >  
> > +/**
> > + * Allow another thread to run on the same CPU core.
> > + *
> > + * Lower priority threads may not be scheduled.
> > + *
> > + * Especially useful in real-time thread priority
> > + * to schedule other real-time threads.
> > + * @see RTE_THREAD_PRIORITY_REALTIME_CRITICAL
> > + */
> > +__rte_experimental
> > +void rte_thread_yield(void);
> > +
> > +/**
> > + * Unblock a CPU core running busy in a real-time thread.
> > + *
> > + * Especially useful in real-time thread priority
> > + * to avoid a busy loop blocking vital threads on a core.
> > + * @see RTE_THREAD_PRIORITY_REALTIME_CRITICAL
> > + */
> > +__rte_experimental
> > +void rte_thread_yield_realtime(void);
> > +
> >  /**
> >   * Get the id of the calling thread.
> >   *
> > diff --git a/lib/eal/unix/rte_thread.c b/lib/eal/unix/rte_thread.c
> > index 36a21ab2f9..92b4e53adb 100644
> > --- a/lib/eal/unix/rte_thread.c
> > +++ b/lib/eal/unix/rte_thread.c
> > @@ -5,9 +5,11 @@
> >  
> >  #include <errno.h>
> >  #include <pthread.h>
> > +#include <sched.h>
> >  #include <stdbool.h>
> >  #include <stdlib.h>
> >  #include <string.h>
> > +#include <time.h>
> >  
> >  #include <rte_errno.h>
> >  #include <rte_log.h>
> > @@ -227,6 +229,20 @@ rte_thread_detach(rte_thread_t thread_id)
> >  	return pthread_detach((pthread_t)thread_id.opaque_id);
> >  }
> >  
> > +void
> > +rte_thread_yield(void)
> > +{
> > +	sched_yield();
> > +}
> > +
> > +void
> > +rte_thread_yield_realtime(void)
> > +{
> > +	/* A simple yield may not be enough to schedule kernel threads. */
> > +	struct timespec wait = {.tv_nsec = 1};
> > +	nanosleep(&wait, NULL);
> > +}
> > +
> While I realise we discussed this earlier, and I also was the original
> suggester of using sched_yield, I think just having just one function using
> sleep is probably best after all.

I think there is a value to have a simple yield function
for scheduling between multiple real-time threads
without sleep overhead (not sure about the overhead).
If there is not much overhead, then a single function is OK.

Note I'm preparing a new version with a simple yield implemented
with the lighter SwitchToThread() on Windows.