Hi Thierry,

sorry for the delayed reply - I'm finally going through the
documentation patches in preparation for the upcoming next version
patchset mail-out.

On Wed, Sep 22, 2021 at 11:39 AM Thierry Delisle [off-list ref] wrote:

On 2021-09-17 2:03 p.m., Peter Oskolkov wrote:
 > [...]
 > +SYS_UMCG_WAIT()
 > +
 > +int sys_umcg_wait(uint32_t flags, uint64_t abs_timeout) operates on
 > +registered UMCG servers and workers: struct umcg_task *self provided to
 > +sys_umcg_ctl() when registering the current task is consulted in
addition
 > +to flags and abs_timeout parameters.
 > +
 > +The function can be used to perform one of the three operations:
 > +
 > +* wait: if self->next_tid is zero, sys_umcg_wait() puts the current
 > +  server or worker to sleep;

I believe this description is misleading but I might be wrong.
 From the example
     * worker to server context switch (worker "yields"):
       S:IDLE+W:RUNNING => +S:RUNNING+W:IDLE

It seems to me that when a worker goes from running to idle, it should
*not* set the next_tid to 0, it should preserve the next_tid as-is,
which is expected to point to its current server. This is consistent
with my understanding of the umcg_wait implementation. This operation
is effectively a direct context-switch to the server.

The documentation here outlines what sys_umcg_wait does, and it does
put the current task to sleep without context switching if next_tid is
zero. The question of whether this behavior is or is not appropriate
for a worker wishing to yield/park itself is at a "policy" level, if
you wish, and this "policy" level is described in "state transitions"
section later in the document. sys_umcg_wait() does not enforce this
"policy" directly, in order to make it simpler and easier to describe
and reason about.

With that said, I'm a little confused by the usage of "yields" in that
example. I would expect workers yielding to behave like kernel threads
calling sched_yield(), i.e., context switch to the server but also be
immediately added to the idle_workers_ptr.

 From my understanding of the umcg_wait call, "worker to server context
switch" does not have analogous behaviour to sched_yield. Am I correct?
If so, I suggest using "park" instead of "yield" in the description
of that example. I believe the naming of wait/wake as park/unpark is
consistent with Java[1] and Rust[2], but I don't know if that naming
is used in contexts closer to the linux kernel.

[1]
https://docs.oracle.com/javase/8/docs/api/java/util/concurrent/locks/LockSupport.html
[2] https://doc.rust-lang.org/std/thread/fn.park.html

I'm not a fan of arguing about how to name things. If the maintainers
ask me to rename wait/wake to park/unpark, I'll do that. But it seems
they are OK with this terminology, I believe because wait/wake is a
relatively well understood pair of verbs in the kernel context;
futexes, for example, have wait/wake operations. A higher level
library in the userspace may later expose park/unpark functions that
at the lower level call sys_umcg_wait...