[dpdk-dev] [RFC 3/5] eal: lcore state FINISHED is not required

Ananyev, Konstantin konstantin.ananyev at intel.com
Fri Mar 19 14:42:39 CET 2021


Hi everyone,

> <snip>
> 
> > >
> > > > > Subject: [RFC 3/5] eal: lcore state FINISHED is not required
> > > > >
> > > > > FINISHED state seems to be used to indicate that the worker's
> > > > > update of the 'state' is not visible to other threads. There seems
> > > > > to be no requirement to have such a state.
> > > >
> > > > I am not sure "FINISHED" is necessary to be removed, and I propose
> > > > some of my profiles for discussion.
> > > >  There are three states for lcore now:
> > > > "WAIT": indicate lcore can start working
> > > > "RUNNING": indicate lcore is working
> > > > "FINISHED": indicate lcore has finished its working and wait to be
> > > > reset
> > > If you look at the definitions of "WAIT" and "FINISHED" states, they look
> > similar, except for "wait to be reset" in "FINISHED" state . The code really does
> > not do anything to reset the lcore. It just changes the state to "WAIT".


I agree that 3 states here seems excessive.
Just 2 (RUNNING/IDLE) seems enough.
Though we can't just remove FINISHED here - it will be an Abi breakage.
Might be deprecate FINISHED now and remove in 21.11.

Also need to decide what rte_eal_wait_lcore() should return in that case?
Always zero, or always status of last function called?

> > >
> > > >
> > > > From the description above, we can find "FINISHED" is different from
> > > > "WAIT", it can shows that lcore has done the work and finished it.
> > > > Thus, if we remove "FINISHED", maybe we will not know whether the
> > > > lcore finishes its work or just doesn't start, because this two state has the
> > same tag "WAIT".
> > > Looking at "eal_thread_loop", the worker thread sets the state to "RUNNING"
> > before sending the ack back to main core. After that it is guaranteed that the
> > worker will run the assigned function. Only case where it will not run the
> > assigned function is when the 'write' syscall fails, in which case it results in a
> > panic.
> >
> > Quick note: it should not panic.
> > We must find a way to return an error
> > without crashing the whole application.
> The syscalls are being used to communicate the status back to the main thread. If they fail, it is not possible to communicate the status.
> May be it is better to panic.
> We could change the implementation using shared variables, but it would require polling the memory. May be the syscalls are being used to
> avoid polling. However, this polling would happen during init time (or similar) for a short duration.

AFAIK we use read and write not for status communication, but sort of sleep/ack point.
Though I agree if we can't do read/write from the system pipe then something is totally wrong,
and probably there is no much point to continue. 
 
> >
> >
> > > > Furthermore, consider such a scenario:
> > > > Core 1 need to monitor Core 2 state, if Core 2 finishes one task,
> > > > Core 1 can start its working.
> > > > However, if there is only  one tag "WAIT", Core 1 maybe  start its
> > > > work at the wrong time, when Core 2 still does not start its task at state
> > "WAIT".
> > > > This is just my guess, and at present, there is no similar
> > > > application scenario in dpdk.
> > > To be able to do this effectively, core 1 needs to observe the state change
> > from WAIT->RUNNING->FINISHED. This requires that core 1 should be calling
> > rte_eal_remote_launch and rte_eal_wait_lcore functions. It is not possible to
> > observe this state transition from a 3rd core (for ex: a worker might go from
> > RUNNING->FINISHED->WAIT->RUNNING which a 3rd core might not be able to
> > observe).
> > >
> > > >
> > > > On the other hand, if we decide to remove "FINISHED", please
> > > > consider the following files:
> > > > 1. lib/librte_eal/linux/eal_thread.c: line 31
> > > >     lib/librte_eal/windows/eal_thread.c: line 22
> > > >     lib/librte_eal/freebsd/eal_thread.c: line 31
> > > I have looked at these lines, they do not capture "why" FINISHED state is
> > required.
> > >
> > >  2.
> > > > lib/librte_eal/include/rte_launch.h: line 24, 44, 121, 123, 131 3.
> > > > examples/l2fwd-
> > > > keepalive/main.c: line 510
> > > > rte_eal_wait_lcore(id_core) can be removed. Because the core state
> > > > has been checked as "WAIT", this is a redundant operation
> >
> >



More information about the dev mailing list