Ralf Baechle wrote:
Yeah, that was inevitable. Since the distribution of the previous
message seems to cover concerned developers outside the MIPS community,
let me make one final(?) plea to actually do this right.
Your patch appears correct - and mipsmt_sys_sched_setaffinity() even
more broken than you thought. It duplicates some code from kernel/sched.c
and has gotten out of sync.
The MIPS SMTC support for managing a single FPU context on a processor
with multiple integer TC contexts involves having the system make
automous, real time, decisions about scheduling affinity. It may be a
first, but it can't possibly be the only case, especially as we've
started seeing more and more mainstream multi-core, multi-thread
designs. System and chip resources are going to be "closer" to one
processor or another. The current Linux paradigm is that it's the
responsibility of programs, or users, to know what the optimal placement
of processes should be for a given system platform, and while it's
absolutely appropriate to provide that level of control for the cases
where the user really does know best, it's mildly insane to make that
the only way that thread placement can be optimized. It's really the
OS's job to match demand to resources.
My contention years ago was, and remains, that it would be a bad idea to
burden the main scheduler loop with checks for two different levels of
affinity, system-automatic and user-specified. It would add
non-trivially to the cache footprint and execution overhead of thread
dispatch, and there's no logical need for it. So the model I proposed,
and implemented in the cloned affinity system calls for SMTC, was that a
*single* affinity mask continue to be used by the scheduler, but that
the per-thread data structures carry two: The one requested explicitly
by the user, and the one actually used by the scheduler. The idea is
that normally those two are the same, but that the system - in the MIPS
SMTC case, the FPU emulator - can overlay its constraints with the
user's constraints to come up with an intersection-of-sets constraint
that satisfies both (there was a clause which prevents system affinity
heuristics from restricting the affinity mask to a null set of CPUs,
though of course the user can do that if he really wants to).
There's nothing particularly MIPS-specific about the problem or the
solution. Most of the mechanisms should really be in
platform-independent code, so we don't get the drift of cloned components.
I don't recall who owned the scheduler at the time, but whoever it was
was too busy dealing with mainstream processor problems to even engage
in a dialogue about this. Is it time to raise the question again?