Arnd Bergmann [off-list ref] writes:

On Tue, Oct 28, 2025, at 06:31, Ankur Arora wrote:

quoted

+ */
+#ifndef smp_cond_load_relaxed_timeout
+#define smp_cond_load_relaxed_timeout(ptr, cond_expr, time_check_expr)	\
+({									\
+	typeof(ptr) __PTR = (ptr);					\
+	__unqual_scalar_typeof(*ptr) VAL;				\
+	u32 __n = 0, __spin = SMP_TIMEOUT_POLL_COUNT;			\
+									\
+	for (;;) {							\
+		VAL = READ_ONCE(*__PTR);				\
+		if (cond_expr)						\
+			break;						\
+		cpu_poll_relax(__PTR, VAL);				\
+		if (++__n < __spin)					\
+			continue;					\
+		if (time_check_expr) {					\
+			VAL = READ_ONCE(*__PTR);			\
+			break;						\
+		}							\
+		__n = 0;						\
+	}								\
+	(typeof(*ptr))VAL;						\
+})
+#endif

I'm trying to think of ideas for how this would done on arm64
with FEAT_FWXT in a way that doesn't hurt other architectures.

The best idea I've come up with is to change that inner loop
to combine the cpu_poll_relax() with the timecheck and then
define the 'time_check_expr' so it has to return an approximate
(ceiling) number of nanoseconds of remaining time or zero if
expired.

Agree that it's a pretty good idea :). I came up with something pretty
similar. Though that had taken a bunch of iterations.

The FEAT_WFXT version would then look something like

static inline void __cmpwait_u64_timeout(volatile u64 *ptr, unsigned long val, __u64 ns)
{
   unsigned long tmp;
   asm volatile ("sev; wfe; ldxr; eor; cbnz; wfet; 1:"
        : "=&r" (tmp), "+Q" (*ptr)
        : "r" (val), "r" (ns));
}
#define cpu_poll_relax_timeout_wfet(__PTR, VAL, TIMECHECK) \
({                                                    \
       u64 __t = TIMECHECK;
       if (__t)
            __cmpwait_u64_timeout(__PTR, VAL, __t);
})

while the 'wfe' version would continue to do the timecheck after the
wait.

I think this is a good way to do it if we need the precision
at some point in the future.

I have two lesser concerns with the generic definition here:

- having both a timeout and a spin counter in the same loop
  feels redundant and error-prone, as the behavior in practice
  would likely depend a lot on the platform. What is the reason
  for keeping the counter if we already have a fixed timeout
  condition?

The main reason was that the time check is expensive in power terms.
Which is fine for platforms with a WFE like primitive but others
want to do the time check only infrequently. That's why poll_idle()
introduced a rate limit on polling (which the generic definition
reused here.)

    commit 4dc2375c1a4e88ed2701f6961e0e4f9a7696ad3c
    Author: Rafael J. Wysocki [off-list ref]
    Date:   Tue Mar 27 23:58:45 2018 +0200

    cpuidle: poll_state: Avoid invoking local_clock() too often

    Rik reports that he sees an increase in CPU use in one benchmark
    due to commit 612f1a22f067 "cpuidle: poll_state: Add time limit to
    poll_idle()" that caused poll_idle() to call local_clock() in every
    iteration of the loop.  Utilization increase generally means more
    non-idle time with respect to total CPU time (on the average) which
    implies reduced CPU frequency.

    Doug reports that limiting the rate of local_clock() invocations
    in there causes much less power to be drawn during a CPU-intensive
    parallel workload (with idle states 1 and 2 disabled to enforce more
    state 0 residency).

    These two reports together suggest that executing local_clock() on
    multiple CPUs in parallel at a high rate may cause chips to get hot
    and trigger thermal/power limits on them to kick in, so reduce the
    rate of local_clock() invocations in poll_idle() to avoid that issue.

- I generally dislike the type-agnostic macros like this one,
  it adds a lot of extra complexity here that I feel can be
  completely avoided if we make explicitly 32-bit and 64-bit
  wide versions of these macros. We probably won't be able
  to resolve this as part of your series, but ideally I'd like
  have explicitly-typed versions of cmpxchg(), smp_load_acquire()
  and all the related ones, the same way we do for atomic_*()
  and atomic64_*().

Ah. And the caller uses say smp_load_acquire_long() or whatever, and
that resolves to whatever makes sense for the arch.

The __unqual_scalar_typeof() does look pretty ugly when looking at the
preprocesed version but other than that smp_cond_load() etc look
pretty straight forward. Just for my curiousity could you elaborate on
the complexity?

--
ankur