So this KPI is not safe to use from interrupt handlers ? I think you need to mention it in the man page.

In D25916#574168, @kib wrote:

So this KPI is not safe to use from interrupt handlers ? I think you need to mention it in the man page.

It is no worse from interrupt handlers as existing random(9) KPI, which does not document that deficiency. But we could use spinlock_enter() in place of critical_enter() if that would solve the problem better.

In D25916#574171, @cem wrote:

In D25916#574168, @kib wrote:

So this KPI is not safe to use from interrupt handlers ? I think you need to mention it in the man page.

It is no worse from interrupt handlers as existing random(9) KPI, which does not document that deficiency. But we could use spinlock_enter() in place of critical_enter() if that would solve the problem better.

It is fine for this KPI to not be interrupt-safe. Adding spinlock would make things (much) worse, either spinlock would be global causing scalability issues, or you have to add per-CPU spinlock, which means the overhead of interrupts flag manipulation, which is serializing on some CPUs.

I only asked to be explicit in the man page, so that the answer to this question did not require code spelunking.

In D25916#574172, @kib wrote:

It is fine for this KPI to not be interrupt-safe. Adding spinlock would make things (much) worse, either spinlock would be global causing scalability issues, or you have to add per-CPU spinlock, which means the overhead of interrupts flag manipulation, which is serializing on some CPUs.

I only asked to be explicit in the man page, so that the answer to this question did not require code spelunking.

Ok, sounds good. I will update the manual page.

From the manpages perspective, a tiny AUTHORS and HISTORY section would also be a good addition to the man page. Just saying... ;-)

Otherwise, the manual seems fine!

In D25916#574631, @markm wrote:

I like this general idea very much! I'm a bit leery of the critical regions, but I admit that I don't understand them properly. I presume that in a PCPU environment they are fine.

Thanks! The critical regions are just used as a very cheap and short-term mutex on the per-CPU state. (The exception, as kib points out, is that CPUs may still be interrupted during a critical section, so this is not safe to use in interrupt handlers.) A handful of multiplications and additions are performed "under lock." (Maybe a few more multiplications and two divisions, if using the _bounded routines.) This is very very cheap and should not significantly impact scheduler fairness too badly.