On Tue, May 25, 2021 at 12:08:35PM -0900, Matthew Dillon wrote:
> The documentation is out of date.
>
> pmap_kenter() synchronizes the TLB on all CPUs.
>
> pmap_kenter_quick() only invalidates the TLB on the current cpu.
>
> pmap_kenter_noinval() does not touch the TLB at all.
>
> The latter two functions are only used under controlled circumstances.
> pmap_kenter_quick() is used when the caller only intends to use the mapping
> on the current cpu, or intends to track which cpu's have synchronized the
> mapping. And pmap_kenter_noinval() is used when the caller is entering a
> ton of PTEs into the pmap and will do a global invalidation when done.
>
> pmap_qenter() passes an argument w/regard to what kind of tlb invalidation
> is desired.
>
> I should note that the buffer cache code (struct buf) tracks mapping
> validity with a cpu mask, so operations on the buffer cache do not cause an
> excessive number of global IPIs for SMP invalidations.
>
> -Matt
Thanks for confirming my suspicion, and for the extra detail. I haven't
been looking at the buffer cache but will keep your note in mind.
I have been keeping track of some of the out-of-date comments I come
across, and just compiled some of what I found into a patch (along with
a little bit of new documentation):
https://bugs.dragonflybsd.org/issues/3279
I hope it's helpful. I realize developer time is limited, but if you
think it's worth your time to review things like that I'll probably
keep sending them :-)
--
James
