Opportunistically update the page's dirty field in pmap_ts_referenced().

[alc]

Summary

This change has us opportunistically updating the page's dirty field in pmap_ts_referenced(). I think that the comments explain the benefits.

This won't eliminate pmap_is_modified() calls when the page is not mapped for write access, but those cases are handled efficiently by pmap_is_modified() using the PGA_WRITEABLE flag. Only writeable pages that have not been written to will bear the full cost of pmap_is_modified().

This is a work in progress. I've only written the amd64 implementation as a demonstration.

[alc]

Here is some data from the debug counters.

debug.counter_yes: 59075
debug.counter_no: 246172

These values come after 3 and a fraction iterations of a parallel "buildworld" on a machine configured with 3GB memory.

[markj]

amd64/amd64/pmap.c

5901 ↗

(On Diff #19875)

Is there a reason these assignments are in the loop bodies rather than being lifted to the top of the function?

[alc]

amd64/amd64/pmap.c

5901 ↗

(On Diff #19875)

Yes, they could change from one pmap to the next. Specifically, on Intel processors, we could have both EPT and regular format page tables.

[alc]

Since this is an optimization, I don't need to change the pmap implementations all at once, so I will likely commit this change in a few hours.