On Wed, Oct 2, 2013 at 10:29 AM, david j <[email protected]> wrote: > It's possible I'm missing something though, because I don't see why > batching many remaps per TLB invalidate was such a critical operation for > them. >
...unless, their need was not for batching many VM changes per call, but simply to be able to do several VM mapping/protection transitions, on a single page, spread out in time without implicit TLB invalidates at each one. This is suggested by the C4 paper "Figure 3. Live cycle of a heap page", which has 7 map/protect transitions for a single page during evacuation, only two of which require TLB invalidates (I think). This could perhaps be overcome by simply collapsing some of their unique page states, and adding more cost and complexity to the fault handler, reducing the number of mprotect calls (and thus TLB invalidates). The cost of this change is proportional to the number of faults, which in their system are hopefully low.
_______________________________________________ bitc-dev mailing list [email protected] http://www.coyotos.org/mailman/listinfo/bitc-dev
