On Thu, 2024-08-22 at 09:14 +0200, Christoph Hellwig wrote: > > I'd suggest two things: > > 1) remove the warning. The use case is perfectly valid and everything > using uncached memory is already slow, so people will just have to > deal with it. Maybe offer a trace point instead if people care about > it.
Well... there's slow and extremely slow .. :-) dcbz() on uncached memory will take an exception for every "cache line" of zeros which will then need to be emulated. We are talking about hundreds to thousands of cycle per cache line. (Do we have an optimisation to detect memset in the emulation code and force it to return to a non-dcbz'ing version ? If not, that's doable and would at least limit it to one exception per memset() instead of one per cache-line). Cheers, Ben.
