On Fri, Mar 23, 2007 at 03:47:45PM +0100, Franck Bui-Huu wrote: > On 3/23/07, Ralf Baechle <ralf@xxxxxxxxxxxxxx> wrote: > >The other thing I still need to understand is why nobody actually seems > >to have triggered this bug on MIPS so far. I suppose our implementation > >of flush_dcache_page() may have done a successful job at papering it > >which means there might be some performance getting lost there as well. > > > > Just to understand, doesn't all mappings of shared anonymous pages and > kernel addresses of them share the same cache lines ? That's true only for all userspace mappings and an anonymous page should normally have only a single mapping per mm anyway. But to make things more complicated a page of course also has a kernel space address in KSEG0 or XKPHYS and on a VIPT cache there we frequently have the case where the user address and the kernel address would map to a different cache line. Let me illustrate this with a little example. Assume we have a page at physical address 0x5000, a page size of 4kB, an 8kB direct mapped cache and 32-byte cache lines. Then address bits 0..4 will be the byte index into the cache line, address bits 5..12 will index the cache array. So now let's map our page into userspace, at address 0x12340000. In KSEG0 it is accessible at 0x80005000. Now, compute the cache index for both addresses compare and curse ... Ralf
