On 31/08/18 23:18, Jiri Kosina wrote: > On Wed, 29 Aug 2018, Juergen Gross wrote: > >> While being very unlikely I still believe this is possible. Any >> thoughts? > > So in theory we should somehow test whether the next page is some form of > mmio/gart/... mapping, but I guess that by itself would kill the > performance advantage of the whole load_unaligned_zeropad() trick. > > And yes, the sideffects of reading from mmio mapping is a real thing -- > see for example the issue fixed by 2a3e83c6f9. > > If noone has any clever idea how to work this around (I don't), I am > afraid we'd have to ditch the whole DCACHE_WORD_ACCESS optimization, as > it's silently dangerous.
Are there any architectures which can still use DCACHE_WORD_ACCESS? x86 shouldn't, what about powerpc, arm and arm64? I'm happy to send patches to remove selection of DCACHE_WORD_ACCESS and eventually to remove all of the load_unaligned_zeropad() stuff. Juergen
