On Thu, 2012-05-17 at 18:44 +0200, Peter Zijlstra wrote: > > So the RCU code can from ppc in commit > 267239116987d64850ad2037d8e0f3071dc3b5ce, which has similar behaviour. > Also I suspect the mm_users < 2 test will be incorrect for ARM since > even the one user can be concurrent with your speculation engine. > > Right, last mail, I promise, I've confused myself enough already! :-) OK, so ppc/sparc are special (forgot all about s390) I think by the time they are done with unmap_page_range() their hardware hash-tables are empty and nobody but software page-table walkers will still access the linux page tables. So when we do free_pgtables() to clean up the actual page-tables. Power/Sparc need to RCU free this to allow concurrent software page-table walkers like gup_fast. Thus I don't think they need to tlb flush again because their hardware doesn't actually walk the link page-tables, it walks hash-tables, which by this time are empty. Now if x86/Xen were to use this, it would indeed also need to TLB flush when freeing the page-tables, since its hardware walkers do indeed traverse these pages and we need to sync against them. So my first patch in the tlb-unify tree is actually buggy. Humm,. what to do adding a tlb flush in there might slow down ppc/sparc unnecessarily.. dave/ben? I guess we need more knobs :-( Now its quite possible I've utterly confused myself and everybody reading, apologies for that, I shall rest and purge all from memory and start over before commenting more.. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href