On Thu, 2010-08-19 at 14:31 -0700, Andrew Morton wrote: > On Thu, 19 Aug 2010 22:13:17 +0200 > Peter Zijlstra <a.p.zijlstra@xxxxxxxxx> wrote: > > > > > This patch-set reworks the kmap_atomic API to be a stack based, instead of > > static slot based. Some might remember this from last year, some not ;-) > > > > The advantage is that you no longer need to worry about KM_foo, the > > disadvantage is that kmap_atomic/kunmap_atomic now needs to be strictly > > nested (CONFIG_HIGHMEM_DEBUG should complain in case its not) -- and of > > course its a big massive patch changing a widely used API. > > Nice. That fixes the "use of irq-only slots from interrupts-on > context" bugs which people keep adding. Ah, I should add a: WARN_ON_ONCE(in_irq() && !irqs_disabled()); like check to ensure people don't use kmap_atomic() in nestable IRQ contexts (nestable IRQ context is bad anyway) the old debug code I deleted did something similar. > We don't have any checks in there for the stack overflowing? +#ifdef CONFIG_DEBUG_HIGHMEM + BUG_ON(idx > KM_TYPE_NR); +#endif Seems to be that. > Did you add every runtime check you could possibly think of? > kmap_atomic_idx_push() and pop() don't have much in there. It'd be > good to lard it up with runtime checks for at least a few weeks. Right, so I currently have: - stack size check in push/pop - proper nesting check in pop (verifies that the vaddr you try to unmap is indeed the top most on the stack) Aside from the proposed no irq-nesting thing to avoid unbounded recursion I can't really come up with more creative abuse. > Well, there's that monster conversion patch. How's about you > temporarily do > > #define kmap_atomic(x, arg...) __kmap_atomic(x) > > so for a while, both kmap_atomic(a, KM_foo) and kmap_atomic(a) are > turned into __kmap_atomic(a). Once all the dust has settled, pull that > out again? Ah, that's a nifty trick, let me try that. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href