On Sun, 6 Feb 2005 22:50:20 +0100 Andi Kleen <[EMAIL PROTECTED]> wrote:
> And yes this stuff does matter - i remember i got LM benchmarkable > improvements in signal latency by optimizing __copy_to_user > to use optimized inlines for small stores. Moving access_ok() out-of-line might even improve I-cache access over what we have today, even with the new min-max check. The min-max variables will be in the same cache line in whatever struct we place them into, so whatever cache miss access_ok() gets now will also be the same for the min-max version. This is kind of strange to be arguing about, given that we just put 4-level page tables into the tree, right? That regressed everybody performance wise, even people not using the full 4-level support. But I have not barked at you about this, I undersand why it's needed. And yet you're using lmbench cycle counting to justify your position against this new verification scheme. And it's not just a sparc64 issue. Sparc64 hardware traps the access, but it's a bug regardless of platform to try to do user accesses whilst get_fs()==KERNEL_DS. All the user has to do is pass in a valid kernel address and you have a root exploit. I mean, do folks really disagree with this?
