Re: 2.6.12-rc5-mm1

[Mikael Pettersson <mikpe@csd.uu.se>]

Andrea Arcangeli writes:
 > Speed is not an issue, cr4 would never be tweaked unless you use
 > seccomp, the cr4 flip is in an extreme slow path.
 > 
 > I think there are two ways to solve this race:
 > 
 > 1) why don't we read the cr4 from the cpu? would movl %%cr4, %%eax
 > generate a general protection fault? Can the cr4 be read in ring 0?
 > Why to read it from memory if we've that information in the cpu already?

Yes the kernel can read cr4. I believe the best solution is to modify
__flush_tlb_global() to read cr4's current value and mix it in when
toggling CR4.PGE and uploading mmu_cr4_features.
(But please make it conditional on CONFIG_SECCOMP.)

The code called from __switch_to() would have to set or clear cr4
locally only. That's easy using write_cr4() and read_cr4().

/Mikael