On 03/17/2015 01:22 PM, Denys Vlasenko wrote: > On 03/17/2015 08:39 AM, Borislav Petkov wrote: >> On Tue, Mar 17, 2015 at 08:21:18AM +0100, Ingo Molnar wrote: >>> Assuming this does not fix the regression, could you apply the minimal >>> patch below - which reverts the old_rsp handling change. >>> >>> (The rest of the commit are in a third patch, but those are only >>> comment changes.) >>> >>> So my theory is that this change is what will revert the regression. >> >> Yep, it does. Below is the diff that works (it is the rough revert >> without the comments :-)): >> > ... >> @@ -395,6 +398,8 @@ __switch_to(struct task_struct *prev_p, struct task_struct *next_p) >> /* >> * Switch the PDA and FPU contexts. >> */ >> + prev->usersp = this_cpu_read(old_rsp); >> + this_cpu_write(old_rsp, next->usersp); > > I have a theory. There is a time window when user's sp > is in PER_CPU_VAR(old_rsp) but not yet in pt_regs->sp, > and *interrupts are enabled*: > > ENTRY(system_call) > SWAPGS_UNSAFE_STACK > movq %rsp,PER_CPU_VAR(old_rsp) > movq PER_CPU_VAR(kernel_stack),%rsp > ENABLE_INTERRUPTS(CLBR_NONE) > ALLOC_PT_GPREGS_ON_STACK 8 /* +8: space for orig_ax */ > movq %rcx,RIP(%rsp) > movq PER_CPU_VAR(old_rsp),%rcx > ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ > movq %r11,EFLAGS(%rsp) > movq %rcx,RSP(%rsp) > > Before indicated insn, interrupts are already enabled. > If preempt would hit now, next task can clobber PER_CPU_VAR(old_rsp). > Then, when we return to this task, a bogus user's sp will be stored > in pt_regs, restores on exit to userspace, and next attempt > to, say, execute RETQ will try to pop a bogus, likely noncanonical > address into RIP -> #GP -> SEGV! > > The theory can be tested by just moving interrupt enable a bit down: > > ENTRY(system_call) > SWAPGS_UNSAFE_STACK > movq %rsp,PER_CPU_VAR(old_rsp) > movq PER_CPU_VAR(kernel_stack),%rsp > - ENABLE_INTERRUPTS(CLBR_NONE) > ALLOC_PT_GPREGS_ON_STACK 8 /* +8: space for orig_ax */ > movq %rcx,RIP(%rsp) > movq PER_CPU_VAR(old_rsp),%rcx > movq %r11,EFLAGS(%rsp) > movq %rcx,RSP(%rsp) > + ENABLE_INTERRUPTS(CLBR_NONE) > > If I'm right, segfaults should be gone. > Borislav, can you try this? I managed to reproduce the segfault, and the fix shown above works. I see that Ingo removed the failing commit from his tree. I'll send two patches: one which moves ENABLE_INTERRUPTS(CLBR_NONE) down, and another which tries to remove thread_struct::usersp again. -- To unsubscribe from this list: send the line "unsubscribe linux-tip-commits" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
|