On Fri, Jul 13, 2018 at 3:56 AM, Joerg Roedel <joro@xxxxxxxxxx> wrote: > Hi Andy, > > thanks for you valuable feedback. > > On Thu, Jul 12, 2018 at 02:09:45PM -0700, Andy Lutomirski wrote: >> > On Jul 11, 2018, at 4:29 AM, Joerg Roedel <joro@xxxxxxxxxx> wrote: >> > -.macro SAVE_ALL pt_regs_ax=%eax >> > +.macro SAVE_ALL pt_regs_ax=%eax switch_stacks=0 >> > cld >> > + /* Push segment registers and %eax */ >> > PUSH_GS >> > pushl %fs >> > pushl %es >> > pushl %ds >> > pushl \pt_regs_ax >> > + >> > + /* Load kernel segments */ >> > + movl $(__USER_DS), %eax >> >> If \pt_regs_ax != %eax, then this will behave oddly. Maybe it’s okay. >> But I don’t see why this change was needed at all. > > This is a left-over from a previous approach I tried and then abandoned > later. You are right, it is not needed. > >> > +/* >> > + * Called with pt_regs fully populated and kernel segments loaded, >> > + * so we can access PER_CPU and use the integer registers. >> > + * >> > + * We need to be very careful here with the %esp switch, because an NMI >> > + * can happen everywhere. If the NMI handler finds itself on the >> > + * entry-stack, it will overwrite the task-stack and everything we >> > + * copied there. So allocate the stack-frame on the task-stack and >> > + * switch to it before we do any copying. >> >> Ick, right. Same with machine check, though. You could alternatively >> fix it by running NMIs on an irq stack if the irq count is zero. How >> confident are you that you got #MC right? > > Pretty confident, #MC uses the exception entry path which also handles > entry-stack and user-cr3 correctly. It might go through through the slow > paranoid exit path, but that's okay for #MC I guess. > > And when the #MC happens while we switch to the task stack and do the > copying the same precautions as for NMI apply. > >> > + */ >> > +.macro SWITCH_TO_KERNEL_STACK >> > + >> > + ALTERNATIVE "", "jmp .Lend_\@", X86_FEATURE_XENPV >> > + >> > + /* Are we on the entry stack? Bail out if not! */ >> > + movl PER_CPU_VAR(cpu_entry_area), %edi >> > + addl $CPU_ENTRY_AREA_entry_stack, %edi >> > + cmpl %esp, %edi >> > + jae .Lend_\@ >> >> That’s an alarming assumption about the address space layout. How >> about an xor and an and instead of cmpl? As it stands, if the address >> layout ever changes, the failure may be rather subtle. > > Right, I implement a more restrictive check. But the check needs to be correct or we'll mess up, right? I think the code will be much more robust and easier to review if you check "on the entry stack" instead of ">= the entry stack". (Or <= -- I can never remember how this works in AT&T syntax.) > >> Anyway, wouldn’t it be easier to solve this by just not switching >> stacks on entries from kernel mode and making the entry stack bigger? >> Stick an assertion in the scheduling code that we’re not on an entry >> stack, perhaps. > > That'll save us the check whether we are on the entry stack and replace > it with a check whether we are coming from user/vm86 mode. I don't think > that this will simplify things much and I am a bit afraid that it'll > break unwritten assumptions elsewhere. It is probably something we can > look into later separatly from the basic pti-x32 enablement. > Fair enough. There's also the issue that NMI still has to switch CR3 if it hits with the wrong CR3. I personally much prefer checking whether you came from user mode rather than the stack address, but I'm okay with either approach here.
