Alexander, On Thu, May 05 2022 at 20:04, Alexander Potapenko wrote: > On Tue, May 3, 2022 at 12:00 AM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote: >> > Regarding the regs, you are right. It should be enough to unpoison the >> > regs at idtentry prologue instead. >> > I tried that initially, but IIRC it required patching each of the >> > DEFINE_IDTENTRY_XXX macros, which already use instrumentation_begin(). >> >> Exactly 4 instances :) >> > > Not really, I had to add a call to `kmsan_unpoison_memory(regs, > sizeof(*regs));` to the following places in > arch/x86/include/asm/idtentry.h: > - DEFINE_IDTENTRY() > - DEFINE_IDTENTRY_ERRORCODE() > - DEFINE_IDTENTRY_RAW() > - DEFINE_IDTENTRY_RAW_ERRORCODE() > - DEFINE_IDTENTRY_IRQ() > - DEFINE_IDTENTRY_SYSVEC() > - DEFINE_IDTENTRY_SYSVEC_SIMPLE() > - DEFINE_IDTENTRY_DF() > > , but even that wasn't enough. For some reason I also had to unpoison > pt_regs directly in > DEFINE_IDTENTRY_SYSVEC(sysvec_apic_timer_interrupt) and > DEFINE_IDTENTRY_IRQ(common_interrupt). > In the latter case, this could have been caused by > asm_common_interrupt being entered from irq_entries_start(), but I am > not sure what is so special about sysvec_apic_timer_interrupt(). > > Ideally, it would be great to find that single point where pt_regs are > set up before being passed to all IDT entries. > I used to do that by inserting calls to kmsan_unpoison_memory right > into arch/x86/entry/entry_64.S > (https://github.com/google/kmsan/commit/3b0583f45f74f3a09f4c7e0e0588169cef918026), > but that required storing/restoring all GP registers. Maybe there's a > better way? Yes. Something like this should cover all exceptions and syscalls before anything instrumentable can touch @regs. Anything up to those points is either off-limit for instrumentation or does not deal with @regs. --- a/kernel/entry/common.c +++ b/kernel/entry/common.c @@ -24,6 +24,7 @@ static __always_inline void __enter_from user_exit_irqoff(); instrumentation_begin(); + unpoison(regs); trace_hardirqs_off_finish(); instrumentation_end(); } @@ -352,6 +353,7 @@ noinstr irqentry_state_t irqentry_enter( lockdep_hardirqs_off(CALLER_ADDR0); rcu_irq_enter(); instrumentation_begin(); + unpoison(regs); trace_hardirqs_off_finish(); instrumentation_end(); @@ -367,6 +369,7 @@ noinstr irqentry_state_t irqentry_enter( */ lockdep_hardirqs_off(CALLER_ADDR0); instrumentation_begin(); + unpoison(regs); rcu_irq_enter_check_tick(); trace_hardirqs_off_finish(); instrumentation_end(); @@ -452,6 +455,7 @@ irqentry_state_t noinstr irqentry_nmi_en rcu_nmi_enter(); instrumentation_begin(); + unpoison(regs); trace_hardirqs_off_finish(); ftrace_nmi_enter(); instrumentation_end(); As I said: 4 places :) > Fortunately, I don't think we need to insert extra instrumentation > into instrumentation_begin()/instrumentation_end() regions. > > What I have in mind is adding a bool flag to kmsan_context_state, that > the instrumentation sets to true before the function call. > When entering an instrumented function, KMSAN would check that flag > and set it to false, so that the context state can be only used once. > If a function is called from another instrumented function, the > context state is properly set up, and there is nothing to worry about. > If it is called from non-instrumented code (either noinstr or the > skipped files that have KMSAN_SANITIZE:=n), KMSAN would detect that > and wipe the context state before use. Sounds good. Thanks, tglx
