On Fri, May 6, 2022 at 6:14 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote: > > On Fri, May 06 2022 at 16:52, Alexander Potapenko wrote: > > On Thu, May 5, 2022 at 11:56 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote: > >> @@ -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 :) > > > > These four instances still do not look sufficient. > > Right now I am seeing e.g. reports with the following stack trace: > > > > BUG: KMSAN: uninit-value in irqtime_account_process_tick+0x255/0x580 > > kernel/sched/cputime.c:382 > > irqtime_account_process_tick+0x255/0x580 kernel/sched/cputime.c:382 > > account_process_tick+0x98/0x450 kernel/sched/cputime.c:476 > > update_process_times+0xe4/0x3e0 kernel/time/timer.c:1786 > > tick_sched_handle kernel/time/tick-sched.c:243 > > tick_sched_timer+0x83e/0x9e0 kernel/time/tick-sched.c:1473 > > __run_hrtimer+0x518/0xe50 kernel/time/hrtimer.c:1685 > > __hrtimer_run_queues kernel/time/hrtimer.c:1749 > > hrtimer_interrupt+0x838/0x15a0 kernel/time/hrtimer.c:1811 > > local_apic_timer_interrupt arch/x86/kernel/apic/apic.c:1086 > > __sysvec_apic_timer_interrupt+0x1ae/0x680 arch/x86/kernel/apic/apic.c:1103 > > sysvec_apic_timer_interrupt+0x95/0xc0 arch/x86/kernel/apic/apic.c:1097 > > ... > > (uninit creation stack trace is irrelevant here, because it is some > > random value from the stack) > > > > sysvec_apic_timer_interrupt() receives struct pt_regs from > > uninstrumented code, so regs can be partially uninitialized. > > They are not passed down the call stack directly, but are instead > > saved by set_irq_regs() in sysvec_apic_timer_interrupt() and loaded by > > get_irq_regs() in tick_sched_timer(). > > sysvec_apic_timer_interrupt() invokes irqentry_enter() _before_ > set_irq_regs() and irqentry_enter() unpoisons @reg. > > Confused... As far as I can tell in this case sysvect_apic_timer_interrupt() is called by the following code in arch/x86/kernel/idt.c: INTG(LOCAL_TIMER_VECTOR, asm_sysvec_apic_timer_interrupt), , which does not use IDTENTRY_SYSVEC framework and thus does not call irqentry_enter(). I guess handling those will require wrapping every interrupt gate into a function that performs register unpoisoning? By the way, if it helps, I think we don't necessarily have to call kmsan_unpoison_memory() from within the instrumentation_begin()/instrumentation_end() region? We could move the call to the beginning of irqentry_enter(), removing unnecessary duplication.
