Messages in this thread | | | From | Dmitry Vyukov <> | Date | Mon, 11 Oct 2021 19:10:20 +0200 | Subject | Re: [syzbot] KASAN: stack-out-of-bounds Read in profile_pc |
| |
On Mon, 11 Oct 2021 at 16:43, Steven Rostedt <rostedt@goodmis.org> wrote: > > On Mon, 11 Oct 2021 14:07:15 +0100 > Lee Jones <lee.jones@linaro.org> wrote: > > > On Thu, 03 Jun 2021, Andi Kleen wrote: > > > > > > > > > True, ftrace does have function profiling (function_profile_enabled). > > > > > > > > Steve, is there a way to enable that on the kernel cmdline? > > > > > > That's not really comparable. function profiling has a lot more overhead. > > > Also there is various code which has ftrace instrumentation disabled. > > > > > > I don't think why you want to kill the old profiler. It's rarely used, but > > > when you need it usually works. It's always good to have simple fall backs. > > > And it's not that it's a lot of difficult code. > > > > sysbot is still sending out reports on this: > > > > https://syzkaller.appspot.com/bug?id=00c965d957410afc0d40cac5343064e0a98b9ecd > > > > Are you guys still planning on sending out a fix? > > > > Is there anything I can do to help? > > > > According to the above: > > ================================================================== > BUG: KASAN: stack-out-of-bounds in profile_pc+0xa4/0xe0 arch/x86/kernel/time.c:42 > Read of size 8 at addr ffffc90001c0f7a0 by task systemd-udevd/12323 > > CPU: 1 PID: 12323 Comm: systemd-udevd Not tainted 5.13.0-rc3-syzkaller #0 > Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 > Call Trace: > <IRQ> > __dump_stack lib/dump_stack.c:79 [inline] > dump_stack+0x202/0x31e lib/dump_stack.c:120 > print_address_description+0x5f/0x3b0 mm/kasan/report.c:233 > __kasan_report mm/kasan/report.c:419 [inline] > kasan_report+0x15c/0x200 mm/kasan/report.c:436 > profile_pc+0xa4/0xe0 arch/x86/kernel/time.c:42 > profile_tick+0xcd/0x120 kernel/profile.c:408 > tick_sched_handle kernel/time/tick-sched.c:227 [inline] > tick_sched_timer+0x287/0x420 kernel/time/tick-sched.c:1373 > __run_hrtimer kernel/time/hrtimer.c:1537 [inline] > __hrtimer_run_queues+0x4cb/0xa60 kernel/time/hrtimer.c:1601 > hrtimer_interrupt+0x3b3/0x1040 kernel/time/hrtimer.c:1663 > local_apic_timer_interrupt arch/x86/kernel/apic/apic.c:1089 [inline] > __sysvec_apic_timer_interrupt+0xf9/0x270 arch/x86/kernel/apic/apic.c:1106 > sysvec_apic_timer_interrupt+0x8c/0xb0 arch/x86/kernel/apic/apic.c:1100 > > And the code has: > > profile_pc+0xa4/0xe0 arch/x86/kernel/time.c:42 > > unsigned long profile_pc(struct pt_regs *regs) > { > unsigned long pc = instruction_pointer(regs); > > if (!user_mode(regs) && in_lock_functions(pc)) { > #ifdef CONFIG_FRAME_POINTER > return *(unsigned long *)(regs->bp + sizeof(long)); > #else > unsigned long *sp = (unsigned long *)regs->sp; > /* > * Return address is either directly at stack pointer > * or above a saved flags. Eflags has bits 22-31 zero, > * kernel addresses don't. > */ > if (sp[0] >> 22) > return sp[0]; <== line 42 > if (sp[1] >> 22) > return sp[1]; > #endif > } > return pc; > } > EXPORT_SYMBOL(profile_pc); > > > It looks to me that the profiler is doing a trick to read the contents of > the stack when the interrupt went off, but this triggers the KASAN > instrumentation to think it's a mistake when it's not. aka "false positive". > > How does one tell KASAN that it wants to go outside the stack, because it > knows what its doing? > > Should that just be converted to a "copy_from_kernel_nofault()"? That is, > does this fix it? (not even compiled tested) > > -- Steve > > > diff --git a/arch/x86/kernel/time.c b/arch/x86/kernel/time.c > index e42faa792c07..cc6ec29aa14d 100644 > --- a/arch/x86/kernel/time.c > +++ b/arch/x86/kernel/time.c > @@ -34,15 +34,18 @@ unsigned long profile_pc(struct pt_regs *regs) > return *(unsigned long *)(regs->bp + sizeof(long)); > #else > unsigned long *sp = (unsigned long *)regs->sp; > + unsigned long retaddr; > /* > * Return address is either directly at stack pointer > * or above a saved flags. Eflags has bits 22-31 zero, > * kernel addresses don't. > */ > - if (sp[0] >> 22) > - return sp[0]; > - if (sp[1] >> 22) > - return sp[1]; > + if (!copy_from_kernel_nofault(&retaddr, sp, sizeof(long)) && > + retaddr >> 22) > + return retaddr; > + if (!copy_from_kernel_nofault(&retaddr, sp + 1, sizeof(long)) && > + retaddr >> 22) > + return retaddr; > #endif > } > return pc;
This looks like a fit for READ_ONCE_NOCHECK(). It's already used in a number of similar places like unwinders: E.g.: https://elixir.bootlin.com/linux/v5.15-rc4/source/arch/x86/kernel/process.c#L984 https://elixir.bootlin.com/linux/v5.15-rc4/A/ident/READ_ONCE_NOCHECK
| |