Messages in this thread | | | From | Andy Lutomirski <> | Subject | Re: Perf: WARNING: arch/x86/entry/common.c:624 idtentry_exit_cond_rcu+0x92/0xc0 | Date | Fri, 12 Jun 2020 02:18:07 -0700 |
| |
> On Jun 12, 2020, at 2:01 AM, Thomas Gleixner <tglx@linutronix.de> wrote: > > Andy Lutomirski <luto@kernel.org> writes: >> On Thu, Jun 11, 2020 at 4:22 PM Andy Lutomirski <luto@kernel.org> wrote: >> >> Two bugs here. >> >> 1. We had an issue with WARN. Patch sent. > > Grabbed it > >> 2. idtentry.h has, for x86_32: >> >> # define DEFINE_IDTENTRY_IST DEFINE_IDTENTRY >> >> This is nonsense. It's getting late over here and I'd rather focus on >> the more interesting RCU issue, so that's all from me today. > > Well, this might be nonsense, but it's exactly matching the current code > in mainline which, e.g. for #DB does: > > SYM_CODE_START(debug) > /* > * Entry from sysenter is now handled in common_exception > */ > ASM_CLAC > pushl $0 > pushl $do_debug > jmp common_exception > SYM_CODE_END(debug) > > There is no IST on 32bit, never was. We do software stack switching for > device interrupts, but that's a different story. >
DEFINE_IDTENTRY does the idtentry_enter_cond_rcu() dance, which isn’t intended to be safe from NMI context. It should probably map to DEFINE_IDTENTRY_RAW() instead. The specific issue is that NMI ends up there, and at least DEFINE_IDTENTRY_NMI should be raw.
I haven’t tried this at all, nor have I dug through all the users of these macros to check what they expect. Perhaps we should not have the _IST one defined at all on 32 bit and rename it to DEFINE_IDTENTRY_IST_RAW on 64 bit to make it more clear what’s going on when reading the C code.
Or maybe I’m too sleepy and I’m nuts. But I don’t think I am.
> Thanks, > > tglx
| |