On 24/03/2021 18.33, Peter Zijlstra wrote: > On Wed, Mar 24, 2021 at 05:45:52PM +0100, Rasmus Villemoes wrote: >> Sorry, I think I misread the code. The static calls are indeed >> initialized with a function with the right prototype. Try adding >> "preempt=full" on the command line so that we exercise these lines >> >> static_call_update(cond_resched, >> (typeof(&__cond_resched)) __static_call_return0); >> static_call_update(might_resched, >> (typeof(&__cond_resched)) __static_call_return0); >> >> I would expect that to blow up, since we end up calling a long (*)(void) >> function using a function pointer of type int (*)(void). > > Note that on x86 there won't actually be any calling of function > pointers. See what arch/x86/kernel/static_call.c does :-)
I know, but so far x86 is the only one with HAVE_STATIC_CALL, so for arm64 which is where CFI seems to be targeted initially, static_calls are function pointers. And unless CFI ignores the return type, I'd really expect the above to fail. > But I think some of this code might need some __va_function() love when > combined with CFI. Well, that was also my first thought when reading through the CFI patches, I hoped that might salvage my reduce-boilerplate-and-get-better-type-safety proposal for the devm_*_action APIs [1]. But I don't think that would help at all; storing __va_function(__static_call_return0) instead of &__static_call_return0 (i.e., __static_call_return0 instead of __static_call_return0.cfi_jt) doesn't help the call sites of that static_call at all, neither address belongs to the range of jump table entries corresponding to the prototype "int (*)(void)". So I think it would be the static_call macro that would somehow need to grow a way to suppress the cfi checking. Rasmus [1] https://lore.kernel.org/lkml/20210309235917.2134565-1-li...@rasmusvillemoes.dk/