Re: [PATCH v4 09/17] perf/core: Use static_call to optimize perf_guest_info_callbacks

From: Sean Christopherson <seanjc@google.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: linux-kernel@vger.kernel.org,
	linux-arm-kernel@lists.infradead.org,
	kvmarm@lists.cs.columbia.edu, kvm@vger.kernel.org,
	Will McVicker <willmcvicker@google.com>,
	Sami Tolvanen <samitolvanen@google.com>
Subject: Re: [PATCH v4 09/17] perf/core: Use static_call to optimize perf_guest_info_callbacks
Date: Wed, 2 Feb 2022 18:43:26 +0000	[thread overview]
Message-ID: <YfrQzoIWyv9lNljh@google.com> (raw)
In-Reply-To: <20211111020738.2512932-10-seanjc@google.com>

+Will and Sami, -most everyone else

On Thu, Nov 11, 2021, Sean Christopherson wrote:
> Use static_call to optimize perf's guest callbacks on arm64 and x86,
> which are now the only architectures that define the callbacks.  Use
> DEFINE_STATIC_CALL_RET0 as the default/NULL for all guest callbacks, as
> the callback semantics are that a return value '0' means "not in guest".
> 
> static_call obviously avoids the overhead of CONFIG_RETPOLINE=y, but is
> also advantageous versus other solutions, e.g. per-cpu callbacks, in that
> a per-cpu memory load is not needed to detect the !guest case.
> 
> Based on code from Peter and Like.
> 
> Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> Cc: Like Xu <like.xu.linux@gmail.com>
> Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
> Signed-off-by: Sean Christopherson <seanjc@google.com>
> ---

...

> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 1c8d341ecc77..b4fd928e4ff8 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -6524,12 +6524,23 @@ static void perf_pending_event(struct irq_work *entry)
>  #ifdef CONFIG_GUEST_PERF_EVENTS
>  struct perf_guest_info_callbacks __rcu *perf_guest_cbs;
>  
> +DEFINE_STATIC_CALL_RET0(__perf_guest_state, *perf_guest_cbs->state);
> +DEFINE_STATIC_CALL_RET0(__perf_guest_get_ip, *perf_guest_cbs->get_ip);
> +DEFINE_STATIC_CALL_RET0(__perf_guest_handle_intel_pt_intr, *perf_guest_cbs->handle_intel_pt_intr);

Using __static_call_return0() makes clang's CFI sad on arm64 due to the resulting
function prototype mistmatch, which IIUC, is verified by clang's __cfi_check()
for indirect calls, i.e. architectures without CONFIG_HAVE_STATIC_CALL.

We could fudge around the issue by using stubs, massaging prototypes, etc..., but
that means doing that for every arch-agnostic user of __static_call_return0().

Any clever ideas?  Can we do something like generate a unique function for every
DEFINE_STATIC_CALL_RET0 for CONFIG_HAVE_STATIC_CALL=n, e.g. using typeof() to
get the prototype?

  Kernel panic - not syncing: CFI failure (target: __static_call_return0+0x0/0x8)
  CPU: 0 PID: 1625 Comm: batterystats-wo Tainted: G        W  OE     5.16.0-mainline #1$
  Hardware name: Raven EVT 1.1 (DT)$
  Call trace:$
   dump_backtrace+0xf0/0x130$
   show_stack+0x1c/0x2c$
   dump_stack_lvl+0x68/0x98$
   panic+0x168/0x420$
   __cfi_check_fail+0x58/0x5c$
   __cfi_slowpath_diag+0x150/0x1a4$
   perf_misc_flags+0x74/0xa4$
   perf_prepare_sample+0x50/0x44c$
   perf_event_output_forward+0x5c/0xcc$
   __perf_event_overflow+0xc8/0x188$
   perf_swevent_event+0x7c/0x10c$
   perf_tp_event+0x168/0x298$
   perf_trace_run_bpf_submit+0x8c/0xdc$
   perf_trace_sched_switch+0x180/0x1cc$
   __schedule+0x850/0x924$
   schedule+0x98/0xe0$
   binder_wait_for_work+0x158/0x368$
   binder_thread_read+0x278/0x243c$
   binder_ioctl_write_read+0x120/0x45c$
   binder_ioctl+0x1ac/0xc34$
   __arm64_sys_ioctl+0xa8/0x118$
   invoke_syscall+0x64/0x178$
   el0_svc_common+0x8c/0x100$
   do_el0_svc+0x28/0xa0$
   el0_svc+0x24/0x84$
   el0t_64_sync_handler+0x88/0xec$
   el0t_64_sync+0x1b4/0x1b8$