• Sebastian Andrzej Siewior's avatar
    x86/percpu: Remove volatile from arch_raw_cpu_ptr(). · 1c1e7e3c
    Sebastian Andrzej Siewior authored
    The volatile attribute in the inline assembly of arch_raw_cpu_ptr()
    forces the compiler to always generate the code, even if the compiler
    can decide upfront that its result is not needed.
    
    For instance invoking __intel_pmu_disable_all(false) (like
    intel_pmu_snapshot_arch_branch_stack() does) leads to loading the
    address of &cpu_hw_events into the register while compiler knows that it
    has no need for it. This ends up with code like:
    
    |	movq	$cpu_hw_events, %rax			#, tcp_ptr__
    |	add	%gs:this_cpu_off(%rip), %rax		# this_cpu_off, tcp_ptr__
    |	xorl	%eax, %eax				# tmp93
    
    It also creates additional code within local_lock() with !RT &&
    !LOCKDEP which is not desired.
    
    By removing the volatile attribute the compiler can place the
    function freely and avoid it if it is not needed in the end.
    By using the function twice the compiler properly caches only the
    variable offset and always loads the CPU-offset.
    
    this_cpu_ptr() also remains properly placed within a preempt_disable()
    sections because
    - arch_raw_cpu_ptr() assembly has a memory input ("m" (this_cpu_off))
    - prempt_{dis,en}able() fundamentally has a 'barrier()' in it
    
    Therefore this_cpu_ptr() is already properly serialized and does not
    rely on the 'volatile' attribute.
    
    Remove volatile from arch_raw_cpu_ptr().
    
    [ bigeasy: Added Linus' explanation why this_cpu_ptr() is not moved out
      of a preempt_disable() section without the 'volatile' attribute. ]
    Suggested-by: default avatarLinus Torvalds <torvalds@linux-foundation.org>
    Signed-off-by: default avatarSebastian Andrzej Siewior <bigeasy@linutronix.de>
    Signed-off-by: default avatarPeter Zijlstra (Intel) <peterz@infradead.org>
    Link: https://lore.kernel.org/r/20220328145810.86783-2-bigeasy@linutronix.de
    1c1e7e3c
percpu.h 16.6 KB