[ Upstream commit 87c3a5893e865739ce78aa7192d36011022e0af7 ]
Except on x86, preempt_count is always accessed with READ_ONCE().
Repeated invocations in macros like irq_count() produce repeated loads.
These redundant instructions appear in various fast paths. In the one
shown below, for example, irq_count() is evaluated during kernel entry
if !tick_nohz_full_cpu(smp_processor_id()).
0001ed0a <irq_enter_rcu>:
1ed0a: 4e56 0000 linkw %fp,#0
1ed0e: 200f movel %sp,%d0
1ed10: 0280 ffff e000 andil #-8192,%d0
1ed16: 2040 moveal %d0,%a0
1ed18: 2028 0008 movel %a0@(8),%d0
1ed1c: 0680 0001 0000 addil #65536,%d0
1ed22: 2140 0008 movel %d0,%a0@(8)
1ed26: 082a 0001 000f btst #1,%a2@(15)
1ed2c: 670c beqs 1ed3a <irq_enter_rcu+0x30>
1ed2e: 2028 0008 movel %a0@(8),%d0
1ed32: 2028 0008 movel %a0@(8),%d0
1ed36: 2028 0008 movel %a0@(8),%d0
1ed3a: 4e5e unlk %fp
1ed3c: 4e75 rts
This patch doesn't prevent the pointless btst and beqs instructions
above, but it does eliminate 2 of the 3 pointless move instructions
here and elsewhere.
On x86, preempt_count is per-cpu data and the problem does not arise
presumably because the compiler is free to optimize more effectively.
This patch was tested on m68k and x86. I was expecting no changes
to object code for x86 and mostly that's what I saw. However, there
were a few places where code generation was perturbed for some reason.
The performance issue addressed here is minor on uniprocessor m68k. I
got a 0.01% improvement from this patch for a simple "find /sys -false"
benchmark. For architectures and workloads susceptible to cache line bounce
the improvement is expected to be larger. The only SMP architecture I have
is x86, and as x86 unaffected I have not done any further measurements.
Fixes: 15115830c8 ("preempt: Cleanup the macro maze a bit")
Signed-off-by: Finn Thain <fthain@linux-m68k.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/0a403120a682a525e6db2d81d1a3ffcc137c3742.1694756831.git.fthain@linux-m68k.org
Signed-off-by: Sasha Levin <sashal@kernel.org>