On 02/24/2010 07:53 AM, Ralf Baechle wrote:
On Thu, Feb 04, 2010 at 11:31:49AM -0800, David Daney wrote:
The current locking mechanism uses a ll/sc sequence to release a
spinlock. This is slower than a wmb() followed by a store to unlock.
The branching forward to .subsection 2 on sc failure slows down the
contended case. So we get rid of that part too.
Since we are now working on naturally aligned u16 values, we can get
rid of a masking operation as the LHU already does the right thing.
The ANDI are reversed for better scheduling on multi-issue CPUs
On a 12 CPU 750MHz Octeon cn5750 this patch improves ipv4 UDP packet
forwarding rates from 3.58*10^6 PPS to 3.99*10^6 PPS, or about 11%.
And in your benchmarking patch you wrote:
spin_single spin_multi
base 106885 247941
spinlock_patch 75194 219465
I did some benchmarking on an IP27 (180MHz, 2 CPU, needs LL/SC workaround):
spin_single spin_multi
base 229341 3505690
spinlock_patch 177847 3615326
So about 22% speedup for spin_single but 3% slowdown for spin_multi.
It is possible that by choosing a better nudge_writes() implementation
for R10K, that the 3% degradation could be erased. Perhaps:
#define nudge_writes() do { } while (0)
Basically you want something that is fast, but that also forces the
write to be globally visible as soon as possible. Some processors have
a prefetch instruction that does this. On other processors a NOP is
optimal as they don't combine writes in the write back buffer.
There is a wbflush() function that could potentially be used, but its
implementation is too heavy on Octeon.
Disabling the R10k LL/SC workaround btw. gives another 23% speedup for
spin_single and marginal 0.3% for spin_multi; the latter may well be
statistical noise.
Ralf
