[PATCH] MIPS: lib: prefetch ahead in csum_partial()

[Rosen Penev]

csum_partial() reads every byte of the buffer once. Cores like the 74K
have no hardware prefetcher, so every cache line is a demand miss, and
after a non-coherent DMA the received data is never in the cache. With
GRO enabled, validating the checksum of forwarded TCP traffic made
csum_partial() the top entry in the profile on a BCM4716 (74Kc at
480 MHz, no L2, ~80 cycles per DRAM miss): 13-16% of all cycles.

Prefetch the 128 byte block two iterations ahead in the main loop.
Unlike memcpy(), which disables prefetching on DMA_NONCOHERENT
entirely, only prefetch blocks that lie inside the buffer: the 74K does
not need a post-DMA invalidate, so a line prefetched past the end could
belong to a buffer the device is still writing, and the CPU would later
read the stale copy.

On an Asus RT-N16 (BCM4716, bgmac, single TCP stream unless noted,
iperf3 median of 3, A/B in the same boot by toggling the prefetch,
two rounds each):

                       without      with      (Mbit/s)
  local rx / tx        220 / 194    243 / 212
  routed up / down     169 / 140    187 / 154
  routed, 4 streams    65 / 79      69 / 84
  flow offload bidir   253          292

Assisted-by: LLM
Signed-off-by: Rosen Penev <rosenp@xxxxxxxxx>
---
 arch/mips/lib/csum_partial.S | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S
index 3d2ff4118d79..f0ec47259ef4 100644
--- a/arch/mips/lib/csum_partial.S
+++ b/arch/mips/lib/csum_partial.S
@@ -190,6 +190,21 @@ EXPORT_SYMBOL(csum_partial)
 	 andi	t2, a1, 0x40
 
 .Lmove_128bytes:
+#ifdef CONFIG_CPU_HAS_PREFETCH
+	/*
+	 * Fetch the block two iterations ahead, but only while it lies
+	 * inside the buffer: on non-coherent systems a line prefetched
+	 * past the end could belong to a buffer the device still owns.
+	 */
+	sltiu	t5, t8, 3
+	bnez	t5, 2f
+	 nop
+	pref	0, 0x100(src)
+	pref	0, 0x120(src)
+	pref	0, 0x140(src)
+	pref	0, 0x160(src)
+2:
+#endif
 	CSUM_BIGCHUNK(src, 0x00, sum, t0, t1, t3, t4)
 	CSUM_BIGCHUNK(src, 0x20, sum, t0, t1, t3, t4)
 	CSUM_BIGCHUNK(src, 0x40, sum, t0, t1, t3, t4)
-- 
2.56.0