X-Git-Url: https://git.lttng.org/?p=urcu.git;a=blobdiff_plain;f=include%2Furcu%2Farch%2Fppc.h;fp=include%2Furcu%2Farch%2Fppc.h;h=1285b6d2211f3239b850a73f8a3fb277a3516990;hp=be9d15857a8bf4d14027f783a38accfe7892621b;hb=5307e3abcb6668ceb84114e1c62b635cb995f4c7;hpb=8c5aef691745bc06750886e10bfda84f2b749ed8

diff --git a/include/urcu/arch/ppc.h b/include/urcu/arch/ppc.h
index be9d158..1285b6d 100644
--- a/include/urcu/arch/ppc.h
+++ b/include/urcu/arch/ppc.h
@@ -19,7 +19,30 @@
 extern "C" {
 #endif
 
-/* Include size of POWER5+ L3 cache lines: 256 bytes */
+/*
+ * Most powerpc machines have 128 bytes cache lines, but to make sure
+ * there is no false sharing on all known Power hardware, use the
+ * largest known cache line size, which is the physical size of POWER5
+ * L3 cache lines (256 bytes).
+ *
+ * "Each slice [of the L3] is 12-way set-associative, with 4,096
+ * congruence classes of 256-byte lines managed as two 128-byte sectors
+ * to match the L2 line size."
+ *
+ * From: "POWER5 system microarchitecture",
+ *       IBM Journal of Research & Development,
+ *       vol. 49, no. 4/5, July/September 2005
+ *       https://www.eecg.utoronto.ca/~moshovos/ACA08/readings/power5.pdf
+ *
+ * This value is a compile-time constant, which prevents us from
+ * querying the processor for the cache line size at runtime. We
+ * therefore need to be pessimistic and assume the largest known cache
+ * line size.
+ *
+ * This value is exposed through public headers, so tuning it for
+ * specific environments is a concern for ABI compatibility between
+ * applications and liburcu.
+ */
 #define CAA_CACHE_LINE_SIZE	256
 
 #ifdef __NO_LWSYNC__