aboutsummaryrefslogtreecommitdiff
path: root/arch/tile/include/asm/cache.h
diff options
context:
space:
mode:
Diffstat (limited to 'arch/tile/include/asm/cache.h')
-rw-r--r--arch/tile/include/asm/cache.h25
1 files changed, 19 insertions, 6 deletions
diff --git a/arch/tile/include/asm/cache.h b/arch/tile/include/asm/cache.h
index 392e5333dd8..6160761d5f6 100644
--- a/arch/tile/include/asm/cache.h
+++ b/arch/tile/include/asm/cache.h
@@ -27,11 +27,17 @@
#define L2_CACHE_ALIGN(x) (((x)+(L2_CACHE_BYTES-1)) & -L2_CACHE_BYTES)
/*
- * TILE-Gx is fully coherent so we don't need to define ARCH_DMA_MINALIGN.
+ * TILEPro I/O is not always coherent (networking typically uses coherent
+ * I/O, but PCI traffic does not) and setting ARCH_DMA_MINALIGN to the
+ * L2 cacheline size helps ensure that kernel heap allocations are aligned.
+ * TILE-Gx I/O is always coherent when used on hash-for-home pages.
+ *
+ * However, it's possible at runtime to request not to use hash-for-home
+ * for the kernel heap, in which case the kernel will use flush-and-inval
+ * to manage coherence. As a result, we use L2_CACHE_BYTES for the
+ * DMA minimum alignment to avoid false sharing in the kernel heap.
*/
-#ifndef __tilegx__
#define ARCH_DMA_MINALIGN L2_CACHE_BYTES
-#endif
/* use the cache line size for the L2, which is where it counts */
#define SMP_CACHE_BYTES_SHIFT L2_CACHE_SHIFT
@@ -43,9 +49,16 @@
#define __read_mostly __attribute__((__section__(".data..read_mostly")))
/*
- * Attribute for data that is kept read/write coherent until the end of
- * initialization, then bumped to read/only incoherent for performance.
+ * Originally we used small TLB pages for kernel data and grouped some
+ * things together as "write once", enforcing the property at the end
+ * of initialization by making those pages read-only and non-coherent.
+ * This allowed better cache utilization since cache inclusion did not
+ * need to be maintained. However, to do this requires an extra TLB
+ * entry, which on balance is more of a performance hit than the
+ * non-coherence is a performance gain, so we now just make "read
+ * mostly" and "write once" be synonyms. We keep the attribute
+ * separate in case we change our minds at a future date.
*/
-#define __write_once __attribute__((__section__(".w1data")))
+#define __write_once __read_mostly
#endif /* _ASM_TILE_CACHE_H */