tilegx pci: support I/O to arbitrarily-cached pages

The tilegx PCI root complex support (currently only in linux-next)
is limited to pages that are homed on cached in the default manner,
i.e. "hash-for-home".  This change supports delivery of I/O data to
pages that are cached in other ways (locally on a particular core,
uncached, user-managed incoherent, etc.).

A large part of the change is supporting flushing pages from cache
on particular homes so that we can transition the data that we are
delivering to or from the device appropriately.  The new homecache_finv*
routines handle this.

Some changes to page_table_range_init() were also required to make
the fixmap code work correctly on tilegx; it hadn't been used there
before.

We also remove some stub mark_caches_evicted_*() routines that
were just no-ops anyway.

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>

4 files changed, 36 insertions, 16 deletions

diff --git a/arch/tile/include/asm/cache.h b/arch/tile/include/asm/cache.h
index 392e5333dd8b..a9a529964e07 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
diff --git a/arch/tile/include/asm/fixmap.h b/arch/tile/include/asm/fixmap.h
index c66f7933beaa..e16dbf929cb5 100644
--- a/arch/tile/include/asm/fixmap.h
+++ b/arch/tile/include/asm/fixmap.h
@@ -45,15 +45,23 @@
  *
  * TLB entries of such buffers will not be flushed across
  * task switches.
- *
- * We don't bother with a FIX_HOLE since above the fixmaps
- * is unmapped memory in any case.
  */
 enum fixed_addresses {
+#ifdef __tilegx__
+	/*
+	 * TILEPro has unmapped memory above so the hole isn't needed,
+	 * and in any case the hole pushes us over a single 16MB pmd.
+	 */
+	FIX_HOLE,
+#endif
 #ifdef CONFIG_HIGHMEM
 	FIX_KMAP_BEGIN,	/* reserved pte's for temporary kernel mappings */
 	FIX_KMAP_END = FIX_KMAP_BEGIN+(KM_TYPE_NR*NR_CPUS)-1,
 #endif
+#ifdef __tilegx__  /* see homecache.c */
+	FIX_HOMECACHE_BEGIN,
+	FIX_HOMECACHE_END = FIX_HOMECACHE_BEGIN+(NR_CPUS)-1,
+#endif
 	__end_of_permanent_fixed_addresses,
 
 	/*
diff --git a/arch/tile/include/asm/homecache.h b/arch/tile/include/asm/homecache.h
index a8243865d49e..7b7771328642 100644
--- a/arch/tile/include/asm/homecache.h
+++ b/arch/tile/include/asm/homecache.h
@@ -79,10 +79,17 @@ extern void homecache_change_page_home(struct page *, int order, int home);
 /*
  * Flush a page out of whatever cache(s) it is in.
  * This is more than just finv, since it properly handles waiting
- * for the data to reach memory on tilepro, but it can be quite
- * heavyweight, particularly on hash-for-home memory.
+ * for the data to reach memory, but it can be quite
+ * heavyweight, particularly on incoherent or immutable memory.
  */
-extern void homecache_flush_cache(struct page *, int order);
+extern void homecache_finv_page(struct page *);
+
+/*
+ * Flush a page out of the specified home cache.
+ * Note that the specified home need not be the actual home of the page,
+ * as for example might be the case when coordinating with I/O devices.
+ */
+extern void homecache_finv_map_page(struct page *, int home);
 
 /*
  * Allocate a page with the given GFP flags, home, and optionally
@@ -104,10 +111,10 @@ extern struct page *homecache_alloc_pages_node(int nid, gfp_t gfp_mask,
  * routines use homecache_change_page_home() to reset the home
  * back to the default before returning the page to the allocator.
  */
+void __homecache_free_pages(struct page *, unsigned int order);
 void homecache_free_pages(unsigned long addr, unsigned int order);
-#define homecache_free_page(page) \
-  homecache_free_pages((page), 0)
-
+#define __homecache_free_page(page) __homecache_free_pages((page), 0)
+#define homecache_free_page(page) homecache_free_pages((page), 0)
 
 
 /*
diff --git a/arch/tile/include/asm/page.h b/arch/tile/include/asm/page.h
index 9d9131e5c552..dd033a4fd627 100644
--- a/arch/tile/include/asm/page.h
+++ b/arch/tile/include/asm/page.h
@@ -174,7 +174,9 @@ static inline __attribute_const__ int get_order(unsigned long size)
 #define MEM_LOW_END		(HALF_VA_SPACE - 1)         /* low half */
 #define MEM_HIGH_START		(-HALF_VA_SPACE)            /* high half */
 #define PAGE_OFFSET		MEM_HIGH_START
-#define _VMALLOC_START		_AC(0xfffffff500000000, UL) /* 4 GB */
+#define FIXADDR_BASE		_AC(0xfffffff400000000, UL) /* 4 GB */
+#define FIXADDR_TOP		_AC(0xfffffff500000000, UL) /* 4 GB */
+#define _VMALLOC_START		FIXADDR_TOP
 #define HUGE_VMAP_BASE		_AC(0xfffffff600000000, UL) /* 4 GB */
 #define MEM_SV_START		_AC(0xfffffff700000000, UL) /* 256 MB */
 #define MEM_SV_INTRPT		MEM_SV_START
@@ -185,9 +187,6 @@ static inline __attribute_const__ int get_order(unsigned long size)
 /* Highest DTLB address we will use */
 #define KERNEL_HIGH_VADDR	MEM_SV_START
 
-/* Since we don't currently provide any fixmaps, we use an impossible VA. */
-#define FIXADDR_TOP             MEM_HV_START
-
 #else /* !__tilegx__ */
 
 /*