[PATCH] [25/31] CPA: Always do full TLB flush when splitting large pages

[Andi Kleen]

Various CPUs have errata when using INVLPG to flush large pages.
This includes Intel Penryn (AV2) and AMD K7 (#16 in Athlon 4) 
While these happen only in specific circumstances it is still
a little risky and it's not clear the kernel can avoid them all.

Avoid this can of worms by always flushing the full TLB (but 
not the full cache) when splitting a large page. This should
not be that expensive anyways and initial splitting should be
hopefully infrequent.

This also removes the previously hard coded workaround for K7 
Athlon on 32bit.

Signed-off-by: Andi Kleen <ak@xxxxxxx>

---
 arch/x86/mm/pageattr_32.c |   17 +++++++++++------
 arch/x86/mm/pageattr_64.c |   12 ++++++++++--
 2 files changed, 21 insertions(+), 8 deletions(-)

Index: linux/arch/x86/mm/pageattr_32.c
===================================================================
--- linux.orig/arch/x86/mm/pageattr_32.c
+++ linux/arch/x86/mm/pageattr_32.c
@@ -113,10 +113,7 @@ static void flush_kernel_map(void *arg)
 		}
 	}
 
-	/* Handle errata with flushing large pages in early Athlons */
-	if (a->full_flush ||
-	    (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
-	     boot_cpu_data.x86 == 7))
+	if (a->full_flush)
 		__flush_tlb_all();
 
 	/*
@@ -198,7 +195,7 @@ static int cache_attr_changed(pte_t pte,
  * data structure to keep track of the flush. This has the added bonus that
  * it will work for MMIO holes without mem_map too.
  */
-static void set_tlb_flush(unsigned long address, int cache)
+static void set_tlb_flush(unsigned long address, int cache, int large)
 {
 	enum flush_mode mode = cache ? FLUSH_CACHE : FLUSH_TLB;
 	struct flush *f = kmalloc(sizeof(struct flush), GFP_KERNEL);
@@ -210,6 +207,13 @@ static void set_tlb_flush(unsigned long 
 	f->addr = address;
 	f->mode = mode;
 	list_add_tail(&f->l, &flush_pages);
+
+	/*
+	 * Work around large page INVLPG bugs in early K7 and in Penryn.
+	 * When we split a large page always do a full TLB flush.
+	 */
+	if (large && full_flush < FLUSH_TLB)
+		full_flush = FLUSH_TLB;
 }
 
 static int
@@ -230,7 +234,8 @@ __change_page_attr(struct page *page, pg
 	BUG_ON(PageLRU(kpte_page));
 	BUG_ON(PageCompound(kpte_page));
 
-	set_tlb_flush(address, cache_attr_changed(*kpte, prot, level));
+	set_tlb_flush(address, cache_attr_changed(*kpte, prot, level),
+			level < 3);
 
 	if (pgprot_val(prot) != pgprot_val(PAGE_KERNEL)) { 
 		if (level == 3) {
Index: linux/arch/x86/mm/pageattr_64.c
===================================================================
--- linux.orig/arch/x86/mm/pageattr_64.c
+++ linux/arch/x86/mm/pageattr_64.c
@@ -164,7 +164,7 @@ static void revert_page(unsigned long ad
  * data structure to keep track of the flush. This has the added bonus that
  * it will work for MMIO holes without mem_map too.
  */
-static void set_tlb_flush(unsigned long address, int cache)
+static void set_tlb_flush(unsigned long address, int cache, int large)
 {
 	enum flush_mode mode = cache ? FLUSH_CACHE : FLUSH_TLB;
 	struct flush *f = kmalloc(sizeof(struct flush), GFP_KERNEL);
@@ -177,6 +177,13 @@ static void set_tlb_flush(unsigned long 
 	f->addr = address;
 	f->mode = mode;
 	list_add_tail(&f->l, &flush_pages);
+
+	/*
+	 * Work around large page INVLPG bugs in early K7 and in Penryn.
+	 * When we split a large page always do a full TLB flush.
+	 */
+	if (large && full_flush < FLUSH_TLB)
+		full_flush = FLUSH_TLB;
 }
 
 static unsigned short pat_bit[5] = {
@@ -208,7 +215,8 @@ __change_page_attr(unsigned long address
 	BUG_ON(PageCompound(kpte_page));
 	BUG_ON(PageLRU(kpte_page));
 
-	set_tlb_flush(address, cache_attr_changed(*kpte, prot, level));
+	set_tlb_flush(address, cache_attr_changed(*kpte, prot, level),
+			level < 4);
 
 	if (pgprot_val(prot) != pgprot_val(ref_prot)) { 
 		if (level == 4) {
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/