x86, pat: Update the page flags for memtype atomically instead of using memtype_lock

While testing an application using the xpmem (out of kernel) driver, we
noticed a significant page fault rate reduction of x86_64 with respect
to ia64. For one test running with 32 cpus, one thread per cpu, it
took 01:08 for each of the threads to vm_insert_pfn 2GB worth of pages.
For the same test running on 256 cpus, one thread per cpu, it took 14:48
to vm_insert_pfn 2 GB worth of pages.

The slowdown was tracked to lookup_memtype which acquires the
spinlock memtype_lock. This heavily contended lock was slowing down
vm_insert_pfn().

With the cmpxchg on page->flags method, both the 32 cpu and 256 cpu
cases take approx 00:01.3 seconds to complete.

Signed-off-by: Robin Holt <[e-mail]>
LKML-Reference: <[e-mail]>
Cc: Venkatesh Pallipadi <[e-mail]>
Cc: Rafael Wysocki <[e-mail]>
Reviewed-by: Suresh Siddha <[e-mail]>
Signed-off-by: H. Peter Anvin <[e-mail]>

arch/x86/include/asm/cacheflush.h

···
 	memcpy(dst, src, len);
 }
 
-#define PG_WC				PG_arch_1
-PAGEFLAG(WC, WC)
-
 #ifdef CONFIG_X86_PAT
 /*
  * X86 PAT uses page flags WC and Uncached together to keep track of
···
  * _PAGE_CACHE_UC_MINUS and fourth state where page's memory type has not
  * been changed from its default (value of -1 used to denote this).
  * Note we do not support _PAGE_CACHE_UC here.
- *
- * Caller must hold memtype_lock for atomicity.
  */
+
+#define _PGMT_DEFAULT		0
+#define _PGMT_WC		(1UL << PG_arch_1)
+#define _PGMT_UC_MINUS		(1UL << PG_uncached)
+#define _PGMT_WB		(1UL << PG_uncached | 1UL << PG_arch_1)
+#define _PGMT_MASK		(1UL << PG_uncached | 1UL << PG_arch_1)
+#define _PGMT_CLEAR_MASK	(~_PGMT_MASK)
+
 static inline unsigned long get_page_memtype(struct page *pg)
 {
-	if (!PageUncached(pg) && !PageWC(pg))
+	unsigned long pg_flags = pg->flags & _PGMT_MASK;
+
+	if (pg_flags == _PGMT_DEFAULT)
 		return -1;
-	else if (!PageUncached(pg) && PageWC(pg))
+	else if (pg_flags == _PGMT_WC)
 		return _PAGE_CACHE_WC;
-	else if (PageUncached(pg) && !PageWC(pg))
+	else if (pg_flags == _PGMT_UC_MINUS)
 		return _PAGE_CACHE_UC_MINUS;
 	else
 		return _PAGE_CACHE_WB;
···
 
 static inline void set_page_memtype(struct page *pg, unsigned long memtype)
 {
+	unsigned long memtype_flags = _PGMT_DEFAULT;
+	unsigned long old_flags;
+	unsigned long new_flags;
+
 	switch (memtype) {
 	case _PAGE_CACHE_WC:
-		ClearPageUncached(pg);
-		SetPageWC(pg);
+		memtype_flags = _PGMT_WC;
 		break;
 	case _PAGE_CACHE_UC_MINUS:
-		SetPageUncached(pg);
-		ClearPageWC(pg);
+		memtype_flags = _PGMT_UC_MINUS;
 		break;
 	case _PAGE_CACHE_WB:
-		SetPageUncached(pg);
-		SetPageWC(pg);
-		break;
-	default:
-	case -1:
-		ClearPageUncached(pg);
-		ClearPageWC(pg);
+		memtype_flags = _PGMT_WB;
 		break;
 	}
+
+	do {
+		old_flags = pg->flags;
+		new_flags = (old_flags & _PGMT_CLEAR_MASK) | memtype_flags;
+	} while (cmpxchg(&pg->flags, old_flags, new_flags) != old_flags);
 }
 #else
 static inline unsigned long get_page_memtype(struct page *pg) { return -1; }

arch/x86/mm/pat.c

···
  * Here we do two pass:
  * - Find the memtype of all the pages in the range, look for any conflicts
  * - In case of no conflicts, set the new memtype for pages in the range
- *
- * Caller must hold memtype_lock for atomicity.
  */
 static int reserve_ram_pages_type(u64 start, u64 end, unsigned long req_type,
 				  unsigned long *new_type)
···
 	is_range_ram = pat_pagerange_is_ram(start, end);
 	if (is_range_ram == 1) {
 
-		spin_lock(&memtype_lock);
 		err = reserve_ram_pages_type(start, end, req_type, new_type);
-		spin_unlock(&memtype_lock);
 
 		return err;
 	} else if (is_range_ram < 0) {
···
 	is_range_ram = pat_pagerange_is_ram(start, end);
 	if (is_range_ram == 1) {
 
-		spin_lock(&memtype_lock);
 		err = free_ram_pages_type(start, end);
-		spin_unlock(&memtype_lock);
 
 		return err;
 	} else if (is_range_ram < 0) {
···
 
 	if (pat_pagerange_is_ram(paddr, paddr + PAGE_SIZE)) {
 		struct page *page;
-		spin_lock(&memtype_lock);
 		page = pfn_to_page(paddr >> PAGE_SHIFT);
 		rettype = get_page_memtype(page);
-		spin_unlock(&memtype_lock);
 		/*
 		 * -1 from get_page_memtype() implies RAM page is in its
 		 * default state and not reserved, and hence of type WB