---

yaml
---
r: 311106
b: refs/heads/master
c: e4eed03
h: refs/heads/master
v: v3

[refs]

@@ -1,2 +1,2 @@
 ---
-refs/heads/master: abca7c4965845924f65d40e0aa1092bdd895e314
+refs/heads/master: e4eed03fd06578571c01d4f1478c874bb432c815

trunk/arch/x86/include/asm/pgtable-3level.h

@@ -47,16 +47,26 @@ static inline void native_set_pte(pte_t *ptep, pte_t pte)
  * they can run pmd_offset_map_lock or pmd_trans_huge or other pmd
  * operations.
  *
- * Without THP if the mmap_sem is hold for reading, the
- * pmd can only transition from null to not null while pmd_read_atomic runs.
- * So there's no need of literally reading it atomically.
+ * Without THP if the mmap_sem is hold for reading, the pmd can only
+ * transition from null to not null while pmd_read_atomic runs. So
+ * we can always return atomic pmd values with this function.
  *
  * With THP if the mmap_sem is hold for reading, the pmd can become
- * THP or null or point to a pte (and in turn become "stable") at any
- * time under pmd_read_atomic, so it's mandatory to read it atomically
- * with cmpxchg8b.
+ * trans_huge or none or point to a pte (and in turn become "stable")
+ * at any time under pmd_read_atomic. We could read it really
+ * atomically here with a atomic64_read for the THP enabled case (and
+ * it would be a whole lot simpler), but to avoid using cmpxchg8b we
+ * only return an atomic pmdval if the low part of the pmdval is later
+ * found stable (i.e. pointing to a pte). And we're returning a none
+ * pmdval if the low part of the pmd is none. In some cases the high
+ * and low part of the pmdval returned may not be consistent if THP is
+ * enabled (the low part may point to previously mapped hugepage,
+ * while the high part may point to a more recently mapped hugepage),
+ * but pmd_none_or_trans_huge_or_clear_bad() only needs the low part
+ * of the pmd to be read atomically to decide if the pmd is unstable
+ * or not, with the only exception of when the low part of the pmd is
+ * zero in which case we return a none pmd.
  */
-#ifndef CONFIG_TRANSPARENT_HUGEPAGE
 static inline pmd_t pmd_read_atomic(pmd_t *pmdp)
 {
 	pmdval_t ret;
@@ -74,12 +84,6 @@ static inline pmd_t pmd_read_atomic(pmd_t *pmdp)
 
 	return (pmd_t) { ret };
 }
-#else /* CONFIG_TRANSPARENT_HUGEPAGE */
-static inline pmd_t pmd_read_atomic(pmd_t *pmdp)
-{
-	return (pmd_t) { atomic64_read((atomic64_t *)pmdp) };
-}
-#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 
 static inline void native_set_pte_atomic(pte_t *ptep, pte_t pte)
 {

trunk/include/asm-generic/pgtable.h

@@ -484,6 +484,16 @@ static inline int pmd_none_or_trans_huge_or_clear_bad(pmd_t *pmd)
 	/*
 	 * The barrier will stabilize the pmdval in a register or on
 	 * the stack so that it will stop changing under the code.
+	 *
+	 * When CONFIG_TRANSPARENT_HUGEPAGE=y on x86 32bit PAE,
+	 * pmd_read_atomic is allowed to return a not atomic pmdval
+	 * (for example pointing to an hugepage that has never been
+	 * mapped in the pmd). The below checks will only care about
+	 * the low part of the pmd with 32bit PAE x86 anyway, with the
+	 * exception of pmd_none(). So the important thing is that if
+	 * the low part of the pmd is found null, the high part will
+	 * be also null or the pmd_none() check below would be
+	 * confused.
 	 */
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 	barrier();