Remove __GFP_NOFAIL in kmem and retry internally.

As of 2.6.31 it's clear __GFP_NOFAIL should no longer be used and it
may disappear from the kernel at any time.  To handle this I have simply
added *_nofail wrappers in the kmem implementation which perform the
retry for non-atomic allocations.

From linux-2.6.31 mm/page_alloc.c:1166
/*
 * __GFP_NOFAIL is not to be used in new code.
 *
 * All __GFP_NOFAIL callers should be fixed so that they
 * properly detect and handle allocation failures.
 *
 * We most definitely don't want callers attempting to
 * allocate greater than order-1 page units with
 * __GFP_NOFAIL.
 */
WARN_ON_ONCE(order > 1);

2 files changed, 59 insertions, 14 deletions

diff --git a/include/sys/kmem.h b/include/sys/kmem.h
index 3e5eb204f..fdeba70d1 100644
--- a/include/sys/kmem.h
+++ b/include/sys/kmem.h
@@ -49,7 +49,7 @@ extern "C" {
 /*
  * Memory allocation interfaces
  */
-#define KM_SLEEP                        (GFP_KERNEL | __GFP_NOFAIL)
+#define KM_SLEEP                        GFP_KERNEL
 #define KM_NOSLEEP                      GFP_ATOMIC
 #undef  KM_PANIC                        /* No linux analog */
 #define KM_PUSHPAGE                     (KM_SLEEP | __GFP_HIGH)
@@ -63,6 +63,51 @@ extern "C" {
 # define __GFP_ZERO                     0x8000
 #endif
 
+/*
+ * __GFP_NOFAIL looks like it will be removed from the kernel perhaps as
+ * early as 2.6.32.  To avoid this issue when it occurs in upstream kernels
+ * we retry the allocation here as long as it is not __GFP_WAIT (GFP_ATOMIC).
+ * I would prefer the caller handle the failure case cleanly but we are
+ * trying to emulate Solaris and those are not the Solaris semantics.
+ */
+static inline void *
+kmalloc_nofail(size_t size, gfp_t flags)
+{
+	void *ptr;
+
+	do {
+		ptr = kmalloc(size, flags);
+	} while (ptr == NULL && (flags & __GFP_WAIT));
+
+	return ptr;
+}
+
+static inline void *
+kzalloc_nofail(size_t size, gfp_t flags)
+{
+	void *ptr;
+
+	do {
+		ptr = kzalloc(size, flags);
+	} while (ptr == NULL && (flags & __GFP_WAIT));
+
+	return ptr;
+}
+
+#ifdef HAVE_KMALLOC_NODE
+static inline void *
+kmalloc_node_nofail(size_t size, gfp_t flags, int node)
+{
+	void *ptr;
+
+	do {
+		ptr = kmalloc_node(size, flags, node);
+	} while (ptr == NULL && (flags & __GFP_WAIT));
+
+	return ptr;
+}
+#endif /* HAVE_KMALLOC_NODE */
+
 #ifdef DEBUG_KMEM
 
 extern atomic64_t kmem_alloc_used;
@@ -125,16 +170,16 @@ extern void vmem_free_debug(void *ptr, size_t size);
 
 #else /* DEBUG_KMEM */
 
-# define kmem_alloc(size, flags)              kmalloc((size), (flags))
-# define kmem_zalloc(size, flags)             kzalloc((size), (flags))
+# define kmem_alloc(size, flags)              kmalloc_nofail((size), (flags))
+# define kmem_zalloc(size, flags)             kzalloc_nofail((size), (flags))
 # define kmem_free(ptr, size)                 ((void)(size), kfree(ptr))
 
 # ifdef HAVE_KMALLOC_NODE
 #  define kmem_alloc_node(size, flags, node)                                  \
-          kmalloc_node((size), (flags), (node))
+          kmalloc_node_nofail((size), (flags), (node))
 # else
 #  define kmem_alloc_node(size, flags, node)                                  \
-          kmalloc((size), (flags))
+          kmalloc_nofail((size), (flags))
 # endif
 
 # define vmem_alloc(size, flags)              __vmalloc((size), ((flags) |    \
diff --git a/module/spl/spl-kmem.c b/module/spl/spl-kmem.c
index 79a7028c2..438f7c6d3 100644
--- a/module/spl/spl-kmem.c
+++ b/module/spl/spl-kmem.c
@@ -380,7 +380,7 @@ kmem_alloc_track(size_t size, int flags, const char *func, int line,
 	unsigned long irq_flags;
 	ENTRY;
 
-	dptr = (kmem_debug_t *) kmalloc(sizeof(kmem_debug_t),
+	dptr = (kmem_debug_t *) kmalloc_nofail(sizeof(kmem_debug_t),
 	    flags & ~__GFP_ZERO);
 
 	if (dptr == NULL) {
@@ -409,11 +409,11 @@ kmem_alloc_track(size_t size, int flags, const char *func, int line,
 		/* Use the correct allocator */
 		if (node_alloc) {
 			ASSERT(!(flags & __GFP_ZERO));
-			ptr = kmalloc_node(size, flags, node);
+			ptr = kmalloc_node_nofail(size, flags, node);
 		} else if (flags & __GFP_ZERO) {
-			ptr = kzalloc(size, flags & ~__GFP_ZERO);
+			ptr = kzalloc_nofail(size, flags & ~__GFP_ZERO);
 		} else {
-			ptr = kmalloc(size, flags);
+			ptr = kmalloc_nofail(size, flags);
 		}
 
 		if (unlikely(ptr == NULL)) {
@@ -500,7 +500,7 @@ vmem_alloc_track(size_t size, int flags, const char *func, int line)
 
 	ASSERT(flags & KM_SLEEP);
 
-	dptr = (kmem_debug_t *) kmalloc(sizeof(kmem_debug_t), flags);
+	dptr = (kmem_debug_t *) kmalloc_nofail(sizeof(kmem_debug_t), flags);
 	if (dptr == NULL) {
 		CWARN("vmem_alloc(%ld, 0x%x) debug failed\n",
 		    sizeof(kmem_debug_t), flags);
@@ -614,11 +614,11 @@ kmem_alloc_debug(size_t size, int flags, const char *func, int line,
 	/* Use the correct allocator */
 	if (node_alloc) {
 		ASSERT(!(flags & __GFP_ZERO));
-		ptr = kmalloc_node(size, flags, node);
+		ptr = kmalloc_node_nofail(size, flags, node);
 	} else if (flags & __GFP_ZERO) {
-		ptr = kzalloc(size, flags & (~__GFP_ZERO));
+		ptr = kzalloc_nofail(size, flags & (~__GFP_ZERO));
 	} else {
-		ptr = kmalloc(size, flags);
+		ptr = kmalloc_nofail(size, flags);
 	}
 
 	if (ptr == NULL) {
@@ -1077,7 +1077,7 @@ spl_magazine_alloc(spl_kmem_cache_t *skc, int node)
 	           sizeof(void *) * skc->skc_mag_size;
 	ENTRY;
 
-	skm = kmem_alloc_node(size, GFP_KERNEL | __GFP_NOFAIL, node);
+	skm = kmem_alloc_node(size, KM_SLEEP, node);
 	if (skm) {
 		skm->skm_magic = SKM_MAGIC;
 		skm->skm_avail = 0;