btrfs-progs: fix unaligned accesses v2

PATCH In review Ben Peddell 2 participants 2 messages Jan 27, 2013 – Jan 28, 2013 Btrfs file system

lore

Share & tools

lore

Ben Peddell

author

2013-01-27 06:43 UTC #0 lore raw

PATCH 1/1 btrfs-progs: fix unaligned accesses v2

gcc optimizes out the memcpy calls at -O2 and -Os.

Replacing memcpy with memmove does't work - gcc treats memmove the same way it treats memcpy.

This patch brings in {get|put}_unaligned_le{16|32|64} (using the packed struct method), and uses them in the failing get/set calls.

On architectures where unaligned accesses are cheap, these unaligned macros should be optimized out by the compiler.

Signed-off-by Ben Peddell <[email protected]>

---
 ctree.h      |    8 ++------
 kerncompat.h |   13 +++++++++++++
 2 files changed, 15 insertions(+), 6 deletions(-)

ctree.h +2 -6

index 0675989..1f5a795 100644
@@ -1086,19 +1086,15 @@ static inline u##bits btrfs_##name(struct extent_buffer *eb,		\
 				   type *s)				\
 {									\
 	unsigned long offset = (unsigned long)s;			\
-	u##bits m;							\
 	type *p = (type *) (eb->data + offset);				\
-	memcpy(&m, &p->member, sizeof(m));				\
-	return le##bits##_to_cpu(m);					\
+	return get_unaligned_le##bits(&p->member);			\
 }									\
 static inline void btrfs_set_##name(struct extent_buffer *eb,		\
 				    type *s, u##bits val)		\
 {									\
 	unsigned long offset = (unsigned long)s;			\
-	u##bits m;							\
 	type *p = (type *) (eb->data + offset);				\
-	m = cpu_to_le##bits(val);					\
-	memcpy(&p->member, &m, sizeof(m));				\
+	put_unaligned_le##bits(val, &p->member);			\
 }
 
 #define BTRFS_SETGET_STACK_FUNCS(name, type, member, bits)		\

kerncompat.h +13 -0

index d60f722..a38a9b0 100644
@@ -267,6 +267,19 @@ typedef u64 __bitwise __be64;
 #define cpu_to_le16(x) ((__force __le16)(u16)(x))
 #define le16_to_cpu(x) ((__force u16)(__le16)(x))
 #endif
+
+struct __una_u16 { u16 x; } __attribute__((__packed__));
+struct __una_u32 { u32 x; } __attribute__((__packed__));
+struct __una_u64 { u64 x; } __attribute__((__packed__));
+
+#define get_unaligned_le8(p) (*((u8 *)(p)))
+#define put_unaligned_le8(val,p) ((*((u8 *)(p))) = (val))
+#define get_unaligned_le16(p) le16_to_cpu(((const struct __una_u16 *)(p))->x)
+#define put_unaligned_le16(val,p) (((struct __una_u16 *)(p))->x = cpu_to_le16(val))
+#define get_unaligned_le32(p) le32_to_cpu(((const struct __una_u32 *)(p))->x)
+#define put_unaligned_le32(val,p) (((struct __una_u32 *)(p))->x = cpu_to_le32(val))
+#define get_unaligned_le64(p) le64_to_cpu(((const struct __una_u64 *)(p))->x)
+#define put_unaligned_le64(val,p) (((struct __una_u64 *)(p))->x = cpu_to_le64(val))
 #endif
 
 #ifndef noinline

Signature · 1 line

1.7.3.4

David Sterba

2013-01-28 17:18 UTC #1 lore raw

Quoted 10 lines · Ben Peddell

gcc optimizes out the memcpy calls at -O2 and -Os.
Replacing memcpy with memmove does't work - gcc treats memmove the same way it treats memcpy.
This patch brings in {get|put}_unaligned_le{16|32|64} (using the packed struct method), and uses them in the failing get/set calls.
On architectures where unaligned accesses are cheap, these unaligned macros should be optimized out by the compiler.

Thanks, queued for integration. These tricky compilers ...

david