Skip to contentSkip to conversation

bcachefs: omit alignment attribute on big endian struct bkey

PATCH Superseded Thomas Bertschinger 3 participants 4 messages Feb 15, 2024 – Feb 16, 2024

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 bcachefs: omit alignment attribute on big endian struct bkey

This is needed for building Rust bindings on big endian architectures like s390x. Currently this is only done in userspace, but it might happen in-kernel in the future. When creating a Rust binding for struct bkey, the "packed" attribute is needed to get a type with the correct member offsets. However, rustc does not allow types to have both a "packed" and "align" attribute. Thus, in order to get a Rust type compatible with the C type, on big endian systems we must omit the "aligned" attribute in C.

This does not affect the struct's size or member offsets, only its toplevel alignment, which should be an acceptable impact.

---
 fs/bcachefs/bcachefs_format.h | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
fs/bcachefs/bcachefs_format.h +5 -1
index 1bb24aa73528..fc12efe8f8cd 100644
@@ -222,7 +222,11 @@ struct bkey {
 
 	__u8		pad[1];
 #endif
-} __packed __aligned(8);
+} __packed
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+__aligned(8)
+#endif
+;
 
 struct bkey_packed {
 	__u64		_data[0];
Signature · 1 line
2.43.0
#1 lore raw
Quoted 22 lines · Thomas Bertschinger

This is needed for building Rust bindings on big endian architectures like s390x. Currently this is only done in userspace, but it might happen in-kernel in the future. When creating a Rust binding for struct bkey, the "packed" attribute is needed to get a type with the correct member offsets. However, rustc does not allow types to have both a "packed" and "align" attribute. Thus, in order to get a Rust type compatible with the C type, on big endian systems we must omit the "aligned" attribute in C.

This does not affect the struct's size or member offsets, only its toplevel alignment, which should be an acceptable impact.

---
 fs/bcachefs/bcachefs_format.h | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
fs/bcachefs/bcachefs_format.h +0 -0
index 1bb24aa73528..fc12efe8f8cd 100644

fs/bcachefs/bcachefs_format.h struct bkey { · line 225

Quoted 5 lines
fs/bcachefs/bcachefs_format.h +1 -1
@@ -222,7 +222,11 @@ struct bkey {
 
 	__u8		pad[1];
 #endif
-} __packed __aligned(8);
+} __packed

I can't really comment on the Rust context, but could you add a comment to explain why the ifdef is here? I.e., even something as simple as "Rust disallows packed and aligned on ..." is useful.

Also out of curiosity, is there some reason these two attributes together presumably isn't a problem for LE?

Brian

Quoted 10 lines
fs/bcachefs/bcachefs_format.h +4 -0
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+__aligned(8)
+#endif
+;
 
 struct bkey_packed {
 	__u64		_data[0];
Signature · 1 line
2.43.0
#2 lore raw
Quoted 5 lines · Brian Foster

I can't really comment on the Rust context, but could you add a comment to explain why the ifdef is here? I.e., even something as simple as "Rust disallows packed and aligned on ..." is useful.

Brian

I'll work on a v2 with an explanatory comment.

Quoted 2 lines

Also out of curiosity, is there some reason these two attributes together presumably isn't a problem for LE?

The root of the problem is that rustc won't compile types with both packed and align attributes. However, rust-bindgen tries to be helpful and when generating Rust bindings for C types with both attrs, will avoid placing one or the other attribute if it can do so without changing the type's layout.

For LE, the "packed" attr is unneeded because the struct is "naturally packed", but the "align(8)" attr is needed because the struct's natural alignment is 4.

For BE, the "packed" attr is needed because there are some members that are not aligned (e.g., size has as offset of 23 but wants to be at 24), so rust-bindgen cannot omit it. Thus we have to remove "align(8)" to get a type that rustc can compile.

- Thomas Bertschinger

#3 lore raw
Quoted 18 lines · Thomas Bertschinger
Quoted 5 lines · Brian Foster

I can't really comment on the Rust context, but could you add a comment to explain why the ifdef is here? I.e., even something as simple as "Rust disallows packed and aligned on ..." is useful.

Brian

I'll work on a v2 with an explanatory comment.

Quoted 2 lines

Also out of curiosity, is there some reason these two attributes together presumably isn't a problem for LE?

The root of the problem is that rustc won't compile types with both packed and align attributes. However, rust-bindgen tries to be helpful and when generating Rust bindings for C types with both attrs, will avoid placing one or the other attribute if it can do so without changing the type's layout.

Quoted 8 lines

For LE, the "packed" attr is unneeded because the struct is "naturally packed", but the "align(8)" attr is needed because the struct's natural alignment is 4.

For BE, the "packed" attr is needed because there are some members that are not aligned (e.g., size has as offset of 23 but wants to be at 24), so rust-bindgen cannot omit it. Thus we have to remove "align(8)" to get a type that rustc can compile.

Specifically, when i was designing bkey, I wanted the header to be no bigger than necessary so that bkey_packed could use the rest. That means that decently offten extent keys will fit into only 8 bytes, instead of spilling over to 16.

But packed_bkey treats the part after the header - the packed section - as a single multi word, variable length integer. And bkey, the unpacked version, is just a special case version of a bkey_packed; all the packed bkey code will work on keys in any packed format, the in-memory representation of an unpacked key also is just one type of packed key...

So that constrains the key part of a bkig endian bkey to start right after the header...

So if we ever do a bkey_v2 and need to expand the hedaer by another byte for some reason, that would clean up that wart :)