Kbuild: enable -fms-extensions, make btrfs the first user
Subsystems (3)
cover Kbuild: enable -fms-extensions, make btrfs the first user
Since -fms-extensions once again came up as potentially useful, Linus suggested that we bite the bullet and enable it.
https://lore.kernel.org/lkml/CAHk-=wjeZwww6Zswn6F_iZTpUihTSNKYppLqj36iQDDhfntuEw@mail.gmail.com/
So that's what patch 1 does, and patch 2 puts it to use in the btrfs case.
Compile-tested only, with gcc (15.2.1) and clang (20.1.8).
Rasmus Villemoes (2):
Kbuild: enable -fms-extensions
btrfs: send: make use of -fms-extensions for defining struct fs_path
Makefile | 9 +++++++++ fs/btrfs/send.c | 39 ++++++++++++++++++++------------------- 2 files changed, 29 insertions(+), 19 deletions(-)
Signature · 1 line
2.51.0
PATCH 1/2 Kbuild: enable -fms-extensions
Once in a while, it turns out that enabling -fms-extensions could allow some slightly prettier code. But every time it has come up, the code that had to be used instead has been deemed "not too awful" and not worth introducing another compiler flag for.
That's probably true for each individual case, but then it's somewhat of a chicken/egg situation.
If we just "bite the bullet" as Linus says and enable it once and for all, it is available whenever a use case turns up, and no individual case has to justify it.
A lore.kernel.org search provides these examples:
- https://lore.kernel.org/lkml/[email protected]/
- https://lore.kernel.org/lkml/[email protected]/
- https://lore.kernel.org/lkml/[email protected]/
- https://lore.kernel.org/lkml/[email protected]/
- https://lore.kernel.org/lkml/CAHk-=wjeZwww6Zswn6F_iZTpUihTSNKYppLqj36iQDDhfntuEw@mail.gmail.com/
Undoubtedly, there are more places in the code where this could also be used but where -fms-extensions just didn't come up in any discussion.
--- Makefile | 9 +++++++++ 1 file changed, 9 insertions(+)
Makefile +9 -0
@@ -1061,6 +1061,15 @@ NOSTDINC_FLAGS += -nostdinc # perform bounds checking. KBUILD_CFLAGS += $(call cc-option, -fstrict-flex-arrays=3) +# Allow including a tagged struct or union anonymously in another struct/union. +KBUILD_CFLAGS += -fms-extensions + +# For clang, the -fms-extensions flag is apparently not enough to +# express one's intention to make use of those extensions. +ifdef CONFIG_CC_IS_CLANG +KBUILD_CFLAGS += -Wno-microsoft-anon-tag +endif + # disable invalid "can't wrap" optimizations for signed / pointers KBUILD_CFLAGS += -fno-strict-overflow
Signature · 1 line
2.51.0
Nathan Chancellor
maintainerRe: [PATCH 1/2] Kbuild: enable -fms-extensions
Quoted 35 lines · Rasmus Villemoes
Once in a while, it turns out that enabling -fms-extensions could allow some slightly prettier code. But every time it has come up, the code that had to be used instead has been deemed "not too awful" and not worth introducing another compiler flag for.
That's probably true for each individual case, but then it's somewhat of a chicken/egg situation.
If we just "bite the bullet" as Linus says and enable it once and for all, it is available whenever a use case turns up, and no individual case has to justify it.
A lore.kernel.org search provides these examples:
- https://lore.kernel.org/lkml/[email protected]/
- https://lore.kernel.org/lkml/[email protected]/
- https://lore.kernel.org/lkml/[email protected]/
- https://lore.kernel.org/lkml/[email protected]/
- https://lore.kernel.org/lkml/CAHk-=wjeZwww6Zswn6F_iZTpUihTSNKYppLqj36iQDDhfntuEw@mail.gmail.com/Undoubtedly, there are more places in the code where this could also be used but where -fms-extensions just didn't come up in any discussion.
Signed-off-by Rasmus Villemoes <[email protected]>--- Makefile | 9 +++++++++ 1 file changed, 9 insertions(+)Makefile +0
-0@@ -1061,6 +1061,15 @@ NOSTDINC_FLAGS += -nostdinc # perform bounds checking.
Makefile NOSTDINC_FLAGS += -nostdinc · line 1071
Quoted 10 lines
Makefile +8
-0KBUILD_CFLAGS += $(call cc-option, -fstrict-flex-arrays=3) +# Allow including a tagged struct or union anonymously in another struct/union. +KBUILD_CFLAGS += -fms-extensions + +# For clang, the -fms-extensions flag is apparently not enough to +# express one's intention to make use of those extensions. +ifdef CONFIG_CC_IS_CLANG +KBUILD_CFLAGS += -Wno-microsoft-anon-tag +endif
I think this should go in the first 'ifdef CONFIG_CC_IS_CLANG' block in scripts/Makefile.extrawarn below '-Wno-gnu' with a comment that is similar in nature, which could even be combined like
# The kernel builds with '-std-gnu11' and '-fms-extensions' so the use
# of GNU and Microsoft extensions is acceptable.
Other than that, this seems fine to me.
Quoted 6 lines
Makefile +0
-0# disable invalid "can't wrap" optimizations for signed / pointers KBUILD_CFLAGS += -fno-strict-overflowSignature · 1 line
2.51.0
Re: [PATCH 1/2] Kbuild: enable -fms-extensions
Quoted 16 lines · Nathan Chancellor
Makefile NOSTDINC_FLAGS += -nostdinc · line 1071
Quoted 8 lines · Rasmus Villemoes
Makefile +8
-0+# Allow including a tagged struct or union anonymously in another struct/union. +KBUILD_CFLAGS += -fms-extensions + +# For clang, the -fms-extensions flag is apparently not enough to +# express one's intention to make use of those extensions. +ifdef CONFIG_CC_IS_CLANG +KBUILD_CFLAGS += -Wno-microsoft-anon-tag +endifI think this should go in the first 'ifdef
CONFIG_CC_IS_CLANG' block inscripts/Makefile.extrawarnbelow '-Wno-gnu' with a comment that is similar in nature, which could even be combined like
Quoted 4 lines
# The kernel builds with '-std-gnu11' and '-fms-extensions' so the use
# of GNU and Microsoft extensions is acceptable.Other than that, this seems fine to me.
I honestly had no idea where it was best to put these, and your suggestion sounds quite reasonable. I didn't think to look in that Makefile.extrawarn as the name suggested that was only about what happens with W=1 and higher.
Feel free to tweak when applying.
Rasmus
Nathan Chancellor
maintainerRe: [PATCH 1/2] Kbuild: enable -fms-extensions
Quoted 20 lines · Rasmus Villemoes
Quoted 19 lines · Nathan Chancellor
Makefile NOSTDINC_FLAGS += -nostdinc · line 1071
Quoted 8 lines · Rasmus Villemoes
Makefile +8
-0+# Allow including a tagged struct or union anonymously in another struct/union. +KBUILD_CFLAGS += -fms-extensions + +# For clang, the -fms-extensions flag is apparently not enough to +# express one's intention to make use of those extensions. +ifdef CONFIG_CC_IS_CLANG +KBUILD_CFLAGS += -Wno-microsoft-anon-tag +endifI think this should go in the first 'ifdef
CONFIG_CC_IS_CLANG' block inscripts/Makefile.extrawarnbelow '-Wno-gnu' with a comment that is similar in nature, which could even be combined like# The kernel builds with '-std-gnu11' and '-fms-extensions' so the use
# of GNU and Microsoft extensions is acceptable.
Quoted 6 lines
Quoted 1 line
Other than that, this seems fine to me.
I honestly had no idea where it was best to put these, and your suggestion sounds quite reasonable. I didn't think to look in that Makefile.extrawarn as the name suggested that was only about what happens with W=1 and higher.
Yeah, we may want to rename that to just Makefile.warn since it encompasses all warnings since commit e88ca24319e4 ("kbuild: consolidate warning flags in scripts/Makefile.extrawarn"), which is actually feedback I gave on the original change:
https://lore.kernel.org/[email protected]/
I'll send a patch for that soon.
Quoted 1 line
Feel free to tweak when applying.
I have tentatively applied this to kbuild-next so that it can spend most of the cycle in -next to try and catch all potential problems.
Cheers,
Nathan
Nathan Chancellor
maintainerRe: [PATCH 1/2] Kbuild: enable -fms-extensions
On Wed, Oct 22, 2025 at 11:11:38PM +0200, Nathan Chancellor wrote:
...
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
(quoted diff) +2
-0+# Allow including a tagged struct or union anonymously in another struct/union. +KBUILD_CFLAGS += -fms-extensions
...
Quoted 2 lines
I have tentatively applied this to kbuild-next so that it can spend most of the cycle in -next to try and catch all potential problems.
One side effect that has been found in my testing so far is clang's '-fms-extensions' turns '_inline' into a keyword, which breaks fs/jfs:
In file included from fs/jfs/jfs_unicode.c:8:
fs/jfs/jfs_incore.h:86:13: error: type name does not allow function specifier to be specified
86 | unchar _inline[128];
| ^
fs/jfs/jfs_incore.h:86:20: error: expected member name or ';' after declaration specifiers
86 | unchar _inline[128];
| ~~~~~~~~~~~~~~^There appear to be other similar keywords (ones with KEYMS in the linke below) but my personal distribution configuration does not show any instances in the build where they matter (I did not test allmodconfig yet).
Something like this is all it takes to resolve the issue, so I will send a patch for formal review/acking but I wanted to bring it up ahead of time in case this is unpalpable and we should throw these changes out of -next instead of forward fixing.
Cheers,
Nathan
fs/jfs/jfs_incore.h +3 -3
@@ -76,14 +76,14 @@ struct jfs_inode_info {struct {unchar _unused[16 ];/* 16: */ dxd_t _dxd;/* 16: */ -/* _inline may overflow into _inline_ea when needed */ +/* _inline_sym may overflow into _inline_ea when needed */ /* _inline_ea may overlay the last part of * file._xtroot if maxentry = XTROOTINITSLOT */ union {struct {/* 128: inline symlink */ -unchar _inline[128 ]; +unchar _inline_sym[128 ];/* 128: inline extended attr */ unchar _inline_ea[128 ]; };
@@ -101,7 +101,7 @@ struct jfs_inode_info {#define i_imap u.file._imap#define i_dirtable u.dir._table#define i_dtroot u.dir._dtroot -#define i_inline u.link._inline +#define i_inline u.link._inline_sym#define i_inline_ea u.link._inline_ea#define i_inline_all u.link._inline_all
Dave Kleikamp
Acked-byRe: [PATCH 1/2] Kbuild: enable -fms-extensions
Quoted 24 lines · Nathan Chancellor
On Wed, Oct 22, 2025 at 11:11:38PM +0200, Nathan Chancellor wrote:
...Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
(quoted diff) +2
-0+# Allow including a tagged struct or union anonymously in another struct/union. +KBUILD_CFLAGS += -fms-extensions...
Quoted 2 lines
I have tentatively applied this to kbuild-next so that it can spend most of the cycle in -next to try and catch all potential problems.
One side effect that has been found in my testing so far is clang's '-fms-extensions' turns '
_inline' into a keyword, which breaks fs/jfs:In file included from
fs/jfs/jfs_unicode.c:8:fs/jfs/jfs_incore.h:86:13: error: type name does not allow function specifier to be specified 86 | unchar _inline[128]; | ^ fs/jfs/jfs_incore.h:86:20: error: expected member name or ';' after declaration specifiers 86 | unchar _inline[128]; | ~~~~~~~~~~~~~~^There appear to be other similar keywords (ones with KEYMS in the linke below) but my personal distribution configuration does not show any instances in the build where they matter (I did not test allmodconfig yet).
Quoted 6 lines
https://github.com/llvm/llvm-project/blob/249883d0c5883996bed038cd82a8999f342994c9/clang/include/clang/Basic/TokenKinds.def#L744-L794Something like this is all it takes to resolve the issue, so I will send a patch for formal review/acking but I wanted to bring it up ahead of time in case this is unpalpable and we should throw these changes out of -next instead of forward fixing.
I'm on vacation now, so I may be slow to respond to a future patch, so I'll go ahead and give you my ack to this.
Quoted 34 lines
Cheers,
Nathanfs/jfs/jfs_incore.h +3
-3@@ -76,14 +76,14 @@ struct jfs_inode_info {struct {unchar _unused[16 ];/* 16: */ dxd_t _dxd;/* 16: */ -/* _inline may overflow into _inline_ea when needed */ +/* _inline_sym may overflow into _inline_ea when needed */ /* _inline_ea may overlay the last part of * file._xtroot if maxentry = XTROOTINITSLOT */ union {struct {/* 128: inline symlink */ -unchar _inline[128 ]; +unchar _inline_sym[128 ];/* 128: inline extended attr */ unchar _inline_ea[128 ]; };@@ -101,7 +101,7 @@ struct jfs_inode_info {#define i_imap u.file._imap#define i_dirtable u.dir._table#define i_dtroot u.dir._dtroot -#define i_inline u.link._inline +#define i_inline u.link._inline_sym#define i_inline_ea u.link._inline_ea#define i_inline_all u.link._inline_all
Nathan Chancellor
maintainerRe: [PATCH 1/2] Kbuild: enable -fms-extensions
Quoted 10 lines · Dave Kleikamp
Quoted 4 lines · Nathan Chancellor
Something like this is all it takes to resolve the issue, so I will send a patch for formal review/acking but I wanted to bring it up ahead of time in case this is unpalpable and we should throw these changes out of -next instead of forward fixing.
I'm on vacation now, so I may be slow to respond to a future patch, so I'll go ahead and give you my ack to this.
Acked-by Dave Kleikamp <[email protected]>
Great, thanks for the tag and taking a look quickly!
Quoted 30 lines
Quoted 29 lines
fs/jfs/jfs_incore.h +3
-3@@ -76,14 +76,14 @@ struct jfs_inode_info {struct {unchar _unused[16 ];/* 16: */ dxd_t _dxd;/* 16: */ -/* _inline may overflow into _inline_ea when needed */ +/* _inline_sym may overflow into _inline_ea when needed */ /* _inline_ea may overlay the last part of * file._xtroot if maxentry = XTROOTINITSLOT */ union {struct {/* 128: inline symlink */ -unchar _inline[128 ]; +unchar _inline_sym[128 ];/* 128: inline extended attr */ unchar _inline_ea[128 ]; };@@ -101,7 +101,7 @@ struct jfs_inode_info {#define i_imap u.file._imap#define i_dirtable u.dir._table#define i_dtroot u.dir._dtroot -#define i_inline u.link._inline +#define i_inline u.link._inline_sym#define i_inline_ea u.link._inline_ea#define i_inline_all u.link._inline_all
PATCH 2/2 btrfs: send: make use of -fms-extensions for defining struct fs_path
In reply to Rasmus VillemoesThe newly introduced -fms-extensions compiler flag allows defining struct fs_path in such a way that inline_buf becomes a proper array with a size known to the compiler.
This also makes the problem fixed by 8aec9dbf2db2 ("btrfs: send: fix -Wflex-array-member-not-at-end warning in struct send_ctx") go away. Whether cur_inode_path should be put back to its original place in struct send_ctx I don't know, but at least the comment no longer applies.
--- fs/btrfs/send.c | 39 ++++++++++++++++++++------------------- 1 file changed, 20 insertions(+), 19 deletions(-)
fs/btrfs/send.c +20 -19
@@ -47,28 +47,30 @@* It allows fast adding of path elements on the right side (normal path) and * fast adding to the left side (reversed path). A reversed path can also be * unreversed if needed. +* +* The definition of struct fs_path relies on -fms-extensions to allow +* including a tagged struct as an anonymous member. */ +struct __fs_path { +char *start; +char *end; + +char *buf; +unsigned short buf_len:15 ; +unsigned short reversed:1 ; +}; +static_assert (sizeof (struct __fs_path ) <256 );struct fs_path { -union { -struct { -char *start; -char *end; - -char *buf; -unsigned short buf_len:15 ; -unsigned short reversed:1 ; -char inline_buf[]; - }; -/* -* Average path length does not exceed 200 bytes, we'll have -* better packing in the slab and higher chance to satisfy -* an allocation later during send. -*/ -char pad[256 ]; - }; +struct __fs_path ; +/* +* Average path length does not exceed 200 bytes, we'll have +* better packing in the slab and higher chance to satisfy +* an allocation later during send. +*/ +char inline_buf[256 -sizeof (struct __fs_path )]; };#define FS_PATH_INLINE_SIZE \ - (sizeof(struct fs_path) - offsetof(struct fs_path, inline_buf)) + sizeof_field(struct fs_path, inline_buf)/* reused for each extent */
@@ -305,7 +307,6 @@ struct send_ctx {struct btrfs_lru_cache dir_created_cache;struct btrfs_lru_cache dir_utimes_cache; -/* Must be last as it ends in a flexible-array member. */ struct fs_path cur_inode_path; };
Signature · 1 line
2.51.0
Re: [PATCH 2/2] btrfs: send: make use of -fms-extensions for defining struct fs_path
Quoted 9 lines · Rasmus Villemoes
fs/btrfs/send.c +8
-0+struct __fs_path { +char *start; +char *end; + +char *buf; +unsigned short buf_len:15 ; +unsigned short reversed:1 ; +};
fs/btrfs/send.c line 71
Quoted 10 lines
fs/btrfs/send.c +8
-0+static_assert (sizeof (struct __fs_path ) <256 );struct fs_path { +struct __fs_path ; +/* +* Average path length does not exceed 200 bytes, we'll have +* better packing in the slab and higher chance to satisfy +* an allocation later during send. +*/ +char inline_buf[256 -sizeof (struct __fs_path )]; };
It strikes me that this won't pack as well as it used to before the change.
On 64-bit architectrures, 'struct __fs_path' will be 8-byte aligned due to the pointers in it, and that means that the size of it will also be aligned: it will be 32 bytes in size.
So you'll get 256-32 bytes of inline_buf.
And it *used* to be that 'inline_buf[]' was packed righ after the 16-bit buf_len / reversed bits, so it used to get an extra six bytes.
I think it could be fixed with a "__packed" thing on that inner struct, but that also worries me a bit because we'd certainly never want the compiler to generate the code for unaligned accesses (on the broken architectures that would do that). You'd then have to mark the containing structure as being aligned to make compilers generate good code.
So either you lose some inline buffer space, or you end up having to add extra packing stuff. Either way is a bit of a bother.
Linus
Re: [PATCH 2/2] btrfs: send: make use of -fms-extensions for defining struct fs_path
Quoted 39 lines · Linus Torvalds
Quoted 9 lines · Rasmus Villemoes
fs/btrfs/send.c +8
-0+struct __fs_path { +char *start; +char *end; + +char *buf; +unsigned short buf_len:15 ; +unsigned short reversed:1 ; +};fs/btrfs/send.c line 71
Quoted 10 lines
fs/btrfs/send.c +8
-0+static_assert (sizeof (struct __fs_path ) <256 );struct fs_path { +struct __fs_path ; +/* +* Average path length does not exceed 200 bytes, we'll have +* better packing in the slab and higher chance to satisfy +* an allocation later during send. +*/ +char inline_buf[256 -sizeof (struct __fs_path )]; };It strikes me that this won't pack as well as it used to before the change.
On 64-bit architectrures, 'struct
__fs_path' will be 8-byte aligned due to the pointers in it, and that means that the size of it will also be aligned: it will be 32 bytes in size.So you'll get 256-32 bytes of
inline_buf.And it *used* to be that 'inline_buf[]' was packed righ after the 16-bit
buf_len/ reversed bits, so it used to get an extra six bytes.I think it could be fixed with a "
__packed" thing on that inner struct, but that also worries me a bit because we'd certainly never want the compiler to generate the code for unaligned accesses (on the broken architectures that would do that). You'd then have to mark the containing structure as being aligned to make compilers generate good code.
Quoted 2 lines
So either you lose some inline buffer space, or you end up having to add extra packing stuff. Either way is a bit of a bother.
For the inline path buffer losing 6 bytes is acceptable, I did some stats on my system and full paths under /.snapshots from snapper are all uner 200 bytes, lengths on another random data partition goes up to 380. Users can of course have path of any length but there's a fallback and the allocations are done in the steps of slab bucket sizes so it's reasonably effective.
Quoted 5 lines · Rasmus Villemoes
Since -fms-extensions once again came up as potentially useful, Linus suggested that we bite the bullet and enable it.
https://lore.kernel.org/lkml/CAHk-=wjeZwww6Zswn6F_iZTpUihTSNKYppLqj36iQDDhfntuEw@mail.gmail.com/
Quoted 8 lines
So that's what patch 1 does, and patch 2 puts it to use in the btrfs case.
Compile-tested only, with gcc (15.2.1) and clang (20.1.8).
Rasmus Villemoes (2):
Kbuild: enable -fms-extensions
btrfs: send: make use of -fms-extensions for defining structfs_path
For the btrfs part
I think it makes more sense to take the patches via the kbuild tree so it's in linux-next for build coverage and eventual tweaks to the Kbuild files. Or I can take the patches into btrfs for-next.
Nathan Chancellor
maintainer+ Nicolas for kbuild
Quoted 15 lines · David Sterba
Quoted 5 lines · Rasmus Villemoes
Since -fms-extensions once again came up as potentially useful, Linus suggested that we bite the bullet and enable it.
https://lore.kernel.org/lkml/CAHk-=wjeZwww6Zswn6F_iZTpUihTSNKYppLqj36iQDDhfntuEw@mail.gmail.com/
Quoted 8 lines
So that's what patch 1 does, and patch 2 puts it to use in the btrfs case.
Compile-tested only, with gcc (15.2.1) and clang (20.1.8).
Rasmus Villemoes (2):
Kbuild: enable -fms-extensions
btrfs: send: make use of -fms-extensions for defining structfs_path
Quoted 7 lines
For the btrfs part
Acked-by David Sterba <[email protected]>I think it makes more sense to take the patches via the kbuild tree so it's in linux-next for build coverage and eventual tweaks to the
Kbuildfiles. Or I can take the patches into btrfs for-next.
Yeah, we should be able to take these into Kbuild. I had just one small nit on patch 1 that we could probably fix up at application time unless folks disagree with it. I hope there will be no follow up fixes needed but it would make more sense in our tree than yours.
Cheers,
Nathan