On 11/9/24 4:26 PM, Ralph Boehme via samba-technical wrote:
On 10/9/24 9:43 PM, Jeremy Allison wrote:Can we just map (access_mask (FILE_APPEND_DATA|SYNCHRONIZE)) == (FILE_APPEND_DATA|SYNCHRONIZE))to O_APPEND, regardless of POSIX mode ?thinking about this a bit more, this seems doable, albeit only for POSIX mode. For non-POSIX mode we could potentially break WIndows application that open only with FILE_APPEND_DATA: I checked with a torture test that Windows doesn't enforce append behaviour for FILE_APPEND_DATA|SYNCHRONIZE.For POSIX opens we should also allow combinations like FILE_READ_ATTRIBUTES|FILE_APPEND_DATA to map to O_APPEND, so clients can write in append mode to the handle and still are able to fstat() it.https://gitlab.com/samba-team/samba/-/merge_requests/3863
in addition, I guess for SMB3 POSIX we might want to require that the client sets offset=-2 in SMB2-WRITES on a handle opened with append semantics, reusing the logic from MS-FSA.
This avoids bringing in the ambiguity of POSIX vs Linux pwrite() O_APPEND behaviour into the protocol by clearly specifying: if a handle is opened in append mode, all writes are using write() with O_APPEND semantics which are clearly specified.
Looking at possible server implementations, it seems both Linux and FreeBSD note in man pwrite() under BUGS, that a pwrite() on a handle opened with O_APPEND ignores the offset, so we in Samba we can continue to use pwrite() in SMB_VFS_PWRITE() default backend.
To avoid relying on BUGS, we can check for and use pwritev2() with flags=RWF_APPEND, at least on Linux.
-slow
Attachment:
OpenPGP_signature.asc
Description: OpenPGP digital signature
