[Samba] Direct I/O

[Jeremy Allison]

On Tue, Jan 25, 2022 at 02:02:05PM -0700, Christof Schmitt via samba wrote:
>On Tue, Jan 25, 2022 at 12:45:07PM -0800, Jeremy Allison wrote:
>> On Tue, Jan 25, 2022 at 09:09:36PM +0100, Ralph Boehme wrote:
>> > On 1/25/22 01:50, Christof Schmitt wrote:
>> > > FWIW, preadv2() has RWF_DSYNC and RWF_SYNC which provide O_DSYNC and
>> > > O_SYNC semantics per write. Maybe one of those aligns more closely with
>> > > WRITE_UNBUFFERED.
>> >
>> > perfect! Any takers? :)
>
>Strictly speaking, WRITE_THROUGH would match the SYNC flag (write the
>data to stable storage). WRITE_UNBUFFERED implies writing to storage,
>but ensure that there is no data buffered, which has no Linux/POSIX
>equivalent call (to my knowledge). So smbd sets write_through for both
>of the flags.
>
>> Hmmm. How to add the in_flags field for the SMB_VFS_PREAD_SEND/SMB_VFS_PWRITE_SEND ?
>>
>> Both take an fsp, so is this something we should be adding
>> to the files_struct as a structure element (or a one element
>> in the fsp_flags bitfield) or as a separate 'flags' parameter
>> to the VFS calls ?
>>
>> Personally I prefer adding to the files_struct as that way it's
>> available to all handle-based VFS calls and we don't have
>> to change SMB_VFS_PREAD_SEND/SMB_VFS_PWRITE_SEND, and can
>> just make use of the feature inside vfs_default.c on
>> platforms that have it.
>
>This flag is specific for the individual write request (passed through
>smbd_smb2_write_state to schedule_aio_smb2_write). No other operation
>than the write has the flag, so having it in files_struct seems a bit
>misplaced. Maybe adding a flags field to pwrite_send would be better?

Yes you are right of course. As soon as Ralph mentioned it
I realized what a silly idea adding it to fsp was :-).