Re: [PATCH v2 1/2] fs,block: Introduce RWF_ZONE_APPEND and handling in direct IO path

[Kanchan Joshi <joshi.k@samsung.com>]

[-- Attachment #1: Type: text/plain, Size: 2283 bytes --]

On Fri, Jun 26, 2020 at 09:58:46AM +0100, Christoph Hellwig wrote:
>To restate my previous NAK:
>
>A low-level protocol detail like RWF_ZONE_APPEND has absolutely no
>business being exposed in the Linux file system interface.
>
>And as mentioned before I think the idea of returning the actual
>position written for O_APPEND writes totally makes sense, and actually
>is generalizable to all files.  Together with zonefs that gives you a
>perfect interface for zone append.
>
>On Thu, Jun 25, 2020 at 10:45:48PM +0530, Kanchan Joshi wrote:
>> Introduce RWF_ZONE_APPEND flag to represent zone-append.
>
>And no one but us select few even know what zone append is, nevermind
>what the detailed semantics are.  If you add a userspace API you need
>to very clearly document the semantics inluding errors and corner cases.

For block IO path (which is the scope of this patchset) there is no
probelm in using RWF_APPEND for zone-append, because it does not do
anything for block device. We can use that, avoiding introduction of
RWF_ZONE_APPEND in user-space.

In kernel, will it be fine to keep IOCB_ZONE_APPEND apart from
IOCB_APPEND? Reason being, this can help to isolate the code meant only
for zone-append from the one that is already present for conventional
append.

Snippet from quick reference -

static inline int kiocb_set_rw_flags(struct kiocb *ki, rwf_t flags)
        ki->ki_flags |= (IOCB_DSYNC | IOCB_SYNC);
        if (flags & RWF_APPEND)
                ki->ki_flags |= IOCB_APPEND;
+       if (flags & RWF_ZONE_APPEND) {
+               /* currently support block device only */
+               umode_t mode = file_inode(ki->ki_filp)->i_mode;
+
+               if (!(S_ISBLK(mode)))
+                       return -EOPNOTSUPP;
+               ki->ki_flags |= IOCB_ZONE_APPEND;
+       }


As for file I/O in future, I see a potential problem with RWF_APPEND.
In io_uring, zone-append requires bit of pre/post processing, which
ideally should be done only for zone-append case. A ZoneFS file using
RWF_APPEND as a mean to invoke zone-append vs a regular file (hosted on
some other FS) requiring conventional RWF_APPEND - both will execute
that processing.
Is there a good way to differentiate ZoneFS file from another file which
only wants use conventional file-append?

[-- Attachment #2: Type: text/plain, Size: 0 bytes --]