Re: [PATCH v5 05/10] xfs: Iomap SW-based atomic write support

[Christoph Hellwig <hch@infradead.org>]

On Wed, Mar 12, 2025 at 02:57:38PM +0000, John Garry wrote:
> > > I do admit that the checks are a bit uneven, i.e. check vs
> > > IOMAP_DIO_ATOMIC_SW and IOCB_ATOMIC
> > > 
> > > If we want a flag to set REQ_ATOMIC from the FS then we need
> > > IOMAP_DIO_BIO_ATOMIC, and that would set IOMAP_BIO_ATOMIC. Is that better?
> > 
> > My expectation from a very cursory view is that iomap would be that
> > there is a IOMAP_F_REQ_ATOMIC that is set in ->iomap_begin and which
> > would make the core iomap code set REQ_ATOMIC on the bio for that
> > iteration.
> 
> but we still need to tell ->iomap_begin about IOCB_ATOMIC, hence

Yeah, ->iomap_begin  can't directly look at the iocb.

> IOMAP_DIO_BIO_ATOMIC which sets IOMAP_BIO_ATOMIC.

> 
> We can't allow __iomap_dio_rw() check IOCB_ATOMIC only (and set
> IOMAP_BIO_ATOMIC), as this is the common path for COW and regular atomic
> write

Well, I'd imagine __iomap_dio_rw just sets IOMAP_ATOMIC from IOCB_ATOMIC
and then it's up to file system internal state if it wants to set
IOMAP_F_REQ_ATOMIC based on that, i.e. the actual setting of
IOMAP_F_REQ_ATOMIC is fully controlled by the file system and not
by the iomap core.