Re: [PATCH v4 5/6] xfs: write stream based AG placement

[Christoph Hellwig <hch@lst.de>]

On Fri, Aug 21, 2026 at 06:03:29PM +0530, Kanchan Joshi wrote:
> On 8/19/2026 12:59 PM, Christoph Hellwig wrote:
> > On Fri, Jul 17, 2026 at 06:25:37PM +0530, Kanchan Joshi wrote:
> >> If AGs are fewer than write streams, write streams are distributed into
> >> available AGs in round robin fashion.
> > 
> > I think we should just limit the number of available write streams
> > to the number of AG, as otherwise we'll get into a bit of a mess.
> 
> To confirm: given above quoting, for the other two cases (i.e., Streams 
> <= AGs), are you open to stream-based AG-sets modulo inode hashing (that 
> you suggested to skip in previous patch)?

Yes.  Even inode hashing might be ok, but I think it mostly papers
over other issues.

> And is this point only for the case when Streams > AGs?

Yes.

> This case can happen for hw streams, but for pure software streams, we 
> will not hit this case because the filesystem determines the stream 
> count based on the AG count.

Right.

> 
> > And for the case of the hardware write streams query them in mkfs
> > to align the AG geometry to that.
> 
> If this is only for the streams > AGs case, the filesystem can just 
> clamp down to match, turning it into a 'streams = AGs' case.
> And that means no changes in mkfs. Is that fine?

Well, it is fine.  But number of available write streams might be
a very useful hint to mkfs as mkfs tries to guess the available
concurrent, and the write streams are how the device exposes that.