On Fri, Feb 07, 2025 at 11:31:17AM -0800, Darrick J. Wong wrote:
> On Fri, Feb 07, 2025 at 11:26:20AM -0800, Luis Chamberlain wrote:
> > On Thu, Feb 06, 2025 at 02:27:16PM -0800, Darrick J. Wong wrote:
> > > NAME MIN-IO
> > > sda 512
> > > ├─sda1 512
> > > ├─sda2 512
> > > │ └─node0.boot 512
> > > ├─sda3 512
> > > │ └─node0.swap 512
> > > └─sda4 512
> > > └─node0.lvm 512
> > > └─node0-root 512
> > > sdb 4096
> > > └─sdb1 4096
> > > nvme1n1 512
> > > └─md0 524288
> > > └─node0.raid 524288
> > > └─node0_raid-storage 524288
> > > nvme0n1 512
> > > └─md0 524288
> > > └─node0.raid 524288
> > > └─node0_raid-storage 524288
> > > nvme2n1 512
> > > └─md0 524288
> > > └─node0.raid 524288
> > > └─node0_raid-storage 524288
> > > nvme3n1 512
> > > └─md0 524288
> > > └─node0.raid 524288
> > > └─node0_raid-storage 524288
> >
> > Can you try this for each of these:
> >
> > stat --print=%o
> >
> > I believe that without that new patch I posted [0] you will get 4 KiB
> > here. Then the blocksize passed won't be the min-io until that patch
> > gets applied.
>
> Yes, that returns 4K on 6.13.0 for every device in the list. I think
> you're saying that stat will start returning 512K for the blocksize if
> your patch is merged?
Yes, as that *is* min-io for block devices.
> > The above is:
> >
> > statx(AT_FDCWD, "/dev/nvme0n1", AT_STATX_SYNC_AS_STAT|AT_SYMLINK_NOFOLLOW|AT_NO_AUTOMOUNT, 0,
> > {stx_mask=STATX_BASIC_STATS|STATX_MNT_ID, stx_attributes=0,
> > stx_mode=S_IFBLK|0660, stx_size=0, ...}) = 0
> >
> > So if we use this instead at mkfs, then even older kernels will get 4
> > KiB, and if distros want to automatically lift the value at mkfs, they
> > could cherry pick that simple patch.
>
> How well does that work if the gold master image creator machine has a
> new kernel and a RAID setup, but the kernel written into the gold master
> image is something older than a 6.12 kernel?
I think you're asking about an image creator for an old release and that
old release uses an old kernel. Isn't that the same concern as what if
the same new kernel creates a filesystem on an image for an old release?
If so the risk is the new kernel defaults take precedence.
And the only thing I can think of for this case an option to ignore this
heuristic for blocksize, which could be used on configs to override
defaults for older target kernels.
Luis
