Re: Regression v6.11 booting cannot mount harddisks (xfs)

[Jesper Dangaard Brouer]

On 10/09/2024 15.06, Damien Le Moal wrote:
> On 2024/09/10 21:19, Jesper Dangaard Brouer wrote:
> > Hi Linus,
> >
> > My testlab kernel devel server isn't booting correctly on v6.11 branches
> > (e.g. net-next at 6.11.0-rc5)
> > I just confirmed this also happens on your tree tag: v6.11-rc7.
> >
> > The symptom/issue is that harddisk dev names (e.g /dev/sda, /dev/sdb,
> > /dev/sdc) gets reordered.  I switched /etc/fstab to use UUID's instead
> > (which boots on v6.10) but on 6.11 it still cannot mount harddisks and
> > doesn't fully boot.
>
> Parallel SCSI device scanning has been around for a long time... This is
> controlled with CONFIG_SCSI_SCAN_ASYNC. And yes, that can cause disk names to
> change, which is why it is never a good idea to rely on them but instead use
> /dev/disk/by-* names. Disabling CONFIG_SCSI_SCAN_ASYNC will likely not guarantee
> that disk names will be constant, given that you seem to have 2 AHCI adapters on
> your host and PCI device scanning is done in parallel.
>
> > E.g. errors:
> >     systemd[1]: Expecting device
> > dev-disk-by\x2duuid-0c2b348d\x2de013\x2d482b\x2da91c\x2d029640ec427a.device
> > - /dev/disk/by-uuid/0c2b348d-e013-482b-a91c-029640ec42
> > 7a...
> >     [DEPEND] Dependency failed for var-lib.mount - /var/lib.
> >     [...]
> >     [ TIME ] Timed out waiting for device
> > dev-d…499e46-b40d-4067-afd4-5f6ad09fcff2.
> >     [DEPEND] Dependency failed for boot.mount - /boot.
> >
> > That corresponds to fstab's:
> >    - UUID=8b499e46-b40d-4067-afd4-5f6ad09fcff2 /boot     xfs defaults 0 0
> >    - UUID=0c2b348d-e013-482b-a91c-029640ec427a /var/lib/ xfs defaults 0 0
> >
> > It looks like disk controller initialization happens in *parallel* on
> > these newer kernels as dmesg shows init printk's overlapping:
> >
> >    [    5.683393] scsi 5:0:0:0: Direct-Access     ATA      SAMSUNG
> > MZ7KM120 003Q PQ: 0 ANSI: 5
> >    [    5.683641] scsi 7:0:0:0: Direct-Access     ATA      SAMSUNG
> > MZ7KM120 003Q PQ: 0 ANSI: 5
> >    [    5.683797] scsi 8:0:0:0: Direct-Access     ATA      Samsung SSD
> > 840  BB0Q PQ: 0 ANSI: 5
> >    [...]
> >    [    7.057376] sd 5:0:0:0: [sda] 234441648 512-byte logical blocks:
> > (120 GB/112 GiB)
> >    [    7.062279] sd 7:0:0:0: [sdb] 234441648 512-byte logical blocks:
> > (120 GB/112 GiB)
> >    [    7.070628] sd 5:0:0:0: [sda] Write Protect is off
> >    [    7.070701] sd 8:0:0:0: [sdc] 488397168 512-byte logical blocks:
> > (250 GB/233 GiB)
> >
> > Perhaps this could be a hint to what changed?
>
> See above. The disk /dev/sdX names not being reliable is rather normal.
> Are you sure you have the correct UUIDs of your FSes on the disks ? You can
> check them with "blkid /dev/sdX[n]"

I have checked that I use the correct UUIDs.

I checked my /etc/fstab have the UUID entries under /dev/disk/by-uuid/
via this oneliner, which needs to have a /etc/fstab entry under each
UUID. We can see I have one partition that I'm not using
(0fd3bc38-6496-401f-87f2-87e09532de53), which is expected.

$ for UUID in $(ls /dev/disk/by-uuid/); do echo $UUID; grep -H $UUID 
/etc/fstab; done
09e8c15f-80d2-47e3-8e73-d3fdfcf33eef
/etc/fstab:UUID=09e8c15f-80d2-47e3-8e73-d3fdfcf33eef / 
    xfs     defaults        0 0
0c2b348d-e013-482b-a91c-029640ec427a
/etc/fstab:UUID=0c2b348d-e013-482b-a91c-029640ec427a	/var/lib/		xfs 
defaults 0 0
0fd3bc38-6496-401f-87f2-87e09532de53
581920da-1ccb-4b25-856c-036310032a74
/etc/fstab:UUID=581920da-1ccb-4b25-856c-036310032a74	/nix			xfs	defaults 0 0
8b499e46-b40d-4067-afd4-5f6ad09fcff2
/etc/fstab:UUID=8b499e46-b40d-4067-afd4-5f6ad09fcff2 /boot 
    xfs     defaults        0 0
cd409a50-0371-47ca-9213-49a2bc7b9317
/etc/fstab:UUID=cd409a50-0371-47ca-9213-49a2bc7b9317 swap 
    swap    defaults        0 0


> > Any hints what commit I should try to test revert?
> > Or good starting point for bisecting?
>
> You said that 6.10 works, so maybe start from there ?

I tested I could boot tag v6.10, and have started bisection.

I've not tried to deselect CONFIG_SCSI_SCAN_ASYNC as the kernel that
worked on tag v6.10 also had this CONFIG_SCSI_SCAN_ASYNC enabled. So, it
is likely not related to the async controller init.

--Jesper