2:55 p.m.
On 4/10/20 2:06 PM, Andrea Bolognani wrote:
On Thu, 2020-04-09 at 12:30 +0200, Shalini Chellathurai Saroja
wrote:
> The ZPCI address validation or autogeneration does not work as
> expected in the following scenarios
> 1. uid = 0 and fid = 0
> 2. uid = 0 and fid > 0
> 3. uid = 0 and fid not specified
> 4. uid not specified and fid > 0
> 5. 2 ZPCI devices with uid > 0 and fid not specified.
>
> This is because of the following reasons
> 1. If uid = 0 or fid = 0 the code assumes that user has not specified
> the corresponding address
> 2. If either uid or fid is provided, the code assumes that both uid
> and fid addresses are specified by the user.
I'd have to dig up the old threads, but based on what I remember the
behaviors you describe are entirely intentional.
For PCI addresses, setting all parts of the address to zero or not
setting it at all is equivalent, and we wanted to be consistent with
that behavior for ZPCI; additionally, zero is not a valid value for
uid so of course neither is the address uid=0 fid=0, which means that
we're not preventing the user from specifying a valid address by
conflating the all-zero address with the unspecified address.
For partially-specified addresses, the behavior is also the same as
PCI: any part you don't specify is considered to be zero, which
results in
uid=0 fid=0 -> uid=0 fid=0 -> address gets autogenerated
uid=0 fid=x -> uid=0 fid=x -> address is rejected as invalid
uid=0 -> uid=0 fid=0 -> address gets autogenerated
fid=x -> uid=0 fid=x -> address is rejected as invalid
uid=x -> uid=x fid=0 -> address is accepted
So, just like for PCI addresses, you have basically two reasonable
options: either don't specify any zPCI address and leave allocation
entirely up to libvirt, or specify all of the addresses completely:
anything in between will likely not work as you'd expect or want.
Again, this is based purely on my recollection of design discussions
that happened one and a half years ago, so I might have gotten some
of the details wrong - in which case by all means call me out on
that O:-)
Hi Andrea,
sorry for the delayed answer. I (and some others as well) lost some
emails on my IMAP account and I just found your answer today.
I can remember that you had a discussion with the original author of the
zpci code. There are a few issues with the currently implemented "rules"
which partially are not even working as you outlined above in more
complex scenarios.
First: Setting uid=0 or uid='0x0000'
The architecture allows to do that BUT if you do than you are NOT using
the uid mode which results for the guest OS to use the legacy mode for
assigning PCI addresses starting with 0 increasing by one following an
unpredictable order by which the pci device are presented to guest OS.
Since we never ever wanted to support the legacy mode in KVM guests we
decided to never allow uid=0. Allowing the uid to be set to 0 is a
contradiction.
Actually the user can also set uid='0x0000' which I consider very
specific and one would end up with something like uid='0x0001' and even
more confusing is that suddenly setting uid='0x0000' on more than one
PCI device is allowed.
Besides that the current zpci code still contains at least one flaw that
is simply caused by the fact that there is no knowledge about which
value was specified by the user.
In Shalini's and your list it is case 5: This scenario runs into errors
when another PCI device already has a fid set to 0 OR another PCI device
exists specified with a uid > 0 and without a fid. The user gets the
error message for something he did not specify:
error: Failed to define domain from pci-addr-test.xml
error: internal error: zPCI fid 0 is already reserved
Regarding setting fid=0 or fid='0x00000000'
Since it is a legal value for fid specifying it should not be considered
as a wildcard or set equivalent to not specifying it at all.
Doing so things like this happen and for the user it certainly seems
like a bug:
Specify this in the domain:
pcidev1: uid='0x0000' fid='0x00000000'
pcidev2: uid='0x0000'
Results in a defined domain:
pcidev1: uid='0x0002' fid='0x00000001'
pcidev2: uid='0x0001' fid='0x00000000'
Another example:
Specify this in the domain:
pcidev1: fid='0x00000000'
pcidev2: fid='0x00000000'
Results in a defined domain:
pcidev1: uid='0x0002' fid='0x00000001'
pcidev2: uid='0x0001' fid='0x00000000'
BUT
Specify this in the domain:
pcidev1: uid='0x0002' fid='0x00000000'
pcidev2: uid='0x0001' fid='0x00000000'
Results in error:
error: Failed to define domain from pci-addr-test.xml
error: internal error: zPCI fid 0 is already reserved
(Btw remove one of the fids results in the flaw described above.)
I think that Shalini's patch series improves the zpci address generation
to better meet the users expected behavior. It also removes a
correlation between uid and fid that does not really exist.
--
Mit freundlichen Grüßen/Kind regards
Boris Fiuczynski
IBM Deutschland Research & Development GmbH
Vorsitzender des Aufsichtsrats: Gregor Pillen
Geschäftsführung: Dirk Wittkopp
Sitz der Gesellschaft: Böblingen
Registergericht: Amtsgericht Stuttgart, HRB 243294