Re: attaching device to PV guest broken by your rework of libxl's PCI handling?

["Durrant, Paul" <xadimgnik@xxxxxxxxx>]

On 09/12/2021 04:17, Jan Beulich wrote:
> Paul,
>
> in 0fdb48ffe7a1 ("libxl: Make sure devices added by pci-attach are
> reflected in the config") you've moved down the invocation of
> libxl__create_pci_backend() from libxl__device_pci_add_xenstore().
> In the PV case, soon after the original invocation place there is
>
>      if (!starting && domtype == LIBXL_DOMAIN_TYPE_PV) {
>          if (libxl__wait_for_backend(gc, be_path, GCSPRINTF("%d", 
> XenbusStateConnected)) < 0)
>              return ERROR_FAIL;
>      }
>
> Afaict the only way this wait could succeed is if the backend was
> created up front. The lack thereof does, I think, explain a report
> we've had:
>
> vh015:~ # xl -vvv pci-attach sles-15-sp4-64-pv-def-net 63:11.4
> libxl: debug: libxl_pci.c:1561:libxl_device_pci_add: Domain 18:ao 
> 0x55a517704170: create: how=(nil) callback=(nil) poller=0x55a517704210
> libxl: debug: libxl_qmp.c:1921:libxl__ev_qmp_dispose:  ev 0x55a5177047e8
> libxl: error: libxl_device.c:1393:libxl__wait_for_backend: Backend 
> /local/domain/0/backend/pci/18/0 does not exist
> libxl: error: libxl_pci.c:1779:device_pci_add_done: Domain 
> 18:libxl__device_pci_add failed for PCI device 0:63:11.4 (rc -3)
> libxl: error: libxl_device.c:1420:device_addrm_aocomplete: unable to add device
Wow. It must be a year since those patches went in... Most of the 
context has disappeared from my mind.
> Since I don't fully understand what that commit as a whole is
> doing, and since the specific change in the sequence of operations
> also isn't explained in the description (or at least not in a way
> for me to recognize the connection), I'm afraid I can't see how a
> possible solution to this could look like. The best guess I could
> come up with so far is that the code quoted above may also need
> moving down, but I can't tell at all whether doing this after the
> various other intermediate steps wouldn't be too late. Your input
> (or even better a patch) would be highly appreciated.
The commit comment explains the problem that it is trying to fix but I 
agree that it does not call out the new sequence. The issue IIRC was in 
what happened before the call to device_add_domain_config() and what 
happened afterwards. In fixing that I guess I missed this immediate use 
of xenstore.
I *think* the correct fix would be to move the wait into the end of 
libxl__create_pci_backend(), which is where the frontend and backend 
state nodes are now set.
  Paul


> Thanks, Jan