Re: [PATCH v4 09/10] vpci/msi: Free MSI resources when init_msi() fails

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Thu, May 22, 2025 at 02:21:16AM +0000, Chen, Jiqian wrote:
> On 2025/5/21 19:23, Roger Pau Monné wrote:
> > On Wed, May 21, 2025 at 07:00:37AM +0000, Chen, Jiqian wrote:
> >> On 2025/5/20 17:43, Roger Pau Monné wrote:
> >>> On Tue, May 20, 2025 at 11:14:27AM +0200, Jan Beulich wrote:
> >>>> On 20.05.2025 11:09, Roger Pau Monné wrote:
> >>>>> On Tue, May 20, 2025 at 08:40:28AM +0200, Jan Beulich wrote:
> >>>>>> On 09.05.2025 11:05, Jiqian Chen wrote:
> >>>>>>> When init_msi() fails, the previous new changes will hide MSI
> >>>>>>> capability, it can't rely on vpci_deassign_device() to remove
> >>>>>>> all MSI related resources anymore, those resources must be
> >>>>>>> removed in cleanup function of MSI.
> >>>>>>
> >>>>>> That's because vpci_deassign_device() simply isn't called anymore?
> >>>>>> Could do with wording along these lines then. But (also applicable
> >>>>>> to the previous patch) - doesn't this need to come earlier? And is
> >>>>>> it sufficient to simply remove the register intercepts? Don't you
> >>>>>> need to put in place ones dropping all writes and making all reads
> >>>>>> return either 0 or ~0 (covering in particular Dom0, while for DomU-s
> >>>>>> this may already be the case by default behavior)?
> >>>>>
> >>>>> For domUs this is already the default behavior.
> >>>>>
> >>>>> For dom0 I think it should be enough to hide the capability from the
> >>>>> linked list, but not hide all the capability related
> >>>>> registers.  IMO a well behaved dom0 won't try to access capabilities
> >>>>> disconnected from the linked list,
> >>>>
> >>>> Just that I've seen drivers knowing where their device has certain
> >>>> capabilities, thus not bothering to look up the respective
> >>>> capability.
> >>>
> >>> OK, so let's make the control register read-only in case of failure.
> >>>
> >>> If MSI(-X) is already enabled we should also make the entries
> >>> read-only, and while that's not very complicated for MSI, it does get
> >>> more convoluted for MSI-X.  I'm fine with just making the control
> >>> register read-only for the time being.
> >> If I understand correctly, I need to avoid control register being removed 
> >> and set the write hook of control register to be vpci_ignored_write and 
> >> avoid freeing vpci->msi?
> >>
> >> "
> >>      if ( !msi_pos || !vpci->msi )
> >>          return;
> >>
> >> +    spin_lock(&vpci->lock);
> >> +    control = vpci_get_register(vpci, msi_control_reg(msi_pos), 2);
> >> +    if ( control )
> >> +        control->write = vpci_ignored_write;
> >> +    spin_unlock(&vpci->lock);
> >> +
> >>      if ( vpci->msi->masking )
> >>          end = msi_pending_bits_reg(msi_pos, vpci->msi->address64);
> >>      else
> >>          end = msi_mask_bits_reg(msi_pos, vpci->msi->address64) - 2;
> >>
> >> -    size = end - msi_control_reg(msi_pos);
> >> +    start = msi_control_reg(msi_pos) + 2;
> >> +    size = end - start;
> >>
> >> -    vpci_remove_registers(vpci, msi_control_reg(msi_pos), size);
> >> -    XFREE(vpci->msi);
> >> +    vpci_remove_registers(vpci, start, size);
> > 
> > I think you want to first purge all the MSI range, and then add the
> > control register, also you want to keep the XFREE(), and set the
> > register as:
> Understood.
> 
> > 
> > vpci_add_register(vpci, vpci_hw_read16, NULL, msi_control_reg(msi_pos),
> >                   2, NULL);
> And one more question, how do I process return value of vpci_add_register 
> since definition of cleanup hook is "void"?
> Print a error message if fail?

Well, we should consider the cleanup function returning an error code.
vpci_remove_registers() can also fail, and the error is currently
ignored.  Both cases should result in failing to assign the device to
the domain IMO.

Thanks, Roger.