On 13.10.2021 12:02, Bertrand Marquis wrote: >> On 13 Oct 2021, at 07:10, Jan Beulich <jbeul...@suse.com> wrote: >> On 12.10.2021 23:37, Stefano Stabellini wrote: >>> Good point about MSIs not being setup before the traps. We should remove >>> the call to pci_cleanup_msi() in the error path then. >> >> Your reply makes me fear you didn't pay enough attention to the "partial" >> in my earlier reply. The traps for the various registers can't all be set >> up atomically, so there may be a transient period where enough traps are >> already in place for a cunning guest to arrange for setup. Unless, as >> said, there are further setup steps needed before a guest could succeed >> in doing so. >> >> But even if partial trap setup alone was sufficient, I think the cleaning >> up of MSI then might still better go on the error path there than on that >> of pci_add_device(). > > I think I should put the msi_cleanup in the exit path if pdev is not null but > we got a non null ret (in an else if ( pdev ) ). > This would cover all exit paths, especially as I will move the add_handler > before the iommu init. > > Would that be ok for everyone ?
Sounds reasonable at the first glance. Jan
