On Mon, Sep 30, 2013 at 12:43:20PM +0300, Marcel Apfelbaum wrote: > On Mon, 2013-09-30 at 12:14 +0300, Michael S. Tsirkin wrote: > > On Mon, Sep 30, 2013 at 11:02:06AM +0200, Paolo Bonzini wrote: > > > Il 30/09/2013 10:58, Michael S. Tsirkin ha scritto: > > > >>> > > As a next step, can we make pci_set_irq non-inline and make > > > >>> > > it call pci_irq_handler directly, and get rid of the irq field? > > > >> > What irq field? > > > > /* IRQ objects for the INTA-INTD pins. */ > > > > qemu_irq *irq; > > > > > > > > > > That's still used by devices that use common code for PCI and sysbus > > > versions (e.g. USB OHCI and EHCI). > > > > > > Paolo > > > > Well this work wouldn't be complete without > > addressing them anyway. > > > > These devices would have to create their own > > irq in pci-specific code, along the lines of: > > This irq field is used also in places where pci_set_irq(PCIDevice dev, level) > can't infer the INTx: > - PCIExpress: > qemu_set_irq(dev->irq[dev->exp.hpev_intx],dev->exp.hpev_notified);
Well the spec says, explicitly: 6.7.3.4. Software Notification of Hot-Plug Events ... Note that all other interrupt sources within the same Function will assert the same virtual INTx wire when requesting service. I read this to mean that this is a bug, and it should simply use pci_set_irq like all other devices. > - vmxnet3 device: qemu_set_irq(d->irq[int_idx], 1); > > What approach should be used here? > > Thanks, > Marcel > > - s->irq = dev->irq[3]; > > + s->irq = qemu_allocate_irqs(pci_set_irq, dev, 1); > > > > > > If there's more than one device like this, we should add > > > > /* Return an irq that calls pci_set_irq internally */ > > qemu_irq *pci_allocate_irq(PCIDevice *); > > > >
