On Tue, Nov 30 2021 at 22:23, Thomas Gleixner wrote:
> On Tue, Nov 30 2021 at 16:28, Jason Gunthorpe wrote:
>
> The real problem is where to store the MSI descriptors because the PCI
> device has its own real PCI/MSI-X interrupts which means it still shares
> the storage space.
Bah. I confused myself by staring at the existing code instead of
looking at how this NTB stuff actually works.
So if I understand it correctly then the end result looks like this:
1) PCIe device (switchtec)
The device has 4 MSI[X] interrupts: event, dma_rpc, message,
doorbell. The event and dma_rpc interrupts are requested by the
switchtec PCI driver itself.
2) Switchtec character device
The switchtec PCI driver creates a character device which is exposed
for device specific IOCTLs
The device belongs to the switchtec_class device class.
3) Switchtec NTB device
The ntb_hw_switchtec driver registers the switchtec_class class
interface.
So when #2 is registered with the driver core the switchtec class
interface add_dev() function is invoked. That function creates a NTB
device, requests the message and the doorbell interrupts which have
been allocated by the underlying PCIe device driver (#1) and
registers the NTB device with the NTB core.
4) The NTB core then tries to use the virtual MSI vectors which have
been allocated by the switchtec driver in #1 and requires the msg
write intercept to actually expose it to the peers.
So we really can go and create a MSI irqdomain and stick the pointer
into stdev->dev.irqdomain. The parent domain of this irqdomain is
stdev->pdev.dev.irqdomain->parent
which is either the irq remapping domain or the vector domain. Which is
pretty much what I proposed as general facility for IMS/IDXD. I need to
go back and polish that up on top of the current pile.
Along with that have an irq chip implementation which exposes:
static struct irq_chip ntb_chip = {
.name = "ntb",
.irq_ack = irq_chip_ack_parent,
.irq_write_msi_msg = ntb_msi_write_msg,
#ifdef CONFIG_SMP
.irq_set_affinity = irq_chip_set_affinity_parent,
#endif
};
We just need some reasonable solution for the DMA/remap problem Jason
mentioned vs. msi_desc::dev, but that wants to be cleaned up in any
case for all the aliasing muck.
Thanks,
tglx
