Re: [PATCH v2] piix: fix regression during unplug in Xen HVM domUs

[Olaf Hering]

Mon, 26 Jun 2023 23:19:01 +0200 Olaf Hering <olaf@aepfle.de>:

> So far I was unable to decipher how the pci_set_word calls can
> possibly affect the outcome and the owner of memory_region_ops_read.

It is enough to return from piix_ide_reset right after
pci_set_word(pci_conf + PCI_COMMAND, 0) to trigger the issue.


One thing which was not mentioned yet: the order in which kernel drivers
are loaded matters. Usually it is xen-platform-pci/uhci-hcd/ata_piix.

When uhci loads, it scans the USB bus, finds the tablet, loads usbhid.
While this happens, ata_piix loads. It finds the PCI device in state
disabled. The PCI code enables the device. On the qemu side this ends
up in pci_default_write_config for PCI device "piix3-ide" with addr=4,
val=1, len=2. This calls pci_update_mappings, which for region #4
changes the addr from 0xc120 to 0xc100. This causes the issue. Now
usbhid tries to use the USB bus, but uhci_irq fails.

If ata_piix is not loaded, uhci works.
If ata_piix is loaded before uhci-hcd, the USB bus can not be scanned,
udev is killed after a timeout and boot proceeds.
If usbhid is loaded before ata_piix, USB bus discovery usually finishes
before ata_piix enables its PCI device, boot proceeds.


Olaf

pgpnZMpYn4Jqb.pgp
Description: Digitale Signatur von OpenPGP