[PATCH v2] PCI: Identifying detached virtual functions
From: Matthew Rosato
Date: Wed Aug 12 2020 - 15:21:30 EST
Changes for v2:
- Added code to vfio_basic_config_read() and vfio_config_init() to
extend emulation to userspace
- Added detached_vf check to vfio_bar_restore()
- @Niklas/@Pierre, I removed your review tags since I made changes,
please have another look
As discussed previously in a qemu-devel thread:
https://www.mail-archive.com/qemu-devel@xxxxxxxxxx/msg725141.html
s390x has the notion of unlinked VFs being available at the LPAR-level
(Virtual Functions where the kernel does not have access to the associated
Physical Function). These devices are currently not marked as is_virtfn.
There seems to be some precedent (ex: in powerpc, eeh_debugfs_break_device())
where pdev->is_virtfn && pdev->physfn == 0 is used to detect these sort of
detached VFs. We toyed with the idea of doing this but it causes additional
fallout as various other areas of kernel code have an expectation that
is_virtfn=1 implies there is a linked PF available to the kernel.
In the s390x case, the firmware layer underneath handles the VF emulation
as it still has access to the PF that the LPAR (and thus the kernel) cannot
see. But one thing this firmware layer does not do is emulate the
PCI_COMMAND_MEMORY bit, which was OK until vfio-pci started enforcing it
via abafbc55. The vfio-pci check is waived for VFs as of ebfa440c, but
vfio-pci can't actually tell that these particular devices are VFs.
The proposed patch attempts to identify these detached VFs and subsequently
provide this information to vfio-pci so that it knows to also accept the
lack of PCI_COMMAND_MEMORY for these sorts of devices. For now the bit is
only set for s390x but other architectures could opt in to it as well if
needed.
Matthew Rosato (1):
PCI: Introduce flag for detached virtual functions
arch/s390/pci/pci.c | 8 ++++++++
drivers/vfio/pci/vfio_pci_config.c | 11 +++++++----
include/linux/pci.h | 1 +
3 files changed, 16 insertions(+), 4 deletions(-)
--
1.8.3.1