On Tue, May 28, 2019 at 2:09 PM Shawn Anastasio <shawn@xxxxxxxxxx> wrote: > > > > On 5/27/19 11:01 PM, Oliver wrote: > > On Tue, May 28, 2019 at 8:56 AM Shawn Anastasio <shawn@xxxxxxxxxx> wrote: > >> > >> Hello all, > >> > >> This patch set implements support for user-specified PCI resource > >> alignment on the pseries platform for hotplugged PCI devices. > >> Currently on pseries, PCI resource alignments specified with the > >> pci=resource_alignment commandline argument are ignored, since > >> the firmware is in charge of managing the PCI resources. In the > >> case of hotplugged devices, though, the kernel is in charge of > >> configuring the resources and should obey alignment requirements. > > > > Are you using hotplug to work around SLOF (the OF we use under qemu) > > not aligning BARs to 64K? It looks like there is a commit in SLOF to > > fix that (https://git.qemu.org/?p=SLOF.git;a=commit;f=board-qemu/slof/pci-phb.fs;h=1903174472f8800caf50c959b304501b4c01153c). > > > > No, my application actually requires PCI hotplug at run-time. > > >> The current behavior of ignoring the alignment for hotplugged devices > >> results in sub-page BARs landing between page boundaries and > >> becoming un-mappable from userspace via the VFIO framework. > >> This issue was observed on a pseries KVM guest with hotplugged > >> ivshmem devices. > > > >> With these changes, users can specify an appropriate > >> pci=resource_alignment argument on boot for devices they wish to use > >> with VFIO. > >> > >> In the future, this could be extended to provide page-aligned > >> resources by default for hotplugged devices, similar to what is done > >> on powernv by commit 382746376993 ("powerpc/powernv: Override > >> pcibios_default_alignment() to force PCI devices to be page aligned"). > > > > Can we make aligning the BARs to PAGE_SIZE the default behaviour? The > > BAR assignment process is complex enough as-is so I'd rather we didn't > > add another platform hack into the mix. > > Absolutely. This will still require the existing changes so that the > custom alignment isn't flat-out ignored on pseries, but I can set > it to default to PAGE_SIZE as well, similar to how it's done on PowerNV. > I've just pushed a v3 to fix a typo and I'll incorporate this change > in v4. I was thinking we could get rid of the ppcmd callback and do it in kernel/pci-common.c. PowerNV is the only platform that implements the callback and the pseries implementation is going to be identical so I don't think there's much of point in keeping the callback. > >> Feedback is appreciated. > >> > >> Thanks, > >> Shawn > >> > >> Shawn Anastasio (3): > >> PCI: Introduce pcibios_ignore_alignment_request > >> powerpc/64: Enable pcibios_after_init hook on ppc64 > >> powerpc/pseries: Allow user-specified PCI resource alignment after > >> init > >> > >> arch/powerpc/include/asm/machdep.h | 6 ++++-- > >> arch/powerpc/kernel/pci-common.c | 9 +++++++++ > >> arch/powerpc/kernel/pci_64.c | 4 ++++ > >> arch/powerpc/platforms/pseries/setup.c | 22 ++++++++++++++++++++++ > >> drivers/pci/pci.c | 9 +++++++-- > >> 5 files changed, 46 insertions(+), 4 deletions(-) > >> > >> -- > >> 2.20.1 > >>
