On Tue, Aug 13, 2019 at 12:55:59PM +0000, Haiyang Zhang wrote: > > > > -----Original Message----- > > From: Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx> > > Sent: Tuesday, August 13, 2019 6:14 AM > > To: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx> > > Cc: sashal@xxxxxxxxxx; bhelgaas@xxxxxxxxxx; linux- > > hyperv@xxxxxxxxxxxxxxx; linux-pci@xxxxxxxxxxxxxxx; KY Srinivasan > > <kys@xxxxxxxxxxxxx>; Stephen Hemminger <sthemmin@xxxxxxxxxxxxx>; > > olaf@xxxxxxxxx; vkuznets <vkuznets@xxxxxxxxxx>; linux- > > kernel@xxxxxxxxxxxxxxx > > Subject: Re: [PATCH v3] PCI: hv: Detect and fix Hyper-V PCI domain number > > collision > > > > On Mon, Aug 12, 2019 at 06:20:53PM +0000, Haiyang Zhang wrote: > > > Currently in Azure cloud, for passthrough devices including GPU, the host > > > sets the device instance ID's bytes 8 - 15 to a value derived from the host > > > HWID, which is the same on all devices in a VM. So, the device instance > > > ID's bytes 8 and 9 provided by the host are no longer unique. This can > > > cause device passthrough to VMs to fail because the bytes 8 and 9 are used > > > as PCI domain number. Collision of domain numbers will cause the second > > > device with the same domain number fail to load. > > > > > > As recommended by Azure host team, the bytes 4, 5 have more uniqueness > > > (info entropy) than bytes 8, 9. So now we use bytes 4, 5 as the PCI domain > > > numbers. On older hosts, bytes 4, 5 can also be used -- no backward > > > compatibility issues here. The chance of collision is greatly reduced. In > > > the rare cases of collision, we will detect and find another number that is > > > not in use. > > > > I have not explained what I meant correctly. This patch fixes an > > issue and the "find another number" fallback can be also applied > > to the current kernel without changing the bytes you use for > > domain numbers. > > > > This patch would leave old kernels susceptible to breakage. > > > > Again, I have no Azure knowledge but it seems better to me to > > add a fallback "find another number" allocation on top of mainline > > and send it to stable kernels. Then we can add another patch to > > change the bytes you use to reduce the number of collision. > > > > Please let me know what you think, thanks. > > Thanks for your clarification. > Actually, I hope the stable kernel will be patched to use bytes 4,5 too, > because host provided numbers are persistent across reboots, we like > to use them if possible. > > I think we can either -- > 1) Apply this patch for mainline and stable kernels as well. > 2) Or, break this patch into two patches, and apply both of them for > Mainline and stable kernels. (2) since one patch is a fix and the other one an (optional - however important it is) change. This way if the optional change needs reverting we still have a working kernel. In the end it is up to you - I am just expressing what I think is the most sensible way forward. Lorenzo
