On Wed, Nov 18, 2020 at 12:36:26PM -0700, David Ahern wrote: > On 11/18/20 11:38 AM, Jason Gunthorpe wrote: > > On Wed, Nov 18, 2020 at 11:03:24AM -0700, David Ahern wrote: > > > >> With Connectx-4 Lx for example the netdev can have at most 63 queues > > > > What netdev calls a queue is really a "can the device deliver > > interrupts and packets to a given per-CPU queue" and covers a whole > > spectrum of smaller limits like RSS scheme, # of available interrupts, > > ability of the device to create queues, etc. > > > > CX4Lx can create a huge number of queues, but hits one of these limits > > that mean netdev's specific usage can't scale up. Other stuff like > > RDMA doesn't have the same limits, and has tonnes of queues. > > > > What seems to be needed is a resource controller concept like cgroup > > has for processes. The system is really organized into a tree: > > > > physical device > > mlx5_core > > / | \ \ (aux bus) > > netdev rdma vdpa SF etc > > | (aux bus) > > mlx5_core > > / \ (aux bus) > > netdev vdpa > > > > And it does make a lot of sense to start to talk about limits at each > > tree level. > > > > eg the top of the tree may have 128 physical interrupts. With 128 CPU > > cores that isn't enough interrupts to support all of those things > > concurrently. > > > > So the user may want to configure: > > - The first level netdev only gets 64, > > - 3rd level mlx5_core gets 32 > > - Final level vdpa gets 8 > > > > Other stuff has to fight it out with the remaining shared interrupts. > > > > In netdev land # of interrupts governs # of queues > > > > For RDMA # of interrupts limits the CPU affinities for queues > > > > VPDA limits the # of VMs that can use VT-d > > > > The same story repeats for other less general resources, mlx5 also > > has consumption of limited BAR space, and consumption of some limited > > memory elements. These numbers are much bigger and may not need > > explicit governing, but the general concept holds. > > > > It would be very nice if the limit could be injected when the aux > > device is created but before the driver is bound. I'm not sure how to > > manage that though.. > > > > I assume other devices will be different, maybe some devices have a > > limit on the number of total queues, or a limit on the number of > > VDPA or RDMA devices. > > A lot of low level resource details that need to be summarized into a > nicer user / config perspective to specify limits / allocations. Well, now that we have the aux bus stuff there is a nice natural place to put things.. The aux bus owner device (mlx5_core) could have a list of available resources Each aux bus device (netdev/rdma/vdpa) could have a list of consumed resources Some API to place a limit on the consumed resources at each aux bus device. The tricky bit is the auto-probing/configure. By the time the user has a chance to apply a limit the drivers are already bound and have already done their setup. So each subsystem has to support dynamically imposing a limit.. And I simplified things a bit above too, we actually have two kinds of interrupt demand: sharable and dedicated. The actual need is to carve out a bunch of dedicated interrupts and only allow subsystems that are doing VT-d guest interrupt assignment to consume them (eg VDPA) Jason
