On Sun, Jan 17, 2021 at 07:16:30PM -0800, Alexander Duyck wrote: > On Sat, Jan 16, 2021 at 12:20 AM Leon Romanovsky <leon@xxxxxxxxxx> wrote: > > > > On Fri, Jan 15, 2021 at 05:48:59PM -0800, Alexander Duyck wrote: > > > On Fri, Jan 15, 2021 at 7:53 AM Leon Romanovsky <leon@xxxxxxxxxx> wrote: > > > > > > > > On Fri, Jan 15, 2021 at 10:06:19AM -0400, Jason Gunthorpe wrote: > > > > > On Thu, Jan 14, 2021 at 05:56:20PM -0800, Alexander Duyck wrote: > > > > > > > > > > > That said, it only works at the driver level. So if the firmware is > > > > > > the one that is having to do this it also occured to me that if this > > > > > > update happened on FLR that would probably be preferred. > > > > > > > > > > FLR is not free, I'd prefer not to require it just for some > > > > > philosophical reason. > > > > > > > > > > > Since the mlx5 already supports devlink I don't see any reason why the > > > > > > driver couldn't be extended to also support the devlink resource > > > > > > interface and apply it to interrupts. > > > > > > > > > > So you are OK with the PF changing the VF as long as it is devlink not > > > > > sysfs? Seems rather arbitary? > > > > > > > > > > Leon knows best, but if I recall devlink becomes wonky when the VF > > > > > driver doesn't provide a devlink instance. How does it do reload of a > > > > > VF then? > > > > > > > > > > I think you end up with essentially the same logic as presented here > > > > > with sysfs. > > > > > > > > The reasons why I decided to go with sysfs are: > > > > 1. This MSI-X table size change is applicable to ALL devices in the world, > > > > and not only netdev. > > > > > > In the PCI world MSI-X table size is a read only value. That is why I > > > am pushing back on this as a PCI interface. > > > > And it stays read-only. > > Only if you come at it directly. What this is adding is a back door > that is visible as a part of the VF sysfs. > > > > > > > > 2. This is purely PCI field and apply equally with same logic to all > > > > subsystems and not to netdev only. > > > > > > Again, calling this "purely PCI" is the sort of wording that has me > > > concerned. I would prefer it if we avoid that wording. There is much > > > more to this than just modifying the table size field. The firmware is > > > having to shift resources between devices and this potentially has an > > > effect on the entire part, not just one VF. > > > > It is internal to HW implementation, dumb device can solve it differently. > > That is my point. I am worried about "dumb devices" that may follow. I > would like to see the steps that should be taken to prevent these sort > of things called out specifically. Basically this isn't just modifying > the PCIe config space, it is actually resizing the PBA and MSI-X > table. Exactly the last line the dumb device can implement differently. The request is simple - configure MSI-X table size to be the new size. > > > > > > > > 3. The sysfs interface is the standard way of configuring PCI/core, not > > > > devlink. > > > > > > This isn't PCI core that is being configured. It is the firmware for > > > the device. You are working with resources that are shared between > > > multiple functions. > > > > I'm ensuring that "lspci -vv .." will work correctly after such change. > > It is PCI core responsibility. > > The current code doesn't work on anything with a driver loaded on it. The problem that no one care about this case, because in opposite to other devices that usually operates in the hypervisor and probed during the boot, the VFs are used differently. They run in VMs, probed there and (usually) not needed in hypervisor. The driver reload would make sense if PF MSI-X table was changed. > In addition the messaging provided is fairly minimal which results in > an interface that will be difficult to understand when it doesn't > work. I'm fond of simple interfaces: 0, EBUSY and EINVAL are common way to inform user. We must remember that this interface is for low-level PCI property and is needed for expert users who needs to squeeze maximum for their VMs out of expensive high speed network card that supports SR-IOV. According to the ebay, the CX6 card costs between 1000 and 1700 USDs, not really home equipment. > In addition there is currently only one piece of hardware that > works with this interface which is the mlx5. It is not different from any other feature, someone should be first. This has very clear purpose, scoped well and understandable when and why it is needed. Kernel is full of devices and features that exist in one device only. > My concern is this is > adding overhead to all VFs that will not be used by most SR-IOV > capable devices. In my view it would make much more sense to have a > top-down approach instead of bottom-up where the PF is registering > interfaces for the VFs. > > If you want yet another compromise I would be much happier with the PF > registering the sysfs interfaces on the VFs rather than the VFs > registering the interface and hoping the PF supports it. At least with > that you are guaranteed the PF will respond to the interface when it > is registered. Thanks a lot, I appreciate it, will take a look now. > > > > > > > > 4. This is how orchestration software provisioning VFs already. It fits > > > > real world usage of SR-IOV, not the artificial one that is proposed during > > > > the discussion. > > > > > > What do you mean this is how they are doing it already? Do you have > > > something out-of-tree and that is why you are fighting to keep the > > > sysfs? If so that isn't a valid argument. > > > > I have Kubernetes and OpenStack, indeed they are not part of the kernel tree. > > They already use sriov_driver_autoprobe sysfs knob to disable autobind > > before even starting. They configure MACs and bind VFs through sysfs/netlink > > already. For them, the read/write of sysfs that is going to be bound to > > the already created VM with known CPU properties, fits perfectly. > > By that argument the same could be said about netlink. What I don't > get is why it is okay to configure the MAC through netlink but > suddenly when we are talking about interrupts it is out of the > question. They belong to different subsystems, while MAC is applicable to the netdev (both PF and VFs), MSI-X is applicable to all devices. I'm not arguing about netlink vs. sysfs, just saying that devlink doesn't fit here. > As far as the binding that is the driver interface which is > more or less grandfathered in anyway as there aren't too many ways to > deal with them as there isn't an alternate interface for the drivers > to define support. > > > > > > > > So the idea to use devlink just because mlx5 supports it, sound really > > > > wrong to me. If it was other driver from another subsystem without > > > > devlink support, the request to use devlink won't never come. > > > > > > > > Thanks > > > > > > I am suggesting the devlink resources interface because it would be a > > > VERY good fit for something like this. By the definition of it: > > > ``devlink`` provides the ability for drivers to register resources, which > > > can allow administrators to see the device restrictions for a given > > > resource, as well as how much of the given resource is currently > > > in use. Additionally, these resources can optionally have configurable size. > > > This could enable the administrator to limit the number of resources that > > > are used. > > > > It is not resource, but HW objects. The devlink doesn't even see the VFs > > as long as they are not bound to the drivers. > > > > This is an example: > > > > [root@vm ~]# echo 0 > /sys/bus/pci/devices/0000\:01\:00.0/sriov_drivers_autoprobe > > [root@vm ~]# echo 0 > /sys/bus/pci/devices/0000\:01\:00.0/sriov_numvfs > > [ 2370.579711] mlx5_core 0000:01:00.0: E-Switch: Disable: mode(LEGACY), nvfs(2), active vports(3) > > [root@vm ~]# echo 2 > /sys/bus/pci/devices/0000\:01\:00.0/sriov_numvfs > > [ 2377.663666] mlx5_core 0000:01:00.0: E-Switch: Enable: mode(LEGACY), nvfs(2), active vports(3) > > [ 2377.777010] pci 0000:01:00.1: [15b3:101c] type 00 class 0x020000 > > [ 2377.784903] pci 0000:01:00.2: [15b3:101c] type 00 class 0x020000 > > [root@vm ~]# devlink dev > > pci/0000:01:00.0 > > [root@vm ~]# lspci |grep nox > > 01:00.0 Ethernet controller: Mellanox Technologies MT28908 Family [ConnectX-6] > > 01:00.1 Ethernet controller: Mellanox Technologies MT28908 Family [ConnectX-6 Virtual Function] > > 01:00.2 Ethernet controller: Mellanox Technologies MT28908 Family [ConnectX-6 Virtual Function] > > > > So despite us having 2 VFs ready to be given to VMs, administrator doesn't > > see them as devices. > > The MSI-X vectors are a resource assigned to hardware objects. It just > depends on how you want to look at things. Right now you have the VFs > register an interface on behalf of the PF. I am arguing it would be > better to have the PF register an interface on behalf of the VFs. > Ultimately the PF is responsible for creating the VFs in the first > place. I don't see it as that much of a leap to have the > mlx5_sriov_enable call register interfaces for the VFs so that you can > configure the MSI-X vectors from the PF, and then tear them down > before it frees the VFs. Having the VFs do the work seems error prone > since it is assuming the interfaces are there on the PF when in all > cases but one (mlx5) it currently isn't. I'm not sure that I understood your last sentence correctly. If VF device is not on hypervisor, it will mean the device is already probed and change of MSI-X table is prohibited. I don't know how you can configure VF devices to be passthrough to the VMs without SR-IOV enable call first. I would say that all devices start their life at the same place where PF is located. > > > > > > > Even looking over the example usage I don't see there being much to > > > prevent you from applying it to this issue. In addition it has the > > > idea of handling changes that cannot be immediately applied already > > > included. Your current solution doesn't have a good way of handling > > > that and instead just aborts with an error. > > > > Yes, because it is HW resource that should be applied immediately to > > make sure that it is honored, before it is committed to the users. > > The problem is you cannot do that at all if the driver is already > loaded. One advantage of using something like devlink is that you > could potentially have the VF driver help to coordinate things so you > could have the case where the VF has the mlx5 driver loaded work > correctly where you would update the MSI-X vector count and then > trigger the driver reload via devlink. The thing is that it is not needed for VFs at all. > > > It is very tempting to use devlink everywhere, but it is really wrong > > tool for this scenario. > > We can agree to disagree there. I am not a fan of sysfs being applied > everywhere either. The problem is it is an easy goto when someone is > looking for a quick and dirty solution and often leads to more > problems later as it usually misses critical path locking issues and > the like. It is fun that you mentioned that devlink as an example of good locking scheme. Without going into to much details, right now Parav and myself are trying to fix devlink locking around reload functionality. It was close to DOA for me when I worked on auxiliary bus patches. So no, devlink is not better. It is another (good) tool that needs more love and care to be real PCI configuration utility. At lest, it should step out of netdev shadow. The block subsystem built whole stack around sysfs and they doesn't seem upset about it. Thanks
