Message-ID: <Y3OCHLiCzOUKLlHa@unreal>
Date:   Tue, 15 Nov 2022 14:12:12 +0200
From:   Leon Romanovsky <leon@...nel.org>
To:     Michal Swiatkowski <michal.swiatkowski@...ux.intel.com>
Cc:     "Samudrala, Sridhar" <sridhar.samudrala@...el.com>,
        netdev@...r.kernel.org, davem@...emloft.net, kuba@...nel.org,
        pabeni@...hat.com, edumazet@...gle.com,
        intel-wired-lan@...ts.osuosl.org, jiri@...dia.com,
        anthony.l.nguyen@...el.com, alexandr.lobakin@...el.com,
        wojciech.drewek@...el.com, lukasz.czapnik@...el.com,
        shiraz.saleem@...el.com, jesse.brandeburg@...el.com,
        mustafa.ismail@...el.com, przemyslaw.kitszel@...el.com,
        piotr.raczynski@...el.com, jacob.e.keller@...el.com,
        david.m.ertman@...el.com, leszek.kaliszczuk@...el.com
Subject: Re: [PATCH net-next 00/13] resource management using devlink reload

On Tue, Nov 15, 2022 at 11:16:58AM +0100, Michal Swiatkowski wrote:
> On Tue, Nov 15, 2022 at 11:32:14AM +0200, Leon Romanovsky wrote:
> > On Tue, Nov 15, 2022 at 10:04:49AM +0100, Michal Swiatkowski wrote:
> > > On Tue, Nov 15, 2022 at 10:11:10AM +0200, Leon Romanovsky wrote:
> > > > On Tue, Nov 15, 2022 at 08:12:52AM +0100, Michal Swiatkowski wrote:
> > > > > On Mon, Nov 14, 2022 at 07:07:54PM +0200, Leon Romanovsky wrote:
> > > > > > On Mon, Nov 14, 2022 at 09:31:11AM -0600, Samudrala, Sridhar wrote:
> > > > > > > On 11/14/2022 7:23 AM, Leon Romanovsky wrote:
> > > > > > > > On Mon, Nov 14, 2022 at 01:57:42PM +0100, Michal Swiatkowski wrote:
> > > > > > > > > Currently the default value for number of PF vectors is number of CPUs.
> > > > > > > > > Because of that there are cases when all vectors are used for PF
> > > > > > > > > and user can't create more VFs. It is hard to set default number of
> > > > > > > > > CPUs right for all different use cases. Instead allow user to choose
> > > > > > > > > how many vectors should be used for various features. After implementing
> > > > > > > > > subdevices this mechanism will be also used to set number of vectors
> > > > > > > > > for subfunctions.
> > > > > > > > > 
> > > > > > > > > The idea is to set vectors for eth or VFs using devlink resource API.
> > > > > > > > > New value of vectors will be used after devlink reinit. Example
> > > > > > > > > commands:
> > > > > > > > > $ sudo devlink resource set pci/0000:31:00.0 path msix/msix_eth size 16
> > > > > > > > > $ sudo devlink dev reload pci/0000:31:00.0
> > > > > > > > > After reload driver will work with 16 vectors used for eth instead of
> > > > > > > > > num_cpus.
> > > > > > > > By saying "vectors", are you referring to MSI-X vectors?
> > > > > > > > If yes, you have specific interface for that.
> > > > > > > > https://lore.kernel.org/linux-pci/20210314124256.70253-1-leon@kernel.org/
> > > > > > > 
> > > > > > > This patch series is exposing a resources API to split the device level MSI-X vectors
> > > > > > > across the different functions supported by the device (PF, RDMA, SR-IOV VFs and
> > > > > > > in future subfunctions). Today this is all hidden in a policy implemented within
> > > > > > > the PF driver.
> > > > > > 
> > > > > > Maybe we are talking about different VFs, but if you refer to PCI VFs,
> > > > > > the amount of MSI-X comes from PCI config space for that specific VF.
> > > > > > 
> > > > > > You shouldn't set any value through netdev as it will cause to
> > > > > > difference in output between lspci (which doesn't require any driver)
> > > > > > and your newly set number.
> > > > > 
> > > > > If I understand correctly, lspci shows the MSI-X number for individual
> > > > > VF. Value set via devlink is the total number of MSI-X that can be used
> > > > > when creating VFs. 
> > > > 
> > > > Yes and no, lspci shows how much MSI-X vectors exist from HW point of
> > > > view. Driver can use less than that. It is exactly as your proposed
> > > > devlink interface.
> > > > 
> > > > 
> > > 
> > > Ok, I have to take a closer look at it. So, are You saing that we should
> > > drop this devlink solution and use sysfs interface fo VFs or are You
> > > fine with having both? What with MSI-X allocation for subfunction?
> > 
> > You should drop for VFs and PFs and keep it for SFs only.
> > 
> 
> I understand that MSI-X for VFs can be set via sysfs interface, but what
> with PFs? 

PFs are even more tricker than VFs, as you are changing that number
while driver is bound. This makes me wonder what will be lspci output,
as you will need to show right number before driver starts to load.

You need to present right value if user decided to unbind driver from PF too.

> Should we always allow max MSI-X for PFs? So hw_max - used -
> sfs? Is it save to call pci_enable_msix always with max vectors
> supported by device?

I'm not sure. I think that it won't give you much if you enable
more than num_online_cpu().

> 
> I added the value for PFs, because we faced a problem with MSI-X
> allocation on 8 port device. Our default value (num_cpus) was too big
> (not enough vectors in hw). Changing the amount of vectors that can be
> used on PFs was solving the issue.

We had something similar for mlx5 SFs, where we don't have enough vectors.
Our solution is simply to move to automatic shared MSI-X mode. I would
advocate for that for you as well. 

> 
> Let me write an example. As default MSI-X for PF is set to num_cpus, the
> platform have 128 CPUs, we have 8 port device installed there and still
> have 1024 vectors in HW (I simplified because I don't count additional
> interrupts). We run out of vectors, there is 0 vectors that can be used
> for VFs. Sure, it is easy to handle, we can divide PFs interrupts by 2
> and will end with 512 vectors for VFs. I assume that with current sysfs
> interface in this situation MSI-X for VFs can be set from 0 to 512? What
> if user wants more? If there is a PFs MSI-X value which can be set by
> user, user can decrease the value and use more vectors for VFs. Is it
> possible in current VFs sysfs interface? I mean, setting VFs MSI-X
> vectors to value that will need to decrease MSI-X for PFs.

You can't do it and this limitation is because PF is bound. You can't change
that number while driver is running. AFAIR, such change will be PCI spec
violation.

> 
> > > 
> > > > > As Jake said I will fix the code to track both values. Thanks for pointing the patch.
> > > > > 
> > > > > > 
> > > > > > Also in RDMA case, it is not clear what will you achieve by this
> > > > > > setting too.
> > > > > >
> > > > > 
> > > > > We have limited number of MSI-X (1024) in the device. Because of that
> > > > > the amount of MSI-X for each feature is set to the best values. Half for
> > > > > ethernet, half for RDMA. This patchset allow user to change this values.
> > > > > If he wants more MSI-X for ethernet, he can decrease MSI-X for RDMA.
> > > > 
> > > > RDMA devices doesn't have PCI logic and everything is controlled through
> > > > you main core module. It means that when you create RDMA auxiliary device,
> > > > it will be connected to netdev (RoCE and iWARP) and that netdev should
> > > > deal with vectors. So I still don't understand what does it mean "half
> > > > for RDMA".
> > > >
> > > 
> > > Yes, it is controlled by module, but during probe, MSI-X vectors for RDMA
> > > are reserved and can't be used by ethernet. For example I have
> > > 64 CPUs, when loading I get 64 vectors from HW for ethernet and 64 for
> > > RDMA. The vectors for RDMA will be consumed by irdma driver, so I won't
> > > be able to use it in ethernet and vice versa.
> > > 
> > > By saing it can't be used I mean that irdma driver received the MSI-X
> > > vectors number and it is using them (connected them with RDMA interrupts).
> > > 
> > > Devlink resource is a way to change the number of MSI-X vectors that
> > > will be reserved for RDMA. You wrote that netdev should deal with
> > > vectors, but how netdev will know how many vectors should go to RDMA aux
> > > device? Does there an interface for setting the vectors amount for RDMA
> > > device?
> > 
> > When RDMA core adds device, it calls to irdma_init_rdma_device() and
> > num_comp_vectors is actually the number of MSI-X vectors which you want
> > to give to that device.
> > 
> > I'm trying to say that probably you shouldn't reserve specific vectors
> > for both ethernet and RDMA and simply share same vectors. RDMA applications
> > that care about performance set comp_vector through user space verbs anyway.
> > 
> 
> Thanks for explanation, appriciate that. In our driver num_comp_vectors for
> RDMA is set during driver probe. Do we have any interface to change
> num_comp_vectors while driver is working?

No, as this number is indirectly exposed to the user space.

> 
> Sorry, I am not fully familiar with RDMA. Can user app for RDMA set
> comp_vector to any value or only to max which is num_comp_vectors given
> for RDMA while creating aux device?

comp_vector logically equal to IRQ number and this is how RDMA
applications controls to which interrupt deliver completions.

" The CQ will use the completion vector comp_vector for signaling
  completion events; it must be at least zero and less than
  context->num_comp_vectors. "
https://man7.org/linux/man-pages/man3/ibv_create_cq.3.html

> 
> Assuming that I gave 64 MSI-X for RDMA by setting num_comp_vectors to
> 64, how I will know if I can or can't use these vectors in ethernet?

Why should you need to know? Vectors are not exclusive and they can be
used by many applications at the same time. The thing is that it is far
fetch to except that high performance RDMA applications and high
performance ethernet can coexist on same device at the same time.

Thanks

> 
> Thanks
> 
> > Thanks
> > 
> > > 
> > > Thanks
> > > 
> > > > Thanks

Powered by blists - more mailing lists