Re: [PATCH] PCI/LINK: Account for BW notification in vector calculation

From: Alex Williamson <alex.williamson@redhat.com>
To: Alex G <mr.nuke.me@gmail.com>
Cc: bhelgaas@google.com, helgaas@kernel.org,
	linux-pci@vger.kernel.org, austin_bolen@dell.com,
	alex_gagniuc@dellteam.com, keith.busch@intel.com,
	Shyam_Iyer@Dell.com, lukas@wunner.de, okaya@kernel.org,
	torvalds@linux-foundation.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] PCI/LINK: Account for BW notification in vector calculation
Date: Tue, 23 Apr 2019 10:22:53 -0600	[thread overview]
Message-ID: <20190423102253.4fd9a019@x1.home> (raw)
In-Reply-To: <88bd2962-bce4-8259-c38f-1a7e9fdde300@gmail.com>

On Tue, 23 Apr 2019 11:03:04 -0500
Alex G <mr.nuke.me@gmail.com> wrote:

> On 4/23/19 10:34 AM, Alex Williamson wrote:
> > On Tue, 23 Apr 2019 09:33:53 -0500
> > Alex G <mr.nuke.me@gmail.com> wrote:
> >   
> >> On 4/22/19 7:33 PM, Alex Williamson wrote:  
> >>> On Mon, 22 Apr 2019 19:05:57 -0500
> >>> Alex G <mr.nuke.me@gmail.com> wrote:  
> >>>> echo 0000:07:00.0:pcie010 |
> >>>> sudo tee /sys/bus/pci_express/drivers/pcie_bw_notification/unbind  
> >>>
> >>> That's a bad solution for users, this is meaningless tracking of a
> >>> device whose driver is actively managing the link bandwidth for power
> >>> purposes.  
> >>
> >> 0.5W savings on a 100+W GPU? I agree it's meaningless.  
> > 
> > Evidence?  Regardless, I don't have control of the driver that's making
> > these changes, but the claim seems unfounded and irrelevant.  
> 
> The number of 5mW/Gb/lane doesn't ring a bell? [1] [2]. Your GPU 
> supports 5Gb/s, so likely using an older, more power hungry process. I 
> suspect it's still within the same order of magnitude.

This doesn't necessarily imply the overall power savings to the
endpoint as a whole though, and it's still irrelevant to the discussion
here.  The driver is doing something reasonable that's generating host
dmesg spam.

> > I'm assigning a device to a VM [snip]
> > I can see why we might want to be notified of degraded links due to signal issues,
> > but what I'm reporting is that there are also entirely normal reasons
> > [snip] we can't seem to tell the difference  
> 
> Unfortunately, there is no way in PCI-Express to distinguish between an 
> expected link bandwidth change and one due to error.

Then assuming every link speed change is an error seems like the wrong
approach.  Should we instead have a callback that drivers can
optionally register to receive link change notifications?  If a driver
doesn't register such a callback then a generic message can be posted,
but if they do, the driver can decide whether this is an error.

> If you're using virt-manager to configure the VM, then virt-manager 
> could have a checkbox to disable link bandwidth management messages. I'd 

What makes us think that this is the only case where such link speed
changes will occur?  Hand waving that a userspace management utility
should go unbind drivers that over-zealously report errors is a poor
solution.

> rather we avoid kernel-side heuristics (like Lukas suggested). If you're 
> confident that your link will operate as intended, and don't want 
> messages about it, that's your call as a user -- we shouldn't decide 
> this in the kernel.

Nor should pci-core decide what link speed changes are intended or
errors.  Minimally we should be enabling drivers to receive this
feedback.  Thanks,

Alex