On Thu, Aug 16, 2018 at 08:50:42PM +0000, Derrick, Jonathan wrote: > On Thu, 2018-08-16 at 15:31 -0500, Bjorn Helgaas wrote: > > On Thu, Aug 16, 2018 at 03:50:47PM +0000, Derrick, Jonathan wrote: > > > On Thu, 2018-08-16 at 08:49 -0700, Matthew Wilcox wrote: > > > > On Wed, Aug 15, 2018 at 03:26:39PM -0600, Jon Derrick wrote: > > > > > Some users may want to disable downstream port containment > > > > > (DPC), > > > > > so > > > > > give them this option > > > > > > > > Is it possible they might only want to disable DPC on a subset of > > > > the > > > > hierarchy rather than globally? > > > > > > Absolutely. I was hoping Logan's pci dev_str would land because I > > > have > > > a few others I want to convert to that api for granular tuning > > > > What's the use case here? I acknowledge there are cases where we > > need > > them, but I'm not a fan of kernel parameters in general because > > they're a real hassle for users. > > > > Is there something wrong with DPC? Is there some way we can make it > > smarter so it does the right thing automatically? > I've encountered some BIOS or switches (not sure who) who've appeared > to have enabled DPC by default, prior to kernel setup. Some users may > just not want this, but it was primarily intended for debugging when I > conceived it. > > There's also the ongoing thread in linux-pci about err handling in PPC > EEH, who may also desire to disable DPC until those issues are > resolved. I haven't caught up on that thread yet. If DPC is incompatible with PPC EEH, there's always the possibility of a switch in the code to disable DPC automatically on PPC. > It can also be disabled with setpci, but is that any less of a hassle? > Genuine question to understand your point of view. Keith already answered here; setpci is primarily for debugging and shouldn't be recommended as normal practice. > > I'm more OK with a blanket "nodpc" switch intended for debugging. > > If we add the complexity of subsets of the hierarchy it starts > > sounding like an administrative thing that makes me more hesitant. > ... > To again understand your point of view, is there anything wrong with > administrative things in kernel boot parameters? There will be cases > where we may want to deviate from default or global pci=* parameters > for certain hierarchies and they can't necessarily be set after the > fact (ex, hpmemsize). "There will be cases" sounds like we're doing something that *might* be useful in the future. It's better if we wait until we actually discover a need for something. There's also a tendency among users to trip over a problem, discover a boot parameter like "pci=nomsi", and conclude that the problem is "fixed". In fact, we want the bug report so we can fix the kernel so no boot parameter is needed. I agree there are things that can't be set after boot. Is this one of them? This seems like something that could be controlled at run-time. But even there, I will ask what the requirement for it is, because we don't want to clutter sysfs with things only needed for debugging. Boot parameters are a hassle because it's hard to build a user interface on top of them, and they require a reboot to take effect.
