On Friday, June 24, 2011, Ram Pai wrote: > On Thu, Jun 23, 2011 at 10:42:29PM +0200, Rafael J. Wysocki wrote: > > On Thursday, June 23, 2011, Jesse Barnes wrote: > > > On Tue, 21 Jun 2011 17:48:16 -0700 > > > Ram Pai <linuxram@xxxxxxxxxx> wrote: > > > > I assume majority of the platforms will have enough resources to satisfy all > > > > the resource requests, and their BIOS would have done a decent job. > > > > > > > > Even if the BIOS has not done a decent job, and there are enough resources > > > > available we should not see a regression. > > > > > > > > The only platforms that would expose a regression is when resources are under > > > > contention and the BIOS has assigned enough resource to the cardbus bridge but > > > > not to some other device. It will be hard to find such a platform, but I am > > > > sure there is one out somewhere there. > > > > > > > > I am sure we will see; some day, reports of regression because that platform > > > > would have the exact right characteristics to expose the issue. But then that > > > > platform is a highly constrained platform in the first place. Its debatable if > > > > that should be characterised as a regression, or a platform that was riding on > > > > good luck till now. > > > > > > > > Even Oliver's platform is a highly constrained platform, and we probably can > > > > treat his platform as 'riding on good luck till now'. > > > > > > > > We won't be able to satisfy all the platforms with resource constraints. I > > > > think our probable choice is to select a solution that breaks least number of > > > > platforms and special case those broken platforms through kernel command line > > > > parameters. > > > > > > Another option is to hide the new allocation behavior behind a kernel > > > parameter. I know Bjorn has opposed this in the past because really > > > this sort of thing should "just work". But so far it hasn't, and we've > > > had to revert both Bjorn's resource tracking changes as well as the > > > re-allocation code. > > > > > > Hiding it behind a boot option would at least let us improve things > > > over time and potentially switch over to new resource code in the > > > future... > > > > > > Thoughts? > > > > Do I understand correctly that at the moment we have two set of systems, > > one of which works with the new code and doesn't work with the old code > > and the other one conversely? > > Here is the current state: > > (a) As of 2.6.39, for platforms whose BIOS have not allocated enough resources to its > devices, those devices will **continue to not work**. An example of such a platform is > the one whose BIOS has not allocated enough resources to SRIOV BARs. > > (b) With Yinghai's patch > the commit "PCI: update bridge resources to get more big ranges when allocating space (again)" > http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=da7822e5ad71ec9b745b412639f1e5e0ba795a20 > Most of the platforms that were not working in (a) will start working, but will break a few platforms, that > have resource constraints and whose BIOS has not allocated enough resources to some of its devices. > Oliver's and Ben Hutching's platform are two of the known platforms; as of now. > > (c) with my patch all the above platforms will start working. But the 4th patch in the series > raises a genuine concern that it might break resource-constrained platforms with cardbus bridges. > > The question is which one of these is a lesser-evil :) > > Personally I think we should merge all the patches except the 4th patch, and support > Oliver's platform through kernel command line parameter. And I think we should > revert Yinghai's patch for now and merge it with all other patches in the 3.0.1 > timeframe after thorough testing. Well, I think in that case the default behavior should be as for 2.6.39 and there may be a command line switch turning on some new behavior for whoever needs that. This way we'll avoid introducing regressions on systems that don't use the new command line switch and we'' allow systems that don't work without it to be handled. What the new behavior should be is to be determined I guess. Thanks, Rafael -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
