Messages in this thread | | | From | Bjorn Helgaas <> | Date | Mon, 6 Feb 2012 09:14:26 -0800 | Subject | Re: [PATCH 02/15] PCI: Add iobusn_resource |
| |
On Fri, Feb 3, 2012 at 6:30 PM, Yinghai Lu <yinghai@kernel.org> wrote: > On Fri, Feb 3, 2012 at 3:38 PM, Benjamin Herrenschmidt > <benh@kernel.crashing.org> wrote: >> On Fri, 2012-02-03 at 08:36 -0800, Bjorn Helgaas wrote: >>> On Wed, Feb 1, 2012 at 2:43 PM, Yinghai Lu <yinghai@kernel.org> wrote: >>> > also add busn_res into struct pci_bus. >>> > >>> > will use them to have bus number resource tree.
>> To be honest that whole business with bus numbers in struct resource >> seems like gratuituous bloat & over engineering to me ... > > ah, i thought it is simple enough, and should be done before already. > >> Does it actually solve a specific problem or serve a purpose ? > > very beginning is for one IBM x3950... > > https://bugzilla.novell.com/show_bug.cgi?id=735909 > > ++max searching valid bus number range is out of boundary of peer root > bus range. > > later, found the code could help: pcie hotplug, pci bus rescan and > even remove some hacks for/from cardbus field. > > now with this patchset (+ one patch that is not sent out yet), I could > even use setpci/pci rescan to move pci bus around.
The current bus number allocation in pci_scan_bridge() is very simple -- we just add one to the highest-numbered bridge we've seen. That works in many cases, but it means any holes in the range of used bus numbers are just wasted.
Here's a case I tripped over where the current scheme fails. The machine boots with bridge 00:03.0 to [bus 05-19]. Now we hot-add a hierarchy that is currently configured like this:
05:00.0 bridge to [bus 06-1e] 06:05.0 bridge to [bus 1c] 06:06.0 bridge to [bus 1d] 1c:00.0 NIC0 1d:00.0 NIC1
[bus 1a-1e] is inaccessible because the 00:03.0 aperture only includes [bus 05-19]. There's plenty of space to reconfigure the hierarchy, e.g., to make 05:00.0 lead to [bus 06-19], 06:05.0 lead to [bus 18], and 06:06.0 lead to [bus 19]. But we aren't smart enough to do that.
So I think it's important to improve bus number allocation to make hot plug more robust, but I don't think the current series is understandable or maintainable yet.
Bjorn
| |