Re: [Bug 84761] New: LSI controller not found when specifying pci=assign-busses

[Bjorn Helgaas <bhelgaas@google.com>]

[+cc Andreas, linux-pci, thanks for the bugzilla; please continue
discussion in email]

On Wed, Sep 17, 2014 at 7:48 AM,  <bugzilla-daemon@bugzilla.kernel.org> wrote:
> https://bugzilla.kernel.org/show_bug.cgi?id=84761
>
>             Bug ID: 84761
>            Summary: LSI controller not found when specifying
>                     pci=assign-busses
>            Product: Drivers
>            Version: 2.5
>     Kernel Version: linux-3.17.0-rc2
>           Hardware: All
>                 OS: Linux
>               Tree: Mainline
>             Status: NEW
>           Severity: normal
>           Priority: P1
>          Component: PCI
>           Assignee: drivers_pci@kernel-bugs.osdl.org
>           Reporter: dmilburn@redhat.com
>         Regression: No
>
> Created attachment 150651
>   --> https://bugzilla.kernel.org/attachment.cgi?id=150651&action=edit
> Patch to change pcibios_assign_all_busses check in pci_scan_bridge()
>
> When booting with kernel command line option "pci=assign-busses", LSI
> controller
> is no longer found.
>
> 05:00.0 SCSI storage controller: LSI Logic / Symbios Logic SAS1068E PCI-Express
> Fusion-MPT SAS (rev 08)
>
> It seems the problem is in pci_scan_bridge (drivers/pci/probe.c):
>
> [    2.542563] PCI_READ_BRIDGE_BASES:
> [    2.545953] PCI_SCAN_BRIDGE:
> [    2.548823] scanning [bus 01-01] behind bridge, pass 0
> [    2.553947] PCI_SCAN_BRIDGE:
> [    2.556818] scanning [bus 05-05] behind bridge, pass 0 <=========PROBLEM
> [    2.561942] PCI_SCAN_BRIDGE:
> [    2.564812] scanning [bus 06-06] behind bridge, pass 0
> [    2.569936] PCI_SCAN_BRIDGE:
> [    2.572806] scanning [bus 03-03] behind bridge, pass 0
> [    2.577930] PCI_SCAN_BRIDGE:
> [    2.580800] scanning [bus 02-02] behind bridge, pass 0
> [    2.585923] PCI_SCAN_BRIDGE:
> [    2.588793] scanning [bus 04-04] behind bridge, pass 0
>
> If I change the pass 0 check from !pcibios_assign_all_busses() to
> pcibios_assign_all_busses() it finds the LSI controller; however, it looks
> like pci_scan_bridge has checked !pcibios_assisng_all_bussses() for a
> very long time.
>
> (changed code in pci_scan_bridge, causes driver to head down that first path)
>   if ((secondary || subordinate) && pcibios_assign_all_busses() &&
>       !is_cardbus && !broken) {
>
> [    2.560225] PCI_SCAN_BRIDGE: secondary 1 subordinate 1 is_cardbus 0 broken 0
> [    2.567252] PCI_SCAN_BRIDGE: !pcibios_assign_all_busses() 0
>  .
>  .
> [    2.849864] scanning [bus 05-05] behind bridge, pass 0   <=====SCANNING
> [    2.854988] PCI_SCAN_BRIDGE: CHECKING FOR ASSIGN_ALL_BUSSES
> [    2.860544] PCI_SCAN_BRIDGE: secondary 5 subordinate 5 is_cardbus 0 broken 0
> [    2.867572] PCI_SCAN_BRIDGE: !pcibios_assign_all_busses() 0
> [    2.873126] PCI_SCAN_BRIDGE: ASSIGN_ALL_BUSSES child           (null)
> [    2.879546] PCI_ADD_NEW_BUS busnr 5:
> [    2.883109] PCI_ALLOC_CHILD_BUS:
> [    2.886327] PCI_SET_BUS_SPEED:
> [    2.889407] PCI_BUS_INSERT_BUSN_RES:
> [    2.892971] PCI_SCAN_CHILD_BUS:
> [    2.896100] PCI_SCAN_CHILD_BUS: scanning bus
> [    2.900357] PCI_SCAN_SLOT:
> [    2.903053] PCI_SCAN_SINGLE_DEVICE:
> [    2.906529] PCI_SCAN_DEVICE: devfn 0
> [    2.910093] PCI_BUS_READ_DEV_VENDOR_ID: devfn 0
> [    2.914608] PCI_ALLOC_DEV:
> [    2.917305] PCI_SETUP_DEVICE:
> [    2.920262] SET_PCIE_PORT_TYPE:       =====DEVICE FOUND BELOW=====
> [    2.923399] pci 0000:05:00.0: [1000:0058] type 00 class 0x010000
> [    2.923401] [1000:0058] type 00 class 0x010000
> [    2.927841] pci 0000:05:00.0: reg 0x10: [io  0xec00-0xecff]
> [    2.927852] pci 0000:05:00.0: reg 0x14: [mem 0xde2ec000-0xde2effff 64bit]
> [    2.927863] pci 0000:05:00.0: reg 0x1c: [mem 0xde2f0000-0xde2fffff 64bit]
> [    2.927877] pci 0000:05:00.0: reg 0x30: [mem 0xde100000-0xde1fffff pref]
> [    2.927879] PCI_DEVICE_ADD:

Tangent: why do you need "pci=assign-busses"?  If that's necessary to
make some device work, I think that's a different bug in itself.

Is this a regression?

Can you attach complete dmesg logs with and without "pci=assign-busses"?

We have a couple patches that are candidates for reversion before
v3.17 because of similar issues (I think they're good patches, but we
may need some additional work to fix some other problems they
exposed).  If you want to try them, they're here:
https://git.kernel.org/cgit/linux/kernel/git/helgaas/pci.git/log/?h=pci/reverts

Bjorn