Skip to contentSkip to conversation

New: PCI can't map correct memory resource if the BAR is 64-bit and then leads to system hang during booting up

Bug 203485 Bjorn Helgaas 2 participants 2 messages May 15, 2019 – May 16, 2019 1 message not in the archive PCI

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]/

#1 lore raw

I should have forwarded this to the list earlier. Maybe somebody wants to work on this?

I *think* the problem is:

- BIOS sets MTRR for 0x40_00000000-0x5f_ffffffff to be write-combining (WC)
- That covers part of this PCI aperture: root bus resource [mem
0x40_00000000-0x7f_ffffffff window]
- That aperture includes a frame buffer: pci 0000:00:02.0: reg 0x18:
[mem 0x40_00000000-0x40_0fffffff 64bit pref]
- Linux assigned an LPSS BAR to the WC area: pci 0000:00:15.0: BAR
0: assigned [mem 0x40_10000000-0x40_10000fff 64bit]
- The LPSS device doesn't work if MMIO accesses to it are combined via WC

I'm not an x86/MTRR/PAT expert, but I think we should be able to make this work correctly by changing that MTRR from WC to UC (I don't know if there are BIOS/OS interface implications with this), or maybe using PAT to override the MTRR WC setting to be UC for part or all of that aperture. The frame buffer driver should be smart enough to request WC if it wants it.

---------- Forwarded message ---------
From: <[email protected]>
Date: Fri, May 3, 2019 at 3:31 AM
Subject: [Bug 203485] New: PCI can't map correct memory resource if the BAR is 64-bit and then leads to system hang during booting up
To: <[email protected]>

https://bugzilla.kernel.org/show_bug.cgi?id=203485

Bug ID: 203485
Summary: PCI can't map correct memory resource if the BAR is
64-bit and then leads to system hang during booting up
Product: Drivers
Version: 2.5
Kernel Version: v5.1-rc7
Hardware: All
OS: Linux
Tree: Mainline
Status: NEW
Severity: normal
Priority: P1
Component: PCI
Assignee: [email protected]
Reporter: [email protected]
Regression: No

Created attachment 282597
--> https://bugzilla.kernel.org/attachment.cgi?id=282597&action=edit
dmesg.log

This issue has been found on loading intel-lpss-pci driver, the pdev->resource passing to its probe function is not correct. And it leads to system hangs while loading the driver.

This is the memory addresses driver got, the start and end are more like the region declared in BAR

[    0.718591] pci 0000:00:15.0: BAR 0: assigned [mem 0x4010000000-0x4010000fff
64bit]
[    0.718613] pci 0000:00:15.1: BAR 0: assigned [mem 0x4010001000-0x4010001fff
64bit]
[    0.718626] pci 0000:00:1d.0: BAR 13: assigned [io  0x4000-0x6fff]
[    0.718634] pci 0000:00:1e.0: BAR 0: assigned [mem 0x4010002000-0x4010002fff

64bit]

[  154.417550] [/home/acelan/linux/drivers/mfd/intel-lpss-pci.c:31]
intel_lpss_pci_probe(): 0000:00:15.0 - mem = 0x0000000047753398, mem->start =
0x0000000010000000, mem->end = 0x0000000010000FFF
[  154.417571] [/home/acelan/linux/drivers/mfd/intel-lpss-pci.c:31]
intel_lpss_pci_probe(): 0000:00:15.1 - mem = 0x0000000047754398, mem->start =
0x0000000010001000, mem->end = 0x0000000010001FFF
[  154.417587] [/home/acelan/linux/drivers/mfd/intel-lpss-pci.c:31]

intel_lpss_pci_probe(): 0000:00:1e.0 - mem = 0x000000004775A398, mem->start = 0x0000000010002000, mem->end = 0x0000000010002FFF

Below is the correct one I copy from 32-Bit BAR

[    3.511100] [/home/acelan/bionic/drivers/mfd/intel-lpss-pci.c:47]

intel_lpss_pci_probe(): 0000:00:15.0 - mem = 0x00000000AB282380, mem->start = 0x00000000ED137000, mem->end = 0x00000000ED137FFF

--
You are receiving this mail because:
You are watching the assignee of the bug.

#2 lore raw
Quoted 14 lines · Bjorn Helgaas

I should have forwarded this to the list earlier. Maybe somebody wants to work on this?

I *think* the problem is:

- BIOS sets MTRR for 0x40_00000000-0x5f_ffffffff to be write-combining (WC)
- That covers part of this PCI aperture: root bus resource [mem
0x40_00000000-0x7f_ffffffff window]
- That aperture includes a frame buffer: pci 0000:00:02.0: reg 0x18:
[mem 0x40_00000000-0x40_0fffffff 64bit pref]
- Linux assigned an LPSS BAR to the WC area: pci 0000:00:15.0: BAR
0: assigned [mem 0x40_10000000-0x40_10000fff 64bit]
- The LPSS device doesn't work if MMIO accesses to it are combined via WC

Quoted 6 lines

I'm not an x86/MTRR/PAT expert, but I think we should be able to make this work correctly by changing that MTRR from WC to UC (I don't know if there are BIOS/OS interface implications with this), or maybe using PAT to override the MTRR WC setting to be UC for part or all of that aperture. The frame buffer driver should be smart enough to request WC if it wants it.

The window for the PCI resources, provided by PCI Root Bridge, may be split to many parts as described by _CRS in ACPI, and BIOS should be consistent in providing both.

I believe there is no issue with Linux in current state. Be "smart" here might give a downside(s) like reducing coverage to test exactly something as above BIOS issue.

My 2 cents.

Signature · 2 lines
With Best Regards,
Andy Shevchenko