4.4 BCM5301X ARM regression "External imprecise Data abort"

[Lucas Stach]

Am Freitag, den 08.04.2016, 08:45 +0200 schrieb Rafał Miłecki:
> On 4 April 2016 at 23:23, Hauke Mehrtens <hauke at hauke-m.de> wrote:
> > On 04/04/2016 11:08 PM, Scott Branden wrote:
> >> On 16-04-03 11:13 PM, Rafał Miłecki wrote:
> >>> I got regression reports from Netgear R8000 (BCM4709A0) users and did
> >>> some testing & regression tracking with Aditya.
> >>>
> >>> It happens that Linux 4.4 doesn't boot due to the following commits:
> >>> bbeb920 ("ARM: 8422/1: enable imprecise aborts during early kernel
> >>> startup")
> >>> 9254970 ("ARM: 8447/1: catch pending imprecise abort on unmask")
> >>> 937b123 ("ARM: BCM5301X: remove workaround imprecise abort fault
> >>> handler")
> >>>
> >>> In kernel 4.3 we got that abort workaround which was resulting in:
> >>> [    5.007128] Freeing unused kernel memory: 212K (c0435000 - c046a000)
> >>> [    5.694632] init: Console is alive
> >>> [    5.698169] init: - watchdog -
> >>> [    5.701470] External imprecise Data abort at addr=0x0, fsr=0x1406
> >>> ignored.
> >>> As you can see, this abort was happening soon after freeing unused
> >>> memory and ignoring it *once* did the trick. It was never appearing
> >>> again.
> >
> > I assume it only can throw one of these and if it is deactivated it will
> > ignore the next one or overwrite it. So it could be that more than one
> > is thrown here.
> >
> >>> With 4.4 similar (or the same?) abort happens earlier (during PCI host
> >>> driver init) and doesn't get ignored:
> >>> [    2.478461] pci 0000:00:00.0: PCI bridge to [bus 01]
> >>> [    2.483451] pci 0000:00:00.0:   bridge window [mem
> >>> 0x08000000-0x085fffff]
> >>> [    2.599449] pcie_iproc_bcma bcma0:8: PCI host bridge to bus 0001:00
> >>> [    2.605744] pci_bus 0001:00: root bus resource [mem
> >>> 0x40000000-0x47ffffff]
> >>> [    2.612657] pcie_iproc_bcma bcma0:8: link: UP
> >>> [    2.617241] PCI: bus0: Fast back to back transfers disabled
> >>> [    2.622845] pci 0001:00:00.0: bridge configuration invalid ([bus
> >>> 00-00]), reconfiguring
> >>> [    2.631297] PCI: bus1: Fast back to back transfers disabled
> >>> [    2.636887] pci 0001:01:00.0: bridge configuration invalid ([bus
> >>> 00-00]), reconfiguring
> >>> [    2.645035] Unhandled fault: imprecise external abort (0x1406) at
> >>> 0x00000000
> >>> (see 4.4.txt for the backtrace)
> >>>
> >>> At first I was hoping that we simply need to re-add the removed
> >>> workaround. I tried it but it appeared that one abort is immediately
> >>> followed by another:
> >>> [    2.936895] pci 0001:01:00.0: bridge configuration invalid ([bus
> >>> 00-00]), reconfiguring
> >>> [    2.945053] External imprecise Data abort at addr=0x0, fsr=0x1406
> >>> ignored.
> >>> [    2.951966] Unhandled fault: imprecise external abort (0x1406) at
> >>> 0x00000000
> >>>
> >>> So it seems that commits bbeb920 and 9254970 broke something in PCI
> >>> host initialization (or maybe just exposed another bug?). Instead of
> >>> getting an abort once and late we are getting now many of them and a
> >>> bit earlier.
> >
> > These commits mad the kernel earlier "listen" to such errors, so that
> > they will be shown at the time they occur and not sometime later.
> 
> So AFAIU with kernel 4.3:
> 1) Aborts were masked (silent) until "Freeing unused kernel memory"
> 2) There was one (silent) abort caused by a bootloader
> 3) There were likely multiple aborts (silent) during early PCI init
> 4) After unmasking we got only a single abort reported and we were ignoring it
> 
> With kernel 4.4:
> 1) All aborts are reported immediately
> 2) Abort caused by a bootloader gets ignored by ARM code:
> "Hit pending asynchronous external abort (FSR=0x00001c06) during first unmask"
> thanks to 9254970 ("ARM: 8447/1: catch pending imprecise abort on unmask")
> 3) There are still multiple aborts during PCI init (reported immediately now)
> 4) To work as before (in 4.3) we should ignore all aborts, not only the 1st one
> 
> Of course proposed solution is an ugly workaround, we should have no
> aborts reported in the first place.
> 
A master abort on the PCI bus during probe of the PCI config space
(device enumeration) is expected. Most host bridges ignore those errors
and just return 0 for the read transaction.

Some bridges forward the error onto the AXI/AMBA bus and thus cause
imprecise external aborts on the ARM core. If your host bridge doesn't
have a way to disable error forwarding during PCI bus probe you need to
install an abort handler. Most implementations based on the designware
PCIe core do this already.

Regards,
Lucas