Dear Bjorn Helgaas,
> On Mon, Oct 21, 2013 at 2:16 AM, Pratyush Anand
>
> <pratyush.anand@xxxxxxxxx> wrote:
> > Hi,
> >
> > On Sat, Oct 19, 2013 at 7:03 AM, Marek Vasut <marex@xxxxxxx> wrote:
> >> Dear Tim Harvey,
> >>
> >> > An imprecise abort is triggered when a port behind a switch is
> >> > accessed and no device is present. At enumeration, imprecise aborts
> >> > are not enabled
> >
> > I do not know the complete history. But I do not understand, why
> > imprecise aborts are not enabled during kernel enumeration.
> > What happens when PCIe switch is already connected before boot?
> > pci_common_init will try to read vendor/product ID for a non-existing
> > function during kernel enumeration itself, which should result in
> > imprecise external abort. In fact, I had observed with my system that if
> > a switch is connected to PCIe RC, then abort occurs during kernel
> > initialization itself. So a proper handling was needed. We called
> > hook_fault_code from
> > a subsys_init pcie_init (or similar) function, which insured that abort
> > handler is called whenever, software tries to read vendor/product id of
> > a non-existing function
> >
> >> > thus this ends up getting deferred until the kernel has completed
> >> > init. At
> >> > that point we must not adjust PC - the handler must do nothing, but a
> >> > handler must exist.
> >> >
> >> > This fixes random crashes that occur right after freeing init.
> >> > This is against linux-pci/host-imx6.
> >> >
> >> > Acked-by: Marek Vasut <marex@xxxxxxx>
> >> > Tested-by: Marek Vasut <marex@xxxxxxx>
> >> > Signed-off-by: Tim Harvey <tharvey@xxxxxxxxxxxxx>
> >>
> >> Expanding CC a bit, let's have more eyes on this.
> >>
> >> > ---
> >> >
> >> > drivers/pci/host/pci-imx6.c | 6 ------
> >> > 1 file changed, 6 deletions(-)
> >> >
> >> > diff --git a/drivers/pci/host/pci-imx6.c b/drivers/pci/host/pci-imx6.c
> >> > index 966bac6..90fce05 100644
> >> > --- a/drivers/pci/host/pci-imx6.c
> >> > +++ b/drivers/pci/host/pci-imx6.c
> >> > @@ -200,12 +200,6 @@ static int pcie_phy_write(void __iomem *dbi_base,
> >> > int
> >> > addr, int data) static int imx6q_pcie_abort_handler(unsigned long
> >> > addr,
> >> >
> >> > unsigned int fsr, struct pt_regs *regs)
> >> >
> >> > {
> >> >
> >> > - /*
> >> > - * If it was an imprecise abort, then we need to correct the
> >> > - * return address to be _after_ the instruction.
> >> > - */
> >> > - if (fsr & (1 << 10))
> >> > - regs->ARM_pc += 4;
> >
> > Just for knowledge, is there any case other than read of IDs which might
> > cause imprecise abort? If not, then this handler can be modified a bit to
> > insure
> > that it does not handle any unintended imprecise abort.
> >
> > Regards
> > Pratyush
> >
> >> > return 0;
> >> >
> >> > }
>
> I'm sort of dubious about imx6q_pcie_abort_handler() to begin with --
> it seems like it assumes that imprecise aborts are either enabled or
> disabled. Isn't there some way it can *check* whether that's the
> case, so it won't break if those aborts are enabled at a different
> point in the future?
The best way here would be to find a way to prevent the DWC controller from
generating data abort on missing device _at_all_ and just make the read from
that portion of config space be 0xffffffff or zeroes . I didn't find any
configuration bit that would be able to do it though :( Richard, can you maybe
check with FSL if such an option exists on the controller?
> For now, I'll apply this to my pci/host-imx6 branch for v3.13, but let
> me know if you want to tweak it somehow.
>
> Bjorn
--
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
