Re: [PATCH v2 2/5] MIPS: ralink: implement 'pcibios_root_bridge_prepare()'

[Sergio Paracuellos]

On Thu, Dec 2, 2021 at 6:02 PM Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
>
> On 12/2/21 7:50 AM, Sergio Paracuellos wrote:
> > Hi Guenter,
> >
> > On Thu, Dec 2, 2021 at 4:06 PM Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> >>
> >> On 12/2/21 12:29 AM, Sergio Paracuellos wrote:
> >>> Hi Guenter,
> >>>
> >>> On Wed, Dec 1, 2021 at 11:17 PM Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> >>>>
> >>>> On 12/1/21 1:51 PM, Sergio Paracuellos wrote:
> >>>>> PCI core code call 'pcibios_root_bridge_prepare()' function inside function
> >>>>> 'pci_register_host_bridge()'. This point is very good way to properly enter
> >>>>> into this MIPS ralink specific code to properly setup I/O coherency units
> >>>>> with PCI memory addresses.
> >>>>>
> >>>>> Signed-off-by: Sergio Paracuellos <sergio.paracuellos@xxxxxxxxx>
> >>>>> ---
> >>>>>     arch/mips/ralink/mt7621.c | 30 ++++++++++++++++++++++++++++++
> >>>>>     1 file changed, 30 insertions(+)
> >>>>>
> >>>>> diff --git a/arch/mips/ralink/mt7621.c b/arch/mips/ralink/mt7621.c
> >>>>> index bd71f5b14238..7649416c1cd7 100644
> >>>>> --- a/arch/mips/ralink/mt7621.c
> >>>>> +++ b/arch/mips/ralink/mt7621.c
> >>>>> @@ -10,6 +10,7 @@
> >>>>>     #include <linux/slab.h>
> >>>>>     #include <linux/sys_soc.h>
> >>>>>     #include <linux/memblock.h>
> >>>>> +#include <linux/pci.h>
> >>>>>
> >>>>>     #include <asm/bootinfo.h>
> >>>>>     #include <asm/mipsregs.h>
> >>>>> @@ -22,6 +23,35 @@
> >>>>>
> >>>>>     static void *detect_magic __initdata = detect_memory_region;
> >>>>>
> >>>>> +int pcibios_root_bridge_prepare(struct pci_host_bridge *bridge)
> >>>>> +{
> >>>>> +     struct resource_entry *entry;
> >>>>> +     resource_size_t mask;
> >>>>> +
> >>>>> +     entry = resource_list_first_type(&bridge->windows, IORESOURCE_MEM);
> >>>>> +     if (!entry) {
> >>>>> +             pr_err("Cannot get memory resource\n");
> >>>>> +             return -EINVAL;
> >>>>> +     }
> >>>>> +
> >>>>> +     if (mips_cps_numiocu(0)) {
> >>>>> +             /*
> >>>>> +              * FIXME: hardware doesn't accept mask values with 1s after
> >>>>> +              * 0s (e.g. 0xffef), so it would be great to warn if that's
> >>>>> +              * about to happen
> >>>>> +              */ > +         mask = ~(entry->res->end - entry->res->start);
> >>>>> +
> >>>>
> >>>> Try something like this:
> >>>>                   WARN_ON((mask != ~0UL && BIT(ffz(mask)) - 1 != mask);
> >>>
> >>> Thanks for the tip. The following works for me:
> >>>
> >>>                     WARN_ON(mask != ~0UL && ~(BIT(__ffs(mask)) - 1) != mask);
> >>
> >> Are you sure ? __ffs() returns the first bit set, which isn't useful
> >> for this test.
> >
> > My mask is calculated as follows:
> >   mask = ~(entry->res->end - entry->res->start);
> >
> > Where for normal memory resource:
> >   - entry->res->end = 0x6fffffff;
> >   - entry->res->start = 0x60000000;
> >
> > So I end up with a mask: 0xf0000000.
> >
> > So applying ~(BIT(__ffs(mask)) - 1) I get a good '0xf0000000' for this
> > particular case which looks correct.
> >
> > Suppose an invalid case with the mask being 0xffef0000.
> >
> > Applying ~(BIT(__ffs(mask)) - 1) will be 0xffff0000 which will trigger
> > the WARN_ON since 0xffff0000 != 0xffef0000
> >
> > So I think this is correct... Am I missing something?
> >
>
> Your description says "hardware doesn't accept mask values with 1s after 0s
> (e.g. 0xffef)". 0xf0000000 has 1s after 0s.
>
> Your version works (I think) as long as the upper mask bits are all 1s.
> It will fail, for example, if the mask value is 0xf000000 and
> sizeof(long) == 8. Your test is the equivalent of "no mask value
> with 0s after 1s", assuming that sizeof(resource_size_t) == sizeof(long).
> As far as I can see with test code, it fails if sizeof(resource_size_t)
> != sizeof(long). Also, it returns an error if mask == 0. I guess that is
> a corner case, but it would be interesting to know if it is theoretically
> valid.

Thanks a lot for the clear explanation. I was assuming MIPS ralink
arch so sizeof(long) and sizeof(resource_size_t) are equal and upper
mask bits are all 1s. But you are right, my version will fail if this
sizeof(long) were eight.

>
> I _think_ the following works even if sizeof(resource_size_t) != sizeof(long).
>
>         WARN_ON(mask && BIT(ffz(~mask)) - 1 != ~mask);

This works for me also and looks like it does the right thing for any
case, thanks.

>
> or, alternatively, something like
>
>         mask2 = entry->res->end - entry->res->start;
>         mask = ~mask2;
>         WARN_ON(mask && BIT(ffz(mask2)) - 1 != mask2);
>
> though that looks a bit weird.

Agreed, using two variables here looks weird also for me.

Best regards,
    Sergio Paracuellos

>
> Thanks,
> Guenter
>
> > Thanks,
> >      Sergio Paracuellos
> >>
> >> Guenter
> >>
> >>>
> >>> I will send this as a different patch, though.
> >>>
> >>> Best regards,
> >>>       Sergio Paracuellos
> >>>
> >>>>
> >>>>> +             write_gcr_reg1_base(entry->res->start);
> >>>>> +             write_gcr_reg1_mask(mask | CM_GCR_REGn_MASK_CMTGT_IOCU0);
> >>>>> +             pr_info("PCI coherence region base: 0x%08llx, mask/settings: 0x%08llx\n",
> >>>>> +                     (unsigned long long)read_gcr_reg1_base(),
> >>>>> +                     (unsigned long long)read_gcr_reg1_mask());
> >>>>> +     }
> >>>>> +
> >>>>> +     return 0;
> >>>>> +}
> >>>>> +
> >>>>>     phys_addr_t mips_cpc_default_phys_base(void)
> >>>>>     {
> >>>>>         panic("Cannot detect cpc address");
> >>>>>
> >>>>
> >>
>