Re: [PATCH] ppc/pnv: Set P10 core xscom region size to match hardware

[Joel Stanley <joel@jms.id.au>]

On Wed, 5 Jul 2023 at 10:02, Cédric Le Goater <clg@kaod.org> wrote:
>
> On 7/5/23 04:05, Joel Stanley wrote:
> > On Wed, 5 Jul 2023 at 01:27, Nicholas Piggin <npiggin@gmail.com> wrote:
> >>
> >> The P10 core xscom memory regions overlap because the size is wrong.
> >> The P10 core+L2 xscom region size is allocated as 0x1000 (with some
> >> unused ranges). "EC" is used as a closer match, as "EX" includes L3
> >> which has a disjoint xscom range that would require a different
> >> region if it were implemented.
> >>
> >> Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
> >
> > Nice, that looks better:
> >
> > 0000000100000000-00000001000fffff (prio 0, i/o): xscom-quad.0: 0x100000
> > 0000000100108000-000000010010ffff (prio 0, i/o): xscom-core.3: 0x8000
> > 0000000100110000-0000000100117fff (prio 0, i/o): xscom-core.2: 0x8000
> > 0000000100120000-0000000100127fff (prio 0, i/o): xscom-core.1: 0x8000
> > 0000000100140000-0000000100147fff (prio 0, i/o): xscom-core.0: 0x8000
> > 0000000108000000-00000001080fffff (prio 0, i/o): xscom-quad.4: 0x100000
> > 0000000108108000-000000010810ffff (prio 0, i/o): xscom-core.7: 0x8000
> > 0000000108110000-0000000108117fff (prio 0, i/o): xscom-core.6: 0x8000
> > 0000000108120000-0000000108127fff (prio 0, i/o): xscom-core.5: 0x8000
> > 0000000108140000-0000000108147fff (prio 0, i/o): xscom-core.4: 0x8000
> >
> > Reviewed-by: Joel Stanley <joel@jms.id.au>
>
> It'd interesting to add some dummy SLW handlers to get rid of the
> XSCOM errors at boot and shutdown on P10 :
>
> [ 4824.393446266,3] XSCOM: write error gcid=0x0 pcb_addr=0x200e883c stat=0x0
> [ 4824.393588777,5] Unable to log error
> [ 4824.393650582,3] XSCOM: Write failed, ret =  -6
> [ 4824.394124623,3] Could not set special wakeup on 0:0: Unable to write QME_SPWU_HYP.
> [ 4824.394368459,3] XSCOM: write error gcid=0x0 pcb_addr=0x200e883c stat=0x0
> [ 4824.394382007,5] Unable to log error
> [ 4824.394384603,3] XSCOM: Write failed, ret =  -6

Yes. I was looking at this yesterday. We need to figure out how to do
the xscom addressing for the QME. It sets (different) bits in order to
address a given core.

For a -smp 4 machine, the P10_QME_SPWU_HYP read comes in on these addresses:

    case 0x200e883c:
    case 0x200e483c:
    case 0x200e283c:
    case 0x200e183c:

ie, the fourth nibble selects the core.

For a -smp 8 machine, the address now has bit 24 set to select the
second quad, so we need to cover these addresses:

    case 0x210e883c:
    case 0x210e483c:
    case 0x210e283c:
    case 0x210e183c:

I am thinking about how to map this into an address range that a model
can claim.

Cheers,

Joel

PS. For reference, this is sufficient to silence xscom errors with
skiboot and -M powernv10 -smp4. A different set of hacks is required
for p9.

--- a/hw/ppc/pnv_xscom.c
+++ b/hw/ppc/pnv_xscom.c
@@ -106,6 +106,26 @@ static uint64_t xscom_read_default(PnvChip *chip,
uint32_t pcba)
     case 0x401082a:
     case 0x4010828:
         return 0;
+
+    /* P10_QME_SPWU_HYP */
+    case 0x200e883c:
+    case 0x200e483c:
+    case 0x200e283c:
+    case 0x200e183c:
+        return 0;
+
+    /* P10_QME_SSH_HYP */
+    case 0x200e882c:
+    case 0x200e482c:
+    case 0x200e282c:
+    case 0x200e182c:
+        return 0;
+
+    /* XPEC_P10_PCI_CPLT_CONF1 */
+    case 0x08000009:
+    case 0x09000009:
+        return 0;
+
     default:
         return -1;
     }
@@ -152,6 +172,13 @@ static bool xscom_write_default(PnvChip *chip,
uint32_t pcba, uint64_t val)
     case PRD_P8_IPOLL_REG_STATUS:
     case PRD_P9_IPOLL_REG_MASK:
     case PRD_P9_IPOLL_REG_STATUS:
+
+    /* P10_QME_SPWU_HYP */
+    case 0x200e883c:
+    case 0x200e483c:
+    case 0x200e283c:
+    case 0x200e183c:
+
         return true;
     default:
         return false;