Re: MSI-X I/O space clash

[Jason Thorpe <thorpej%me.com@localhost>]

This seems perfectly reasonable.

-- thorpej
Sent from my iPhone.

> On Sep 28, 2019, at 10:24 AM, Tobias Nygren <tnn%netbsd.org@localhost> wrote:
> 
> Hi,
> 
> For some PCI-e devices the MSI-X vector table is embedded in BAR0.
> If the driver maps the whole of BAR0 then the x86 MSI-X MD code can't
> map the table because the bus space range is already reserved, and
> we fail to establish interrupt handlers.
> We already have some hacks in drivers to deal with this, for
> example in xhci and nvme:
> 
> https://github.com/NetBSD/src/commit/c44e99c3172714cc4b9952bca60a00756fcf5123#diff-518473c01ea8c7a44eabd6c8b2c8c694
> 
> mcx(4) also has this problem but it's worse because the table
> is not at the end of the BAR but in the middle of it, which makes
> it convoluted to work around.
> 
> jmcneill@ suggested to make the following change to the MD code
> instead:
> 
> --- sys/arch/x86/pci/msipic.c    26 Jun 2019 10:20:06 -0000    1.17
> +++ sys/arch/x86/pci/msipic.c    28 Sep 2019 17:14:13 -0000
> @@ -711,7 +711,7 @@ msipic_construct_msix_pic(const struct p
>        flags &= ~BUS_SPACE_MAP_PREFETCHABLE;
>    }
>    bssize = roundup(table_size, PAGE_SIZE);
> -    err = bus_space_map(pa->pa_memt, memaddr + table_offset, bssize, flags,
> +    err = _x86_memio_map(pa->pa_memt, memaddr + table_offset, bssize, flags,
>        &bshandle);
>    bstag = pa->pa_memt;
> #endif
> @@ -739,8 +739,8 @@ msipic_destruct_msix_pic(struct pic *msi
>    KASSERT(msix_pic->pic_type == PIC_MSIX);
> 
>    msipic = msix_pic->pic_msipic;
> -    bus_space_unmap(msipic->mp_bstag, msipic->mp_bshandle,
> -        msipic->mp_bssize);
> +    _x86_memio_unmap(msipic->mp_bstag, msipic->mp_bshandle,
> +        msipic->mp_bssize, NULL);
> 
>    msipic_destruct_common_msi_pic(msix_pic);
> }
> 
> 
> I'd like to make this change and pull it up to netbsd-9 along with
> mcx(4) attachment for amd64, unless there are objections.
> 
> Kind regards,
> -Tobias