Re: [PATCH v3 2/2] asm-generic: Add new pci.h and use it

[Rob Herring <robh@kernel.org>]

On Fri, Jul 22, 2022 at 9:27 AM Palmer Dabbelt <palmer@dabbelt.com> wrote:
>
> On Thu, 21 Jul 2022 16:06:52 PDT (-0700), Rob Herring wrote:
> > On Tue, Jul 19, 2022 at 9:59 AM Palmer Dabbelt <palmer@dabbelt.com> wrote:
> >>
> >> On Sun, 17 Jul 2022 17:41:14 PDT (-0700), shorne@gmail.com wrote:
> >> > The asm/pci.h used for many newer architectures share similar
> >> > definitions.  Move the common parts to asm-generic/pci.h to allow for
> >> > sharing code.
> >> >
> >> > Two things to note are:
> >> >
> >> >  - isa_dma_bridge_buggy, traditionally this is defined in asm/dma.h but
> >> >    these architectures avoid creating that file and add the definition
> >> >    to asm/pci.h.
> >> >  - ARCH_GENERIC_PCI_MMAP_RESOURCE, csky does not define this so we
> >> >    undefine it after including asm-generic/pci.h.  Why doesn't csky
> >> >    define it?
> >> >  - pci_get_legacy_ide_irq, This function is only used on architectures
> >> >    that support PNP.  It is only maintained for arm64, in other
> >> >    architectures it is removed.
> >> >
> >> > Suggested-by: Arnd Bergmann <arnd@arndb.de>
> >> > Link: https://lore.kernel.org/lkml/CAK8P3a0JmPeczfmMBE__vn=Jbvf=nkbpVaZCycyv40pZNCJJXQ@mail.gmail.com/
> >> > Signed-off-by: Stafford Horne <shorne@gmail.com>
> >> > ---
> >> > Second note on isa_dma_bridge_buggy, this is set on x86 but it it also set in
> >> > pci/quirks.c.  We discussed limiting it only to x86 though as its a general
> >> > quick triggered by pci ids I think it will be more tricky than we thought so I
> >> > will leave as is.  It might be nice to move it out of asm/dma.h and into
> >> > asm/pci.h though.
> >> >
> >> > Since v2:
> >> >  - Nothing
> >> > Since v1:
> >> >  - Remove definition of pci_get_legacy_ide_irq
> >> >
> >> >  arch/arm64/include/asm/pci.h | 12 +++---------
> >> >  arch/csky/include/asm/pci.h  | 24 ++++--------------------
> >> >  arch/riscv/include/asm/pci.h | 25 +++----------------------
> >> >  arch/um/include/asm/pci.h    | 24 ++----------------------
> >> >  include/asm-generic/pci.h    | 36 ++++++++++++++++++++++++++++++++++++
> >> >  5 files changed, 48 insertions(+), 73 deletions(-)
> >> >  create mode 100644 include/asm-generic/pci.h
> >> >
> >> > diff --git a/arch/arm64/include/asm/pci.h b/arch/arm64/include/asm/pci.h
> >> > index b33ca260e3c9..1180e83712f5 100644
> >> > --- a/arch/arm64/include/asm/pci.h
> >> > +++ b/arch/arm64/include/asm/pci.h
> >> > @@ -9,7 +9,6 @@
> >> >  #include <asm/io.h>
> >> >
> >> >  #define PCIBIOS_MIN_IO               0x1000
> >> > -#define PCIBIOS_MIN_MEM              0
> >> >
> >> >  /*
> >> >   * Set to 1 if the kernel should re-assign all PCI bus numbers
> >> > @@ -18,9 +17,6 @@
> >> >       (pci_has_flag(PCI_REASSIGN_ALL_BUS))
> >> >
> >> >  #define arch_can_pci_mmap_wc() 1
> >> > -#define ARCH_GENERIC_PCI_MMAP_RESOURCE       1
> >> > -
> >> > -extern int isa_dma_bridge_buggy;
> >> >
> >> >  #ifdef CONFIG_PCI
> >> >  static inline int pci_get_legacy_ide_irq(struct pci_dev *dev, int channel)
> >> > @@ -28,11 +24,9 @@ static inline int pci_get_legacy_ide_irq(struct pci_dev *dev, int channel)
> >> >       /* no legacy IRQ on arm64 */
> >> >       return -ENODEV;
> >> >  }
> >> > -
> >> > -static inline int pci_proc_domain(struct pci_bus *bus)
> >> > -{
> >> > -     return 1;
> >> > -}
> >> >  #endif  /* CONFIG_PCI */
> >> >
> >> > +/* Generic PCI */
> >> > +#include <asm-generic/pci.h>
> >> > +
> >> >  #endif  /* __ASM_PCI_H */
> >> > diff --git a/arch/csky/include/asm/pci.h b/arch/csky/include/asm/pci.h
> >> > index ebc765b1f78b..44866c1ad461 100644
> >> > --- a/arch/csky/include/asm/pci.h
> >> > +++ b/arch/csky/include/asm/pci.h
> >> > @@ -9,26 +9,10 @@
> >> >
> >> >  #include <asm/io.h>
> >> >
> >> > -#define PCIBIOS_MIN_IO               0
> >> > -#define PCIBIOS_MIN_MEM              0
> >> > +/* Generic PCI */
> >> > +#include <asm-generic/pci.h>
> >> >
> >> > -/* C-SKY shim does not initialize PCI bus */
> >> > -#define pcibios_assign_all_busses() 1
> >> > -
> >> > -extern int isa_dma_bridge_buggy;
> >> > -
> >> > -#ifdef CONFIG_PCI
> >> > -static inline int pci_get_legacy_ide_irq(struct pci_dev *dev, int channel)
> >> > -{
> >> > -     /* no legacy IRQ on csky */
> >> > -     return -ENODEV;
> >> > -}
> >> > -
> >> > -static inline int pci_proc_domain(struct pci_bus *bus)
> >> > -{
> >> > -     /* always show the domain in /proc */
> >> > -     return 1;
> >> > -}
> >> > -#endif  /* CONFIG_PCI */
> >> > +/* csky doesn't use generic pci resource mapping */
> >> > +#undef ARCH_GENERIC_PCI_MMAP_RESOURCE
> >> >
> >> >  #endif  /* __ASM_CSKY_PCI_H */
> >> > diff --git a/arch/riscv/include/asm/pci.h b/arch/riscv/include/asm/pci.h
> >> > index 7fd52a30e605..12ce8150cfb0 100644
> >> > --- a/arch/riscv/include/asm/pci.h
> >> > +++ b/arch/riscv/include/asm/pci.h
> >> > @@ -12,29 +12,7 @@
> >> >
> >> >  #include <asm/io.h>
> >> >
> >> > -#define PCIBIOS_MIN_IO               0
> >> > -#define PCIBIOS_MIN_MEM              0
> >>
> >> My for-next changes these in bb356ddb78b2 ("RISC-V: PCI: Avoid handing
> >> out address 0 to devices").  Do you mind either splitting out the
> >> arch/riscv bits or having this in via some sort of shared tag?
> >
> > Shouldn't the values not matter here if the IO and mem resources are
> > described in the DT (and don't use 0)? The values of 4 and 16 look
> > odd.
>
> The linked thread has a fairly long discussion
> <https://lore.kernel.org/all/alpine.DEB.2.21.2202260044180.25061@angie.orcam.me.uk/>.
> I agree it's odd to have this in arch code: "don't hand out address 0"
> isn't really a RISC-V constraint (ie, we don't have architecture-defined
> limitations on these address spaces) but a constraint that comes from
> the generic port I/O functions and some other related PCI/resource code
> where the value 0 is a sentinel.

If you look at arm32, we have a variable for PCIBIOS_MIN_MEM because
pre-DT what platforms required was all over the place. Nothing using
DT needs to set that variable. And arm64 uses 0 without problems. In
all those platforms, none of them have the same restrictions? So it is
still curious to me how PCIBIOS_MIN_MEM matters for Risc-V.

I/O is different as Arnd said, but I'd imagine we could just set the
min to 4 in the generic header and be done with it.

> Maybe the right thing to do here is actually to make the default
> definitions of these macros non-zero, or to add some sort of ARCH_
> flavor of them and move that non-zero requirement closer to where it
> comes from?  From the look of it any port that uses the generic port I/O
> functions and has 0 for these will be broken in the same way.
>
> That said, I'm not really a PCI guy so maybe Bjorn or Maciej has a
> better idea?

From fu740:
                       ranges = <0x81000000  0x0 0x60080000  0x0
0x60080000 0x0 0x10000>,      /* I/O */
                                 <0x82000000  0x0 0x60090000  0x0
0x60090000 0x0 0xff70000>,    /* mem */
                                 <0x82000000  0x0 0x70000000  0x0
0x70000000 0x0 0x1000000>,    /* mem */
                                 <0xc3000000 0x20 0x00000000 0x20
0x00000000 0x20 0x00000000>;  /* mem prefetchable */

So again, how does one get a 0 address handed out when that's not even
a valid region according to DT? Is there some legacy stuff that
ignores the bridge windows?

Rob