Re: [PATCH v4 03/14] PCI: cadence: Add support to use custom read and write accessors

[Rob Herring <robh@kernel.org>]

On Thu, May 21, 2020 at 7:33 AM Kishon Vijay Abraham I <kishon@ti.com> wrote:
>
> Hi Rob,
>
> On 5/21/2020 3:37 AM, Rob Herring wrote:
> > On Wed, May 06, 2020 at 08:44:18PM +0530, Kishon Vijay Abraham I wrote:
> >> Add support to use custom read and write accessors. Platforms that
> >> don't support half word or byte access or any other constraint
> >> while accessing registers can use this feature to populate custom
> >> read and write accessors. These custom accessors are used for both
> >> standard register access and configuration space register access of
> >> the PCIe host bridge.
> >>
> >> Signed-off-by: Kishon Vijay Abraham I <kishon@ti.com>
> >> ---
> >>  drivers/pci/controller/cadence/pcie-cadence.h | 107 +++++++++++++++---
> >>  1 file changed, 94 insertions(+), 13 deletions(-)
> >
> > Actually, take back my R-by...
> >
> >>
> >> diff --git a/drivers/pci/controller/cadence/pcie-cadence.h b/drivers/pci/controller/cadence/pcie-cadence.h
> >> index df14ad002fe9..70b6b25153e8 100644
> >> --- a/drivers/pci/controller/cadence/pcie-cadence.h
> >> +++ b/drivers/pci/controller/cadence/pcie-cadence.h
> >> @@ -223,6 +223,11 @@ enum cdns_pcie_msg_routing {
> >>      MSG_ROUTING_GATHER,
> >>  };
> >>
> >> +struct cdns_pcie_ops {
> >> +    u32     (*read)(void __iomem *addr, int size);
> >> +    void    (*write)(void __iomem *addr, int size, u32 value);
> >> +};
> >> +
> >>  /**
> >>   * struct cdns_pcie - private data for Cadence PCIe controller drivers
> >>   * @reg_base: IO mapped register base
> >> @@ -239,7 +244,7 @@ struct cdns_pcie {
> >>      int                     phy_count;
> >>      struct phy              **phy;
> >>      struct device_link      **link;
> >> -    const struct cdns_pcie_common_ops *ops;
> >> +    const struct cdns_pcie_ops *ops;
> >>  };
> >>
> >>  /**
> >> @@ -299,69 +304,145 @@ struct cdns_pcie_ep {
> >>  /* Register access */
> >>  static inline void cdns_pcie_writeb(struct cdns_pcie *pcie, u32 reg, u8 value)
> >>  {
> >> -    writeb(value, pcie->reg_base + reg);
> >> +    void __iomem *addr = pcie->reg_base + reg;
> >> +
> >> +    if (pcie->ops && pcie->ops->write) {
> >> +            pcie->ops->write(addr, 0x1, value);
> >> +            return;
> >> +    }
> >> +
> >> +    writeb(value, addr);
> >>  }
> >>
> >>  static inline void cdns_pcie_writew(struct cdns_pcie *pcie, u32 reg, u16 value)
> >>  {
> >> -    writew(value, pcie->reg_base + reg);
> >> +    void __iomem *addr = pcie->reg_base + reg;
> >> +
> >> +    if (pcie->ops && pcie->ops->write) {
> >> +            pcie->ops->write(addr, 0x2, value);
> >> +            return;
> >> +    }
> >> +
> >> +    writew(value, addr);
> >>  }
> >
> > cdns_pcie_writeb and cdns_pcie_writew are used, so remove them.
> >
> >>
> >>  static inline void cdns_pcie_writel(struct cdns_pcie *pcie, u32 reg, u32 value)
> >>  {
> >> -    writel(value, pcie->reg_base + reg);
> >> +    void __iomem *addr = pcie->reg_base + reg;
> >> +
> >> +    if (pcie->ops && pcie->ops->write) {
> >> +            pcie->ops->write(addr, 0x4, value);
> >> +            return;
> >> +    }
> >> +
> >> +    writel(value, addr);
> >
> > writel isn't broken for you, so you don't need this either.
> >
> >>  }
> >>
> >>  static inline u32 cdns_pcie_readl(struct cdns_pcie *pcie, u32 reg)
> >>  {
> >> -    return readl(pcie->reg_base + reg);
> >> +    void __iomem *addr = pcie->reg_base + reg;
> >> +
> >> +    if (pcie->ops && pcie->ops->read)
> >> +            return pcie->ops->read(addr, 0x4);
> >> +
> >> +    return readl(addr);
> >
> > And neither is readl.
>
> Sure, I'll remove all the unused functions and avoid using ops for readl and
> writel.
> >
> >>  }
> >>
> >>  /* Root Port register access */
> >>  static inline void cdns_pcie_rp_writeb(struct cdns_pcie *pcie,
> >>                                     u32 reg, u8 value)
> >>  {
> >> -    writeb(value, pcie->reg_base + CDNS_PCIE_RP_BASE + reg);
> >> +    void __iomem *addr = pcie->reg_base + CDNS_PCIE_RP_BASE + reg;
> >> +
> >> +    if (pcie->ops && pcie->ops->write) {
> >> +            pcie->ops->write(addr, 0x1, value);
> >> +            return;
> >> +    }
> >> +
> >> +    writeb(value, addr);
> >>  }
> >>
> >>  static inline void cdns_pcie_rp_writew(struct cdns_pcie *pcie,
> >>                                     u32 reg, u16 value)
> >>  {
> >> -    writew(value, pcie->reg_base + CDNS_PCIE_RP_BASE + reg);
> >> +    void __iomem *addr = pcie->reg_base + CDNS_PCIE_RP_BASE + reg;
> >> +
> >> +    if (pcie->ops && pcie->ops->write) {
> >> +            pcie->ops->write(addr, 0x2, value);
> >> +            return;
> >> +    }
> >> +
> >> +    writew(value, addr);
> >
> > You removed 2 out of 3 calls to this. I think I'd just make the root
> > port writes always be 32-bit. It is all just one time init stuff
> > anyways.
> >
> > Either rework the calls to assemble the data into 32-bits or keep these
> > functions and do the RMW here.
>
> The problem with assembling data into 32-bits is we have to read/write with
> different offsets. We'll give PCI_VENDOR_ID offset for modifying deviceID,
> PCI_INTERRUPT_LINE for modifying INTERRUPT_PIN which might get non-intuitive.
> Similarly in endpoint we read and write to MSI_FLAGS (which is at offset 2) we
> have to directly use MSI capability offset.
>
> And doing RMW in the accessors would mean the same RMW op is repeated. So if we
> just have cdns_pcie_rp_writeb() and cdns_pcie_rp_writew(), the same code will
> be repeated here twice.

Why repeated? Just copy what the config accessors do:

static inline void cdns_pcie_write_sz(u32 val, void __iomem *addr, int size)
{
  u32 tmp, mask, where = (unsigned long)addr & 0x3;

  addr -= where;

  mask = ~(((1 << (size * 8)) - 1) << (where * 8));
  tmp = readl(addr) & mask;
  tmp |= val << (where * 8);
  writel(tmp, addr);
}

/* Root Port register access */
static inline void cdns_pcie_rp_writeb(struct cdns_pcie *pcie,
       u32 reg, u8 value)
{
  cdns_pcie_write_sz(value, pcie->reg_base + CDNS_PCIE_RP_BASE + reg, 1);
}

static inline void cdns_pcie_rp_writew(struct cdns_pcie *pcie,
       u32 reg, u16 value)
{
  cdns_pcie_write_sz(value, pcie->reg_base + CDNS_PCIE_RP_BASE + reg, 2);
}

>
> IMHO using ops is a lot cleaner for these cases. IMHO except for removing
> unused functions and not changing readl/writel, others should use ops.

Trying to read the DW PCI driver I don't agree...

Again, unless doing RMW is fundamentally broken (like it is for config
space at runtime), then you only want to do it on broken h/w and ops
would be appropriate. It is all mostly one time setup, so doing a few
extra reads isn't a big deal. If you really care about speed on that,
probably should use the _relaxed accessor variants.

I'm being hopeful the Cadence IP doesn't become the mess that DW is.

Rob