Re: [v4,2/3] PCI: mediatek: Add new generation controller support

[Rob Herring <robh+dt@kernel.org>]

On Sun, Nov 22, 2020 at 11:45 PM Jianjun Wang <jianjun.wang@mediatek.com> wrote:
>
> On Thu, 2020-11-19 at 14:28 -0600, Bjorn Helgaas wrote:
> > "Add new generation" really contains no information.  And "mediatek"
> > is already used for the pcie-mediatek.c driver, so we should have a
> > new tag for this new driver.  Include useful information in the
> > subject, e.g.,
> >
> >   PCI: mediatek-gen3: Add MediaTek Gen3 driver for MT8192
> >
> > On Wed, Nov 18, 2020 at 04:29:34PM +0800, Jianjun Wang wrote:
> > > MediaTek's PCIe host controller has three generation HWs, the new
> > > generation HW is an individual bridge, it supoorts Gen3 speed and
> > > up to 256 MSI interrupt numbers for multi-function devices.
> >
> > s/supoorts/supports/
> >
> > > Add support for new Gen3 controller which can be found on MT8192.
> > >
> > > Signed-off-by: Jianjun Wang <jianjun.wang@mediatek.com>
> > > Acked-by: Ryder Lee <ryder.lee@mediatek.com>

[...]

> > > +static int mtk_pcie_config_read(struct pci_bus *bus, unsigned int devfn,
> > > +                               int where, int size, u32 *val)
> > > +{
> > > +   struct mtk_pcie_port *port = bus->sysdata;
> > > +   int bytes;
> > > +
> > > +   bytes = ((1 << size) - 1) << (where & 0x3);
> >
> > This seems like some unusual bit twiddling; at least, I don't remember
> > seeing this before.  Can you skim other drivers and see if others do
> > the same thing, and adopt a common style if they do?
>
> Hi Bjorn,
>
> Thanks for your review, I will fix it in the next version.
> >
> > > +   writel(PCIE_CFG_HEADER_FORCE_BE(devfn, bus->number, bytes),
> > > +          port->base + PCIE_CFGNUM_REG);
> > > +
> > > +   *val = readl(port->base + PCIE_CFG_OFFSET_ADDR + (where & ~0x3));
> >
> > These look like they need to be atomic, since you need a writel()
> > followed by a readl().
> >
> > pci_lock_config() (used in pci_bus_read_config_*(), etc) uses the
> > global pci_lock for this unless CONFIG_PCI_LOCKLESS_CONFIG is set.
> >
> > But I would like to eventually move away from this implicit dependency
> > on pci_lock.  If you need to make this atomic, can you add the
> > explicit locking here, so there's a clear connection between the lock
> > and the things it protects?
>
> Sure, I will split it to a map_bus() function and use the standard
> pci_generic_config_read32/write32 functions as Rob's suggestion. I think
> the potential risks of atomic read/write can be avoided.

The generic functions have no effect on atomicity, but using them does
make it easier to find the non-atomic cases.

I'm not sure that having host drivers do their own locking is the best
approach. That's a recipe for more cleanups. It's a common enough
issue that I think it's better if we have locking done in 1 place.
Then host drivers can simply say if they need locking or not via some
bus flag.

Rob