Re: [PATCH wireless-drivers-next] bcma: get SoC device struct & copy its DMA params to the subdevices

["Rafał Miłecki" <zajec5@gmail.com>]

On Mon, 21 Jan 2019 at 15:46, Christoph Hellwig <hch@lst.de> wrote:
> On Mon, Jan 21, 2019 at 11:11:21AM +0100, Rafał Miłecki wrote:
> > From: Rafał Miłecki <rafal@milecki.pl>
> >
> > For bus devices to be fully usable it's required to set their DMA
> > parameters.
> >
> > For years it has been missing and remained unnoticed because of
> > mips_dma_alloc_coherent() silently handling the empty coherent_dma_mask.
> > Kernel 4.19 came with a lot of DMA changes and caused a regression on
> > the bcm47xx. Starting with the commit f8c55dc6e828 ("MIPS: use generic
> > dma noncoherent ops for simple noncoherent platforms") DMA coherent
> > allocations just fail. Example:
> > [    1.114914] bgmac_bcma bcma0:2: Allocation of TX ring 0x200 failed
> > [    1.121215] bgmac_bcma bcma0:2: Unable to alloc memory for DMA
> > [    1.127626] bgmac_bcma: probe of bcma0:2 failed with error -12
> > [    1.133838] bgmac_bcma: Broadcom 47xx GBit MAC driver loaded
> >
> > This change fixes above regression in addition to the MIPS bcm47xx
> > commit 321c46b91550 ("MIPS: BCM47XX: Setup struct device for the SoC").
> >
> > It also fixes another *old* GPIO regression caused by a parent pointing
> > to the NULL:
> > [    0.157054] missing gpiochip .dev parent pointer
> > [    0.157287] bcma: bus0: Error registering GPIO driver: -22
> > introduced by the commit 74f4e0cc6108 ("bcma: switch GPIO portions to
> > use GPIOLIB_IRQCHIP").
> >
> > Fixes: f8c55dc6e828 ("MIPS: use generic dma noncoherent ops for simple noncoherent platforms")
> > Fixes: 74f4e0cc6108 ("bcma: switch GPIO portions to use GPIOLIB_IRQCHIP")
> > Cc: linux-mips@linux-mips.org
> > Cc: Christoph Hellwig <hch@lst.de>
> > Cc: Linus Walleij <linus.walleij@linaro.org>
> > Signed-off-by: Rafał Miłecki <rafal@milecki.pl>
> > ---
> > While this patch is a regression fix, it depends on a change present in
> > the wireless-drivers-next.git:
> > bcma: keep a direct pointer to the struct device
> >
> > That's why I suggest pushing it into the wireless-drivers-next.git and I
> > can take care of picking it for the stable@kernel.org later.
> >
> > Another option would be cherry-picking commit 5a1c18b761dd ("bcma: keep
> > a direct pointer to the struct device") to the wireless-drivers.git but
> > I don't think it's a common practice.
> > ---
> >  drivers/bcma/host_soc.c       | 2 ++
> >  drivers/bcma/main.c           | 6 +++++-
> >  include/linux/bcma/bcma_soc.h | 1 +
> >  3 files changed, 8 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/bcma/host_soc.c b/drivers/bcma/host_soc.c
> > index c8073b509a2b..1fdfb704f22d 100644
> > --- a/drivers/bcma/host_soc.c
> > +++ b/drivers/bcma/host_soc.c
> > @@ -191,6 +191,8 @@ int __init bcma_host_soc_init(struct bcma_soc *soc)
> >       struct bcma_bus *bus = &soc->bus;
> >       int err;
> >
> > +     bus->dev = soc->dev;
> > +
> >       /* Scan bus and initialize it */
> >       err = bcma_bus_early_register(bus);
> >       if (err)
> > diff --git a/drivers/bcma/main.c b/drivers/bcma/main.c
> > index 6535614a7dc1..433ca5e2ed2c 100644
> > --- a/drivers/bcma/main.c
> > +++ b/drivers/bcma/main.c
> > @@ -236,12 +236,16 @@ EXPORT_SYMBOL(bcma_core_irq);
> >
> >  void bcma_prepare_core(struct bcma_bus *bus, struct bcma_device *core)
> >  {
> > +     struct device *dev = &core->dev;
> > +
> >       core->dev.release = bcma_release_core_dev;
> >       core->dev.bus = &bcma_bus_type;
> >       dev_set_name(&core->dev, "bcma%d:%d", bus->num, core->core_index);
> >       core->dev.parent = bus->dev;
> > -     if (bus->dev)
> > +     if (bus->dev) {
> >               bcma_of_fill_device(bus->dev, core);
> > +             dma_coerce_mask_and_coherent(dev, bus->dev->coherent_dma_mask);
> > +     }
>
> I don't think this does the right thing for bcma devices behind
> PCI/PCIe, as those might need a IOMMU which relies on having a device
> it has properly enumerated to be passed into the DMA API.  In other
> words:  I think you need to change the layering so that the DMA API
> is always called on the underlying PCI/PCIe/platform device, not of
> the child bus.

This is already implemented (we have dma_dev pointer), but it may be
indeed more correct to apply this change for the SoC code only. I'll
send V2.

-- 
Rafał