Re: [PATCH] mtd: bcm47xxsflash: use ioremap_cachable() instead of KSEG0ADDR()

[Brian Norris]

On Mon, Jan 25, 2016 at 07:15:49PM +0000, Maciej W. Rozycki wrote:
> On Mon, 25 Jan 2016, Brian Norris wrote:
> 
> > > > +	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
> > > > +	if (!res) {
> > > > +		dev_err(dev, "invalid resource\n");
> > > > +		return -EINVAL;
> > > > +	}
> > > > +	if (!devm_request_mem_region(dev, res->start, resource_size(res),
> > > > +				     res->name)) {
> > > > +		dev_err(dev, "can't request region for resource %pR\n", res);
> > > > +		return -EBUSY;
> > > > +	}
> > > > +	b47s->window = ioremap_cachable(res->start, resource_size(res));
> > > > +	if (!b47s->window) {
> > > > +		dev_err(dev, "ioremap failed for resource %pR\n", res);
> > > 
> > >  You need to call `devm_release_mem_region' in this case.
> > 
> > No he doesn't. devm_* functions automatically release their resources
> > when either the device is removed, or the probe() fails. So the whole
> > point is that we don't have to explicitly manage the error case.
> 
>  Why does `devm_ioremap_resource' (in lib/devres.c) do that manually then?
> 
> 	dest_ptr = devm_ioremap(dev, res->start, size);
> 	if (!dest_ptr) {
> 		dev_err(dev, "ioremap failed for resource %pR\n", res);
> 		devm_release_mem_region(dev, res->start, size);
> 		dest_ptr = IOMEM_ERR_PTR(-ENOMEM);
> 	}
> 
> Is this an oversight or was it a deliberate design decision?

I didn't design it, but I suspect it's because the API for
devm_ioremap_resource() doesn't assume that a failed
devm_ioremap_resource() means probe() will exit. Perhaps the driver will
have some alternative action instead. It's still good to release the
resources implicitly requested by devm_ioremap_resource().

The difference for us is that we *know* we're exiting the probe(), so we
can rely on the automatic cleanup.

> > But...since he's not using a devm_* version of ioremap (there isn't one
> > for ioremap_cachable()), we actually need to add an iounmap() for the
> > case where mtd_device_parse_register() fails. If we fix that one, I can
> > apply this.
> 
>  As from 4.5-rc1 we now have `ioremap_cache' available for MIPS as well 
> (thanks, Ralf, for a quick action on that!), so you can use that instead 
> to make your code generic.

OK, but there's still not a devm_* version of ioremap_cache(), so we'd still need to do the iounmap() ourselves.

Also, it looks like kernel/memremap.c is suggesting we use memremap()
instead of ioremap_cache()... But I'm not sure if that's really what we
want.

Brian