[RFC 3/9] MIPS: add virt_to_phys() and phys_to_virt()

[Peter Mamonov]

On Mon, 7 Dec 2015 11:27:44 +0100
Sascha Hauer <s.hauer at pengutronix.de> wrote:

> On Sun, Dec 06, 2015 at 05:50:51PM +0300, Antony Pavlov wrote:
> > On Fri, 4 Sep 2015 08:20:48 +0200
> > Sascha Hauer <s.hauer at pengutronix.de> wrote:
> > 
> > > On Fri, Aug 28, 2015 at 06:46:14PM +0300, Antony Pavlov wrote:
> > > > On Fri, 28 Aug 2015 08:34:32 +0200
> > > > Sascha Hauer <s.hauer at pengutronix.de> wrote:
> > > > 
> > > > > On Fri, Aug 28, 2015 at 01:24:04AM +0300, Antony Pavlov wrote:
> > > > > > N.B. phys_to_virt() translates phys address
> > > > > > to KSEG1 (uncached) address as barebox mips
> > > > > > has no cache support.
> > > > > 
> > > > > What would it take to implement cache support for mips?
> > > > lack of the cache support is critical problem for current
> > > > barebox mips support. I'm planning to add cache support in
> > > > several weeks. This task needs much test efforts for different
> > > > boards.
> > > > 
> > > > Anyway I can't carry out cache adding work at one.
> > > > But adding virt_to_phys and DMA support will help to add cache
> > > > support one day anyway.
> > > 
> > > Looking at this again the virt_to_phys/phys_to_virt macros are not
> > > necessary. dma_alloc_coherent() already returns both the virtual
> > > address and the DMA address. It should be possible to replace
> > > DMA_ADDRESS_BROKEN in the ehci driver with a real pointer and use
> > > it appropriatly in the driver.
> > > 
> > 
> > I have tried to get physical address from dma_alloc_coherent(),
> > here is a small part of my patch (just for demonstration):
> > 
> > --- a/drivers/usb/host/ehci-hcd.c
> > +++ b/drivers/usb/host/ehci-hcd.c
> > @@ -41,6 +41,7 @@ struct ehci_priv {
> >  	struct ehci_hcor *hcor;
> >  	struct usb_host host;
> >  	struct QH *qh_list;
> > +	dma_addr_t qh_list_dma;
> >  	struct qTD *td;
> >  	int portreset;
> >  	unsigned long flags;
> > @@ -403,7 +410,7 @@ ehci_submit_async(struct usb_device *dev,
> > unsigned long pipe, void *buffer, goto fail;
> >  	}
> >  
> > -	ehci->qh_list->qh_link =
> > cpu_to_hc32((uint32_t)ehci->qh_list | QH_LINK_TYPE_QH);
> > +	ehci->qh_list->qh_link = cpu_to_hc32(ehci->qh_list_dma |
> > QH_LINK_TYPE_QH); 
> >  	token = hc32_to_cpu(qh->qt_token);
> >  	if (!(token & 0x80)) {
> > @@ -1306,7 +1313,7 @@ int ehci_register(struct device_d *dev,
> > struct ehci_data *data) ehci->post_init = data->post_init;
> >  
> >  	ehci->qh_list = dma_alloc_coherent(sizeof(struct QH) *
> > NUM_TD,
> > -					   DMA_ADDRESS_BROKEN);
> > +					   &ehci->qh_list_dma);
> >  	ehci->periodic_queue = dma_alloc_coherent(sizeof(struct
> > QH), DMA_ADDRESS_BROKEN);
> >  	ehci->td = dma_alloc_coherent(sizeof(struct qTD) * NUM_TD,
> > 
> > 
> > However ehci_td_buffer() gets buf pointer outside of ehci driver
> > (e.g. via usb_bulk_msg()), so it is difficult to avoid
> > virt_to_phys() in ehci_td_buffer(), e.g.:
> > 
> > @@ -195,7 +197,7 @@ static int ehci_td_buffer(struct qTD *td, void
> > *buf, size_t sz) 
> >         idx = 0;
> >         while (idx < 5) {
> > -               td->qt_buffer[idx] = cpu_to_hc32(addr);
> > +               td->qt_buffer[idx] =
> > cpu_to_hc32(virt_to_phys(addr)); next = (addr + 4096) & ~4095;
> >                 delta = next - addr;
> >                 if (delta >= sz)
> > 
> > Have you any idea?
> 
> No. Translating this addres into a physical address seems unavoidable
> here. 

The other solution is to allocate private buffer using
dma_alloc_coherent(), use the returned dma address to fill the qt_buffer
field, copy data from the supplied buffer to the private buffer before
the transaction and copy back after the transaction is done.

> I would still prefer a real cache implementation for MIPS
> though. Have you looked further into that? The current MIPS
> cached/uncached memory windows seem to be a constant source of
> problems we could avoid.
> 
> Sascha
> 
>