Message ID | 1436823096-24059-1-git-send-email-ddaney.cavm@gmail.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On Mon, Jul 13, 2015 at 10:31:36PM +0100, David Daney wrote: > From: David Daney <david.daney@cavium.com> > > Needed to make pci_iomap() work. Care to elaborate? AFAICT, mapping an IO bar on arm64 just gives you back a VA into our PCI_IOBASE region and the ioreadX accessors will just call the readX macros, so there should be no need for further port adjustment. Will
On 07/14/2015 04:00 AM, Will Deacon wrote: > On Mon, Jul 13, 2015 at 10:31:36PM +0100, David Daney wrote: >> From: David Daney <david.daney@cavium.com> >> >> Needed to make pci_iomap() work. > > Care to elaborate? > I should have explained what I am doing here a little better. Systems based on the Cavium ThunderX processor may have up to 8 independent PCIe root complexes. The I/O space on each bus occupies an independent physical address window. So, in order to be able to map all of these (semi) contiguously, we need a lot more virtual address space than is supplied by the default values for all these constants. The option I chose here was to unconditionally expand the I/O ranges for all arm64 systems. If you think this breaks existing systems/drivers, I will have to look for other options. David Daney > AFAICT, mapping an IO bar on arm64 just gives you back a VA into our > PCI_IOBASE region and the ioreadX accessors will just call the readX > macros, so there should be no need for further port adjustment. > > Will >
On Tue, Jul 14, 2015 at 05:12:57PM +0100, David Daney wrote: > On 07/14/2015 04:00 AM, Will Deacon wrote: > > On Mon, Jul 13, 2015 at 10:31:36PM +0100, David Daney wrote: > >> From: David Daney <david.daney@cavium.com> > >> > >> Needed to make pci_iomap() work. > > > > Care to elaborate? > > > > I should have explained what I am doing here a little better. Yeah, thanks. > Systems based on the Cavium ThunderX processor may have up to 8 > independent PCIe root complexes. The I/O space on each bus occupies an > independent physical address window. Hmm, so do you have 64k of I/O space per-bus? That gives 8x256x64k = 128M IIUC, so not sure what your 32MB is for. > So, in order to be able to map all of these (semi) contiguously, we need > a lot more virtual address space than is supplied by the default values > for all these constants. > > The option I chose here was to unconditionally expand the I/O ranges for > all arm64 systems. If you think this breaks existing systems/drivers, I > will have to look for other options. Hmm, but pci_iomap winds up calling __pci_ioport_map, which expands to ioport_map which just does: return PCI_IOBASE + (port & IO_SPACE_LIMIT); so I'm struggling to see what your patch achieves. Will
On 07/14/2015 09:29 AM, Will Deacon wrote: > On Tue, Jul 14, 2015 at 05:12:57PM +0100, David Daney wrote: >> On 07/14/2015 04:00 AM, Will Deacon wrote: >>> On Mon, Jul 13, 2015 at 10:31:36PM +0100, David Daney wrote: >>>> From: David Daney <david.daney@cavium.com> >>>> >>>> Needed to make pci_iomap() work. >>> >>> Care to elaborate? >>> >> >> I should have explained what I am doing here a little better. > > Yeah, thanks. > >> Systems based on the Cavium ThunderX processor may have up to 8 >> independent PCIe root complexes. The I/O space on each bus occupies an >> independent physical address window. > > Hmm, so do you have 64k of I/O space per-bus? That gives 8x256x64k = 128M > IIUC, so not sure what your 32MB is for. I don't understand where your 256 came from there. Actually, my current implementation has 1M per bus(which is overkill). For 8 buses I need 8M, which fits within the PCI_IO_SIZE... > >> So, in order to be able to map all of these (semi) contiguously, we need >> a lot more virtual address space than is supplied by the default values >> for all these constants. >> >> The option I chose here was to unconditionally expand the I/O ranges for >> all arm64 systems. If you think this breaks existing systems/drivers, I >> will have to look for other options. > > Hmm, but pci_iomap winds up calling __pci_ioport_map, which expands to > ioport_map which just does: > > return PCI_IOBASE + (port & IO_SPACE_LIMIT); > > so I'm struggling to see what your patch achieves. Here is ioport_map (from lib/iomap.c): void __iomem *ioport_map(unsigned long port, unsigned int nr) { if (port > PIO_MASK) return NULL; return (void __iomem *) (unsigned long) (port + PIO_OFFSET); } With the default value of PIO_MASK (64K), I cannot map any I/O ports on my PCIe RC 1..7 The values I supplied in my patch may be sub-optimal, but I think something is needed. I will look into this in a little more detail today. Thanks, David Daney > > Will >
Hi David, On Tue, Jul 14, 2015 at 05:58:20PM +0100, David Daney wrote: > On 07/14/2015 09:29 AM, Will Deacon wrote: > > On Tue, Jul 14, 2015 at 05:12:57PM +0100, David Daney wrote: > >> Systems based on the Cavium ThunderX processor may have up to 8 > >> independent PCIe root complexes. The I/O space on each bus occupies an > >> independent physical address window. > > > > Hmm, so do you have 64k of I/O space per-bus? That gives 8x256x64k = 128M > > IIUC, so not sure what your 32MB is for. > > I don't understand where your 256 came from there. Sorry, sounds like we're mixing up buses and root complexes. Which is it? > Actually, my current implementation has 1M per bus(which is overkill). > For 8 buses I need 8M, which fits within the PCI_IO_SIZE... So there's no problem? > >> So, in order to be able to map all of these (semi) contiguously, we need > >> a lot more virtual address space than is supplied by the default values > >> for all these constants. > >> > >> The option I chose here was to unconditionally expand the I/O ranges for > >> all arm64 systems. If you think this breaks existing systems/drivers, I > >> will have to look for other options. > > > > Hmm, but pci_iomap winds up calling __pci_ioport_map, which expands to > > ioport_map which just does: > > > > return PCI_IOBASE + (port & IO_SPACE_LIMIT); > > > > so I'm struggling to see what your patch achieves. > > Here is ioport_map (from lib/iomap.c): > > > void __iomem *ioport_map(unsigned long port, unsigned int nr) > { > if (port > PIO_MASK) > return NULL; > return (void __iomem *) (unsigned long) (port + PIO_OFFSET); > } We only get this definition if CONFIG_GENERIC_IOMAP=y. Why is that selected? Will
On 07/14/2015 10:04 AM, Will Deacon wrote: > Hi David, > > On Tue, Jul 14, 2015 at 05:58:20PM +0100, David Daney wrote: [...] > We only get this definition if CONFIG_GENERIC_IOMAP=y. Why is that > selected? > OK, Good question. The original patch was against 3.18, but commit 09a5723983e (arm64: Use include/asm-generic/io.h) changed some of this, so it may need reconsideration. Let me reevaluate this patch. Thanks, David Daney
diff --git a/arch/arm64/include/asm/io.h b/arch/arm64/include/asm/io.h index 540f7c0..8ef78d5 100644 --- a/arch/arm64/include/asm/io.h +++ b/arch/arm64/include/asm/io.h @@ -149,6 +149,11 @@ static inline u64 __raw_readq(const volatile void __iomem *addr) #define IO_SPACE_LIMIT (PCI_IO_SIZE - 1) #define PCI_IOBASE ((void __iomem *)PCI_IO_START) +#define HAVE_ARCH_PIO_SIZE 1 +#define PIO_RESERVED SZ_32M +#define PIO_OFFSET 0 +#define PIO_MASK (PIO_RESERVED - 1) + /* * String version of I/O memory access operations. */