| Message ID | 1507834211-24922-2-git-send-email-vidyas@nvidia.com (mailing list archive) |
|---|---|
| State | New, archived |
| Delegated to: | Bjorn Helgaas |
| Headers | show |
On Fri, Oct 13, 2017 at 12:20:06AM +0530, Vidya Sagar wrote: > use only 4K space from available 1GB PCIe aperture to access > end points configuration space by dynamically moving AFI_AXI_BAR > base address and always making sure that the desired location > to be accessed for generating required config space access falls > in the 4K space reserved for this purpose. This would give more > space for mapping end point device's BARs on some of Tegra platforms Do you now have a revlock issue between this driver change and the corresponding DT changes? I.e., what happens when you run this driver change with an older DT that specifies a 256MB config aperture? What happens if you run the older driver with a new DT that only specifies a 4KB config aperture? > Signed-off-by: Vidya Sagar <vidyas@nvidia.com> > --- > drivers/pci/host/pci-tegra.c | 85 ++++++++++++-------------------------------- > 1 file changed, 23 insertions(+), 62 deletions(-) > > diff --git a/drivers/pci/host/pci-tegra.c b/drivers/pci/host/pci-tegra.c > index 9c40da54f88a..ebdcfca10e17 100644 > --- a/drivers/pci/host/pci-tegra.c > +++ b/drivers/pci/host/pci-tegra.c > @@ -269,6 +269,8 @@ struct tegra_pcie { > struct list_head buses; > struct resource *cs; > > + void __iomem *cfg_va_base; > + > struct resource io; > struct resource pio; > struct resource mem; > @@ -317,7 +319,6 @@ struct tegra_pcie_port { > }; > > struct tegra_pcie_bus { > - struct vm_struct *area; > struct list_head list; > unsigned int nr; > }; > @@ -357,34 +358,16 @@ static inline u32 pads_readl(struct tegra_pcie *pcie, unsigned long offset) > * > * Mapping the whole extended configuration space would require 256 MiB of > * virtual address space, only a small part of which will actually be used. > - * To work around this, a 1 MiB of virtual addresses are allocated per bus > - * when the bus is first accessed. When the physical range is mapped, the > - * the bus number bits are hidden so that the extended register number bits > - * appear as bits [19:16]. Therefore the virtual mapping looks like this: > - * > - * [19:16] extended register number > - * [15:11] device number > - * [10: 8] function number > - * [ 7: 0] register number > - * > - * This is achieved by stitching together 16 chunks of 64 KiB of physical > - * address space via the MMU. > + * To work around this, a 4K of region is used to generate required > + * configuration transaction with relevant B:D:F values. This is achieved by > + * dynamically programming base address and size of AFI_AXI_BAR used for > + * end point config space mapping to make sure that the address (access to > + * which generates correct config transaction) falls in this 4K region > */ > -static unsigned long tegra_pcie_conf_offset(unsigned int devfn, int where) > -{ > - return ((where & 0xf00) << 8) | (PCI_SLOT(devfn) << 11) | > - (PCI_FUNC(devfn) << 8) | (where & 0xfc); > -} > - > static struct tegra_pcie_bus *tegra_pcie_bus_alloc(struct tegra_pcie *pcie, > unsigned int busnr) > { > - struct device *dev = pcie->dev; > - pgprot_t prot = pgprot_noncached(PAGE_KERNEL); > - phys_addr_t cs = pcie->cs->start; > struct tegra_pcie_bus *bus; > - unsigned int i; > - int err; > > bus = kzalloc(sizeof(*bus), GFP_KERNEL); > if (!bus) > @@ -393,33 +376,16 @@ static struct tegra_pcie_bus *tegra_pcie_bus_alloc(struct tegra_pcie *pcie, > INIT_LIST_HEAD(&bus->list); > bus->nr = busnr; > > - /* allocate 1 MiB of virtual addresses */ > - bus->area = get_vm_area(SZ_1M, VM_IOREMAP); > - if (!bus->area) { > - err = -ENOMEM; > - goto free; > - } > - > - /* map each of the 16 chunks of 64 KiB each */ > - for (i = 0; i < 16; i++) { > - unsigned long virt = (unsigned long)bus->area->addr + > - i * SZ_64K; > - phys_addr_t phys = cs + i * SZ_16M + busnr * SZ_64K; > - > - err = ioremap_page_range(virt, virt + SZ_64K, phys, prot); > - if (err < 0) { > - dev_err(dev, "ioremap_page_range() failed: %d\n", err); > - goto unmap; > + if (!pcie->cfg_va_base) { > + pcie->cfg_va_base = ioremap(pcie->cs->start, SZ_4K); > + if (!pcie->cfg_va_base) { > + dev_err(pcie->dev, "failed to ioremap config space\n"); > + kfree(bus); > + bus = (struct tegra_pcie_bus *)-ENOMEM; tegra_pcie_bus_alloc() used to do some bus-specific work -- it allocated VM space and mapped it. This pcie->cfg_va_base ioremap() is not bus-specific, so I don't think it fits here. It looks like something that should be done in the tegra_pcie_probe() path instead. If you move that, then I think you might as well inline the kzalloc() stuff into tegra_pcie_add_bus(). > } > } > > return bus; > - > -unmap: > - vunmap(bus->area->addr); > -free: > - kfree(bus); > - return ERR_PTR(err); > } > > static int tegra_pcie_add_bus(struct pci_bus *bus) > @@ -445,12 +411,12 @@ static void tegra_pcie_remove_bus(struct pci_bus *child) > > list_for_each_entry_safe(bus, tmp, &pcie->buses, list) { > if (bus->nr == child->number) { > - vunmap(bus->area->addr); > list_del(&bus->list); > kfree(bus); > break; > } > } > + iounmap(pcie->cfg_va_base); > } > > static void __iomem *tegra_pcie_map_bus(struct pci_bus *bus, > @@ -459,8 +425,8 @@ static void __iomem *tegra_pcie_map_bus(struct pci_bus *bus, > { > struct pci_host_bridge *host = pci_find_host_bridge(bus); > struct tegra_pcie *pcie = pci_host_bridge_priv(host); > - struct device *dev = pcie->dev; > void __iomem *addr = NULL; > + u32 val = 0; > > if (bus->number == 0) { > unsigned int slot = PCI_SLOT(devfn); > @@ -473,19 +439,14 @@ static void __iomem *tegra_pcie_map_bus(struct pci_bus *bus, > } > } > } else { > - struct tegra_pcie_bus *b; > - > - list_for_each_entry(b, &pcie->buses, list) > - if (b->nr == bus->number) > - addr = (void __iomem *)b->area->addr; > - > - if (!addr) { > - dev_err(dev, "failed to map cfg. space for bus %u\n", > - bus->number); > - return NULL; > - } > - > - addr += tegra_pcie_conf_offset(devfn, where); > + addr = pcie->cfg_va_base; > + val = ((((u32)where & 0xf00) >> 8) << 24) | > + (bus->number << 16) | (PCI_SLOT(devfn) << 11) | > + (PCI_FUNC(devfn) << 8) | (where & 0xff); I think it would be useful to keep tegra_pcie_conf_offset() (extending it to take the bus number) near the comment describing the mapping. > + addr = (val & (SZ_4K - 1)) + addr; > + val = val & ~(SZ_4K - 1); val &= ~(SZ_4K - 1); I *think* what you're doing is computing a config space offset and mapping the 4KB window that contains it. "addr" is not used at all in the AFI_AXI_BAR0 programming, so this would be clearer as: val = ... afi_writel(...) afi_writel(...) addr = pcie->cfg_va_base + (val & (SZ_4K - 1)); } return addr; > + afi_writel(pcie, pcie->cs->start - val, AFI_AXI_BAR0_START); > + afi_writel(pcie, (val + SZ_4K) >> 12, AFI_AXI_BAR0_SZ); > } > > return addr; > -- > 2.7.4 >
On Saturday 21 October 2017 12:38 AM, Bjorn Helgaas wrote: > On Fri, Oct 13, 2017 at 12:20:06AM +0530, Vidya Sagar wrote: >> use only 4K space from available 1GB PCIe aperture to access >> end points configuration space by dynamically moving AFI_AXI_BAR >> base address and always making sure that the desired location >> to be accessed for generating required config space access falls >> in the 4K space reserved for this purpose. This would give more >> space for mapping end point device's BARs on some of Tegra platforms > Do you now have a revlock issue between this driver change and the > corresponding DT changes? I.e., what happens when you run this driver > change with an older DT that specifies a 256MB config aperture? What > happens if you run the older driver with a new DT that only specifies > a 4KB config aperture? Yes. Both driver and DT changes must go together. New driver and old DT (or) Old driver and new DT would break config space access. > >> Signed-off-by: Vidya Sagar <vidyas@nvidia.com> >> --- >> drivers/pci/host/pci-tegra.c | 85 ++++++++++++-------------------------------- >> 1 file changed, 23 insertions(+), 62 deletions(-) >> >> diff --git a/drivers/pci/host/pci-tegra.c b/drivers/pci/host/pci-tegra.c >> index 9c40da54f88a..ebdcfca10e17 100644 >> --- a/drivers/pci/host/pci-tegra.c >> +++ b/drivers/pci/host/pci-tegra.c >> @@ -269,6 +269,8 @@ struct tegra_pcie { >> struct list_head buses; >> struct resource *cs; >> >> + void __iomem *cfg_va_base; >> + >> struct resource io; >> struct resource pio; >> struct resource mem; >> @@ -317,7 +319,6 @@ struct tegra_pcie_port { >> }; >> >> struct tegra_pcie_bus { >> - struct vm_struct *area; >> struct list_head list; >> unsigned int nr; >> }; >> @@ -357,34 +358,16 @@ static inline u32 pads_readl(struct tegra_pcie *pcie, unsigned long offset) >> * >> * Mapping the whole extended configuration space would require 256 MiB of >> * virtual address space, only a small part of which will actually be used. >> - * To work around this, a 1 MiB of virtual addresses are allocated per bus >> - * when the bus is first accessed. When the physical range is mapped, the >> - * the bus number bits are hidden so that the extended register number bits >> - * appear as bits [19:16]. Therefore the virtual mapping looks like this: >> - * >> - * [19:16] extended register number >> - * [15:11] device number >> - * [10: 8] function number >> - * [ 7: 0] register number >> - * >> - * This is achieved by stitching together 16 chunks of 64 KiB of physical >> - * address space via the MMU. >> + * To work around this, a 4K of region is used to generate required >> + * configuration transaction with relevant B:D:F values. This is achieved by >> + * dynamically programming base address and size of AFI_AXI_BAR used for >> + * end point config space mapping to make sure that the address (access to >> + * which generates correct config transaction) falls in this 4K region >> */ >> -static unsigned long tegra_pcie_conf_offset(unsigned int devfn, int where) >> -{ >> - return ((where & 0xf00) << 8) | (PCI_SLOT(devfn) << 11) | >> - (PCI_FUNC(devfn) << 8) | (where & 0xfc); >> -} >> - >> static struct tegra_pcie_bus *tegra_pcie_bus_alloc(struct tegra_pcie *pcie, >> unsigned int busnr) >> { >> - struct device *dev = pcie->dev; >> - pgprot_t prot = pgprot_noncached(PAGE_KERNEL); >> - phys_addr_t cs = pcie->cs->start; >> struct tegra_pcie_bus *bus; >> - unsigned int i; >> - int err; >> >> bus = kzalloc(sizeof(*bus), GFP_KERNEL); >> if (!bus) >> @@ -393,33 +376,16 @@ static struct tegra_pcie_bus *tegra_pcie_bus_alloc(struct tegra_pcie *pcie, >> INIT_LIST_HEAD(&bus->list); >> bus->nr = busnr; >> >> - /* allocate 1 MiB of virtual addresses */ >> - bus->area = get_vm_area(SZ_1M, VM_IOREMAP); >> - if (!bus->area) { >> - err = -ENOMEM; >> - goto free; >> - } >> - >> - /* map each of the 16 chunks of 64 KiB each */ >> - for (i = 0; i < 16; i++) { >> - unsigned long virt = (unsigned long)bus->area->addr + >> - i * SZ_64K; >> - phys_addr_t phys = cs + i * SZ_16M + busnr * SZ_64K; >> - >> - err = ioremap_page_range(virt, virt + SZ_64K, phys, prot); >> - if (err < 0) { >> - dev_err(dev, "ioremap_page_range() failed: %d\n", err); >> - goto unmap; >> + if (!pcie->cfg_va_base) { >> + pcie->cfg_va_base = ioremap(pcie->cs->start, SZ_4K); >> + if (!pcie->cfg_va_base) { >> + dev_err(pcie->dev, "failed to ioremap config space\n"); >> + kfree(bus); >> + bus = (struct tegra_pcie_bus *)-ENOMEM; > tegra_pcie_bus_alloc() used to do some bus-specific work -- it allocated VM > space and mapped it. This pcie->cfg_va_base ioremap() is not bus-specific, > so I don't think it fits here. It looks like something that should be done > in the tegra_pcie_probe() path instead. > > If you move that, then I think you might as well inline the kzalloc() stuff > into tegra_pcie_add_bus(). I'll take care of it in next patch. >> } >> } >> >> return bus; >> - >> -unmap: >> - vunmap(bus->area->addr); >> -free: >> - kfree(bus); >> - return ERR_PTR(err); >> } >> >> static int tegra_pcie_add_bus(struct pci_bus *bus) >> @@ -445,12 +411,12 @@ static void tegra_pcie_remove_bus(struct pci_bus *child) >> >> list_for_each_entry_safe(bus, tmp, &pcie->buses, list) { >> if (bus->nr == child->number) { >> - vunmap(bus->area->addr); >> list_del(&bus->list); >> kfree(bus); >> break; >> } >> } >> + iounmap(pcie->cfg_va_base); >> } >> >> static void __iomem *tegra_pcie_map_bus(struct pci_bus *bus, >> @@ -459,8 +425,8 @@ static void __iomem *tegra_pcie_map_bus(struct pci_bus *bus, >> { >> struct pci_host_bridge *host = pci_find_host_bridge(bus); >> struct tegra_pcie *pcie = pci_host_bridge_priv(host); >> - struct device *dev = pcie->dev; >> void __iomem *addr = NULL; >> + u32 val = 0; >> >> if (bus->number == 0) { >> unsigned int slot = PCI_SLOT(devfn); >> @@ -473,19 +439,14 @@ static void __iomem *tegra_pcie_map_bus(struct pci_bus *bus, >> } >> } >> } else { >> - struct tegra_pcie_bus *b; >> - >> - list_for_each_entry(b, &pcie->buses, list) >> - if (b->nr == bus->number) >> - addr = (void __iomem *)b->area->addr; >> - >> - if (!addr) { >> - dev_err(dev, "failed to map cfg. space for bus %u\n", >> - bus->number); >> - return NULL; >> - } >> - >> - addr += tegra_pcie_conf_offset(devfn, where); >> + addr = pcie->cfg_va_base; >> + val = ((((u32)where & 0xf00) >> 8) << 24) | >> + (bus->number << 16) | (PCI_SLOT(devfn) << 11) | >> + (PCI_FUNC(devfn) << 8) | (where & 0xff); > I think it would be useful to keep tegra_pcie_conf_offset() (extending > it to take the bus number) near the comment describing the mapping. I'll take care of it in next patch. > >> + addr = (val & (SZ_4K - 1)) + addr; >> + val = val & ~(SZ_4K - 1); > val &= ~(SZ_4K - 1); > > I *think* what you're doing is computing a config space offset and > mapping the 4KB window that contains it. > > "addr" is not used at all in the AFI_AXI_BAR0 programming, so this > would be clearer as: > > val = ... > afi_writel(...) > afi_writel(...) > > addr = pcie->cfg_va_base + (val & (SZ_4K - 1)); but, addr is using the intermediate 'val' i.e. before 'val &= ~(SZ_4K - 1)' operation is performed on it. > } > > return addr; > >> + afi_writel(pcie, pcie->cs->start - val, AFI_AXI_BAR0_START); >> + afi_writel(pcie, (val + SZ_4K) >> 12, AFI_AXI_BAR0_SZ); >> } >> >> return addr; >> -- >> 2.7.4 >>
[+cc Rob, devicetree list] On Mon, Oct 23, 2017 at 11:09:50AM +0530, Vidya Sagar wrote: > On Saturday 21 October 2017 12:38 AM, Bjorn Helgaas wrote: > >On Fri, Oct 13, 2017 at 12:20:06AM +0530, Vidya Sagar wrote: > >>use only 4K space from available 1GB PCIe aperture to access > >>end points configuration space by dynamically moving AFI_AXI_BAR > >>base address and always making sure that the desired location > >>to be accessed for generating required config space access falls > >>in the 4K space reserved for this purpose. This would give more > >>space for mapping end point device's BARs on some of Tegra platforms > > > >Do you now have a revlock issue between this driver change and the > >corresponding DT changes? I.e., what happens when you run this driver > >change with an older DT that specifies a 256MB config aperture? What > >happens if you run the older driver with a new DT that only specifies > >a 4KB config aperture? > > Yes. Both driver and DT changes must go together. New driver and old DT (or) > Old driver and new DT would break config space access. This revlock sounds like a pretty major issue. I don't know if the DT conventions even allow for this. If they do, it definitely needs to be mentioned in the changelog. Looking for acks from Thierry and Rob. > >>- addr += tegra_pcie_conf_offset(devfn, where); > >>+ addr = pcie->cfg_va_base; > >>+ val = ((((u32)where & 0xf00) >> 8) << 24) | > >>+ (bus->number << 16) | (PCI_SLOT(devfn) << 11) | > >>+ (PCI_FUNC(devfn) << 8) | (where & 0xff); > >I think it would be useful to keep tegra_pcie_conf_offset() (extending > >it to take the bus number) near the comment describing the mapping. > I'll take care of it in next patch. > > > >>+ addr = (val & (SZ_4K - 1)) + addr; > >>+ val = val & ~(SZ_4K - 1); > > val &= ~(SZ_4K - 1); > > > >I *think* what you're doing is computing a config space offset and > >mapping the 4KB window that contains it. > > > >"addr" is not used at all in the AFI_AXI_BAR0 programming, so this > >would be clearer as: > > > > val = ... > > afi_writel(...) > > afi_writel(...) > > > > addr = pcie->cfg_va_base + (val & (SZ_4K - 1)); > but, addr is using the intermediate 'val' i.e. before 'val &= > ~(SZ_4K - 1)' operation is performed on it. Ah, right. I think adding another temporary variable here would make it clearer. > > } > > > > return addr; > > > >>+ afi_writel(pcie, pcie->cs->start - val, AFI_AXI_BAR0_START); > >>+ afi_writel(pcie, (val + SZ_4K) >> 12, AFI_AXI_BAR0_SZ); > >> } > >> return addr; > >>-- > >>2.7.4 > >> >
diff --git a/drivers/pci/host/pci-tegra.c b/drivers/pci/host/pci-tegra.c index 9c40da54f88a..ebdcfca10e17 100644 --- a/drivers/pci/host/pci-tegra.c +++ b/drivers/pci/host/pci-tegra.c @@ -269,6 +269,8 @@ struct tegra_pcie { struct list_head buses; struct resource *cs; + void __iomem *cfg_va_base; + struct resource io; struct resource pio; struct resource mem; @@ -317,7 +319,6 @@ struct tegra_pcie_port { }; struct tegra_pcie_bus { - struct vm_struct *area; struct list_head list; unsigned int nr; }; @@ -357,34 +358,16 @@ static inline u32 pads_readl(struct tegra_pcie *pcie, unsigned long offset) * * Mapping the whole extended configuration space would require 256 MiB of * virtual address space, only a small part of which will actually be used. - * To work around this, a 1 MiB of virtual addresses are allocated per bus - * when the bus is first accessed. When the physical range is mapped, the - * the bus number bits are hidden so that the extended register number bits - * appear as bits [19:16]. Therefore the virtual mapping looks like this: - * - * [19:16] extended register number - * [15:11] device number - * [10: 8] function number - * [ 7: 0] register number - * - * This is achieved by stitching together 16 chunks of 64 KiB of physical - * address space via the MMU. + * To work around this, a 4K of region is used to generate required + * configuration transaction with relevant B:D:F values. This is achieved by + * dynamically programming base address and size of AFI_AXI_BAR used for + * end point config space mapping to make sure that the address (access to + * which generates correct config transaction) falls in this 4K region */ -static unsigned long tegra_pcie_conf_offset(unsigned int devfn, int where) -{ - return ((where & 0xf00) << 8) | (PCI_SLOT(devfn) << 11) | - (PCI_FUNC(devfn) << 8) | (where & 0xfc); -} - static struct tegra_pcie_bus *tegra_pcie_bus_alloc(struct tegra_pcie *pcie, unsigned int busnr) { - struct device *dev = pcie->dev; - pgprot_t prot = pgprot_noncached(PAGE_KERNEL); - phys_addr_t cs = pcie->cs->start; struct tegra_pcie_bus *bus; - unsigned int i; - int err; bus = kzalloc(sizeof(*bus), GFP_KERNEL); if (!bus) @@ -393,33 +376,16 @@ static struct tegra_pcie_bus *tegra_pcie_bus_alloc(struct tegra_pcie *pcie, INIT_LIST_HEAD(&bus->list); bus->nr = busnr; - /* allocate 1 MiB of virtual addresses */ - bus->area = get_vm_area(SZ_1M, VM_IOREMAP); - if (!bus->area) { - err = -ENOMEM; - goto free; - } - - /* map each of the 16 chunks of 64 KiB each */ - for (i = 0; i < 16; i++) { - unsigned long virt = (unsigned long)bus->area->addr + - i * SZ_64K; - phys_addr_t phys = cs + i * SZ_16M + busnr * SZ_64K; - - err = ioremap_page_range(virt, virt + SZ_64K, phys, prot); - if (err < 0) { - dev_err(dev, "ioremap_page_range() failed: %d\n", err); - goto unmap; + if (!pcie->cfg_va_base) { + pcie->cfg_va_base = ioremap(pcie->cs->start, SZ_4K); + if (!pcie->cfg_va_base) { + dev_err(pcie->dev, "failed to ioremap config space\n"); + kfree(bus); + bus = (struct tegra_pcie_bus *)-ENOMEM; } } return bus; - -unmap: - vunmap(bus->area->addr); -free: - kfree(bus); - return ERR_PTR(err); } static int tegra_pcie_add_bus(struct pci_bus *bus) @@ -445,12 +411,12 @@ static void tegra_pcie_remove_bus(struct pci_bus *child) list_for_each_entry_safe(bus, tmp, &pcie->buses, list) { if (bus->nr == child->number) { - vunmap(bus->area->addr); list_del(&bus->list); kfree(bus); break; } } + iounmap(pcie->cfg_va_base); } static void __iomem *tegra_pcie_map_bus(struct pci_bus *bus, @@ -459,8 +425,8 @@ static void __iomem *tegra_pcie_map_bus(struct pci_bus *bus, { struct pci_host_bridge *host = pci_find_host_bridge(bus); struct tegra_pcie *pcie = pci_host_bridge_priv(host); - struct device *dev = pcie->dev; void __iomem *addr = NULL; + u32 val = 0; if (bus->number == 0) { unsigned int slot = PCI_SLOT(devfn); @@ -473,19 +439,14 @@ static void __iomem *tegra_pcie_map_bus(struct pci_bus *bus, } } } else { - struct tegra_pcie_bus *b; - - list_for_each_entry(b, &pcie->buses, list) - if (b->nr == bus->number) - addr = (void __iomem *)b->area->addr; - - if (!addr) { - dev_err(dev, "failed to map cfg. space for bus %u\n", - bus->number); - return NULL; - } - - addr += tegra_pcie_conf_offset(devfn, where); + addr = pcie->cfg_va_base; + val = ((((u32)where & 0xf00) >> 8) << 24) | + (bus->number << 16) | (PCI_SLOT(devfn) << 11) | + (PCI_FUNC(devfn) << 8) | (where & 0xff); + addr = (val & (SZ_4K - 1)) + addr; + val = val & ~(SZ_4K - 1); + afi_writel(pcie, pcie->cs->start - val, AFI_AXI_BAR0_START); + afi_writel(pcie, (val + SZ_4K) >> 12, AFI_AXI_BAR0_SZ); } return addr;
use only 4K space from available 1GB PCIe aperture to access end points configuration space by dynamically moving AFI_AXI_BAR base address and always making sure that the desired location to be accessed for generating required config space access falls in the 4K space reserved for this purpose. This would give more space for mapping end point device's BARs on some of Tegra platforms Signed-off-by: Vidya Sagar <vidyas@nvidia.com> --- drivers/pci/host/pci-tegra.c | 85 ++++++++++++-------------------------------- 1 file changed, 23 insertions(+), 62 deletions(-)