From 57f0f512b273f60d52568b8c6b77e17f5636edc0 Mon Sep 17 00:00:00 2001 From: AndrĂ© Fabian Silva Delgado Date: Wed, 5 Aug 2015 17:04:01 -0300 Subject: Initial import --- arch/ia64/sn/pci/Makefile | 12 + arch/ia64/sn/pci/pci_dma.c | 487 ++++++++++++++ arch/ia64/sn/pci/pcibr/Makefile | 13 + arch/ia64/sn/pci/pcibr/pcibr_ate.c | 177 ++++++ arch/ia64/sn/pci/pcibr/pcibr_dma.c | 413 ++++++++++++ arch/ia64/sn/pci/pcibr/pcibr_provider.c | 265 ++++++++ arch/ia64/sn/pci/pcibr/pcibr_reg.c | 285 +++++++++ arch/ia64/sn/pci/tioca_provider.c | 677 ++++++++++++++++++++ arch/ia64/sn/pci/tioce_provider.c | 1062 +++++++++++++++++++++++++++++++ 9 files changed, 3391 insertions(+) create mode 100644 arch/ia64/sn/pci/Makefile create mode 100644 arch/ia64/sn/pci/pci_dma.c create mode 100644 arch/ia64/sn/pci/pcibr/Makefile create mode 100644 arch/ia64/sn/pci/pcibr/pcibr_ate.c create mode 100644 arch/ia64/sn/pci/pcibr/pcibr_dma.c create mode 100644 arch/ia64/sn/pci/pcibr/pcibr_provider.c create mode 100644 arch/ia64/sn/pci/pcibr/pcibr_reg.c create mode 100644 arch/ia64/sn/pci/tioca_provider.c create mode 100644 arch/ia64/sn/pci/tioce_provider.c (limited to 'arch/ia64/sn/pci') diff --git a/arch/ia64/sn/pci/Makefile b/arch/ia64/sn/pci/Makefile new file mode 100644 index 000000000..df2a90145 --- /dev/null +++ b/arch/ia64/sn/pci/Makefile @@ -0,0 +1,12 @@ +# +# This file is subject to the terms and conditions of the GNU General Public +# License. See the file "COPYING" in the main directory of this archive +# for more details. +# +# Copyright (C) 2000-2004 Silicon Graphics, Inc. All Rights Reserved. +# +# Makefile for the sn pci general routines. + +ccflags-y := -Iarch/ia64/sn/include + +obj-y := pci_dma.o tioca_provider.o tioce_provider.o pcibr/ diff --git a/arch/ia64/sn/pci/pci_dma.c b/arch/ia64/sn/pci/pci_dma.c new file mode 100644 index 000000000..d0853e8e8 --- /dev/null +++ b/arch/ia64/sn/pci/pci_dma.c @@ -0,0 +1,487 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * Copyright (C) 2000,2002-2005 Silicon Graphics, Inc. All rights reserved. + * + * Routines for PCI DMA mapping. See Documentation/DMA-API.txt for + * a description of how these routines should be used. + */ + +#include +#include +#include +#include +#include +#include +#include +#include + +#define SG_ENT_VIRT_ADDRESS(sg) (sg_virt((sg))) +#define SG_ENT_PHYS_ADDRESS(SG) virt_to_phys(SG_ENT_VIRT_ADDRESS(SG)) + +/** + * sn_dma_supported - test a DMA mask + * @dev: device to test + * @mask: DMA mask to test + * + * Return whether the given PCI device DMA address mask can be supported + * properly. For example, if your device can only drive the low 24-bits + * during PCI bus mastering, then you would pass 0x00ffffff as the mask to + * this function. Of course, SN only supports devices that have 32 or more + * address bits when using the PMU. + */ +static int sn_dma_supported(struct device *dev, u64 mask) +{ + BUG_ON(!dev_is_pci(dev)); + + if (mask < 0x7fffffff) + return 0; + return 1; +} + +/** + * sn_dma_set_mask - set the DMA mask + * @dev: device to set + * @dma_mask: new mask + * + * Set @dev's DMA mask if the hw supports it. + */ +int sn_dma_set_mask(struct device *dev, u64 dma_mask) +{ + BUG_ON(!dev_is_pci(dev)); + + if (!sn_dma_supported(dev, dma_mask)) + return 0; + + *dev->dma_mask = dma_mask; + return 1; +} +EXPORT_SYMBOL(sn_dma_set_mask); + +/** + * sn_dma_alloc_coherent - allocate memory for coherent DMA + * @dev: device to allocate for + * @size: size of the region + * @dma_handle: DMA (bus) address + * @flags: memory allocation flags + * + * dma_alloc_coherent() returns a pointer to a memory region suitable for + * coherent DMA traffic to/from a PCI device. On SN platforms, this means + * that @dma_handle will have the %PCIIO_DMA_CMD flag set. + * + * This interface is usually used for "command" streams (e.g. the command + * queue for a SCSI controller). See Documentation/DMA-API.txt for + * more information. + */ +static void *sn_dma_alloc_coherent(struct device *dev, size_t size, + dma_addr_t * dma_handle, gfp_t flags, + struct dma_attrs *attrs) +{ + void *cpuaddr; + unsigned long phys_addr; + int node; + struct pci_dev *pdev = to_pci_dev(dev); + struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev); + + BUG_ON(!dev_is_pci(dev)); + + /* + * Allocate the memory. + */ + node = pcibus_to_node(pdev->bus); + if (likely(node >=0)) { + struct page *p = alloc_pages_exact_node(node, + flags, get_order(size)); + + if (likely(p)) + cpuaddr = page_address(p); + else + return NULL; + } else + cpuaddr = (void *)__get_free_pages(flags, get_order(size)); + + if (unlikely(!cpuaddr)) + return NULL; + + memset(cpuaddr, 0x0, size); + + /* physical addr. of the memory we just got */ + phys_addr = __pa(cpuaddr); + + /* + * 64 bit address translations should never fail. + * 32 bit translations can fail if there are insufficient mapping + * resources. + */ + + *dma_handle = provider->dma_map_consistent(pdev, phys_addr, size, + SN_DMA_ADDR_PHYS); + if (!*dma_handle) { + printk(KERN_ERR "%s: out of ATEs\n", __func__); + free_pages((unsigned long)cpuaddr, get_order(size)); + return NULL; + } + + return cpuaddr; +} + +/** + * sn_pci_free_coherent - free memory associated with coherent DMAable region + * @dev: device to free for + * @size: size to free + * @cpu_addr: kernel virtual address to free + * @dma_handle: DMA address associated with this region + * + * Frees the memory allocated by dma_alloc_coherent(), potentially unmapping + * any associated IOMMU mappings. + */ +static void sn_dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, + dma_addr_t dma_handle, struct dma_attrs *attrs) +{ + struct pci_dev *pdev = to_pci_dev(dev); + struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev); + + BUG_ON(!dev_is_pci(dev)); + + provider->dma_unmap(pdev, dma_handle, 0); + free_pages((unsigned long)cpu_addr, get_order(size)); +} + +/** + * sn_dma_map_single_attrs - map a single page for DMA + * @dev: device to map for + * @cpu_addr: kernel virtual address of the region to map + * @size: size of the region + * @direction: DMA direction + * @attrs: optional dma attributes + * + * Map the region pointed to by @cpu_addr for DMA and return the + * DMA address. + * + * We map this to the one step pcibr_dmamap_trans interface rather than + * the two step pcibr_dmamap_alloc/pcibr_dmamap_addr because we have + * no way of saving the dmamap handle from the alloc to later free + * (which is pretty much unacceptable). + * + * mappings with the DMA_ATTR_WRITE_BARRIER get mapped with + * dma_map_consistent() so that writes force a flush of pending DMA. + * (See "SGI Altix Architecture Considerations for Linux Device Drivers", + * Document Number: 007-4763-001) + * + * TODO: simplify our interface; + * figure out how to save dmamap handle so can use two step. + */ +static dma_addr_t sn_dma_map_page(struct device *dev, struct page *page, + unsigned long offset, size_t size, + enum dma_data_direction dir, + struct dma_attrs *attrs) +{ + void *cpu_addr = page_address(page) + offset; + dma_addr_t dma_addr; + unsigned long phys_addr; + struct pci_dev *pdev = to_pci_dev(dev); + struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev); + int dmabarr; + + dmabarr = dma_get_attr(DMA_ATTR_WRITE_BARRIER, attrs); + + BUG_ON(!dev_is_pci(dev)); + + phys_addr = __pa(cpu_addr); + if (dmabarr) + dma_addr = provider->dma_map_consistent(pdev, phys_addr, + size, SN_DMA_ADDR_PHYS); + else + dma_addr = provider->dma_map(pdev, phys_addr, size, + SN_DMA_ADDR_PHYS); + + if (!dma_addr) { + printk(KERN_ERR "%s: out of ATEs\n", __func__); + return 0; + } + return dma_addr; +} + +/** + * sn_dma_unmap_single_attrs - unamp a DMA mapped page + * @dev: device to sync + * @dma_addr: DMA address to sync + * @size: size of region + * @direction: DMA direction + * @attrs: optional dma attributes + * + * This routine is supposed to sync the DMA region specified + * by @dma_handle into the coherence domain. On SN, we're always cache + * coherent, so we just need to free any ATEs associated with this mapping. + */ +static void sn_dma_unmap_page(struct device *dev, dma_addr_t dma_addr, + size_t size, enum dma_data_direction dir, + struct dma_attrs *attrs) +{ + struct pci_dev *pdev = to_pci_dev(dev); + struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev); + + BUG_ON(!dev_is_pci(dev)); + + provider->dma_unmap(pdev, dma_addr, dir); +} + +/** + * sn_dma_unmap_sg - unmap a DMA scatterlist + * @dev: device to unmap + * @sg: scatterlist to unmap + * @nhwentries: number of scatterlist entries + * @direction: DMA direction + * @attrs: optional dma attributes + * + * Unmap a set of streaming mode DMA translations. + */ +static void sn_dma_unmap_sg(struct device *dev, struct scatterlist *sgl, + int nhwentries, enum dma_data_direction dir, + struct dma_attrs *attrs) +{ + int i; + struct pci_dev *pdev = to_pci_dev(dev); + struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev); + struct scatterlist *sg; + + BUG_ON(!dev_is_pci(dev)); + + for_each_sg(sgl, sg, nhwentries, i) { + provider->dma_unmap(pdev, sg->dma_address, dir); + sg->dma_address = (dma_addr_t) NULL; + sg->dma_length = 0; + } +} + +/** + * sn_dma_map_sg - map a scatterlist for DMA + * @dev: device to map for + * @sg: scatterlist to map + * @nhwentries: number of entries + * @direction: direction of the DMA transaction + * @attrs: optional dma attributes + * + * mappings with the DMA_ATTR_WRITE_BARRIER get mapped with + * dma_map_consistent() so that writes force a flush of pending DMA. + * (See "SGI Altix Architecture Considerations for Linux Device Drivers", + * Document Number: 007-4763-001) + * + * Maps each entry of @sg for DMA. + */ +static int sn_dma_map_sg(struct device *dev, struct scatterlist *sgl, + int nhwentries, enum dma_data_direction dir, + struct dma_attrs *attrs) +{ + unsigned long phys_addr; + struct scatterlist *saved_sg = sgl, *sg; + struct pci_dev *pdev = to_pci_dev(dev); + struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev); + int i; + int dmabarr; + + dmabarr = dma_get_attr(DMA_ATTR_WRITE_BARRIER, attrs); + + BUG_ON(!dev_is_pci(dev)); + + /* + * Setup a DMA address for each entry in the scatterlist. + */ + for_each_sg(sgl, sg, nhwentries, i) { + dma_addr_t dma_addr; + phys_addr = SG_ENT_PHYS_ADDRESS(sg); + if (dmabarr) + dma_addr = provider->dma_map_consistent(pdev, + phys_addr, + sg->length, + SN_DMA_ADDR_PHYS); + else + dma_addr = provider->dma_map(pdev, phys_addr, + sg->length, + SN_DMA_ADDR_PHYS); + + sg->dma_address = dma_addr; + if (!sg->dma_address) { + printk(KERN_ERR "%s: out of ATEs\n", __func__); + + /* + * Free any successfully allocated entries. + */ + if (i > 0) + sn_dma_unmap_sg(dev, saved_sg, i, dir, attrs); + return 0; + } + + sg->dma_length = sg->length; + } + + return nhwentries; +} + +static void sn_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle, + size_t size, enum dma_data_direction dir) +{ + BUG_ON(!dev_is_pci(dev)); +} + +static void sn_dma_sync_single_for_device(struct device *dev, dma_addr_t dma_handle, + size_t size, + enum dma_data_direction dir) +{ + BUG_ON(!dev_is_pci(dev)); +} + +static void sn_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg, + int nelems, enum dma_data_direction dir) +{ + BUG_ON(!dev_is_pci(dev)); +} + +static void sn_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg, + int nelems, enum dma_data_direction dir) +{ + BUG_ON(!dev_is_pci(dev)); +} + +static int sn_dma_mapping_error(struct device *dev, dma_addr_t dma_addr) +{ + return 0; +} + +u64 sn_dma_get_required_mask(struct device *dev) +{ + return DMA_BIT_MASK(64); +} +EXPORT_SYMBOL_GPL(sn_dma_get_required_mask); + +char *sn_pci_get_legacy_mem(struct pci_bus *bus) +{ + if (!SN_PCIBUS_BUSSOFT(bus)) + return ERR_PTR(-ENODEV); + + return (char *)(SN_PCIBUS_BUSSOFT(bus)->bs_legacy_mem | __IA64_UNCACHED_OFFSET); +} + +int sn_pci_legacy_read(struct pci_bus *bus, u16 port, u32 *val, u8 size) +{ + unsigned long addr; + int ret; + struct ia64_sal_retval isrv; + + /* + * First, try the SN_SAL_IOIF_PCI_SAFE SAL call which can work + * around hw issues at the pci bus level. SGI proms older than + * 4.10 don't implement this. + */ + + SAL_CALL(isrv, SN_SAL_IOIF_PCI_SAFE, + pci_domain_nr(bus), bus->number, + 0, /* io */ + 0, /* read */ + port, size, __pa(val)); + + if (isrv.status == 0) + return size; + + /* + * If the above failed, retry using the SAL_PROBE call which should + * be present in all proms (but which cannot work round PCI chipset + * bugs). This code is retained for compatibility with old + * pre-4.10 proms, and should be removed at some point in the future. + */ + + if (!SN_PCIBUS_BUSSOFT(bus)) + return -ENODEV; + + addr = SN_PCIBUS_BUSSOFT(bus)->bs_legacy_io | __IA64_UNCACHED_OFFSET; + addr += port; + + ret = ia64_sn_probe_mem(addr, (long)size, (void *)val); + + if (ret == 2) + return -EINVAL; + + if (ret == 1) + *val = -1; + + return size; +} + +int sn_pci_legacy_write(struct pci_bus *bus, u16 port, u32 val, u8 size) +{ + int ret = size; + unsigned long paddr; + unsigned long *addr; + struct ia64_sal_retval isrv; + + /* + * First, try the SN_SAL_IOIF_PCI_SAFE SAL call which can work + * around hw issues at the pci bus level. SGI proms older than + * 4.10 don't implement this. + */ + + SAL_CALL(isrv, SN_SAL_IOIF_PCI_SAFE, + pci_domain_nr(bus), bus->number, + 0, /* io */ + 1, /* write */ + port, size, __pa(&val)); + + if (isrv.status == 0) + return size; + + /* + * If the above failed, retry using the SAL_PROBE call which should + * be present in all proms (but which cannot work round PCI chipset + * bugs). This code is retained for compatibility with old + * pre-4.10 proms, and should be removed at some point in the future. + */ + + if (!SN_PCIBUS_BUSSOFT(bus)) { + ret = -ENODEV; + goto out; + } + + /* Put the phys addr in uncached space */ + paddr = SN_PCIBUS_BUSSOFT(bus)->bs_legacy_io | __IA64_UNCACHED_OFFSET; + paddr += port; + addr = (unsigned long *)paddr; + + switch (size) { + case 1: + *(volatile u8 *)(addr) = (u8)(val); + break; + case 2: + *(volatile u16 *)(addr) = (u16)(val); + break; + case 4: + *(volatile u32 *)(addr) = (u32)(val); + break; + default: + ret = -EINVAL; + break; + } + out: + return ret; +} + +static struct dma_map_ops sn_dma_ops = { + .alloc = sn_dma_alloc_coherent, + .free = sn_dma_free_coherent, + .map_page = sn_dma_map_page, + .unmap_page = sn_dma_unmap_page, + .map_sg = sn_dma_map_sg, + .unmap_sg = sn_dma_unmap_sg, + .sync_single_for_cpu = sn_dma_sync_single_for_cpu, + .sync_sg_for_cpu = sn_dma_sync_sg_for_cpu, + .sync_single_for_device = sn_dma_sync_single_for_device, + .sync_sg_for_device = sn_dma_sync_sg_for_device, + .mapping_error = sn_dma_mapping_error, + .dma_supported = sn_dma_supported, +}; + +void sn_dma_init(void) +{ + dma_ops = &sn_dma_ops; +} diff --git a/arch/ia64/sn/pci/pcibr/Makefile b/arch/ia64/sn/pci/pcibr/Makefile new file mode 100644 index 000000000..396bcae36 --- /dev/null +++ b/arch/ia64/sn/pci/pcibr/Makefile @@ -0,0 +1,13 @@ +# +# This file is subject to the terms and conditions of the GNU General Public +# License. See the file "COPYING" in the main directory of this archive +# for more details. +# +# Copyright (C) 2002-2004 Silicon Graphics, Inc. All Rights Reserved. +# +# Makefile for the sn2 io routines. + +ccflags-y := -Iarch/ia64/sn/include + +obj-y += pcibr_dma.o pcibr_reg.o \ + pcibr_ate.o pcibr_provider.o diff --git a/arch/ia64/sn/pci/pcibr/pcibr_ate.c b/arch/ia64/sn/pci/pcibr/pcibr_ate.c new file mode 100644 index 000000000..5bc34eac9 --- /dev/null +++ b/arch/ia64/sn/pci/pcibr/pcibr_ate.c @@ -0,0 +1,177 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * Copyright (C) 2001-2006 Silicon Graphics, Inc. All rights reserved. + */ + +#include +#include +#include +#include +#include + +int pcibr_invalidate_ate; /* by default don't invalidate ATE on free */ + +/* + * mark_ate: Mark the ate as either free or inuse. + */ +static void mark_ate(struct ate_resource *ate_resource, int start, int number, + u64 value) +{ + u64 *ate = ate_resource->ate; + int index; + int length = 0; + + for (index = start; length < number; index++, length++) + ate[index] = value; +} + +/* + * find_free_ate: Find the first free ate index starting from the given + * index for the desired consecutive count. + */ +static int find_free_ate(struct ate_resource *ate_resource, int start, + int count) +{ + u64 *ate = ate_resource->ate; + int index; + int start_free; + + for (index = start; index < ate_resource->num_ate;) { + if (!ate[index]) { + int i; + int free; + free = 0; + start_free = index; /* Found start free ate */ + for (i = start_free; i < ate_resource->num_ate; i++) { + if (!ate[i]) { /* This is free */ + if (++free == count) + return start_free; + } else { + index = i + 1; + break; + } + } + if (i >= ate_resource->num_ate) + return -1; + } else + index++; /* Try next ate */ + } + + return -1; +} + +/* + * free_ate_resource: Free the requested number of ATEs. + */ +static inline void free_ate_resource(struct ate_resource *ate_resource, + int start) +{ + mark_ate(ate_resource, start, ate_resource->ate[start], 0); + if ((ate_resource->lowest_free_index > start) || + (ate_resource->lowest_free_index < 0)) + ate_resource->lowest_free_index = start; +} + +/* + * alloc_ate_resource: Allocate the requested number of ATEs. + */ +static inline int alloc_ate_resource(struct ate_resource *ate_resource, + int ate_needed) +{ + int start_index; + + /* + * Check for ate exhaustion. + */ + if (ate_resource->lowest_free_index < 0) + return -1; + + /* + * Find the required number of free consecutive ates. + */ + start_index = + find_free_ate(ate_resource, ate_resource->lowest_free_index, + ate_needed); + if (start_index >= 0) + mark_ate(ate_resource, start_index, ate_needed, ate_needed); + + ate_resource->lowest_free_index = + find_free_ate(ate_resource, ate_resource->lowest_free_index, 1); + + return start_index; +} + +/* + * Allocate "count" contiguous Bridge Address Translation Entries + * on the specified bridge to be used for PCI to XTALK mappings. + * Indices in rm map range from 1..num_entries. Indices returned + * to caller range from 0..num_entries-1. + * + * Return the start index on success, -1 on failure. + */ +int pcibr_ate_alloc(struct pcibus_info *pcibus_info, int count) +{ + int status; + unsigned long flags; + + spin_lock_irqsave(&pcibus_info->pbi_lock, flags); + status = alloc_ate_resource(&pcibus_info->pbi_int_ate_resource, count); + spin_unlock_irqrestore(&pcibus_info->pbi_lock, flags); + + return status; +} + +/* + * Setup an Address Translation Entry as specified. Use either the Bridge + * internal maps or the external map RAM, as appropriate. + */ +static inline u64 __iomem *pcibr_ate_addr(struct pcibus_info *pcibus_info, + int ate_index) +{ + if (ate_index < pcibus_info->pbi_int_ate_size) { + return pcireg_int_ate_addr(pcibus_info, ate_index); + } + panic("pcibr_ate_addr: invalid ate_index 0x%x", ate_index); +} + +/* + * Update the ate. + */ +void inline +ate_write(struct pcibus_info *pcibus_info, int ate_index, int count, + volatile u64 ate) +{ + while (count-- > 0) { + if (ate_index < pcibus_info->pbi_int_ate_size) { + pcireg_int_ate_set(pcibus_info, ate_index, ate); + } else { + panic("ate_write: invalid ate_index 0x%x", ate_index); + } + ate_index++; + ate += IOPGSIZE; + } + + pcireg_tflush_get(pcibus_info); /* wait until Bridge PIO complete */ +} + +void pcibr_ate_free(struct pcibus_info *pcibus_info, int index) +{ + + volatile u64 ate; + int count; + unsigned long flags; + + if (pcibr_invalidate_ate) { + /* For debugging purposes, clear the valid bit in the ATE */ + ate = *pcibr_ate_addr(pcibus_info, index); + count = pcibus_info->pbi_int_ate_resource.ate[index]; + ate_write(pcibus_info, index, count, (ate & ~PCI32_ATE_V)); + } + + spin_lock_irqsave(&pcibus_info->pbi_lock, flags); + free_ate_resource(&pcibus_info->pbi_int_ate_resource, index); + spin_unlock_irqrestore(&pcibus_info->pbi_lock, flags); +} diff --git a/arch/ia64/sn/pci/pcibr/pcibr_dma.c b/arch/ia64/sn/pci/pcibr/pcibr_dma.c new file mode 100644 index 000000000..1e863b277 --- /dev/null +++ b/arch/ia64/sn/pci/pcibr/pcibr_dma.c @@ -0,0 +1,413 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * Copyright (C) 2001-2005 Silicon Graphics, Inc. All rights reserved. + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "tio.h" +#include "xtalk/xwidgetdev.h" +#include "xtalk/hubdev.h" + +extern int sn_ioif_inited; + +/* ===================================================================== + * DMA MANAGEMENT + * + * The Bridge ASIC provides three methods of doing DMA: via a "direct map" + * register available in 32-bit PCI space (which selects a contiguous 2G + * address space on some other widget), via "direct" addressing via 64-bit + * PCI space (all destination information comes from the PCI address, + * including transfer attributes), and via a "mapped" region that allows + * a bunch of different small mappings to be established with the PMU. + * + * For efficiency, we most prefer to use the 32bit direct mapping facility, + * since it requires no resource allocations. The advantage of using the + * PMU over the 64-bit direct is that single-cycle PCI addressing can be + * used; the advantage of using 64-bit direct over PMU addressing is that + * we do not have to allocate entries in the PMU. + */ + +static dma_addr_t +pcibr_dmamap_ate32(struct pcidev_info *info, + u64 paddr, size_t req_size, u64 flags, int dma_flags) +{ + + struct pcidev_info *pcidev_info = info->pdi_host_pcidev_info; + struct pcibus_info *pcibus_info = (struct pcibus_info *)pcidev_info-> + pdi_pcibus_info; + u8 internal_device = (PCI_SLOT(pcidev_info->pdi_host_pcidev_info-> + pdi_linux_pcidev->devfn)) - 1; + int ate_count; + int ate_index; + u64 ate_flags = flags | PCI32_ATE_V; + u64 ate; + u64 pci_addr; + u64 xio_addr; + u64 offset; + + /* PIC in PCI-X mode does not supports 32bit PageMap mode */ + if (IS_PIC_SOFT(pcibus_info) && IS_PCIX(pcibus_info)) { + return 0; + } + + /* Calculate the number of ATEs needed. */ + if (!(MINIMAL_ATE_FLAG(paddr, req_size))) { + ate_count = IOPG((IOPGSIZE - 1) /* worst case start offset */ + +req_size /* max mapping bytes */ + - 1) + 1; /* round UP */ + } else { /* assume requested target is page aligned */ + ate_count = IOPG(req_size /* max mapping bytes */ + - 1) + 1; /* round UP */ + } + + /* Get the number of ATEs required. */ + ate_index = pcibr_ate_alloc(pcibus_info, ate_count); + if (ate_index < 0) + return 0; + + /* In PCI-X mode, Prefetch not supported */ + if (IS_PCIX(pcibus_info)) + ate_flags &= ~(PCI32_ATE_PREF); + + if (SN_DMA_ADDRTYPE(dma_flags == SN_DMA_ADDR_PHYS)) + xio_addr = IS_PIC_SOFT(pcibus_info) ? PHYS_TO_DMA(paddr) : + PHYS_TO_TIODMA(paddr); + else + xio_addr = paddr; + + offset = IOPGOFF(xio_addr); + ate = ate_flags | (xio_addr - offset); + + /* If PIC, put the targetid in the ATE */ + if (IS_PIC_SOFT(pcibus_info)) { + ate |= (pcibus_info->pbi_hub_xid << PIC_ATE_TARGETID_SHFT); + } + + /* + * If we're mapping for MSI, set the MSI bit in the ATE. If it's a + * TIOCP based pci bus, we also need to set the PIO bit in the ATE. + */ + if (dma_flags & SN_DMA_MSI) { + ate |= PCI32_ATE_MSI; + if (IS_TIOCP_SOFT(pcibus_info)) + ate |= PCI32_ATE_PIO; + } + + ate_write(pcibus_info, ate_index, ate_count, ate); + + /* + * Set up the DMA mapped Address. + */ + pci_addr = PCI32_MAPPED_BASE + offset + IOPGSIZE * ate_index; + + /* + * If swap was set in device in pcibr_endian_set() + * we need to turn swapping on. + */ + if (pcibus_info->pbi_devreg[internal_device] & PCIBR_DEV_SWAP_DIR) + ATE_SWAP_ON(pci_addr); + + + return pci_addr; +} + +static dma_addr_t +pcibr_dmatrans_direct64(struct pcidev_info * info, u64 paddr, + u64 dma_attributes, int dma_flags) +{ + struct pcibus_info *pcibus_info = (struct pcibus_info *) + ((info->pdi_host_pcidev_info)->pdi_pcibus_info); + u64 pci_addr; + + /* Translate to Crosstalk View of Physical Address */ + if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS) + pci_addr = IS_PIC_SOFT(pcibus_info) ? + PHYS_TO_DMA(paddr) : + PHYS_TO_TIODMA(paddr); + else + pci_addr = paddr; + pci_addr |= dma_attributes; + + /* Handle Bus mode */ + if (IS_PCIX(pcibus_info)) + pci_addr &= ~PCI64_ATTR_PREF; + + /* Handle Bridge Chipset differences */ + if (IS_PIC_SOFT(pcibus_info)) { + pci_addr |= + ((u64) pcibus_info-> + pbi_hub_xid << PIC_PCI64_ATTR_TARG_SHFT); + } else + pci_addr |= (dma_flags & SN_DMA_MSI) ? + TIOCP_PCI64_CMDTYPE_MSI : + TIOCP_PCI64_CMDTYPE_MEM; + + /* If PCI mode, func zero uses VCHAN0, every other func uses VCHAN1 */ + if (!IS_PCIX(pcibus_info) && PCI_FUNC(info->pdi_linux_pcidev->devfn)) + pci_addr |= PCI64_ATTR_VIRTUAL; + + return pci_addr; +} + +static dma_addr_t +pcibr_dmatrans_direct32(struct pcidev_info * info, + u64 paddr, size_t req_size, u64 flags, int dma_flags) +{ + struct pcidev_info *pcidev_info = info->pdi_host_pcidev_info; + struct pcibus_info *pcibus_info = (struct pcibus_info *)pcidev_info-> + pdi_pcibus_info; + u64 xio_addr; + + u64 xio_base; + u64 offset; + u64 endoff; + + if (IS_PCIX(pcibus_info)) { + return 0; + } + + if (dma_flags & SN_DMA_MSI) + return 0; + + if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS) + xio_addr = IS_PIC_SOFT(pcibus_info) ? PHYS_TO_DMA(paddr) : + PHYS_TO_TIODMA(paddr); + else + xio_addr = paddr; + + xio_base = pcibus_info->pbi_dir_xbase; + offset = xio_addr - xio_base; + endoff = req_size + offset; + if ((req_size > (1ULL << 31)) || /* Too Big */ + (xio_addr < xio_base) || /* Out of range for mappings */ + (endoff > (1ULL << 31))) { /* Too Big */ + return 0; + } + + return PCI32_DIRECT_BASE | offset; +} + +/* + * Wrapper routine for freeing DMA maps + * DMA mappings for Direct 64 and 32 do not have any DMA maps. + */ +void +pcibr_dma_unmap(struct pci_dev *hwdev, dma_addr_t dma_handle, int direction) +{ + struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev); + struct pcibus_info *pcibus_info = + (struct pcibus_info *)pcidev_info->pdi_pcibus_info; + + if (IS_PCI32_MAPPED(dma_handle)) { + int ate_index; + + ate_index = + IOPG((ATE_SWAP_OFF(dma_handle) - PCI32_MAPPED_BASE)); + pcibr_ate_free(pcibus_info, ate_index); + } +} + +/* + * On SN systems there is a race condition between a PIO read response and + * DMA's. In rare cases, the read response may beat the DMA, causing the + * driver to think that data in memory is complete and meaningful. This code + * eliminates that race. This routine is called by the PIO read routines + * after doing the read. For PIC this routine then forces a fake interrupt + * on another line, which is logically associated with the slot that the PIO + * is addressed to. It then spins while watching the memory location that + * the interrupt is targeted to. When the interrupt response arrives, we + * are sure that the DMA has landed in memory and it is safe for the driver + * to proceed. For TIOCP use the Device(x) Write Request Buffer Flush + * Bridge register since it ensures the data has entered the coherence domain, + * unlike the PIC Device(x) Write Request Buffer Flush register. + */ + +void sn_dma_flush(u64 addr) +{ + nasid_t nasid; + int is_tio; + int wid_num; + int i, j; + unsigned long flags; + u64 itte; + struct hubdev_info *hubinfo; + struct sn_flush_device_kernel *p; + struct sn_flush_device_common *common; + struct sn_flush_nasid_entry *flush_nasid_list; + + if (!sn_ioif_inited) + return; + + nasid = NASID_GET(addr); + if (-1 == nasid_to_cnodeid(nasid)) + return; + + hubinfo = (NODEPDA(nasid_to_cnodeid(nasid)))->pdinfo; + + BUG_ON(!hubinfo); + + flush_nasid_list = &hubinfo->hdi_flush_nasid_list; + if (flush_nasid_list->widget_p == NULL) + return; + + is_tio = (nasid & 1); + if (is_tio) { + int itte_index; + + if (TIO_HWIN(addr)) + itte_index = 0; + else if (TIO_BWIN_WINDOWNUM(addr)) + itte_index = TIO_BWIN_WINDOWNUM(addr); + else + itte_index = -1; + + if (itte_index >= 0) { + itte = flush_nasid_list->iio_itte[itte_index]; + if (! TIO_ITTE_VALID(itte)) + return; + wid_num = TIO_ITTE_WIDGET(itte); + } else + wid_num = TIO_SWIN_WIDGETNUM(addr); + } else { + if (BWIN_WINDOWNUM(addr)) { + itte = flush_nasid_list->iio_itte[BWIN_WINDOWNUM(addr)]; + wid_num = IIO_ITTE_WIDGET(itte); + } else + wid_num = SWIN_WIDGETNUM(addr); + } + if (flush_nasid_list->widget_p[wid_num] == NULL) + return; + p = &flush_nasid_list->widget_p[wid_num][0]; + + /* find a matching BAR */ + for (i = 0; i < DEV_PER_WIDGET; i++,p++) { + common = p->common; + for (j = 0; j < PCI_ROM_RESOURCE; j++) { + if (common->sfdl_bar_list[j].start == 0) + break; + if (addr >= common->sfdl_bar_list[j].start + && addr <= common->sfdl_bar_list[j].end) + break; + } + if (j < PCI_ROM_RESOURCE && common->sfdl_bar_list[j].start != 0) + break; + } + + /* if no matching BAR, return without doing anything. */ + if (i == DEV_PER_WIDGET) + return; + + /* + * For TIOCP use the Device(x) Write Request Buffer Flush Bridge + * register since it ensures the data has entered the coherence + * domain, unlike PIC. + */ + if (is_tio) { + /* + * Note: devices behind TIOCE should never be matched in the + * above code, and so the following code is PIC/CP centric. + * If CE ever needs the sn_dma_flush mechanism, we will have + * to account for that here and in tioce_bus_fixup(). + */ + u32 tio_id = HUB_L(TIO_IOSPACE_ADDR(nasid, TIO_NODE_ID)); + u32 revnum = XWIDGET_PART_REV_NUM(tio_id); + + /* TIOCP BRINGUP WAR (PV907516): Don't write buffer flush reg */ + if ((1 << XWIDGET_PART_REV_NUM_REV(revnum)) & PV907516) { + return; + } else { + pcireg_wrb_flush_get(common->sfdl_pcibus_info, + (common->sfdl_slot - 1)); + } + } else { + spin_lock_irqsave(&p->sfdl_flush_lock, flags); + *common->sfdl_flush_addr = 0; + + /* force an interrupt. */ + *(volatile u32 *)(common->sfdl_force_int_addr) = 1; + + /* wait for the interrupt to come back. */ + while (*(common->sfdl_flush_addr) != 0x10f) + cpu_relax(); + + /* okay, everything is synched up. */ + spin_unlock_irqrestore(&p->sfdl_flush_lock, flags); + } + return; +} + +/* + * DMA interfaces. Called from pci_dma.c routines. + */ + +dma_addr_t +pcibr_dma_map(struct pci_dev * hwdev, unsigned long phys_addr, size_t size, int dma_flags) +{ + dma_addr_t dma_handle; + struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev); + + /* SN cannot support DMA addresses smaller than 32 bits. */ + if (hwdev->dma_mask < 0x7fffffff) { + return 0; + } + + if (hwdev->dma_mask == ~0UL) { + /* + * Handle the most common case: 64 bit cards. This + * call should always succeed. + */ + + dma_handle = pcibr_dmatrans_direct64(pcidev_info, phys_addr, + PCI64_ATTR_PREF, dma_flags); + } else { + /* Handle 32-63 bit cards via direct mapping */ + dma_handle = pcibr_dmatrans_direct32(pcidev_info, phys_addr, + size, 0, dma_flags); + if (!dma_handle) { + /* + * It is a 32 bit card and we cannot do direct mapping, + * so we use an ATE. + */ + + dma_handle = pcibr_dmamap_ate32(pcidev_info, phys_addr, + size, PCI32_ATE_PREF, + dma_flags); + } + } + + return dma_handle; +} + +dma_addr_t +pcibr_dma_map_consistent(struct pci_dev * hwdev, unsigned long phys_addr, + size_t size, int dma_flags) +{ + dma_addr_t dma_handle; + struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev); + + if (hwdev->dev.coherent_dma_mask == ~0UL) { + dma_handle = pcibr_dmatrans_direct64(pcidev_info, phys_addr, + PCI64_ATTR_BAR, dma_flags); + } else { + dma_handle = (dma_addr_t) pcibr_dmamap_ate32(pcidev_info, + phys_addr, size, + PCI32_ATE_BAR, dma_flags); + } + + return dma_handle; +} + +EXPORT_SYMBOL(sn_dma_flush); diff --git a/arch/ia64/sn/pci/pcibr/pcibr_provider.c b/arch/ia64/sn/pci/pcibr/pcibr_provider.c new file mode 100644 index 000000000..8dbbef4a4 --- /dev/null +++ b/arch/ia64/sn/pci/pcibr/pcibr_provider.c @@ -0,0 +1,265 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * Copyright (C) 2001-2004, 2006 Silicon Graphics, Inc. All rights reserved. + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "xtalk/xwidgetdev.h" +#include "xtalk/hubdev.h" + +int +sal_pcibr_slot_enable(struct pcibus_info *soft, int device, void *resp, + char **ssdt) +{ + struct ia64_sal_retval ret_stuff; + u64 busnum; + u64 segment; + + ret_stuff.status = 0; + ret_stuff.v0 = 0; + + segment = soft->pbi_buscommon.bs_persist_segment; + busnum = soft->pbi_buscommon.bs_persist_busnum; + SAL_CALL_NOLOCK(ret_stuff, (u64) SN_SAL_IOIF_SLOT_ENABLE, segment, + busnum, (u64) device, (u64) resp, (u64)ia64_tpa(ssdt), + 0, 0); + + return (int)ret_stuff.v0; +} + +int +sal_pcibr_slot_disable(struct pcibus_info *soft, int device, int action, + void *resp) +{ + struct ia64_sal_retval ret_stuff; + u64 busnum; + u64 segment; + + ret_stuff.status = 0; + ret_stuff.v0 = 0; + + segment = soft->pbi_buscommon.bs_persist_segment; + busnum = soft->pbi_buscommon.bs_persist_busnum; + SAL_CALL_NOLOCK(ret_stuff, (u64) SN_SAL_IOIF_SLOT_DISABLE, + segment, busnum, (u64) device, (u64) action, + (u64) resp, 0, 0); + + return (int)ret_stuff.v0; +} + +static int sal_pcibr_error_interrupt(struct pcibus_info *soft) +{ + struct ia64_sal_retval ret_stuff; + u64 busnum; + int segment; + ret_stuff.status = 0; + ret_stuff.v0 = 0; + + segment = soft->pbi_buscommon.bs_persist_segment; + busnum = soft->pbi_buscommon.bs_persist_busnum; + SAL_CALL_NOLOCK(ret_stuff, + (u64) SN_SAL_IOIF_ERROR_INTERRUPT, + (u64) segment, (u64) busnum, 0, 0, 0, 0, 0); + + return (int)ret_stuff.v0; +} + +u16 sn_ioboard_to_pci_bus(struct pci_bus *pci_bus) +{ + long rc; + u16 uninitialized_var(ioboard); /* GCC be quiet */ + nasid_t nasid = NASID_GET(SN_PCIBUS_BUSSOFT(pci_bus)->bs_base); + + rc = ia64_sn_sysctl_ioboard_get(nasid, &ioboard); + if (rc) { + printk(KERN_WARNING "ia64_sn_sysctl_ioboard_get failed: %ld\n", + rc); + return 0; + } + + return ioboard; +} + +/* + * PCI Bridge Error interrupt handler. Gets invoked whenever a PCI + * bridge sends an error interrupt. + */ +static irqreturn_t +pcibr_error_intr_handler(int irq, void *arg) +{ + struct pcibus_info *soft = arg; + + if (sal_pcibr_error_interrupt(soft) < 0) + panic("pcibr_error_intr_handler(): Fatal Bridge Error"); + + return IRQ_HANDLED; +} + +void * +pcibr_bus_fixup(struct pcibus_bussoft *prom_bussoft, struct pci_controller *controller) +{ + int nasid, cnode, j; + struct hubdev_info *hubdev_info; + struct pcibus_info *soft; + struct sn_flush_device_kernel *sn_flush_device_kernel; + struct sn_flush_device_common *common; + + if (! IS_PCI_BRIDGE_ASIC(prom_bussoft->bs_asic_type)) { + return NULL; + } + + /* + * Allocate kernel bus soft and copy from prom. + */ + + soft = kmemdup(prom_bussoft, sizeof(struct pcibus_info), GFP_KERNEL); + if (!soft) { + return NULL; + } + + soft->pbi_buscommon.bs_base = (unsigned long) + ioremap(REGION_OFFSET(soft->pbi_buscommon.bs_base), + sizeof(struct pic)); + + spin_lock_init(&soft->pbi_lock); + + /* + * register the bridge's error interrupt handler + */ + if (request_irq(SGI_PCIASIC_ERROR, pcibr_error_intr_handler, + IRQF_SHARED, "PCIBR error", (void *)(soft))) { + printk(KERN_WARNING + "pcibr cannot allocate interrupt for error handler\n"); + } + irq_set_handler(SGI_PCIASIC_ERROR, handle_level_irq); + sn_set_err_irq_affinity(SGI_PCIASIC_ERROR); + + /* + * Update the Bridge with the "kernel" pagesize + */ + if (PAGE_SIZE < 16384) { + pcireg_control_bit_clr(soft, PCIBR_CTRL_PAGE_SIZE); + } else { + pcireg_control_bit_set(soft, PCIBR_CTRL_PAGE_SIZE); + } + + nasid = NASID_GET(soft->pbi_buscommon.bs_base); + cnode = nasid_to_cnodeid(nasid); + hubdev_info = (struct hubdev_info *)(NODEPDA(cnode)->pdinfo); + + if (hubdev_info->hdi_flush_nasid_list.widget_p) { + sn_flush_device_kernel = hubdev_info->hdi_flush_nasid_list. + widget_p[(int)soft->pbi_buscommon.bs_xid]; + if (sn_flush_device_kernel) { + for (j = 0; j < DEV_PER_WIDGET; + j++, sn_flush_device_kernel++) { + common = sn_flush_device_kernel->common; + if (common->sfdl_slot == -1) + continue; + if ((common->sfdl_persistent_segment == + soft->pbi_buscommon.bs_persist_segment) && + (common->sfdl_persistent_busnum == + soft->pbi_buscommon.bs_persist_busnum)) + common->sfdl_pcibus_info = + soft; + } + } + } + + /* Setup the PMU ATE map */ + soft->pbi_int_ate_resource.lowest_free_index = 0; + soft->pbi_int_ate_resource.ate = + kzalloc(soft->pbi_int_ate_size * sizeof(u64), GFP_KERNEL); + + if (!soft->pbi_int_ate_resource.ate) { + kfree(soft); + return NULL; + } + + return soft; +} + +void pcibr_force_interrupt(struct sn_irq_info *sn_irq_info) +{ + struct pcidev_info *pcidev_info; + struct pcibus_info *pcibus_info; + int bit = sn_irq_info->irq_int_bit; + + if (! sn_irq_info->irq_bridge) + return; + + pcidev_info = (struct pcidev_info *)sn_irq_info->irq_pciioinfo; + if (pcidev_info) { + pcibus_info = + (struct pcibus_info *)pcidev_info->pdi_host_pcidev_info-> + pdi_pcibus_info; + pcireg_force_intr_set(pcibus_info, bit); + } +} + +void pcibr_target_interrupt(struct sn_irq_info *sn_irq_info) +{ + struct pcidev_info *pcidev_info; + struct pcibus_info *pcibus_info; + int bit = sn_irq_info->irq_int_bit; + u64 xtalk_addr = sn_irq_info->irq_xtalkaddr; + + pcidev_info = (struct pcidev_info *)sn_irq_info->irq_pciioinfo; + if (pcidev_info) { + pcibus_info = + (struct pcibus_info *)pcidev_info->pdi_host_pcidev_info-> + pdi_pcibus_info; + + /* Disable the device's IRQ */ + pcireg_intr_enable_bit_clr(pcibus_info, (1 << bit)); + + /* Change the device's IRQ */ + pcireg_intr_addr_addr_set(pcibus_info, bit, xtalk_addr); + + /* Re-enable the device's IRQ */ + pcireg_intr_enable_bit_set(pcibus_info, (1 << bit)); + + pcibr_force_interrupt(sn_irq_info); + } +} + +/* + * Provider entries for PIC/CP + */ + +struct sn_pcibus_provider pcibr_provider = { + .dma_map = pcibr_dma_map, + .dma_map_consistent = pcibr_dma_map_consistent, + .dma_unmap = pcibr_dma_unmap, + .bus_fixup = pcibr_bus_fixup, + .force_interrupt = pcibr_force_interrupt, + .target_interrupt = pcibr_target_interrupt +}; + +int +pcibr_init_provider(void) +{ + sn_pci_provider[PCIIO_ASIC_TYPE_PIC] = &pcibr_provider; + sn_pci_provider[PCIIO_ASIC_TYPE_TIOCP] = &pcibr_provider; + + return 0; +} + +EXPORT_SYMBOL_GPL(sal_pcibr_slot_enable); +EXPORT_SYMBOL_GPL(sal_pcibr_slot_disable); +EXPORT_SYMBOL_GPL(sn_ioboard_to_pci_bus); diff --git a/arch/ia64/sn/pci/pcibr/pcibr_reg.c b/arch/ia64/sn/pci/pcibr/pcibr_reg.c new file mode 100644 index 000000000..8b8bbd51d --- /dev/null +++ b/arch/ia64/sn/pci/pcibr/pcibr_reg.c @@ -0,0 +1,285 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * Copyright (C) 2004 Silicon Graphics, Inc. All rights reserved. + */ + +#include +#include +#include +#include +#include +#include +#include +#include + +union br_ptr { + struct tiocp tio; + struct pic pic; +}; + +/* + * Control Register Access -- Read/Write 0000_0020 + */ +void pcireg_control_bit_clr(struct pcibus_info *pcibus_info, u64 bits) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + __sn_clrq_relaxed(&ptr->tio.cp_control, bits); + break; + case PCIBR_BRIDGETYPE_PIC: + __sn_clrq_relaxed(&ptr->pic.p_wid_control, bits); + break; + default: + panic + ("pcireg_control_bit_clr: unknown bridgetype bridge 0x%p", + ptr); + } + } +} + +void pcireg_control_bit_set(struct pcibus_info *pcibus_info, u64 bits) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + __sn_setq_relaxed(&ptr->tio.cp_control, bits); + break; + case PCIBR_BRIDGETYPE_PIC: + __sn_setq_relaxed(&ptr->pic.p_wid_control, bits); + break; + default: + panic + ("pcireg_control_bit_set: unknown bridgetype bridge 0x%p", + ptr); + } + } +} + +/* + * PCI/PCIX Target Flush Register Access -- Read Only 0000_0050 + */ +u64 pcireg_tflush_get(struct pcibus_info *pcibus_info) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + u64 ret = 0; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + ret = __sn_readq_relaxed(&ptr->tio.cp_tflush); + break; + case PCIBR_BRIDGETYPE_PIC: + ret = __sn_readq_relaxed(&ptr->pic.p_wid_tflush); + break; + default: + panic + ("pcireg_tflush_get: unknown bridgetype bridge 0x%p", + ptr); + } + } + + /* Read of the Target Flush should always return zero */ + if (ret != 0) + panic("pcireg_tflush_get:Target Flush failed\n"); + + return ret; +} + +/* + * Interrupt Status Register Access -- Read Only 0000_0100 + */ +u64 pcireg_intr_status_get(struct pcibus_info * pcibus_info) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + u64 ret = 0; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + ret = __sn_readq_relaxed(&ptr->tio.cp_int_status); + break; + case PCIBR_BRIDGETYPE_PIC: + ret = __sn_readq_relaxed(&ptr->pic.p_int_status); + break; + default: + panic + ("pcireg_intr_status_get: unknown bridgetype bridge 0x%p", + ptr); + } + } + return ret; +} + +/* + * Interrupt Enable Register Access -- Read/Write 0000_0108 + */ +void pcireg_intr_enable_bit_clr(struct pcibus_info *pcibus_info, u64 bits) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + __sn_clrq_relaxed(&ptr->tio.cp_int_enable, bits); + break; + case PCIBR_BRIDGETYPE_PIC: + __sn_clrq_relaxed(&ptr->pic.p_int_enable, bits); + break; + default: + panic + ("pcireg_intr_enable_bit_clr: unknown bridgetype bridge 0x%p", + ptr); + } + } +} + +void pcireg_intr_enable_bit_set(struct pcibus_info *pcibus_info, u64 bits) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + __sn_setq_relaxed(&ptr->tio.cp_int_enable, bits); + break; + case PCIBR_BRIDGETYPE_PIC: + __sn_setq_relaxed(&ptr->pic.p_int_enable, bits); + break; + default: + panic + ("pcireg_intr_enable_bit_set: unknown bridgetype bridge 0x%p", + ptr); + } + } +} + +/* + * Intr Host Address Register (int_addr) -- Read/Write 0000_0130 - 0000_0168 + */ +void pcireg_intr_addr_addr_set(struct pcibus_info *pcibus_info, int int_n, + u64 addr) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + __sn_clrq_relaxed(&ptr->tio.cp_int_addr[int_n], + TIOCP_HOST_INTR_ADDR); + __sn_setq_relaxed(&ptr->tio.cp_int_addr[int_n], + (addr & TIOCP_HOST_INTR_ADDR)); + break; + case PCIBR_BRIDGETYPE_PIC: + __sn_clrq_relaxed(&ptr->pic.p_int_addr[int_n], + PIC_HOST_INTR_ADDR); + __sn_setq_relaxed(&ptr->pic.p_int_addr[int_n], + (addr & PIC_HOST_INTR_ADDR)); + break; + default: + panic + ("pcireg_intr_addr_addr_get: unknown bridgetype bridge 0x%p", + ptr); + } + } +} + +/* + * Force Interrupt Register Access -- Write Only 0000_01C0 - 0000_01F8 + */ +void pcireg_force_intr_set(struct pcibus_info *pcibus_info, int int_n) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + writeq(1, &ptr->tio.cp_force_pin[int_n]); + break; + case PCIBR_BRIDGETYPE_PIC: + writeq(1, &ptr->pic.p_force_pin[int_n]); + break; + default: + panic + ("pcireg_force_intr_set: unknown bridgetype bridge 0x%p", + ptr); + } + } +} + +/* + * Device(x) Write Buffer Flush Reg Access -- Read Only 0000_0240 - 0000_0258 + */ +u64 pcireg_wrb_flush_get(struct pcibus_info *pcibus_info, int device) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + u64 ret = 0; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + ret = + __sn_readq_relaxed(&ptr->tio.cp_wr_req_buf[device]); + break; + case PCIBR_BRIDGETYPE_PIC: + ret = + __sn_readq_relaxed(&ptr->pic.p_wr_req_buf[device]); + break; + default: + panic("pcireg_wrb_flush_get: unknown bridgetype bridge 0x%p", ptr); + } + + } + /* Read of the Write Buffer Flush should always return zero */ + return ret; +} + +void pcireg_int_ate_set(struct pcibus_info *pcibus_info, int ate_index, + u64 val) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + writeq(val, &ptr->tio.cp_int_ate_ram[ate_index]); + break; + case PCIBR_BRIDGETYPE_PIC: + writeq(val, &ptr->pic.p_int_ate_ram[ate_index]); + break; + default: + panic + ("pcireg_int_ate_set: unknown bridgetype bridge 0x%p", + ptr); + } + } +} + +u64 __iomem *pcireg_int_ate_addr(struct pcibus_info *pcibus_info, int ate_index) +{ + union br_ptr __iomem *ptr = (union br_ptr __iomem *)pcibus_info->pbi_buscommon.bs_base; + u64 __iomem *ret = NULL; + + if (pcibus_info) { + switch (pcibus_info->pbi_bridge_type) { + case PCIBR_BRIDGETYPE_TIOCP: + ret = &ptr->tio.cp_int_ate_ram[ate_index]; + break; + case PCIBR_BRIDGETYPE_PIC: + ret = &ptr->pic.p_int_ate_ram[ate_index]; + break; + default: + panic + ("pcireg_int_ate_addr: unknown bridgetype bridge 0x%p", + ptr); + } + } + return ret; +} diff --git a/arch/ia64/sn/pci/tioca_provider.c b/arch/ia64/sn/pci/tioca_provider.c new file mode 100644 index 000000000..a70b11fd5 --- /dev/null +++ b/arch/ia64/sn/pci/tioca_provider.c @@ -0,0 +1,677 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * Copyright (C) 2003-2005 Silicon Graphics, Inc. All Rights Reserved. + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +u32 tioca_gart_found; +EXPORT_SYMBOL(tioca_gart_found); /* used by agp-sgi */ + +LIST_HEAD(tioca_list); +EXPORT_SYMBOL(tioca_list); /* used by agp-sgi */ + +static int tioca_gart_init(struct tioca_kernel *); + +/** + * tioca_gart_init - Initialize SGI TIOCA GART + * @tioca_common: ptr to common prom/kernel struct identifying the + * + * If the indicated tioca has devices present, initialize its associated + * GART MMR's and kernel memory. + */ +static int +tioca_gart_init(struct tioca_kernel *tioca_kern) +{ + u64 ap_reg; + u64 offset; + struct page *tmp; + struct tioca_common *tioca_common; + struct tioca __iomem *ca_base; + + tioca_common = tioca_kern->ca_common; + ca_base = (struct tioca __iomem *)tioca_common->ca_common.bs_base; + + if (list_empty(tioca_kern->ca_devices)) + return 0; + + ap_reg = 0; + + /* + * Validate aperature size + */ + + switch (CA_APERATURE_SIZE >> 20) { + case 4: + ap_reg |= (0x3ff << CA_GART_AP_SIZE_SHFT); /* 4MB */ + break; + case 8: + ap_reg |= (0x3fe << CA_GART_AP_SIZE_SHFT); /* 8MB */ + break; + case 16: + ap_reg |= (0x3fc << CA_GART_AP_SIZE_SHFT); /* 16MB */ + break; + case 32: + ap_reg |= (0x3f8 << CA_GART_AP_SIZE_SHFT); /* 32 MB */ + break; + case 64: + ap_reg |= (0x3f0 << CA_GART_AP_SIZE_SHFT); /* 64 MB */ + break; + case 128: + ap_reg |= (0x3e0 << CA_GART_AP_SIZE_SHFT); /* 128 MB */ + break; + case 256: + ap_reg |= (0x3c0 << CA_GART_AP_SIZE_SHFT); /* 256 MB */ + break; + case 512: + ap_reg |= (0x380 << CA_GART_AP_SIZE_SHFT); /* 512 MB */ + break; + case 1024: + ap_reg |= (0x300 << CA_GART_AP_SIZE_SHFT); /* 1GB */ + break; + case 2048: + ap_reg |= (0x200 << CA_GART_AP_SIZE_SHFT); /* 2GB */ + break; + case 4096: + ap_reg |= (0x000 << CA_GART_AP_SIZE_SHFT); /* 4 GB */ + break; + default: + printk(KERN_ERR "%s: Invalid CA_APERATURE_SIZE " + "0x%lx\n", __func__, (ulong) CA_APERATURE_SIZE); + return -1; + } + + /* + * Set up other aperature parameters + */ + + if (PAGE_SIZE >= 16384) { + tioca_kern->ca_ap_pagesize = 16384; + ap_reg |= CA_GART_PAGE_SIZE; + } else { + tioca_kern->ca_ap_pagesize = 4096; + } + + tioca_kern->ca_ap_size = CA_APERATURE_SIZE; + tioca_kern->ca_ap_bus_base = CA_APERATURE_BASE; + tioca_kern->ca_gart_entries = + tioca_kern->ca_ap_size / tioca_kern->ca_ap_pagesize; + + ap_reg |= (CA_GART_AP_ENB_AGP | CA_GART_AP_ENB_PCI); + ap_reg |= tioca_kern->ca_ap_bus_base; + + /* + * Allocate and set up the GART + */ + + tioca_kern->ca_gart_size = tioca_kern->ca_gart_entries * sizeof(u64); + tmp = + alloc_pages_node(tioca_kern->ca_closest_node, + GFP_KERNEL | __GFP_ZERO, + get_order(tioca_kern->ca_gart_size)); + + if (!tmp) { + printk(KERN_ERR "%s: Could not allocate " + "%llu bytes (order %d) for GART\n", + __func__, + tioca_kern->ca_gart_size, + get_order(tioca_kern->ca_gart_size)); + return -ENOMEM; + } + + tioca_kern->ca_gart = page_address(tmp); + tioca_kern->ca_gart_coretalk_addr = + PHYS_TO_TIODMA(virt_to_phys(tioca_kern->ca_gart)); + + /* + * Compute PCI/AGP convenience fields + */ + + offset = CA_PCI32_MAPPED_BASE - CA_APERATURE_BASE; + tioca_kern->ca_pciap_base = CA_PCI32_MAPPED_BASE; + tioca_kern->ca_pciap_size = CA_PCI32_MAPPED_SIZE; + tioca_kern->ca_pcigart_start = offset / tioca_kern->ca_ap_pagesize; + tioca_kern->ca_pcigart_base = + tioca_kern->ca_gart_coretalk_addr + offset; + tioca_kern->ca_pcigart = + &tioca_kern->ca_gart[tioca_kern->ca_pcigart_start]; + tioca_kern->ca_pcigart_entries = + tioca_kern->ca_pciap_size / tioca_kern->ca_ap_pagesize; + tioca_kern->ca_pcigart_pagemap = + kzalloc(tioca_kern->ca_pcigart_entries / 8, GFP_KERNEL); + if (!tioca_kern->ca_pcigart_pagemap) { + free_pages((unsigned long)tioca_kern->ca_gart, + get_order(tioca_kern->ca_gart_size)); + return -1; + } + + offset = CA_AGP_MAPPED_BASE - CA_APERATURE_BASE; + tioca_kern->ca_gfxap_base = CA_AGP_MAPPED_BASE; + tioca_kern->ca_gfxap_size = CA_AGP_MAPPED_SIZE; + tioca_kern->ca_gfxgart_start = offset / tioca_kern->ca_ap_pagesize; + tioca_kern->ca_gfxgart_base = + tioca_kern->ca_gart_coretalk_addr + offset; + tioca_kern->ca_gfxgart = + &tioca_kern->ca_gart[tioca_kern->ca_gfxgart_start]; + tioca_kern->ca_gfxgart_entries = + tioca_kern->ca_gfxap_size / tioca_kern->ca_ap_pagesize; + + /* + * various control settings: + * use agp op-combining + * use GET semantics to fetch memory + * participate in coherency domain + * DISABLE GART PREFETCHING due to hw bug tracked in SGI PV930029 + */ + + __sn_setq_relaxed(&ca_base->ca_control1, + CA_AGPDMA_OP_ENB_COMBDELAY); /* PV895469 ? */ + __sn_clrq_relaxed(&ca_base->ca_control2, CA_GART_MEM_PARAM); + __sn_setq_relaxed(&ca_base->ca_control2, + (0x2ull << CA_GART_MEM_PARAM_SHFT)); + tioca_kern->ca_gart_iscoherent = 1; + __sn_clrq_relaxed(&ca_base->ca_control2, + (CA_GART_WR_PREFETCH_ENB | CA_GART_RD_PREFETCH_ENB)); + + /* + * Unmask GART fetch error interrupts. Clear residual errors first. + */ + + writeq(CA_GART_FETCH_ERR, &ca_base->ca_int_status_alias); + writeq(CA_GART_FETCH_ERR, &ca_base->ca_mult_error_alias); + __sn_clrq_relaxed(&ca_base->ca_int_mask, CA_GART_FETCH_ERR); + + /* + * Program the aperature and gart registers in TIOCA + */ + + writeq(ap_reg, &ca_base->ca_gart_aperature); + writeq(tioca_kern->ca_gart_coretalk_addr|1, &ca_base->ca_gart_ptr_table); + + return 0; +} + +/** + * tioca_fastwrite_enable - enable AGP FW for a tioca and its functions + * @tioca_kernel: structure representing the CA + * + * Given a CA, scan all attached functions making sure they all support + * FastWrite. If so, enable FastWrite for all functions and the CA itself. + */ + +void +tioca_fastwrite_enable(struct tioca_kernel *tioca_kern) +{ + int cap_ptr; + u32 reg; + struct tioca __iomem *tioca_base; + struct pci_dev *pdev; + struct tioca_common *common; + + common = tioca_kern->ca_common; + + /* + * Scan all vga controllers on this bus making sure they all + * support FW. If not, return. + */ + + list_for_each_entry(pdev, tioca_kern->ca_devices, bus_list) { + if (pdev->class != (PCI_CLASS_DISPLAY_VGA << 8)) + continue; + + cap_ptr = pci_find_capability(pdev, PCI_CAP_ID_AGP); + if (!cap_ptr) + return; /* no AGP CAP means no FW */ + + pci_read_config_dword(pdev, cap_ptr + PCI_AGP_STATUS, ®); + if (!(reg & PCI_AGP_STATUS_FW)) + return; /* function doesn't support FW */ + } + + /* + * Set fw for all vga fn's + */ + + list_for_each_entry(pdev, tioca_kern->ca_devices, bus_list) { + if (pdev->class != (PCI_CLASS_DISPLAY_VGA << 8)) + continue; + + cap_ptr = pci_find_capability(pdev, PCI_CAP_ID_AGP); + pci_read_config_dword(pdev, cap_ptr + PCI_AGP_COMMAND, ®); + reg |= PCI_AGP_COMMAND_FW; + pci_write_config_dword(pdev, cap_ptr + PCI_AGP_COMMAND, reg); + } + + /* + * Set ca's fw to match + */ + + tioca_base = (struct tioca __iomem*)common->ca_common.bs_base; + __sn_setq_relaxed(&tioca_base->ca_control1, CA_AGP_FW_ENABLE); +} + +EXPORT_SYMBOL(tioca_fastwrite_enable); /* used by agp-sgi */ + +/** + * tioca_dma_d64 - create a DMA mapping using 64-bit direct mode + * @paddr: system physical address + * + * Map @paddr into 64-bit CA bus space. No device context is necessary. + * Bits 53:0 come from the coretalk address. We just need to mask in the + * following optional bits of the 64-bit pci address: + * + * 63:60 - Coretalk Packet Type - 0x1 for Mem Get/Put (coherent) + * 0x2 for PIO (non-coherent) + * We will always use 0x1 + * 55:55 - Swap bytes Currently unused + */ +static u64 +tioca_dma_d64(unsigned long paddr) +{ + dma_addr_t bus_addr; + + bus_addr = PHYS_TO_TIODMA(paddr); + + BUG_ON(!bus_addr); + BUG_ON(bus_addr >> 54); + + /* Set upper nibble to Cache Coherent Memory op */ + bus_addr |= (1UL << 60); + + return bus_addr; +} + +/** + * tioca_dma_d48 - create a DMA mapping using 48-bit direct mode + * @pdev: linux pci_dev representing the function + * @paddr: system physical address + * + * Map @paddr into 64-bit bus space of the CA associated with @pcidev_info. + * + * The CA agp 48 bit direct address falls out as follows: + * + * When direct mapping AGP addresses, the 48 bit AGP address is + * constructed as follows: + * + * [47:40] - Low 8 bits of the page Node ID extracted from coretalk + * address [47:40]. The upper 8 node bits are fixed + * and come from the xxx register bits [5:0] + * [39:38] - Chiplet ID extracted from coretalk address [39:38] + * [37:00] - node offset extracted from coretalk address [37:00] + * + * Since the node id in general will be non-zero, and the chiplet id + * will always be non-zero, it follows that the device must support + * a dma mask of at least 0xffffffffff (40 bits) to target node 0 + * and in general should be 0xffffffffffff (48 bits) to target nodes + * up to 255. Nodes above 255 need the support of the xxx register, + * and so a given CA can only directly target nodes in the range + * xxx - xxx+255. + */ +static u64 +tioca_dma_d48(struct pci_dev *pdev, u64 paddr) +{ + struct tioca_common *tioca_common; + struct tioca __iomem *ca_base; + u64 ct_addr; + dma_addr_t bus_addr; + u32 node_upper; + u64 agp_dma_extn; + struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(pdev); + + tioca_common = (struct tioca_common *)pcidev_info->pdi_pcibus_info; + ca_base = (struct tioca __iomem *)tioca_common->ca_common.bs_base; + + ct_addr = PHYS_TO_TIODMA(paddr); + if (!ct_addr) + return 0; + + bus_addr = (dma_addr_t) (ct_addr & 0xffffffffffffUL); + node_upper = ct_addr >> 48; + + if (node_upper > 64) { + printk(KERN_ERR "%s: coretalk addr 0x%p node id out " + "of range\n", __func__, (void *)ct_addr); + return 0; + } + + agp_dma_extn = __sn_readq_relaxed(&ca_base->ca_agp_dma_addr_extn); + if (node_upper != (agp_dma_extn >> CA_AGP_DMA_NODE_ID_SHFT)) { + printk(KERN_ERR "%s: coretalk upper node (%u) " + "mismatch with ca_agp_dma_addr_extn (%llu)\n", + __func__, + node_upper, (agp_dma_extn >> CA_AGP_DMA_NODE_ID_SHFT)); + return 0; + } + + return bus_addr; +} + +/** + * tioca_dma_mapped - create a DMA mapping using a CA GART + * @pdev: linux pci_dev representing the function + * @paddr: host physical address to map + * @req_size: len (bytes) to map + * + * Map @paddr into CA address space using the GART mechanism. The mapped + * dma_addr_t is guaranteed to be contiguous in CA bus space. + */ +static dma_addr_t +tioca_dma_mapped(struct pci_dev *pdev, unsigned long paddr, size_t req_size) +{ + int ps, ps_shift, entry, entries, mapsize; + u64 xio_addr, end_xio_addr; + struct tioca_common *tioca_common; + struct tioca_kernel *tioca_kern; + dma_addr_t bus_addr = 0; + struct tioca_dmamap *ca_dmamap; + void *map; + unsigned long flags; + struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(pdev); + + tioca_common = (struct tioca_common *)pcidev_info->pdi_pcibus_info; + tioca_kern = (struct tioca_kernel *)tioca_common->ca_kernel_private; + + xio_addr = PHYS_TO_TIODMA(paddr); + if (!xio_addr) + return 0; + + spin_lock_irqsave(&tioca_kern->ca_lock, flags); + + /* + * allocate a map struct + */ + + ca_dmamap = kzalloc(sizeof(struct tioca_dmamap), GFP_ATOMIC); + if (!ca_dmamap) + goto map_return; + + /* + * Locate free entries that can hold req_size. Account for + * unaligned start/length when allocating. + */ + + ps = tioca_kern->ca_ap_pagesize; /* will be power of 2 */ + ps_shift = ffs(ps) - 1; + end_xio_addr = xio_addr + req_size - 1; + + entries = (end_xio_addr >> ps_shift) - (xio_addr >> ps_shift) + 1; + + map = tioca_kern->ca_pcigart_pagemap; + mapsize = tioca_kern->ca_pcigart_entries; + + entry = bitmap_find_next_zero_area(map, mapsize, 0, entries, 0); + if (entry >= mapsize) { + kfree(ca_dmamap); + goto map_return; + } + + bitmap_set(map, entry, entries); + + bus_addr = tioca_kern->ca_pciap_base + (entry * ps); + + ca_dmamap->cad_dma_addr = bus_addr; + ca_dmamap->cad_gart_size = entries; + ca_dmamap->cad_gart_entry = entry; + list_add(&ca_dmamap->cad_list, &tioca_kern->ca_dmamaps); + + if (xio_addr % ps) { + tioca_kern->ca_pcigart[entry] = tioca_paddr_to_gart(xio_addr); + bus_addr += xio_addr & (ps - 1); + xio_addr &= ~(ps - 1); + xio_addr += ps; + entry++; + } + + while (xio_addr < end_xio_addr) { + tioca_kern->ca_pcigart[entry] = tioca_paddr_to_gart(xio_addr); + xio_addr += ps; + entry++; + } + + tioca_tlbflush(tioca_kern); + +map_return: + spin_unlock_irqrestore(&tioca_kern->ca_lock, flags); + return bus_addr; +} + +/** + * tioca_dma_unmap - release CA mapping resources + * @pdev: linux pci_dev representing the function + * @bus_addr: bus address returned by an earlier tioca_dma_map + * @dir: mapping direction (unused) + * + * Locate mapping resources associated with @bus_addr and release them. + * For mappings created using the direct modes (64 or 48) there are no + * resources to release. + */ +static void +tioca_dma_unmap(struct pci_dev *pdev, dma_addr_t bus_addr, int dir) +{ + int i, entry; + struct tioca_common *tioca_common; + struct tioca_kernel *tioca_kern; + struct tioca_dmamap *map; + struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(pdev); + unsigned long flags; + + tioca_common = (struct tioca_common *)pcidev_info->pdi_pcibus_info; + tioca_kern = (struct tioca_kernel *)tioca_common->ca_kernel_private; + + /* return straight away if this isn't be a mapped address */ + + if (bus_addr < tioca_kern->ca_pciap_base || + bus_addr >= (tioca_kern->ca_pciap_base + tioca_kern->ca_pciap_size)) + return; + + spin_lock_irqsave(&tioca_kern->ca_lock, flags); + + list_for_each_entry(map, &tioca_kern->ca_dmamaps, cad_list) + if (map->cad_dma_addr == bus_addr) + break; + + BUG_ON(map == NULL); + + entry = map->cad_gart_entry; + + for (i = 0; i < map->cad_gart_size; i++, entry++) { + clear_bit(entry, tioca_kern->ca_pcigart_pagemap); + tioca_kern->ca_pcigart[entry] = 0; + } + tioca_tlbflush(tioca_kern); + + list_del(&map->cad_list); + spin_unlock_irqrestore(&tioca_kern->ca_lock, flags); + kfree(map); +} + +/** + * tioca_dma_map - map pages for PCI DMA + * @pdev: linux pci_dev representing the function + * @paddr: host physical address to map + * @byte_count: bytes to map + * + * This is the main wrapper for mapping host physical pages to CA PCI space. + * The mapping mode used is based on the devices dma_mask. As a last resort + * use the GART mapped mode. + */ +static u64 +tioca_dma_map(struct pci_dev *pdev, unsigned long paddr, size_t byte_count, int dma_flags) +{ + u64 mapaddr; + + /* + * Not supported for now ... + */ + if (dma_flags & SN_DMA_MSI) + return 0; + + /* + * If card is 64 or 48 bit addressable, use a direct mapping. 32 + * bit direct is so restrictive w.r.t. where the memory resides that + * we don't use it even though CA has some support. + */ + + if (pdev->dma_mask == ~0UL) + mapaddr = tioca_dma_d64(paddr); + else if (pdev->dma_mask == 0xffffffffffffUL) + mapaddr = tioca_dma_d48(pdev, paddr); + else + mapaddr = 0; + + /* Last resort ... use PCI portion of CA GART */ + + if (mapaddr == 0) + mapaddr = tioca_dma_mapped(pdev, paddr, byte_count); + + return mapaddr; +} + +/** + * tioca_error_intr_handler - SGI TIO CA error interrupt handler + * @irq: unused + * @arg: pointer to tioca_common struct for the given CA + * + * Handle a CA error interrupt. Simply a wrapper around a SAL call which + * defers processing to the SGI prom. + */ +static irqreturn_t +tioca_error_intr_handler(int irq, void *arg) +{ + struct tioca_common *soft = arg; + struct ia64_sal_retval ret_stuff; + u64 segment; + u64 busnum; + ret_stuff.status = 0; + ret_stuff.v0 = 0; + + segment = soft->ca_common.bs_persist_segment; + busnum = soft->ca_common.bs_persist_busnum; + + SAL_CALL_NOLOCK(ret_stuff, + (u64) SN_SAL_IOIF_ERROR_INTERRUPT, + segment, busnum, 0, 0, 0, 0, 0); + + return IRQ_HANDLED; +} + +/** + * tioca_bus_fixup - perform final PCI fixup for a TIO CA bus + * @prom_bussoft: Common prom/kernel struct representing the bus + * + * Replicates the tioca_common pointed to by @prom_bussoft in kernel + * space. Allocates and initializes a kernel-only area for a given CA, + * and sets up an irq for handling CA error interrupts. + * + * On successful setup, returns the kernel version of tioca_common back to + * the caller. + */ +static void * +tioca_bus_fixup(struct pcibus_bussoft *prom_bussoft, struct pci_controller *controller) +{ + struct tioca_common *tioca_common; + struct tioca_kernel *tioca_kern; + struct pci_bus *bus; + + /* sanity check prom rev */ + + if (is_shub1() && sn_sal_rev() < 0x0406) { + printk + (KERN_ERR "%s: SGI prom rev 4.06 or greater required " + "for tioca support\n", __func__); + return NULL; + } + + /* + * Allocate kernel bus soft and copy from prom. + */ + + tioca_common = kmemdup(prom_bussoft, sizeof(struct tioca_common), + GFP_KERNEL); + if (!tioca_common) + return NULL; + + tioca_common->ca_common.bs_base = (unsigned long) + ioremap(REGION_OFFSET(tioca_common->ca_common.bs_base), + sizeof(struct tioca_common)); + + /* init kernel-private area */ + + tioca_kern = kzalloc(sizeof(struct tioca_kernel), GFP_KERNEL); + if (!tioca_kern) { + kfree(tioca_common); + return NULL; + } + + tioca_kern->ca_common = tioca_common; + spin_lock_init(&tioca_kern->ca_lock); + INIT_LIST_HEAD(&tioca_kern->ca_dmamaps); + tioca_kern->ca_closest_node = + nasid_to_cnodeid(tioca_common->ca_closest_nasid); + tioca_common->ca_kernel_private = (u64) tioca_kern; + + bus = pci_find_bus(tioca_common->ca_common.bs_persist_segment, + tioca_common->ca_common.bs_persist_busnum); + BUG_ON(!bus); + tioca_kern->ca_devices = &bus->devices; + + /* init GART */ + + if (tioca_gart_init(tioca_kern) < 0) { + kfree(tioca_kern); + kfree(tioca_common); + return NULL; + } + + tioca_gart_found++; + list_add(&tioca_kern->ca_list, &tioca_list); + + if (request_irq(SGI_TIOCA_ERROR, + tioca_error_intr_handler, + IRQF_SHARED, "TIOCA error", (void *)tioca_common)) + printk(KERN_WARNING + "%s: Unable to get irq %d. " + "Error interrupts won't be routed for TIOCA bus %d\n", + __func__, SGI_TIOCA_ERROR, + (int)tioca_common->ca_common.bs_persist_busnum); + + irq_set_handler(SGI_TIOCA_ERROR, handle_level_irq); + sn_set_err_irq_affinity(SGI_TIOCA_ERROR); + + /* Setup locality information */ + controller->node = tioca_kern->ca_closest_node; + return tioca_common; +} + +static struct sn_pcibus_provider tioca_pci_interfaces = { + .dma_map = tioca_dma_map, + .dma_map_consistent = tioca_dma_map, + .dma_unmap = tioca_dma_unmap, + .bus_fixup = tioca_bus_fixup, + .force_interrupt = NULL, + .target_interrupt = NULL +}; + +/** + * tioca_init_provider - init SN PCI provider ops for TIO CA + */ +int +tioca_init_provider(void) +{ + sn_pci_provider[PCIIO_ASIC_TYPE_TIOCA] = &tioca_pci_interfaces; + return 0; +} diff --git a/arch/ia64/sn/pci/tioce_provider.c b/arch/ia64/sn/pci/tioce_provider.c new file mode 100644 index 000000000..46d3df4b0 --- /dev/null +++ b/arch/ia64/sn/pci/tioce_provider.c @@ -0,0 +1,1062 @@ +/* + * This file is subject to the terms and conditions of the GNU General Public + * License. See the file "COPYING" in the main directory of this archive + * for more details. + * + * Copyright (C) 2003-2006 Silicon Graphics, Inc. All Rights Reserved. + */ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +/* + * 1/26/2006 + * + * WAR for SGI PV 944642. For revA TIOCE, need to use the following recipe + * (taken from the above PV) before and after accessing tioce internal MMR's + * to avoid tioce lockups. + * + * The recipe as taken from the PV: + * + * if(mmr address < 0x45000) { + * if(mmr address == 0 or 0x80) + * mmr wrt or read address 0xc0 + * else if(mmr address == 0x148 or 0x200) + * mmr wrt or read address 0x28 + * else + * mmr wrt or read address 0x158 + * + * do desired mmr access (rd or wrt) + * + * if(mmr address == 0x100) + * mmr wrt or read address 0x38 + * mmr wrt or read address 0xb050 + * } else + * do desired mmr access + * + * According to hw, we can use reads instead of writes to the above address + * + * Note this WAR can only to be used for accessing internal MMR's in the + * TIOCE Coretalk Address Range 0x0 - 0x07ff_ffff. This includes the + * "Local CE Registers and Memories" and "PCI Compatible Config Space" address + * spaces from table 2-1 of the "CE Programmer's Reference Overview" document. + * + * All registers defined in struct tioce will meet that criteria. + */ + +static void inline +tioce_mmr_war_pre(struct tioce_kernel *kern, void __iomem *mmr_addr) +{ + u64 mmr_base; + u64 mmr_offset; + + if (kern->ce_common->ce_rev != TIOCE_REV_A) + return; + + mmr_base = kern->ce_common->ce_pcibus.bs_base; + mmr_offset = (unsigned long)mmr_addr - mmr_base; + + if (mmr_offset < 0x45000) { + u64 mmr_war_offset; + + if (mmr_offset == 0 || mmr_offset == 0x80) + mmr_war_offset = 0xc0; + else if (mmr_offset == 0x148 || mmr_offset == 0x200) + mmr_war_offset = 0x28; + else + mmr_war_offset = 0x158; + + readq_relaxed((void __iomem *)(mmr_base + mmr_war_offset)); + } +} + +static void inline +tioce_mmr_war_post(struct tioce_kernel *kern, void __iomem *mmr_addr) +{ + u64 mmr_base; + u64 mmr_offset; + + if (kern->ce_common->ce_rev != TIOCE_REV_A) + return; + + mmr_base = kern->ce_common->ce_pcibus.bs_base; + mmr_offset = (unsigned long)mmr_addr - mmr_base; + + if (mmr_offset < 0x45000) { + if (mmr_offset == 0x100) + readq_relaxed((void __iomem *)(mmr_base + 0x38)); + readq_relaxed((void __iomem *)(mmr_base + 0xb050)); + } +} + +/* load mmr contents into a variable */ +#define tioce_mmr_load(kern, mmrp, varp) do {\ + tioce_mmr_war_pre(kern, mmrp); \ + *(varp) = readq_relaxed(mmrp); \ + tioce_mmr_war_post(kern, mmrp); \ +} while (0) + +/* store variable contents into mmr */ +#define tioce_mmr_store(kern, mmrp, varp) do {\ + tioce_mmr_war_pre(kern, mmrp); \ + writeq(*varp, mmrp); \ + tioce_mmr_war_post(kern, mmrp); \ +} while (0) + +/* store immediate value into mmr */ +#define tioce_mmr_storei(kern, mmrp, val) do {\ + tioce_mmr_war_pre(kern, mmrp); \ + writeq(val, mmrp); \ + tioce_mmr_war_post(kern, mmrp); \ +} while (0) + +/* set bits (immediate value) into mmr */ +#define tioce_mmr_seti(kern, mmrp, bits) do {\ + u64 tmp; \ + tioce_mmr_load(kern, mmrp, &tmp); \ + tmp |= (bits); \ + tioce_mmr_store(kern, mmrp, &tmp); \ +} while (0) + +/* clear bits (immediate value) into mmr */ +#define tioce_mmr_clri(kern, mmrp, bits) do { \ + u64 tmp; \ + tioce_mmr_load(kern, mmrp, &tmp); \ + tmp &= ~(bits); \ + tioce_mmr_store(kern, mmrp, &tmp); \ +} while (0) + +/** + * Bus address ranges for the 5 flavors of TIOCE DMA + */ + +#define TIOCE_D64_MIN 0x8000000000000000UL +#define TIOCE_D64_MAX 0xffffffffffffffffUL +#define TIOCE_D64_ADDR(a) ((a) >= TIOCE_D64_MIN) + +#define TIOCE_D32_MIN 0x0000000080000000UL +#define TIOCE_D32_MAX 0x00000000ffffffffUL +#define TIOCE_D32_ADDR(a) ((a) >= TIOCE_D32_MIN && (a) <= TIOCE_D32_MAX) + +#define TIOCE_M32_MIN 0x0000000000000000UL +#define TIOCE_M32_MAX 0x000000007fffffffUL +#define TIOCE_M32_ADDR(a) ((a) >= TIOCE_M32_MIN && (a) <= TIOCE_M32_MAX) + +#define TIOCE_M40_MIN 0x0000004000000000UL +#define TIOCE_M40_MAX 0x0000007fffffffffUL +#define TIOCE_M40_ADDR(a) ((a) >= TIOCE_M40_MIN && (a) <= TIOCE_M40_MAX) + +#define TIOCE_M40S_MIN 0x0000008000000000UL +#define TIOCE_M40S_MAX 0x000000ffffffffffUL +#define TIOCE_M40S_ADDR(a) ((a) >= TIOCE_M40S_MIN && (a) <= TIOCE_M40S_MAX) + +/* + * ATE manipulation macros. + */ + +#define ATE_PAGESHIFT(ps) (__ffs(ps)) +#define ATE_PAGEMASK(ps) ((ps)-1) + +#define ATE_PAGE(x, ps) ((x) >> ATE_PAGESHIFT(ps)) +#define ATE_NPAGES(start, len, pagesize) \ + (ATE_PAGE((start)+(len)-1, pagesize) - ATE_PAGE(start, pagesize) + 1) + +#define ATE_VALID(ate) ((ate) & (1UL << 63)) +#define ATE_MAKE(addr, ps, msi) \ + (((addr) & ~ATE_PAGEMASK(ps)) | (1UL << 63) | ((msi)?(1UL << 62):0)) + +/* + * Flavors of ate-based mapping supported by tioce_alloc_map() + */ + +#define TIOCE_ATE_M32 1 +#define TIOCE_ATE_M40 2 +#define TIOCE_ATE_M40S 3 + +#define KB(x) ((u64)(x) << 10) +#define MB(x) ((u64)(x) << 20) +#define GB(x) ((u64)(x) << 30) + +/** + * tioce_dma_d64 - create a DMA mapping using 64-bit direct mode + * @ct_addr: system coretalk address + * + * Map @ct_addr into 64-bit CE bus space. No device context is necessary + * and no CE mapping are consumed. + * + * Bits 53:0 come from the coretalk address. The remaining bits are set as + * follows: + * + * 63 - must be 1 to indicate d64 mode to CE hardware + * 62 - barrier bit ... controlled with tioce_dma_barrier() + * 61 - msi bit ... specified through dma_flags + * 60:54 - reserved, MBZ + */ +static u64 +tioce_dma_d64(unsigned long ct_addr, int dma_flags) +{ + u64 bus_addr; + + bus_addr = ct_addr | (1UL << 63); + if (dma_flags & SN_DMA_MSI) + bus_addr |= (1UL << 61); + + return bus_addr; +} + +/** + * pcidev_to_tioce - return misc ce related pointers given a pci_dev + * @pci_dev: pci device context + * @base: ptr to store struct tioce_mmr * for the CE holding this device + * @kernel: ptr to store struct tioce_kernel * for the CE holding this device + * @port: ptr to store the CE port number that this device is on + * + * Return pointers to various CE-related structures for the CE upstream of + * @pci_dev. + */ +static inline void +pcidev_to_tioce(struct pci_dev *pdev, struct tioce __iomem **base, + struct tioce_kernel **kernel, int *port) +{ + struct pcidev_info *pcidev_info; + struct tioce_common *ce_common; + struct tioce_kernel *ce_kernel; + + pcidev_info = SN_PCIDEV_INFO(pdev); + ce_common = (struct tioce_common *)pcidev_info->pdi_pcibus_info; + ce_kernel = (struct tioce_kernel *)ce_common->ce_kernel_private; + + if (base) + *base = (struct tioce __iomem *)ce_common->ce_pcibus.bs_base; + if (kernel) + *kernel = ce_kernel; + + /* + * we use port as a zero-based value internally, even though the + * documentation is 1-based. + */ + if (port) + *port = + (pdev->bus->number < ce_kernel->ce_port1_secondary) ? 0 : 1; +} + +/** + * tioce_alloc_map - Given a coretalk address, map it to pcie bus address + * space using one of the various ATE-based address modes. + * @ce_kern: tioce context + * @type: map mode to use + * @port: 0-based port that the requesting device is downstream of + * @ct_addr: the coretalk address to map + * @len: number of bytes to map + * + * Given the addressing type, set up various parameters that define the + * ATE pool to use. Search for a contiguous block of entries to cover the + * length, and if enough resources exist, fill in the ATEs and construct a + * tioce_dmamap struct to track the mapping. + */ +static u64 +tioce_alloc_map(struct tioce_kernel *ce_kern, int type, int port, + u64 ct_addr, int len, int dma_flags) +{ + int i; + int j; + int first; + int last; + int entries; + int nates; + u64 pagesize; + int msi_capable, msi_wanted; + u64 *ate_shadow; + u64 __iomem *ate_reg; + u64 addr; + struct tioce __iomem *ce_mmr; + u64 bus_base; + struct tioce_dmamap *map; + + ce_mmr = (struct tioce __iomem *)ce_kern->ce_common->ce_pcibus.bs_base; + + switch (type) { + case TIOCE_ATE_M32: + /* + * The first 64 entries of the ate3240 pool are dedicated to + * super-page (TIOCE_ATE_M40S) mode. + */ + first = 64; + entries = TIOCE_NUM_M3240_ATES - 64; + ate_shadow = ce_kern->ce_ate3240_shadow; + ate_reg = ce_mmr->ce_ure_ate3240; + pagesize = ce_kern->ce_ate3240_pagesize; + bus_base = TIOCE_M32_MIN; + msi_capable = 1; + break; + case TIOCE_ATE_M40: + first = 0; + entries = TIOCE_NUM_M40_ATES; + ate_shadow = ce_kern->ce_ate40_shadow; + ate_reg = ce_mmr->ce_ure_ate40; + pagesize = MB(64); + bus_base = TIOCE_M40_MIN; + msi_capable = 0; + break; + case TIOCE_ATE_M40S: + /* + * ate3240 entries 0-31 are dedicated to port1 super-page + * mappings. ate3240 entries 32-63 are dedicated to port2. + */ + first = port * 32; + entries = 32; + ate_shadow = ce_kern->ce_ate3240_shadow; + ate_reg = ce_mmr->ce_ure_ate3240; + pagesize = GB(16); + bus_base = TIOCE_M40S_MIN; + msi_capable = 0; + break; + default: + return 0; + } + + msi_wanted = dma_flags & SN_DMA_MSI; + if (msi_wanted && !msi_capable) + return 0; + + nates = ATE_NPAGES(ct_addr, len, pagesize); + if (nates > entries) + return 0; + + last = first + entries - nates; + for (i = first; i <= last; i++) { + if (ATE_VALID(ate_shadow[i])) + continue; + + for (j = i; j < i + nates; j++) + if (ATE_VALID(ate_shadow[j])) + break; + + if (j >= i + nates) + break; + } + + if (i > last) + return 0; + + map = kzalloc(sizeof(struct tioce_dmamap), GFP_ATOMIC); + if (!map) + return 0; + + addr = ct_addr; + for (j = 0; j < nates; j++) { + u64 ate; + + ate = ATE_MAKE(addr, pagesize, msi_wanted); + ate_shadow[i + j] = ate; + tioce_mmr_storei(ce_kern, &ate_reg[i + j], ate); + addr += pagesize; + } + + map->refcnt = 1; + map->nbytes = nates * pagesize; + map->ct_start = ct_addr & ~ATE_PAGEMASK(pagesize); + map->pci_start = bus_base + (i * pagesize); + map->ate_hw = &ate_reg[i]; + map->ate_shadow = &ate_shadow[i]; + map->ate_count = nates; + + list_add(&map->ce_dmamap_list, &ce_kern->ce_dmamap_list); + + return (map->pci_start + (ct_addr - map->ct_start)); +} + +/** + * tioce_dma_d32 - create a DMA mapping using 32-bit direct mode + * @pdev: linux pci_dev representing the function + * @paddr: system physical address + * + * Map @paddr into 32-bit bus space of the CE associated with @pcidev_info. + */ +static u64 +tioce_dma_d32(struct pci_dev *pdev, u64 ct_addr, int dma_flags) +{ + int dma_ok; + int port; + struct tioce __iomem *ce_mmr; + struct tioce_kernel *ce_kern; + u64 ct_upper; + u64 ct_lower; + dma_addr_t bus_addr; + + if (dma_flags & SN_DMA_MSI) + return 0; + + ct_upper = ct_addr & ~0x3fffffffUL; + ct_lower = ct_addr & 0x3fffffffUL; + + pcidev_to_tioce(pdev, &ce_mmr, &ce_kern, &port); + + if (ce_kern->ce_port[port].dirmap_refcnt == 0) { + u64 tmp; + + ce_kern->ce_port[port].dirmap_shadow = ct_upper; + tioce_mmr_storei(ce_kern, &ce_mmr->ce_ure_dir_map[port], + ct_upper); + tmp = ce_mmr->ce_ure_dir_map[port]; + dma_ok = 1; + } else + dma_ok = (ce_kern->ce_port[port].dirmap_shadow == ct_upper); + + if (dma_ok) { + ce_kern->ce_port[port].dirmap_refcnt++; + bus_addr = TIOCE_D32_MIN + ct_lower; + } else + bus_addr = 0; + + return bus_addr; +} + +/** + * tioce_dma_barrier - swizzle a TIOCE bus address to include or exclude + * the barrier bit. + * @bus_addr: bus address to swizzle + * + * Given a TIOCE bus address, set the appropriate bit to indicate barrier + * attributes. + */ +static u64 +tioce_dma_barrier(u64 bus_addr, int on) +{ + u64 barrier_bit; + + /* barrier not supported in M40/M40S mode */ + if (TIOCE_M40_ADDR(bus_addr) || TIOCE_M40S_ADDR(bus_addr)) + return bus_addr; + + if (TIOCE_D64_ADDR(bus_addr)) + barrier_bit = (1UL << 62); + else /* must be m32 or d32 */ + barrier_bit = (1UL << 30); + + return (on) ? (bus_addr | barrier_bit) : (bus_addr & ~barrier_bit); +} + +/** + * tioce_dma_unmap - release CE mapping resources + * @pdev: linux pci_dev representing the function + * @bus_addr: bus address returned by an earlier tioce_dma_map + * @dir: mapping direction (unused) + * + * Locate mapping resources associated with @bus_addr and release them. + * For mappings created using the direct modes there are no resources + * to release. + */ +void +tioce_dma_unmap(struct pci_dev *pdev, dma_addr_t bus_addr, int dir) +{ + int i; + int port; + struct tioce_kernel *ce_kern; + struct tioce __iomem *ce_mmr; + unsigned long flags; + + bus_addr = tioce_dma_barrier(bus_addr, 0); + pcidev_to_tioce(pdev, &ce_mmr, &ce_kern, &port); + + /* nothing to do for D64 */ + + if (TIOCE_D64_ADDR(bus_addr)) + return; + + spin_lock_irqsave(&ce_kern->ce_lock, flags); + + if (TIOCE_D32_ADDR(bus_addr)) { + if (--ce_kern->ce_port[port].dirmap_refcnt == 0) { + ce_kern->ce_port[port].dirmap_shadow = 0; + tioce_mmr_storei(ce_kern, &ce_mmr->ce_ure_dir_map[port], + 0); + } + } else { + struct tioce_dmamap *map; + + list_for_each_entry(map, &ce_kern->ce_dmamap_list, + ce_dmamap_list) { + u64 last; + + last = map->pci_start + map->nbytes - 1; + if (bus_addr >= map->pci_start && bus_addr <= last) + break; + } + + if (&map->ce_dmamap_list == &ce_kern->ce_dmamap_list) { + printk(KERN_WARNING + "%s: %s - no map found for bus_addr 0x%llx\n", + __func__, pci_name(pdev), bus_addr); + } else if (--map->refcnt == 0) { + for (i = 0; i < map->ate_count; i++) { + map->ate_shadow[i] = 0; + tioce_mmr_storei(ce_kern, &map->ate_hw[i], 0); + } + + list_del(&map->ce_dmamap_list); + kfree(map); + } + } + + spin_unlock_irqrestore(&ce_kern->ce_lock, flags); +} + +/** + * tioce_do_dma_map - map pages for PCI DMA + * @pdev: linux pci_dev representing the function + * @paddr: host physical address to map + * @byte_count: bytes to map + * + * This is the main wrapper for mapping host physical pages to CE PCI space. + * The mapping mode used is based on the device's dma_mask. + */ +static u64 +tioce_do_dma_map(struct pci_dev *pdev, u64 paddr, size_t byte_count, + int barrier, int dma_flags) +{ + unsigned long flags; + u64 ct_addr; + u64 mapaddr = 0; + struct tioce_kernel *ce_kern; + struct tioce_dmamap *map; + int port; + u64 dma_mask; + + dma_mask = (barrier) ? pdev->dev.coherent_dma_mask : pdev->dma_mask; + + /* cards must be able to address at least 31 bits */ + if (dma_mask < 0x7fffffffUL) + return 0; + + if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS) + ct_addr = PHYS_TO_TIODMA(paddr); + else + ct_addr = paddr; + + /* + * If the device can generate 64 bit addresses, create a D64 map. + */ + if (dma_mask == ~0UL) { + mapaddr = tioce_dma_d64(ct_addr, dma_flags); + if (mapaddr) + goto dma_map_done; + } + + pcidev_to_tioce(pdev, NULL, &ce_kern, &port); + + spin_lock_irqsave(&ce_kern->ce_lock, flags); + + /* + * D64 didn't work ... See if we have an existing map that covers + * this address range. Must account for devices dma_mask here since + * an existing map might have been done in a mode using more pci + * address bits than this device can support. + */ + list_for_each_entry(map, &ce_kern->ce_dmamap_list, ce_dmamap_list) { + u64 last; + + last = map->ct_start + map->nbytes - 1; + if (ct_addr >= map->ct_start && + ct_addr + byte_count - 1 <= last && + map->pci_start <= dma_mask) { + map->refcnt++; + mapaddr = map->pci_start + (ct_addr - map->ct_start); + break; + } + } + + /* + * If we don't have a map yet, and the card can generate 40 + * bit addresses, try the M40/M40S modes. Note these modes do not + * support a barrier bit, so if we need a consistent map these + * won't work. + */ + if (!mapaddr && !barrier && dma_mask >= 0xffffffffffUL) { + /* + * We have two options for 40-bit mappings: 16GB "super" ATEs + * and 64MB "regular" ATEs. We'll try both if needed for a + * given mapping but which one we try first depends on the + * size. For requests >64MB, prefer to use a super page with + * regular as the fallback. Otherwise, try in the reverse order. + */ + + if (byte_count > MB(64)) { + mapaddr = tioce_alloc_map(ce_kern, TIOCE_ATE_M40S, + port, ct_addr, byte_count, + dma_flags); + if (!mapaddr) + mapaddr = + tioce_alloc_map(ce_kern, TIOCE_ATE_M40, -1, + ct_addr, byte_count, + dma_flags); + } else { + mapaddr = tioce_alloc_map(ce_kern, TIOCE_ATE_M40, -1, + ct_addr, byte_count, + dma_flags); + if (!mapaddr) + mapaddr = + tioce_alloc_map(ce_kern, TIOCE_ATE_M40S, + port, ct_addr, byte_count, + dma_flags); + } + } + + /* + * 32-bit direct is the next mode to try + */ + if (!mapaddr && dma_mask >= 0xffffffffUL) + mapaddr = tioce_dma_d32(pdev, ct_addr, dma_flags); + + /* + * Last resort, try 32-bit ATE-based map. + */ + if (!mapaddr) + mapaddr = + tioce_alloc_map(ce_kern, TIOCE_ATE_M32, -1, ct_addr, + byte_count, dma_flags); + + spin_unlock_irqrestore(&ce_kern->ce_lock, flags); + +dma_map_done: + if (mapaddr && barrier) + mapaddr = tioce_dma_barrier(mapaddr, 1); + + return mapaddr; +} + +/** + * tioce_dma - standard pci dma map interface + * @pdev: pci device requesting the map + * @paddr: system physical address to map into pci space + * @byte_count: # bytes to map + * + * Simply call tioce_do_dma_map() to create a map with the barrier bit clear + * in the address. + */ +static u64 +tioce_dma(struct pci_dev *pdev, unsigned long paddr, size_t byte_count, int dma_flags) +{ + return tioce_do_dma_map(pdev, paddr, byte_count, 0, dma_flags); +} + +/** + * tioce_dma_consistent - consistent pci dma map interface + * @pdev: pci device requesting the map + * @paddr: system physical address to map into pci space + * @byte_count: # bytes to map + * + * Simply call tioce_do_dma_map() to create a map with the barrier bit set + * in the address. + */ +static u64 +tioce_dma_consistent(struct pci_dev *pdev, unsigned long paddr, size_t byte_count, int dma_flags) +{ + return tioce_do_dma_map(pdev, paddr, byte_count, 1, dma_flags); +} + +/** + * tioce_error_intr_handler - SGI TIO CE error interrupt handler + * @irq: unused + * @arg: pointer to tioce_common struct for the given CE + * + * Handle a CE error interrupt. Simply a wrapper around a SAL call which + * defers processing to the SGI prom. + */ +static irqreturn_t +tioce_error_intr_handler(int irq, void *arg) +{ + struct tioce_common *soft = arg; + struct ia64_sal_retval ret_stuff; + ret_stuff.status = 0; + ret_stuff.v0 = 0; + + SAL_CALL_NOLOCK(ret_stuff, (u64) SN_SAL_IOIF_ERROR_INTERRUPT, + soft->ce_pcibus.bs_persist_segment, + soft->ce_pcibus.bs_persist_busnum, 0, 0, 0, 0, 0); + + if (ret_stuff.v0) + panic("tioce_error_intr_handler: Fatal TIOCE error"); + + return IRQ_HANDLED; +} + +/** + * tioce_reserve_m32 - reserve M32 ATEs for the indicated address range + * @tioce_kernel: TIOCE context to reserve ATEs for + * @base: starting bus address to reserve + * @limit: last bus address to reserve + * + * If base/limit falls within the range of bus space mapped through the + * M32 space, reserve the resources corresponding to the range. + */ +static void +tioce_reserve_m32(struct tioce_kernel *ce_kern, u64 base, u64 limit) +{ + int ate_index, last_ate, ps; + struct tioce __iomem *ce_mmr; + + ce_mmr = (struct tioce __iomem *)ce_kern->ce_common->ce_pcibus.bs_base; + ps = ce_kern->ce_ate3240_pagesize; + ate_index = ATE_PAGE(base, ps); + last_ate = ate_index + ATE_NPAGES(base, limit-base+1, ps) - 1; + + if (ate_index < 64) + ate_index = 64; + + if (last_ate >= TIOCE_NUM_M3240_ATES) + last_ate = TIOCE_NUM_M3240_ATES - 1; + + while (ate_index <= last_ate) { + u64 ate; + + ate = ATE_MAKE(0xdeadbeef, ps, 0); + ce_kern->ce_ate3240_shadow[ate_index] = ate; + tioce_mmr_storei(ce_kern, &ce_mmr->ce_ure_ate3240[ate_index], + ate); + ate_index++; + } +} + +/** + * tioce_kern_init - init kernel structures related to a given TIOCE + * @tioce_common: ptr to a cached tioce_common struct that originated in prom + */ +static struct tioce_kernel * +tioce_kern_init(struct tioce_common *tioce_common) +{ + int i; + int ps; + int dev; + u32 tmp; + unsigned int seg, bus; + struct tioce __iomem *tioce_mmr; + struct tioce_kernel *tioce_kern; + + tioce_kern = kzalloc(sizeof(struct tioce_kernel), GFP_KERNEL); + if (!tioce_kern) { + return NULL; + } + + tioce_kern->ce_common = tioce_common; + spin_lock_init(&tioce_kern->ce_lock); + INIT_LIST_HEAD(&tioce_kern->ce_dmamap_list); + tioce_common->ce_kernel_private = (u64) tioce_kern; + + /* + * Determine the secondary bus number of the port2 logical PPB. + * This is used to decide whether a given pci device resides on + * port1 or port2. Note: We don't have enough plumbing set up + * here to use pci_read_config_xxx() so use raw_pci_read(). + */ + + seg = tioce_common->ce_pcibus.bs_persist_segment; + bus = tioce_common->ce_pcibus.bs_persist_busnum; + + raw_pci_read(seg, bus, PCI_DEVFN(2, 0), PCI_SECONDARY_BUS, 1,&tmp); + tioce_kern->ce_port1_secondary = (u8) tmp; + + /* + * Set PMU pagesize to the largest size available, and zero out + * the ATEs. + */ + + tioce_mmr = (struct tioce __iomem *)tioce_common->ce_pcibus.bs_base; + tioce_mmr_clri(tioce_kern, &tioce_mmr->ce_ure_page_map, + CE_URE_PAGESIZE_MASK); + tioce_mmr_seti(tioce_kern, &tioce_mmr->ce_ure_page_map, + CE_URE_256K_PAGESIZE); + ps = tioce_kern->ce_ate3240_pagesize = KB(256); + + for (i = 0; i < TIOCE_NUM_M40_ATES; i++) { + tioce_kern->ce_ate40_shadow[i] = 0; + tioce_mmr_storei(tioce_kern, &tioce_mmr->ce_ure_ate40[i], 0); + } + + for (i = 0; i < TIOCE_NUM_M3240_ATES; i++) { + tioce_kern->ce_ate3240_shadow[i] = 0; + tioce_mmr_storei(tioce_kern, &tioce_mmr->ce_ure_ate3240[i], 0); + } + + /* + * Reserve ATEs corresponding to reserved address ranges. These + * include: + * + * Memory space covered by each PPB mem base/limit register + * Memory space covered by each PPB prefetch base/limit register + * + * These bus ranges are for pio (downstream) traffic only, and so + * cannot be used for DMA. + */ + + for (dev = 1; dev <= 2; dev++) { + u64 base, limit; + + /* mem base/limit */ + + raw_pci_read(seg, bus, PCI_DEVFN(dev, 0), + PCI_MEMORY_BASE, 2, &tmp); + base = (u64)tmp << 16; + + raw_pci_read(seg, bus, PCI_DEVFN(dev, 0), + PCI_MEMORY_LIMIT, 2, &tmp); + limit = (u64)tmp << 16; + limit |= 0xfffffUL; + + if (base < limit) + tioce_reserve_m32(tioce_kern, base, limit); + + /* + * prefetch mem base/limit. The tioce ppb's have 64-bit + * decoders, so read the upper portions w/o checking the + * attributes. + */ + + raw_pci_read(seg, bus, PCI_DEVFN(dev, 0), + PCI_PREF_MEMORY_BASE, 2, &tmp); + base = ((u64)tmp & PCI_PREF_RANGE_MASK) << 16; + + raw_pci_read(seg, bus, PCI_DEVFN(dev, 0), + PCI_PREF_BASE_UPPER32, 4, &tmp); + base |= (u64)tmp << 32; + + raw_pci_read(seg, bus, PCI_DEVFN(dev, 0), + PCI_PREF_MEMORY_LIMIT, 2, &tmp); + + limit = ((u64)tmp & PCI_PREF_RANGE_MASK) << 16; + limit |= 0xfffffUL; + + raw_pci_read(seg, bus, PCI_DEVFN(dev, 0), + PCI_PREF_LIMIT_UPPER32, 4, &tmp); + limit |= (u64)tmp << 32; + + if ((base < limit) && TIOCE_M32_ADDR(base)) + tioce_reserve_m32(tioce_kern, base, limit); + } + + return tioce_kern; +} + +/** + * tioce_force_interrupt - implement altix force_interrupt() backend for CE + * @sn_irq_info: sn asic irq that we need an interrupt generated for + * + * Given an sn_irq_info struct, set the proper bit in ce_adm_force_int to + * force a secondary interrupt to be generated. This is to work around an + * asic issue where there is a small window of opportunity for a legacy device + * interrupt to be lost. + */ +static void +tioce_force_interrupt(struct sn_irq_info *sn_irq_info) +{ + struct pcidev_info *pcidev_info; + struct tioce_common *ce_common; + struct tioce_kernel *ce_kern; + struct tioce __iomem *ce_mmr; + u64 force_int_val; + + if (!sn_irq_info->irq_bridge) + return; + + if (sn_irq_info->irq_bridge_type != PCIIO_ASIC_TYPE_TIOCE) + return; + + pcidev_info = (struct pcidev_info *)sn_irq_info->irq_pciioinfo; + if (!pcidev_info) + return; + + ce_common = (struct tioce_common *)pcidev_info->pdi_pcibus_info; + ce_mmr = (struct tioce __iomem *)ce_common->ce_pcibus.bs_base; + ce_kern = (struct tioce_kernel *)ce_common->ce_kernel_private; + + /* + * TIOCE Rev A workaround (PV 945826), force an interrupt by writing + * the TIO_INTx register directly (1/26/2006) + */ + if (ce_common->ce_rev == TIOCE_REV_A) { + u64 int_bit_mask = (1ULL << sn_irq_info->irq_int_bit); + u64 status; + + tioce_mmr_load(ce_kern, &ce_mmr->ce_adm_int_status, &status); + if (status & int_bit_mask) { + u64 force_irq = (1 << 8) | sn_irq_info->irq_irq; + u64 ctalk = sn_irq_info->irq_xtalkaddr; + u64 nasid, offset; + + nasid = (ctalk & CTALK_NASID_MASK) >> CTALK_NASID_SHFT; + offset = (ctalk & CTALK_NODE_OFFSET); + HUB_S(TIO_IOSPACE_ADDR(nasid, offset), force_irq); + } + + return; + } + + /* + * irq_int_bit is originally set up by prom, and holds the interrupt + * bit shift (not mask) as defined by the bit definitions in the + * ce_adm_int mmr. These shifts are not the same for the + * ce_adm_force_int register, so do an explicit mapping here to make + * things clearer. + */ + + switch (sn_irq_info->irq_int_bit) { + case CE_ADM_INT_PCIE_PORT1_DEV_A_SHFT: + force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT1_DEV_A_SHFT; + break; + case CE_ADM_INT_PCIE_PORT1_DEV_B_SHFT: + force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT1_DEV_B_SHFT; + break; + case CE_ADM_INT_PCIE_PORT1_DEV_C_SHFT: + force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT1_DEV_C_SHFT; + break; + case CE_ADM_INT_PCIE_PORT1_DEV_D_SHFT: + force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT1_DEV_D_SHFT; + break; + case CE_ADM_INT_PCIE_PORT2_DEV_A_SHFT: + force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT2_DEV_A_SHFT; + break; + case CE_ADM_INT_PCIE_PORT2_DEV_B_SHFT: + force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT2_DEV_B_SHFT; + break; + case CE_ADM_INT_PCIE_PORT2_DEV_C_SHFT: + force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT2_DEV_C_SHFT; + break; + case CE_ADM_INT_PCIE_PORT2_DEV_D_SHFT: + force_int_val = 1UL << CE_ADM_FORCE_INT_PCIE_PORT2_DEV_D_SHFT; + break; + default: + return; + } + tioce_mmr_storei(ce_kern, &ce_mmr->ce_adm_force_int, force_int_val); +} + +/** + * tioce_target_interrupt - implement set_irq_affinity for tioce resident + * functions. Note: only applies to line interrupts, not MSI's. + * + * @sn_irq_info: SN IRQ context + * + * Given an sn_irq_info, set the associated CE device's interrupt destination + * register. Since the interrupt destination registers are on a per-ce-slot + * basis, this will retarget line interrupts for all functions downstream of + * the slot. + */ +static void +tioce_target_interrupt(struct sn_irq_info *sn_irq_info) +{ + struct pcidev_info *pcidev_info; + struct tioce_common *ce_common; + struct tioce_kernel *ce_kern; + struct tioce __iomem *ce_mmr; + int bit; + u64 vector; + + pcidev_info = (struct pcidev_info *)sn_irq_info->irq_pciioinfo; + if (!pcidev_info) + return; + + ce_common = (struct tioce_common *)pcidev_info->pdi_pcibus_info; + ce_mmr = (struct tioce __iomem *)ce_common->ce_pcibus.bs_base; + ce_kern = (struct tioce_kernel *)ce_common->ce_kernel_private; + + bit = sn_irq_info->irq_int_bit; + + tioce_mmr_seti(ce_kern, &ce_mmr->ce_adm_int_mask, (1UL << bit)); + vector = (u64)sn_irq_info->irq_irq << INTR_VECTOR_SHFT; + vector |= sn_irq_info->irq_xtalkaddr; + tioce_mmr_storei(ce_kern, &ce_mmr->ce_adm_int_dest[bit], vector); + tioce_mmr_clri(ce_kern, &ce_mmr->ce_adm_int_mask, (1UL << bit)); + + tioce_force_interrupt(sn_irq_info); +} + +/** + * tioce_bus_fixup - perform final PCI fixup for a TIO CE bus + * @prom_bussoft: Common prom/kernel struct representing the bus + * + * Replicates the tioce_common pointed to by @prom_bussoft in kernel + * space. Allocates and initializes a kernel-only area for a given CE, + * and sets up an irq for handling CE error interrupts. + * + * On successful setup, returns the kernel version of tioce_common back to + * the caller. + */ +static void * +tioce_bus_fixup(struct pcibus_bussoft *prom_bussoft, struct pci_controller *controller) +{ + struct tioce_common *tioce_common; + struct tioce_kernel *tioce_kern; + struct tioce __iomem *tioce_mmr; + + /* + * Allocate kernel bus soft and copy from prom. + */ + + tioce_common = kzalloc(sizeof(struct tioce_common), GFP_KERNEL); + if (!tioce_common) + return NULL; + + memcpy(tioce_common, prom_bussoft, sizeof(struct tioce_common)); + tioce_common->ce_pcibus.bs_base = (unsigned long) + ioremap(REGION_OFFSET(tioce_common->ce_pcibus.bs_base), + sizeof(struct tioce_common)); + + tioce_kern = tioce_kern_init(tioce_common); + if (tioce_kern == NULL) { + kfree(tioce_common); + return NULL; + } + + /* + * Clear out any transient errors before registering the error + * interrupt handler. + */ + + tioce_mmr = (struct tioce __iomem *)tioce_common->ce_pcibus.bs_base; + tioce_mmr_seti(tioce_kern, &tioce_mmr->ce_adm_int_status_alias, ~0ULL); + tioce_mmr_seti(tioce_kern, &tioce_mmr->ce_adm_error_summary_alias, + ~0ULL); + tioce_mmr_seti(tioce_kern, &tioce_mmr->ce_dre_comp_err_addr, 0ULL); + + if (request_irq(SGI_PCIASIC_ERROR, + tioce_error_intr_handler, + IRQF_SHARED, "TIOCE error", (void *)tioce_common)) + printk(KERN_WARNING + "%s: Unable to get irq %d. " + "Error interrupts won't be routed for " + "TIOCE bus %04x:%02x\n", + __func__, SGI_PCIASIC_ERROR, + tioce_common->ce_pcibus.bs_persist_segment, + tioce_common->ce_pcibus.bs_persist_busnum); + + irq_set_handler(SGI_PCIASIC_ERROR, handle_level_irq); + sn_set_err_irq_affinity(SGI_PCIASIC_ERROR); + return tioce_common; +} + +static struct sn_pcibus_provider tioce_pci_interfaces = { + .dma_map = tioce_dma, + .dma_map_consistent = tioce_dma_consistent, + .dma_unmap = tioce_dma_unmap, + .bus_fixup = tioce_bus_fixup, + .force_interrupt = tioce_force_interrupt, + .target_interrupt = tioce_target_interrupt +}; + +/** + * tioce_init_provider - init SN PCI provider ops for TIO CE + */ +int +tioce_init_provider(void) +{ + sn_pci_provider[PCIIO_ASIC_TYPE_TIOCE] = &tioce_pci_interfaces; + return 0; +} -- cgit v1.2.3-54-g00ecf