arch/tile: tilegx PCI root complex support

This change implements PCIe root complex support for tilegx using
the kernel support layer for accessing the TRIO hardware shim.

Reviewed-by: Bjorn Helgaas <bhelgaas@google.com> [changes in 07487f3]
Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>

arch/tile/Kconfig

@@ -356,6 +356,9 @@ config PCI
 	default y
 	select PCI_DOMAINS
 	select GENERIC_PCI_IOMAP
+	select TILE_GXIO_TRIO if TILEGX
+	select ARCH_SUPPORTS_MSI if TILEGX
+	select PCI_MSI if TILEGX
 	---help---
 	  Enable PCI root complex support, so PCIe endpoint devices can
 	  be attached to the Tile chip.  Many, but not all, PCI devices

arch/tile/include/asm/pci.h

@@ -16,8 +16,11 @@
 #define _ASM_TILE_PCI_H
 
 #include <linux/pci.h>
+#include <linux/numa.h>
 #include <asm-generic/pci_iomap.h>
 
+#ifndef __tilegx__
+
 /*
  * Structure of a PCI controller (host bridge)
  */
@@ -40,6 +43,91 @@ struct pci_controller {
 	struct resource mem_resources[3];
 };
 
+/*
+ * This flag tells if the platform is TILEmpower that needs
+ * special configuration for the PLX switch chip.
+ */
+extern int tile_plx_gen1;
+
+static inline void pci_iounmap(struct pci_dev *dev, void __iomem *addr) {}
+
+#define	TILE_NUM_PCIE	2
+
+#else
+
+#include <asm/page.h>
+#include <gxio/trio.h>
+
+/**
+ * We reserve the hugepage-size address range at the top of the 64-bit address
+ * space to serve as the PCI window, emulating the BAR0 space of an endpoint
+ * device. This window is used by the chip-to-chip applications running on
+ * the RC node. The reason for carving out this window is that Mem-Maps that
+ * back up this window will not overlap with those that map the real physical
+ * memory.
+ */
+#define PCIE_HOST_BAR0_SIZE		HPAGE_SIZE
+#define PCIE_HOST_BAR0_START		HPAGE_MASK
+
+/**
+ * The first PAGE_SIZE of the above "BAR" window is mapped to the
+ * gxpci_host_regs structure.
+ */
+#define PCIE_HOST_REGS_SIZE		PAGE_SIZE
+
+/*
+ * This is the PCI address where the Mem-Map interrupt regions start.
+ * We use the 2nd to the last huge page of the 64-bit address space.
+ * The last huge page is used for the rootcomplex "bar", for C2C purpose.
+ */
+#define	MEM_MAP_INTR_REGIONS_BASE	(HPAGE_MASK - HPAGE_SIZE)
+
+/*
+ * Each Mem-Map interrupt region occupies 4KB.
+ */
+#define	MEM_MAP_INTR_REGION_SIZE	(1<< TRIO_MAP_MEM_LIM__ADDR_SHIFT)
+
+/*
+ * Structure of a PCI controller (host bridge) on Gx.
+ */
+struct pci_controller {
+
+	/* Pointer back to the TRIO that this PCIe port is connected to. */
+	gxio_trio_context_t *trio;
+	int mac;		/* PCIe mac index on the TRIO shim */
+	int trio_index;		/* Index of TRIO shim that contains the MAC. */
+
+	int pio_mem_index;	/* PIO region index for memory access */
+
+	/*
+	 * Mem-Map regions for all the memory controllers so that Linux can
+	 * map all of its physical memory space to the PCI bus.
+	 */
+	int mem_maps[MAX_NUMNODES];
+
+	int index;		/* PCI domain number */
+	struct pci_bus *root_bus;
+
+	int last_busno;
+
+	struct pci_ops *ops;
+
+	/* Table that maps the INTx numbers to Linux irq numbers. */
+	int irq_intx_table[4];
+
+	struct resource mem_space;
+
+	/* Address ranges that are routed to this controller/bridge. */
+	struct resource mem_resources[3];
+};
+
+extern struct pci_controller pci_controllers[TILEGX_NUM_TRIO * TILEGX_TRIO_PCIES];
+extern gxio_trio_context_t trio_contexts[TILEGX_NUM_TRIO];
+
+extern void pci_iounmap(struct pci_dev *dev, void __iomem *);
+
+#endif /* __tilegx__ */
+
 /*
  * The hypervisor maps the entirety of CPA-space as bus addresses, so
  * bus addresses are physical addresses.  The networking and block
@@ -50,12 +138,8 @@ struct pci_controller {
 int __init tile_pci_init(void);
 int __init pcibios_init(void);
 
-static inline void pci_iounmap(struct pci_dev *dev, void __iomem *addr) {}
-
 void __devinit pcibios_fixup_bus(struct pci_bus *bus);
 
-#define	TILE_NUM_PCIE	2
-
 #define pci_domain_nr(bus) (((struct pci_controller *)(bus)->sysdata)->index)
 
 /*
@@ -79,12 +163,6 @@ static inline int pcibios_assign_all_busses(void)
 #define PCIBIOS_MIN_MEM		0
 #define PCIBIOS_MIN_IO		0
 
-/*
- * This flag tells if the platform is TILEmpower that needs
- * special configuration for the PLX switch chip.
- */
-extern int tile_plx_gen1;
-
 /* Use any cpu for PCI. */
 #define cpumask_of_pcibus(bus) cpu_online_mask
 

arch/tile/kernel/Makefile

@@ -14,4 +14,8 @@ obj-$(CONFIG_SMP)		+= smpboot.o smp.o tlb.o
 obj-$(CONFIG_MODULES)		+= module.o
 obj-$(CONFIG_EARLY_PRINTK)	+= early_printk.o
 obj-$(CONFIG_KEXEC)		+= machine_kexec.o relocate_kernel_$(BITS).o
+ifdef CONFIG_TILEGX
+obj-$(CONFIG_PCI)		+= pci_gx.o
+else
 obj-$(CONFIG_PCI)		+= pci.o
+endif

arch/tile/kernel/pci_gx.c

+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   This program is free software; you can redistribute it and/or
+ *   modify it under the terms of the GNU General Public License
+ *   as published by the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/kernel.h>
+#include <linux/mmzone.h>
+#include <linux/pci.h>
+#include <linux/delay.h>
+#include <linux/string.h>
+#include <linux/init.h>
+#include <linux/capability.h>
+#include <linux/sched.h>
+#include <linux/errno.h>
+#include <linux/irq.h>
+#include <linux/msi.h>
+#include <linux/io.h>
+#include <linux/uaccess.h>
+#include <linux/ctype.h>
+
+#include <asm/processor.h>
+#include <asm/sections.h>
+#include <asm/byteorder.h>
+
+#include <gxio/iorpc_globals.h>
+#include <gxio/kiorpc.h>
+#include <gxio/trio.h>
+#include <gxio/iorpc_trio.h>
+#include <hv/drv_trio_intf.h>
+
+#include <arch/sim.h>
+
+/*
+ * Initialization flow and process
+ * -------------------------------
+ *
+ * This files containes the routines to search for PCI buses,
+ * enumerate the buses, and configure any attached devices.
+ *
+ * There are two entry points here:
+ * 1) tile_pci_init
+ *    This sets up the pci_controller structs, and opens the
+ *    FDs to the hypervisor.  This is called from setup_arch() early
+ *    in the boot process.
+ * 2) pcibios_init
+ *    This probes the PCI bus(es) for any attached hardware.  It's
+ *    called by subsys_initcall.  All of the real work is done by the
+ *    generic Linux PCI layer.
+ *
+ */
+
+#define DEBUG_PCI_CFG	0
+
+#if DEBUG_PCI_CFG
+#define TRACE_CFG_WR(size, val, bus, dev, func, offset) \
+	pr_info("CFG WR %d-byte VAL %#x to bus %d dev %d func %d addr %u\n", \
+		size, val, bus, dev, func, offset & 0xFFF);
+#define TRACE_CFG_RD(size, val, bus, dev, func, offset) \
+	pr_info("CFG RD %d-byte VAL %#x from bus %d dev %d func %d addr %u\n", \
+		size, val, bus, dev, func, offset & 0xFFF);
+#else
+#define TRACE_CFG_WR(...)
+#define TRACE_CFG_RD(...)
+#endif
+
+static int __devinitdata pci_probe = 1;
+
+/* Information on the PCIe RC ports configuration. */
+static int __devinitdata pcie_rc[TILEGX_NUM_TRIO][TILEGX_TRIO_PCIES];
+
+/*
+ * On some platforms with one or more Gx endpoint ports, we need to
+ * delay the PCIe RC port probe for a few seconds to work around
+ * a HW PCIe link-training bug. The exact delay is specified with
+ * a kernel boot argument in the form of "pcie_rc_delay=T,P,S",
+ * where T is the TRIO instance number, P is the port number and S is
+ * the delay in seconds. If the delay is not provided, the value
+ * will be DEFAULT_RC_DELAY.
+ */
+static int __devinitdata rc_delay[TILEGX_NUM_TRIO][TILEGX_TRIO_PCIES];
+
+/* Default number of seconds that the PCIe RC port probe can be delayed. */
+#define DEFAULT_RC_DELAY	10
+
+/* Max number of seconds that the PCIe RC port probe can be delayed. */
+#define MAX_RC_DELAY		20
+
+/* Array of the PCIe ports configuration info obtained from the BIB. */
+struct pcie_port_property pcie_ports[TILEGX_NUM_TRIO][TILEGX_TRIO_PCIES];
+
+/* All drivers share the TRIO contexts defined here. */
+gxio_trio_context_t trio_contexts[TILEGX_NUM_TRIO];
+
+/* Pointer to an array of PCIe RC controllers. */
+struct pci_controller pci_controllers[TILEGX_NUM_TRIO * TILEGX_TRIO_PCIES];
+int num_rc_controllers;
+static int num_ep_controllers;
+
+static struct pci_ops tile_cfg_ops;
+
+/* Mask of CPUs that should receive PCIe interrupts. */
+static struct cpumask intr_cpus_map;
+
+/*
+ * We don't need to worry about the alignment of resources.
+ */
+resource_size_t pcibios_align_resource(void *data, const struct resource *res,
+				resource_size_t size, resource_size_t align)
+{
+	return res->start;
+}
+EXPORT_SYMBOL(pcibios_align_resource);
+
+
+/*
+ * Pick a CPU to receive and handle the PCIe interrupts, based on the IRQ #.
+ * For now, we simply send interrupts to non-dataplane CPUs.
+ * We may implement methods to allow user to specify the target CPUs,
+ * e.g. via boot arguments.
+ */
+static int tile_irq_cpu(int irq)
+{
+	unsigned int count;
+	int i = 0;
+	int cpu;
+
+	count = cpumask_weight(&intr_cpus_map);
+	if (unlikely(count == 0)) {
+		pr_warning("intr_cpus_map empty, interrupts will be"
+			   " delievered to dataplane tiles\n");
+		return irq % (smp_height * smp_width);
+	}
+
+	count = irq % count;
+	for_each_cpu(cpu, &intr_cpus_map) {
+		if (i++ == count)
+			break;
+	}
+	return cpu;
+}
+
+/*
+ * Open a file descriptor to the TRIO shim.
+ */
+static int __devinit tile_pcie_open(int trio_index)
+{
+	gxio_trio_context_t *context = &trio_contexts[trio_index];
+	int ret;
+
+	/*
+	 * This opens a file descriptor to the TRIO shim.
+	 */
+	ret = gxio_trio_init(context, trio_index);
+	if (ret < 0)
+		return ret;
+
+	/*
+	 * Allocate an ASID for the kernel.
+	 */
+	ret = gxio_trio_alloc_asids(context, 1, 0, 0);
+	if (ret < 0) {
+		pr_err("PCI: ASID alloc failure on TRIO %d, give up\n",
+			trio_index);
+		goto asid_alloc_failure;
+	}
+
+	context->asid = ret;
+
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+	/*
+	 * Alloc a PIO region for config access, shared by all MACs per TRIO.
+	 * This shouldn't fail since the kernel is supposed to the first
+	 * client of the TRIO's PIO regions.
+	 */
+	ret = gxio_trio_alloc_pio_regions(context, 1, 0, 0);
+	if (ret < 0) {
+		pr_err("PCI: CFG PIO alloc failure on TRIO %d, give up\n",
+			trio_index);
+		goto pio_alloc_failure;
+	}
+
+	context->pio_cfg_index = ret;
+
+	/*
+	 * For PIO CFG, the bus_address_hi parameter is 0. The mac parameter
+	 * is also 0 because it is specified in PIO_REGION_SETUP_CFG_ADDR.
+	 */
+	ret = gxio_trio_init_pio_region_aux(context, context->pio_cfg_index,
+		0, 0, HV_TRIO_PIO_FLAG_CONFIG_SPACE);
+	if (ret < 0) {
+		pr_err("PCI: CFG PIO init failure on TRIO %d, give up\n",
+			trio_index);
+		goto pio_alloc_failure;
+	}
+#endif
+
+	return ret;
+
+asid_alloc_failure:
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+pio_alloc_failure:
+#endif
+	hv_dev_close(context->fd);
+
+	return ret;
+}
+
+static void
+tilegx_legacy_irq_ack(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_EVENT_RESET_K, 1UL << d->irq);
+}
+
+static void
+tilegx_legacy_irq_mask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_SET_K, 1UL << d->irq);
+}
+
+static void
+tilegx_legacy_irq_unmask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_RESET_K, 1UL << d->irq);
+}
+
+static struct irq_chip tilegx_legacy_irq_chip = {
+	.name			= "tilegx_legacy_irq",
+	.irq_ack		= tilegx_legacy_irq_ack,
+	.irq_mask		= tilegx_legacy_irq_mask,
+	.irq_unmask		= tilegx_legacy_irq_unmask,
+
+	/* TBD: support set_affinity. */
+};
+
+/*
+ * This is a wrapper function of the kernel level-trigger interrupt
+ * handler handle_level_irq() for PCI legacy interrupts. The TRIO
+ * is configured such that only INTx Assert interrupts are proxied
+ * to Linux which just calls handle_level_irq() after clearing the
+ * MAC INTx Assert status bit associated with this interrupt.
+ */
+static void
+trio_handle_level_irq(unsigned int irq, struct irq_desc *desc)
+{
+	struct pci_controller *controller = irq_desc_get_handler_data(desc);
+	gxio_trio_context_t *trio_context = controller->trio;
+	uint64_t intx = (uint64_t)irq_desc_get_chip_data(desc);
+	int mac = controller->mac;
+	unsigned int reg_offset;
+	uint64_t level_mask;
+
+	handle_level_irq(irq, desc);
+
+	/*
+	 * Clear the INTx Level status, otherwise future interrupts are
+	 * not sent.
+	 */
+	reg_offset = (TRIO_PCIE_INTFC_MAC_INT_STS <<
+		TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+		TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	level_mask = TRIO_PCIE_INTFC_MAC_INT_STS__INT_LEVEL_MASK << intx;
+
+	__gxio_mmio_write(trio_context->mmio_base_mac + reg_offset, level_mask);
+}
+
+/*
+ * Create kernel irqs and set up the handlers for the legacy interrupts.
+ * Also some minimum initialization for the MSI support.
+ */
+static int __devinit tile_init_irqs(struct pci_controller *controller)
+{
+	int i;
+	int j;
+	int irq;
+	int result;
+
+	cpumask_copy(&intr_cpus_map, cpu_online_mask);
+
+
+	for (i = 0; i < 4; i++) {
+		gxio_trio_context_t *context = controller->trio;
+		int cpu;
+
+		/* Ask the kernel to allocate an IRQ. */
+		irq = create_irq();
+		if (irq < 0) {
+			pr_err("PCI: no free irq vectors, failed for %d\n", i);
+
+			goto free_irqs;
+		}
+		controller->irq_intx_table[i] = irq;
+
+		/* Distribute the 4 IRQs to different tiles. */
+		cpu = tile_irq_cpu(irq);
+
+		/* Configure the TRIO intr binding for this IRQ. */
+		result = gxio_trio_config_legacy_intr(context, cpu_x(cpu),
+						      cpu_y(cpu), KERNEL_PL,
+						      irq, controller->mac, i);
+		if (result < 0) {
+			pr_err("PCI: MAC intx config failed for %d\n", i);
+
+			goto free_irqs;
+		}
+
+		/*
+		 * Register the IRQ handler with the kernel.
+		 */
+		irq_set_chip_and_handler(irq, &tilegx_legacy_irq_chip,
+					trio_handle_level_irq);
+		irq_set_chip_data(irq, (void *)(uint64_t)i);
+		irq_set_handler_data(irq, controller);
+	}
+
+	return 0;
+
+free_irqs:
+	for (j = 0; j < i; j++)
+		destroy_irq(controller->irq_intx_table[j]);
+
+	return -1;
+}
+
+/*
+ * First initialization entry point, called from setup_arch().
+ *
+ * Find valid controllers and fill in pci_controller structs for each
+ * of them.
+ *
+ * Returns the number of controllers discovered.
+ */
+int __init tile_pci_init(void)
+{
+	int num_trio_shims = 0;
+	int ctl_index = 0;
+	int i, j;
+
+	if (!pci_probe) {
+		pr_info("PCI: disabled by boot argument\n");
+		return 0;
+	}
+
+	pr_info("PCI: Searching for controllers...\n");
+
+	/*
+	 * We loop over all the TRIO shims.
+	 */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		int ret;
+
+		ret = tile_pcie_open(i);
+		if (ret < 0)
+			continue;
+
+		num_trio_shims++;
+	}
+
+	if (num_trio_shims == 0 || sim_is_simulator())
+		return 0;
+
+	/*
+	 * Now determine which PCIe ports are configured to operate in RC mode.
+	 * We look at the Board Information Block first and then see if there
+	 * are any overriding configuration by the HW strapping pin.
+	 */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		gxio_trio_context_t *context = &trio_contexts[i];
+		int ret;
+
+		if (context->fd < 0)
+			continue;
+
+		ret = hv_dev_pread(context->fd, 0,
+			(HV_VirtAddr)&pcie_ports[i][0],
+			sizeof(struct pcie_port_property) * TILEGX_TRIO_PCIES,
+			GXIO_TRIO_OP_GET_PORT_PROPERTY);
+		if (ret < 0) {
+			pr_err("PCI: PCIE_GET_PORT_PROPERTY failure, error %d,"
+				" on TRIO %d\n", ret, i);
+			continue;
+		}
+
+		for (j = 0; j < TILEGX_TRIO_PCIES; j++) {
+			if (pcie_ports[i][j].allow_rc) {
+				pcie_rc[i][j] = 1;
+				num_rc_controllers++;
+			}
+			else if (pcie_ports[i][j].allow_ep) {
+				num_ep_controllers++;
+			}
+		}
+	}
+
+	/*
+	 * Return if no PCIe ports are configured to operate in RC mode.
+	 */
+	if (num_rc_controllers == 0)
+		return 0;
+
+	/*
+	 * Set the TRIO pointer and MAC index for each PCIe RC port.
+	 */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		for (j = 0; j < TILEGX_TRIO_PCIES; j++) {
+			if (pcie_rc[i][j]) {
+				pci_controllers[ctl_index].trio =
+					&trio_contexts[i];
+				pci_controllers[ctl_index].mac = j;
+				pci_controllers[ctl_index].trio_index = i;
+				ctl_index++;
+				if (ctl_index == num_rc_controllers)
+					goto out;
+			}
+		}
+	}
+
+out:
+	/*
+	 * Configure each PCIe RC port.
+	 */
+	for (i = 0; i < num_rc_controllers; i++) {
+		/*
+		 * Configure the PCIe MAC to run in RC mode.
+		 */
+
+		struct pci_controller *controller = &pci_controllers[i];
+
+		controller->index = i;
+		controller->last_busno = 0xff;
+		controller->ops = &tile_cfg_ops;
+
+	}
+
+	return num_rc_controllers;
+}
+
+/*
+ * (pin - 1) converts from the PCI standard's [1:4] convention to
+ * a normal [0:3] range.
+ */
+static int tile_map_irq(const struct pci_dev *dev, u8 device, u8 pin)
+{
+	struct pci_controller *controller =
+		(struct pci_controller *)dev->sysdata;
+	return controller->irq_intx_table[pin - 1];
+}
+
+
+static void __devinit fixup_read_and_payload_sizes(struct pci_controller *
+						controller)
+{
+	gxio_trio_context_t *trio_context = controller->trio;
+	struct pci_bus *root_bus = controller->root_bus;
+	TRIO_PCIE_RC_DEVICE_CONTROL_t dev_control;
+	TRIO_PCIE_RC_DEVICE_CAP_t rc_dev_cap;
+	unsigned int reg_offset;
+	struct pci_bus *child;
+	int mac;
+	int err;
+
+	mac = controller->mac;
+
+	/*
+	 * Set our max read request size to be 4KB.
+	 */
+	reg_offset =
+		(TRIO_PCIE_RC_DEVICE_CONTROL <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	dev_control.word = __gxio_mmio_read32(trio_context->mmio_base_mac +
+						reg_offset);
+	dev_control.max_read_req_sz = 5;
+	__gxio_mmio_write32(trio_context->mmio_base_mac + reg_offset,
+						dev_control.word);
+
+	/*
+	 * Set the max payload size supported by this Gx PCIe MAC.
+	 * Though Gx PCIe supports Max Payload Size of up to 1024 bytes,
+	 * experiments have shown that setting MPS to 256 yields the
+	 * best performance.
+	 */
+	reg_offset =
+		(TRIO_PCIE_RC_DEVICE_CAP <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	rc_dev_cap.word = __gxio_mmio_read32(trio_context->mmio_base_mac +
+						reg_offset);
+	rc_dev_cap.mps_sup = 1;
+	__gxio_mmio_write32(trio_context->mmio_base_mac + reg_offset,
+						rc_dev_cap.word);
+
+	/* Configure PCI Express MPS setting. */
+	list_for_each_entry(child, &root_bus->children, node) {
+		struct pci_dev *self = child->self;
+		if (!self)
+			continue;
+
+		pcie_bus_configure_settings(child, self->pcie_mpss);
+	}
+
+	/*
+	 * Set the mac_config register in trio based on the MPS/MRS of the link.
+	 */
+	reg_offset =
+		(TRIO_PCIE_RC_DEVICE_CONTROL <<
+			TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+		(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+			TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+		(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+	dev_control.word = __gxio_mmio_read32(trio_context->mmio_base_mac +
+						reg_offset);
+
+	err = gxio_trio_set_mps_mrs(trio_context,
+				    dev_control.max_payload_size,
+				    dev_control.max_read_req_sz,
+				    mac);
+        if (err < 0) {
+		pr_err("PCI: PCIE_CONFIGURE_MAC_MPS_MRS failure, "
+			"MAC %d on TRIO %d\n",
+			mac, controller->trio_index);
+	}
+}
+
+static int __devinit setup_pcie_rc_delay(char *str)
+{
+	unsigned long delay = 0;
+	unsigned long trio_index;
+	unsigned long mac;
+
+	if (str == NULL || !isdigit(*str))
+		return -EINVAL;
+	trio_index = simple_strtoul(str, (char **)&str, 10);
+	if (trio_index >= TILEGX_NUM_TRIO)
+		return -EINVAL;
+
+	if (*str != ',')
+		return -EINVAL;
+
+	str++;
+	if (!isdigit(*str))
+		return -EINVAL;
+	mac = simple_strtoul(str, (char **)&str, 10);
+	if (mac >= TILEGX_TRIO_PCIES)
+		return -EINVAL;
+
+	if (*str != '\0') {
+		if (*str != ',')
+			return -EINVAL;
+
+		str++;
+		if (!isdigit(*str))
+			return -EINVAL;
+		delay = simple_strtoul(str, (char **)&str, 10);
+		if (delay > MAX_RC_DELAY)
+			return -EINVAL;
+	}
+
+	rc_delay[trio_index][mac] = delay ? : DEFAULT_RC_DELAY;
+	pr_info("Delaying PCIe RC link training for %u sec"
+		" on MAC %lu on TRIO %lu\n", rc_delay[trio_index][mac],
+		mac, trio_index);
+	return 0;
+}
+early_param("pcie_rc_delay", setup_pcie_rc_delay);
+
+/*
+ * Second PCI initialization entry point, called by subsys_initcall.
+ *
+ * The controllers have been set up by the time we get here, by a call to
+ * tile_pci_init.
+ */
+int __init pcibios_init(void)
+{
+	resource_size_t offset;
+	LIST_HEAD(resources);
+	int i;
+
+	if (num_rc_controllers == 0 && num_ep_controllers == 0)
+		return 0;
+
+	pr_info("PCI: Probing PCI hardware\n");
+
+	/*
+	 * We loop over all the TRIO shims and set up the MMIO mappings.
+	 * This step can't be done in tile_pci_init because the MM subsystem
+	 * hasn't been initialized then.
+	 */
+	for (i = 0; i < TILEGX_NUM_TRIO; i++) {
+		gxio_trio_context_t *context = &trio_contexts[i];
+
+		if (context->fd < 0)
+			continue;
+
+		/*
+		 * Map in the MMIO space for the MAC.
+		 */
+		offset = 0;
+		context->mmio_base_mac =
+			iorpc_ioremap(context->fd, offset,
+				      HV_TRIO_CONFIG_IOREMAP_SIZE);
+		if (context->mmio_base_mac == NULL) {
+			pr_err("PCI: MAC map failure on TRIO %d\n", i);
+
+			hv_dev_close(context->fd);
+			context->fd = -1;
+			continue;
+		}
+	}
+
+	/*
+	 * Delay a bit in case devices aren't ready.  Some devices are
+	 * known to require at least 20ms here, but we use a more
+	 * conservative value.
+	 */
+	msleep(250);
+
+	/* Scan all of the recorded PCI controllers.  */
+	for (i = 0; i < num_rc_controllers; i++) {
+		struct pci_controller *controller = &pci_controllers[i];
+		gxio_trio_context_t *trio_context = controller->trio;
+		TRIO_PCIE_INTFC_PORT_CONFIG_t port_config;
+		TRIO_PCIE_INTFC_PORT_STATUS_t port_status;
+		TRIO_PCIE_INTFC_TX_FIFO_CTL_t tx_fifo_ctl;
+		struct pci_bus *bus;
+		unsigned int reg_offset;
+		unsigned int class_code_revision;
+		int trio_index;
+		int mac;
+#ifndef USE_SHARED_PCIE_CONFIG_REGION
+		int ret;
+#endif
+
+		if (trio_context->fd < 0)
+			continue;
+
+		trio_index = controller->trio_index;
+		mac = controller->mac;
+
+		/*
+		 * Check the port strap state which will override the BIB
+		 * setting.
+		 */
+
+		reg_offset =
+			(TRIO_PCIE_INTFC_PORT_CONFIG <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		port_config.word =
+			__gxio_mmio_read(trio_context->mmio_base_mac +
+					 reg_offset);
+
+		if ((port_config.strap_state !=
+			TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_RC) &&
+			(port_config.strap_state !=
+			TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_RC_G1)) {
+			/*
+			 * If this is really intended to be an EP port,
+			 * record it so that the endpoint driver will know about it.
+			 */
+			if (port_config.strap_state ==
+			TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_ENDPOINT ||
+			port_config.strap_state ==
+			TRIO_PCIE_INTFC_PORT_CONFIG__STRAP_STATE_VAL_AUTO_CONFIG_ENDPOINT_G1)
+				pcie_ports[trio_index][mac].allow_ep = 1;
+
+			continue;
+		}
+
+		/*
+		 * Delay the RC link training if needed.
+		 */
+		if (rc_delay[trio_index][mac])
+			msleep(rc_delay[trio_index][mac] * 1000);
+
+		ret = gxio_trio_force_rc_link_up(trio_context, mac);
+		if (ret < 0)
+			pr_err("PCI: PCIE_FORCE_LINK_UP failure, "
+				"MAC %d on TRIO %d\n", mac, trio_index);
+
+		pr_info("PCI: Found PCI controller #%d on TRIO %d MAC %d\n", i,
+			trio_index, controller->mac);
+
+		/*
+		 * Wait a bit here because some EP devices take longer
+		 * to come up.
+		 */
+		msleep(1000);
+
+		/*
+		 * Check for PCIe link-up status.
+		 */
+
+		reg_offset =
+			(TRIO_PCIE_INTFC_PORT_STATUS <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		port_status.word =
+			__gxio_mmio_read(trio_context->mmio_base_mac +
+					 reg_offset);
+		if (!port_status.dl_up) {
+			pr_err("PCI: link is down, MAC %d on TRIO %d\n",
+				mac, trio_index);
+			continue;
+		}
+
+		/*
+		 * Ensure that the link can come out of L1 power down state.
+		 * Strictly speaking, this is needed only in the case of
+		 * heavy RC-initiated DMAs.
+		 */
+		reg_offset =
+			(TRIO_PCIE_INTFC_TX_FIFO_CTL <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_INTERFACE <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+		tx_fifo_ctl.word =
+			__gxio_mmio_read(trio_context->mmio_base_mac +
+					 reg_offset);
+		tx_fifo_ctl.min_p_credits = 0;
+		__gxio_mmio_write(trio_context->mmio_base_mac + reg_offset,
+				  tx_fifo_ctl.word);
+
+		/*
+		 * Change the device ID so that Linux bus crawl doesn't confuse
+		 * the internal bridge with any Tilera endpoints.
+		 */
+
+		reg_offset =
+			(TRIO_PCIE_RC_DEVICE_ID_VEN_ID <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		__gxio_mmio_write32(trio_context->mmio_base_mac + reg_offset,
+				    (TILERA_GX36_RC_DEV_ID <<
+				    TRIO_PCIE_RC_DEVICE_ID_VEN_ID__DEV_ID_SHIFT) |
+				    TILERA_VENDOR_ID);
+
+		/*
+		 * Set the internal P2P bridge class code.
+		 */
+
+		reg_offset =
+			(TRIO_PCIE_RC_REVISION_ID <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+			(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_STANDARD <<
+				TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+			(mac << TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+		class_code_revision =
+			__gxio_mmio_read32(trio_context->mmio_base_mac +
+					   reg_offset);
+		class_code_revision = (class_code_revision & 0xff ) |
+					(PCI_CLASS_BRIDGE_PCI << 16);
+
+		__gxio_mmio_write32(trio_context->mmio_base_mac +
+				    reg_offset, class_code_revision);
+
+#ifdef USE_SHARED_PCIE_CONFIG_REGION
+
+		/*
+		 * Map in the MMIO space for the PIO region.
+		 */
+		offset = HV_TRIO_PIO_OFFSET(trio_context->pio_cfg_index) |
+			(((unsigned long long)mac) <<
+			TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT);
+
+#else
+
+		/*
+		 * Alloc a PIO region for PCI config access per MAC.
+		 */
+		ret = gxio_trio_alloc_pio_regions(trio_context, 1, 0, 0);
+		if (ret < 0) {
+			pr_err("PCI: PCI CFG PIO alloc failure for mac %d "
+				"on TRIO %d, give up\n", mac, trio_index);
+
+			/* TBD: cleanup ... */
+
+			continue;
+		}
+
+		trio_context->pio_cfg_index[mac] = ret;
+
+		/*
+		 * For PIO CFG, the bus_address_hi parameter is 0.
+		 */
+		ret = gxio_trio_init_pio_region_aux(trio_context,
+			trio_context->pio_cfg_index[mac],
+			mac, 0, HV_TRIO_PIO_FLAG_CONFIG_SPACE);
+		if (ret < 0) {
+			pr_err("PCI: PCI CFG PIO init failure for mac %d "
+				"on TRIO %d, give up\n", mac, trio_index);
+
+			/* TBD: cleanup ... */
+
+			continue;
+		}
+
+		offset = HV_TRIO_PIO_OFFSET(trio_context->pio_cfg_index[mac]) |
+			(((unsigned long long)mac) <<
+			TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT);
+
+#endif
+
+		trio_context->mmio_base_pio_cfg[mac] =
+			iorpc_ioremap(trio_context->fd, offset,
+			(1 << TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR__MAC_SHIFT));
+		if (trio_context->mmio_base_pio_cfg[mac] == NULL) {
+			pr_err("PCI: PIO map failure for mac %d on TRIO %d\n",
+				mac, trio_index);
+
+			/* TBD: cleanup ... */
+
+			continue;
+		}
+
+		/*
+		 * Initialize the PCIe interrupts.
+		 */
+		if (tile_init_irqs(controller)) {
+			pr_err("PCI: IRQs init failure for mac %d on TRIO %d\n",
+				mac, trio_index);
+
+			continue;
+		}
+
+		pci_add_resource(&resources, &iomem_resource);
+		bus = pci_scan_root_bus(NULL, 0, controller->ops,
+					controller, &resources);
+		controller->root_bus = bus;
+		controller->last_busno = bus->subordinate;
+
+	}
+
+	/* Do machine dependent PCI interrupt routing */
+	pci_fixup_irqs(pci_common_swizzle, tile_map_irq);
+
+	/*
+	 * This comes from the generic Linux PCI driver.
+	 *
+	 * It allocates all of the resources (I/O memory, etc)
+	 * associated with the devices read in above.
+	 */
+
+	pci_assign_unassigned_resources();
+
+	/* Record the I/O resources in the PCI controller structure. */
+	for (i = 0; i < num_rc_controllers; i++) {
+		struct pci_controller *controller = &pci_controllers[i];
+		gxio_trio_context_t *trio_context = controller->trio;
+		struct pci_bus *root_bus = pci_controllers[i].root_bus;
+		struct pci_bus *next_bus;
+		uint32_t bus_address_hi;
+		struct pci_dev *dev;
+		int ret;
+		int j;
+
+		/*
+		 * Skip controllers that are not properly initialized or
+		 * have down links.
+		 */
+		if (root_bus == NULL)
+			continue;
+
+		/* Configure the max_payload_size values for this domain. */
+		fixup_read_and_payload_sizes(controller);
+
+		list_for_each_entry(dev, &root_bus->devices, bus_list) {
+			/* Find the PCI host controller, ie. the 1st bridge. */
+			if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI &&
+				(PCI_SLOT(dev->devfn) == 0)) {
+				next_bus = dev->subordinate;
+				pci_controllers[i].mem_resources[0] =
+					*next_bus->resource[0];
+				pci_controllers[i].mem_resources[1] =
+					 *next_bus->resource[1];
+				pci_controllers[i].mem_resources[2] =
+					 *next_bus->resource[2];
+
+				break;
+			}
+		}
+
+		if (pci_controllers[i].mem_resources[1].flags & IORESOURCE_MEM)
+			bus_address_hi =
+				pci_controllers[i].mem_resources[1].start >> 32;
+		else if (pci_controllers[i].mem_resources[2].flags & IORESOURCE_PREFETCH)
+			bus_address_hi =
+				pci_controllers[i].mem_resources[2].start >> 32;
+		else {
+			/* This is unlikely. */
+			pr_err("PCI: no memory resources on TRIO %d mac %d\n",
+				controller->trio_index, controller->mac);
+			continue;
+		}
+
+		/*
+		 * We always assign 32-bit PCI bus BAR ranges.
+		 */
+		BUG_ON(bus_address_hi != 0);
+
+		/*
+		 * Alloc a PIO region for PCI memory access for each RC port.
+		 */
+		ret = gxio_trio_alloc_pio_regions(trio_context, 1, 0, 0);
+		if (ret < 0) {
+			pr_err("PCI: MEM PIO alloc failure on TRIO %d mac %d, "
+				"give up\n", controller->trio_index,
+				controller->mac);
+
+			/* TBD: cleanup ... */
+
+			continue;
+		}
+
+		controller->pio_mem_index = ret;
+
+		/*
+		 * For PIO MEM, the bus_address_hi parameter is hard-coded 0
+		 * because we always assign 32-bit PCI bus BAR ranges.
+		 */
+		ret = gxio_trio_init_pio_region_aux(trio_context,
+						    controller->pio_mem_index,
+						    controller->mac,
+						    bus_address_hi,
+						    0);
+		if (ret < 0) {
+			pr_err("PCI: MEM PIO init failure on TRIO %d mac %d, "
+				"give up\n", controller->trio_index,
+				controller->mac);
+
+			/* TBD: cleanup ... */
+
+			continue;
+		}
+
+		/*
+		 * Configure a Mem-Map region for each memory controller so
+		 * that Linux can map all of its PA space to the PCI bus.
+		 * Use the IOMMU to handle hash-for-home memory.
+		 */
+		for_each_online_node(j) {
+			unsigned long start_pfn = node_start_pfn[j];
+			unsigned long end_pfn = node_end_pfn[j];
+			unsigned long nr_pages = end_pfn - start_pfn;
+
+			ret = gxio_trio_alloc_memory_maps(trio_context, 1, 0,
+							  0);
+			if (ret < 0) {
+				pr_err("PCI: Mem-Map alloc failure on TRIO %d "
+					"mac %d for MC %d, give up\n",
+					controller->trio_index,
+					controller->mac, j);
+
+				/* TBD: cleanup ... */
+
+				goto alloc_mem_map_failed;
+			}
+
+			controller->mem_maps[j] = ret;
+
+			/*
+			 * Initialize the Mem-Map and the I/O MMU so that all
+			 * the physical memory can be accessed by the endpoint
+			 * devices. The base bus address is set to the base CPA
+			 * of this memory controller, so is the base VA. The
+			 * I/O MMU table essentially translates the CPA to
+			 * the real PA.
+			 */
+			ret = gxio_trio_init_memory_map_mmu_aux(trio_context,
+				controller->mem_maps[j],
+				start_pfn << PAGE_SHIFT,
+				nr_pages << PAGE_SHIFT,
+				trio_context->asid,
+				controller->mac,
+				start_pfn << PAGE_SHIFT,
+				j,
+				GXIO_TRIO_ORDER_MODE_UNORDERED);
+			if (ret < 0) {
+				pr_err("PCI: Mem-Map init failure on TRIO %d "
+					"mac %d for MC %d, give up\n",
+					controller->trio_index,
+					controller->mac, j);
+
+				/* TBD: cleanup ... */
+
+				goto alloc_mem_map_failed;
+			}
+
+			continue;
+
+alloc_mem_map_failed:
+			break;
+		}
+
+	}
+
+	return 0;
+}
+subsys_initcall(pcibios_init);
+
+/*
+ * No bus fixups needed.
+ */
+void __devinit pcibios_fixup_bus(struct pci_bus *bus)
+{
+	/* Nothing needs to be done. */
+}
+
+/*
+ * This can be called from the generic PCI layer, but doesn't need to
+ * do anything.
+ */
+char __devinit *pcibios_setup(char *str)
+{
+	if (!strcmp(str, "off")) {
+		pci_probe = 0;
+		return NULL;
+	}
+	return str;
+}
+
+/*
+ * This is called from the generic Linux layer.
+ */
+void __devinit pcibios_update_irq(struct pci_dev *dev, int irq)
+{
+	pci_write_config_byte(dev, PCI_INTERRUPT_LINE, irq);
+}
+
+/*
+ * Enable memory address decoding, as appropriate, for the
+ * device described by the 'dev' struct. The I/O decoding
+ * is disabled, though the TILE-Gx supports I/O addressing.
+ *
+ * This is called from the generic PCI layer, and can be called
+ * for bridges or endpoints.
+ */
+int pcibios_enable_device(struct pci_dev *dev, int mask)
+{
+	return pci_enable_resources(dev, mask);
+}
+
+/* Map a PCI MMIO bus address into VA space. */
+void __iomem *ioremap(resource_size_t phys_addr, unsigned long size)
+{
+	struct pci_controller *controller = NULL;
+	resource_size_t bar_start;
+	resource_size_t bar_end;
+	resource_size_t offset;
+	resource_size_t start;
+	resource_size_t end;
+	int trio_fd;
+	int i, j;
+
+	start = phys_addr;
+	end = phys_addr + size - 1;
+
+	/*
+	 * In the following, each PCI controller's mem_resources[1]
+	 * represents its (non-prefetchable) PCI memory resource and
+	 * mem_resources[2] refers to its prefetchable PCI memory resource.
+	 * By searching phys_addr in each controller's mem_resources[], we can
+	 * determine the controller that should accept the PCI memory access.
+	 */
+
+	for (i = 0; i < num_rc_controllers; i++) {
+		/*
+		 * Skip controllers that are not properly initialized or
+		 * have down links.
+		 */
+		if (pci_controllers[i].root_bus == NULL)
+			continue;
+
+		for (j = 1; j < 3; j++) {
+			bar_start =
+				pci_controllers[i].mem_resources[j].start;
+			bar_end =
+				pci_controllers[i].mem_resources[j].end;
+
+			if ((start >= bar_start) && (end <= bar_end)) {
+
+				controller = &pci_controllers[i];
+
+				goto got_it;
+			}
+		}
+	}
+
+	if (controller == NULL)
+		return NULL;
+
+got_it:
+	trio_fd = controller->trio->fd;
+
+	offset = HV_TRIO_PIO_OFFSET(controller->pio_mem_index) + phys_addr;
+
+	/*
+	 * We need to keep the PCI bus address's in-page offset in the VA.
+	 */
+	return iorpc_ioremap(trio_fd, offset, size) +
+					(phys_addr & (PAGE_SIZE - 1));
+}
+EXPORT_SYMBOL(ioremap);
+
+void pci_iounmap(struct pci_dev *dev, void __iomem *addr)
+{
+	iounmap(addr);
+}
+EXPORT_SYMBOL(pci_iounmap);
+
+/****************************************************************
+ *
+ * Tile PCI config space read/write routines
+ *
+ ****************************************************************/
+
+/*
+ * These are the normal read and write ops
+ * These are expanded with macros from  pci_bus_read_config_byte() etc.
+ *
+ * devfn is the combined PCI device & function.
+ *
+ * offset is in bytes, from the start of config space for the
+ * specified bus & device.
+ */
+
+static int __devinit tile_cfg_read(struct pci_bus *bus,
+				   unsigned int devfn,
+				   int offset,
+				   int size,
+				   u32 *val)
+{
+	struct pci_controller *controller = bus->sysdata;
+	gxio_trio_context_t *trio_context = controller->trio;
+	int busnum = bus->number & 0xff;
+	int device = PCI_SLOT(devfn);
+	int function = PCI_FUNC(devfn);
+	int config_type = 1;
+	TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR_t cfg_addr;
+	void *mmio_addr;
+
+	/*
+	 * Map all accesses to the local device (bus == 0) into the
+	 * MMIO space of the MAC. Accesses to the downstream devices
+	 * go to the PIO space.
+	 */
+	if (busnum == 0) {
+		if (device == 0) {
+			/*
+			 * This is the internal downstream P2P bridge,
+			 * access directly.
+			 */
+			unsigned int reg_offset;
+
+			reg_offset = ((offset & 0xFFF) <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+				(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_PROTECTED
+				<< TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+				(controller->mac <<
+					TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+			mmio_addr = trio_context->mmio_base_mac + reg_offset;
+
+			goto valid_device;
+
+		} else {
+			/*
+			 * We fake an empty device for (device > 0),
+			 * since there is only one device on bus 0.
+			 */
+			goto invalid_device;
+		}
+	}
+
+	/*
+	 * Accesses to the directly attached device (bus == 1) have to be
+	 * sent as type-0 configs.
+	 */
+
+	if (busnum == 1) {
+		/*
+		 * There is only one device off of our built-in P2P bridge.
+		 */
+		if (device != 0)
+			goto invalid_device;
+
+		config_type = 0;
+	}
+
+	cfg_addr.word = 0;
+	cfg_addr.reg_addr = (offset & 0xFFF);
+	cfg_addr.fn = function;
+	cfg_addr.dev = device;
+	cfg_addr.bus = busnum;
+	cfg_addr.type = config_type;
+
+	/*
+	 * Note that we don't set the mac field in cfg_addr because the
+	 * mapping is per port.
+	 */
+
+	mmio_addr = trio_context->mmio_base_pio_cfg[controller->mac] +
+			cfg_addr.word;
+
+valid_device:
+
+	switch (size) {
+	case 4:
+		*val = __gxio_mmio_read32(mmio_addr);
+		break;
+
+	case 2:
+		*val = __gxio_mmio_read16(mmio_addr);
+		break;
+
+	case 1:
+		*val = __gxio_mmio_read8(mmio_addr);
+		break;
+
+	default:
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+	}
+
+	TRACE_CFG_RD(size, *val, busnum, device, function, offset);
+
+	return 0;
+
+invalid_device:
+
+	switch (size) {
+	case 4:
+		*val = 0xFFFFFFFF;
+		break;
+
+	case 2:
+		*val = 0xFFFF;
+		break;
+
+	case 1:
+		*val = 0xFF;
+		break;
+
+	default:
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+	}
+
+	return 0;
+}
+
+
+/*
+ * See tile_cfg_read() for relevent comments.
+ * Note that "val" is the value to write, not a pointer to that value.
+ */
+static int __devinit tile_cfg_write(struct pci_bus *bus,
+				    unsigned int devfn,
+				    int offset,
+				    int size,
+				    u32 val)
+{
+	struct pci_controller *controller = bus->sysdata;
+	gxio_trio_context_t *trio_context = controller->trio;
+	int busnum = bus->number & 0xff;
+	int device = PCI_SLOT(devfn);
+	int function = PCI_FUNC(devfn);
+	int config_type = 1;
+	TRIO_TILE_PIO_REGION_SETUP_CFG_ADDR_t cfg_addr;
+	void *mmio_addr;
+	u32 val_32 = (u32)val;
+	u16 val_16 = (u16)val;
+	u8 val_8 = (u8)val;
+
+	/*
+	 * Map all accesses to the local device (bus == 0) into the
+	 * MMIO space of the MAC. Accesses to the downstream devices
+	 * go to the PIO space.
+	 */
+	if (busnum == 0) {
+		if (device == 0) {
+			/*
+			 * This is the internal downstream P2P bridge,
+			 * access directly.
+			 */
+			unsigned int reg_offset;
+
+			reg_offset = ((offset & 0xFFF) <<
+				TRIO_CFG_REGION_ADDR__REG_SHIFT) |
+				(TRIO_CFG_REGION_ADDR__INTFC_VAL_MAC_PROTECTED
+				<< TRIO_CFG_REGION_ADDR__INTFC_SHIFT ) |
+				(controller->mac <<
+					TRIO_CFG_REGION_ADDR__MAC_SEL_SHIFT);
+
+			mmio_addr = trio_context->mmio_base_mac + reg_offset;
+
+			goto valid_device;
+
+		} else {
+			/*
+			 * We fake an empty device for (device > 0),
+			 * since there is only one device on bus 0.
+			 */
+			goto invalid_device;
+		}
+	}
+
+	/*
+	 * Accesses to the directly attached device (bus == 1) have to be
+	 * sent as type-0 configs.
+	 */
+
+	if (busnum == 1) {
+		/*
+		 * There is only one device off of our built-in P2P bridge.
+		 */
+		if (device != 0)
+			goto invalid_device;
+
+		config_type = 0;
+	}
+
+	cfg_addr.word = 0;
+	cfg_addr.reg_addr = (offset & 0xFFF);
+	cfg_addr.fn = function;
+	cfg_addr.dev = device;
+	cfg_addr.bus = busnum;
+	cfg_addr.type = config_type;
+
+	/*
+	 * Note that we don't set the mac field in cfg_addr because the
+	 * mapping is per port.
+	 */
+
+	mmio_addr = trio_context->mmio_base_pio_cfg[controller->mac] +
+			cfg_addr.word;
+
+valid_device:
+
+	switch (size) {
+	case 4:
+		__gxio_mmio_write32(mmio_addr, val_32);
+		TRACE_CFG_WR(size, val_32, busnum, device, function, offset);
+		break;
+
+	case 2:
+		__gxio_mmio_write16(mmio_addr, val_16);
+		TRACE_CFG_WR(size, val_16, busnum, device, function, offset);
+		break;
+
+	case 1:
+		__gxio_mmio_write8(mmio_addr, val_8);
+		TRACE_CFG_WR(size, val_8, busnum, device, function, offset);
+		break;
+
+	default:
+		return PCIBIOS_FUNC_NOT_SUPPORTED;
+	}
+
+invalid_device:
+
+	return 0;
+}
+
+
+static struct pci_ops tile_cfg_ops = {
+	.read =         tile_cfg_read,
+	.write =        tile_cfg_write,
+};
+
+
+/*
+ * MSI support starts here.
+ */
+static unsigned int
+tilegx_msi_startup(struct irq_data *d)
+{
+	if (d->msi_desc)
+		unmask_msi_irq(d);
+
+	return 0;
+}
+
+static void
+tilegx_msi_ack(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_EVENT_RESET_K, 1UL << d->irq);
+}
+
+static void
+tilegx_msi_mask(struct irq_data *d)
+{
+	mask_msi_irq(d);
+	__insn_mtspr(SPR_IPI_MASK_SET_K, 1UL << d->irq);
+}
+
+static void
+tilegx_msi_unmask(struct irq_data *d)
+{
+	__insn_mtspr(SPR_IPI_MASK_RESET_K, 1UL << d->irq);
+	unmask_msi_irq(d);
+}
+
+static struct irq_chip tilegx_msi_chip = {
+	.name			= "tilegx_msi",
+	.irq_startup		= tilegx_msi_startup,
+	.irq_ack		= tilegx_msi_ack,
+	.irq_mask		= tilegx_msi_mask,
+	.irq_unmask		= tilegx_msi_unmask,
+
+	/* TBD: support set_affinity. */
+};
+
+int arch_setup_msi_irq(struct pci_dev *pdev, struct msi_desc *desc)
+{
+	struct pci_controller *controller;
+	gxio_trio_context_t *trio_context;
+	struct msi_msg msg;
+	int default_irq;
+	uint64_t mem_map_base;
+	uint64_t mem_map_limit;
+	u64 msi_addr;
+	int mem_map;
+	int cpu;
+	int irq;
+	int ret;
+
+	irq = create_irq();
+	if (irq < 0)
+		return irq;
+
+	/*
+	 * Since we use a 64-bit Mem-Map to accept the MSI write, we fail
+	 * devices that are not capable of generating a 64-bit message address.
+	 * These devices will fall back to using the legacy interrupts.
+	 * Most PCIe endpoint devices do support 64-bit message addressing.
+	 */
+	if (desc->msi_attrib.is_64 == 0) {
+		dev_printk(KERN_INFO, &pdev->dev,
+			"64-bit MSI message address not supported, "
+			"falling back to legacy interrupts.\n");
+
+		ret = -ENOMEM;
+		goto is_64_failure;
+	}
+
+	default_irq = desc->msi_attrib.default_irq;
+	controller = irq_get_handler_data(default_irq);
+
+	BUG_ON(!controller);
+
+	trio_context = controller->trio;
+
+	/*
+	 * Allocate the Mem-Map that will accept the MSI write and
+	 * trigger the TILE-side interrupts.
+	 */
+	mem_map = gxio_trio_alloc_memory_maps(trio_context, 1, 0, 0);
+	if (mem_map < 0) {
+		dev_printk(KERN_INFO, &pdev->dev,
+			"%s Mem-Map alloc failure. "
+			"Failed to initialize MSI interrupts. "
+			"Falling back to legacy interrupts.\n",
+			desc->msi_attrib.is_msix ? "MSI-X" : "MSI");
+
+		ret = -ENOMEM;
+		goto msi_mem_map_alloc_failure;
+	}
+
+	/* We try to distribute different IRQs to different tiles. */
+	cpu = tile_irq_cpu(irq);
+
+	/*
+	 * Now call up to the HV to configure the Mem-Map interrupt and
+	 * set up the IPI binding.
+	 */
+	mem_map_base = MEM_MAP_INTR_REGIONS_BASE +
+		mem_map * MEM_MAP_INTR_REGION_SIZE;
+	mem_map_limit = mem_map_base + MEM_MAP_INTR_REGION_SIZE - 1;
+
+	ret = gxio_trio_config_msi_intr(trio_context, cpu_x(cpu), cpu_y(cpu),
+					KERNEL_PL, irq, controller->mac,
+					mem_map, mem_map_base, mem_map_limit,
+					trio_context->asid);
+	if (ret < 0) {
+		dev_printk(KERN_INFO, &pdev->dev, "HV MSI config failed.\n");
+
+		goto hv_msi_config_failure;
+	}
+
+	irq_set_msi_desc(irq, desc);
+
+	msi_addr = mem_map_base + TRIO_MAP_MEM_REG_INT3 - TRIO_MAP_MEM_REG_INT0;
+
+	msg.address_hi = msi_addr >> 32;
+	msg.address_lo = msi_addr & 0xffffffff;
+
+	msg.data = mem_map;
+
+	write_msi_msg(irq, &msg);
+	irq_set_chip_and_handler(irq, &tilegx_msi_chip, handle_level_irq);
+	irq_set_handler_data(irq, controller);
+
+	return 0;
+
+hv_msi_config_failure:
+	/* Free mem-map */
+msi_mem_map_alloc_failure:
+is_64_failure:
+	destroy_irq(irq);
+	return ret;
+}
+
+void arch_teardown_msi_irq(unsigned int irq)
+{
+	destroy_irq(irq);
+}

arch/tile/kernel/setup.c

@@ -1344,6 +1344,7 @@ void __init setup_arch(char **cmdline_p)
 
 
 #ifdef CONFIG_PCI
+#if !defined (__tilegx__)
 	/*
 	 * Initialize the PCI structures.  This is done before memory
 	 * setup so that we know whether or not a pci_reserve region
@@ -1351,6 +1352,7 @@ void __init setup_arch(char **cmdline_p)
 	 */
 	if (tile_pci_init() == 0)
 		pci_reserve_mb = 0;
+#endif
 
 	/* PCI systems reserve a region just below 4GB for mapping iomem. */
 	pci_reserve_end_pfn  = (1 << (32 - PAGE_SHIFT));
@@ -1379,6 +1381,10 @@ void __init setup_arch(char **cmdline_p)
 	setup_cpu(1);
 	setup_clock();
 	load_hv_initrd();
+
+#if defined(CONFIG_PCI) && defined (__tilegx__)
+	tile_pci_init();
+#endif
 }
 
 

arch/tile/mm/pgtable.c

@@ -575,13 +575,6 @@ void __iomem *ioremap_prot(resource_size_t phys_addr, unsigned long size,
 }
 EXPORT_SYMBOL(ioremap_prot);
 
-/* Map a PCI MMIO bus address into VA space. */
-void __iomem *ioremap(resource_size_t phys_addr, unsigned long size)
-{
-	panic("ioremap for PCI MMIO is not supported");
-}
-EXPORT_SYMBOL(ioremap);
-
 /* Unmap an MMIO VA mapping. */
 void iounmap(volatile void __iomem *addr_in)
 {

drivers/pci/quirks.c

@@ -2143,9 +2143,9 @@ DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_82865_HB,
 DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_82875_HB,
 			quirk_unhide_mch_dev6);
 
-#ifdef CONFIG_TILE
+#ifdef CONFIG_TILEPRO
 /*
- * The Tilera TILEmpower platform needs to set the link speed
+ * The Tilera TILEmpower tilepro platform needs to set the link speed
  * to 2.5GT(Giga-Transfers)/s (Gen 1). The default link speed
  * setting is 5GT/s (Gen 2). 0x98 is the Link Control2 PCIe
  * capability register of the PEX8624 PCIe switch. The switch
@@ -2160,7 +2160,7 @@ static void __devinit quirk_tile_plx_gen1(struct pci_dev *dev)
 	}
 }
 DECLARE_PCI_FIXUP_EARLY(PCI_VENDOR_ID_PLX, 0x8624, quirk_tile_plx_gen1);
-#endif /* CONFIG_TILE */
+#endif /* CONFIG_TILEPRO */
 
 #ifdef CONFIG_PCI_MSI
 /* Some chipsets do not support MSI. We cannot easily rely on setting