IB/ipath: chip initialisation code, and diag support
authorBryan O'Sullivan <bos@pathscale.com>
Wed, 29 Mar 2006 23:23:28 +0000 (15:23 -0800)
committerRoland Dreier <rolandd@cisco.com>
Fri, 31 Mar 2006 21:14:19 +0000 (13:14 -0800)
ipath_init_chip.c sets up an InfiniPath device for use.

ipath_diag.c permits userspace diagnostic tools to read and write a
chip's registers.  It is different in purpose from the mmap interfaces
to the /sys/bus/pci resource files.

Signed-off-by: Bryan O'Sullivan <bos@pathscale.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/ipath/ipath_diag.c [new file with mode: 0644]
drivers/infiniband/hw/ipath/ipath_init_chip.c [new file with mode: 0644]

diff --git a/drivers/infiniband/hw/ipath/ipath_diag.c b/drivers/infiniband/hw/ipath/ipath_diag.c
new file mode 100644 (file)
index 0000000..cd533cf
--- /dev/null
@@ -0,0 +1,379 @@
+/*
+ * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+/*
+ * This file contains support for diagnostic functions.  It is accessed by
+ * opening the ipath_diag device, normally minor number 129.  Diagnostic use
+ * of the InfiniPath chip may render the chip or board unusable until the
+ * driver is unloaded, or in some cases, until the system is rebooted.
+ *
+ * Accesses to the chip through this interface are not similar to going
+ * through the /sys/bus/pci resource mmap interface.
+ */
+
+#include <linux/pci.h>
+#include <asm/uaccess.h>
+
+#include "ipath_common.h"
+#include "ipath_kernel.h"
+#include "ips_common.h"
+#include "ipath_layer.h"
+
+int ipath_diag_inuse;
+static int diag_set_link;
+
+static int ipath_diag_open(struct inode *in, struct file *fp);
+static int ipath_diag_release(struct inode *in, struct file *fp);
+static ssize_t ipath_diag_read(struct file *fp, char __user *data,
+                              size_t count, loff_t *off);
+static ssize_t ipath_diag_write(struct file *fp, const char __user *data,
+                               size_t count, loff_t *off);
+
+static struct file_operations diag_file_ops = {
+       .owner = THIS_MODULE,
+       .write = ipath_diag_write,
+       .read = ipath_diag_read,
+       .open = ipath_diag_open,
+       .release = ipath_diag_release
+};
+
+static struct cdev *diag_cdev;
+static struct class_device *diag_class_dev;
+
+int ipath_diag_init(void)
+{
+       return ipath_cdev_init(IPATH_DIAG_MINOR, "ipath_diag",
+                              &diag_file_ops, &diag_cdev, &diag_class_dev);
+}
+
+void ipath_diag_cleanup(void)
+{
+       ipath_cdev_cleanup(&diag_cdev, &diag_class_dev);
+}
+
+/**
+ * ipath_read_umem64 - read a 64-bit quantity from the chip into user space
+ * @dd: the infinipath device
+ * @uaddr: the location to store the data in user memory
+ * @caddr: the source chip address (full pointer, not offset)
+ * @count: number of bytes to copy (multiple of 32 bits)
+ *
+ * This function also localizes all chip memory accesses.
+ * The copy should be written such that we read full cacheline packets
+ * from the chip.  This is usually used for a single qword
+ *
+ * NOTE:  This assumes the chip address is 64-bit aligned.
+ */
+static int ipath_read_umem64(struct ipath_devdata *dd, void __user *uaddr,
+                            const void __iomem *caddr, size_t count)
+{
+       const u64 __iomem *reg_addr = caddr;
+       const u64 __iomem *reg_end = reg_addr + (count / sizeof(u64));
+       int ret;
+
+       /* not very efficient, but it works for now */
+       if (reg_addr < dd->ipath_kregbase ||
+           reg_end > dd->ipath_kregend) {
+               ret = -EINVAL;
+               goto bail;
+       }
+       while (reg_addr < reg_end) {
+               u64 data = readq(reg_addr);
+               if (copy_to_user(uaddr, &data, sizeof(u64))) {
+                       ret = -EFAULT;
+                       goto bail;
+               }
+               reg_addr++;
+               uaddr++;
+       }
+       ret = 0;
+bail:
+       return ret;
+}
+
+/**
+ * ipath_write_umem64 - write a 64-bit quantity to the chip from user space
+ * @dd: the infinipath device
+ * @caddr: the destination chip address (full pointer, not offset)
+ * @uaddr: the source of the data in user memory
+ * @count: the number of bytes to copy (multiple of 32 bits)
+ *
+ * This is usually used for a single qword
+ * NOTE:  This assumes the chip address is 64-bit aligned.
+ */
+
+static int ipath_write_umem64(struct ipath_devdata *dd, void __iomem *caddr,
+                             const void __user *uaddr, size_t count)
+{
+       u64 __iomem *reg_addr = caddr;
+       const u64 __iomem *reg_end = reg_addr + (count / sizeof(u64));
+       int ret;
+
+       /* not very efficient, but it works for now */
+       if (reg_addr < dd->ipath_kregbase ||
+           reg_end > dd->ipath_kregend) {
+               ret = -EINVAL;
+               goto bail;
+       }
+       while (reg_addr < reg_end) {
+               u64 data;
+               if (copy_from_user(&data, uaddr, sizeof(data))) {
+                       ret = -EFAULT;
+                       goto bail;
+               }
+               writeq(data, reg_addr);
+
+               reg_addr++;
+               uaddr++;
+       }
+       ret = 0;
+bail:
+       return ret;
+}
+
+/**
+ * ipath_read_umem32 - read a 32-bit quantity from the chip into user space
+ * @dd: the infinipath device
+ * @uaddr: the location to store the data in user memory
+ * @caddr: the source chip address (full pointer, not offset)
+ * @count: number of bytes to copy
+ *
+ * read 32 bit values, not 64 bit; for memories that only
+ * support 32 bit reads; usually a single dword.
+ */
+static int ipath_read_umem32(struct ipath_devdata *dd, void __user *uaddr,
+                            const void __iomem *caddr, size_t count)
+{
+       const u32 __iomem *reg_addr = caddr;
+       const u32 __iomem *reg_end = reg_addr + (count / sizeof(u32));
+       int ret;
+
+       if (reg_addr < (u32 __iomem *) dd->ipath_kregbase ||
+           reg_end > (u32 __iomem *) dd->ipath_kregend) {
+               ret = -EINVAL;
+               goto bail;
+       }
+       /* not very efficient, but it works for now */
+       while (reg_addr < reg_end) {
+               u32 data = readl(reg_addr);
+               if (copy_to_user(uaddr, &data, sizeof(data))) {
+                       ret = -EFAULT;
+                       goto bail;
+               }
+
+               reg_addr++;
+               uaddr++;
+       }
+       ret = 0;
+bail:
+       return ret;
+}
+
+/**
+ * ipath_write_umem32 - write a 32-bit quantity to the chip from user space
+ * @dd: the infinipath device
+ * @caddr: the destination chip address (full pointer, not offset)
+ * @uaddr: the source of the data in user memory
+ * @count: number of bytes to copy
+ *
+ * write 32 bit values, not 64 bit; for memories that only
+ * support 32 bit write; usually a single dword.
+ */
+
+static int ipath_write_umem32(struct ipath_devdata *dd, void __iomem *caddr,
+                             const void __user *uaddr, size_t count)
+{
+       u32 __iomem *reg_addr = caddr;
+       const u32 __iomem *reg_end = reg_addr + (count / sizeof(u32));
+       int ret;
+
+       if (reg_addr < (u32 __iomem *) dd->ipath_kregbase ||
+           reg_end > (u32 __iomem *) dd->ipath_kregend) {
+               ret = -EINVAL;
+               goto bail;
+       }
+       while (reg_addr < reg_end) {
+               u32 data;
+               if (copy_from_user(&data, uaddr, sizeof(data))) {
+                       ret = -EFAULT;
+                       goto bail;
+               }
+               writel(data, reg_addr);
+
+               reg_addr++;
+               uaddr++;
+       }
+       ret = 0;
+bail:
+       return ret;
+}
+
+static int ipath_diag_open(struct inode *in, struct file *fp)
+{
+       struct ipath_devdata *dd;
+       int unit = 0; /* XXX this is bogus */
+       unsigned long flags;
+       int ret;
+
+       dd = ipath_lookup(unit);
+
+       mutex_lock(&ipath_mutex);
+       spin_lock_irqsave(&ipath_devs_lock, flags);
+
+       if (ipath_diag_inuse) {
+               ret = -EBUSY;
+               goto bail;
+       }
+
+       list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
+               /*
+                * we need at least one infinipath device to be present
+                * (don't use INITTED, because we want to be able to open
+                * even if device is in freeze mode, which cleared INITTED).
+                * There is a small amount of risk to this, which is why we
+                * also verify kregbase is set.
+                */
+
+               if (!(dd->ipath_flags & IPATH_PRESENT) ||
+                   !dd->ipath_kregbase)
+                       continue;
+
+               ipath_diag_inuse = 1;
+               diag_set_link = 0;
+               ret = 0;
+               goto bail;
+       }
+
+       ret = -ENODEV;
+
+bail:
+       spin_unlock_irqrestore(&ipath_devs_lock, flags);
+       mutex_unlock(&ipath_mutex);
+
+       /* Only expose a way to reset the device if we
+          make it into diag mode. */
+       if (ret == 0)
+               ipath_expose_reset(&dd->pcidev->dev);
+
+       return ret;
+}
+
+static int ipath_diag_release(struct inode *i, struct file *f)
+{
+       mutex_lock(&ipath_mutex);
+       ipath_diag_inuse = 0;
+       mutex_unlock(&ipath_mutex);
+       return 0;
+}
+
+static ssize_t ipath_diag_read(struct file *fp, char __user *data,
+                              size_t count, loff_t *off)
+{
+       int unit = 0; /* XXX provide for reads on other units some day */
+       struct ipath_devdata *dd;
+       void __iomem *kreg_base;
+       ssize_t ret;
+
+       dd = ipath_lookup(unit);
+       if (!dd) {
+               ret = -ENODEV;
+               goto bail;
+       }
+
+       kreg_base = dd->ipath_kregbase;
+
+       if (count == 0)
+               ret = 0;
+       else if ((count % 4) || (*off % 4))
+               /* address or length is not 32-bit aligned, hence invalid */
+               ret = -EINVAL;
+       else if ((count % 8) || (*off % 8))
+               /* address or length not 64-bit aligned; do 32-bit reads */
+               ret = ipath_read_umem32(dd, data, kreg_base + *off, count);
+       else
+               ret = ipath_read_umem64(dd, data, kreg_base + *off, count);
+
+       if (ret >= 0) {
+               *off += count;
+               ret = count;
+       }
+
+bail:
+       return ret;
+}
+
+static ssize_t ipath_diag_write(struct file *fp, const char __user *data,
+                               size_t count, loff_t *off)
+{
+       int unit = 0; /* XXX this is bogus */
+       struct ipath_devdata *dd;
+       void __iomem *kreg_base;
+       ssize_t ret;
+
+       dd = ipath_lookup(unit);
+       if (!dd) {
+               ret = -ENODEV;
+               goto bail;
+       }
+       kreg_base = dd->ipath_kregbase;
+
+       if (count == 0)
+               ret = 0;
+       else if ((count % 4) || (*off % 4))
+               /* address or length is not 32-bit aligned, hence invalid */
+               ret = -EINVAL;
+       else if ((count % 8) || (*off % 8))
+               /* address or length not 64-bit aligned; do 32-bit writes */
+               ret = ipath_write_umem32(dd, kreg_base + *off, data, count);
+       else
+               ret = ipath_write_umem64(dd, kreg_base + *off, data, count);
+
+       if (ret >= 0) {
+               *off += count;
+               ret = count;
+       }
+
+bail:
+       return ret;
+}
+
+void ipath_diag_bringup_link(struct ipath_devdata *dd)
+{
+       if (diag_set_link || (dd->ipath_flags & IPATH_LINKACTIVE))
+               return;
+
+       diag_set_link = 1;
+       ipath_cdbg(VERBOSE, "Trying to set to set link active for "
+                  "diag pkt\n");
+       ipath_layer_set_linkstate(dd, IPATH_IB_LINKARM);
+       ipath_layer_set_linkstate(dd, IPATH_IB_LINKACTIVE);
+}
diff --git a/drivers/infiniband/hw/ipath/ipath_init_chip.c b/drivers/infiniband/hw/ipath/ipath_init_chip.c
new file mode 100644 (file)
index 0000000..2823ff9
--- /dev/null
@@ -0,0 +1,951 @@
+/*
+ * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
+ *
+ * This software is available to you under a choice of one of two
+ * licenses.  You may choose to be licensed under the terms of the GNU
+ * General Public License (GPL) Version 2, available from the file
+ * COPYING in the main directory of this source tree, or the
+ * OpenIB.org BSD license below:
+ *
+ *     Redistribution and use in source and binary forms, with or
+ *     without modification, are permitted provided that the following
+ *     conditions are met:
+ *
+ *      - Redistributions of source code must retain the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer.
+ *
+ *      - Redistributions in binary form must reproduce the above
+ *        copyright notice, this list of conditions and the following
+ *        disclaimer in the documentation and/or other materials
+ *        provided with the distribution.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include <linux/pci.h>
+#include <linux/netdevice.h>
+#include <linux/vmalloc.h>
+
+#include "ipath_kernel.h"
+#include "ips_common.h"
+
+/*
+ * min buffers we want to have per port, after driver
+ */
+#define IPATH_MIN_USER_PORT_BUFCNT 8
+
+/*
+ * Number of ports we are configured to use (to allow for more pio
+ * buffers per port, etc.)  Zero means use chip value.
+ */
+static ushort ipath_cfgports;
+
+module_param_named(cfgports, ipath_cfgports, ushort, S_IRUGO);
+MODULE_PARM_DESC(cfgports, "Set max number of ports to use");
+
+/*
+ * Number of buffers reserved for driver (layered drivers and SMA
+ * send).  Reserved at end of buffer list.
+ */
+static ushort ipath_kpiobufs = 32;
+
+static int ipath_set_kpiobufs(const char *val, struct kernel_param *kp);
+
+module_param_call(kpiobufs, ipath_set_kpiobufs, param_get_uint,
+                 &ipath_kpiobufs, S_IWUSR | S_IRUGO);
+MODULE_PARM_DESC(kpiobufs, "Set number of PIO buffers for driver");
+
+/**
+ * create_port0_egr - allocate the eager TID buffers
+ * @dd: the infinipath device
+ *
+ * This code is now quite different for user and kernel, because
+ * the kernel uses skb's, for the accelerated network performance.
+ * This is the kernel (port0) version.
+ *
+ * Allocate the eager TID buffers and program them into infinipath.
+ * We use the network layer alloc_skb() allocator to allocate the
+ * memory, and either use the buffers as is for things like SMA
+ * packets, or pass the buffers up to the ipath layered driver and
+ * thence the network layer, replacing them as we do so (see
+ * ipath_rcv_layer()).
+ */
+static int create_port0_egr(struct ipath_devdata *dd)
+{
+       unsigned e, egrcnt;
+       struct sk_buff **skbs;
+       int ret;
+
+       egrcnt = dd->ipath_rcvegrcnt;
+
+       skbs = vmalloc(sizeof(*dd->ipath_port0_skbs) * egrcnt);
+       if (skbs == NULL) {
+               ipath_dev_err(dd, "allocation error for eager TID "
+                             "skb array\n");
+               ret = -ENOMEM;
+               goto bail;
+       }
+       for (e = 0; e < egrcnt; e++) {
+               /*
+                * This is a bit tricky in that we allocate extra
+                * space for 2 bytes of the 14 byte ethernet header.
+                * These two bytes are passed in the ipath header so
+                * the rest of the data is word aligned.  We allocate
+                * 4 bytes so that the data buffer stays word aligned.
+                * See ipath_kreceive() for more details.
+                */
+               skbs[e] = ipath_alloc_skb(dd, GFP_KERNEL);
+               if (!skbs[e]) {
+                       ipath_dev_err(dd, "SKB allocation error for "
+                                     "eager TID %u\n", e);
+                       while (e != 0)
+                               dev_kfree_skb(skbs[--e]);
+                       ret = -ENOMEM;
+                       goto bail;
+               }
+       }
+       /*
+        * After loop above, so we can test non-NULL to see if ready
+        * to use at receive, etc.
+        */
+       dd->ipath_port0_skbs = skbs;
+
+       for (e = 0; e < egrcnt; e++) {
+               unsigned long phys =
+                       virt_to_phys(dd->ipath_port0_skbs[e]->data);
+               dd->ipath_f_put_tid(dd, e + (u64 __iomem *)
+                                   ((char __iomem *) dd->ipath_kregbase +
+                                    dd->ipath_rcvegrbase), 0, phys);
+       }
+
+       ret = 0;
+
+bail:
+       return ret;
+}
+
+static int bringup_link(struct ipath_devdata *dd)
+{
+       u64 val, ibc;
+       int ret = 0;
+
+       /* hold IBC in reset */
+       dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
+                        dd->ipath_control);
+
+       /*
+        * Note that prior to try 14 or 15 of IB, the credit scaling
+        * wasn't working, because it was swapped for writes with the
+        * 1 bit default linkstate field
+        */
+
+       /* ignore pbc and align word */
+       val = dd->ipath_piosize2k - 2 * sizeof(u32);
+       /*
+        * for ICRC, which we only send in diag test pkt mode, and we
+        * don't need to worry about that for mtu
+        */
+       val += 1;
+       /*
+        * Set the IBC maxpktlength to the size of our pio buffers the
+        * maxpktlength is in words.  This is *not* the IB data MTU.
+        */
+       ibc = (val / sizeof(u32)) << INFINIPATH_IBCC_MAXPKTLEN_SHIFT;
+       /* in KB */
+       ibc |= 0x5ULL << INFINIPATH_IBCC_FLOWCTRLWATERMARK_SHIFT;
+       /*
+        * How often flowctrl sent.  More or less in usecs; balance against
+        * watermark value, so that in theory senders always get a flow
+        * control update in time to not let the IB link go idle.
+        */
+       ibc |= 0x3ULL << INFINIPATH_IBCC_FLOWCTRLPERIOD_SHIFT;
+       /* max error tolerance */
+       ibc |= 0xfULL << INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT;
+       /* use "real" buffer space for */
+       ibc |= 4ULL << INFINIPATH_IBCC_CREDITSCALE_SHIFT;
+       /* IB credit flow control. */
+       ibc |= 0xfULL << INFINIPATH_IBCC_OVERRUNTHRESHOLD_SHIFT;
+       /* initially come up waiting for TS1, without sending anything. */
+       dd->ipath_ibcctrl = ibc;
+       /*
+        * Want to start out with both LINKCMD and LINKINITCMD in NOP
+        * (0 and 0).  Don't put linkinitcmd in ipath_ibcctrl, want that
+        * to stay a NOP
+        */
+       ibc |= INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
+               INFINIPATH_IBCC_LINKINITCMD_SHIFT;
+       ipath_cdbg(VERBOSE, "Writing 0x%llx to ibcctrl\n",
+                  (unsigned long long) ibc);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl, ibc);
+
+       // be sure chip saw it
+       val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
+
+       ret = dd->ipath_f_bringup_serdes(dd);
+
+       if (ret)
+               dev_info(&dd->pcidev->dev, "Could not initialize SerDes, "
+                        "not usable\n");
+       else {
+               /* enable IBC */
+               dd->ipath_control |= INFINIPATH_C_LINKENABLE;
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
+                                dd->ipath_control);
+       }
+
+       return ret;
+}
+
+static int init_chip_first(struct ipath_devdata *dd,
+                          struct ipath_portdata **pdp)
+{
+       struct ipath_portdata *pd = NULL;
+       int ret = 0;
+       u64 val;
+
+       /*
+        * skip cfgports stuff because we are not allocating memory,
+        * and we don't want problems if the portcnt changed due to
+        * cfgports.  We do still check and report a difference, if
+        * not same (should be impossible).
+        */
+       dd->ipath_portcnt =
+               ipath_read_kreg32(dd, dd->ipath_kregs->kr_portcnt);
+       if (!ipath_cfgports)
+               dd->ipath_cfgports = dd->ipath_portcnt;
+       else if (ipath_cfgports <= dd->ipath_portcnt) {
+               dd->ipath_cfgports = ipath_cfgports;
+               ipath_dbg("Configured to use %u ports out of %u in chip\n",
+                         dd->ipath_cfgports, dd->ipath_portcnt);
+       } else {
+               dd->ipath_cfgports = dd->ipath_portcnt;
+               ipath_dbg("Tried to configured to use %u ports; chip "
+                         "only supports %u\n", ipath_cfgports,
+                         dd->ipath_portcnt);
+       }
+       dd->ipath_pd = kzalloc(sizeof(*dd->ipath_pd) * dd->ipath_cfgports,
+                              GFP_KERNEL);
+
+       if (!dd->ipath_pd) {
+               ipath_dev_err(dd, "Unable to allocate portdata array, "
+                             "failing\n");
+               ret = -ENOMEM;
+               goto done;
+       }
+
+       dd->ipath_lastegrheads = kzalloc(sizeof(*dd->ipath_lastegrheads)
+                                        * dd->ipath_cfgports,
+                                        GFP_KERNEL);
+       dd->ipath_lastrcvhdrqtails =
+               kzalloc(sizeof(*dd->ipath_lastrcvhdrqtails)
+                       * dd->ipath_cfgports, GFP_KERNEL);
+
+       if (!dd->ipath_lastegrheads || !dd->ipath_lastrcvhdrqtails) {
+               ipath_dev_err(dd, "Unable to allocate head arrays, "
+                             "failing\n");
+               ret = -ENOMEM;
+               goto done;
+       }
+
+       dd->ipath_pd[0] = kzalloc(sizeof(*pd), GFP_KERNEL);
+
+       if (!dd->ipath_pd[0]) {
+               ipath_dev_err(dd, "Unable to allocate portdata for port "
+                             "0, failing\n");
+               ret = -ENOMEM;
+               goto done;
+       }
+       pd = dd->ipath_pd[0];
+       pd->port_dd = dd;
+       pd->port_port = 0;
+       pd->port_cnt = 1;
+       /* The port 0 pkey table is used by the layer interface. */
+       pd->port_pkeys[0] = IPS_DEFAULT_P_KEY;
+       dd->ipath_rcvtidcnt =
+               ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
+       dd->ipath_rcvtidbase =
+               ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
+       dd->ipath_rcvegrcnt =
+               ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
+       dd->ipath_rcvegrbase =
+               ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
+       dd->ipath_palign =
+               ipath_read_kreg32(dd, dd->ipath_kregs->kr_pagealign);
+       dd->ipath_piobufbase =
+               ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufbase);
+       val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiosize);
+       dd->ipath_piosize2k = val & ~0U;
+       dd->ipath_piosize4k = val >> 32;
+       dd->ipath_ibmtu = 4096; /* default to largest legal MTU */
+       val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufcnt);
+       dd->ipath_piobcnt2k = val & ~0U;
+       dd->ipath_piobcnt4k = val >> 32;
+       dd->ipath_pio2kbase =
+               (u32 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
+                                (dd->ipath_piobufbase & 0xffffffff));
+       if (dd->ipath_piobcnt4k) {
+               dd->ipath_pio4kbase = (u32 __iomem *)
+                       (((char __iomem *) dd->ipath_kregbase) +
+                        (dd->ipath_piobufbase >> 32));
+               /*
+                * 4K buffers take 2 pages; we use roundup just to be
+                * paranoid; we calculate it once here, rather than on
+                * ever buf allocate
+                */
+               dd->ipath_4kalign = ALIGN(dd->ipath_piosize4k,
+                                         dd->ipath_palign);
+               ipath_dbg("%u 2k(%x) piobufs @ %p, %u 4k(%x) @ %p "
+                         "(%x aligned)\n",
+                         dd->ipath_piobcnt2k, dd->ipath_piosize2k,
+                         dd->ipath_pio2kbase, dd->ipath_piobcnt4k,
+                         dd->ipath_piosize4k, dd->ipath_pio4kbase,
+                         dd->ipath_4kalign);
+       }
+       else ipath_dbg("%u 2k piobufs @ %p\n",
+                      dd->ipath_piobcnt2k, dd->ipath_pio2kbase);
+
+       spin_lock_init(&dd->ipath_tid_lock);
+
+done:
+       *pdp = pd;
+       return ret;
+}
+
+/**
+ * init_chip_reset - re-initialize after a reset, or enable
+ * @dd: the infinipath device
+ * @pdp: output for port data
+ *
+ * sanity check at least some of the values after reset, and
+ * ensure no receive or transmit (explictly, in case reset
+ * failed
+ */
+static int init_chip_reset(struct ipath_devdata *dd,
+                          struct ipath_portdata **pdp)
+{
+       struct ipath_portdata *pd;
+       u32 rtmp;
+
+       *pdp = pd = dd->ipath_pd[0];
+       /* ensure chip does no sends or receives while we re-initialize */
+       dd->ipath_control = dd->ipath_sendctrl = dd->ipath_rcvctrl = 0U;
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, 0);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, 0);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0);
+
+       rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_portcnt);
+       if (dd->ipath_portcnt != rtmp)
+               dev_info(&dd->pcidev->dev, "portcnt was %u before "
+                        "reset, now %u, using original\n",
+                        dd->ipath_portcnt, rtmp);
+       rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
+       if (rtmp != dd->ipath_rcvtidcnt)
+               dev_info(&dd->pcidev->dev, "tidcnt was %u before "
+                        "reset, now %u, using original\n",
+                        dd->ipath_rcvtidcnt, rtmp);
+       rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
+       if (rtmp != dd->ipath_rcvtidbase)
+               dev_info(&dd->pcidev->dev, "tidbase was %u before "
+                        "reset, now %u, using original\n",
+                        dd->ipath_rcvtidbase, rtmp);
+       rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
+       if (rtmp != dd->ipath_rcvegrcnt)
+               dev_info(&dd->pcidev->dev, "egrcnt was %u before "
+                        "reset, now %u, using original\n",
+                        dd->ipath_rcvegrcnt, rtmp);
+       rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
+       if (rtmp != dd->ipath_rcvegrbase)
+               dev_info(&dd->pcidev->dev, "egrbase was %u before "
+                        "reset, now %u, using original\n",
+                        dd->ipath_rcvegrbase, rtmp);
+
+       return 0;
+}
+
+static int init_pioavailregs(struct ipath_devdata *dd)
+{
+       int ret;
+
+       dd->ipath_pioavailregs_dma = dma_alloc_coherent(
+               &dd->pcidev->dev, PAGE_SIZE, &dd->ipath_pioavailregs_phys,
+               GFP_KERNEL);
+       if (!dd->ipath_pioavailregs_dma) {
+               ipath_dev_err(dd, "failed to allocate PIOavail reg area "
+                             "in memory\n");
+               ret = -ENOMEM;
+               goto done;
+       }
+
+       /*
+        * we really want L2 cache aligned, but for current CPUs of
+        * interest, they are the same.
+        */
+       dd->ipath_statusp = (u64 *)
+               ((char *)dd->ipath_pioavailregs_dma +
+                ((2 * L1_CACHE_BYTES +
+                  dd->ipath_pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES));
+       /* copy the current value now that it's really allocated */
+       *dd->ipath_statusp = dd->_ipath_status;
+       /*
+        * setup buffer to hold freeze msg, accessible to apps,
+        * following statusp
+        */
+       dd->ipath_freezemsg = (char *)&dd->ipath_statusp[1];
+       /* and its length */
+       dd->ipath_freezelen = L1_CACHE_BYTES - sizeof(dd->ipath_statusp[0]);
+
+       if (dd->ipath_unit * 64 > (IPATH_PORT0_RCVHDRTAIL_SIZE - 64)) {
+               ipath_dev_err(dd, "unit %u too large for port 0 "
+                             "rcvhdrtail buffer size\n", dd->ipath_unit);
+               ret = -ENODEV;
+       }
+       else
+               ret = 0;
+
+       /* so we can get current tail in ipath_kreceive(), per chip */
+       dd->ipath_hdrqtailptr = &ipath_port0_rcvhdrtail[
+               dd->ipath_unit * (64 / sizeof(*ipath_port0_rcvhdrtail))];
+done:
+       return ret;
+}
+
+/**
+ * init_shadow_tids - allocate the shadow TID array
+ * @dd: the infinipath device
+ *
+ * allocate the shadow TID array, so we can ipath_munlock previous
+ * entries.  It may make more sense to move the pageshadow to the
+ * port data structure, so we only allocate memory for ports actually
+ * in use, since we at 8k per port, now.
+ */
+static void init_shadow_tids(struct ipath_devdata *dd)
+{
+       dd->ipath_pageshadow = (struct page **)
+               vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
+                       sizeof(struct page *));
+       if (!dd->ipath_pageshadow)
+               ipath_dev_err(dd, "failed to allocate shadow page * "
+                             "array, no expected sends!\n");
+       else
+               memset(dd->ipath_pageshadow, 0,
+                      dd->ipath_cfgports * dd->ipath_rcvtidcnt *
+                      sizeof(struct page *));
+}
+
+static void enable_chip(struct ipath_devdata *dd,
+                       struct ipath_portdata *pd, int reinit)
+{
+       u32 val;
+       int i;
+
+       if (!reinit) {
+               init_waitqueue_head(&ipath_sma_state_wait);
+       }
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
+                        dd->ipath_rcvctrl);
+
+       /* Enable PIO send, and update of PIOavail regs to memory. */
+       dd->ipath_sendctrl = INFINIPATH_S_PIOENABLE |
+               INFINIPATH_S_PIOBUFAVAILUPD;
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
+                        dd->ipath_sendctrl);
+
+       /*
+        * enable port 0 receive, and receive interrupt.  other ports
+        * done as user opens and inits them.
+        */
+       dd->ipath_rcvctrl = INFINIPATH_R_TAILUPD |
+               (1ULL << INFINIPATH_R_PORTENABLE_SHIFT) |
+               (1ULL << INFINIPATH_R_INTRAVAIL_SHIFT);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
+                        dd->ipath_rcvctrl);
+
+       /*
+        * now ready for use.  this should be cleared whenever we
+        * detect a reset, or initiate one.
+        */
+       dd->ipath_flags |= IPATH_INITTED;
+
+       /*
+        * init our shadow copies of head from tail values, and write
+        * head values to match.
+        */
+       val = ipath_read_ureg32(dd, ur_rcvegrindextail, 0);
+       (void)ipath_write_ureg(dd, ur_rcvegrindexhead, val, 0);
+       dd->ipath_port0head = ipath_read_ureg32(dd, ur_rcvhdrtail, 0);
+
+       /* Initialize so we interrupt on next packet received */
+       (void)ipath_write_ureg(dd, ur_rcvhdrhead,
+                              dd->ipath_rhdrhead_intr_off |
+                              dd->ipath_port0head, 0);
+
+       /*
+        * by now pioavail updates to memory should have occurred, so
+        * copy them into our working/shadow registers; this is in
+        * case something went wrong with abort, but mostly to get the
+        * initial values of the generation bit correct.
+        */
+       for (i = 0; i < dd->ipath_pioavregs; i++) {
+               __le64 val;
+
+               /*
+                * Chip Errata bug 6641; even and odd qwords>3 are swapped.
+                */
+               if (i > 3) {
+                       if (i & 1)
+                               val = dd->ipath_pioavailregs_dma[i - 1];
+                       else
+                               val = dd->ipath_pioavailregs_dma[i + 1];
+               }
+               else
+                       val = dd->ipath_pioavailregs_dma[i];
+               dd->ipath_pioavailshadow[i] = le64_to_cpu(val);
+       }
+       /* can get counters, stats, etc. */
+       dd->ipath_flags |= IPATH_PRESENT;
+}
+
+static int init_housekeeping(struct ipath_devdata *dd,
+                            struct ipath_portdata **pdp, int reinit)
+{
+       char boardn[32];
+       int ret = 0;
+
+       /*
+        * have to clear shadow copies of registers at init that are
+        * not otherwise set here, or all kinds of bizarre things
+        * happen with driver on chip reset
+        */
+       dd->ipath_rcvhdrsize = 0;
+
+       /*
+        * Don't clear ipath_flags as 8bit mode was set before
+        * entering this func. However, we do set the linkstate to
+        * unknown, so we can watch for a transition.
+        */
+       dd->ipath_flags |= IPATH_LINKUNK;
+       dd->ipath_flags &= ~(IPATH_LINKACTIVE | IPATH_LINKARMED |
+                            IPATH_LINKDOWN | IPATH_LINKINIT);
+
+       ipath_cdbg(VERBOSE, "Try to read spc chip revision\n");
+       dd->ipath_revision =
+               ipath_read_kreg64(dd, dd->ipath_kregs->kr_revision);
+
+       /*
+        * set up fundamental info we need to use the chip; we assume
+        * if the revision reg and these regs are OK, we don't need to
+        * special case the rest
+        */
+       dd->ipath_sregbase =
+               ipath_read_kreg32(dd, dd->ipath_kregs->kr_sendregbase);
+       dd->ipath_cregbase =
+               ipath_read_kreg32(dd, dd->ipath_kregs->kr_counterregbase);
+       dd->ipath_uregbase =
+               ipath_read_kreg32(dd, dd->ipath_kregs->kr_userregbase);
+       ipath_cdbg(VERBOSE, "ipath_kregbase %p, sendbase %x usrbase %x, "
+                  "cntrbase %x\n", dd->ipath_kregbase, dd->ipath_sregbase,
+                  dd->ipath_uregbase, dd->ipath_cregbase);
+       if ((dd->ipath_revision & 0xffffffff) == 0xffffffff
+           || (dd->ipath_sregbase & 0xffffffff) == 0xffffffff
+           || (dd->ipath_cregbase & 0xffffffff) == 0xffffffff
+           || (dd->ipath_uregbase & 0xffffffff) == 0xffffffff) {
+               ipath_dev_err(dd, "Register read failures from chip, "
+                             "giving up initialization\n");
+               ret = -ENODEV;
+               goto done;
+       }
+
+       /* clear the initial reset flag, in case first driver load */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
+                        INFINIPATH_E_RESET);
+
+       if (reinit)
+               ret = init_chip_reset(dd, pdp);
+       else
+               ret = init_chip_first(dd, pdp);
+
+       if (ret)
+               goto done;
+
+       ipath_cdbg(VERBOSE, "Revision %llx (PCI %x), %u ports, %u tids, "
+                  "%u egrtids\n", (unsigned long long) dd->ipath_revision,
+                  dd->ipath_pcirev, dd->ipath_portcnt, dd->ipath_rcvtidcnt,
+                  dd->ipath_rcvegrcnt);
+
+       if (((dd->ipath_revision >> INFINIPATH_R_SOFTWARE_SHIFT) &
+            INFINIPATH_R_SOFTWARE_MASK) != IPATH_CHIP_SWVERSION) {
+               ipath_dev_err(dd, "Driver only handles version %d, "
+                             "chip swversion is %d (%llx), failng\n",
+                             IPATH_CHIP_SWVERSION,
+                             (int)(dd->ipath_revision >>
+                                   INFINIPATH_R_SOFTWARE_SHIFT) &
+                             INFINIPATH_R_SOFTWARE_MASK,
+                             (unsigned long long) dd->ipath_revision);
+               ret = -ENOSYS;
+               goto done;
+       }
+       dd->ipath_majrev = (u8) ((dd->ipath_revision >>
+                                 INFINIPATH_R_CHIPREVMAJOR_SHIFT) &
+                                INFINIPATH_R_CHIPREVMAJOR_MASK);
+       dd->ipath_minrev = (u8) ((dd->ipath_revision >>
+                                 INFINIPATH_R_CHIPREVMINOR_SHIFT) &
+                                INFINIPATH_R_CHIPREVMINOR_MASK);
+       dd->ipath_boardrev = (u8) ((dd->ipath_revision >>
+                                   INFINIPATH_R_BOARDID_SHIFT) &
+                                  INFINIPATH_R_BOARDID_MASK);
+
+       ret = dd->ipath_f_get_boardname(dd, boardn, sizeof boardn);
+
+       snprintf(dd->ipath_boardversion, sizeof(dd->ipath_boardversion),
+                "Driver %u.%u, %s, InfiniPath%u %u.%u, PCI %u, "
+                "SW Compat %u\n",
+                IPATH_CHIP_VERS_MAJ, IPATH_CHIP_VERS_MIN, boardn,
+                (unsigned)(dd->ipath_revision >> INFINIPATH_R_ARCH_SHIFT) &
+                INFINIPATH_R_ARCH_MASK,
+                dd->ipath_majrev, dd->ipath_minrev, dd->ipath_pcirev,
+                (unsigned)(dd->ipath_revision >>
+                           INFINIPATH_R_SOFTWARE_SHIFT) &
+                INFINIPATH_R_SOFTWARE_MASK);
+
+       ipath_dbg("%s", dd->ipath_boardversion);
+
+done:
+       return ret;
+}
+
+
+/**
+ * ipath_init_chip - do the actual initialization sequence on the chip
+ * @dd: the infinipath device
+ * @reinit: reinitializing, so don't allocate new memory
+ *
+ * Do the actual initialization sequence on the chip.  This is done
+ * both from the init routine called from the PCI infrastructure, and
+ * when we reset the chip, or detect that it was reset internally,
+ * or it's administratively re-enabled.
+ *
+ * Memory allocation here and in called routines is only done in
+ * the first case (reinit == 0).  We have to be careful, because even
+ * without memory allocation, we need to re-write all the chip registers
+ * TIDs, etc. after the reset or enable has completed.
+ */
+int ipath_init_chip(struct ipath_devdata *dd, int reinit)
+{
+       int ret = 0, i;
+       u32 val32, kpiobufs;
+       u64 val, atmp;
+       struct ipath_portdata *pd = NULL; /* keep gcc4 happy */
+
+       ret = init_housekeeping(dd, &pd, reinit);
+       if (ret)
+               goto done;
+
+       /*
+        * we ignore most issues after reporting them, but have to specially
+        * handle hardware-disabled chips.
+        */
+       if (ret == 2) {
+               /* unique error, known to ipath_init_one */
+               ret = -EPERM;
+               goto done;
+       }
+
+       /*
+        * We could bump this to allow for full rcvegrcnt + rcvtidcnt,
+        * but then it no longer nicely fits power of two, and since
+        * we now use routines that backend onto __get_free_pages, the
+        * rest would be wasted.
+        */
+       dd->ipath_rcvhdrcnt = dd->ipath_rcvegrcnt;
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrcnt,
+                        dd->ipath_rcvhdrcnt);
+
+       /*
+        * Set up the shadow copies of the piobufavail registers,
+        * which we compare against the chip registers for now, and
+        * the in memory DMA'ed copies of the registers.  This has to
+        * be done early, before we calculate lastport, etc.
+        */
+       val = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
+       /*
+        * calc number of pioavail registers, and save it; we have 2
+        * bits per buffer.
+        */
+       dd->ipath_pioavregs = ALIGN(val, sizeof(u64) * BITS_PER_BYTE / 2)
+               / (sizeof(u64) * BITS_PER_BYTE / 2);
+       if (!ipath_kpiobufs)    /* have to have at least 1, for SMA */
+               kpiobufs = ipath_kpiobufs = 1;
+       else if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) <
+                (dd->ipath_cfgports * IPATH_MIN_USER_PORT_BUFCNT)) {
+               dev_info(&dd->pcidev->dev, "Too few PIO buffers (%u) "
+                        "for %u ports to have %u each!\n",
+                        dd->ipath_piobcnt2k + dd->ipath_piobcnt4k,
+                        dd->ipath_cfgports, IPATH_MIN_USER_PORT_BUFCNT);
+               kpiobufs = 1;   /* reserve just the minimum for SMA/ether */
+       } else
+               kpiobufs = ipath_kpiobufs;
+
+       if (kpiobufs >
+           (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
+            (dd->ipath_cfgports * IPATH_MIN_USER_PORT_BUFCNT))) {
+               i = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
+                       (dd->ipath_cfgports * IPATH_MIN_USER_PORT_BUFCNT);
+               if (i < 0)
+                       i = 0;
+               dev_info(&dd->pcidev->dev, "Allocating %d PIO bufs for "
+                        "kernel leaves too few for %d user ports "
+                        "(%d each); using %u\n", kpiobufs,
+                        dd->ipath_cfgports - 1,
+                        IPATH_MIN_USER_PORT_BUFCNT, i);
+               /*
+                * shouldn't change ipath_kpiobufs, because could be
+                * different for different devices...
+                */
+               kpiobufs = i;
+       }
+       dd->ipath_lastport_piobuf =
+               dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - kpiobufs;
+       dd->ipath_pbufsport = dd->ipath_cfgports > 1
+               ? dd->ipath_lastport_piobuf / (dd->ipath_cfgports - 1)
+               : 0;
+       val32 = dd->ipath_lastport_piobuf -
+               (dd->ipath_pbufsport * (dd->ipath_cfgports - 1));
+       if (val32 > 0) {
+               ipath_dbg("allocating %u pbufs/port leaves %u unused, "
+                         "add to kernel\n", dd->ipath_pbufsport, val32);
+               dd->ipath_lastport_piobuf -= val32;
+               ipath_dbg("%u pbufs/port leaves %u unused, add to kernel\n",
+                         dd->ipath_pbufsport, val32);
+       }
+       dd->ipath_lastpioindex = dd->ipath_lastport_piobuf;
+       ipath_cdbg(VERBOSE, "%d PIO bufs for kernel out of %d total %u "
+                  "each for %u user ports\n", kpiobufs,
+                  dd->ipath_piobcnt2k + dd->ipath_piobcnt4k,
+                  dd->ipath_pbufsport, dd->ipath_cfgports - 1);
+
+       dd->ipath_f_early_init(dd);
+
+       /* early_init sets rcvhdrentsize and rcvhdrsize, so this must be
+        * done after early_init */
+       dd->ipath_hdrqlast =
+               dd->ipath_rcvhdrentsize * (dd->ipath_rcvhdrcnt - 1);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrentsize,
+                        dd->ipath_rcvhdrentsize);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
+                        dd->ipath_rcvhdrsize);
+
+       if (!reinit) {
+               ret = init_pioavailregs(dd);
+               init_shadow_tids(dd);
+               if (ret)
+                       goto done;
+       }
+
+       (void)ipath_write_kreg(dd, dd->ipath_kregs->kr_sendpioavailaddr,
+                              dd->ipath_pioavailregs_phys);
+       /*
+        * this is to detect s/w errors, which the h/w works around by
+        * ignoring the low 6 bits of address, if it wasn't aligned.
+        */
+       val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpioavailaddr);
+       if (val != dd->ipath_pioavailregs_phys) {
+               ipath_dev_err(dd, "Catastrophic software error, "
+                             "SendPIOAvailAddr written as %lx, "
+                             "read back as %llx\n",
+                             (unsigned long) dd->ipath_pioavailregs_phys,
+                             (unsigned long long) val);
+               ret = -EINVAL;
+               goto done;
+       }
+
+       val = ipath_port0_rcvhdrtail_dma + dd->ipath_unit * 64;
+
+       /* verify that the alignment requirement was met */
+       ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
+                             0, val);
+       atmp = ipath_read_kreg64_port(
+               dd, dd->ipath_kregs->kr_rcvhdrtailaddr, 0);
+       if (val != atmp) {
+               ipath_dev_err(dd, "Catastrophic software error, "
+                             "RcvHdrTailAddr0 written as %llx, "
+                             "read back as %llx from %x\n",
+                             (unsigned long long) val,
+                             (unsigned long long) atmp,
+                             dd->ipath_kregs->kr_rcvhdrtailaddr);
+               ret = -EINVAL;
+               goto done;
+       }
+
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvbthqp, IPATH_KD_QP);
+
+       /*
+        * make sure we are not in freeze, and PIO send enabled, so
+        * writes to pbc happen
+        */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask, 0ULL);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
+                        ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0ULL);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
+                        INFINIPATH_S_PIOENABLE);
+
+       /*
+        * before error clears, since we expect serdes pll errors during
+        * this, the first time after reset
+        */
+       if (bringup_link(dd)) {
+               dev_info(&dd->pcidev->dev, "Failed to bringup IB link\n");
+               ret = -ENETDOWN;
+               goto done;
+       }
+
+       /*
+        * clear any "expected" hwerrs from reset and/or initialization
+        * clear any that aren't enabled (at least this once), and then
+        * set the enable mask
+        */
+       dd->ipath_f_init_hwerrors(dd);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
+                        ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
+                        dd->ipath_hwerrmask);
+
+       dd->ipath_maskederrs = dd->ipath_ignorederrs;
+       /* clear all */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
+       /* enable errors that are masked, at least this first time. */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
+                        ~dd->ipath_maskederrs);
+       /* clear any interrups up to this point (ints still not enabled) */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
+
+       ipath_stats.sps_lid[dd->ipath_unit] = dd->ipath_lid;
+
+       /*
+        * Set up the port 0 (kernel) rcvhdr q and egr TIDs.  If doing
+        * re-init, the simplest way to handle this is to free
+        * existing, and re-allocate.
+        */
+       if (reinit)
+               ipath_free_pddata(dd, 0, 0);
+       dd->ipath_f_tidtemplate(dd);
+       ret = ipath_create_rcvhdrq(dd, pd);
+       if (!ret)
+               ret = create_port0_egr(dd);
+       if (ret)
+               ipath_dev_err(dd, "failed to allocate port 0 (kernel) "
+                             "rcvhdrq and/or egr bufs\n");
+       else
+               enable_chip(dd, pd, reinit);
+
+       /*
+        * cause retrigger of pending interrupts ignored during init,
+        * even if we had errors
+        */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
+
+       if(!dd->ipath_stats_timer_active) {
+               /*
+                * first init, or after an admin disable/enable
+                * set up stats retrieval timer, even if we had errors
+                * in last portion of setup
+                */
+               init_timer(&dd->ipath_stats_timer);
+               dd->ipath_stats_timer.function = ipath_get_faststats;
+               dd->ipath_stats_timer.data = (unsigned long) dd;
+               /* every 5 seconds; */
+               dd->ipath_stats_timer.expires = jiffies + 5 * HZ;
+               /* takes ~16 seconds to overflow at full IB 4x bandwdith */
+               add_timer(&dd->ipath_stats_timer);
+               dd->ipath_stats_timer_active = 1;
+       }
+
+done:
+       if (!ret) {
+               ipath_get_guid(dd);
+               *dd->ipath_statusp |= IPATH_STATUS_CHIP_PRESENT;
+               if (!dd->ipath_f_intrsetup(dd)) {
+                       /* now we can enable all interrupts from the chip */
+                       ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
+                                        -1LL);
+                       /* force re-interrupt of any pending interrupts. */
+                       ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear,
+                                        0ULL);
+                       /* chip is usable; mark it as initialized */
+                       *dd->ipath_statusp |= IPATH_STATUS_INITTED;
+               } else
+                       ipath_dev_err(dd, "No interrupts enabled, couldn't "
+                                     "setup interrupt address\n");
+
+               if (dd->ipath_cfgports > ipath_stats.sps_nports)
+                       /*
+                        * sps_nports is a global, so, we set it to
+                        * the highest number of ports of any of the
+                        * chips we find; we never decrement it, at
+                        * least for now.  Since this might have changed
+                        * over disable/enable or prior to reset, always
+                        * do the check and potentially adjust.
+                        */
+                       ipath_stats.sps_nports = dd->ipath_cfgports;
+       } else
+               ipath_dbg("Failed (%d) to initialize chip\n", ret);
+
+       /* if ret is non-zero, we probably should do some cleanup
+          here... */
+       return ret;
+}
+
+static int ipath_set_kpiobufs(const char *str, struct kernel_param *kp)
+{
+       struct ipath_devdata *dd;
+       unsigned long flags;
+       unsigned short val;
+       int ret;
+
+       ret = ipath_parse_ushort(str, &val);
+
+       spin_lock_irqsave(&ipath_devs_lock, flags);
+
+       if (ret < 0)
+               goto bail;
+
+       if (val == 0) {
+               ret = -EINVAL;
+               goto bail;
+       }
+
+       list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
+               if (dd->ipath_kregbase)
+                       continue;
+               if (val > (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
+                          (dd->ipath_cfgports *
+                           IPATH_MIN_USER_PORT_BUFCNT)))
+               {
+                       ipath_dev_err(
+                               dd,
+                               "Allocating %d PIO bufs for kernel leaves "
+                               "too few for %d user ports (%d each)\n",
+                               val, dd->ipath_cfgports - 1,
+                               IPATH_MIN_USER_PORT_BUFCNT);
+                       ret = -EINVAL;
+                       goto bail;
+               }
+               dd->ipath_lastport_piobuf =
+                       dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - val;
+       }
+
+       ret = 0;
+bail:
+       spin_unlock_irqrestore(&ipath_devs_lock, flags);
+
+       return ret;
+}