This commit was manufactured by cvs2svn to create branch 'vserver'.
[linux-2.6.git] / drivers / infiniband / hw / ipath / ipath_iba6120.c
diff --git a/drivers/infiniband/hw/ipath/ipath_iba6120.c b/drivers/infiniband/hw/ipath/ipath_iba6120.c
new file mode 100644 (file)
index 0000000..ae8bf99
--- /dev/null
@@ -0,0 +1,1373 @@
+/*
+ * Copyright (c) 2006 QLogic, Inc. All rights reserved.
+ * 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 all of the code that is specific to the
+ * InfiniPath PCIe chip.
+ */
+
+#include <linux/interrupt.h>
+#include <linux/pci.h>
+#include <linux/delay.h>
+
+
+#include "ipath_kernel.h"
+#include "ipath_registers.h"
+
+/*
+ * This file contains all the chip-specific register information and
+ * access functions for the QLogic InfiniPath PCI-Express chip.
+ *
+ * This lists the InfiniPath registers, in the actual chip layout.
+ * This structure should never be directly accessed.
+ */
+struct _infinipath_do_not_use_kernel_regs {
+       unsigned long long Revision;
+       unsigned long long Control;
+       unsigned long long PageAlign;
+       unsigned long long PortCnt;
+       unsigned long long DebugPortSelect;
+       unsigned long long Reserved0;
+       unsigned long long SendRegBase;
+       unsigned long long UserRegBase;
+       unsigned long long CounterRegBase;
+       unsigned long long Scratch;
+       unsigned long long Reserved1;
+       unsigned long long Reserved2;
+       unsigned long long IntBlocked;
+       unsigned long long IntMask;
+       unsigned long long IntStatus;
+       unsigned long long IntClear;
+       unsigned long long ErrorMask;
+       unsigned long long ErrorStatus;
+       unsigned long long ErrorClear;
+       unsigned long long HwErrMask;
+       unsigned long long HwErrStatus;
+       unsigned long long HwErrClear;
+       unsigned long long HwDiagCtrl;
+       unsigned long long MDIO;
+       unsigned long long IBCStatus;
+       unsigned long long IBCCtrl;
+       unsigned long long ExtStatus;
+       unsigned long long ExtCtrl;
+       unsigned long long GPIOOut;
+       unsigned long long GPIOMask;
+       unsigned long long GPIOStatus;
+       unsigned long long GPIOClear;
+       unsigned long long RcvCtrl;
+       unsigned long long RcvBTHQP;
+       unsigned long long RcvHdrSize;
+       unsigned long long RcvHdrCnt;
+       unsigned long long RcvHdrEntSize;
+       unsigned long long RcvTIDBase;
+       unsigned long long RcvTIDCnt;
+       unsigned long long RcvEgrBase;
+       unsigned long long RcvEgrCnt;
+       unsigned long long RcvBufBase;
+       unsigned long long RcvBufSize;
+       unsigned long long RxIntMemBase;
+       unsigned long long RxIntMemSize;
+       unsigned long long RcvPartitionKey;
+       unsigned long long Reserved3;
+       unsigned long long RcvPktLEDCnt;
+       unsigned long long Reserved4[8];
+       unsigned long long SendCtrl;
+       unsigned long long SendPIOBufBase;
+       unsigned long long SendPIOSize;
+       unsigned long long SendPIOBufCnt;
+       unsigned long long SendPIOAvailAddr;
+       unsigned long long TxIntMemBase;
+       unsigned long long TxIntMemSize;
+       unsigned long long Reserved5;
+       unsigned long long PCIeRBufTestReg0;
+       unsigned long long PCIeRBufTestReg1;
+       unsigned long long Reserved51[6];
+       unsigned long long SendBufferError;
+       unsigned long long SendBufferErrorCONT1;
+       unsigned long long Reserved6SBE[6];
+       unsigned long long RcvHdrAddr0;
+       unsigned long long RcvHdrAddr1;
+       unsigned long long RcvHdrAddr2;
+       unsigned long long RcvHdrAddr3;
+       unsigned long long RcvHdrAddr4;
+       unsigned long long Reserved7RHA[11];
+       unsigned long long RcvHdrTailAddr0;
+       unsigned long long RcvHdrTailAddr1;
+       unsigned long long RcvHdrTailAddr2;
+       unsigned long long RcvHdrTailAddr3;
+       unsigned long long RcvHdrTailAddr4;
+       unsigned long long Reserved8RHTA[11];
+       unsigned long long Reserved9SW[8];
+       unsigned long long SerdesConfig0;
+       unsigned long long SerdesConfig1;
+       unsigned long long SerdesStatus;
+       unsigned long long XGXSConfig;
+       unsigned long long IBPLLCfg;
+       unsigned long long Reserved10SW2[3];
+       unsigned long long PCIEQ0SerdesConfig0;
+       unsigned long long PCIEQ0SerdesConfig1;
+       unsigned long long PCIEQ0SerdesStatus;
+       unsigned long long Reserved11;
+       unsigned long long PCIEQ1SerdesConfig0;
+       unsigned long long PCIEQ1SerdesConfig1;
+       unsigned long long PCIEQ1SerdesStatus;
+       unsigned long long Reserved12;
+};
+
+#define IPATH_KREG_OFFSET(field) (offsetof(struct \
+    _infinipath_do_not_use_kernel_regs, field) / sizeof(u64))
+#define IPATH_CREG_OFFSET(field) (offsetof( \
+    struct infinipath_counters, field) / sizeof(u64))
+
+static const struct ipath_kregs ipath_pe_kregs = {
+       .kr_control = IPATH_KREG_OFFSET(Control),
+       .kr_counterregbase = IPATH_KREG_OFFSET(CounterRegBase),
+       .kr_debugportselect = IPATH_KREG_OFFSET(DebugPortSelect),
+       .kr_errorclear = IPATH_KREG_OFFSET(ErrorClear),
+       .kr_errormask = IPATH_KREG_OFFSET(ErrorMask),
+       .kr_errorstatus = IPATH_KREG_OFFSET(ErrorStatus),
+       .kr_extctrl = IPATH_KREG_OFFSET(ExtCtrl),
+       .kr_extstatus = IPATH_KREG_OFFSET(ExtStatus),
+       .kr_gpio_clear = IPATH_KREG_OFFSET(GPIOClear),
+       .kr_gpio_mask = IPATH_KREG_OFFSET(GPIOMask),
+       .kr_gpio_out = IPATH_KREG_OFFSET(GPIOOut),
+       .kr_gpio_status = IPATH_KREG_OFFSET(GPIOStatus),
+       .kr_hwdiagctrl = IPATH_KREG_OFFSET(HwDiagCtrl),
+       .kr_hwerrclear = IPATH_KREG_OFFSET(HwErrClear),
+       .kr_hwerrmask = IPATH_KREG_OFFSET(HwErrMask),
+       .kr_hwerrstatus = IPATH_KREG_OFFSET(HwErrStatus),
+       .kr_ibcctrl = IPATH_KREG_OFFSET(IBCCtrl),
+       .kr_ibcstatus = IPATH_KREG_OFFSET(IBCStatus),
+       .kr_intblocked = IPATH_KREG_OFFSET(IntBlocked),
+       .kr_intclear = IPATH_KREG_OFFSET(IntClear),
+       .kr_intmask = IPATH_KREG_OFFSET(IntMask),
+       .kr_intstatus = IPATH_KREG_OFFSET(IntStatus),
+       .kr_mdio = IPATH_KREG_OFFSET(MDIO),
+       .kr_pagealign = IPATH_KREG_OFFSET(PageAlign),
+       .kr_partitionkey = IPATH_KREG_OFFSET(RcvPartitionKey),
+       .kr_portcnt = IPATH_KREG_OFFSET(PortCnt),
+       .kr_rcvbthqp = IPATH_KREG_OFFSET(RcvBTHQP),
+       .kr_rcvbufbase = IPATH_KREG_OFFSET(RcvBufBase),
+       .kr_rcvbufsize = IPATH_KREG_OFFSET(RcvBufSize),
+       .kr_rcvctrl = IPATH_KREG_OFFSET(RcvCtrl),
+       .kr_rcvegrbase = IPATH_KREG_OFFSET(RcvEgrBase),
+       .kr_rcvegrcnt = IPATH_KREG_OFFSET(RcvEgrCnt),
+       .kr_rcvhdrcnt = IPATH_KREG_OFFSET(RcvHdrCnt),
+       .kr_rcvhdrentsize = IPATH_KREG_OFFSET(RcvHdrEntSize),
+       .kr_rcvhdrsize = IPATH_KREG_OFFSET(RcvHdrSize),
+       .kr_rcvintmembase = IPATH_KREG_OFFSET(RxIntMemBase),
+       .kr_rcvintmemsize = IPATH_KREG_OFFSET(RxIntMemSize),
+       .kr_rcvtidbase = IPATH_KREG_OFFSET(RcvTIDBase),
+       .kr_rcvtidcnt = IPATH_KREG_OFFSET(RcvTIDCnt),
+       .kr_revision = IPATH_KREG_OFFSET(Revision),
+       .kr_scratch = IPATH_KREG_OFFSET(Scratch),
+       .kr_sendbuffererror = IPATH_KREG_OFFSET(SendBufferError),
+       .kr_sendctrl = IPATH_KREG_OFFSET(SendCtrl),
+       .kr_sendpioavailaddr = IPATH_KREG_OFFSET(SendPIOAvailAddr),
+       .kr_sendpiobufbase = IPATH_KREG_OFFSET(SendPIOBufBase),
+       .kr_sendpiobufcnt = IPATH_KREG_OFFSET(SendPIOBufCnt),
+       .kr_sendpiosize = IPATH_KREG_OFFSET(SendPIOSize),
+       .kr_sendregbase = IPATH_KREG_OFFSET(SendRegBase),
+       .kr_txintmembase = IPATH_KREG_OFFSET(TxIntMemBase),
+       .kr_txintmemsize = IPATH_KREG_OFFSET(TxIntMemSize),
+       .kr_userregbase = IPATH_KREG_OFFSET(UserRegBase),
+       .kr_serdesconfig0 = IPATH_KREG_OFFSET(SerdesConfig0),
+       .kr_serdesconfig1 = IPATH_KREG_OFFSET(SerdesConfig1),
+       .kr_serdesstatus = IPATH_KREG_OFFSET(SerdesStatus),
+       .kr_xgxsconfig = IPATH_KREG_OFFSET(XGXSConfig),
+       .kr_ibpllcfg = IPATH_KREG_OFFSET(IBPLLCfg),
+
+       /*
+        * These should not be used directly via ipath_read_kreg64(),
+        * use them with ipath_read_kreg64_port()
+        */
+       .kr_rcvhdraddr = IPATH_KREG_OFFSET(RcvHdrAddr0),
+       .kr_rcvhdrtailaddr = IPATH_KREG_OFFSET(RcvHdrTailAddr0),
+
+       /* The rcvpktled register controls one of the debug port signals, so
+        * a packet activity LED can be connected to it. */
+       .kr_rcvpktledcnt = IPATH_KREG_OFFSET(RcvPktLEDCnt),
+       .kr_pcierbuftestreg0 = IPATH_KREG_OFFSET(PCIeRBufTestReg0),
+       .kr_pcierbuftestreg1 = IPATH_KREG_OFFSET(PCIeRBufTestReg1),
+       .kr_pcieq0serdesconfig0 = IPATH_KREG_OFFSET(PCIEQ0SerdesConfig0),
+       .kr_pcieq0serdesconfig1 = IPATH_KREG_OFFSET(PCIEQ0SerdesConfig1),
+       .kr_pcieq0serdesstatus = IPATH_KREG_OFFSET(PCIEQ0SerdesStatus),
+       .kr_pcieq1serdesconfig0 = IPATH_KREG_OFFSET(PCIEQ1SerdesConfig0),
+       .kr_pcieq1serdesconfig1 = IPATH_KREG_OFFSET(PCIEQ1SerdesConfig1),
+       .kr_pcieq1serdesstatus = IPATH_KREG_OFFSET(PCIEQ1SerdesStatus)
+};
+
+static const struct ipath_cregs ipath_pe_cregs = {
+       .cr_badformatcnt = IPATH_CREG_OFFSET(RxBadFormatCnt),
+       .cr_erricrccnt = IPATH_CREG_OFFSET(RxICRCErrCnt),
+       .cr_errlinkcnt = IPATH_CREG_OFFSET(RxLinkProblemCnt),
+       .cr_errlpcrccnt = IPATH_CREG_OFFSET(RxLPCRCErrCnt),
+       .cr_errpkey = IPATH_CREG_OFFSET(RxPKeyMismatchCnt),
+       .cr_errrcvflowctrlcnt = IPATH_CREG_OFFSET(RxFlowCtrlErrCnt),
+       .cr_err_rlencnt = IPATH_CREG_OFFSET(RxLenErrCnt),
+       .cr_errslencnt = IPATH_CREG_OFFSET(TxLenErrCnt),
+       .cr_errtidfull = IPATH_CREG_OFFSET(RxTIDFullErrCnt),
+       .cr_errtidvalid = IPATH_CREG_OFFSET(RxTIDValidErrCnt),
+       .cr_errvcrccnt = IPATH_CREG_OFFSET(RxVCRCErrCnt),
+       .cr_ibstatuschange = IPATH_CREG_OFFSET(IBStatusChangeCnt),
+       .cr_intcnt = IPATH_CREG_OFFSET(LBIntCnt),
+       .cr_invalidrlencnt = IPATH_CREG_OFFSET(RxMaxMinLenErrCnt),
+       .cr_invalidslencnt = IPATH_CREG_OFFSET(TxMaxMinLenErrCnt),
+       .cr_lbflowstallcnt = IPATH_CREG_OFFSET(LBFlowStallCnt),
+       .cr_pktrcvcnt = IPATH_CREG_OFFSET(RxDataPktCnt),
+       .cr_pktrcvflowctrlcnt = IPATH_CREG_OFFSET(RxFlowPktCnt),
+       .cr_pktsendcnt = IPATH_CREG_OFFSET(TxDataPktCnt),
+       .cr_pktsendflowcnt = IPATH_CREG_OFFSET(TxFlowPktCnt),
+       .cr_portovflcnt = IPATH_CREG_OFFSET(RxP0HdrEgrOvflCnt),
+       .cr_rcvebpcnt = IPATH_CREG_OFFSET(RxEBPCnt),
+       .cr_rcvovflcnt = IPATH_CREG_OFFSET(RxBufOvflCnt),
+       .cr_senddropped = IPATH_CREG_OFFSET(TxDroppedPktCnt),
+       .cr_sendstallcnt = IPATH_CREG_OFFSET(TxFlowStallCnt),
+       .cr_sendunderruncnt = IPATH_CREG_OFFSET(TxUnderrunCnt),
+       .cr_wordrcvcnt = IPATH_CREG_OFFSET(RxDwordCnt),
+       .cr_wordsendcnt = IPATH_CREG_OFFSET(TxDwordCnt),
+       .cr_unsupvlcnt = IPATH_CREG_OFFSET(TxUnsupVLErrCnt),
+       .cr_rxdroppktcnt = IPATH_CREG_OFFSET(RxDroppedPktCnt),
+       .cr_iblinkerrrecovcnt = IPATH_CREG_OFFSET(IBLinkErrRecoveryCnt),
+       .cr_iblinkdowncnt = IPATH_CREG_OFFSET(IBLinkDownedCnt),
+       .cr_ibsymbolerrcnt = IPATH_CREG_OFFSET(IBSymbolErrCnt)
+};
+
+/* kr_intstatus, kr_intclear, kr_intmask bits */
+#define INFINIPATH_I_RCVURG_MASK ((1U<<5)-1)
+#define INFINIPATH_I_RCVAVAIL_MASK ((1U<<5)-1)
+
+/* kr_hwerrclear, kr_hwerrmask, kr_hwerrstatus, bits */
+#define INFINIPATH_HWE_PCIEMEMPARITYERR_MASK  0x000000000000003fULL
+#define INFINIPATH_HWE_PCIEMEMPARITYERR_SHIFT 0
+#define INFINIPATH_HWE_PCIEPOISONEDTLP      0x0000000010000000ULL
+#define INFINIPATH_HWE_PCIECPLTIMEOUT       0x0000000020000000ULL
+#define INFINIPATH_HWE_PCIEBUSPARITYXTLH    0x0000000040000000ULL
+#define INFINIPATH_HWE_PCIEBUSPARITYXADM    0x0000000080000000ULL
+#define INFINIPATH_HWE_PCIEBUSPARITYRADM    0x0000000100000000ULL
+#define INFINIPATH_HWE_COREPLL_FBSLIP       0x0080000000000000ULL
+#define INFINIPATH_HWE_COREPLL_RFSLIP       0x0100000000000000ULL
+#define INFINIPATH_HWE_PCIE1PLLFAILED       0x0400000000000000ULL
+#define INFINIPATH_HWE_PCIE0PLLFAILED       0x0800000000000000ULL
+#define INFINIPATH_HWE_SERDESPLLFAILED      0x1000000000000000ULL
+
+/* kr_extstatus bits */
+#define INFINIPATH_EXTS_FREQSEL 0x2
+#define INFINIPATH_EXTS_SERDESSEL 0x4
+#define INFINIPATH_EXTS_MEMBIST_ENDTEST     0x0000000000004000
+#define INFINIPATH_EXTS_MEMBIST_FOUND       0x0000000000008000
+
+#define _IPATH_GPIO_SDA_NUM 1
+#define _IPATH_GPIO_SCL_NUM 0
+
+#define IPATH_GPIO_SDA (1ULL << \
+       (_IPATH_GPIO_SDA_NUM+INFINIPATH_EXTC_GPIOOE_SHIFT))
+#define IPATH_GPIO_SCL (1ULL << \
+       (_IPATH_GPIO_SCL_NUM+INFINIPATH_EXTC_GPIOOE_SHIFT))
+
+/*
+ * Rev2 silicon allows suppressing check for ArmLaunch errors.
+ * this can speed up short packet sends on systems that do
+ * not guaranteee write-order.
+ */
+#define INFINIPATH_XGXS_SUPPRESS_ARMLAUNCH_ERR (1ULL<<63)
+
+/* 6120 specific hardware errors... */
+static const struct ipath_hwerror_msgs ipath_6120_hwerror_msgs[] = {
+       INFINIPATH_HWE_MSG(PCIEPOISONEDTLP, "PCIe Poisoned TLP"),
+       INFINIPATH_HWE_MSG(PCIECPLTIMEOUT, "PCIe completion timeout"),
+       /*
+        * In practice, it's unlikely wthat we'll see PCIe PLL, or bus
+        * parity or memory parity error failures, because most likely we
+        * won't be able to talk to the core of the chip.  Nonetheless, we
+        * might see them, if they are in parts of the PCIe core that aren't
+        * essential.
+        */
+       INFINIPATH_HWE_MSG(PCIE1PLLFAILED, "PCIePLL1"),
+       INFINIPATH_HWE_MSG(PCIE0PLLFAILED, "PCIePLL0"),
+       INFINIPATH_HWE_MSG(PCIEBUSPARITYXTLH, "PCIe XTLH core parity"),
+       INFINIPATH_HWE_MSG(PCIEBUSPARITYXADM, "PCIe ADM TX core parity"),
+       INFINIPATH_HWE_MSG(PCIEBUSPARITYRADM, "PCIe ADM RX core parity"),
+       INFINIPATH_HWE_MSG(RXDSYNCMEMPARITYERR, "Rx Dsync"),
+       INFINIPATH_HWE_MSG(SERDESPLLFAILED, "SerDes PLL"),
+};
+
+/**
+ * ipath_pe_handle_hwerrors - display hardware errors.
+ * @dd: the infinipath device
+ * @msg: the output buffer
+ * @msgl: the size of the output buffer
+ *
+ * Use same msg buffer as regular errors to avoid excessive stack
+ * use.  Most hardware errors are catastrophic, but for right now,
+ * we'll print them and continue.  We reuse the same message buffer as
+ * ipath_handle_errors() to avoid excessive stack usage.
+ */
+static void ipath_pe_handle_hwerrors(struct ipath_devdata *dd, char *msg,
+                                    size_t msgl)
+{
+       ipath_err_t hwerrs;
+       u32 bits, ctrl;
+       int isfatal = 0;
+       char bitsmsg[64];
+
+       hwerrs = ipath_read_kreg64(dd, dd->ipath_kregs->kr_hwerrstatus);
+       if (!hwerrs) {
+               /*
+                * better than printing cofusing messages
+                * This seems to be related to clearing the crc error, or
+                * the pll error during init.
+                */
+               ipath_cdbg(VERBOSE, "Called but no hardware errors set\n");
+               return;
+       } else if (hwerrs == ~0ULL) {
+               ipath_dev_err(dd, "Read of hardware error status failed "
+                             "(all bits set); ignoring\n");
+               return;
+       }
+       ipath_stats.sps_hwerrs++;
+
+       /* Always clear the error status register, except MEMBISTFAIL,
+        * regardless of whether we continue or stop using the chip.
+        * We want that set so we know it failed, even across driver reload.
+        * We'll still ignore it in the hwerrmask.  We do this partly for
+        * diagnostics, but also for support */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
+                        hwerrs&~INFINIPATH_HWE_MEMBISTFAILED);
+
+       hwerrs &= dd->ipath_hwerrmask;
+
+       /*
+        * make sure we get this much out, unless told to be quiet,
+        * or it's occurred within the last 5 seconds
+        */
+       if ((hwerrs & ~(dd->ipath_lasthwerror |
+                       ((INFINIPATH_HWE_TXEMEMPARITYERR_PIOBUF |
+                         INFINIPATH_HWE_TXEMEMPARITYERR_PIOPBC)
+                        << INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT))) ||
+           (ipath_debug & __IPATH_VERBDBG))
+               dev_info(&dd->pcidev->dev, "Hardware error: hwerr=0x%llx "
+                        "(cleared)\n", (unsigned long long) hwerrs);
+       dd->ipath_lasthwerror |= hwerrs;
+
+       if (hwerrs & ~dd->ipath_hwe_bitsextant)
+               ipath_dev_err(dd, "hwerror interrupt with unknown errors "
+                             "%llx set\n", (unsigned long long)
+                             (hwerrs & ~dd->ipath_hwe_bitsextant));
+
+       ctrl = ipath_read_kreg32(dd, dd->ipath_kregs->kr_control);
+       if (ctrl & INFINIPATH_C_FREEZEMODE) {
+               /*
+                * parity errors in send memory are recoverable,
+                * just cancel the send (if indicated in * sendbuffererror),
+                * count the occurrence, unfreeze (if no other handled
+                * hardware error bits are set), and continue. They can
+                * occur if a processor speculative read is done to the PIO
+                * buffer while we are sending a packet, for example.
+                */
+               if (hwerrs & ((INFINIPATH_HWE_TXEMEMPARITYERR_PIOBUF |
+                              INFINIPATH_HWE_TXEMEMPARITYERR_PIOPBC)
+                             << INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT)) {
+                       ipath_stats.sps_txeparity++;
+                       ipath_dbg("Recovering from TXE parity error (%llu), "
+                                 "hwerrstatus=%llx\n",
+                                 (unsigned long long) ipath_stats.sps_txeparity,
+                                 (unsigned long long) hwerrs);
+                       ipath_disarm_senderrbufs(dd);
+                       hwerrs &= ~((INFINIPATH_HWE_TXEMEMPARITYERR_PIOBUF |
+                                    INFINIPATH_HWE_TXEMEMPARITYERR_PIOPBC)
+                                   << INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT);
+                       if (!hwerrs) { /* else leave in freeze mode */
+                               ipath_write_kreg(dd,
+                                                dd->ipath_kregs->kr_control,
+                                                dd->ipath_control);
+                           return;
+                       }
+               }
+               if (hwerrs) {
+                       /*
+                        * if any set that we aren't ignoring only make the
+                        * complaint once, in case it's stuck or recurring,
+                        * and we get here multiple times
+                        */
+                       if (dd->ipath_flags & IPATH_INITTED) {
+                               ipath_dev_err(dd, "Fatal Hardware Error (freeze "
+                                             "mode), no longer usable, SN %.16s\n",
+                                                 dd->ipath_serial);
+                               isfatal = 1;
+                       }
+                       /*
+                        * Mark as having had an error for driver, and also
+                        * for /sys and status word mapped to user programs.
+                        * This marks unit as not usable, until reset
+                        */
+                       *dd->ipath_statusp &= ~IPATH_STATUS_IB_READY;
+                       *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
+                       dd->ipath_flags &= ~IPATH_INITTED;
+               } else {
+                       ipath_dbg("Clearing freezemode on ignored hardware "
+                                 "error\n");
+                       ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
+                                        dd->ipath_control);
+               }
+       }
+
+       *msg = '\0';
+
+       if (hwerrs & INFINIPATH_HWE_MEMBISTFAILED) {
+               strlcat(msg, "[Memory BIST test failed, InfiniPath hardware unusable]",
+                       msgl);
+               /* ignore from now on, so disable until driver reloaded */
+               *dd->ipath_statusp |= IPATH_STATUS_HWERROR;
+               dd->ipath_hwerrmask &= ~INFINIPATH_HWE_MEMBISTFAILED;
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
+                                dd->ipath_hwerrmask);
+       }
+
+       ipath_format_hwerrors(hwerrs,
+                             ipath_6120_hwerror_msgs,
+                             sizeof(ipath_6120_hwerror_msgs)/
+                             sizeof(ipath_6120_hwerror_msgs[0]),
+                             msg, msgl);
+
+       if (hwerrs & (INFINIPATH_HWE_PCIEMEMPARITYERR_MASK
+                     << INFINIPATH_HWE_PCIEMEMPARITYERR_SHIFT)) {
+               bits = (u32) ((hwerrs >>
+                              INFINIPATH_HWE_PCIEMEMPARITYERR_SHIFT) &
+                             INFINIPATH_HWE_PCIEMEMPARITYERR_MASK);
+               snprintf(bitsmsg, sizeof bitsmsg,
+                        "[PCIe Mem Parity Errs %x] ", bits);
+               strlcat(msg, bitsmsg, msgl);
+       }
+
+#define _IPATH_PLL_FAIL (INFINIPATH_HWE_COREPLL_FBSLIP |       \
+                        INFINIPATH_HWE_COREPLL_RFSLIP )
+
+       if (hwerrs & _IPATH_PLL_FAIL) {
+               snprintf(bitsmsg, sizeof bitsmsg,
+                        "[PLL failed (%llx), InfiniPath hardware unusable]",
+                        (unsigned long long) hwerrs & _IPATH_PLL_FAIL);
+               strlcat(msg, bitsmsg, msgl);
+               /* ignore from now on, so disable until driver reloaded */
+               dd->ipath_hwerrmask &= ~(hwerrs & _IPATH_PLL_FAIL);
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
+                                dd->ipath_hwerrmask);
+       }
+
+       if (hwerrs & INFINIPATH_HWE_SERDESPLLFAILED) {
+               /*
+                * If it occurs, it is left masked since the eternal
+                * interface is unused
+                */
+               dd->ipath_hwerrmask &= ~INFINIPATH_HWE_SERDESPLLFAILED;
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
+                                dd->ipath_hwerrmask);
+       }
+
+       ipath_dev_err(dd, "%s hardware error\n", msg);
+       if (isfatal && !ipath_diag_inuse && dd->ipath_freezemsg) {
+               /*
+                * for /sys status file ; if no trailing } is copied, we'll
+                * know it was truncated.
+                */
+               snprintf(dd->ipath_freezemsg, dd->ipath_freezelen,
+                        "{%s}", msg);
+       }
+}
+
+/**
+ * ipath_pe_boardname - fill in the board name
+ * @dd: the infinipath device
+ * @name: the output buffer
+ * @namelen: the size of the output buffer
+ *
+ * info is based on the board revision register
+ */
+static int ipath_pe_boardname(struct ipath_devdata *dd, char *name,
+                             size_t namelen)
+{
+       char *n = NULL;
+       u8 boardrev = dd->ipath_boardrev;
+       int ret;
+
+       switch (boardrev) {
+       case 0:
+               n = "InfiniPath_Emulation";
+               break;
+       case 1:
+               n = "InfiniPath_QLE7140-Bringup";
+               break;
+       case 2:
+               n = "InfiniPath_QLE7140";
+               break;
+       case 3:
+               n = "InfiniPath_QMI7140";
+               break;
+       case 4:
+               n = "InfiniPath_QEM7140";
+               break;
+       case 5:
+               n = "InfiniPath_QMH7140";
+               break;
+       case 6:
+               n = "InfiniPath_QLE7142";
+               break;
+       default:
+               ipath_dev_err(dd,
+                             "Don't yet know about board with ID %u\n",
+                             boardrev);
+               snprintf(name, namelen, "Unknown_InfiniPath_PCIe_%u",
+                        boardrev);
+               break;
+       }
+       if (n)
+               snprintf(name, namelen, "%s", n);
+
+       if (dd->ipath_majrev != 4 || !dd->ipath_minrev || dd->ipath_minrev>2) {
+               ipath_dev_err(dd, "Unsupported InfiniPath hardware revision %u.%u!\n",
+                             dd->ipath_majrev, dd->ipath_minrev);
+               ret = 1;
+       } else
+               ret = 0;
+
+       return ret;
+}
+
+/**
+ * ipath_pe_init_hwerrors - enable hardware errors
+ * @dd: the infinipath device
+ *
+ * now that we have finished initializing everything that might reasonably
+ * cause a hardware error, and cleared those errors bits as they occur,
+ * we can enable hardware errors in the mask (potentially enabling
+ * freeze mode), and enable hardware errors as errors (along with
+ * everything else) in errormask
+ */
+static void ipath_pe_init_hwerrors(struct ipath_devdata *dd)
+{
+       ipath_err_t val;
+       u64 extsval;
+
+       extsval = ipath_read_kreg64(dd, dd->ipath_kregs->kr_extstatus);
+
+       if (!(extsval & INFINIPATH_EXTS_MEMBIST_ENDTEST))
+               ipath_dev_err(dd, "MemBIST did not complete!\n");
+
+       val = ~0ULL;    /* barring bugs, all hwerrors become interrupts, */
+
+       if (!dd->ipath_boardrev)        // no PLL for Emulator
+               val &= ~INFINIPATH_HWE_SERDESPLLFAILED;
+
+       if (dd->ipath_minrev < 2) {
+               /* workaround bug 9460 in internal interface bus parity
+                * checking. Fixed (HW bug 9490) in Rev2.
+                */
+               val &= ~INFINIPATH_HWE_PCIEBUSPARITYRADM;
+       }
+       dd->ipath_hwerrmask = val;
+}
+
+/**
+ * ipath_pe_bringup_serdes - bring up the serdes
+ * @dd: the infinipath device
+ */
+static int ipath_pe_bringup_serdes(struct ipath_devdata *dd)
+{
+       u64 val, config1, prev_val;
+       int ret = 0;
+
+       ipath_dbg("Trying to bringup serdes\n");
+
+       if (ipath_read_kreg64(dd, dd->ipath_kregs->kr_hwerrstatus) &
+           INFINIPATH_HWE_SERDESPLLFAILED) {
+               ipath_dbg("At start, serdes PLL failed bit set "
+                         "in hwerrstatus, clearing and continuing\n");
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
+                                INFINIPATH_HWE_SERDESPLLFAILED);
+       }
+
+       val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_serdesconfig0);
+       config1 = ipath_read_kreg64(dd, dd->ipath_kregs->kr_serdesconfig1);
+
+       ipath_cdbg(VERBOSE, "SerDes status config0=%llx config1=%llx, "
+                  "xgxsconfig %llx\n", (unsigned long long) val,
+                  (unsigned long long) config1, (unsigned long long)
+                  ipath_read_kreg64(dd, dd->ipath_kregs->kr_xgxsconfig));
+
+       /*
+        * Force reset on, also set rxdetect enable.  Must do before reading
+        * serdesstatus at least for simulation, or some of the bits in
+        * serdes status will come back as undefined and cause simulation
+        * failures
+        */
+       val |= INFINIPATH_SERDC0_RESET_PLL | INFINIPATH_SERDC0_RXDETECT_EN
+               | INFINIPATH_SERDC0_L1PWR_DN;
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_serdesconfig0, val);
+       /* be sure chip saw it */
+       ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
+       udelay(5);              /* need pll reset set at least for a bit */
+       /*
+        * after PLL is reset, set the per-lane Resets and TxIdle and
+        * clear the PLL reset and rxdetect (to get falling edge).
+        * Leave L1PWR bits set (permanently)
+        */
+       val &= ~(INFINIPATH_SERDC0_RXDETECT_EN | INFINIPATH_SERDC0_RESET_PLL
+                | INFINIPATH_SERDC0_L1PWR_DN);
+       val |= INFINIPATH_SERDC0_RESET_MASK | INFINIPATH_SERDC0_TXIDLE;
+       ipath_cdbg(VERBOSE, "Clearing pll reset and setting lane resets "
+                  "and txidle (%llx)\n", (unsigned long long) val);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_serdesconfig0, val);
+       /* be sure chip saw it */
+       ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
+       /* need PLL reset clear for at least 11 usec before lane
+        * resets cleared; give it a few more to be sure */
+       udelay(15);
+       val &= ~(INFINIPATH_SERDC0_RESET_MASK | INFINIPATH_SERDC0_TXIDLE);
+
+       ipath_cdbg(VERBOSE, "Clearing lane resets and txidle "
+                  "(writing %llx)\n", (unsigned long long) val);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_serdesconfig0, val);
+       /* be sure chip saw it */
+       val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
+
+       val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_xgxsconfig);
+       prev_val = val;
+       if (((val >> INFINIPATH_XGXS_MDIOADDR_SHIFT) &
+            INFINIPATH_XGXS_MDIOADDR_MASK) != 3) {
+               val &=
+                       ~(INFINIPATH_XGXS_MDIOADDR_MASK <<
+                         INFINIPATH_XGXS_MDIOADDR_SHIFT);
+               /* MDIO address 3 */
+               val |= 3ULL << INFINIPATH_XGXS_MDIOADDR_SHIFT;
+       }
+       if (val & INFINIPATH_XGXS_RESET) {
+               val &= ~INFINIPATH_XGXS_RESET;
+       }
+       if (((val >> INFINIPATH_XGXS_RX_POL_SHIFT) &
+            INFINIPATH_XGXS_RX_POL_MASK) != dd->ipath_rx_pol_inv ) {
+               /* need to compensate for Tx inversion in partner */
+               val &= ~(INFINIPATH_XGXS_RX_POL_MASK <<
+                        INFINIPATH_XGXS_RX_POL_SHIFT);
+               val |= dd->ipath_rx_pol_inv <<
+                       INFINIPATH_XGXS_RX_POL_SHIFT;
+       }
+       if (dd->ipath_minrev >= 2) {
+               /* Rev 2. can tolerate multiple writes to PBC, and
+                * allowing them can provide lower latency on some
+                * CPUs, but this feature is off by default, only
+                * turned on by setting D63 of XGXSconfig reg.
+                * May want to make this conditional more
+                * fine-grained in future. This is not exactly
+                * related to XGXS, but where the bit ended up.
+                */
+               val |= INFINIPATH_XGXS_SUPPRESS_ARMLAUNCH_ERR;
+       }
+       if (val != prev_val)
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_xgxsconfig, val);
+
+       val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_serdesconfig0);
+
+       /* clear current and de-emphasis bits */
+       config1 &= ~0x0ffffffff00ULL;
+       /* set current to 20ma */
+       config1 |= 0x00000000000ULL;
+       /* set de-emphasis to -5.68dB */
+       config1 |= 0x0cccc000000ULL;
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_serdesconfig1, config1);
+
+       ipath_cdbg(VERBOSE, "done: SerDes status config0=%llx "
+                  "config1=%llx, sstatus=%llx xgxs=%llx\n",
+                  (unsigned long long) val, (unsigned long long) config1,
+                  (unsigned long long)
+                  ipath_read_kreg64(dd, dd->ipath_kregs->kr_serdesstatus),
+                  (unsigned long long)
+                  ipath_read_kreg64(dd, dd->ipath_kregs->kr_xgxsconfig));
+
+       if (!ipath_waitfor_mdio_cmdready(dd)) {
+               ipath_write_kreg(
+                       dd, dd->ipath_kregs->kr_mdio,
+                       ipath_mdio_req(IPATH_MDIO_CMD_READ, 31,
+                                      IPATH_MDIO_CTRL_XGXS_REG_8, 0));
+               if (ipath_waitfor_complete(dd, dd->ipath_kregs->kr_mdio,
+                                          IPATH_MDIO_DATAVALID, &val))
+                       ipath_dbg("Never got MDIO data for XGXS "
+                                 "status read\n");
+               else
+                       ipath_cdbg(VERBOSE, "MDIO Read reg8, "
+                                  "'bank' 31 %x\n", (u32) val);
+       } else
+               ipath_dbg("Never got MDIO cmdready for XGXS status read\n");
+
+       return ret;
+}
+
+/**
+ * ipath_pe_quiet_serdes - set serdes to txidle
+ * @dd: the infinipath device
+ * Called when driver is being unloaded
+ */
+static void ipath_pe_quiet_serdes(struct ipath_devdata *dd)
+{
+       u64 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_serdesconfig0);
+
+       val |= INFINIPATH_SERDC0_TXIDLE;
+       ipath_dbg("Setting TxIdleEn on serdes (config0 = %llx)\n",
+                 (unsigned long long) val);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_serdesconfig0, val);
+}
+
+static int ipath_pe_intconfig(struct ipath_devdata *dd)
+{
+       u64 val;
+       u32 chiprev;
+
+       /*
+        * If the chip supports added error indication via GPIO pins,
+        * enable interrupts on those bits so the interrupt routine
+        * can count the events. Also set flag so interrupt routine
+        * can know they are expected.
+        */
+       chiprev = dd->ipath_revision >> INFINIPATH_R_CHIPREVMINOR_SHIFT;
+       if ((chiprev & INFINIPATH_R_CHIPREVMINOR_MASK) > 1) {
+               /* Rev2+ reports extra errors via internal GPIO pins */
+               dd->ipath_flags |= IPATH_GPIO_ERRINTRS;
+               val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_gpio_mask);
+               val |= IPATH_GPIO_ERRINTR_MASK;
+               ipath_write_kreg( dd, dd->ipath_kregs->kr_gpio_mask, val);
+       }
+       return 0;
+}
+
+/**
+ * ipath_setup_pe_setextled - set the state of the two external LEDs
+ * @dd: the infinipath device
+ * @lst: the L state
+ * @ltst: the LT state
+
+ * These LEDs indicate the physical and logical state of IB link.
+ * For this chip (at least with recommended board pinouts), LED1
+ * is Yellow (logical state) and LED2 is Green (physical state),
+ *
+ * Note:  We try to match the Mellanox HCA LED behavior as best
+ * we can.  Green indicates physical link state is OK (something is
+ * plugged in, and we can train).
+ * Amber indicates the link is logically up (ACTIVE).
+ * Mellanox further blinks the amber LED to indicate data packet
+ * activity, but we have no hardware support for that, so it would
+ * require waking up every 10-20 msecs and checking the counters
+ * on the chip, and then turning the LED off if appropriate.  That's
+ * visible overhead, so not something we will do.
+ *
+ */
+static void ipath_setup_pe_setextled(struct ipath_devdata *dd, u64 lst,
+                                    u64 ltst)
+{
+       u64 extctl;
+
+       /* the diags use the LED to indicate diag info, so we leave
+        * the external LED alone when the diags are running */
+       if (ipath_diag_inuse)
+               return;
+
+       extctl = dd->ipath_extctrl & ~(INFINIPATH_EXTC_LED1PRIPORT_ON |
+                                      INFINIPATH_EXTC_LED2PRIPORT_ON);
+
+       if (ltst & INFINIPATH_IBCS_LT_STATE_LINKUP)
+               extctl |= INFINIPATH_EXTC_LED2PRIPORT_ON;
+       if (lst == INFINIPATH_IBCS_L_STATE_ACTIVE)
+               extctl |= INFINIPATH_EXTC_LED1PRIPORT_ON;
+       dd->ipath_extctrl = extctl;
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_extctrl, extctl);
+}
+
+/**
+ * ipath_setup_pe_cleanup - clean up any per-chip chip-specific stuff
+ * @dd: the infinipath device
+ *
+ * This is called during driver unload.
+ * We do the pci_disable_msi here, not in generic code, because it
+ * isn't used for the HT chips. If we do end up needing pci_enable_msi
+ * at some point in the future for HT, we'll move the call back
+ * into the main init_one code.
+ */
+static void ipath_setup_pe_cleanup(struct ipath_devdata *dd)
+{
+       dd->ipath_msi_lo = 0;   /* just in case unload fails */
+       pci_disable_msi(dd->pcidev);
+}
+
+/**
+ * ipath_setup_pe_config - setup PCIe config related stuff
+ * @dd: the infinipath device
+ * @pdev: the PCI device
+ *
+ * The pci_enable_msi() call will fail on systems with MSI quirks
+ * such as those with AMD8131, even if the device of interest is not
+ * attached to that device, (in the 2.6.13 - 2.6.15 kernels, at least, fixed
+ * late in 2.6.16).
+ * All that can be done is to edit the kernel source to remove the quirk
+ * check until that is fixed.
+ * We do not need to call enable_msi() for our HyperTransport chip,
+ * even though it uses MSI, and we want to avoid the quirk warning, so
+ * So we call enable_msi only for PCIe.  If we do end up needing
+ * pci_enable_msi at some point in the future for HT, we'll move the
+ * call back into the main init_one code.
+ * We save the msi lo and hi values, so we can restore them after
+ * chip reset (the kernel PCI infrastructure doesn't yet handle that
+ * correctly).
+ */
+static int ipath_setup_pe_config(struct ipath_devdata *dd,
+                                struct pci_dev *pdev)
+{
+       int pos, ret;
+
+       dd->ipath_msi_lo = 0;   /* used as a flag during reset processing */
+       ret = pci_enable_msi(dd->pcidev);
+       if (ret)
+               ipath_dev_err(dd, "pci_enable_msi failed: %d, "
+                             "interrupts may not work\n", ret);
+       /* continue even if it fails, we may still be OK... */
+       dd->ipath_irq = pdev->irq;
+
+       if ((pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI))) {
+               u16 control;
+               pci_read_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
+                                     &dd->ipath_msi_lo);
+               pci_read_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
+                                     &dd->ipath_msi_hi);
+               pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
+                                    &control);
+               /* now save the data (vector) info */
+               pci_read_config_word(dd->pcidev,
+                                    pos + ((control & PCI_MSI_FLAGS_64BIT)
+                                           ? 12 : 8),
+                                    &dd->ipath_msi_data);
+               ipath_cdbg(VERBOSE, "Read msi data 0x%x from config offset "
+                          "0x%x, control=0x%x\n", dd->ipath_msi_data,
+                          pos + ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
+                          control);
+               /* we save the cachelinesize also, although it doesn't
+                * really matter */
+               pci_read_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE,
+                                    &dd->ipath_pci_cacheline);
+       } else
+               ipath_dev_err(dd, "Can't find MSI capability, "
+                             "can't save MSI settings for reset\n");
+       if ((pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_EXP))) {
+               u16 linkstat;
+               pci_read_config_word(dd->pcidev, pos + PCI_EXP_LNKSTA,
+                                    &linkstat);
+               linkstat >>= 4;
+               linkstat &= 0x1f;
+               if (linkstat != 8)
+                       ipath_dev_err(dd, "PCIe width %u, "
+                                     "performance reduced\n", linkstat);
+       }
+       else
+               ipath_dev_err(dd, "Can't find PCI Express "
+                             "capability!\n");
+       return 0;
+}
+
+static void ipath_init_pe_variables(struct ipath_devdata *dd)
+{
+       /*
+        * bits for selecting i2c direction and values,
+        * used for I2C serial flash
+        */
+       dd->ipath_gpio_sda_num = _IPATH_GPIO_SDA_NUM;
+       dd->ipath_gpio_scl_num = _IPATH_GPIO_SCL_NUM;
+       dd->ipath_gpio_sda = IPATH_GPIO_SDA;
+       dd->ipath_gpio_scl = IPATH_GPIO_SCL;
+
+       /* variables for sanity checking interrupt and errors */
+       dd->ipath_hwe_bitsextant =
+               (INFINIPATH_HWE_RXEMEMPARITYERR_MASK <<
+                INFINIPATH_HWE_RXEMEMPARITYERR_SHIFT) |
+               (INFINIPATH_HWE_TXEMEMPARITYERR_MASK <<
+                INFINIPATH_HWE_TXEMEMPARITYERR_SHIFT) |
+               (INFINIPATH_HWE_PCIEMEMPARITYERR_MASK <<
+                INFINIPATH_HWE_PCIEMEMPARITYERR_SHIFT) |
+               INFINIPATH_HWE_PCIE1PLLFAILED |
+               INFINIPATH_HWE_PCIE0PLLFAILED |
+               INFINIPATH_HWE_PCIEPOISONEDTLP |
+               INFINIPATH_HWE_PCIECPLTIMEOUT |
+               INFINIPATH_HWE_PCIEBUSPARITYXTLH |
+               INFINIPATH_HWE_PCIEBUSPARITYXADM |
+               INFINIPATH_HWE_PCIEBUSPARITYRADM |
+               INFINIPATH_HWE_MEMBISTFAILED |
+               INFINIPATH_HWE_COREPLL_FBSLIP |
+               INFINIPATH_HWE_COREPLL_RFSLIP |
+               INFINIPATH_HWE_SERDESPLLFAILED |
+               INFINIPATH_HWE_IBCBUSTOSPCPARITYERR |
+               INFINIPATH_HWE_IBCBUSFRSPCPARITYERR;
+       dd->ipath_i_bitsextant =
+               (INFINIPATH_I_RCVURG_MASK << INFINIPATH_I_RCVURG_SHIFT) |
+               (INFINIPATH_I_RCVAVAIL_MASK <<
+                INFINIPATH_I_RCVAVAIL_SHIFT) |
+               INFINIPATH_I_ERROR | INFINIPATH_I_SPIOSENT |
+               INFINIPATH_I_SPIOBUFAVAIL | INFINIPATH_I_GPIO;
+       dd->ipath_e_bitsextant =
+               INFINIPATH_E_RFORMATERR | INFINIPATH_E_RVCRC |
+               INFINIPATH_E_RICRC | INFINIPATH_E_RMINPKTLEN |
+               INFINIPATH_E_RMAXPKTLEN | INFINIPATH_E_RLONGPKTLEN |
+               INFINIPATH_E_RSHORTPKTLEN | INFINIPATH_E_RUNEXPCHAR |
+               INFINIPATH_E_RUNSUPVL | INFINIPATH_E_REBP |
+               INFINIPATH_E_RIBFLOW | INFINIPATH_E_RBADVERSION |
+               INFINIPATH_E_RRCVEGRFULL | INFINIPATH_E_RRCVHDRFULL |
+               INFINIPATH_E_RBADTID | INFINIPATH_E_RHDRLEN |
+               INFINIPATH_E_RHDR | INFINIPATH_E_RIBLOSTLINK |
+               INFINIPATH_E_SMINPKTLEN | INFINIPATH_E_SMAXPKTLEN |
+               INFINIPATH_E_SUNDERRUN | INFINIPATH_E_SPKTLEN |
+               INFINIPATH_E_SDROPPEDSMPPKT | INFINIPATH_E_SDROPPEDDATAPKT |
+               INFINIPATH_E_SPIOARMLAUNCH | INFINIPATH_E_SUNEXPERRPKTNUM |
+               INFINIPATH_E_SUNSUPVL | INFINIPATH_E_IBSTATUSCHANGED |
+               INFINIPATH_E_INVALIDADDR | INFINIPATH_E_RESET |
+               INFINIPATH_E_HARDWARE;
+
+       dd->ipath_i_rcvavail_mask = INFINIPATH_I_RCVAVAIL_MASK;
+       dd->ipath_i_rcvurg_mask = INFINIPATH_I_RCVURG_MASK;
+}
+
+/* setup the MSI stuff again after a reset.  I'd like to just call
+ * pci_enable_msi() and request_irq() again, but when I do that,
+ * the MSI enable bit doesn't get set in the command word, and
+ * we switch to to a different interrupt vector, which is confusing,
+ * so I instead just do it all inline.  Perhaps somehow can tie this
+ * into the PCIe hotplug support at some point
+ * Note, because I'm doing it all here, I don't call pci_disable_msi()
+ * or free_irq() at the start of ipath_setup_pe_reset().
+ */
+static int ipath_reinit_msi(struct ipath_devdata *dd)
+{
+       int pos;
+       u16 control;
+       int ret;
+
+       if (!dd->ipath_msi_lo) {
+               dev_info(&dd->pcidev->dev, "Can't restore MSI config, "
+                        "initial setup failed?\n");
+               ret = 0;
+               goto bail;
+       }
+
+       if (!(pos = pci_find_capability(dd->pcidev, PCI_CAP_ID_MSI))) {
+               ipath_dev_err(dd, "Can't find MSI capability, "
+                             "can't restore MSI settings\n");
+               ret = 0;
+               goto bail;
+       }
+       ipath_cdbg(VERBOSE, "Writing msi_lo 0x%x to config offset 0x%x\n",
+                  dd->ipath_msi_lo, pos + PCI_MSI_ADDRESS_LO);
+       pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_LO,
+                              dd->ipath_msi_lo);
+       ipath_cdbg(VERBOSE, "Writing msi_lo 0x%x to config offset 0x%x\n",
+                  dd->ipath_msi_hi, pos + PCI_MSI_ADDRESS_HI);
+       pci_write_config_dword(dd->pcidev, pos + PCI_MSI_ADDRESS_HI,
+                              dd->ipath_msi_hi);
+       pci_read_config_word(dd->pcidev, pos + PCI_MSI_FLAGS, &control);
+       if (!(control & PCI_MSI_FLAGS_ENABLE)) {
+               ipath_cdbg(VERBOSE, "MSI control at off %x was %x, "
+                          "setting MSI enable (%x)\n", pos + PCI_MSI_FLAGS,
+                          control, control | PCI_MSI_FLAGS_ENABLE);
+               control |= PCI_MSI_FLAGS_ENABLE;
+               pci_write_config_word(dd->pcidev, pos + PCI_MSI_FLAGS,
+                                     control);
+       }
+       /* now rewrite the data (vector) info */
+       pci_write_config_word(dd->pcidev, pos +
+                             ((control & PCI_MSI_FLAGS_64BIT) ? 12 : 8),
+                             dd->ipath_msi_data);
+       /* we restore the cachelinesize also, although it doesn't really
+        * matter */
+       pci_write_config_byte(dd->pcidev, PCI_CACHE_LINE_SIZE,
+                             dd->ipath_pci_cacheline);
+       /* and now set the pci master bit again */
+       pci_set_master(dd->pcidev);
+       ret = 1;
+
+bail:
+       return ret;
+}
+
+/* This routine sleeps, so it can only be called from user context, not
+ * from interrupt context.  If we need interrupt context, we can split
+ * it into two routines.
+*/
+static int ipath_setup_pe_reset(struct ipath_devdata *dd)
+{
+       u64 val;
+       int i;
+       int ret;
+
+       /* Use ERROR so it shows up in logs, etc. */
+       ipath_dev_err(dd, "Resetting InfiniPath unit %u\n", dd->ipath_unit);
+       /* keep chip from being accessed in a few places */
+       dd->ipath_flags &= ~(IPATH_INITTED|IPATH_PRESENT);
+       val = dd->ipath_control | INFINIPATH_C_RESET;
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_control, val);
+       mb();
+
+       for (i = 1; i <= 5; i++) {
+               int r;
+               /* allow MBIST, etc. to complete; longer on each retry.
+                * We sometimes get machine checks from bus timeout if no
+                * response, so for now, make it *really* long.
+                */
+               msleep(1000 + (1 + i) * 2000);
+               if ((r =
+                    pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_0,
+                                           dd->ipath_pcibar0)))
+                       ipath_dev_err(dd, "rewrite of BAR0 failed: %d\n",
+                                     r);
+               if ((r =
+                    pci_write_config_dword(dd->pcidev, PCI_BASE_ADDRESS_1,
+                                           dd->ipath_pcibar1)))
+                       ipath_dev_err(dd, "rewrite of BAR1 failed: %d\n",
+                                     r);
+               /* now re-enable memory access */
+               if ((r = pci_enable_device(dd->pcidev)))
+                       ipath_dev_err(dd, "pci_enable_device failed after "
+                                     "reset: %d\n", r);
+               /* whether it worked or not, mark as present, again */
+               dd->ipath_flags |= IPATH_PRESENT;
+               val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_revision);
+               if (val == dd->ipath_revision) {
+                       ipath_cdbg(VERBOSE, "Got matching revision "
+                                  "register %llx on try %d\n",
+                                  (unsigned long long) val, i);
+                       ret = ipath_reinit_msi(dd);
+                       goto bail;
+               }
+               /* Probably getting -1 back */
+               ipath_dbg("Didn't get expected revision register, "
+                         "got %llx, try %d\n", (unsigned long long) val,
+                         i + 1);
+       }
+       ret = 0; /* failed */
+
+bail:
+       return ret;
+}
+
+/**
+ * ipath_pe_put_tid - write a TID in chip
+ * @dd: the infinipath device
+ * @tidptr: pointer to the expected TID (in chip) to udpate
+ * @tidtype: 0 for eager, 1 for expected
+ * @pa: physical address of in memory buffer; ipath_tidinvalid if freeing
+ *
+ * This exists as a separate routine to allow for special locking etc.
+ * It's used for both the full cleanup on exit, as well as the normal
+ * setup and teardown.
+ */
+static void ipath_pe_put_tid(struct ipath_devdata *dd, u64 __iomem *tidptr,
+                            u32 type, unsigned long pa)
+{
+       u32 __iomem *tidp32 = (u32 __iomem *)tidptr;
+       unsigned long flags = 0; /* keep gcc quiet */
+
+       if (pa != dd->ipath_tidinvalid) {
+               if (pa & ((1U << 11) - 1)) {
+                       dev_info(&dd->pcidev->dev, "BUG: physaddr %lx "
+                                "not 4KB aligned!\n", pa);
+                       return;
+               }
+               pa >>= 11;
+               /* paranoia check */
+               if (pa & (7<<29))
+                       ipath_dev_err(dd,
+                                     "BUG: Physical page address 0x%lx "
+                                     "has bits set in 31-29\n", pa);
+
+               if (type == 0)
+                       pa |= dd->ipath_tidtemplate;
+               else /* for now, always full 4KB page */
+                       pa |= 2 << 29;
+       }
+
+       /* workaround chip bug 9437 by writing each TID twice
+        * and holding a spinlock around the writes, so they don't
+        * intermix with other TID (eager or expected) writes
+        * Unfortunately, this call can be done from interrupt level
+        * for the port 0 eager TIDs, so we have to use irqsave
+        */
+       spin_lock_irqsave(&dd->ipath_tid_lock, flags);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_scratch, 0xfeeddeaf);
+       if (dd->ipath_kregbase)
+               writel(pa, tidp32);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_scratch, 0xdeadbeef);
+       mmiowb();
+       spin_unlock_irqrestore(&dd->ipath_tid_lock, flags);
+}
+/**
+ * ipath_pe_put_tid_2 - write a TID in chip, Revision 2 or higher
+ * @dd: the infinipath device
+ * @tidptr: pointer to the expected TID (in chip) to udpate
+ * @tidtype: 0 for eager, 1 for expected
+ * @pa: physical address of in memory buffer; ipath_tidinvalid if freeing
+ *
+ * This exists as a separate routine to allow for selection of the
+ * appropriate "flavor". The static calls in cleanup just use the
+ * revision-agnostic form, as they are not performance critical.
+ */
+static void ipath_pe_put_tid_2(struct ipath_devdata *dd, u64 __iomem *tidptr,
+                            u32 type, unsigned long pa)
+{
+       u32 __iomem *tidp32 = (u32 __iomem *)tidptr;
+
+       if (pa != dd->ipath_tidinvalid) {
+               if (pa & ((1U << 11) - 1)) {
+                       dev_info(&dd->pcidev->dev, "BUG: physaddr %lx "
+                                "not 2KB aligned!\n", pa);
+                       return;
+               }
+               pa >>= 11;
+               /* paranoia check */
+               if (pa & (7<<29))
+                       ipath_dev_err(dd,
+                                     "BUG: Physical page address 0x%lx "
+                                     "has bits set in 31-29\n", pa);
+
+               if (type == 0)
+                       pa |= dd->ipath_tidtemplate;
+               else /* for now, always full 4KB page */
+                       pa |= 2 << 29;
+       }
+       if (dd->ipath_kregbase)
+               writel(pa, tidp32);
+       mmiowb();
+}
+
+
+/**
+ * ipath_pe_clear_tid - clear all TID entries for a port, expected and eager
+ * @dd: the infinipath device
+ * @port: the port
+ *
+ * clear all TID entries for a port, expected and eager.
+ * Used from ipath_close().  On this chip, TIDs are only 32 bits,
+ * not 64, but they are still on 64 bit boundaries, so tidbase
+ * is declared as u64 * for the pointer math, even though we write 32 bits
+ */
+static void ipath_pe_clear_tids(struct ipath_devdata *dd, unsigned port)
+{
+       u64 __iomem *tidbase;
+       unsigned long tidinv;
+       int i;
+
+       if (!dd->ipath_kregbase)
+               return;
+
+       ipath_cdbg(VERBOSE, "Invalidate TIDs for port %u\n", port);
+
+       tidinv = dd->ipath_tidinvalid;
+       tidbase = (u64 __iomem *)
+               ((char __iomem *)(dd->ipath_kregbase) +
+                dd->ipath_rcvtidbase +
+                port * dd->ipath_rcvtidcnt * sizeof(*tidbase));
+
+       for (i = 0; i < dd->ipath_rcvtidcnt; i++)
+               ipath_pe_put_tid(dd, &tidbase[i], 0, tidinv);
+
+       tidbase = (u64 __iomem *)
+               ((char __iomem *)(dd->ipath_kregbase) +
+                dd->ipath_rcvegrbase +
+                port * dd->ipath_rcvegrcnt * sizeof(*tidbase));
+
+       for (i = 0; i < dd->ipath_rcvegrcnt; i++)
+               ipath_pe_put_tid(dd, &tidbase[i], 1, tidinv);
+}
+
+/**
+ * ipath_pe_tidtemplate - setup constants for TID updates
+ * @dd: the infinipath device
+ *
+ * We setup stuff that we use a lot, to avoid calculating each time
+ */
+static void ipath_pe_tidtemplate(struct ipath_devdata *dd)
+{
+       u32 egrsize = dd->ipath_rcvegrbufsize;
+
+       /* For now, we always allocate 4KB buffers (at init) so we can
+        * receive max size packets.  We may want a module parameter to
+        * specify 2KB or 4KB and/or make be per port instead of per device
+        * for those who want to reduce memory footprint.  Note that the
+        * ipath_rcvhdrentsize size must be large enough to hold the largest
+        * IB header (currently 96 bytes) that we expect to handle (plus of
+        * course the 2 dwords of RHF).
+        */
+       if (egrsize == 2048)
+               dd->ipath_tidtemplate = 1U << 29;
+       else if (egrsize == 4096)
+               dd->ipath_tidtemplate = 2U << 29;
+       else {
+               egrsize = 4096;
+               dev_info(&dd->pcidev->dev, "BUG: unsupported egrbufsize "
+                        "%u, using %u\n", dd->ipath_rcvegrbufsize,
+                        egrsize);
+               dd->ipath_tidtemplate = 2U << 29;
+       }
+       dd->ipath_tidinvalid = 0;
+}
+
+static int ipath_pe_early_init(struct ipath_devdata *dd)
+{
+       dd->ipath_flags |= IPATH_4BYTE_TID;
+
+       /*
+        * For openfabrics, we need to be able to handle an IB header of
+        * 24 dwords.  HT chip has arbitrary sized receive buffers, so we
+        * made them the same size as the PIO buffers.  This chip does not
+        * handle arbitrary size buffers, so we need the header large enough
+        * to handle largest IB header, but still have room for a 2KB MTU
+        * standard IB packet.
+        */
+       dd->ipath_rcvhdrentsize = 24;
+       dd->ipath_rcvhdrsize = IPATH_DFLT_RCVHDRSIZE;
+
+       /*
+        * To truly support a 4KB MTU (for usermode), we need to
+        * bump this to a larger value.  For now, we use them for
+        * the kernel only.
+        */
+       dd->ipath_rcvegrbufsize = 2048;
+       /*
+        * the min() check here is currently a nop, but it may not always
+        * be, depending on just how we do ipath_rcvegrbufsize
+        */
+       dd->ipath_ibmaxlen = min(dd->ipath_piosize2k,
+                                dd->ipath_rcvegrbufsize +
+                                (dd->ipath_rcvhdrentsize << 2));
+       dd->ipath_init_ibmaxlen = dd->ipath_ibmaxlen;
+
+       /*
+        * We can request a receive interrupt for 1 or
+        * more packets from current offset.  For now, we set this
+        * up for a single packet.
+        */
+       dd->ipath_rhdrhead_intr_off = 1ULL<<32;
+
+       ipath_get_eeprom_info(dd);
+
+       return 0;
+}
+
+int __attribute__((weak)) ipath_unordered_wc(void)
+{
+       return 0;
+}
+
+/**
+ * ipath_init_pe_get_base_info - set chip-specific flags for user code
+ * @pd: the infinipath port
+ * @kbase: ipath_base_info pointer
+ *
+ * We set the PCIE flag because the lower bandwidth on PCIe vs
+ * HyperTransport can affect some user packet algorithims.
+ */
+static int ipath_pe_get_base_info(struct ipath_portdata *pd, void *kbase)
+{
+       struct ipath_base_info *kinfo = kbase;
+       struct ipath_devdata *dd;
+
+       if (ipath_unordered_wc()) {
+               kinfo->spi_runtime_flags |= IPATH_RUNTIME_FORCE_WC_ORDER;
+               ipath_cdbg(PROC, "Intel processor, forcing WC order\n");
+       }
+       else
+               ipath_cdbg(PROC, "Not Intel processor, WC ordered\n");
+
+       if (pd == NULL)
+               goto done;
+
+       dd = pd->port_dd;
+
+       if (dd != NULL && dd->ipath_minrev >= 2) {
+               ipath_cdbg(PROC, "IBA6120 Rev2, allow multiple PBC write\n");
+               kinfo->spi_runtime_flags |= IPATH_RUNTIME_PBC_REWRITE;
+               ipath_cdbg(PROC, "IBA6120 Rev2, allow loose DMA alignment\n");
+               kinfo->spi_runtime_flags |= IPATH_RUNTIME_LOOSE_DMA_ALIGN;
+       }
+
+done:
+       kinfo->spi_runtime_flags |= IPATH_RUNTIME_PCIE;
+       return 0;
+}
+
+static void ipath_pe_free_irq(struct ipath_devdata *dd)
+{
+       free_irq(dd->ipath_irq, dd);
+       dd->ipath_irq = 0;
+}
+
+/**
+ * ipath_init_iba6120_funcs - set up the chip-specific function pointers
+ * @dd: the infinipath device
+ *
+ * This is global, and is called directly at init to set up the
+ * chip-specific function pointers for later use.
+ */
+void ipath_init_iba6120_funcs(struct ipath_devdata *dd)
+{
+       dd->ipath_f_intrsetup = ipath_pe_intconfig;
+       dd->ipath_f_bus = ipath_setup_pe_config;
+       dd->ipath_f_reset = ipath_setup_pe_reset;
+       dd->ipath_f_get_boardname = ipath_pe_boardname;
+       dd->ipath_f_init_hwerrors = ipath_pe_init_hwerrors;
+       dd->ipath_f_early_init = ipath_pe_early_init;
+       dd->ipath_f_handle_hwerrors = ipath_pe_handle_hwerrors;
+       dd->ipath_f_quiet_serdes = ipath_pe_quiet_serdes;
+       dd->ipath_f_bringup_serdes = ipath_pe_bringup_serdes;
+       dd->ipath_f_clear_tids = ipath_pe_clear_tids;
+       if (dd->ipath_minrev >= 2)
+               dd->ipath_f_put_tid = ipath_pe_put_tid_2;
+       else
+               dd->ipath_f_put_tid = ipath_pe_put_tid;
+       dd->ipath_f_cleanup = ipath_setup_pe_cleanup;
+       dd->ipath_f_setextled = ipath_setup_pe_setextled;
+       dd->ipath_f_get_base_info = ipath_pe_get_base_info;
+       dd->ipath_f_free_irq = ipath_pe_free_irq;
+
+       /* initialize chip-specific variables */
+       dd->ipath_f_tidtemplate = ipath_pe_tidtemplate;
+
+       /*
+        * setup the register offsets, since they are different for each
+        * chip
+        */
+       dd->ipath_kregs = &ipath_pe_kregs;
+       dd->ipath_cregs = &ipath_pe_cregs;
+
+       ipath_init_pe_variables(dd);
+}
+