This commit was manufactured by cvs2svn to create branch 'vserver'.
[linux-2.6.git] / drivers / infiniband / hw / ipath / ipath_mad.c
diff --git a/drivers/infiniband/hw/ipath/ipath_mad.c b/drivers/infiniband/hw/ipath/ipath_mad.c
new file mode 100644 (file)
index 0000000..f7f8391
--- /dev/null
@@ -0,0 +1,1352 @@
+/*
+ * Copyright (c) 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 <rdma/ib_smi.h>
+
+#include "ipath_kernel.h"
+#include "ipath_verbs.h"
+#include "ips_common.h"
+
+#define IB_SMP_UNSUP_VERSION   __constant_htons(0x0004)
+#define IB_SMP_UNSUP_METHOD    __constant_htons(0x0008)
+#define IB_SMP_UNSUP_METH_ATTR __constant_htons(0x000C)
+#define IB_SMP_INVALID_FIELD   __constant_htons(0x001C)
+
+static int reply(struct ib_smp *smp)
+{
+       /*
+        * The verbs framework will handle the directed/LID route
+        * packet changes.
+        */
+       smp->method = IB_MGMT_METHOD_GET_RESP;
+       if (smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
+               smp->status |= IB_SMP_DIRECTION;
+       return IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY;
+}
+
+static int recv_subn_get_nodedescription(struct ib_smp *smp,
+                                        struct ib_device *ibdev)
+{
+       if (smp->attr_mod)
+               smp->status |= IB_SMP_INVALID_FIELD;
+
+       strncpy(smp->data, ibdev->node_desc, sizeof(smp->data));
+
+       return reply(smp);
+}
+
+struct nodeinfo {
+       u8 base_version;
+       u8 class_version;
+       u8 node_type;
+       u8 num_ports;
+       __be64 sys_guid;
+       __be64 node_guid;
+       __be64 port_guid;
+       __be16 partition_cap;
+       __be16 device_id;
+       __be32 revision;
+       u8 local_port_num;
+       u8 vendor_id[3];
+} __attribute__ ((packed));
+
+static int recv_subn_get_nodeinfo(struct ib_smp *smp,
+                                 struct ib_device *ibdev, u8 port)
+{
+       struct nodeinfo *nip = (struct nodeinfo *)&smp->data;
+       struct ipath_devdata *dd = to_idev(ibdev)->dd;
+       u32 vendor, boardid, majrev, minrev;
+
+       if (smp->attr_mod)
+               smp->status |= IB_SMP_INVALID_FIELD;
+
+       nip->base_version = 1;
+       nip->class_version = 1;
+       nip->node_type = 1;     /* channel adapter */
+       /*
+        * XXX The num_ports value will need a layer function to get
+        * the value if we ever have more than one IB port on a chip.
+        * We will also need to get the GUID for the port.
+        */
+       nip->num_ports = ibdev->phys_port_cnt;
+       /* This is already in network order */
+       nip->sys_guid = to_idev(ibdev)->sys_image_guid;
+       nip->node_guid = ipath_layer_get_guid(dd);
+       nip->port_guid = nip->sys_guid;
+       nip->partition_cap = cpu_to_be16(ipath_layer_get_npkeys(dd));
+       nip->device_id = cpu_to_be16(ipath_layer_get_deviceid(dd));
+       ipath_layer_query_device(dd, &vendor, &boardid, &majrev, &minrev);
+       nip->revision = cpu_to_be32((majrev << 16) | minrev);
+       nip->local_port_num = port;
+       nip->vendor_id[0] = 0;
+       nip->vendor_id[1] = vendor >> 8;
+       nip->vendor_id[2] = vendor;
+
+       return reply(smp);
+}
+
+static int recv_subn_get_guidinfo(struct ib_smp *smp,
+                                 struct ib_device *ibdev)
+{
+       u32 startgx = 8 * be32_to_cpu(smp->attr_mod);
+       __be64 *p = (__be64 *) smp->data;
+
+       /* 32 blocks of 8 64-bit GUIDs per block */
+
+       memset(smp->data, 0, sizeof(smp->data));
+
+       /*
+        * We only support one GUID for now.  If this changes, the
+        * portinfo.guid_cap field needs to be updated too.
+        */
+       if (startgx == 0)
+               /* The first is a copy of the read-only HW GUID. */
+               *p = ipath_layer_get_guid(to_idev(ibdev)->dd);
+       else
+               smp->status |= IB_SMP_INVALID_FIELD;
+
+       return reply(smp);
+}
+
+struct port_info {
+       __be64 mkey;
+       __be64 gid_prefix;
+       __be16 lid;
+       __be16 sm_lid;
+       __be32 cap_mask;
+       __be16 diag_code;
+       __be16 mkey_lease_period;
+       u8 local_port_num;
+       u8 link_width_enabled;
+       u8 link_width_supported;
+       u8 link_width_active;
+       u8 linkspeed_portstate;                 /* 4 bits, 4 bits */
+       u8 portphysstate_linkdown;              /* 4 bits, 4 bits */
+       u8 mkeyprot_resv_lmc;                   /* 2 bits, 3, 3 */
+       u8 linkspeedactive_enabled;             /* 4 bits, 4 bits */
+       u8 neighbormtu_mastersmsl;              /* 4 bits, 4 bits */
+       u8 vlcap_inittype;                      /* 4 bits, 4 bits */
+       u8 vl_high_limit;
+       u8 vl_arb_high_cap;
+       u8 vl_arb_low_cap;
+       u8 inittypereply_mtucap;                /* 4 bits, 4 bits */
+       u8 vlstallcnt_hoqlife;                  /* 3 bits, 5 bits */
+       u8 operationalvl_pei_peo_fpi_fpo;       /* 4 bits, 1, 1, 1, 1 */
+       __be16 mkey_violations;
+       __be16 pkey_violations;
+       __be16 qkey_violations;
+       u8 guid_cap;
+       u8 clientrereg_resv_subnetto;           /* 1 bit, 2 bits, 5 */
+       u8 resv_resptimevalue;                  /* 3 bits, 5 bits */
+       u8 localphyerrors_overrunerrors;        /* 4 bits, 4 bits */
+       __be16 max_credit_hint;
+       u8 resv;
+       u8 link_roundtrip_latency[3];
+} __attribute__ ((packed));
+
+static int recv_subn_get_portinfo(struct ib_smp *smp,
+                                 struct ib_device *ibdev, u8 port)
+{
+       struct ipath_ibdev *dev;
+       struct port_info *pip = (struct port_info *)smp->data;
+       u16 lid;
+       u8 ibcstat;
+       u8 mtu;
+       int ret;
+
+       if (be32_to_cpu(smp->attr_mod) > ibdev->phys_port_cnt) {
+               smp->status |= IB_SMP_INVALID_FIELD;
+               ret = reply(smp);
+               goto bail;
+       }
+
+       dev = to_idev(ibdev);
+
+       /* Clear all fields.  Only set the non-zero fields. */
+       memset(smp->data, 0, sizeof(smp->data));
+
+       /* Only return the mkey if the protection field allows it. */
+       if (smp->method == IB_MGMT_METHOD_SET || dev->mkey == smp->mkey ||
+           (dev->mkeyprot_resv_lmc >> 6) == 0)
+               pip->mkey = dev->mkey;
+       pip->gid_prefix = dev->gid_prefix;
+       lid = ipath_layer_get_lid(dev->dd);
+       pip->lid = lid ? cpu_to_be16(lid) : IB_LID_PERMISSIVE;
+       pip->sm_lid = cpu_to_be16(dev->sm_lid);
+       pip->cap_mask = cpu_to_be32(dev->port_cap_flags);
+       /* pip->diag_code; */
+       pip->mkey_lease_period = cpu_to_be16(dev->mkey_lease_period);
+       pip->local_port_num = port;
+       pip->link_width_enabled = dev->link_width_enabled;
+       pip->link_width_supported = 3;  /* 1x or 4x */
+       pip->link_width_active = 2;     /* 4x */
+       pip->linkspeed_portstate = 0x10;        /* 2.5Gbps */
+       ibcstat = ipath_layer_get_lastibcstat(dev->dd);
+       pip->linkspeed_portstate |= ((ibcstat >> 4) & 0x3) + 1;
+       pip->portphysstate_linkdown =
+               (ipath_cvt_physportstate[ibcstat & 0xf] << 4) |
+               (ipath_layer_get_linkdowndefaultstate(dev->dd) ? 1 : 2);
+       pip->mkeyprot_resv_lmc = dev->mkeyprot_resv_lmc;
+       pip->linkspeedactive_enabled = 0x11;    /* 2.5Gbps, 2.5Gbps */
+       switch (ipath_layer_get_ibmtu(dev->dd)) {
+       case 4096:
+               mtu = IB_MTU_4096;
+               break;
+       case 2048:
+               mtu = IB_MTU_2048;
+               break;
+       case 1024:
+               mtu = IB_MTU_1024;
+               break;
+       case 512:
+               mtu = IB_MTU_512;
+               break;
+       case 256:
+               mtu = IB_MTU_256;
+               break;
+       default:                /* oops, something is wrong */
+               mtu = IB_MTU_2048;
+               break;
+       }
+       pip->neighbormtu_mastersmsl = (mtu << 4) | dev->sm_sl;
+       pip->vlcap_inittype = 0x10;     /* VLCap = VL0, InitType = 0 */
+       pip->vl_high_limit = dev->vl_high_limit;
+       /* pip->vl_arb_high_cap; // only one VL */
+       /* pip->vl_arb_low_cap; // only one VL */
+       /* InitTypeReply = 0 */
+       pip->inittypereply_mtucap = IB_MTU_4096;
+       // HCAs ignore VLStallCount and HOQLife
+       /* pip->vlstallcnt_hoqlife; */
+       pip->operationalvl_pei_peo_fpi_fpo = 0x10;      /* OVLs = 1 */
+       pip->mkey_violations = cpu_to_be16(dev->mkey_violations);
+       /* P_KeyViolations are counted by hardware. */
+       pip->pkey_violations =
+               cpu_to_be16((ipath_layer_get_cr_errpkey(dev->dd) -
+                            dev->n_pkey_violations) & 0xFFFF);
+       pip->qkey_violations = cpu_to_be16(dev->qkey_violations);
+       /* Only the hardware GUID is supported for now */
+       pip->guid_cap = 1;
+       pip->clientrereg_resv_subnetto = dev->subnet_timeout;
+       /* 32.768 usec. response time (guessing) */
+       pip->resv_resptimevalue = 3;
+       pip->localphyerrors_overrunerrors =
+               (ipath_layer_get_phyerrthreshold(dev->dd) << 4) |
+               ipath_layer_get_overrunthreshold(dev->dd);
+       /* pip->max_credit_hint; */
+       /* pip->link_roundtrip_latency[3]; */
+
+       ret = reply(smp);
+
+bail:
+       return ret;
+}
+
+static int recv_subn_get_pkeytable(struct ib_smp *smp,
+                                  struct ib_device *ibdev)
+{
+       u32 startpx = 32 * (be32_to_cpu(smp->attr_mod) & 0xffff);
+       u16 *p = (u16 *) smp->data;
+       __be16 *q = (__be16 *) smp->data;
+
+       /* 64 blocks of 32 16-bit P_Key entries */
+
+       memset(smp->data, 0, sizeof(smp->data));
+       if (startpx == 0) {
+               struct ipath_ibdev *dev = to_idev(ibdev);
+               unsigned i, n = ipath_layer_get_npkeys(dev->dd);
+
+               ipath_layer_get_pkeys(dev->dd, p);
+
+               for (i = 0; i < n; i++)
+                       q[i] = cpu_to_be16(p[i]);
+       } else
+               smp->status |= IB_SMP_INVALID_FIELD;
+
+       return reply(smp);
+}
+
+static int recv_subn_set_guidinfo(struct ib_smp *smp,
+                                 struct ib_device *ibdev)
+{
+       /* The only GUID we support is the first read-only entry. */
+       return recv_subn_get_guidinfo(smp, ibdev);
+}
+
+/**
+ * recv_subn_set_portinfo - set port information
+ * @smp: the incoming SM packet
+ * @ibdev: the infiniband device
+ * @port: the port on the device
+ *
+ * Set Portinfo (see ch. 14.2.5.6).
+ */
+static int recv_subn_set_portinfo(struct ib_smp *smp,
+                                 struct ib_device *ibdev, u8 port)
+{
+       struct port_info *pip = (struct port_info *)smp->data;
+       struct ib_event event;
+       struct ipath_ibdev *dev;
+       u32 flags;
+       char clientrereg = 0;
+       u16 lid, smlid;
+       u8 lwe;
+       u8 lse;
+       u8 state;
+       u16 lstate;
+       u32 mtu;
+       int ret;
+
+       if (be32_to_cpu(smp->attr_mod) > ibdev->phys_port_cnt)
+               goto err;
+
+       dev = to_idev(ibdev);
+       event.device = ibdev;
+       event.element.port_num = port;
+
+       dev->mkey = pip->mkey;
+       dev->gid_prefix = pip->gid_prefix;
+       dev->mkey_lease_period = be16_to_cpu(pip->mkey_lease_period);
+
+       lid = be16_to_cpu(pip->lid);
+       if (lid != ipath_layer_get_lid(dev->dd)) {
+               /* Must be a valid unicast LID address. */
+               if (lid == 0 || lid >= IPS_MULTICAST_LID_BASE)
+                       goto err;
+               ipath_set_sps_lid(dev->dd, lid, pip->mkeyprot_resv_lmc & 7);
+               event.event = IB_EVENT_LID_CHANGE;
+               ib_dispatch_event(&event);
+       }
+
+       smlid = be16_to_cpu(pip->sm_lid);
+       if (smlid != dev->sm_lid) {
+               /* Must be a valid unicast LID address. */
+               if (smlid == 0 || smlid >= IPS_MULTICAST_LID_BASE)
+                       goto err;
+               dev->sm_lid = smlid;
+               event.event = IB_EVENT_SM_CHANGE;
+               ib_dispatch_event(&event);
+       }
+
+       /* Only 4x supported but allow 1x or 4x to be set (see 14.2.6.6). */
+       lwe = pip->link_width_enabled;
+       if ((lwe >= 4 && lwe <= 8) || (lwe >= 0xC && lwe <= 0xFE))
+               goto err;
+       if (lwe == 0xFF)
+               dev->link_width_enabled = 3;    /* 1x or 4x */
+       else if (lwe)
+               dev->link_width_enabled = lwe;
+
+       /* Only 2.5 Gbs supported. */
+       lse = pip->linkspeedactive_enabled & 0xF;
+       if (lse >= 2 && lse <= 0xE)
+               goto err;
+
+       /* Set link down default state. */
+       switch (pip->portphysstate_linkdown & 0xF) {
+       case 0: /* NOP */
+               break;
+       case 1: /* SLEEP */
+               if (ipath_layer_set_linkdowndefaultstate(dev->dd, 1))
+                       goto err;
+               break;
+       case 2: /* POLL */
+               if (ipath_layer_set_linkdowndefaultstate(dev->dd, 0))
+                       goto err;
+               break;
+       default:
+               goto err;
+       }
+
+       dev->mkeyprot_resv_lmc = pip->mkeyprot_resv_lmc;
+       dev->vl_high_limit = pip->vl_high_limit;
+
+       switch ((pip->neighbormtu_mastersmsl >> 4) & 0xF) {
+       case IB_MTU_256:
+               mtu = 256;
+               break;
+       case IB_MTU_512:
+               mtu = 512;
+               break;
+       case IB_MTU_1024:
+               mtu = 1024;
+               break;
+       case IB_MTU_2048:
+               mtu = 2048;
+               break;
+       case IB_MTU_4096:
+               mtu = 4096;
+               break;
+       default:
+               /* XXX We have already partially updated our state! */
+               goto err;
+       }
+       ipath_layer_set_mtu(dev->dd, mtu);
+
+       dev->sm_sl = pip->neighbormtu_mastersmsl & 0xF;
+
+       /* We only support VL0 */
+       if (((pip->operationalvl_pei_peo_fpi_fpo >> 4) & 0xF) > 1)
+               goto err;
+
+       if (pip->mkey_violations == 0)
+               dev->mkey_violations = 0;
+
+       /*
+        * Hardware counter can't be reset so snapshot and subtract
+        * later.
+        */
+       if (pip->pkey_violations == 0)
+               dev->n_pkey_violations =
+                       ipath_layer_get_cr_errpkey(dev->dd);
+
+       if (pip->qkey_violations == 0)
+               dev->qkey_violations = 0;
+
+       if (ipath_layer_set_phyerrthreshold(
+                   dev->dd,
+                   (pip->localphyerrors_overrunerrors >> 4) & 0xF))
+               goto err;
+
+       if (ipath_layer_set_overrunthreshold(
+                   dev->dd,
+                   (pip->localphyerrors_overrunerrors & 0xF)))
+               goto err;
+
+       dev->subnet_timeout = pip->clientrereg_resv_subnetto & 0x1F;
+
+       if (pip->clientrereg_resv_subnetto & 0x80) {
+               clientrereg = 1;
+               event.event = IB_EVENT_LID_CHANGE;
+               ib_dispatch_event(&event);
+       }
+
+       /*
+        * Do the port state change now that the other link parameters
+        * have been set.
+        * Changing the port physical state only makes sense if the link
+        * is down or is being set to down.
+        */
+       state = pip->linkspeed_portstate & 0xF;
+       flags = ipath_layer_get_flags(dev->dd);
+       lstate = (pip->portphysstate_linkdown >> 4) & 0xF;
+       if (lstate && !(state == IB_PORT_DOWN || state == IB_PORT_NOP))
+               goto err;
+
+       /*
+        * Only state changes of DOWN, ARM, and ACTIVE are valid
+        * and must be in the correct state to take effect (see 7.2.6).
+        */
+       switch (state) {
+       case IB_PORT_NOP:
+               if (lstate == 0)
+                       break;
+               /* FALLTHROUGH */
+       case IB_PORT_DOWN:
+               if (lstate == 0)
+                       if (ipath_layer_get_linkdowndefaultstate(dev->dd))
+                               lstate = IPATH_IB_LINKDOWN_SLEEP;
+                       else
+                               lstate = IPATH_IB_LINKDOWN;
+               else if (lstate == 1)
+                       lstate = IPATH_IB_LINKDOWN_SLEEP;
+               else if (lstate == 2)
+                       lstate = IPATH_IB_LINKDOWN;
+               else if (lstate == 3)
+                       lstate = IPATH_IB_LINKDOWN_DISABLE;
+               else
+                       goto err;
+               ipath_layer_set_linkstate(dev->dd, lstate);
+               if (flags & IPATH_LINKACTIVE) {
+                       event.event = IB_EVENT_PORT_ERR;
+                       ib_dispatch_event(&event);
+               }
+               break;
+       case IB_PORT_ARMED:
+               if (!(flags & (IPATH_LINKINIT | IPATH_LINKACTIVE)))
+                       break;
+               ipath_layer_set_linkstate(dev->dd, IPATH_IB_LINKARM);
+               if (flags & IPATH_LINKACTIVE) {
+                       event.event = IB_EVENT_PORT_ERR;
+                       ib_dispatch_event(&event);
+               }
+               break;
+       case IB_PORT_ACTIVE:
+               if (!(flags & IPATH_LINKARMED))
+                       break;
+               ipath_layer_set_linkstate(dev->dd, IPATH_IB_LINKACTIVE);
+               event.event = IB_EVENT_PORT_ACTIVE;
+               ib_dispatch_event(&event);
+               break;
+       default:
+               /* XXX We have already partially updated our state! */
+               goto err;
+       }
+
+       ret = recv_subn_get_portinfo(smp, ibdev, port);
+
+       if (clientrereg)
+               pip->clientrereg_resv_subnetto |= 0x80;
+
+       goto done;
+
+err:
+       smp->status |= IB_SMP_INVALID_FIELD;
+       ret = recv_subn_get_portinfo(smp, ibdev, port);
+
+done:
+       return ret;
+}
+
+static int recv_subn_set_pkeytable(struct ib_smp *smp,
+                                  struct ib_device *ibdev)
+{
+       u32 startpx = 32 * (be32_to_cpu(smp->attr_mod) & 0xffff);
+       __be16 *p = (__be16 *) smp->data;
+       u16 *q = (u16 *) smp->data;
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       unsigned i, n = ipath_layer_get_npkeys(dev->dd);
+
+       for (i = 0; i < n; i++)
+               q[i] = be16_to_cpu(p[i]);
+
+       if (startpx != 0 ||
+           ipath_layer_set_pkeys(dev->dd, q) != 0)
+               smp->status |= IB_SMP_INVALID_FIELD;
+
+       return recv_subn_get_pkeytable(smp, ibdev);
+}
+
+#define IB_PMA_CLASS_PORT_INFO         __constant_htons(0x0001)
+#define IB_PMA_PORT_SAMPLES_CONTROL    __constant_htons(0x0010)
+#define IB_PMA_PORT_SAMPLES_RESULT     __constant_htons(0x0011)
+#define IB_PMA_PORT_COUNTERS           __constant_htons(0x0012)
+#define IB_PMA_PORT_COUNTERS_EXT       __constant_htons(0x001D)
+#define IB_PMA_PORT_SAMPLES_RESULT_EXT __constant_htons(0x001E)
+
+struct ib_perf {
+       u8 base_version;
+       u8 mgmt_class;
+       u8 class_version;
+       u8 method;
+       __be16 status;
+       __be16 unused;
+       __be64 tid;
+       __be16 attr_id;
+       __be16 resv;
+       __be32 attr_mod;
+       u8 reserved[40];
+       u8 data[192];
+} __attribute__ ((packed));
+
+struct ib_pma_classportinfo {
+       u8 base_version;
+       u8 class_version;
+       __be16 cap_mask;
+       u8 reserved[3];
+       u8 resp_time_value;     /* only lower 5 bits */
+       union ib_gid redirect_gid;
+       __be32 redirect_tc_sl_fl;       /* 8, 4, 20 bits respectively */
+       __be16 redirect_lid;
+       __be16 redirect_pkey;
+       __be32 redirect_qp;     /* only lower 24 bits */
+       __be32 redirect_qkey;
+       union ib_gid trap_gid;
+       __be32 trap_tc_sl_fl;   /* 8, 4, 20 bits respectively */
+       __be16 trap_lid;
+       __be16 trap_pkey;
+       __be32 trap_hl_qp;      /* 8, 24 bits respectively */
+       __be32 trap_qkey;
+} __attribute__ ((packed));
+
+struct ib_pma_portsamplescontrol {
+       u8 opcode;
+       u8 port_select;
+       u8 tick;
+       u8 counter_width;       /* only lower 3 bits */
+       __be32 counter_mask0_9; /* 2, 10 * 3, bits */
+       __be16 counter_mask10_14;       /* 1, 5 * 3, bits */
+       u8 sample_mechanisms;
+       u8 sample_status;       /* only lower 2 bits */
+       __be64 option_mask;
+       __be64 vendor_mask;
+       __be32 sample_start;
+       __be32 sample_interval;
+       __be16 tag;
+       __be16 counter_select[15];
+} __attribute__ ((packed));
+
+struct ib_pma_portsamplesresult {
+       __be16 tag;
+       __be16 sample_status;   /* only lower 2 bits */
+       __be32 counter[15];
+} __attribute__ ((packed));
+
+struct ib_pma_portsamplesresult_ext {
+       __be16 tag;
+       __be16 sample_status;   /* only lower 2 bits */
+       __be32 extended_width;  /* only upper 2 bits */
+       __be64 counter[15];
+} __attribute__ ((packed));
+
+struct ib_pma_portcounters {
+       u8 reserved;
+       u8 port_select;
+       __be16 counter_select;
+       __be16 symbol_error_counter;
+       u8 link_error_recovery_counter;
+       u8 link_downed_counter;
+       __be16 port_rcv_errors;
+       __be16 port_rcv_remphys_errors;
+       __be16 port_rcv_switch_relay_errors;
+       __be16 port_xmit_discards;
+       u8 port_xmit_constraint_errors;
+       u8 port_rcv_constraint_errors;
+       u8 reserved1;
+       u8 lli_ebor_errors;     /* 4, 4, bits */
+       __be16 reserved2;
+       __be16 vl15_dropped;
+       __be32 port_xmit_data;
+       __be32 port_rcv_data;
+       __be32 port_xmit_packets;
+       __be32 port_rcv_packets;
+} __attribute__ ((packed));
+
+#define IB_PMA_SEL_SYMBOL_ERROR                        __constant_htons(0x0001)
+#define IB_PMA_SEL_LINK_ERROR_RECOVERY         __constant_htons(0x0002)
+#define IB_PMA_SEL_LINK_DOWNED                 __constant_htons(0x0004)
+#define IB_PMA_SEL_PORT_RCV_ERRORS             __constant_htons(0x0008)
+#define IB_PMA_SEL_PORT_RCV_REMPHYS_ERRORS     __constant_htons(0x0010)
+#define IB_PMA_SEL_PORT_XMIT_DISCARDS          __constant_htons(0x0040)
+#define IB_PMA_SEL_PORT_XMIT_DATA              __constant_htons(0x1000)
+#define IB_PMA_SEL_PORT_RCV_DATA               __constant_htons(0x2000)
+#define IB_PMA_SEL_PORT_XMIT_PACKETS           __constant_htons(0x4000)
+#define IB_PMA_SEL_PORT_RCV_PACKETS            __constant_htons(0x8000)
+
+struct ib_pma_portcounters_ext {
+       u8 reserved;
+       u8 port_select;
+       __be16 counter_select;
+       __be32 reserved1;
+       __be64 port_xmit_data;
+       __be64 port_rcv_data;
+       __be64 port_xmit_packets;
+       __be64 port_rcv_packets;
+       __be64 port_unicast_xmit_packets;
+       __be64 port_unicast_rcv_packets;
+       __be64 port_multicast_xmit_packets;
+       __be64 port_multicast_rcv_packets;
+} __attribute__ ((packed));
+
+#define IB_PMA_SELX_PORT_XMIT_DATA             __constant_htons(0x0001)
+#define IB_PMA_SELX_PORT_RCV_DATA              __constant_htons(0x0002)
+#define IB_PMA_SELX_PORT_XMIT_PACKETS          __constant_htons(0x0004)
+#define IB_PMA_SELX_PORT_RCV_PACKETS           __constant_htons(0x0008)
+#define IB_PMA_SELX_PORT_UNI_XMIT_PACKETS      __constant_htons(0x0010)
+#define IB_PMA_SELX_PORT_UNI_RCV_PACKETS       __constant_htons(0x0020)
+#define IB_PMA_SELX_PORT_MULTI_XMIT_PACKETS    __constant_htons(0x0040)
+#define IB_PMA_SELX_PORT_MULTI_RCV_PACKETS     __constant_htons(0x0080)
+
+static int recv_pma_get_classportinfo(struct ib_perf *pmp)
+{
+       struct ib_pma_classportinfo *p =
+               (struct ib_pma_classportinfo *)pmp->data;
+
+       memset(pmp->data, 0, sizeof(pmp->data));
+
+       if (pmp->attr_mod != 0)
+               pmp->status |= IB_SMP_INVALID_FIELD;
+
+       /* Indicate AllPortSelect is valid (only one port anyway) */
+       p->cap_mask = __constant_cpu_to_be16(1 << 8);
+       p->base_version = 1;
+       p->class_version = 1;
+       /*
+        * Expected response time is 4.096 usec. * 2^18 == 1.073741824
+        * sec.
+        */
+       p->resp_time_value = 18;
+
+       return reply((struct ib_smp *) pmp);
+}
+
+/*
+ * The PortSamplesControl.CounterMasks field is an array of 3 bit fields
+ * which specify the N'th counter's capabilities. See ch. 16.1.3.2.
+ * We support 5 counters which only count the mandatory quantities.
+ */
+#define COUNTER_MASK(q, n) (q << ((9 - n) * 3))
+#define COUNTER_MASK0_9 \
+       __constant_cpu_to_be32(COUNTER_MASK(1, 0) | \
+                              COUNTER_MASK(1, 1) | \
+                              COUNTER_MASK(1, 2) | \
+                              COUNTER_MASK(1, 3) | \
+                              COUNTER_MASK(1, 4))
+
+static int recv_pma_get_portsamplescontrol(struct ib_perf *pmp,
+                                          struct ib_device *ibdev, u8 port)
+{
+       struct ib_pma_portsamplescontrol *p =
+               (struct ib_pma_portsamplescontrol *)pmp->data;
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       unsigned long flags;
+       u8 port_select = p->port_select;
+
+       memset(pmp->data, 0, sizeof(pmp->data));
+
+       p->port_select = port_select;
+       if (pmp->attr_mod != 0 ||
+           (port_select != port && port_select != 0xFF))
+               pmp->status |= IB_SMP_INVALID_FIELD;
+       /*
+        * Ticks are 10x the link transfer period which for 2.5Gbs is 4
+        * nsec.  0 == 4 nsec., 1 == 8 nsec., ..., 255 == 1020 nsec.  Sample
+        * intervals are counted in ticks.  Since we use Linux timers, that
+        * count in jiffies, we can't sample for less than 1000 ticks if HZ
+        * == 1000 (4000 ticks if HZ is 250).
+        */
+       /* XXX This is WRONG. */
+       p->tick = 250;          /* 1 usec. */
+       p->counter_width = 4;   /* 32 bit counters */
+       p->counter_mask0_9 = COUNTER_MASK0_9;
+       spin_lock_irqsave(&dev->pending_lock, flags);
+       p->sample_status = dev->pma_sample_status;
+       p->sample_start = cpu_to_be32(dev->pma_sample_start);
+       p->sample_interval = cpu_to_be32(dev->pma_sample_interval);
+       p->tag = cpu_to_be16(dev->pma_tag);
+       p->counter_select[0] = dev->pma_counter_select[0];
+       p->counter_select[1] = dev->pma_counter_select[1];
+       p->counter_select[2] = dev->pma_counter_select[2];
+       p->counter_select[3] = dev->pma_counter_select[3];
+       p->counter_select[4] = dev->pma_counter_select[4];
+       spin_unlock_irqrestore(&dev->pending_lock, flags);
+
+       return reply((struct ib_smp *) pmp);
+}
+
+static int recv_pma_set_portsamplescontrol(struct ib_perf *pmp,
+                                          struct ib_device *ibdev, u8 port)
+{
+       struct ib_pma_portsamplescontrol *p =
+               (struct ib_pma_portsamplescontrol *)pmp->data;
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       unsigned long flags;
+       u32 start;
+       int ret;
+
+       if (pmp->attr_mod != 0 ||
+           (p->port_select != port && p->port_select != 0xFF)) {
+               pmp->status |= IB_SMP_INVALID_FIELD;
+               ret = reply((struct ib_smp *) pmp);
+               goto bail;
+       }
+
+       start = be32_to_cpu(p->sample_start);
+       if (start != 0) {
+               spin_lock_irqsave(&dev->pending_lock, flags);
+               if (dev->pma_sample_status == IB_PMA_SAMPLE_STATUS_DONE) {
+                       dev->pma_sample_status =
+                               IB_PMA_SAMPLE_STATUS_STARTED;
+                       dev->pma_sample_start = start;
+                       dev->pma_sample_interval =
+                               be32_to_cpu(p->sample_interval);
+                       dev->pma_tag = be16_to_cpu(p->tag);
+                       if (p->counter_select[0])
+                               dev->pma_counter_select[0] =
+                                       p->counter_select[0];
+                       if (p->counter_select[1])
+                               dev->pma_counter_select[1] =
+                                       p->counter_select[1];
+                       if (p->counter_select[2])
+                               dev->pma_counter_select[2] =
+                                       p->counter_select[2];
+                       if (p->counter_select[3])
+                               dev->pma_counter_select[3] =
+                                       p->counter_select[3];
+                       if (p->counter_select[4])
+                               dev->pma_counter_select[4] =
+                                       p->counter_select[4];
+               }
+               spin_unlock_irqrestore(&dev->pending_lock, flags);
+       }
+       ret = recv_pma_get_portsamplescontrol(pmp, ibdev, port);
+
+bail:
+       return ret;
+}
+
+static u64 get_counter(struct ipath_ibdev *dev, __be16 sel)
+{
+       u64 ret;
+
+       switch (sel) {
+       case IB_PMA_PORT_XMIT_DATA:
+               ret = dev->ipath_sword;
+               break;
+       case IB_PMA_PORT_RCV_DATA:
+               ret = dev->ipath_rword;
+               break;
+       case IB_PMA_PORT_XMIT_PKTS:
+               ret = dev->ipath_spkts;
+               break;
+       case IB_PMA_PORT_RCV_PKTS:
+               ret = dev->ipath_rpkts;
+               break;
+       case IB_PMA_PORT_XMIT_WAIT:
+               ret = dev->ipath_xmit_wait;
+               break;
+       default:
+               ret = 0;
+       }
+
+       return ret;
+}
+
+static int recv_pma_get_portsamplesresult(struct ib_perf *pmp,
+                                         struct ib_device *ibdev)
+{
+       struct ib_pma_portsamplesresult *p =
+               (struct ib_pma_portsamplesresult *)pmp->data;
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       int i;
+
+       memset(pmp->data, 0, sizeof(pmp->data));
+       p->tag = cpu_to_be16(dev->pma_tag);
+       p->sample_status = cpu_to_be16(dev->pma_sample_status);
+       for (i = 0; i < ARRAY_SIZE(dev->pma_counter_select); i++)
+               p->counter[i] = cpu_to_be32(
+                       get_counter(dev, dev->pma_counter_select[i]));
+
+       return reply((struct ib_smp *) pmp);
+}
+
+static int recv_pma_get_portsamplesresult_ext(struct ib_perf *pmp,
+                                             struct ib_device *ibdev)
+{
+       struct ib_pma_portsamplesresult_ext *p =
+               (struct ib_pma_portsamplesresult_ext *)pmp->data;
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       int i;
+
+       memset(pmp->data, 0, sizeof(pmp->data));
+       p->tag = cpu_to_be16(dev->pma_tag);
+       p->sample_status = cpu_to_be16(dev->pma_sample_status);
+       /* 64 bits */
+       p->extended_width = __constant_cpu_to_be32(0x80000000);
+       for (i = 0; i < ARRAY_SIZE(dev->pma_counter_select); i++)
+               p->counter[i] = cpu_to_be64(
+                       get_counter(dev, dev->pma_counter_select[i]));
+
+       return reply((struct ib_smp *) pmp);
+}
+
+static int recv_pma_get_portcounters(struct ib_perf *pmp,
+                                    struct ib_device *ibdev, u8 port)
+{
+       struct ib_pma_portcounters *p = (struct ib_pma_portcounters *)
+               pmp->data;
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       struct ipath_layer_counters cntrs;
+       u8 port_select = p->port_select;
+
+       ipath_layer_get_counters(dev->dd, &cntrs);
+
+       /* Adjust counters for any resets done. */
+       cntrs.symbol_error_counter -= dev->n_symbol_error_counter;
+       cntrs.link_error_recovery_counter -=
+               dev->n_link_error_recovery_counter;
+       cntrs.link_downed_counter -= dev->n_link_downed_counter;
+       cntrs.port_rcv_errors += dev->rcv_errors;
+       cntrs.port_rcv_errors -= dev->n_port_rcv_errors;
+       cntrs.port_rcv_remphys_errors -= dev->n_port_rcv_remphys_errors;
+       cntrs.port_xmit_discards -= dev->n_port_xmit_discards;
+       cntrs.port_xmit_data -= dev->n_port_xmit_data;
+       cntrs.port_rcv_data -= dev->n_port_rcv_data;
+       cntrs.port_xmit_packets -= dev->n_port_xmit_packets;
+       cntrs.port_rcv_packets -= dev->n_port_rcv_packets;
+
+       memset(pmp->data, 0, sizeof(pmp->data));
+
+       p->port_select = port_select;
+       if (pmp->attr_mod != 0 ||
+           (port_select != port && port_select != 0xFF))
+               pmp->status |= IB_SMP_INVALID_FIELD;
+
+       if (cntrs.symbol_error_counter > 0xFFFFUL)
+               p->symbol_error_counter = __constant_cpu_to_be16(0xFFFF);
+       else
+               p->symbol_error_counter =
+                       cpu_to_be16((u16)cntrs.symbol_error_counter);
+       if (cntrs.link_error_recovery_counter > 0xFFUL)
+               p->link_error_recovery_counter = 0xFF;
+       else
+               p->link_error_recovery_counter =
+                       (u8)cntrs.link_error_recovery_counter;
+       if (cntrs.link_downed_counter > 0xFFUL)
+               p->link_downed_counter = 0xFF;
+       else
+               p->link_downed_counter = (u8)cntrs.link_downed_counter;
+       if (cntrs.port_rcv_errors > 0xFFFFUL)
+               p->port_rcv_errors = __constant_cpu_to_be16(0xFFFF);
+       else
+               p->port_rcv_errors =
+                       cpu_to_be16((u16) cntrs.port_rcv_errors);
+       if (cntrs.port_rcv_remphys_errors > 0xFFFFUL)
+               p->port_rcv_remphys_errors = __constant_cpu_to_be16(0xFFFF);
+       else
+               p->port_rcv_remphys_errors =
+                       cpu_to_be16((u16)cntrs.port_rcv_remphys_errors);
+       if (cntrs.port_xmit_discards > 0xFFFFUL)
+               p->port_xmit_discards = __constant_cpu_to_be16(0xFFFF);
+       else
+               p->port_xmit_discards =
+                       cpu_to_be16((u16)cntrs.port_xmit_discards);
+       if (cntrs.port_xmit_data > 0xFFFFFFFFUL)
+               p->port_xmit_data = __constant_cpu_to_be32(0xFFFFFFFF);
+       else
+               p->port_xmit_data = cpu_to_be32((u32)cntrs.port_xmit_data);
+       if (cntrs.port_rcv_data > 0xFFFFFFFFUL)
+               p->port_rcv_data = __constant_cpu_to_be32(0xFFFFFFFF);
+       else
+               p->port_rcv_data = cpu_to_be32((u32)cntrs.port_rcv_data);
+       if (cntrs.port_xmit_packets > 0xFFFFFFFFUL)
+               p->port_xmit_packets = __constant_cpu_to_be32(0xFFFFFFFF);
+       else
+               p->port_xmit_packets =
+                       cpu_to_be32((u32)cntrs.port_xmit_packets);
+       if (cntrs.port_rcv_packets > 0xFFFFFFFFUL)
+               p->port_rcv_packets = __constant_cpu_to_be32(0xFFFFFFFF);
+       else
+               p->port_rcv_packets =
+                       cpu_to_be32((u32) cntrs.port_rcv_packets);
+
+       return reply((struct ib_smp *) pmp);
+}
+
+static int recv_pma_get_portcounters_ext(struct ib_perf *pmp,
+                                        struct ib_device *ibdev, u8 port)
+{
+       struct ib_pma_portcounters_ext *p =
+               (struct ib_pma_portcounters_ext *)pmp->data;
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       u64 swords, rwords, spkts, rpkts, xwait;
+       u8 port_select = p->port_select;
+
+       ipath_layer_snapshot_counters(dev->dd, &swords, &rwords, &spkts,
+                                     &rpkts, &xwait);
+
+       /* Adjust counters for any resets done. */
+       swords -= dev->n_port_xmit_data;
+       rwords -= dev->n_port_rcv_data;
+       spkts -= dev->n_port_xmit_packets;
+       rpkts -= dev->n_port_rcv_packets;
+
+       memset(pmp->data, 0, sizeof(pmp->data));
+
+       p->port_select = port_select;
+       if (pmp->attr_mod != 0 ||
+           (port_select != port && port_select != 0xFF))
+               pmp->status |= IB_SMP_INVALID_FIELD;
+
+       p->port_xmit_data = cpu_to_be64(swords);
+       p->port_rcv_data = cpu_to_be64(rwords);
+       p->port_xmit_packets = cpu_to_be64(spkts);
+       p->port_rcv_packets = cpu_to_be64(rpkts);
+       p->port_unicast_xmit_packets = cpu_to_be64(dev->n_unicast_xmit);
+       p->port_unicast_rcv_packets = cpu_to_be64(dev->n_unicast_rcv);
+       p->port_multicast_xmit_packets = cpu_to_be64(dev->n_multicast_xmit);
+       p->port_multicast_rcv_packets = cpu_to_be64(dev->n_multicast_rcv);
+
+       return reply((struct ib_smp *) pmp);
+}
+
+static int recv_pma_set_portcounters(struct ib_perf *pmp,
+                                    struct ib_device *ibdev, u8 port)
+{
+       struct ib_pma_portcounters *p = (struct ib_pma_portcounters *)
+               pmp->data;
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       struct ipath_layer_counters cntrs;
+
+       /*
+        * Since the HW doesn't support clearing counters, we save the
+        * current count and subtract it from future responses.
+        */
+       ipath_layer_get_counters(dev->dd, &cntrs);
+
+       if (p->counter_select & IB_PMA_SEL_SYMBOL_ERROR)
+               dev->n_symbol_error_counter = cntrs.symbol_error_counter;
+
+       if (p->counter_select & IB_PMA_SEL_LINK_ERROR_RECOVERY)
+               dev->n_link_error_recovery_counter =
+                       cntrs.link_error_recovery_counter;
+
+       if (p->counter_select & IB_PMA_SEL_LINK_DOWNED)
+               dev->n_link_downed_counter = cntrs.link_downed_counter;
+
+       if (p->counter_select & IB_PMA_SEL_PORT_RCV_ERRORS)
+               dev->n_port_rcv_errors =
+                       cntrs.port_rcv_errors + dev->rcv_errors;
+
+       if (p->counter_select & IB_PMA_SEL_PORT_RCV_REMPHYS_ERRORS)
+               dev->n_port_rcv_remphys_errors =
+                       cntrs.port_rcv_remphys_errors;
+
+       if (p->counter_select & IB_PMA_SEL_PORT_XMIT_DISCARDS)
+               dev->n_port_xmit_discards = cntrs.port_xmit_discards;
+
+       if (p->counter_select & IB_PMA_SEL_PORT_XMIT_DATA)
+               dev->n_port_xmit_data = cntrs.port_xmit_data;
+
+       if (p->counter_select & IB_PMA_SEL_PORT_RCV_DATA)
+               dev->n_port_rcv_data = cntrs.port_rcv_data;
+
+       if (p->counter_select & IB_PMA_SEL_PORT_XMIT_PACKETS)
+               dev->n_port_xmit_packets = cntrs.port_xmit_packets;
+
+       if (p->counter_select & IB_PMA_SEL_PORT_RCV_PACKETS)
+               dev->n_port_rcv_packets = cntrs.port_rcv_packets;
+
+       return recv_pma_get_portcounters(pmp, ibdev, port);
+}
+
+static int recv_pma_set_portcounters_ext(struct ib_perf *pmp,
+                                        struct ib_device *ibdev, u8 port)
+{
+       struct ib_pma_portcounters *p = (struct ib_pma_portcounters *)
+               pmp->data;
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       u64 swords, rwords, spkts, rpkts, xwait;
+
+       ipath_layer_snapshot_counters(dev->dd, &swords, &rwords, &spkts,
+                                     &rpkts, &xwait);
+
+       if (p->counter_select & IB_PMA_SELX_PORT_XMIT_DATA)
+               dev->n_port_xmit_data = swords;
+
+       if (p->counter_select & IB_PMA_SELX_PORT_RCV_DATA)
+               dev->n_port_rcv_data = rwords;
+
+       if (p->counter_select & IB_PMA_SELX_PORT_XMIT_PACKETS)
+               dev->n_port_xmit_packets = spkts;
+
+       if (p->counter_select & IB_PMA_SELX_PORT_RCV_PACKETS)
+               dev->n_port_rcv_packets = rpkts;
+
+       if (p->counter_select & IB_PMA_SELX_PORT_UNI_XMIT_PACKETS)
+               dev->n_unicast_xmit = 0;
+
+       if (p->counter_select & IB_PMA_SELX_PORT_UNI_RCV_PACKETS)
+               dev->n_unicast_rcv = 0;
+
+       if (p->counter_select & IB_PMA_SELX_PORT_MULTI_XMIT_PACKETS)
+               dev->n_multicast_xmit = 0;
+
+       if (p->counter_select & IB_PMA_SELX_PORT_MULTI_RCV_PACKETS)
+               dev->n_multicast_rcv = 0;
+
+       return recv_pma_get_portcounters_ext(pmp, ibdev, port);
+}
+
+static int process_subn(struct ib_device *ibdev, int mad_flags,
+                       u8 port_num, struct ib_mad *in_mad,
+                       struct ib_mad *out_mad)
+{
+       struct ib_smp *smp = (struct ib_smp *)out_mad;
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       int ret;
+
+       *out_mad = *in_mad;
+       if (smp->class_version != 1) {
+               smp->status |= IB_SMP_UNSUP_VERSION;
+               ret = reply(smp);
+               goto bail;
+       }
+
+       /* Is the mkey in the process of expiring? */
+       if (dev->mkey_lease_timeout && jiffies >= dev->mkey_lease_timeout) {
+               /* Clear timeout and mkey protection field. */
+               dev->mkey_lease_timeout = 0;
+               dev->mkeyprot_resv_lmc &= 0x3F;
+       }
+
+       /*
+        * M_Key checking depends on
+        * Portinfo:M_Key_protect_bits
+        */
+       if ((mad_flags & IB_MAD_IGNORE_MKEY) == 0 && dev->mkey != 0 &&
+           dev->mkey != smp->mkey &&
+           (smp->method == IB_MGMT_METHOD_SET ||
+            (smp->method == IB_MGMT_METHOD_GET &&
+             (dev->mkeyprot_resv_lmc >> 7) != 0))) {
+               if (dev->mkey_violations != 0xFFFF)
+                       ++dev->mkey_violations;
+               if (dev->mkey_lease_timeout ||
+                   dev->mkey_lease_period == 0) {
+                       ret = IB_MAD_RESULT_SUCCESS |
+                               IB_MAD_RESULT_CONSUMED;
+                       goto bail;
+               }
+               dev->mkey_lease_timeout = jiffies +
+                       dev->mkey_lease_period * HZ;
+               /* Future: Generate a trap notice. */
+               ret = IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED;
+               goto bail;
+       } else if (dev->mkey_lease_timeout)
+               dev->mkey_lease_timeout = 0;
+
+       switch (smp->method) {
+       case IB_MGMT_METHOD_GET:
+               switch (smp->attr_id) {
+               case IB_SMP_ATTR_NODE_DESC:
+                       ret = recv_subn_get_nodedescription(smp, ibdev);
+                       goto bail;
+               case IB_SMP_ATTR_NODE_INFO:
+                       ret = recv_subn_get_nodeinfo(smp, ibdev, port_num);
+                       goto bail;
+               case IB_SMP_ATTR_GUID_INFO:
+                       ret = recv_subn_get_guidinfo(smp, ibdev);
+                       goto bail;
+               case IB_SMP_ATTR_PORT_INFO:
+                       ret = recv_subn_get_portinfo(smp, ibdev, port_num);
+                       goto bail;
+               case IB_SMP_ATTR_PKEY_TABLE:
+                       ret = recv_subn_get_pkeytable(smp, ibdev);
+                       goto bail;
+               case IB_SMP_ATTR_SM_INFO:
+                       if (dev->port_cap_flags & IB_PORT_SM_DISABLED) {
+                               ret = IB_MAD_RESULT_SUCCESS |
+                                       IB_MAD_RESULT_CONSUMED;
+                               goto bail;
+                       }
+                       if (dev->port_cap_flags & IB_PORT_SM) {
+                               ret = IB_MAD_RESULT_SUCCESS;
+                               goto bail;
+                       }
+                       /* FALLTHROUGH */
+               default:
+                       smp->status |= IB_SMP_UNSUP_METH_ATTR;
+                       ret = reply(smp);
+                       goto bail;
+               }
+
+       case IB_MGMT_METHOD_SET:
+               switch (smp->attr_id) {
+               case IB_SMP_ATTR_GUID_INFO:
+                       ret = recv_subn_set_guidinfo(smp, ibdev);
+                       goto bail;
+               case IB_SMP_ATTR_PORT_INFO:
+                       ret = recv_subn_set_portinfo(smp, ibdev, port_num);
+                       goto bail;
+               case IB_SMP_ATTR_PKEY_TABLE:
+                       ret = recv_subn_set_pkeytable(smp, ibdev);
+                       goto bail;
+               case IB_SMP_ATTR_SM_INFO:
+                       if (dev->port_cap_flags & IB_PORT_SM_DISABLED) {
+                               ret = IB_MAD_RESULT_SUCCESS |
+                                       IB_MAD_RESULT_CONSUMED;
+                               goto bail;
+                       }
+                       if (dev->port_cap_flags & IB_PORT_SM) {
+                               ret = IB_MAD_RESULT_SUCCESS;
+                               goto bail;
+                       }
+                       /* FALLTHROUGH */
+               default:
+                       smp->status |= IB_SMP_UNSUP_METH_ATTR;
+                       ret = reply(smp);
+                       goto bail;
+               }
+
+       case IB_MGMT_METHOD_GET_RESP:
+               /*
+                * The ib_mad module will call us to process responses
+                * before checking for other consumers.
+                * Just tell the caller to process it normally.
+                */
+               ret = IB_MAD_RESULT_FAILURE;
+               goto bail;
+       default:
+               smp->status |= IB_SMP_UNSUP_METHOD;
+               ret = reply(smp);
+       }
+
+bail:
+       return ret;
+}
+
+static int process_perf(struct ib_device *ibdev, u8 port_num,
+                       struct ib_mad *in_mad,
+                       struct ib_mad *out_mad)
+{
+       struct ib_perf *pmp = (struct ib_perf *)out_mad;
+       int ret;
+
+       *out_mad = *in_mad;
+       if (pmp->class_version != 1) {
+               pmp->status |= IB_SMP_UNSUP_VERSION;
+               ret = reply((struct ib_smp *) pmp);
+               goto bail;
+       }
+
+       switch (pmp->method) {
+       case IB_MGMT_METHOD_GET:
+               switch (pmp->attr_id) {
+               case IB_PMA_CLASS_PORT_INFO:
+                       ret = recv_pma_get_classportinfo(pmp);
+                       goto bail;
+               case IB_PMA_PORT_SAMPLES_CONTROL:
+                       ret = recv_pma_get_portsamplescontrol(pmp, ibdev,
+                                                             port_num);
+                       goto bail;
+               case IB_PMA_PORT_SAMPLES_RESULT:
+                       ret = recv_pma_get_portsamplesresult(pmp, ibdev);
+                       goto bail;
+               case IB_PMA_PORT_SAMPLES_RESULT_EXT:
+                       ret = recv_pma_get_portsamplesresult_ext(pmp,
+                                                                ibdev);
+                       goto bail;
+               case IB_PMA_PORT_COUNTERS:
+                       ret = recv_pma_get_portcounters(pmp, ibdev,
+                                                       port_num);
+                       goto bail;
+               case IB_PMA_PORT_COUNTERS_EXT:
+                       ret = recv_pma_get_portcounters_ext(pmp, ibdev,
+                                                           port_num);
+                       goto bail;
+               default:
+                       pmp->status |= IB_SMP_UNSUP_METH_ATTR;
+                       ret = reply((struct ib_smp *) pmp);
+                       goto bail;
+               }
+
+       case IB_MGMT_METHOD_SET:
+               switch (pmp->attr_id) {
+               case IB_PMA_PORT_SAMPLES_CONTROL:
+                       ret = recv_pma_set_portsamplescontrol(pmp, ibdev,
+                                                             port_num);
+                       goto bail;
+               case IB_PMA_PORT_COUNTERS:
+                       ret = recv_pma_set_portcounters(pmp, ibdev,
+                                                       port_num);
+                       goto bail;
+               case IB_PMA_PORT_COUNTERS_EXT:
+                       ret = recv_pma_set_portcounters_ext(pmp, ibdev,
+                                                           port_num);
+                       goto bail;
+               default:
+                       pmp->status |= IB_SMP_UNSUP_METH_ATTR;
+                       ret = reply((struct ib_smp *) pmp);
+                       goto bail;
+               }
+
+       case IB_MGMT_METHOD_GET_RESP:
+               /*
+                * The ib_mad module will call us to process responses
+                * before checking for other consumers.
+                * Just tell the caller to process it normally.
+                */
+               ret = IB_MAD_RESULT_FAILURE;
+               goto bail;
+       default:
+               pmp->status |= IB_SMP_UNSUP_METHOD;
+               ret = reply((struct ib_smp *) pmp);
+       }
+
+bail:
+       return ret;
+}
+
+/**
+ * ipath_process_mad - process an incoming MAD packet
+ * @ibdev: the infiniband device this packet came in on
+ * @mad_flags: MAD flags
+ * @port_num: the port number this packet came in on
+ * @in_wc: the work completion entry for this packet
+ * @in_grh: the global route header for this packet
+ * @in_mad: the incoming MAD
+ * @out_mad: any outgoing MAD reply
+ *
+ * Returns IB_MAD_RESULT_SUCCESS if this is a MAD that we are not
+ * interested in processing.
+ *
+ * Note that the verbs framework has already done the MAD sanity checks,
+ * and hop count/pointer updating for IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
+ * MADs.
+ *
+ * This is called by the ib_mad module.
+ */
+int ipath_process_mad(struct ib_device *ibdev, int mad_flags, u8 port_num,
+                     struct ib_wc *in_wc, struct ib_grh *in_grh,
+                     struct ib_mad *in_mad, struct ib_mad *out_mad)
+{
+       struct ipath_ibdev *dev = to_idev(ibdev);
+       int ret;
+
+       /*
+        * Snapshot current HW counters to "clear" them.
+        * This should be done when the driver is loaded except that for
+        * some reason we get a zillion errors when brining up the link.
+        */
+       if (dev->rcv_errors == 0) {
+               struct ipath_layer_counters cntrs;
+
+               ipath_layer_get_counters(to_idev(ibdev)->dd, &cntrs);
+               dev->rcv_errors++;
+               dev->n_symbol_error_counter = cntrs.symbol_error_counter;
+               dev->n_link_error_recovery_counter =
+                       cntrs.link_error_recovery_counter;
+               dev->n_link_downed_counter = cntrs.link_downed_counter;
+               dev->n_port_rcv_errors = cntrs.port_rcv_errors + 1;
+               dev->n_port_rcv_remphys_errors =
+                       cntrs.port_rcv_remphys_errors;
+               dev->n_port_xmit_discards = cntrs.port_xmit_discards;
+               dev->n_port_xmit_data = cntrs.port_xmit_data;
+               dev->n_port_rcv_data = cntrs.port_rcv_data;
+               dev->n_port_xmit_packets = cntrs.port_xmit_packets;
+               dev->n_port_rcv_packets = cntrs.port_rcv_packets;
+       }
+       switch (in_mad->mad_hdr.mgmt_class) {
+       case IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE:
+       case IB_MGMT_CLASS_SUBN_LID_ROUTED:
+               ret = process_subn(ibdev, mad_flags, port_num,
+                                  in_mad, out_mad);
+               goto bail;
+       case IB_MGMT_CLASS_PERF_MGMT:
+               ret = process_perf(ibdev, port_num, in_mad, out_mad);
+               goto bail;
+       default:
+               ret = IB_MAD_RESULT_SUCCESS;
+       }
+
+bail:
+       return ret;
+}