Whamcloud - gitweb
LU-5570 lnet: check router aliveness timestamp
[fs/lustre-release.git] / lnet / lnet / router_proc.c
index fce91c2..95d34c9 100644 (file)
@@ -1,7 +1,7 @@
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
+/*
+ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  *
- * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2011, 2013, Intel Corporation.
  *
  *   This file is part of Portals
  *   http://sourceforge.net/projects/sandiaportals/
 
 #if defined(__KERNEL__) && defined(LNET_ROUTER)
 
-/* this is really lnet_proc.c */
+/* This is really lnet_proc.c. You might need to update sanity test 215
+ * if any file format is changed. */
 
-static cfs_sysctl_table_header_t *lnet_table_header = NULL;
+static struct ctl_table_header *lnet_table_header = NULL;
 
-#ifndef HAVE_SYSCTL_UNNUMBERED
-#define CTL_LNET         (0x100)
-enum {
-        PSDEV_LNET_STATS = 100,
-        PSDEV_LNET_ROUTES,
-        PSDEV_LNET_ROUTERS,
-        PSDEV_LNET_PEERS,
-        PSDEV_LNET_BUFFERS,
-        PSDEV_LNET_NIS,
-};
-#else
-#define CTL_LNET           CTL_UNNUMBERED
-#define PSDEV_LNET_STATS   CTL_UNNUMBERED
-#define PSDEV_LNET_ROUTES  CTL_UNNUMBERED
-#define PSDEV_LNET_ROUTERS CTL_UNNUMBERED
-#define PSDEV_LNET_PEERS   CTL_UNNUMBERED
-#define PSDEV_LNET_BUFFERS CTL_UNNUMBERED
-#define PSDEV_LNET_NIS     CTL_UNNUMBERED
-#endif
+#define LNET_LOFFT_BITS                (sizeof(loff_t) * 8)
+/*
+ * NB: max allowed LNET_CPT_BITS is 8 on 64-bit system and 2 on 32-bit system
+ */
+#define LNET_PROC_CPT_BITS     (LNET_CPT_BITS + 1)
+/* change version, 16 bits or 8 bits */
+#define LNET_PROC_VER_BITS     MAX(((MIN(LNET_LOFFT_BITS, 64)) / 4), 8)
+
+#define LNET_PROC_HASH_BITS    LNET_PEER_HASH_BITS
+/*
+ * bits for peer hash offset
+ * NB: we don't use the highest bit of *ppos because it's signed
+ */
+#define LNET_PROC_HOFF_BITS    (LNET_LOFFT_BITS -       \
+                                LNET_PROC_CPT_BITS -    \
+                                LNET_PROC_VER_BITS -    \
+                                LNET_PROC_HASH_BITS - 1)
+/* bits for hash index + position */
+#define LNET_PROC_HPOS_BITS    (LNET_PROC_HASH_BITS + LNET_PROC_HOFF_BITS)
+/* bits for peer hash table + hash version */
+#define LNET_PROC_VPOS_BITS    (LNET_PROC_HPOS_BITS + LNET_PROC_VER_BITS)
+
+#define LNET_PROC_CPT_MASK     ((1ULL << LNET_PROC_CPT_BITS) - 1)
+#define LNET_PROC_VER_MASK     ((1ULL << LNET_PROC_VER_BITS) - 1)
+#define LNET_PROC_HASH_MASK    ((1ULL << LNET_PROC_HASH_BITS) - 1)
+#define LNET_PROC_HOFF_MASK    ((1ULL << LNET_PROC_HOFF_BITS) - 1)
+
+#define LNET_PROC_CPT_GET(pos)                         \
+       (int)(((pos) >> LNET_PROC_VPOS_BITS) & LNET_PROC_CPT_MASK)
+
+#define LNET_PROC_VER_GET(pos)                         \
+       (int)(((pos) >> LNET_PROC_HPOS_BITS) & LNET_PROC_VER_MASK)
+
+#define LNET_PROC_HASH_GET(pos)                                \
+       (int)(((pos) >> LNET_PROC_HOFF_BITS) & LNET_PROC_HASH_MASK)
+
+#define LNET_PROC_HOFF_GET(pos)                                \
+       (int)((pos) & LNET_PROC_HOFF_MASK)
+
+#define LNET_PROC_POS_MAKE(cpt, ver, hash, off)                \
+       (((((loff_t)(cpt)) & LNET_PROC_CPT_MASK) << LNET_PROC_VPOS_BITS) |   \
+       ((((loff_t)(ver)) & LNET_PROC_VER_MASK) << LNET_PROC_HPOS_BITS) |   \
+       ((((loff_t)(hash)) & LNET_PROC_HASH_MASK) << LNET_PROC_HOFF_BITS) | \
+       ((off) & LNET_PROC_HOFF_MASK))
+
+#define LNET_PROC_VERSION(v)   ((unsigned int)((v) & LNET_PROC_VER_MASK))
 
 static int __proc_lnet_stats(void *data, int write,
-                             loff_t pos, void *buffer, int nob)
+                            loff_t pos, void __user *buffer, int nob)
 {
         int              rc;
         lnet_counters_t *ctrs;
@@ -61,10 +89,8 @@ static int __proc_lnet_stats(void *data, int write,
         const int        tmpsiz = 256; /* 7 %u and 4 LPU64 */
 
         if (write) {
-                LNET_LOCK();
-                memset(&the_lnet.ln_counters, 0, sizeof(the_lnet.ln_counters));
-                LNET_UNLOCK();
-                return 0;
+               lnet_counters_reset();
+               return 0;
         }
 
         /* read */
@@ -79,9 +105,7 @@ static int __proc_lnet_stats(void *data, int write,
                 return -ENOMEM;
         }
 
-        LNET_LOCK();
-        *ctrs = the_lnet.ln_counters;
-        LNET_UNLOCK();
+       lnet_counters_get(ctrs);
 
         len = snprintf(tmpstr, tmpsiz,
                        "%u %u %u %u %u %u %u "LPU64" "LPU64" "
@@ -96,26 +120,38 @@ static int __proc_lnet_stats(void *data, int write,
         if (pos >= min_t(int, len, strlen(tmpstr)))
                 rc = 0;
         else
-                rc = trace_copyout_string(buffer, nob,
-                                          tmpstr + pos, "\n");
+                rc = cfs_trace_copyout_string(buffer, nob,
+                                              tmpstr + pos, "\n");
 
         LIBCFS_FREE(tmpstr, tmpsiz);
         LIBCFS_FREE(ctrs, sizeof(*ctrs));
         return rc;
 }
 
-DECLARE_PROC_HANDLER(proc_lnet_stats);
+static int
+proc_lnet_stats(struct ctl_table *table, int write, void __user *buffer,
+               size_t *lenp, loff_t *ppos)
+{
+       return lprocfs_call_handler(table->data, write, ppos, buffer, lenp,
+                                   __proc_lnet_stats);
+}
 
-int LL_PROC_PROTO(proc_lnet_routes)
+static int
+proc_lnet_routes(struct ctl_table *table, int write, void __user *buffer,
+                size_t *lenp, loff_t *ppos)
 {
-        int        rc     = 0;
-        char      *tmpstr;
-        char      *s;
-        const int  tmpsiz = 256;
-        int        len;
-        int       *ver_p  = (unsigned int *)(&filp->private_data);
+       const int       tmpsiz = 256;
+       char            *tmpstr;
+       char            *s;
+       int             rc = 0;
+       int             len;
+       int             ver;
+       int             off;
+
+       CLASSERT(sizeof(loff_t) >= 4);
 
-        DECLARE_LL_PROC_PPOS_DECL;
+       off = LNET_PROC_HOFF_GET(*ppos);
+       ver = LNET_PROC_VER_GET(*ppos);
 
         LASSERT (!write);
 
@@ -133,63 +169,77 @@ int LL_PROC_PROTO(proc_lnet_routes)
                               the_lnet.ln_routing ? "enabled" : "disabled");
                 LASSERT (tmpstr + tmpsiz - s > 0);
 
-                s += snprintf(s, tmpstr + tmpsiz - s, "%-8s %4s %7s %s\n",
-                              "net", "hops", "state", "router");
+               s += snprintf(s, tmpstr + tmpsiz - s, "%-8s %4s %8s %7s %s\n",
+                             "net", "hops", "priority", "state", "router");
                 LASSERT (tmpstr + tmpsiz - s > 0);
 
-                LNET_LOCK();
-                *ver_p = (unsigned int)the_lnet.ln_remote_nets_version;
-                LNET_UNLOCK();
-        } else {
-                struct list_head  *n;
-                struct list_head  *r;
-                lnet_route_t      *route = NULL;
-                lnet_remotenet_t  *rnet  = NULL;
-                int                skip  = *ppos - 1;
-
-                LNET_LOCK();
-
-                if (*ver_p != (unsigned int)the_lnet.ln_remote_nets_version) {
-                        LNET_UNLOCK();
-                        LIBCFS_FREE(tmpstr, tmpsiz);
-                        return -ESTALE;
-                }
-
-                n = the_lnet.ln_remote_nets.next;
-
-                while (n != &the_lnet.ln_remote_nets && route == NULL) {
-                        rnet = list_entry(n, lnet_remotenet_t, lrn_list);
-
-                        r = rnet->lrn_routes.next;
-
-                        while (r != &rnet->lrn_routes) {
-                                lnet_route_t *re = list_entry(r, lnet_route_t,
-                                                              lr_list);
-                                if (skip == 0) {
-                                        route = re;
-                                        break;
-                                } else
-                                        skip--;
-
-                                r = r->next;
-                        }
-
-                        n = n->next;
-                }
+               lnet_net_lock(0);
+               ver = (unsigned int)the_lnet.ln_remote_nets_version;
+               lnet_net_unlock(0);
+               *ppos = LNET_PROC_POS_MAKE(0, ver, 0, off);
+       } else {
+               struct list_head        *n;
+               struct list_head        *r;
+               lnet_route_t            *route = NULL;
+               lnet_remotenet_t        *rnet  = NULL;
+               int                     skip  = off - 1;
+               struct list_head        *rn_list;
+               int                     i;
+
+               lnet_net_lock(0);
+
+               if (ver != LNET_PROC_VERSION(the_lnet.ln_remote_nets_version)) {
+                       lnet_net_unlock(0);
+                       LIBCFS_FREE(tmpstr, tmpsiz);
+                       return -ESTALE;
+               }
+
+               for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE && route == NULL;
+                    i++) {
+                       rn_list = &the_lnet.ln_remote_nets_hash[i];
+
+                       n = rn_list->next;
+
+                       while (n != rn_list && route == NULL) {
+                               rnet = list_entry(n, lnet_remotenet_t,
+                                                 lrn_list);
+
+                               r = rnet->lrn_routes.next;
+
+                               while (r != &rnet->lrn_routes) {
+                                       lnet_route_t *re =
+                                               list_entry(r, lnet_route_t,
+                                                          lr_list);
+                                       if (skip == 0) {
+                                               route = re;
+                                               break;
+                                       }
+
+                                       skip--;
+                                       r = r->next;
+                               }
+
+                               n = n->next;
+                       }
+               }
 
                 if (route != NULL) {
-                        __u32        net   = rnet->lrn_net;
-                        unsigned int hops  = route->lr_hops;
-                        lnet_nid_t   nid   = route->lr_gateway->lp_nid;
-                        int          alive = route->lr_gateway->lp_alive;
-
-                        s += snprintf(s, tmpstr + tmpsiz - s, "%-8s %4u %7s %s\n",
-                                      libcfs_net2str(net), hops,
-                                      alive ? "up" : "down", libcfs_nid2str(nid));
-                        LASSERT (tmpstr + tmpsiz - s > 0);
-                }
-
-                LNET_UNLOCK();
+                       __u32        net        = rnet->lrn_net;
+                       unsigned int hops       = route->lr_hops;
+                       unsigned int priority   = route->lr_priority;
+                       lnet_nid_t   nid        = route->lr_gateway->lp_nid;
+                       int          alive      = lnet_is_route_alive(route);
+
+                       s += snprintf(s, tmpstr + tmpsiz - s,
+                                     "%-8s %4u %8u %7s %s\n",
+                                     libcfs_net2str(net), hops,
+                                     priority,
+                                     alive ? "up" : "down",
+                                     libcfs_nid2str(nid));
+                       LASSERT(tmpstr + tmpsiz - s > 0);
+               }
+
+               lnet_net_unlock(0);
         }
 
         len = s - tmpstr;     /* how many bytes was written */
@@ -197,10 +247,12 @@ int LL_PROC_PROTO(proc_lnet_routes)
         if (len > *lenp) {    /* linux-supplied buffer is too small */
                 rc = -EINVAL;
         } else if (len > 0) { /* wrote something */
-                if (copy_to_user(buffer, tmpstr, len))
+               if (copy_to_user(buffer, tmpstr, len))
                         rc = -EFAULT;
-                else
-                        *ppos += 1;
+                else {
+                       off += 1;
+                       *ppos = LNET_PROC_POS_MAKE(0, ver, 0, off);
+                }
         }
 
         LIBCFS_FREE(tmpstr, tmpsiz);
@@ -211,16 +263,20 @@ int LL_PROC_PROTO(proc_lnet_routes)
         return rc;
 }
 
-int LL_PROC_PROTO(proc_lnet_routers)
+static int
+proc_lnet_routers(struct ctl_table *table, int write, void __user *buffer,
+                 size_t *lenp, loff_t *ppos)
 {
         int        rc = 0;
         char      *tmpstr;
         char      *s;
         const int  tmpsiz = 256;
         int        len;
-        int       *ver_p = (unsigned int *)(&filp->private_data);
+        int        ver;
+       int        off;
 
-        DECLARE_LL_PROC_PPOS_DECL;
+       off = LNET_PROC_HOFF_GET(*ppos);
+       ver = LNET_PROC_VER_GET(*ppos);
 
         LASSERT (!write);
 
@@ -234,40 +290,43 @@ int LL_PROC_PROTO(proc_lnet_routers)
         s = tmpstr; /* points to current position in tmpstr[] */
 
         if (*ppos == 0) {
-                s += snprintf(s, tmpstr + tmpsiz - s,
-                              "%-4s %7s %9s %6s %12s %9s %8s %7s %s\n",
-                              "ref", "rtr_ref", "alive_cnt", "state", "last_ping",
-                              "ping_sent", "deadline", "down_ni", "router");
-                LASSERT (tmpstr + tmpsiz - s > 0);
-
-                LNET_LOCK();
-                *ver_p = (unsigned int)the_lnet.ln_routers_version;
-                LNET_UNLOCK();
-        } else {
-                struct list_head  *r;
-                lnet_peer_t       *peer = NULL;
-                int                skip = *ppos - 1;
+               s += snprintf(s, tmpstr + tmpsiz - s,
+                             "%-4s %7s %9s %6s %12s %9s %8s %7s %s\n",
+                             "ref", "rtr_ref", "alive_cnt", "state",
+                             "last_ping", "ping_sent", "deadline",
+                             "down_ni", "router");
+               LASSERT(tmpstr + tmpsiz - s > 0);
+
+               lnet_net_lock(0);
+               ver = (unsigned int)the_lnet.ln_routers_version;
+               lnet_net_unlock(0);
+               *ppos = LNET_PROC_POS_MAKE(0, ver, 0, off);
+       } else {
+               struct list_head *r;
+               struct lnet_peer *peer = NULL;
+               int               skip = off - 1;
+
+               lnet_net_lock(0);
+
+               if (ver != LNET_PROC_VERSION(the_lnet.ln_routers_version)) {
+                       lnet_net_unlock(0);
 
-                LNET_LOCK();
-
-                if (*ver_p != (unsigned int)the_lnet.ln_routers_version) {
-                        LNET_UNLOCK();
                         LIBCFS_FREE(tmpstr, tmpsiz);
                         return -ESTALE;
                 }
 
                 r = the_lnet.ln_routers.next;
 
-                while (r != &the_lnet.ln_routers) {
-                        lnet_peer_t *lp = list_entry(r, lnet_peer_t,
-                                                     lp_rtr_list);
+               while (r != &the_lnet.ln_routers) {
+                       lnet_peer_t *lp = list_entry(r, lnet_peer_t,
+                                                    lp_rtr_list);
 
                         if (skip == 0) {
                                 peer = lp;
-                                        break;
-                                } else
-                                        skip--;
+                                break;
+                        }
 
+                        skip--;
                         r = r->next;
                 }
 
@@ -280,8 +339,23 @@ int LL_PROC_PROTO(proc_lnet_routers)
                         int alive_cnt = peer->lp_alive_count;
                         int alive     = peer->lp_alive;
                         int pingsent  = !peer->lp_ping_notsent;
-                        int last_ping = cfs_duration_sec(now - peer->lp_ping_timestamp);
-                        int down_ni   = lnet_router_down_ni(peer, LNET_NIDNET(LNET_NID_ANY));
+                        int last_ping = cfs_duration_sec(cfs_time_sub(now,
+                                                     peer->lp_ping_timestamp));
+                       int down_ni   = 0;
+                       lnet_route_t *rtr;
+
+                       if ((peer->lp_ping_feats &
+                            LNET_PING_FEAT_NI_STATUS) != 0) {
+                               list_for_each_entry(rtr, &peer->lp_routes,
+                                                   lr_gwlist) {
+                                       /* downis on any route should be the
+                                        * number of downis on the gateway */
+                                       if (rtr->lr_downis != 0) {
+                                               down_ni = rtr->lr_downis;
+                                               break;
+                                       }
+                               }
+                       }
 
                         if (deadline == 0)
                                 s += snprintf(s, tmpstr + tmpsiz - s,
@@ -296,12 +370,12 @@ int LL_PROC_PROTO(proc_lnet_routers)
                                               nrefs, nrtrrefs, alive_cnt,
                                               alive ? "up" : "down", last_ping,
                                               pingsent,
-                                              cfs_duration_sec(deadline - now),
+                                              cfs_duration_sec(cfs_time_sub(deadline, now)),
                                               down_ni, libcfs_nid2str(nid));
                         LASSERT (tmpstr + tmpsiz - s > 0);
                 }
 
-                LNET_UNLOCK();
+               lnet_net_unlock(0);
         }
 
         len = s - tmpstr;     /* how many bytes was written */
@@ -309,10 +383,12 @@ int LL_PROC_PROTO(proc_lnet_routers)
         if (len > *lenp) {    /* linux-supplied buffer is too small */
                 rc = -EINVAL;
         } else if (len > 0) { /* wrote something */
-                if (copy_to_user(buffer, tmpstr, len))
+               if (copy_to_user(buffer, tmpstr, len))
                         rc = -EFAULT;
-                else
-                        *ppos += 1;
+                else {
+                       off += 1;
+                       *ppos = LNET_PROC_POS_MAKE(0, ver, 0, off);
+                }
         }
 
         LIBCFS_FREE(tmpstr, tmpsiz);
@@ -323,45 +399,31 @@ int LL_PROC_PROTO(proc_lnet_routers)
         return rc;
 }
 
-/*
- * NB: we don't use the highest bit of *ppos because it's signed;
- *     next 9 bits is used to stash idx (assuming that
- *     LNET_PEER_HASHSIZE < 512)
- */
-#define LNET_LOFFT_BITS (sizeof(loff_t) * 8)
-#define LNET_PHASH_BITS 9
-#define LNET_PHASH_IDX_MASK (((1ULL << LNET_PHASH_BITS) - 1) <<               \
-                             (LNET_LOFFT_BITS - LNET_PHASH_BITS - 1))
-#define LNET_PHASH_NUM_MASK ((1ULL <<                                         \
-                              (LNET_LOFFT_BITS - LNET_PHASH_BITS -1)) - 1)
-#define LNET_PHASH_IDX_GET(pos) (int)(((pos) & LNET_PHASH_IDX_MASK) >>  \
-                                      (LNET_LOFFT_BITS - LNET_PHASH_BITS -1))
-#define LNET_PHASH_NUM_GET(pos) (int)((pos) & LNET_PHASH_NUM_MASK)
-#define LNET_PHASH_POS_MAKE(idx, num) ((((loff_t)idx) << (LNET_LOFFT_BITS -   \
-                                                  LNET_PHASH_BITS -1)) | (num))
-
-int LL_PROC_PROTO(proc_lnet_peers)
+static int
+proc_lnet_peers(struct ctl_table *table, int write, void __user *buffer,
+               size_t *lenp, loff_t *ppos)
 {
-        int        rc = 0;
-        char      *tmpstr;
-        char      *s;
-        const int  tmpsiz      = 256;
-        int        len;
-        int       *ver_p       = (unsigned int *)(&filp->private_data);
-        int        idx;
-        int        num;
-
-        DECLARE_LL_PROC_PPOS_DECL;
-
-        idx = LNET_PHASH_IDX_GET(*ppos);
-        num = LNET_PHASH_NUM_GET(*ppos);
-
-        CLASSERT ((1 << LNET_PHASH_BITS) > LNET_PEER_HASHSIZE);
-
-        LASSERT (!write);
-
-        if (*lenp == 0)
-                return 0;
+       const int               tmpsiz  = 256;
+       struct lnet_peer_table  *ptable;
+       char                    *tmpstr;
+       char                    *s;
+       int                     cpt  = LNET_PROC_CPT_GET(*ppos);
+       int                     ver  = LNET_PROC_VER_GET(*ppos);
+       int                     hash = LNET_PROC_HASH_GET(*ppos);
+       int                     hoff = LNET_PROC_HOFF_GET(*ppos);
+       int                     rc = 0;
+       int                     len;
+
+       CLASSERT(LNET_PROC_HASH_BITS >= LNET_PEER_HASH_BITS);
+       LASSERT(!write);
+
+       if (*lenp == 0)
+               return 0;
+
+       if (cpt >= LNET_CPT_NUMBER) {
+               *lenp = 0;
+               return 0;
+       }
 
         LIBCFS_ALLOC(tmpstr, tmpsiz);
         if (tmpstr == NULL)
@@ -371,53 +433,57 @@ int LL_PROC_PROTO(proc_lnet_peers)
 
         if (*ppos == 0) {
                 s += snprintf(s, tmpstr + tmpsiz - s,
-                              "%-24s %4s %5s %5s %5s %5s %5s %5s %s\n",
-                              "nid", "refs", "state", "max",
+                              "%-24s %4s %5s %5s %5s %5s %5s %5s %5s %s\n",
+                              "nid", "refs", "state", "last", "max",
                               "rtr", "min", "tx", "min", "queue");
                 LASSERT (tmpstr + tmpsiz - s > 0);
 
-                LNET_LOCK();
-                *ver_p  = (unsigned int)the_lnet.ln_peertable_version;
-                LNET_UNLOCK();
-
-                num++;
-        } else {
-                struct list_head  *p    = NULL;
-                lnet_peer_t       *peer = NULL;
-                int                skip = num - 1;
-
-                LNET_LOCK();
-
-                if (*ver_p != (unsigned int)the_lnet.ln_peertable_version) {
-                        LNET_UNLOCK();
-                        LIBCFS_FREE(tmpstr, tmpsiz);
-                        return -ESTALE;
-                }
-
-                while (idx < LNET_PEER_HASHSIZE) {
-                        if (p == NULL)
-                                p = the_lnet.ln_peer_hash[idx].next;
-
-                        while (p != &the_lnet.ln_peer_hash[idx]) {
-                                lnet_peer_t *lp = list_entry(p, lnet_peer_t,
-                                                             lp_hashlist);
+               hoff++;
+       } else {
+               struct lnet_peer        *peer;
+               struct list_head        *p;
+               int                     skip;
+ again:
+               p = NULL;
+               peer = NULL;
+               skip = hoff - 1;
+
+               lnet_net_lock(cpt);
+               ptable = the_lnet.ln_peer_tables[cpt];
+               if (hoff == 1)
+                       ver = LNET_PROC_VERSION(ptable->pt_version);
+
+               if (ver != LNET_PROC_VERSION(ptable->pt_version)) {
+                       lnet_net_unlock(cpt);
+                       LIBCFS_FREE(tmpstr, tmpsiz);
+                       return -ESTALE;
+               }
+
+               while (hash < LNET_PEER_HASH_SIZE) {
+                       if (p == NULL)
+                               p = ptable->pt_hash[hash].next;
+
+                       while (p != &ptable->pt_hash[hash]) {
+                               lnet_peer_t *lp = list_entry(p, lnet_peer_t,
+                                                            lp_hashlist);
                                 if (skip == 0) {
                                         peer = lp;
 
-                                        /* minor optimiztion: start from idx+1
+                                        /* minor optimization: start from idx+1
                                          * on next iteration if we've just
                                          * drained lp_hashlist */
-                                        if (lp->lp_hashlist.next ==
-                                            &the_lnet.ln_peer_hash[idx]) {
-                                                num = 1;
-                                                idx++;
-                                        } else
-                                                num++;
+                                       if (lp->lp_hashlist.next ==
+                                           &ptable->pt_hash[hash]) {
+                                               hoff = 1;
+                                               hash++;
+                                       } else {
+                                               hoff++;
+                                        }
 
                                         break;
-                                } else
-                                        skip--;
+                                }
 
+                                skip--;
                                 p = lp->lp_hashlist.next;
                         }
 
@@ -425,13 +491,14 @@ int LL_PROC_PROTO(proc_lnet_peers)
                                 break;
 
                         p = NULL;
-                        num = 1;
-                        idx++;
+                       hoff = 1;
+                       hash++;
                 }
 
                 if (peer != NULL) {
                         lnet_nid_t nid       = peer->lp_nid;
                         int        nrefs     = peer->lp_refcount;
+                        int        lastalive = -1;
                         char      *aliveness = "NA";
                         int        maxcr     = peer->lp_ni->ni_peertxcredits;
                         int        txcr      = peer->lp_txcredits;
@@ -440,19 +507,42 @@ int LL_PROC_PROTO(proc_lnet_peers)
                         int        minrtrcr  = peer->lp_minrtrcredits;
                         int        txqnob    = peer->lp_txqnob;
 
-                        if (lnet_isrouter(peer) ||
-                            peer->lp_ni->ni_peertimeout > 0)
-                                aliveness = peer->lp_alive ? "up" : "down";
+                        if (lnet_peer_aliveness_enabled(peer)) {
+                               cfs_time_t     now = cfs_time_current();
+                               cfs_duration_t delta;
+
+                               aliveness = peer->lp_alive ? "up" : "down";
+                                delta = cfs_time_sub(now, peer->lp_last_alive);
+                                lastalive = cfs_duration_sec(delta);
+
+                                /* No need to mess up peers contents with
+                                 * arbitrarily long integers - it suffices to
+                                 * know that lastalive is more than 10000s old
+                                 */
+                                if (lastalive >= 10000)
+                                        lastalive = 9999;
+                        }
+
+                       lnet_net_unlock(cpt);
 
                         s += snprintf(s, tmpstr + tmpsiz - s,
-                                      "%-24s %4d %5s %5d %5d %5d %5d %5d %d\n",
+                                      "%-24s %4d %5s %5d %5d %5d %5d %5d %5d %d\n",
                                       libcfs_nid2str(nid), nrefs, aliveness,
-                                      maxcr, rtrcr, minrtrcr, txcr,
+                                      lastalive, maxcr, rtrcr, minrtrcr, txcr,
                                       mintxcr, txqnob);
                         LASSERT (tmpstr + tmpsiz - s > 0);
-                }
 
-                LNET_UNLOCK();
+               } else { /* peer is NULL */
+                       lnet_net_unlock(cpt);
+               }
+
+               if (hash == LNET_PEER_HASH_SIZE) {
+                       cpt++;
+                       hash = 0;
+                       hoff = 1;
+                       if (peer == NULL && cpt < LNET_CPT_NUMBER)
+                               goto again;
+               }
         }
 
         len = s - tmpstr;     /* how many bytes was written */
@@ -460,10 +550,10 @@ int LL_PROC_PROTO(proc_lnet_peers)
         if (len > *lenp) {    /* linux-supplied buffer is too small */
                 rc = -EINVAL;
         } else if (len > 0) { /* wrote something */
-                if (copy_to_user(buffer, tmpstr, len))
+               if (copy_to_user(buffer, tmpstr, len))
                         rc = -EFAULT;
                 else
-                        *ppos = LNET_PHASH_POS_MAKE(idx, num);
+                       *ppos = LNET_PROC_POS_MAKE(cpt, ver, hash, hoff);
         }
 
         LIBCFS_FREE(tmpstr, tmpsiz);
@@ -475,18 +565,20 @@ int LL_PROC_PROTO(proc_lnet_peers)
 }
 
 static int __proc_lnet_buffers(void *data, int write,
-                               loff_t pos, void *buffer, int nob)
+                              loff_t pos, void __user *buffer, int nob)
 {
-
-        int              rc;
-        int              len;
-        char            *s;
-        char            *tmpstr;
-        const int        tmpsiz = 64 * (LNET_NRBPOOLS + 1); /* (4 %d) * 4 */
-        int              idx;
-
-        LASSERT (!write);
-
+       char            *s;
+       char            *tmpstr;
+       int             tmpsiz;
+       int             idx;
+       int             len;
+       int             rc;
+       int             i;
+
+       LASSERT(!write);
+
+       /* (4 %d) * 4 * LNET_CPT_NUMBER */
+       tmpsiz = 64 * (LNET_NRBPOOLS + 1) * LNET_CPT_NUMBER;
         LIBCFS_ALLOC(tmpstr, tmpsiz);
         if (tmpstr == NULL)
                 return -ENOMEM;
@@ -498,48 +590,56 @@ static int __proc_lnet_buffers(void *data, int write,
                       "pages", "count", "credits", "min");
         LASSERT (tmpstr + tmpsiz - s > 0);
 
-        LNET_LOCK();
-
-        for (idx = 0; idx < LNET_NRBPOOLS; idx++) {
-                lnet_rtrbufpool_t *rbp = &the_lnet.ln_rtrpools[idx];
-
-                int npages = rbp->rbp_npages;
-                int nbuf   = rbp->rbp_nbuffers;
-                int cr     = rbp->rbp_credits;
-                int mincr  = rbp->rbp_mincredits;
-
-                s += snprintf(s, tmpstr + tmpsiz - s,
-                              "%5d %5d %7d %7d\n",
-                              npages, nbuf, cr, mincr);
-                LASSERT (tmpstr + tmpsiz - s > 0);
-        }
-
-        LNET_UNLOCK();
-
+       if (the_lnet.ln_rtrpools == NULL)
+               goto out; /* I'm not a router */
+
+       for (idx = 0; idx < LNET_NRBPOOLS; idx++) {
+               lnet_rtrbufpool_t *rbp;
+
+               lnet_net_lock(LNET_LOCK_EX);
+               cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
+                       s += snprintf(s, tmpstr + tmpsiz - s,
+                                     "%5d %5d %7d %7d\n",
+                                     rbp[idx].rbp_npages,
+                                     rbp[idx].rbp_nbuffers,
+                                     rbp[idx].rbp_credits,
+                                     rbp[idx].rbp_mincredits);
+                       LASSERT(tmpstr + tmpsiz - s > 0);
+               }
+               lnet_net_unlock(LNET_LOCK_EX);
+       }
+
+ out:
         len = s - tmpstr;
 
         if (pos >= min_t(int, len, strlen(tmpstr)))
                 rc = 0;
         else
-                rc = trace_copyout_string(buffer, nob,
-                                          tmpstr + pos, NULL);
+                rc = cfs_trace_copyout_string(buffer, nob,
+                                              tmpstr + pos, NULL);
 
         LIBCFS_FREE(tmpstr, tmpsiz);
         return rc;
 }
 
-DECLARE_PROC_HANDLER(proc_lnet_buffers);
+static int
+proc_lnet_buffers(struct ctl_table *table, int write, void __user *buffer,
+                 size_t *lenp, loff_t *ppos)
+{
+       return lprocfs_call_handler(table->data, write, ppos, buffer, lenp,
+                                   __proc_lnet_buffers);
+}
 
-int LL_PROC_PROTO(proc_lnet_nis)
+static int
+proc_lnet_nis(struct ctl_table *table, int write, void __user *buffer,
+             size_t *lenp, loff_t *ppos)
 {
+       int     tmpsiz = 128 * LNET_CPT_NUMBER;
         int        rc = 0;
         char      *tmpstr;
         char      *s;
-        const int  tmpsiz = 256;
         int        len;
 
-        DECLARE_LL_PROC_PPOS_DECL;
-
         LASSERT (!write);
 
         if (*lenp == 0)
@@ -558,56 +658,77 @@ int LL_PROC_PROTO(proc_lnet_nis)
                               "rtr", "max", "tx", "min");
                 LASSERT (tmpstr + tmpsiz - s > 0);
         } else {
-                struct list_head  *n;
+               struct list_head  *n;
                 lnet_ni_t         *ni   = NULL;
                 int                skip = *ppos - 1;
 
-                LNET_LOCK();
+               lnet_net_lock(0);
 
                 n = the_lnet.ln_nis.next;
 
-                while (n != &the_lnet.ln_nis) {
-                        lnet_ni_t *a_ni = list_entry(n, lnet_ni_t, ni_list);
+               while (n != &the_lnet.ln_nis) {
+                       lnet_ni_t *a_ni = list_entry(n, lnet_ni_t, ni_list);
 
                         if (skip == 0) {
                                 ni = a_ni;
                                 break;
-                        } else
-                                skip--;
+                        }
 
+                        skip--;
                         n = n->next;
                 }
 
                 if (ni != NULL) {
-                        cfs_time_t now = cfs_time_current();
-                        int        maxtxcr = ni->ni_maxtxcredits;
-                        int        txcr = ni->ni_txcredits;
-                        int        mintxcr = ni->ni_mintxcredits;
-                        int        npeertxcr = ni->ni_peertxcredits;
-                        int        npeerrtrcr = ni->ni_peerrtrcredits;
-                        lnet_nid_t nid = ni->ni_nid;
-                        int        nref = ni->ni_refcount;
-                        int        last_alive;
-                        char      *stat;
-
-                        last_alive = (the_lnet.ln_routing) ?
-                                 cfs_duration_sec(now - ni->ni_last_alive) : -1;
-                        if (ni->ni_lnd->lnd_type == LOLND)  /* @lo forever alive */
-                                last_alive = 0;
-
-                        LASSERT (ni->ni_status != NULL);
-                        stat = (ni->ni_status->ns_status == LNET_NI_STATUS_UP) ?
-                                                                  "up" : "down";
-
-                        s += snprintf(s, tmpstr + tmpsiz - s,
-                                      "%-24s %6s %5d %4d %4d %4d %5d %5d %5d\n",
-                                      libcfs_nid2str(nid), stat, last_alive, nref,
-                                      npeertxcr, npeerrtrcr, maxtxcr,
-                                      txcr, mintxcr);
-                        LASSERT (tmpstr + tmpsiz - s > 0);
-                }
-
-                LNET_UNLOCK();
+                       struct lnet_tx_queue    *tq;
+                       char    *stat;
+                       long    now = cfs_time_current_sec();
+                       int     last_alive = -1;
+                       int     i;
+                       int     j;
+
+                       if (the_lnet.ln_routing)
+                               last_alive = now - ni->ni_last_alive;
+
+                       /* @lo forever alive */
+                       if (ni->ni_lnd->lnd_type == LOLND)
+                               last_alive = 0;
+
+                       lnet_ni_lock(ni);
+                       LASSERT(ni->ni_status != NULL);
+                       stat = (ni->ni_status->ns_status ==
+                               LNET_NI_STATUS_UP) ? "up" : "down";
+                       lnet_ni_unlock(ni);
+
+                       /* we actually output credits information for
+                        * TX queue of each partition */
+                       cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
+                               for (j = 0; ni->ni_cpts != NULL &&
+                                    j < ni->ni_ncpts; j++) {
+                                       if (i == ni->ni_cpts[j])
+                                               break;
+                               }
+
+                               if (j == ni->ni_ncpts)
+                                       continue;
+
+                               if (i != 0)
+                                       lnet_net_lock(i);
+
+                               s += snprintf(s, tmpstr + tmpsiz - s,
+                                     "%-24s %6s %5d %4d %4d %4d %5d %5d %5d\n",
+                                     libcfs_nid2str(ni->ni_nid), stat,
+                                     last_alive, *ni->ni_refs[i],
+                                     ni->ni_peertxcredits,
+                                     ni->ni_peerrtrcredits,
+                                     tq->tq_credits_max,
+                                     tq->tq_credits, tq->tq_credits_min);
+                               if (i != 0)
+                                       lnet_net_unlock(i);
+                       }
+                       LASSERT(tmpstr + tmpsiz - s > 0);
+               }
+
+               lnet_net_unlock(0);
         }
 
         len = s - tmpstr;     /* how many bytes was written */
@@ -615,7 +736,7 @@ int LL_PROC_PROTO(proc_lnet_nis)
         if (len > *lenp) {    /* linux-supplied buffer is too small */
                 rc = -EINVAL;
         } else if (len > 0) { /* wrote something */
-                if (copy_to_user(buffer, tmpstr, len))
+               if (copy_to_user(buffer, tmpstr, len))
                         rc = -EFAULT;
                 else
                         *ppos += 1;
@@ -629,70 +750,183 @@ int LL_PROC_PROTO(proc_lnet_nis)
         return rc;
 }
 
-static cfs_sysctl_table_t lnet_table[] = {
-        /*
-         * NB No .strategy entries have been provided since sysctl(8) prefers
-         * to go via /proc for portability.
-         */
-        {
-                .ctl_name = PSDEV_LNET_STATS,
-                .procname = "stats",
-                .mode     = 0644,
-                .proc_handler = &proc_lnet_stats,
-        },
-        {
-                .ctl_name = PSDEV_LNET_ROUTES,
-                .procname = "routes",
-                .mode     = 0444,
-                .proc_handler = &proc_lnet_routes,
-        },
-        {
-                .ctl_name = PSDEV_LNET_ROUTERS,
-                .procname = "routers",
-                .mode     = 0444,
-                .proc_handler = &proc_lnet_routers,
-        },
-        {
-                .ctl_name = PSDEV_LNET_PEERS,
-                .procname = "peers",
-                .mode     = 0444,
-                .proc_handler = &proc_lnet_peers,
-        },
-        {
-                .ctl_name = PSDEV_LNET_PEERS,
-                .procname = "buffers",
-                .mode     = 0444,
-                .proc_handler = &proc_lnet_buffers,
-        },
-        {
-                .ctl_name = PSDEV_LNET_NIS,
-                .procname = "nis",
-                .mode     = 0444,
-                .proc_handler = &proc_lnet_nis,
-        },
-        {0}
+struct lnet_portal_rotors {
+       int             pr_value;
+       const char      *pr_name;
+       const char      *pr_desc;
 };
 
-static cfs_sysctl_table_t top_table[] = {
-        {
-                .ctl_name = CTL_LNET,
-                .procname = "lnet",
-                .mode     = 0555,
-                .data     = NULL,
-                .maxlen   = 0,
-                .child    = lnet_table,
-        },
-        {
-                .ctl_name = 0
-        }
+static struct lnet_portal_rotors       portal_rotors[] = {
+       {
+               .pr_value = LNET_PTL_ROTOR_OFF,
+               .pr_name  = "OFF",
+               .pr_desc  = "Turn off message rotor for wildcard portals"
+       },
+       {
+               .pr_value = LNET_PTL_ROTOR_ON,
+               .pr_name  = "ON",
+               .pr_desc  = "round-robin dispatch all PUT messages for "
+                           "wildcard portals"
+       },
+       {
+               .pr_value = LNET_PTL_ROTOR_RR_RT,
+               .pr_name  = "RR_RT",
+               .pr_desc  = "round-robin dispatch routed PUT message for "
+                           "wildcard portals"
+       },
+       {
+               .pr_value = LNET_PTL_ROTOR_HASH_RT,
+               .pr_name  = "HASH_RT",
+               .pr_desc  = "dispatch routed PUT message by hashing source "
+                           "NID for wildcard portals"
+       },
+       {
+               .pr_value = -1,
+               .pr_name  = NULL,
+               .pr_desc  = NULL
+       },
+};
+
+extern int portal_rotor;
+
+static int __proc_lnet_portal_rotor(void *data, int write,
+                                   loff_t pos, void __user *buffer, int nob)
+{
+       const int       buf_len = 128;
+       char            *buf;
+       char            *tmp;
+       int             rc;
+       int             i;
+
+       LIBCFS_ALLOC(buf, buf_len);
+       if (buf == NULL)
+               return -ENOMEM;
+
+       if (!write) {
+               lnet_res_lock(0);
+
+               for (i = 0; portal_rotors[i].pr_value >= 0; i++) {
+                       if (portal_rotors[i].pr_value == portal_rotor)
+                               break;
+               }
+
+               LASSERT(portal_rotors[i].pr_value == portal_rotor);
+               lnet_res_unlock(0);
+
+               rc = snprintf(buf, buf_len,
+                             "{\n\tportals: all\n"
+                             "\trotor: %s\n\tdescription: %s\n}",
+                             portal_rotors[i].pr_name,
+                             portal_rotors[i].pr_desc);
+
+               if (pos >= min_t(int, rc, buf_len)) {
+                       rc = 0;
+               } else {
+                       rc = cfs_trace_copyout_string(buffer, nob,
+                                       buf + pos, "\n");
+               }
+               goto out;
+       }
+
+       rc = cfs_trace_copyin_string(buf, buf_len, buffer, nob);
+       if (rc < 0)
+               goto out;
+
+       tmp = cfs_trimwhite(buf);
+
+       rc = -EINVAL;
+       lnet_res_lock(0);
+       for (i = 0; portal_rotors[i].pr_name != NULL; i++) {
+               if (strncasecmp(portal_rotors[i].pr_name, tmp,
+                               strlen(portal_rotors[i].pr_name)) == 0) {
+                       portal_rotor = portal_rotors[i].pr_value;
+                       rc = 0;
+                       break;
+               }
+       }
+       lnet_res_unlock(0);
+out:
+       LIBCFS_FREE(buf, buf_len);
+       return rc;
+}
+
+static int
+proc_lnet_portal_rotor(struct ctl_table *table, int write, void __user *buffer,
+                      size_t *lenp, loff_t *ppos)
+{
+       return lprocfs_call_handler(table->data, write, ppos, buffer, lenp,
+                                   __proc_lnet_portal_rotor);
+}
+
+
+static struct ctl_table lnet_table[] = {
+       /*
+        * NB No .strategy entries have been provided since sysctl(8) prefers
+        * to go via /proc for portability.
+        */
+       {
+               INIT_CTL_NAME
+               .procname       = "stats",
+               .mode           = 0644,
+               .proc_handler   = &proc_lnet_stats,
+       },
+       {
+               INIT_CTL_NAME
+               .procname       = "routes",
+               .mode           = 0444,
+               .proc_handler   = &proc_lnet_routes,
+       },
+       {
+               INIT_CTL_NAME
+               .procname       = "routers",
+               .mode           = 0444,
+               .proc_handler   = &proc_lnet_routers,
+       },
+       {
+               INIT_CTL_NAME
+               .procname       = "peers",
+               .mode           = 0444,
+               .proc_handler   = &proc_lnet_peers,
+       },
+       {
+               INIT_CTL_NAME
+               .procname       = "buffers",
+               .mode           = 0444,
+               .proc_handler   = &proc_lnet_buffers,
+       },
+       {
+               INIT_CTL_NAME
+               .procname       = "nis",
+               .mode           = 0444,
+               .proc_handler   = &proc_lnet_nis,
+       },
+       {
+               INIT_CTL_NAME
+               .procname       = "portal_rotor",
+               .mode           = 0644,
+               .proc_handler   = &proc_lnet_portal_rotor,
+       },
+       { 0 }
+};
+
+static struct ctl_table top_table[] = {
+       {
+               INIT_CTL_NAME
+               .procname       = "lnet",
+               .mode           = 0555,
+               .data           = NULL,
+               .maxlen         = 0,
+               .child          = lnet_table,
+       },
+       { 0 }
 };
 
 void
 lnet_proc_init(void)
 {
 #ifdef CONFIG_SYSCTL
-        if (lnet_table_header == NULL)
-                lnet_table_header = cfs_register_sysctl_table(top_table, 0);
+       if (lnet_table_header == NULL)
+               lnet_table_header = register_sysctl_table(top_table);
 #endif
 }
 
@@ -700,10 +934,10 @@ void
 lnet_proc_fini(void)
 {
 #ifdef CONFIG_SYSCTL
-        if (lnet_table_header != NULL)
-                cfs_unregister_sysctl_table(lnet_table_header);
+       if (lnet_table_header != NULL)
+               unregister_sysctl_table(lnet_table_header);
 
-        lnet_table_header = NULL;
+       lnet_table_header = NULL;
 #endif
 }