Whamcloud - gitweb
LU-2133 lnet: wrong peer state reported
[fs/lustre-release.git] / lnet / include / lnet / lib-types.h
index 9efa5da..2233638 100644 (file)
@@ -1,6 +1,4 @@
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
  * GPL HEADER START
  *
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  * GPL HEADER END
  */
 /*
- * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
+ *
+ * Copyright (c) 2012, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -48,6 +48,8 @@
 #include <lnet/darwin/lib-types.h>
 #elif defined(__WINNT__)
 #include <lnet/winnt/lib-types.h>
+#include <libcfs/libcfs_pack.h>
+#include <libcfs/libcfs_unpack.h>
 #else
 #error Unsupported Operating System
 #endif
 
 #define WIRE_ATTR       __attribute__((packed))
 
+/* Packed version of lnet_process_id_t to transfer via network */
+typedef struct {
+        lnet_nid_t nid;
+        lnet_pid_t pid;   /* node id / process id */
+} WIRE_ATTR lnet_process_id_packed_t;
+
 /* The wire handle's interface cookie only matches one network interface in
  * one epoch (i.e. new cookie when the interface restarts or the node
  * reboots).  The object cookie only matches one object on that interface
  * during that object's lifetime (i.e. no cookie re-use). */
-#include <libcfs/libcfs_pack.h>
 typedef struct {
         __u64 wh_interface_cookie;
         __u64 wh_object_cookie;
 } WIRE_ATTR lnet_handle_wire_t;
-#include <libcfs/libcfs_unpack.h>
 
 typedef enum {
         LNET_MSG_ACK = 0,
@@ -81,7 +87,6 @@ typedef enum {
  * byte boundary in the message header.  Note that all types used in these
  * wire structs MUST be fixed size and the smaller types are placed at the
  * end. */
-#include <libcfs/libcfs_pack.h>
 typedef struct lnet_ack {
         lnet_handle_wire_t  dst_wmd;
         __u64               match_bits;
@@ -147,14 +152,11 @@ typedef struct {
 
 /* PROTO MAGIC for LNDs */
 #define LNET_PROTO_IB_MAGIC                 0x0be91b91
-#define LNET_PROTO_OPENIB_MAGIC             LNET_PROTO_IB_MAGIC
-#define LNET_PROTO_IIB_MAGIC                LNET_PROTO_IB_MAGIC
-#define LNET_PROTO_VIB_MAGIC                LNET_PROTO_IB_MAGIC
 #define LNET_PROTO_RA_MAGIC                 0x0be91b92
 #define LNET_PROTO_QSW_MAGIC                0x0be91b93
+#define LNET_PROTO_GNI_MAGIC                0xb00fbabe /* ask Kim */
 #define LNET_PROTO_TCP_MAGIC                0xeebc0ded
 #define LNET_PROTO_PTL_MAGIC                0x50746C4E /* 'PtlN' unique magic */
-#define LNET_PROTO_GM_MAGIC                 0x6d797269 /* 'myri'! */
 #define LNET_PROTO_MX_MAGIC                 0x4d583130 /* 'MX10'! */
 #define LNET_PROTO_ACCEPTOR_MAGIC           0xacce7100
 #define LNET_PROTO_PING_MAGIC               0x70696E67 /* 'ping' */
@@ -174,7 +176,6 @@ typedef struct {
         __u32       acr_version;                /* protocol version */
         __u64       acr_nid;                    /* target NID */
 } WIRE_ATTR lnet_acceptor_connreq_t;
-#include <libcfs/libcfs_unpack.h>
 
 #define LNET_PROTO_ACCEPTOR_VERSION       1
 
@@ -182,87 +183,105 @@ typedef struct {
 struct lnet_libmd;
 
 typedef struct lnet_msg {
-        struct list_head    msg_activelist;
-        struct list_head    msg_list;           /* Q for credits/MD */
-
-        lnet_process_id_t   msg_target;
-        __u32               msg_type;
-
-        unsigned int        msg_target_is_router:1; /* sending to a router */
-        unsigned int        msg_routing:1;      /* being forwarded */
-        unsigned int        msg_ack:1;          /* ack on finalize (PUT) */
-        unsigned int        msg_sending:1;      /* outgoing message */
-        unsigned int        msg_receiving:1;    /* being received */
-        unsigned int        msg_delayed:1;      /* had to Q for buffer or tx credit */
-        unsigned int        msg_txcredit:1;     /* taken an NI send credit */
-        unsigned int        msg_peertxcredit:1; /* taken a peer send credit */
-        unsigned int        msg_rtrcredit:1;    /* taken a globel router credit */
-        unsigned int        msg_peerrtrcredit:1; /* taken a peer router credit */
-        unsigned int        msg_onactivelist:1; /* on the activelist */
-
-        struct lnet_peer   *msg_txpeer;         /* peer I'm sending to */
-        struct lnet_peer   *msg_rxpeer;         /* peer I received from */
-
-        void               *msg_private;
-        struct lnet_libmd  *msg_md;
-
-        unsigned int        msg_len;
-        unsigned int        msg_wanted;
-        unsigned int        msg_offset;
-        unsigned int        msg_niov;
-        struct iovec       *msg_iov;
-        lnet_kiov_t        *msg_kiov;
-
-        lnet_event_t        msg_ev;
-        lnet_hdr_t          msg_hdr;
+        cfs_list_t            msg_activelist;
+        cfs_list_t            msg_list;           /* Q for credits/MD */
+
+        lnet_process_id_t     msg_target;
+       /* where is it from, it's only for building event */
+       lnet_nid_t              msg_from;
+       __u32                   msg_type;
+
+       /* commited for sending */
+       unsigned int            msg_tx_committed:1;
+       /* CPT # this message committed for sending */
+       unsigned int            msg_tx_cpt:15;
+       /* commited for receiving */
+       unsigned int            msg_rx_committed:1;
+       /* CPT # this message committed for receiving */
+       unsigned int            msg_rx_cpt:15;
+       /* queued for tx credit */
+       unsigned int            msg_tx_delayed:1;
+       /* queued for RX buffer */
+       unsigned int            msg_rx_delayed:1;
+       /* ready for pending on RX delay list */
+       unsigned int            msg_rx_ready_delay:1;
+
+        unsigned int          msg_vmflush:1;      /* VM trying to free memory */
+        unsigned int          msg_target_is_router:1; /* sending to a router */
+        unsigned int          msg_routing:1;      /* being forwarded */
+        unsigned int          msg_ack:1;          /* ack on finalize (PUT) */
+        unsigned int          msg_sending:1;      /* outgoing message */
+        unsigned int          msg_receiving:1;    /* being received */
+        unsigned int          msg_txcredit:1;     /* taken an NI send credit */
+        unsigned int          msg_peertxcredit:1; /* taken a peer send credit */
+        unsigned int          msg_rtrcredit:1;    /* taken a globel router credit */
+        unsigned int          msg_peerrtrcredit:1; /* taken a peer router credit */
+        unsigned int          msg_onactivelist:1; /* on the activelist */
+
+        struct lnet_peer     *msg_txpeer;         /* peer I'm sending to */
+        struct lnet_peer     *msg_rxpeer;         /* peer I received from */
+
+        void                 *msg_private;
+        struct lnet_libmd    *msg_md;
+
+        unsigned int          msg_len;
+        unsigned int          msg_wanted;
+        unsigned int          msg_offset;
+        unsigned int          msg_niov;
+        struct iovec         *msg_iov;
+        lnet_kiov_t          *msg_kiov;
+
+        lnet_event_t          msg_ev;
+        lnet_hdr_t            msg_hdr;
 } lnet_msg_t;
 
 
 typedef struct lnet_libhandle {
-        struct list_head  lh_hash_chain;
-        __u64             lh_cookie;
+        cfs_list_t            lh_hash_chain;
+        __u64                 lh_cookie;
 } lnet_libhandle_t;
 
 #define lh_entry(ptr, type, member) \
         ((type *)((char *)(ptr)-(char *)(&((type *)0)->member)))
 
 typedef struct lnet_eq {
-        struct list_head  eq_list;
-        lnet_libhandle_t  eq_lh;
-        lnet_seq_t        eq_enq_seq;
-        lnet_seq_t        eq_deq_seq;
-        unsigned int      eq_size;
-        lnet_event_t     *eq_events;
-        int               eq_refcount;
-        lnet_eq_handler_t eq_callback;
+       cfs_list_t              eq_list;
+       lnet_libhandle_t        eq_lh;
+       lnet_seq_t              eq_enq_seq;
+       lnet_seq_t              eq_deq_seq;
+       unsigned int            eq_size;
+       lnet_eq_handler_t       eq_callback;
+       lnet_event_t            *eq_events;
+       int                     **eq_refs;      /* percpt refcount for EQ */
 } lnet_eq_t;
 
 typedef struct lnet_me {
-        struct list_head   me_list;
-        lnet_libhandle_t   me_lh;
-        lnet_process_id_t  me_match_id;
-        unsigned int       me_portal;
-        __u64              me_match_bits;
-        __u64              me_ignore_bits;
-        lnet_unlink_t      me_unlink;
-        struct lnet_libmd *me_md;
+        cfs_list_t             me_list;
+        lnet_libhandle_t       me_lh;
+        lnet_process_id_t      me_match_id;
+        unsigned int           me_portal;
+       unsigned int           me_pos;          /* hash offset in mt_hash */
+        __u64                  me_match_bits;
+        __u64                  me_ignore_bits;
+        lnet_unlink_t          me_unlink;
+        struct lnet_libmd     *me_md;
 } lnet_me_t;
 
 typedef struct lnet_libmd {
-        struct list_head  md_list;
-        lnet_libhandle_t  md_lh;
-        lnet_me_t        *md_me;
-        char             *md_start;
-        unsigned int      md_offset;
-        unsigned int      md_length;
-        unsigned int      md_max_size;
-        int               md_threshold;
-        int               md_refcount;
-        unsigned int      md_options;
-        unsigned int      md_flags;
-        void             *md_user_ptr;
-        lnet_eq_t        *md_eq;
-        unsigned int      md_niov;                /* # frags */
+        cfs_list_t            md_list;
+        lnet_libhandle_t      md_lh;
+        lnet_me_t            *md_me;
+        char                 *md_start;
+        unsigned int          md_offset;
+        unsigned int          md_length;
+        unsigned int          md_max_size;
+        int                   md_threshold;
+        int                   md_refcount;
+        unsigned int          md_options;
+        unsigned int          md_flags;
+        void                 *md_user_ptr;
+        lnet_eq_t            *md_eq;
+        unsigned int          md_niov;                /* # frags */
         union {
                 struct iovec  iov[LNET_MAX_IOV];
                 lnet_kiov_t   kiov[LNET_MAX_IOV];
@@ -275,43 +294,42 @@ typedef struct lnet_libmd {
 #ifdef LNET_USE_LIB_FREELIST
 typedef struct
 {
-        void              *fl_objs;             /* single contiguous array of objects */
-        int                fl_nobjs;            /* the number of them */
-        int                fl_objsize;          /* the size (including overhead) of each of them */
-        struct list_head   fl_list;             /* where they are enqueued */
+        void                  *fl_objs;          /* single contiguous array of objects */
+        int                    fl_nobjs;         /* the number of them */
+        int                    fl_objsize;       /* the size (including overhead) of each of them */
+        cfs_list_t             fl_list;          /* where they are enqueued */
 } lnet_freelist_t;
 
 typedef struct
 {
-        struct list_head   fo_list;             /* enqueue on fl_list */
-        void              *fo_contents;         /* aligned contents */
+        cfs_list_t             fo_list;             /* enqueue on fl_list */
+        void                  *fo_contents;         /* aligned contents */
 } lnet_freeobj_t;
 #endif
 
 typedef struct {
         /* info about peers we are trying to fail */
-        struct list_head   tp_list;             /* ln_test_peers */
-        lnet_nid_t         tp_nid;              /* matching nid */
-        unsigned int       tp_threshold;        /* # failures to simulate */
+        cfs_list_t             tp_list;             /* ln_test_peers */
+        lnet_nid_t             tp_nid;              /* matching nid */
+        unsigned int           tp_threshold;        /* # failures to simulate */
 } lnet_test_peer_t;
 
 #define LNET_COOKIE_TYPE_MD    1
 #define LNET_COOKIE_TYPE_ME    2
 #define LNET_COOKIE_TYPE_EQ    3
-#define LNET_COOKIE_TYPES      4
-/* LNET_COOKIE_TYPES must be a power of 2, so the cookie type can be
- * extracted by masking with (LNET_COOKIE_TYPES - 1) */
+#define LNET_COOKIE_TYPE_BITS  2
+#define LNET_COOKIE_MASK       ((1ULL << LNET_COOKIE_TYPE_BITS) - 1ULL)
 
 struct lnet_ni;                                  /* forward ref */
 
 typedef struct lnet_lnd
 {
         /* fields managed by portals */
-        struct list_head  lnd_list;             /* stash in the LND table */
-        int               lnd_refcount;         /* # active instances */
+        cfs_list_t            lnd_list;             /* stash in the LND table */
+        int                   lnd_refcount;         /* # active instances */
 
         /* fields initialised by the LND */
-        unsigned int      lnd_type;
+        unsigned int          lnd_type;
 
         int  (*lnd_startup) (struct lnet_ni *ni);
         void (*lnd_shutdown) (struct lnet_ni *ni);
@@ -383,51 +401,82 @@ typedef struct {
         __u32      ns_unused;
 } WIRE_ATTR lnet_ni_status_t;
 
+struct lnet_tx_queue {
+       int                     tq_credits;     /* # tx credits free */
+       int                     tq_credits_min; /* lowest it's been */
+       int                     tq_credits_max; /* total # tx credits */
+       cfs_list_t              tq_delayed;     /* delayed TXs */
+};
+
 #define LNET_MAX_INTERFACES   16
 
 typedef struct lnet_ni {
-        struct list_head  ni_list;              /* chain on ln_nis */
-        struct list_head  ni_txq;               /* messages waiting for tx credits */
-        int               ni_maxtxcredits;      /* # tx credits  */
-        int               ni_txcredits;         /* # tx credits free */
-        int               ni_mintxcredits;      /* lowest it's been */
-        int               ni_peertxcredits;     /* # per-peer send credits */
-        int               ni_peerrtrcredits;    /* # per-peer router buffer credits */
-        int               ni_peertimeout;       /* seconds to consider peer dead */
-        lnet_nid_t        ni_nid;               /* interface's NID */
-        void             *ni_data;              /* instance-specific data */
-        lnd_t            *ni_lnd;               /* procedural interface */
-        int               ni_refcount;          /* reference count */
-        cfs_time_t        ni_last_alive;        /* when I was last alive */
-        lnet_ni_status_t *ni_status;            /* my health status */
-        char             *ni_interfaces[LNET_MAX_INTERFACES]; /* equivalent interfaces to use */
+#ifdef __KERNEL__
+       spinlock_t              ni_lock;
+#else
+# ifndef HAVE_LIBPTHREAD
+       int                     ni_lock;
+# else
+       pthread_mutex_t         ni_lock;
+# endif
+#endif
+       cfs_list_t              ni_list;        /* chain on ln_nis */
+       cfs_list_t              ni_cptlist;     /* chain on ln_nis_cpt */
+       int                     ni_maxtxcredits; /* # tx credits  */
+       /* # per-peer send credits */
+       int                     ni_peertxcredits;
+       /* # per-peer router buffer credits */
+       int                     ni_peerrtrcredits;
+       /* seconds to consider peer dead */
+       int                     ni_peertimeout;
+       int                     ni_ncpts;       /* number of CPTs */
+       __u32                   *ni_cpts;       /* bond NI on some CPTs */
+       lnet_nid_t              ni_nid;         /* interface's NID */
+       void                    *ni_data;       /* instance-specific data */
+       lnd_t                   *ni_lnd;        /* procedural interface */
+       struct lnet_tx_queue    **ni_tx_queues; /* percpt TX queues */
+       int                     **ni_refs;      /* percpt reference count */
+       long                    ni_last_alive;  /* when I was last alive */
+       lnet_ni_status_t        *ni_status;     /* my health status */
+       /* equivalent interfaces to use */
+       char                    *ni_interfaces[LNET_MAX_INTERFACES];
 } lnet_ni_t;
 
-#define LNET_PROTO_PING_MATCHBITS     0x8000000000000000LL
-#define LNET_PROTO_PING_VERSION       2
-#define LNET_PROTO_PING_VERSION1      1
+#define LNET_PROTO_PING_MATCHBITS      0x8000000000000000LL
+
+/* NB: value of these features equal to LNET_PROTO_PING_VERSION_x
+ * of old LNet, so there shouldn't be any compatibility issue */
+#define LNET_PING_FEAT_INVAL           (0)             /* no feature */
+#define LNET_PING_FEAT_BASE            (1 << 0)        /* just a ping */
+#define LNET_PING_FEAT_NI_STATUS       (1 << 1)        /* return NI status */
+
+#define LNET_PING_FEAT_MASK            (LNET_PING_FEAT_BASE | \
+                                        LNET_PING_FEAT_NI_STATUS)
+
 typedef struct {
-        __u32            pi_magic;
-        __u32            pi_version;
-        lnet_pid_t       pi_pid;
-        __u32            pi_nnis;
-        lnet_ni_status_t pi_ni[0];
+       __u32                   pi_magic;
+       __u32                   pi_features;
+       lnet_pid_t              pi_pid;
+       __u32                   pi_nnis;
+       lnet_ni_status_t        pi_ni[0];
 } WIRE_ATTR lnet_ping_info_t;
 
 /* router checker data, per router */
 #define LNET_MAX_RTR_NIS   16
 #define LNET_PINGINFO_SIZE offsetof(lnet_ping_info_t, pi_ni[LNET_MAX_RTR_NIS])
 typedef struct {
-        struct list_head  rcd_list;             /* chain on the_lnet.ln_zombie_rcd */
-        lnet_handle_md_t  rcd_mdh;              /* ping buffer MD */
-        lnet_ping_info_t *rcd_pinginfo;         /* ping buffer */
+       /* chain on the_lnet.ln_zombie_rcd or ln_deathrow_rcd */
+       cfs_list_t              rcd_list;
+       lnet_handle_md_t        rcd_mdh;        /* ping buffer MD */
+       struct lnet_peer        *rcd_gateway;   /* reference to gateway */
+       lnet_ping_info_t        *rcd_pinginfo;  /* ping buffer */
 } lnet_rc_data_t;
 
 typedef struct lnet_peer {
-        struct list_head  lp_hashlist;          /* chain on peer hash */
-        struct list_head  lp_txq;               /* messages blocking for tx credits */
-        struct list_head  lp_rtrq;              /* messages blocking for router credits */
-        struct list_head  lp_rtr_list;          /* chain on router list */
+        cfs_list_t        lp_hashlist;          /* chain on peer hash */
+        cfs_list_t        lp_txq;               /* messages blocking for tx credits */
+        cfs_list_t        lp_rtrq;              /* messages blocking for router credits */
+        cfs_list_t        lp_rtr_list;          /* chain on router list */
         int               lp_txcredits;         /* # tx credits available */
         int               lp_mintxcredits;      /* low water mark */
         int               lp_rtrcredits;        /* # router credits */
@@ -447,38 +496,68 @@ typedef struct lnet_peer {
         lnet_ni_t        *lp_ni;                /* interface peer is on */
         lnet_nid_t        lp_nid;               /* peer's NID */
         int               lp_refcount;          /* # refs */
-        int               lp_rtr_refcount;      /* # refs from lnet_route_t::lr_gateway */
-        lnet_rc_data_t   *lp_rcd;               /* router checker state */
+       int                     lp_cpt;         /* CPT this peer attached on */
+       /* # refs from lnet_route_t::lr_gateway */
+       int                     lp_rtr_refcount;
+       /* returned RC ping features */
+       unsigned int            lp_ping_feats;
+       cfs_list_t              lp_routes;      /* routers on this peer */
+       lnet_rc_data_t          *lp_rcd;        /* router checker state */
 } lnet_peer_t;
 
+
+/* peer hash size */
+#define LNET_PEER_HASH_BITS     9
+#define LNET_PEER_HASH_SIZE     (1 << LNET_PEER_HASH_BITS)
+
+/* peer hash table */
+struct lnet_peer_table {
+       int                     pt_version;     /* /proc validity stamp */
+       int                     pt_number;      /* # peers extant */
+       cfs_list_t              pt_deathrow;    /* zombie peers */
+       cfs_list_t              *pt_hash;       /* NID->peer hash */
+};
+
+/* peer aliveness is enabled only on routers for peers in a network where the
+ * lnet_ni_t::ni_peertimeout has been set to a positive value */
+#define lnet_peer_aliveness_enabled(lp) (the_lnet.ln_routing != 0 && \
+                                        (lp)->lp_ni->ni_peertimeout > 0)
+
 typedef struct {
-        struct list_head  lr_list;              /* chain on net */
-        lnet_peer_t      *lr_gateway;           /* router node */
+       cfs_list_t              lr_list;        /* chain on net */
+       cfs_list_t              lr_gwlist;      /* chain on gateway */
+       lnet_peer_t             *lr_gateway;    /* router node */
+       __u32                   lr_net;         /* remote network number */
+       int                     lr_seq;         /* sequence for round-robin */
+       unsigned int            lr_downis;      /* number of down NIs */
+       unsigned int            lr_hops;        /* how far I am */
 } lnet_route_t;
 
+#define LNET_REMOTE_NETS_HASH_DEFAULT  (1U << 7)
+#define LNET_REMOTE_NETS_HASH_MAX      (1U << 16)
+#define LNET_REMOTE_NETS_HASH_SIZE     (1 << the_lnet.ln_remote_nets_hbits)
+
 typedef struct {
-        struct list_head        lrn_list;       /* chain on ln_remote_nets */
-        struct list_head        lrn_routes;     /* routes to me */
+        cfs_list_t              lrn_list;       /* chain on ln_remote_nets_hash */
+        cfs_list_t              lrn_routes;     /* routes to me */
         __u32                   lrn_net;        /* my net number */
-        unsigned int            lrn_hops;       /* how far I am */
 } lnet_remotenet_t;
 
 typedef struct {
-        struct list_head  rbp_bufs;             /* my free buffer pool */
-        struct list_head  rbp_msgs;             /* messages blocking for a buffer */
-        int               rbp_npages;           /* # pages in each buffer */
-        int               rbp_nbuffers;         /* # buffers */
-        int               rbp_credits;          /* # free buffers / blocked messages */
-        int               rbp_mincredits;       /* low water mark */
+        cfs_list_t rbp_bufs;             /* my free buffer pool */
+        cfs_list_t rbp_msgs;             /* messages blocking for a buffer */
+        int        rbp_npages;           /* # pages in each buffer */
+        int        rbp_nbuffers;         /* # buffers */
+        int        rbp_credits;          /* # free buffers / blocked messages */
+        int        rbp_mincredits;       /* low water mark */
 } lnet_rtrbufpool_t;
 
 typedef struct {
-        struct list_head   rb_list;             /* chain on rbp_bufs */
-        lnet_rtrbufpool_t *rb_pool;             /* owning pool */
-        lnet_kiov_t        rb_kiov[0];          /* the buffer space */
+        cfs_list_t             rb_list;             /* chain on rbp_bufs */
+        lnet_rtrbufpool_t     *rb_pool;             /* owning pool */
+        lnet_kiov_t            rb_kiov[0];          /* the buffer space */
 } lnet_rtrbuf_t;
 
-#include <libcfs/libcfs_pack.h>
 typedef struct {
         __u32        msgs_alloc;
         __u32        msgs_max;
@@ -492,133 +571,255 @@ typedef struct {
         __u64        route_length;
         __u64        drop_length;
 } WIRE_ATTR lnet_counters_t;
-#include <libcfs/libcfs_unpack.h>
 
 #define LNET_PEER_HASHSIZE   503                /* prime! */
 
 #define LNET_NRBPOOLS         3                 /* # different router buffer pools */
 
+enum {
+       /* Didn't match anything */
+       LNET_MATCHMD_NONE       = (1 << 0),
+       /* Matched OK */
+       LNET_MATCHMD_OK         = (1 << 1),
+       /* Must be discarded */
+       LNET_MATCHMD_DROP       = (1 << 2),
+       /* match and buffer is exhausted */
+       LNET_MATCHMD_EXHAUSTED  = (1 << 3),
+       /* match or drop */
+       LNET_MATCHMD_FINISH     = (LNET_MATCHMD_OK | LNET_MATCHMD_DROP),
+};
+
 /* Options for lnet_portal_t::ptl_options */
 #define LNET_PTL_LAZY               (1 << 0)
-typedef struct {
-        struct list_head ptl_ml;  /* match list */
-        struct list_head ptl_msgq; /* messages blocking for MD */
-        __u64            ptl_ml_version;    /* validity stamp, only changed for new attached MD */
-        __u64            ptl_msgq_version;  /* validity stamp */
-        unsigned int     ptl_options;
-} lnet_portal_t;
-
-/* Router Checker states */
-#define LNET_RC_STATE_SHUTDOWN     0            /* not started */
-#define LNET_RC_STATE_RUNNING      1            /* started up OK */
-#define LNET_RC_STATE_STOPTHREAD   2            /* telling thread to stop */
-#define LNET_RC_STATE_UNLINKING    3            /* unlinking RC MD */
-#define LNET_RC_STATE_UNLINKED     4            /* RC's MD has been unlinked */
-
-typedef struct
-{
-        /* Stuff initialised at LNetInit() */
-        int                ln_init;             /* LNetInit() called? */
-        int                ln_refcount;         /* LNetNIInit/LNetNIFini counter */
-        int                ln_niinit_self;      /* Have I called LNetNIInit myself? */
-
-        struct list_head   ln_lnds;             /* registered LNDs */
-
+#define LNET_PTL_MATCH_UNIQUE       (1 << 1)    /* unique match, for RDMA */
+#define LNET_PTL_MATCH_WILDCARD     (1 << 2)    /* wildcard match, request portal */
+
+/* parameter for matching operations (GET, PUT) */
+struct lnet_match_info {
+       __u64                   mi_mbits;
+       lnet_process_id_t       mi_id;
+       unsigned int            mi_opc;
+       unsigned int            mi_portal;
+       unsigned int            mi_rlength;
+       unsigned int            mi_roffset;
+};
+
+/* ME hash of RDMA portal */
+#define LNET_MT_HASH_BITS              8
+#define LNET_MT_HASH_SIZE              (1 << LNET_MT_HASH_BITS)
+#define LNET_MT_HASH_MASK              (LNET_MT_HASH_SIZE - 1)
+/* we allocate (LNET_MT_HASH_SIZE + 1) entries for lnet_match_table::mt_hash,
+ * the last entry is reserved for MEs with ignore-bits */
+#define LNET_MT_HASH_IGNORE            LNET_MT_HASH_SIZE
+/* __u64 has 2^6 bits, so need 2^(LNET_MT_HASH_BITS - LNET_MT_BITS_U64) which
+ * is 4 __u64s as bit-map, and add an extra __u64 (only use one bit) for the
+ * ME-list with ignore-bits, which is mtable::mt_hash[LNET_MT_HASH_IGNORE] */
+#define LNET_MT_BITS_U64               6       /* 2^6 bits */
+#define LNET_MT_EXHAUSTED_BITS         (LNET_MT_HASH_BITS - LNET_MT_BITS_U64)
+#define LNET_MT_EXHAUSTED_BMAP         ((1 << LNET_MT_EXHAUSTED_BITS) + 1)
+
+/* portal match table */
+struct lnet_match_table {
+       /* reserved for upcoming patches, CPU partition ID */
+       unsigned int            mt_cpt;
+       unsigned int            mt_portal;      /* portal index */
+       /* match table is set as "enabled" if there's non-exhausted MD
+        * attached on mt_mhash, it's only valide for wildcard portal */
+       unsigned int            mt_enabled;
+       /* bitmap to flag whether MEs on mt_hash are exhausted or not */
+       __u64                   mt_exhausted[LNET_MT_EXHAUSTED_BMAP];
+       cfs_list_t              *mt_mhash;      /* matching hash */
+};
+
+/* these are only useful for wildcard portal */
+/* Turn off message rotor for wildcard portals */
+#define        LNET_PTL_ROTOR_OFF      0
+/* round-robin dispatch all PUT messages for wildcard portals */
+#define        LNET_PTL_ROTOR_ON       1
+/* round-robin dispatch routed PUT message for wildcard portals */
+#define        LNET_PTL_ROTOR_RR_RT    2
+/* dispatch routed PUT message by hashing source NID for wildcard portals */
+#define        LNET_PTL_ROTOR_HASH_RT  3
+
+typedef struct lnet_portal {
 #ifdef __KERNEL__
-        spinlock_t         ln_lock;
-        cfs_waitq_t        ln_waitq;
-        struct semaphore   ln_api_mutex;
-        struct semaphore   ln_lnd_mutex;
+       spinlock_t              ptl_lock;
 #else
 # ifndef HAVE_LIBPTHREAD
-        int                ln_lock;
-        int                ln_api_mutex;
-        int                ln_lnd_mutex;
+       int                     ptl_lock;
 # else
-        pthread_cond_t     ln_cond;
-        pthread_mutex_t    ln_lock;
-        pthread_mutex_t    ln_api_mutex;
-        pthread_mutex_t    ln_lnd_mutex;
+       pthread_mutex_t         ptl_lock;
 # endif
 #endif
+       unsigned int            ptl_index;      /* portal ID, reserved */
+       /* flags on this portal: lazy, unique... */
+       unsigned int            ptl_options;
+       /* list of messags which are stealing buffer */
+       cfs_list_t              ptl_msg_stealing;
+       /* messages blocking for MD */
+       cfs_list_t              ptl_msg_delayed;
+       /* Match table for each CPT */
+       struct lnet_match_table **ptl_mtables;
+       /* spread rotor of incoming "PUT" */
+       int                     ptl_rotor;
+       /* # active entries for this portal */
+       int                     ptl_mt_nmaps;
+       /* array of active entries' cpu-partition-id */
+       int                     ptl_mt_maps[0];
+} lnet_portal_t;
 
-        /* Stuff initialised at LNetNIInit() */
-
-        int                ln_shutdown;         /* shutdown in progress */
-        int                ln_nportals;         /* # portals */
-        lnet_portal_t     *ln_portals;          /* the vector of portals */
-
-        lnet_pid_t         ln_pid;              /* requested pid */
-
-        struct list_head   ln_nis;              /* LND instances */
-        lnet_ni_t         *ln_loni;             /* the loopback NI */
-        lnet_ni_t         *ln_eqwaitni;         /* NI to wait for events in */
-        struct list_head   ln_zombie_nis;       /* dying LND instances */
-        int                ln_nzombie_nis;      /* # of NIs to wait for */
-
-        struct list_head   ln_remote_nets;      /* remote networks with routes to them */
-        __u64              ln_remote_nets_version; /* validity stamp */
-
-        struct list_head   ln_routers;          /* list of all known routers */
-        __u64              ln_routers_version;  /* validity stamp */
-
-        struct list_head  *ln_peer_hash;        /* NID->peer hash */
-        int                ln_npeers;           /* # peers extant */
-        int                ln_peertable_version; /* /proc validity stamp */
-
-        int                ln_routing;          /* am I a router? */
-        lnet_rtrbufpool_t  ln_rtrpools[LNET_NRBPOOLS]; /* router buffer pools */
-
-        int                ln_lh_hash_size;     /* size of lib handle hash table */
-        struct list_head  *ln_lh_hash_table;    /* all extant lib handles, this interface */
-        __u64              ln_next_object_cookie; /* cookie generator */
-        __u64              ln_interface_cookie; /* uniquely identifies this ni in this epoch */
+#define LNET_LH_HASH_BITS      12
+#define LNET_LH_HASH_SIZE      (1ULL << LNET_LH_HASH_BITS)
+#define LNET_LH_HASH_MASK      (LNET_LH_HASH_SIZE - 1)
 
-        char              *ln_network_tokens;   /* space for network names */
-        int                ln_network_tokens_nob;
+/* resource container (ME, MD, EQ) */
+struct lnet_res_container {
+       unsigned int            rec_type;       /* container type */
+       __u64                   rec_lh_cookie;  /* cookie generator */
+       cfs_list_t              rec_active;     /* active resource list */
+       cfs_list_t              *rec_lh_hash;   /* handle hash */
+#ifdef LNET_USE_LIB_FREELIST
+       lnet_freelist_t         rec_freelist;   /* freelist for resources */
+#endif
+};
+
+/* message container */
+struct lnet_msg_container {
+       int                     msc_init;       /* initialized or not */
+       /* max # threads finalizing */
+       int                     msc_nfinalizers;
+       /* msgs waiting to complete finalizing */
+       cfs_list_t              msc_finalizing;
+       cfs_list_t              msc_active;     /* active message list */
+       /* threads doing finalization */
+       void                    **msc_finalizers;
+#ifdef LNET_USE_LIB_FREELIST
+       lnet_freelist_t         msc_freelist;   /* freelist for messages */
+#endif
+};
 
-        int                ln_testprotocompat;  /* test protocol compatibility flags */
+/* Router Checker states */
+#define LNET_RC_STATE_SHUTDOWN         0       /* not started */
+#define LNET_RC_STATE_RUNNING          1       /* started up OK */
+#define LNET_RC_STATE_STOPPING         2       /* telling thread to stop */
 
-        struct list_head   ln_finalizeq;        /* msgs waiting to complete finalizing */
+typedef struct
+{
+       /* CPU partition table of LNet */
+       struct cfs_cpt_table            *ln_cpt_table;
+       /* number of CPTs in ln_cpt_table */
+       unsigned int                    ln_cpt_number;
+       unsigned int                    ln_cpt_bits;
+
+       /* protect LNet resources (ME/MD/EQ) */
+       struct cfs_percpt_lock          *ln_res_lock;
+       /* # portals */
+       int                             ln_nportals;
+       /* the vector of portals */
+       lnet_portal_t                   **ln_portals;
+       /* percpt ME containers */
+       struct lnet_res_container       **ln_me_containers;
+       /* percpt MD container */
+       struct lnet_res_container       **ln_md_containers;
+
+       /* Event Queue container */
+       struct lnet_res_container       ln_eq_container;
 #ifdef __KERNEL__
-        void             **ln_finalizers;       /* threads doing finalization */
-        int                ln_nfinalizers;      /* max # threads finalizing */
+       cfs_waitq_t                     ln_eq_waitq;
+       spinlock_t                      ln_eq_wait_lock;
 #else
-        int                ln_finalizing;
+# ifndef HAVE_LIBPTHREAD
+       int                             ln_eq_wait_lock;
+# else
+       pthread_cond_t                  ln_eq_cond;
+       pthread_mutex_t                 ln_eq_wait_lock;
+# endif
 #endif
-        struct list_head   ln_test_peers;       /* failure simulation */
-
-        lnet_handle_md_t   ln_ping_target_md;
-        lnet_handle_eq_t   ln_ping_target_eq;
-        lnet_ping_info_t  *ln_ping_info;
-
+       unsigned int                    ln_remote_nets_hbits;
+
+       /* protect NI, peer table, credits, routers, rtrbuf... */
+       struct cfs_percpt_lock          *ln_net_lock;
+       /* percpt message containers for active/finalizing/freed message */
+       struct lnet_msg_container       **ln_msg_containers;
+       lnet_counters_t                 **ln_counters;
+       struct lnet_peer_table          **ln_peer_tables;
+       /* failure simulation */
+       cfs_list_t                      ln_test_peers;
+
+       cfs_list_t                      ln_nis;         /* LND instances */
+       /* NIs bond on specific CPT(s) */
+       cfs_list_t                      ln_nis_cpt;
+       /* dying LND instances */
+       cfs_list_t                      ln_nis_zombie;
+       lnet_ni_t                       *ln_loni;       /* the loopback NI */
+       /* NI to wait for events in */
+       lnet_ni_t                       *ln_eq_waitni;
+
+       /* remote networks with routes to them */
+       cfs_list_t                      *ln_remote_nets_hash;
+       /* validity stamp */
+       __u64                           ln_remote_nets_version;
+       /* list of all known routers */
+       cfs_list_t                      ln_routers;
+       /* validity stamp */
+       __u64                           ln_routers_version;
+       /* percpt router buffer pools */
+       lnet_rtrbufpool_t               **ln_rtrpools;
+
+       lnet_handle_md_t                ln_ping_target_md;
+       lnet_handle_eq_t                ln_ping_target_eq;
+       lnet_ping_info_t                *ln_ping_info;
+
+       /* router checker startup/shutdown state */
+       int                             ln_rc_state;
+       /* router checker's event queue */
+       lnet_handle_eq_t                ln_rc_eqh;
+       /* rcd still pending on net */
+       cfs_list_t                      ln_rcd_deathrow;
+       /* rcd ready for free */
+       cfs_list_t                      ln_rcd_zombie;
 #ifdef __KERNEL__
-        struct semaphore   ln_rc_signal;        /* serialise startup/shutdown */
-#endif
-        int                ln_rc_state;         /* router checker startup/shutdown state */
-        lnet_handle_eq_t   ln_rc_eqh;           /* router checker's event queue */
-        lnet_handle_md_t   ln_rc_mdh;
-        struct list_head   ln_zombie_rcd;
+       /* serialise startup/shutdown */
+       struct semaphore                ln_rc_signal;
 
-#ifdef LNET_USE_LIB_FREELIST
-        lnet_freelist_t    ln_free_mes;
-        lnet_freelist_t    ln_free_msgs;
-        lnet_freelist_t    ln_free_mds;
-        lnet_freelist_t    ln_free_eqs;
+       struct mutex                    ln_api_mutex;
+       struct mutex                    ln_lnd_mutex;
+#else
+# ifndef HAVE_LIBPTHREAD
+       int                             ln_api_mutex;
+       int                             ln_lnd_mutex;
+# else
+       pthread_mutex_t                 ln_api_mutex;
+       pthread_mutex_t                 ln_lnd_mutex;
+# endif
 #endif
-        struct list_head   ln_active_msgs;
-        struct list_head   ln_active_mds;
-        struct list_head   ln_active_eqs;
-
-        lnet_counters_t    ln_counters;
+       int                             ln_init;        /* LNetInit() called? */
+       /* Have I called LNetNIInit myself? */
+       int                             ln_niinit_self;
+       /* LNetNIInit/LNetNIFini counter */
+       int                             ln_refcount;
+       /* shutdown in progress */
+       int                             ln_shutdown;
+
+       int                             ln_routing;     /* am I a router? */
+       lnet_pid_t                      ln_pid;         /* requested pid */
+       /* uniquely identifies this ni in this epoch */
+       __u64                           ln_interface_cookie;
+       /* registered LNDs */
+       cfs_list_t                      ln_lnds;
+
+       /* space for network names */
+       char                            *ln_network_tokens;
+       int                             ln_network_tokens_nob;
+       /* test protocol compatibility flags */
+       int                             ln_testprotocompat;
 
 #ifndef __KERNEL__
-        /* Temporary workaround to allow uOSS and test programs force
-         * server mode in userspace. The only place where we use it is
-         * lnet_prepare(). The only way to turn this flag on is to
-         * call lnet_server_mode() */
-
-        int                ln_server_mode_flag;
+       /* Temporary workaround to allow uOSS and test programs force
+        * server mode in userspace. The only place where we use it is
+        * lnet_prepare(). The only way to turn this flag on is to
+        * call lnet_server_mode() */
+       int                             ln_server_mode_flag;
 #endif
 } lnet_t;