4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Seagate, Inc.
32 * lnet/include/lnet/lib-types.h
35 #ifndef __LNET_LIB_TYPES_H__
36 #define __LNET_LIB_TYPES_H__
38 #include <linux/kthread.h>
39 #include <linux/uio.h>
40 #include <linux/types.h>
41 #include <linux/completion.h>
46 /* Max payload size */
47 #define LNET_MAX_PAYLOAD CONFIG_LNET_MAX_PAYLOAD
48 #if (LNET_MAX_PAYLOAD < LNET_MTU)
49 # error "LNET_MAX_PAYLOAD too small - error in configure --with-max-payload-mb"
50 #elif (LNET_MAX_PAYLOAD > (PAGE_SIZE * LNET_MAX_IOV))
51 # error "LNET_MAX_PAYLOAD too large - error in configure --with-max-payload-mb"
57 typedef struct lnet_msg {
58 struct list_head msg_activelist;
59 struct list_head msg_list; /* Q for credits/MD */
61 lnet_process_id_t msg_target;
62 /* where is it from, it's only for building event */
66 /* committed for sending */
67 unsigned int msg_tx_committed:1;
68 /* CPT # this message committed for sending */
69 unsigned int msg_tx_cpt:15;
70 /* committed for receiving */
71 unsigned int msg_rx_committed:1;
72 /* CPT # this message committed for receiving */
73 unsigned int msg_rx_cpt:15;
74 /* queued for tx credit */
75 unsigned int msg_tx_delayed:1;
76 /* queued for RX buffer */
77 unsigned int msg_rx_delayed:1;
78 /* ready for pending on RX delay list */
79 unsigned int msg_rx_ready_delay:1;
81 unsigned int msg_vmflush:1; /* VM trying to free memory */
82 unsigned int msg_target_is_router:1; /* sending to a router */
83 unsigned int msg_routing:1; /* being forwarded */
84 unsigned int msg_ack:1; /* ack on finalize (PUT) */
85 unsigned int msg_sending:1; /* outgoing message */
86 unsigned int msg_receiving:1; /* being received */
87 unsigned int msg_txcredit:1; /* taken an NI send credit */
88 unsigned int msg_peertxcredit:1; /* taken a peer send credit */
89 unsigned int msg_rtrcredit:1; /* taken a global router credit */
90 unsigned int msg_peerrtrcredit:1; /* taken a peer router credit */
91 unsigned int msg_onactivelist:1; /* on the activelist */
92 unsigned int msg_rdma_get:1;
94 struct lnet_peer *msg_txpeer; /* peer I'm sending to */
95 struct lnet_peer *msg_rxpeer; /* peer I received from */
98 struct lnet_libmd *msg_md;
100 unsigned int msg_len;
101 unsigned int msg_wanted;
102 unsigned int msg_offset;
103 unsigned int msg_niov;
104 struct kvec *msg_iov;
105 lnet_kiov_t *msg_kiov;
111 typedef struct lnet_libhandle {
112 struct list_head lh_hash_chain;
116 #define lh_entry(ptr, type, member) \
117 ((type *)((char *)(ptr) - (char *)(&((type *)0)->member)))
119 typedef struct lnet_eq {
120 struct list_head eq_list;
121 lnet_libhandle_t eq_lh;
122 lnet_seq_t eq_enq_seq;
123 lnet_seq_t eq_deq_seq;
124 unsigned int eq_size;
125 lnet_eq_handler_t eq_callback;
126 lnet_event_t *eq_events;
127 int **eq_refs; /* percpt refcount for EQ */
130 typedef struct lnet_me {
131 struct list_head me_list;
132 lnet_libhandle_t me_lh;
133 lnet_process_id_t me_match_id;
134 unsigned int me_portal;
135 unsigned int me_pos; /* hash offset in mt_hash */
137 __u64 me_ignore_bits;
138 lnet_unlink_t me_unlink;
139 struct lnet_libmd *me_md;
142 typedef struct lnet_libmd {
143 struct list_head md_list;
144 lnet_libhandle_t md_lh;
147 unsigned int md_offset;
148 unsigned int md_length;
149 unsigned int md_max_size;
152 unsigned int md_options;
153 unsigned int md_flags;
156 unsigned int md_niov; /* # frags */
158 struct kvec iov[LNET_MAX_IOV];
159 lnet_kiov_t kiov[LNET_MAX_IOV];
163 #define LNET_MD_FLAG_ZOMBIE (1 << 0)
164 #define LNET_MD_FLAG_AUTO_UNLINK (1 << 1)
165 #define LNET_MD_FLAG_ABORTED (1 << 2)
168 /* info about peers we are trying to fail */
169 struct list_head tp_list; /* ln_test_peers */
170 lnet_nid_t tp_nid; /* matching nid */
171 unsigned int tp_threshold; /* # failures to simulate */
174 #define LNET_COOKIE_TYPE_MD 1
175 #define LNET_COOKIE_TYPE_ME 2
176 #define LNET_COOKIE_TYPE_EQ 3
177 #define LNET_COOKIE_TYPE_BITS 2
178 #define LNET_COOKIE_MASK ((1ULL << LNET_COOKIE_TYPE_BITS) - 1ULL)
180 struct lnet_ni; /* forward ref */
182 typedef struct lnet_lnd {
183 /* fields managed by portals */
184 struct list_head lnd_list; /* stash in the LND table */
185 int lnd_refcount; /* # active instances */
187 /* fields initialised by the LND */
190 int (*lnd_startup)(struct lnet_ni *ni);
191 void (*lnd_shutdown)(struct lnet_ni *ni);
192 int (*lnd_ctl)(struct lnet_ni *ni, unsigned int cmd, void *arg);
195 * In data movement APIs below, payload buffers are described as a set
196 * of 'niov' fragments which are...
198 * in virtual memory (struct iovec *iov != NULL)
200 * in pages (kernel only: plt_kiov_t *kiov != NULL).
201 * The LND may NOT overwrite these fragment descriptors.
202 * An 'offset' and may specify a byte offset within the set of
203 * fragments to start from
207 * Start sending a preformatted message. 'private' is NULL for PUT and
208 * GET messages; otherwise this is a response to an incoming message
209 * and 'private' is the 'private' passed to lnet_parse(). Return
210 * non-zero for immediate failure, otherwise complete later with
213 int (*lnd_send)(struct lnet_ni *ni, void *private, lnet_msg_t *msg);
216 * Start receiving 'mlen' bytes of payload data, skipping the following
217 * 'rlen' - 'mlen' bytes. 'private' is the 'private' passed to
218 * lnet_parse(). Return non-zero for immediate failure, otherwise
219 * complete later with lnet_finalize(). This also gives back a receive
220 * credit if the LND does flow control.
222 int (*lnd_recv)(struct lnet_ni *ni, void *private, lnet_msg_t *msg,
223 int delayed, struct iov_iter *to, unsigned int rlen);
226 * lnet_parse() has had to delay processing of this message
227 * (e.g. waiting for a forwarding buffer or send credits). Give the
228 * LND a chance to free urgently needed resources. If called, return 0
229 * for success and do NOT give back a receive credit; that has to wait
230 * until lnd_recv() gets called. On failure return < 0 and
231 * release resources; lnd_recv() will not be called.
233 int (*lnd_eager_recv)(struct lnet_ni *ni, void *private,
234 lnet_msg_t *msg, void **new_privatep);
236 /* notification of peer health */
237 void (*lnd_notify)(struct lnet_ni *ni, lnet_nid_t peer, int alive);
239 /* query of peer aliveness */
240 void (*lnd_query)(struct lnet_ni *ni, lnet_nid_t peer,
241 unsigned long *when);
243 /* accept a new connection */
244 int (*lnd_accept)(struct lnet_ni *ni, struct socket *sock);
247 struct lnet_tx_queue {
248 int tq_credits; /* # tx credits free */
249 int tq_credits_min; /* lowest it's been */
250 int tq_credits_max; /* total # tx credits */
251 struct list_head tq_delayed; /* delayed TXs */
254 typedef struct lnet_ni {
256 struct list_head ni_list; /* chain on ln_nis */
257 struct list_head ni_cptlist; /* chain on ln_nis_cpt */
258 int ni_maxtxcredits; /* # tx credits */
259 /* # per-peer send credits */
260 int ni_peertxcredits;
261 /* # per-peer router buffer credits */
262 int ni_peerrtrcredits;
263 /* seconds to consider peer dead */
265 int ni_ncpts; /* number of CPTs */
266 __u32 *ni_cpts; /* bond NI on some CPTs */
267 lnet_nid_t ni_nid; /* interface's NID */
268 void *ni_data; /* instance-specific data */
269 lnd_t *ni_lnd; /* procedural interface */
270 struct lnet_tx_queue **ni_tx_queues; /* percpt TX queues */
271 int **ni_refs; /* percpt reference count */
272 time64_t ni_last_alive;/* when I was last alive */
273 lnet_ni_status_t *ni_status; /* my health status */
274 /* per NI LND tunables */
275 struct lnet_ioctl_config_lnd_tunables *ni_lnd_tunables;
276 /* equivalent interfaces to use */
277 char *ni_interfaces[LNET_MAX_INTERFACES];
278 /* original net namespace */
279 struct net *ni_net_ns;
282 #define LNET_PROTO_PING_MATCHBITS 0x8000000000000000LL
285 * NB: value of these features equal to LNET_PROTO_PING_VERSION_x
286 * of old LNet, so there shouldn't be any compatibility issue
288 #define LNET_PING_FEAT_INVAL (0) /* no feature */
289 #define LNET_PING_FEAT_BASE (1 << 0) /* just a ping */
290 #define LNET_PING_FEAT_NI_STATUS (1 << 1) /* return NI status */
291 #define LNET_PING_FEAT_RTE_DISABLED (1 << 2) /* Routing enabled */
293 #define LNET_PING_FEAT_MASK (LNET_PING_FEAT_BASE | \
294 LNET_PING_FEAT_NI_STATUS)
296 /* router checker data, per router */
297 #define LNET_MAX_RTR_NIS 16
298 #define LNET_PINGINFO_SIZE offsetof(lnet_ping_info_t, pi_ni[LNET_MAX_RTR_NIS])
300 /* chain on the_lnet.ln_zombie_rcd or ln_deathrow_rcd */
301 struct list_head rcd_list;
302 lnet_handle_md_t rcd_mdh; /* ping buffer MD */
303 struct lnet_peer *rcd_gateway; /* reference to gateway */
304 lnet_ping_info_t *rcd_pinginfo; /* ping buffer */
307 typedef struct lnet_peer {
308 struct list_head lp_hashlist; /* chain on peer hash */
309 struct list_head lp_txq; /* messages blocking for
311 struct list_head lp_rtrq; /* messages blocking for
313 struct list_head lp_rtr_list; /* chain on router list */
314 int lp_txcredits; /* # tx credits available */
315 int lp_mintxcredits; /* low water mark */
316 int lp_rtrcredits; /* # router credits */
317 int lp_minrtrcredits; /* low water mark */
318 unsigned int lp_alive:1; /* alive/dead? */
319 unsigned int lp_notify:1; /* notification outstanding? */
320 unsigned int lp_notifylnd:1;/* outstanding notification
322 unsigned int lp_notifying:1; /* some thread is handling
324 unsigned int lp_ping_notsent;/* SEND event outstanding
326 int lp_alive_count; /* # times router went
328 long lp_txqnob; /* bytes queued for sending */
329 unsigned long lp_timestamp; /* time of last aliveness
331 unsigned long lp_ping_timestamp;/* time of last ping
333 unsigned long lp_ping_deadline; /* != 0 if ping reply
335 unsigned long lp_last_alive; /* when I was last alive */
336 unsigned long lp_last_query; /* when lp_ni was queried
338 lnet_ni_t *lp_ni; /* interface peer is on */
339 lnet_nid_t lp_nid; /* peer's NID */
340 int lp_refcount; /* # refs */
341 int lp_cpt; /* CPT this peer attached on */
342 /* # refs from lnet_route_t::lr_gateway */
344 /* returned RC ping features */
345 unsigned int lp_ping_feats;
346 struct list_head lp_routes; /* routers on this peer */
347 lnet_rc_data_t *lp_rcd; /* router checker state */
351 #define LNET_PEER_HASH_BITS 9
352 #define LNET_PEER_HASH_SIZE (1 << LNET_PEER_HASH_BITS)
354 /* peer hash table */
355 struct lnet_peer_table {
356 int pt_version; /* /proc validity stamp */
357 int pt_number; /* # peers extant */
358 /* # zombies to go to deathrow (and not there yet) */
360 struct list_head pt_deathrow; /* zombie peers */
361 struct list_head *pt_hash; /* NID->peer hash */
365 * peer aliveness is enabled only on routers for peers in a network where the
366 * lnet_ni_t::ni_peertimeout has been set to a positive value
368 #define lnet_peer_aliveness_enabled(lp) (the_lnet.ln_routing && \
369 (lp)->lp_ni->ni_peertimeout > 0)
372 struct list_head lr_list; /* chain on net */
373 struct list_head lr_gwlist; /* chain on gateway */
374 lnet_peer_t *lr_gateway; /* router node */
375 __u32 lr_net; /* remote network number */
376 int lr_seq; /* sequence for round-robin */
377 unsigned int lr_downis; /* number of down NIs */
378 __u32 lr_hops; /* how far I am */
379 unsigned int lr_priority; /* route priority */
382 #define LNET_REMOTE_NETS_HASH_DEFAULT (1U << 7)
383 #define LNET_REMOTE_NETS_HASH_MAX (1U << 16)
384 #define LNET_REMOTE_NETS_HASH_SIZE (1 << the_lnet.ln_remote_nets_hbits)
387 struct list_head lrn_list; /* chain on
388 ln_remote_nets_hash */
389 struct list_head lrn_routes; /* routes to me */
390 __u32 lrn_net; /* my net number */
393 /** lnet message has credit and can be submitted to lnd for send/receive */
394 #define LNET_CREDIT_OK 0
395 /** lnet message is waiting for credit */
396 #define LNET_CREDIT_WAIT 1
399 struct list_head rbp_bufs; /* my free buffer pool */
400 struct list_head rbp_msgs; /* messages blocking
402 int rbp_npages; /* # pages in each buffer */
403 /* requested number of buffers */
404 int rbp_req_nbuffers;
405 /* # buffers actually allocated */
407 int rbp_credits; /* # free buffers /
409 int rbp_mincredits; /* low water mark */
413 struct list_head rb_list; /* chain on rbp_bufs */
414 lnet_rtrbufpool_t *rb_pool; /* owning pool */
415 lnet_kiov_t rb_kiov[0]; /* the buffer space */
418 #define LNET_PEER_HASHSIZE 503 /* prime! */
420 #define LNET_TINY_BUF_IDX 0
421 #define LNET_SMALL_BUF_IDX 1
422 #define LNET_LARGE_BUF_IDX 2
424 /* # different router buffer pools */
425 #define LNET_NRBPOOLS (LNET_LARGE_BUF_IDX + 1)
428 /* Didn't match anything */
429 LNET_MATCHMD_NONE = (1 << 0),
431 LNET_MATCHMD_OK = (1 << 1),
432 /* Must be discarded */
433 LNET_MATCHMD_DROP = (1 << 2),
434 /* match and buffer is exhausted */
435 LNET_MATCHMD_EXHAUSTED = (1 << 3),
437 LNET_MATCHMD_FINISH = (LNET_MATCHMD_OK | LNET_MATCHMD_DROP),
440 /* Options for lnet_portal_t::ptl_options */
441 #define LNET_PTL_LAZY (1 << 0)
442 #define LNET_PTL_MATCH_UNIQUE (1 << 1) /* unique match, for RDMA */
443 #define LNET_PTL_MATCH_WILDCARD (1 << 2) /* wildcard match,
446 /* parameter for matching operations (GET, PUT) */
447 struct lnet_match_info {
449 lnet_process_id_t mi_id;
451 unsigned int mi_portal;
452 unsigned int mi_rlength;
453 unsigned int mi_roffset;
456 /* ME hash of RDMA portal */
457 #define LNET_MT_HASH_BITS 8
458 #define LNET_MT_HASH_SIZE (1 << LNET_MT_HASH_BITS)
459 #define LNET_MT_HASH_MASK (LNET_MT_HASH_SIZE - 1)
461 * we allocate (LNET_MT_HASH_SIZE + 1) entries for lnet_match_table::mt_hash,
462 * the last entry is reserved for MEs with ignore-bits
464 #define LNET_MT_HASH_IGNORE LNET_MT_HASH_SIZE
466 * __u64 has 2^6 bits, so need 2^(LNET_MT_HASH_BITS - LNET_MT_BITS_U64) which
467 * is 4 __u64s as bit-map, and add an extra __u64 (only use one bit) for the
468 * ME-list with ignore-bits, which is mtable::mt_hash[LNET_MT_HASH_IGNORE]
470 #define LNET_MT_BITS_U64 6 /* 2^6 bits */
471 #define LNET_MT_EXHAUSTED_BITS (LNET_MT_HASH_BITS - LNET_MT_BITS_U64)
472 #define LNET_MT_EXHAUSTED_BMAP ((1 << LNET_MT_EXHAUSTED_BITS) + 1)
474 /* portal match table */
475 struct lnet_match_table {
476 /* reserved for upcoming patches, CPU partition ID */
478 unsigned int mt_portal; /* portal index */
480 * match table is set as "enabled" if there's non-exhausted MD
481 * attached on mt_mhash, it's only valid for wildcard portal
483 unsigned int mt_enabled;
484 /* bitmap to flag whether MEs on mt_hash are exhausted or not */
485 __u64 mt_exhausted[LNET_MT_EXHAUSTED_BMAP];
486 struct list_head *mt_mhash; /* matching hash */
489 /* these are only useful for wildcard portal */
490 /* Turn off message rotor for wildcard portals */
491 #define LNET_PTL_ROTOR_OFF 0
492 /* round-robin dispatch all PUT messages for wildcard portals */
493 #define LNET_PTL_ROTOR_ON 1
494 /* round-robin dispatch routed PUT message for wildcard portals */
495 #define LNET_PTL_ROTOR_RR_RT 2
496 /* dispatch routed PUT message by hashing source NID for wildcard portals */
497 #define LNET_PTL_ROTOR_HASH_RT 3
499 typedef struct lnet_portal {
501 unsigned int ptl_index; /* portal ID, reserved */
502 /* flags on this portal: lazy, unique... */
503 unsigned int ptl_options;
504 /* list of messages which are stealing buffer */
505 struct list_head ptl_msg_stealing;
506 /* messages blocking for MD */
507 struct list_head ptl_msg_delayed;
508 /* Match table for each CPT */
509 struct lnet_match_table **ptl_mtables;
510 /* spread rotor of incoming "PUT" */
511 unsigned int ptl_rotor;
512 /* # active entries for this portal */
514 /* array of active entries' cpu-partition-id */
518 #define LNET_LH_HASH_BITS 12
519 #define LNET_LH_HASH_SIZE (1ULL << LNET_LH_HASH_BITS)
520 #define LNET_LH_HASH_MASK (LNET_LH_HASH_SIZE - 1)
522 /* resource container (ME, MD, EQ) */
523 struct lnet_res_container {
524 unsigned int rec_type; /* container type */
525 __u64 rec_lh_cookie; /* cookie generator */
526 struct list_head rec_active; /* active resource list */
527 struct list_head *rec_lh_hash; /* handle hash */
530 /* message container */
531 struct lnet_msg_container {
532 int msc_init; /* initialized or not */
533 /* max # threads finalizing */
535 /* msgs waiting to complete finalizing */
536 struct list_head msc_finalizing;
537 struct list_head msc_active; /* active message list */
538 /* threads doing finalization */
539 void **msc_finalizers;
542 /* Router Checker states */
543 #define LNET_RC_STATE_SHUTDOWN 0 /* not started */
544 #define LNET_RC_STATE_RUNNING 1 /* started up OK */
545 #define LNET_RC_STATE_STOPPING 2 /* telling thread to stop */
548 /* CPU partition table of LNet */
549 struct cfs_cpt_table *ln_cpt_table;
550 /* number of CPTs in ln_cpt_table */
551 unsigned int ln_cpt_number;
552 unsigned int ln_cpt_bits;
554 /* protect LNet resources (ME/MD/EQ) */
555 struct cfs_percpt_lock *ln_res_lock;
558 /* the vector of portals */
559 lnet_portal_t **ln_portals;
560 /* percpt ME containers */
561 struct lnet_res_container **ln_me_containers;
562 /* percpt MD container */
563 struct lnet_res_container **ln_md_containers;
565 /* Event Queue container */
566 struct lnet_res_container ln_eq_container;
567 wait_queue_head_t ln_eq_waitq;
568 spinlock_t ln_eq_wait_lock;
569 unsigned int ln_remote_nets_hbits;
571 /* protect NI, peer table, credits, routers, rtrbuf... */
572 struct cfs_percpt_lock *ln_net_lock;
573 /* percpt message containers for active/finalizing/freed message */
574 struct lnet_msg_container **ln_msg_containers;
575 lnet_counters_t **ln_counters;
576 struct lnet_peer_table **ln_peer_tables;
577 /* failure simulation */
578 struct list_head ln_test_peers;
579 struct list_head ln_drop_rules;
580 struct list_head ln_delay_rules;
582 struct list_head ln_nis; /* LND instances */
583 /* NIs bond on specific CPT(s) */
584 struct list_head ln_nis_cpt;
585 /* dying LND instances */
586 struct list_head ln_nis_zombie;
587 lnet_ni_t *ln_loni; /* the loopback NI */
589 /* remote networks with routes to them */
590 struct list_head *ln_remote_nets_hash;
592 __u64 ln_remote_nets_version;
593 /* list of all known routers */
594 struct list_head ln_routers;
596 __u64 ln_routers_version;
597 /* percpt router buffer pools */
598 lnet_rtrbufpool_t **ln_rtrpools;
600 lnet_handle_md_t ln_ping_target_md;
601 lnet_handle_eq_t ln_ping_target_eq;
602 lnet_ping_info_t *ln_ping_info;
604 /* router checker startup/shutdown state */
606 /* router checker's event queue */
607 lnet_handle_eq_t ln_rc_eqh;
608 /* rcd still pending on net */
609 struct list_head ln_rcd_deathrow;
610 /* rcd ready for free */
611 struct list_head ln_rcd_zombie;
612 /* serialise startup/shutdown */
613 struct completion ln_rc_signal;
615 struct mutex ln_api_mutex;
616 struct mutex ln_lnd_mutex;
617 struct mutex ln_delay_mutex;
618 /* Have I called LNetNIInit myself? */
620 /* LNetNIInit/LNetNIFini counter */
622 /* shutdown in progress */
625 int ln_routing; /* am I a router? */
626 lnet_pid_t ln_pid; /* requested pid */
627 /* uniquely identifies this ni in this epoch */
628 __u64 ln_interface_cookie;
629 /* registered LNDs */
630 struct list_head ln_lnds;
632 /* test protocol compatibility flags */
633 int ln_testprotocompat;
636 * 0 - load the NIs from the mod params
637 * 1 - do not load the NIs from the mod params
638 * Reverse logic to ensure that other calls to LNetNIInit
641 bool ln_nis_from_mod_params;
644 * waitq for router checker. As long as there are no routes in
645 * the list, the router checker will sleep on this queue. when
646 * routes are added the thread will wake up
648 wait_queue_head_t ln_rc_waitq;