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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/include/lnet/lib-types.h
38 * Types used by the library side routines that do not need to be
39 * exposed to the user application
42 #ifndef __LNET_LIB_TYPES_H__
43 #define __LNET_LIB_TYPES_H__
46 # error This include is only for kernel use.
49 #include <linux/kthread.h>
50 #include <linux/uio.h>
51 #include <linux/types.h>
53 #include <lnet/lnetctl.h>
55 /* Max payload size */
56 #ifndef CONFIG_LNET_MAX_PAYLOAD
57 # error "CONFIG_LNET_MAX_PAYLOAD must be defined in config.h"
60 #define LNET_MAX_PAYLOAD CONFIG_LNET_MAX_PAYLOAD
61 #if (LNET_MAX_PAYLOAD < LNET_MTU)
62 # error "LNET_MAX_PAYLOAD too small - error in configure --with-max-payload-mb"
63 #elif (LNET_MAX_PAYLOAD > (PAGE_SIZE * LNET_MAX_IOV))
64 # error "LNET_MAX_PAYLOAD too large - error in configure --with-max-payload-mb"
70 typedef struct lnet_msg {
71 struct list_head msg_activelist;
72 struct list_head msg_list; /* Q for credits/MD */
74 lnet_process_id_t msg_target;
75 /* where is it from, it's only for building event */
79 /* committed for sending */
80 unsigned int msg_tx_committed:1;
81 /* CPT # this message committed for sending */
82 unsigned int msg_tx_cpt:15;
83 /* committed for receiving */
84 unsigned int msg_rx_committed:1;
85 /* CPT # this message committed for receiving */
86 unsigned int msg_rx_cpt:15;
87 /* queued for tx credit */
88 unsigned int msg_tx_delayed:1;
89 /* queued for RX buffer */
90 unsigned int msg_rx_delayed:1;
91 /* ready for pending on RX delay list */
92 unsigned int msg_rx_ready_delay:1;
94 unsigned int msg_vmflush:1; /* VM trying to free memory */
95 unsigned int msg_target_is_router:1; /* sending to a router */
96 unsigned int msg_routing:1; /* being forwarded */
97 unsigned int msg_ack:1; /* ack on finalize (PUT) */
98 unsigned int msg_sending:1; /* outgoing message */
99 unsigned int msg_receiving:1; /* being received */
100 unsigned int msg_txcredit:1; /* taken an NI send credit */
101 unsigned int msg_peertxcredit:1; /* taken a peer send credit */
102 unsigned int msg_rtrcredit:1; /* taken a globel router credit */
103 unsigned int msg_peerrtrcredit:1; /* taken a peer router credit */
104 unsigned int msg_onactivelist:1; /* on the activelist */
105 unsigned int msg_rdma_get:1;
107 struct lnet_peer *msg_txpeer; /* peer I'm sending to */
108 struct lnet_peer *msg_rxpeer; /* peer I received from */
111 struct lnet_libmd *msg_md;
113 unsigned int msg_len;
114 unsigned int msg_wanted;
115 unsigned int msg_offset;
116 unsigned int msg_niov;
117 struct kvec *msg_iov;
118 lnet_kiov_t *msg_kiov;
125 typedef struct lnet_libhandle {
126 struct list_head lh_hash_chain;
130 #define lh_entry(ptr, type, member) \
131 ((type *)((char *)(ptr)-(char *)(&((type *)0)->member)))
133 typedef struct lnet_eq {
134 struct list_head eq_list;
135 lnet_libhandle_t eq_lh;
136 lnet_seq_t eq_enq_seq;
137 lnet_seq_t eq_deq_seq;
138 unsigned int eq_size;
139 lnet_eq_handler_t eq_callback;
140 lnet_event_t *eq_events;
141 int **eq_refs; /* percpt refcount for EQ */
144 typedef struct lnet_me {
145 struct list_head me_list;
146 lnet_libhandle_t me_lh;
147 lnet_process_id_t me_match_id;
148 unsigned int me_portal;
149 unsigned int me_pos; /* hash offset in mt_hash */
151 __u64 me_ignore_bits;
152 lnet_unlink_t me_unlink;
153 struct lnet_libmd *me_md;
156 typedef struct lnet_libmd {
157 struct list_head md_list;
158 lnet_libhandle_t md_lh;
161 unsigned int md_offset;
162 unsigned int md_length;
163 unsigned int md_max_size;
166 unsigned int md_options;
167 unsigned int md_flags;
170 unsigned int md_niov; /* # frags */
172 struct kvec iov[LNET_MAX_IOV];
173 lnet_kiov_t kiov[LNET_MAX_IOV];
177 #define LNET_MD_FLAG_ZOMBIE (1 << 0)
178 #define LNET_MD_FLAG_AUTO_UNLINK (1 << 1)
179 #define LNET_MD_FLAG_ABORTED (1 << 2)
182 /* info about peers we are trying to fail */
183 struct list_head tp_list; /* ln_test_peers */
184 lnet_nid_t tp_nid; /* matching nid */
185 unsigned int tp_threshold; /* # failures to simulate */
188 #define LNET_COOKIE_TYPE_MD 1
189 #define LNET_COOKIE_TYPE_ME 2
190 #define LNET_COOKIE_TYPE_EQ 3
191 #define LNET_COOKIE_TYPE_BITS 2
192 #define LNET_COOKIE_MASK ((1ULL << LNET_COOKIE_TYPE_BITS) - 1ULL)
194 struct lnet_ni; /* forward ref */
197 typedef struct lnet_lnd
199 /* fields managed by portals */
200 struct list_head lnd_list; /* stash in the LND table */
201 int lnd_refcount; /* # active instances */
203 /* fields initialized by the LND */
206 int (*lnd_startup) (struct lnet_ni *ni);
207 void (*lnd_shutdown) (struct lnet_ni *ni);
208 int (*lnd_ctl)(struct lnet_ni *ni, unsigned int cmd, void *arg);
210 /* In data movement APIs below, payload buffers are described as a set
211 * of 'niov' fragments which are...
213 * in virtual memory (struct kvec *iov != NULL)
215 * in pages (kernel only: plt_kiov_t *kiov != NULL).
216 * The LND may NOT overwrite these fragment descriptors.
217 * An 'offset' and may specify a byte offset within the set of
218 * fragments to start from
221 /* Start sending a preformatted message. 'private' is NULL for PUT and
222 * GET messages; otherwise this is a response to an incoming message
223 * and 'private' is the 'private' passed to lnet_parse(). Return
224 * non-zero for immediate failure, otherwise complete later with
226 int (*lnd_send)(struct lnet_ni *ni, void *private, lnet_msg_t *msg);
228 /* Start receiving 'mlen' bytes of payload data, skipping the following
229 * 'rlen' - 'mlen' bytes. 'private' is the 'private' passed to
230 * lnet_parse(). Return non-zero for immedaite failure, otherwise
231 * complete later with lnet_finalize(). This also gives back a receive
232 * credit if the LND does flow control. */
233 int (*lnd_recv)(struct lnet_ni *ni, void *private, lnet_msg_t *msg,
234 int delayed, unsigned int niov,
235 struct kvec *iov, lnet_kiov_t *kiov,
236 unsigned int offset, unsigned int mlen, unsigned int rlen);
238 /* lnet_parse() has had to delay processing of this message
239 * (e.g. waiting for a forwarding buffer or send credits). Give the
240 * LND a chance to free urgently needed resources. If called, return 0
241 * for success and do NOT give back a receive credit; that has to wait
242 * until lnd_recv() gets called. On failure return < 0 and
243 * release resources; lnd_recv() will not be called. */
244 int (*lnd_eager_recv)(struct lnet_ni *ni, void *private, lnet_msg_t *msg,
245 void **new_privatep);
247 /* notification of peer health */
248 void (*lnd_notify)(struct lnet_ni *ni, lnet_nid_t peer, int alive);
250 /* query of peer aliveness */
251 void (*lnd_query)(struct lnet_ni *ni, lnet_nid_t peer, cfs_time_t *when);
253 /* accept a new connection */
254 int (*lnd_accept)(struct lnet_ni *ni, struct socket *sock);
261 } WIRE_ATTR lnet_ni_status_t;
263 struct lnet_tx_queue {
264 int tq_credits; /* # tx credits free */
265 int tq_credits_min; /* lowest it's been */
266 int tq_credits_max; /* total # tx credits */
267 struct list_head tq_delayed; /* delayed TXs */
270 typedef struct lnet_ni {
272 struct list_head ni_list; /* chain on ln_nis */
273 struct list_head ni_cptlist; /* chain on ln_nis_cpt */
274 int ni_maxtxcredits; /* # tx credits */
275 /* # per-peer send credits */
276 int ni_peertxcredits;
277 /* # per-peer router buffer credits */
278 int ni_peerrtrcredits;
279 /* seconds to consider peer dead */
281 int ni_ncpts; /* number of CPTs */
282 __u32 *ni_cpts; /* bond NI on some CPTs */
283 lnet_nid_t ni_nid; /* interface's NID */
284 void *ni_data; /* instance-specific data */
285 lnd_t *ni_lnd; /* procedural interface */
286 struct lnet_tx_queue **ni_tx_queues; /* percpt TX queues */
287 int **ni_refs; /* percpt reference count */
288 long ni_last_alive; /* when I was last alive */
289 lnet_ni_status_t *ni_status; /* my health status */
290 /* equivalent interfaces to use */
291 char *ni_interfaces[LNET_MAX_INTERFACES];
294 #define LNET_PROTO_PING_MATCHBITS 0x8000000000000000LL
296 /* NB: value of these features equal to LNET_PROTO_PING_VERSION_x
297 * of old LNet, so there shouldn't be any compatibility issue */
298 #define LNET_PING_FEAT_INVAL (0) /* no feature */
299 #define LNET_PING_FEAT_BASE (1 << 0) /* just a ping */
300 #define LNET_PING_FEAT_NI_STATUS (1 << 1) /* return NI status */
301 #define LNET_PING_FEAT_RTE_DISABLED (1 << 2) /* Routing enabled */
303 #define LNET_PING_FEAT_MASK (LNET_PING_FEAT_BASE | \
304 LNET_PING_FEAT_NI_STATUS)
311 lnet_ni_status_t pi_ni[0];
312 } WIRE_ATTR lnet_ping_info_t;
314 /* router checker data, per router */
315 #define LNET_MAX_RTR_NIS 16
316 #define LNET_PINGINFO_SIZE offsetof(lnet_ping_info_t, pi_ni[LNET_MAX_RTR_NIS])
318 /* chain on the_lnet.ln_zombie_rcd or ln_deathrow_rcd */
319 struct list_head rcd_list;
320 lnet_handle_md_t rcd_mdh; /* ping buffer MD */
321 struct lnet_peer *rcd_gateway; /* reference to gateway */
322 lnet_ping_info_t *rcd_pinginfo; /* ping buffer */
325 typedef struct lnet_peer {
326 /* chain on peer hash */
327 struct list_head lp_hashlist;
328 /* messages blocking for tx credits */
329 struct list_head lp_txq;
330 /* messages blocking for router credits */
331 struct list_head lp_rtrq;
332 /* chain on router list */
333 struct list_head lp_rtr_list;
334 /* # tx credits available */
338 /* # router credits */
341 int lp_minrtrcredits;
343 unsigned int lp_alive:1;
344 /* notification outstanding? */
345 unsigned int lp_notify:1;
346 /* outstanding notification for LND? */
347 unsigned int lp_notifylnd:1;
348 /* some thread is handling notification */
349 unsigned int lp_notifying:1;
350 /* SEND event outstanding from ping */
351 unsigned int lp_ping_notsent;
352 /* # times router went dead<->alive */
354 /* bytes queued for sending */
356 /* time of last aliveness news */
357 cfs_time_t lp_timestamp;
358 /* time of last ping attempt */
359 cfs_time_t lp_ping_timestamp;
360 /* != 0 if ping reply expected */
361 cfs_time_t lp_ping_deadline;
362 /* when I was last alive */
363 cfs_time_t lp_last_alive;
364 /* when lp_ni was queried last time */
365 cfs_time_t lp_last_query;
366 /* interface peer is on */
368 lnet_nid_t lp_nid; /* peer's NID */
369 int lp_refcount; /* # refs */
370 int lp_cpt; /* CPT this peer attached on */
371 /* # refs from lnet_route_t::lr_gateway */
373 /* returned RC ping features */
374 unsigned int lp_ping_feats;
375 struct list_head lp_routes; /* routers on this peer */
376 lnet_rc_data_t *lp_rcd; /* router checker state */
380 #define LNET_PEER_HASH_BITS 9
381 #define LNET_PEER_HASH_SIZE (1 << LNET_PEER_HASH_BITS)
383 /* peer hash table */
384 struct lnet_peer_table {
385 int pt_version; /* /proc validity stamp */
386 int pt_number; /* # peers extant */
387 int pt_zombies; /* # zombies to go to deathrow
388 * (and not there yet) */
389 struct list_head pt_deathrow; /* zombie peers */
390 struct list_head *pt_hash; /* NID->peer hash */
393 /* peer aliveness is enabled only on routers for peers in a network where the
394 * lnet_ni_t::ni_peertimeout has been set to a positive value */
395 #define lnet_peer_aliveness_enabled(lp) (the_lnet.ln_routing != 0 && \
396 (lp)->lp_ni->ni_peertimeout > 0)
399 struct list_head lr_list; /* chain on net */
400 struct list_head lr_gwlist; /* chain on gateway */
401 lnet_peer_t *lr_gateway; /* router node */
402 __u32 lr_net; /* remote network number */
403 int lr_seq; /* sequence for round-robin */
404 unsigned int lr_downis; /* number of down NIs */
405 __u32 lr_hops; /* how far I am */
406 unsigned int lr_priority; /* route priority */
409 #define LNET_REMOTE_NETS_HASH_DEFAULT (1U << 7)
410 #define LNET_REMOTE_NETS_HASH_MAX (1U << 16)
411 #define LNET_REMOTE_NETS_HASH_SIZE (1 << the_lnet.ln_remote_nets_hbits)
414 /* chain on ln_remote_nets_hash */
415 struct list_head lrn_list;
417 struct list_head lrn_routes;
422 /** lnet message has credit and can be submitted to lnd for send/receive */
423 #define LNET_CREDIT_OK 0
424 /** lnet message is waiting for credit */
425 #define LNET_CREDIT_WAIT 1
428 /* my free buffer pool */
429 struct list_head rbp_bufs;
430 /* messages blocking for a buffer */
431 struct list_head rbp_msgs;
432 /* # pages in each buffer */
434 /* requested number of buffers */
435 int rbp_req_nbuffers;
436 /* # buffers actually allocated */
438 /* # free buffers / blocked messages */
445 struct list_head rb_list; /* chain on rbp_bufs */
446 lnet_rtrbufpool_t *rb_pool; /* owning pool */
447 lnet_kiov_t rb_kiov[0]; /* the buffer space */
450 #define LNET_PEER_HASHSIZE 503 /* prime! */
453 /* Didn't match anything */
454 LNET_MATCHMD_NONE = (1 << 0),
456 LNET_MATCHMD_OK = (1 << 1),
457 /* Must be discarded */
458 LNET_MATCHMD_DROP = (1 << 2),
459 /* match and buffer is exhausted */
460 LNET_MATCHMD_EXHAUSTED = (1 << 3),
462 LNET_MATCHMD_FINISH = (LNET_MATCHMD_OK | LNET_MATCHMD_DROP),
465 /* Options for lnet_portal_t::ptl_options */
466 #define LNET_PTL_LAZY (1 << 0)
467 #define LNET_PTL_MATCH_UNIQUE (1 << 1) /* unique match, for RDMA */
468 #define LNET_PTL_MATCH_WILDCARD (1 << 2) /* wildcard match, request portal */
470 /* parameter for matching operations (GET, PUT) */
471 struct lnet_match_info {
473 lnet_process_id_t mi_id;
475 unsigned int mi_portal;
476 unsigned int mi_rlength;
477 unsigned int mi_roffset;
480 /* ME hash of RDMA portal */
481 #define LNET_MT_HASH_BITS 8
482 #define LNET_MT_HASH_SIZE (1 << LNET_MT_HASH_BITS)
483 #define LNET_MT_HASH_MASK (LNET_MT_HASH_SIZE - 1)
484 /* we allocate (LNET_MT_HASH_SIZE + 1) entries for lnet_match_table::mt_hash,
485 * the last entry is reserved for MEs with ignore-bits */
486 #define LNET_MT_HASH_IGNORE LNET_MT_HASH_SIZE
487 /* __u64 has 2^6 bits, so need 2^(LNET_MT_HASH_BITS - LNET_MT_BITS_U64) which
488 * is 4 __u64s as bit-map, and add an extra __u64 (only use one bit) for the
489 * ME-list with ignore-bits, which is mtable::mt_hash[LNET_MT_HASH_IGNORE] */
490 #define LNET_MT_BITS_U64 6 /* 2^6 bits */
491 #define LNET_MT_EXHAUSTED_BITS (LNET_MT_HASH_BITS - LNET_MT_BITS_U64)
492 #define LNET_MT_EXHAUSTED_BMAP ((1 << LNET_MT_EXHAUSTED_BITS) + 1)
494 /* portal match table */
495 struct lnet_match_table {
496 /* reserved for upcoming patches, CPU partition ID */
498 unsigned int mt_portal; /* portal index */
499 /* match table is set as "enabled" if there's non-exhausted MD
500 * attached on mt_mhash, it's only valid for wildcard portal */
501 unsigned int mt_enabled;
502 /* bitmap to flag whether MEs on mt_hash are exhausted or not */
503 __u64 mt_exhausted[LNET_MT_EXHAUSTED_BMAP];
504 struct list_head *mt_mhash; /* matching hash */
507 /* these are only useful for wildcard portal */
508 /* Turn off message rotor for wildcard portals */
509 #define LNET_PTL_ROTOR_OFF 0
510 /* round-robin dispatch all PUT messages for wildcard portals */
511 #define LNET_PTL_ROTOR_ON 1
512 /* round-robin dispatch routed PUT message for wildcard portals */
513 #define LNET_PTL_ROTOR_RR_RT 2
514 /* dispatch routed PUT message by hashing source NID for wildcard portals */
515 #define LNET_PTL_ROTOR_HASH_RT 3
517 typedef struct lnet_portal {
519 unsigned int ptl_index; /* portal ID, reserved */
520 /* flags on this portal: lazy, unique... */
521 unsigned int ptl_options;
522 /* list of messages which are stealing buffer */
523 struct list_head ptl_msg_stealing;
524 /* messages blocking for MD */
525 struct list_head ptl_msg_delayed;
526 /* Match table for each CPT */
527 struct lnet_match_table **ptl_mtables;
528 /* spread rotor of incoming "PUT" */
529 unsigned int ptl_rotor;
530 /* # active entries for this portal */
532 /* array of active entries' cpu-partition-id */
536 #define LNET_LH_HASH_BITS 12
537 #define LNET_LH_HASH_SIZE (1ULL << LNET_LH_HASH_BITS)
538 #define LNET_LH_HASH_MASK (LNET_LH_HASH_SIZE - 1)
540 /* resource container (ME, MD, EQ) */
541 struct lnet_res_container {
542 unsigned int rec_type; /* container type */
543 __u64 rec_lh_cookie; /* cookie generator */
544 struct list_head rec_active; /* active resource list */
545 struct list_head *rec_lh_hash; /* handle hash */
548 /* message container */
549 struct lnet_msg_container {
550 int msc_init; /* initialized or not */
551 /* max # threads finalizing */
553 /* msgs waiting to complete finalizing */
554 struct list_head msc_finalizing;
555 struct list_head msc_active; /* active message list */
556 /* threads doing finalization */
557 void **msc_finalizers;
560 /* Router Checker states */
561 #define LNET_RC_STATE_SHUTDOWN 0 /* not started */
562 #define LNET_RC_STATE_RUNNING 1 /* started up OK */
563 #define LNET_RC_STATE_STOPPING 2 /* telling thread to stop */
567 /* CPU partition table of LNet */
568 struct cfs_cpt_table *ln_cpt_table;
569 /* number of CPTs in ln_cpt_table */
570 unsigned int ln_cpt_number;
571 unsigned int ln_cpt_bits;
573 /* protect LNet resources (ME/MD/EQ) */
574 struct cfs_percpt_lock *ln_res_lock;
577 /* the vector of portals */
578 lnet_portal_t **ln_portals;
579 /* percpt ME containers */
580 struct lnet_res_container **ln_me_containers;
581 /* percpt MD container */
582 struct lnet_res_container **ln_md_containers;
584 /* Event Queue container */
585 struct lnet_res_container ln_eq_container;
586 wait_queue_head_t ln_eq_waitq;
587 spinlock_t ln_eq_wait_lock;
589 unsigned int ln_remote_nets_hbits;
591 /* protect NI, peer table, credits, routers, rtrbuf... */
592 struct cfs_percpt_lock *ln_net_lock;
593 /* percpt message containers for active/finalizing/freed message */
594 struct lnet_msg_container **ln_msg_containers;
595 lnet_counters_t **ln_counters;
596 struct lnet_peer_table **ln_peer_tables;
597 /* failure simulation */
598 struct list_head ln_test_peers;
599 struct list_head ln_drop_rules;
600 struct list_head ln_delay_rules;
602 struct list_head ln_nis; /* LND instances */
603 /* NIs bond on specific CPT(s) */
604 struct list_head ln_nis_cpt;
605 /* dying LND instances */
606 struct list_head ln_nis_zombie;
607 lnet_ni_t *ln_loni; /* the loopback NI */
609 /* remote networks with routes to them */
610 struct list_head *ln_remote_nets_hash;
612 __u64 ln_remote_nets_version;
613 /* list of all known routers */
614 struct list_head ln_routers;
616 __u64 ln_routers_version;
617 /* percpt router buffer pools */
618 lnet_rtrbufpool_t **ln_rtrpools;
620 lnet_handle_md_t ln_ping_target_md;
621 lnet_handle_eq_t ln_ping_target_eq;
622 lnet_ping_info_t *ln_ping_info;
624 /* router checker startup/shutdown state */
626 /* router checker's event queue */
627 lnet_handle_eq_t ln_rc_eqh;
628 /* rcd still pending on net */
629 struct list_head ln_rcd_deathrow;
630 /* rcd ready for free */
631 struct list_head ln_rcd_zombie;
632 /* serialise startup/shutdown */
633 struct semaphore ln_rc_signal;
635 struct mutex ln_api_mutex;
636 struct mutex ln_lnd_mutex;
637 /* Have I called LNetNIInit myself? */
639 /* LNetNIInit/LNetNIFini counter */
641 /* shutdown in progress */
644 int ln_routing; /* am I a router? */
645 lnet_pid_t ln_pid; /* requested pid */
646 /* uniquely identifies this ni in this epoch */
647 __u64 ln_interface_cookie;
648 /* registered LNDs */
649 struct list_head ln_lnds;
651 /* test protocol compatibility flags */
652 int ln_testprotocompat;
654 /* 0 - load the NIs from the mod params
655 * 1 - do not load the NIs from the mod params
656 * Reverse logic to ensure that other calls to LNetNIInit
659 bool ln_nis_from_mod_params;
661 /* waitq for router checker. As long as there are no routes in
662 * the list, the router checker will sleep on this queue. when
663 * routes are added the thread will wake up */
664 wait_queue_head_t ln_rc_waitq;