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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/ptllnd/ptllnd.h
38 * Author: PJ Kirner <pjkirner@clusterfs.com>
41 #include <linux/module.h>
42 #include <linux/kernel.h>
44 #include <linux/string.h>
45 #include <linux/stat.h>
46 #include <linux/errno.h>
47 #include <linux/unistd.h>
48 #include <linux/uio.h>
50 #include <asm/uaccess.h>
53 #include <linux/init.h>
55 #include <linux/file.h>
56 #include <linux/stat.h>
57 #include <linux/list.h>
58 #include <linux/kmod.h>
59 #include <linux/sysctl.h>
60 #include <linux/random.h>
66 #define DEBUG_SUBSYSTEM S_LND
68 #include <libcfs/libcfs.h>
69 #include <lnet/lnet.h>
70 #include <lnet/lib-lnet.h>
71 #include <lnet/lnet-sysctl.h>
72 #include <portals/p30.h>
73 #include <lnet/ptllnd.h> /* Depends on portals/p30.h */
76 * Define this to enable console debug logging
79 //#define PJK_DEBUGGING
82 # define PTLLND_N_SCHED cfs_num_online_cpus() /* # schedulers */
84 # define PTLLND_N_SCHED 1 /* # schedulers */
87 #define PTLLND_CREDIT_HIGHWATER ((*kptllnd_tunables.kptl_peertxcredits)-1)
88 /* when eagerly to return credits */
92 int *kptl_ntx; /* # tx descs to pre-allocate */
93 int *kptl_max_nodes; /* max # nodes all talking to me */
94 int *kptl_max_procs_per_node; /* max # processes per node */
95 int *kptl_checksum; /* checksum kptl_msg_t? */
96 int *kptl_timeout; /* comms timeout (seconds) */
97 int *kptl_portal; /* portal number */
98 int *kptl_pid; /* portals PID (self + kernel peers) */
99 int *kptl_rxb_npages; /* number of pages for rx buffer */
100 int *kptl_rxb_nspare; /* number of spare rx buffers */
101 int *kptl_credits; /* number of credits */
102 int *kptl_peertxcredits; /* number of peer tx credits */
103 int *kptl_peerrtrcredits; /* number of peer router credits */
104 int *kptl_max_msg_size; /* max immd message size*/
105 int *kptl_peer_hash_table_size; /* # slots in peer hash table */
106 int *kptl_reschedule_loops; /* scheduler yield loops */
107 int *kptl_ack_puts; /* make portals ack PUTs */
109 int *kptl_simulation_bitmap;/* simulation bitmap */
112 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
113 cfs_sysctl_table_header_t *kptl_sysctl; /* sysctl interface */
117 #include "lnet/ptllnd_wire.h"
119 /***********************************************************************/
121 typedef struct kptl_data kptl_data_t;
122 typedef struct kptl_net kptl_net_t;
123 typedef struct kptl_rx_buffer kptl_rx_buffer_t;
124 typedef struct kptl_peer kptl_peer_t;
130 #define PTLLND_EVENTARG_TYPE_MSG 0x1
131 #define PTLLND_EVENTARG_TYPE_RDMA 0x2
132 #define PTLLND_EVENTARG_TYPE_BUF 0x3
134 typedef struct kptl_rx /* receive message */
136 cfs_list_t rx_list; /* queue for attention */
137 kptl_rx_buffer_t *rx_rxb; /* the rx buffer pointer */
138 kptl_msg_t *rx_msg; /* received message */
139 int rx_nob; /* received message size */
140 unsigned long rx_treceived; /* time received */
141 ptl_process_id_t rx_initiator; /* sender's address */
142 kptl_peer_t *rx_peer; /* pointer to peer */
143 char rx_space[0]; /* copy of incoming request */
146 #define PTLLND_POSTRX_DONT_POST 0 /* don't post */
147 #define PTLLND_POSTRX_NO_CREDIT 1 /* post: no credits */
148 #define PTLLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
150 typedef struct kptl_rx_buffer_pool
152 spinlock_t rxbp_lock;
153 cfs_list_t rxbp_list; /* all allocated buffers */
154 int rxbp_count; /* # allocated buffers */
155 int rxbp_reserved; /* # requests to buffer */
156 int rxbp_shutdown; /* shutdown flag */
157 } kptl_rx_buffer_pool_t;
159 struct kptl_rx_buffer
161 kptl_rx_buffer_pool_t *rxb_pool;
162 cfs_list_t rxb_list; /* for the rxb_pool list */
163 cfs_list_t rxb_repost_list;/* for the kptl_sched_rxbq list */
164 int rxb_posted:1; /* on the net */
165 int rxb_idle:1; /* all done */
166 kptl_eventarg_t rxb_eventarg; /* event->md.user_ptr */
167 int rxb_refcount; /* reference count */
168 ptl_handle_md_t rxb_mdh; /* the portals memory descriptor (MD) handle */
169 char *rxb_buffer; /* the buffer */
175 TX_TYPE_RESERVED = 0,
176 TX_TYPE_SMALL_MESSAGE = 1,
177 TX_TYPE_PUT_REQUEST = 2,
178 TX_TYPE_GET_REQUEST = 3,
179 TX_TYPE_PUT_RESPONSE = 4,
180 TX_TYPE_GET_RESPONSE = 5,
184 #ifdef _USING_LUSTRE_PORTALS_
185 struct iovec iov[PTL_MD_MAX_IOV];
186 lnet_kiov_t kiov[PTL_MD_MAX_IOV];
188 ptl_md_iovec_t iov[PTL_MD_MAX_IOV];
192 typedef struct kptl_tx /* transmit message */
194 cfs_list_t tx_list; /* queue on idle_txs etc */
195 cfs_atomic_t tx_refcount; /* reference count*/
196 enum kptl_tx_type tx_type; /* small msg/{put,get}{req,resp} */
197 int tx_active:1; /* queued on the peer */
198 int tx_idle:1; /* on the free list */
199 int tx_acked:1; /* portals ACK wanted (for debug only) */
200 kptl_eventarg_t tx_msg_eventarg; /* event->md.user_ptr */
201 kptl_eventarg_t tx_rdma_eventarg; /* event->md.user_ptr */
202 int tx_status; /* the status of this tx descriptor */
203 ptl_handle_md_t tx_rdma_mdh; /* RDMA buffer */
204 ptl_handle_md_t tx_msg_mdh; /* the portals MD handle for the initial message */
205 lnet_msg_t *tx_lnet_msg; /* LNET message to finalize */
206 lnet_msg_t *tx_lnet_replymsg; /* LNET reply message to finalize */
207 kptl_msg_t *tx_msg; /* the message data */
208 kptl_peer_t *tx_peer; /* the peer this is waiting on */
209 unsigned long tx_deadline; /* deadline */
210 unsigned long tx_tposted; /* time posted */
211 ptl_md_t tx_rdma_md; /* rdma descriptor */
212 kptl_fragvec_t *tx_frags; /* buffer fragments */
215 enum kptllnd_peer_state
217 PEER_STATE_UNINITIALIZED = 0,
218 PEER_STATE_ALLOCATED = 1,
219 PEER_STATE_WAITING_HELLO = 2,
220 PEER_STATE_ACTIVE = 3,
221 PEER_STATE_CLOSING = 4,
222 PEER_STATE_ZOMBIE = 5,
227 cfs_list_t peer_list;
228 cfs_atomic_t peer_refcount; /* The current references */
229 enum kptllnd_peer_state peer_state;
230 spinlock_t peer_lock; /* serialize */
231 cfs_list_t peer_noops; /* PTLLND_MSG_TYPE_NOOP txs */
232 cfs_list_t peer_sendq; /* txs waiting for mh handles */
233 cfs_list_t peer_activeq; /* txs awaiting completion */
234 lnet_process_id_t peer_id; /* Peer's LNET id */
235 ptl_process_id_t peer_ptlid; /* Peer's portals id */
236 __u64 peer_incarnation; /* peer's incarnation */
237 __u64 peer_myincarnation; /* my incarnation at HELLO */
238 int peer_sent_hello; /* have I sent HELLO? */
239 int peer_credits; /* number of send credits */
240 int peer_outstanding_credits;/* number of peer credits to return */
241 int peer_sent_credits; /* #msg buffers posted for peer */
242 int peer_max_msg_size; /* peer's rx buffer size */
243 int peer_error; /* errno on closing this peer */
244 int peer_retry_noop; /* need to retry returning credits */
245 int peer_check_stamp; /* watchdog check stamp */
246 cfs_time_t peer_last_alive; /* when (in jiffies) I was last alive */
247 __u64 peer_next_matchbits; /* Next value to register RDMA from peer */
248 __u64 peer_last_matchbits_seen; /* last matchbits used to RDMA to peer */
253 int kptl_init; /* initialisation state */
254 volatile int kptl_shutdown; /* shut down? */
255 cfs_atomic_t kptl_nthreads; /* # live threads */
256 ptl_handle_ni_t kptl_nih; /* network inteface handle */
257 ptl_process_id_t kptl_portals_id; /* Portals ID of interface */
258 __u64 kptl_incarnation; /* which one am I */
259 ptl_handle_eq_t kptl_eqh; /* Event Queue (EQ) */
261 rwlock_t kptl_net_rw_lock; /* serialise... */
262 cfs_list_t kptl_nets; /* kptl_net instance*/
264 spinlock_t kptl_sched_lock; /* serialise... */
265 cfs_waitq_t kptl_sched_waitq; /* schedulers sleep here */
266 cfs_list_t kptl_sched_txq; /* tx requiring attention */
267 cfs_list_t kptl_sched_rxq; /* rx requiring attention */
268 cfs_list_t kptl_sched_rxbq; /* rxb requiring reposting */
270 cfs_waitq_t kptl_watchdog_waitq; /* watchdog sleeps here */
272 kptl_rx_buffer_pool_t kptl_rx_buffer_pool; /* rx buffer pool */
273 cfs_mem_cache_t* kptl_rx_cache; /* rx descripter cache */
275 cfs_atomic_t kptl_ntx; /* # tx descs allocated */
276 spinlock_t kptl_tx_lock; /* serialise idle tx list*/
277 cfs_list_t kptl_idle_txs; /* idle tx descriptors */
279 rwlock_t kptl_peer_rw_lock; /* lock for peer table */
280 cfs_list_t *kptl_peers; /* hash table of all my known peers */
281 cfs_list_t kptl_closing_peers; /* peers being closed */
282 cfs_list_t kptl_zombie_peers; /* peers waiting for refs to drain */
283 int kptl_peer_hash_size; /* size of kptl_peers */
284 int kptl_npeers; /* # peers extant */
285 int kptl_n_active_peers; /* # active peers */
286 int kptl_expected_peers; /* # peers I can buffer HELLOs from */
288 kptl_msg_t *kptl_nak_msg; /* common NAK message */
289 spinlock_t kptl_ptlid2str_lock; /* serialise str ops */
294 cfs_list_t net_list; /* chain on kptl_data:: kptl_nets */
296 cfs_atomic_t net_refcount; /* # current references */
297 int net_shutdown; /* lnd_shutdown called */
302 PTLLND_INIT_NOTHING = 0,
307 extern kptl_tunables_t kptllnd_tunables;
308 extern kptl_data_t kptllnd_data;
310 static inline lnet_nid_t
311 kptllnd_ptl2lnetnid(lnet_nid_t ni_nid, ptl_nid_t ptl_nid)
313 #ifdef _USING_LUSTRE_PORTALS_
314 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(ptl_nid));
316 return LNET_MKNID(LNET_NIDNET(ni_nid), ptl_nid);
320 static inline ptl_nid_t
321 kptllnd_lnet2ptlnid(lnet_nid_t lnet_nid)
323 #ifdef _USING_LUSTRE_PORTALS_
324 return LNET_MKNID(LNET_NIDNET(kptllnd_data.kptl_portals_id.nid),
325 LNET_NIDADDR(lnet_nid));
327 return LNET_NIDADDR(lnet_nid);
331 int kptllnd_startup(lnet_ni_t *ni);
332 void kptllnd_shutdown(lnet_ni_t *ni);
333 int kptllnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
334 void kptllnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
335 int kptllnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
336 int kptllnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
337 int delayed, unsigned int niov,
338 struct iovec *iov, lnet_kiov_t *kiov,
339 unsigned int offset, unsigned int mlen, unsigned int rlen);
340 int kptllnd_eager_recv(struct lnet_ni *ni, void *private,
341 lnet_msg_t *msg, void **new_privatep);
342 void kptllnd_eq_callback(ptl_event_t *evp);
343 int kptllnd_scheduler(void *arg);
344 int kptllnd_watchdog(void *arg);
345 int kptllnd_thread_start(int (*fn)(void *arg), void *arg);
346 int kptllnd_tunables_init(void);
347 void kptllnd_tunables_fini(void);
349 const char *kptllnd_evtype2str(int evtype);
350 const char *kptllnd_msgtype2str(int msgtype);
351 const char *kptllnd_errtype2str(int errtype);
354 kptllnd_eventarg2obj (kptl_eventarg_t *eva)
356 switch (eva->eva_type) {
359 case PTLLND_EVENTARG_TYPE_BUF:
360 return cfs_list_entry(eva, kptl_rx_buffer_t, rxb_eventarg);
361 case PTLLND_EVENTARG_TYPE_RDMA:
362 return cfs_list_entry(eva, kptl_tx_t, tx_rdma_eventarg);
363 case PTLLND_EVENTARG_TYPE_MSG:
364 return cfs_list_entry(eva, kptl_tx_t, tx_msg_eventarg);
369 * RX BUFFER SUPPORT FUNCTIONS
371 void kptllnd_rx_buffer_pool_init(kptl_rx_buffer_pool_t *rxbp);
372 void kptllnd_rx_buffer_pool_fini(kptl_rx_buffer_pool_t *rxbp);
373 int kptllnd_rx_buffer_pool_reserve(kptl_rx_buffer_pool_t *rxbp, int count);
374 void kptllnd_rx_buffer_pool_unreserve(kptl_rx_buffer_pool_t *rxbp, int count);
375 void kptllnd_rx_buffer_callback(ptl_event_t *ev);
376 void kptllnd_rx_buffer_post(kptl_rx_buffer_t *rxb);
379 kptllnd_rx_buffer_size(void)
381 return PAGE_SIZE * (*kptllnd_tunables.kptl_rxb_npages);
385 kptllnd_rx_buffer_addref(kptl_rx_buffer_t *rxb)
389 spin_lock_irqsave(&rxb->rxb_pool->rxbp_lock, flags);
391 spin_unlock_irqrestore(&rxb->rxb_pool->rxbp_lock, flags);
395 kptllnd_rx_buffer_decref_locked(kptl_rx_buffer_t *rxb)
397 if (--(rxb->rxb_refcount) == 0) {
398 spin_lock(&kptllnd_data.kptl_sched_lock);
400 cfs_list_add_tail(&rxb->rxb_repost_list,
401 &kptllnd_data.kptl_sched_rxbq);
402 cfs_waitq_signal(&kptllnd_data.kptl_sched_waitq);
404 spin_unlock(&kptllnd_data.kptl_sched_lock);
409 kptllnd_rx_buffer_decref(kptl_rx_buffer_t *rxb)
414 spin_lock_irqsave(&rxb->rxb_pool->rxbp_lock, flags);
415 count = --(rxb->rxb_refcount);
416 spin_unlock_irqrestore(&rxb->rxb_pool->rxbp_lock, flags);
419 kptllnd_rx_buffer_post(rxb);
423 * RX SUPPORT FUNCTIONS
425 void kptllnd_rx_parse(kptl_rx_t *rx);
426 void kptllnd_rx_done(kptl_rx_t *rx, int post_credit);
429 * PEER SUPPORT FUNCTIONS
431 int kptllnd_get_peer_info(int index,
432 lnet_process_id_t *id,
433 int *state, int *sent_hello,
434 int *refcount, __u64 *incarnation,
435 __u64 *next_matchbits, __u64 *last_matchbits_seen,
436 int *nsendq, int *nactiveq,
437 int *credits, int *outstanding_credits);
438 void kptllnd_peer_destroy(kptl_peer_t *peer);
439 int kptllnd_peer_del(lnet_process_id_t id);
440 void kptllnd_peer_close_locked(kptl_peer_t *peer, int why);
441 void kptllnd_peer_close(kptl_peer_t *peer, int why);
442 void kptllnd_handle_closing_peers(void);
443 int kptllnd_peer_connect(kptl_tx_t *tx, lnet_nid_t nid);
444 void kptllnd_peer_check_sends(kptl_peer_t *peer);
445 void kptllnd_peer_check_bucket(int idx, int stamp);
446 void kptllnd_tx_launch(kptl_peer_t *peer, kptl_tx_t *tx, int nfrag);
447 int kptllnd_find_target(kptl_net_t *net, lnet_process_id_t target,
448 kptl_peer_t **peerp);
449 kptl_peer_t *kptllnd_peer_handle_hello(kptl_net_t *net,
450 ptl_process_id_t initiator,
452 kptl_peer_t *kptllnd_id2peer_locked(lnet_process_id_t id);
453 void kptllnd_peer_alive(kptl_peer_t *peer);
456 kptllnd_peer_addref (kptl_peer_t *peer)
458 cfs_atomic_inc(&peer->peer_refcount);
462 kptllnd_peer_decref (kptl_peer_t *peer)
464 if (cfs_atomic_dec_and_test(&peer->peer_refcount))
465 kptllnd_peer_destroy(peer);
469 kptllnd_net_addref (kptl_net_t *net)
471 LASSERT (cfs_atomic_read(&net->net_refcount) > 0);
472 cfs_atomic_inc(&net->net_refcount);
476 kptllnd_net_decref (kptl_net_t *net)
478 LASSERT (cfs_atomic_read(&net->net_refcount) > 0);
479 cfs_atomic_dec(&net->net_refcount);
483 kptllnd_set_tx_peer(kptl_tx_t *tx, kptl_peer_t *peer)
485 LASSERT (tx->tx_peer == NULL);
487 kptllnd_peer_addref(peer);
491 static inline cfs_list_t *
492 kptllnd_nid2peerlist(lnet_nid_t nid)
494 /* Only one copy of peer state for all logical peers, so the net part
495 * of NIDs is ignored; e.g. A@ptl0 and A@ptl2 share peer state */
496 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) %
497 kptllnd_data.kptl_peer_hash_size;
499 return &kptllnd_data.kptl_peers[hash];
502 static inline kptl_peer_t *
503 kptllnd_id2peer(lnet_process_id_t id)
508 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
509 peer = kptllnd_id2peer_locked(id);
510 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
516 kptllnd_reserve_buffers(int n)
518 return kptllnd_rx_buffer_pool_reserve(&kptllnd_data.kptl_rx_buffer_pool,
523 kptllnd_peer_reserve_buffers(void)
525 return kptllnd_reserve_buffers(*kptllnd_tunables.kptl_peertxcredits);
529 kptllnd_peer_unreserve_buffers(void)
531 kptllnd_rx_buffer_pool_unreserve(&kptllnd_data.kptl_rx_buffer_pool,
532 *kptllnd_tunables.kptl_peertxcredits);
536 * TX SUPPORT FUNCTIONS
538 int kptllnd_setup_tx_descs(void);
539 void kptllnd_cleanup_tx_descs(void);
540 void kptllnd_tx_fini(kptl_tx_t *tx);
541 void kptllnd_cancel_txlist(cfs_list_t *peerq, cfs_list_t *txs);
542 void kptllnd_restart_txs(kptl_net_t *net, lnet_process_id_t id,
543 cfs_list_t *restarts);
544 kptl_tx_t *kptllnd_get_idle_tx(enum kptl_tx_type purpose);
545 void kptllnd_tx_callback(ptl_event_t *ev);
546 const char *kptllnd_tx_typestr(int type);
549 kptllnd_tx_addref(kptl_tx_t *tx)
551 cfs_atomic_inc(&tx->tx_refcount);
555 kptllnd_tx_decref(kptl_tx_t *tx)
557 LASSERT (!cfs_in_interrupt()); /* Thread context only */
559 if (cfs_atomic_dec_and_test(&tx->tx_refcount))
564 * MESSAGE SUPPORT FUNCTIONS
566 void kptllnd_init_msg(kptl_msg_t *msg, int type,
567 lnet_process_id_t target, int body_nob);
568 void kptllnd_msg_pack(kptl_msg_t *msg, kptl_peer_t *peer);
569 int kptllnd_msg_unpack(kptl_msg_t *msg, int nob);
572 * MISC SUPPORT FUNCTIONS
574 void kptllnd_init_rdma_md(kptl_tx_t *tx, unsigned int niov,
575 struct iovec *iov, lnet_kiov_t *kiov,
576 unsigned int offset, unsigned int nob);
577 char *kptllnd_ptlid2str(ptl_process_id_t id);
579 void kptllnd_init_ptltrace(void);
580 void kptllnd_dump_ptltrace(void);
583 #define SIMULATION_FAIL_TX_PUT_ALLOC 0 /* 0x00000001 */
584 #define SIMULATION_FAIL_TX_GET_ALLOC 1 /* 0x00000002 */
585 #define SIMULATION_FAIL_TX 2 /* 0x00000004 */
586 #define SIMULATION_FAIL_RX_ALLOC 3 /* 0x00000008 */
588 #define IS_SIMULATION_ENABLED(x) \
589 (((*kptllnd_tunables.kptl_simulation_bitmap) & 1<< SIMULATION_##x) != 0)
591 #define IS_SIMULATION_ENABLED(x) 0