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.
31 * This file is part of Lustre, http://www.lustre.org/
32 * Lustre is a trademark of Sun Microsystems, Inc.
34 * lnet/klnds/ptllnd/ptllnd.h
36 * Author: PJ Kirner <pjkirner@clusterfs.com>
39 #ifndef AUTOCONF_INCLUDED
40 #include <linux/config.h>
42 #include <linux/module.h>
43 #include <linux/kernel.h>
45 #include <linux/string.h>
46 #include <linux/stat.h>
47 #include <linux/errno.h>
48 #include <linux/smp_lock.h>
49 #include <linux/unistd.h>
50 #include <linux/uio.h>
52 #include <asm/system.h>
53 #include <asm/uaccess.h>
56 #include <linux/init.h>
58 #include <linux/file.h>
59 #include <linux/stat.h>
60 #include <linux/list.h>
61 #include <linux/kmod.h>
62 #include <linux/sysctl.h>
63 #include <linux/random.h>
69 #define DEBUG_SUBSYSTEM S_LND
71 #include <libcfs/libcfs.h>
72 #include <lnet/lnet.h>
73 #include <lnet/lib-lnet.h>
74 #include <lnet/lnet-sysctl.h>
75 #include <portals/p30.h>
76 #include <lnet/ptllnd.h> /* Depends on portals/p30.h */
79 * Define this to enable console debug logging
82 //#define PJK_DEBUGGING
85 # define PTLLND_N_SCHED cfs_num_online_cpus() /* # schedulers */
87 # define PTLLND_N_SCHED 1 /* # schedulers */
90 #define PTLLND_CREDIT_HIGHWATER ((*kptllnd_tunables.kptl_peertxcredits)-1)
91 /* when eagerly to return credits */
95 int *kptl_ntx; /* # tx descs to pre-allocate */
96 int *kptl_max_nodes; /* max # nodes all talking to me */
97 int *kptl_max_procs_per_node; /* max # processes per node */
98 int *kptl_checksum; /* checksum kptl_msg_t? */
99 int *kptl_timeout; /* comms timeout (seconds) */
100 int *kptl_portal; /* portal number */
101 int *kptl_pid; /* portals PID (self + kernel peers) */
102 int *kptl_rxb_npages; /* number of pages for rx buffer */
103 int *kptl_rxb_nspare; /* number of spare rx buffers */
104 int *kptl_credits; /* number of credits */
105 int *kptl_peertxcredits; /* number of peer tx credits */
106 int *kptl_peerrtrcredits; /* number of peer router credits */
107 int *kptl_max_msg_size; /* max immd message size*/
108 int *kptl_peer_hash_table_size; /* # slots in peer hash table */
109 int *kptl_reschedule_loops; /* scheduler yield loops */
110 int *kptl_ack_puts; /* make portals ack PUTs */
112 int *kptl_simulation_bitmap;/* simulation bitmap */
115 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
116 cfs_sysctl_table_header_t *kptl_sysctl; /* sysctl interface */
120 #include "lnet/ptllnd_wire.h"
122 /***********************************************************************/
124 typedef struct kptl_data kptl_data_t;
125 typedef struct kptl_net kptl_net_t;
126 typedef struct kptl_rx_buffer kptl_rx_buffer_t;
127 typedef struct kptl_peer kptl_peer_t;
133 #define PTLLND_EVENTARG_TYPE_MSG 0x1
134 #define PTLLND_EVENTARG_TYPE_RDMA 0x2
135 #define PTLLND_EVENTARG_TYPE_BUF 0x3
137 typedef struct kptl_rx /* receive message */
139 cfs_list_t rx_list; /* queue for attention */
140 kptl_rx_buffer_t *rx_rxb; /* the rx buffer pointer */
141 kptl_msg_t *rx_msg; /* received message */
142 int rx_nob; /* received message size */
143 unsigned long rx_treceived; /* time received */
144 ptl_process_id_t rx_initiator; /* sender's address */
145 kptl_peer_t *rx_peer; /* pointer to peer */
146 char rx_space[0]; /* copy of incoming request */
149 #define PTLLND_POSTRX_DONT_POST 0 /* don't post */
150 #define PTLLND_POSTRX_NO_CREDIT 1 /* post: no credits */
151 #define PTLLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
153 typedef struct kptl_rx_buffer_pool
155 cfs_spinlock_t rxbp_lock;
156 cfs_list_t rxbp_list; /* all allocated buffers */
157 int rxbp_count; /* # allocated buffers */
158 int rxbp_reserved; /* # requests to buffer */
159 int rxbp_shutdown; /* shutdown flag */
160 } kptl_rx_buffer_pool_t;
162 struct kptl_rx_buffer
164 kptl_rx_buffer_pool_t *rxb_pool;
165 cfs_list_t rxb_list; /* for the rxb_pool list */
166 cfs_list_t rxb_repost_list;/* for the kptl_sched_rxbq list */
167 int rxb_posted:1; /* on the net */
168 int rxb_idle:1; /* all done */
169 kptl_eventarg_t rxb_eventarg; /* event->md.user_ptr */
170 int rxb_refcount; /* reference count */
171 ptl_handle_md_t rxb_mdh; /* the portals memory descriptor (MD) handle */
172 char *rxb_buffer; /* the buffer */
178 TX_TYPE_RESERVED = 0,
179 TX_TYPE_SMALL_MESSAGE = 1,
180 TX_TYPE_PUT_REQUEST = 2,
181 TX_TYPE_GET_REQUEST = 3,
182 TX_TYPE_PUT_RESPONSE = 4,
183 TX_TYPE_GET_RESPONSE = 5,
187 #ifdef _USING_LUSTRE_PORTALS_
188 struct iovec iov[PTL_MD_MAX_IOV];
189 lnet_kiov_t kiov[PTL_MD_MAX_IOV];
191 ptl_md_iovec_t iov[PTL_MD_MAX_IOV];
195 typedef struct kptl_tx /* transmit message */
197 cfs_list_t tx_list; /* queue on idle_txs etc */
198 cfs_atomic_t tx_refcount; /* reference count*/
199 enum kptl_tx_type tx_type; /* small msg/{put,get}{req,resp} */
200 int tx_active:1; /* queued on the peer */
201 int tx_idle:1; /* on the free list */
202 int tx_acked:1; /* portals ACK wanted (for debug only) */
203 kptl_eventarg_t tx_msg_eventarg; /* event->md.user_ptr */
204 kptl_eventarg_t tx_rdma_eventarg; /* event->md.user_ptr */
205 int tx_status; /* the status of this tx descriptor */
206 ptl_handle_md_t tx_rdma_mdh; /* RDMA buffer */
207 ptl_handle_md_t tx_msg_mdh; /* the portals MD handle for the initial message */
208 lnet_msg_t *tx_lnet_msg; /* LNET message to finalize */
209 lnet_msg_t *tx_lnet_replymsg; /* LNET reply message to finalize */
210 kptl_msg_t *tx_msg; /* the message data */
211 kptl_peer_t *tx_peer; /* the peer this is waiting on */
212 unsigned long tx_deadline; /* deadline */
213 unsigned long tx_tposted; /* time posted */
214 ptl_md_t tx_rdma_md; /* rdma descriptor */
215 kptl_fragvec_t *tx_frags; /* buffer fragments */
218 enum kptllnd_peer_state
220 PEER_STATE_UNINITIALIZED = 0,
221 PEER_STATE_ALLOCATED = 1,
222 PEER_STATE_WAITING_HELLO = 2,
223 PEER_STATE_ACTIVE = 3,
224 PEER_STATE_CLOSING = 4,
225 PEER_STATE_ZOMBIE = 5,
230 cfs_list_t peer_list;
231 cfs_atomic_t peer_refcount; /* The current references */
232 enum kptllnd_peer_state peer_state;
233 cfs_spinlock_t peer_lock; /* serialize */
234 cfs_list_t peer_noops; /* PTLLND_MSG_TYPE_NOOP txs */
235 cfs_list_t peer_sendq; /* txs waiting for mh handles */
236 cfs_list_t peer_activeq; /* txs awaiting completion */
237 lnet_process_id_t peer_id; /* Peer's LNET id */
238 ptl_process_id_t peer_ptlid; /* Peer's portals id */
239 __u64 peer_incarnation; /* peer's incarnation */
240 __u64 peer_myincarnation; /* my incarnation at HELLO */
241 int peer_sent_hello; /* have I sent HELLO? */
242 int peer_credits; /* number of send credits */
243 int peer_outstanding_credits;/* number of peer credits to return */
244 int peer_sent_credits; /* #msg buffers posted for peer */
245 int peer_max_msg_size; /* peer's rx buffer size */
246 int peer_error; /* errno on closing this peer */
247 int peer_retry_noop; /* need to retry returning credits */
248 int peer_check_stamp; /* watchdog check stamp */
249 cfs_time_t peer_last_alive; /* when (in jiffies) I was last alive */
250 __u64 peer_next_matchbits; /* Next value to register RDMA from peer */
251 __u64 peer_last_matchbits_seen; /* last matchbits used to RDMA to peer */
256 int kptl_init; /* initialisation state */
257 volatile int kptl_shutdown; /* shut down? */
258 cfs_atomic_t kptl_nthreads; /* # live threads */
259 ptl_handle_ni_t kptl_nih; /* network inteface handle */
260 ptl_process_id_t kptl_portals_id; /* Portals ID of interface */
261 __u64 kptl_incarnation; /* which one am I */
262 ptl_handle_eq_t kptl_eqh; /* Event Queue (EQ) */
264 cfs_rwlock_t kptl_net_rw_lock; /* serialise... */
265 cfs_list_t kptl_nets; /* kptl_net instances */
267 cfs_spinlock_t kptl_sched_lock; /* serialise... */
268 cfs_waitq_t kptl_sched_waitq; /* schedulers sleep here */
269 cfs_list_t kptl_sched_txq; /* tx requiring attention */
270 cfs_list_t kptl_sched_rxq; /* rx requiring attention */
271 cfs_list_t kptl_sched_rxbq; /* rxb requiring reposting */
273 cfs_waitq_t kptl_watchdog_waitq; /* watchdog sleeps here */
275 kptl_rx_buffer_pool_t kptl_rx_buffer_pool; /* rx buffer pool */
276 cfs_mem_cache_t* kptl_rx_cache; /* rx descripter cache */
278 cfs_atomic_t kptl_ntx; /* # tx descs allocated */
279 cfs_spinlock_t kptl_tx_lock; /* serialise idle tx list*/
280 cfs_list_t kptl_idle_txs; /* idle tx descriptors */
282 cfs_rwlock_t kptl_peer_rw_lock; /* lock for peer table */
283 cfs_list_t *kptl_peers; /* hash table of all my known peers */
284 cfs_list_t kptl_closing_peers; /* peers being closed */
285 cfs_list_t kptl_zombie_peers; /* peers waiting for refs to drain */
286 int kptl_peer_hash_size; /* size of kptl_peers */
287 int kptl_npeers; /* # peers extant */
288 int kptl_n_active_peers; /* # active peers */
289 int kptl_expected_peers; /* # peers I can buffer HELLOs from */
291 kptl_msg_t *kptl_nak_msg; /* common NAK message */
292 cfs_spinlock_t kptl_ptlid2str_lock; /* serialise str ops */
297 cfs_list_t net_list; /* chain on kptl_data:: kptl_nets */
299 cfs_atomic_t net_refcount; /* # current references */
300 int net_shutdown; /* lnd_shutdown called */
305 PTLLND_INIT_NOTHING = 0,
310 extern kptl_tunables_t kptllnd_tunables;
311 extern kptl_data_t kptllnd_data;
313 static inline lnet_nid_t
314 kptllnd_ptl2lnetnid(lnet_nid_t ni_nid, ptl_nid_t ptl_nid)
316 #ifdef _USING_LUSTRE_PORTALS_
317 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(ptl_nid));
319 return LNET_MKNID(LNET_NIDNET(ni_nid), ptl_nid);
323 static inline ptl_nid_t
324 kptllnd_lnet2ptlnid(lnet_nid_t lnet_nid)
326 #ifdef _USING_LUSTRE_PORTALS_
327 return LNET_MKNID(LNET_NIDNET(kptllnd_data.kptl_portals_id.nid),
328 LNET_NIDADDR(lnet_nid));
330 return LNET_NIDADDR(lnet_nid);
334 int kptllnd_startup(lnet_ni_t *ni);
335 void kptllnd_shutdown(lnet_ni_t *ni);
336 int kptllnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
337 void kptllnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
338 int kptllnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
339 int kptllnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
340 int delayed, unsigned int niov,
341 struct iovec *iov, lnet_kiov_t *kiov,
342 unsigned int offset, unsigned int mlen, unsigned int rlen);
343 int kptllnd_eager_recv(struct lnet_ni *ni, void *private,
344 lnet_msg_t *msg, void **new_privatep);
345 void kptllnd_eq_callback(ptl_event_t *evp);
346 int kptllnd_scheduler(void *arg);
347 int kptllnd_watchdog(void *arg);
348 int kptllnd_thread_start(int (*fn)(void *arg), void *arg);
349 int kptllnd_tunables_init(void);
350 void kptllnd_tunables_fini(void);
352 const char *kptllnd_evtype2str(int evtype);
353 const char *kptllnd_msgtype2str(int msgtype);
354 const char *kptllnd_errtype2str(int errtype);
357 kptllnd_eventarg2obj (kptl_eventarg_t *eva)
359 switch (eva->eva_type) {
362 case PTLLND_EVENTARG_TYPE_BUF:
363 return cfs_list_entry(eva, kptl_rx_buffer_t, rxb_eventarg);
364 case PTLLND_EVENTARG_TYPE_RDMA:
365 return cfs_list_entry(eva, kptl_tx_t, tx_rdma_eventarg);
366 case PTLLND_EVENTARG_TYPE_MSG:
367 return cfs_list_entry(eva, kptl_tx_t, tx_msg_eventarg);
372 * RX BUFFER SUPPORT FUNCTIONS
374 void kptllnd_rx_buffer_pool_init(kptl_rx_buffer_pool_t *rxbp);
375 void kptllnd_rx_buffer_pool_fini(kptl_rx_buffer_pool_t *rxbp);
376 int kptllnd_rx_buffer_pool_reserve(kptl_rx_buffer_pool_t *rxbp, int count);
377 void kptllnd_rx_buffer_pool_unreserve(kptl_rx_buffer_pool_t *rxbp, int count);
378 void kptllnd_rx_buffer_callback(ptl_event_t *ev);
379 void kptllnd_rx_buffer_post(kptl_rx_buffer_t *rxb);
382 kptllnd_rx_buffer_size(void)
384 return PAGE_SIZE * (*kptllnd_tunables.kptl_rxb_npages);
388 kptllnd_rx_buffer_addref(kptl_rx_buffer_t *rxb)
392 cfs_spin_lock_irqsave(&rxb->rxb_pool->rxbp_lock, flags);
394 cfs_spin_unlock_irqrestore(&rxb->rxb_pool->rxbp_lock, flags);
398 kptllnd_rx_buffer_decref_locked(kptl_rx_buffer_t *rxb)
400 if (--(rxb->rxb_refcount) == 0) {
401 cfs_spin_lock(&kptllnd_data.kptl_sched_lock);
403 cfs_list_add_tail(&rxb->rxb_repost_list,
404 &kptllnd_data.kptl_sched_rxbq);
405 cfs_waitq_signal(&kptllnd_data.kptl_sched_waitq);
407 cfs_spin_unlock(&kptllnd_data.kptl_sched_lock);
412 kptllnd_rx_buffer_decref(kptl_rx_buffer_t *rxb)
417 cfs_spin_lock_irqsave(&rxb->rxb_pool->rxbp_lock, flags);
418 count = --(rxb->rxb_refcount);
419 cfs_spin_unlock_irqrestore(&rxb->rxb_pool->rxbp_lock, flags);
422 kptllnd_rx_buffer_post(rxb);
426 * RX SUPPORT FUNCTIONS
428 void kptllnd_rx_parse(kptl_rx_t *rx);
429 void kptllnd_rx_done(kptl_rx_t *rx, int post_credit);
432 * PEER SUPPORT FUNCTIONS
434 int kptllnd_get_peer_info(int index,
435 lnet_process_id_t *id,
436 int *state, int *sent_hello,
437 int *refcount, __u64 *incarnation,
438 __u64 *next_matchbits, __u64 *last_matchbits_seen,
439 int *nsendq, int *nactiveq,
440 int *credits, int *outstanding_credits);
441 void kptllnd_peer_destroy(kptl_peer_t *peer);
442 int kptllnd_peer_del(lnet_process_id_t id);
443 void kptllnd_peer_close_locked(kptl_peer_t *peer, int why);
444 void kptllnd_peer_close(kptl_peer_t *peer, int why);
445 void kptllnd_handle_closing_peers(void);
446 int kptllnd_peer_connect(kptl_tx_t *tx, lnet_nid_t nid);
447 void kptllnd_peer_check_sends(kptl_peer_t *peer);
448 void kptllnd_peer_check_bucket(int idx, int stamp);
449 void kptllnd_tx_launch(kptl_peer_t *peer, kptl_tx_t *tx, int nfrag);
450 int kptllnd_find_target(kptl_net_t *net, lnet_process_id_t target,
451 kptl_peer_t **peerp);
452 kptl_peer_t *kptllnd_peer_handle_hello(kptl_net_t *net,
453 ptl_process_id_t initiator,
455 kptl_peer_t *kptllnd_id2peer_locked(lnet_process_id_t id);
456 void kptllnd_peer_alive(kptl_peer_t *peer);
459 kptllnd_peer_addref (kptl_peer_t *peer)
461 cfs_atomic_inc(&peer->peer_refcount);
465 kptllnd_peer_decref (kptl_peer_t *peer)
467 if (cfs_atomic_dec_and_test(&peer->peer_refcount))
468 kptllnd_peer_destroy(peer);
472 kptllnd_net_addref (kptl_net_t *net)
474 LASSERT (cfs_atomic_read(&net->net_refcount) > 0);
475 cfs_atomic_inc(&net->net_refcount);
479 kptllnd_net_decref (kptl_net_t *net)
481 LASSERT (cfs_atomic_read(&net->net_refcount) > 0);
482 cfs_atomic_dec(&net->net_refcount);
486 kptllnd_set_tx_peer(kptl_tx_t *tx, kptl_peer_t *peer)
488 LASSERT (tx->tx_peer == NULL);
490 kptllnd_peer_addref(peer);
494 static inline cfs_list_t *
495 kptllnd_nid2peerlist(lnet_nid_t nid)
497 /* Only one copy of peer state for all logical peers, so the net part
498 * of NIDs is ignored; e.g. A@ptl0 and A@ptl2 share peer state */
499 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) %
500 kptllnd_data.kptl_peer_hash_size;
502 return &kptllnd_data.kptl_peers[hash];
505 static inline kptl_peer_t *
506 kptllnd_id2peer(lnet_process_id_t id)
511 cfs_read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
512 peer = kptllnd_id2peer_locked(id);
513 cfs_read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
519 kptllnd_reserve_buffers(int n)
521 return kptllnd_rx_buffer_pool_reserve(&kptllnd_data.kptl_rx_buffer_pool,
526 kptllnd_peer_reserve_buffers(void)
528 return kptllnd_reserve_buffers(*kptllnd_tunables.kptl_peertxcredits);
532 kptllnd_peer_unreserve_buffers(void)
534 kptllnd_rx_buffer_pool_unreserve(&kptllnd_data.kptl_rx_buffer_pool,
535 *kptllnd_tunables.kptl_peertxcredits);
539 * TX SUPPORT FUNCTIONS
541 int kptllnd_setup_tx_descs(void);
542 void kptllnd_cleanup_tx_descs(void);
543 void kptllnd_tx_fini(kptl_tx_t *tx);
544 void kptllnd_cancel_txlist(cfs_list_t *peerq, cfs_list_t *txs);
545 void kptllnd_restart_txs(kptl_net_t *net, lnet_process_id_t id,
546 cfs_list_t *restarts);
547 kptl_tx_t *kptllnd_get_idle_tx(enum kptl_tx_type purpose);
548 void kptllnd_tx_callback(ptl_event_t *ev);
549 const char *kptllnd_tx_typestr(int type);
552 kptllnd_tx_addref(kptl_tx_t *tx)
554 cfs_atomic_inc(&tx->tx_refcount);
558 kptllnd_tx_decref(kptl_tx_t *tx)
560 LASSERT (!cfs_in_interrupt()); /* Thread context only */
562 if (cfs_atomic_dec_and_test(&tx->tx_refcount))
567 * MESSAGE SUPPORT FUNCTIONS
569 void kptllnd_init_msg(kptl_msg_t *msg, int type,
570 lnet_process_id_t target, int body_nob);
571 void kptllnd_msg_pack(kptl_msg_t *msg, kptl_peer_t *peer);
572 int kptllnd_msg_unpack(kptl_msg_t *msg, int nob);
575 * MISC SUPPORT FUNCTIONS
577 void kptllnd_init_rdma_md(kptl_tx_t *tx, unsigned int niov,
578 struct iovec *iov, lnet_kiov_t *kiov,
579 unsigned int offset, unsigned int nob);
580 char *kptllnd_ptlid2str(ptl_process_id_t id);
582 void kptllnd_init_ptltrace(void);
583 void kptllnd_dump_ptltrace(void);
586 #define SIMULATION_FAIL_TX_PUT_ALLOC 0 /* 0x00000001 */
587 #define SIMULATION_FAIL_TX_GET_ALLOC 1 /* 0x00000002 */
588 #define SIMULATION_FAIL_TX 2 /* 0x00000004 */
589 #define SIMULATION_FAIL_RX_ALLOC 3 /* 0x00000008 */
591 #define IS_SIMULATION_ENABLED(x) \
592 (((*kptllnd_tunables.kptl_simulation_bitmap) & 1<< SIMULATION_##x) != 0)
594 #define IS_SIMULATION_ENABLED(x) 0