1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
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>
42 # define EXPORT_SYMTAB
44 #ifndef AUTOCONF_INCLUDED
45 #include <linux/config.h>
47 #include <linux/module.h>
48 #include <linux/kernel.h>
50 #include <linux/string.h>
51 #include <linux/stat.h>
52 #include <linux/errno.h>
53 #include <linux/smp_lock.h>
54 #include <linux/unistd.h>
55 #include <linux/uio.h>
57 #include <asm/system.h>
58 #include <asm/uaccess.h>
61 #include <linux/init.h>
63 #include <linux/file.h>
64 #include <linux/stat.h>
65 #include <linux/list.h>
66 #include <linux/kmod.h>
67 #include <linux/sysctl.h>
68 #include <linux/random.h>
74 #define DEBUG_SUBSYSTEM S_LND
76 #include <libcfs/libcfs.h>
77 #include <lnet/lnet.h>
78 #include <lnet/lib-lnet.h>
79 #include <lnet/lnet-sysctl.h>
80 #include <portals/p30.h>
82 #include <portals/ptltrace.h>
84 #include <lnet/ptllnd.h> /* Depends on portals/p30.h */
87 * Define this to enable console debug logging
90 //#define PJK_DEBUGGING
93 # define PTLLND_N_SCHED num_online_cpus() /* # schedulers */
95 # define PTLLND_N_SCHED 1 /* # schedulers */
98 #define PTLLND_CREDIT_HIGHWATER ((*kptllnd_tunables.kptl_peertxcredits)-1)
99 /* when eagerly to return credits */
103 int *kptl_ntx; /* # tx descs to pre-allocate */
104 int *kptl_max_nodes; /* max # nodes all talking to me */
105 int *kptl_max_procs_per_node; /* max # processes per node */
106 int *kptl_checksum; /* checksum kptl_msg_t? */
107 int *kptl_timeout; /* comms timeout (seconds) */
108 int *kptl_portal; /* portal number */
109 int *kptl_pid; /* portals PID (self + kernel peers) */
110 int *kptl_rxb_npages; /* number of pages for rx buffer */
111 int *kptl_rxb_nspare; /* number of spare rx buffers */
112 int *kptl_credits; /* number of credits */
113 int *kptl_peertxcredits; /* number of peer tx credits */
114 int *kptl_peerrtrcredits; /* number of peer router credits */
115 int *kptl_max_msg_size; /* max immd message size*/
116 int *kptl_peer_hash_table_size; /* # slots in peer hash table */
117 int *kptl_reschedule_loops; /* scheduler yield loops */
118 int *kptl_ack_puts; /* make portals ack PUTs */
120 int *kptl_ptltrace_on_timeout; /* dump pltrace on timeout? */
121 int *kptl_ptltrace_on_fail; /* dump pltrace on PTL_NAL_FAILED? */
122 char **kptl_ptltrace_basename; /* ptltrace dump file basename */
125 int *kptl_simulation_bitmap;/* simulation bitmap */
128 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
129 cfs_sysctl_table_header_t *kptl_sysctl; /* sysctl interface */
133 #include "lnet/ptllnd_wire.h"
135 /***********************************************************************/
137 typedef struct kptl_data kptl_data_t;
138 typedef struct kptl_net kptl_net_t;
139 typedef struct kptl_rx_buffer kptl_rx_buffer_t;
140 typedef struct kptl_peer kptl_peer_t;
146 #define PTLLND_EVENTARG_TYPE_MSG 0x1
147 #define PTLLND_EVENTARG_TYPE_RDMA 0x2
148 #define PTLLND_EVENTARG_TYPE_BUF 0x3
150 typedef struct kptl_rx /* receive message */
152 struct list_head rx_list; /* queue for attention */
153 kptl_rx_buffer_t *rx_rxb; /* the rx buffer pointer */
154 kptl_msg_t *rx_msg; /* received message */
155 int rx_nob; /* received message size */
156 unsigned long rx_treceived; /* time received */
157 ptl_process_id_t rx_initiator; /* sender's address */
159 ptl_uid_t rx_uid; /* sender's uid */
161 kptl_peer_t *rx_peer; /* pointer to peer */
162 char rx_space[0]; /* copy of incoming request */
165 #define PTLLND_POSTRX_DONT_POST 0 /* don't post */
166 #define PTLLND_POSTRX_NO_CREDIT 1 /* post: no credits */
167 #define PTLLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
169 typedef struct kptl_rx_buffer_pool
171 spinlock_t rxbp_lock;
172 struct list_head rxbp_list; /* all allocated buffers */
173 int rxbp_count; /* # allocated buffers */
174 int rxbp_reserved; /* # requests to buffer */
175 int rxbp_shutdown; /* shutdown flag */
176 } kptl_rx_buffer_pool_t;
178 struct kptl_rx_buffer
180 kptl_rx_buffer_pool_t *rxb_pool;
181 struct list_head rxb_list; /* for the rxb_pool list */
182 struct list_head rxb_repost_list;/* for the kptl_sched_rxbq list */
183 int rxb_posted:1; /* on the net */
184 int rxb_idle:1; /* all done */
185 kptl_eventarg_t rxb_eventarg; /* event->md.user_ptr */
186 int rxb_refcount; /* reference count */
187 ptl_handle_md_t rxb_mdh; /* the portals memory descriptor (MD) handle */
188 char *rxb_buffer; /* the buffer */
194 TX_TYPE_RESERVED = 0,
195 TX_TYPE_SMALL_MESSAGE = 1,
196 TX_TYPE_PUT_REQUEST = 2,
197 TX_TYPE_GET_REQUEST = 3,
198 TX_TYPE_PUT_RESPONSE = 4,
199 TX_TYPE_GET_RESPONSE = 5,
203 #ifdef _USING_LUSTRE_PORTALS_
204 struct iovec iov[PTL_MD_MAX_IOV];
205 lnet_kiov_t kiov[PTL_MD_MAX_IOV];
207 ptl_md_iovec_t iov[PTL_MD_MAX_IOV];
211 typedef struct kptl_tx /* transmit message */
213 struct list_head tx_list; /* queue on idle_txs etc */
214 atomic_t tx_refcount; /* reference count*/
215 enum kptl_tx_type tx_type; /* small msg/{put,get}{req,resp} */
216 int tx_active:1; /* queued on the peer */
217 int tx_idle:1; /* on the free list */
218 int tx_acked:1; /* portals ACK wanted (for debug only) */
219 kptl_eventarg_t tx_msg_eventarg; /* event->md.user_ptr */
220 kptl_eventarg_t tx_rdma_eventarg; /* event->md.user_ptr */
221 int tx_status; /* the status of this tx descriptor */
222 ptl_handle_md_t tx_rdma_mdh; /* RDMA buffer */
223 ptl_handle_md_t tx_msg_mdh; /* the portals MD handle for the initial message */
224 lnet_msg_t *tx_lnet_msg; /* LNET message to finalize */
225 lnet_msg_t *tx_lnet_replymsg; /* LNET reply message to finalize */
226 kptl_msg_t *tx_msg; /* the message data */
227 kptl_peer_t *tx_peer; /* the peer this is waiting on */
228 unsigned long tx_deadline; /* deadline */
229 unsigned long tx_tposted; /* time posted */
230 ptl_md_t tx_rdma_md; /* rdma descriptor */
231 kptl_fragvec_t *tx_frags; /* buffer fragments */
234 enum kptllnd_peer_state
236 PEER_STATE_UNINITIALIZED = 0,
237 PEER_STATE_ALLOCATED = 1,
238 PEER_STATE_WAITING_HELLO = 2,
239 PEER_STATE_ACTIVE = 3,
240 PEER_STATE_CLOSING = 4,
241 PEER_STATE_ZOMBIE = 5,
246 struct list_head peer_list;
247 atomic_t peer_refcount; /* The current references */
248 enum kptllnd_peer_state peer_state;
249 spinlock_t peer_lock; /* serialize */
250 struct list_head peer_noops; /* PTLLND_MSG_TYPE_NOOP txs */
251 struct list_head peer_sendq; /* txs waiting for mh handles */
252 struct list_head peer_activeq; /* txs awaiting completion */
253 lnet_process_id_t peer_id; /* Peer's LNET id */
254 ptl_process_id_t peer_ptlid; /* Peer's portals id */
255 __u64 peer_incarnation; /* peer's incarnation */
256 __u64 peer_myincarnation; /* my incarnation at HELLO */
257 int peer_sent_hello; /* have I sent HELLO? */
258 int peer_credits; /* number of send credits */
259 int peer_outstanding_credits;/* number of peer credits to return */
260 int peer_sent_credits; /* #msg buffers posted for peer */
261 int peer_max_msg_size; /* peer's rx buffer size */
262 int peer_error; /* errno on closing this peer */
263 int peer_retry_noop; /* need to retry returning credits */
264 int peer_check_stamp; /* watchdog check stamp */
265 cfs_time_t peer_last_alive; /* when (in jiffies) I was last alive */
266 __u64 peer_next_matchbits; /* Next value to register RDMA from peer */
267 __u64 peer_last_matchbits_seen; /* last matchbits used to RDMA to peer */
272 int kptl_init; /* initialisation state */
273 volatile int kptl_shutdown; /* shut down? */
274 atomic_t kptl_nthreads; /* # live threads */
275 ptl_handle_ni_t kptl_nih; /* network inteface handle */
276 ptl_process_id_t kptl_portals_id; /* Portals ID of interface */
277 __u64 kptl_incarnation; /* which one am I */
278 ptl_handle_eq_t kptl_eqh; /* Event Queue (EQ) */
280 rwlock_t kptl_net_rw_lock; /* serialise... */
281 struct list_head kptl_nets; /* kptl_net instances */
283 spinlock_t kptl_sched_lock; /* serialise... */
284 wait_queue_head_t kptl_sched_waitq; /* schedulers sleep here */
285 struct list_head kptl_sched_txq; /* tx requiring attention */
286 struct list_head kptl_sched_rxq; /* rx requiring attention */
287 struct list_head kptl_sched_rxbq; /* rxb requiring reposting */
289 wait_queue_head_t kptl_watchdog_waitq; /* watchdog sleeps here */
290 atomic_t kptl_needs_ptltrace; /* watchdog thread to dump ptltrace */
292 kptl_rx_buffer_pool_t kptl_rx_buffer_pool; /* rx buffer pool */
293 cfs_mem_cache_t* kptl_rx_cache; /* rx descripter cache */
295 atomic_t kptl_ntx; /* # tx descs allocated */
296 spinlock_t kptl_tx_lock; /* serialise idle tx list*/
297 struct list_head kptl_idle_txs; /* idle tx descriptors */
299 rwlock_t kptl_peer_rw_lock; /* lock for peer table */
300 struct list_head *kptl_peers; /* hash table of all my known peers */
301 struct list_head kptl_closing_peers; /* peers being closed */
302 struct list_head kptl_zombie_peers; /* peers waiting for refs to drain */
303 int kptl_peer_hash_size; /* size of kptl_peers */
304 int kptl_npeers; /* # peers extant */
305 int kptl_n_active_peers; /* # active peers */
306 int kptl_expected_peers; /* # peers I can buffer HELLOs from */
308 kptl_msg_t *kptl_nak_msg; /* common NAK message */
309 spinlock_t kptl_ptlid2str_lock; /* serialise str ops */
314 struct list_head net_list; /* chain on kptl_data:: kptl_nets */
316 atomic_t net_refcount; /* # current references */
317 int net_shutdown; /* lnd_shutdown called */
322 PTLLND_INIT_NOTHING = 0,
327 extern kptl_tunables_t kptllnd_tunables;
328 extern kptl_data_t kptllnd_data;
330 static inline lnet_nid_t
331 kptllnd_ptl2lnetnid(lnet_nid_t ni_nid, ptl_nid_t ptl_nid)
333 #ifdef _USING_LUSTRE_PORTALS_
334 return LNET_MKNID(LNET_NIDNET(ni_nid), LNET_NIDADDR(ptl_nid));
336 return LNET_MKNID(LNET_NIDNET(ni_nid), ptl_nid);
340 static inline ptl_nid_t
341 kptllnd_lnet2ptlnid(lnet_nid_t lnet_nid)
343 #ifdef _USING_LUSTRE_PORTALS_
344 return LNET_MKNID(LNET_NIDNET(kptllnd_data.kptl_portals_id.nid),
345 LNET_NIDADDR(lnet_nid));
347 return LNET_NIDADDR(lnet_nid);
352 kptllnd_schedule_ptltrace_dump (void)
355 if (*kptllnd_tunables.kptl_ptltrace_on_fail) {
356 atomic_inc(&kptllnd_data.kptl_needs_ptltrace);
357 wake_up(&kptllnd_data.kptl_watchdog_waitq);
362 int kptllnd_startup(lnet_ni_t *ni);
363 void kptllnd_shutdown(lnet_ni_t *ni);
364 int kptllnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
365 void kptllnd_query (struct lnet_ni *ni, lnet_nid_t nid, time_t *when);
366 int kptllnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
367 int kptllnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
368 int delayed, unsigned int niov,
369 struct iovec *iov, lnet_kiov_t *kiov,
370 unsigned int offset, unsigned int mlen, unsigned int rlen);
371 int kptllnd_eager_recv(struct lnet_ni *ni, void *private,
372 lnet_msg_t *msg, void **new_privatep);
373 void kptllnd_eq_callback(ptl_event_t *evp);
374 int kptllnd_scheduler(void *arg);
375 int kptllnd_watchdog(void *arg);
376 int kptllnd_thread_start(int (*fn)(void *arg), void *arg);
377 int kptllnd_tunables_init(void);
378 void kptllnd_tunables_fini(void);
380 const char *kptllnd_evtype2str(int evtype);
381 const char *kptllnd_msgtype2str(int msgtype);
382 const char *kptllnd_errtype2str(int errtype);
385 kptllnd_eventarg2obj (kptl_eventarg_t *eva)
387 switch (eva->eva_type) {
390 case PTLLND_EVENTARG_TYPE_BUF:
391 return list_entry(eva, kptl_rx_buffer_t, rxb_eventarg);
392 case PTLLND_EVENTARG_TYPE_RDMA:
393 return list_entry(eva, kptl_tx_t, tx_rdma_eventarg);
394 case PTLLND_EVENTARG_TYPE_MSG:
395 return list_entry(eva, kptl_tx_t, tx_msg_eventarg);
400 * RX BUFFER SUPPORT FUNCTIONS
402 void kptllnd_rx_buffer_pool_init(kptl_rx_buffer_pool_t *rxbp);
403 void kptllnd_rx_buffer_pool_fini(kptl_rx_buffer_pool_t *rxbp);
404 int kptllnd_rx_buffer_pool_reserve(kptl_rx_buffer_pool_t *rxbp, int count);
405 void kptllnd_rx_buffer_pool_unreserve(kptl_rx_buffer_pool_t *rxbp, int count);
406 void kptllnd_rx_buffer_callback(ptl_event_t *ev);
407 void kptllnd_rx_buffer_post(kptl_rx_buffer_t *rxb);
410 kptllnd_rx_buffer_size(void)
412 return PAGE_SIZE * (*kptllnd_tunables.kptl_rxb_npages);
416 kptllnd_rx_buffer_addref(kptl_rx_buffer_t *rxb)
420 spin_lock_irqsave(&rxb->rxb_pool->rxbp_lock, flags);
422 spin_unlock_irqrestore(&rxb->rxb_pool->rxbp_lock, flags);
426 kptllnd_rx_buffer_decref_locked(kptl_rx_buffer_t *rxb)
428 if (--(rxb->rxb_refcount) == 0) {
429 spin_lock(&kptllnd_data.kptl_sched_lock);
431 list_add_tail(&rxb->rxb_repost_list,
432 &kptllnd_data.kptl_sched_rxbq);
433 wake_up(&kptllnd_data.kptl_sched_waitq);
435 spin_unlock(&kptllnd_data.kptl_sched_lock);
440 kptllnd_rx_buffer_decref(kptl_rx_buffer_t *rxb)
445 spin_lock_irqsave(&rxb->rxb_pool->rxbp_lock, flags);
446 count = --(rxb->rxb_refcount);
447 spin_unlock_irqrestore(&rxb->rxb_pool->rxbp_lock, flags);
450 kptllnd_rx_buffer_post(rxb);
454 * RX SUPPORT FUNCTIONS
456 void kptllnd_rx_parse(kptl_rx_t *rx);
457 void kptllnd_rx_done(kptl_rx_t *rx, int post_credit);
460 * PEER SUPPORT FUNCTIONS
462 int kptllnd_get_peer_info(int index,
463 lnet_process_id_t *id,
464 int *state, int *sent_hello,
465 int *refcount, __u64 *incarnation,
466 __u64 *next_matchbits, __u64 *last_matchbits_seen,
467 int *nsendq, int *nactiveq,
468 int *credits, int *outstanding_credits);
469 void kptllnd_peer_destroy(kptl_peer_t *peer);
470 int kptllnd_peer_del(lnet_process_id_t id);
471 void kptllnd_peer_close_locked(kptl_peer_t *peer, int why);
472 void kptllnd_peer_close(kptl_peer_t *peer, int why);
473 void kptllnd_handle_closing_peers(void);
474 int kptllnd_peer_connect(kptl_tx_t *tx, lnet_nid_t nid);
475 void kptllnd_peer_check_sends(kptl_peer_t *peer);
476 void kptllnd_peer_check_bucket(int idx, int stamp);
477 void kptllnd_tx_launch(kptl_peer_t *peer, kptl_tx_t *tx, int nfrag);
478 int kptllnd_find_target(kptl_net_t *net, lnet_process_id_t target,
479 kptl_peer_t **peerp);
480 kptl_peer_t *kptllnd_peer_handle_hello(kptl_net_t *net,
481 ptl_process_id_t initiator,
483 kptl_peer_t *kptllnd_id2peer_locked(lnet_process_id_t id);
484 void kptllnd_peer_alive(kptl_peer_t *peer);
487 kptllnd_peer_addref (kptl_peer_t *peer)
489 atomic_inc(&peer->peer_refcount);
493 kptllnd_peer_decref (kptl_peer_t *peer)
495 if (atomic_dec_and_test(&peer->peer_refcount))
496 kptllnd_peer_destroy(peer);
500 kptllnd_net_addref (kptl_net_t *net)
502 LASSERT (atomic_read(&net->net_refcount) > 0);
503 atomic_inc(&net->net_refcount);
507 kptllnd_net_decref (kptl_net_t *net)
509 LASSERT (atomic_read(&net->net_refcount) > 0);
510 atomic_dec(&net->net_refcount);
514 kptllnd_set_tx_peer(kptl_tx_t *tx, kptl_peer_t *peer)
516 LASSERT (tx->tx_peer == NULL);
518 kptllnd_peer_addref(peer);
522 static inline struct list_head *
523 kptllnd_nid2peerlist(lnet_nid_t nid)
525 /* Only one copy of peer state for all logical peers, so the net part
526 * of NIDs is ignored; e.g. A@ptl0 and A@ptl2 share peer state */
527 unsigned int hash = ((unsigned int)LNET_NIDADDR(nid)) %
528 kptllnd_data.kptl_peer_hash_size;
530 return &kptllnd_data.kptl_peers[hash];
533 static inline kptl_peer_t *
534 kptllnd_id2peer(lnet_process_id_t id)
539 read_lock_irqsave(&kptllnd_data.kptl_peer_rw_lock, flags);
540 peer = kptllnd_id2peer_locked(id);
541 read_unlock_irqrestore(&kptllnd_data.kptl_peer_rw_lock, flags);
547 kptllnd_reserve_buffers(int n)
549 return kptllnd_rx_buffer_pool_reserve(&kptllnd_data.kptl_rx_buffer_pool,
554 kptllnd_peer_reserve_buffers(void)
556 return kptllnd_reserve_buffers(*kptllnd_tunables.kptl_peertxcredits);
560 kptllnd_peer_unreserve_buffers(void)
562 kptllnd_rx_buffer_pool_unreserve(&kptllnd_data.kptl_rx_buffer_pool,
563 *kptllnd_tunables.kptl_peertxcredits);
567 * TX SUPPORT FUNCTIONS
569 int kptllnd_setup_tx_descs(void);
570 void kptllnd_cleanup_tx_descs(void);
571 void kptllnd_tx_fini(kptl_tx_t *tx);
572 void kptllnd_cancel_txlist(struct list_head *peerq, struct list_head *txs);
573 void kptllnd_restart_txs(kptl_net_t *net, lnet_process_id_t id, struct list_head *restarts);
574 kptl_tx_t *kptllnd_get_idle_tx(enum kptl_tx_type purpose);
575 void kptllnd_tx_callback(ptl_event_t *ev);
576 const char *kptllnd_tx_typestr(int type);
579 kptllnd_tx_addref(kptl_tx_t *tx)
581 atomic_inc(&tx->tx_refcount);
585 kptllnd_tx_decref(kptl_tx_t *tx)
587 LASSERT (!in_interrupt()); /* Thread context only */
589 if (atomic_dec_and_test(&tx->tx_refcount))
594 * MESSAGE SUPPORT FUNCTIONS
596 void kptllnd_init_msg(kptl_msg_t *msg, int type,
597 lnet_process_id_t target, int body_nob);
598 void kptllnd_msg_pack(kptl_msg_t *msg, kptl_peer_t *peer);
599 int kptllnd_msg_unpack(kptl_msg_t *msg, int nob);
602 * MISC SUPPORT FUNCTIONS
604 void kptllnd_init_rdma_md(kptl_tx_t *tx, unsigned int niov,
605 struct iovec *iov, lnet_kiov_t *kiov,
606 unsigned int offset, unsigned int nob);
607 char *kptllnd_ptlid2str(ptl_process_id_t id);
609 void kptllnd_init_ptltrace(void);
610 void kptllnd_dump_ptltrace(void);
613 #define SIMULATION_FAIL_TX_PUT_ALLOC 0 /* 0x00000001 */
614 #define SIMULATION_FAIL_TX_GET_ALLOC 1 /* 0x00000002 */
615 #define SIMULATION_FAIL_TX 2 /* 0x00000004 */
616 #define SIMULATION_FAIL_RX_ALLOC 3 /* 0x00000008 */
618 #define IS_SIMULATION_ENABLED(x) \
619 (((*kptllnd_tunables.kptl_simulation_bitmap) & 1<< SIMULATION_##x) != 0)
621 #define IS_SIMULATION_ENABLED(x) 0