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) 2004, 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/ralnd/ralnd.h
36 * Author: Eric Barton <eric@bartonsoftware.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/unistd.h>
49 #include <linux/uio.h>
51 #include <asm/system.h>
52 #include <asm/uaccess.h>
55 #include <linux/init.h>
57 #include <linux/file.h>
58 #include <linux/stat.h>
59 #include <linux/list.h>
60 #include <linux/kmod.h>
61 #include <linux/sysctl.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>
75 /* tunables determined at compile time */
76 #define RANAL_RESCHED 100 /* # scheduler loops before reschedule */
78 #define RANAL_PEER_HASH_SIZE 101 /* # peer lists */
79 #define RANAL_CONN_HASH_SIZE 101 /* # conn lists */
81 #define RANAL_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
82 #define RANAL_TIMEOUT2KEEPALIVE(t) (((t)+1)/2) /* timeout -> keepalive interval */
85 #define RANAL_MAXDEVS 2 /* max # devices RapidArray supports */
86 #define RANAL_FMA_MAX_PREFIX 232 /* max bytes in FMA "Prefix" we can use */
87 #define RANAL_FMA_MAX_DATA ((7<<10)-256) /* Max FMA MSG is 7K including prefix */
92 int *kra_n_connd; /* # connection daemons */
93 int *kra_min_reconnect_interval; /* first failed connection retry... */
94 int *kra_max_reconnect_interval; /* ...exponentially increasing to this */
95 int *kra_ntx; /* # tx descs */
96 int *kra_credits; /* # concurrent sends */
97 int *kra_peercredits; /* # concurrent sends to 1 peer */
98 int *kra_fma_cq_size; /* # entries in receive CQ */
99 int *kra_timeout; /* comms timeout (seconds) */
100 int *kra_max_immediate; /* immediate payload breakpoint */
102 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
103 cfs_sysctl_table_header_t *kra_sysctl; /* sysctl interface */
109 RAP_PVOID rad_handle; /* device handle */
110 RAP_PVOID rad_fma_cqh; /* FMA completion queue handle */
111 RAP_PVOID rad_rdma_cqh; /* rdma completion queue handle */
112 int rad_id; /* device id */
113 int rad_idx; /* index in kra_devices */
114 int rad_ready; /* set by device callback */
115 cfs_list_t rad_ready_conns;/* connections ready to tx/rx */
116 cfs_list_t rad_new_conns; /* new connections to complete */
117 cfs_waitq_t rad_waitq; /* scheduler waits here */
118 cfs_spinlock_t rad_lock; /* serialise */
119 void *rad_scheduler; /* scheduling thread */
120 unsigned int rad_nphysmap; /* # phys mappings */
121 unsigned int rad_nppphysmap;/* # phys pages mapped */
122 unsigned int rad_nvirtmap; /* # virt mappings */
123 unsigned long rad_nobvirtmap;/* # virt bytes mapped */
128 int kra_init; /* initialisation state */
129 int kra_shutdown; /* shut down? */
130 cfs_atomic_t kra_nthreads; /* # live threads */
131 lnet_ni_t *kra_ni; /* _the_ nal instance */
133 kra_device_t kra_devices[RANAL_MAXDEVS]; /* device/ptag/cq */
134 int kra_ndevs; /* # devices */
136 cfs_rwlock_t kra_global_lock; /* stabilize peer/conn ops */
138 cfs_list_t *kra_peers; /* hash table of all my known peers */
139 int kra_peer_hash_size; /* size of kra_peers */
140 cfs_atomic_t kra_npeers; /* # peers extant */
141 int kra_nonewpeers; /* prevent new peers */
143 cfs_list_t *kra_conns; /* conns hashed by cqid */
144 int kra_conn_hash_size; /* size of kra_conns */
145 __u64 kra_peerstamp; /* when I started up */
146 __u64 kra_connstamp; /* conn stamp generator */
147 int kra_next_cqid; /* cqid generator */
148 cfs_atomic_t kra_nconns; /* # connections extant */
150 long kra_new_min_timeout; /* minimum timeout on any new conn */
151 cfs_waitq_t kra_reaper_waitq; /* reaper sleeps here */
152 cfs_spinlock_t kra_reaper_lock; /* serialise */
154 cfs_list_t kra_connd_peers; /* peers waiting for a connection */
155 cfs_list_t kra_connd_acceptq; /* accepted sockets to handshake */
156 cfs_waitq_t kra_connd_waitq; /* connection daemons sleep here */
157 cfs_spinlock_t kra_connd_lock; /* serialise */
159 cfs_list_t kra_idle_txs; /* idle tx descriptors */
160 __u64 kra_next_tx_cookie; /* RDMA completion cookie */
161 cfs_spinlock_t kra_tx_lock; /* serialise */
164 #define RANAL_INIT_NOTHING 0
165 #define RANAL_INIT_DATA 1
166 #define RANAL_INIT_ALL 2
168 typedef struct kra_acceptsock /* accepted socket queued for connd */
170 cfs_list_t ras_list; /* queue for attention */
171 struct socket *ras_sock; /* the accepted socket */
174 /************************************************************************
175 * Wire message structs. These are sent in sender's byte order
176 * (i.e. receiver checks magic and flips if required).
179 typedef struct kra_connreq /* connection request/response */
180 { /* (sent via socket) */
181 __u32 racr_magic; /* I'm an ranal connreq */
182 __u16 racr_version; /* this is my version number */
183 __u16 racr_devid; /* sender's device ID */
184 __u64 racr_srcnid; /* sender's NID */
185 __u64 racr_dstnid; /* who sender expects to listen */
186 __u64 racr_peerstamp; /* sender's instance stamp */
187 __u64 racr_connstamp; /* sender's connection stamp */
188 __u32 racr_timeout; /* sender's timeout */
189 RAP_RI_PARAMETERS racr_riparams; /* sender's endpoint info */
194 RAP_MEM_KEY rard_key;
195 RAP_PVOID64 rard_addr;
201 lnet_hdr_t raim_hdr; /* portals header */
202 /* Portals payload is in FMA "Message Data" */
203 } kra_immediate_msg_t;
207 lnet_hdr_t raprm_hdr; /* portals header */
208 __u64 raprm_cookie; /* opaque completion cookie */
213 __u64 rapam_src_cookie; /* reflected completion cookie */
214 __u64 rapam_dst_cookie; /* opaque completion cookie */
215 kra_rdma_desc_t rapam_desc; /* sender's sink buffer */
220 lnet_hdr_t ragm_hdr; /* portals header */
221 __u64 ragm_cookie; /* opaque completion cookie */
222 kra_rdma_desc_t ragm_desc; /* sender's sink buffer */
227 __u64 racm_cookie; /* reflected completion cookie */
228 } kra_completion_msg_t;
230 typedef struct /* NB must fit in FMA "Prefix" */
232 __u32 ram_magic; /* I'm an ranal message */
233 __u16 ram_version; /* this is my version number */
234 __u16 ram_type; /* msg type */
235 __u64 ram_srcnid; /* sender's NID */
236 __u64 ram_connstamp; /* sender's connection stamp */
238 kra_immediate_msg_t immediate;
239 kra_putreq_msg_t putreq;
240 kra_putack_msg_t putack;
242 kra_completion_msg_t completion;
244 __u32 ram_seq; /* incrementing sequence number */
247 #define RANAL_MSG_MAGIC LNET_PROTO_RA_MAGIC /* unique magic */
248 #define RANAL_MSG_VERSION 1 /* current protocol version */
250 #define RANAL_MSG_FENCE 0x80 /* fence RDMA */
252 #define RANAL_MSG_NONE 0x00 /* illegal message */
253 #define RANAL_MSG_NOOP 0x01 /* empty ram_u (keepalive) */
254 #define RANAL_MSG_IMMEDIATE 0x02 /* ram_u.immediate */
255 #define RANAL_MSG_PUT_REQ 0x03 /* ram_u.putreq (src->sink) */
256 #define RANAL_MSG_PUT_NAK 0x04 /* ram_u.completion (no PUT match: sink->src) */
257 #define RANAL_MSG_PUT_ACK 0x05 /* ram_u.putack (PUT matched: sink->src) */
258 #define RANAL_MSG_PUT_DONE 0x86 /* ram_u.completion (src->sink) */
259 #define RANAL_MSG_GET_REQ 0x07 /* ram_u.get (sink->src) */
260 #define RANAL_MSG_GET_NAK 0x08 /* ram_u.completion (no GET match: src->sink) */
261 #define RANAL_MSG_GET_DONE 0x89 /* ram_u.completion (src->sink) */
262 #define RANAL_MSG_CLOSE 0x8a /* empty ram_u */
264 /***********************************************************************/
266 typedef struct kra_tx /* message descriptor */
268 cfs_list_t tx_list; /* queue on idle_txs/rac_sendq/rac_waitq */
269 struct kra_conn *tx_conn; /* owning conn */
270 lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
271 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
272 int tx_nob; /* # bytes of payload */
273 int tx_buftype; /* payload buffer type */
274 void *tx_buffer; /* source/sink buffer */
275 int tx_phys_offset; /* first page offset (if phys) */
276 int tx_phys_npages; /* # physical pages */
277 RAP_PHYS_REGION *tx_phys; /* page descriptors */
278 RAP_MEM_KEY tx_map_key; /* mapping key */
279 RAP_RDMA_DESCRIPTOR tx_rdma_desc; /* rdma descriptor */
280 __u64 tx_cookie; /* identify this tx to peer */
281 kra_msg_t tx_msg; /* FMA message buffer */
284 #define RANAL_BUF_NONE 0 /* buffer type not set */
285 #define RANAL_BUF_IMMEDIATE 1 /* immediate data */
286 #define RANAL_BUF_PHYS_UNMAPPED 2 /* physical: not mapped yet */
287 #define RANAL_BUF_PHYS_MAPPED 3 /* physical: mapped already */
288 #define RANAL_BUF_VIRT_UNMAPPED 4 /* virtual: not mapped yet */
289 #define RANAL_BUF_VIRT_MAPPED 5 /* virtual: mapped already */
291 typedef struct kra_conn
293 struct kra_peer *rac_peer; /* owning peer */
294 cfs_list_t rac_list; /* stash on peer's conn list */
295 cfs_list_t rac_hashlist; /* stash in connection hash table */
296 cfs_list_t rac_schedlist; /* schedule (on rad_???_conns) for attention */
297 cfs_list_t rac_fmaq; /* txs queued for FMA */
298 cfs_list_t rac_rdmaq; /* txs awaiting RDMA completion */
299 cfs_list_t rac_replyq; /* txs awaiting replies */
300 __u64 rac_peerstamp; /* peer's unique stamp */
301 __u64 rac_peer_connstamp;/* peer's unique connection stamp */
302 __u64 rac_my_connstamp; /* my unique connection stamp */
303 unsigned long rac_last_tx; /* when I last sent an FMA message (jiffies) */
304 unsigned long rac_last_rx; /* when I last received an FMA messages (jiffies) */
305 long rac_keepalive; /* keepalive interval (seconds) */
306 long rac_timeout; /* infer peer death if no rx for this many seconds */
307 __u32 rac_cqid; /* my completion callback id (non-unique) */
308 __u32 rac_tx_seq; /* tx msg sequence number */
309 __u32 rac_rx_seq; /* rx msg sequence number */
310 cfs_atomic_t rac_refcount; /* # users */
311 unsigned int rac_close_sent; /* I've sent CLOSE */
312 unsigned int rac_close_recvd; /* I've received CLOSE */
313 unsigned int rac_state; /* connection state */
314 unsigned int rac_scheduled; /* being attented to */
315 cfs_spinlock_t rac_lock; /* serialise */
316 kra_device_t *rac_device; /* which device */
317 RAP_PVOID rac_rihandle; /* RA endpoint */
318 kra_msg_t *rac_rxmsg; /* incoming message (FMA prefix) */
319 kra_msg_t rac_msg; /* keepalive/CLOSE message buffer */
322 #define RANAL_CONN_ESTABLISHED 0
323 #define RANAL_CONN_CLOSING 1
324 #define RANAL_CONN_CLOSED 2
326 typedef struct kra_peer
328 cfs_list_t rap_list; /* stash on global peer list */
329 cfs_list_t rap_connd_list; /* schedule on kra_connd_peers */
330 cfs_list_t rap_conns; /* all active connections */
331 cfs_list_t rap_tx_queue; /* msgs waiting for a conn */
332 lnet_nid_t rap_nid; /* who's on the other end(s) */
333 __u32 rap_ip; /* IP address of peer */
334 int rap_port; /* port on which peer listens */
335 cfs_atomic_t rap_refcount; /* # users */
336 int rap_persistence; /* "known" peer refs */
337 int rap_connecting; /* connection forming */
338 unsigned long rap_reconnect_time; /* CURRENT_SECONDS when reconnect OK */
339 unsigned long rap_reconnect_interval; /* exponential backoff */
342 extern kra_data_t kranal_data;
343 extern kra_tunables_t kranal_tunables;
345 extern void kranal_destroy_peer(kra_peer_t *peer);
346 extern void kranal_destroy_conn(kra_conn_t *conn);
349 kranal_peer_addref(kra_peer_t *peer)
351 CDEBUG(D_NET, "%p->%s\n", peer, libcfs_nid2str(peer->rap_nid));
352 LASSERT(cfs_atomic_read(&peer->rap_refcount) > 0);
353 cfs_atomic_inc(&peer->rap_refcount);
357 kranal_peer_decref(kra_peer_t *peer)
359 CDEBUG(D_NET, "%p->%s\n", peer, libcfs_nid2str(peer->rap_nid));
360 LASSERT(cfs_atomic_read(&peer->rap_refcount) > 0);
361 if (cfs_atomic_dec_and_test(&peer->rap_refcount))
362 kranal_destroy_peer(peer);
365 static inline cfs_list_t *
366 kranal_nid2peerlist (lnet_nid_t nid)
368 unsigned int hash = ((unsigned int)nid) % kranal_data.kra_peer_hash_size;
370 return (&kranal_data.kra_peers[hash]);
374 kranal_peer_active(kra_peer_t *peer)
376 /* Am I in the peer hash table? */
377 return (!cfs_list_empty(&peer->rap_list));
381 kranal_conn_addref(kra_conn_t *conn)
383 CDEBUG(D_NET, "%p->%s\n", conn,
384 libcfs_nid2str(conn->rac_peer->rap_nid));
385 LASSERT(cfs_atomic_read(&conn->rac_refcount) > 0);
386 cfs_atomic_inc(&conn->rac_refcount);
390 kranal_conn_decref(kra_conn_t *conn)
392 CDEBUG(D_NET, "%p->%s\n", conn,
393 libcfs_nid2str(conn->rac_peer->rap_nid));
394 LASSERT(cfs_atomic_read(&conn->rac_refcount) > 0);
395 if (cfs_atomic_dec_and_test(&conn->rac_refcount))
396 kranal_destroy_conn(conn);
399 static inline cfs_list_t *
400 kranal_cqid2connlist (__u32 cqid)
402 unsigned int hash = cqid % kranal_data.kra_conn_hash_size;
404 return (&kranal_data.kra_conns [hash]);
407 static inline kra_conn_t *
408 kranal_cqid2conn_locked (__u32 cqid)
410 cfs_list_t *conns = kranal_cqid2connlist(cqid);
414 cfs_list_for_each(tmp, conns) {
415 conn = cfs_list_entry(tmp, kra_conn_t, rac_hashlist);
417 if (conn->rac_cqid == cqid)
425 kranal_tx_mapped (kra_tx_t *tx)
427 return (tx->tx_buftype == RANAL_BUF_VIRT_MAPPED ||
428 tx->tx_buftype == RANAL_BUF_PHYS_MAPPED);
431 int kranal_startup (lnet_ni_t *ni);
432 void kranal_shutdown (lnet_ni_t *ni);
433 int kranal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
434 int kranal_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
435 int kranal_eager_recv(lnet_ni_t *ni, void *private,
436 lnet_msg_t *lntmsg, void **new_private);
437 int kranal_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
438 int delayed, unsigned int niov,
439 struct iovec *iov, lnet_kiov_t *kiov,
440 unsigned int offset, unsigned int mlen, unsigned int rlen);
441 int kranal_accept(lnet_ni_t *ni, struct socket *sock);
443 extern void kranal_free_acceptsock (kra_acceptsock_t *ras);
444 extern int kranal_listener_procint (cfs_sysctl_table_t *table,
445 int write, struct file *filp,
446 void *buffer, size_t *lenp);
447 extern void kranal_update_reaper_timeout (long timeout);
448 extern void kranal_tx_done (kra_tx_t *tx, int completion);
449 extern void kranal_unlink_peer_locked (kra_peer_t *peer);
450 extern void kranal_schedule_conn (kra_conn_t *conn);
451 extern int kranal_create_peer (kra_peer_t **peerp, lnet_nid_t nid);
452 extern int kranal_add_persistent_peer (lnet_nid_t nid, __u32 ip, int port);
453 extern kra_peer_t *kranal_find_peer_locked (lnet_nid_t nid);
454 extern void kranal_post_fma (kra_conn_t *conn, kra_tx_t *tx);
455 extern int kranal_del_peer (lnet_nid_t nid);
456 extern void kranal_device_callback (RAP_INT32 devid, RAP_PVOID arg);
457 extern int kranal_thread_start (int(*fn)(void *arg), void *arg);
458 extern int kranal_connd (void *arg);
459 extern int kranal_reaper (void *arg);
460 extern int kranal_scheduler (void *arg);
461 extern void kranal_close_conn_locked (kra_conn_t *conn, int error);
462 extern void kranal_close_conn (kra_conn_t *conn, int error);
463 extern void kranal_terminate_conn_locked (kra_conn_t *conn);
464 extern void kranal_connect (kra_peer_t *peer);
465 extern int kranal_conn_handshake (struct socket *sock, kra_peer_t *peer);
466 extern int kranal_tunables_init(void);
467 extern void kranal_tunables_fini(void);
468 extern void kranal_init_msg(kra_msg_t *msg, int type);