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 [sun.com URL with a
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/ralnd/ralnd.h
38 * Author: Eric Barton <eric@bartonsoftware.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>
72 #define DEBUG_SUBSYSTEM S_LND
74 #include <libcfs/libcfs.h>
75 #include <lnet/lnet.h>
76 #include <lnet/lib-lnet.h>
80 /* tunables determined at compile time */
81 #define RANAL_RESCHED 100 /* # scheduler loops before reschedule */
83 #define RANAL_PEER_HASH_SIZE 101 /* # peer lists */
84 #define RANAL_CONN_HASH_SIZE 101 /* # conn lists */
86 #define RANAL_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
87 #define RANAL_TIMEOUT2KEEPALIVE(t) (((t)+1)/2) /* timeout -> keepalive interval */
90 #define RANAL_MAXDEVS 2 /* max # devices RapidArray supports */
91 #define RANAL_FMA_MAX_PREFIX 232 /* max bytes in FMA "Prefix" we can use */
92 #define RANAL_FMA_MAX_DATA ((7<<10)-256) /* Max FMA MSG is 7K including prefix */
97 int *kra_n_connd; /* # connection daemons */
98 int *kra_min_reconnect_interval; /* first failed connection retry... */
99 int *kra_max_reconnect_interval; /* ...exponentially increasing to this */
100 int *kra_ntx; /* # tx descs */
101 int *kra_credits; /* # concurrent sends */
102 int *kra_peercredits; /* # concurrent sends to 1 peer */
103 int *kra_fma_cq_size; /* # entries in receive CQ */
104 int *kra_timeout; /* comms timeout (seconds) */
105 int *kra_max_immediate; /* immediate payload breakpoint */
107 #if CONFIG_SYSCTL && !CFS_SYSFS_MODULE_PARM
108 cfs_sysctl_table_header_t *kra_sysctl; /* sysctl interface */
114 RAP_PVOID rad_handle; /* device handle */
115 RAP_PVOID rad_fma_cqh; /* FMA completion queue handle */
116 RAP_PVOID rad_rdma_cqh; /* rdma completion queue handle */
117 int rad_id; /* device id */
118 int rad_idx; /* index in kra_devices */
119 int rad_ready; /* set by device callback */
120 struct list_head rad_ready_conns;/* connections ready to tx/rx */
121 struct list_head rad_new_conns; /* new connections to complete */
122 wait_queue_head_t rad_waitq; /* scheduler waits here */
123 spinlock_t rad_lock; /* serialise */
124 void *rad_scheduler; /* scheduling thread */
125 unsigned int rad_nphysmap; /* # phys mappings */
126 unsigned int rad_nppphysmap; /* # phys pages mapped */
127 unsigned int rad_nvirtmap; /* # virt mappings */
128 unsigned long rad_nobvirtmap; /* # virt bytes mapped */
133 int kra_init; /* initialisation state */
134 int kra_shutdown; /* shut down? */
135 atomic_t kra_nthreads; /* # live threads */
136 lnet_ni_t *kra_ni; /* _the_ nal instance */
138 kra_device_t kra_devices[RANAL_MAXDEVS]; /* device/ptag/cq etc */
139 int kra_ndevs; /* # devices */
141 rwlock_t kra_global_lock; /* stabilize peer/conn ops */
143 struct list_head *kra_peers; /* hash table of all my known peers */
144 int kra_peer_hash_size; /* size of kra_peers */
145 atomic_t kra_npeers; /* # peers extant */
146 int kra_nonewpeers; /* prevent new peers */
148 struct list_head *kra_conns; /* conns hashed by cqid */
149 int kra_conn_hash_size; /* size of kra_conns */
150 __u64 kra_peerstamp; /* when I started up */
151 __u64 kra_connstamp; /* conn stamp generator */
152 int kra_next_cqid; /* cqid generator */
153 atomic_t kra_nconns; /* # connections extant */
155 long kra_new_min_timeout; /* minimum timeout on any new conn */
156 wait_queue_head_t kra_reaper_waitq; /* reaper sleeps here */
157 spinlock_t kra_reaper_lock; /* serialise */
159 struct list_head kra_connd_peers; /* peers waiting for a connection */
160 struct list_head kra_connd_acceptq; /* accepted sockets to handshake */
161 wait_queue_head_t kra_connd_waitq; /* connection daemons sleep here */
162 spinlock_t kra_connd_lock; /* serialise */
164 struct list_head kra_idle_txs; /* idle tx descriptors */
165 __u64 kra_next_tx_cookie; /* RDMA completion cookie */
166 spinlock_t kra_tx_lock; /* serialise */
169 #define RANAL_INIT_NOTHING 0
170 #define RANAL_INIT_DATA 1
171 #define RANAL_INIT_ALL 2
173 typedef struct kra_acceptsock /* accepted socket queued for connd */
175 struct list_head ras_list; /* queue for attention */
176 struct socket *ras_sock; /* the accepted socket */
179 /************************************************************************
180 * Wire message structs. These are sent in sender's byte order
181 * (i.e. receiver checks magic and flips if required).
184 typedef struct kra_connreq /* connection request/response */
185 { /* (sent via socket) */
186 __u32 racr_magic; /* I'm an ranal connreq */
187 __u16 racr_version; /* this is my version number */
188 __u16 racr_devid; /* sender's device ID */
189 __u64 racr_srcnid; /* sender's NID */
190 __u64 racr_dstnid; /* who sender expects to listen */
191 __u64 racr_peerstamp; /* sender's instance stamp */
192 __u64 racr_connstamp; /* sender's connection stamp */
193 __u32 racr_timeout; /* sender's timeout */
194 RAP_RI_PARAMETERS racr_riparams; /* sender's endpoint info */
199 RAP_MEM_KEY rard_key;
200 RAP_PVOID64 rard_addr;
206 lnet_hdr_t raim_hdr; /* portals header */
207 /* Portals payload is in FMA "Message Data" */
208 } kra_immediate_msg_t;
212 lnet_hdr_t raprm_hdr; /* portals header */
213 __u64 raprm_cookie; /* opaque completion cookie */
218 __u64 rapam_src_cookie; /* reflected completion cookie */
219 __u64 rapam_dst_cookie; /* opaque completion cookie */
220 kra_rdma_desc_t rapam_desc; /* sender's sink buffer */
225 lnet_hdr_t ragm_hdr; /* portals header */
226 __u64 ragm_cookie; /* opaque completion cookie */
227 kra_rdma_desc_t ragm_desc; /* sender's sink buffer */
232 __u64 racm_cookie; /* reflected completion cookie */
233 } kra_completion_msg_t;
235 typedef struct /* NB must fit in FMA "Prefix" */
237 __u32 ram_magic; /* I'm an ranal message */
238 __u16 ram_version; /* this is my version number */
239 __u16 ram_type; /* msg type */
240 __u64 ram_srcnid; /* sender's NID */
241 __u64 ram_connstamp; /* sender's connection stamp */
243 kra_immediate_msg_t immediate;
244 kra_putreq_msg_t putreq;
245 kra_putack_msg_t putack;
247 kra_completion_msg_t completion;
249 __u32 ram_seq; /* incrementing sequence number */
252 #define RANAL_MSG_MAGIC LNET_PROTO_RA_MAGIC /* unique magic */
253 #define RANAL_MSG_VERSION 1 /* current protocol version */
255 #define RANAL_MSG_FENCE 0x80 /* fence RDMA */
257 #define RANAL_MSG_NONE 0x00 /* illegal message */
258 #define RANAL_MSG_NOOP 0x01 /* empty ram_u (keepalive) */
259 #define RANAL_MSG_IMMEDIATE 0x02 /* ram_u.immediate */
260 #define RANAL_MSG_PUT_REQ 0x03 /* ram_u.putreq (src->sink) */
261 #define RANAL_MSG_PUT_NAK 0x04 /* ram_u.completion (no PUT match: sink->src) */
262 #define RANAL_MSG_PUT_ACK 0x05 /* ram_u.putack (PUT matched: sink->src) */
263 #define RANAL_MSG_PUT_DONE 0x86 /* ram_u.completion (src->sink) */
264 #define RANAL_MSG_GET_REQ 0x07 /* ram_u.get (sink->src) */
265 #define RANAL_MSG_GET_NAK 0x08 /* ram_u.completion (no GET match: src->sink) */
266 #define RANAL_MSG_GET_DONE 0x89 /* ram_u.completion (src->sink) */
267 #define RANAL_MSG_CLOSE 0x8a /* empty ram_u */
269 /***********************************************************************/
271 typedef struct kra_tx /* message descriptor */
273 struct list_head tx_list; /* queue on idle_txs/rac_sendq/rac_waitq */
274 struct kra_conn *tx_conn; /* owning conn */
275 lnet_msg_t *tx_lntmsg[2]; /* ptl msgs to finalize on completion */
276 unsigned long tx_qtime; /* when tx started to wait for something (jiffies) */
277 int tx_nob; /* # bytes of payload */
278 int tx_buftype; /* payload buffer type */
279 void *tx_buffer; /* source/sink buffer */
280 int tx_phys_offset; /* first page offset (if phys) */
281 int tx_phys_npages; /* # physical pages */
282 RAP_PHYS_REGION *tx_phys; /* page descriptors */
283 RAP_MEM_KEY tx_map_key; /* mapping key */
284 RAP_RDMA_DESCRIPTOR tx_rdma_desc; /* rdma descriptor */
285 __u64 tx_cookie; /* identify this tx to peer */
286 kra_msg_t tx_msg; /* FMA message buffer */
289 #define RANAL_BUF_NONE 0 /* buffer type not set */
290 #define RANAL_BUF_IMMEDIATE 1 /* immediate data */
291 #define RANAL_BUF_PHYS_UNMAPPED 2 /* physical: not mapped yet */
292 #define RANAL_BUF_PHYS_MAPPED 3 /* physical: mapped already */
293 #define RANAL_BUF_VIRT_UNMAPPED 4 /* virtual: not mapped yet */
294 #define RANAL_BUF_VIRT_MAPPED 5 /* virtual: mapped already */
296 typedef struct kra_conn
298 struct kra_peer *rac_peer; /* owning peer */
299 struct list_head rac_list; /* stash on peer's conn list */
300 struct list_head rac_hashlist; /* stash in connection hash table */
301 struct list_head rac_schedlist; /* schedule (on rad_???_conns) for attention */
302 struct list_head rac_fmaq; /* txs queued for FMA */
303 struct list_head rac_rdmaq; /* txs awaiting RDMA completion */
304 struct list_head rac_replyq; /* txs awaiting replies */
305 __u64 rac_peerstamp; /* peer's unique stamp */
306 __u64 rac_peer_connstamp; /* peer's unique connection stamp */
307 __u64 rac_my_connstamp; /* my unique connection stamp */
308 unsigned long rac_last_tx; /* when I last sent an FMA message (jiffies) */
309 unsigned long rac_last_rx; /* when I last received an FMA messages (jiffies) */
310 long rac_keepalive; /* keepalive interval (seconds) */
311 long rac_timeout; /* infer peer death if no rx for this many seconds */
312 __u32 rac_cqid; /* my completion callback id (non-unique) */
313 __u32 rac_tx_seq; /* tx msg sequence number */
314 __u32 rac_rx_seq; /* rx msg sequence number */
315 atomic_t rac_refcount; /* # users */
316 unsigned int rac_close_sent; /* I've sent CLOSE */
317 unsigned int rac_close_recvd; /* I've received CLOSE */
318 unsigned int rac_state; /* connection state */
319 unsigned int rac_scheduled; /* being attented to */
320 spinlock_t rac_lock; /* serialise */
321 kra_device_t *rac_device; /* which device */
322 RAP_PVOID rac_rihandle; /* RA endpoint */
323 kra_msg_t *rac_rxmsg; /* incoming message (FMA prefix) */
324 kra_msg_t rac_msg; /* keepalive/CLOSE message buffer */
327 #define RANAL_CONN_ESTABLISHED 0
328 #define RANAL_CONN_CLOSING 1
329 #define RANAL_CONN_CLOSED 2
331 typedef struct kra_peer
333 struct list_head rap_list; /* stash on global peer list */
334 struct list_head rap_connd_list; /* schedule on kra_connd_peers */
335 struct list_head rap_conns; /* all active connections */
336 struct list_head rap_tx_queue; /* msgs waiting for a conn */
337 lnet_nid_t rap_nid; /* who's on the other end(s) */
338 __u32 rap_ip; /* IP address of peer */
339 int rap_port; /* port on which peer listens */
340 atomic_t rap_refcount; /* # users */
341 int rap_persistence; /* "known" peer refs */
342 int rap_connecting; /* connection forming */
343 unsigned long rap_reconnect_time; /* CURRENT_SECONDS when reconnect OK */
344 unsigned long rap_reconnect_interval; /* exponential backoff */
347 extern kra_data_t kranal_data;
348 extern kra_tunables_t kranal_tunables;
350 extern void kranal_destroy_peer(kra_peer_t *peer);
351 extern void kranal_destroy_conn(kra_conn_t *conn);
354 kranal_peer_addref(kra_peer_t *peer)
356 CDEBUG(D_NET, "%p->%s\n", peer, libcfs_nid2str(peer->rap_nid));
357 LASSERT(atomic_read(&peer->rap_refcount) > 0);
358 atomic_inc(&peer->rap_refcount);
362 kranal_peer_decref(kra_peer_t *peer)
364 CDEBUG(D_NET, "%p->%s\n", peer, libcfs_nid2str(peer->rap_nid));
365 LASSERT(atomic_read(&peer->rap_refcount) > 0);
366 if (atomic_dec_and_test(&peer->rap_refcount))
367 kranal_destroy_peer(peer);
370 static inline struct list_head *
371 kranal_nid2peerlist (lnet_nid_t nid)
373 unsigned int hash = ((unsigned int)nid) % kranal_data.kra_peer_hash_size;
375 return (&kranal_data.kra_peers[hash]);
379 kranal_peer_active(kra_peer_t *peer)
381 /* Am I in the peer hash table? */
382 return (!list_empty(&peer->rap_list));
386 kranal_conn_addref(kra_conn_t *conn)
388 CDEBUG(D_NET, "%p->%s\n", conn,
389 libcfs_nid2str(conn->rac_peer->rap_nid));
390 LASSERT(atomic_read(&conn->rac_refcount) > 0);
391 atomic_inc(&conn->rac_refcount);
395 kranal_conn_decref(kra_conn_t *conn)
397 CDEBUG(D_NET, "%p->%s\n", conn,
398 libcfs_nid2str(conn->rac_peer->rap_nid));
399 LASSERT(atomic_read(&conn->rac_refcount) > 0);
400 if (atomic_dec_and_test(&conn->rac_refcount))
401 kranal_destroy_conn(conn);
404 static inline struct list_head *
405 kranal_cqid2connlist (__u32 cqid)
407 unsigned int hash = cqid % kranal_data.kra_conn_hash_size;
409 return (&kranal_data.kra_conns [hash]);
412 static inline kra_conn_t *
413 kranal_cqid2conn_locked (__u32 cqid)
415 struct list_head *conns = kranal_cqid2connlist(cqid);
416 struct list_head *tmp;
419 list_for_each(tmp, conns) {
420 conn = list_entry(tmp, kra_conn_t, rac_hashlist);
422 if (conn->rac_cqid == cqid)
430 kranal_tx_mapped (kra_tx_t *tx)
432 return (tx->tx_buftype == RANAL_BUF_VIRT_MAPPED ||
433 tx->tx_buftype == RANAL_BUF_PHYS_MAPPED);
436 int kranal_startup (lnet_ni_t *ni);
437 void kranal_shutdown (lnet_ni_t *ni);
438 int kranal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
439 int kranal_send (lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
440 int kranal_eager_recv(lnet_ni_t *ni, void *private,
441 lnet_msg_t *lntmsg, void **new_private);
442 int kranal_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg,
443 int delayed, unsigned int niov,
444 struct iovec *iov, lnet_kiov_t *kiov,
445 unsigned int offset, unsigned int mlen, unsigned int rlen);
446 int kranal_accept(lnet_ni_t *ni, struct socket *sock);
448 extern void kranal_free_acceptsock (kra_acceptsock_t *ras);
449 extern int kranal_listener_procint (cfs_sysctl_table_t *table,
450 int write, struct file *filp,
451 void *buffer, size_t *lenp);
452 extern void kranal_update_reaper_timeout (long timeout);
453 extern void kranal_tx_done (kra_tx_t *tx, int completion);
454 extern void kranal_unlink_peer_locked (kra_peer_t *peer);
455 extern void kranal_schedule_conn (kra_conn_t *conn);
456 extern int kranal_create_peer (kra_peer_t **peerp, lnet_nid_t nid);
457 extern int kranal_add_persistent_peer (lnet_nid_t nid, __u32 ip, int port);
458 extern kra_peer_t *kranal_find_peer_locked (lnet_nid_t nid);
459 extern void kranal_post_fma (kra_conn_t *conn, kra_tx_t *tx);
460 extern int kranal_del_peer (lnet_nid_t nid);
461 extern void kranal_device_callback (RAP_INT32 devid, RAP_PVOID arg);
462 extern int kranal_thread_start (int(*fn)(void *arg), void *arg);
463 extern int kranal_connd (void *arg);
464 extern int kranal_reaper (void *arg);
465 extern int kranal_scheduler (void *arg);
466 extern void kranal_close_conn_locked (kra_conn_t *conn, int error);
467 extern void kranal_close_conn (kra_conn_t *conn, int error);
468 extern void kranal_terminate_conn_locked (kra_conn_t *conn);
469 extern void kranal_connect (kra_peer_t *peer);
470 extern int kranal_conn_handshake (struct socket *sock, kra_peer_t *peer);
471 extern int kranal_tunables_init(void);
472 extern void kranal_tunables_fini(void);
473 extern void kranal_init_msg(kra_msg_t *msg, int type);