1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2004 Cluster File Systems, Inc.
5 * Author: Eric Barton <eric@bartonsoftware.com>
7 * This file is part of Lustre, http://www.lustre.org.
9 * Lustre is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Lustre is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Lustre; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 # define EXPORT_SYMTAB
28 #include <linux/config.h>
29 #include <linux/module.h>
30 #include <linux/kernel.h>
32 #include <linux/string.h>
33 #include <linux/stat.h>
34 #include <linux/errno.h>
35 #include <linux/smp_lock.h>
36 #include <linux/unistd.h>
37 #include <linux/uio.h>
39 #include <asm/system.h>
40 #include <asm/uaccess.h>
43 #include <linux/init.h>
45 #include <linux/file.h>
46 #include <linux/stat.h>
47 #include <linux/list.h>
48 #include <linux/kmod.h>
49 #include <linux/sysctl.h>
54 #define DEBUG_SUBSYSTEM S_NAL
56 #include <linux/kp30.h>
57 #include <linux/portals_compat25.h>
58 #include <portals/p30.h>
59 #include <portals/lib-p30.h>
60 #include <portals/nal.h>
64 #define RANAL_MAXDEVS 2 /* max # devices RapidArray supports */
66 #define RANAL_N_CONND 4 /* # connection daemons */
68 #define RANAL_MIN_RECONNECT_INTERVAL 1 /* first failed connection retry (seconds)... */
69 #define RANAL_MAX_RECONNECT_INTERVAL 60 /* ...exponentially increasing to this */
71 #define RANAL_FMA_MAX_PREFIX 232 /* max size of FMA "Prefix" */
72 #define RANAL_FMA_MAX_DATA ((7<<10)-256) /* Max FMA MSG is 7K including prefix */
74 #define RANAL_PEER_HASH_SIZE 101 /* # peer lists */
75 #define RANAL_CONN_HASH_SIZE 101 /* # conn lists */
77 #define RANAL_NTX 64 /* # tx descs */
78 #define RANAL_NTX_NBLK 256 /* # reserved tx descs */
80 #define RANAL_FMA_CQ_SIZE 8192 /* # entries in receive CQ
81 * (overflow is a performance hit) */
83 #define RANAL_RESCHED 100 /* # scheduler loops before reschedule */
85 #define RANAL_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
86 #define RANAL_TIMEOUT2KEEPALIVE(t) (((t)+1)/2) /* timeout -> keepalive interval */
88 /* default vals for runtime tunables */
89 #define RANAL_TIMEOUT 30 /* comms timeout (seconds) */
90 #define RANAL_LISTENER_TIMEOUT 5 /* listener timeout (seconds) */
91 #define RANAL_BACKLOG 127 /* listener's backlog */
92 #define RANAL_PORT 988 /* listener's port */
93 #define RANAL_MAX_IMMEDIATE (2<<10) /* immediate payload breakpoint */
97 int kra_timeout; /* comms timeout (seconds) */
98 int kra_listener_timeout; /* max time the listener can block */
99 int kra_backlog; /* listener's backlog */
100 int kra_port; /* listener's TCP/IP port */
101 int kra_max_immediate; /* immediate payload breakpoint */
103 struct ctl_table_header *kra_sysctl; /* sysctl interface */
108 RAP_PVOID rad_handle; /* device handle */
109 RAP_PVOID rad_fma_cqh; /* FMA completion queue handle */
110 RAP_PVOID rad_rdma_cqh; /* rdma completion queue handle */
111 int rad_id; /* device id */
112 int rad_idx; /* index in kra_devices */
113 int rad_ready; /* set by device callback */
114 struct list_head rad_connq; /* connections requiring attention */
115 struct list_head rad_zombies; /* connections to free */
116 wait_queue_head_t rad_waitq; /* scheduler waits here */
117 spinlock_t rad_lock; /* serialise */
118 void *rad_scheduler; /* scheduling thread */
123 int kra_init; /* initialisation state */
124 int kra_shutdown; /* shut down? */
125 atomic_t kra_nthreads; /* # live threads */
127 struct semaphore kra_nid_mutex; /* serialise NID/listener ops */
128 struct semaphore kra_listener_signal; /* block for listener startup/shutdown */
129 struct socket *kra_listener_sock; /* listener's socket */
130 int kra_listener_shutdown; /* ask listener to close */
132 kra_device_t kra_devices[RANAL_MAXDEVS]; /* device/ptag/cq etc */
133 int kra_ndevs; /* # devices */
135 rwlock_t kra_global_lock; /* stabilize peer/conn ops */
137 struct list_head *kra_peers; /* hash table of all my known peers */
138 int kra_peer_hash_size; /* size of kra_peers */
139 atomic_t kra_npeers; /* # peers extant */
141 struct list_head *kra_conns; /* conns hashed by cqid */
142 int kra_conn_hash_size; /* size of kra_conns */
143 __u64 kra_peerstamp; /* when I started up */
144 __u64 kra_connstamp; /* conn stamp generator */
145 int kra_next_cqid; /* cqid generator */
146 atomic_t kra_nconns; /* # connections extant */
148 long kra_new_min_timeout; /* minimum timeout on any new conn */
149 wait_queue_head_t kra_reaper_waitq; /* reaper sleeps here */
150 spinlock_t kra_reaper_lock; /* serialise */
152 struct list_head kra_connd_peers; /* peers waiting for a connection */
153 struct list_head kra_connd_acceptq; /* accepted sockets to handshake */
154 wait_queue_head_t kra_connd_waitq; /* connection daemons sleep here */
155 spinlock_t kra_connd_lock; /* serialise */
157 struct list_head kra_idle_txs; /* idle tx descriptors */
158 struct list_head kra_idle_nblk_txs; /* idle reserved tx descriptors */
159 __u64 kra_next_tx_cookie; /* RDMA completion cookie */
160 wait_queue_head_t kra_idle_tx_waitq; /* block here for tx descriptor */
161 spinlock_t kra_tx_lock; /* serialise */
164 #define RANAL_INIT_NOTHING 0
165 #define RANAL_INIT_DATA 1
166 #define RANAL_INIT_LIB 2
167 #define RANAL_INIT_ALL 3
169 typedef struct kra_acceptsock /* accepted socket queued for connd */
171 struct list_head ras_list; /* queue for attention */
172 struct socket *ras_sock; /* the accepted socket */
175 /************************************************************************
176 * Wire message structs. These are sent in sender's byte order
177 * (i.e. receiver checks magic and flips if required).
180 typedef struct kra_connreq /* connection request/response */
181 { /* (sent via socket) */
182 __u32 racr_magic; /* I'm an ranal connreq */
183 __u16 racr_version; /* this is my version number */
184 __u16 racr_devid; /* sender's device ID */
185 __u64 racr_srcnid; /* sender's NID */
186 __u64 racr_dstnid; /* who sender expects to listen */
187 __u64 racr_peerstamp; /* sender's instance stamp */
188 __u64 racr_connstamp; /* sender's connection stamp */
189 __u32 racr_timeout; /* sender's timeout */
190 RAP_RI_PARAMETERS racr_riparams; /* sender's endpoint info */
195 RAP_MEM_KEY rard_key;
196 RAP_PVOID64 rard_addr;
202 ptl_hdr_t raim_hdr; /* portals header */
203 /* Portals payload is in FMA "Message Data" */
204 } kra_immediate_msg_t;
208 ptl_hdr_t raprm_hdr; /* portals header */
209 __u64 raprm_cookie; /* opaque completion cookie */
214 __u64 rapam_src_cookie; /* reflected completion cookie */
215 __u64 rapam_dst_cookie; /* opaque completion cookie */
216 kra_rdma_desc_t rapam_desc; /* sender's sink buffer */
221 ptl_hdr_t ragm_hdr; /* portals header */
222 __u64 ragm_cookie; /* opaque completion cookie */
223 kra_rdma_desc_t ragm_desc; /* sender's sink buffer */
228 __u64 racm_cookie; /* reflected completion cookie */
229 } kra_completion_msg_t;
231 typedef struct /* NB must fit in FMA "Prefix" */
233 __u32 ram_magic; /* I'm an ranal message */
234 __u16 ram_version; /* this is my version number */
235 __u16 ram_type; /* msg type */
236 __u64 ram_srcnid; /* sender's NID */
237 __u64 ram_connstamp; /* sender's connection stamp */
239 kra_immediate_msg_t immediate;
240 kra_putreq_msg_t putreq;
241 kra_putack_msg_t putack;
243 kra_completion_msg_t completion;
245 __u32 ram_seq; /* incrementing sequence number */
248 #define RANAL_MSG_MAGIC 0x0be91b92 /* unique magic */
249 #define RANAL_MSG_VERSION 1 /* current protocol version */
251 #define RANAL_MSG_FENCE 0x80 /* fence RDMA */
253 #define RANAL_MSG_NONE 0x00 /* illegal message */
254 #define RANAL_MSG_NOOP 0x01 /* empty ram_u (keepalive) */
255 #define RANAL_MSG_IMMEDIATE 0x02 /* ram_u.immediate */
256 #define RANAL_MSG_PUT_REQ 0x03 /* ram_u.putreq (src->sink) */
257 #define RANAL_MSG_PUT_NAK 0x04 /* ram_u.completion (no PUT match: sink->src) */
258 #define RANAL_MSG_PUT_ACK 0x05 /* ram_u.putack (PUT matched: sink->src) */
259 #define RANAL_MSG_PUT_DONE 0x86 /* ram_u.completion (src->sink) */
260 #define RANAL_MSG_GET_REQ 0x07 /* ram_u.get (sink->src) */
261 #define RANAL_MSG_GET_NAK 0x08 /* ram_u.completion (no GET match: src->sink) */
262 #define RANAL_MSG_GET_DONE 0x89 /* ram_u.completion (src->sink) */
263 #define RANAL_MSG_CLOSE 0x8a /* empty ram_u */
265 /***********************************************************************/
267 typedef struct kra_tx /* message descriptor */
269 struct list_head tx_list; /* queue on idle_txs/rac_sendq/rac_waitq */
270 struct kra_conn *tx_conn; /* owning conn */
271 lib_msg_t *tx_libmsg[2]; /* lib msgs to finalize on completion */
272 unsigned long tx_qtime; /* when tx started to wait for something */
273 int tx_isnblk; /* I'm reserved for non-blocking sends */
274 int tx_nob; /* # bytes of payload */
275 int tx_buftype; /* payload buffer type */
276 void *tx_buffer; /* source/sink buffer */
277 int tx_phys_offset; /* first page offset (if phys) */
278 int tx_phys_npages; /* # physical pages */
279 RAP_PHYS_REGION *tx_phys; /* page descriptors */
280 RAP_MEM_KEY tx_map_key; /* mapping key */
281 RAP_RDMA_DESCRIPTOR tx_rdma_desc; /* rdma descriptor */
282 __u64 tx_cookie; /* identify this tx to peer */
283 kra_msg_t tx_msg; /* FMA message buffer */
286 #define RANAL_BUF_NONE 0 /* buffer type not set */
287 #define RANAL_BUF_IMMEDIATE 1 /* immediate data */
288 #define RANAL_BUF_PHYS_UNMAPPED 2 /* physical: not mapped yet */
289 #define RANAL_BUF_PHYS_MAPPED 3 /* physical: mapped already */
290 #define RANAL_BUF_VIRT_UNMAPPED 4 /* virtual: not mapped yet */
291 #define RANAL_BUF_VIRT_MAPPED 5 /* virtual: mapped already */
293 #define RANAL_TX_IDLE 0x00 /* on freelist */
294 #define RANAL_TX_SIMPLE 0x10 /* about to send a simple message */
295 #define RANAL_TX_PUTI_REQ 0x20 /* PUT initiator about to send PUT_REQ */
296 #define RANAL_TX_PUTI_WAIT_ACK 0x21 /* PUT initiator waiting for PUT_ACK */
297 #define RANAL_TX_PUTI_RDMA 0x22 /* PUT initiator waiting for RDMA to complete */
298 #define RANAL_TX_PUTI_DONE 0x23 /* PUT initiator about to send PUT_DONE */
299 #define RANAL_TX_PUTT_NAK 0x30 /* PUT target about to send PUT_NAK */
300 #define RANAL_TX_PUTT_ACK 0x30 /* PUT target about to send PUT_ACK */
301 #define RANAL_TX_PUTT_WAIT_DONE 0x31 /* PUT target waiting for PUT_DONE */
302 #define RANAL_TX_GETI_REQ 0x40 /* GET initiator about to send GET_REQ */
303 #define RANAL_TX_GETI_WAIT_DONE 0x41 /* GET initiator waiting for GET_DONE */
304 #define RANAL_TX_GETT_NAK 0x50 /* GET target about to send PUT_NAK */
305 #define RANAL_TX_GETT_RDMA 0x51 /* GET target waiting for RDMA to complete */
306 #define RANAL_TX_GETT_DONE 0x52 /* GET target about to send GET_DONE */
308 typedef struct kra_conn
310 struct kra_peer *rac_peer; /* owning peer */
311 struct list_head rac_list; /* stash on peer's conn list */
312 struct list_head rac_hashlist; /* stash in connection hash table */
313 struct list_head rac_schedlist; /* schedule (on rad_connq) for attention */
314 struct list_head rac_fmaq; /* txs queued for FMA */
315 struct list_head rac_rdmaq; /* txs awaiting RDMA completion */
316 struct list_head rac_replyq; /* txs awaiting replies */
317 __u64 rac_peerstamp; /* peer's unique stamp */
318 __u64 rac_peer_connstamp; /* peer's unique connection stamp */
319 __u64 rac_my_connstamp; /* my unique connection stamp */
320 unsigned long rac_last_tx; /* when I last sent an FMA message */
321 unsigned long rac_last_rx; /* when I last received an FMA messages */
322 long rac_keepalive; /* keepalive interval */
323 long rac_timeout; /* infer peer death on (last_rx + timout > now) */
324 __u32 rac_cqid; /* my completion callback id (non-unique) */
325 __u32 rac_tx_seq; /* tx msg sequence number */
326 __u32 rac_rx_seq; /* rx msg sequence number */
327 atomic_t rac_refcount; /* # users */
328 unsigned int rac_close_sent; /* I've sent CLOSE */
329 unsigned int rac_close_recvd; /* I've received CLOSE */
330 unsigned int rac_state; /* connection state */
331 unsigned int rac_scheduled; /* being attented to */
332 spinlock_t rac_lock; /* serialise */
333 kra_device_t *rac_device; /* which device */
334 RAP_PVOID rac_rihandle; /* RA endpoint */
335 kra_msg_t *rac_rxmsg; /* incoming message (FMA prefix) */
336 kra_msg_t rac_msg; /* keepalive/CLOSE message buffer */
339 #define RANAL_CONN_ESTABLISHED 0
340 #define RANAL_CONN_CLOSING 1
341 #define RANAL_CONN_CLOSED 2
343 typedef struct kra_peer
345 struct list_head rap_list; /* stash on global peer list */
346 struct list_head rap_connd_list; /* schedule on kra_connd_peers */
347 struct list_head rap_conns; /* all active connections */
348 struct list_head rap_tx_queue; /* msgs waiting for a conn */
349 ptl_nid_t rap_nid; /* who's on the other end(s) */
350 __u32 rap_ip; /* IP address of peer */
351 int rap_port; /* port on which peer listens */
352 atomic_t rap_refcount; /* # users */
353 int rap_persistence; /* "known" peer refs */
354 int rap_connecting; /* connection forming */
355 unsigned long rap_reconnect_time; /* CURRENT_SECONDS when reconnect OK */
356 unsigned long rap_reconnect_interval; /* exponential backoff */
359 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,0))
360 # define sk_allocation allocation
361 # define sk_data_ready data_ready
362 # define sk_write_space write_space
363 # define sk_user_data user_data
364 # define sk_prot prot
365 # define sk_sndbuf sndbuf
366 # define sk_socket socket
367 # define sk_wmem_queued wmem_queued
369 # define sk_sleep sleep
372 extern lib_nal_t kranal_lib;
373 extern kra_data_t kranal_data;
374 extern kra_tunables_t kranal_tunables;
376 extern void kranal_destroy_peer(kra_peer_t *peer);
377 extern void kranal_destroy_conn(kra_conn_t *conn);
380 kranal_peer_addref(kra_peer_t *peer)
382 CDEBUG(D_NET, "%p->"LPX64"\n", peer, peer->rap_nid);
383 LASSERT(atomic_read(&peer->rap_refcount) > 0);
384 atomic_inc(&peer->rap_refcount);
388 kranal_peer_decref(kra_peer_t *peer)
390 CDEBUG(D_NET, "%p->"LPX64"\n", peer, peer->rap_nid);
391 LASSERT(atomic_read(&peer->rap_refcount) > 0);
392 if (atomic_dec_and_test(&peer->rap_refcount))
393 kranal_destroy_peer(peer);
396 static inline struct list_head *
397 kranal_nid2peerlist (ptl_nid_t nid)
399 unsigned int hash = ((unsigned int)nid) % kranal_data.kra_peer_hash_size;
401 return (&kranal_data.kra_peers[hash]);
405 kranal_peer_active(kra_peer_t *peer)
407 /* Am I in the peer hash table? */
408 return (!list_empty(&peer->rap_list));
412 kranal_conn_addref(kra_conn_t *conn)
414 CDEBUG(D_NET, "%p->"LPX64"\n", conn, conn->rac_peer->rap_nid);
415 LASSERT(atomic_read(&conn->rac_refcount) > 0);
416 atomic_inc(&conn->rac_refcount);
420 kranal_conn_decref(kra_conn_t *conn)
422 CDEBUG(D_NET, "%p->"LPX64"\n", conn, conn->rac_peer->rap_nid);
423 LASSERT(atomic_read(&conn->rac_refcount) > 0);
424 if (atomic_dec_and_test(&conn->rac_refcount))
425 kranal_destroy_conn(conn);
428 static inline struct list_head *
429 kranal_cqid2connlist (__u32 cqid)
431 unsigned int hash = cqid % kranal_data.kra_conn_hash_size;
433 return (&kranal_data.kra_conns [hash]);
436 static inline kra_conn_t *
437 kranal_cqid2conn_locked (__u32 cqid)
439 struct list_head *conns = kranal_cqid2connlist(cqid);
440 struct list_head *tmp;
443 list_for_each(tmp, conns) {
444 conn = list_entry(tmp, kra_conn_t, rac_hashlist);
446 if (conn->rac_cqid == cqid)
454 kranal_tx_mapped (kra_tx_t *tx)
456 return (tx->tx_buftype == RANAL_BUF_VIRT_MAPPED ||
457 tx->tx_buftype == RANAL_BUF_PHYS_MAPPED);
461 kranal_page2phys (struct page *p)
463 return page_to_phys(p);
466 extern void kranal_free_acceptsock (kra_acceptsock_t *ras);
467 extern int kranal_listener_procint (ctl_table *table,
468 int write, struct file *filp,
469 void *buffer, size_t *lenp);
470 extern void kranal_update_reaper_timeout (long timeout);
471 extern void kranal_tx_done (kra_tx_t *tx, int completion);
472 extern void kranal_unlink_peer_locked (kra_peer_t *peer);
473 extern void kranal_schedule_conn (kra_conn_t *conn);
474 extern kra_peer_t *kranal_create_peer (ptl_nid_t nid);
475 extern kra_peer_t *kranal_find_peer_locked (ptl_nid_t nid);
476 extern void kranal_post_fma (kra_conn_t *conn, kra_tx_t *tx);
477 extern int kranal_del_peer (ptl_nid_t nid, int single_share);
478 extern void kranal_device_callback (RAP_INT32 devid, RAP_PVOID arg);
479 extern int kranal_thread_start (int(*fn)(void *arg), void *arg);
480 extern int kranal_connd (void *arg);
481 extern int kranal_reaper (void *arg);
482 extern int kranal_scheduler (void *arg);
483 extern void kranal_close_conn_locked (kra_conn_t *conn, int error);
484 extern void kranal_terminate_conn_locked (kra_conn_t *conn);
485 extern void kranal_connect (kra_peer_t *peer);
486 extern int kranal_conn_handshake (struct socket *sock, kra_peer_t *peer);