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>
51 #define DEBUG_SUBSYSTEM S_NAL
53 #include <linux/kp30.h>
54 #include <portals/p30.h>
55 #include <portals/lib-p30.h>
56 #include <portals/nal.h>
61 # define RANAL_N_SCHED num_online_cpus() /* # schedulers */
63 # define RANAL_N_SCHED 1 /* # schedulers */
66 #define RANAL_MAXDEVS 2 /* max # devices RapidArray supports */
68 #define RANAL_N_CONND 4 /* # connection daemons */
70 #define RANAL_MIN_RECONNECT_INTERVAL 1 /* first failed connection retry (seconds)... */
71 #define RANAL_MAX_RECONNECT_INTERVAL 60 /* ...exponentially increasing to this */
73 #define RANAL_FMA_PREFIX_LEN 232 /* size of FMA "Prefix" */
74 #define RANAL_FMA_MAX_DATA_LEN ((7<<10)-256) /* Max FMA MSG is 7K including prefix */
76 #define RANAL_PEER_HASH_SIZE 101 /* # peer lists */
77 #define RANAL_CONN_HASH_SIZE 101 /* # conn lists */
79 #define RANAL_NTX 64 /* # tx descs */
80 #define RANAL_NTX_NBLK 256 /* # reserved tx descs */
82 #define RANAL_RX_CQ_SIZE 1024 /* # entries in receive CQ
83 * (overflow is a performance hit) */
85 #define RANAL_RESCHED 100 /* # scheduler loops before reschedule */
87 #define RANAL_MIN_TIMEOUT 5 /* minimum timeout interval (seconds) */
88 #define RANAL_TIMEOUT2KEEPALIVE(t) (((t)+1)/2) /* timeout -> keepalive interval */
90 /* default vals for runtime tunables */
91 #define RANAL_TIMEOUT 30 /* comms timeout (seconds) */
92 #define RANAL_LISTENER_TIMEOUT 5 /* listener timeout (seconds) */
93 #define RANAL_MAX_IMMEDIATE (2<<10) /* biggest immediate payload */
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; /* biggest immediate payload */
102 struct ctl_table_header *kra_sysctl; /* sysctl interface */
107 RAP_PVOID rad_handle; /* device handle */
108 RAP_PROTECTION_HANDLE rad_ptag; /* protection tag */
109 RAP_CQ_HANDLE rad_fma_cq; /* FMA (small message) completion queue */
110 RAP_CQ_HANDLE rad_rdma_cq; /* rdma completion queue */
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 wait_queue_head_t rad_waitq; /* scheduler waits here */
116 spinlock_t rad_lock; /* serialise */
121 int kra_init; /* initialisation state */
122 int kra_shutdown; /* shut down? */
123 atomic_t kra_nthreads; /* # live threads */
125 struct semaphore kra_nid_mutex; /* serialise NID/listener ops */
126 struct semaphore kra_listener_signal; /* block for listener startup/shutdown */
127 struct socket *kra_listener_sock; /* listener's socket */
128 int kra_listener_shutdown; /* ask listener to close */
130 kra_device_t kra_devices[RANAL_MAXDEVS]; /* device/ptag/cq etc */
131 int kra_ndevs; /* # devices */
133 rwlock_t kra_global_lock; /* stabilize peer/conn ops */
135 struct list_head *kra_peers; /* hash table of all my known peers */
136 int kra_peer_hash_size; /* size of kra_peers */
137 atomic_t kra_npeers; /* # peers extant */
139 struct list_head *kra_conns; /* conns hashed by cqid */
140 int kra_conn_hash_size; /* size of kra_conns */
141 __u64 kra_next_incarnation; /* conn incarnation # generator */
142 int kra_next_cqid; /* cqid generator */
143 atomic_t kra_nconns; /* # connections extant */
145 long kra_new_min_timeout; /* minimum timeout on any new conn */
146 wait_queue_head_t kra_reaper_waitq; /* reaper sleeps here */
147 spinlock_t kra_reaper_lock; /* serialise */
149 struct list_head kra_connd_peers; /* peers waiting for a connection */
150 wait_queue_head_t kra_connd_waitq; /* connection daemons sleep here */
151 spinlock_t kra_connd_lock; /* serialise */
153 struct list_head kra_idle_txs; /* idle tx descriptors */
154 struct list_head kra_idle_nblk_txs; /* idle reserved tx descriptors */
155 __u64 kra_next_tx_cookie; /* RDMA completion cookie */
156 wait_queue_head_t kra_idle_tx_waitq; /* block here for tx descriptor */
157 spinlock_t kra_tx_lock; /* serialise */
160 #define RANAL_INIT_NOTHING 0
161 #define RANAL_INIT_DATA 1
163 #define RANAL_INIT_ALL 7
165 /************************************************************************
166 * Wire message structs. These are sent in sender's byte order
167 * (i.e. receiver checks magic and flips if required).
170 typedef struct kra_connreq /* connection request/response */
171 { /* (sent via socket) */
172 __u32 racr_magic; /* I'm an ranal connreq */
173 __u16 racr_version; /* this is my version number */
174 __u16 racr_devid; /* which device to connect on */
175 __u64 racr_nid; /* my NID */
176 __u64 racr_incarnation; /* my incarnation */
177 __u32 racr_timeout; /* my timeout */
178 RAP_RI_PARAMETERS racr_riparams; /* my endpoint info */
183 RAP_MEM_KEY rard_key;
184 RAP_PVOID64 rard_addr;
190 ptl_hdr_t raim_hdr; /* portals header */
191 /* Portals payload is in FMA "Message Data" */
192 } kra_immediate_msg_t;
196 ptl_hdr_t raprm_hdr; /* portals header */
197 __u64 raprm_cookie; /* opaque completion cookie */
202 __u64 rapam_src_cookie; /* reflected completion cookie */
203 __u64 rapam_dst_cookie; /* opaque completion cookie */
204 kra_rdma_desc_t rapam_desc; /* sender's sink buffer */
209 ptl_hdr_t ragm_hdr; /* portals header */
210 __u64 ragm_cookie; /* opaque completion cookie */
211 kra_rdma_desc_t ragm_desc; /* sender's sink buffer */
216 __u64 racm_cookie; /* reflected completion cookie */
217 } kra_completion_msg_t;
219 typedef struct /* NB must fit in FMA "Prefix" */
221 __u32 ram_magic; /* I'm an ranal message */
222 __u16 ram_version; /* this is my version number */
223 __u16 ram_type; /* msg type */
224 __u64 ram_srcnid; /* sender's NID */
225 __u64 ram_incarnation; /* sender's connection incarnation */
227 kra_immediate_msg_t immediate;
228 kra_putreq_msg_t putreq;
229 kra_putack_msg_t putack;
231 kra_completion_msg_t completion;
233 __u32 ram_seq; /* incrementing sequence number */
236 #define RANAL_MSG_MAGIC 0x0be91b92 /* unique magic */
237 #define RANAL_MSG_VERSION 1 /* current protocol version */
239 #define RANAL_MSG_FENCE 0x80 /* fence RDMA */
241 #define RANAL_MSG_NONE 0x00 /* illegal message */
242 #define RANAL_MSG_NOOP 0x01 /* empty ram_u (keepalive) */
243 #define RANAL_MSG_IMMEDIATE 0x02 /* ram_u.immediate */
244 #define RANAL_MSG_PUT_REQ 0x03 /* ram_u.putreq (src->sink) */
245 #define RANAL_MSG_PUT_NAK 0x04 /* ram_u.completion (no PUT match: sink->src) */
246 #define RANAL_MSG_PUT_ACK 0x05 /* ram_u.putack (PUT matched: sink->src) */
247 #define RANAL_MSG_PUT_DONE 0x86 /* ram_u.completion (src->sink) */
248 #define RANAL_MSG_GET_REQ 0x07 /* ram_u.get (sink->src) */
249 #define RANAL_MSG_GET_NAK 0x08 /* ram_u.completion (no GET match: src->sink) */
250 #define RANAL_MSG_GET_DONE 0x89 /* ram_u.completion (src->sink) */
251 #define RANAL_MSG_CLOSE 0x8a /* empty ram_u */
253 /***********************************************************************/
255 typedef struct kra_tx /* message descriptor */
257 struct list_head tx_list; /* queue on idle_txs/rac_sendq/rac_waitq */
258 struct kra_conn *tx_conn; /* owning conn */
259 lib_msg_t *tx_libmsg[2]; /* lib msgs to finalize on completion */
260 unsigned long tx_qtime; /* when tx started to wait for something */
261 int tx_isnblk; /* I'm reserved for non-blocking sends */
262 int tx_nob; /* # bytes of payload */
263 int tx_buftype; /* payload buffer type */
264 void *tx_buffer; /* source/sink buffer */
265 int tx_phys_offset; /* first page offset (if phys) */
266 int tx_phys_npages; /* # physical pages */
267 RAP_PHYS_REGION *tx_phys; /* page descriptors */
268 RAP_MEM_KEY tx_map_key; /* mapping key */
269 RAP_RDMA_DESCRIPTOR tx_rdma_desc; /* rdma descriptor */
270 __u64 tx_cookie; /* identify this tx to peer */
271 kra_msg_t tx_msg; /* FMA message buffer */
274 #define RANAL_BUF_NONE 0 /* buffer type not set */
275 #define RANAL_BUF_IMMEDIATE 1 /* immediate data */
276 #define RANAL_BUF_PHYS_UNMAPPED 2 /* physical: not mapped yet */
277 #define RANAL_BUF_PHYS_MAPPED 3 /* physical: mapped already */
278 #define RANAL_BUF_VIRT_UNMAPPED 4 /* virtual: not mapped yet */
279 #define RANAL_BUF_VIRT_MAPPED 5 /* virtual: mapped already */
281 #define RANAL_TX_IDLE 0x00 /* on freelist */
282 #define RANAL_TX_SIMPLE 0x10 /* about to send a simple message */
283 #define RANAL_TX_PUTI_REQ 0x20 /* PUT initiator about to send PUT_REQ */
284 #define RANAL_TX_PUTI_WAIT_ACK 0x21 /* PUT initiator waiting for PUT_ACK */
285 #define RANAL_TX_PUTI_RDMA 0x22 /* PUT initiator waiting for RDMA to complete */
286 #define RANAL_TX_PUTI_DONE 0x23 /* PUT initiator about to send PUT_DONE */
287 #define RANAL_TX_PUTT_NAK 0x30 /* PUT target about to send PUT_NAK */
288 #define RANAL_TX_PUTT_ACK 0x30 /* PUT target about to send PUT_ACK */
289 #define RANAL_TX_PUTT_WAIT_DONE 0x31 /* PUT target waiting for PUT_DONE */
290 #define RANAL_TX_GETI_REQ 0x40 /* GET initiator about to send GET_REQ */
291 #define RANAL_TX_GETI_WAIT_DONE 0x41 /* GET initiator waiting for GET_DONE */
292 #define RANAL_TX_GETT_NAK 0x50 /* GET target about to send PUT_NAK */
293 #define RANAL_TX_GETT_RDMA 0x51 /* GET target waiting for RDMA to complete */
294 #define RANAL_TX_GETT_DONE 0x52 /* GET target about to send GET_DONE */
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; /* queue for scheduler */
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_peer_incarnation; /* peer's unique connection stamp */
306 __u64 rac_my_incarnation; /* my unique connection stamp */
307 unsigned long rac_last_tx; /* when I last sent an FMA message */
308 unsigned long rac_last_rx; /* when I last received an FMA messages */
309 long rac_keepalive; /* keepalive interval */
310 long rac_timeout; /* infer peer death on (last_rx + timout > now) */
311 __u32 rac_cqid; /* my completion callback id (non-unique) */
312 __u32 rac_tx_seq; /* tx msg sequence number */
313 __u32 rac_rx_seq; /* rx msg sequence number */
314 atomic_t rac_refcount; /* # users */
315 unsigned int rac_close_sent; /* I've sent CLOSE */
316 unsigned int rac_close_recvd; /* I've received CLOSE */
317 unsigned int rac_closing; /* connection being torn down */
318 unsigned int rac_scheduled; /* being attented to */
319 spinlock_t rac_lock; /* serialise */
320 kra_device_t *rac_device; /* which device */
321 RAP_PVOID rac_rihandle; /* RA endpoint */
322 kra_msg_t *rac_rxmsg; /* incoming message (FMA prefix) */
323 kra_msg_t rac_msg; /* keepalive/CLOSE message buffer */
326 typedef struct kra_peer
328 struct list_head rap_list; /* stash on global peer list */
329 struct list_head rap_connd_list; /* schedule on kra_connd_peers */
330 struct list_head rap_conns; /* all active connections */
331 struct list_head rap_tx_queue; /* msgs waiting for a conn */
332 ptl_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 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_TIME when reconnect OK */
339 unsigned long rap_reconnect_interval; /* exponential backoff */
343 extern lib_nal_t kranal_lib;
344 extern kra_data_t kranal_data;
345 extern kra_tunables_t kranal_tunables;
348 kranal_peer_addref(kra_peer_t *peer)
350 CDEBUG(D_NET, "%p->"LPX64"\n", peer, peer->rap_nid);
351 LASSERT(atomic_read(&peer->rap_refcount) > 0);
352 atomic_inc(&peer->rap_refcount);
356 kranal_peer_decref(kra_peer_t *peer)
358 CDEBUG(D_NET, "%p->"LPX64"\n", peer, peer->rap_nid);
359 LASSERT(atomic_read(&peer->rap_refcount) > 0);
360 if (atomic_dec_and_test(&peer->rap_refcount))
361 __kranal_peer_decref(peer);
364 static inline struct list_head *
365 kranal_nid2peerlist (ptl_nid_t nid)
367 unsigned int hash = ((unsigned int)nid) % kranal_data.kra_peer_hash_size;
369 return (&kranal_data.kra_peers [hash]);
373 kranal_peer_active(kra_peer_t *peer)
375 /* Am I in the peer hash table? */
376 return (!list_empty(&peer->rap_list));
380 kranal_conn_addref(kra_conn_t *conn)
382 CDEBUG(D_NET, "%p->"LPX64"\n", conn, conn->rac_peer->rap_nid);
383 LASSERT(atomic_read(&conn->rac_refcount) > 0);
384 atomic_inc(&conn->rac_refcount);
388 kranal_conn_decref(kra_conn_t *conn)
390 CDEBUG(D_NET, "%p->"LPX64"\n", conn, conn->rac_peer->rap_nid);
391 LASSERT(atomic_read(&conn->rac_refcount) > 0);
392 if (atomic_dec_and_test(&conn->rac_refcount))
393 __kranal_conn_decref(conn);
396 static inline struct list_head *
397 kranal_cqid2connlist (__u32 cqid)
399 unsigned int hash = cqid % kranal_data.kra_conn_hash_size;
401 return (&kranal_data.kra_conns [hash]);
404 static inline kra_conn_t *
405 kranal_cqid2conn_locked (__u32 cqid)
407 struct list_head conns = kranal_cqid2connlist(cqid);
408 struct list_head *tmp;
410 list_for_each(tmp, conns) {
411 conn = list_entry(tmp, kra_conn_t, rac_hashlist);
413 if (conn->rac_cqid == cqid)
421 kranal_tx_mapped (kra_tx_t *tx)
423 return (tx->tx_buftype == RANAL_BUF_VIRT_MAPPED ||
424 tx->tx_buftype == RANAL_BUF_PHYS_MAPPED);
429 kranal_page2phys (struct page *p)
431 __u64 page_number = p - mem_map;
433 return (page_number << PAGE_SHIFT);
436 # error "no page->phys"