projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'wr-cleanup' into k.o/for-4.4
[deliverable/linux.git] / drivers / staging / lustre / lnet / klnds / o2iblnd / o2iblnd.h
1 /*
2 * GPL HEADER START
3 *
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
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.
9 *
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).
15 *
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
19 *
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
22 * have any questions.
23 *
24 * GPL HEADER END
25 */
26 /*
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
29 *
30 * Copyright (c) 2011, 2012, Intel Corporation.
31 */
32 /*
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
35 *
36 * lnet/klnds/o2iblnd/o2iblnd.h
37 *
38 * Author: Eric Barton <eric@bartonsoftware.com>
39 */
40
41 #include <linux/module.h>
42 #include <linux/kernel.h>
43 #include <linux/mm.h>
44 #include <linux/string.h>
45 #include <linux/stat.h>
46 #include <linux/errno.h>
47 #include <linux/unistd.h>
48 #include <linux/uio.h>
49 #include <linux/uaccess.h>
50
51 #include <linux/io.h>
52
53 #include <linux/fs.h>
54 #include <linux/file.h>
55 #include <linux/list.h>
56 #include <linux/kmod.h>
57 #include <linux/sysctl.h>
58 #include <linux/pci.h>
59
60 #include <net/sock.h>
61 #include <linux/in.h>
62
63 #define DEBUG_SUBSYSTEM S_LND
64
65 #include "../../../include/linux/libcfs/libcfs.h"
66 #include "../../../include/linux/lnet/lnet.h"
67 #include "../../../include/linux/lnet/lib-lnet.h"
68
69 #include <rdma/rdma_cm.h>
70 #include <rdma/ib_cm.h>
71 #include <rdma/ib_verbs.h>
72 #include <rdma/ib_fmr_pool.h>
73
74 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
75 /* # scheduler loops before reschedule */
76 #define IBLND_RESCHED 100
77
78 #define IBLND_N_SCHED 2
79 #define IBLND_N_SCHED_HIGH 4
80
81 typedef struct {
82 int *kib_dev_failover; /* HCA failover */
83 unsigned int *kib_service; /* IB service number */
84 int *kib_min_reconnect_interval; /* first failed connection
85 * retry... */
86 int *kib_max_reconnect_interval; /* ...exponentially increasing
87 * to this */
88 int *kib_cksum; /* checksum kib_msg_t? */
89 int *kib_timeout; /* comms timeout (seconds) */
90 int *kib_keepalive; /* keepalive timeout (seconds) */
91 int *kib_ntx; /* # tx descs */
92 int *kib_credits; /* # concurrent sends */
93 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
94 int *kib_peerrtrcredits; /* # per-peer router buffer
95 * credits */
96 int *kib_peercredits_hiw; /* # when eagerly to return
97 * credits */
98 int *kib_peertimeout; /* seconds to consider peer dead */
99 char **kib_default_ipif; /* default IPoIB interface */
100 int *kib_retry_count;
101 int *kib_rnr_retry_count;
102 int *kib_concurrent_sends; /* send work queue sizing */
103 int *kib_ib_mtu; /* IB MTU */
104 int *kib_map_on_demand; /* map-on-demand if RD has more
105 * fragments than this value, 0
106 * disable map-on-demand */
107 int *kib_fmr_pool_size; /* # FMRs in pool */
108 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
109 int *kib_fmr_cache; /* enable FMR pool cache? */
110 int *kib_require_priv_port; /* accept only privileged ports */
111 int *kib_use_priv_port; /* use privileged port for active
112 * connect */
113 int *kib_nscheds; /* # threads on each CPT */
114 } kib_tunables_t;
115
116 extern kib_tunables_t kiblnd_tunables;
117
118 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
119 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
120
121 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
122 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t *) 0)->ibm_credits)) - 1) /* Max # of peer credits */
123
124 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
125 IBLND_MSG_QUEUE_SIZE_V1 : \
126 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
127 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
128 IBLND_CREDIT_HIGHWATER_V1 : \
129 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
130
131 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(&init_net, \
132 cb, dev, \
133 ps, qpt)
134
135 static inline int
136 kiblnd_concurrent_sends_v1(void)
137 {
138 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
139 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
140
141 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
142 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
143
144 return *kiblnd_tunables.kib_concurrent_sends;
145 }
146
147 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
148 kiblnd_concurrent_sends_v1() : \
149 *kiblnd_tunables.kib_concurrent_sends)
150 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
151 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
152 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
153
154 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
155 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
156 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
157 *kiblnd_tunables.kib_map_on_demand : \
158 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
159 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
160 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
161
162 /************************/
163 /* derived constants... */
164 /* Pools (shared by connections on each CPT) */
165 /* These pools can grow at runtime, so don't need give a very large value */
166 #define IBLND_TX_POOL 256
167 #define IBLND_FMR_POOL 256
168 #define IBLND_FMR_POOL_FLUSH 192
169
170 /* TX messages (shared by all connections) */
171 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
172
173 /* RX messages (per connection) */
174 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
175 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
176 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
177
178 /* WRs and CQEs (per connection) */
179 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
180 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
181 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
182
183 struct kib_hca_dev;
184
185 /* o2iblnd can run over aliased interface */
186 #ifdef IFALIASZ
187 #define KIB_IFNAME_SIZE IFALIASZ
188 #else
189 #define KIB_IFNAME_SIZE 256
190 #endif
191
192 typedef struct {
193 struct list_head ibd_list; /* chain on kib_devs */
194 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
195 __u32 ibd_ifip; /* IPoIB interface IP */
196
197 /* IPoIB interface name */
198 char ibd_ifname[KIB_IFNAME_SIZE];
199 int ibd_nnets; /* # nets extant */
200
201 unsigned long ibd_next_failover;
202 int ibd_failed_failover; /* # failover failures */
203 unsigned int ibd_failover; /* failover in progress */
204 unsigned int ibd_can_failover; /* IPoIB interface is a bonding
205 * master */
206 struct list_head ibd_nets;
207 struct kib_hca_dev *ibd_hdev;
208 } kib_dev_t;
209
210 typedef struct kib_hca_dev {
211 struct rdma_cm_id *ibh_cmid; /* listener cmid */
212 struct ib_device *ibh_ibdev; /* IB device */
213 int ibh_page_shift; /* page shift of current HCA */
214 int ibh_page_size; /* page size of current HCA */
215 __u64 ibh_page_mask; /* page mask of current HCA */
216 int ibh_mr_shift; /* bits shift of max MR size */
217 __u64 ibh_mr_size; /* size of MR */
218 int ibh_nmrs; /* # of global MRs */
219 struct ib_mr **ibh_mrs; /* global MR */
220 struct ib_pd *ibh_pd; /* PD */
221 kib_dev_t *ibh_dev; /* owner */
222 atomic_t ibh_ref; /* refcount */
223 } kib_hca_dev_t;
224
225 /** # of seconds to keep pool alive */
226 #define IBLND_POOL_DEADLINE 300
227 /** # of seconds to retry if allocation failed */
228 #define IBLND_POOL_RETRY 1
229
230 typedef struct {
231 int ibp_npages; /* # pages */
232 struct page *ibp_pages[0]; /* page array */
233 } kib_pages_t;
234
235 struct kib_pool;
236 struct kib_poolset;
237
238 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
239 int inc, struct kib_pool **pp_po);
240 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
241 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
242 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
243
244 struct kib_net;
245
246 #define IBLND_POOL_NAME_LEN 32
247
248 typedef struct kib_poolset {
249 spinlock_t ps_lock; /* serialize */
250 struct kib_net *ps_net; /* network it belongs to */
251 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
252 struct list_head ps_pool_list; /* list of pools */
253 struct list_head ps_failed_pool_list;/* failed pool list */
254 unsigned long ps_next_retry; /* time stamp for retry if
255 * failed to allocate */
256 int ps_increasing; /* is allocating new pool */
257 int ps_pool_size; /* new pool size */
258 int ps_cpt; /* CPT id */
259
260 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
261 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
262 kib_ps_node_init_t ps_node_init; /* initialize new allocated
263 * node */
264 kib_ps_node_fini_t ps_node_fini; /* finalize node */
265 } kib_poolset_t;
266
267 typedef struct kib_pool {
268 struct list_head po_list; /* chain on pool list */
269 struct list_head po_free_list; /* pre-allocated node */
270 kib_poolset_t *po_owner; /* pool_set of this pool */
271 unsigned long po_deadline; /* deadline of this pool */
272 int po_allocated; /* # of elements in use */
273 int po_failed; /* pool is created on failed
274 * HCA */
275 int po_size; /* # of pre-allocated elements */
276 } kib_pool_t;
277
278 typedef struct {
279 kib_poolset_t tps_poolset; /* pool-set */
280 __u64 tps_next_tx_cookie; /* cookie of TX */
281 } kib_tx_poolset_t;
282
283 typedef struct {
284 kib_pool_t tpo_pool; /* pool */
285 struct kib_hca_dev *tpo_hdev; /* device for this pool */
286 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
287 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
288 } kib_tx_pool_t;
289
290 typedef struct {
291 spinlock_t fps_lock; /* serialize */
292 struct kib_net *fps_net; /* IB network */
293 struct list_head fps_pool_list; /* FMR pool list */
294 struct list_head fps_failed_pool_list;/* FMR pool list */
295 __u64 fps_version; /* validity stamp */
296 int fps_cpt; /* CPT id */
297 int fps_pool_size;
298 int fps_flush_trigger;
299 int fps_increasing; /* is allocating new pool */
300 unsigned long fps_next_retry; /* time stamp for retry if
301 * failed to allocate */
302 } kib_fmr_poolset_t;
303
304 typedef struct {
305 struct list_head fpo_list; /* chain on pool list */
306 struct kib_hca_dev *fpo_hdev; /* device for this pool */
307 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
308 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
309 unsigned long fpo_deadline; /* deadline of this pool */
310 int fpo_failed; /* fmr pool is failed */
311 int fpo_map_count; /* # of mapped FMR */
312 } kib_fmr_pool_t;
313
314 typedef struct {
315 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
316 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
317 } kib_fmr_t;
318
319 typedef struct kib_net {
320 struct list_head ibn_list; /* chain on kib_dev_t::ibd_nets */
321 __u64 ibn_incarnation;/* my epoch */
322 int ibn_init; /* initialisation state */
323 int ibn_shutdown; /* shutting down? */
324
325 atomic_t ibn_npeers; /* # peers extant */
326 atomic_t ibn_nconns; /* # connections extant */
327
328 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
329 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
330
331 kib_dev_t *ibn_dev; /* underlying IB device */
332 } kib_net_t;
333
334 #define KIB_THREAD_SHIFT 16
335 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
336 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
337 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
338
339 struct kib_sched_info {
340 spinlock_t ibs_lock; /* serialise */
341 wait_queue_head_t ibs_waitq; /* schedulers sleep here */
342 struct list_head ibs_conns; /* conns to check for rx completions */
343 int ibs_nthreads; /* number of scheduler threads */
344 int ibs_nthreads_max; /* max allowed scheduler threads */
345 int ibs_cpt; /* CPT id */
346 };
347
348 typedef struct {
349 int kib_init; /* initialisation state */
350 int kib_shutdown; /* shut down? */
351 struct list_head kib_devs; /* IB devices extant */
352 struct list_head kib_failed_devs; /* list head of failed
353 * devices */
354 wait_queue_head_t kib_failover_waitq; /* schedulers sleep here */
355 atomic_t kib_nthreads; /* # live threads */
356 rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn
357 * ops */
358 struct list_head *kib_peers; /* hash table of all my known
359 * peers */
360 int kib_peer_hash_size; /* size of kib_peers */
361 void *kib_connd; /* the connd task
362 * (serialisation assertions)
363 */
364 struct list_head kib_connd_conns; /* connections to
365 * setup/teardown */
366 struct list_head kib_connd_zombies; /* connections with zero
367 * refcount */
368 wait_queue_head_t kib_connd_waitq; /* connection daemon sleeps
369 * here */
370 spinlock_t kib_connd_lock; /* serialise */
371 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
372 struct kib_sched_info **kib_scheds; /* percpt data for schedulers
373 */
374 } kib_data_t;
375
376 #define IBLND_INIT_NOTHING 0
377 #define IBLND_INIT_DATA 1
378 #define IBLND_INIT_ALL 2
379
380 /************************************************************************
381 * IB Wire message format.
382 * These are sent in sender's byte order (i.e. receiver flips).
383 */
384
385 typedef struct kib_connparams {
386 __u16 ibcp_queue_depth;
387 __u16 ibcp_max_frags;
388 __u32 ibcp_max_msg_size;
389 } WIRE_ATTR kib_connparams_t;
390
391 typedef struct {
392 lnet_hdr_t ibim_hdr; /* portals header */
393 char ibim_payload[0]; /* piggy-backed payload */
394 } WIRE_ATTR kib_immediate_msg_t;
395
396 typedef struct {
397 __u32 rf_nob; /* # bytes this frag */
398 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
399 } WIRE_ATTR kib_rdma_frag_t;
400
401 typedef struct {
402 __u32 rd_key; /* local/remote key */
403 __u32 rd_nfrags; /* # fragments */
404 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
405 } WIRE_ATTR kib_rdma_desc_t;
406
407 typedef struct {
408 lnet_hdr_t ibprm_hdr; /* portals header */
409 __u64 ibprm_cookie; /* opaque completion cookie */
410 } WIRE_ATTR kib_putreq_msg_t;
411
412 typedef struct {
413 __u64 ibpam_src_cookie; /* reflected completion cookie */
414 __u64 ibpam_dst_cookie; /* opaque completion cookie */
415 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
416 } WIRE_ATTR kib_putack_msg_t;
417
418 typedef struct {
419 lnet_hdr_t ibgm_hdr; /* portals header */
420 __u64 ibgm_cookie; /* opaque completion cookie */
421 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
422 } WIRE_ATTR kib_get_msg_t;
423
424 typedef struct {
425 __u64 ibcm_cookie; /* opaque completion cookie */
426 __s32 ibcm_status; /* < 0 failure: >= 0 length */
427 } WIRE_ATTR kib_completion_msg_t;
428
429 typedef struct {
430 /* First 2 fields fixed FOR ALL TIME */
431 __u32 ibm_magic; /* I'm an ibnal message */
432 __u16 ibm_version; /* this is my version number */
433
434 __u8 ibm_type; /* msg type */
435 __u8 ibm_credits; /* returned credits */
436 __u32 ibm_nob; /* # bytes in whole message */
437 __u32 ibm_cksum; /* checksum (0 == no checksum) */
438 __u64 ibm_srcnid; /* sender's NID */
439 __u64 ibm_srcstamp; /* sender's incarnation */
440 __u64 ibm_dstnid; /* destination's NID */
441 __u64 ibm_dststamp; /* destination's incarnation */
442
443 union {
444 kib_connparams_t connparams;
445 kib_immediate_msg_t immediate;
446 kib_putreq_msg_t putreq;
447 kib_putack_msg_t putack;
448 kib_get_msg_t get;
449 kib_completion_msg_t completion;
450 } WIRE_ATTR ibm_u;
451 } WIRE_ATTR kib_msg_t;
452
453 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
454
455 #define IBLND_MSG_VERSION_1 0x11
456 #define IBLND_MSG_VERSION_2 0x12
457 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
458
459 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
460 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
461 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
462 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
463 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
464 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
465 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
466 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
467 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
468 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
469
470 typedef struct {
471 __u32 ibr_magic; /* sender's magic */
472 __u16 ibr_version; /* sender's version */
473 __u8 ibr_why; /* reject reason */
474 __u8 ibr_padding; /* padding */
475 __u64 ibr_incarnation; /* incarnation of peer */
476 kib_connparams_t ibr_cp; /* connection parameters */
477 } WIRE_ATTR kib_rej_t;
478
479 /* connection rejection reasons */
480 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
481 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
482 #define IBLND_REJECT_FATAL 3 /* Anything else */
483 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
484 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
485 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match
486 * mine */
487 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't
488 * match mine */
489
490 /***********************************************************************/
491
492 typedef struct kib_rx /* receive message */
493 {
494 struct list_head rx_list; /* queue for attention */
495 struct kib_conn *rx_conn; /* owning conn */
496 int rx_nob; /* # bytes received (-1 while
497 * posted) */
498 enum ib_wc_status rx_status; /* completion status */
499 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
500 __u64 rx_msgaddr; /* message buffer (I/O addr) */
501 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap); /* for dma_unmap_single() */
502 struct ib_recv_wr rx_wrq; /* receive work item... */
503 struct ib_sge rx_sge; /* ...and its memory */
504 } kib_rx_t;
505
506 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
507 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
508 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
509 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved
510 * credit */
511
512 typedef struct kib_tx /* transmit message */
513 {
514 struct list_head tx_list; /* queue on idle_txs ibc_tx_queue
515 * etc. */
516 kib_tx_pool_t *tx_pool; /* pool I'm from */
517 struct kib_conn *tx_conn; /* owning conn */
518 short tx_sending; /* # tx callbacks outstanding */
519 short tx_queued; /* queued for sending */
520 short tx_waiting; /* waiting for peer */
521 int tx_status; /* LNET completion status */
522 unsigned long tx_deadline; /* completion deadline */
523 __u64 tx_cookie; /* completion cookie */
524 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on
525 * completion */
526 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
527 __u64 tx_msgaddr; /* message buffer (I/O addr) */
528 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap); /* for dma_unmap_single() */
529 int tx_nwrq; /* # send work items */
530 struct ib_rdma_wr *tx_wrq; /* send work items... */
531 struct ib_sge *tx_sge; /* ...and their memory */
532 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
533 int tx_nfrags; /* # entries in... */
534 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
535 __u64 *tx_pages; /* rdma phys page addrs */
536 kib_fmr_t fmr; /* FMR */
537 int tx_dmadir; /* dma direction */
538 } kib_tx_t;
539
540 typedef struct kib_connvars {
541 kib_msg_t cv_msg; /* connection-in-progress variables */
542 } kib_connvars_t;
543
544 typedef struct kib_conn {
545 struct kib_sched_info *ibc_sched; /* scheduler information */
546 struct kib_peer *ibc_peer; /* owning peer */
547 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
548 struct list_head ibc_list; /* stash on peer's conn
549 * list */
550 struct list_head ibc_sched_list; /* schedule for attention */
551 __u16 ibc_version; /* version of connection */
552 __u64 ibc_incarnation; /* which instance of the
553 * peer */
554 atomic_t ibc_refcount; /* # users */
555 int ibc_state; /* what's happening */
556 int ibc_nsends_posted; /* # uncompleted sends */
557 int ibc_noops_posted; /* # uncompleted NOOPs */
558 int ibc_credits; /* # credits I have */
559 int ibc_outstanding_credits; /* # credits to return */
560 int ibc_reserved_credits; /* # ACK/DONE msg credits */
561 int ibc_comms_error; /* set on comms error */
562 unsigned int ibc_nrx:16; /* receive buffers owned */
563 unsigned int ibc_scheduled:1; /* scheduled for attention
564 */
565 unsigned int ibc_ready:1; /* CQ callback fired */
566 unsigned long ibc_last_send; /* time of last send */
567 struct list_head ibc_connd_list; /* link chain for
568 * kiblnd_check_conns only
569 */
570 struct list_head ibc_early_rxs; /* rxs completed before
571 * ESTABLISHED */
572 struct list_head ibc_tx_noops; /* IBLND_MSG_NOOPs for
573 * IBLND_MSG_VERSION_1 */
574 struct list_head ibc_tx_queue; /* sends that need a credit
575 */
576 struct list_head ibc_tx_queue_nocred; /* sends that don't need a
577 * credit */
578 struct list_head ibc_tx_queue_rsrvd; /* sends that need to
579 * reserve an ACK/DONE msg
580 */
581 struct list_head ibc_active_txs; /* active tx awaiting
582 * completion */
583 spinlock_t ibc_lock; /* serialise */
584 kib_rx_t *ibc_rxs; /* the rx descs */
585 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
586
587 struct rdma_cm_id *ibc_cmid; /* CM id */
588 struct ib_cq *ibc_cq; /* completion queue */
589
590 kib_connvars_t *ibc_connvars; /* in-progress connection
591 * state */
592 } kib_conn_t;
593
594 #define IBLND_CONN_INIT 0 /* being initialised */
595 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
596 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
597 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
598 #define IBLND_CONN_CLOSING 4 /* being closed */
599 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
600
601 typedef struct kib_peer {
602 struct list_head ibp_list; /* stash on global peer list */
603 lnet_nid_t ibp_nid; /* who's on the other end(s) */
604 lnet_ni_t *ibp_ni; /* LNet interface */
605 atomic_t ibp_refcount; /* # users */
606 struct list_head ibp_conns; /* all active connections */
607 struct list_head ibp_tx_queue; /* msgs waiting for a conn */
608 __u16 ibp_version; /* version of peer */
609 __u64 ibp_incarnation; /* incarnation of peer */
610 int ibp_connecting; /* current active connection attempts
611 */
612 int ibp_accepting; /* current passive connection attempts
613 */
614 int ibp_error; /* errno on closing this peer */
615 unsigned long ibp_last_alive; /* when (in jiffies) I was last alive
616 */
617 } kib_peer_t;
618
619 extern kib_data_t kiblnd_data;
620
621 void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
622
623 static inline void
624 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
625 {
626 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
627 atomic_inc(&hdev->ibh_ref);
628 }
629
630 static inline void
631 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
632 {
633 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
634 if (atomic_dec_and_test(&hdev->ibh_ref))
635 kiblnd_hdev_destroy(hdev);
636 }
637
638 static inline int
639 kiblnd_dev_can_failover(kib_dev_t *dev)
640 {
641 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
642 return 0;
643
644 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
645 return 0;
646
647 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
648 return 1;
649
650 return dev->ibd_can_failover;
651 }
652
653 #define kiblnd_conn_addref(conn) \
654 do { \
655 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
656 (conn), atomic_read(&(conn)->ibc_refcount)); \
657 atomic_inc(&(conn)->ibc_refcount); \
658 } while (0)
659
660 #define kiblnd_conn_decref(conn) \
661 do { \
662 unsigned long flags; \
663 \
664 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
665 (conn), atomic_read(&(conn)->ibc_refcount)); \
666 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
667 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
668 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
669 list_add_tail(&(conn)->ibc_list, \
670 &kiblnd_data.kib_connd_zombies); \
671 wake_up(&kiblnd_data.kib_connd_waitq); \
672 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
673 } \
674 } while (0)
675
676 #define kiblnd_peer_addref(peer) \
677 do { \
678 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
679 (peer), libcfs_nid2str((peer)->ibp_nid), \
680 atomic_read(&(peer)->ibp_refcount)); \
681 atomic_inc(&(peer)->ibp_refcount); \
682 } while (0)
683
684 #define kiblnd_peer_decref(peer) \
685 do { \
686 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
687 (peer), libcfs_nid2str((peer)->ibp_nid), \
688 atomic_read(&(peer)->ibp_refcount)); \
689 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
690 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
691 kiblnd_destroy_peer(peer); \
692 } while (0)
693
694 static inline struct list_head *
695 kiblnd_nid2peerlist(lnet_nid_t nid)
696 {
697 unsigned int hash =
698 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
699
700 return &kiblnd_data.kib_peers[hash];
701 }
702
703 static inline int
704 kiblnd_peer_active(kib_peer_t *peer)
705 {
706 /* Am I in the peer hash table? */
707 return !list_empty(&peer->ibp_list);
708 }
709
710 static inline kib_conn_t *
711 kiblnd_get_conn_locked(kib_peer_t *peer)
712 {
713 LASSERT(!list_empty(&peer->ibp_conns));
714
715 /* just return the first connection */
716 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
717 }
718
719 static inline int
720 kiblnd_send_keepalive(kib_conn_t *conn)
721 {
722 return (*kiblnd_tunables.kib_keepalive > 0) &&
723 cfs_time_after(jiffies, conn->ibc_last_send +
724 *kiblnd_tunables.kib_keepalive*HZ);
725 }
726
727 static inline int
728 kiblnd_need_noop(kib_conn_t *conn)
729 {
730 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
731
732 if (conn->ibc_outstanding_credits <
733 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
734 !kiblnd_send_keepalive(conn))
735 return 0; /* No need to send NOOP */
736
737 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
738 if (!list_empty(&conn->ibc_tx_queue_nocred))
739 return 0; /* NOOP can be piggybacked */
740
741 /* No tx to piggyback NOOP onto or no credit to send a tx */
742 return (list_empty(&conn->ibc_tx_queue) ||
743 conn->ibc_credits == 0);
744 }
745
746 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
747 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
748 conn->ibc_credits == 0) /* no credit */
749 return 0;
750
751 if (conn->ibc_credits == 1 && /* last credit reserved for */
752 conn->ibc_outstanding_credits == 0) /* giving back credits */
753 return 0;
754
755 /* No tx to piggyback NOOP onto or no credit to send a tx */
756 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
757 }
758
759 static inline void
760 kiblnd_abort_receives(kib_conn_t *conn)
761 {
762 ib_modify_qp(conn->ibc_cmid->qp,
763 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
764 }
765
766 static inline const char *
767 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
768 {
769 if (q == &conn->ibc_tx_queue)
770 return "tx_queue";
771
772 if (q == &conn->ibc_tx_queue_rsrvd)
773 return "tx_queue_rsrvd";
774
775 if (q == &conn->ibc_tx_queue_nocred)
776 return "tx_queue_nocred";
777
778 if (q == &conn->ibc_active_txs)
779 return "active_txs";
780
781 LBUG();
782 return NULL;
783 }
784
785 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
786 * lowest bits of the work request id to stash the work item type. */
787
788 #define IBLND_WID_TX 0
789 #define IBLND_WID_RDMA 1
790 #define IBLND_WID_RX 2
791 #define IBLND_WID_MASK 3UL
792
793 static inline __u64
794 kiblnd_ptr2wreqid(void *ptr, int type)
795 {
796 unsigned long lptr = (unsigned long)ptr;
797
798 LASSERT((lptr & IBLND_WID_MASK) == 0);
799 LASSERT((type & ~IBLND_WID_MASK) == 0);
800 return (__u64)(lptr | type);
801 }
802
803 static inline void *
804 kiblnd_wreqid2ptr(__u64 wreqid)
805 {
806 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
807 }
808
809 static inline int
810 kiblnd_wreqid2type(__u64 wreqid)
811 {
812 return wreqid & IBLND_WID_MASK;
813 }
814
815 static inline void
816 kiblnd_set_conn_state(kib_conn_t *conn, int state)
817 {
818 conn->ibc_state = state;
819 mb();
820 }
821
822 static inline void
823 kiblnd_init_msg(kib_msg_t *msg, int type, int body_nob)
824 {
825 msg->ibm_type = type;
826 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
827 }
828
829 static inline int
830 kiblnd_rd_size(kib_rdma_desc_t *rd)
831 {
832 int i;
833 int size;
834
835 for (i = size = 0; i < rd->rd_nfrags; i++)
836 size += rd->rd_frags[i].rf_nob;
837
838 return size;
839 }
840
841 static inline __u64
842 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
843 {
844 return rd->rd_frags[index].rf_addr;
845 }
846
847 static inline __u32
848 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
849 {
850 return rd->rd_frags[index].rf_nob;
851 }
852
853 static inline __u32
854 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
855 {
856 return rd->rd_key;
857 }
858
859 static inline int
860 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
861 {
862 if (nob < rd->rd_frags[index].rf_nob) {
863 rd->rd_frags[index].rf_addr += nob;
864 rd->rd_frags[index].rf_nob -= nob;
865 } else {
866 index++;
867 }
868
869 return index;
870 }
871
872 static inline int
873 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
874 {
875 LASSERT(msgtype == IBLND_MSG_GET_REQ ||
876 msgtype == IBLND_MSG_PUT_ACK);
877
878 return msgtype == IBLND_MSG_GET_REQ ?
879 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
880 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
881 }
882
883
884 static inline __u64
885 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
886 {
887 return ib_dma_mapping_error(dev, dma_addr);
888 }
889
890 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
891 void *msg, size_t size,
892 enum dma_data_direction direction)
893 {
894 return ib_dma_map_single(dev, msg, size, direction);
895 }
896
897 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
898 __u64 addr, size_t size,
899 enum dma_data_direction direction)
900 {
901 ib_dma_unmap_single(dev, addr, size, direction);
902 }
903
904 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
905 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
906
907 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
908 struct scatterlist *sg, int nents,
909 enum dma_data_direction direction)
910 {
911 return ib_dma_map_sg(dev, sg, nents, direction);
912 }
913
914 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
915 struct scatterlist *sg, int nents,
916 enum dma_data_direction direction)
917 {
918 ib_dma_unmap_sg(dev, sg, nents, direction);
919 }
920
921 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
922 struct scatterlist *sg)
923 {
924 return ib_sg_dma_address(dev, sg);
925 }
926
927 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
928 struct scatterlist *sg)
929 {
930 return ib_sg_dma_len(dev, sg);
931 }
932
933 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
934 * right because OFED1.2 defines it as const, to use it we have to add
935 * (void *) cast to overcome "const" */
936
937 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
938 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
939
940
941 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
942 kib_rdma_desc_t *rd);
943 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
944 __u64 addr, __u64 size);
945 void kiblnd_map_rx_descs(kib_conn_t *conn);
946 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
947 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
948 kib_rdma_desc_t *rd, int nfrags);
949 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
950 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
951 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
952
953 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
954 int npages, __u64 iov, kib_fmr_t *fmr);
955 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
956
957 int kiblnd_startup(lnet_ni_t *ni);
958 void kiblnd_shutdown(lnet_ni_t *ni);
959 int kiblnd_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg);
960 void kiblnd_query(struct lnet_ni *ni, lnet_nid_t nid, unsigned long *when);
961
962 int kiblnd_tunables_init(void);
963 void kiblnd_tunables_fini(void);
964
965 int kiblnd_connd(void *arg);
966 int kiblnd_scheduler(void *arg);
967 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
968 int kiblnd_failover_thread(void *arg);
969
970 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
971 void kiblnd_free_pages(kib_pages_t *p);
972
973 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
974 struct rdma_cm_event *event);
975 int kiblnd_translate_mtu(int value);
976
977 int kiblnd_dev_failover(kib_dev_t *dev);
978 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
979 void kiblnd_destroy_peer(kib_peer_t *peer);
980 void kiblnd_destroy_dev(kib_dev_t *dev);
981 void kiblnd_unlink_peer_locked(kib_peer_t *peer);
982 void kiblnd_peer_alive(kib_peer_t *peer);
983 kib_peer_t *kiblnd_find_peer_locked(lnet_nid_t nid);
984 void kiblnd_peer_connect_failed(kib_peer_t *peer, int active, int error);
985 int kiblnd_close_stale_conns_locked(kib_peer_t *peer,
986 int version, __u64 incarnation);
987 int kiblnd_close_peer_conns_locked(kib_peer_t *peer, int why);
988
989 void kiblnd_connreq_done(kib_conn_t *conn, int status);
990 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
991 int state, int version);
992 void kiblnd_destroy_conn(kib_conn_t *conn);
993 void kiblnd_close_conn(kib_conn_t *conn, int error);
994 void kiblnd_close_conn_locked(kib_conn_t *conn, int error);
995
996 int kiblnd_init_rdma(kib_conn_t *conn, kib_tx_t *tx, int type,
997 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
998
999 void kiblnd_launch_tx(lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1000 void kiblnd_queue_tx_locked(kib_tx_t *tx, kib_conn_t *conn);
1001 void kiblnd_queue_tx(kib_tx_t *tx, kib_conn_t *conn);
1002 void kiblnd_init_tx_msg(lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1003 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist,
1004 int status);
1005 void kiblnd_check_sends (kib_conn_t *conn);
1006
1007 void kiblnd_qp_event(struct ib_event *event, void *arg);
1008 void kiblnd_cq_event(struct ib_event *event, void *arg);
1009 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1010
1011 void kiblnd_pack_msg(lnet_ni_t *ni, kib_msg_t *msg, int version,
1012 int credits, lnet_nid_t dstnid, __u64 dststamp);
1013 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1014 int kiblnd_post_rx(kib_rx_t *rx, int credit);
1015
1016 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1017 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1018 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1019 unsigned int offset, unsigned int mlen, unsigned int rlen);
Linux kernel - Deliverables
This page took 0.067784 seconds and 5 git commands to generate.