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) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd.h
38 * Author: Eric Barton <eric@bartonsoftware.com>
41 #include <linux/module.h>
42 #include <linux/kernel.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>
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>
63 #define DEBUG_SUBSYSTEM S_LND
65 #include "../../../include/linux/libcfs/libcfs.h"
66 #include "../../../include/linux/lnet/lnet.h"
67 #include "../../../include/linux/lnet/lib-lnet.h"
68 #include "../../../include/linux/lnet/lnet-sysctl.h"
70 #include <rdma/rdma_cm.h>
71 #include <rdma/ib_cm.h>
72 #include <rdma/ib_verbs.h>
73 #include <rdma/ib_fmr_pool.h>
75 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
76 /* # scheduler loops before reschedule */
77 #define IBLND_RESCHED 100
79 #define IBLND_N_SCHED 2
80 #define IBLND_N_SCHED_HIGH 4
83 int *kib_dev_failover; /* HCA failover */
84 unsigned int *kib_service; /* IB service number */
85 int *kib_min_reconnect_interval; /* first failed connection retry... */
86 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
87 int *kib_cksum; /* checksum kib_msg_t? */
88 int *kib_timeout; /* comms timeout (seconds) */
89 int *kib_keepalive; /* keepalive timeout (seconds) */
90 int *kib_ntx; /* # tx descs */
91 int *kib_credits; /* # concurrent sends */
92 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
93 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
94 int *kib_peercredits_hiw; /* # when eagerly to return credits */
95 int *kib_peertimeout; /* seconds to consider peer dead */
96 char **kib_default_ipif; /* default IPoIB interface */
98 int *kib_rnr_retry_count;
99 int *kib_concurrent_sends; /* send work queue sizing */
100 int *kib_ib_mtu; /* IB MTU */
101 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
102 * than this value, 0 disable map-on-demand */
103 int *kib_pmr_pool_size; /* # physical MR in pool */
104 int *kib_fmr_pool_size; /* # FMRs in pool */
105 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
106 int *kib_fmr_cache; /* enable FMR pool cache? */
107 int *kib_require_priv_port;/* accept only privileged ports */
108 int *kib_use_priv_port; /* use privileged port for active connect */
109 /* # threads on each CPT */
113 extern kib_tunables_t kiblnd_tunables;
115 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
116 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
118 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
119 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
121 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
122 IBLND_MSG_QUEUE_SIZE_V1 : \
123 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
124 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
125 IBLND_CREDIT_HIGHWATER_V1 : \
126 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
128 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
131 kiblnd_concurrent_sends_v1(void)
133 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
134 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
136 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
137 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
139 return *kiblnd_tunables.kib_concurrent_sends;
142 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
143 kiblnd_concurrent_sends_v1() : \
144 *kiblnd_tunables.kib_concurrent_sends)
145 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
146 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
147 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
149 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
150 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
151 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
152 *kiblnd_tunables.kib_map_on_demand : \
153 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
154 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
155 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
157 /************************/
158 /* derived constants... */
159 /* Pools (shared by connections on each CPT) */
160 /* These pools can grow at runtime, so don't need give a very large value */
161 #define IBLND_TX_POOL 256
162 #define IBLND_PMR_POOL 256
163 #define IBLND_FMR_POOL 256
164 #define IBLND_FMR_POOL_FLUSH 192
166 /* TX messages (shared by all connections) */
167 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
169 /* RX messages (per connection) */
170 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
171 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
172 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
174 /* WRs and CQEs (per connection) */
175 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
176 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
177 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
181 /* o2iblnd can run over aliased interface */
183 #define KIB_IFNAME_SIZE IFALIASZ
185 #define KIB_IFNAME_SIZE 256
189 struct list_head ibd_list; /* chain on kib_devs */
190 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
191 __u32 ibd_ifip; /* IPoIB interface IP */
192 /** IPoIB interface name */
193 char ibd_ifname[KIB_IFNAME_SIZE];
194 int ibd_nnets; /* # nets extant */
196 unsigned long ibd_next_failover;
197 int ibd_failed_failover; /* # failover failures */
198 unsigned int ibd_failover; /* failover in progress */
199 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
200 struct list_head ibd_nets;
201 struct kib_hca_dev *ibd_hdev;
204 typedef struct kib_hca_dev {
205 struct rdma_cm_id *ibh_cmid; /* listener cmid */
206 struct ib_device *ibh_ibdev; /* IB device */
207 int ibh_page_shift; /* page shift of current HCA */
208 int ibh_page_size; /* page size of current HCA */
209 __u64 ibh_page_mask; /* page mask of current HCA */
210 int ibh_mr_shift; /* bits shift of max MR size */
211 __u64 ibh_mr_size; /* size of MR */
212 int ibh_nmrs; /* # of global MRs */
213 struct ib_mr **ibh_mrs; /* global MR */
214 struct ib_pd *ibh_pd; /* PD */
215 kib_dev_t *ibh_dev; /* owner */
216 atomic_t ibh_ref; /* refcount */
219 /** # of seconds to keep pool alive */
220 #define IBLND_POOL_DEADLINE 300
221 /** # of seconds to retry if allocation failed */
222 #define IBLND_POOL_RETRY 1
225 int ibp_npages; /* # pages */
226 struct page *ibp_pages[0]; /* page array */
232 struct list_head pmr_list; /* chain node */
233 struct ib_phys_buf *pmr_ipb; /* physical buffer */
234 struct ib_mr *pmr_mr; /* IB MR */
235 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
236 __u64 pmr_iova; /* Virtual I/O address */
237 int pmr_refcount; /* reference count */
243 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
244 int inc, struct kib_pool **pp_po);
245 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
246 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
247 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
251 #define IBLND_POOL_NAME_LEN 32
253 typedef struct kib_poolset {
254 spinlock_t ps_lock; /* serialize */
255 struct kib_net *ps_net; /* network it belongs to */
256 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
257 struct list_head ps_pool_list; /* list of pools */
258 struct list_head ps_failed_pool_list; /* failed pool list */
259 unsigned long ps_next_retry; /* time stamp for retry if failed to allocate */
260 int ps_increasing; /* is allocating new pool */
261 int ps_pool_size; /* new pool size */
262 int ps_cpt; /* CPT id */
264 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
265 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
266 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
267 kib_ps_node_fini_t ps_node_fini; /* finalize node */
270 typedef struct kib_pool {
271 struct list_head po_list; /* chain on pool list */
272 struct list_head po_free_list; /* pre-allocated node */
273 kib_poolset_t *po_owner; /* pool_set of this pool */
274 unsigned long po_deadline; /* deadline of this pool */
275 int po_allocated; /* # of elements in use */
276 int po_failed; /* pool is created on failed HCA */
277 int po_size; /* # of pre-allocated elements */
281 kib_poolset_t tps_poolset; /* pool-set */
282 __u64 tps_next_tx_cookie; /* cookie of TX */
286 kib_pool_t tpo_pool; /* pool */
287 struct kib_hca_dev *tpo_hdev; /* device for this pool */
288 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
289 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
293 kib_poolset_t pps_poolset; /* pool-set */
296 typedef struct kib_pmr_pool {
297 struct kib_hca_dev *ppo_hdev; /* device for this pool */
298 kib_pool_t ppo_pool; /* pool */
302 spinlock_t fps_lock; /* serialize */
303 struct kib_net *fps_net; /* IB network */
304 struct list_head fps_pool_list; /* FMR pool list */
305 struct list_head fps_failed_pool_list; /* FMR pool list */
306 __u64 fps_version; /* validity stamp */
307 int fps_cpt; /* CPT id */
309 int fps_flush_trigger;
310 /* is allocating new pool */
312 /* time stamp for retry if failed to allocate */
313 unsigned long fps_next_retry;
317 struct list_head fpo_list; /* chain on pool list */
318 struct kib_hca_dev *fpo_hdev; /* device for this pool */
319 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
320 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
321 unsigned long fpo_deadline; /* deadline of this pool */
322 int fpo_failed; /* fmr pool is failed */
323 int fpo_map_count; /* # of mapped FMR */
327 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
328 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
331 typedef struct kib_net {
332 struct list_head ibn_list; /* chain on kib_dev_t::ibd_nets */
333 __u64 ibn_incarnation; /* my epoch */
334 int ibn_init; /* initialisation state */
335 int ibn_shutdown; /* shutting down? */
337 atomic_t ibn_npeers; /* # peers extant */
338 atomic_t ibn_nconns; /* # connections extant */
340 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
341 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
342 kib_pmr_poolset_t **ibn_pmr_ps; /* pmr pool-set */
344 kib_dev_t *ibn_dev; /* underlying IB device */
347 #define KIB_THREAD_SHIFT 16
348 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
349 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
350 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
352 struct kib_sched_info {
355 /* schedulers sleep here */
356 wait_queue_head_t ibs_waitq;
357 /* conns to check for rx completions */
358 struct list_head ibs_conns;
359 /* number of scheduler threads */
361 /* max allowed scheduler threads */
362 int ibs_nthreads_max;
363 int ibs_cpt; /* CPT id */
367 int kib_init; /* initialisation state */
368 int kib_shutdown; /* shut down? */
369 struct list_head kib_devs; /* IB devices extant */
370 /* list head of failed devices */
371 struct list_head kib_failed_devs;
372 /* schedulers sleep here */
373 wait_queue_head_t kib_failover_waitq;
374 atomic_t kib_nthreads; /* # live threads */
375 /* stabilize net/dev/peer/conn ops */
376 rwlock_t kib_global_lock;
377 /* hash table of all my known peers */
378 struct list_head *kib_peers;
379 /* size of kib_peers */
380 int kib_peer_hash_size;
381 /* the connd task (serialisation assertions) */
383 /* connections to setup/teardown */
384 struct list_head kib_connd_conns;
385 /* connections with zero refcount */
386 struct list_head kib_connd_zombies;
387 /* connection daemon sleeps here */
388 wait_queue_head_t kib_connd_waitq;
389 spinlock_t kib_connd_lock; /* serialise */
390 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
391 /* percpt data for schedulers */
392 struct kib_sched_info **kib_scheds;
395 #define IBLND_INIT_NOTHING 0
396 #define IBLND_INIT_DATA 1
397 #define IBLND_INIT_ALL 2
399 /************************************************************************
400 * IB Wire message format.
401 * These are sent in sender's byte order (i.e. receiver flips).
404 typedef struct kib_connparams {
405 __u16 ibcp_queue_depth;
406 __u16 ibcp_max_frags;
407 __u32 ibcp_max_msg_size;
408 } WIRE_ATTR kib_connparams_t;
411 lnet_hdr_t ibim_hdr; /* portals header */
412 char ibim_payload[0]; /* piggy-backed payload */
413 } WIRE_ATTR kib_immediate_msg_t;
416 __u32 rf_nob; /* # bytes this frag */
417 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
418 } WIRE_ATTR kib_rdma_frag_t;
421 __u32 rd_key; /* local/remote key */
422 __u32 rd_nfrags; /* # fragments */
423 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
424 } WIRE_ATTR kib_rdma_desc_t;
427 lnet_hdr_t ibprm_hdr; /* portals header */
428 __u64 ibprm_cookie; /* opaque completion cookie */
429 } WIRE_ATTR kib_putreq_msg_t;
432 __u64 ibpam_src_cookie; /* reflected completion cookie */
433 __u64 ibpam_dst_cookie; /* opaque completion cookie */
434 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
435 } WIRE_ATTR kib_putack_msg_t;
438 lnet_hdr_t ibgm_hdr; /* portals header */
439 __u64 ibgm_cookie; /* opaque completion cookie */
440 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
441 } WIRE_ATTR kib_get_msg_t;
444 __u64 ibcm_cookie; /* opaque completion cookie */
445 __s32 ibcm_status; /* < 0 failure: >= 0 length */
446 } WIRE_ATTR kib_completion_msg_t;
449 /* First 2 fields fixed FOR ALL TIME */
450 __u32 ibm_magic; /* I'm an ibnal message */
451 __u16 ibm_version; /* this is my version number */
453 __u8 ibm_type; /* msg type */
454 __u8 ibm_credits; /* returned credits */
455 __u32 ibm_nob; /* # bytes in whole message */
456 __u32 ibm_cksum; /* checksum (0 == no checksum) */
457 __u64 ibm_srcnid; /* sender's NID */
458 __u64 ibm_srcstamp; /* sender's incarnation */
459 __u64 ibm_dstnid; /* destination's NID */
460 __u64 ibm_dststamp; /* destination's incarnation */
463 kib_connparams_t connparams;
464 kib_immediate_msg_t immediate;
465 kib_putreq_msg_t putreq;
466 kib_putack_msg_t putack;
468 kib_completion_msg_t completion;
470 } WIRE_ATTR kib_msg_t;
472 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
474 #define IBLND_MSG_VERSION_1 0x11
475 #define IBLND_MSG_VERSION_2 0x12
476 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
478 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
479 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
480 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
481 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
482 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
483 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
484 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
485 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
486 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
487 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
490 __u32 ibr_magic; /* sender's magic */
491 __u16 ibr_version; /* sender's version */
492 __u8 ibr_why; /* reject reason */
493 __u8 ibr_padding; /* padding */
494 __u64 ibr_incarnation; /* incarnation of peer */
495 kib_connparams_t ibr_cp; /* connection parameters */
496 } WIRE_ATTR kib_rej_t;
498 /* connection rejection reasons */
499 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
500 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
501 #define IBLND_REJECT_FATAL 3 /* Anything else */
503 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
504 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
506 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
507 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
509 /***********************************************************************/
511 typedef struct kib_rx /* receive message */
513 struct list_head rx_list; /* queue for attention */
514 struct kib_conn *rx_conn; /* owning conn */
515 int rx_nob; /* # bytes received (-1 while posted) */
516 enum ib_wc_status rx_status; /* completion status */
517 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
518 __u64 rx_msgaddr; /* message buffer (I/O addr) */
519 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
520 struct ib_recv_wr rx_wrq; /* receive work item... */
521 struct ib_sge rx_sge; /* ...and its memory */
524 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
525 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
526 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
527 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
529 typedef struct kib_tx /* transmit message */
531 struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */
532 kib_tx_pool_t *tx_pool; /* pool I'm from */
533 struct kib_conn *tx_conn; /* owning conn */
534 short tx_sending; /* # tx callbacks outstanding */
535 short tx_queued; /* queued for sending */
536 short tx_waiting; /* waiting for peer */
537 int tx_status; /* LNET completion status */
538 unsigned long tx_deadline; /* completion deadline */
539 __u64 tx_cookie; /* completion cookie */
540 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
541 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
542 __u64 tx_msgaddr; /* message buffer (I/O addr) */
543 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
544 int tx_nwrq; /* # send work items */
545 struct ib_send_wr *tx_wrq; /* send work items... */
546 struct ib_sge *tx_sge; /* ...and their memory */
547 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
548 int tx_nfrags; /* # entries in... */
549 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
550 __u64 *tx_pages; /* rdma phys page addrs */
552 kib_phys_mr_t *pmr; /* MR for physical buffer */
553 kib_fmr_t fmr; /* FMR */
555 int tx_dmadir; /* dma direction */
558 typedef struct kib_connvars {
559 /* connection-in-progress variables */
563 typedef struct kib_conn {
564 struct kib_sched_info *ibc_sched; /* scheduler information */
565 struct kib_peer *ibc_peer; /* owning peer */
566 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
567 struct list_head ibc_list; /* stash on peer's conn list */
568 struct list_head ibc_sched_list; /* schedule for attention */
569 __u16 ibc_version; /* version of connection */
570 __u64 ibc_incarnation; /* which instance of the peer */
571 atomic_t ibc_refcount; /* # users */
572 int ibc_state; /* what's happening */
573 int ibc_nsends_posted; /* # uncompleted sends */
574 int ibc_noops_posted; /* # uncompleted NOOPs */
575 int ibc_credits; /* # credits I have */
576 int ibc_outstanding_credits; /* # credits to return */
577 int ibc_reserved_credits;/* # ACK/DONE msg credits */
578 int ibc_comms_error; /* set on comms error */
579 unsigned int ibc_nrx:16; /* receive buffers owned */
580 unsigned int ibc_scheduled:1; /* scheduled for attention */
581 unsigned int ibc_ready:1; /* CQ callback fired */
582 /* time of last send */
583 unsigned long ibc_last_send;
584 /** link chain for kiblnd_check_conns only */
585 struct list_head ibc_connd_list;
586 /** rxs completed before ESTABLISHED */
587 struct list_head ibc_early_rxs;
588 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
589 struct list_head ibc_tx_noops;
590 struct list_head ibc_tx_queue; /* sends that need a credit */
591 struct list_head ibc_tx_queue_nocred;/* sends that don't need a credit */
592 struct list_head ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
593 struct list_head ibc_active_txs; /* active tx awaiting completion */
594 spinlock_t ibc_lock; /* serialise */
595 kib_rx_t *ibc_rxs; /* the rx descs */
596 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
598 struct rdma_cm_id *ibc_cmid; /* CM id */
599 struct ib_cq *ibc_cq; /* completion queue */
601 kib_connvars_t *ibc_connvars; /* in-progress connection state */
604 #define IBLND_CONN_INIT 0 /* being initialised */
605 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
606 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
607 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
608 #define IBLND_CONN_CLOSING 4 /* being closed */
609 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
611 typedef struct kib_peer {
612 struct list_head ibp_list; /* stash on global peer list */
613 lnet_nid_t ibp_nid; /* who's on the other end(s) */
614 lnet_ni_t *ibp_ni; /* LNet interface */
615 atomic_t ibp_refcount; /* # users */
616 struct list_head ibp_conns; /* all active connections */
617 struct list_head ibp_tx_queue; /* msgs waiting for a conn */
618 __u16 ibp_version; /* version of peer */
619 __u64 ibp_incarnation; /* incarnation of peer */
620 int ibp_connecting; /* current active connection attempts */
621 int ibp_accepting; /* current passive connection attempts */
622 int ibp_error; /* errno on closing this peer */
623 unsigned long ibp_last_alive; /* when (in jiffies) I was last alive */
626 extern kib_data_t kiblnd_data;
628 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
631 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
633 LASSERT (atomic_read(&hdev->ibh_ref) > 0);
634 atomic_inc(&hdev->ibh_ref);
638 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
640 LASSERT (atomic_read(&hdev->ibh_ref) > 0);
641 if (atomic_dec_and_test(&hdev->ibh_ref))
642 kiblnd_hdev_destroy(hdev);
646 kiblnd_dev_can_failover(kib_dev_t *dev)
648 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
651 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
654 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
657 return dev->ibd_can_failover;
660 #define kiblnd_conn_addref(conn) \
662 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
663 (conn), atomic_read(&(conn)->ibc_refcount)); \
664 atomic_inc(&(conn)->ibc_refcount); \
667 #define kiblnd_conn_decref(conn) \
669 unsigned long flags; \
671 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
672 (conn), atomic_read(&(conn)->ibc_refcount)); \
673 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
674 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
675 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
676 list_add_tail(&(conn)->ibc_list, \
677 &kiblnd_data.kib_connd_zombies); \
678 wake_up(&kiblnd_data.kib_connd_waitq); \
679 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
683 #define kiblnd_peer_addref(peer) \
685 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
686 (peer), libcfs_nid2str((peer)->ibp_nid), \
687 atomic_read (&(peer)->ibp_refcount)); \
688 atomic_inc(&(peer)->ibp_refcount); \
691 #define kiblnd_peer_decref(peer) \
693 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
694 (peer), libcfs_nid2str((peer)->ibp_nid), \
695 atomic_read (&(peer)->ibp_refcount)); \
696 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
697 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
698 kiblnd_destroy_peer(peer); \
701 static inline struct list_head *
702 kiblnd_nid2peerlist (lnet_nid_t nid)
705 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
707 return (&kiblnd_data.kib_peers [hash]);
711 kiblnd_peer_active (kib_peer_t *peer)
713 /* Am I in the peer hash table? */
714 return (!list_empty(&peer->ibp_list));
717 static inline kib_conn_t *
718 kiblnd_get_conn_locked (kib_peer_t *peer)
720 LASSERT (!list_empty(&peer->ibp_conns));
722 /* just return the first connection */
723 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
727 kiblnd_send_keepalive(kib_conn_t *conn)
729 return (*kiblnd_tunables.kib_keepalive > 0) &&
730 cfs_time_after(jiffies, conn->ibc_last_send +
731 *kiblnd_tunables.kib_keepalive*HZ);
735 kiblnd_need_noop(kib_conn_t *conn)
737 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
739 if (conn->ibc_outstanding_credits <
740 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
741 !kiblnd_send_keepalive(conn))
742 return 0; /* No need to send NOOP */
744 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
745 if (!list_empty(&conn->ibc_tx_queue_nocred))
746 return 0; /* NOOP can be piggybacked */
748 /* No tx to piggyback NOOP onto or no credit to send a tx */
749 return (list_empty(&conn->ibc_tx_queue) ||
750 conn->ibc_credits == 0);
753 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
754 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
755 conn->ibc_credits == 0) /* no credit */
758 if (conn->ibc_credits == 1 && /* last credit reserved for */
759 conn->ibc_outstanding_credits == 0) /* giving back credits */
762 /* No tx to piggyback NOOP onto or no credit to send a tx */
763 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
767 kiblnd_abort_receives(kib_conn_t *conn)
769 ib_modify_qp(conn->ibc_cmid->qp,
770 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
773 static inline const char *
774 kiblnd_queue2str (kib_conn_t *conn, struct list_head *q)
776 if (q == &conn->ibc_tx_queue)
779 if (q == &conn->ibc_tx_queue_rsrvd)
780 return "tx_queue_rsrvd";
782 if (q == &conn->ibc_tx_queue_nocred)
783 return "tx_queue_nocred";
785 if (q == &conn->ibc_active_txs)
792 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
793 * lowest bits of the work request id to stash the work item type. */
795 #define IBLND_WID_TX 0
796 #define IBLND_WID_RDMA 1
797 #define IBLND_WID_RX 2
798 #define IBLND_WID_MASK 3UL
801 kiblnd_ptr2wreqid (void *ptr, int type)
803 unsigned long lptr = (unsigned long)ptr;
805 LASSERT ((lptr & IBLND_WID_MASK) == 0);
806 LASSERT ((type & ~IBLND_WID_MASK) == 0);
807 return (__u64)(lptr | type);
811 kiblnd_wreqid2ptr (__u64 wreqid)
813 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
817 kiblnd_wreqid2type (__u64 wreqid)
819 return (wreqid & IBLND_WID_MASK);
823 kiblnd_set_conn_state (kib_conn_t *conn, int state)
825 conn->ibc_state = state;
830 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
832 msg->ibm_type = type;
833 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
837 kiblnd_rd_size (kib_rdma_desc_t *rd)
842 for (i = size = 0; i < rd->rd_nfrags; i++)
843 size += rd->rd_frags[i].rf_nob;
849 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
851 return rd->rd_frags[index].rf_addr;
855 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
857 return rd->rd_frags[index].rf_nob;
861 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
867 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
869 if (nob < rd->rd_frags[index].rf_nob) {
870 rd->rd_frags[index].rf_addr += nob;
871 rd->rd_frags[index].rf_nob -= nob;
880 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
882 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
883 msgtype == IBLND_MSG_PUT_ACK);
885 return msgtype == IBLND_MSG_GET_REQ ?
886 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
887 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
892 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
894 return ib_dma_mapping_error(dev, dma_addr);
897 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
898 void *msg, size_t size,
899 enum dma_data_direction direction)
901 return ib_dma_map_single(dev, msg, size, direction);
904 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
905 __u64 addr, size_t size,
906 enum dma_data_direction direction)
908 ib_dma_unmap_single(dev, addr, size, direction);
911 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
912 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
914 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
915 struct scatterlist *sg, int nents,
916 enum dma_data_direction direction)
918 return ib_dma_map_sg(dev, sg, nents, direction);
921 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
922 struct scatterlist *sg, int nents,
923 enum dma_data_direction direction)
925 ib_dma_unmap_sg(dev, sg, nents, direction);
928 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
929 struct scatterlist *sg)
931 return ib_sg_dma_address(dev, sg);
934 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
935 struct scatterlist *sg)
937 return ib_sg_dma_len(dev, sg);
940 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
941 * right because OFED1.2 defines it as const, to use it we have to add
942 * (void *) cast to overcome "const" */
944 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
945 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
948 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
949 kib_rdma_desc_t *rd);
950 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
951 __u64 addr, __u64 size);
952 void kiblnd_map_rx_descs(kib_conn_t *conn);
953 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
954 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
955 kib_rdma_desc_t *rd, int nfrags);
956 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
957 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
958 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
960 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
961 int npages, __u64 iov, kib_fmr_t *fmr);
962 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
964 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
965 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
966 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
968 int kiblnd_startup (lnet_ni_t *ni);
969 void kiblnd_shutdown (lnet_ni_t *ni);
970 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
971 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, unsigned long *when);
973 int kiblnd_tunables_init(void);
974 void kiblnd_tunables_fini(void);
976 int kiblnd_connd (void *arg);
977 int kiblnd_scheduler(void *arg);
978 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
979 int kiblnd_failover_thread (void *arg);
981 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
982 void kiblnd_free_pages (kib_pages_t *p);
984 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
985 struct rdma_cm_event *event);
986 int kiblnd_translate_mtu(int value);
988 int kiblnd_dev_failover(kib_dev_t *dev);
989 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
990 void kiblnd_destroy_peer (kib_peer_t *peer);
991 void kiblnd_destroy_dev (kib_dev_t *dev);
992 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
993 void kiblnd_peer_alive (kib_peer_t *peer);
994 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
995 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
996 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
997 int version, __u64 incarnation);
998 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1000 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1001 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1002 int state, int version);
1003 void kiblnd_destroy_conn (kib_conn_t *conn);
1004 void kiblnd_close_conn (kib_conn_t *conn, int error);
1005 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1007 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1008 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1010 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1011 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1012 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1013 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1014 void kiblnd_txlist_done (lnet_ni_t *ni, struct list_head *txlist,
1016 void kiblnd_check_sends (kib_conn_t *conn);
1018 void kiblnd_qp_event(struct ib_event *event, void *arg);
1019 void kiblnd_cq_event(struct ib_event *event, void *arg);
1020 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1022 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1023 int credits, lnet_nid_t dstnid, __u64 dststamp);
1024 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1025 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1027 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1028 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1029 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1030 unsigned int offset, unsigned int mlen, unsigned int rlen);