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>
50 #include <asm/uaccess.h>
53 #include <linux/init.h>
55 #include <linux/file.h>
56 #include <linux/list.h>
57 #include <linux/kmod.h>
58 #include <linux/sysctl.h>
59 #include <linux/pci.h>
64 #define DEBUG_SUBSYSTEM S_LND
66 #include <linux/libcfs/libcfs.h>
67 #include <linux/lnet/lnet.h>
68 #include <linux/lnet/lib-lnet.h>
69 #include <linux/lnet/lnet-sysctl.h>
71 #include <rdma/rdma_cm.h>
72 #include <rdma/ib_cm.h>
73 #include <rdma/ib_verbs.h>
74 #include <rdma/ib_fmr_pool.h>
76 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
77 /* # scheduler loops before reschedule */
78 #define IBLND_RESCHED 100
80 #define IBLND_N_SCHED 2
81 #define IBLND_N_SCHED_HIGH 4
85 int *kib_dev_failover; /* HCA failover */
86 unsigned int *kib_service; /* IB service number */
87 int *kib_min_reconnect_interval; /* first failed connection retry... */
88 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
89 int *kib_cksum; /* checksum kib_msg_t? */
90 int *kib_timeout; /* comms timeout (seconds) */
91 int *kib_keepalive; /* keepalive timeout (seconds) */
92 int *kib_ntx; /* # tx descs */
93 int *kib_credits; /* # concurrent sends */
94 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
95 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
96 int *kib_peercredits_hiw; /* # when eagerly to return credits */
97 int *kib_peertimeout; /* seconds to consider peer dead */
98 char **kib_default_ipif; /* default IPoIB interface */
100 int *kib_rnr_retry_count;
101 int *kib_concurrent_sends; /* send work queue sizing */
102 int *kib_ib_mtu; /* IB MTU */
103 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
104 * than this value, 0 disable map-on-demand */
105 int *kib_pmr_pool_size; /* # physical MR in pool */
106 int *kib_fmr_pool_size; /* # FMRs in pool */
107 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
108 int *kib_fmr_cache; /* enable FMR pool cache? */
109 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
110 ctl_table_header_t *kib_sysctl; /* sysctl interface */
112 int *kib_require_priv_port;/* accept only privileged ports */
113 int *kib_use_priv_port; /* use privileged port for active connect */
114 /* # threads on each CPT */
118 extern kib_tunables_t kiblnd_tunables;
120 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
121 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
123 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
124 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
126 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
127 IBLND_MSG_QUEUE_SIZE_V1 : \
128 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
129 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
130 IBLND_CREDIT_HIGHWATER_V1 : \
131 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
133 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
136 kiblnd_concurrent_sends_v1(void)
138 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
139 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
141 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
142 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
144 return *kiblnd_tunables.kib_concurrent_sends;
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)
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)
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_PMR_POOL 256
168 #define IBLND_FMR_POOL 256
169 #define IBLND_FMR_POOL_FLUSH 192
171 /* TX messages (shared by all connections) */
172 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
174 /* RX messages (per connection) */
175 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
176 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
177 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
179 /* WRs and CQEs (per connection) */
180 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
181 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
182 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
186 /* o2iblnd can run over aliased interface */
188 #define KIB_IFNAME_SIZE IFALIASZ
190 #define KIB_IFNAME_SIZE 256
195 struct list_head ibd_list; /* chain on kib_devs */
196 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
197 __u32 ibd_ifip; /* IPoIB interface IP */
198 /** IPoIB interface name */
199 char ibd_ifname[KIB_IFNAME_SIZE];
200 int ibd_nnets; /* # nets extant */
202 cfs_time_t ibd_next_failover;
203 int ibd_failed_failover; /* # failover failures */
204 unsigned int ibd_failover; /* failover in progress */
205 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
206 struct list_head ibd_nets;
207 struct kib_hca_dev *ibd_hdev;
210 typedef struct kib_hca_dev
212 struct rdma_cm_id *ibh_cmid; /* listener cmid */
213 struct ib_device *ibh_ibdev; /* IB device */
214 int ibh_page_shift; /* page shift of current HCA */
215 int ibh_page_size; /* page size of current HCA */
216 __u64 ibh_page_mask; /* page mask of current HCA */
217 int ibh_mr_shift; /* bits shift of max MR size */
218 __u64 ibh_mr_size; /* size of MR */
219 int ibh_nmrs; /* # of global MRs */
220 struct ib_mr **ibh_mrs; /* global MR */
221 struct ib_pd *ibh_pd; /* PD */
222 kib_dev_t *ibh_dev; /* owner */
223 atomic_t ibh_ref; /* refcount */
226 /** # of seconds to keep pool alive */
227 #define IBLND_POOL_DEADLINE 300
228 /** # of seconds to retry if allocation failed */
229 #define IBLND_POOL_RETRY 1
233 int ibp_npages; /* # pages */
234 struct page *ibp_pages[0]; /* page array */
240 struct list_head pmr_list; /* chain node */
241 struct ib_phys_buf *pmr_ipb; /* physical buffer */
242 struct ib_mr *pmr_mr; /* IB MR */
243 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
244 __u64 pmr_iova; /* Virtual I/O address */
245 int pmr_refcount; /* reference count */
251 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
252 int inc, struct kib_pool **pp_po);
253 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
254 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
255 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
259 #define IBLND_POOL_NAME_LEN 32
261 typedef struct kib_poolset
263 spinlock_t ps_lock; /* serialize */
264 struct kib_net *ps_net; /* network it belongs to */
265 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
266 struct list_head ps_pool_list; /* list of pools */
267 struct list_head ps_failed_pool_list; /* failed pool list */
268 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
269 int ps_increasing; /* is allocating new pool */
270 int ps_pool_size; /* new pool size */
271 int ps_cpt; /* CPT id */
273 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
274 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
275 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
276 kib_ps_node_fini_t ps_node_fini; /* finalize node */
279 typedef struct kib_pool
281 struct list_head po_list; /* chain on pool list */
282 struct list_head po_free_list; /* pre-allocated node */
283 kib_poolset_t *po_owner; /* pool_set of this pool */
284 cfs_time_t po_deadline; /* deadline of this pool */
285 int po_allocated; /* # of elements in use */
286 int po_failed; /* pool is created on failed HCA */
287 int po_size; /* # of pre-allocated elements */
291 kib_poolset_t tps_poolset; /* pool-set */
292 __u64 tps_next_tx_cookie; /* cookie of TX */
296 kib_pool_t tpo_pool; /* pool */
297 struct kib_hca_dev *tpo_hdev; /* device for this pool */
298 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
299 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
303 kib_poolset_t pps_poolset; /* pool-set */
306 typedef struct kib_pmr_pool {
307 struct kib_hca_dev *ppo_hdev; /* device for this pool */
308 kib_pool_t ppo_pool; /* pool */
313 spinlock_t fps_lock; /* serialize */
314 struct kib_net *fps_net; /* IB network */
315 struct list_head fps_pool_list; /* FMR pool list */
316 struct list_head fps_failed_pool_list; /* FMR pool list */
317 __u64 fps_version; /* validity stamp */
318 int fps_cpt; /* CPT id */
320 int fps_flush_trigger;
321 /* is allocating new pool */
323 /* time stamp for retry if failed to allocate */
324 cfs_time_t fps_next_retry;
329 struct list_head fpo_list; /* chain on pool list */
330 struct kib_hca_dev *fpo_hdev; /* device for this pool */
331 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
332 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
333 cfs_time_t fpo_deadline; /* deadline of this pool */
334 int fpo_failed; /* fmr pool is failed */
335 int fpo_map_count; /* # of mapped FMR */
339 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
340 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
343 typedef struct kib_net
345 struct list_head ibn_list; /* chain on kib_dev_t::ibd_nets */
346 __u64 ibn_incarnation; /* my epoch */
347 int ibn_init; /* initialisation state */
348 int ibn_shutdown; /* shutting down? */
350 atomic_t ibn_npeers; /* # peers extant */
351 atomic_t ibn_nconns; /* # connections extant */
353 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
354 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
355 kib_pmr_poolset_t **ibn_pmr_ps; /* pmr pool-set */
357 kib_dev_t *ibn_dev; /* underlying IB device */
360 #define KIB_THREAD_SHIFT 16
361 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
362 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
363 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
365 struct kib_sched_info {
368 /* schedulers sleep here */
369 wait_queue_head_t ibs_waitq;
370 /* conns to check for rx completions */
371 struct list_head ibs_conns;
372 /* number of scheduler threads */
374 /* max allowed scheduler threads */
375 int ibs_nthreads_max;
376 int ibs_cpt; /* CPT id */
381 int kib_init; /* initialisation state */
382 int kib_shutdown; /* shut down? */
383 struct list_head kib_devs; /* IB devices extant */
384 /* list head of failed devices */
385 struct list_head kib_failed_devs;
386 /* schedulers sleep here */
387 wait_queue_head_t kib_failover_waitq;
388 atomic_t kib_nthreads; /* # live threads */
389 /* stabilize net/dev/peer/conn ops */
390 rwlock_t kib_global_lock;
391 /* hash table of all my known peers */
392 struct list_head *kib_peers;
393 /* size of kib_peers */
394 int kib_peer_hash_size;
395 /* the connd task (serialisation assertions) */
397 /* connections to setup/teardown */
398 struct list_head kib_connd_conns;
399 /* connections with zero refcount */
400 struct list_head kib_connd_zombies;
401 /* connection daemon sleeps here */
402 wait_queue_head_t kib_connd_waitq;
403 spinlock_t kib_connd_lock; /* serialise */
404 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
405 /* percpt data for schedulers */
406 struct kib_sched_info **kib_scheds;
409 #define IBLND_INIT_NOTHING 0
410 #define IBLND_INIT_DATA 1
411 #define IBLND_INIT_ALL 2
413 /************************************************************************
414 * IB Wire message format.
415 * These are sent in sender's byte order (i.e. receiver flips).
418 typedef struct kib_connparams
420 __u16 ibcp_queue_depth;
421 __u16 ibcp_max_frags;
422 __u32 ibcp_max_msg_size;
423 } WIRE_ATTR kib_connparams_t;
427 lnet_hdr_t ibim_hdr; /* portals header */
428 char ibim_payload[0]; /* piggy-backed payload */
429 } WIRE_ATTR kib_immediate_msg_t;
433 __u32 rf_nob; /* # bytes this frag */
434 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
435 } WIRE_ATTR kib_rdma_frag_t;
439 __u32 rd_key; /* local/remote key */
440 __u32 rd_nfrags; /* # fragments */
441 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
442 } WIRE_ATTR kib_rdma_desc_t;
446 lnet_hdr_t ibprm_hdr; /* portals header */
447 __u64 ibprm_cookie; /* opaque completion cookie */
448 } WIRE_ATTR kib_putreq_msg_t;
452 __u64 ibpam_src_cookie; /* reflected completion cookie */
453 __u64 ibpam_dst_cookie; /* opaque completion cookie */
454 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
455 } WIRE_ATTR kib_putack_msg_t;
459 lnet_hdr_t ibgm_hdr; /* portals header */
460 __u64 ibgm_cookie; /* opaque completion cookie */
461 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
462 } WIRE_ATTR kib_get_msg_t;
466 __u64 ibcm_cookie; /* opaque completion cookie */
467 __s32 ibcm_status; /* < 0 failure: >= 0 length */
468 } WIRE_ATTR kib_completion_msg_t;
472 /* First 2 fields fixed FOR ALL TIME */
473 __u32 ibm_magic; /* I'm an ibnal message */
474 __u16 ibm_version; /* this is my version number */
476 __u8 ibm_type; /* msg type */
477 __u8 ibm_credits; /* returned credits */
478 __u32 ibm_nob; /* # bytes in whole message */
479 __u32 ibm_cksum; /* checksum (0 == no checksum) */
480 __u64 ibm_srcnid; /* sender's NID */
481 __u64 ibm_srcstamp; /* sender's incarnation */
482 __u64 ibm_dstnid; /* destination's NID */
483 __u64 ibm_dststamp; /* destination's incarnation */
486 kib_connparams_t connparams;
487 kib_immediate_msg_t immediate;
488 kib_putreq_msg_t putreq;
489 kib_putack_msg_t putack;
491 kib_completion_msg_t completion;
493 } WIRE_ATTR kib_msg_t;
495 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
497 #define IBLND_MSG_VERSION_1 0x11
498 #define IBLND_MSG_VERSION_2 0x12
499 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
501 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
502 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
503 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
504 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
505 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
506 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
507 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
508 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
509 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
510 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
513 __u32 ibr_magic; /* sender's magic */
514 __u16 ibr_version; /* sender's version */
515 __u8 ibr_why; /* reject reason */
516 __u8 ibr_padding; /* padding */
517 __u64 ibr_incarnation; /* incarnation of peer */
518 kib_connparams_t ibr_cp; /* connection parameters */
519 } WIRE_ATTR kib_rej_t;
521 /* connection rejection reasons */
522 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
523 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
524 #define IBLND_REJECT_FATAL 3 /* Anything else */
526 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
527 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
529 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
530 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
532 /***********************************************************************/
534 typedef struct kib_rx /* receive message */
536 struct list_head rx_list; /* queue for attention */
537 struct kib_conn *rx_conn; /* owning conn */
538 int rx_nob; /* # bytes received (-1 while posted) */
539 enum ib_wc_status rx_status; /* completion status */
540 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
541 __u64 rx_msgaddr; /* message buffer (I/O addr) */
542 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
543 struct ib_recv_wr rx_wrq; /* receive work item... */
544 struct ib_sge rx_sge; /* ...and its memory */
547 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
548 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
549 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
550 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
552 typedef struct kib_tx /* transmit message */
554 struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */
555 kib_tx_pool_t *tx_pool; /* pool I'm from */
556 struct kib_conn *tx_conn; /* owning conn */
557 short tx_sending; /* # tx callbacks outstanding */
558 short tx_queued; /* queued for sending */
559 short tx_waiting; /* waiting for peer */
560 int tx_status; /* LNET completion status */
561 unsigned long tx_deadline; /* completion deadline */
562 __u64 tx_cookie; /* completion cookie */
563 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
564 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
565 __u64 tx_msgaddr; /* message buffer (I/O addr) */
566 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
567 int tx_nwrq; /* # send work items */
568 struct ib_send_wr *tx_wrq; /* send work items... */
569 struct ib_sge *tx_sge; /* ...and their memory */
570 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
571 int tx_nfrags; /* # entries in... */
572 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
573 __u64 *tx_pages; /* rdma phys page addrs */
575 kib_phys_mr_t *pmr; /* MR for physical buffer */
576 kib_fmr_t fmr; /* FMR */
578 int tx_dmadir; /* dma direction */
581 typedef struct kib_connvars
583 /* connection-in-progress variables */
587 typedef struct kib_conn
589 struct kib_sched_info *ibc_sched; /* scheduler information */
590 struct kib_peer *ibc_peer; /* owning peer */
591 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
592 struct list_head ibc_list; /* stash on peer's conn list */
593 struct list_head ibc_sched_list; /* schedule for attention */
594 __u16 ibc_version; /* version of connection */
595 __u64 ibc_incarnation; /* which instance of the peer */
596 atomic_t ibc_refcount; /* # users */
597 int ibc_state; /* what's happening */
598 int ibc_nsends_posted; /* # uncompleted sends */
599 int ibc_noops_posted; /* # uncompleted NOOPs */
600 int ibc_credits; /* # credits I have */
601 int ibc_outstanding_credits; /* # credits to return */
602 int ibc_reserved_credits;/* # ACK/DONE msg credits */
603 int ibc_comms_error; /* set on comms error */
604 unsigned int ibc_nrx:16; /* receive buffers owned */
605 unsigned int ibc_scheduled:1; /* scheduled for attention */
606 unsigned int ibc_ready:1; /* CQ callback fired */
607 /* time of last send */
608 unsigned long ibc_last_send;
609 /** link chain for kiblnd_check_conns only */
610 struct list_head ibc_connd_list;
611 /** rxs completed before ESTABLISHED */
612 struct list_head ibc_early_rxs;
613 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
614 struct list_head ibc_tx_noops;
615 struct list_head ibc_tx_queue; /* sends that need a credit */
616 struct list_head ibc_tx_queue_nocred;/* sends that don't need a credit */
617 struct list_head ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
618 struct list_head ibc_active_txs; /* active tx awaiting completion */
619 spinlock_t ibc_lock; /* serialise */
620 kib_rx_t *ibc_rxs; /* the rx descs */
621 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
623 struct rdma_cm_id *ibc_cmid; /* CM id */
624 struct ib_cq *ibc_cq; /* completion queue */
626 kib_connvars_t *ibc_connvars; /* in-progress connection state */
629 #define IBLND_CONN_INIT 0 /* being initialised */
630 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
631 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
632 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
633 #define IBLND_CONN_CLOSING 4 /* being closed */
634 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
636 typedef struct kib_peer
638 struct list_head ibp_list; /* stash on global peer list */
639 lnet_nid_t ibp_nid; /* who's on the other end(s) */
640 lnet_ni_t *ibp_ni; /* LNet interface */
641 atomic_t ibp_refcount; /* # users */
642 struct list_head ibp_conns; /* all active connections */
643 struct list_head ibp_tx_queue; /* msgs waiting for a conn */
644 __u16 ibp_version; /* version of peer */
645 __u64 ibp_incarnation; /* incarnation of peer */
646 int ibp_connecting; /* current active connection attempts */
647 int ibp_accepting; /* current passive connection attempts */
648 int ibp_error; /* errno on closing this peer */
649 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
652 extern kib_data_t kiblnd_data;
654 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
657 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
659 LASSERT (atomic_read(&hdev->ibh_ref) > 0);
660 atomic_inc(&hdev->ibh_ref);
664 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
666 LASSERT (atomic_read(&hdev->ibh_ref) > 0);
667 if (atomic_dec_and_test(&hdev->ibh_ref))
668 kiblnd_hdev_destroy(hdev);
672 kiblnd_dev_can_failover(kib_dev_t *dev)
674 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
677 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
680 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
683 return dev->ibd_can_failover;
686 #define kiblnd_conn_addref(conn) \
688 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
689 (conn), atomic_read(&(conn)->ibc_refcount)); \
690 atomic_inc(&(conn)->ibc_refcount); \
693 #define kiblnd_conn_decref(conn) \
695 unsigned long flags; \
697 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
698 (conn), atomic_read(&(conn)->ibc_refcount)); \
699 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
700 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
701 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
702 list_add_tail(&(conn)->ibc_list, \
703 &kiblnd_data.kib_connd_zombies); \
704 wake_up(&kiblnd_data.kib_connd_waitq); \
705 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
709 #define kiblnd_peer_addref(peer) \
711 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
712 (peer), libcfs_nid2str((peer)->ibp_nid), \
713 atomic_read (&(peer)->ibp_refcount)); \
714 atomic_inc(&(peer)->ibp_refcount); \
717 #define kiblnd_peer_decref(peer) \
719 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
720 (peer), libcfs_nid2str((peer)->ibp_nid), \
721 atomic_read (&(peer)->ibp_refcount)); \
722 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
723 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
724 kiblnd_destroy_peer(peer); \
727 static inline struct list_head *
728 kiblnd_nid2peerlist (lnet_nid_t nid)
731 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
733 return (&kiblnd_data.kib_peers [hash]);
737 kiblnd_peer_active (kib_peer_t *peer)
739 /* Am I in the peer hash table? */
740 return (!list_empty(&peer->ibp_list));
743 static inline kib_conn_t *
744 kiblnd_get_conn_locked (kib_peer_t *peer)
746 LASSERT (!list_empty(&peer->ibp_conns));
748 /* just return the first connection */
749 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
753 kiblnd_send_keepalive(kib_conn_t *conn)
755 return (*kiblnd_tunables.kib_keepalive > 0) &&
756 cfs_time_after(jiffies, conn->ibc_last_send +
757 *kiblnd_tunables.kib_keepalive*HZ);
761 kiblnd_need_noop(kib_conn_t *conn)
763 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
765 if (conn->ibc_outstanding_credits <
766 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
767 !kiblnd_send_keepalive(conn))
768 return 0; /* No need to send NOOP */
770 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
771 if (!list_empty(&conn->ibc_tx_queue_nocred))
772 return 0; /* NOOP can be piggybacked */
774 /* No tx to piggyback NOOP onto or no credit to send a tx */
775 return (list_empty(&conn->ibc_tx_queue) ||
776 conn->ibc_credits == 0);
779 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
780 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
781 conn->ibc_credits == 0) /* no credit */
784 if (conn->ibc_credits == 1 && /* last credit reserved for */
785 conn->ibc_outstanding_credits == 0) /* giving back credits */
788 /* No tx to piggyback NOOP onto or no credit to send a tx */
789 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
793 kiblnd_abort_receives(kib_conn_t *conn)
795 ib_modify_qp(conn->ibc_cmid->qp,
796 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
799 static inline const char *
800 kiblnd_queue2str (kib_conn_t *conn, struct list_head *q)
802 if (q == &conn->ibc_tx_queue)
805 if (q == &conn->ibc_tx_queue_rsrvd)
806 return "tx_queue_rsrvd";
808 if (q == &conn->ibc_tx_queue_nocred)
809 return "tx_queue_nocred";
811 if (q == &conn->ibc_active_txs)
818 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
819 * lowest bits of the work request id to stash the work item type. */
821 #define IBLND_WID_TX 0
822 #define IBLND_WID_RDMA 1
823 #define IBLND_WID_RX 2
824 #define IBLND_WID_MASK 3UL
827 kiblnd_ptr2wreqid (void *ptr, int type)
829 unsigned long lptr = (unsigned long)ptr;
831 LASSERT ((lptr & IBLND_WID_MASK) == 0);
832 LASSERT ((type & ~IBLND_WID_MASK) == 0);
833 return (__u64)(lptr | type);
837 kiblnd_wreqid2ptr (__u64 wreqid)
839 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
843 kiblnd_wreqid2type (__u64 wreqid)
845 return (wreqid & IBLND_WID_MASK);
849 kiblnd_set_conn_state (kib_conn_t *conn, int state)
851 conn->ibc_state = state;
856 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
858 msg->ibm_type = type;
859 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
863 kiblnd_rd_size (kib_rdma_desc_t *rd)
868 for (i = size = 0; i < rd->rd_nfrags; i++)
869 size += rd->rd_frags[i].rf_nob;
875 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
877 return rd->rd_frags[index].rf_addr;
881 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
883 return rd->rd_frags[index].rf_nob;
887 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
893 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
895 if (nob < rd->rd_frags[index].rf_nob) {
896 rd->rd_frags[index].rf_addr += nob;
897 rd->rd_frags[index].rf_nob -= nob;
906 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
908 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
909 msgtype == IBLND_MSG_PUT_ACK);
911 return msgtype == IBLND_MSG_GET_REQ ?
912 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
913 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
918 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
920 return ib_dma_mapping_error(dev, dma_addr);
923 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
924 void *msg, size_t size,
925 enum dma_data_direction direction)
927 return ib_dma_map_single(dev, msg, size, direction);
930 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
931 __u64 addr, size_t size,
932 enum dma_data_direction direction)
934 ib_dma_unmap_single(dev, addr, size, direction);
937 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
938 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
940 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
941 struct scatterlist *sg, int nents,
942 enum dma_data_direction direction)
944 return ib_dma_map_sg(dev, sg, nents, direction);
947 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
948 struct scatterlist *sg, int nents,
949 enum dma_data_direction direction)
951 ib_dma_unmap_sg(dev, sg, nents, direction);
954 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
955 struct scatterlist *sg)
957 return ib_sg_dma_address(dev, sg);
960 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
961 struct scatterlist *sg)
963 return ib_sg_dma_len(dev, sg);
966 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
967 * right because OFED1.2 defines it as const, to use it we have to add
968 * (void *) cast to overcome "const" */
970 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
971 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
974 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
975 kib_rdma_desc_t *rd);
976 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
977 __u64 addr, __u64 size);
978 void kiblnd_map_rx_descs(kib_conn_t *conn);
979 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
980 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
981 kib_rdma_desc_t *rd, int nfrags);
982 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
983 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
984 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
986 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
987 int npages, __u64 iov, kib_fmr_t *fmr);
988 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
990 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
991 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
992 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
994 int kiblnd_startup (lnet_ni_t *ni);
995 void kiblnd_shutdown (lnet_ni_t *ni);
996 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
997 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
999 int kiblnd_tunables_init(void);
1000 void kiblnd_tunables_fini(void);
1002 int kiblnd_connd (void *arg);
1003 int kiblnd_scheduler(void *arg);
1004 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1005 int kiblnd_failover_thread (void *arg);
1007 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1008 void kiblnd_free_pages (kib_pages_t *p);
1010 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1011 struct rdma_cm_event *event);
1012 int kiblnd_translate_mtu(int value);
1014 int kiblnd_dev_failover(kib_dev_t *dev);
1015 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1016 void kiblnd_destroy_peer (kib_peer_t *peer);
1017 void kiblnd_destroy_dev (kib_dev_t *dev);
1018 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1019 void kiblnd_peer_alive (kib_peer_t *peer);
1020 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1021 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1022 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1023 int version, __u64 incarnation);
1024 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1026 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1027 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1028 int state, int version);
1029 void kiblnd_destroy_conn (kib_conn_t *conn);
1030 void kiblnd_close_conn (kib_conn_t *conn, int error);
1031 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1033 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1034 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1036 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1037 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1038 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1039 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1040 void kiblnd_txlist_done (lnet_ni_t *ni, struct list_head *txlist,
1042 void kiblnd_check_sends (kib_conn_t *conn);
1044 void kiblnd_qp_event(struct ib_event *event, void *arg);
1045 void kiblnd_cq_event(struct ib_event *event, void *arg);
1046 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1048 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1049 int credits, lnet_nid_t dstnid, __u64 dststamp);
1050 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1051 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1053 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1054 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1055 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1056 unsigned int offset, unsigned int mlen, unsigned int rlen);