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, 2015, 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 #ifdef HAVE_COMPAT_RDMA
42 #include <linux/compat-2.6.h>
45 #include <linux/version.h>
46 #include <linux/module.h>
47 #include <linux/kernel.h>
48 #include <linux/kthread.h>
50 #include <linux/string.h>
51 #include <linux/stat.h>
52 #include <linux/errno.h>
53 #include <linux/unistd.h>
54 #include <linux/uio.h>
56 #include <asm/uaccess.h>
59 #include <linux/init.h>
61 #include <linux/file.h>
62 #include <linux/stat.h>
63 #include <linux/list.h>
64 #include <linux/kmod.h>
65 #include <linux/sysctl.h>
66 #include <linux/pci.h>
67 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,32)
68 #include <linux/pci-dma.h>
74 #include <rdma/rdma_cm.h>
75 #include <rdma/ib_cm.h>
76 #include <rdma/ib_verbs.h>
77 #include <rdma/ib_fmr_pool.h>
79 #define DEBUG_SUBSYSTEM S_LND
81 #include <libcfs/libcfs.h>
82 #include <lnet/lnet.h>
83 #include <lnet/lib-lnet.h>
85 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
86 /* # scheduler loops before reschedule */
87 #define IBLND_RESCHED 100
89 #define IBLND_N_SCHED 2
90 #define IBLND_N_SCHED_HIGH 4
94 int *kib_dev_failover; /* HCA failover */
95 unsigned int *kib_service; /* IB service number */
96 int *kib_min_reconnect_interval; /* first failed connection retry... */
97 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
98 int *kib_cksum; /* checksum kib_msg_t? */
99 int *kib_timeout; /* comms timeout (seconds) */
100 int *kib_keepalive; /* keepalive timeout (seconds) */
101 int *kib_ntx; /* # tx descs */
102 char **kib_default_ipif; /* default IPoIB interface */
103 int *kib_retry_count;
104 int *kib_rnr_retry_count;
105 int *kib_ib_mtu; /* IB MTU */
106 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
107 struct ctl_table_header *kib_sysctl; /* sysctl interface */
109 int *kib_require_priv_port;/* accept only privileged ports */
110 int *kib_use_priv_port; /* use privileged port for active connect */
111 /* # threads on each CPT */
115 extern kib_tunables_t kiblnd_tunables;
117 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
118 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
120 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
121 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
123 /* when eagerly to return credits */
124 #define IBLND_CREDITS_HIGHWATER(t, v) ((v) == IBLND_MSG_VERSION_1 ? \
125 IBLND_CREDIT_HIGHWATER_V1 : \
126 t->lnd_peercredits_hiw)
128 #ifdef HAVE_RDMA_CREATE_ID_5ARG
129 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(current->nsproxy->net_ns, \
133 # ifdef HAVE_RDMA_CREATE_ID_4ARG
134 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, \
137 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
141 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
142 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
143 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
145 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
146 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
148 /************************/
149 /* derived constants... */
150 /* Pools (shared by connections on each CPT) */
151 /* These pools can grow at runtime, so don't need give a very large value */
152 #define IBLND_TX_POOL 256
153 #define IBLND_FMR_POOL 256
154 #define IBLND_FMR_POOL_FLUSH 192
156 /* RX messages (per connection) */
157 #define IBLND_RX_MSGS(c) \
158 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
159 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
160 #define IBLND_RX_MSG_PAGES(c) \
161 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
163 /* WRs and CQEs (per connection) */
164 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
165 #define IBLND_SEND_WRS(c) \
166 ((c->ibc_max_frags + 1) * kiblnd_concurrent_sends(c->ibc_version, \
167 c->ibc_peer->ibp_ni))
168 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + IBLND_SEND_WRS(c))
172 /* o2iblnd can run over aliased interface */
174 #define KIB_IFNAME_SIZE IFALIASZ
176 #define KIB_IFNAME_SIZE 256
181 struct list_head ibd_list; /* chain on kib_devs */
182 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
183 __u32 ibd_ifip; /* IPoIB interface IP */
184 /** IPoIB interface name */
185 char ibd_ifname[KIB_IFNAME_SIZE];
186 int ibd_nnets; /* # nets extant */
188 cfs_time_t ibd_next_failover;
189 /* # failover failures */
190 int ibd_failed_failover;
191 /* failover in progress */
192 unsigned int ibd_failover;
193 /* IPoIB interface is a bonding master */
194 unsigned int ibd_can_failover;
195 struct list_head ibd_nets;
196 struct kib_hca_dev *ibd_hdev;
199 typedef struct kib_hca_dev
201 struct rdma_cm_id *ibh_cmid; /* listener cmid */
202 struct ib_device *ibh_ibdev; /* IB device */
203 int ibh_page_shift; /* page shift of current HCA */
204 int ibh_page_size; /* page size of current HCA */
205 __u64 ibh_page_mask; /* page mask of current HCA */
206 int ibh_mr_shift; /* bits shift of max MR size */
207 __u64 ibh_mr_size; /* size of MR */
208 struct ib_mr *ibh_mrs; /* global MR */
209 struct ib_pd *ibh_pd; /* PD */
210 kib_dev_t *ibh_dev; /* owner */
211 atomic_t ibh_ref; /* refcount */
214 /** # of seconds to keep pool alive */
215 #define IBLND_POOL_DEADLINE 300
216 /** # of seconds to retry if allocation failed */
217 #define IBLND_POOL_RETRY 1
221 int ibp_npages; /* # pages */
222 struct page *ibp_pages[0]; /* page array */
228 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
229 int inc, struct kib_pool **pp_po);
230 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
231 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
232 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
236 #define IBLND_POOL_NAME_LEN 32
238 typedef struct kib_poolset
242 /* network it belongs to */
243 struct kib_net *ps_net;
245 char ps_name[IBLND_POOL_NAME_LEN];
247 struct list_head ps_pool_list;
248 /* failed pool list */
249 struct list_head ps_failed_pool_list;
250 /* time stamp for retry if failed to allocate */
251 cfs_time_t ps_next_retry;
252 /* is allocating new pool */
259 /* create a new pool */
260 kib_ps_pool_create_t ps_pool_create;
262 kib_ps_pool_destroy_t ps_pool_destroy;
263 /* initialize new allocated node */
264 kib_ps_node_init_t ps_node_init;
266 kib_ps_node_fini_t ps_node_fini;
269 typedef struct kib_pool
271 /* chain on pool list */
272 struct list_head po_list;
273 /* pre-allocated node */
274 struct list_head po_free_list;
275 /* pool_set of this pool */
276 kib_poolset_t *po_owner;
277 /* deadline of this pool */
278 cfs_time_t po_deadline;
279 /* # of elements in use */
281 /* pool is created on failed HCA */
283 /* # of pre-allocated elements */
288 kib_poolset_t tps_poolset; /* pool-set */
289 __u64 tps_next_tx_cookie; /* cookie of TX */
293 kib_pool_t tpo_pool; /* pool */
294 struct kib_hca_dev *tpo_hdev; /* device for this pool */
295 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
296 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
301 spinlock_t fps_lock; /* serialize */
302 struct kib_net *fps_net; /* IB network */
303 struct list_head fps_pool_list; /* FMR pool list */
304 struct list_head fps_failed_pool_list; /* FMR pool list */
305 __u64 fps_version; /* validity stamp */
306 int fps_cpt; /* CPT id */
308 int fps_flush_trigger;
310 /* is allocating new pool */
312 /* time stamp for retry if failed to allocate */
313 cfs_time_t fps_next_retry;
316 struct kib_fast_reg_descriptor { /* For fast registration */
317 struct list_head frd_list;
318 struct ib_send_wr frd_inv_wr;
319 struct ib_send_wr frd_fastreg_wr;
320 struct ib_mr *frd_mr;
321 struct ib_fast_reg_page_list *frd_frpl;
327 struct list_head fpo_list; /* chain on pool list */
328 struct kib_hca_dev *fpo_hdev; /* device for this pool */
329 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
332 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
334 struct { /* For fast registration */
335 struct list_head fpo_pool_list;
339 cfs_time_t fpo_deadline; /* deadline of this pool */
340 int fpo_failed; /* fmr pool is failed */
341 int fpo_map_count; /* # of mapped FMR */
346 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
347 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
348 struct kib_fast_reg_descriptor *fmr_frd;
352 typedef struct kib_net
354 /* chain on kib_dev_t::ibd_nets */
355 struct list_head ibn_list;
356 __u64 ibn_incarnation;/* my epoch */
357 int ibn_init; /* initialisation state */
358 int ibn_shutdown; /* shutting down? */
360 atomic_t ibn_npeers; /* # peers extant */
361 atomic_t ibn_nconns; /* # connections extant */
363 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
364 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
366 kib_dev_t *ibn_dev; /* underlying IB device */
369 #define KIB_THREAD_SHIFT 16
370 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
371 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
372 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
374 struct kib_sched_info {
377 /* schedulers sleep here */
378 wait_queue_head_t ibs_waitq;
379 /* conns to check for rx completions */
380 struct list_head ibs_conns;
381 /* number of scheduler threads */
383 /* max allowed scheduler threads */
384 int ibs_nthreads_max;
385 int ibs_cpt; /* CPT id */
390 int kib_init; /* initialisation state */
391 int kib_shutdown; /* shut down? */
392 struct list_head kib_devs; /* IB devices extant */
393 /* list head of failed devices */
394 struct list_head kib_failed_devs;
395 /* schedulers sleep here */
396 wait_queue_head_t kib_failover_waitq;
397 atomic_t kib_nthreads; /* # live threads */
398 /* stabilize net/dev/peer/conn ops */
399 rwlock_t kib_global_lock;
400 /* hash table of all my known peers */
401 struct list_head *kib_peers;
402 /* size of kib_peers */
403 int kib_peer_hash_size;
404 /* the connd task (serialisation assertions) */
406 /* connections to setup/teardown */
407 struct list_head kib_connd_conns;
408 /* connections with zero refcount */
409 struct list_head kib_connd_zombies;
410 /* connections to reconnect */
411 struct list_head kib_reconn_list;
412 /* peers wait for reconnection */
413 struct list_head kib_reconn_wait;
415 * The second that peers are pulled out from \a kib_reconn_wait
418 unsigned int kib_reconn_sec;
419 /* connection daemon sleeps here */
420 wait_queue_head_t kib_connd_waitq;
421 spinlock_t kib_connd_lock; /* serialise */
422 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
423 /* percpt data for schedulers */
424 struct kib_sched_info **kib_scheds;
427 #define IBLND_INIT_NOTHING 0
428 #define IBLND_INIT_DATA 1
429 #define IBLND_INIT_ALL 2
431 /************************************************************************
432 * IB Wire message format.
433 * These are sent in sender's byte order (i.e. receiver flips).
436 typedef struct kib_connparams
438 __u16 ibcp_queue_depth;
439 __u16 ibcp_max_frags;
440 __u32 ibcp_max_msg_size;
441 } WIRE_ATTR kib_connparams_t;
445 lnet_hdr_t ibim_hdr; /* portals header */
446 char ibim_payload[0]; /* piggy-backed payload */
447 } WIRE_ATTR kib_immediate_msg_t;
451 __u32 rf_nob; /* # bytes this frag */
452 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
453 } WIRE_ATTR kib_rdma_frag_t;
457 __u32 rd_key; /* local/remote key */
458 __u32 rd_nfrags; /* # fragments */
459 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
460 } WIRE_ATTR kib_rdma_desc_t;
464 lnet_hdr_t ibprm_hdr; /* portals header */
465 __u64 ibprm_cookie; /* opaque completion cookie */
466 } WIRE_ATTR kib_putreq_msg_t;
470 __u64 ibpam_src_cookie; /* reflected completion cookie */
471 __u64 ibpam_dst_cookie; /* opaque completion cookie */
472 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
473 } WIRE_ATTR kib_putack_msg_t;
477 lnet_hdr_t ibgm_hdr; /* portals header */
478 __u64 ibgm_cookie; /* opaque completion cookie */
479 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
480 } WIRE_ATTR kib_get_msg_t;
484 __u64 ibcm_cookie; /* opaque completion cookie */
485 __s32 ibcm_status; /* < 0 failure: >= 0 length */
486 } WIRE_ATTR kib_completion_msg_t;
490 /* First 2 fields fixed FOR ALL TIME */
491 __u32 ibm_magic; /* I'm an ibnal message */
492 __u16 ibm_version; /* this is my version number */
494 __u8 ibm_type; /* msg type */
495 __u8 ibm_credits; /* returned credits */
496 __u32 ibm_nob; /* # bytes in whole message */
497 __u32 ibm_cksum; /* checksum (0 == no checksum) */
498 __u64 ibm_srcnid; /* sender's NID */
499 __u64 ibm_srcstamp; /* sender's incarnation */
500 __u64 ibm_dstnid; /* destination's NID */
501 __u64 ibm_dststamp; /* destination's incarnation */
504 kib_connparams_t connparams;
505 kib_immediate_msg_t immediate;
506 kib_putreq_msg_t putreq;
507 kib_putack_msg_t putack;
509 kib_completion_msg_t completion;
511 } WIRE_ATTR kib_msg_t;
513 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
515 #define IBLND_MSG_VERSION_1 0x11
516 #define IBLND_MSG_VERSION_2 0x12
517 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
519 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
520 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
521 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
522 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
523 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
524 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
525 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
526 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
527 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
528 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
531 __u32 ibr_magic; /* sender's magic */
532 __u16 ibr_version; /* sender's version */
533 __u8 ibr_why; /* reject reason */
534 __u8 ibr_padding; /* padding */
535 __u64 ibr_incarnation; /* incarnation of peer */
536 kib_connparams_t ibr_cp; /* connection parameters */
537 } WIRE_ATTR kib_rej_t;
539 /* connection rejection reasons */
540 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
541 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
542 #define IBLND_REJECT_FATAL 3 /* Anything else */
544 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
545 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
547 /* peer's rdma frags doesn't match mine */
548 #define IBLND_REJECT_RDMA_FRAGS 6
549 /* peer's msg queue size doesn't match mine */
550 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
552 /***********************************************************************/
554 typedef struct kib_rx /* receive message */
556 /* queue for attention */
557 struct list_head rx_list;
559 struct kib_conn *rx_conn;
560 /* # bytes received (-1 while posted) */
562 /* completion status */
563 enum ib_wc_status rx_status;
564 /* message buffer (host vaddr) */
566 /* message buffer (I/O addr) */
568 /* for dma_unmap_single() */
569 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap);
570 /* receive work item... */
571 struct ib_recv_wr rx_wrq;
572 /* ...and its memory */
573 struct ib_sge rx_sge;
576 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
577 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
578 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
579 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
581 typedef struct kib_tx /* transmit message */
583 /* queue on idle_txs ibc_tx_queue etc. */
584 struct list_head tx_list;
586 kib_tx_pool_t *tx_pool;
588 struct kib_conn *tx_conn;
589 /* # tx callbacks outstanding */
591 /* queued for sending */
593 /* waiting for peer */
595 /* LNET completion status */
597 /* completion deadline */
598 unsigned long tx_deadline;
599 /* completion cookie */
601 /* lnet msgs to finalize on completion */
602 lnet_msg_t *tx_lntmsg[2];
603 /* message buffer (host vaddr) */
605 /* message buffer (I/O addr) */
607 /* for dma_unmap_single() */
608 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap);
609 /* # send work items */
611 /* send work items... */
612 struct ib_send_wr *tx_wrq;
613 /* ...and their memory */
614 struct ib_sge *tx_sge;
615 /* rdma descriptor */
616 kib_rdma_desc_t *tx_rd;
617 /* # entries in... */
619 /* dma_map_sg descriptor */
620 struct scatterlist *tx_frags;
621 /* rdma phys page addrs */
629 typedef struct kib_connvars
631 /* connection-in-progress variables */
635 typedef struct kib_conn
637 /* scheduler information */
638 struct kib_sched_info *ibc_sched;
640 struct kib_peer *ibc_peer;
642 kib_hca_dev_t *ibc_hdev;
643 /* stash on peer's conn list */
644 struct list_head ibc_list;
645 /* schedule for attention */
646 struct list_head ibc_sched_list;
647 /* version of connection */
649 /* reconnect later */
650 __u16 ibc_reconnect:1;
651 /* which instance of the peer */
652 __u64 ibc_incarnation;
654 atomic_t ibc_refcount;
655 /* what's happening */
657 /* # uncompleted sends */
658 int ibc_nsends_posted;
659 /* # uncompleted NOOPs */
660 int ibc_noops_posted;
661 /* # credits I have */
663 /* # credits to return */
664 int ibc_outstanding_credits;
665 /* # ACK/DONE msg credits */
666 int ibc_reserved_credits;
667 /* set on comms error */
669 /* connections queue depth */
670 __u16 ibc_queue_depth;
671 /* connections max frags */
673 /* receive buffers owned */
674 unsigned int ibc_nrx:16;
675 /* scheduled for attention */
676 unsigned int ibc_scheduled:1;
677 /* CQ callback fired */
678 unsigned int ibc_ready:1;
679 /* time of last send */
680 unsigned long ibc_last_send;
681 /** link chain for kiblnd_check_conns only */
682 struct list_head ibc_connd_list;
683 /** rxs completed before ESTABLISHED */
684 struct list_head ibc_early_rxs;
685 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
686 struct list_head ibc_tx_noops;
687 /* sends that need a credit */
688 struct list_head ibc_tx_queue;
689 /* sends that don't need a credit */
690 struct list_head ibc_tx_queue_nocred;
691 /* sends that need to reserve an ACK/DONE msg */
692 struct list_head ibc_tx_queue_rsrvd;
693 /* active tx awaiting completion */
694 struct list_head ibc_active_txs;
699 /* premapped rx msg pages */
700 kib_pages_t *ibc_rx_pages;
703 struct rdma_cm_id *ibc_cmid;
704 /* completion queue */
705 struct ib_cq *ibc_cq;
707 /* in-progress connection state */
708 kib_connvars_t *ibc_connvars;
711 #define IBLND_CONN_INIT 0 /* being initialised */
712 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
713 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
714 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
715 #define IBLND_CONN_CLOSING 4 /* being closed */
716 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
718 typedef struct kib_peer
720 /* stash on global peer list */
721 struct list_head ibp_list;
722 /* who's on the other end(s) */
726 /* all active connections */
727 struct list_head ibp_conns;
728 /* msgs waiting for a conn */
729 struct list_head ibp_tx_queue;
730 /* incarnation of peer */
731 __u64 ibp_incarnation;
732 /* when (in jiffies) I was last alive */
733 cfs_time_t ibp_last_alive;
735 atomic_t ibp_refcount;
736 /* version of peer */
738 /* current passive connection attempts */
739 unsigned short ibp_accepting;
740 /* current active connection attempts */
741 unsigned short ibp_connecting;
742 /* reconnect this peer later */
743 unsigned short ibp_reconnecting:1;
744 /* # consecutive reconnection attempts to this peer */
745 unsigned int ibp_reconnected;
746 /* errno on closing this peer */
748 /* max map_on_demand */
750 /* max_peer_credits */
751 __u16 ibp_queue_depth;
754 #ifndef HAVE_IB_INC_RKEY
756 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
757 * for calculating a new rkey for type 2 memory windows.
758 * @rkey - the rkey to increment.
760 static inline u32 ib_inc_rkey(u32 rkey)
762 const u32 mask = 0x000000ff;
763 return ((rkey + 1) & mask) | (rkey & ~mask);
767 extern kib_data_t kiblnd_data;
769 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
771 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
773 /* max # of fragments configured by user */
775 kiblnd_cfg_rdma_frags(struct lnet_ni *ni)
777 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
780 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
781 mod = tunables->lnd_map_on_demand;
782 return mod != 0 ? mod : IBLND_MAX_RDMA_FRAGS;
786 kiblnd_rdma_frags(int version, struct lnet_ni *ni)
788 return version == IBLND_MSG_VERSION_1 ?
789 IBLND_MAX_RDMA_FRAGS :
790 kiblnd_cfg_rdma_frags(ni);
794 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
796 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
797 int concurrent_sends;
799 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
800 concurrent_sends = tunables->lnd_concurrent_sends;
802 if (version == IBLND_MSG_VERSION_1) {
803 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
804 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
806 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
807 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
810 return concurrent_sends;
814 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
816 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
817 atomic_inc(&hdev->ibh_ref);
821 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
823 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
824 if (atomic_dec_and_test(&hdev->ibh_ref))
825 kiblnd_hdev_destroy(hdev);
829 kiblnd_dev_can_failover(kib_dev_t *dev)
831 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
834 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
837 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
840 return dev->ibd_can_failover;
843 #define kiblnd_conn_addref(conn) \
845 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
846 (conn), atomic_read(&(conn)->ibc_refcount)); \
847 atomic_inc(&(conn)->ibc_refcount); \
850 #define kiblnd_conn_decref(conn) \
852 unsigned long flags; \
854 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
855 (conn), atomic_read(&(conn)->ibc_refcount)); \
856 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
857 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
858 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
859 list_add_tail(&(conn)->ibc_list, \
860 &kiblnd_data.kib_connd_zombies); \
861 wake_up(&kiblnd_data.kib_connd_waitq); \
862 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
866 #define kiblnd_peer_addref(peer) \
868 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
869 (peer), libcfs_nid2str((peer)->ibp_nid), \
870 atomic_read (&(peer)->ibp_refcount)); \
871 atomic_inc(&(peer)->ibp_refcount); \
874 #define kiblnd_peer_decref(peer) \
876 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
877 (peer), libcfs_nid2str((peer)->ibp_nid), \
878 atomic_read (&(peer)->ibp_refcount)); \
879 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
880 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
881 kiblnd_destroy_peer(peer); \
885 kiblnd_peer_connecting(kib_peer_t *peer)
887 return peer->ibp_connecting != 0 ||
888 peer->ibp_reconnecting != 0 ||
889 peer->ibp_accepting != 0;
893 kiblnd_peer_idle(kib_peer_t *peer)
895 return !kiblnd_peer_connecting(peer) && list_empty(&peer->ibp_conns);
898 static inline struct list_head *
899 kiblnd_nid2peerlist (lnet_nid_t nid)
902 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
904 return &kiblnd_data.kib_peers[hash];
908 kiblnd_peer_active (kib_peer_t *peer)
910 /* Am I in the peer hash table? */
911 return !list_empty(&peer->ibp_list);
914 static inline kib_conn_t *
915 kiblnd_get_conn_locked (kib_peer_t *peer)
917 LASSERT(!list_empty(&peer->ibp_conns));
919 /* just return the first connection */
920 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
924 kiblnd_send_keepalive(kib_conn_t *conn)
926 return (*kiblnd_tunables.kib_keepalive > 0) &&
927 cfs_time_after(jiffies, conn->ibc_last_send +
928 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
933 kiblnd_need_noop(kib_conn_t *conn)
935 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
936 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
938 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
939 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
941 if (conn->ibc_outstanding_credits <
942 IBLND_CREDITS_HIGHWATER(tunables, conn->ibc_version) &&
943 !kiblnd_send_keepalive(conn))
944 return 0; /* No need to send NOOP */
946 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
947 if (!list_empty(&conn->ibc_tx_queue_nocred))
948 return 0; /* NOOP can be piggybacked */
950 /* No tx to piggyback NOOP onto or no credit to send a tx */
951 return (list_empty(&conn->ibc_tx_queue) ||
952 conn->ibc_credits == 0);
955 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
956 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
957 conn->ibc_credits == 0) /* no credit */
960 if (conn->ibc_credits == 1 && /* last credit reserved for */
961 conn->ibc_outstanding_credits == 0) /* giving back credits */
964 /* No tx to piggyback NOOP onto or no credit to send a tx */
965 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
969 kiblnd_abort_receives(kib_conn_t *conn)
971 ib_modify_qp(conn->ibc_cmid->qp,
972 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
975 static inline const char *
976 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
978 if (q == &conn->ibc_tx_queue)
981 if (q == &conn->ibc_tx_queue_rsrvd)
982 return "tx_queue_rsrvd";
984 if (q == &conn->ibc_tx_queue_nocred)
985 return "tx_queue_nocred";
987 if (q == &conn->ibc_active_txs)
994 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
995 * lowest bits of the work request id to stash the work item type. */
997 #define IBLND_WID_INVAL 0
998 #define IBLND_WID_TX 1
999 #define IBLND_WID_RX 2
1000 #define IBLND_WID_RDMA 3
1001 #define IBLND_WID_MR 4
1002 #define IBLND_WID_MASK 7UL
1005 kiblnd_ptr2wreqid (void *ptr, int type)
1007 unsigned long lptr = (unsigned long)ptr;
1009 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1010 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1011 return (__u64)(lptr | type);
1014 static inline void *
1015 kiblnd_wreqid2ptr (__u64 wreqid)
1017 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1021 kiblnd_wreqid2type (__u64 wreqid)
1023 return (wreqid & IBLND_WID_MASK);
1027 kiblnd_set_conn_state (kib_conn_t *conn, int state)
1029 conn->ibc_state = state;
1034 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
1036 msg->ibm_type = type;
1037 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1041 kiblnd_rd_size (kib_rdma_desc_t *rd)
1046 for (i = size = 0; i < rd->rd_nfrags; i++)
1047 size += rd->rd_frags[i].rf_nob;
1053 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
1055 return rd->rd_frags[index].rf_addr;
1059 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
1061 return rd->rd_frags[index].rf_nob;
1065 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
1071 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
1073 if (nob < rd->rd_frags[index].rf_nob) {
1074 rd->rd_frags[index].rf_addr += nob;
1075 rd->rd_frags[index].rf_nob -= nob;
1084 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
1086 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1087 msgtype == IBLND_MSG_PUT_ACK);
1089 return msgtype == IBLND_MSG_GET_REQ ?
1090 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
1091 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
1095 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1097 return ib_dma_mapping_error(dev, dma_addr);
1100 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1101 void *msg, size_t size,
1102 enum dma_data_direction direction)
1104 return ib_dma_map_single(dev, msg, size, direction);
1107 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1108 __u64 addr, size_t size,
1109 enum dma_data_direction direction)
1111 ib_dma_unmap_single(dev, addr, size, direction);
1114 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1115 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1117 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1118 struct scatterlist *sg, int nents,
1119 enum dma_data_direction direction)
1121 return ib_dma_map_sg(dev, sg, nents, direction);
1124 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1125 struct scatterlist *sg, int nents,
1126 enum dma_data_direction direction)
1128 ib_dma_unmap_sg(dev, sg, nents, direction);
1131 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1132 struct scatterlist *sg)
1134 return ib_sg_dma_address(dev, sg);
1137 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1138 struct scatterlist *sg)
1140 return ib_sg_dma_len(dev, sg);
1143 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1144 * right because OFED1.2 defines it as const, to use it we have to add
1145 * (void *) cast to overcome "const" */
1147 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1148 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1150 struct ib_mr *kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1151 int negotiated_nfrags);
1152 void kiblnd_map_rx_descs(kib_conn_t *conn);
1153 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1154 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
1155 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1157 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1158 __u32 nob, __u64 iov, bool is_rx, kib_fmr_t *fmr);
1159 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1161 int kiblnd_tunables_setup(struct lnet_ni *ni);
1162 int kiblnd_tunables_init(void);
1163 void kiblnd_tunables_fini(void);
1165 int kiblnd_connd (void *arg);
1166 int kiblnd_scheduler(void *arg);
1167 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1168 int kiblnd_failover_thread (void *arg);
1170 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1172 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1173 struct rdma_cm_event *event);
1174 int kiblnd_translate_mtu(int value);
1176 int kiblnd_dev_failover(kib_dev_t *dev);
1177 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1178 void kiblnd_destroy_peer (kib_peer_t *peer);
1179 bool kiblnd_reconnect_peer(kib_peer_t *peer);
1180 void kiblnd_destroy_dev (kib_dev_t *dev);
1181 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1182 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1183 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1184 int version, __u64 incarnation);
1185 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1187 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
1188 int state, int version);
1189 void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn);
1190 void kiblnd_close_conn (kib_conn_t *conn, int error);
1191 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1193 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1194 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status);
1196 void kiblnd_qp_event(struct ib_event *event, void *arg);
1197 void kiblnd_cq_event(struct ib_event *event, void *arg);
1198 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1200 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1201 int credits, lnet_nid_t dstnid, __u64 dststamp);
1202 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1203 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1205 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1206 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1207 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1208 unsigned int offset, unsigned int mlen, unsigned int rlen);