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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lnet/klnds/o2iblnd/o2iblnd.h
34 * Author: Eric Barton <eric@bartonsoftware.com>
37 #ifdef HAVE_COMPAT_RDMA
38 #include <linux/compat-2.6.h>
40 #ifdef LINUX_3_17_COMPAT_H
41 #undef NEED_KTIME_GET_REAL_NS
46 #include <linux/version.h>
47 #include <linux/module.h>
48 #include <linux/kernel.h>
49 #include <linux/kthread.h>
51 #include <linux/string.h>
52 #include <linux/stat.h>
53 #include <linux/errno.h>
54 #include <linux/unistd.h>
55 #include <linux/uio.h>
57 #include <asm/uaccess.h>
60 #include <linux/init.h>
62 #include <linux/file.h>
63 #include <linux/stat.h>
64 #include <linux/list.h>
65 #include <linux/kmod.h>
66 #include <linux/sysctl.h>
67 #include <linux/pci.h>
68 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,32)
69 #include <linux/pci-dma.h>
75 #include <rdma/rdma_cm.h>
76 #include <rdma/ib_cm.h>
77 #include <rdma/ib_verbs.h>
78 #include <rdma/ib_fmr_pool.h>
80 #define DEBUG_SUBSYSTEM S_LND
82 #include <libcfs/libcfs.h>
83 #include <lnet/lib-lnet.h>
85 #define IBLND_PEER_HASH_SIZE 101 /* # peer_ni lists */
86 /* # scheduler loops before reschedule */
87 #define IBLND_RESCHED 100
89 #define IBLND_N_SCHED 2
90 #define IBLND_N_SCHED_HIGH 4
92 #define IBLND_DEV_CAPS_FASTREG_ENABLED 0x1
93 #define IBLND_DEV_CAPS_FASTREG_GAPS_SUPPORT 0x2
94 #define IBLND_DEV_CAPS_FMR_ENABLED 0x4
98 int *kib_dev_failover; /* HCA failover */
99 unsigned int *kib_service; /* IB service number */
100 int *kib_min_reconnect_interval; /* first failed connection retry... */
101 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
102 int *kib_cksum; /* checksum kib_msg_t? */
103 int *kib_timeout; /* comms timeout (seconds) */
104 int *kib_keepalive; /* keepalive timeout (seconds) */
105 int *kib_ntx; /* # tx descs */
106 char **kib_default_ipif; /* default IPoIB interface */
107 int *kib_retry_count;
108 int *kib_rnr_retry_count;
109 int *kib_ib_mtu; /* IB MTU */
110 int *kib_require_priv_port;/* accept only privileged ports */
111 int *kib_use_priv_port; /* use privileged port for active connect */
112 /* # threads on each CPT */
114 int *kib_wrq_sge; /* # sg elements per wrq */
117 extern kib_tunables_t kiblnd_tunables;
119 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
120 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
122 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer_ni credits */
123 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer_ni credits */
125 /* when eagerly to return credits */
126 #define IBLND_CREDITS_HIGHWATER(t, v) ((v) == IBLND_MSG_VERSION_1 ? \
127 IBLND_CREDIT_HIGHWATER_V1 : \
128 t->lnd_peercredits_hiw)
130 #ifdef HAVE_RDMA_CREATE_ID_5ARG
131 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(current->nsproxy->net_ns, \
135 # ifdef HAVE_RDMA_CREATE_ID_4ARG
136 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, \
139 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
143 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
144 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
145 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
147 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
148 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
150 /************************/
151 /* derived constants... */
152 /* Pools (shared by connections on each CPT) */
153 /* These pools can grow at runtime, so don't need give a very large value */
154 #define IBLND_TX_POOL 256
155 #define IBLND_FMR_POOL 256
156 #define IBLND_FMR_POOL_FLUSH 192
158 /* RX messages (per connection) */
159 #define IBLND_RX_MSGS(c) \
160 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
161 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
162 #define IBLND_RX_MSG_PAGES(c) \
163 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
165 /* WRs and CQEs (per connection) */
166 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
167 #define IBLND_SEND_WRS(c) \
168 ((c->ibc_max_frags + 1) * kiblnd_concurrent_sends(c->ibc_version, \
169 c->ibc_peer->ibp_ni))
171 /* 2 = LNet msg + Transfer chain */
172 #define IBLND_CQ_ENTRIES(c) \
173 (IBLND_RECV_WRS(c) + 2 * kiblnd_concurrent_sends(c->ibc_version, \
174 c->ibc_peer->ibp_ni))
178 /* o2iblnd can run over aliased interface */
180 #define KIB_IFNAME_SIZE IFALIASZ
182 #define KIB_IFNAME_SIZE 256
187 struct list_head ibd_list; /* chain on kib_devs */
188 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
189 __u32 ibd_ifip; /* IPoIB interface IP */
190 /** IPoIB interface name */
191 char ibd_ifname[KIB_IFNAME_SIZE];
192 int ibd_nnets; /* # nets extant */
194 cfs_time_t ibd_next_failover;
195 /* # failover failures */
196 int ibd_failed_failover;
197 /* failover in progress */
198 unsigned int ibd_failover;
199 /* IPoIB interface is a bonding master */
200 unsigned int ibd_can_failover;
201 struct list_head ibd_nets;
202 struct kib_hca_dev *ibd_hdev;
206 typedef struct kib_hca_dev
208 struct rdma_cm_id *ibh_cmid; /* listener cmid */
209 struct ib_device *ibh_ibdev; /* IB device */
210 int ibh_page_shift; /* page shift of current HCA */
211 int ibh_page_size; /* page size of current HCA */
212 __u64 ibh_page_mask; /* page mask of current HCA */
213 int ibh_mr_shift; /* bits shift of max MR size */
214 __u64 ibh_mr_size; /* size of MR */
215 #ifdef HAVE_IB_GET_DMA_MR
216 struct ib_mr *ibh_mrs; /* global MR */
218 struct ib_pd *ibh_pd; /* PD */
219 kib_dev_t *ibh_dev; /* owner */
220 atomic_t ibh_ref; /* refcount */
223 /** # of seconds to keep pool alive */
224 #define IBLND_POOL_DEADLINE 300
225 /** # of seconds to retry if allocation failed */
226 #define IBLND_POOL_RETRY 1
230 int ibp_npages; /* # pages */
231 struct page *ibp_pages[0]; /* page array */
237 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
238 int inc, struct kib_pool **pp_po);
239 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
240 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
241 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
245 #define IBLND_POOL_NAME_LEN 32
247 typedef struct kib_poolset
251 /* network it belongs to */
252 struct kib_net *ps_net;
254 char ps_name[IBLND_POOL_NAME_LEN];
256 struct list_head ps_pool_list;
257 /* failed pool list */
258 struct list_head ps_failed_pool_list;
259 /* time stamp for retry if failed to allocate */
260 cfs_time_t ps_next_retry;
261 /* is allocating new pool */
268 /* create a new pool */
269 kib_ps_pool_create_t ps_pool_create;
271 kib_ps_pool_destroy_t ps_pool_destroy;
272 /* initialize new allocated node */
273 kib_ps_node_init_t ps_node_init;
275 kib_ps_node_fini_t ps_node_fini;
278 typedef struct kib_pool
280 /* chain on pool list */
281 struct list_head po_list;
282 /* pre-allocated node */
283 struct list_head po_free_list;
284 /* pool_set of this pool */
285 kib_poolset_t *po_owner;
286 /* deadline of this pool */
287 cfs_time_t po_deadline;
288 /* # of elements in use */
290 /* pool is created on failed HCA */
292 /* # of pre-allocated elements */
297 kib_poolset_t tps_poolset; /* pool-set */
298 __u64 tps_next_tx_cookie; /* cookie of TX */
302 kib_pool_t tpo_pool; /* pool */
303 struct kib_hca_dev *tpo_hdev; /* device for this pool */
304 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
305 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
310 spinlock_t fps_lock; /* serialize */
311 struct kib_net *fps_net; /* IB network */
312 struct list_head fps_pool_list; /* FMR pool list */
313 struct list_head fps_failed_pool_list; /* FMR pool list */
314 __u64 fps_version; /* validity stamp */
315 int fps_cpt; /* CPT id */
317 int fps_flush_trigger;
319 /* is allocating new pool */
321 /* time stamp for retry if failed to allocate */
322 cfs_time_t fps_next_retry;
325 #ifndef HAVE_IB_RDMA_WR
327 struct ib_send_wr wr;
331 struct kib_fast_reg_descriptor { /* For fast registration */
332 struct list_head frd_list;
333 struct ib_rdma_wr frd_inv_wr;
334 #ifdef HAVE_IB_MAP_MR_SG
335 struct ib_reg_wr frd_fastreg_wr;
337 struct ib_rdma_wr frd_fastreg_wr;
338 struct ib_fast_reg_page_list *frd_frpl;
340 struct ib_mr *frd_mr;
346 struct list_head fpo_list; /* chain on pool list */
347 struct kib_hca_dev *fpo_hdev; /* device for this pool */
348 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
351 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
353 struct { /* For fast registration */
354 struct list_head fpo_pool_list;
358 cfs_time_t fpo_deadline; /* deadline of this pool */
359 int fpo_failed; /* fmr pool is failed */
360 int fpo_map_count; /* # of mapped FMR */
361 bool fpo_is_fmr; /* True if FMR pools allocated */
365 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
366 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
367 struct kib_fast_reg_descriptor *fmr_frd;
371 typedef struct kib_net
373 /* chain on kib_dev_t::ibd_nets */
374 struct list_head ibn_list;
375 __u64 ibn_incarnation;/* my epoch */
376 int ibn_init; /* initialisation state */
377 int ibn_shutdown; /* shutting down? */
379 atomic_t ibn_npeers; /* # peers extant */
380 atomic_t ibn_nconns; /* # connections extant */
382 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
383 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
385 kib_dev_t *ibn_dev; /* underlying IB device */
388 #define KIB_THREAD_SHIFT 16
389 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
390 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
391 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
393 struct kib_sched_info {
396 /* schedulers sleep here */
397 wait_queue_head_t ibs_waitq;
398 /* conns to check for rx completions */
399 struct list_head ibs_conns;
400 /* number of scheduler threads */
402 /* max allowed scheduler threads */
403 int ibs_nthreads_max;
404 int ibs_cpt; /* CPT id */
409 int kib_init; /* initialisation state */
410 int kib_shutdown; /* shut down? */
411 struct list_head kib_devs; /* IB devices extant */
412 /* list head of failed devices */
413 struct list_head kib_failed_devs;
414 /* schedulers sleep here */
415 wait_queue_head_t kib_failover_waitq;
416 atomic_t kib_nthreads; /* # live threads */
417 /* stabilize net/dev/peer_ni/conn ops */
418 rwlock_t kib_global_lock;
419 /* hash table of all my known peers */
420 struct list_head *kib_peers;
421 /* size of kib_peers */
422 int kib_peer_hash_size;
423 /* the connd task (serialisation assertions) */
425 /* connections to setup/teardown */
426 struct list_head kib_connd_conns;
427 /* connections with zero refcount */
428 struct list_head kib_connd_zombies;
429 /* connections to reconnect */
430 struct list_head kib_reconn_list;
431 /* peers wait for reconnection */
432 struct list_head kib_reconn_wait;
434 * The second that peers are pulled out from \a kib_reconn_wait
437 time64_t kib_reconn_sec;
438 /* connection daemon sleeps here */
439 wait_queue_head_t kib_connd_waitq;
440 spinlock_t kib_connd_lock; /* serialise */
441 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
442 /* percpt data for schedulers */
443 struct kib_sched_info **kib_scheds;
446 #define IBLND_INIT_NOTHING 0
447 #define IBLND_INIT_DATA 1
448 #define IBLND_INIT_ALL 2
450 /************************************************************************
451 * IB Wire message format.
452 * These are sent in sender's byte order (i.e. receiver flips).
455 typedef struct kib_connparams
457 __u16 ibcp_queue_depth;
458 __u16 ibcp_max_frags;
459 __u32 ibcp_max_msg_size;
460 } WIRE_ATTR kib_connparams_t;
464 struct lnet_hdr ibim_hdr; /* portals header */
465 char ibim_payload[0];/* piggy-backed payload */
466 } WIRE_ATTR kib_immediate_msg_t;
470 __u32 rf_nob; /* # bytes this frag */
471 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
472 } WIRE_ATTR kib_rdma_frag_t;
476 __u32 rd_key; /* local/remote key */
477 __u32 rd_nfrags; /* # fragments */
478 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
479 } WIRE_ATTR kib_rdma_desc_t;
483 struct lnet_hdr ibprm_hdr; /* portals header */
484 __u64 ibprm_cookie; /* opaque completion cookie */
485 } WIRE_ATTR kib_putreq_msg_t;
489 __u64 ibpam_src_cookie; /* reflected completion cookie */
490 __u64 ibpam_dst_cookie; /* opaque completion cookie */
491 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
492 } WIRE_ATTR kib_putack_msg_t;
496 struct lnet_hdr ibgm_hdr; /* portals header */
497 __u64 ibgm_cookie; /* opaque completion cookie */
498 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
499 } WIRE_ATTR kib_get_msg_t;
503 __u64 ibcm_cookie; /* opaque completion cookie */
504 __s32 ibcm_status; /* < 0 failure: >= 0 length */
505 } WIRE_ATTR kib_completion_msg_t;
509 /* First 2 fields fixed FOR ALL TIME */
510 __u32 ibm_magic; /* I'm an ibnal message */
511 __u16 ibm_version; /* this is my version number */
513 __u8 ibm_type; /* msg type */
514 __u8 ibm_credits; /* returned credits */
515 __u32 ibm_nob; /* # bytes in whole message */
516 __u32 ibm_cksum; /* checksum (0 == no checksum) */
517 __u64 ibm_srcnid; /* sender's NID */
518 __u64 ibm_srcstamp; /* sender's incarnation */
519 __u64 ibm_dstnid; /* destination's NID */
520 __u64 ibm_dststamp; /* destination's incarnation */
523 kib_connparams_t connparams;
524 kib_immediate_msg_t immediate;
525 kib_putreq_msg_t putreq;
526 kib_putack_msg_t putack;
528 kib_completion_msg_t completion;
530 } WIRE_ATTR kib_msg_t;
532 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
534 #define IBLND_MSG_VERSION_1 0x11
535 #define IBLND_MSG_VERSION_2 0x12
536 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
538 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
539 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
540 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
541 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
542 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
543 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
544 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
545 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
546 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
547 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
550 __u32 ibr_magic; /* sender's magic */
551 __u16 ibr_version; /* sender's version */
552 __u8 ibr_why; /* reject reason */
553 __u8 ibr_padding; /* padding */
554 __u64 ibr_incarnation; /* incarnation of peer_ni */
555 kib_connparams_t ibr_cp; /* connection parameters */
556 } WIRE_ATTR kib_rej_t;
558 /* connection rejection reasons */
559 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
560 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
561 #define IBLND_REJECT_FATAL 3 /* Anything else */
563 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer_ni */
564 #define IBLND_REJECT_CONN_STALE 5 /* stale peer_ni */
566 /* peer_ni's rdma frags doesn't match mine */
567 #define IBLND_REJECT_RDMA_FRAGS 6
568 /* peer_ni's msg queue size doesn't match mine */
569 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
570 #define IBLND_REJECT_INVALID_SRV_ID 8
572 /***********************************************************************/
574 typedef struct kib_rx /* receive message */
576 /* queue for attention */
577 struct list_head rx_list;
579 struct kib_conn *rx_conn;
580 /* # bytes received (-1 while posted) */
582 /* completion status */
583 enum ib_wc_status rx_status;
584 /* message buffer (host vaddr) */
586 /* message buffer (I/O addr) */
588 /* for dma_unmap_single() */
589 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap);
590 /* receive work item... */
591 struct ib_recv_wr rx_wrq;
592 /* ...and its memory */
593 struct ib_sge rx_sge;
596 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
597 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
598 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer_ni back 1 credit */
599 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
601 typedef struct kib_tx /* transmit message */
603 /* queue on idle_txs ibc_tx_queue etc. */
604 struct list_head tx_list;
606 kib_tx_pool_t *tx_pool;
608 struct kib_conn *tx_conn;
609 /* # tx callbacks outstanding */
611 /* queued for sending */
613 /* waiting for peer_ni */
615 /* LNET completion status */
617 /* completion deadline */
618 unsigned long tx_deadline;
619 /* completion cookie */
621 /* lnet msgs to finalize on completion */
622 struct lnet_msg *tx_lntmsg[2];
623 /* message buffer (host vaddr) */
625 /* message buffer (I/O addr) */
627 /* for dma_unmap_single() */
628 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap);
629 /** sge for tx_msgaddr */
630 struct ib_sge tx_msgsge;
631 /* # send work items */
633 /* # used scatter/gather elements */
635 /* send work items... */
636 struct ib_rdma_wr *tx_wrq;
637 /* ...and their memory */
638 struct ib_sge *tx_sge;
639 /* rdma descriptor */
640 kib_rdma_desc_t *tx_rd;
641 /* # entries in... */
643 /* dma_map_sg descriptor */
644 struct scatterlist *tx_frags;
645 /* rdma phys page addrs */
647 /* gaps in fragments */
655 typedef struct kib_connvars
657 /* connection-in-progress variables */
661 typedef struct kib_conn
663 /* scheduler information */
664 struct kib_sched_info *ibc_sched;
666 struct kib_peer *ibc_peer;
668 kib_hca_dev_t *ibc_hdev;
669 /* stash on peer_ni's conn list */
670 struct list_head ibc_list;
671 /* schedule for attention */
672 struct list_head ibc_sched_list;
673 /* version of connection */
675 /* reconnect later */
676 __u16 ibc_reconnect:1;
677 /* which instance of the peer */
678 __u64 ibc_incarnation;
680 atomic_t ibc_refcount;
681 /* what's happening */
683 /* # uncompleted sends */
684 int ibc_nsends_posted;
685 /* # uncompleted NOOPs */
686 int ibc_noops_posted;
687 /* # credits I have */
689 /* # credits to return */
690 int ibc_outstanding_credits;
691 /* # ACK/DONE msg credits */
692 int ibc_reserved_credits;
693 /* set on comms error */
695 /* connections queue depth */
696 __u16 ibc_queue_depth;
697 /* connections max frags */
699 /* receive buffers owned */
700 unsigned int ibc_nrx:16;
701 /* scheduled for attention */
702 unsigned int ibc_scheduled:1;
703 /* CQ callback fired */
704 unsigned int ibc_ready:1;
705 /* time of last send */
706 unsigned long ibc_last_send;
707 /** link chain for kiblnd_check_conns only */
708 struct list_head ibc_connd_list;
709 /** rxs completed before ESTABLISHED */
710 struct list_head ibc_early_rxs;
711 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
712 struct list_head ibc_tx_noops;
713 /* sends that need a credit */
714 struct list_head ibc_tx_queue;
715 /* sends that don't need a credit */
716 struct list_head ibc_tx_queue_nocred;
717 /* sends that need to reserve an ACK/DONE msg */
718 struct list_head ibc_tx_queue_rsrvd;
719 /* active tx awaiting completion */
720 struct list_head ibc_active_txs;
725 /* premapped rx msg pages */
726 kib_pages_t *ibc_rx_pages;
729 struct rdma_cm_id *ibc_cmid;
730 /* completion queue */
731 struct ib_cq *ibc_cq;
733 /* in-progress connection state */
734 kib_connvars_t *ibc_connvars;
737 #define IBLND_CONN_INIT 0 /* being initialised */
738 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
739 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
740 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
741 #define IBLND_CONN_CLOSING 4 /* being closed */
742 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
744 typedef struct kib_peer
746 /* stash on global peer_ni list */
747 struct list_head ibp_list;
748 /* who's on the other end(s) */
751 struct lnet_ni *ibp_ni;
752 /* all active connections */
753 struct list_head ibp_conns;
754 /* next connection to send on for round robin */
755 struct kib_conn *ibp_next_conn;
756 /* msgs waiting for a conn */
757 struct list_head ibp_tx_queue;
758 /* incarnation of peer_ni */
759 __u64 ibp_incarnation;
760 /* when (in jiffies) I was last alive */
761 cfs_time_t ibp_last_alive;
763 atomic_t ibp_refcount;
764 /* version of peer_ni */
766 /* current passive connection attempts */
767 unsigned short ibp_accepting;
768 /* current active connection attempts */
769 unsigned short ibp_connecting;
770 /* reconnect this peer_ni later */
771 unsigned char ibp_reconnecting;
772 /* counter of how many times we triggered a conn race */
773 unsigned char ibp_races;
774 /* # consecutive reconnection attempts to this peer */
775 unsigned int ibp_reconnected;
776 /* errno on closing this peer_ni */
778 /* max map_on_demand */
780 /* max_peer_credits */
781 __u16 ibp_queue_depth;
784 #ifndef HAVE_IB_INC_RKEY
786 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
787 * for calculating a new rkey for type 2 memory windows.
788 * @rkey - the rkey to increment.
790 static inline u32 ib_inc_rkey(u32 rkey)
792 const u32 mask = 0x000000ff;
793 return ((rkey + 1) & mask) | (rkey & ~mask);
797 extern kib_data_t kiblnd_data;
799 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
801 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
804 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
806 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
807 int concurrent_sends;
809 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
810 concurrent_sends = tunables->lnd_concurrent_sends;
812 if (version == IBLND_MSG_VERSION_1) {
813 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
814 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
816 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
817 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
820 return concurrent_sends;
824 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
826 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
827 atomic_inc(&hdev->ibh_ref);
831 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
833 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
834 if (atomic_dec_and_test(&hdev->ibh_ref))
835 kiblnd_hdev_destroy(hdev);
839 kiblnd_dev_can_failover(kib_dev_t *dev)
841 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
844 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
847 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
850 return dev->ibd_can_failover;
853 #define kiblnd_conn_addref(conn) \
855 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
856 (conn), atomic_read(&(conn)->ibc_refcount)); \
857 atomic_inc(&(conn)->ibc_refcount); \
860 #define kiblnd_conn_decref(conn) \
862 unsigned long flags; \
864 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
865 (conn), atomic_read(&(conn)->ibc_refcount)); \
866 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
867 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
868 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
869 list_add_tail(&(conn)->ibc_list, \
870 &kiblnd_data.kib_connd_zombies); \
871 wake_up(&kiblnd_data.kib_connd_waitq); \
872 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
876 #define kiblnd_peer_addref(peer_ni) \
878 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)++\n", \
879 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
880 atomic_read (&(peer_ni)->ibp_refcount)); \
881 atomic_inc(&(peer_ni)->ibp_refcount); \
884 #define kiblnd_peer_decref(peer_ni) \
886 CDEBUG(D_NET, "peer_ni[%p] -> %s (%d)--\n", \
887 (peer_ni), libcfs_nid2str((peer_ni)->ibp_nid), \
888 atomic_read (&(peer_ni)->ibp_refcount)); \
889 LASSERT_ATOMIC_POS(&(peer_ni)->ibp_refcount); \
890 if (atomic_dec_and_test(&(peer_ni)->ibp_refcount)) \
891 kiblnd_destroy_peer(peer_ni); \
895 kiblnd_peer_connecting(kib_peer_ni_t *peer_ni)
897 return peer_ni->ibp_connecting != 0 ||
898 peer_ni->ibp_reconnecting != 0 ||
899 peer_ni->ibp_accepting != 0;
903 kiblnd_peer_idle(kib_peer_ni_t *peer_ni)
905 return !kiblnd_peer_connecting(peer_ni) && list_empty(&peer_ni->ibp_conns);
908 static inline struct list_head *
909 kiblnd_nid2peerlist (lnet_nid_t nid)
912 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
914 return &kiblnd_data.kib_peers[hash];
918 kiblnd_peer_active (kib_peer_ni_t *peer_ni)
920 /* Am I in the peer_ni hash table? */
921 return !list_empty(&peer_ni->ibp_list);
924 static inline struct kib_conn *
925 kiblnd_get_conn_locked (kib_peer_ni_t *peer_ni)
927 struct list_head *next;
929 LASSERT(!list_empty(&peer_ni->ibp_conns));
931 /* Advance to next connection, be sure to skip the head node */
932 if (!peer_ni->ibp_next_conn ||
933 peer_ni->ibp_next_conn->ibc_list.next == &peer_ni->ibp_conns)
934 next = peer_ni->ibp_conns.next;
936 next = peer_ni->ibp_next_conn->ibc_list.next;
937 peer_ni->ibp_next_conn = list_entry(next, struct kib_conn, ibc_list);
939 return peer_ni->ibp_next_conn;
943 kiblnd_send_keepalive(kib_conn_t *conn)
945 return (*kiblnd_tunables.kib_keepalive > 0) &&
946 cfs_time_after(jiffies, conn->ibc_last_send +
947 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
952 kiblnd_need_noop(kib_conn_t *conn)
954 struct lnet_ni *ni = conn->ibc_peer->ibp_ni;
955 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
957 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
958 tunables = &ni->ni_lnd_tunables.lnd_tun_u.lnd_o2ib;
960 if (conn->ibc_outstanding_credits <
961 IBLND_CREDITS_HIGHWATER(tunables, conn->ibc_version) &&
962 !kiblnd_send_keepalive(conn))
963 return 0; /* No need to send NOOP */
965 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
966 if (!list_empty(&conn->ibc_tx_queue_nocred))
967 return 0; /* NOOP can be piggybacked */
969 /* No tx to piggyback NOOP onto or no credit to send a tx */
970 return (list_empty(&conn->ibc_tx_queue) ||
971 conn->ibc_credits == 0);
974 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
975 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
976 conn->ibc_credits == 0) /* no credit */
979 if (conn->ibc_credits == 1 && /* last credit reserved for */
980 conn->ibc_outstanding_credits == 0) /* giving back credits */
983 /* No tx to piggyback NOOP onto or no credit to send a tx */
984 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
988 kiblnd_abort_receives(kib_conn_t *conn)
990 ib_modify_qp(conn->ibc_cmid->qp,
991 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
994 static inline const char *
995 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
997 if (q == &conn->ibc_tx_queue)
1000 if (q == &conn->ibc_tx_queue_rsrvd)
1001 return "tx_queue_rsrvd";
1003 if (q == &conn->ibc_tx_queue_nocred)
1004 return "tx_queue_nocred";
1006 if (q == &conn->ibc_active_txs)
1007 return "active_txs";
1013 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
1014 * lowest bits of the work request id to stash the work item type. */
1016 #define IBLND_WID_INVAL 0
1017 #define IBLND_WID_TX 1
1018 #define IBLND_WID_RX 2
1019 #define IBLND_WID_RDMA 3
1020 #define IBLND_WID_MR 4
1021 #define IBLND_WID_MASK 7UL
1024 kiblnd_ptr2wreqid (void *ptr, int type)
1026 unsigned long lptr = (unsigned long)ptr;
1028 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1029 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1030 return (__u64)(lptr | type);
1033 static inline void *
1034 kiblnd_wreqid2ptr (__u64 wreqid)
1036 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1040 kiblnd_wreqid2type (__u64 wreqid)
1042 return (wreqid & IBLND_WID_MASK);
1046 kiblnd_set_conn_state (kib_conn_t *conn, int state)
1048 conn->ibc_state = state;
1053 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
1055 msg->ibm_type = type;
1056 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1060 kiblnd_rd_size (kib_rdma_desc_t *rd)
1065 for (i = size = 0; i < rd->rd_nfrags; i++)
1066 size += rd->rd_frags[i].rf_nob;
1072 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
1074 return rd->rd_frags[index].rf_addr;
1078 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
1080 return rd->rd_frags[index].rf_nob;
1084 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
1090 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
1092 if (nob < rd->rd_frags[index].rf_nob) {
1093 rd->rd_frags[index].rf_addr += nob;
1094 rd->rd_frags[index].rf_nob -= nob;
1103 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
1105 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1106 msgtype == IBLND_MSG_PUT_ACK);
1108 return msgtype == IBLND_MSG_GET_REQ ?
1109 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
1110 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
1114 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1116 return ib_dma_mapping_error(dev, dma_addr);
1119 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1120 void *msg, size_t size,
1121 enum dma_data_direction direction)
1123 return ib_dma_map_single(dev, msg, size, direction);
1126 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1127 __u64 addr, size_t size,
1128 enum dma_data_direction direction)
1130 ib_dma_unmap_single(dev, addr, size, direction);
1133 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1134 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1136 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1137 struct scatterlist *sg, int nents,
1138 enum dma_data_direction direction)
1140 return ib_dma_map_sg(dev, sg, nents, direction);
1143 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1144 struct scatterlist *sg, int nents,
1145 enum dma_data_direction direction)
1147 ib_dma_unmap_sg(dev, sg, nents, direction);
1150 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1151 struct scatterlist *sg)
1153 return ib_sg_dma_address(dev, sg);
1156 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1157 struct scatterlist *sg)
1159 return ib_sg_dma_len(dev, sg);
1162 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1163 * right because OFED1.2 defines it as const, to use it we have to add
1164 * (void *) cast to overcome "const" */
1166 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1167 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1169 void kiblnd_map_rx_descs(kib_conn_t *conn);
1170 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1171 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
1172 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1174 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx,
1175 kib_rdma_desc_t *rd, __u32 nob, __u64 iov,
1177 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1179 int kiblnd_tunables_setup(struct lnet_ni *ni);
1180 int kiblnd_tunables_init(void);
1182 int kiblnd_connd (void *arg);
1183 int kiblnd_scheduler(void *arg);
1184 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1185 int kiblnd_failover_thread (void *arg);
1187 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1189 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1190 struct rdma_cm_event *event);
1191 int kiblnd_translate_mtu(int value);
1193 int kiblnd_dev_failover(kib_dev_t *dev);
1194 int kiblnd_create_peer(struct lnet_ni *ni, kib_peer_ni_t **peerp,
1196 void kiblnd_destroy_peer (kib_peer_ni_t *peer);
1197 bool kiblnd_reconnect_peer(kib_peer_ni_t *peer);
1198 void kiblnd_destroy_dev (kib_dev_t *dev);
1199 void kiblnd_unlink_peer_locked (kib_peer_ni_t *peer_ni);
1200 kib_peer_ni_t *kiblnd_find_peer_locked(struct lnet_ni *ni, lnet_nid_t nid);
1201 int kiblnd_close_stale_conns_locked (kib_peer_ni_t *peer_ni,
1202 int version, __u64 incarnation);
1203 int kiblnd_close_peer_conns_locked (kib_peer_ni_t *peer_ni, int why);
1205 kib_conn_t *kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
1206 int state, int version);
1207 void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn);
1208 void kiblnd_close_conn (kib_conn_t *conn, int error);
1209 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1211 void kiblnd_launch_tx(struct lnet_ni *ni, kib_tx_t *tx, lnet_nid_t nid);
1212 void kiblnd_txlist_done(struct list_head *txlist, int status);
1214 void kiblnd_qp_event(struct ib_event *event, void *arg);
1215 void kiblnd_cq_event(struct ib_event *event, void *arg);
1216 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1218 void kiblnd_pack_msg(struct lnet_ni *ni, kib_msg_t *msg, int version,
1219 int credits, lnet_nid_t dstnid, __u64 dststamp);
1220 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1221 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1223 int kiblnd_send(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg);
1224 int kiblnd_recv(struct lnet_ni *ni, void *private, struct lnet_msg *lntmsg,
1225 int delayed, unsigned int niov, struct kvec *iov,
1226 lnet_kiov_t *kiov, unsigned int offset, unsigned int mlen,