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>
41 #include <linux/version.h>
42 #include <linux/module.h>
43 #include <linux/kernel.h>
44 #include <linux/kthread.h>
46 #include <linux/string.h>
47 #include <linux/stat.h>
48 #include <linux/errno.h>
49 #include <linux/unistd.h>
50 #include <linux/uio.h>
52 #include <asm/uaccess.h>
55 #include <linux/init.h>
57 #include <linux/file.h>
58 #include <linux/stat.h>
59 #include <linux/list.h>
60 #include <linux/kmod.h>
61 #include <linux/sysctl.h>
62 #include <linux/pci.h>
63 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,32)
64 #include <linux/pci-dma.h>
70 #include <rdma/rdma_cm.h>
71 #include <rdma/ib_cm.h>
72 #include <rdma/ib_verbs.h>
73 #include <rdma/ib_fmr_pool.h>
75 #define DEBUG_SUBSYSTEM S_LND
77 #include <libcfs/libcfs.h>
78 #include <lnet/lnet.h>
79 #include <lnet/lib-lnet.h>
81 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
82 /* # scheduler loops before reschedule */
83 #define IBLND_RESCHED 100
85 #define IBLND_N_SCHED 2
86 #define IBLND_N_SCHED_HIGH 4
90 int *kib_dev_failover; /* HCA failover */
91 unsigned int *kib_service; /* IB service number */
92 int *kib_min_reconnect_interval; /* first failed connection retry... */
93 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
94 int *kib_cksum; /* checksum kib_msg_t? */
95 int *kib_timeout; /* comms timeout (seconds) */
96 int *kib_keepalive; /* keepalive timeout (seconds) */
97 int *kib_ntx; /* # tx descs */
98 char **kib_default_ipif; /* default IPoIB interface */
100 int *kib_rnr_retry_count;
101 int *kib_ib_mtu; /* IB MTU */
102 int *kib_require_priv_port;/* accept only privileged ports */
103 int *kib_use_priv_port; /* use privileged port for active connect */
104 /* # threads on each CPT */
108 extern kib_tunables_t kiblnd_tunables;
110 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
111 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
113 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
114 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
116 /* when eagerly to return credits */
117 #define IBLND_CREDITS_HIGHWATER(t, v) ((v) == IBLND_MSG_VERSION_1 ? \
118 IBLND_CREDIT_HIGHWATER_V1 : \
119 t->lnd_peercredits_hiw)
121 #ifdef HAVE_RDMA_CREATE_ID_5ARG
122 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(current->nsproxy->net_ns, \
126 # ifdef HAVE_RDMA_CREATE_ID_4ARG
127 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, \
130 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
134 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
135 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
136 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
138 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
139 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
141 /************************/
142 /* derived constants... */
143 /* Pools (shared by connections on each CPT) */
144 /* These pools can grow at runtime, so don't need give a very large value */
145 #define IBLND_TX_POOL 256
146 #define IBLND_FMR_POOL 256
147 #define IBLND_FMR_POOL_FLUSH 192
149 /* RX messages (per connection) */
150 #define IBLND_RX_MSGS(c) \
151 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
152 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
153 #define IBLND_RX_MSG_PAGES(c) \
154 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
156 /* WRs and CQEs (per connection) */
157 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
158 #define IBLND_SEND_WRS(c) \
159 ((c->ibc_max_frags + 1) * kiblnd_concurrent_sends(c->ibc_version, \
160 c->ibc_peer->ibp_ni))
161 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + IBLND_SEND_WRS(c))
165 /* o2iblnd can run over aliased interface */
167 #define KIB_IFNAME_SIZE IFALIASZ
169 #define KIB_IFNAME_SIZE 256
174 struct list_head ibd_list; /* chain on kib_devs */
175 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
176 __u32 ibd_ifip; /* IPoIB interface IP */
177 /** IPoIB interface name */
178 char ibd_ifname[KIB_IFNAME_SIZE];
179 int ibd_nnets; /* # nets extant */
181 cfs_time_t ibd_next_failover;
182 /* # failover failures */
183 int ibd_failed_failover;
184 /* failover in progress */
185 unsigned int ibd_failover;
186 /* IPoIB interface is a bonding master */
187 unsigned int ibd_can_failover;
188 struct list_head ibd_nets;
189 struct kib_hca_dev *ibd_hdev;
192 typedef struct kib_hca_dev
194 struct rdma_cm_id *ibh_cmid; /* listener cmid */
195 struct ib_device *ibh_ibdev; /* IB device */
196 int ibh_page_shift; /* page shift of current HCA */
197 int ibh_page_size; /* page size of current HCA */
198 __u64 ibh_page_mask; /* page mask of current HCA */
199 int ibh_mr_shift; /* bits shift of max MR size */
200 __u64 ibh_mr_size; /* size of MR */
201 struct ib_mr *ibh_mrs; /* global MR */
202 struct ib_pd *ibh_pd; /* PD */
203 kib_dev_t *ibh_dev; /* owner */
204 atomic_t ibh_ref; /* refcount */
207 /** # of seconds to keep pool alive */
208 #define IBLND_POOL_DEADLINE 300
209 /** # of seconds to retry if allocation failed */
210 #define IBLND_POOL_RETRY 1
214 int ibp_npages; /* # pages */
215 struct page *ibp_pages[0]; /* page array */
221 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
222 int inc, struct kib_pool **pp_po);
223 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
224 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
225 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
229 #define IBLND_POOL_NAME_LEN 32
231 typedef struct kib_poolset
235 /* network it belongs to */
236 struct kib_net *ps_net;
238 char ps_name[IBLND_POOL_NAME_LEN];
240 struct list_head ps_pool_list;
241 /* failed pool list */
242 struct list_head ps_failed_pool_list;
243 /* time stamp for retry if failed to allocate */
244 cfs_time_t ps_next_retry;
245 /* is allocating new pool */
252 /* create a new pool */
253 kib_ps_pool_create_t ps_pool_create;
255 kib_ps_pool_destroy_t ps_pool_destroy;
256 /* initialize new allocated node */
257 kib_ps_node_init_t ps_node_init;
259 kib_ps_node_fini_t ps_node_fini;
262 typedef struct kib_pool
264 /* chain on pool list */
265 struct list_head po_list;
266 /* pre-allocated node */
267 struct list_head po_free_list;
268 /* pool_set of this pool */
269 kib_poolset_t *po_owner;
270 /* deadline of this pool */
271 cfs_time_t po_deadline;
272 /* # of elements in use */
274 /* pool is created on failed HCA */
276 /* # of pre-allocated elements */
281 kib_poolset_t tps_poolset; /* pool-set */
282 __u64 tps_next_tx_cookie; /* cookie of TX */
286 kib_pool_t tpo_pool; /* pool */
287 struct kib_hca_dev *tpo_hdev; /* device for this pool */
288 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
289 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
294 spinlock_t fps_lock; /* serialize */
295 struct kib_net *fps_net; /* IB network */
296 struct list_head fps_pool_list; /* FMR pool list */
297 struct list_head fps_failed_pool_list; /* FMR pool list */
298 __u64 fps_version; /* validity stamp */
299 int fps_cpt; /* CPT id */
301 int fps_flush_trigger;
303 /* is allocating new pool */
305 /* time stamp for retry if failed to allocate */
306 cfs_time_t fps_next_retry;
309 #ifndef HAVE_IB_RDMA_WR
311 struct ib_send_wr wr;
315 struct kib_fast_reg_descriptor { /* For fast registration */
316 struct list_head frd_list;
317 struct ib_rdma_wr frd_inv_wr;
318 #ifdef HAVE_IB_MAP_MR_SG
319 struct ib_reg_wr frd_fastreg_wr;
321 struct ib_rdma_wr frd_fastreg_wr;
322 struct ib_fast_reg_page_list *frd_frpl;
324 struct ib_mr *frd_mr;
330 struct list_head fpo_list; /* chain on pool list */
331 struct kib_hca_dev *fpo_hdev; /* device for this pool */
332 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
335 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
337 struct { /* For fast registration */
338 struct list_head fpo_pool_list;
342 cfs_time_t fpo_deadline; /* deadline of this pool */
343 int fpo_failed; /* fmr pool is failed */
344 int fpo_map_count; /* # of mapped FMR */
349 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
350 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
351 struct kib_fast_reg_descriptor *fmr_frd;
355 typedef struct kib_net
357 /* chain on kib_dev_t::ibd_nets */
358 struct list_head ibn_list;
359 __u64 ibn_incarnation;/* my epoch */
360 int ibn_init; /* initialisation state */
361 int ibn_shutdown; /* shutting down? */
363 atomic_t ibn_npeers; /* # peers extant */
364 atomic_t ibn_nconns; /* # connections extant */
366 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
367 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
369 kib_dev_t *ibn_dev; /* underlying IB device */
372 #define KIB_THREAD_SHIFT 16
373 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
374 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
375 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
377 struct kib_sched_info {
380 /* schedulers sleep here */
381 wait_queue_head_t ibs_waitq;
382 /* conns to check for rx completions */
383 struct list_head ibs_conns;
384 /* number of scheduler threads */
386 /* max allowed scheduler threads */
387 int ibs_nthreads_max;
388 int ibs_cpt; /* CPT id */
393 int kib_init; /* initialisation state */
394 int kib_shutdown; /* shut down? */
395 struct list_head kib_devs; /* IB devices extant */
396 /* list head of failed devices */
397 struct list_head kib_failed_devs;
398 /* schedulers sleep here */
399 wait_queue_head_t kib_failover_waitq;
400 atomic_t kib_nthreads; /* # live threads */
401 /* stabilize net/dev/peer/conn ops */
402 rwlock_t kib_global_lock;
403 /* hash table of all my known peers */
404 struct list_head *kib_peers;
405 /* size of kib_peers */
406 int kib_peer_hash_size;
407 /* the connd task (serialisation assertions) */
409 /* connections to setup/teardown */
410 struct list_head kib_connd_conns;
411 /* connections with zero refcount */
412 struct list_head kib_connd_zombies;
413 /* connections to reconnect */
414 struct list_head kib_reconn_list;
415 /* peers wait for reconnection */
416 struct list_head kib_reconn_wait;
418 * The second that peers are pulled out from \a kib_reconn_wait
421 time64_t kib_reconn_sec;
422 /* connection daemon sleeps here */
423 wait_queue_head_t kib_connd_waitq;
424 spinlock_t kib_connd_lock; /* serialise */
425 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
426 /* percpt data for schedulers */
427 struct kib_sched_info **kib_scheds;
430 #define IBLND_INIT_NOTHING 0
431 #define IBLND_INIT_DATA 1
432 #define IBLND_INIT_ALL 2
434 /************************************************************************
435 * IB Wire message format.
436 * These are sent in sender's byte order (i.e. receiver flips).
439 typedef struct kib_connparams
441 __u16 ibcp_queue_depth;
442 __u16 ibcp_max_frags;
443 __u32 ibcp_max_msg_size;
444 } WIRE_ATTR kib_connparams_t;
448 lnet_hdr_t ibim_hdr; /* portals header */
449 char ibim_payload[0]; /* piggy-backed payload */
450 } WIRE_ATTR kib_immediate_msg_t;
454 __u32 rf_nob; /* # bytes this frag */
455 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
456 } WIRE_ATTR kib_rdma_frag_t;
460 __u32 rd_key; /* local/remote key */
461 __u32 rd_nfrags; /* # fragments */
462 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
463 } WIRE_ATTR kib_rdma_desc_t;
467 lnet_hdr_t ibprm_hdr; /* portals header */
468 __u64 ibprm_cookie; /* opaque completion cookie */
469 } WIRE_ATTR kib_putreq_msg_t;
473 __u64 ibpam_src_cookie; /* reflected completion cookie */
474 __u64 ibpam_dst_cookie; /* opaque completion cookie */
475 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
476 } WIRE_ATTR kib_putack_msg_t;
480 lnet_hdr_t ibgm_hdr; /* portals header */
481 __u64 ibgm_cookie; /* opaque completion cookie */
482 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
483 } WIRE_ATTR kib_get_msg_t;
487 __u64 ibcm_cookie; /* opaque completion cookie */
488 __s32 ibcm_status; /* < 0 failure: >= 0 length */
489 } WIRE_ATTR kib_completion_msg_t;
493 /* First 2 fields fixed FOR ALL TIME */
494 __u32 ibm_magic; /* I'm an ibnal message */
495 __u16 ibm_version; /* this is my version number */
497 __u8 ibm_type; /* msg type */
498 __u8 ibm_credits; /* returned credits */
499 __u32 ibm_nob; /* # bytes in whole message */
500 __u32 ibm_cksum; /* checksum (0 == no checksum) */
501 __u64 ibm_srcnid; /* sender's NID */
502 __u64 ibm_srcstamp; /* sender's incarnation */
503 __u64 ibm_dstnid; /* destination's NID */
504 __u64 ibm_dststamp; /* destination's incarnation */
507 kib_connparams_t connparams;
508 kib_immediate_msg_t immediate;
509 kib_putreq_msg_t putreq;
510 kib_putack_msg_t putack;
512 kib_completion_msg_t completion;
514 } WIRE_ATTR kib_msg_t;
516 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
518 #define IBLND_MSG_VERSION_1 0x11
519 #define IBLND_MSG_VERSION_2 0x12
520 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
522 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
523 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
524 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
525 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
526 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
527 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
528 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
529 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
530 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
531 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
534 __u32 ibr_magic; /* sender's magic */
535 __u16 ibr_version; /* sender's version */
536 __u8 ibr_why; /* reject reason */
537 __u8 ibr_padding; /* padding */
538 __u64 ibr_incarnation; /* incarnation of peer */
539 kib_connparams_t ibr_cp; /* connection parameters */
540 } WIRE_ATTR kib_rej_t;
542 /* connection rejection reasons */
543 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
544 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
545 #define IBLND_REJECT_FATAL 3 /* Anything else */
547 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
548 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
550 /* peer's rdma frags doesn't match mine */
551 #define IBLND_REJECT_RDMA_FRAGS 6
552 /* peer's msg queue size doesn't match mine */
553 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
555 /***********************************************************************/
557 typedef struct kib_rx /* receive message */
559 /* queue for attention */
560 struct list_head rx_list;
562 struct kib_conn *rx_conn;
563 /* # bytes received (-1 while posted) */
565 /* completion status */
566 enum ib_wc_status rx_status;
567 /* message buffer (host vaddr) */
569 /* message buffer (I/O addr) */
571 /* for dma_unmap_single() */
572 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap);
573 /* receive work item... */
574 struct ib_recv_wr rx_wrq;
575 /* ...and its memory */
576 struct ib_sge rx_sge;
579 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
580 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
581 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
582 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
584 typedef struct kib_tx /* transmit message */
586 /* queue on idle_txs ibc_tx_queue etc. */
587 struct list_head tx_list;
589 kib_tx_pool_t *tx_pool;
591 struct kib_conn *tx_conn;
592 /* # tx callbacks outstanding */
594 /* queued for sending */
596 /* waiting for peer */
598 /* LNET completion status */
600 /* completion deadline */
601 unsigned long tx_deadline;
602 /* completion cookie */
604 /* lnet msgs to finalize on completion */
605 lnet_msg_t *tx_lntmsg[2];
606 /* message buffer (host vaddr) */
608 /* message buffer (I/O addr) */
610 /* for dma_unmap_single() */
611 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap);
612 /* # send work items */
614 /* send work items... */
615 struct ib_rdma_wr *tx_wrq;
616 /* ...and their memory */
617 struct ib_sge *tx_sge;
618 /* rdma descriptor */
619 kib_rdma_desc_t *tx_rd;
620 /* # entries in... */
622 /* dma_map_sg descriptor */
623 struct scatterlist *tx_frags;
624 /* rdma phys page addrs */
632 typedef struct kib_connvars
634 /* connection-in-progress variables */
638 typedef struct kib_conn
640 /* scheduler information */
641 struct kib_sched_info *ibc_sched;
643 struct kib_peer *ibc_peer;
645 kib_hca_dev_t *ibc_hdev;
646 /* stash on peer's conn list */
647 struct list_head ibc_list;
648 /* schedule for attention */
649 struct list_head ibc_sched_list;
650 /* version of connection */
652 /* reconnect later */
653 __u16 ibc_reconnect:1;
654 /* which instance of the peer */
655 __u64 ibc_incarnation;
657 atomic_t ibc_refcount;
658 /* what's happening */
660 /* # uncompleted sends */
661 int ibc_nsends_posted;
662 /* # uncompleted NOOPs */
663 int ibc_noops_posted;
664 /* # credits I have */
666 /* # credits to return */
667 int ibc_outstanding_credits;
668 /* # ACK/DONE msg credits */
669 int ibc_reserved_credits;
670 /* set on comms error */
672 /* connections queue depth */
673 __u16 ibc_queue_depth;
674 /* connections max frags */
676 /* receive buffers owned */
677 unsigned int ibc_nrx:16;
678 /* scheduled for attention */
679 unsigned int ibc_scheduled:1;
680 /* CQ callback fired */
681 unsigned int ibc_ready:1;
682 /* time of last send */
683 unsigned long ibc_last_send;
684 /** link chain for kiblnd_check_conns only */
685 struct list_head ibc_connd_list;
686 /** rxs completed before ESTABLISHED */
687 struct list_head ibc_early_rxs;
688 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
689 struct list_head ibc_tx_noops;
690 /* sends that need a credit */
691 struct list_head ibc_tx_queue;
692 /* sends that don't need a credit */
693 struct list_head ibc_tx_queue_nocred;
694 /* sends that need to reserve an ACK/DONE msg */
695 struct list_head ibc_tx_queue_rsrvd;
696 /* active tx awaiting completion */
697 struct list_head ibc_active_txs;
702 /* premapped rx msg pages */
703 kib_pages_t *ibc_rx_pages;
706 struct rdma_cm_id *ibc_cmid;
707 /* completion queue */
708 struct ib_cq *ibc_cq;
710 /* in-progress connection state */
711 kib_connvars_t *ibc_connvars;
714 #define IBLND_CONN_INIT 0 /* being initialised */
715 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
716 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
717 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
718 #define IBLND_CONN_CLOSING 4 /* being closed */
719 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
721 typedef struct kib_peer
723 /* stash on global peer list */
724 struct list_head ibp_list;
725 /* who's on the other end(s) */
729 /* all active connections */
730 struct list_head ibp_conns;
731 /* msgs waiting for a conn */
732 struct list_head ibp_tx_queue;
733 /* incarnation of peer */
734 __u64 ibp_incarnation;
735 /* when (in jiffies) I was last alive */
736 cfs_time_t ibp_last_alive;
738 atomic_t ibp_refcount;
739 /* version of peer */
741 /* current passive connection attempts */
742 unsigned short ibp_accepting;
743 /* current active connection attempts */
744 unsigned short ibp_connecting;
745 /* reconnect this peer later */
746 unsigned short ibp_reconnecting:1;
747 /* counter of how many times we triggered a conn race */
748 unsigned char ibp_races;
749 /* # consecutive reconnection attempts to this peer */
750 unsigned int ibp_reconnected;
751 /* errno on closing this peer */
753 /* max map_on_demand */
755 /* max_peer_credits */
756 __u16 ibp_queue_depth;
759 #ifndef HAVE_IB_INC_RKEY
761 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
762 * for calculating a new rkey for type 2 memory windows.
763 * @rkey - the rkey to increment.
765 static inline u32 ib_inc_rkey(u32 rkey)
767 const u32 mask = 0x000000ff;
768 return ((rkey + 1) & mask) | (rkey & ~mask);
772 extern kib_data_t kiblnd_data;
774 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
776 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
778 /* max # of fragments configured by user */
780 kiblnd_cfg_rdma_frags(struct lnet_ni *ni)
782 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
785 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
786 mod = tunables->lnd_map_on_demand;
787 return mod != 0 ? mod : IBLND_MAX_RDMA_FRAGS;
791 kiblnd_rdma_frags(int version, struct lnet_ni *ni)
793 return version == IBLND_MSG_VERSION_1 ?
794 IBLND_MAX_RDMA_FRAGS :
795 kiblnd_cfg_rdma_frags(ni);
799 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
801 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
802 int concurrent_sends;
804 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
805 concurrent_sends = tunables->lnd_concurrent_sends;
807 if (version == IBLND_MSG_VERSION_1) {
808 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
809 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
811 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
812 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
815 return concurrent_sends;
819 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
821 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
822 atomic_inc(&hdev->ibh_ref);
826 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
828 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
829 if (atomic_dec_and_test(&hdev->ibh_ref))
830 kiblnd_hdev_destroy(hdev);
834 kiblnd_dev_can_failover(kib_dev_t *dev)
836 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
839 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
842 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
845 return dev->ibd_can_failover;
848 #define kiblnd_conn_addref(conn) \
850 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
851 (conn), atomic_read(&(conn)->ibc_refcount)); \
852 atomic_inc(&(conn)->ibc_refcount); \
855 #define kiblnd_conn_decref(conn) \
857 unsigned long flags; \
859 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
860 (conn), atomic_read(&(conn)->ibc_refcount)); \
861 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
862 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
863 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
864 list_add_tail(&(conn)->ibc_list, \
865 &kiblnd_data.kib_connd_zombies); \
866 wake_up(&kiblnd_data.kib_connd_waitq); \
867 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
871 #define kiblnd_peer_addref(peer) \
873 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
874 (peer), libcfs_nid2str((peer)->ibp_nid), \
875 atomic_read (&(peer)->ibp_refcount)); \
876 atomic_inc(&(peer)->ibp_refcount); \
879 #define kiblnd_peer_decref(peer) \
881 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
882 (peer), libcfs_nid2str((peer)->ibp_nid), \
883 atomic_read (&(peer)->ibp_refcount)); \
884 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
885 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
886 kiblnd_destroy_peer(peer); \
890 kiblnd_peer_connecting(kib_peer_t *peer)
892 return peer->ibp_connecting != 0 ||
893 peer->ibp_reconnecting != 0 ||
894 peer->ibp_accepting != 0;
898 kiblnd_peer_idle(kib_peer_t *peer)
900 return !kiblnd_peer_connecting(peer) && list_empty(&peer->ibp_conns);
903 static inline struct list_head *
904 kiblnd_nid2peerlist (lnet_nid_t nid)
907 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
909 return &kiblnd_data.kib_peers[hash];
913 kiblnd_peer_active (kib_peer_t *peer)
915 /* Am I in the peer hash table? */
916 return !list_empty(&peer->ibp_list);
919 static inline kib_conn_t *
920 kiblnd_get_conn_locked (kib_peer_t *peer)
922 LASSERT(!list_empty(&peer->ibp_conns));
924 /* just return the first connection */
925 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
929 kiblnd_send_keepalive(kib_conn_t *conn)
931 return (*kiblnd_tunables.kib_keepalive > 0) &&
932 cfs_time_after(jiffies, conn->ibc_last_send +
933 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
938 kiblnd_need_noop(kib_conn_t *conn)
940 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
941 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
943 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
944 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
946 if (conn->ibc_outstanding_credits <
947 IBLND_CREDITS_HIGHWATER(tunables, conn->ibc_version) &&
948 !kiblnd_send_keepalive(conn))
949 return 0; /* No need to send NOOP */
951 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
952 if (!list_empty(&conn->ibc_tx_queue_nocred))
953 return 0; /* NOOP can be piggybacked */
955 /* No tx to piggyback NOOP onto or no credit to send a tx */
956 return (list_empty(&conn->ibc_tx_queue) ||
957 conn->ibc_credits == 0);
960 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
961 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
962 conn->ibc_credits == 0) /* no credit */
965 if (conn->ibc_credits == 1 && /* last credit reserved for */
966 conn->ibc_outstanding_credits == 0) /* giving back credits */
969 /* No tx to piggyback NOOP onto or no credit to send a tx */
970 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
974 kiblnd_abort_receives(kib_conn_t *conn)
976 ib_modify_qp(conn->ibc_cmid->qp,
977 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
980 static inline const char *
981 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
983 if (q == &conn->ibc_tx_queue)
986 if (q == &conn->ibc_tx_queue_rsrvd)
987 return "tx_queue_rsrvd";
989 if (q == &conn->ibc_tx_queue_nocred)
990 return "tx_queue_nocred";
992 if (q == &conn->ibc_active_txs)
999 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
1000 * lowest bits of the work request id to stash the work item type. */
1002 #define IBLND_WID_INVAL 0
1003 #define IBLND_WID_TX 1
1004 #define IBLND_WID_RX 2
1005 #define IBLND_WID_RDMA 3
1006 #define IBLND_WID_MR 4
1007 #define IBLND_WID_MASK 7UL
1010 kiblnd_ptr2wreqid (void *ptr, int type)
1012 unsigned long lptr = (unsigned long)ptr;
1014 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1015 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1016 return (__u64)(lptr | type);
1019 static inline void *
1020 kiblnd_wreqid2ptr (__u64 wreqid)
1022 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1026 kiblnd_wreqid2type (__u64 wreqid)
1028 return (wreqid & IBLND_WID_MASK);
1032 kiblnd_set_conn_state (kib_conn_t *conn, int state)
1034 conn->ibc_state = state;
1039 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
1041 msg->ibm_type = type;
1042 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1046 kiblnd_rd_size (kib_rdma_desc_t *rd)
1051 for (i = size = 0; i < rd->rd_nfrags; i++)
1052 size += rd->rd_frags[i].rf_nob;
1058 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
1060 return rd->rd_frags[index].rf_addr;
1064 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
1066 return rd->rd_frags[index].rf_nob;
1070 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
1076 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
1078 if (nob < rd->rd_frags[index].rf_nob) {
1079 rd->rd_frags[index].rf_addr += nob;
1080 rd->rd_frags[index].rf_nob -= nob;
1089 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
1091 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1092 msgtype == IBLND_MSG_PUT_ACK);
1094 return msgtype == IBLND_MSG_GET_REQ ?
1095 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
1096 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
1100 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1102 return ib_dma_mapping_error(dev, dma_addr);
1105 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1106 void *msg, size_t size,
1107 enum dma_data_direction direction)
1109 return ib_dma_map_single(dev, msg, size, direction);
1112 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1113 __u64 addr, size_t size,
1114 enum dma_data_direction direction)
1116 ib_dma_unmap_single(dev, addr, size, direction);
1119 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1120 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1122 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1123 struct scatterlist *sg, int nents,
1124 enum dma_data_direction direction)
1126 return ib_dma_map_sg(dev, sg, nents, direction);
1129 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1130 struct scatterlist *sg, int nents,
1131 enum dma_data_direction direction)
1133 ib_dma_unmap_sg(dev, sg, nents, direction);
1136 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1137 struct scatterlist *sg)
1139 return ib_sg_dma_address(dev, sg);
1142 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1143 struct scatterlist *sg)
1145 return ib_sg_dma_len(dev, sg);
1148 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1149 * right because OFED1.2 defines it as const, to use it we have to add
1150 * (void *) cast to overcome "const" */
1152 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1153 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1155 struct ib_mr *kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1156 int negotiated_nfrags);
1157 void kiblnd_map_rx_descs(kib_conn_t *conn);
1158 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1159 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
1160 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1162 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx,
1163 kib_rdma_desc_t *rd, __u32 nob, __u64 iov,
1165 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1167 int kiblnd_tunables_setup(struct lnet_ni *ni);
1168 int kiblnd_tunables_init(void);
1170 int kiblnd_connd (void *arg);
1171 int kiblnd_scheduler(void *arg);
1172 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1173 int kiblnd_failover_thread (void *arg);
1175 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1177 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1178 struct rdma_cm_event *event);
1179 int kiblnd_translate_mtu(int value);
1181 int kiblnd_dev_failover(kib_dev_t *dev);
1182 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1183 void kiblnd_destroy_peer (kib_peer_t *peer);
1184 bool kiblnd_reconnect_peer(kib_peer_t *peer);
1185 void kiblnd_destroy_dev (kib_dev_t *dev);
1186 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1187 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1188 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1189 int version, __u64 incarnation);
1190 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1192 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
1193 int state, int version);
1194 void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn);
1195 void kiblnd_close_conn (kib_conn_t *conn, int error);
1196 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1198 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1199 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status);
1201 void kiblnd_qp_event(struct ib_event *event, void *arg);
1202 void kiblnd_cq_event(struct ib_event *event, void *arg);
1203 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1205 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1206 int credits, lnet_nid_t dstnid, __u64 dststamp);
1207 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1208 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1210 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1211 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1212 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1213 unsigned int offset, unsigned int mlen, unsigned int rlen);