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 int *kib_require_priv_port;/* accept only privileged ports */
107 int *kib_use_priv_port; /* use privileged port for active connect */
108 /* # threads on each CPT */
112 extern kib_tunables_t kiblnd_tunables;
114 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
115 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
117 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
118 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
120 /* when eagerly to return credits */
121 #define IBLND_CREDITS_HIGHWATER(t, v) ((v) == IBLND_MSG_VERSION_1 ? \
122 IBLND_CREDIT_HIGHWATER_V1 : \
123 t->lnd_peercredits_hiw)
125 #ifdef HAVE_RDMA_CREATE_ID_5ARG
126 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(current->nsproxy->net_ns, \
130 # ifdef HAVE_RDMA_CREATE_ID_4ARG
131 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, \
134 # define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
138 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
139 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
140 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
142 #define IBLND_FRAG_SHIFT (PAGE_SHIFT - 12) /* frag size on wire is in 4K units */
143 #define IBLND_MSG_SIZE (4 << 10) /* max size of queued messages (inc hdr) */
144 #define IBLND_MAX_RDMA_FRAGS (LNET_MAX_PAYLOAD >> 12)/* max # of fragments supported in 4K size */
146 /************************/
147 /* derived constants... */
148 /* Pools (shared by connections on each CPT) */
149 /* These pools can grow at runtime, so don't need give a very large value */
150 #define IBLND_TX_POOL 256
151 #define IBLND_FMR_POOL 256
152 #define IBLND_FMR_POOL_FLUSH 192
154 /* RX messages (per connection) */
155 #define IBLND_RX_MSGS(c) \
156 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
157 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
158 #define IBLND_RX_MSG_PAGES(c) \
159 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
161 /* WRs and CQEs (per connection) */
162 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
163 #define IBLND_SEND_WRS(c) \
164 (((c->ibc_max_frags + 1) << IBLND_FRAG_SHIFT) * \
165 kiblnd_concurrent_sends(c->ibc_version, c->ibc_peer->ibp_ni))
166 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + IBLND_SEND_WRS(c))
170 /* o2iblnd can run over aliased interface */
172 #define KIB_IFNAME_SIZE IFALIASZ
174 #define KIB_IFNAME_SIZE 256
179 struct list_head ibd_list; /* chain on kib_devs */
180 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
181 __u32 ibd_ifip; /* IPoIB interface IP */
182 /** IPoIB interface name */
183 char ibd_ifname[KIB_IFNAME_SIZE];
184 int ibd_nnets; /* # nets extant */
186 cfs_time_t ibd_next_failover;
187 /* # failover failures */
188 int ibd_failed_failover;
189 /* failover in progress */
190 unsigned int ibd_failover;
191 /* IPoIB interface is a bonding master */
192 unsigned int ibd_can_failover;
193 struct list_head ibd_nets;
194 struct kib_hca_dev *ibd_hdev;
197 typedef struct kib_hca_dev
199 struct rdma_cm_id *ibh_cmid; /* listener cmid */
200 struct ib_device *ibh_ibdev; /* IB device */
201 int ibh_page_shift; /* page shift of current HCA */
202 int ibh_page_size; /* page size of current HCA */
203 __u64 ibh_page_mask; /* page mask of current HCA */
204 int ibh_mr_shift; /* bits shift of max MR size */
205 __u64 ibh_mr_size; /* size of MR */
206 struct ib_mr *ibh_mrs; /* global MR */
207 struct ib_pd *ibh_pd; /* PD */
208 kib_dev_t *ibh_dev; /* owner */
209 atomic_t ibh_ref; /* refcount */
212 /** # of seconds to keep pool alive */
213 #define IBLND_POOL_DEADLINE 300
214 /** # of seconds to retry if allocation failed */
215 #define IBLND_POOL_RETRY 1
219 int ibp_npages; /* # pages */
220 struct page *ibp_pages[0]; /* page array */
226 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
227 int inc, struct kib_pool **pp_po);
228 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
229 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
230 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
234 #define IBLND_POOL_NAME_LEN 32
236 typedef struct kib_poolset
240 /* network it belongs to */
241 struct kib_net *ps_net;
243 char ps_name[IBLND_POOL_NAME_LEN];
245 struct list_head ps_pool_list;
246 /* failed pool list */
247 struct list_head ps_failed_pool_list;
248 /* time stamp for retry if failed to allocate */
249 cfs_time_t ps_next_retry;
250 /* is allocating new pool */
257 /* create a new pool */
258 kib_ps_pool_create_t ps_pool_create;
260 kib_ps_pool_destroy_t ps_pool_destroy;
261 /* initialize new allocated node */
262 kib_ps_node_init_t ps_node_init;
264 kib_ps_node_fini_t ps_node_fini;
267 typedef struct kib_pool
269 /* chain on pool list */
270 struct list_head po_list;
271 /* pre-allocated node */
272 struct list_head po_free_list;
273 /* pool_set of this pool */
274 kib_poolset_t *po_owner;
275 /* deadline of this pool */
276 cfs_time_t po_deadline;
277 /* # of elements in use */
279 /* pool is created on failed HCA */
281 /* # of pre-allocated elements */
286 kib_poolset_t tps_poolset; /* pool-set */
287 __u64 tps_next_tx_cookie; /* cookie of TX */
291 kib_pool_t tpo_pool; /* pool */
292 struct kib_hca_dev *tpo_hdev; /* device for this pool */
293 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
294 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
299 spinlock_t fps_lock; /* serialize */
300 struct kib_net *fps_net; /* IB network */
301 struct list_head fps_pool_list; /* FMR pool list */
302 struct list_head fps_failed_pool_list; /* FMR pool list */
303 __u64 fps_version; /* validity stamp */
304 int fps_cpt; /* CPT id */
306 int fps_flush_trigger;
308 /* is allocating new pool */
310 /* time stamp for retry if failed to allocate */
311 cfs_time_t fps_next_retry;
314 #ifndef HAVE_IB_RDMA_WR
316 struct ib_send_wr wr;
320 struct kib_fast_reg_descriptor { /* For fast registration */
321 struct list_head frd_list;
322 struct ib_rdma_wr frd_inv_wr;
323 #ifdef HAVE_IB_MAP_MR_SG
324 struct ib_reg_wr frd_fastreg_wr;
326 struct ib_rdma_wr frd_fastreg_wr;
327 struct ib_fast_reg_page_list *frd_frpl;
329 struct ib_mr *frd_mr;
335 struct list_head fpo_list; /* chain on pool list */
336 struct kib_hca_dev *fpo_hdev; /* device for this pool */
337 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
340 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
342 struct { /* For fast registration */
343 struct list_head fpo_pool_list;
347 cfs_time_t fpo_deadline; /* deadline of this pool */
348 int fpo_failed; /* fmr pool is failed */
349 int fpo_map_count; /* # of mapped FMR */
354 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
355 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
356 struct kib_fast_reg_descriptor *fmr_frd;
360 typedef struct kib_net
362 /* chain on kib_dev_t::ibd_nets */
363 struct list_head ibn_list;
364 __u64 ibn_incarnation;/* my epoch */
365 int ibn_init; /* initialisation state */
366 int ibn_shutdown; /* shutting down? */
368 atomic_t ibn_npeers; /* # peers extant */
369 atomic_t ibn_nconns; /* # connections extant */
371 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
372 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
374 kib_dev_t *ibn_dev; /* underlying IB device */
377 #define KIB_THREAD_SHIFT 16
378 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
379 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
380 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
382 struct kib_sched_info {
385 /* schedulers sleep here */
386 wait_queue_head_t ibs_waitq;
387 /* conns to check for rx completions */
388 struct list_head ibs_conns;
389 /* number of scheduler threads */
391 /* max allowed scheduler threads */
392 int ibs_nthreads_max;
393 int ibs_cpt; /* CPT id */
398 int kib_init; /* initialisation state */
399 int kib_shutdown; /* shut down? */
400 struct list_head kib_devs; /* IB devices extant */
401 /* list head of failed devices */
402 struct list_head kib_failed_devs;
403 /* schedulers sleep here */
404 wait_queue_head_t kib_failover_waitq;
405 atomic_t kib_nthreads; /* # live threads */
406 /* stabilize net/dev/peer/conn ops */
407 rwlock_t kib_global_lock;
408 /* hash table of all my known peers */
409 struct list_head *kib_peers;
410 /* size of kib_peers */
411 int kib_peer_hash_size;
412 /* the connd task (serialisation assertions) */
414 /* connections to setup/teardown */
415 struct list_head kib_connd_conns;
416 /* connections with zero refcount */
417 struct list_head kib_connd_zombies;
418 /* connections to reconnect */
419 struct list_head kib_reconn_list;
420 /* peers wait for reconnection */
421 struct list_head kib_reconn_wait;
423 * The second that peers are pulled out from \a kib_reconn_wait
426 unsigned int kib_reconn_sec;
427 /* connection daemon sleeps here */
428 wait_queue_head_t kib_connd_waitq;
429 spinlock_t kib_connd_lock; /* serialise */
430 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
431 /* percpt data for schedulers */
432 struct kib_sched_info **kib_scheds;
435 #define IBLND_INIT_NOTHING 0
436 #define IBLND_INIT_DATA 1
437 #define IBLND_INIT_ALL 2
439 /************************************************************************
440 * IB Wire message format.
441 * These are sent in sender's byte order (i.e. receiver flips).
444 typedef struct kib_connparams
446 __u16 ibcp_queue_depth;
447 __u16 ibcp_max_frags;
448 __u32 ibcp_max_msg_size;
449 } WIRE_ATTR kib_connparams_t;
453 lnet_hdr_t ibim_hdr; /* portals header */
454 char ibim_payload[0]; /* piggy-backed payload */
455 } WIRE_ATTR kib_immediate_msg_t;
459 __u32 rf_nob; /* # bytes this frag */
460 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
461 } WIRE_ATTR kib_rdma_frag_t;
465 __u32 rd_key; /* local/remote key */
466 __u32 rd_nfrags; /* # fragments */
467 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
468 } WIRE_ATTR kib_rdma_desc_t;
472 lnet_hdr_t ibprm_hdr; /* portals header */
473 __u64 ibprm_cookie; /* opaque completion cookie */
474 } WIRE_ATTR kib_putreq_msg_t;
478 __u64 ibpam_src_cookie; /* reflected completion cookie */
479 __u64 ibpam_dst_cookie; /* opaque completion cookie */
480 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
481 } WIRE_ATTR kib_putack_msg_t;
485 lnet_hdr_t ibgm_hdr; /* portals header */
486 __u64 ibgm_cookie; /* opaque completion cookie */
487 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
488 } WIRE_ATTR kib_get_msg_t;
492 __u64 ibcm_cookie; /* opaque completion cookie */
493 __s32 ibcm_status; /* < 0 failure: >= 0 length */
494 } WIRE_ATTR kib_completion_msg_t;
498 /* First 2 fields fixed FOR ALL TIME */
499 __u32 ibm_magic; /* I'm an ibnal message */
500 __u16 ibm_version; /* this is my version number */
502 __u8 ibm_type; /* msg type */
503 __u8 ibm_credits; /* returned credits */
504 __u32 ibm_nob; /* # bytes in whole message */
505 __u32 ibm_cksum; /* checksum (0 == no checksum) */
506 __u64 ibm_srcnid; /* sender's NID */
507 __u64 ibm_srcstamp; /* sender's incarnation */
508 __u64 ibm_dstnid; /* destination's NID */
509 __u64 ibm_dststamp; /* destination's incarnation */
512 kib_connparams_t connparams;
513 kib_immediate_msg_t immediate;
514 kib_putreq_msg_t putreq;
515 kib_putack_msg_t putack;
517 kib_completion_msg_t completion;
519 } WIRE_ATTR kib_msg_t;
521 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
523 #define IBLND_MSG_VERSION_1 0x11
524 #define IBLND_MSG_VERSION_2 0x12
525 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
527 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
528 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
529 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
530 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
531 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
532 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
533 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
534 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
535 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
536 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
539 __u32 ibr_magic; /* sender's magic */
540 __u16 ibr_version; /* sender's version */
541 __u8 ibr_why; /* reject reason */
542 __u8 ibr_padding; /* padding */
543 __u64 ibr_incarnation; /* incarnation of peer */
544 kib_connparams_t ibr_cp; /* connection parameters */
545 } WIRE_ATTR kib_rej_t;
547 /* connection rejection reasons */
548 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
549 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
550 #define IBLND_REJECT_FATAL 3 /* Anything else */
552 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
553 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
555 /* peer's rdma frags doesn't match mine */
556 #define IBLND_REJECT_RDMA_FRAGS 6
557 /* peer's msg queue size doesn't match mine */
558 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
560 /***********************************************************************/
562 typedef struct kib_rx /* receive message */
564 /* queue for attention */
565 struct list_head rx_list;
567 struct kib_conn *rx_conn;
568 /* # bytes received (-1 while posted) */
570 /* completion status */
571 enum ib_wc_status rx_status;
572 /* message buffer (host vaddr) */
574 /* message buffer (I/O addr) */
576 /* for dma_unmap_single() */
577 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap);
578 /* receive work item... */
579 struct ib_recv_wr rx_wrq;
580 /* ...and its memory */
581 struct ib_sge rx_sge;
584 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
585 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
586 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
587 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
589 typedef struct kib_tx /* transmit message */
591 /* queue on idle_txs ibc_tx_queue etc. */
592 struct list_head tx_list;
594 kib_tx_pool_t *tx_pool;
596 struct kib_conn *tx_conn;
597 /* # tx callbacks outstanding */
599 /* queued for sending */
601 /* waiting for peer */
603 /* LNET completion status */
605 /* completion deadline */
606 unsigned long tx_deadline;
607 /* completion cookie */
609 /* lnet msgs to finalize on completion */
610 lnet_msg_t *tx_lntmsg[2];
611 /* message buffer (host vaddr) */
613 /* message buffer (I/O addr) */
615 /* for dma_unmap_single() */
616 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap);
617 /* # send work items */
619 /* send work items... */
620 struct ib_rdma_wr *tx_wrq;
621 /* ...and their memory */
622 struct ib_sge *tx_sge;
623 /* rdma descriptor */
624 kib_rdma_desc_t *tx_rd;
625 /* # entries in... */
627 /* dma_map_sg descriptor */
628 struct scatterlist *tx_frags;
629 /* rdma phys page addrs */
637 typedef struct kib_connvars
639 /* connection-in-progress variables */
643 typedef struct kib_conn
645 /* scheduler information */
646 struct kib_sched_info *ibc_sched;
648 struct kib_peer *ibc_peer;
650 kib_hca_dev_t *ibc_hdev;
651 /* stash on peer's conn list */
652 struct list_head ibc_list;
653 /* schedule for attention */
654 struct list_head ibc_sched_list;
655 /* version of connection */
657 /* reconnect later */
658 __u16 ibc_reconnect:1;
659 /* which instance of the peer */
660 __u64 ibc_incarnation;
662 atomic_t ibc_refcount;
663 /* what's happening */
665 /* # uncompleted sends */
666 int ibc_nsends_posted;
667 /* # uncompleted NOOPs */
668 int ibc_noops_posted;
669 /* # credits I have */
671 /* # credits to return */
672 int ibc_outstanding_credits;
673 /* # ACK/DONE msg credits */
674 int ibc_reserved_credits;
675 /* set on comms error */
677 /* connections queue depth */
678 __u16 ibc_queue_depth;
679 /* connections max frags */
681 /* receive buffers owned */
682 unsigned int ibc_nrx:16;
683 /* scheduled for attention */
684 unsigned int ibc_scheduled:1;
685 /* CQ callback fired */
686 unsigned int ibc_ready:1;
687 /* time of last send */
688 unsigned long ibc_last_send;
689 /** link chain for kiblnd_check_conns only */
690 struct list_head ibc_connd_list;
691 /** rxs completed before ESTABLISHED */
692 struct list_head ibc_early_rxs;
693 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
694 struct list_head ibc_tx_noops;
695 /* sends that need a credit */
696 struct list_head ibc_tx_queue;
697 /* sends that don't need a credit */
698 struct list_head ibc_tx_queue_nocred;
699 /* sends that need to reserve an ACK/DONE msg */
700 struct list_head ibc_tx_queue_rsrvd;
701 /* active tx awaiting completion */
702 struct list_head ibc_active_txs;
707 /* premapped rx msg pages */
708 kib_pages_t *ibc_rx_pages;
711 struct rdma_cm_id *ibc_cmid;
712 /* completion queue */
713 struct ib_cq *ibc_cq;
715 /* in-progress connection state */
716 kib_connvars_t *ibc_connvars;
719 #define IBLND_CONN_INIT 0 /* being initialised */
720 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
721 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
722 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
723 #define IBLND_CONN_CLOSING 4 /* being closed */
724 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
726 typedef struct kib_peer
728 /* stash on global peer list */
729 struct list_head ibp_list;
730 /* who's on the other end(s) */
734 /* all active connections */
735 struct list_head ibp_conns;
736 /* msgs waiting for a conn */
737 struct list_head ibp_tx_queue;
738 /* incarnation of peer */
739 __u64 ibp_incarnation;
740 /* when (in jiffies) I was last alive */
741 cfs_time_t ibp_last_alive;
743 atomic_t ibp_refcount;
744 /* version of peer */
746 /* current passive connection attempts */
747 unsigned short ibp_accepting;
748 /* current active connection attempts */
749 unsigned short ibp_connecting;
750 /* reconnect this peer later */
751 unsigned short ibp_reconnecting:1;
752 /* counter of how many times we triggered a conn race */
753 unsigned char ibp_races;
754 /* # consecutive reconnection attempts to this peer */
755 unsigned int ibp_reconnected;
756 /* errno on closing this peer */
758 /* max map_on_demand */
760 /* max_peer_credits */
761 __u16 ibp_queue_depth;
764 #ifndef HAVE_IB_INC_RKEY
766 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
767 * for calculating a new rkey for type 2 memory windows.
768 * @rkey - the rkey to increment.
770 static inline u32 ib_inc_rkey(u32 rkey)
772 const u32 mask = 0x000000ff;
773 return ((rkey + 1) & mask) | (rkey & ~mask);
777 extern kib_data_t kiblnd_data;
779 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
781 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
783 /* max # of fragments configured by user */
785 kiblnd_cfg_rdma_frags(struct lnet_ni *ni)
787 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
790 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
791 mod = tunables->lnd_map_on_demand;
792 return mod != 0 ? mod : IBLND_MAX_RDMA_FRAGS >> IBLND_FRAG_SHIFT;
796 kiblnd_rdma_frags(int version, struct lnet_ni *ni)
798 return version == IBLND_MSG_VERSION_1 ?
799 (IBLND_MAX_RDMA_FRAGS >> IBLND_FRAG_SHIFT) :
800 kiblnd_cfg_rdma_frags(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->lt_tun_u.lt_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) \
878 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
879 (peer), libcfs_nid2str((peer)->ibp_nid), \
880 atomic_read (&(peer)->ibp_refcount)); \
881 atomic_inc(&(peer)->ibp_refcount); \
884 #define kiblnd_peer_decref(peer) \
886 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
887 (peer), libcfs_nid2str((peer)->ibp_nid), \
888 atomic_read (&(peer)->ibp_refcount)); \
889 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
890 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
891 kiblnd_destroy_peer(peer); \
895 kiblnd_peer_connecting(kib_peer_t *peer)
897 return peer->ibp_connecting != 0 ||
898 peer->ibp_reconnecting != 0 ||
899 peer->ibp_accepting != 0;
903 kiblnd_peer_idle(kib_peer_t *peer)
905 return !kiblnd_peer_connecting(peer) && list_empty(&peer->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_t *peer)
920 /* Am I in the peer hash table? */
921 return !list_empty(&peer->ibp_list);
924 static inline kib_conn_t *
925 kiblnd_get_conn_locked (kib_peer_t *peer)
927 LASSERT(!list_empty(&peer->ibp_conns));
929 /* just return the first connection */
930 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
934 kiblnd_send_keepalive(kib_conn_t *conn)
936 return (*kiblnd_tunables.kib_keepalive > 0) &&
937 cfs_time_after(jiffies, conn->ibc_last_send +
938 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
943 kiblnd_need_noop(kib_conn_t *conn)
945 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
946 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
948 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
949 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
951 if (conn->ibc_outstanding_credits <
952 IBLND_CREDITS_HIGHWATER(tunables, conn->ibc_version) &&
953 !kiblnd_send_keepalive(conn))
954 return 0; /* No need to send NOOP */
956 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
957 if (!list_empty(&conn->ibc_tx_queue_nocred))
958 return 0; /* NOOP can be piggybacked */
960 /* No tx to piggyback NOOP onto or no credit to send a tx */
961 return (list_empty(&conn->ibc_tx_queue) ||
962 conn->ibc_credits == 0);
965 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
966 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
967 conn->ibc_credits == 0) /* no credit */
970 if (conn->ibc_credits == 1 && /* last credit reserved for */
971 conn->ibc_outstanding_credits == 0) /* giving back credits */
974 /* No tx to piggyback NOOP onto or no credit to send a tx */
975 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
979 kiblnd_abort_receives(kib_conn_t *conn)
981 ib_modify_qp(conn->ibc_cmid->qp,
982 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
985 static inline const char *
986 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
988 if (q == &conn->ibc_tx_queue)
991 if (q == &conn->ibc_tx_queue_rsrvd)
992 return "tx_queue_rsrvd";
994 if (q == &conn->ibc_tx_queue_nocred)
995 return "tx_queue_nocred";
997 if (q == &conn->ibc_active_txs)
1004 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
1005 * lowest bits of the work request id to stash the work item type. */
1007 #define IBLND_WID_INVAL 0
1008 #define IBLND_WID_TX 1
1009 #define IBLND_WID_RX 2
1010 #define IBLND_WID_RDMA 3
1011 #define IBLND_WID_MR 4
1012 #define IBLND_WID_MASK 7UL
1015 kiblnd_ptr2wreqid (void *ptr, int type)
1017 unsigned long lptr = (unsigned long)ptr;
1019 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1020 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1021 return (__u64)(lptr | type);
1024 static inline void *
1025 kiblnd_wreqid2ptr (__u64 wreqid)
1027 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1031 kiblnd_wreqid2type (__u64 wreqid)
1033 return (wreqid & IBLND_WID_MASK);
1037 kiblnd_set_conn_state (kib_conn_t *conn, int state)
1039 conn->ibc_state = state;
1044 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
1046 msg->ibm_type = type;
1047 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1051 kiblnd_rd_size (kib_rdma_desc_t *rd)
1056 for (i = size = 0; i < rd->rd_nfrags; i++)
1057 size += rd->rd_frags[i].rf_nob;
1063 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
1065 return rd->rd_frags[index].rf_addr;
1069 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
1071 return rd->rd_frags[index].rf_nob;
1075 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
1081 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
1083 if (nob < rd->rd_frags[index].rf_nob) {
1084 rd->rd_frags[index].rf_addr += nob;
1085 rd->rd_frags[index].rf_nob -= nob;
1094 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
1096 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1097 msgtype == IBLND_MSG_PUT_ACK);
1099 return msgtype == IBLND_MSG_GET_REQ ?
1100 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
1101 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
1105 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1107 return ib_dma_mapping_error(dev, dma_addr);
1110 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1111 void *msg, size_t size,
1112 enum dma_data_direction direction)
1114 return ib_dma_map_single(dev, msg, size, direction);
1117 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1118 __u64 addr, size_t size,
1119 enum dma_data_direction direction)
1121 ib_dma_unmap_single(dev, addr, size, direction);
1124 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1125 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1127 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1128 struct scatterlist *sg, int nents,
1129 enum dma_data_direction direction)
1131 return ib_dma_map_sg(dev, sg, nents, direction);
1134 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1135 struct scatterlist *sg, int nents,
1136 enum dma_data_direction direction)
1138 ib_dma_unmap_sg(dev, sg, nents, direction);
1141 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1142 struct scatterlist *sg)
1144 return ib_sg_dma_address(dev, sg);
1147 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1148 struct scatterlist *sg)
1150 return ib_sg_dma_len(dev, sg);
1153 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1154 * right because OFED1.2 defines it as const, to use it we have to add
1155 * (void *) cast to overcome "const" */
1157 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1158 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1160 struct ib_mr *kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1161 int negotiated_nfrags);
1162 void kiblnd_map_rx_descs(kib_conn_t *conn);
1163 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1164 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
1165 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1167 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx,
1168 kib_rdma_desc_t *rd, __u32 nob, __u64 iov,
1170 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1172 int kiblnd_tunables_setup(struct lnet_ni *ni);
1173 int kiblnd_tunables_init(void);
1175 int kiblnd_connd (void *arg);
1176 int kiblnd_scheduler(void *arg);
1177 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1178 int kiblnd_failover_thread (void *arg);
1180 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1182 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1183 struct rdma_cm_event *event);
1184 int kiblnd_translate_mtu(int value);
1186 int kiblnd_dev_failover(kib_dev_t *dev);
1187 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1188 void kiblnd_destroy_peer (kib_peer_t *peer);
1189 bool kiblnd_reconnect_peer(kib_peer_t *peer);
1190 void kiblnd_destroy_dev (kib_dev_t *dev);
1191 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1192 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1193 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1194 int version, __u64 incarnation);
1195 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1197 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
1198 int state, int version);
1199 void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn);
1200 void kiblnd_close_conn (kib_conn_t *conn, int error);
1201 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1203 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1204 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status);
1206 void kiblnd_qp_event(struct ib_event *event, void *arg);
1207 void kiblnd_cq_event(struct ib_event *event, void *arg);
1208 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1210 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1211 int credits, lnet_nid_t dstnid, __u64 dststamp);
1212 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1213 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1215 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1216 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1217 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1218 unsigned int offset, unsigned int mlen, unsigned int rlen);