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_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
143 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
145 /************************/
146 /* derived constants... */
147 /* Pools (shared by connections on each CPT) */
148 /* These pools can grow at runtime, so don't need give a very large value */
149 #define IBLND_TX_POOL 256
150 #define IBLND_FMR_POOL 256
151 #define IBLND_FMR_POOL_FLUSH 192
153 /* RX messages (per connection) */
154 #define IBLND_RX_MSGS(c) \
155 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
156 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
157 #define IBLND_RX_MSG_PAGES(c) \
158 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
160 /* WRs and CQEs (per connection) */
161 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
162 #define IBLND_SEND_WRS(c) \
163 ((c->ibc_max_frags + 1) * kiblnd_concurrent_sends(c->ibc_version, \
164 c->ibc_peer->ibp_ni))
165 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + IBLND_SEND_WRS(c))
169 /* o2iblnd can run over aliased interface */
171 #define KIB_IFNAME_SIZE IFALIASZ
173 #define KIB_IFNAME_SIZE 256
178 struct list_head ibd_list; /* chain on kib_devs */
179 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
180 __u32 ibd_ifip; /* IPoIB interface IP */
181 /** IPoIB interface name */
182 char ibd_ifname[KIB_IFNAME_SIZE];
183 int ibd_nnets; /* # nets extant */
185 cfs_time_t ibd_next_failover;
186 /* # failover failures */
187 int ibd_failed_failover;
188 /* failover in progress */
189 unsigned int ibd_failover;
190 /* IPoIB interface is a bonding master */
191 unsigned int ibd_can_failover;
192 struct list_head ibd_nets;
193 struct kib_hca_dev *ibd_hdev;
196 typedef struct kib_hca_dev
198 struct rdma_cm_id *ibh_cmid; /* listener cmid */
199 struct ib_device *ibh_ibdev; /* IB device */
200 int ibh_page_shift; /* page shift of current HCA */
201 int ibh_page_size; /* page size of current HCA */
202 __u64 ibh_page_mask; /* page mask of current HCA */
203 int ibh_mr_shift; /* bits shift of max MR size */
204 __u64 ibh_mr_size; /* size of MR */
205 struct ib_mr *ibh_mrs; /* global MR */
206 struct ib_pd *ibh_pd; /* PD */
207 kib_dev_t *ibh_dev; /* owner */
208 atomic_t ibh_ref; /* refcount */
211 /** # of seconds to keep pool alive */
212 #define IBLND_POOL_DEADLINE 300
213 /** # of seconds to retry if allocation failed */
214 #define IBLND_POOL_RETRY 1
218 int ibp_npages; /* # pages */
219 struct page *ibp_pages[0]; /* page array */
225 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
226 int inc, struct kib_pool **pp_po);
227 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
228 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
229 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
233 #define IBLND_POOL_NAME_LEN 32
235 typedef struct kib_poolset
239 /* network it belongs to */
240 struct kib_net *ps_net;
242 char ps_name[IBLND_POOL_NAME_LEN];
244 struct list_head ps_pool_list;
245 /* failed pool list */
246 struct list_head ps_failed_pool_list;
247 /* time stamp for retry if failed to allocate */
248 cfs_time_t ps_next_retry;
249 /* is allocating new pool */
256 /* create a new pool */
257 kib_ps_pool_create_t ps_pool_create;
259 kib_ps_pool_destroy_t ps_pool_destroy;
260 /* initialize new allocated node */
261 kib_ps_node_init_t ps_node_init;
263 kib_ps_node_fini_t ps_node_fini;
266 typedef struct kib_pool
268 /* chain on pool list */
269 struct list_head po_list;
270 /* pre-allocated node */
271 struct list_head po_free_list;
272 /* pool_set of this pool */
273 kib_poolset_t *po_owner;
274 /* deadline of this pool */
275 cfs_time_t po_deadline;
276 /* # of elements in use */
278 /* pool is created on failed HCA */
280 /* # of pre-allocated elements */
285 kib_poolset_t tps_poolset; /* pool-set */
286 __u64 tps_next_tx_cookie; /* cookie of TX */
290 kib_pool_t tpo_pool; /* pool */
291 struct kib_hca_dev *tpo_hdev; /* device for this pool */
292 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
293 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
298 spinlock_t fps_lock; /* serialize */
299 struct kib_net *fps_net; /* IB network */
300 struct list_head fps_pool_list; /* FMR pool list */
301 struct list_head fps_failed_pool_list; /* FMR pool list */
302 __u64 fps_version; /* validity stamp */
303 int fps_cpt; /* CPT id */
305 int fps_flush_trigger;
307 /* is allocating new pool */
309 /* time stamp for retry if failed to allocate */
310 cfs_time_t fps_next_retry;
313 struct kib_fast_reg_descriptor { /* For fast registration */
314 struct list_head frd_list;
315 struct ib_send_wr frd_inv_wr;
316 struct ib_send_wr frd_fastreg_wr;
317 struct ib_mr *frd_mr;
318 struct ib_fast_reg_page_list *frd_frpl;
324 struct list_head fpo_list; /* chain on pool list */
325 struct kib_hca_dev *fpo_hdev; /* device for this pool */
326 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
329 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
331 struct { /* For fast registration */
332 struct list_head fpo_pool_list;
336 cfs_time_t fpo_deadline; /* deadline of this pool */
337 int fpo_failed; /* fmr pool is failed */
338 int fpo_map_count; /* # of mapped FMR */
343 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
344 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
345 struct kib_fast_reg_descriptor *fmr_frd;
349 typedef struct kib_net
351 /* chain on kib_dev_t::ibd_nets */
352 struct list_head ibn_list;
353 __u64 ibn_incarnation;/* my epoch */
354 int ibn_init; /* initialisation state */
355 int ibn_shutdown; /* shutting down? */
357 atomic_t ibn_npeers; /* # peers extant */
358 atomic_t ibn_nconns; /* # connections extant */
360 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
361 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
363 kib_dev_t *ibn_dev; /* underlying IB device */
366 #define KIB_THREAD_SHIFT 16
367 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
368 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
369 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
371 struct kib_sched_info {
374 /* schedulers sleep here */
375 wait_queue_head_t ibs_waitq;
376 /* conns to check for rx completions */
377 struct list_head ibs_conns;
378 /* number of scheduler threads */
380 /* max allowed scheduler threads */
381 int ibs_nthreads_max;
382 int ibs_cpt; /* CPT id */
387 int kib_init; /* initialisation state */
388 int kib_shutdown; /* shut down? */
389 struct list_head kib_devs; /* IB devices extant */
390 /* list head of failed devices */
391 struct list_head kib_failed_devs;
392 /* schedulers sleep here */
393 wait_queue_head_t kib_failover_waitq;
394 atomic_t kib_nthreads; /* # live threads */
395 /* stabilize net/dev/peer/conn ops */
396 rwlock_t kib_global_lock;
397 /* hash table of all my known peers */
398 struct list_head *kib_peers;
399 /* size of kib_peers */
400 int kib_peer_hash_size;
401 /* the connd task (serialisation assertions) */
403 /* connections to setup/teardown */
404 struct list_head kib_connd_conns;
405 /* connections with zero refcount */
406 struct list_head kib_connd_zombies;
407 /* connections to reconnect */
408 struct list_head kib_reconn_list;
409 /* peers wait for reconnection */
410 struct list_head kib_reconn_wait;
412 * The second that peers are pulled out from \a kib_reconn_wait
415 unsigned int kib_reconn_sec;
416 /* connection daemon sleeps here */
417 wait_queue_head_t kib_connd_waitq;
418 spinlock_t kib_connd_lock; /* serialise */
419 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
420 /* percpt data for schedulers */
421 struct kib_sched_info **kib_scheds;
424 #define IBLND_INIT_NOTHING 0
425 #define IBLND_INIT_DATA 1
426 #define IBLND_INIT_ALL 2
428 /************************************************************************
429 * IB Wire message format.
430 * These are sent in sender's byte order (i.e. receiver flips).
433 typedef struct kib_connparams
435 __u16 ibcp_queue_depth;
436 __u16 ibcp_max_frags;
437 __u32 ibcp_max_msg_size;
438 } WIRE_ATTR kib_connparams_t;
442 lnet_hdr_t ibim_hdr; /* portals header */
443 char ibim_payload[0]; /* piggy-backed payload */
444 } WIRE_ATTR kib_immediate_msg_t;
448 __u32 rf_nob; /* # bytes this frag */
449 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
450 } WIRE_ATTR kib_rdma_frag_t;
454 __u32 rd_key; /* local/remote key */
455 __u32 rd_nfrags; /* # fragments */
456 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
457 } WIRE_ATTR kib_rdma_desc_t;
461 lnet_hdr_t ibprm_hdr; /* portals header */
462 __u64 ibprm_cookie; /* opaque completion cookie */
463 } WIRE_ATTR kib_putreq_msg_t;
467 __u64 ibpam_src_cookie; /* reflected completion cookie */
468 __u64 ibpam_dst_cookie; /* opaque completion cookie */
469 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
470 } WIRE_ATTR kib_putack_msg_t;
474 lnet_hdr_t ibgm_hdr; /* portals header */
475 __u64 ibgm_cookie; /* opaque completion cookie */
476 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
477 } WIRE_ATTR kib_get_msg_t;
481 __u64 ibcm_cookie; /* opaque completion cookie */
482 __s32 ibcm_status; /* < 0 failure: >= 0 length */
483 } WIRE_ATTR kib_completion_msg_t;
487 /* First 2 fields fixed FOR ALL TIME */
488 __u32 ibm_magic; /* I'm an ibnal message */
489 __u16 ibm_version; /* this is my version number */
491 __u8 ibm_type; /* msg type */
492 __u8 ibm_credits; /* returned credits */
493 __u32 ibm_nob; /* # bytes in whole message */
494 __u32 ibm_cksum; /* checksum (0 == no checksum) */
495 __u64 ibm_srcnid; /* sender's NID */
496 __u64 ibm_srcstamp; /* sender's incarnation */
497 __u64 ibm_dstnid; /* destination's NID */
498 __u64 ibm_dststamp; /* destination's incarnation */
501 kib_connparams_t connparams;
502 kib_immediate_msg_t immediate;
503 kib_putreq_msg_t putreq;
504 kib_putack_msg_t putack;
506 kib_completion_msg_t completion;
508 } WIRE_ATTR kib_msg_t;
510 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
512 #define IBLND_MSG_VERSION_1 0x11
513 #define IBLND_MSG_VERSION_2 0x12
514 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
516 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
517 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
518 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
519 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
520 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
521 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
522 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
523 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
524 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
525 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
528 __u32 ibr_magic; /* sender's magic */
529 __u16 ibr_version; /* sender's version */
530 __u8 ibr_why; /* reject reason */
531 __u8 ibr_padding; /* padding */
532 __u64 ibr_incarnation; /* incarnation of peer */
533 kib_connparams_t ibr_cp; /* connection parameters */
534 } WIRE_ATTR kib_rej_t;
536 /* connection rejection reasons */
537 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
538 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
539 #define IBLND_REJECT_FATAL 3 /* Anything else */
541 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
542 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
544 /* peer's rdma frags doesn't match mine */
545 #define IBLND_REJECT_RDMA_FRAGS 6
546 /* peer's msg queue size doesn't match mine */
547 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
549 /***********************************************************************/
551 typedef struct kib_rx /* receive message */
553 /* queue for attention */
554 struct list_head rx_list;
556 struct kib_conn *rx_conn;
557 /* # bytes received (-1 while posted) */
559 /* completion status */
560 enum ib_wc_status rx_status;
561 /* message buffer (host vaddr) */
563 /* message buffer (I/O addr) */
565 /* for dma_unmap_single() */
566 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap);
567 /* receive work item... */
568 struct ib_recv_wr rx_wrq;
569 /* ...and its memory */
570 struct ib_sge rx_sge;
573 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
574 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
575 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
576 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
578 typedef struct kib_tx /* transmit message */
580 /* queue on idle_txs ibc_tx_queue etc. */
581 struct list_head tx_list;
583 kib_tx_pool_t *tx_pool;
585 struct kib_conn *tx_conn;
586 /* # tx callbacks outstanding */
588 /* queued for sending */
590 /* waiting for peer */
592 /* LNET completion status */
594 /* completion deadline */
595 unsigned long tx_deadline;
596 /* completion cookie */
598 /* lnet msgs to finalize on completion */
599 lnet_msg_t *tx_lntmsg[2];
600 /* message buffer (host vaddr) */
602 /* message buffer (I/O addr) */
604 /* for dma_unmap_single() */
605 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap);
606 /* # send work items */
608 /* send work items... */
609 struct ib_send_wr *tx_wrq;
610 /* ...and their memory */
611 struct ib_sge *tx_sge;
612 /* rdma descriptor */
613 kib_rdma_desc_t *tx_rd;
614 /* # entries in... */
616 /* dma_map_sg descriptor */
617 struct scatterlist *tx_frags;
618 /* rdma phys page addrs */
626 typedef struct kib_connvars
628 /* connection-in-progress variables */
632 typedef struct kib_conn
634 /* scheduler information */
635 struct kib_sched_info *ibc_sched;
637 struct kib_peer *ibc_peer;
639 kib_hca_dev_t *ibc_hdev;
640 /* stash on peer's conn list */
641 struct list_head ibc_list;
642 /* schedule for attention */
643 struct list_head ibc_sched_list;
644 /* version of connection */
646 /* reconnect later */
647 __u16 ibc_reconnect:1;
648 /* which instance of the peer */
649 __u64 ibc_incarnation;
651 atomic_t ibc_refcount;
652 /* what's happening */
654 /* # uncompleted sends */
655 int ibc_nsends_posted;
656 /* # uncompleted NOOPs */
657 int ibc_noops_posted;
658 /* # credits I have */
660 /* # credits to return */
661 int ibc_outstanding_credits;
662 /* # ACK/DONE msg credits */
663 int ibc_reserved_credits;
664 /* set on comms error */
666 /* connections queue depth */
667 __u16 ibc_queue_depth;
668 /* connections max frags */
670 /* receive buffers owned */
671 unsigned int ibc_nrx:16;
672 /* scheduled for attention */
673 unsigned int ibc_scheduled:1;
674 /* CQ callback fired */
675 unsigned int ibc_ready:1;
676 /* time of last send */
677 unsigned long ibc_last_send;
678 /** link chain for kiblnd_check_conns only */
679 struct list_head ibc_connd_list;
680 /** rxs completed before ESTABLISHED */
681 struct list_head ibc_early_rxs;
682 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
683 struct list_head ibc_tx_noops;
684 /* sends that need a credit */
685 struct list_head ibc_tx_queue;
686 /* sends that don't need a credit */
687 struct list_head ibc_tx_queue_nocred;
688 /* sends that need to reserve an ACK/DONE msg */
689 struct list_head ibc_tx_queue_rsrvd;
690 /* active tx awaiting completion */
691 struct list_head ibc_active_txs;
696 /* premapped rx msg pages */
697 kib_pages_t *ibc_rx_pages;
700 struct rdma_cm_id *ibc_cmid;
701 /* completion queue */
702 struct ib_cq *ibc_cq;
704 /* in-progress connection state */
705 kib_connvars_t *ibc_connvars;
708 #define IBLND_CONN_INIT 0 /* being initialised */
709 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
710 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
711 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
712 #define IBLND_CONN_CLOSING 4 /* being closed */
713 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
715 typedef struct kib_peer
717 /* stash on global peer list */
718 struct list_head ibp_list;
719 /* who's on the other end(s) */
723 /* all active connections */
724 struct list_head ibp_conns;
725 /* msgs waiting for a conn */
726 struct list_head ibp_tx_queue;
727 /* incarnation of peer */
728 __u64 ibp_incarnation;
729 /* when (in jiffies) I was last alive */
730 cfs_time_t ibp_last_alive;
732 atomic_t ibp_refcount;
733 /* version of peer */
735 /* current passive connection attempts */
736 unsigned short ibp_accepting;
737 /* current active connection attempts */
738 unsigned short ibp_connecting;
739 /* reconnect this peer later */
740 unsigned short ibp_reconnecting:1;
741 /* # consecutive reconnection attempts to this peer */
742 unsigned int ibp_reconnected;
743 /* errno on closing this peer */
745 /* max map_on_demand */
747 /* max_peer_credits */
748 __u16 ibp_queue_depth;
751 #ifndef HAVE_IB_INC_RKEY
753 * ib_inc_rkey - increments the key portion of the given rkey. Can be used
754 * for calculating a new rkey for type 2 memory windows.
755 * @rkey - the rkey to increment.
757 static inline u32 ib_inc_rkey(u32 rkey)
759 const u32 mask = 0x000000ff;
760 return ((rkey + 1) & mask) | (rkey & ~mask);
764 extern kib_data_t kiblnd_data;
766 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
768 int kiblnd_msg_queue_size(int version, struct lnet_ni *ni);
770 /* max # of fragments configured by user */
772 kiblnd_cfg_rdma_frags(struct lnet_ni *ni)
774 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
777 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
778 mod = tunables->lnd_map_on_demand;
779 return mod != 0 ? mod : IBLND_MAX_RDMA_FRAGS;
783 kiblnd_rdma_frags(int version, struct lnet_ni *ni)
785 return version == IBLND_MSG_VERSION_1 ?
786 IBLND_MAX_RDMA_FRAGS :
787 kiblnd_cfg_rdma_frags(ni);
791 kiblnd_concurrent_sends(int version, struct lnet_ni *ni)
793 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
794 int concurrent_sends;
796 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
797 concurrent_sends = tunables->lnd_concurrent_sends;
799 if (version == IBLND_MSG_VERSION_1) {
800 if (concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
801 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
803 if (concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
804 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
807 return concurrent_sends;
811 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
813 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
814 atomic_inc(&hdev->ibh_ref);
818 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
820 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
821 if (atomic_dec_and_test(&hdev->ibh_ref))
822 kiblnd_hdev_destroy(hdev);
826 kiblnd_dev_can_failover(kib_dev_t *dev)
828 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
831 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
834 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
837 return dev->ibd_can_failover;
840 #define kiblnd_conn_addref(conn) \
842 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
843 (conn), atomic_read(&(conn)->ibc_refcount)); \
844 atomic_inc(&(conn)->ibc_refcount); \
847 #define kiblnd_conn_decref(conn) \
849 unsigned long flags; \
851 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
852 (conn), atomic_read(&(conn)->ibc_refcount)); \
853 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
854 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
855 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
856 list_add_tail(&(conn)->ibc_list, \
857 &kiblnd_data.kib_connd_zombies); \
858 wake_up(&kiblnd_data.kib_connd_waitq); \
859 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
863 #define kiblnd_peer_addref(peer) \
865 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
866 (peer), libcfs_nid2str((peer)->ibp_nid), \
867 atomic_read (&(peer)->ibp_refcount)); \
868 atomic_inc(&(peer)->ibp_refcount); \
871 #define kiblnd_peer_decref(peer) \
873 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
874 (peer), libcfs_nid2str((peer)->ibp_nid), \
875 atomic_read (&(peer)->ibp_refcount)); \
876 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
877 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
878 kiblnd_destroy_peer(peer); \
882 kiblnd_peer_connecting(kib_peer_t *peer)
884 return peer->ibp_connecting != 0 ||
885 peer->ibp_reconnecting != 0 ||
886 peer->ibp_accepting != 0;
890 kiblnd_peer_idle(kib_peer_t *peer)
892 return !kiblnd_peer_connecting(peer) && list_empty(&peer->ibp_conns);
895 static inline struct list_head *
896 kiblnd_nid2peerlist (lnet_nid_t nid)
899 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
901 return &kiblnd_data.kib_peers[hash];
905 kiblnd_peer_active (kib_peer_t *peer)
907 /* Am I in the peer hash table? */
908 return !list_empty(&peer->ibp_list);
911 static inline kib_conn_t *
912 kiblnd_get_conn_locked (kib_peer_t *peer)
914 LASSERT(!list_empty(&peer->ibp_conns));
916 /* just return the first connection */
917 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
921 kiblnd_send_keepalive(kib_conn_t *conn)
923 return (*kiblnd_tunables.kib_keepalive > 0) &&
924 cfs_time_after(jiffies, conn->ibc_last_send +
925 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
930 kiblnd_need_noop(kib_conn_t *conn)
932 lnet_ni_t *ni = conn->ibc_peer->ibp_ni;
933 struct lnet_ioctl_config_o2iblnd_tunables *tunables;
935 LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
936 tunables = &ni->ni_lnd_tunables->lt_tun_u.lt_o2ib;
938 if (conn->ibc_outstanding_credits <
939 IBLND_CREDITS_HIGHWATER(tunables, conn->ibc_version) &&
940 !kiblnd_send_keepalive(conn))
941 return 0; /* No need to send NOOP */
943 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
944 if (!list_empty(&conn->ibc_tx_queue_nocred))
945 return 0; /* NOOP can be piggybacked */
947 /* No tx to piggyback NOOP onto or no credit to send a tx */
948 return (list_empty(&conn->ibc_tx_queue) ||
949 conn->ibc_credits == 0);
952 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
953 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
954 conn->ibc_credits == 0) /* no credit */
957 if (conn->ibc_credits == 1 && /* last credit reserved for */
958 conn->ibc_outstanding_credits == 0) /* giving back credits */
961 /* No tx to piggyback NOOP onto or no credit to send a tx */
962 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
966 kiblnd_abort_receives(kib_conn_t *conn)
968 ib_modify_qp(conn->ibc_cmid->qp,
969 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
972 static inline const char *
973 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
975 if (q == &conn->ibc_tx_queue)
978 if (q == &conn->ibc_tx_queue_rsrvd)
979 return "tx_queue_rsrvd";
981 if (q == &conn->ibc_tx_queue_nocred)
982 return "tx_queue_nocred";
984 if (q == &conn->ibc_active_txs)
991 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
992 * lowest bits of the work request id to stash the work item type. */
994 #define IBLND_WID_INVAL 0
995 #define IBLND_WID_TX 1
996 #define IBLND_WID_RX 2
997 #define IBLND_WID_RDMA 3
998 #define IBLND_WID_MR 4
999 #define IBLND_WID_MASK 7UL
1002 kiblnd_ptr2wreqid (void *ptr, int type)
1004 unsigned long lptr = (unsigned long)ptr;
1006 LASSERT ((lptr & IBLND_WID_MASK) == 0);
1007 LASSERT ((type & ~IBLND_WID_MASK) == 0);
1008 return (__u64)(lptr | type);
1011 static inline void *
1012 kiblnd_wreqid2ptr (__u64 wreqid)
1014 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
1018 kiblnd_wreqid2type (__u64 wreqid)
1020 return (wreqid & IBLND_WID_MASK);
1024 kiblnd_set_conn_state (kib_conn_t *conn, int state)
1026 conn->ibc_state = state;
1031 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
1033 msg->ibm_type = type;
1034 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1038 kiblnd_rd_size (kib_rdma_desc_t *rd)
1043 for (i = size = 0; i < rd->rd_nfrags; i++)
1044 size += rd->rd_frags[i].rf_nob;
1050 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
1052 return rd->rd_frags[index].rf_addr;
1056 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
1058 return rd->rd_frags[index].rf_nob;
1062 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
1068 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
1070 if (nob < rd->rd_frags[index].rf_nob) {
1071 rd->rd_frags[index].rf_addr += nob;
1072 rd->rd_frags[index].rf_nob -= nob;
1081 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
1083 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1084 msgtype == IBLND_MSG_PUT_ACK);
1086 return msgtype == IBLND_MSG_GET_REQ ?
1087 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
1088 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
1092 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1094 return ib_dma_mapping_error(dev, dma_addr);
1097 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1098 void *msg, size_t size,
1099 enum dma_data_direction direction)
1101 return ib_dma_map_single(dev, msg, size, direction);
1104 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1105 __u64 addr, size_t size,
1106 enum dma_data_direction direction)
1108 ib_dma_unmap_single(dev, addr, size, direction);
1111 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1112 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1114 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1115 struct scatterlist *sg, int nents,
1116 enum dma_data_direction direction)
1118 return ib_dma_map_sg(dev, sg, nents, direction);
1121 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1122 struct scatterlist *sg, int nents,
1123 enum dma_data_direction direction)
1125 ib_dma_unmap_sg(dev, sg, nents, direction);
1128 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1129 struct scatterlist *sg)
1131 return ib_sg_dma_address(dev, sg);
1134 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1135 struct scatterlist *sg)
1137 return ib_sg_dma_len(dev, sg);
1140 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1141 * right because OFED1.2 defines it as const, to use it we have to add
1142 * (void *) cast to overcome "const" */
1144 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1145 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1147 struct ib_mr *kiblnd_find_rd_dma_mr(struct lnet_ni *ni, kib_rdma_desc_t *rd,
1148 int negotiated_nfrags);
1149 void kiblnd_map_rx_descs(kib_conn_t *conn);
1150 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1151 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
1152 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1154 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages, int npages,
1155 __u32 nob, __u64 iov, bool is_rx, kib_fmr_t *fmr);
1156 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1158 int kiblnd_tunables_setup(struct lnet_ni *ni);
1159 int kiblnd_tunables_init(void);
1161 int kiblnd_connd (void *arg);
1162 int kiblnd_scheduler(void *arg);
1163 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1164 int kiblnd_failover_thread (void *arg);
1166 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1168 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1169 struct rdma_cm_event *event);
1170 int kiblnd_translate_mtu(int value);
1172 int kiblnd_dev_failover(kib_dev_t *dev);
1173 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1174 void kiblnd_destroy_peer (kib_peer_t *peer);
1175 bool kiblnd_reconnect_peer(kib_peer_t *peer);
1176 void kiblnd_destroy_dev (kib_dev_t *dev);
1177 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1178 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1179 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1180 int version, __u64 incarnation);
1181 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1183 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
1184 int state, int version);
1185 void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn);
1186 void kiblnd_close_conn (kib_conn_t *conn, int error);
1187 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1189 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1190 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status);
1192 void kiblnd_qp_event(struct ib_event *event, void *arg);
1193 void kiblnd_cq_event(struct ib_event *event, void *arg);
1194 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1196 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1197 int credits, lnet_nid_t dstnid, __u64 dststamp);
1198 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1199 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1201 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1202 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1203 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1204 unsigned int offset, unsigned int mlen, unsigned int rlen);