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 int *kib_credits; /* # concurrent sends */
103 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
104 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
105 int *kib_peercredits_hiw; /* # when eagerly to return credits */
106 int *kib_peertimeout; /* seconds to consider peer dead */
107 char **kib_default_ipif; /* default IPoIB interface */
108 int *kib_retry_count;
109 int *kib_rnr_retry_count;
110 int *kib_concurrent_sends; /* send work queue sizing */
111 int *kib_ib_mtu; /* IB MTU */
112 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
113 * than this value, 0 disable map-on-demand */
114 int *kib_fmr_pool_size; /* # FMRs in pool */
115 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
116 int *kib_fmr_cache; /* enable FMR pool cache? */
117 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
118 struct ctl_table_header *kib_sysctl; /* sysctl interface */
120 int *kib_require_priv_port;/* accept only privileged ports */
121 int *kib_use_priv_port; /* use privileged port for active connect */
122 /* # threads on each CPT */
126 extern kib_tunables_t kiblnd_tunables;
128 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
129 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
131 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
132 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
134 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
135 IBLND_MSG_QUEUE_SIZE_V1 : \
136 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
137 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
138 IBLND_CREDIT_HIGHWATER_V1 : \
139 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
141 #ifdef HAVE_RDMA_CREATE_ID_4ARG
142 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
144 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
148 kiblnd_concurrent_sends_v1(void)
150 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
151 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
153 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
154 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
156 return *kiblnd_tunables.kib_concurrent_sends;
159 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
160 kiblnd_concurrent_sends_v1() : \
161 *kiblnd_tunables.kib_concurrent_sends)
162 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
163 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
164 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
166 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
167 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
168 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
169 *kiblnd_tunables.kib_map_on_demand : \
170 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
171 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
172 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
174 /************************/
175 /* derived constants... */
176 /* Pools (shared by connections on each CPT) */
177 /* These pools can grow at runtime, so don't need give a very large value */
178 #define IBLND_TX_POOL 256
179 #define IBLND_FMR_POOL 256
180 #define IBLND_FMR_POOL_FLUSH 192
182 /* RX messages (per connection) */
183 #define IBLND_RX_MSGS(c) \
184 ((c->ibc_queue_depth) * 2 + IBLND_OOB_MSGS(c->ibc_version))
185 #define IBLND_RX_MSG_BYTES(c) (IBLND_RX_MSGS(c) * IBLND_MSG_SIZE)
186 #define IBLND_RX_MSG_PAGES(c) \
187 ((IBLND_RX_MSG_BYTES(c) + PAGE_SIZE - 1) / PAGE_SIZE)
189 /* WRs and CQEs (per connection) */
190 #define IBLND_RECV_WRS(c) IBLND_RX_MSGS(c)
191 #define IBLND_SEND_WRS(c) \
192 ((c->ibc_max_frags + 1) * IBLND_CONCURRENT_SENDS(c->ibc_version))
193 #define IBLND_CQ_ENTRIES(c) (IBLND_RECV_WRS(c) + IBLND_SEND_WRS(c))
197 /* o2iblnd can run over aliased interface */
199 #define KIB_IFNAME_SIZE IFALIASZ
201 #define KIB_IFNAME_SIZE 256
206 struct list_head ibd_list; /* chain on kib_devs */
207 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
208 __u32 ibd_ifip; /* IPoIB interface IP */
209 /** IPoIB interface name */
210 char ibd_ifname[KIB_IFNAME_SIZE];
211 int ibd_nnets; /* # nets extant */
213 cfs_time_t ibd_next_failover;
214 /* # failover failures */
215 int ibd_failed_failover;
216 /* failover in progress */
217 unsigned int ibd_failover;
218 /* IPoIB interface is a bonding master */
219 unsigned int ibd_can_failover;
220 struct list_head ibd_nets;
221 struct kib_hca_dev *ibd_hdev;
224 typedef struct kib_hca_dev
226 struct rdma_cm_id *ibh_cmid; /* listener cmid */
227 struct ib_device *ibh_ibdev; /* IB device */
228 int ibh_page_shift; /* page shift of current HCA */
229 int ibh_page_size; /* page size of current HCA */
230 __u64 ibh_page_mask; /* page mask of current HCA */
231 int ibh_mr_shift; /* bits shift of max MR size */
232 __u64 ibh_mr_size; /* size of MR */
233 struct ib_mr *ibh_mrs; /* global MR */
234 struct ib_pd *ibh_pd; /* PD */
235 kib_dev_t *ibh_dev; /* owner */
236 atomic_t ibh_ref; /* refcount */
239 /** # of seconds to keep pool alive */
240 #define IBLND_POOL_DEADLINE 300
241 /** # of seconds to retry if allocation failed */
242 #define IBLND_POOL_RETRY 1
246 int ibp_npages; /* # pages */
247 struct page *ibp_pages[0]; /* page array */
253 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
254 int inc, struct kib_pool **pp_po);
255 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
256 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
257 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
261 #define IBLND_POOL_NAME_LEN 32
263 typedef struct kib_poolset
267 /* network it belongs to */
268 struct kib_net *ps_net;
270 char ps_name[IBLND_POOL_NAME_LEN];
272 struct list_head ps_pool_list;
273 /* failed pool list */
274 struct list_head ps_failed_pool_list;
275 /* time stamp for retry if failed to allocate */
276 cfs_time_t ps_next_retry;
277 /* is allocating new pool */
284 /* create a new pool */
285 kib_ps_pool_create_t ps_pool_create;
287 kib_ps_pool_destroy_t ps_pool_destroy;
288 /* initialize new allocated node */
289 kib_ps_node_init_t ps_node_init;
291 kib_ps_node_fini_t ps_node_fini;
294 typedef struct kib_pool
296 /* chain on pool list */
297 struct list_head po_list;
298 /* pre-allocated node */
299 struct list_head po_free_list;
300 /* pool_set of this pool */
301 kib_poolset_t *po_owner;
302 /* deadline of this pool */
303 cfs_time_t po_deadline;
304 /* # of elements in use */
306 /* pool is created on failed HCA */
308 /* # of pre-allocated elements */
313 kib_poolset_t tps_poolset; /* pool-set */
314 __u64 tps_next_tx_cookie; /* cookie of TX */
318 kib_pool_t tpo_pool; /* pool */
319 struct kib_hca_dev *tpo_hdev; /* device for this pool */
320 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
321 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
326 spinlock_t fps_lock; /* serialize */
327 struct kib_net *fps_net; /* IB network */
328 struct list_head fps_pool_list; /* FMR pool list */
329 struct list_head fps_failed_pool_list; /* FMR pool list */
330 __u64 fps_version; /* validity stamp */
331 int fps_cpt; /* CPT id */
333 int fps_flush_trigger;
334 /* is allocating new pool */
336 /* time stamp for retry if failed to allocate */
337 cfs_time_t fps_next_retry;
342 struct list_head fpo_list; /* chain on pool list */
343 struct kib_hca_dev *fpo_hdev; /* device for this pool */
344 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
345 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
346 cfs_time_t fpo_deadline; /* deadline of this pool */
347 int fpo_failed; /* fmr pool is failed */
348 int fpo_map_count; /* # of mapped FMR */
352 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
353 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
356 typedef struct kib_net
358 /* chain on kib_dev_t::ibd_nets */
359 struct list_head ibn_list;
360 __u64 ibn_incarnation;/* my epoch */
361 int ibn_init; /* initialisation state */
362 int ibn_shutdown; /* shutting down? */
364 atomic_t ibn_npeers; /* # peers extant */
365 atomic_t ibn_nconns; /* # connections extant */
367 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
368 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
370 kib_dev_t *ibn_dev; /* underlying IB device */
373 #define KIB_THREAD_SHIFT 16
374 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
375 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
376 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
378 struct kib_sched_info {
381 /* schedulers sleep here */
382 wait_queue_head_t ibs_waitq;
383 /* conns to check for rx completions */
384 struct list_head ibs_conns;
385 /* number of scheduler threads */
387 /* max allowed scheduler threads */
388 int ibs_nthreads_max;
389 int ibs_cpt; /* CPT id */
394 int kib_init; /* initialisation state */
395 int kib_shutdown; /* shut down? */
396 struct list_head kib_devs; /* IB devices extant */
397 /* list head of failed devices */
398 struct list_head kib_failed_devs;
399 /* schedulers sleep here */
400 wait_queue_head_t kib_failover_waitq;
401 atomic_t kib_nthreads; /* # live threads */
402 /* stabilize net/dev/peer/conn ops */
403 rwlock_t kib_global_lock;
404 /* hash table of all my known peers */
405 struct list_head *kib_peers;
406 /* size of kib_peers */
407 int kib_peer_hash_size;
408 /* the connd task (serialisation assertions) */
410 /* connections to setup/teardown */
411 struct list_head kib_connd_conns;
412 /* connections with zero refcount */
413 struct list_head kib_connd_zombies;
414 /* connections to reconnect */
415 struct list_head kib_reconn_list;
416 /* peers wait for reconnection */
417 struct list_head kib_reconn_wait;
419 * The second that peers are pulled out from \a kib_reconn_wait
422 unsigned int kib_reconn_sec;
423 /* connection daemon sleeps here */
424 wait_queue_head_t kib_connd_waitq;
425 spinlock_t kib_connd_lock; /* serialise */
426 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
427 /* percpt data for schedulers */
428 struct kib_sched_info **kib_scheds;
431 #define IBLND_INIT_NOTHING 0
432 #define IBLND_INIT_DATA 1
433 #define IBLND_INIT_ALL 2
435 /************************************************************************
436 * IB Wire message format.
437 * These are sent in sender's byte order (i.e. receiver flips).
440 typedef struct kib_connparams
442 __u16 ibcp_queue_depth;
443 __u16 ibcp_max_frags;
444 __u32 ibcp_max_msg_size;
445 } WIRE_ATTR kib_connparams_t;
449 lnet_hdr_t ibim_hdr; /* portals header */
450 char ibim_payload[0]; /* piggy-backed payload */
451 } WIRE_ATTR kib_immediate_msg_t;
455 __u32 rf_nob; /* # bytes this frag */
456 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
457 } WIRE_ATTR kib_rdma_frag_t;
461 __u32 rd_key; /* local/remote key */
462 __u32 rd_nfrags; /* # fragments */
463 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
464 } WIRE_ATTR kib_rdma_desc_t;
468 lnet_hdr_t ibprm_hdr; /* portals header */
469 __u64 ibprm_cookie; /* opaque completion cookie */
470 } WIRE_ATTR kib_putreq_msg_t;
474 __u64 ibpam_src_cookie; /* reflected completion cookie */
475 __u64 ibpam_dst_cookie; /* opaque completion cookie */
476 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
477 } WIRE_ATTR kib_putack_msg_t;
481 lnet_hdr_t ibgm_hdr; /* portals header */
482 __u64 ibgm_cookie; /* opaque completion cookie */
483 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
484 } WIRE_ATTR kib_get_msg_t;
488 __u64 ibcm_cookie; /* opaque completion cookie */
489 __s32 ibcm_status; /* < 0 failure: >= 0 length */
490 } WIRE_ATTR kib_completion_msg_t;
494 /* First 2 fields fixed FOR ALL TIME */
495 __u32 ibm_magic; /* I'm an ibnal message */
496 __u16 ibm_version; /* this is my version number */
498 __u8 ibm_type; /* msg type */
499 __u8 ibm_credits; /* returned credits */
500 __u32 ibm_nob; /* # bytes in whole message */
501 __u32 ibm_cksum; /* checksum (0 == no checksum) */
502 __u64 ibm_srcnid; /* sender's NID */
503 __u64 ibm_srcstamp; /* sender's incarnation */
504 __u64 ibm_dstnid; /* destination's NID */
505 __u64 ibm_dststamp; /* destination's incarnation */
508 kib_connparams_t connparams;
509 kib_immediate_msg_t immediate;
510 kib_putreq_msg_t putreq;
511 kib_putack_msg_t putack;
513 kib_completion_msg_t completion;
515 } WIRE_ATTR kib_msg_t;
517 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
519 #define IBLND_MSG_VERSION_1 0x11
520 #define IBLND_MSG_VERSION_2 0x12
521 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
523 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
524 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
525 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
526 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
527 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
528 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
529 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
530 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
531 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
532 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
534 /* flag to show a peer can understand connrace protocol */
535 #define IBLND_CONNREQ_NOOP (1U << 0)
536 /* want to win connrace */
537 #define IBLND_CONNREQ_WIN_WISH (1U << 1)
540 __u32 ibr_magic; /* sender's magic */
541 __u16 ibr_version; /* sender's version */
542 __u8 ibr_why; /* reject reason */
543 __u8 ibr_padding; /* padding */
544 __u64 ibr_incarnation; /* incarnation of peer */
545 kib_connparams_t ibr_cp; /* connection parameters */
546 } WIRE_ATTR kib_rej_t;
548 /* connection rejection reasons */
549 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
550 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
551 #define IBLND_REJECT_FATAL 3 /* Anything else */
553 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
554 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
556 /* peer's rdma frags doesn't match mine */
557 #define IBLND_REJECT_RDMA_FRAGS 6
558 /* peer's msg queue size doesn't match mine */
559 #define IBLND_REJECT_MSG_QUEUE_SIZE 7
561 /***********************************************************************/
563 typedef struct kib_rx /* receive message */
565 /* queue for attention */
566 struct list_head rx_list;
568 struct kib_conn *rx_conn;
569 /* # bytes received (-1 while posted) */
571 /* completion status */
572 enum ib_wc_status rx_status;
573 /* message buffer (host vaddr) */
575 /* message buffer (I/O addr) */
577 /* for dma_unmap_single() */
578 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap);
579 /* receive work item... */
580 struct ib_recv_wr rx_wrq;
581 /* ...and its memory */
582 struct ib_sge rx_sge;
585 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
586 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
587 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
588 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
590 typedef struct kib_tx /* transmit message */
592 /* queue on idle_txs ibc_tx_queue etc. */
593 struct list_head tx_list;
595 kib_tx_pool_t *tx_pool;
597 struct kib_conn *tx_conn;
598 /* # tx callbacks outstanding */
600 /* queued for sending */
602 /* waiting for peer */
604 /* LNET completion status */
606 /* completion deadline */
607 unsigned long tx_deadline;
608 /* completion cookie */
610 /* lnet msgs to finalize on completion */
611 lnet_msg_t *tx_lntmsg[2];
612 /* message buffer (host vaddr) */
614 /* message buffer (I/O addr) */
616 /* for dma_unmap_single() */
617 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap);
618 /* # send work items */
620 /* send work items... */
621 struct ib_send_wr *tx_wrq;
622 /* ...and their memory */
623 struct ib_sge *tx_sge;
624 /* rdma descriptor */
625 kib_rdma_desc_t *tx_rd;
626 /* # entries in... */
628 /* dma_map_sg descriptor */
629 struct scatterlist *tx_frags;
630 /* rdma phys page addrs */
638 typedef struct kib_connvars
640 /* connection-in-progress variables */
644 typedef struct kib_conn
646 /* scheduler information */
647 struct kib_sched_info *ibc_sched;
649 struct kib_peer *ibc_peer;
651 kib_hca_dev_t *ibc_hdev;
652 /* stash on peer's conn list */
653 struct list_head ibc_list;
654 /* schedule for attention */
655 struct list_head ibc_sched_list;
656 /* version of connection */
658 /* reconnect later */
659 __u16 ibc_reconnect:1;
660 /* rejected by connrace */
661 __u16 ibc_connrace:1;
662 /* which instance of the peer */
663 __u64 ibc_incarnation;
665 atomic_t ibc_refcount;
666 /* what's happening */
668 /* # uncompleted sends */
669 int ibc_nsends_posted;
670 /* # uncompleted NOOPs */
671 int ibc_noops_posted;
672 /* # credits I have */
674 /* # credits to return */
675 int ibc_outstanding_credits;
676 /* # ACK/DONE msg credits */
677 int ibc_reserved_credits;
678 /* set on comms error */
680 /* connections queue depth */
681 __u16 ibc_queue_depth;
682 /* connections max frags */
684 /* receive buffers owned */
685 unsigned int ibc_nrx:16;
686 /* scheduled for attention */
687 unsigned int ibc_scheduled:1;
688 /* CQ callback fired */
689 unsigned int ibc_ready:1;
690 /* time of last send */
691 unsigned long ibc_last_send;
692 /** link chain for kiblnd_check_conns only */
693 struct list_head ibc_connd_list;
694 /** rxs completed before ESTABLISHED */
695 struct list_head ibc_early_rxs;
696 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
697 struct list_head ibc_tx_noops;
698 /* sends that need a credit */
699 struct list_head ibc_tx_queue;
700 /* sends that don't need a credit */
701 struct list_head ibc_tx_queue_nocred;
702 /* sends that need to reserve an ACK/DONE msg */
703 struct list_head ibc_tx_queue_rsrvd;
704 /* active tx awaiting completion */
705 struct list_head ibc_active_txs;
710 /* premapped rx msg pages */
711 kib_pages_t *ibc_rx_pages;
714 struct rdma_cm_id *ibc_cmid;
715 /* completion queue */
716 struct ib_cq *ibc_cq;
718 /* in-progress connection state */
719 kib_connvars_t *ibc_connvars;
722 #define IBLND_CONN_INIT 0 /* being initialised */
723 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
724 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
725 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
726 #define IBLND_CONN_CLOSING 4 /* being closed */
727 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
729 typedef struct kib_peer
731 /* stash on global peer list */
732 struct list_head ibp_list;
733 /* who's on the other end(s) */
737 /* all active connections */
738 struct list_head ibp_conns;
739 /* msgs waiting for a conn */
740 struct list_head ibp_tx_queue;
741 /* incarnation of peer */
742 __u64 ibp_incarnation;
743 /* when (in jiffies) I was last alive */
744 cfs_time_t ibp_last_alive;
746 atomic_t ibp_refcount;
747 /* version of peer */
749 /* current passive connection attempts */
750 unsigned short ibp_accepting;
751 /* current active connection attempts */
752 unsigned short ibp_connecting;
753 /* reconnect this peer later */
754 unsigned short ibp_reconnecting:1;
755 /* wish to win the connrace */
756 unsigned short ibp_connrace_win:1;
757 /* # consecutive reconnection attempts to this peer */
758 unsigned int ibp_reconnected;
759 /* errno on closing this peer */
761 /* max map_on_demand */
763 /* max_peer_credits */
764 __u16 ibp_queue_depth;
767 extern kib_data_t kiblnd_data;
769 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
772 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
774 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
775 atomic_inc(&hdev->ibh_ref);
779 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
781 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
782 if (atomic_dec_and_test(&hdev->ibh_ref))
783 kiblnd_hdev_destroy(hdev);
787 kiblnd_dev_can_failover(kib_dev_t *dev)
789 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
792 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
795 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
798 return dev->ibd_can_failover;
801 #define kiblnd_conn_addref(conn) \
803 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
804 (conn), atomic_read(&(conn)->ibc_refcount)); \
805 atomic_inc(&(conn)->ibc_refcount); \
808 #define kiblnd_conn_decref(conn) \
810 unsigned long flags; \
812 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
813 (conn), atomic_read(&(conn)->ibc_refcount)); \
814 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
815 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
816 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
817 list_add_tail(&(conn)->ibc_list, \
818 &kiblnd_data.kib_connd_zombies); \
819 wake_up(&kiblnd_data.kib_connd_waitq); \
820 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
824 #define kiblnd_peer_addref(peer) \
826 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
827 (peer), libcfs_nid2str((peer)->ibp_nid), \
828 atomic_read (&(peer)->ibp_refcount)); \
829 atomic_inc(&(peer)->ibp_refcount); \
832 #define kiblnd_peer_decref(peer) \
834 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
835 (peer), libcfs_nid2str((peer)->ibp_nid), \
836 atomic_read (&(peer)->ibp_refcount)); \
837 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
838 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
839 kiblnd_destroy_peer(peer); \
843 kiblnd_peer_win_race(kib_peer_t *peer, kib_msg_t *msg)
845 if (!peer->ibp_connecting)
846 return true; /* no race */
848 if (msg->ibm_credits & IBLND_CONNREQ_NOOP) {
849 /* peer can understand connrace protocol */
851 if (msg->ibm_credits & IBLND_CONNREQ_WIN_WISH)
852 return true; /* peer has win wish */
854 if (peer->ibp_connrace_win)
855 return false; /* I wish to win, reject peer */
857 /* tie-break connection race in favour of the higher NID */
858 return peer->ibp_nid > peer->ibp_ni->ni_nid;
862 kiblnd_peer_connecting(kib_peer_t *peer)
864 return peer->ibp_connecting != 0 ||
865 peer->ibp_reconnecting != 0 ||
866 peer->ibp_accepting != 0;
870 kiblnd_peer_idle(kib_peer_t *peer)
872 return !kiblnd_peer_connecting(peer) && list_empty(&peer->ibp_conns);
875 static inline struct list_head *
876 kiblnd_nid2peerlist (lnet_nid_t nid)
879 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
881 return &kiblnd_data.kib_peers[hash];
885 kiblnd_peer_active (kib_peer_t *peer)
887 /* Am I in the peer hash table? */
888 return !list_empty(&peer->ibp_list);
891 static inline kib_conn_t *
892 kiblnd_get_conn_locked (kib_peer_t *peer)
894 LASSERT(!list_empty(&peer->ibp_conns));
896 /* just return the first connection */
897 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
901 kiblnd_send_keepalive(kib_conn_t *conn)
903 return (*kiblnd_tunables.kib_keepalive > 0) &&
904 cfs_time_after(jiffies, conn->ibc_last_send +
905 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
910 kiblnd_need_noop(kib_conn_t *conn)
912 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
914 if (conn->ibc_outstanding_credits <
915 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
916 !kiblnd_send_keepalive(conn))
917 return 0; /* No need to send NOOP */
919 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
920 if (!list_empty(&conn->ibc_tx_queue_nocred))
921 return 0; /* NOOP can be piggybacked */
923 /* No tx to piggyback NOOP onto or no credit to send a tx */
924 return (list_empty(&conn->ibc_tx_queue) ||
925 conn->ibc_credits == 0);
928 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
929 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
930 conn->ibc_credits == 0) /* no credit */
933 if (conn->ibc_credits == 1 && /* last credit reserved for */
934 conn->ibc_outstanding_credits == 0) /* giving back credits */
937 /* No tx to piggyback NOOP onto or no credit to send a tx */
938 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
942 kiblnd_abort_receives(kib_conn_t *conn)
944 ib_modify_qp(conn->ibc_cmid->qp,
945 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
948 static inline const char *
949 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
951 if (q == &conn->ibc_tx_queue)
954 if (q == &conn->ibc_tx_queue_rsrvd)
955 return "tx_queue_rsrvd";
957 if (q == &conn->ibc_tx_queue_nocred)
958 return "tx_queue_nocred";
960 if (q == &conn->ibc_active_txs)
967 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
968 * lowest bits of the work request id to stash the work item type. */
970 #define IBLND_WID_INVAL 0
971 #define IBLND_WID_TX 1
972 #define IBLND_WID_RX 2
973 #define IBLND_WID_RDMA 3
974 #define IBLND_WID_MASK 3UL
977 kiblnd_ptr2wreqid (void *ptr, int type)
979 unsigned long lptr = (unsigned long)ptr;
981 LASSERT ((lptr & IBLND_WID_MASK) == 0);
982 LASSERT ((type & ~IBLND_WID_MASK) == 0);
983 return (__u64)(lptr | type);
987 kiblnd_wreqid2ptr (__u64 wreqid)
989 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
993 kiblnd_wreqid2type (__u64 wreqid)
995 return (wreqid & IBLND_WID_MASK);
999 kiblnd_set_conn_state (kib_conn_t *conn, int state)
1001 conn->ibc_state = state;
1006 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
1008 msg->ibm_type = type;
1009 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
1013 kiblnd_rd_size (kib_rdma_desc_t *rd)
1018 for (i = size = 0; i < rd->rd_nfrags; i++)
1019 size += rd->rd_frags[i].rf_nob;
1025 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
1027 return rd->rd_frags[index].rf_addr;
1031 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
1033 return rd->rd_frags[index].rf_nob;
1037 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
1043 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
1045 if (nob < rd->rd_frags[index].rf_nob) {
1046 rd->rd_frags[index].rf_addr += nob;
1047 rd->rd_frags[index].rf_nob -= nob;
1056 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
1058 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1059 msgtype == IBLND_MSG_PUT_ACK);
1061 return msgtype == IBLND_MSG_GET_REQ ?
1062 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
1063 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
1067 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1069 return ib_dma_mapping_error(dev, dma_addr);
1072 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1073 void *msg, size_t size,
1074 enum dma_data_direction direction)
1076 return ib_dma_map_single(dev, msg, size, direction);
1079 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1080 __u64 addr, size_t size,
1081 enum dma_data_direction direction)
1083 ib_dma_unmap_single(dev, addr, size, direction);
1086 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1087 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1089 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1090 struct scatterlist *sg, int nents,
1091 enum dma_data_direction direction)
1093 return ib_dma_map_sg(dev, sg, nents, direction);
1096 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1097 struct scatterlist *sg, int nents,
1098 enum dma_data_direction direction)
1100 ib_dma_unmap_sg(dev, sg, nents, direction);
1103 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1104 struct scatterlist *sg)
1106 return ib_sg_dma_address(dev, sg);
1109 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1110 struct scatterlist *sg)
1112 return ib_sg_dma_len(dev, sg);
1115 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1116 * right because OFED1.2 defines it as const, to use it we have to add
1117 * (void *) cast to overcome "const" */
1119 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1120 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1122 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1123 kib_rdma_desc_t *rd,
1124 int negotiated_nfrags);
1125 void kiblnd_map_rx_descs(kib_conn_t *conn);
1126 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1127 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
1128 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1130 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1131 int npages, __u64 iov, kib_fmr_t *fmr);
1132 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1134 int kiblnd_tunables_init(void);
1135 void kiblnd_tunables_fini(void);
1137 int kiblnd_connd (void *arg);
1138 int kiblnd_scheduler(void *arg);
1139 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1140 int kiblnd_failover_thread (void *arg);
1142 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1144 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1145 struct rdma_cm_event *event);
1146 int kiblnd_translate_mtu(int value);
1148 int kiblnd_dev_failover(kib_dev_t *dev);
1149 int kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1150 void kiblnd_destroy_peer (kib_peer_t *peer);
1151 bool kiblnd_reconnect_peer(kib_peer_t *peer, bool connrace_win);
1152 void kiblnd_destroy_dev (kib_dev_t *dev);
1153 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1154 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1155 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1156 int version, __u64 incarnation);
1157 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1159 kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
1160 int state, int version);
1161 void kiblnd_destroy_conn(kib_conn_t *conn, bool free_conn);
1162 void kiblnd_close_conn (kib_conn_t *conn, int error);
1163 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1165 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1166 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status);
1168 void kiblnd_qp_event(struct ib_event *event, void *arg);
1169 void kiblnd_cq_event(struct ib_event *event, void *arg);
1170 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1172 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1173 int credits, lnet_nid_t dstnid, __u64 dststamp);
1174 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1175 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1177 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1178 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1179 unsigned int niov, struct kvec *iov, lnet_kiov_t *kiov,
1180 unsigned int offset, unsigned int mlen, unsigned int rlen);