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, 2014, 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_pmr_pool_size; /* # physical MR in pool */
115 int *kib_fmr_pool_size; /* # FMRs in pool */
116 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
117 int *kib_fmr_cache; /* enable FMR pool cache? */
118 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
119 struct ctl_table_header *kib_sysctl; /* sysctl interface */
121 int *kib_require_priv_port;/* accept only privileged ports */
122 int *kib_use_priv_port; /* use privileged port for active connect */
123 /* # threads on each CPT */
127 extern kib_tunables_t kiblnd_tunables;
129 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
130 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
132 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
133 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
135 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
136 IBLND_MSG_QUEUE_SIZE_V1 : \
137 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
138 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
139 IBLND_CREDIT_HIGHWATER_V1 : \
140 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
142 #ifdef HAVE_RDMA_CREATE_ID_4ARG
143 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
145 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
149 kiblnd_concurrent_sends_v1(void)
151 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
152 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
154 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
155 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
157 return *kiblnd_tunables.kib_concurrent_sends;
160 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
161 kiblnd_concurrent_sends_v1() : \
162 *kiblnd_tunables.kib_concurrent_sends)
163 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
164 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
165 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
167 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
168 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
169 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
170 *kiblnd_tunables.kib_map_on_demand : \
171 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
172 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
173 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
175 /************************/
176 /* derived constants... */
177 /* Pools (shared by connections on each CPT) */
178 /* These pools can grow at runtime, so don't need give a very large value */
179 #define IBLND_TX_POOL 256
180 #define IBLND_PMR_POOL 256
181 #define IBLND_FMR_POOL 256
182 #define IBLND_FMR_POOL_FLUSH 192
184 /* TX messages (shared by all connections) */
185 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
187 /* RX messages (per connection) */
188 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
189 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
190 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
192 /* WRs and CQEs (per connection) */
193 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
194 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
195 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
199 /* o2iblnd can run over aliased interface */
201 #define KIB_IFNAME_SIZE IFALIASZ
203 #define KIB_IFNAME_SIZE 256
208 struct list_head ibd_list; /* chain on kib_devs */
209 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
210 __u32 ibd_ifip; /* IPoIB interface IP */
211 /** IPoIB interface name */
212 char ibd_ifname[KIB_IFNAME_SIZE];
213 int ibd_nnets; /* # nets extant */
215 cfs_time_t ibd_next_failover;
216 /* # failover failures */
217 int ibd_failed_failover;
218 /* failover in progress */
219 unsigned int ibd_failover;
220 /* IPoIB interface is a bonding master */
221 unsigned int ibd_can_failover;
222 struct list_head ibd_nets;
223 struct kib_hca_dev *ibd_hdev;
226 typedef struct kib_hca_dev
228 struct rdma_cm_id *ibh_cmid; /* listener cmid */
229 struct ib_device *ibh_ibdev; /* IB device */
230 int ibh_page_shift; /* page shift of current HCA */
231 int ibh_page_size; /* page size of current HCA */
232 __u64 ibh_page_mask; /* page mask of current HCA */
233 int ibh_mr_shift; /* bits shift of max MR size */
234 __u64 ibh_mr_size; /* size of MR */
235 int ibh_nmrs; /* # of global MRs */
236 struct ib_mr **ibh_mrs; /* global MR */
237 struct ib_pd *ibh_pd; /* PD */
238 kib_dev_t *ibh_dev; /* owner */
239 atomic_t ibh_ref; /* refcount */
242 /** # of seconds to keep pool alive */
243 #define IBLND_POOL_DEADLINE 300
244 /** # of seconds to retry if allocation failed */
245 #define IBLND_POOL_RETRY 1
249 int ibp_npages; /* # pages */
250 struct page *ibp_pages[0]; /* page array */
256 struct list_head pmr_list; /* chain node */
257 struct ib_phys_buf *pmr_ipb; /* physical buffer */
258 struct ib_mr *pmr_mr; /* IB MR */
259 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
260 __u64 pmr_iova; /* Virtual I/O address */
261 int pmr_refcount; /* reference count */
267 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
268 int inc, struct kib_pool **pp_po);
269 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
270 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
271 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
275 #define IBLND_POOL_NAME_LEN 32
277 typedef struct kib_poolset
281 /* network it belongs to */
282 struct kib_net *ps_net;
284 char ps_name[IBLND_POOL_NAME_LEN];
286 struct list_head ps_pool_list;
287 /* failed pool list */
288 struct list_head ps_failed_pool_list;
289 /* time stamp for retry if failed to allocate */
290 cfs_time_t ps_next_retry;
291 /* is allocating new pool */
298 /* create a new pool */
299 kib_ps_pool_create_t ps_pool_create;
301 kib_ps_pool_destroy_t ps_pool_destroy;
302 /* initialize new allocated node */
303 kib_ps_node_init_t ps_node_init;
305 kib_ps_node_fini_t ps_node_fini;
308 typedef struct kib_pool
310 /* chain on pool list */
311 struct list_head po_list;
312 /* pre-allocated node */
313 struct list_head po_free_list;
314 /* pool_set of this pool */
315 kib_poolset_t *po_owner;
316 /* deadline of this pool */
317 cfs_time_t po_deadline;
318 /* # of elements in use */
320 /* pool is created on failed HCA */
322 /* # of pre-allocated elements */
327 kib_poolset_t tps_poolset; /* pool-set */
328 __u64 tps_next_tx_cookie; /* cookie of TX */
332 kib_pool_t tpo_pool; /* pool */
333 struct kib_hca_dev *tpo_hdev; /* device for this pool */
334 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
335 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
339 kib_poolset_t pps_poolset; /* pool-set */
342 typedef struct kib_pmr_pool {
343 struct kib_hca_dev *ppo_hdev; /* device for this pool */
344 kib_pool_t ppo_pool; /* pool */
349 spinlock_t fps_lock; /* serialize */
350 struct kib_net *fps_net; /* IB network */
351 struct list_head fps_pool_list; /* FMR pool list */
352 struct list_head fps_failed_pool_list; /* FMR pool list */
353 __u64 fps_version; /* validity stamp */
354 int fps_cpt; /* CPT id */
356 int fps_flush_trigger;
357 /* is allocating new pool */
359 /* time stamp for retry if failed to allocate */
360 cfs_time_t fps_next_retry;
365 struct list_head fpo_list; /* chain on pool list */
366 struct kib_hca_dev *fpo_hdev; /* device for this pool */
367 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
368 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
369 cfs_time_t fpo_deadline; /* deadline of this pool */
370 int fpo_failed; /* fmr pool is failed */
371 int fpo_map_count; /* # of mapped FMR */
375 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
376 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
379 typedef struct kib_net
381 /* chain on kib_dev_t::ibd_nets */
382 struct list_head ibn_list;
383 __u64 ibn_incarnation;/* my epoch */
384 int ibn_init; /* initialisation state */
385 int ibn_shutdown; /* shutting down? */
387 atomic_t ibn_npeers; /* # peers extant */
388 atomic_t ibn_nconns; /* # connections extant */
390 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
391 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
392 kib_pmr_poolset_t **ibn_pmr_ps; /* pmr pool-set */
394 kib_dev_t *ibn_dev; /* underlying IB device */
397 #define KIB_THREAD_SHIFT 16
398 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
399 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
400 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
402 struct kib_sched_info {
405 /* schedulers sleep here */
406 wait_queue_head_t ibs_waitq;
407 /* conns to check for rx completions */
408 struct list_head ibs_conns;
409 /* number of scheduler threads */
411 /* max allowed scheduler threads */
412 int ibs_nthreads_max;
413 int ibs_cpt; /* CPT id */
418 int kib_init; /* initialisation state */
419 int kib_shutdown; /* shut down? */
420 struct list_head kib_devs; /* IB devices extant */
421 /* list head of failed devices */
422 struct list_head kib_failed_devs;
423 /* schedulers sleep here */
424 wait_queue_head_t kib_failover_waitq;
425 atomic_t kib_nthreads; /* # live threads */
426 /* stabilize net/dev/peer/conn ops */
427 rwlock_t kib_global_lock;
428 /* hash table of all my known peers */
429 struct list_head *kib_peers;
430 /* size of kib_peers */
431 int kib_peer_hash_size;
432 /* the connd task (serialisation assertions) */
434 /* connections to setup/teardown */
435 struct list_head kib_connd_conns;
436 /* connections with zero refcount */
437 struct list_head kib_connd_zombies;
438 /* connection daemon sleeps here */
439 wait_queue_head_t kib_connd_waitq;
440 spinlock_t kib_connd_lock; /* serialise */
441 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
442 /* percpt data for schedulers */
443 struct kib_sched_info **kib_scheds;
446 #define IBLND_INIT_NOTHING 0
447 #define IBLND_INIT_DATA 1
448 #define IBLND_INIT_ALL 2
450 /************************************************************************
451 * IB Wire message format.
452 * These are sent in sender's byte order (i.e. receiver flips).
455 typedef struct kib_connparams
457 __u16 ibcp_queue_depth;
458 __u16 ibcp_max_frags;
459 __u32 ibcp_max_msg_size;
460 } WIRE_ATTR kib_connparams_t;
464 lnet_hdr_t ibim_hdr; /* portals header */
465 char ibim_payload[0]; /* piggy-backed payload */
466 } WIRE_ATTR kib_immediate_msg_t;
470 __u32 rf_nob; /* # bytes this frag */
471 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
472 } WIRE_ATTR kib_rdma_frag_t;
476 __u32 rd_key; /* local/remote key */
477 __u32 rd_nfrags; /* # fragments */
478 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
479 } WIRE_ATTR kib_rdma_desc_t;
483 lnet_hdr_t ibprm_hdr; /* portals header */
484 __u64 ibprm_cookie; /* opaque completion cookie */
485 } WIRE_ATTR kib_putreq_msg_t;
489 __u64 ibpam_src_cookie; /* reflected completion cookie */
490 __u64 ibpam_dst_cookie; /* opaque completion cookie */
491 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
492 } WIRE_ATTR kib_putack_msg_t;
496 lnet_hdr_t ibgm_hdr; /* portals header */
497 __u64 ibgm_cookie; /* opaque completion cookie */
498 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
499 } WIRE_ATTR kib_get_msg_t;
503 __u64 ibcm_cookie; /* opaque completion cookie */
504 __s32 ibcm_status; /* < 0 failure: >= 0 length */
505 } WIRE_ATTR kib_completion_msg_t;
509 /* First 2 fields fixed FOR ALL TIME */
510 __u32 ibm_magic; /* I'm an ibnal message */
511 __u16 ibm_version; /* this is my version number */
513 __u8 ibm_type; /* msg type */
514 __u8 ibm_credits; /* returned credits */
515 __u32 ibm_nob; /* # bytes in whole message */
516 __u32 ibm_cksum; /* checksum (0 == no checksum) */
517 __u64 ibm_srcnid; /* sender's NID */
518 __u64 ibm_srcstamp; /* sender's incarnation */
519 __u64 ibm_dstnid; /* destination's NID */
520 __u64 ibm_dststamp; /* destination's incarnation */
523 kib_connparams_t connparams;
524 kib_immediate_msg_t immediate;
525 kib_putreq_msg_t putreq;
526 kib_putack_msg_t putack;
528 kib_completion_msg_t completion;
530 } WIRE_ATTR kib_msg_t;
532 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
534 #define IBLND_MSG_VERSION_1 0x11
535 #define IBLND_MSG_VERSION_2 0x12
536 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
538 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
539 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
540 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
541 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
542 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
543 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
544 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
545 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
546 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
547 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
550 __u32 ibr_magic; /* sender's magic */
551 __u16 ibr_version; /* sender's version */
552 __u8 ibr_why; /* reject reason */
553 __u8 ibr_padding; /* padding */
554 __u64 ibr_incarnation; /* incarnation of peer */
555 kib_connparams_t ibr_cp; /* connection parameters */
556 } WIRE_ATTR kib_rej_t;
558 /* connection rejection reasons */
559 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
560 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
561 #define IBLND_REJECT_FATAL 3 /* Anything else */
563 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
564 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
566 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
567 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
569 /***********************************************************************/
571 typedef struct kib_rx /* receive message */
573 /* queue for attention */
574 struct list_head rx_list;
576 struct kib_conn *rx_conn;
577 /* # bytes received (-1 while posted) */
579 /* completion status */
580 enum ib_wc_status rx_status;
581 /* message buffer (host vaddr) */
583 /* message buffer (I/O addr) */
585 /* for dma_unmap_single() */
586 DECLARE_PCI_UNMAP_ADDR(rx_msgunmap);
587 /* receive work item... */
588 struct ib_recv_wr rx_wrq;
589 /* ...and its memory */
590 struct ib_sge rx_sge;
593 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
594 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
595 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
596 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
598 typedef struct kib_tx /* transmit message */
600 /* queue on idle_txs ibc_tx_queue etc. */
601 struct list_head tx_list;
603 kib_tx_pool_t *tx_pool;
605 struct kib_conn *tx_conn;
606 /* # tx callbacks outstanding */
608 /* queued for sending */
610 /* waiting for peer */
612 /* LNET completion status */
614 /* completion deadline */
615 unsigned long tx_deadline;
616 /* completion cookie */
618 /* lnet msgs to finalize on completion */
619 lnet_msg_t *tx_lntmsg[2];
620 /* message buffer (host vaddr) */
622 /* message buffer (I/O addr) */
624 /* for dma_unmap_single() */
625 DECLARE_PCI_UNMAP_ADDR(tx_msgunmap);
626 /* # send work items */
628 /* send work items... */
629 struct ib_send_wr *tx_wrq;
630 /* ...and their memory */
631 struct ib_sge *tx_sge;
632 /* rdma descriptor */
633 kib_rdma_desc_t *tx_rd;
634 /* # entries in... */
636 /* dma_map_sg descriptor */
637 struct scatterlist *tx_frags;
638 /* rdma phys page addrs */
641 /* MR for physical buffer */
650 typedef struct kib_connvars
652 /* connection-in-progress variables */
656 typedef struct kib_conn
658 /* scheduler information */
659 struct kib_sched_info *ibc_sched;
661 struct kib_peer *ibc_peer;
663 kib_hca_dev_t *ibc_hdev;
664 /* stash on peer's conn list */
665 struct list_head ibc_list;
666 /* schedule for attention */
667 struct list_head ibc_sched_list;
668 /* version of connection */
670 /* which instance of the peer */
671 __u64 ibc_incarnation;
673 atomic_t ibc_refcount;
674 /* what's happening */
676 /* # uncompleted sends */
677 int ibc_nsends_posted;
678 /* # uncompleted NOOPs */
679 int ibc_noops_posted;
680 /* # credits I have */
682 /* # credits to return */
683 int ibc_outstanding_credits;
684 /* # ACK/DONE msg credits */
685 int ibc_reserved_credits;
686 /* set on comms error */
688 /* receive buffers owned */
689 unsigned short ibc_nrx;
690 /** rejected by connection race */
691 unsigned short ibc_conn_race:1;
692 /* scheduled for attention */
693 unsigned short ibc_scheduled:1;
694 /* CQ callback fired */
695 unsigned short ibc_ready:1;
696 /* time of last send */
697 unsigned long ibc_last_send;
698 /** link chain for kiblnd_check_conns only */
699 struct list_head ibc_connd_list;
700 /** rxs completed before ESTABLISHED */
701 struct list_head ibc_early_rxs;
702 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
703 struct list_head ibc_tx_noops;
704 /* sends that need a credit */
705 struct list_head ibc_tx_queue;
706 /* sends that don't need a credit */
707 struct list_head ibc_tx_queue_nocred;
708 /* sends that need to reserve an ACK/DONE msg */
709 struct list_head ibc_tx_queue_rsrvd;
710 /* active tx awaiting completion */
711 struct list_head ibc_active_txs;
716 /* premapped rx msg pages */
717 kib_pages_t *ibc_rx_pages;
720 struct rdma_cm_id *ibc_cmid;
721 /* completion queue */
722 struct ib_cq *ibc_cq;
724 /* in-progress connection state */
725 kib_connvars_t *ibc_connvars;
728 #define IBLND_CONN_INIT 0 /* being initialised */
729 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
730 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
731 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
732 #define IBLND_CONN_CLOSING 4 /* being closed */
733 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
735 typedef struct kib_peer
737 /* stash on global peer list */
738 struct list_head ibp_list;
739 /* who's on the other end(s) */
744 atomic_t ibp_refcount;
745 /* all active connections */
746 struct list_head ibp_conns;
747 /* msgs waiting for a conn */
748 struct list_head ibp_tx_queue;
749 /* version of peer */
751 /* incarnation of peer */
752 __u64 ibp_incarnation;
753 /* current active connection attempts */
755 /* current passive connection attempts */
757 /* errno on closing this peer */
759 /* when (in jiffies) I was last alive */
760 cfs_time_t ibp_last_alive;
763 extern kib_data_t kiblnd_data;
765 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
768 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
770 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
771 atomic_inc(&hdev->ibh_ref);
775 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
777 LASSERT(atomic_read(&hdev->ibh_ref) > 0);
778 if (atomic_dec_and_test(&hdev->ibh_ref))
779 kiblnd_hdev_destroy(hdev);
783 kiblnd_dev_can_failover(kib_dev_t *dev)
785 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
788 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
791 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
794 return dev->ibd_can_failover;
797 #define kiblnd_conn_addref(conn) \
799 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
800 (conn), atomic_read(&(conn)->ibc_refcount)); \
801 atomic_inc(&(conn)->ibc_refcount); \
804 #define kiblnd_conn_decref(conn) \
806 unsigned long flags; \
808 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
809 (conn), atomic_read(&(conn)->ibc_refcount)); \
810 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
811 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
812 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
813 list_add_tail(&(conn)->ibc_list, \
814 &kiblnd_data.kib_connd_zombies); \
815 wake_up(&kiblnd_data.kib_connd_waitq); \
816 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
820 #define kiblnd_peer_addref(peer) \
822 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
823 (peer), libcfs_nid2str((peer)->ibp_nid), \
824 atomic_read (&(peer)->ibp_refcount)); \
825 atomic_inc(&(peer)->ibp_refcount); \
828 #define kiblnd_peer_decref(peer) \
830 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
831 (peer), libcfs_nid2str((peer)->ibp_nid), \
832 atomic_read (&(peer)->ibp_refcount)); \
833 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
834 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
835 kiblnd_destroy_peer(peer); \
838 static inline struct list_head *
839 kiblnd_nid2peerlist (lnet_nid_t nid)
842 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
844 return &kiblnd_data.kib_peers[hash];
848 kiblnd_peer_active (kib_peer_t *peer)
850 /* Am I in the peer hash table? */
851 return !list_empty(&peer->ibp_list);
854 static inline kib_conn_t *
855 kiblnd_get_conn_locked (kib_peer_t *peer)
857 LASSERT(!list_empty(&peer->ibp_conns));
859 /* just return the first connection */
860 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
864 kiblnd_send_keepalive(kib_conn_t *conn)
866 return (*kiblnd_tunables.kib_keepalive > 0) &&
867 cfs_time_after(jiffies, conn->ibc_last_send +
868 msecs_to_jiffies(*kiblnd_tunables.kib_keepalive *
873 kiblnd_need_noop(kib_conn_t *conn)
875 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
877 if (conn->ibc_outstanding_credits <
878 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
879 !kiblnd_send_keepalive(conn))
880 return 0; /* No need to send NOOP */
882 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
883 if (!list_empty(&conn->ibc_tx_queue_nocred))
884 return 0; /* NOOP can be piggybacked */
886 /* No tx to piggyback NOOP onto or no credit to send a tx */
887 return (list_empty(&conn->ibc_tx_queue) ||
888 conn->ibc_credits == 0);
891 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
892 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
893 conn->ibc_credits == 0) /* no credit */
896 if (conn->ibc_credits == 1 && /* last credit reserved for */
897 conn->ibc_outstanding_credits == 0) /* giving back credits */
900 /* No tx to piggyback NOOP onto or no credit to send a tx */
901 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
905 kiblnd_abort_receives(kib_conn_t *conn)
907 ib_modify_qp(conn->ibc_cmid->qp,
908 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
911 static inline const char *
912 kiblnd_queue2str(kib_conn_t *conn, struct list_head *q)
914 if (q == &conn->ibc_tx_queue)
917 if (q == &conn->ibc_tx_queue_rsrvd)
918 return "tx_queue_rsrvd";
920 if (q == &conn->ibc_tx_queue_nocred)
921 return "tx_queue_nocred";
923 if (q == &conn->ibc_active_txs)
930 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
931 * lowest bits of the work request id to stash the work item type. */
933 #define IBLND_WID_TX 0
934 #define IBLND_WID_RDMA 1
935 #define IBLND_WID_RX 2
936 #define IBLND_WID_MASK 3UL
939 kiblnd_ptr2wreqid (void *ptr, int type)
941 unsigned long lptr = (unsigned long)ptr;
943 LASSERT ((lptr & IBLND_WID_MASK) == 0);
944 LASSERT ((type & ~IBLND_WID_MASK) == 0);
945 return (__u64)(lptr | type);
949 kiblnd_wreqid2ptr (__u64 wreqid)
951 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
955 kiblnd_wreqid2type (__u64 wreqid)
957 return (wreqid & IBLND_WID_MASK);
961 kiblnd_set_conn_state (kib_conn_t *conn, int state)
963 conn->ibc_state = state;
968 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
970 msg->ibm_type = type;
971 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
975 kiblnd_rd_size (kib_rdma_desc_t *rd)
980 for (i = size = 0; i < rd->rd_nfrags; i++)
981 size += rd->rd_frags[i].rf_nob;
987 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
989 return rd->rd_frags[index].rf_addr;
993 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
995 return rd->rd_frags[index].rf_nob;
999 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
1005 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
1007 if (nob < rd->rd_frags[index].rf_nob) {
1008 rd->rd_frags[index].rf_addr += nob;
1009 rd->rd_frags[index].rf_nob -= nob;
1018 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
1020 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
1021 msgtype == IBLND_MSG_PUT_ACK);
1023 return msgtype == IBLND_MSG_GET_REQ ?
1024 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
1025 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
1029 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
1031 return ib_dma_mapping_error(dev, dma_addr);
1034 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
1035 void *msg, size_t size,
1036 enum dma_data_direction direction)
1038 return ib_dma_map_single(dev, msg, size, direction);
1041 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1042 __u64 addr, size_t size,
1043 enum dma_data_direction direction)
1045 ib_dma_unmap_single(dev, addr, size, direction);
1048 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
1049 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
1051 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1052 struct scatterlist *sg, int nents,
1053 enum dma_data_direction direction)
1055 return ib_dma_map_sg(dev, sg, nents, direction);
1058 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1059 struct scatterlist *sg, int nents,
1060 enum dma_data_direction direction)
1062 ib_dma_unmap_sg(dev, sg, nents, direction);
1065 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
1066 struct scatterlist *sg)
1068 return ib_sg_dma_address(dev, sg);
1071 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1072 struct scatterlist *sg)
1074 return ib_sg_dma_len(dev, sg);
1077 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
1078 * right because OFED1.2 defines it as const, to use it we have to add
1079 * (void *) cast to overcome "const" */
1081 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
1082 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
1084 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1085 kib_rdma_desc_t *rd);
1086 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
1087 __u64 addr, __u64 size);
1088 void kiblnd_map_rx_descs(kib_conn_t *conn);
1089 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1090 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
1091 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1093 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1094 int npages, __u64 iov, kib_fmr_t *fmr);
1095 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1097 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1098 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1099 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1101 int kiblnd_tunables_init(void);
1102 void kiblnd_tunables_fini(void);
1104 int kiblnd_connd (void *arg);
1105 int kiblnd_scheduler(void *arg);
1106 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1107 int kiblnd_failover_thread (void *arg);
1109 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1111 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1112 struct rdma_cm_event *event);
1113 int kiblnd_translate_mtu(int value);
1115 int kiblnd_dev_failover(kib_dev_t *dev);
1116 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1117 void kiblnd_destroy_peer (kib_peer_t *peer);
1118 void kiblnd_connect_peer(kib_peer_t *peer);
1119 void kiblnd_destroy_dev (kib_dev_t *dev);
1120 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1121 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1122 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1123 int version, __u64 incarnation);
1124 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1126 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1127 int state, int version);
1128 void kiblnd_destroy_conn (kib_conn_t *conn);
1129 void kiblnd_close_conn (kib_conn_t *conn, int error);
1130 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1132 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1133 void kiblnd_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int status);
1135 void kiblnd_qp_event(struct ib_event *event, void *arg);
1136 void kiblnd_cq_event(struct ib_event *event, void *arg);
1137 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1139 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1140 int credits, lnet_nid_t dstnid, __u64 dststamp);
1141 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1142 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1144 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1145 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1146 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1147 unsigned int offset, unsigned int mlen, unsigned int rlen);