1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
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>
42 # define EXPORT_SYMTAB
44 #ifndef AUTOCONF_INCLUDED
45 #include <linux/config.h>
47 #include <linux/module.h>
48 #include <linux/kernel.h>
50 #include <linux/string.h>
51 #include <linux/stat.h>
52 #include <linux/errno.h>
53 #include <linux/smp_lock.h>
54 #include <linux/unistd.h>
55 #include <linux/uio.h>
57 #include <asm/system.h>
58 #include <asm/uaccess.h>
61 #include <linux/init.h>
63 #include <linux/file.h>
64 #include <linux/stat.h>
65 #include <linux/list.h>
66 #include <linux/kmod.h>
67 #include <linux/sysctl.h>
68 #include <linux/pci.h>
73 #define DEBUG_SUBSYSTEM S_LND
75 #include <libcfs/libcfs.h>
76 #include <lnet/lnet.h>
77 #include <lnet/lib-lnet.h>
78 #include <lnet/lnet-sysctl.h>
84 #include <rdma/rdma_cm.h>
85 #include <rdma/ib_cm.h>
86 #include <rdma/ib_verbs.h>
87 #include <rdma/ib_fmr_pool.h>
89 /* tunables fixed at compile time */
91 # define IBLND_N_SCHED cfs_num_online_cpus() /* # schedulers */
93 # define IBLND_N_SCHED 1 /* # schedulers */
96 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
97 #define IBLND_RESCHED 100 /* # scheduler loops before reschedule */
101 int *kib_dev_failover; /* HCA failover */
102 unsigned int *kib_service; /* IB service number */
103 int *kib_min_reconnect_interval; /* first failed connection retry... */
104 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
105 int *kib_cksum; /* checksum kib_msg_t? */
106 int *kib_timeout; /* comms timeout (seconds) */
107 int *kib_keepalive; /* keepalive timeout (seconds) */
108 int *kib_ntx; /* # tx descs */
109 int *kib_credits; /* # concurrent sends */
110 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
111 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
112 int *kib_peercredits_hiw; /* # when eagerly to return credits */
113 int *kib_peertimeout; /* seconds to consider peer dead */
114 char **kib_default_ipif; /* default IPoIB interface */
115 int *kib_retry_count;
116 int *kib_rnr_retry_count;
117 int *kib_concurrent_sends; /* send work queue sizing */
118 int *kib_ib_mtu; /* IB MTU */
119 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
120 * than this value, 0 disable map-on-demand */
121 int *kib_pmr_pool_size; /* # physical MR in pool */
122 int *kib_fmr_pool_size; /* # FMRs in pool */
123 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
124 int *kib_fmr_cache; /* enable FMR pool cache? */
125 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
126 cfs_sysctl_table_header_t *kib_sysctl; /* sysctl interface */
128 int *kib_require_priv_port;/* accept only privileged ports */
129 int *kib_use_priv_port; /* use privileged port for active connect */
132 extern kib_tunables_t kiblnd_tunables;
134 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
135 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
137 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
138 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
140 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
141 IBLND_MSG_QUEUE_SIZE_V1 : \
142 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
143 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
144 IBLND_CREDIT_HIGHWATER_V1 : \
145 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
147 #ifdef HAVE_RDMA_CREATE_ID_4ARG
148 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
150 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
154 kiblnd_concurrent_sends_v1(void)
156 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
157 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
159 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
160 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
162 return *kiblnd_tunables.kib_concurrent_sends;
165 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
166 kiblnd_concurrent_sends_v1() : \
167 *kiblnd_tunables.kib_concurrent_sends)
168 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
169 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
170 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
172 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
173 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
174 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
175 *kiblnd_tunables.kib_map_on_demand : \
176 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
177 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
178 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
180 /************************/
181 /* derived constants... */
183 /* TX messages (shared by all connections) */
184 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
186 /* RX messages (per connection) */
187 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
188 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
189 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
191 /* WRs and CQEs (per connection) */
192 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
193 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
194 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
200 cfs_list_t ibd_list; /* chain on kib_devs */
201 cfs_list_t ibd_fail_list; /* chain on kib_failed_devs */
202 __u32 ibd_ifip; /* IPoIB interface IP */
203 char ibd_ifname[32]; /* IPoIB interface name */
204 int ibd_nnets; /* # nets extant */
206 cfs_time_t ibd_next_failover;
207 int ibd_failed_failover; /* # failover failures */
208 unsigned int ibd_failover; /* failover in progress */
209 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
211 struct kib_hca_dev *ibd_hdev;
214 typedef struct kib_hca_dev
216 struct rdma_cm_id *ibh_cmid; /* listener cmid */
217 struct ib_device *ibh_ibdev; /* IB device */
218 int ibh_page_shift; /* page shift of current HCA */
219 int ibh_page_size; /* page size of current HCA */
220 __u64 ibh_page_mask; /* page mask of current HCA */
221 int ibh_mr_shift; /* bits shift of max MR size */
222 __u64 ibh_mr_size; /* size of MR */
223 int ibh_nmrs; /* # of global MRs */
224 struct ib_mr **ibh_mrs; /* global MR */
225 struct ib_pd *ibh_pd; /* PD */
226 kib_dev_t *ibh_dev; /* owner */
227 cfs_atomic_t ibh_ref; /* refcount */
230 /** # of seconds to keep pool alive */
231 #define IBLND_POOL_DEADLINE 300
232 /** # of seconds to retry if allocation failed */
233 #define IBLND_POOL_RETRY 1
237 int ibp_npages; /* # pages */
238 struct page *ibp_pages[0]; /* page array */
244 cfs_list_t pmr_list; /* chain node */
245 struct ib_phys_buf *pmr_ipb; /* physical buffer */
246 struct ib_mr *pmr_mr; /* IB MR */
247 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
248 __u64 pmr_iova; /* Virtual I/O address */
249 int pmr_refcount; /* reference count */
255 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps, int inc, struct kib_pool **pp_po);
256 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
257 typedef void (*kib_ps_node_init_t)(struct kib_pool *po,
259 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po,
264 #define IBLND_POOL_NAME_LEN 32
266 typedef struct kib_poolset
268 cfs_spinlock_t ps_lock; /* serialize */
269 struct kib_net *ps_net; /* network it belongs to */
270 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
271 cfs_list_t ps_pool_list; /* list of pools */
272 cfs_list_t ps_failed_pool_list; /* failed pool list */
273 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
274 int ps_increasing; /* is allocating new pool */
275 int ps_pool_size; /* new pool size */
277 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
278 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
279 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
280 kib_ps_node_fini_t ps_node_fini; /* finalize node */
283 typedef struct kib_pool
285 cfs_list_t po_list; /* chain on pool list */
286 cfs_list_t po_free_list; /* pre-allocated node */
287 kib_poolset_t *po_owner; /* pool_set of this pool */
288 cfs_time_t po_deadline; /* deadline of this pool */
289 int po_allocated; /* # of elements in use */
290 int po_failed; /* pool is created on failed HCA */
291 int po_size; /* # of pre-allocated elements */
295 kib_poolset_t tps_poolset; /* pool-set */
296 __u64 tps_next_tx_cookie; /* cookie of TX */
300 kib_pool_t tpo_pool; /* pool */
301 struct kib_hca_dev *tpo_hdev; /* device for this pool */
302 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
303 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
307 kib_poolset_t pps_poolset; /* pool-set */
310 typedef struct kib_pmr_pool {
311 struct kib_hca_dev *ppo_hdev; /* device for this pool */
312 kib_pool_t ppo_pool; /* pool */
317 cfs_spinlock_t fps_lock; /* serialize */
318 struct kib_net *fps_net; /* IB network */
319 cfs_list_t fps_pool_list; /* FMR pool list */
320 cfs_list_t fps_failed_pool_list; /* FMR pool list */
321 __u64 fps_version; /* validity stamp */
322 int fps_increasing; /* is allocating new pool */
323 cfs_time_t fps_next_retry; /* time stamp for retry if failed to allocate */
328 cfs_list_t fpo_list; /* chain on pool list */
329 struct kib_hca_dev *fpo_hdev; /* device for this pool */
330 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
331 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
332 cfs_time_t fpo_deadline; /* deadline of this pool */
333 int fpo_failed; /* fmr pool is failed */
334 int fpo_map_count; /* # of mapped FMR */
338 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
339 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
342 typedef struct kib_net
344 cfs_list_t ibn_list; /* chain on kib_dev_t::ibd_nets */
345 __u64 ibn_incarnation; /* my epoch */
346 int ibn_init; /* initialisation state */
347 int ibn_shutdown; /* shutting down? */
348 unsigned int ibn_with_fmr:1; /* FMR? */
349 unsigned int ibn_with_pmr:1; /* PMR? */
351 cfs_atomic_t ibn_npeers; /* # peers extant */
352 cfs_atomic_t ibn_nconns; /* # connections extant */
354 kib_tx_poolset_t ibn_tx_ps; /* tx pool-set */
355 kib_fmr_poolset_t ibn_fmr_ps; /* fmr pool-set */
356 kib_pmr_poolset_t ibn_pmr_ps; /* pmr pool-set */
358 kib_dev_t *ibn_dev; /* underlying IB device */
363 int kib_init; /* initialisation state */
364 int kib_shutdown; /* shut down? */
365 cfs_list_t kib_devs; /* IB devices extant */
366 cfs_list_t kib_failed_devs; /* list head of failed devices */
367 cfs_atomic_t kib_nthreads; /* # live threads */
368 cfs_rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
370 cfs_list_t *kib_peers; /* hash table of all my known peers */
371 int kib_peer_hash_size;/* size of kib_peers */
373 void *kib_connd; /* the connd task (serialisation assertions) */
374 cfs_list_t kib_connd_conns; /* connections to setup/teardown */
375 cfs_list_t kib_connd_zombies;/* connections with zero refcount */
376 cfs_waitq_t kib_connd_waitq; /* connection daemon sleeps here */
377 cfs_spinlock_t kib_connd_lock; /* serialise */
379 cfs_waitq_t kib_sched_waitq; /* schedulers sleep here */
380 cfs_list_t kib_sched_conns; /* conns to check for rx completions */
381 cfs_spinlock_t kib_sched_lock; /* serialise */
382 cfs_waitq_t kib_failover_waitq; /* schedulers sleep here */
384 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
387 #define IBLND_INIT_NOTHING 0
388 #define IBLND_INIT_DATA 1
389 #define IBLND_INIT_ALL 2
391 /************************************************************************
392 * IB Wire message format.
393 * These are sent in sender's byte order (i.e. receiver flips).
396 typedef struct kib_connparams
398 __u16 ibcp_queue_depth;
399 __u16 ibcp_max_frags;
400 __u32 ibcp_max_msg_size;
401 } WIRE_ATTR kib_connparams_t;
405 lnet_hdr_t ibim_hdr; /* portals header */
406 char ibim_payload[0]; /* piggy-backed payload */
407 } WIRE_ATTR kib_immediate_msg_t;
411 __u32 rf_nob; /* # bytes this frag */
412 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
413 } WIRE_ATTR kib_rdma_frag_t;
417 __u32 rd_key; /* local/remote key */
418 __u32 rd_nfrags; /* # fragments */
419 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
420 } WIRE_ATTR kib_rdma_desc_t;
424 lnet_hdr_t ibprm_hdr; /* portals header */
425 __u64 ibprm_cookie; /* opaque completion cookie */
426 } WIRE_ATTR kib_putreq_msg_t;
430 __u64 ibpam_src_cookie; /* reflected completion cookie */
431 __u64 ibpam_dst_cookie; /* opaque completion cookie */
432 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
433 } WIRE_ATTR kib_putack_msg_t;
437 lnet_hdr_t ibgm_hdr; /* portals header */
438 __u64 ibgm_cookie; /* opaque completion cookie */
439 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
440 } WIRE_ATTR kib_get_msg_t;
444 __u64 ibcm_cookie; /* opaque completion cookie */
445 __s32 ibcm_status; /* < 0 failure: >= 0 length */
446 } WIRE_ATTR kib_completion_msg_t;
450 /* First 2 fields fixed FOR ALL TIME */
451 __u32 ibm_magic; /* I'm an ibnal message */
452 __u16 ibm_version; /* this is my version number */
454 __u8 ibm_type; /* msg type */
455 __u8 ibm_credits; /* returned credits */
456 __u32 ibm_nob; /* # bytes in whole message */
457 __u32 ibm_cksum; /* checksum (0 == no checksum) */
458 __u64 ibm_srcnid; /* sender's NID */
459 __u64 ibm_srcstamp; /* sender's incarnation */
460 __u64 ibm_dstnid; /* destination's NID */
461 __u64 ibm_dststamp; /* destination's incarnation */
464 kib_connparams_t connparams;
465 kib_immediate_msg_t immediate;
466 kib_putreq_msg_t putreq;
467 kib_putack_msg_t putack;
469 kib_completion_msg_t completion;
471 } WIRE_ATTR kib_msg_t;
473 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
475 #define IBLND_MSG_VERSION_1 0x11
476 #define IBLND_MSG_VERSION_2 0x12
477 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
479 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
480 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
481 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
482 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
483 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
484 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
485 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
486 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
487 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
488 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
491 __u32 ibr_magic; /* sender's magic */
492 __u16 ibr_version; /* sender's version */
493 __u8 ibr_why; /* reject reason */
494 __u8 ibr_padding; /* padding */
495 __u64 ibr_incarnation; /* incarnation of peer */
496 kib_connparams_t ibr_cp; /* connection parameters */
497 } WIRE_ATTR kib_rej_t;
499 /* connection rejection reasons */
500 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
501 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
502 #define IBLND_REJECT_FATAL 3 /* Anything else */
504 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
505 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
507 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
508 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
510 /***********************************************************************/
512 typedef struct kib_rx /* receive message */
514 cfs_list_t rx_list; /* queue for attention */
515 struct kib_conn *rx_conn; /* owning conn */
516 int rx_nob; /* # bytes received (-1 while posted) */
517 enum ib_wc_status rx_status; /* completion status */
518 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
519 __u64 rx_msgaddr; /* message buffer (I/O addr) */
520 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
521 struct ib_recv_wr rx_wrq; /* receive work item... */
522 struct ib_sge rx_sge; /* ...and its memory */
525 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
526 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
527 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
528 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
530 typedef struct kib_tx /* transmit message */
532 cfs_list_t tx_list; /* queue on idle_txs ibc_tx_queue etc. */
533 kib_tx_pool_t *tx_pool; /* pool I'm from */
534 struct kib_conn *tx_conn; /* owning conn */
535 short tx_sending; /* # tx callbacks outstanding */
536 short tx_queued; /* queued for sending */
537 short tx_waiting; /* waiting for peer */
538 int tx_status; /* LNET completion status */
539 unsigned long tx_deadline; /* completion deadline */
540 __u64 tx_cookie; /* completion cookie */
541 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
542 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
543 __u64 tx_msgaddr; /* message buffer (I/O addr) */
544 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
545 int tx_nwrq; /* # send work items */
546 struct ib_send_wr *tx_wrq; /* send work items... */
547 struct ib_sge *tx_sge; /* ...and their memory */
548 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
549 int tx_nfrags; /* # entries in... */
550 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
551 __u64 *tx_pages; /* rdma phys page addrs */
553 kib_phys_mr_t *pmr; /* MR for physical buffer */
554 kib_fmr_t fmr; /* FMR */
556 int tx_dmadir; /* dma direction */
559 typedef struct kib_connvars
561 /* connection-in-progress variables */
565 typedef struct kib_conn
567 struct kib_peer *ibc_peer; /* owning peer */
568 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
569 cfs_list_t ibc_list; /* stash on peer's conn list */
570 cfs_list_t ibc_sched_list; /* schedule for attention */
571 __u16 ibc_version; /* version of connection */
572 __u64 ibc_incarnation; /* which instance of the peer */
573 cfs_atomic_t ibc_refcount; /* # users */
574 int ibc_state; /* what's happening */
575 int ibc_nsends_posted; /* # uncompleted sends */
576 int ibc_noops_posted; /* # uncompleted NOOPs */
577 int ibc_credits; /* # credits I have */
578 int ibc_outstanding_credits; /* # credits to return */
579 int ibc_reserved_credits;/* # ACK/DONE msg credits */
580 int ibc_comms_error; /* set on comms error */
581 int ibc_nrx:16; /* receive buffers owned */
582 int ibc_scheduled:1; /* scheduled for attention */
583 int ibc_ready:1; /* CQ callback fired */
584 unsigned long ibc_last_send; /* time of last send */
585 cfs_list_t ibc_early_rxs; /* rxs completed before ESTABLISHED */
586 cfs_list_t ibc_tx_noops; /* IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
587 cfs_list_t ibc_tx_queue; /* sends that need a credit */
588 cfs_list_t ibc_tx_queue_nocred;/* sends that don't need a credit */
589 cfs_list_t ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
590 cfs_list_t ibc_active_txs; /* active tx awaiting completion */
591 cfs_spinlock_t ibc_lock; /* serialise */
592 kib_rx_t *ibc_rxs; /* the rx descs */
593 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
595 struct rdma_cm_id *ibc_cmid; /* CM id */
596 struct ib_cq *ibc_cq; /* completion queue */
598 kib_connvars_t *ibc_connvars; /* in-progress connection state */
601 #define IBLND_CONN_INIT 0 /* being intialised */
602 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
603 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
604 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
605 #define IBLND_CONN_CLOSING 4 /* being closed */
606 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
608 typedef struct kib_peer
610 cfs_list_t ibp_list; /* stash on global peer list */
611 lnet_nid_t ibp_nid; /* who's on the other end(s) */
612 lnet_ni_t *ibp_ni; /* LNet interface */
613 cfs_atomic_t ibp_refcount; /* # users */
614 cfs_list_t ibp_conns; /* all active connections */
615 cfs_list_t ibp_tx_queue; /* msgs waiting for a conn */
616 __u16 ibp_version; /* version of peer */
617 __u64 ibp_incarnation; /* incarnation of peer */
618 int ibp_connecting; /* current active connection attempts */
619 int ibp_accepting; /* current passive connection attempts */
620 int ibp_error; /* errno on closing this peer */
621 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
624 extern kib_data_t kiblnd_data;
626 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
629 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
631 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
632 cfs_atomic_inc(&hdev->ibh_ref);
636 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
638 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
639 if (cfs_atomic_dec_and_test(&hdev->ibh_ref))
640 kiblnd_hdev_destroy(hdev);
644 kiblnd_dev_can_failover(kib_dev_t *dev)
646 if (!cfs_list_empty(&dev->ibd_fail_list)) /* already scheduled */
649 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
652 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
655 return dev->ibd_can_failover;
658 #define kiblnd_conn_addref(conn) \
660 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
661 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
662 cfs_atomic_inc(&(conn)->ibc_refcount); \
665 #define kiblnd_conn_decref(conn) \
667 unsigned long flags; \
669 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
670 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
671 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
672 if (cfs_atomic_dec_and_test(&(conn)->ibc_refcount)) { \
673 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
674 cfs_list_add_tail(&(conn)->ibc_list, \
675 &kiblnd_data.kib_connd_zombies); \
676 cfs_waitq_signal(&kiblnd_data.kib_connd_waitq); \
677 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
681 #define kiblnd_peer_addref(peer) \
683 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
684 (peer), libcfs_nid2str((peer)->ibp_nid), \
685 cfs_atomic_read (&(peer)->ibp_refcount)); \
686 cfs_atomic_inc(&(peer)->ibp_refcount); \
689 #define kiblnd_peer_decref(peer) \
691 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
692 (peer), libcfs_nid2str((peer)->ibp_nid), \
693 cfs_atomic_read (&(peer)->ibp_refcount)); \
694 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
695 if (cfs_atomic_dec_and_test(&(peer)->ibp_refcount)) \
696 kiblnd_destroy_peer(peer); \
699 static inline cfs_list_t *
700 kiblnd_nid2peerlist (lnet_nid_t nid)
703 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
705 return (&kiblnd_data.kib_peers [hash]);
709 kiblnd_peer_active (kib_peer_t *peer)
711 /* Am I in the peer hash table? */
712 return (!cfs_list_empty(&peer->ibp_list));
715 static inline kib_conn_t *
716 kiblnd_get_conn_locked (kib_peer_t *peer)
718 LASSERT (!cfs_list_empty(&peer->ibp_conns));
720 /* just return the first connection */
721 return cfs_list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
725 kiblnd_send_keepalive(kib_conn_t *conn)
727 return (*kiblnd_tunables.kib_keepalive > 0) &&
728 cfs_time_after(jiffies, conn->ibc_last_send +
729 *kiblnd_tunables.kib_keepalive*CFS_HZ);
733 kiblnd_send_noop(kib_conn_t *conn)
735 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
737 if (conn->ibc_outstanding_credits <
738 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
739 !kiblnd_send_keepalive(conn))
740 return 0; /* No need to send NOOP */
742 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
743 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred))
744 return 0; /* NOOP can be piggybacked */
746 /* No tx to piggyback NOOP onto or no credit to send a tx */
747 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 0);
750 if (!cfs_list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
751 !cfs_list_empty(&conn->ibc_tx_queue_nocred) || /* can be piggybacked */
752 conn->ibc_credits == 0) /* no credit */
755 if (conn->ibc_credits == 1 && /* last credit reserved for */
756 conn->ibc_outstanding_credits == 0) /* giving back credits */
759 /* No tx to piggyback NOOP onto or no credit to send a tx */
760 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
764 kiblnd_abort_receives(kib_conn_t *conn)
766 ib_modify_qp(conn->ibc_cmid->qp,
767 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
770 static inline const char *
771 kiblnd_queue2str (kib_conn_t *conn, cfs_list_t *q)
773 if (q == &conn->ibc_tx_queue)
776 if (q == &conn->ibc_tx_queue_rsrvd)
777 return "tx_queue_rsrvd";
779 if (q == &conn->ibc_tx_queue_nocred)
780 return "tx_queue_nocred";
782 if (q == &conn->ibc_active_txs)
789 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
790 * lowest bits of the work request id to stash the work item type. */
792 #define IBLND_WID_TX 0
793 #define IBLND_WID_RDMA 1
794 #define IBLND_WID_RX 2
795 #define IBLND_WID_MASK 3UL
798 kiblnd_ptr2wreqid (void *ptr, int type)
800 unsigned long lptr = (unsigned long)ptr;
802 LASSERT ((lptr & IBLND_WID_MASK) == 0);
803 LASSERT ((type & ~IBLND_WID_MASK) == 0);
804 return (__u64)(lptr | type);
808 kiblnd_wreqid2ptr (__u64 wreqid)
810 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
814 kiblnd_wreqid2type (__u64 wreqid)
816 return (wreqid & IBLND_WID_MASK);
820 kiblnd_set_conn_state (kib_conn_t *conn, int state)
822 conn->ibc_state = state;
827 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
829 msg->ibm_type = type;
830 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
834 kiblnd_rd_size (kib_rdma_desc_t *rd)
839 for (i = size = 0; i < rd->rd_nfrags; i++)
840 size += rd->rd_frags[i].rf_nob;
846 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
848 return rd->rd_frags[index].rf_addr;
852 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
854 return rd->rd_frags[index].rf_nob;
858 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
864 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
866 if (nob < rd->rd_frags[index].rf_nob) {
867 rd->rd_frags[index].rf_addr += nob;
868 rd->rd_frags[index].rf_nob -= nob;
877 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
879 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
880 msgtype == IBLND_MSG_PUT_ACK);
882 return msgtype == IBLND_MSG_GET_REQ ?
883 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
884 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
887 #ifdef HAVE_OFED_IB_DMA_MAP
890 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
892 return ib_dma_mapping_error(dev, dma_addr);
895 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
896 void *msg, size_t size,
897 enum dma_data_direction direction)
899 return ib_dma_map_single(dev, msg, size, direction);
902 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
903 __u64 addr, size_t size,
904 enum dma_data_direction direction)
906 ib_dma_unmap_single(dev, addr, size, direction);
909 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
910 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
912 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
913 struct scatterlist *sg, int nents,
914 enum dma_data_direction direction)
916 return ib_dma_map_sg(dev, sg, nents, direction);
919 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
920 struct scatterlist *sg, int nents,
921 enum dma_data_direction direction)
923 ib_dma_unmap_sg(dev, sg, nents, direction);
926 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
927 struct scatterlist *sg)
929 return ib_sg_dma_address(dev, sg);
932 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
933 struct scatterlist *sg)
935 return ib_sg_dma_len(dev, sg);
938 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
939 * right because OFED1.2 defines it as const, to use it we have to add
940 * (void *) cast to overcome "const" */
942 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
943 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
948 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
950 return dma_mapping_error(dma_addr);
953 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
954 void *msg, size_t size,
955 enum dma_data_direction direction)
957 return dma_map_single(dev->dma_device, msg, size, direction);
960 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
961 dma_addr_t addr, size_t size,
962 enum dma_data_direction direction)
964 dma_unmap_single(dev->dma_device, addr, size, direction);
967 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
968 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
970 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
971 struct scatterlist *sg, int nents,
972 enum dma_data_direction direction)
974 return dma_map_sg(dev->dma_device, sg, nents, direction);
977 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
978 struct scatterlist *sg, int nents,
979 enum dma_data_direction direction)
981 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
985 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
986 struct scatterlist *sg)
988 return sg_dma_address(sg);
992 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
993 struct scatterlist *sg)
995 return sg_dma_len(sg);
998 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
999 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
1003 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1004 kib_rdma_desc_t *rd);
1005 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
1006 __u64 addr, __u64 size);
1007 void kiblnd_map_rx_descs(kib_conn_t *conn);
1008 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1009 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
1010 kib_rdma_desc_t *rd, int nfrags);
1011 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
1012 void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node);
1013 cfs_list_t *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1015 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1016 int npages, __u64 iov, kib_fmr_t *fmr);
1017 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1019 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1020 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1021 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1023 int kiblnd_startup (lnet_ni_t *ni);
1024 void kiblnd_shutdown (lnet_ni_t *ni);
1025 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
1026 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1028 int kiblnd_tunables_init(void);
1029 void kiblnd_tunables_fini(void);
1031 int kiblnd_connd (void *arg);
1032 int kiblnd_scheduler(void *arg);
1033 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
1034 int kiblnd_failover_thread (void *arg);
1036 int kiblnd_alloc_pages (kib_pages_t **pp, int npages);
1037 void kiblnd_free_pages (kib_pages_t *p);
1039 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1040 struct rdma_cm_event *event);
1041 int kiblnd_translate_mtu(int value);
1043 int kiblnd_dev_failover(kib_dev_t *dev);
1044 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1045 void kiblnd_destroy_peer (kib_peer_t *peer);
1046 void kiblnd_destroy_dev (kib_dev_t *dev);
1047 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1048 void kiblnd_peer_alive (kib_peer_t *peer);
1049 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1050 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1051 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1052 int version, __u64 incarnation);
1053 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1055 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1056 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1057 int state, int version);
1058 void kiblnd_destroy_conn (kib_conn_t *conn);
1059 void kiblnd_close_conn (kib_conn_t *conn, int error);
1060 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1062 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1063 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1065 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1066 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1067 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1068 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1069 void kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist,
1071 void kiblnd_check_sends (kib_conn_t *conn);
1073 void kiblnd_qp_event(struct ib_event *event, void *arg);
1074 void kiblnd_cq_event(struct ib_event *event, void *arg);
1075 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1077 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1078 int credits, lnet_nid_t dstnid, __u64 dststamp);
1079 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1080 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1082 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1083 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1084 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1085 unsigned int offset, unsigned int mlen, unsigned int rlen);