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.
32 * Copyright (c) 2011, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * lnet/klnds/o2iblnd/o2iblnd.h
40 * Author: Eric Barton <eric@bartonsoftware.com>
44 # define EXPORT_SYMTAB
46 #ifndef AUTOCONF_INCLUDED
47 #include <linux/config.h>
49 #include <linux/module.h>
50 #include <linux/kernel.h>
52 #include <linux/string.h>
53 #include <linux/stat.h>
54 #include <linux/errno.h>
55 #include <linux/smp_lock.h>
56 #include <linux/unistd.h>
57 #include <linux/uio.h>
59 #include <asm/system.h>
60 #include <asm/uaccess.h>
63 #include <linux/init.h>
65 #include <linux/file.h>
66 #include <linux/stat.h>
67 #include <linux/list.h>
68 #include <linux/kmod.h>
69 #include <linux/sysctl.h>
70 #include <linux/pci.h>
75 #define DEBUG_SUBSYSTEM S_LND
77 #include <libcfs/libcfs.h>
78 #include <lnet/lnet.h>
79 #include <lnet/lib-lnet.h>
80 #include <lnet/lnet-sysctl.h>
86 #include <rdma/rdma_cm.h>
87 #include <rdma/ib_cm.h>
88 #include <rdma/ib_verbs.h>
89 #include <rdma/ib_fmr_pool.h>
91 /* tunables fixed at compile time */
93 # define IBLND_N_SCHED cfs_num_online_cpus() /* # schedulers */
95 # define IBLND_N_SCHED 1 /* # schedulers */
98 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
99 #define IBLND_RESCHED 100 /* # scheduler loops before reschedule */
103 int *kib_dev_failover; /* HCA failover */
104 unsigned int *kib_service; /* IB service number */
105 int *kib_min_reconnect_interval; /* first failed connection retry... */
106 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
107 int *kib_cksum; /* checksum kib_msg_t? */
108 int *kib_timeout; /* comms timeout (seconds) */
109 int *kib_keepalive; /* keepalive timeout (seconds) */
110 int *kib_ntx; /* # tx descs */
111 int *kib_credits; /* # concurrent sends */
112 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
113 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
114 int *kib_peercredits_hiw; /* # when eagerly to return credits */
115 int *kib_peertimeout; /* seconds to consider peer dead */
116 char **kib_default_ipif; /* default IPoIB interface */
117 int *kib_retry_count;
118 int *kib_rnr_retry_count;
119 int *kib_concurrent_sends; /* send work queue sizing */
120 int *kib_ib_mtu; /* IB MTU */
121 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
122 * than this value, 0 disable map-on-demand */
123 int *kib_pmr_pool_size; /* # physical MR in pool */
124 int *kib_fmr_pool_size; /* # FMRs in pool */
125 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
126 int *kib_fmr_cache; /* enable FMR pool cache? */
127 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
128 cfs_sysctl_table_header_t *kib_sysctl; /* sysctl interface */
130 int *kib_require_priv_port;/* accept only privileged ports */
131 int *kib_use_priv_port; /* use privileged port for active connect */
134 extern kib_tunables_t kiblnd_tunables;
136 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
137 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
139 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
140 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
142 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
143 IBLND_MSG_QUEUE_SIZE_V1 : \
144 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
145 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
146 IBLND_CREDIT_HIGHWATER_V1 : \
147 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
149 #ifdef HAVE_RDMA_CREATE_ID_4ARG
150 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
152 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
156 kiblnd_concurrent_sends_v1(void)
158 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
159 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
161 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
162 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
164 return *kiblnd_tunables.kib_concurrent_sends;
167 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
168 kiblnd_concurrent_sends_v1() : \
169 *kiblnd_tunables.kib_concurrent_sends)
170 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
171 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
172 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
174 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
175 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
176 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
177 *kiblnd_tunables.kib_map_on_demand : \
178 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
179 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
180 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
182 /************************/
183 /* derived constants... */
185 /* TX messages (shared by all connections) */
186 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
188 /* RX messages (per connection) */
189 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
190 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
191 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
193 /* WRs and CQEs (per connection) */
194 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
195 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
196 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
200 /* o2iblnd can run over aliased interface */
202 #define KIB_IFNAME_SIZE IFALIASZ
204 #define KIB_IFNAME_SIZE 256
209 cfs_list_t ibd_list; /* chain on kib_devs */
210 cfs_list_t ibd_fail_list; /* chain on kib_failed_devs */
211 __u32 ibd_ifip; /* IPoIB interface IP */
212 /** IPoIB interface name */
213 char ibd_ifname[KIB_IFNAME_SIZE];
214 int ibd_nnets; /* # nets extant */
216 cfs_time_t ibd_next_failover;
217 int ibd_failed_failover; /* # failover failures */
218 unsigned int ibd_failover; /* failover in progress */
219 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
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 int ibh_nmrs; /* # of global MRs */
234 struct ib_mr **ibh_mrs; /* global MR */
235 struct ib_pd *ibh_pd; /* PD */
236 kib_dev_t *ibh_dev; /* owner */
237 cfs_atomic_t ibh_ref; /* refcount */
240 /** # of seconds to keep pool alive */
241 #define IBLND_POOL_DEADLINE 300
242 /** # of seconds to retry if allocation failed */
243 #define IBLND_POOL_RETRY 1
247 int ibp_npages; /* # pages */
248 struct page *ibp_pages[0]; /* page array */
254 cfs_list_t pmr_list; /* chain node */
255 struct ib_phys_buf *pmr_ipb; /* physical buffer */
256 struct ib_mr *pmr_mr; /* IB MR */
257 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
258 __u64 pmr_iova; /* Virtual I/O address */
259 int pmr_refcount; /* reference count */
265 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps, int inc, struct kib_pool **pp_po);
266 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
267 typedef void (*kib_ps_node_init_t)(struct kib_pool *po,
269 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po,
274 #define IBLND_POOL_NAME_LEN 32
276 typedef struct kib_poolset
278 cfs_spinlock_t ps_lock; /* serialize */
279 struct kib_net *ps_net; /* network it belongs to */
280 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
281 cfs_list_t ps_pool_list; /* list of pools */
282 cfs_list_t ps_failed_pool_list; /* failed pool list */
283 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
284 int ps_increasing; /* is allocating new pool */
285 int ps_pool_size; /* new pool size */
287 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
288 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
289 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
290 kib_ps_node_fini_t ps_node_fini; /* finalize node */
293 typedef struct kib_pool
295 cfs_list_t po_list; /* chain on pool list */
296 cfs_list_t po_free_list; /* pre-allocated node */
297 kib_poolset_t *po_owner; /* pool_set of this pool */
298 cfs_time_t po_deadline; /* deadline of this pool */
299 int po_allocated; /* # of elements in use */
300 int po_failed; /* pool is created on failed HCA */
301 int po_size; /* # of pre-allocated elements */
305 kib_poolset_t tps_poolset; /* pool-set */
306 __u64 tps_next_tx_cookie; /* cookie of TX */
310 kib_pool_t tpo_pool; /* pool */
311 struct kib_hca_dev *tpo_hdev; /* device for this pool */
312 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
313 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
317 kib_poolset_t pps_poolset; /* pool-set */
320 typedef struct kib_pmr_pool {
321 struct kib_hca_dev *ppo_hdev; /* device for this pool */
322 kib_pool_t ppo_pool; /* pool */
327 cfs_spinlock_t fps_lock; /* serialize */
328 struct kib_net *fps_net; /* IB network */
329 cfs_list_t fps_pool_list; /* FMR pool list */
330 cfs_list_t fps_failed_pool_list; /* FMR pool list */
331 __u64 fps_version; /* validity stamp */
332 int fps_increasing; /* is allocating new pool */
333 cfs_time_t fps_next_retry; /* time stamp for retry if failed to allocate */
338 cfs_list_t fpo_list; /* chain on pool list */
339 struct kib_hca_dev *fpo_hdev; /* device for this pool */
340 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
341 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
342 cfs_time_t fpo_deadline; /* deadline of this pool */
343 int fpo_failed; /* fmr pool is failed */
344 int fpo_map_count; /* # of mapped FMR */
348 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
349 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
352 typedef struct kib_net
354 cfs_list_t ibn_list; /* chain on kib_dev_t::ibd_nets */
355 __u64 ibn_incarnation; /* my epoch */
356 int ibn_init; /* initialisation state */
357 int ibn_shutdown; /* shutting down? */
358 unsigned int ibn_with_fmr:1; /* FMR? */
359 unsigned int ibn_with_pmr:1; /* PMR? */
361 cfs_atomic_t ibn_npeers; /* # peers extant */
362 cfs_atomic_t ibn_nconns; /* # connections extant */
364 kib_tx_poolset_t ibn_tx_ps; /* tx pool-set */
365 kib_fmr_poolset_t ibn_fmr_ps; /* fmr pool-set */
366 kib_pmr_poolset_t ibn_pmr_ps; /* pmr pool-set */
368 kib_dev_t *ibn_dev; /* underlying IB device */
373 int kib_init; /* initialisation state */
374 int kib_shutdown; /* shut down? */
375 cfs_list_t kib_devs; /* IB devices extant */
376 cfs_list_t kib_failed_devs; /* list head of failed devices */
377 cfs_atomic_t kib_nthreads; /* # live threads */
378 cfs_rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
380 cfs_list_t *kib_peers; /* hash table of all my known peers */
381 int kib_peer_hash_size;/* size of kib_peers */
383 void *kib_connd; /* the connd task (serialisation assertions) */
384 cfs_list_t kib_connd_conns; /* connections to setup/teardown */
385 cfs_list_t kib_connd_zombies;/* connections with zero refcount */
386 cfs_waitq_t kib_connd_waitq; /* connection daemon sleeps here */
387 cfs_spinlock_t kib_connd_lock; /* serialise */
389 cfs_waitq_t kib_sched_waitq; /* schedulers sleep here */
390 cfs_list_t kib_sched_conns; /* conns to check for rx completions */
391 cfs_spinlock_t kib_sched_lock; /* serialise */
392 cfs_waitq_t kib_failover_waitq; /* schedulers sleep here */
394 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
397 #define IBLND_INIT_NOTHING 0
398 #define IBLND_INIT_DATA 1
399 #define IBLND_INIT_ALL 2
401 /************************************************************************
402 * IB Wire message format.
403 * These are sent in sender's byte order (i.e. receiver flips).
406 typedef struct kib_connparams
408 __u16 ibcp_queue_depth;
409 __u16 ibcp_max_frags;
410 __u32 ibcp_max_msg_size;
411 } WIRE_ATTR kib_connparams_t;
415 lnet_hdr_t ibim_hdr; /* portals header */
416 char ibim_payload[0]; /* piggy-backed payload */
417 } WIRE_ATTR kib_immediate_msg_t;
421 __u32 rf_nob; /* # bytes this frag */
422 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
423 } WIRE_ATTR kib_rdma_frag_t;
427 __u32 rd_key; /* local/remote key */
428 __u32 rd_nfrags; /* # fragments */
429 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
430 } WIRE_ATTR kib_rdma_desc_t;
434 lnet_hdr_t ibprm_hdr; /* portals header */
435 __u64 ibprm_cookie; /* opaque completion cookie */
436 } WIRE_ATTR kib_putreq_msg_t;
440 __u64 ibpam_src_cookie; /* reflected completion cookie */
441 __u64 ibpam_dst_cookie; /* opaque completion cookie */
442 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
443 } WIRE_ATTR kib_putack_msg_t;
447 lnet_hdr_t ibgm_hdr; /* portals header */
448 __u64 ibgm_cookie; /* opaque completion cookie */
449 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
450 } WIRE_ATTR kib_get_msg_t;
454 __u64 ibcm_cookie; /* opaque completion cookie */
455 __s32 ibcm_status; /* < 0 failure: >= 0 length */
456 } WIRE_ATTR kib_completion_msg_t;
460 /* First 2 fields fixed FOR ALL TIME */
461 __u32 ibm_magic; /* I'm an ibnal message */
462 __u16 ibm_version; /* this is my version number */
464 __u8 ibm_type; /* msg type */
465 __u8 ibm_credits; /* returned credits */
466 __u32 ibm_nob; /* # bytes in whole message */
467 __u32 ibm_cksum; /* checksum (0 == no checksum) */
468 __u64 ibm_srcnid; /* sender's NID */
469 __u64 ibm_srcstamp; /* sender's incarnation */
470 __u64 ibm_dstnid; /* destination's NID */
471 __u64 ibm_dststamp; /* destination's incarnation */
474 kib_connparams_t connparams;
475 kib_immediate_msg_t immediate;
476 kib_putreq_msg_t putreq;
477 kib_putack_msg_t putack;
479 kib_completion_msg_t completion;
481 } WIRE_ATTR kib_msg_t;
483 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
485 #define IBLND_MSG_VERSION_1 0x11
486 #define IBLND_MSG_VERSION_2 0x12
487 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
489 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
490 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
491 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
492 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
493 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
494 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
495 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
496 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
497 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
498 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
501 __u32 ibr_magic; /* sender's magic */
502 __u16 ibr_version; /* sender's version */
503 __u8 ibr_why; /* reject reason */
504 __u8 ibr_padding; /* padding */
505 __u64 ibr_incarnation; /* incarnation of peer */
506 kib_connparams_t ibr_cp; /* connection parameters */
507 } WIRE_ATTR kib_rej_t;
509 /* connection rejection reasons */
510 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
511 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
512 #define IBLND_REJECT_FATAL 3 /* Anything else */
514 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
515 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
517 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
518 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
520 /***********************************************************************/
522 typedef struct kib_rx /* receive message */
524 cfs_list_t rx_list; /* queue for attention */
525 struct kib_conn *rx_conn; /* owning conn */
526 int rx_nob; /* # bytes received (-1 while posted) */
527 enum ib_wc_status rx_status; /* completion status */
528 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
529 __u64 rx_msgaddr; /* message buffer (I/O addr) */
530 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
531 struct ib_recv_wr rx_wrq; /* receive work item... */
532 struct ib_sge rx_sge; /* ...and its memory */
535 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
536 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
537 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
538 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
540 typedef struct kib_tx /* transmit message */
542 cfs_list_t tx_list; /* queue on idle_txs ibc_tx_queue etc. */
543 kib_tx_pool_t *tx_pool; /* pool I'm from */
544 struct kib_conn *tx_conn; /* owning conn */
545 short tx_sending; /* # tx callbacks outstanding */
546 short tx_queued; /* queued for sending */
547 short tx_waiting; /* waiting for peer */
548 int tx_status; /* LNET completion status */
549 unsigned long tx_deadline; /* completion deadline */
550 __u64 tx_cookie; /* completion cookie */
551 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
552 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
553 __u64 tx_msgaddr; /* message buffer (I/O addr) */
554 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
555 int tx_nwrq; /* # send work items */
556 struct ib_send_wr *tx_wrq; /* send work items... */
557 struct ib_sge *tx_sge; /* ...and their memory */
558 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
559 int tx_nfrags; /* # entries in... */
560 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
561 __u64 *tx_pages; /* rdma phys page addrs */
563 kib_phys_mr_t *pmr; /* MR for physical buffer */
564 kib_fmr_t fmr; /* FMR */
566 int tx_dmadir; /* dma direction */
569 typedef struct kib_connvars
571 /* connection-in-progress variables */
575 typedef struct kib_conn
577 struct kib_peer *ibc_peer; /* owning peer */
578 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
579 cfs_list_t ibc_list; /* stash on peer's conn list */
580 cfs_list_t ibc_sched_list; /* schedule for attention */
581 __u16 ibc_version; /* version of connection */
582 __u64 ibc_incarnation; /* which instance of the peer */
583 cfs_atomic_t ibc_refcount; /* # users */
584 int ibc_state; /* what's happening */
585 int ibc_nsends_posted; /* # uncompleted sends */
586 int ibc_noops_posted; /* # uncompleted NOOPs */
587 int ibc_credits; /* # credits I have */
588 int ibc_outstanding_credits; /* # credits to return */
589 int ibc_reserved_credits;/* # ACK/DONE msg credits */
590 int ibc_comms_error; /* set on comms error */
591 int ibc_nrx:16; /* receive buffers owned */
592 int ibc_scheduled:1; /* scheduled for attention */
593 int ibc_ready:1; /* CQ callback fired */
594 /* time of last send */
595 unsigned long ibc_last_send;
596 /** link chain for kiblnd_check_conns only */
597 cfs_list_t ibc_connd_list;
598 /** rxs completed before ESTABLISHED */
599 cfs_list_t ibc_early_rxs;
600 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
601 cfs_list_t ibc_tx_noops;
602 cfs_list_t ibc_tx_queue; /* sends that need a credit */
603 cfs_list_t ibc_tx_queue_nocred;/* sends that don't need a credit */
604 cfs_list_t ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
605 cfs_list_t ibc_active_txs; /* active tx awaiting completion */
606 cfs_spinlock_t ibc_lock; /* serialise */
607 kib_rx_t *ibc_rxs; /* the rx descs */
608 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
610 struct rdma_cm_id *ibc_cmid; /* CM id */
611 struct ib_cq *ibc_cq; /* completion queue */
613 kib_connvars_t *ibc_connvars; /* in-progress connection state */
616 #define IBLND_CONN_INIT 0 /* being initialised */
617 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
618 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
619 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
620 #define IBLND_CONN_CLOSING 4 /* being closed */
621 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
623 typedef struct kib_peer
625 cfs_list_t ibp_list; /* stash on global peer list */
626 lnet_nid_t ibp_nid; /* who's on the other end(s) */
627 lnet_ni_t *ibp_ni; /* LNet interface */
628 cfs_atomic_t ibp_refcount; /* # users */
629 cfs_list_t ibp_conns; /* all active connections */
630 cfs_list_t ibp_tx_queue; /* msgs waiting for a conn */
631 __u16 ibp_version; /* version of peer */
632 __u64 ibp_incarnation; /* incarnation of peer */
633 int ibp_connecting; /* current active connection attempts */
634 int ibp_accepting; /* current passive connection attempts */
635 int ibp_error; /* errno on closing this peer */
636 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
639 extern kib_data_t kiblnd_data;
641 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
644 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
646 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
647 cfs_atomic_inc(&hdev->ibh_ref);
651 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
653 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
654 if (cfs_atomic_dec_and_test(&hdev->ibh_ref))
655 kiblnd_hdev_destroy(hdev);
659 kiblnd_dev_can_failover(kib_dev_t *dev)
661 if (!cfs_list_empty(&dev->ibd_fail_list)) /* already scheduled */
664 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
667 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
670 return dev->ibd_can_failover;
673 #define kiblnd_conn_addref(conn) \
675 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
676 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
677 cfs_atomic_inc(&(conn)->ibc_refcount); \
680 #define kiblnd_conn_decref(conn) \
682 unsigned long flags; \
684 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
685 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
686 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
687 if (cfs_atomic_dec_and_test(&(conn)->ibc_refcount)) { \
688 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
689 cfs_list_add_tail(&(conn)->ibc_list, \
690 &kiblnd_data.kib_connd_zombies); \
691 cfs_waitq_signal(&kiblnd_data.kib_connd_waitq); \
692 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
696 #define kiblnd_peer_addref(peer) \
698 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
699 (peer), libcfs_nid2str((peer)->ibp_nid), \
700 cfs_atomic_read (&(peer)->ibp_refcount)); \
701 cfs_atomic_inc(&(peer)->ibp_refcount); \
704 #define kiblnd_peer_decref(peer) \
706 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
707 (peer), libcfs_nid2str((peer)->ibp_nid), \
708 cfs_atomic_read (&(peer)->ibp_refcount)); \
709 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
710 if (cfs_atomic_dec_and_test(&(peer)->ibp_refcount)) \
711 kiblnd_destroy_peer(peer); \
714 static inline cfs_list_t *
715 kiblnd_nid2peerlist (lnet_nid_t nid)
718 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
720 return (&kiblnd_data.kib_peers [hash]);
724 kiblnd_peer_active (kib_peer_t *peer)
726 /* Am I in the peer hash table? */
727 return (!cfs_list_empty(&peer->ibp_list));
730 static inline kib_conn_t *
731 kiblnd_get_conn_locked (kib_peer_t *peer)
733 LASSERT (!cfs_list_empty(&peer->ibp_conns));
735 /* just return the first connection */
736 return cfs_list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
740 kiblnd_send_keepalive(kib_conn_t *conn)
742 return (*kiblnd_tunables.kib_keepalive > 0) &&
743 cfs_time_after(jiffies, conn->ibc_last_send +
744 *kiblnd_tunables.kib_keepalive*CFS_HZ);
748 kiblnd_need_noop(kib_conn_t *conn)
750 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
752 if (conn->ibc_outstanding_credits <
753 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
754 !kiblnd_send_keepalive(conn))
755 return 0; /* No need to send NOOP */
757 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
758 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred))
759 return 0; /* NOOP can be piggybacked */
761 /* No tx to piggyback NOOP onto or no credit to send a tx */
762 return (cfs_list_empty(&conn->ibc_tx_queue) ||
763 conn->ibc_credits == 0);
766 if (!cfs_list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
767 !cfs_list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
768 conn->ibc_credits == 0) /* no credit */
771 if (conn->ibc_credits == 1 && /* last credit reserved for */
772 conn->ibc_outstanding_credits == 0) /* giving back credits */
775 /* No tx to piggyback NOOP onto or no credit to send a tx */
776 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
780 kiblnd_abort_receives(kib_conn_t *conn)
782 ib_modify_qp(conn->ibc_cmid->qp,
783 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
786 static inline const char *
787 kiblnd_queue2str (kib_conn_t *conn, cfs_list_t *q)
789 if (q == &conn->ibc_tx_queue)
792 if (q == &conn->ibc_tx_queue_rsrvd)
793 return "tx_queue_rsrvd";
795 if (q == &conn->ibc_tx_queue_nocred)
796 return "tx_queue_nocred";
798 if (q == &conn->ibc_active_txs)
805 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
806 * lowest bits of the work request id to stash the work item type. */
808 #define IBLND_WID_TX 0
809 #define IBLND_WID_RDMA 1
810 #define IBLND_WID_RX 2
811 #define IBLND_WID_MASK 3UL
814 kiblnd_ptr2wreqid (void *ptr, int type)
816 unsigned long lptr = (unsigned long)ptr;
818 LASSERT ((lptr & IBLND_WID_MASK) == 0);
819 LASSERT ((type & ~IBLND_WID_MASK) == 0);
820 return (__u64)(lptr | type);
824 kiblnd_wreqid2ptr (__u64 wreqid)
826 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
830 kiblnd_wreqid2type (__u64 wreqid)
832 return (wreqid & IBLND_WID_MASK);
836 kiblnd_set_conn_state (kib_conn_t *conn, int state)
838 conn->ibc_state = state;
843 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
845 msg->ibm_type = type;
846 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
850 kiblnd_rd_size (kib_rdma_desc_t *rd)
855 for (i = size = 0; i < rd->rd_nfrags; i++)
856 size += rd->rd_frags[i].rf_nob;
862 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
864 return rd->rd_frags[index].rf_addr;
868 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
870 return rd->rd_frags[index].rf_nob;
874 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
880 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
882 if (nob < rd->rd_frags[index].rf_nob) {
883 rd->rd_frags[index].rf_addr += nob;
884 rd->rd_frags[index].rf_nob -= nob;
893 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
895 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
896 msgtype == IBLND_MSG_PUT_ACK);
898 return msgtype == IBLND_MSG_GET_REQ ?
899 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
900 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
903 #ifdef HAVE_OFED_IB_DMA_MAP
906 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
908 return ib_dma_mapping_error(dev, dma_addr);
911 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
912 void *msg, size_t size,
913 enum dma_data_direction direction)
915 return ib_dma_map_single(dev, msg, size, direction);
918 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
919 __u64 addr, size_t size,
920 enum dma_data_direction direction)
922 ib_dma_unmap_single(dev, addr, size, direction);
925 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
926 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
928 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
929 struct scatterlist *sg, int nents,
930 enum dma_data_direction direction)
932 return ib_dma_map_sg(dev, sg, nents, direction);
935 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
936 struct scatterlist *sg, int nents,
937 enum dma_data_direction direction)
939 ib_dma_unmap_sg(dev, sg, nents, direction);
942 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
943 struct scatterlist *sg)
945 return ib_sg_dma_address(dev, sg);
948 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
949 struct scatterlist *sg)
951 return ib_sg_dma_len(dev, sg);
954 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
955 * right because OFED1.2 defines it as const, to use it we have to add
956 * (void *) cast to overcome "const" */
958 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
959 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
964 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
966 return dma_mapping_error(dma_addr);
969 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
970 void *msg, size_t size,
971 enum dma_data_direction direction)
973 return dma_map_single(dev->dma_device, msg, size, direction);
976 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
977 dma_addr_t addr, size_t size,
978 enum dma_data_direction direction)
980 dma_unmap_single(dev->dma_device, addr, size, direction);
983 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
984 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
986 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
987 struct scatterlist *sg, int nents,
988 enum dma_data_direction direction)
990 return dma_map_sg(dev->dma_device, sg, nents, direction);
993 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
994 struct scatterlist *sg, int nents,
995 enum dma_data_direction direction)
997 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
1001 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
1002 struct scatterlist *sg)
1004 return sg_dma_address(sg);
1008 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1009 struct scatterlist *sg)
1011 return sg_dma_len(sg);
1014 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
1015 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
1019 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1020 kib_rdma_desc_t *rd);
1021 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
1022 __u64 addr, __u64 size);
1023 void kiblnd_map_rx_descs(kib_conn_t *conn);
1024 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1025 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
1026 kib_rdma_desc_t *rd, int nfrags);
1027 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
1028 void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node);
1029 cfs_list_t *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1031 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1032 int npages, __u64 iov, kib_fmr_t *fmr);
1033 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1035 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1036 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1037 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1039 int kiblnd_startup (lnet_ni_t *ni);
1040 void kiblnd_shutdown (lnet_ni_t *ni);
1041 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
1042 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1044 int kiblnd_tunables_init(void);
1045 void kiblnd_tunables_fini(void);
1047 int kiblnd_connd (void *arg);
1048 int kiblnd_scheduler(void *arg);
1049 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
1050 int kiblnd_failover_thread (void *arg);
1052 int kiblnd_alloc_pages (kib_pages_t **pp, int npages);
1053 void kiblnd_free_pages (kib_pages_t *p);
1055 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1056 struct rdma_cm_event *event);
1057 int kiblnd_translate_mtu(int value);
1059 int kiblnd_dev_failover(kib_dev_t *dev);
1060 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1061 void kiblnd_destroy_peer (kib_peer_t *peer);
1062 void kiblnd_destroy_dev (kib_dev_t *dev);
1063 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1064 void kiblnd_peer_alive (kib_peer_t *peer);
1065 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1066 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1067 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1068 int version, __u64 incarnation);
1069 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1071 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1072 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1073 int state, int version);
1074 void kiblnd_destroy_conn (kib_conn_t *conn);
1075 void kiblnd_close_conn (kib_conn_t *conn, int error);
1076 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1078 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1079 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1081 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1082 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1083 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1084 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1085 void kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist,
1087 void kiblnd_check_sends (kib_conn_t *conn);
1089 void kiblnd_qp_event(struct ib_event *event, void *arg);
1090 void kiblnd_cq_event(struct ib_event *event, void *arg);
1091 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1093 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1094 int credits, lnet_nid_t dstnid, __u64 dststamp);
1095 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1096 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1098 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1099 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1100 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1101 unsigned int offset, unsigned int mlen, unsigned int rlen);