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, 2012, 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 #include <linux/module.h>
42 #include <linux/kernel.h>
44 #include <linux/string.h>
45 #include <linux/stat.h>
46 #include <linux/errno.h>
47 #include <linux/unistd.h>
48 #include <linux/uio.h>
50 #include <asm/uaccess.h>
53 #include <linux/init.h>
55 #include <linux/file.h>
56 #include <linux/stat.h>
57 #include <linux/list.h>
58 #include <linux/kmod.h>
59 #include <linux/sysctl.h>
60 #include <linux/pci.h>
65 #define DEBUG_SUBSYSTEM S_LND
67 #include <linux/libcfs/libcfs.h>
68 #include <linux/lnet/lnet.h>
69 #include <linux/lnet/lib-lnet.h>
70 #include <linux/lnet/lnet-sysctl.h>
72 #include <rdma/rdma_cm.h>
73 #include <rdma/ib_cm.h>
74 #include <rdma/ib_verbs.h>
75 #include <rdma/ib_fmr_pool.h>
77 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
78 /* # scheduler loops before reschedule */
79 #define IBLND_RESCHED 100
81 #define IBLND_N_SCHED 2
82 #define IBLND_N_SCHED_HIGH 4
86 int *kib_dev_failover; /* HCA failover */
87 unsigned int *kib_service; /* IB service number */
88 int *kib_min_reconnect_interval; /* first failed connection retry... */
89 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
90 int *kib_cksum; /* checksum kib_msg_t? */
91 int *kib_timeout; /* comms timeout (seconds) */
92 int *kib_keepalive; /* keepalive timeout (seconds) */
93 int *kib_ntx; /* # tx descs */
94 int *kib_credits; /* # concurrent sends */
95 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
96 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
97 int *kib_peercredits_hiw; /* # when eagerly to return credits */
98 int *kib_peertimeout; /* seconds to consider peer dead */
99 char **kib_default_ipif; /* default IPoIB interface */
100 int *kib_retry_count;
101 int *kib_rnr_retry_count;
102 int *kib_concurrent_sends; /* send work queue sizing */
103 int *kib_ib_mtu; /* IB MTU */
104 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
105 * than this value, 0 disable map-on-demand */
106 int *kib_pmr_pool_size; /* # physical MR in pool */
107 int *kib_fmr_pool_size; /* # FMRs in pool */
108 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
109 int *kib_fmr_cache; /* enable FMR pool cache? */
110 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
111 ctl_table_header_t *kib_sysctl; /* sysctl interface */
113 int *kib_require_priv_port;/* accept only privileged ports */
114 int *kib_use_priv_port; /* use privileged port for active connect */
115 /* # threads on each CPT */
119 extern kib_tunables_t kiblnd_tunables;
121 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
122 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
124 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
125 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
127 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
128 IBLND_MSG_QUEUE_SIZE_V1 : \
129 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
130 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
131 IBLND_CREDIT_HIGHWATER_V1 : \
132 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
134 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
137 kiblnd_concurrent_sends_v1(void)
139 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
140 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
142 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
143 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
145 return *kiblnd_tunables.kib_concurrent_sends;
148 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
149 kiblnd_concurrent_sends_v1() : \
150 *kiblnd_tunables.kib_concurrent_sends)
151 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
152 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
153 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
155 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
156 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
157 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
158 *kiblnd_tunables.kib_map_on_demand : \
159 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
160 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
161 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
163 /************************/
164 /* derived constants... */
165 /* Pools (shared by connections on each CPT) */
166 /* These pools can grow at runtime, so don't need give a very large value */
167 #define IBLND_TX_POOL 256
168 #define IBLND_PMR_POOL 256
169 #define IBLND_FMR_POOL 256
170 #define IBLND_FMR_POOL_FLUSH 192
172 /* TX messages (shared by all connections) */
173 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
175 /* RX messages (per connection) */
176 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
177 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
178 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
180 /* WRs and CQEs (per connection) */
181 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
182 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
183 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
187 /* o2iblnd can run over aliased interface */
189 #define KIB_IFNAME_SIZE IFALIASZ
191 #define KIB_IFNAME_SIZE 256
196 struct list_head ibd_list; /* chain on kib_devs */
197 struct list_head ibd_fail_list; /* chain on kib_failed_devs */
198 __u32 ibd_ifip; /* IPoIB interface IP */
199 /** IPoIB interface name */
200 char ibd_ifname[KIB_IFNAME_SIZE];
201 int ibd_nnets; /* # nets extant */
203 cfs_time_t ibd_next_failover;
204 int ibd_failed_failover; /* # failover failures */
205 unsigned int ibd_failover; /* failover in progress */
206 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
207 struct list_head ibd_nets;
208 struct kib_hca_dev *ibd_hdev;
211 typedef struct kib_hca_dev
213 struct rdma_cm_id *ibh_cmid; /* listener cmid */
214 struct ib_device *ibh_ibdev; /* IB device */
215 int ibh_page_shift; /* page shift of current HCA */
216 int ibh_page_size; /* page size of current HCA */
217 __u64 ibh_page_mask; /* page mask of current HCA */
218 int ibh_mr_shift; /* bits shift of max MR size */
219 __u64 ibh_mr_size; /* size of MR */
220 int ibh_nmrs; /* # of global MRs */
221 struct ib_mr **ibh_mrs; /* global MR */
222 struct ib_pd *ibh_pd; /* PD */
223 kib_dev_t *ibh_dev; /* owner */
224 atomic_t ibh_ref; /* refcount */
227 /** # of seconds to keep pool alive */
228 #define IBLND_POOL_DEADLINE 300
229 /** # of seconds to retry if allocation failed */
230 #define IBLND_POOL_RETRY 1
234 int ibp_npages; /* # pages */
235 struct page *ibp_pages[0]; /* page array */
241 struct list_head pmr_list; /* chain node */
242 struct ib_phys_buf *pmr_ipb; /* physical buffer */
243 struct ib_mr *pmr_mr; /* IB MR */
244 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
245 __u64 pmr_iova; /* Virtual I/O address */
246 int pmr_refcount; /* reference count */
252 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
253 int inc, struct kib_pool **pp_po);
254 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
255 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
256 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
260 #define IBLND_POOL_NAME_LEN 32
262 typedef struct kib_poolset
264 spinlock_t ps_lock; /* serialize */
265 struct kib_net *ps_net; /* network it belongs to */
266 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
267 struct list_head ps_pool_list; /* list of pools */
268 struct list_head ps_failed_pool_list; /* failed pool list */
269 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
270 int ps_increasing; /* is allocating new pool */
271 int ps_pool_size; /* new pool size */
272 int ps_cpt; /* CPT id */
274 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
275 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
276 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
277 kib_ps_node_fini_t ps_node_fini; /* finalize node */
280 typedef struct kib_pool
282 struct list_head po_list; /* chain on pool list */
283 struct list_head po_free_list; /* pre-allocated node */
284 kib_poolset_t *po_owner; /* pool_set of this pool */
285 cfs_time_t po_deadline; /* deadline of this pool */
286 int po_allocated; /* # of elements in use */
287 int po_failed; /* pool is created on failed HCA */
288 int po_size; /* # of pre-allocated elements */
292 kib_poolset_t tps_poolset; /* pool-set */
293 __u64 tps_next_tx_cookie; /* cookie of TX */
297 kib_pool_t tpo_pool; /* pool */
298 struct kib_hca_dev *tpo_hdev; /* device for this pool */
299 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
300 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
304 kib_poolset_t pps_poolset; /* pool-set */
307 typedef struct kib_pmr_pool {
308 struct kib_hca_dev *ppo_hdev; /* device for this pool */
309 kib_pool_t ppo_pool; /* pool */
314 spinlock_t fps_lock; /* serialize */
315 struct kib_net *fps_net; /* IB network */
316 struct list_head fps_pool_list; /* FMR pool list */
317 struct list_head fps_failed_pool_list; /* FMR pool list */
318 __u64 fps_version; /* validity stamp */
319 int fps_cpt; /* CPT id */
321 int fps_flush_trigger;
322 /* is allocating new pool */
324 /* time stamp for retry if failed to allocate */
325 cfs_time_t fps_next_retry;
330 struct list_head fpo_list; /* chain on pool list */
331 struct kib_hca_dev *fpo_hdev; /* device for this pool */
332 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
333 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
334 cfs_time_t fpo_deadline; /* deadline of this pool */
335 int fpo_failed; /* fmr pool is failed */
336 int fpo_map_count; /* # of mapped FMR */
340 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
341 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
344 typedef struct kib_net
346 struct list_head ibn_list; /* chain on kib_dev_t::ibd_nets */
347 __u64 ibn_incarnation; /* my epoch */
348 int ibn_init; /* initialisation state */
349 int ibn_shutdown; /* shutting down? */
351 atomic_t ibn_npeers; /* # peers extant */
352 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 */
361 #define KIB_THREAD_SHIFT 16
362 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
363 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
364 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
366 struct kib_sched_info {
369 /* schedulers sleep here */
370 wait_queue_head_t ibs_waitq;
371 /* conns to check for rx completions */
372 struct list_head ibs_conns;
373 /* number of scheduler threads */
375 /* max allowed scheduler threads */
376 int ibs_nthreads_max;
377 int ibs_cpt; /* CPT id */
382 int kib_init; /* initialisation state */
383 int kib_shutdown; /* shut down? */
384 struct list_head kib_devs; /* IB devices extant */
385 /* list head of failed devices */
386 struct list_head kib_failed_devs;
387 /* schedulers sleep here */
388 wait_queue_head_t kib_failover_waitq;
389 atomic_t kib_nthreads; /* # live threads */
390 /* stabilize net/dev/peer/conn ops */
391 rwlock_t kib_global_lock;
392 /* hash table of all my known peers */
393 struct list_head *kib_peers;
394 /* size of kib_peers */
395 int kib_peer_hash_size;
396 /* the connd task (serialisation assertions) */
398 /* connections to setup/teardown */
399 struct list_head kib_connd_conns;
400 /* connections with zero refcount */
401 struct list_head kib_connd_zombies;
402 /* connection daemon sleeps here */
403 wait_queue_head_t kib_connd_waitq;
404 spinlock_t kib_connd_lock; /* serialise */
405 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
406 /* percpt data for schedulers */
407 struct kib_sched_info **kib_scheds;
410 #define IBLND_INIT_NOTHING 0
411 #define IBLND_INIT_DATA 1
412 #define IBLND_INIT_ALL 2
414 /************************************************************************
415 * IB Wire message format.
416 * These are sent in sender's byte order (i.e. receiver flips).
419 typedef struct kib_connparams
421 __u16 ibcp_queue_depth;
422 __u16 ibcp_max_frags;
423 __u32 ibcp_max_msg_size;
424 } WIRE_ATTR kib_connparams_t;
428 lnet_hdr_t ibim_hdr; /* portals header */
429 char ibim_payload[0]; /* piggy-backed payload */
430 } WIRE_ATTR kib_immediate_msg_t;
434 __u32 rf_nob; /* # bytes this frag */
435 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
436 } WIRE_ATTR kib_rdma_frag_t;
440 __u32 rd_key; /* local/remote key */
441 __u32 rd_nfrags; /* # fragments */
442 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
443 } WIRE_ATTR kib_rdma_desc_t;
447 lnet_hdr_t ibprm_hdr; /* portals header */
448 __u64 ibprm_cookie; /* opaque completion cookie */
449 } WIRE_ATTR kib_putreq_msg_t;
453 __u64 ibpam_src_cookie; /* reflected completion cookie */
454 __u64 ibpam_dst_cookie; /* opaque completion cookie */
455 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
456 } WIRE_ATTR kib_putack_msg_t;
460 lnet_hdr_t ibgm_hdr; /* portals header */
461 __u64 ibgm_cookie; /* opaque completion cookie */
462 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
463 } WIRE_ATTR kib_get_msg_t;
467 __u64 ibcm_cookie; /* opaque completion cookie */
468 __s32 ibcm_status; /* < 0 failure: >= 0 length */
469 } WIRE_ATTR kib_completion_msg_t;
473 /* First 2 fields fixed FOR ALL TIME */
474 __u32 ibm_magic; /* I'm an ibnal message */
475 __u16 ibm_version; /* this is my version number */
477 __u8 ibm_type; /* msg type */
478 __u8 ibm_credits; /* returned credits */
479 __u32 ibm_nob; /* # bytes in whole message */
480 __u32 ibm_cksum; /* checksum (0 == no checksum) */
481 __u64 ibm_srcnid; /* sender's NID */
482 __u64 ibm_srcstamp; /* sender's incarnation */
483 __u64 ibm_dstnid; /* destination's NID */
484 __u64 ibm_dststamp; /* destination's incarnation */
487 kib_connparams_t connparams;
488 kib_immediate_msg_t immediate;
489 kib_putreq_msg_t putreq;
490 kib_putack_msg_t putack;
492 kib_completion_msg_t completion;
494 } WIRE_ATTR kib_msg_t;
496 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
498 #define IBLND_MSG_VERSION_1 0x11
499 #define IBLND_MSG_VERSION_2 0x12
500 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
502 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
503 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
504 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
505 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
506 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
507 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
508 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
509 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
510 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
511 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
514 __u32 ibr_magic; /* sender's magic */
515 __u16 ibr_version; /* sender's version */
516 __u8 ibr_why; /* reject reason */
517 __u8 ibr_padding; /* padding */
518 __u64 ibr_incarnation; /* incarnation of peer */
519 kib_connparams_t ibr_cp; /* connection parameters */
520 } WIRE_ATTR kib_rej_t;
522 /* connection rejection reasons */
523 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
524 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
525 #define IBLND_REJECT_FATAL 3 /* Anything else */
527 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
528 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
530 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
531 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
533 /***********************************************************************/
535 typedef struct kib_rx /* receive message */
537 struct list_head rx_list; /* queue for attention */
538 struct kib_conn *rx_conn; /* owning conn */
539 int rx_nob; /* # bytes received (-1 while posted) */
540 enum ib_wc_status rx_status; /* completion status */
541 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
542 __u64 rx_msgaddr; /* message buffer (I/O addr) */
543 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
544 struct ib_recv_wr rx_wrq; /* receive work item... */
545 struct ib_sge rx_sge; /* ...and its memory */
548 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
549 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
550 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
551 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
553 typedef struct kib_tx /* transmit message */
555 struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */
556 kib_tx_pool_t *tx_pool; /* pool I'm from */
557 struct kib_conn *tx_conn; /* owning conn */
558 short tx_sending; /* # tx callbacks outstanding */
559 short tx_queued; /* queued for sending */
560 short tx_waiting; /* waiting for peer */
561 int tx_status; /* LNET completion status */
562 unsigned long tx_deadline; /* completion deadline */
563 __u64 tx_cookie; /* completion cookie */
564 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
565 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
566 __u64 tx_msgaddr; /* message buffer (I/O addr) */
567 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
568 int tx_nwrq; /* # send work items */
569 struct ib_send_wr *tx_wrq; /* send work items... */
570 struct ib_sge *tx_sge; /* ...and their memory */
571 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
572 int tx_nfrags; /* # entries in... */
573 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
574 __u64 *tx_pages; /* rdma phys page addrs */
576 kib_phys_mr_t *pmr; /* MR for physical buffer */
577 kib_fmr_t fmr; /* FMR */
579 int tx_dmadir; /* dma direction */
582 typedef struct kib_connvars
584 /* connection-in-progress variables */
588 typedef struct kib_conn
590 struct kib_sched_info *ibc_sched; /* scheduler information */
591 struct kib_peer *ibc_peer; /* owning peer */
592 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
593 struct list_head ibc_list; /* stash on peer's conn list */
594 struct list_head ibc_sched_list; /* schedule for attention */
595 __u16 ibc_version; /* version of connection */
596 __u64 ibc_incarnation; /* which instance of the peer */
597 atomic_t ibc_refcount; /* # users */
598 int ibc_state; /* what's happening */
599 int ibc_nsends_posted; /* # uncompleted sends */
600 int ibc_noops_posted; /* # uncompleted NOOPs */
601 int ibc_credits; /* # credits I have */
602 int ibc_outstanding_credits; /* # credits to return */
603 int ibc_reserved_credits;/* # ACK/DONE msg credits */
604 int ibc_comms_error; /* set on comms error */
605 unsigned int ibc_nrx:16; /* receive buffers owned */
606 unsigned int ibc_scheduled:1; /* scheduled for attention */
607 unsigned int ibc_ready:1; /* CQ callback fired */
608 /* time of last send */
609 unsigned long ibc_last_send;
610 /** link chain for kiblnd_check_conns only */
611 struct list_head ibc_connd_list;
612 /** rxs completed before ESTABLISHED */
613 struct list_head ibc_early_rxs;
614 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
615 struct list_head ibc_tx_noops;
616 struct list_head ibc_tx_queue; /* sends that need a credit */
617 struct list_head ibc_tx_queue_nocred;/* sends that don't need a credit */
618 struct list_head ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
619 struct list_head ibc_active_txs; /* active tx awaiting completion */
620 spinlock_t ibc_lock; /* serialise */
621 kib_rx_t *ibc_rxs; /* the rx descs */
622 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
624 struct rdma_cm_id *ibc_cmid; /* CM id */
625 struct ib_cq *ibc_cq; /* completion queue */
627 kib_connvars_t *ibc_connvars; /* in-progress connection state */
630 #define IBLND_CONN_INIT 0 /* being initialised */
631 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
632 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
633 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
634 #define IBLND_CONN_CLOSING 4 /* being closed */
635 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
637 typedef struct kib_peer
639 struct list_head ibp_list; /* stash on global peer list */
640 lnet_nid_t ibp_nid; /* who's on the other end(s) */
641 lnet_ni_t *ibp_ni; /* LNet interface */
642 atomic_t ibp_refcount; /* # users */
643 struct list_head ibp_conns; /* all active connections */
644 struct list_head ibp_tx_queue; /* msgs waiting for a conn */
645 __u16 ibp_version; /* version of peer */
646 __u64 ibp_incarnation; /* incarnation of peer */
647 int ibp_connecting; /* current active connection attempts */
648 int ibp_accepting; /* current passive connection attempts */
649 int ibp_error; /* errno on closing this peer */
650 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
653 extern kib_data_t kiblnd_data;
655 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
658 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
660 LASSERT (atomic_read(&hdev->ibh_ref) > 0);
661 atomic_inc(&hdev->ibh_ref);
665 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
667 LASSERT (atomic_read(&hdev->ibh_ref) > 0);
668 if (atomic_dec_and_test(&hdev->ibh_ref))
669 kiblnd_hdev_destroy(hdev);
673 kiblnd_dev_can_failover(kib_dev_t *dev)
675 if (!list_empty(&dev->ibd_fail_list)) /* already scheduled */
678 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
681 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
684 return dev->ibd_can_failover;
687 #define kiblnd_conn_addref(conn) \
689 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
690 (conn), atomic_read(&(conn)->ibc_refcount)); \
691 atomic_inc(&(conn)->ibc_refcount); \
694 #define kiblnd_conn_decref(conn) \
696 unsigned long flags; \
698 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
699 (conn), atomic_read(&(conn)->ibc_refcount)); \
700 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
701 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
702 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
703 list_add_tail(&(conn)->ibc_list, \
704 &kiblnd_data.kib_connd_zombies); \
705 wake_up(&kiblnd_data.kib_connd_waitq); \
706 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
710 #define kiblnd_peer_addref(peer) \
712 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
713 (peer), libcfs_nid2str((peer)->ibp_nid), \
714 atomic_read (&(peer)->ibp_refcount)); \
715 atomic_inc(&(peer)->ibp_refcount); \
718 #define kiblnd_peer_decref(peer) \
720 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
721 (peer), libcfs_nid2str((peer)->ibp_nid), \
722 atomic_read (&(peer)->ibp_refcount)); \
723 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
724 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
725 kiblnd_destroy_peer(peer); \
728 static inline struct list_head *
729 kiblnd_nid2peerlist (lnet_nid_t nid)
732 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
734 return (&kiblnd_data.kib_peers [hash]);
738 kiblnd_peer_active (kib_peer_t *peer)
740 /* Am I in the peer hash table? */
741 return (!list_empty(&peer->ibp_list));
744 static inline kib_conn_t *
745 kiblnd_get_conn_locked (kib_peer_t *peer)
747 LASSERT (!list_empty(&peer->ibp_conns));
749 /* just return the first connection */
750 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
754 kiblnd_send_keepalive(kib_conn_t *conn)
756 return (*kiblnd_tunables.kib_keepalive > 0) &&
757 cfs_time_after(jiffies, conn->ibc_last_send +
758 *kiblnd_tunables.kib_keepalive*HZ);
762 kiblnd_need_noop(kib_conn_t *conn)
764 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
766 if (conn->ibc_outstanding_credits <
767 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
768 !kiblnd_send_keepalive(conn))
769 return 0; /* No need to send NOOP */
771 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
772 if (!list_empty(&conn->ibc_tx_queue_nocred))
773 return 0; /* NOOP can be piggybacked */
775 /* No tx to piggyback NOOP onto or no credit to send a tx */
776 return (list_empty(&conn->ibc_tx_queue) ||
777 conn->ibc_credits == 0);
780 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
781 !list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
782 conn->ibc_credits == 0) /* no credit */
785 if (conn->ibc_credits == 1 && /* last credit reserved for */
786 conn->ibc_outstanding_credits == 0) /* giving back credits */
789 /* No tx to piggyback NOOP onto or no credit to send a tx */
790 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
794 kiblnd_abort_receives(kib_conn_t *conn)
796 ib_modify_qp(conn->ibc_cmid->qp,
797 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
800 static inline const char *
801 kiblnd_queue2str (kib_conn_t *conn, struct list_head *q)
803 if (q == &conn->ibc_tx_queue)
806 if (q == &conn->ibc_tx_queue_rsrvd)
807 return "tx_queue_rsrvd";
809 if (q == &conn->ibc_tx_queue_nocred)
810 return "tx_queue_nocred";
812 if (q == &conn->ibc_active_txs)
819 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
820 * lowest bits of the work request id to stash the work item type. */
822 #define IBLND_WID_TX 0
823 #define IBLND_WID_RDMA 1
824 #define IBLND_WID_RX 2
825 #define IBLND_WID_MASK 3UL
828 kiblnd_ptr2wreqid (void *ptr, int type)
830 unsigned long lptr = (unsigned long)ptr;
832 LASSERT ((lptr & IBLND_WID_MASK) == 0);
833 LASSERT ((type & ~IBLND_WID_MASK) == 0);
834 return (__u64)(lptr | type);
838 kiblnd_wreqid2ptr (__u64 wreqid)
840 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
844 kiblnd_wreqid2type (__u64 wreqid)
846 return (wreqid & IBLND_WID_MASK);
850 kiblnd_set_conn_state (kib_conn_t *conn, int state)
852 conn->ibc_state = state;
857 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
859 msg->ibm_type = type;
860 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
864 kiblnd_rd_size (kib_rdma_desc_t *rd)
869 for (i = size = 0; i < rd->rd_nfrags; i++)
870 size += rd->rd_frags[i].rf_nob;
876 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
878 return rd->rd_frags[index].rf_addr;
882 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
884 return rd->rd_frags[index].rf_nob;
888 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
894 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
896 if (nob < rd->rd_frags[index].rf_nob) {
897 rd->rd_frags[index].rf_addr += nob;
898 rd->rd_frags[index].rf_nob -= nob;
907 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
909 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
910 msgtype == IBLND_MSG_PUT_ACK);
912 return msgtype == IBLND_MSG_GET_REQ ?
913 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
914 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
919 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
921 return ib_dma_mapping_error(dev, dma_addr);
924 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
925 void *msg, size_t size,
926 enum dma_data_direction direction)
928 return ib_dma_map_single(dev, msg, size, direction);
931 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
932 __u64 addr, size_t size,
933 enum dma_data_direction direction)
935 ib_dma_unmap_single(dev, addr, size, direction);
938 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
939 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
941 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
942 struct scatterlist *sg, int nents,
943 enum dma_data_direction direction)
945 return ib_dma_map_sg(dev, sg, nents, direction);
948 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
949 struct scatterlist *sg, int nents,
950 enum dma_data_direction direction)
952 ib_dma_unmap_sg(dev, sg, nents, direction);
955 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
956 struct scatterlist *sg)
958 return ib_sg_dma_address(dev, sg);
961 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
962 struct scatterlist *sg)
964 return ib_sg_dma_len(dev, sg);
967 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
968 * right because OFED1.2 defines it as const, to use it we have to add
969 * (void *) cast to overcome "const" */
971 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
972 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
975 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
976 kib_rdma_desc_t *rd);
977 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
978 __u64 addr, __u64 size);
979 void kiblnd_map_rx_descs(kib_conn_t *conn);
980 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
981 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
982 kib_rdma_desc_t *rd, int nfrags);
983 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
984 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
985 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
987 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
988 int npages, __u64 iov, kib_fmr_t *fmr);
989 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
991 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
992 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
993 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
995 int kiblnd_startup (lnet_ni_t *ni);
996 void kiblnd_shutdown (lnet_ni_t *ni);
997 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
998 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1000 int kiblnd_tunables_init(void);
1001 void kiblnd_tunables_fini(void);
1003 int kiblnd_connd (void *arg);
1004 int kiblnd_scheduler(void *arg);
1005 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1006 int kiblnd_failover_thread (void *arg);
1008 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1009 void kiblnd_free_pages (kib_pages_t *p);
1011 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1012 struct rdma_cm_event *event);
1013 int kiblnd_translate_mtu(int value);
1015 int kiblnd_dev_failover(kib_dev_t *dev);
1016 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1017 void kiblnd_destroy_peer (kib_peer_t *peer);
1018 void kiblnd_destroy_dev (kib_dev_t *dev);
1019 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1020 void kiblnd_peer_alive (kib_peer_t *peer);
1021 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1022 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1023 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1024 int version, __u64 incarnation);
1025 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1027 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1028 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1029 int state, int version);
1030 void kiblnd_destroy_conn (kib_conn_t *conn);
1031 void kiblnd_close_conn (kib_conn_t *conn, int error);
1032 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1034 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1035 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1037 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1038 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1039 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1040 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1041 void kiblnd_txlist_done (lnet_ni_t *ni, struct list_head *txlist,
1043 void kiblnd_check_sends (kib_conn_t *conn);
1045 void kiblnd_qp_event(struct ib_event *event, void *arg);
1046 void kiblnd_cq_event(struct ib_event *event, void *arg);
1047 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1049 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1050 int credits, lnet_nid_t dstnid, __u64 dststamp);
1051 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1052 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1054 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1055 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1056 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1057 unsigned int offset, unsigned int mlen, unsigned int rlen);