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, Whamcloud, Inc.
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 #ifndef AUTOCONF_INCLUDED
42 #include <linux/config.h>
44 #include <linux/module.h>
45 #include <linux/kernel.h>
47 #include <linux/string.h>
48 #include <linux/stat.h>
49 #include <linux/errno.h>
50 #include <linux/smp_lock.h>
51 #include <linux/unistd.h>
52 #include <linux/uio.h>
54 #include <asm/system.h>
55 #include <asm/uaccess.h>
58 #include <linux/init.h>
60 #include <linux/file.h>
61 #include <linux/stat.h>
62 #include <linux/list.h>
63 #include <linux/kmod.h>
64 #include <linux/sysctl.h>
65 #include <linux/pci.h>
70 #define DEBUG_SUBSYSTEM S_LND
72 #include <libcfs/libcfs.h>
73 #include <lnet/lnet.h>
74 #include <lnet/lib-lnet.h>
75 #include <lnet/lnet-sysctl.h>
77 #include <rdma/rdma_cm.h>
78 #include <rdma/ib_cm.h>
79 #include <rdma/ib_verbs.h>
80 #include <rdma/ib_fmr_pool.h>
82 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
83 /* # scheduler loops before reschedule */
84 #define IBLND_RESCHED 100
86 #define IBLND_N_SCHED 2
87 #define IBLND_N_SCHED_HIGH 4
91 int *kib_dev_failover; /* HCA failover */
92 unsigned int *kib_service; /* IB service number */
93 int *kib_min_reconnect_interval; /* first failed connection retry... */
94 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
95 int *kib_cksum; /* checksum kib_msg_t? */
96 int *kib_timeout; /* comms timeout (seconds) */
97 int *kib_keepalive; /* keepalive timeout (seconds) */
98 int *kib_ntx; /* # tx descs */
99 int *kib_credits; /* # concurrent sends */
100 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
101 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
102 int *kib_peercredits_hiw; /* # when eagerly to return credits */
103 int *kib_peertimeout; /* seconds to consider peer dead */
104 char **kib_default_ipif; /* default IPoIB interface */
105 int *kib_retry_count;
106 int *kib_rnr_retry_count;
107 int *kib_concurrent_sends; /* send work queue sizing */
108 int *kib_ib_mtu; /* IB MTU */
109 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
110 * than this value, 0 disable map-on-demand */
111 int *kib_pmr_pool_size; /* # physical MR in pool */
112 int *kib_fmr_pool_size; /* # FMRs in pool */
113 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
114 int *kib_fmr_cache; /* enable FMR pool cache? */
115 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
116 cfs_sysctl_table_header_t *kib_sysctl; /* sysctl interface */
118 int *kib_require_priv_port;/* accept only privileged ports */
119 int *kib_use_priv_port; /* use privileged port for active connect */
120 /* # threads on each CPT */
124 extern kib_tunables_t kiblnd_tunables;
126 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
127 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
129 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
130 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
132 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
133 IBLND_MSG_QUEUE_SIZE_V1 : \
134 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
135 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
136 IBLND_CREDIT_HIGHWATER_V1 : \
137 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
139 #ifdef HAVE_RDMA_CREATE_ID_4ARG
140 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
142 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
146 kiblnd_concurrent_sends_v1(void)
148 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
149 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
151 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
152 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
154 return *kiblnd_tunables.kib_concurrent_sends;
157 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
158 kiblnd_concurrent_sends_v1() : \
159 *kiblnd_tunables.kib_concurrent_sends)
160 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
161 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
162 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
164 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
165 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
166 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
167 *kiblnd_tunables.kib_map_on_demand : \
168 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
169 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
170 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
172 /************************/
173 /* derived constants... */
174 /* Pools (shared by connections on each CPT) */
175 /* These pools can grow at runtime, so don't need give a very large value */
176 #define IBLND_TX_POOL 256
177 #define IBLND_PMR_POOL 256
178 #define IBLND_FMR_POOL 256
179 #define IBLND_FMR_POOL_FLUSH 192
181 /* TX messages (shared by all connections) */
182 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
184 /* RX messages (per connection) */
185 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
186 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
187 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
189 /* WRs and CQEs (per connection) */
190 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
191 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
192 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
196 /* o2iblnd can run over aliased interface */
198 #define KIB_IFNAME_SIZE IFALIASZ
200 #define KIB_IFNAME_SIZE 256
205 cfs_list_t ibd_list; /* chain on kib_devs */
206 cfs_list_t ibd_fail_list; /* chain on kib_failed_devs */
207 __u32 ibd_ifip; /* IPoIB interface IP */
208 /** IPoIB interface name */
209 char ibd_ifname[KIB_IFNAME_SIZE];
210 int ibd_nnets; /* # nets extant */
212 cfs_time_t ibd_next_failover;
213 int ibd_failed_failover; /* # failover failures */
214 unsigned int ibd_failover; /* failover in progress */
215 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
217 struct kib_hca_dev *ibd_hdev;
220 typedef struct kib_hca_dev
222 struct rdma_cm_id *ibh_cmid; /* listener cmid */
223 struct ib_device *ibh_ibdev; /* IB device */
224 int ibh_page_shift; /* page shift of current HCA */
225 int ibh_page_size; /* page size of current HCA */
226 __u64 ibh_page_mask; /* page mask of current HCA */
227 int ibh_mr_shift; /* bits shift of max MR size */
228 __u64 ibh_mr_size; /* size of MR */
229 int ibh_nmrs; /* # of global MRs */
230 struct ib_mr **ibh_mrs; /* global MR */
231 struct ib_pd *ibh_pd; /* PD */
232 kib_dev_t *ibh_dev; /* owner */
233 cfs_atomic_t ibh_ref; /* refcount */
236 /** # of seconds to keep pool alive */
237 #define IBLND_POOL_DEADLINE 300
238 /** # of seconds to retry if allocation failed */
239 #define IBLND_POOL_RETRY 1
243 int ibp_npages; /* # pages */
244 struct page *ibp_pages[0]; /* page array */
250 cfs_list_t pmr_list; /* chain node */
251 struct ib_phys_buf *pmr_ipb; /* physical buffer */
252 struct ib_mr *pmr_mr; /* IB MR */
253 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
254 __u64 pmr_iova; /* Virtual I/O address */
255 int pmr_refcount; /* reference count */
261 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
262 int inc, struct kib_pool **pp_po);
263 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
264 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, cfs_list_t *node);
265 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, cfs_list_t *node);
269 #define IBLND_POOL_NAME_LEN 32
271 typedef struct kib_poolset
273 cfs_spinlock_t ps_lock; /* serialize */
274 struct kib_net *ps_net; /* network it belongs to */
275 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
276 cfs_list_t ps_pool_list; /* list of pools */
277 cfs_list_t ps_failed_pool_list; /* failed pool list */
278 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
279 int ps_increasing; /* is allocating new pool */
280 int ps_pool_size; /* new pool size */
281 int ps_cpt; /* CPT id */
283 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
284 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
285 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
286 kib_ps_node_fini_t ps_node_fini; /* finalize node */
289 typedef struct kib_pool
291 cfs_list_t po_list; /* chain on pool list */
292 cfs_list_t po_free_list; /* pre-allocated node */
293 kib_poolset_t *po_owner; /* pool_set of this pool */
294 cfs_time_t po_deadline; /* deadline of this pool */
295 int po_allocated; /* # of elements in use */
296 int po_failed; /* pool is created on failed HCA */
297 int po_size; /* # of pre-allocated elements */
301 kib_poolset_t tps_poolset; /* pool-set */
302 __u64 tps_next_tx_cookie; /* cookie of TX */
306 kib_pool_t tpo_pool; /* pool */
307 struct kib_hca_dev *tpo_hdev; /* device for this pool */
308 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
309 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
313 kib_poolset_t pps_poolset; /* pool-set */
316 typedef struct kib_pmr_pool {
317 struct kib_hca_dev *ppo_hdev; /* device for this pool */
318 kib_pool_t ppo_pool; /* pool */
323 cfs_spinlock_t fps_lock; /* serialize */
324 struct kib_net *fps_net; /* IB network */
325 cfs_list_t fps_pool_list; /* FMR pool list */
326 cfs_list_t fps_failed_pool_list; /* FMR pool list */
327 __u64 fps_version; /* validity stamp */
328 int fps_cpt; /* CPT id */
330 int fps_flush_trigger;
331 /* is allocating new pool */
333 /* time stamp for retry if failed to allocate */
334 cfs_time_t fps_next_retry;
339 cfs_list_t fpo_list; /* chain on pool list */
340 struct kib_hca_dev *fpo_hdev; /* device for this pool */
341 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
342 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
343 cfs_time_t fpo_deadline; /* deadline of this pool */
344 int fpo_failed; /* fmr pool is failed */
345 int fpo_map_count; /* # of mapped FMR */
349 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
350 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
353 typedef struct kib_net
355 cfs_list_t ibn_list; /* chain on kib_dev_t::ibd_nets */
356 __u64 ibn_incarnation; /* my epoch */
357 int ibn_init; /* initialisation state */
358 int ibn_shutdown; /* shutting down? */
360 cfs_atomic_t ibn_npeers; /* # peers extant */
361 cfs_atomic_t ibn_nconns; /* # connections extant */
363 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
364 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
365 kib_pmr_poolset_t **ibn_pmr_ps; /* pmr pool-set */
367 kib_dev_t *ibn_dev; /* underlying IB device */
370 #define KIB_THREAD_SHIFT 16
371 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
372 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
373 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
375 struct kib_sched_info {
377 cfs_spinlock_t ibs_lock;
378 /* schedulers sleep here */
379 cfs_waitq_t ibs_waitq;
380 /* conns to check for rx completions */
381 cfs_list_t ibs_conns;
382 /* number of scheduler threads */
384 /* max allowed scheduler threads */
385 int ibs_nthreads_max;
386 int ibs_cpt; /* CPT id */
391 int kib_init; /* initialisation state */
392 int kib_shutdown; /* shut down? */
393 cfs_list_t kib_devs; /* IB devices extant */
394 /* list head of failed devices */
395 cfs_list_t kib_failed_devs;
396 /* schedulers sleep here */
397 cfs_waitq_t kib_failover_waitq;
398 cfs_atomic_t kib_nthreads; /* # live threads */
399 /* stabilize net/dev/peer/conn ops */
400 cfs_rwlock_t kib_global_lock;
401 /* hash table of all my known peers */
402 cfs_list_t *kib_peers;
403 /* size of kib_peers */
404 int kib_peer_hash_size;
405 /* the connd task (serialisation assertions) */
407 /* connections to setup/teardown */
408 cfs_list_t kib_connd_conns;
409 /* connections with zero refcount */
410 cfs_list_t kib_connd_zombies;
411 /* connection daemon sleeps here */
412 cfs_waitq_t kib_connd_waitq;
413 cfs_spinlock_t kib_connd_lock; /* serialise */
414 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
415 /* percpt data for schedulers */
416 struct kib_sched_info **kib_scheds;
419 #define IBLND_INIT_NOTHING 0
420 #define IBLND_INIT_DATA 1
421 #define IBLND_INIT_ALL 2
423 /************************************************************************
424 * IB Wire message format.
425 * These are sent in sender's byte order (i.e. receiver flips).
428 typedef struct kib_connparams
430 __u16 ibcp_queue_depth;
431 __u16 ibcp_max_frags;
432 __u32 ibcp_max_msg_size;
433 } WIRE_ATTR kib_connparams_t;
437 lnet_hdr_t ibim_hdr; /* portals header */
438 char ibim_payload[0]; /* piggy-backed payload */
439 } WIRE_ATTR kib_immediate_msg_t;
443 __u32 rf_nob; /* # bytes this frag */
444 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
445 } WIRE_ATTR kib_rdma_frag_t;
449 __u32 rd_key; /* local/remote key */
450 __u32 rd_nfrags; /* # fragments */
451 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
452 } WIRE_ATTR kib_rdma_desc_t;
456 lnet_hdr_t ibprm_hdr; /* portals header */
457 __u64 ibprm_cookie; /* opaque completion cookie */
458 } WIRE_ATTR kib_putreq_msg_t;
462 __u64 ibpam_src_cookie; /* reflected completion cookie */
463 __u64 ibpam_dst_cookie; /* opaque completion cookie */
464 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
465 } WIRE_ATTR kib_putack_msg_t;
469 lnet_hdr_t ibgm_hdr; /* portals header */
470 __u64 ibgm_cookie; /* opaque completion cookie */
471 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
472 } WIRE_ATTR kib_get_msg_t;
476 __u64 ibcm_cookie; /* opaque completion cookie */
477 __s32 ibcm_status; /* < 0 failure: >= 0 length */
478 } WIRE_ATTR kib_completion_msg_t;
482 /* First 2 fields fixed FOR ALL TIME */
483 __u32 ibm_magic; /* I'm an ibnal message */
484 __u16 ibm_version; /* this is my version number */
486 __u8 ibm_type; /* msg type */
487 __u8 ibm_credits; /* returned credits */
488 __u32 ibm_nob; /* # bytes in whole message */
489 __u32 ibm_cksum; /* checksum (0 == no checksum) */
490 __u64 ibm_srcnid; /* sender's NID */
491 __u64 ibm_srcstamp; /* sender's incarnation */
492 __u64 ibm_dstnid; /* destination's NID */
493 __u64 ibm_dststamp; /* destination's incarnation */
496 kib_connparams_t connparams;
497 kib_immediate_msg_t immediate;
498 kib_putreq_msg_t putreq;
499 kib_putack_msg_t putack;
501 kib_completion_msg_t completion;
503 } WIRE_ATTR kib_msg_t;
505 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
507 #define IBLND_MSG_VERSION_1 0x11
508 #define IBLND_MSG_VERSION_2 0x12
509 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
511 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
512 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
513 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
514 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
515 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
516 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
517 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
518 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
519 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
520 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
523 __u32 ibr_magic; /* sender's magic */
524 __u16 ibr_version; /* sender's version */
525 __u8 ibr_why; /* reject reason */
526 __u8 ibr_padding; /* padding */
527 __u64 ibr_incarnation; /* incarnation of peer */
528 kib_connparams_t ibr_cp; /* connection parameters */
529 } WIRE_ATTR kib_rej_t;
531 /* connection rejection reasons */
532 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
533 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
534 #define IBLND_REJECT_FATAL 3 /* Anything else */
536 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
537 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
539 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
540 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
542 /***********************************************************************/
544 typedef struct kib_rx /* receive message */
546 cfs_list_t rx_list; /* queue for attention */
547 struct kib_conn *rx_conn; /* owning conn */
548 int rx_nob; /* # bytes received (-1 while posted) */
549 enum ib_wc_status rx_status; /* completion status */
550 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
551 __u64 rx_msgaddr; /* message buffer (I/O addr) */
552 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
553 struct ib_recv_wr rx_wrq; /* receive work item... */
554 struct ib_sge rx_sge; /* ...and its memory */
557 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
558 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
559 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
560 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
562 typedef struct kib_tx /* transmit message */
564 cfs_list_t tx_list; /* queue on idle_txs ibc_tx_queue etc. */
565 kib_tx_pool_t *tx_pool; /* pool I'm from */
566 struct kib_conn *tx_conn; /* owning conn */
567 short tx_sending; /* # tx callbacks outstanding */
568 short tx_queued; /* queued for sending */
569 short tx_waiting; /* waiting for peer */
570 int tx_status; /* LNET completion status */
571 unsigned long tx_deadline; /* completion deadline */
572 __u64 tx_cookie; /* completion cookie */
573 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
574 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
575 __u64 tx_msgaddr; /* message buffer (I/O addr) */
576 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
577 int tx_nwrq; /* # send work items */
578 struct ib_send_wr *tx_wrq; /* send work items... */
579 struct ib_sge *tx_sge; /* ...and their memory */
580 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
581 int tx_nfrags; /* # entries in... */
582 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
583 __u64 *tx_pages; /* rdma phys page addrs */
585 kib_phys_mr_t *pmr; /* MR for physical buffer */
586 kib_fmr_t fmr; /* FMR */
588 int tx_dmadir; /* dma direction */
591 typedef struct kib_connvars
593 /* connection-in-progress variables */
597 typedef struct kib_conn
599 struct kib_sched_info *ibc_sched; /* scheduler information */
600 struct kib_peer *ibc_peer; /* owning peer */
601 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
602 cfs_list_t ibc_list; /* stash on peer's conn list */
603 cfs_list_t ibc_sched_list; /* schedule for attention */
604 __u16 ibc_version; /* version of connection */
605 __u64 ibc_incarnation; /* which instance of the peer */
606 cfs_atomic_t ibc_refcount; /* # users */
607 int ibc_state; /* what's happening */
608 int ibc_nsends_posted; /* # uncompleted sends */
609 int ibc_noops_posted; /* # uncompleted NOOPs */
610 int ibc_credits; /* # credits I have */
611 int ibc_outstanding_credits; /* # credits to return */
612 int ibc_reserved_credits;/* # ACK/DONE msg credits */
613 int ibc_comms_error; /* set on comms error */
614 int ibc_nrx:16; /* receive buffers owned */
615 int ibc_scheduled:1; /* scheduled for attention */
616 int ibc_ready:1; /* CQ callback fired */
617 /* time of last send */
618 unsigned long ibc_last_send;
619 /** link chain for kiblnd_check_conns only */
620 cfs_list_t ibc_connd_list;
621 /** rxs completed before ESTABLISHED */
622 cfs_list_t ibc_early_rxs;
623 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
624 cfs_list_t ibc_tx_noops;
625 cfs_list_t ibc_tx_queue; /* sends that need a credit */
626 cfs_list_t ibc_tx_queue_nocred;/* sends that don't need a credit */
627 cfs_list_t ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
628 cfs_list_t ibc_active_txs; /* active tx awaiting completion */
629 cfs_spinlock_t ibc_lock; /* serialise */
630 kib_rx_t *ibc_rxs; /* the rx descs */
631 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
633 struct rdma_cm_id *ibc_cmid; /* CM id */
634 struct ib_cq *ibc_cq; /* completion queue */
636 kib_connvars_t *ibc_connvars; /* in-progress connection state */
639 #define IBLND_CONN_INIT 0 /* being initialised */
640 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
641 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
642 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
643 #define IBLND_CONN_CLOSING 4 /* being closed */
644 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
646 typedef struct kib_peer
648 cfs_list_t ibp_list; /* stash on global peer list */
649 lnet_nid_t ibp_nid; /* who's on the other end(s) */
650 lnet_ni_t *ibp_ni; /* LNet interface */
651 cfs_atomic_t ibp_refcount; /* # users */
652 cfs_list_t ibp_conns; /* all active connections */
653 cfs_list_t ibp_tx_queue; /* msgs waiting for a conn */
654 __u16 ibp_version; /* version of peer */
655 __u64 ibp_incarnation; /* incarnation of peer */
656 int ibp_connecting; /* current active connection attempts */
657 int ibp_accepting; /* current passive connection attempts */
658 int ibp_error; /* errno on closing this peer */
659 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
662 extern kib_data_t kiblnd_data;
664 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
667 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
669 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
670 cfs_atomic_inc(&hdev->ibh_ref);
674 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
676 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
677 if (cfs_atomic_dec_and_test(&hdev->ibh_ref))
678 kiblnd_hdev_destroy(hdev);
682 kiblnd_dev_can_failover(kib_dev_t *dev)
684 if (!cfs_list_empty(&dev->ibd_fail_list)) /* already scheduled */
687 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
690 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
693 return dev->ibd_can_failover;
696 #define kiblnd_conn_addref(conn) \
698 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
699 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
700 cfs_atomic_inc(&(conn)->ibc_refcount); \
703 #define kiblnd_conn_decref(conn) \
705 unsigned long flags; \
707 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
708 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
709 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
710 if (cfs_atomic_dec_and_test(&(conn)->ibc_refcount)) { \
711 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
712 cfs_list_add_tail(&(conn)->ibc_list, \
713 &kiblnd_data.kib_connd_zombies); \
714 cfs_waitq_signal(&kiblnd_data.kib_connd_waitq); \
715 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
719 #define kiblnd_peer_addref(peer) \
721 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
722 (peer), libcfs_nid2str((peer)->ibp_nid), \
723 cfs_atomic_read (&(peer)->ibp_refcount)); \
724 cfs_atomic_inc(&(peer)->ibp_refcount); \
727 #define kiblnd_peer_decref(peer) \
729 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
730 (peer), libcfs_nid2str((peer)->ibp_nid), \
731 cfs_atomic_read (&(peer)->ibp_refcount)); \
732 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
733 if (cfs_atomic_dec_and_test(&(peer)->ibp_refcount)) \
734 kiblnd_destroy_peer(peer); \
737 static inline cfs_list_t *
738 kiblnd_nid2peerlist (lnet_nid_t nid)
741 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
743 return (&kiblnd_data.kib_peers [hash]);
747 kiblnd_peer_active (kib_peer_t *peer)
749 /* Am I in the peer hash table? */
750 return (!cfs_list_empty(&peer->ibp_list));
753 static inline kib_conn_t *
754 kiblnd_get_conn_locked (kib_peer_t *peer)
756 LASSERT (!cfs_list_empty(&peer->ibp_conns));
758 /* just return the first connection */
759 return cfs_list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
763 kiblnd_send_keepalive(kib_conn_t *conn)
765 return (*kiblnd_tunables.kib_keepalive > 0) &&
766 cfs_time_after(jiffies, conn->ibc_last_send +
767 *kiblnd_tunables.kib_keepalive*CFS_HZ);
771 kiblnd_need_noop(kib_conn_t *conn)
773 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
775 if (conn->ibc_outstanding_credits <
776 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
777 !kiblnd_send_keepalive(conn))
778 return 0; /* No need to send NOOP */
780 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
781 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred))
782 return 0; /* NOOP can be piggybacked */
784 /* No tx to piggyback NOOP onto or no credit to send a tx */
785 return (cfs_list_empty(&conn->ibc_tx_queue) ||
786 conn->ibc_credits == 0);
789 if (!cfs_list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
790 !cfs_list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
791 conn->ibc_credits == 0) /* no credit */
794 if (conn->ibc_credits == 1 && /* last credit reserved for */
795 conn->ibc_outstanding_credits == 0) /* giving back credits */
798 /* No tx to piggyback NOOP onto or no credit to send a tx */
799 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
803 kiblnd_abort_receives(kib_conn_t *conn)
805 ib_modify_qp(conn->ibc_cmid->qp,
806 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
809 static inline const char *
810 kiblnd_queue2str (kib_conn_t *conn, cfs_list_t *q)
812 if (q == &conn->ibc_tx_queue)
815 if (q == &conn->ibc_tx_queue_rsrvd)
816 return "tx_queue_rsrvd";
818 if (q == &conn->ibc_tx_queue_nocred)
819 return "tx_queue_nocred";
821 if (q == &conn->ibc_active_txs)
828 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
829 * lowest bits of the work request id to stash the work item type. */
831 #define IBLND_WID_TX 0
832 #define IBLND_WID_RDMA 1
833 #define IBLND_WID_RX 2
834 #define IBLND_WID_MASK 3UL
837 kiblnd_ptr2wreqid (void *ptr, int type)
839 unsigned long lptr = (unsigned long)ptr;
841 LASSERT ((lptr & IBLND_WID_MASK) == 0);
842 LASSERT ((type & ~IBLND_WID_MASK) == 0);
843 return (__u64)(lptr | type);
847 kiblnd_wreqid2ptr (__u64 wreqid)
849 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
853 kiblnd_wreqid2type (__u64 wreqid)
855 return (wreqid & IBLND_WID_MASK);
859 kiblnd_set_conn_state (kib_conn_t *conn, int state)
861 conn->ibc_state = state;
866 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
868 msg->ibm_type = type;
869 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
873 kiblnd_rd_size (kib_rdma_desc_t *rd)
878 for (i = size = 0; i < rd->rd_nfrags; i++)
879 size += rd->rd_frags[i].rf_nob;
885 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
887 return rd->rd_frags[index].rf_addr;
891 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
893 return rd->rd_frags[index].rf_nob;
897 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
903 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
905 if (nob < rd->rd_frags[index].rf_nob) {
906 rd->rd_frags[index].rf_addr += nob;
907 rd->rd_frags[index].rf_nob -= nob;
916 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
918 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
919 msgtype == IBLND_MSG_PUT_ACK);
921 return msgtype == IBLND_MSG_GET_REQ ?
922 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
923 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
926 #ifdef HAVE_OFED_IB_DMA_MAP
929 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
931 return ib_dma_mapping_error(dev, dma_addr);
934 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
935 void *msg, size_t size,
936 enum dma_data_direction direction)
938 return ib_dma_map_single(dev, msg, size, direction);
941 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
942 __u64 addr, size_t size,
943 enum dma_data_direction direction)
945 ib_dma_unmap_single(dev, addr, size, direction);
948 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
949 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
951 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
952 struct scatterlist *sg, int nents,
953 enum dma_data_direction direction)
955 return ib_dma_map_sg(dev, sg, nents, direction);
958 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
959 struct scatterlist *sg, int nents,
960 enum dma_data_direction direction)
962 ib_dma_unmap_sg(dev, sg, nents, direction);
965 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
966 struct scatterlist *sg)
968 return ib_sg_dma_address(dev, sg);
971 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
972 struct scatterlist *sg)
974 return ib_sg_dma_len(dev, sg);
977 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
978 * right because OFED1.2 defines it as const, to use it we have to add
979 * (void *) cast to overcome "const" */
981 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
982 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
987 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
989 return dma_mapping_error(dma_addr);
992 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
993 void *msg, size_t size,
994 enum dma_data_direction direction)
996 return dma_map_single(dev->dma_device, msg, size, direction);
999 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1000 dma_addr_t addr, size_t size,
1001 enum dma_data_direction direction)
1003 dma_unmap_single(dev->dma_device, addr, size, direction);
1006 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
1007 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
1009 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1010 struct scatterlist *sg, int nents,
1011 enum dma_data_direction direction)
1013 return dma_map_sg(dev->dma_device, sg, nents, direction);
1016 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1017 struct scatterlist *sg, int nents,
1018 enum dma_data_direction direction)
1020 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
1024 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
1025 struct scatterlist *sg)
1027 return sg_dma_address(sg);
1031 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1032 struct scatterlist *sg)
1034 return sg_dma_len(sg);
1037 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
1038 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
1042 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1043 kib_rdma_desc_t *rd);
1044 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
1045 __u64 addr, __u64 size);
1046 void kiblnd_map_rx_descs(kib_conn_t *conn);
1047 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1048 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
1049 kib_rdma_desc_t *rd, int nfrags);
1050 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
1051 void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node);
1052 cfs_list_t *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1054 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1055 int npages, __u64 iov, kib_fmr_t *fmr);
1056 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1058 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1059 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1060 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1062 int kiblnd_startup (lnet_ni_t *ni);
1063 void kiblnd_shutdown (lnet_ni_t *ni);
1064 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
1065 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1067 int kiblnd_tunables_init(void);
1068 void kiblnd_tunables_fini(void);
1070 int kiblnd_connd (void *arg);
1071 int kiblnd_scheduler(void *arg);
1072 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
1073 int kiblnd_failover_thread (void *arg);
1075 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1076 void kiblnd_free_pages (kib_pages_t *p);
1078 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1079 struct rdma_cm_event *event);
1080 int kiblnd_translate_mtu(int value);
1082 int kiblnd_dev_failover(kib_dev_t *dev);
1083 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1084 void kiblnd_destroy_peer (kib_peer_t *peer);
1085 void kiblnd_destroy_dev (kib_dev_t *dev);
1086 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1087 void kiblnd_peer_alive (kib_peer_t *peer);
1088 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1089 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1090 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1091 int version, __u64 incarnation);
1092 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1094 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1095 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1096 int state, int version);
1097 void kiblnd_destroy_conn (kib_conn_t *conn);
1098 void kiblnd_close_conn (kib_conn_t *conn, int error);
1099 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1101 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1102 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1104 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1105 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1106 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1107 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1108 void kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist,
1110 void kiblnd_check_sends (kib_conn_t *conn);
1112 void kiblnd_qp_event(struct ib_event *event, void *arg);
1113 void kiblnd_cq_event(struct ib_event *event, void *arg);
1114 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1116 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1117 int credits, lnet_nid_t dstnid, __u64 dststamp);
1118 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1119 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1121 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1122 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1123 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1124 unsigned int offset, unsigned int mlen, unsigned int rlen);