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 #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 <libcfs/libcfs.h>
68 #include <lnet/lnet.h>
69 #include <lnet/lib-lnet.h>
70 #include <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 cfs_sysctl_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 #ifdef HAVE_RDMA_CREATE_ID_4ARG
135 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
137 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
141 kiblnd_concurrent_sends_v1(void)
143 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
144 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
146 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
147 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
149 return *kiblnd_tunables.kib_concurrent_sends;
152 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
153 kiblnd_concurrent_sends_v1() : \
154 *kiblnd_tunables.kib_concurrent_sends)
155 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
156 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
157 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
159 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
160 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
161 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
162 *kiblnd_tunables.kib_map_on_demand : \
163 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
164 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
165 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
167 /************************/
168 /* derived constants... */
169 /* Pools (shared by connections on each CPT) */
170 /* These pools can grow at runtime, so don't need give a very large value */
171 #define IBLND_TX_POOL 256
172 #define IBLND_PMR_POOL 256
173 #define IBLND_FMR_POOL 256
174 #define IBLND_FMR_POOL_FLUSH 192
176 /* TX messages (shared by all connections) */
177 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
179 /* RX messages (per connection) */
180 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
181 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
182 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
184 /* WRs and CQEs (per connection) */
185 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
186 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
187 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
191 /* o2iblnd can run over aliased interface */
193 #define KIB_IFNAME_SIZE IFALIASZ
195 #define KIB_IFNAME_SIZE 256
200 cfs_list_t ibd_list; /* chain on kib_devs */
201 cfs_list_t ibd_fail_list; /* chain on kib_failed_devs */
202 __u32 ibd_ifip; /* IPoIB interface IP */
203 /** IPoIB interface name */
204 char ibd_ifname[KIB_IFNAME_SIZE];
205 int ibd_nnets; /* # nets extant */
207 cfs_time_t ibd_next_failover;
208 int ibd_failed_failover; /* # failover failures */
209 unsigned int ibd_failover; /* failover in progress */
210 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
212 struct kib_hca_dev *ibd_hdev;
215 typedef struct kib_hca_dev
217 struct rdma_cm_id *ibh_cmid; /* listener cmid */
218 struct ib_device *ibh_ibdev; /* IB device */
219 int ibh_page_shift; /* page shift of current HCA */
220 int ibh_page_size; /* page size of current HCA */
221 __u64 ibh_page_mask; /* page mask of current HCA */
222 int ibh_mr_shift; /* bits shift of max MR size */
223 __u64 ibh_mr_size; /* size of MR */
224 int ibh_nmrs; /* # of global MRs */
225 struct ib_mr **ibh_mrs; /* global MR */
226 struct ib_pd *ibh_pd; /* PD */
227 kib_dev_t *ibh_dev; /* owner */
228 cfs_atomic_t ibh_ref; /* refcount */
231 /** # of seconds to keep pool alive */
232 #define IBLND_POOL_DEADLINE 300
233 /** # of seconds to retry if allocation failed */
234 #define IBLND_POOL_RETRY 1
238 int ibp_npages; /* # pages */
239 struct page *ibp_pages[0]; /* page array */
245 cfs_list_t pmr_list; /* chain node */
246 struct ib_phys_buf *pmr_ipb; /* physical buffer */
247 struct ib_mr *pmr_mr; /* IB MR */
248 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
249 __u64 pmr_iova; /* Virtual I/O address */
250 int pmr_refcount; /* reference count */
256 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
257 int inc, struct kib_pool **pp_po);
258 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
259 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, cfs_list_t *node);
260 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, cfs_list_t *node);
264 #define IBLND_POOL_NAME_LEN 32
266 typedef struct kib_poolset
268 spinlock_t ps_lock; /* serialize */
269 struct kib_net *ps_net; /* network it belongs to */
270 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
271 cfs_list_t ps_pool_list; /* list of pools */
272 cfs_list_t ps_failed_pool_list; /* failed pool list */
273 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
274 int ps_increasing; /* is allocating new pool */
275 int ps_pool_size; /* new pool size */
276 int ps_cpt; /* CPT id */
278 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
279 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
280 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
281 kib_ps_node_fini_t ps_node_fini; /* finalize node */
284 typedef struct kib_pool
286 cfs_list_t po_list; /* chain on pool list */
287 cfs_list_t po_free_list; /* pre-allocated node */
288 kib_poolset_t *po_owner; /* pool_set of this pool */
289 cfs_time_t po_deadline; /* deadline of this pool */
290 int po_allocated; /* # of elements in use */
291 int po_failed; /* pool is created on failed HCA */
292 int po_size; /* # of pre-allocated elements */
296 kib_poolset_t tps_poolset; /* pool-set */
297 __u64 tps_next_tx_cookie; /* cookie of TX */
301 kib_pool_t tpo_pool; /* pool */
302 struct kib_hca_dev *tpo_hdev; /* device for this pool */
303 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
304 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
308 kib_poolset_t pps_poolset; /* pool-set */
311 typedef struct kib_pmr_pool {
312 struct kib_hca_dev *ppo_hdev; /* device for this pool */
313 kib_pool_t ppo_pool; /* pool */
318 spinlock_t fps_lock; /* serialize */
319 struct kib_net *fps_net; /* IB network */
320 cfs_list_t fps_pool_list; /* FMR pool list */
321 cfs_list_t fps_failed_pool_list; /* FMR pool list */
322 __u64 fps_version; /* validity stamp */
323 int fps_cpt; /* CPT id */
325 int fps_flush_trigger;
326 /* is allocating new pool */
328 /* time stamp for retry if failed to allocate */
329 cfs_time_t fps_next_retry;
334 cfs_list_t fpo_list; /* chain on pool list */
335 struct kib_hca_dev *fpo_hdev; /* device for this pool */
336 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
337 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
338 cfs_time_t fpo_deadline; /* deadline of this pool */
339 int fpo_failed; /* fmr pool is failed */
340 int fpo_map_count; /* # of mapped FMR */
344 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
345 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
348 typedef struct kib_net
350 cfs_list_t ibn_list; /* chain on kib_dev_t::ibd_nets */
351 __u64 ibn_incarnation; /* my epoch */
352 int ibn_init; /* initialisation state */
353 int ibn_shutdown; /* shutting down? */
355 cfs_atomic_t ibn_npeers; /* # peers extant */
356 cfs_atomic_t ibn_nconns; /* # connections extant */
358 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
359 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
360 kib_pmr_poolset_t **ibn_pmr_ps; /* pmr pool-set */
362 kib_dev_t *ibn_dev; /* underlying IB device */
365 #define KIB_THREAD_SHIFT 16
366 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
367 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
368 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
370 struct kib_sched_info {
373 /* schedulers sleep here */
374 cfs_waitq_t ibs_waitq;
375 /* conns to check for rx completions */
376 cfs_list_t ibs_conns;
377 /* number of scheduler threads */
379 /* max allowed scheduler threads */
380 int ibs_nthreads_max;
381 int ibs_cpt; /* CPT id */
386 int kib_init; /* initialisation state */
387 int kib_shutdown; /* shut down? */
388 cfs_list_t kib_devs; /* IB devices extant */
389 /* list head of failed devices */
390 cfs_list_t kib_failed_devs;
391 /* schedulers sleep here */
392 cfs_waitq_t kib_failover_waitq;
393 cfs_atomic_t kib_nthreads; /* # live threads */
394 /* stabilize net/dev/peer/conn ops */
395 rwlock_t kib_global_lock;
396 /* hash table of all my known peers */
397 cfs_list_t *kib_peers;
398 /* size of kib_peers */
399 int kib_peer_hash_size;
400 /* the connd task (serialisation assertions) */
402 /* connections to setup/teardown */
403 cfs_list_t kib_connd_conns;
404 /* connections with zero refcount */
405 cfs_list_t kib_connd_zombies;
406 /* connection daemon sleeps here */
407 cfs_waitq_t kib_connd_waitq;
408 spinlock_t kib_connd_lock; /* serialise */
409 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
410 /* percpt data for schedulers */
411 struct kib_sched_info **kib_scheds;
414 #define IBLND_INIT_NOTHING 0
415 #define IBLND_INIT_DATA 1
416 #define IBLND_INIT_ALL 2
418 /************************************************************************
419 * IB Wire message format.
420 * These are sent in sender's byte order (i.e. receiver flips).
423 typedef struct kib_connparams
425 __u16 ibcp_queue_depth;
426 __u16 ibcp_max_frags;
427 __u32 ibcp_max_msg_size;
428 } WIRE_ATTR kib_connparams_t;
432 lnet_hdr_t ibim_hdr; /* portals header */
433 char ibim_payload[0]; /* piggy-backed payload */
434 } WIRE_ATTR kib_immediate_msg_t;
438 __u32 rf_nob; /* # bytes this frag */
439 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
440 } WIRE_ATTR kib_rdma_frag_t;
444 __u32 rd_key; /* local/remote key */
445 __u32 rd_nfrags; /* # fragments */
446 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
447 } WIRE_ATTR kib_rdma_desc_t;
451 lnet_hdr_t ibprm_hdr; /* portals header */
452 __u64 ibprm_cookie; /* opaque completion cookie */
453 } WIRE_ATTR kib_putreq_msg_t;
457 __u64 ibpam_src_cookie; /* reflected completion cookie */
458 __u64 ibpam_dst_cookie; /* opaque completion cookie */
459 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
460 } WIRE_ATTR kib_putack_msg_t;
464 lnet_hdr_t ibgm_hdr; /* portals header */
465 __u64 ibgm_cookie; /* opaque completion cookie */
466 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
467 } WIRE_ATTR kib_get_msg_t;
471 __u64 ibcm_cookie; /* opaque completion cookie */
472 __s32 ibcm_status; /* < 0 failure: >= 0 length */
473 } WIRE_ATTR kib_completion_msg_t;
477 /* First 2 fields fixed FOR ALL TIME */
478 __u32 ibm_magic; /* I'm an ibnal message */
479 __u16 ibm_version; /* this is my version number */
481 __u8 ibm_type; /* msg type */
482 __u8 ibm_credits; /* returned credits */
483 __u32 ibm_nob; /* # bytes in whole message */
484 __u32 ibm_cksum; /* checksum (0 == no checksum) */
485 __u64 ibm_srcnid; /* sender's NID */
486 __u64 ibm_srcstamp; /* sender's incarnation */
487 __u64 ibm_dstnid; /* destination's NID */
488 __u64 ibm_dststamp; /* destination's incarnation */
491 kib_connparams_t connparams;
492 kib_immediate_msg_t immediate;
493 kib_putreq_msg_t putreq;
494 kib_putack_msg_t putack;
496 kib_completion_msg_t completion;
498 } WIRE_ATTR kib_msg_t;
500 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
502 #define IBLND_MSG_VERSION_1 0x11
503 #define IBLND_MSG_VERSION_2 0x12
504 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
506 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
507 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
508 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
509 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
510 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
511 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
512 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
513 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
514 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
515 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
518 __u32 ibr_magic; /* sender's magic */
519 __u16 ibr_version; /* sender's version */
520 __u8 ibr_why; /* reject reason */
521 __u8 ibr_padding; /* padding */
522 __u64 ibr_incarnation; /* incarnation of peer */
523 kib_connparams_t ibr_cp; /* connection parameters */
524 } WIRE_ATTR kib_rej_t;
526 /* connection rejection reasons */
527 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
528 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
529 #define IBLND_REJECT_FATAL 3 /* Anything else */
531 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
532 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
534 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
535 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
537 /***********************************************************************/
539 typedef struct kib_rx /* receive message */
541 cfs_list_t rx_list; /* queue for attention */
542 struct kib_conn *rx_conn; /* owning conn */
543 int rx_nob; /* # bytes received (-1 while posted) */
544 enum ib_wc_status rx_status; /* completion status */
545 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
546 __u64 rx_msgaddr; /* message buffer (I/O addr) */
547 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
548 struct ib_recv_wr rx_wrq; /* receive work item... */
549 struct ib_sge rx_sge; /* ...and its memory */
552 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
553 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
554 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
555 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
557 typedef struct kib_tx /* transmit message */
559 cfs_list_t tx_list; /* queue on idle_txs ibc_tx_queue etc. */
560 kib_tx_pool_t *tx_pool; /* pool I'm from */
561 struct kib_conn *tx_conn; /* owning conn */
562 short tx_sending; /* # tx callbacks outstanding */
563 short tx_queued; /* queued for sending */
564 short tx_waiting; /* waiting for peer */
565 int tx_status; /* LNET completion status */
566 unsigned long tx_deadline; /* completion deadline */
567 __u64 tx_cookie; /* completion cookie */
568 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
569 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
570 __u64 tx_msgaddr; /* message buffer (I/O addr) */
571 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
572 int tx_nwrq; /* # send work items */
573 struct ib_send_wr *tx_wrq; /* send work items... */
574 struct ib_sge *tx_sge; /* ...and their memory */
575 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
576 int tx_nfrags; /* # entries in... */
577 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
578 __u64 *tx_pages; /* rdma phys page addrs */
580 kib_phys_mr_t *pmr; /* MR for physical buffer */
581 kib_fmr_t fmr; /* FMR */
583 int tx_dmadir; /* dma direction */
586 typedef struct kib_connvars
588 /* connection-in-progress variables */
592 typedef struct kib_conn
594 struct kib_sched_info *ibc_sched; /* scheduler information */
595 struct kib_peer *ibc_peer; /* owning peer */
596 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
597 cfs_list_t ibc_list; /* stash on peer's conn list */
598 cfs_list_t ibc_sched_list; /* schedule for attention */
599 __u16 ibc_version; /* version of connection */
600 __u64 ibc_incarnation; /* which instance of the peer */
601 cfs_atomic_t ibc_refcount; /* # users */
602 int ibc_state; /* what's happening */
603 int ibc_nsends_posted; /* # uncompleted sends */
604 int ibc_noops_posted; /* # uncompleted NOOPs */
605 int ibc_credits; /* # credits I have */
606 int ibc_outstanding_credits; /* # credits to return */
607 int ibc_reserved_credits;/* # ACK/DONE msg credits */
608 int ibc_comms_error; /* set on comms error */
609 int ibc_nrx:16; /* receive buffers owned */
610 int ibc_scheduled:1; /* scheduled for attention */
611 int ibc_ready:1; /* CQ callback fired */
612 /* time of last send */
613 unsigned long ibc_last_send;
614 /** link chain for kiblnd_check_conns only */
615 cfs_list_t ibc_connd_list;
616 /** rxs completed before ESTABLISHED */
617 cfs_list_t ibc_early_rxs;
618 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
619 cfs_list_t ibc_tx_noops;
620 cfs_list_t ibc_tx_queue; /* sends that need a credit */
621 cfs_list_t ibc_tx_queue_nocred;/* sends that don't need a credit */
622 cfs_list_t ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
623 cfs_list_t ibc_active_txs; /* active tx awaiting completion */
624 spinlock_t ibc_lock; /* serialise */
625 kib_rx_t *ibc_rxs; /* the rx descs */
626 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
628 struct rdma_cm_id *ibc_cmid; /* CM id */
629 struct ib_cq *ibc_cq; /* completion queue */
631 kib_connvars_t *ibc_connvars; /* in-progress connection state */
634 #define IBLND_CONN_INIT 0 /* being initialised */
635 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
636 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
637 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
638 #define IBLND_CONN_CLOSING 4 /* being closed */
639 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
641 typedef struct kib_peer
643 cfs_list_t ibp_list; /* stash on global peer list */
644 lnet_nid_t ibp_nid; /* who's on the other end(s) */
645 lnet_ni_t *ibp_ni; /* LNet interface */
646 cfs_atomic_t ibp_refcount; /* # users */
647 cfs_list_t ibp_conns; /* all active connections */
648 cfs_list_t ibp_tx_queue; /* msgs waiting for a conn */
649 __u16 ibp_version; /* version of peer */
650 __u64 ibp_incarnation; /* incarnation of peer */
651 int ibp_connecting; /* current active connection attempts */
652 int ibp_accepting; /* current passive connection attempts */
653 int ibp_error; /* errno on closing this peer */
654 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
657 extern kib_data_t kiblnd_data;
659 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
662 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
664 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
665 cfs_atomic_inc(&hdev->ibh_ref);
669 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
671 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
672 if (cfs_atomic_dec_and_test(&hdev->ibh_ref))
673 kiblnd_hdev_destroy(hdev);
677 kiblnd_dev_can_failover(kib_dev_t *dev)
679 if (!cfs_list_empty(&dev->ibd_fail_list)) /* already scheduled */
682 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
685 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
688 return dev->ibd_can_failover;
691 #define kiblnd_conn_addref(conn) \
693 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
694 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
695 cfs_atomic_inc(&(conn)->ibc_refcount); \
698 #define kiblnd_conn_decref(conn) \
700 unsigned long flags; \
702 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
703 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
704 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
705 if (cfs_atomic_dec_and_test(&(conn)->ibc_refcount)) { \
706 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
707 cfs_list_add_tail(&(conn)->ibc_list, \
708 &kiblnd_data.kib_connd_zombies); \
709 cfs_waitq_signal(&kiblnd_data.kib_connd_waitq); \
710 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
714 #define kiblnd_peer_addref(peer) \
716 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
717 (peer), libcfs_nid2str((peer)->ibp_nid), \
718 cfs_atomic_read (&(peer)->ibp_refcount)); \
719 cfs_atomic_inc(&(peer)->ibp_refcount); \
722 #define kiblnd_peer_decref(peer) \
724 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
725 (peer), libcfs_nid2str((peer)->ibp_nid), \
726 cfs_atomic_read (&(peer)->ibp_refcount)); \
727 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
728 if (cfs_atomic_dec_and_test(&(peer)->ibp_refcount)) \
729 kiblnd_destroy_peer(peer); \
732 static inline cfs_list_t *
733 kiblnd_nid2peerlist (lnet_nid_t nid)
736 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
738 return (&kiblnd_data.kib_peers [hash]);
742 kiblnd_peer_active (kib_peer_t *peer)
744 /* Am I in the peer hash table? */
745 return (!cfs_list_empty(&peer->ibp_list));
748 static inline kib_conn_t *
749 kiblnd_get_conn_locked (kib_peer_t *peer)
751 LASSERT (!cfs_list_empty(&peer->ibp_conns));
753 /* just return the first connection */
754 return cfs_list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
758 kiblnd_send_keepalive(kib_conn_t *conn)
760 return (*kiblnd_tunables.kib_keepalive > 0) &&
761 cfs_time_after(jiffies, conn->ibc_last_send +
762 *kiblnd_tunables.kib_keepalive*CFS_HZ);
766 kiblnd_need_noop(kib_conn_t *conn)
768 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
770 if (conn->ibc_outstanding_credits <
771 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
772 !kiblnd_send_keepalive(conn))
773 return 0; /* No need to send NOOP */
775 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
776 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred))
777 return 0; /* NOOP can be piggybacked */
779 /* No tx to piggyback NOOP onto or no credit to send a tx */
780 return (cfs_list_empty(&conn->ibc_tx_queue) ||
781 conn->ibc_credits == 0);
784 if (!cfs_list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
785 !cfs_list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
786 conn->ibc_credits == 0) /* no credit */
789 if (conn->ibc_credits == 1 && /* last credit reserved for */
790 conn->ibc_outstanding_credits == 0) /* giving back credits */
793 /* No tx to piggyback NOOP onto or no credit to send a tx */
794 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
798 kiblnd_abort_receives(kib_conn_t *conn)
800 ib_modify_qp(conn->ibc_cmid->qp,
801 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
804 static inline const char *
805 kiblnd_queue2str (kib_conn_t *conn, cfs_list_t *q)
807 if (q == &conn->ibc_tx_queue)
810 if (q == &conn->ibc_tx_queue_rsrvd)
811 return "tx_queue_rsrvd";
813 if (q == &conn->ibc_tx_queue_nocred)
814 return "tx_queue_nocred";
816 if (q == &conn->ibc_active_txs)
823 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
824 * lowest bits of the work request id to stash the work item type. */
826 #define IBLND_WID_TX 0
827 #define IBLND_WID_RDMA 1
828 #define IBLND_WID_RX 2
829 #define IBLND_WID_MASK 3UL
832 kiblnd_ptr2wreqid (void *ptr, int type)
834 unsigned long lptr = (unsigned long)ptr;
836 LASSERT ((lptr & IBLND_WID_MASK) == 0);
837 LASSERT ((type & ~IBLND_WID_MASK) == 0);
838 return (__u64)(lptr | type);
842 kiblnd_wreqid2ptr (__u64 wreqid)
844 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
848 kiblnd_wreqid2type (__u64 wreqid)
850 return (wreqid & IBLND_WID_MASK);
854 kiblnd_set_conn_state (kib_conn_t *conn, int state)
856 conn->ibc_state = state;
861 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
863 msg->ibm_type = type;
864 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
868 kiblnd_rd_size (kib_rdma_desc_t *rd)
873 for (i = size = 0; i < rd->rd_nfrags; i++)
874 size += rd->rd_frags[i].rf_nob;
880 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
882 return rd->rd_frags[index].rf_addr;
886 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
888 return rd->rd_frags[index].rf_nob;
892 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
898 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
900 if (nob < rd->rd_frags[index].rf_nob) {
901 rd->rd_frags[index].rf_addr += nob;
902 rd->rd_frags[index].rf_nob -= nob;
911 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
913 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
914 msgtype == IBLND_MSG_PUT_ACK);
916 return msgtype == IBLND_MSG_GET_REQ ?
917 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
918 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
921 #ifdef HAVE_OFED_IB_DMA_MAP
924 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
926 return ib_dma_mapping_error(dev, dma_addr);
929 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
930 void *msg, size_t size,
931 enum dma_data_direction direction)
933 return ib_dma_map_single(dev, msg, size, direction);
936 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
937 __u64 addr, size_t size,
938 enum dma_data_direction direction)
940 ib_dma_unmap_single(dev, addr, size, direction);
943 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
944 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
946 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
947 struct scatterlist *sg, int nents,
948 enum dma_data_direction direction)
950 return ib_dma_map_sg(dev, sg, nents, direction);
953 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
954 struct scatterlist *sg, int nents,
955 enum dma_data_direction direction)
957 ib_dma_unmap_sg(dev, sg, nents, direction);
960 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
961 struct scatterlist *sg)
963 return ib_sg_dma_address(dev, sg);
966 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
967 struct scatterlist *sg)
969 return ib_sg_dma_len(dev, sg);
972 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
973 * right because OFED1.2 defines it as const, to use it we have to add
974 * (void *) cast to overcome "const" */
976 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
977 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
982 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
984 return dma_mapping_error(dma_addr);
987 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
988 void *msg, size_t size,
989 enum dma_data_direction direction)
991 return dma_map_single(dev->dma_device, msg, size, direction);
994 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
995 dma_addr_t addr, size_t size,
996 enum dma_data_direction direction)
998 dma_unmap_single(dev->dma_device, addr, size, direction);
1001 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
1002 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
1004 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1005 struct scatterlist *sg, int nents,
1006 enum dma_data_direction direction)
1008 return dma_map_sg(dev->dma_device, sg, nents, direction);
1011 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1012 struct scatterlist *sg, int nents,
1013 enum dma_data_direction direction)
1015 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
1019 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
1020 struct scatterlist *sg)
1022 return sg_dma_address(sg);
1026 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1027 struct scatterlist *sg)
1029 return sg_dma_len(sg);
1032 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
1033 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
1037 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1038 kib_rdma_desc_t *rd);
1039 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
1040 __u64 addr, __u64 size);
1041 void kiblnd_map_rx_descs(kib_conn_t *conn);
1042 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1043 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
1044 kib_rdma_desc_t *rd, int nfrags);
1045 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
1046 void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node);
1047 cfs_list_t *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1049 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1050 int npages, __u64 iov, kib_fmr_t *fmr);
1051 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1053 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1054 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1055 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1057 int kiblnd_startup (lnet_ni_t *ni);
1058 void kiblnd_shutdown (lnet_ni_t *ni);
1059 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
1060 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1062 int kiblnd_tunables_init(void);
1063 void kiblnd_tunables_fini(void);
1065 int kiblnd_connd (void *arg);
1066 int kiblnd_scheduler(void *arg);
1067 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
1068 int kiblnd_failover_thread (void *arg);
1070 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1071 void kiblnd_free_pages (kib_pages_t *p);
1073 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1074 struct rdma_cm_event *event);
1075 int kiblnd_translate_mtu(int value);
1077 int kiblnd_dev_failover(kib_dev_t *dev);
1078 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1079 void kiblnd_destroy_peer (kib_peer_t *peer);
1080 void kiblnd_destroy_dev (kib_dev_t *dev);
1081 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1082 void kiblnd_peer_alive (kib_peer_t *peer);
1083 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1084 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1085 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1086 int version, __u64 incarnation);
1087 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1089 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1090 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1091 int state, int version);
1092 void kiblnd_destroy_conn (kib_conn_t *conn);
1093 void kiblnd_close_conn (kib_conn_t *conn, int error);
1094 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1096 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1097 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1099 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1100 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1101 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1102 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1103 void kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist,
1105 void kiblnd_check_sends (kib_conn_t *conn);
1107 void kiblnd_qp_event(struct ib_event *event, void *arg);
1108 void kiblnd_cq_event(struct ib_event *event, void *arg);
1109 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1111 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1112 int credits, lnet_nid_t dstnid, __u64 dststamp);
1113 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1114 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1116 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1117 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1118 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1119 unsigned int offset, unsigned int mlen, unsigned int rlen);