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/system.h>
51 #include <asm/uaccess.h>
54 #include <linux/init.h>
56 #include <linux/file.h>
57 #include <linux/stat.h>
58 #include <linux/list.h>
59 #include <linux/kmod.h>
60 #include <linux/sysctl.h>
61 #include <linux/pci.h>
66 #define DEBUG_SUBSYSTEM S_LND
68 #include <libcfs/libcfs.h>
69 #include <lnet/lnet.h>
70 #include <lnet/lib-lnet.h>
71 #include <lnet/lnet-sysctl.h>
73 #include <rdma/rdma_cm.h>
74 #include <rdma/ib_cm.h>
75 #include <rdma/ib_verbs.h>
76 #include <rdma/ib_fmr_pool.h>
78 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
79 /* # scheduler loops before reschedule */
80 #define IBLND_RESCHED 100
82 #define IBLND_N_SCHED 2
83 #define IBLND_N_SCHED_HIGH 4
87 int *kib_dev_failover; /* HCA failover */
88 unsigned int *kib_service; /* IB service number */
89 int *kib_min_reconnect_interval; /* first failed connection retry... */
90 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
91 int *kib_cksum; /* checksum kib_msg_t? */
92 int *kib_timeout; /* comms timeout (seconds) */
93 int *kib_keepalive; /* keepalive timeout (seconds) */
94 int *kib_ntx; /* # tx descs */
95 int *kib_credits; /* # concurrent sends */
96 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
97 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
98 int *kib_peercredits_hiw; /* # when eagerly to return credits */
99 int *kib_peertimeout; /* seconds to consider peer dead */
100 char **kib_default_ipif; /* default IPoIB interface */
101 int *kib_retry_count;
102 int *kib_rnr_retry_count;
103 int *kib_concurrent_sends; /* send work queue sizing */
104 int *kib_ib_mtu; /* IB MTU */
105 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
106 * than this value, 0 disable map-on-demand */
107 int *kib_pmr_pool_size; /* # physical MR in pool */
108 int *kib_fmr_pool_size; /* # FMRs in pool */
109 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
110 int *kib_fmr_cache; /* enable FMR pool cache? */
111 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
112 cfs_sysctl_table_header_t *kib_sysctl; /* sysctl interface */
114 int *kib_require_priv_port;/* accept only privileged ports */
115 int *kib_use_priv_port; /* use privileged port for active connect */
116 /* # threads on each CPT */
120 extern kib_tunables_t kiblnd_tunables;
122 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
123 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
125 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
126 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
128 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
129 IBLND_MSG_QUEUE_SIZE_V1 : \
130 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
131 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
132 IBLND_CREDIT_HIGHWATER_V1 : \
133 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
135 #ifdef HAVE_RDMA_CREATE_ID_4ARG
136 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
138 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
142 kiblnd_concurrent_sends_v1(void)
144 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
145 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
147 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
148 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
150 return *kiblnd_tunables.kib_concurrent_sends;
153 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
154 kiblnd_concurrent_sends_v1() : \
155 *kiblnd_tunables.kib_concurrent_sends)
156 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
157 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
158 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
160 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
161 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
162 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
163 *kiblnd_tunables.kib_map_on_demand : \
164 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
165 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
166 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
168 /************************/
169 /* derived constants... */
170 /* Pools (shared by connections on each CPT) */
171 /* These pools can grow at runtime, so don't need give a very large value */
172 #define IBLND_TX_POOL 256
173 #define IBLND_PMR_POOL 256
174 #define IBLND_FMR_POOL 256
175 #define IBLND_FMR_POOL_FLUSH 192
177 /* TX messages (shared by all connections) */
178 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
180 /* RX messages (per connection) */
181 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
182 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
183 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
185 /* WRs and CQEs (per connection) */
186 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
187 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
188 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
192 /* o2iblnd can run over aliased interface */
194 #define KIB_IFNAME_SIZE IFALIASZ
196 #define KIB_IFNAME_SIZE 256
201 cfs_list_t ibd_list; /* chain on kib_devs */
202 cfs_list_t ibd_fail_list; /* chain on kib_failed_devs */
203 __u32 ibd_ifip; /* IPoIB interface IP */
204 /** IPoIB interface name */
205 char ibd_ifname[KIB_IFNAME_SIZE];
206 int ibd_nnets; /* # nets extant */
208 cfs_time_t ibd_next_failover;
209 int ibd_failed_failover; /* # failover failures */
210 unsigned int ibd_failover; /* failover in progress */
211 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
213 struct kib_hca_dev *ibd_hdev;
216 typedef struct kib_hca_dev
218 struct rdma_cm_id *ibh_cmid; /* listener cmid */
219 struct ib_device *ibh_ibdev; /* IB device */
220 int ibh_page_shift; /* page shift of current HCA */
221 int ibh_page_size; /* page size of current HCA */
222 __u64 ibh_page_mask; /* page mask of current HCA */
223 int ibh_mr_shift; /* bits shift of max MR size */
224 __u64 ibh_mr_size; /* size of MR */
225 int ibh_nmrs; /* # of global MRs */
226 struct ib_mr **ibh_mrs; /* global MR */
227 struct ib_pd *ibh_pd; /* PD */
228 kib_dev_t *ibh_dev; /* owner */
229 cfs_atomic_t ibh_ref; /* refcount */
232 /** # of seconds to keep pool alive */
233 #define IBLND_POOL_DEADLINE 300
234 /** # of seconds to retry if allocation failed */
235 #define IBLND_POOL_RETRY 1
239 int ibp_npages; /* # pages */
240 struct page *ibp_pages[0]; /* page array */
246 cfs_list_t pmr_list; /* chain node */
247 struct ib_phys_buf *pmr_ipb; /* physical buffer */
248 struct ib_mr *pmr_mr; /* IB MR */
249 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
250 __u64 pmr_iova; /* Virtual I/O address */
251 int pmr_refcount; /* reference count */
257 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
258 int inc, struct kib_pool **pp_po);
259 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
260 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, cfs_list_t *node);
261 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, cfs_list_t *node);
265 #define IBLND_POOL_NAME_LEN 32
267 typedef struct kib_poolset
269 cfs_spinlock_t ps_lock; /* serialize */
270 struct kib_net *ps_net; /* network it belongs to */
271 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
272 cfs_list_t ps_pool_list; /* list of pools */
273 cfs_list_t ps_failed_pool_list; /* failed pool list */
274 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
275 int ps_increasing; /* is allocating new pool */
276 int ps_pool_size; /* new pool size */
277 int ps_cpt; /* CPT id */
279 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
280 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
281 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
282 kib_ps_node_fini_t ps_node_fini; /* finalize node */
285 typedef struct kib_pool
287 cfs_list_t po_list; /* chain on pool list */
288 cfs_list_t po_free_list; /* pre-allocated node */
289 kib_poolset_t *po_owner; /* pool_set of this pool */
290 cfs_time_t po_deadline; /* deadline of this pool */
291 int po_allocated; /* # of elements in use */
292 int po_failed; /* pool is created on failed HCA */
293 int po_size; /* # of pre-allocated elements */
297 kib_poolset_t tps_poolset; /* pool-set */
298 __u64 tps_next_tx_cookie; /* cookie of TX */
302 kib_pool_t tpo_pool; /* pool */
303 struct kib_hca_dev *tpo_hdev; /* device for this pool */
304 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
305 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
309 kib_poolset_t pps_poolset; /* pool-set */
312 typedef struct kib_pmr_pool {
313 struct kib_hca_dev *ppo_hdev; /* device for this pool */
314 kib_pool_t ppo_pool; /* pool */
319 cfs_spinlock_t fps_lock; /* serialize */
320 struct kib_net *fps_net; /* IB network */
321 cfs_list_t fps_pool_list; /* FMR pool list */
322 cfs_list_t fps_failed_pool_list; /* FMR pool list */
323 __u64 fps_version; /* validity stamp */
324 int fps_cpt; /* CPT id */
326 int fps_flush_trigger;
327 /* is allocating new pool */
329 /* time stamp for retry if failed to allocate */
330 cfs_time_t fps_next_retry;
335 cfs_list_t fpo_list; /* chain on pool list */
336 struct kib_hca_dev *fpo_hdev; /* device for this pool */
337 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
338 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
339 cfs_time_t fpo_deadline; /* deadline of this pool */
340 int fpo_failed; /* fmr pool is failed */
341 int fpo_map_count; /* # of mapped FMR */
345 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
346 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
349 typedef struct kib_net
351 cfs_list_t ibn_list; /* chain on kib_dev_t::ibd_nets */
352 __u64 ibn_incarnation; /* my epoch */
353 int ibn_init; /* initialisation state */
354 int ibn_shutdown; /* shutting down? */
356 cfs_atomic_t ibn_npeers; /* # peers extant */
357 cfs_atomic_t ibn_nconns; /* # connections extant */
359 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
360 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
361 kib_pmr_poolset_t **ibn_pmr_ps; /* pmr pool-set */
363 kib_dev_t *ibn_dev; /* underlying IB device */
366 #define KIB_THREAD_SHIFT 16
367 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
368 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
369 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
371 struct kib_sched_info {
373 cfs_spinlock_t ibs_lock;
374 /* schedulers sleep here */
375 cfs_waitq_t ibs_waitq;
376 /* conns to check for rx completions */
377 cfs_list_t ibs_conns;
378 /* number of scheduler threads */
380 /* max allowed scheduler threads */
381 int ibs_nthreads_max;
382 int ibs_cpt; /* CPT id */
387 int kib_init; /* initialisation state */
388 int kib_shutdown; /* shut down? */
389 cfs_list_t kib_devs; /* IB devices extant */
390 /* list head of failed devices */
391 cfs_list_t kib_failed_devs;
392 /* schedulers sleep here */
393 cfs_waitq_t kib_failover_waitq;
394 cfs_atomic_t kib_nthreads; /* # live threads */
395 /* stabilize net/dev/peer/conn ops */
396 cfs_rwlock_t kib_global_lock;
397 /* hash table of all my known peers */
398 cfs_list_t *kib_peers;
399 /* size of kib_peers */
400 int kib_peer_hash_size;
401 /* the connd task (serialisation assertions) */
403 /* connections to setup/teardown */
404 cfs_list_t kib_connd_conns;
405 /* connections with zero refcount */
406 cfs_list_t kib_connd_zombies;
407 /* connection daemon sleeps here */
408 cfs_waitq_t kib_connd_waitq;
409 cfs_spinlock_t kib_connd_lock; /* serialise */
410 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
411 /* percpt data for schedulers */
412 struct kib_sched_info **kib_scheds;
415 #define IBLND_INIT_NOTHING 0
416 #define IBLND_INIT_DATA 1
417 #define IBLND_INIT_ALL 2
419 /************************************************************************
420 * IB Wire message format.
421 * These are sent in sender's byte order (i.e. receiver flips).
424 typedef struct kib_connparams
426 __u16 ibcp_queue_depth;
427 __u16 ibcp_max_frags;
428 __u32 ibcp_max_msg_size;
429 } WIRE_ATTR kib_connparams_t;
433 lnet_hdr_t ibim_hdr; /* portals header */
434 char ibim_payload[0]; /* piggy-backed payload */
435 } WIRE_ATTR kib_immediate_msg_t;
439 __u32 rf_nob; /* # bytes this frag */
440 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
441 } WIRE_ATTR kib_rdma_frag_t;
445 __u32 rd_key; /* local/remote key */
446 __u32 rd_nfrags; /* # fragments */
447 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
448 } WIRE_ATTR kib_rdma_desc_t;
452 lnet_hdr_t ibprm_hdr; /* portals header */
453 __u64 ibprm_cookie; /* opaque completion cookie */
454 } WIRE_ATTR kib_putreq_msg_t;
458 __u64 ibpam_src_cookie; /* reflected completion cookie */
459 __u64 ibpam_dst_cookie; /* opaque completion cookie */
460 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
461 } WIRE_ATTR kib_putack_msg_t;
465 lnet_hdr_t ibgm_hdr; /* portals header */
466 __u64 ibgm_cookie; /* opaque completion cookie */
467 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
468 } WIRE_ATTR kib_get_msg_t;
472 __u64 ibcm_cookie; /* opaque completion cookie */
473 __s32 ibcm_status; /* < 0 failure: >= 0 length */
474 } WIRE_ATTR kib_completion_msg_t;
478 /* First 2 fields fixed FOR ALL TIME */
479 __u32 ibm_magic; /* I'm an ibnal message */
480 __u16 ibm_version; /* this is my version number */
482 __u8 ibm_type; /* msg type */
483 __u8 ibm_credits; /* returned credits */
484 __u32 ibm_nob; /* # bytes in whole message */
485 __u32 ibm_cksum; /* checksum (0 == no checksum) */
486 __u64 ibm_srcnid; /* sender's NID */
487 __u64 ibm_srcstamp; /* sender's incarnation */
488 __u64 ibm_dstnid; /* destination's NID */
489 __u64 ibm_dststamp; /* destination's incarnation */
492 kib_connparams_t connparams;
493 kib_immediate_msg_t immediate;
494 kib_putreq_msg_t putreq;
495 kib_putack_msg_t putack;
497 kib_completion_msg_t completion;
499 } WIRE_ATTR kib_msg_t;
501 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
503 #define IBLND_MSG_VERSION_1 0x11
504 #define IBLND_MSG_VERSION_2 0x12
505 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
507 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
508 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
509 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
510 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
511 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
512 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
513 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
514 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
515 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
516 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
519 __u32 ibr_magic; /* sender's magic */
520 __u16 ibr_version; /* sender's version */
521 __u8 ibr_why; /* reject reason */
522 __u8 ibr_padding; /* padding */
523 __u64 ibr_incarnation; /* incarnation of peer */
524 kib_connparams_t ibr_cp; /* connection parameters */
525 } WIRE_ATTR kib_rej_t;
527 /* connection rejection reasons */
528 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
529 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
530 #define IBLND_REJECT_FATAL 3 /* Anything else */
532 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
533 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
535 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
536 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
538 /***********************************************************************/
540 typedef struct kib_rx /* receive message */
542 cfs_list_t rx_list; /* queue for attention */
543 struct kib_conn *rx_conn; /* owning conn */
544 int rx_nob; /* # bytes received (-1 while posted) */
545 enum ib_wc_status rx_status; /* completion status */
546 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
547 __u64 rx_msgaddr; /* message buffer (I/O addr) */
548 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
549 struct ib_recv_wr rx_wrq; /* receive work item... */
550 struct ib_sge rx_sge; /* ...and its memory */
553 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
554 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
555 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
556 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
558 typedef struct kib_tx /* transmit message */
560 cfs_list_t tx_list; /* queue on idle_txs ibc_tx_queue etc. */
561 kib_tx_pool_t *tx_pool; /* pool I'm from */
562 struct kib_conn *tx_conn; /* owning conn */
563 short tx_sending; /* # tx callbacks outstanding */
564 short tx_queued; /* queued for sending */
565 short tx_waiting; /* waiting for peer */
566 int tx_status; /* LNET completion status */
567 unsigned long tx_deadline; /* completion deadline */
568 __u64 tx_cookie; /* completion cookie */
569 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
570 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
571 __u64 tx_msgaddr; /* message buffer (I/O addr) */
572 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
573 int tx_nwrq; /* # send work items */
574 struct ib_send_wr *tx_wrq; /* send work items... */
575 struct ib_sge *tx_sge; /* ...and their memory */
576 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
577 int tx_nfrags; /* # entries in... */
578 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
579 __u64 *tx_pages; /* rdma phys page addrs */
581 kib_phys_mr_t *pmr; /* MR for physical buffer */
582 kib_fmr_t fmr; /* FMR */
584 int tx_dmadir; /* dma direction */
587 typedef struct kib_connvars
589 /* connection-in-progress variables */
593 typedef struct kib_conn
595 struct kib_sched_info *ibc_sched; /* scheduler information */
596 struct kib_peer *ibc_peer; /* owning peer */
597 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
598 cfs_list_t ibc_list; /* stash on peer's conn list */
599 cfs_list_t ibc_sched_list; /* schedule for attention */
600 __u16 ibc_version; /* version of connection */
601 __u64 ibc_incarnation; /* which instance of the peer */
602 cfs_atomic_t ibc_refcount; /* # users */
603 int ibc_state; /* what's happening */
604 int ibc_nsends_posted; /* # uncompleted sends */
605 int ibc_noops_posted; /* # uncompleted NOOPs */
606 int ibc_credits; /* # credits I have */
607 int ibc_outstanding_credits; /* # credits to return */
608 int ibc_reserved_credits;/* # ACK/DONE msg credits */
609 int ibc_comms_error; /* set on comms error */
610 int ibc_nrx:16; /* receive buffers owned */
611 int ibc_scheduled:1; /* scheduled for attention */
612 int ibc_ready:1; /* CQ callback fired */
613 /* time of last send */
614 unsigned long ibc_last_send;
615 /** link chain for kiblnd_check_conns only */
616 cfs_list_t ibc_connd_list;
617 /** rxs completed before ESTABLISHED */
618 cfs_list_t ibc_early_rxs;
619 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
620 cfs_list_t ibc_tx_noops;
621 cfs_list_t ibc_tx_queue; /* sends that need a credit */
622 cfs_list_t ibc_tx_queue_nocred;/* sends that don't need a credit */
623 cfs_list_t ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
624 cfs_list_t ibc_active_txs; /* active tx awaiting completion */
625 cfs_spinlock_t ibc_lock; /* serialise */
626 kib_rx_t *ibc_rxs; /* the rx descs */
627 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
629 struct rdma_cm_id *ibc_cmid; /* CM id */
630 struct ib_cq *ibc_cq; /* completion queue */
632 kib_connvars_t *ibc_connvars; /* in-progress connection state */
635 #define IBLND_CONN_INIT 0 /* being initialised */
636 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
637 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
638 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
639 #define IBLND_CONN_CLOSING 4 /* being closed */
640 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
642 typedef struct kib_peer
644 cfs_list_t ibp_list; /* stash on global peer list */
645 lnet_nid_t ibp_nid; /* who's on the other end(s) */
646 lnet_ni_t *ibp_ni; /* LNet interface */
647 cfs_atomic_t ibp_refcount; /* # users */
648 cfs_list_t ibp_conns; /* all active connections */
649 cfs_list_t ibp_tx_queue; /* msgs waiting for a conn */
650 __u16 ibp_version; /* version of peer */
651 __u64 ibp_incarnation; /* incarnation of peer */
652 int ibp_connecting; /* current active connection attempts */
653 int ibp_accepting; /* current passive connection attempts */
654 int ibp_error; /* errno on closing this peer */
655 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
658 extern kib_data_t kiblnd_data;
660 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
663 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
665 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
666 cfs_atomic_inc(&hdev->ibh_ref);
670 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
672 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
673 if (cfs_atomic_dec_and_test(&hdev->ibh_ref))
674 kiblnd_hdev_destroy(hdev);
678 kiblnd_dev_can_failover(kib_dev_t *dev)
680 if (!cfs_list_empty(&dev->ibd_fail_list)) /* already scheduled */
683 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
686 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
689 return dev->ibd_can_failover;
692 #define kiblnd_conn_addref(conn) \
694 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
695 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
696 cfs_atomic_inc(&(conn)->ibc_refcount); \
699 #define kiblnd_conn_decref(conn) \
701 unsigned long flags; \
703 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
704 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
705 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
706 if (cfs_atomic_dec_and_test(&(conn)->ibc_refcount)) { \
707 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
708 cfs_list_add_tail(&(conn)->ibc_list, \
709 &kiblnd_data.kib_connd_zombies); \
710 cfs_waitq_signal(&kiblnd_data.kib_connd_waitq); \
711 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
715 #define kiblnd_peer_addref(peer) \
717 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
718 (peer), libcfs_nid2str((peer)->ibp_nid), \
719 cfs_atomic_read (&(peer)->ibp_refcount)); \
720 cfs_atomic_inc(&(peer)->ibp_refcount); \
723 #define kiblnd_peer_decref(peer) \
725 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
726 (peer), libcfs_nid2str((peer)->ibp_nid), \
727 cfs_atomic_read (&(peer)->ibp_refcount)); \
728 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
729 if (cfs_atomic_dec_and_test(&(peer)->ibp_refcount)) \
730 kiblnd_destroy_peer(peer); \
733 static inline cfs_list_t *
734 kiblnd_nid2peerlist (lnet_nid_t nid)
737 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
739 return (&kiblnd_data.kib_peers [hash]);
743 kiblnd_peer_active (kib_peer_t *peer)
745 /* Am I in the peer hash table? */
746 return (!cfs_list_empty(&peer->ibp_list));
749 static inline kib_conn_t *
750 kiblnd_get_conn_locked (kib_peer_t *peer)
752 LASSERT (!cfs_list_empty(&peer->ibp_conns));
754 /* just return the first connection */
755 return cfs_list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
759 kiblnd_send_keepalive(kib_conn_t *conn)
761 return (*kiblnd_tunables.kib_keepalive > 0) &&
762 cfs_time_after(jiffies, conn->ibc_last_send +
763 *kiblnd_tunables.kib_keepalive*CFS_HZ);
767 kiblnd_need_noop(kib_conn_t *conn)
769 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
771 if (conn->ibc_outstanding_credits <
772 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
773 !kiblnd_send_keepalive(conn))
774 return 0; /* No need to send NOOP */
776 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
777 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred))
778 return 0; /* NOOP can be piggybacked */
780 /* No tx to piggyback NOOP onto or no credit to send a tx */
781 return (cfs_list_empty(&conn->ibc_tx_queue) ||
782 conn->ibc_credits == 0);
785 if (!cfs_list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
786 !cfs_list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
787 conn->ibc_credits == 0) /* no credit */
790 if (conn->ibc_credits == 1 && /* last credit reserved for */
791 conn->ibc_outstanding_credits == 0) /* giving back credits */
794 /* No tx to piggyback NOOP onto or no credit to send a tx */
795 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
799 kiblnd_abort_receives(kib_conn_t *conn)
801 ib_modify_qp(conn->ibc_cmid->qp,
802 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
805 static inline const char *
806 kiblnd_queue2str (kib_conn_t *conn, cfs_list_t *q)
808 if (q == &conn->ibc_tx_queue)
811 if (q == &conn->ibc_tx_queue_rsrvd)
812 return "tx_queue_rsrvd";
814 if (q == &conn->ibc_tx_queue_nocred)
815 return "tx_queue_nocred";
817 if (q == &conn->ibc_active_txs)
824 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
825 * lowest bits of the work request id to stash the work item type. */
827 #define IBLND_WID_TX 0
828 #define IBLND_WID_RDMA 1
829 #define IBLND_WID_RX 2
830 #define IBLND_WID_MASK 3UL
833 kiblnd_ptr2wreqid (void *ptr, int type)
835 unsigned long lptr = (unsigned long)ptr;
837 LASSERT ((lptr & IBLND_WID_MASK) == 0);
838 LASSERT ((type & ~IBLND_WID_MASK) == 0);
839 return (__u64)(lptr | type);
843 kiblnd_wreqid2ptr (__u64 wreqid)
845 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
849 kiblnd_wreqid2type (__u64 wreqid)
851 return (wreqid & IBLND_WID_MASK);
855 kiblnd_set_conn_state (kib_conn_t *conn, int state)
857 conn->ibc_state = state;
862 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
864 msg->ibm_type = type;
865 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
869 kiblnd_rd_size (kib_rdma_desc_t *rd)
874 for (i = size = 0; i < rd->rd_nfrags; i++)
875 size += rd->rd_frags[i].rf_nob;
881 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
883 return rd->rd_frags[index].rf_addr;
887 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
889 return rd->rd_frags[index].rf_nob;
893 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
899 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
901 if (nob < rd->rd_frags[index].rf_nob) {
902 rd->rd_frags[index].rf_addr += nob;
903 rd->rd_frags[index].rf_nob -= nob;
912 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
914 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
915 msgtype == IBLND_MSG_PUT_ACK);
917 return msgtype == IBLND_MSG_GET_REQ ?
918 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
919 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
922 #ifdef HAVE_OFED_IB_DMA_MAP
925 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
927 return ib_dma_mapping_error(dev, dma_addr);
930 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
931 void *msg, size_t size,
932 enum dma_data_direction direction)
934 return ib_dma_map_single(dev, msg, size, direction);
937 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
938 __u64 addr, size_t size,
939 enum dma_data_direction direction)
941 ib_dma_unmap_single(dev, addr, size, direction);
944 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
945 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
947 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
948 struct scatterlist *sg, int nents,
949 enum dma_data_direction direction)
951 return ib_dma_map_sg(dev, sg, nents, direction);
954 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
955 struct scatterlist *sg, int nents,
956 enum dma_data_direction direction)
958 ib_dma_unmap_sg(dev, sg, nents, direction);
961 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
962 struct scatterlist *sg)
964 return ib_sg_dma_address(dev, sg);
967 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
968 struct scatterlist *sg)
970 return ib_sg_dma_len(dev, sg);
973 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
974 * right because OFED1.2 defines it as const, to use it we have to add
975 * (void *) cast to overcome "const" */
977 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
978 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
983 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
985 return dma_mapping_error(dma_addr);
988 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
989 void *msg, size_t size,
990 enum dma_data_direction direction)
992 return dma_map_single(dev->dma_device, msg, size, direction);
995 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
996 dma_addr_t addr, size_t size,
997 enum dma_data_direction direction)
999 dma_unmap_single(dev->dma_device, addr, size, direction);
1002 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
1003 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
1005 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1006 struct scatterlist *sg, int nents,
1007 enum dma_data_direction direction)
1009 return dma_map_sg(dev->dma_device, sg, nents, direction);
1012 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1013 struct scatterlist *sg, int nents,
1014 enum dma_data_direction direction)
1016 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
1020 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
1021 struct scatterlist *sg)
1023 return sg_dma_address(sg);
1027 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1028 struct scatterlist *sg)
1030 return sg_dma_len(sg);
1033 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
1034 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
1038 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1039 kib_rdma_desc_t *rd);
1040 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
1041 __u64 addr, __u64 size);
1042 void kiblnd_map_rx_descs(kib_conn_t *conn);
1043 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1044 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
1045 kib_rdma_desc_t *rd, int nfrags);
1046 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
1047 void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node);
1048 cfs_list_t *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1050 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1051 int npages, __u64 iov, kib_fmr_t *fmr);
1052 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1054 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1055 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1056 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1058 int kiblnd_startup (lnet_ni_t *ni);
1059 void kiblnd_shutdown (lnet_ni_t *ni);
1060 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
1061 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1063 int kiblnd_tunables_init(void);
1064 void kiblnd_tunables_fini(void);
1066 int kiblnd_connd (void *arg);
1067 int kiblnd_scheduler(void *arg);
1068 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
1069 int kiblnd_failover_thread (void *arg);
1071 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1072 void kiblnd_free_pages (kib_pages_t *p);
1074 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1075 struct rdma_cm_event *event);
1076 int kiblnd_translate_mtu(int value);
1078 int kiblnd_dev_failover(kib_dev_t *dev);
1079 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1080 void kiblnd_destroy_peer (kib_peer_t *peer);
1081 void kiblnd_destroy_dev (kib_dev_t *dev);
1082 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1083 void kiblnd_peer_alive (kib_peer_t *peer);
1084 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1085 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1086 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1087 int version, __u64 incarnation);
1088 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1090 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1091 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1092 int state, int version);
1093 void kiblnd_destroy_conn (kib_conn_t *conn);
1094 void kiblnd_close_conn (kib_conn_t *conn, int error);
1095 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1097 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1098 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1100 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1101 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1102 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1103 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1104 void kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist,
1106 void kiblnd_check_sends (kib_conn_t *conn);
1108 void kiblnd_qp_event(struct ib_event *event, void *arg);
1109 void kiblnd_cq_event(struct ib_event *event, void *arg);
1110 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1112 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1113 int credits, lnet_nid_t dstnid, __u64 dststamp);
1114 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1115 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1117 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1118 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1119 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1120 unsigned int offset, unsigned int mlen, unsigned int rlen);