1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * lnet/klnds/o2iblnd/o2iblnd.h
40 * Author: Eric Barton <eric@bartonsoftware.com>
44 # define EXPORT_SYMTAB
46 #ifndef AUTOCONF_INCLUDED
47 #include <linux/config.h>
49 #include <linux/module.h>
50 #include <linux/kernel.h>
52 #include <linux/string.h>
53 #include <linux/stat.h>
54 #include <linux/errno.h>
55 #include <linux/smp_lock.h>
56 #include <linux/unistd.h>
57 #include <linux/uio.h>
59 #include <asm/system.h>
60 #include <asm/uaccess.h>
63 #include <linux/init.h>
65 #include <linux/file.h>
66 #include <linux/stat.h>
67 #include <linux/list.h>
68 #include <linux/kmod.h>
69 #include <linux/sysctl.h>
70 #include <linux/pci.h>
75 #define DEBUG_SUBSYSTEM S_LND
77 #include <libcfs/libcfs.h>
78 #include <lnet/lnet.h>
79 #include <lnet/lib-lnet.h>
80 #include <lnet/lnet-sysctl.h>
82 #include <rdma/rdma_cm.h>
83 #include <rdma/ib_cm.h>
84 #include <rdma/ib_verbs.h>
85 #include <rdma/ib_fmr_pool.h>
87 /* tunables fixed at compile time */
89 # define IBLND_N_SCHED cfs_num_online_cpus() /* # schedulers */
91 # define IBLND_N_SCHED 1 /* # schedulers */
94 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
95 #define IBLND_RESCHED 100 /* # scheduler loops before reschedule */
99 int *kib_dev_failover; /* HCA failover */
100 unsigned int *kib_service; /* IB service number */
101 int *kib_min_reconnect_interval; /* first failed connection retry... */
102 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
103 int *kib_cksum; /* checksum kib_msg_t? */
104 int *kib_timeout; /* comms timeout (seconds) */
105 int *kib_keepalive; /* keepalive timeout (seconds) */
106 int *kib_ntx; /* # tx descs */
107 int *kib_credits; /* # concurrent sends */
108 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
109 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
110 int *kib_peercredits_hiw; /* # when eagerly to return credits */
111 int *kib_peertimeout; /* seconds to consider peer dead */
112 char **kib_default_ipif; /* default IPoIB interface */
113 int *kib_retry_count;
114 int *kib_rnr_retry_count;
115 int *kib_concurrent_sends; /* send work queue sizing */
116 int *kib_ib_mtu; /* IB MTU */
117 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
118 * than this value, 0 disable map-on-demand */
119 int *kib_pmr_pool_size; /* # physical MR in pool */
120 int *kib_fmr_pool_size; /* # FMRs in pool */
121 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
122 int *kib_fmr_cache; /* enable FMR pool cache? */
123 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
124 cfs_sysctl_table_header_t *kib_sysctl; /* sysctl interface */
126 int *kib_require_priv_port;/* accept only privileged ports */
127 int *kib_use_priv_port; /* use privileged port for active connect */
130 extern kib_tunables_t kiblnd_tunables;
132 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
133 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
135 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
136 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
138 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
139 IBLND_MSG_QUEUE_SIZE_V1 : \
140 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
141 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
142 IBLND_CREDIT_HIGHWATER_V1 : \
143 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
145 #ifdef HAVE_RDMA_CREATE_ID_4ARG
146 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
148 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
152 kiblnd_concurrent_sends_v1(void)
154 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
155 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
157 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
158 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
160 return *kiblnd_tunables.kib_concurrent_sends;
163 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
164 kiblnd_concurrent_sends_v1() : \
165 *kiblnd_tunables.kib_concurrent_sends)
166 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
167 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
168 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
170 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
171 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
172 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
173 *kiblnd_tunables.kib_map_on_demand : \
174 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
175 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
176 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
178 /************************/
179 /* derived constants... */
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, int inc, struct kib_pool **pp_po);
262 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
263 typedef void (*kib_ps_node_init_t)(struct kib_pool *po,
265 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po,
270 #define IBLND_POOL_NAME_LEN 32
272 typedef struct kib_poolset
274 cfs_spinlock_t ps_lock; /* serialize */
275 struct kib_net *ps_net; /* network it belongs to */
276 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
277 cfs_list_t ps_pool_list; /* list of pools */
278 cfs_list_t ps_failed_pool_list; /* failed pool list */
279 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
280 int ps_increasing; /* is allocating new pool */
281 int ps_pool_size; /* new pool size */
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_increasing; /* is allocating new pool */
329 cfs_time_t fps_next_retry; /* time stamp for retry if failed to allocate */
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? */
354 unsigned int ibn_with_fmr:1; /* FMR? */
355 unsigned int ibn_with_pmr:1; /* PMR? */
357 cfs_atomic_t ibn_npeers; /* # peers extant */
358 cfs_atomic_t ibn_nconns; /* # connections extant */
360 kib_tx_poolset_t ibn_tx_ps; /* tx pool-set */
361 kib_fmr_poolset_t ibn_fmr_ps; /* fmr pool-set */
362 kib_pmr_poolset_t ibn_pmr_ps; /* pmr pool-set */
364 kib_dev_t *ibn_dev; /* underlying IB device */
369 int kib_init; /* initialisation state */
370 int kib_shutdown; /* shut down? */
371 cfs_list_t kib_devs; /* IB devices extant */
372 cfs_list_t kib_failed_devs; /* list head of failed devices */
373 cfs_atomic_t kib_nthreads; /* # live threads */
374 cfs_rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
376 cfs_list_t *kib_peers; /* hash table of all my known peers */
377 int kib_peer_hash_size;/* size of kib_peers */
379 void *kib_connd; /* the connd task (serialisation assertions) */
380 cfs_list_t kib_connd_conns; /* connections to setup/teardown */
381 cfs_list_t kib_connd_zombies;/* connections with zero refcount */
382 cfs_waitq_t kib_connd_waitq; /* connection daemon sleeps here */
383 cfs_spinlock_t kib_connd_lock; /* serialise */
385 cfs_waitq_t kib_sched_waitq; /* schedulers sleep here */
386 cfs_list_t kib_sched_conns; /* conns to check for rx completions */
387 cfs_spinlock_t kib_sched_lock; /* serialise */
388 cfs_waitq_t kib_failover_waitq; /* schedulers sleep here */
390 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
393 #define IBLND_INIT_NOTHING 0
394 #define IBLND_INIT_DATA 1
395 #define IBLND_INIT_ALL 2
397 /************************************************************************
398 * IB Wire message format.
399 * These are sent in sender's byte order (i.e. receiver flips).
402 typedef struct kib_connparams
404 __u16 ibcp_queue_depth;
405 __u16 ibcp_max_frags;
406 __u32 ibcp_max_msg_size;
407 } WIRE_ATTR kib_connparams_t;
411 lnet_hdr_t ibim_hdr; /* portals header */
412 char ibim_payload[0]; /* piggy-backed payload */
413 } WIRE_ATTR kib_immediate_msg_t;
417 __u32 rf_nob; /* # bytes this frag */
418 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
419 } WIRE_ATTR kib_rdma_frag_t;
423 __u32 rd_key; /* local/remote key */
424 __u32 rd_nfrags; /* # fragments */
425 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
426 } WIRE_ATTR kib_rdma_desc_t;
430 lnet_hdr_t ibprm_hdr; /* portals header */
431 __u64 ibprm_cookie; /* opaque completion cookie */
432 } WIRE_ATTR kib_putreq_msg_t;
436 __u64 ibpam_src_cookie; /* reflected completion cookie */
437 __u64 ibpam_dst_cookie; /* opaque completion cookie */
438 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
439 } WIRE_ATTR kib_putack_msg_t;
443 lnet_hdr_t ibgm_hdr; /* portals header */
444 __u64 ibgm_cookie; /* opaque completion cookie */
445 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
446 } WIRE_ATTR kib_get_msg_t;
450 __u64 ibcm_cookie; /* opaque completion cookie */
451 __s32 ibcm_status; /* < 0 failure: >= 0 length */
452 } WIRE_ATTR kib_completion_msg_t;
456 /* First 2 fields fixed FOR ALL TIME */
457 __u32 ibm_magic; /* I'm an ibnal message */
458 __u16 ibm_version; /* this is my version number */
460 __u8 ibm_type; /* msg type */
461 __u8 ibm_credits; /* returned credits */
462 __u32 ibm_nob; /* # bytes in whole message */
463 __u32 ibm_cksum; /* checksum (0 == no checksum) */
464 __u64 ibm_srcnid; /* sender's NID */
465 __u64 ibm_srcstamp; /* sender's incarnation */
466 __u64 ibm_dstnid; /* destination's NID */
467 __u64 ibm_dststamp; /* destination's incarnation */
470 kib_connparams_t connparams;
471 kib_immediate_msg_t immediate;
472 kib_putreq_msg_t putreq;
473 kib_putack_msg_t putack;
475 kib_completion_msg_t completion;
477 } WIRE_ATTR kib_msg_t;
479 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
481 #define IBLND_MSG_VERSION_1 0x11
482 #define IBLND_MSG_VERSION_2 0x12
483 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
485 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
486 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
487 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
488 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
489 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
490 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
491 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
492 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
493 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
494 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
497 __u32 ibr_magic; /* sender's magic */
498 __u16 ibr_version; /* sender's version */
499 __u8 ibr_why; /* reject reason */
500 __u8 ibr_padding; /* padding */
501 __u64 ibr_incarnation; /* incarnation of peer */
502 kib_connparams_t ibr_cp; /* connection parameters */
503 } WIRE_ATTR kib_rej_t;
505 /* connection rejection reasons */
506 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
507 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
508 #define IBLND_REJECT_FATAL 3 /* Anything else */
510 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
511 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
513 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
514 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
516 /***********************************************************************/
518 typedef struct kib_rx /* receive message */
520 cfs_list_t rx_list; /* queue for attention */
521 struct kib_conn *rx_conn; /* owning conn */
522 int rx_nob; /* # bytes received (-1 while posted) */
523 enum ib_wc_status rx_status; /* completion status */
524 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
525 __u64 rx_msgaddr; /* message buffer (I/O addr) */
526 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
527 struct ib_recv_wr rx_wrq; /* receive work item... */
528 struct ib_sge rx_sge; /* ...and its memory */
531 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
532 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
533 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
534 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
536 typedef struct kib_tx /* transmit message */
538 cfs_list_t tx_list; /* queue on idle_txs ibc_tx_queue etc. */
539 kib_tx_pool_t *tx_pool; /* pool I'm from */
540 struct kib_conn *tx_conn; /* owning conn */
541 short tx_sending; /* # tx callbacks outstanding */
542 short tx_queued; /* queued for sending */
543 short tx_waiting; /* waiting for peer */
544 int tx_status; /* LNET completion status */
545 unsigned long tx_deadline; /* completion deadline */
546 __u64 tx_cookie; /* completion cookie */
547 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
548 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
549 __u64 tx_msgaddr; /* message buffer (I/O addr) */
550 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
551 int tx_nwrq; /* # send work items */
552 struct ib_send_wr *tx_wrq; /* send work items... */
553 struct ib_sge *tx_sge; /* ...and their memory */
554 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
555 int tx_nfrags; /* # entries in... */
556 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
557 __u64 *tx_pages; /* rdma phys page addrs */
559 kib_phys_mr_t *pmr; /* MR for physical buffer */
560 kib_fmr_t fmr; /* FMR */
562 int tx_dmadir; /* dma direction */
565 typedef struct kib_connvars
567 /* connection-in-progress variables */
571 typedef struct kib_conn
573 struct kib_peer *ibc_peer; /* owning peer */
574 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
575 cfs_list_t ibc_list; /* stash on peer's conn list */
576 cfs_list_t ibc_sched_list; /* schedule for attention */
577 __u16 ibc_version; /* version of connection */
578 __u64 ibc_incarnation; /* which instance of the peer */
579 cfs_atomic_t ibc_refcount; /* # users */
580 int ibc_state; /* what's happening */
581 int ibc_nsends_posted; /* # uncompleted sends */
582 int ibc_noops_posted; /* # uncompleted NOOPs */
583 int ibc_credits; /* # credits I have */
584 int ibc_outstanding_credits; /* # credits to return */
585 int ibc_reserved_credits;/* # ACK/DONE msg credits */
586 int ibc_comms_error; /* set on comms error */
587 int ibc_nrx:16; /* receive buffers owned */
588 int ibc_scheduled:1; /* scheduled for attention */
589 int ibc_ready:1; /* CQ callback fired */
590 /* time of last send */
591 unsigned long ibc_last_send;
592 /** link chain for kiblnd_check_conns only */
593 cfs_list_t ibc_connd_list;
594 /** rxs completed before ESTABLISHED */
595 cfs_list_t ibc_early_rxs;
596 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
597 cfs_list_t ibc_tx_noops;
598 cfs_list_t ibc_tx_queue; /* sends that need a credit */
599 cfs_list_t ibc_tx_queue_nocred;/* sends that don't need a credit */
600 cfs_list_t ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
601 cfs_list_t ibc_active_txs; /* active tx awaiting completion */
602 cfs_spinlock_t ibc_lock; /* serialise */
603 kib_rx_t *ibc_rxs; /* the rx descs */
604 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
606 struct rdma_cm_id *ibc_cmid; /* CM id */
607 struct ib_cq *ibc_cq; /* completion queue */
609 kib_connvars_t *ibc_connvars; /* in-progress connection state */
612 #define IBLND_CONN_INIT 0 /* being initialised */
613 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
614 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
615 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
616 #define IBLND_CONN_CLOSING 4 /* being closed */
617 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
619 typedef struct kib_peer
621 cfs_list_t ibp_list; /* stash on global peer list */
622 lnet_nid_t ibp_nid; /* who's on the other end(s) */
623 lnet_ni_t *ibp_ni; /* LNet interface */
624 cfs_atomic_t ibp_refcount; /* # users */
625 cfs_list_t ibp_conns; /* all active connections */
626 cfs_list_t ibp_tx_queue; /* msgs waiting for a conn */
627 __u16 ibp_version; /* version of peer */
628 __u64 ibp_incarnation; /* incarnation of peer */
629 int ibp_connecting; /* current active connection attempts */
630 int ibp_accepting; /* current passive connection attempts */
631 int ibp_error; /* errno on closing this peer */
632 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
635 extern kib_data_t kiblnd_data;
637 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
640 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
642 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
643 cfs_atomic_inc(&hdev->ibh_ref);
647 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
649 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
650 if (cfs_atomic_dec_and_test(&hdev->ibh_ref))
651 kiblnd_hdev_destroy(hdev);
655 kiblnd_dev_can_failover(kib_dev_t *dev)
657 if (!cfs_list_empty(&dev->ibd_fail_list)) /* already scheduled */
660 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
663 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
666 return dev->ibd_can_failover;
669 #define kiblnd_conn_addref(conn) \
671 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
672 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
673 cfs_atomic_inc(&(conn)->ibc_refcount); \
676 #define kiblnd_conn_decref(conn) \
678 unsigned long flags; \
680 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
681 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
682 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
683 if (cfs_atomic_dec_and_test(&(conn)->ibc_refcount)) { \
684 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
685 cfs_list_add_tail(&(conn)->ibc_list, \
686 &kiblnd_data.kib_connd_zombies); \
687 cfs_waitq_signal(&kiblnd_data.kib_connd_waitq); \
688 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
692 #define kiblnd_peer_addref(peer) \
694 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
695 (peer), libcfs_nid2str((peer)->ibp_nid), \
696 cfs_atomic_read (&(peer)->ibp_refcount)); \
697 cfs_atomic_inc(&(peer)->ibp_refcount); \
700 #define kiblnd_peer_decref(peer) \
702 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
703 (peer), libcfs_nid2str((peer)->ibp_nid), \
704 cfs_atomic_read (&(peer)->ibp_refcount)); \
705 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
706 if (cfs_atomic_dec_and_test(&(peer)->ibp_refcount)) \
707 kiblnd_destroy_peer(peer); \
710 static inline cfs_list_t *
711 kiblnd_nid2peerlist (lnet_nid_t nid)
714 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
716 return (&kiblnd_data.kib_peers [hash]);
720 kiblnd_peer_active (kib_peer_t *peer)
722 /* Am I in the peer hash table? */
723 return (!cfs_list_empty(&peer->ibp_list));
726 static inline kib_conn_t *
727 kiblnd_get_conn_locked (kib_peer_t *peer)
729 LASSERT (!cfs_list_empty(&peer->ibp_conns));
731 /* just return the first connection */
732 return cfs_list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
736 kiblnd_send_keepalive(kib_conn_t *conn)
738 return (*kiblnd_tunables.kib_keepalive > 0) &&
739 cfs_time_after(jiffies, conn->ibc_last_send +
740 *kiblnd_tunables.kib_keepalive*CFS_HZ);
744 kiblnd_need_noop(kib_conn_t *conn)
746 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
748 if (conn->ibc_outstanding_credits <
749 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
750 !kiblnd_send_keepalive(conn))
751 return 0; /* No need to send NOOP */
753 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
754 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred))
755 return 0; /* NOOP can be piggybacked */
757 /* No tx to piggyback NOOP onto or no credit to send a tx */
758 return (cfs_list_empty(&conn->ibc_tx_queue) ||
759 conn->ibc_credits == 0);
762 if (!cfs_list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
763 !cfs_list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
764 conn->ibc_credits == 0) /* no credit */
767 if (conn->ibc_credits == 1 && /* last credit reserved for */
768 conn->ibc_outstanding_credits == 0) /* giving back credits */
771 /* No tx to piggyback NOOP onto or no credit to send a tx */
772 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
776 kiblnd_abort_receives(kib_conn_t *conn)
778 ib_modify_qp(conn->ibc_cmid->qp,
779 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
782 static inline const char *
783 kiblnd_queue2str (kib_conn_t *conn, cfs_list_t *q)
785 if (q == &conn->ibc_tx_queue)
788 if (q == &conn->ibc_tx_queue_rsrvd)
789 return "tx_queue_rsrvd";
791 if (q == &conn->ibc_tx_queue_nocred)
792 return "tx_queue_nocred";
794 if (q == &conn->ibc_active_txs)
801 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
802 * lowest bits of the work request id to stash the work item type. */
804 #define IBLND_WID_TX 0
805 #define IBLND_WID_RDMA 1
806 #define IBLND_WID_RX 2
807 #define IBLND_WID_MASK 3UL
810 kiblnd_ptr2wreqid (void *ptr, int type)
812 unsigned long lptr = (unsigned long)ptr;
814 LASSERT ((lptr & IBLND_WID_MASK) == 0);
815 LASSERT ((type & ~IBLND_WID_MASK) == 0);
816 return (__u64)(lptr | type);
820 kiblnd_wreqid2ptr (__u64 wreqid)
822 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
826 kiblnd_wreqid2type (__u64 wreqid)
828 return (wreqid & IBLND_WID_MASK);
832 kiblnd_set_conn_state (kib_conn_t *conn, int state)
834 conn->ibc_state = state;
839 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
841 msg->ibm_type = type;
842 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
846 kiblnd_rd_size (kib_rdma_desc_t *rd)
851 for (i = size = 0; i < rd->rd_nfrags; i++)
852 size += rd->rd_frags[i].rf_nob;
858 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
860 return rd->rd_frags[index].rf_addr;
864 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
866 return rd->rd_frags[index].rf_nob;
870 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
876 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
878 if (nob < rd->rd_frags[index].rf_nob) {
879 rd->rd_frags[index].rf_addr += nob;
880 rd->rd_frags[index].rf_nob -= nob;
889 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
891 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
892 msgtype == IBLND_MSG_PUT_ACK);
894 return msgtype == IBLND_MSG_GET_REQ ?
895 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
896 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
899 #ifdef HAVE_OFED_IB_DMA_MAP
902 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
904 return ib_dma_mapping_error(dev, dma_addr);
907 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
908 void *msg, size_t size,
909 enum dma_data_direction direction)
911 return ib_dma_map_single(dev, msg, size, direction);
914 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
915 __u64 addr, size_t size,
916 enum dma_data_direction direction)
918 ib_dma_unmap_single(dev, addr, size, direction);
921 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
922 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
924 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
925 struct scatterlist *sg, int nents,
926 enum dma_data_direction direction)
928 return ib_dma_map_sg(dev, sg, nents, direction);
931 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
932 struct scatterlist *sg, int nents,
933 enum dma_data_direction direction)
935 ib_dma_unmap_sg(dev, sg, nents, direction);
938 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
939 struct scatterlist *sg)
941 return ib_sg_dma_address(dev, sg);
944 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
945 struct scatterlist *sg)
947 return ib_sg_dma_len(dev, sg);
950 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
951 * right because OFED1.2 defines it as const, to use it we have to add
952 * (void *) cast to overcome "const" */
954 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
955 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
960 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
962 return dma_mapping_error(dma_addr);
965 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
966 void *msg, size_t size,
967 enum dma_data_direction direction)
969 return dma_map_single(dev->dma_device, msg, size, direction);
972 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
973 dma_addr_t addr, size_t size,
974 enum dma_data_direction direction)
976 dma_unmap_single(dev->dma_device, addr, size, direction);
979 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
980 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
982 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
983 struct scatterlist *sg, int nents,
984 enum dma_data_direction direction)
986 return dma_map_sg(dev->dma_device, sg, nents, direction);
989 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
990 struct scatterlist *sg, int nents,
991 enum dma_data_direction direction)
993 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
997 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
998 struct scatterlist *sg)
1000 return sg_dma_address(sg);
1004 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1005 struct scatterlist *sg)
1007 return sg_dma_len(sg);
1010 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
1011 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
1015 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1016 kib_rdma_desc_t *rd);
1017 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
1018 __u64 addr, __u64 size);
1019 void kiblnd_map_rx_descs(kib_conn_t *conn);
1020 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1021 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
1022 kib_rdma_desc_t *rd, int nfrags);
1023 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
1024 void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node);
1025 cfs_list_t *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1027 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1028 int npages, __u64 iov, kib_fmr_t *fmr);
1029 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1031 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1032 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1033 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1035 int kiblnd_startup (lnet_ni_t *ni);
1036 void kiblnd_shutdown (lnet_ni_t *ni);
1037 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
1038 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1040 int kiblnd_tunables_init(void);
1041 void kiblnd_tunables_fini(void);
1043 int kiblnd_connd (void *arg);
1044 int kiblnd_scheduler(void *arg);
1045 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
1046 int kiblnd_failover_thread (void *arg);
1048 int kiblnd_alloc_pages (kib_pages_t **pp, int npages);
1049 void kiblnd_free_pages (kib_pages_t *p);
1051 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1052 struct rdma_cm_event *event);
1053 int kiblnd_translate_mtu(int value);
1055 int kiblnd_dev_failover(kib_dev_t *dev);
1056 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1057 void kiblnd_destroy_peer (kib_peer_t *peer);
1058 void kiblnd_destroy_dev (kib_dev_t *dev);
1059 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1060 void kiblnd_peer_alive (kib_peer_t *peer);
1061 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1062 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1063 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1064 int version, __u64 incarnation);
1065 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1067 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1068 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1069 int state, int version);
1070 void kiblnd_destroy_conn (kib_conn_t *conn);
1071 void kiblnd_close_conn (kib_conn_t *conn, int error);
1072 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1074 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1075 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1077 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1078 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1079 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1080 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1081 void kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist,
1083 void kiblnd_check_sends (kib_conn_t *conn);
1085 void kiblnd_qp_event(struct ib_event *event, void *arg);
1086 void kiblnd_cq_event(struct ib_event *event, void *arg);
1087 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1089 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1090 int credits, lnet_nid_t dstnid, __u64 dststamp);
1091 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1092 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1094 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1095 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1096 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1097 unsigned int offset, unsigned int mlen, unsigned int rlen);