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 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
32 * Copyright (c) 2012, Intel Corporation.
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>
58 #ifdef HAVE_SCATTERLIST_SETPAGE
59 # include <linux/scatterlist.h>
62 #include <asm/system.h>
63 #include <asm/uaccess.h>
66 #include <linux/init.h>
68 #include <linux/file.h>
69 #include <linux/stat.h>
70 #include <linux/list.h>
71 #include <linux/kmod.h>
72 #include <linux/sysctl.h>
73 #include <linux/random.h>
74 #include <linux/pci.h>
79 #define DEBUG_SUBSYSTEM S_LND
81 #include <libcfs/kp30.h>
82 #include <lnet/lnet.h>
83 #include <lnet/lib-lnet.h>
84 #include <lnet/lnet-sysctl.h>
90 #if !defined(HAVE_OFED_BACKPORT_H) && defined(HAVE_SCSI_FC_COMPAT_H)
91 #include <scsi/fc_compat.h>
93 #include <rdma/rdma_cm.h>
94 #include <rdma/ib_cm.h>
95 #include <rdma/ib_verbs.h>
96 #include <rdma/ib_fmr_pool.h>
98 /* tunables fixed at compile time */
100 # define IBLND_N_SCHED num_online_cpus() /* # schedulers */
102 # define IBLND_N_SCHED 1 /* # schedulers */
105 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
106 #define IBLND_RESCHED 100 /* # scheduler loops before reschedule */
110 unsigned int *kib_service; /* IB service number */
111 int *kib_min_reconnect_interval; /* first failed connection retry... */
112 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
113 int *kib_cksum; /* checksum kib_msg_t? */
114 int *kib_timeout; /* comms timeout (seconds) */
115 int *kib_keepalive; /* keepalive timeout (seconds) */
116 int *kib_ntx; /* # tx descs */
117 int *kib_credits; /* # concurrent sends */
118 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
119 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
120 int *kib_peercredits_hiw; /* # when eagerly to return credits */
121 int *kib_peertimeout; /* seconds to consider peer dead */
122 char **kib_default_ipif; /* default IPoIB interface */
123 int *kib_retry_count;
124 int *kib_rnr_retry_count;
125 int *kib_concurrent_sends; /* send work queue sizing */
126 int *kib_ib_mtu; /* IB MTU */
127 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
128 * than this value, 0 disable map-on-demand */
129 int *kib_pmr_pool_size; /* # physical MR in pool */
130 int *kib_fmr_pool_size; /* # FMRs in pool */
131 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
132 int *kib_fmr_cache; /* enable FMR pool cache? */
133 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
134 cfs_sysctl_table_header_t *kib_sysctl; /* sysctl interface */
138 extern kib_tunables_t kiblnd_tunables;
140 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
141 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
143 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
144 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits))-1) /* Max # of peer credits */
146 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
147 IBLND_MSG_QUEUE_SIZE_V1 : \
148 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
149 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
150 IBLND_CREDIT_HIGHWATER_V1 : \
151 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
153 #ifdef HAVE_RDMA_CREATE_ID_4ARG
154 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
156 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
160 kiblnd_concurrent_sends_v1(void)
162 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
163 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
165 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
166 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
168 return *kiblnd_tunables.kib_concurrent_sends;
171 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
172 kiblnd_concurrent_sends_v1() : \
173 *kiblnd_tunables.kib_concurrent_sends)
174 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
175 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
176 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
178 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
179 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
180 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
181 *kiblnd_tunables.kib_map_on_demand : \
182 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
183 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
184 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
186 /************************/
187 /* derived constants... */
189 /* TX messages (shared by all connections) */
190 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
192 /* RX messages (per connection) */
193 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
194 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
195 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
197 /* WRs and CQEs (per connection) */
198 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
199 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
200 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
204 struct list_head ibd_list; /* chain on kib_devs */
205 __u32 ibd_ifip; /* IPoIB interface IP */
206 char ibd_ifname[32]; /* IPoIB interface name */
207 int ibd_nnets; /* # nets extant */
209 struct rdma_cm_id *ibd_cmid; /* IB listener (bound to 1 device) */
210 struct ib_pd *ibd_pd; /* PD for the device */
211 int ibd_page_shift; /* page shift of current HCA */
212 int ibd_page_size; /* page size of current HCA */
213 __u64 ibd_page_mask; /* page mask of current HCA */
214 int ibd_mr_shift; /* bits shift of max MR size */
215 __u64 ibd_mr_size; /* size of MR */
217 int ibd_nmrs; /* # of global MRs */
218 struct ib_mr **ibd_mrs; /* MR for non RDMA I/O */
221 #define IBLND_POOL_DEADLINE 300 /* # of seconds to keep pool alive */
225 struct ib_device *ibp_device; /* device for mapping */
226 int ibp_npages; /* # pages */
227 struct page *ibp_pages[0]; /* page array */
233 struct list_head pmr_list; /* chain node */
234 struct ib_phys_buf *pmr_ipb; /* physical buffer */
235 struct ib_mr *pmr_mr; /* IB MR */
236 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
237 __u64 pmr_iova; /* Virtual I/O address */
238 int pmr_refcount; /* reference count */
244 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps, int inc, struct kib_pool **pp_po);
245 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
246 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, struct list_head *node);
247 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, struct list_head *node);
251 #define IBLND_POOL_NAME_LEN 32
253 typedef struct kib_poolset
255 spinlock_t ps_lock; /* serialize */
256 struct kib_net *ps_net; /* network it belongs to */
257 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
258 struct list_head ps_pool_list; /* list of pools */
259 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
260 int ps_increasing; /* is allocating new pool */
261 int ps_pool_size; /* new pool size */
263 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
264 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
265 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
266 kib_ps_node_fini_t ps_node_fini; /* finalize node */
269 typedef struct kib_pool
271 struct list_head po_list; /* chain on pool list */
272 struct list_head po_free_list; /* pre-allocated node */
273 kib_poolset_t *po_owner; /* pool_set of this pool */
274 cfs_time_t po_deadline; /* deadline of this pool */
275 int po_allocated; /* # of elements in use */
276 int po_size; /* # of pre-allocated elements */
280 kib_poolset_t tps_poolset; /* pool-set */
281 __u64 tps_next_tx_cookie; /* cookie of TX */
285 kib_pool_t tpo_pool; /* pool */
286 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
287 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
291 kib_poolset_t pps_poolset; /* pool-set */
294 typedef struct kib_pmr_pool {
295 kib_pool_t ppo_pool; /* pool */
300 spinlock_t fps_lock; /* serialize */
301 struct kib_net *fps_net; /* IB network */
302 struct list_head fps_pool_list; /* FMR pool list */
303 __u64 fps_version; /* validity stamp */
304 int fps_increasing; /* is allocating new pool */
305 cfs_time_t fps_next_retry; /* time stamp for retry if failed to allocate */
310 struct list_head fpo_list; /* chain on pool list */
311 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
312 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
313 cfs_time_t fpo_deadline; /* deadline of this pool */
314 int fpo_map_count; /* # of mapped FMR */
318 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
319 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
322 typedef struct kib_net
324 __u64 ibn_incarnation; /* my epoch */
325 int ibn_init; /* initialisation state */
326 int ibn_shutdown; /* shutting down? */
327 unsigned int ibn_with_fmr:1; /* FMR? */
328 unsigned int ibn_with_pmr:1; /* PMR? */
330 atomic_t ibn_npeers; /* # peers extant */
331 atomic_t ibn_nconns; /* # connections extant */
333 kib_tx_poolset_t ibn_tx_ps; /* tx pool-set */
334 kib_fmr_poolset_t ibn_fmr_ps; /* fmr pool-set */
335 kib_pmr_poolset_t ibn_pmr_ps; /* pmr pool-set */
337 kib_dev_t *ibn_dev; /* underlying IB device */
342 int kib_init; /* initialisation state */
343 int kib_shutdown; /* shut down? */
344 struct list_head kib_devs; /* IB devices extant */
345 atomic_t kib_nthreads; /* # live threads */
346 rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
348 struct list_head *kib_peers; /* hash table of all my known peers */
349 int kib_peer_hash_size; /* size of kib_peers */
351 void *kib_connd; /* the connd task (serialisation assertions) */
352 struct list_head kib_connd_conns; /* connections to setup/teardown */
353 struct list_head kib_connd_zombies; /* connections with zero refcount */
354 wait_queue_head_t kib_connd_waitq; /* connection daemon sleeps here */
355 spinlock_t kib_connd_lock; /* serialise */
357 wait_queue_head_t kib_sched_waitq; /* schedulers sleep here */
358 struct list_head kib_sched_conns; /* conns to check for rx completions */
359 spinlock_t kib_sched_lock; /* serialise */
361 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
364 #define IBLND_INIT_NOTHING 0
365 #define IBLND_INIT_DATA 1
366 #define IBLND_INIT_ALL 2
368 /************************************************************************
369 * IB Wire message format.
370 * These are sent in sender's byte order (i.e. receiver flips).
373 typedef struct kib_connparams
375 __u16 ibcp_queue_depth;
376 __u16 ibcp_max_frags;
377 __u32 ibcp_max_msg_size;
378 } WIRE_ATTR kib_connparams_t;
382 lnet_hdr_t ibim_hdr; /* portals header */
383 char ibim_payload[0]; /* piggy-backed payload */
384 } WIRE_ATTR kib_immediate_msg_t;
388 __u32 rf_nob; /* # bytes this frag */
389 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
390 } WIRE_ATTR kib_rdma_frag_t;
394 __u32 rd_key; /* local/remote key */
395 __u32 rd_nfrags; /* # fragments */
396 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
397 } WIRE_ATTR kib_rdma_desc_t;
401 lnet_hdr_t ibprm_hdr; /* portals header */
402 __u64 ibprm_cookie; /* opaque completion cookie */
403 } WIRE_ATTR kib_putreq_msg_t;
407 __u64 ibpam_src_cookie; /* reflected completion cookie */
408 __u64 ibpam_dst_cookie; /* opaque completion cookie */
409 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
410 } WIRE_ATTR kib_putack_msg_t;
414 lnet_hdr_t ibgm_hdr; /* portals header */
415 __u64 ibgm_cookie; /* opaque completion cookie */
416 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
417 } WIRE_ATTR kib_get_msg_t;
421 __u64 ibcm_cookie; /* opaque completion cookie */
422 __s32 ibcm_status; /* < 0 failure: >= 0 length */
423 } WIRE_ATTR kib_completion_msg_t;
427 /* First 2 fields fixed FOR ALL TIME */
428 __u32 ibm_magic; /* I'm an openibnal message */
429 __u16 ibm_version; /* this is my version number */
431 __u8 ibm_type; /* msg type */
432 __u8 ibm_credits; /* returned credits */
433 __u32 ibm_nob; /* # bytes in whole message */
434 __u32 ibm_cksum; /* checksum (0 == no checksum) */
435 __u64 ibm_srcnid; /* sender's NID */
436 __u64 ibm_srcstamp; /* sender's incarnation */
437 __u64 ibm_dstnid; /* destination's NID */
438 __u64 ibm_dststamp; /* destination's incarnation */
441 kib_connparams_t connparams;
442 kib_immediate_msg_t immediate;
443 kib_putreq_msg_t putreq;
444 kib_putack_msg_t putack;
446 kib_completion_msg_t completion;
448 } WIRE_ATTR kib_msg_t;
450 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
452 #define IBLND_MSG_VERSION_1 0x11
453 #define IBLND_MSG_VERSION_2 0x12
454 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
456 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
457 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
458 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
459 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
460 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
461 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
462 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
463 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
464 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
465 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
468 __u32 ibr_magic; /* sender's magic */
469 __u16 ibr_version; /* sender's version */
470 __u8 ibr_why; /* reject reason */
471 __u8 ibr_padding; /* padding */
472 __u64 ibr_incarnation; /* incarnation of peer */
473 kib_connparams_t ibr_cp; /* connection parameters */
474 } WIRE_ATTR kib_rej_t;
476 /* connection rejection reasons */
477 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
478 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
479 #define IBLND_REJECT_FATAL 3 /* Anything else */
481 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
482 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
484 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
485 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
487 /***********************************************************************/
489 typedef struct kib_rx /* receive message */
491 struct list_head rx_list; /* queue for attention */
492 struct kib_conn *rx_conn; /* owning conn */
493 int rx_nob; /* # bytes received (-1 while posted) */
494 enum ib_wc_status rx_status; /* completion status */
495 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
496 __u64 rx_msgaddr; /* message buffer (I/O addr) */
497 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
498 struct ib_recv_wr rx_wrq; /* receive work item... */
499 struct ib_sge rx_sge; /* ...and its memory */
502 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
503 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
504 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
505 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
507 typedef struct kib_tx /* transmit message */
509 struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */
510 kib_tx_pool_t *tx_pool; /* pool I'm from */
511 struct kib_conn *tx_conn; /* owning conn */
512 short tx_sending; /* # tx callbacks outstanding */
513 short tx_queued; /* queued for sending */
514 short tx_waiting; /* waiting for peer */
515 int tx_status; /* LNET completion status */
516 unsigned long tx_deadline; /* completion deadline */
517 __u64 tx_cookie; /* completion cookie */
518 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
519 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
520 __u64 tx_msgaddr; /* message buffer (I/O addr) */
521 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
522 int tx_nwrq; /* # send work items */
523 struct ib_send_wr *tx_wrq; /* send work items... */
524 struct ib_sge *tx_sge; /* ...and their memory */
525 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
526 int tx_nfrags; /* # entries in... */
527 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
528 __u64 *tx_pages; /* rdma phys page addrs */
530 kib_phys_mr_t *pmr; /* MR for physical buffer */
531 kib_fmr_t fmr; /* FMR */
533 int tx_dmadir; /* dma direction */
536 typedef struct kib_connvars
538 /* connection-in-progress variables */
542 typedef struct kib_conn
544 struct kib_peer *ibc_peer; /* owning peer */
545 struct list_head ibc_list; /* stash on peer's conn list */
546 struct list_head ibc_sched_list; /* schedule for attention */
547 __u16 ibc_version; /* version of connection */
548 __u64 ibc_incarnation; /* which instance of the peer */
549 atomic_t ibc_refcount; /* # users */
550 int ibc_state; /* what's happening */
551 int ibc_nsends_posted; /* # uncompleted sends */
552 int ibc_noops_posted; /* # uncompleted NOOPs */
553 int ibc_credits; /* # credits I have */
554 int ibc_outstanding_credits; /* # credits to return */
555 int ibc_reserved_credits;/* # ACK/DONE msg credits */
556 int ibc_comms_error; /* set on comms error */
557 int ibc_nrx:16; /* receive buffers owned */
558 int ibc_scheduled:1; /* scheduled for attention */
559 int ibc_ready:1; /* CQ callback fired */
560 unsigned long ibc_last_send; /* time of last send */
561 struct list_head ibc_early_rxs; /* rxs completed before ESTABLISHED */
562 struct list_head ibc_tx_noops; /* IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
563 struct list_head ibc_tx_queue; /* sends that need a credit */
564 struct list_head ibc_tx_queue_nocred;/* sends that don't need a credit */
565 struct list_head ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
566 struct list_head ibc_active_txs; /* active tx awaiting completion */
567 spinlock_t ibc_lock; /* serialise */
568 kib_rx_t *ibc_rxs; /* the rx descs */
569 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
571 struct rdma_cm_id *ibc_cmid; /* CM id */
572 struct ib_cq *ibc_cq; /* completion queue */
574 kib_connvars_t *ibc_connvars; /* in-progress connection state */
577 #define IBLND_CONN_INIT 0 /* being intialised */
578 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
579 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
580 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
581 #define IBLND_CONN_CLOSING 4 /* being closed */
582 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
584 typedef struct kib_peer
586 struct list_head ibp_list; /* stash on global peer list */
587 lnet_nid_t ibp_nid; /* who's on the other end(s) */
588 lnet_ni_t *ibp_ni; /* LNet interface */
589 atomic_t ibp_refcount; /* # users */
590 struct list_head ibp_conns; /* all active connections */
591 struct list_head ibp_tx_queue; /* msgs waiting for a conn */
592 __u16 ibp_version; /* version of peer */
593 __u64 ibp_incarnation; /* incarnation of peer */
594 int ibp_connecting; /* current active connection attempts */
595 int ibp_accepting; /* current passive connection attempts */
596 int ibp_error; /* errno on closing this peer */
597 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
600 extern kib_data_t kiblnd_data;
602 #define kiblnd_conn_addref(conn) \
604 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
605 (conn), atomic_read(&(conn)->ibc_refcount)); \
606 LASSERT(atomic_read(&(conn)->ibc_refcount) > 0); \
607 atomic_inc(&(conn)->ibc_refcount); \
610 #define kiblnd_conn_decref(conn) \
612 unsigned long flags; \
614 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
615 (conn), atomic_read(&(conn)->ibc_refcount)); \
616 LASSERT(atomic_read(&(conn)->ibc_refcount) > 0); \
617 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
618 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
619 list_add_tail(&(conn)->ibc_list, \
620 &kiblnd_data.kib_connd_zombies); \
621 wake_up(&kiblnd_data.kib_connd_waitq); \
622 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags); \
626 #define kiblnd_peer_addref(peer) \
628 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
629 (peer), libcfs_nid2str((peer)->ibp_nid), \
630 atomic_read (&(peer)->ibp_refcount)); \
631 LASSERT(atomic_read(&(peer)->ibp_refcount) > 0); \
632 atomic_inc(&(peer)->ibp_refcount); \
635 #define kiblnd_peer_decref(peer) \
637 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
638 (peer), libcfs_nid2str((peer)->ibp_nid), \
639 atomic_read (&(peer)->ibp_refcount)); \
640 LASSERT(atomic_read(&(peer)->ibp_refcount) > 0); \
641 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
642 kiblnd_destroy_peer(peer); \
645 static inline struct list_head *
646 kiblnd_nid2peerlist (lnet_nid_t nid)
648 unsigned int hash = ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
650 return (&kiblnd_data.kib_peers [hash]);
654 kiblnd_peer_active (kib_peer_t *peer)
656 /* Am I in the peer hash table? */
657 return (!list_empty(&peer->ibp_list));
660 static inline kib_conn_t *
661 kiblnd_get_conn_locked (kib_peer_t *peer)
663 LASSERT (!list_empty(&peer->ibp_conns));
665 /* just return the first connection */
666 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
670 kiblnd_send_keepalive(kib_conn_t *conn)
672 return (*kiblnd_tunables.kib_keepalive > 0) &&
673 time_after(jiffies, conn->ibc_last_send +
674 *kiblnd_tunables.kib_keepalive*HZ);
678 kiblnd_send_noop(kib_conn_t *conn)
680 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
682 if (conn->ibc_outstanding_credits <
683 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
684 !kiblnd_send_keepalive(conn))
685 return 0; /* No need to send NOOP */
687 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
688 if (!list_empty(&conn->ibc_tx_queue_nocred))
689 return 0; /* NOOP can be piggybacked */
691 /* No tx to piggyback NOOP onto or no credit to send a tx */
692 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 0);
695 if (!list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
696 !list_empty(&conn->ibc_tx_queue_nocred) || /* can be piggybacked */
697 conn->ibc_credits == 0) /* no credit */
700 if (conn->ibc_credits == 1 && /* last credit reserved for */
701 conn->ibc_outstanding_credits == 0) /* giving back credits */
704 /* No tx to piggyback NOOP onto or no credit to send a tx */
705 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
709 kiblnd_abort_receives(kib_conn_t *conn)
711 ib_modify_qp(conn->ibc_cmid->qp,
712 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
715 static inline const char *
716 kiblnd_queue2str (kib_conn_t *conn, struct list_head *q)
718 if (q == &conn->ibc_tx_queue)
721 if (q == &conn->ibc_tx_queue_rsrvd)
722 return "tx_queue_rsrvd";
724 if (q == &conn->ibc_tx_queue_nocred)
725 return "tx_queue_nocred";
727 if (q == &conn->ibc_active_txs)
734 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
735 * lowest bits of the work request id to stash the work item type. */
737 #define IBLND_WID_TX 0
738 #define IBLND_WID_RDMA 1
739 #define IBLND_WID_RX 2
740 #define IBLND_WID_MASK 3UL
743 kiblnd_ptr2wreqid (void *ptr, int type)
745 unsigned long lptr = (unsigned long)ptr;
747 LASSERT ((lptr & IBLND_WID_MASK) == 0);
748 LASSERT ((type & ~IBLND_WID_MASK) == 0);
749 return (__u64)(lptr | type);
753 kiblnd_wreqid2ptr (__u64 wreqid)
755 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
759 kiblnd_wreqid2type (__u64 wreqid)
761 return (wreqid & IBLND_WID_MASK);
765 kiblnd_set_conn_state (kib_conn_t *conn, int state)
767 conn->ibc_state = state;
772 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
774 msg->ibm_type = type;
775 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
779 kiblnd_rd_size (kib_rdma_desc_t *rd)
784 for (i = size = 0; i < rd->rd_nfrags; i++)
785 size += rd->rd_frags[i].rf_nob;
791 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
793 return rd->rd_frags[index].rf_addr;
797 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
799 return rd->rd_frags[index].rf_nob;
803 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
809 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
811 if (nob < rd->rd_frags[index].rf_nob) {
812 rd->rd_frags[index].rf_addr += nob;
813 rd->rd_frags[index].rf_nob -= nob;
822 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
824 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
825 msgtype == IBLND_MSG_PUT_ACK);
827 return msgtype == IBLND_MSG_GET_REQ ?
828 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
829 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
832 #ifdef HAVE_OFED_IB_DMA_MAP
835 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
837 return ib_dma_mapping_error(dev, dma_addr);
840 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
841 void *msg, size_t size,
842 enum dma_data_direction direction)
844 return ib_dma_map_single(dev, msg, size, direction);
847 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
848 __u64 addr, size_t size,
849 enum dma_data_direction direction)
851 ib_dma_unmap_single(dev, addr, size, direction);
854 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
855 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
857 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
858 struct scatterlist *sg, int nents,
859 enum dma_data_direction direction)
861 return ib_dma_map_sg(dev, sg, nents, direction);
864 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
865 struct scatterlist *sg, int nents,
866 enum dma_data_direction direction)
868 ib_dma_unmap_sg(dev, sg, nents, direction);
871 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
872 struct scatterlist *sg)
874 return ib_sg_dma_address(dev, sg);
877 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
878 struct scatterlist *sg)
880 return ib_sg_dma_len(dev, sg);
883 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
884 * right because OFED1.2 defines it as const, to use it we have to add
885 * (void *) cast to overcome "const" */
887 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
888 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
893 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
895 return dma_mapping_error(dma_addr);
898 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
899 void *msg, size_t size,
900 enum dma_data_direction direction)
902 return dma_map_single(dev->dma_device, msg, size, direction);
905 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
906 dma_addr_t addr, size_t size,
907 enum dma_data_direction direction)
909 dma_unmap_single(dev->dma_device, addr, size, direction);
912 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
913 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
915 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
916 struct scatterlist *sg, int nents,
917 enum dma_data_direction direction)
919 return dma_map_sg(dev->dma_device, sg, nents, direction);
922 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
923 struct scatterlist *sg, int nents,
924 enum dma_data_direction direction)
926 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
930 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
931 struct scatterlist *sg)
933 return sg_dma_address(sg);
937 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
938 struct scatterlist *sg)
940 return sg_dma_len(sg);
943 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
944 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
948 struct ib_mr *kiblnd_find_rd_dma_mr(kib_net_t *net,
949 kib_rdma_desc_t *rd);
950 struct ib_mr *kiblnd_find_dma_mr(kib_net_t *net,
951 __u64 addr, __u64 size);
952 void kiblnd_map_rx_descs(kib_conn_t *conn);
953 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
954 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
955 kib_rdma_desc_t *rd, int nfrags);
956 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
957 void kiblnd_pool_free_node(kib_pool_t *pool, struct list_head *node);
958 struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
960 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
961 int npages, __u64 iov, kib_fmr_t *fmr);
962 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
964 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd,
965 __u64 *iova, kib_phys_mr_t **pp_pmr);
966 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
968 int kiblnd_startup (lnet_ni_t *ni);
969 void kiblnd_shutdown (lnet_ni_t *ni);
970 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
971 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
973 int kiblnd_tunables_init(void);
974 void kiblnd_tunables_fini(void);
976 int kiblnd_connd (void *arg);
977 int kiblnd_scheduler(void *arg);
978 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
980 int kiblnd_alloc_pages (kib_pages_t **pp, int npages);
981 void kiblnd_free_pages (kib_pages_t *p);
983 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
984 struct rdma_cm_event *event);
985 int kiblnd_translate_mtu(int value);
987 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
988 void kiblnd_destroy_peer (kib_peer_t *peer);
989 void kiblnd_destroy_dev (kib_dev_t *dev);
990 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
991 void kiblnd_peer_alive (kib_peer_t *peer);
992 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
993 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
994 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
995 int version, __u64 incarnation);
996 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
998 void kiblnd_connreq_done(kib_conn_t *conn, int status);
999 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1000 int state, int version);
1001 void kiblnd_destroy_conn (kib_conn_t *conn);
1002 void kiblnd_close_conn (kib_conn_t *conn, int error);
1003 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1005 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1006 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1008 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1009 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1010 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1011 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1012 void kiblnd_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int status);
1013 void kiblnd_check_sends (kib_conn_t *conn);
1015 void kiblnd_qp_event(struct ib_event *event, void *arg);
1016 void kiblnd_cq_event(struct ib_event *event, void *arg);
1017 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1019 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1020 int credits, lnet_nid_t dstnid, __u64 dststamp);
1021 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1022 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1024 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1025 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1026 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1027 unsigned int offset, unsigned int mlen, unsigned int rlen);
1029 /* compat macroses */
1030 #ifndef HAVE_SCATTERLIST_SETPAGE
1031 static inline void sg_set_page(struct scatterlist *sg, struct page *page,
1032 unsigned int len, unsigned int offset)
1035 sg->offset = offset;