4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd.h
38 * Author: Eric Barton <eric@bartonsoftware.com>
41 #ifndef AUTOCONF_INCLUDED
42 #include <linux/config.h>
44 #include <linux/module.h>
45 #include <linux/kernel.h>
47 #include <linux/string.h>
48 #include <linux/stat.h>
49 #include <linux/errno.h>
50 #include <linux/smp_lock.h>
51 #include <linux/unistd.h>
52 #include <linux/uio.h>
54 #include <asm/system.h>
55 #include <asm/uaccess.h>
58 #include <linux/init.h>
60 #include <linux/file.h>
61 #include <linux/stat.h>
62 #include <linux/list.h>
63 #include <linux/kmod.h>
64 #include <linux/sysctl.h>
65 #include <linux/pci.h>
70 #define DEBUG_SUBSYSTEM S_LND
72 #include <libcfs/libcfs.h>
73 #include <lnet/lnet.h>
74 #include <lnet/lib-lnet.h>
75 #include <lnet/lnet-sysctl.h>
77 #include <rdma/rdma_cm.h>
78 #include <rdma/ib_cm.h>
79 #include <rdma/ib_verbs.h>
80 #include <rdma/ib_fmr_pool.h>
82 /* tunables fixed at compile time */
84 # define IBLND_N_SCHED cfs_num_online_cpus() /* # schedulers */
86 # define IBLND_N_SCHED 1 /* # schedulers */
89 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
90 #define IBLND_RESCHED 100 /* # scheduler loops before reschedule */
94 int *kib_dev_failover; /* HCA failover */
95 unsigned int *kib_service; /* IB service number */
96 int *kib_min_reconnect_interval; /* first failed connection retry... */
97 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
98 int *kib_cksum; /* checksum kib_msg_t? */
99 int *kib_timeout; /* comms timeout (seconds) */
100 int *kib_keepalive; /* keepalive timeout (seconds) */
101 int *kib_ntx; /* # tx descs */
102 int *kib_credits; /* # concurrent sends */
103 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
104 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
105 int *kib_peercredits_hiw; /* # when eagerly to return credits */
106 int *kib_peertimeout; /* seconds to consider peer dead */
107 char **kib_default_ipif; /* default IPoIB interface */
108 int *kib_retry_count;
109 int *kib_rnr_retry_count;
110 int *kib_concurrent_sends; /* send work queue sizing */
111 int *kib_ib_mtu; /* IB MTU */
112 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
113 * than this value, 0 disable map-on-demand */
114 int *kib_pmr_pool_size; /* # physical MR in pool */
115 int *kib_fmr_pool_size; /* # FMRs in pool */
116 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
117 int *kib_fmr_cache; /* enable FMR pool cache? */
118 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
119 cfs_sysctl_table_header_t *kib_sysctl; /* sysctl interface */
121 int *kib_require_priv_port;/* accept only privileged ports */
122 int *kib_use_priv_port; /* use privileged port for active connect */
125 extern kib_tunables_t kiblnd_tunables;
127 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
128 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
130 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
131 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
133 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
134 IBLND_MSG_QUEUE_SIZE_V1 : \
135 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
136 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
137 IBLND_CREDIT_HIGHWATER_V1 : \
138 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
140 #ifdef HAVE_RDMA_CREATE_ID_4ARG
141 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps, qpt)
143 #define kiblnd_rdma_create_id(cb, dev, ps, qpt) rdma_create_id(cb, dev, ps)
147 kiblnd_concurrent_sends_v1(void)
149 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
150 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
152 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
153 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
155 return *kiblnd_tunables.kib_concurrent_sends;
158 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
159 kiblnd_concurrent_sends_v1() : \
160 *kiblnd_tunables.kib_concurrent_sends)
161 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
162 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
163 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
165 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
166 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
167 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
168 *kiblnd_tunables.kib_map_on_demand : \
169 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
170 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
171 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
173 /************************/
174 /* derived constants... */
176 /* TX messages (shared by all connections) */
177 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
179 /* RX messages (per connection) */
180 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
181 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
182 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
184 /* WRs and CQEs (per connection) */
185 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
186 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
187 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
191 /* o2iblnd can run over aliased interface */
193 #define KIB_IFNAME_SIZE IFALIASZ
195 #define KIB_IFNAME_SIZE 256
200 cfs_list_t ibd_list; /* chain on kib_devs */
201 cfs_list_t ibd_fail_list; /* chain on kib_failed_devs */
202 __u32 ibd_ifip; /* IPoIB interface IP */
203 /** IPoIB interface name */
204 char ibd_ifname[KIB_IFNAME_SIZE];
205 int ibd_nnets; /* # nets extant */
207 cfs_time_t ibd_next_failover;
208 int ibd_failed_failover; /* # failover failures */
209 unsigned int ibd_failover; /* failover in progress */
210 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
212 struct kib_hca_dev *ibd_hdev;
215 typedef struct kib_hca_dev
217 struct rdma_cm_id *ibh_cmid; /* listener cmid */
218 struct ib_device *ibh_ibdev; /* IB device */
219 int ibh_page_shift; /* page shift of current HCA */
220 int ibh_page_size; /* page size of current HCA */
221 __u64 ibh_page_mask; /* page mask of current HCA */
222 int ibh_mr_shift; /* bits shift of max MR size */
223 __u64 ibh_mr_size; /* size of MR */
224 int ibh_nmrs; /* # of global MRs */
225 struct ib_mr **ibh_mrs; /* global MR */
226 struct ib_pd *ibh_pd; /* PD */
227 kib_dev_t *ibh_dev; /* owner */
228 cfs_atomic_t ibh_ref; /* refcount */
231 /** # of seconds to keep pool alive */
232 #define IBLND_POOL_DEADLINE 300
233 /** # of seconds to retry if allocation failed */
234 #define IBLND_POOL_RETRY 1
238 int ibp_npages; /* # pages */
239 struct page *ibp_pages[0]; /* page array */
245 cfs_list_t pmr_list; /* chain node */
246 struct ib_phys_buf *pmr_ipb; /* physical buffer */
247 struct ib_mr *pmr_mr; /* IB MR */
248 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
249 __u64 pmr_iova; /* Virtual I/O address */
250 int pmr_refcount; /* reference count */
256 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps, int inc, struct kib_pool **pp_po);
257 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
258 typedef void (*kib_ps_node_init_t)(struct kib_pool *po,
260 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po,
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 */
278 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
279 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
280 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
281 kib_ps_node_fini_t ps_node_fini; /* finalize node */
284 typedef struct kib_pool
286 cfs_list_t po_list; /* chain on pool list */
287 cfs_list_t po_free_list; /* pre-allocated node */
288 kib_poolset_t *po_owner; /* pool_set of this pool */
289 cfs_time_t po_deadline; /* deadline of this pool */
290 int po_allocated; /* # of elements in use */
291 int po_failed; /* pool is created on failed HCA */
292 int po_size; /* # of pre-allocated elements */
296 kib_poolset_t tps_poolset; /* pool-set */
297 __u64 tps_next_tx_cookie; /* cookie of TX */
301 kib_pool_t tpo_pool; /* pool */
302 struct kib_hca_dev *tpo_hdev; /* device for this pool */
303 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
304 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
308 kib_poolset_t pps_poolset; /* pool-set */
311 typedef struct kib_pmr_pool {
312 struct kib_hca_dev *ppo_hdev; /* device for this pool */
313 kib_pool_t ppo_pool; /* pool */
318 cfs_spinlock_t fps_lock; /* serialize */
319 struct kib_net *fps_net; /* IB network */
320 cfs_list_t fps_pool_list; /* FMR pool list */
321 cfs_list_t fps_failed_pool_list; /* FMR pool list */
322 __u64 fps_version; /* validity stamp */
323 int fps_increasing; /* is allocating new pool */
324 cfs_time_t fps_next_retry; /* time stamp for retry if failed to allocate */
329 cfs_list_t fpo_list; /* chain on pool list */
330 struct kib_hca_dev *fpo_hdev; /* device for this pool */
331 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
332 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
333 cfs_time_t fpo_deadline; /* deadline of this pool */
334 int fpo_failed; /* fmr pool is failed */
335 int fpo_map_count; /* # of mapped FMR */
339 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
340 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
343 typedef struct kib_net
345 cfs_list_t ibn_list; /* chain on kib_dev_t::ibd_nets */
346 __u64 ibn_incarnation; /* my epoch */
347 int ibn_init; /* initialisation state */
348 int ibn_shutdown; /* shutting down? */
349 unsigned int ibn_with_fmr:1; /* FMR? */
350 unsigned int ibn_with_pmr:1; /* PMR? */
352 cfs_atomic_t ibn_npeers; /* # peers extant */
353 cfs_atomic_t ibn_nconns; /* # connections extant */
355 kib_tx_poolset_t ibn_tx_ps; /* tx pool-set */
356 kib_fmr_poolset_t ibn_fmr_ps; /* fmr pool-set */
357 kib_pmr_poolset_t ibn_pmr_ps; /* pmr pool-set */
359 kib_dev_t *ibn_dev; /* underlying IB device */
364 int kib_init; /* initialisation state */
365 int kib_shutdown; /* shut down? */
366 cfs_list_t kib_devs; /* IB devices extant */
367 cfs_list_t kib_failed_devs; /* list head of failed devices */
368 cfs_atomic_t kib_nthreads; /* # live threads */
369 cfs_rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
371 cfs_list_t *kib_peers; /* hash table of all my known peers */
372 int kib_peer_hash_size;/* size of kib_peers */
374 void *kib_connd; /* the connd task (serialisation assertions) */
375 cfs_list_t kib_connd_conns; /* connections to setup/teardown */
376 cfs_list_t kib_connd_zombies;/* connections with zero refcount */
377 cfs_waitq_t kib_connd_waitq; /* connection daemon sleeps here */
378 cfs_spinlock_t kib_connd_lock; /* serialise */
380 cfs_waitq_t kib_sched_waitq; /* schedulers sleep here */
381 cfs_list_t kib_sched_conns; /* conns to check for rx completions */
382 cfs_spinlock_t kib_sched_lock; /* serialise */
383 cfs_waitq_t kib_failover_waitq; /* schedulers sleep here */
385 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
388 #define IBLND_INIT_NOTHING 0
389 #define IBLND_INIT_DATA 1
390 #define IBLND_INIT_ALL 2
392 /************************************************************************
393 * IB Wire message format.
394 * These are sent in sender's byte order (i.e. receiver flips).
397 typedef struct kib_connparams
399 __u16 ibcp_queue_depth;
400 __u16 ibcp_max_frags;
401 __u32 ibcp_max_msg_size;
402 } WIRE_ATTR kib_connparams_t;
406 lnet_hdr_t ibim_hdr; /* portals header */
407 char ibim_payload[0]; /* piggy-backed payload */
408 } WIRE_ATTR kib_immediate_msg_t;
412 __u32 rf_nob; /* # bytes this frag */
413 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
414 } WIRE_ATTR kib_rdma_frag_t;
418 __u32 rd_key; /* local/remote key */
419 __u32 rd_nfrags; /* # fragments */
420 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
421 } WIRE_ATTR kib_rdma_desc_t;
425 lnet_hdr_t ibprm_hdr; /* portals header */
426 __u64 ibprm_cookie; /* opaque completion cookie */
427 } WIRE_ATTR kib_putreq_msg_t;
431 __u64 ibpam_src_cookie; /* reflected completion cookie */
432 __u64 ibpam_dst_cookie; /* opaque completion cookie */
433 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
434 } WIRE_ATTR kib_putack_msg_t;
438 lnet_hdr_t ibgm_hdr; /* portals header */
439 __u64 ibgm_cookie; /* opaque completion cookie */
440 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
441 } WIRE_ATTR kib_get_msg_t;
445 __u64 ibcm_cookie; /* opaque completion cookie */
446 __s32 ibcm_status; /* < 0 failure: >= 0 length */
447 } WIRE_ATTR kib_completion_msg_t;
451 /* First 2 fields fixed FOR ALL TIME */
452 __u32 ibm_magic; /* I'm an ibnal message */
453 __u16 ibm_version; /* this is my version number */
455 __u8 ibm_type; /* msg type */
456 __u8 ibm_credits; /* returned credits */
457 __u32 ibm_nob; /* # bytes in whole message */
458 __u32 ibm_cksum; /* checksum (0 == no checksum) */
459 __u64 ibm_srcnid; /* sender's NID */
460 __u64 ibm_srcstamp; /* sender's incarnation */
461 __u64 ibm_dstnid; /* destination's NID */
462 __u64 ibm_dststamp; /* destination's incarnation */
465 kib_connparams_t connparams;
466 kib_immediate_msg_t immediate;
467 kib_putreq_msg_t putreq;
468 kib_putack_msg_t putack;
470 kib_completion_msg_t completion;
472 } WIRE_ATTR kib_msg_t;
474 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
476 #define IBLND_MSG_VERSION_1 0x11
477 #define IBLND_MSG_VERSION_2 0x12
478 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
480 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
481 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
482 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
483 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
484 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
485 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
486 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
487 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
488 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
489 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
492 __u32 ibr_magic; /* sender's magic */
493 __u16 ibr_version; /* sender's version */
494 __u8 ibr_why; /* reject reason */
495 __u8 ibr_padding; /* padding */
496 __u64 ibr_incarnation; /* incarnation of peer */
497 kib_connparams_t ibr_cp; /* connection parameters */
498 } WIRE_ATTR kib_rej_t;
500 /* connection rejection reasons */
501 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
502 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
503 #define IBLND_REJECT_FATAL 3 /* Anything else */
505 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
506 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
508 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
509 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
511 /***********************************************************************/
513 typedef struct kib_rx /* receive message */
515 cfs_list_t rx_list; /* queue for attention */
516 struct kib_conn *rx_conn; /* owning conn */
517 int rx_nob; /* # bytes received (-1 while posted) */
518 enum ib_wc_status rx_status; /* completion status */
519 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
520 __u64 rx_msgaddr; /* message buffer (I/O addr) */
521 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
522 struct ib_recv_wr rx_wrq; /* receive work item... */
523 struct ib_sge rx_sge; /* ...and its memory */
526 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
527 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
528 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
529 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
531 typedef struct kib_tx /* transmit message */
533 cfs_list_t tx_list; /* queue on idle_txs ibc_tx_queue etc. */
534 kib_tx_pool_t *tx_pool; /* pool I'm from */
535 struct kib_conn *tx_conn; /* owning conn */
536 short tx_sending; /* # tx callbacks outstanding */
537 short tx_queued; /* queued for sending */
538 short tx_waiting; /* waiting for peer */
539 int tx_status; /* LNET completion status */
540 unsigned long tx_deadline; /* completion deadline */
541 __u64 tx_cookie; /* completion cookie */
542 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
543 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
544 __u64 tx_msgaddr; /* message buffer (I/O addr) */
545 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
546 int tx_nwrq; /* # send work items */
547 struct ib_send_wr *tx_wrq; /* send work items... */
548 struct ib_sge *tx_sge; /* ...and their memory */
549 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
550 int tx_nfrags; /* # entries in... */
551 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
552 __u64 *tx_pages; /* rdma phys page addrs */
554 kib_phys_mr_t *pmr; /* MR for physical buffer */
555 kib_fmr_t fmr; /* FMR */
557 int tx_dmadir; /* dma direction */
560 typedef struct kib_connvars
562 /* connection-in-progress variables */
566 typedef struct kib_conn
568 struct kib_peer *ibc_peer; /* owning peer */
569 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
570 cfs_list_t ibc_list; /* stash on peer's conn list */
571 cfs_list_t ibc_sched_list; /* schedule for attention */
572 __u16 ibc_version; /* version of connection */
573 __u64 ibc_incarnation; /* which instance of the peer */
574 cfs_atomic_t ibc_refcount; /* # users */
575 int ibc_state; /* what's happening */
576 int ibc_nsends_posted; /* # uncompleted sends */
577 int ibc_noops_posted; /* # uncompleted NOOPs */
578 int ibc_credits; /* # credits I have */
579 int ibc_outstanding_credits; /* # credits to return */
580 int ibc_reserved_credits;/* # ACK/DONE msg credits */
581 int ibc_comms_error; /* set on comms error */
582 int ibc_nrx:16; /* receive buffers owned */
583 int ibc_scheduled:1; /* scheduled for attention */
584 int ibc_ready:1; /* CQ callback fired */
585 /* time of last send */
586 unsigned long ibc_last_send;
587 /** link chain for kiblnd_check_conns only */
588 cfs_list_t ibc_connd_list;
589 /** rxs completed before ESTABLISHED */
590 cfs_list_t ibc_early_rxs;
591 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
592 cfs_list_t ibc_tx_noops;
593 cfs_list_t ibc_tx_queue; /* sends that need a credit */
594 cfs_list_t ibc_tx_queue_nocred;/* sends that don't need a credit */
595 cfs_list_t ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
596 cfs_list_t ibc_active_txs; /* active tx awaiting completion */
597 cfs_spinlock_t ibc_lock; /* serialise */
598 kib_rx_t *ibc_rxs; /* the rx descs */
599 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
601 struct rdma_cm_id *ibc_cmid; /* CM id */
602 struct ib_cq *ibc_cq; /* completion queue */
604 kib_connvars_t *ibc_connvars; /* in-progress connection state */
607 #define IBLND_CONN_INIT 0 /* being initialised */
608 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
609 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
610 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
611 #define IBLND_CONN_CLOSING 4 /* being closed */
612 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
614 typedef struct kib_peer
616 cfs_list_t ibp_list; /* stash on global peer list */
617 lnet_nid_t ibp_nid; /* who's on the other end(s) */
618 lnet_ni_t *ibp_ni; /* LNet interface */
619 cfs_atomic_t ibp_refcount; /* # users */
620 cfs_list_t ibp_conns; /* all active connections */
621 cfs_list_t ibp_tx_queue; /* msgs waiting for a conn */
622 __u16 ibp_version; /* version of peer */
623 __u64 ibp_incarnation; /* incarnation of peer */
624 int ibp_connecting; /* current active connection attempts */
625 int ibp_accepting; /* current passive connection attempts */
626 int ibp_error; /* errno on closing this peer */
627 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
630 extern kib_data_t kiblnd_data;
632 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
635 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
637 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
638 cfs_atomic_inc(&hdev->ibh_ref);
642 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
644 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
645 if (cfs_atomic_dec_and_test(&hdev->ibh_ref))
646 kiblnd_hdev_destroy(hdev);
650 kiblnd_dev_can_failover(kib_dev_t *dev)
652 if (!cfs_list_empty(&dev->ibd_fail_list)) /* already scheduled */
655 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
658 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
661 return dev->ibd_can_failover;
664 #define kiblnd_conn_addref(conn) \
666 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
667 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
668 cfs_atomic_inc(&(conn)->ibc_refcount); \
671 #define kiblnd_conn_decref(conn) \
673 unsigned long flags; \
675 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
676 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
677 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
678 if (cfs_atomic_dec_and_test(&(conn)->ibc_refcount)) { \
679 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
680 cfs_list_add_tail(&(conn)->ibc_list, \
681 &kiblnd_data.kib_connd_zombies); \
682 cfs_waitq_signal(&kiblnd_data.kib_connd_waitq); \
683 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
687 #define kiblnd_peer_addref(peer) \
689 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
690 (peer), libcfs_nid2str((peer)->ibp_nid), \
691 cfs_atomic_read (&(peer)->ibp_refcount)); \
692 cfs_atomic_inc(&(peer)->ibp_refcount); \
695 #define kiblnd_peer_decref(peer) \
697 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
698 (peer), libcfs_nid2str((peer)->ibp_nid), \
699 cfs_atomic_read (&(peer)->ibp_refcount)); \
700 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
701 if (cfs_atomic_dec_and_test(&(peer)->ibp_refcount)) \
702 kiblnd_destroy_peer(peer); \
705 static inline cfs_list_t *
706 kiblnd_nid2peerlist (lnet_nid_t nid)
709 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
711 return (&kiblnd_data.kib_peers [hash]);
715 kiblnd_peer_active (kib_peer_t *peer)
717 /* Am I in the peer hash table? */
718 return (!cfs_list_empty(&peer->ibp_list));
721 static inline kib_conn_t *
722 kiblnd_get_conn_locked (kib_peer_t *peer)
724 LASSERT (!cfs_list_empty(&peer->ibp_conns));
726 /* just return the first connection */
727 return cfs_list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
731 kiblnd_send_keepalive(kib_conn_t *conn)
733 return (*kiblnd_tunables.kib_keepalive > 0) &&
734 cfs_time_after(jiffies, conn->ibc_last_send +
735 *kiblnd_tunables.kib_keepalive*CFS_HZ);
739 kiblnd_need_noop(kib_conn_t *conn)
741 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
743 if (conn->ibc_outstanding_credits <
744 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
745 !kiblnd_send_keepalive(conn))
746 return 0; /* No need to send NOOP */
748 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
749 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred))
750 return 0; /* NOOP can be piggybacked */
752 /* No tx to piggyback NOOP onto or no credit to send a tx */
753 return (cfs_list_empty(&conn->ibc_tx_queue) ||
754 conn->ibc_credits == 0);
757 if (!cfs_list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
758 !cfs_list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
759 conn->ibc_credits == 0) /* no credit */
762 if (conn->ibc_credits == 1 && /* last credit reserved for */
763 conn->ibc_outstanding_credits == 0) /* giving back credits */
766 /* No tx to piggyback NOOP onto or no credit to send a tx */
767 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
771 kiblnd_abort_receives(kib_conn_t *conn)
773 ib_modify_qp(conn->ibc_cmid->qp,
774 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
777 static inline const char *
778 kiblnd_queue2str (kib_conn_t *conn, cfs_list_t *q)
780 if (q == &conn->ibc_tx_queue)
783 if (q == &conn->ibc_tx_queue_rsrvd)
784 return "tx_queue_rsrvd";
786 if (q == &conn->ibc_tx_queue_nocred)
787 return "tx_queue_nocred";
789 if (q == &conn->ibc_active_txs)
796 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
797 * lowest bits of the work request id to stash the work item type. */
799 #define IBLND_WID_TX 0
800 #define IBLND_WID_RDMA 1
801 #define IBLND_WID_RX 2
802 #define IBLND_WID_MASK 3UL
805 kiblnd_ptr2wreqid (void *ptr, int type)
807 unsigned long lptr = (unsigned long)ptr;
809 LASSERT ((lptr & IBLND_WID_MASK) == 0);
810 LASSERT ((type & ~IBLND_WID_MASK) == 0);
811 return (__u64)(lptr | type);
815 kiblnd_wreqid2ptr (__u64 wreqid)
817 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
821 kiblnd_wreqid2type (__u64 wreqid)
823 return (wreqid & IBLND_WID_MASK);
827 kiblnd_set_conn_state (kib_conn_t *conn, int state)
829 conn->ibc_state = state;
834 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
836 msg->ibm_type = type;
837 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
841 kiblnd_rd_size (kib_rdma_desc_t *rd)
846 for (i = size = 0; i < rd->rd_nfrags; i++)
847 size += rd->rd_frags[i].rf_nob;
853 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
855 return rd->rd_frags[index].rf_addr;
859 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
861 return rd->rd_frags[index].rf_nob;
865 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
871 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
873 if (nob < rd->rd_frags[index].rf_nob) {
874 rd->rd_frags[index].rf_addr += nob;
875 rd->rd_frags[index].rf_nob -= nob;
884 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
886 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
887 msgtype == IBLND_MSG_PUT_ACK);
889 return msgtype == IBLND_MSG_GET_REQ ?
890 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
891 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
894 #ifdef HAVE_OFED_IB_DMA_MAP
897 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
899 return ib_dma_mapping_error(dev, dma_addr);
902 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
903 void *msg, size_t size,
904 enum dma_data_direction direction)
906 return ib_dma_map_single(dev, msg, size, direction);
909 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
910 __u64 addr, size_t size,
911 enum dma_data_direction direction)
913 ib_dma_unmap_single(dev, addr, size, direction);
916 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
917 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
919 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
920 struct scatterlist *sg, int nents,
921 enum dma_data_direction direction)
923 return ib_dma_map_sg(dev, sg, nents, direction);
926 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
927 struct scatterlist *sg, int nents,
928 enum dma_data_direction direction)
930 ib_dma_unmap_sg(dev, sg, nents, direction);
933 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
934 struct scatterlist *sg)
936 return ib_sg_dma_address(dev, sg);
939 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
940 struct scatterlist *sg)
942 return ib_sg_dma_len(dev, sg);
945 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
946 * right because OFED1.2 defines it as const, to use it we have to add
947 * (void *) cast to overcome "const" */
949 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
950 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
955 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
957 return dma_mapping_error(dma_addr);
960 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
961 void *msg, size_t size,
962 enum dma_data_direction direction)
964 return dma_map_single(dev->dma_device, msg, size, direction);
967 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
968 dma_addr_t addr, size_t size,
969 enum dma_data_direction direction)
971 dma_unmap_single(dev->dma_device, addr, size, direction);
974 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
975 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
977 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
978 struct scatterlist *sg, int nents,
979 enum dma_data_direction direction)
981 return dma_map_sg(dev->dma_device, sg, nents, direction);
984 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
985 struct scatterlist *sg, int nents,
986 enum dma_data_direction direction)
988 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
992 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
993 struct scatterlist *sg)
995 return sg_dma_address(sg);
999 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1000 struct scatterlist *sg)
1002 return sg_dma_len(sg);
1005 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
1006 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
1010 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1011 kib_rdma_desc_t *rd);
1012 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
1013 __u64 addr, __u64 size);
1014 void kiblnd_map_rx_descs(kib_conn_t *conn);
1015 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1016 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
1017 kib_rdma_desc_t *rd, int nfrags);
1018 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
1019 void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node);
1020 cfs_list_t *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1022 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1023 int npages, __u64 iov, kib_fmr_t *fmr);
1024 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1026 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1027 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1028 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1030 int kiblnd_startup (lnet_ni_t *ni);
1031 void kiblnd_shutdown (lnet_ni_t *ni);
1032 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
1033 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1035 int kiblnd_tunables_init(void);
1036 void kiblnd_tunables_fini(void);
1038 int kiblnd_connd (void *arg);
1039 int kiblnd_scheduler(void *arg);
1040 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
1041 int kiblnd_failover_thread (void *arg);
1043 int kiblnd_alloc_pages (kib_pages_t **pp, int npages);
1044 void kiblnd_free_pages (kib_pages_t *p);
1046 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1047 struct rdma_cm_event *event);
1048 int kiblnd_translate_mtu(int value);
1050 int kiblnd_dev_failover(kib_dev_t *dev);
1051 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1052 void kiblnd_destroy_peer (kib_peer_t *peer);
1053 void kiblnd_destroy_dev (kib_dev_t *dev);
1054 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1055 void kiblnd_peer_alive (kib_peer_t *peer);
1056 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1057 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1058 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1059 int version, __u64 incarnation);
1060 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1062 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1063 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1064 int state, int version);
1065 void kiblnd_destroy_conn (kib_conn_t *conn);
1066 void kiblnd_close_conn (kib_conn_t *conn, int error);
1067 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1069 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1070 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1072 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1073 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1074 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1075 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1076 void kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist,
1078 void kiblnd_check_sends (kib_conn_t *conn);
1080 void kiblnd_qp_event(struct ib_event *event, void *arg);
1081 void kiblnd_cq_event(struct ib_event *event, void *arg);
1082 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1084 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1085 int credits, lnet_nid_t dstnid, __u64 dststamp);
1086 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1087 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1089 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1090 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1091 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1092 unsigned int offset, unsigned int mlen, unsigned int rlen);