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.
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>
42 # define EXPORT_SYMTAB
44 #ifndef AUTOCONF_INCLUDED
45 #include <linux/config.h>
47 #include <linux/module.h>
48 #include <linux/kernel.h>
50 #include <linux/string.h>
51 #include <linux/stat.h>
52 #include <linux/errno.h>
53 #include <linux/smp_lock.h>
54 #include <linux/unistd.h>
55 #include <linux/uio.h>
57 #include <asm/system.h>
58 #include <asm/uaccess.h>
61 #include <linux/init.h>
63 #include <linux/file.h>
64 #include <linux/stat.h>
65 #include <linux/list.h>
66 #include <linux/kmod.h>
67 #include <linux/sysctl.h>
68 #include <linux/pci.h>
73 #define DEBUG_SUBSYSTEM S_LND
75 #include <libcfs/libcfs.h>
76 #include <lnet/lnet.h>
77 #include <lnet/lib-lnet.h>
78 #include <lnet/lnet-sysctl.h>
84 #include <rdma/rdma_cm.h>
85 #include <rdma/ib_cm.h>
86 #include <rdma/ib_verbs.h>
87 #include <rdma/ib_fmr_pool.h>
89 /* tunables fixed at compile time */
91 # define IBLND_N_SCHED cfs_num_online_cpus() /* # schedulers */
93 # define IBLND_N_SCHED 1 /* # schedulers */
96 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
97 #define IBLND_RESCHED 100 /* # scheduler loops before reschedule */
101 int *kib_dev_failover; /* HCA failover */
102 unsigned int *kib_service; /* IB service number */
103 int *kib_min_reconnect_interval; /* first failed connection retry... */
104 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
105 int *kib_cksum; /* checksum kib_msg_t? */
106 int *kib_timeout; /* comms timeout (seconds) */
107 int *kib_keepalive; /* keepalive timeout (seconds) */
108 int *kib_ntx; /* # tx descs */
109 int *kib_credits; /* # concurrent sends */
110 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
111 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
112 int *kib_peercredits_hiw; /* # when eagerly to return credits */
113 int *kib_peertimeout; /* seconds to consider peer dead */
114 char **kib_default_ipif; /* default IPoIB interface */
115 int *kib_retry_count;
116 int *kib_rnr_retry_count;
117 int *kib_concurrent_sends; /* send work queue sizing */
118 int *kib_ib_mtu; /* IB MTU */
119 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
120 * than this value, 0 disable map-on-demand */
121 int *kib_pmr_pool_size; /* # physical MR in pool */
122 int *kib_fmr_pool_size; /* # FMRs in pool */
123 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
124 int *kib_fmr_cache; /* enable FMR pool cache? */
125 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
126 cfs_sysctl_table_header_t *kib_sysctl; /* sysctl interface */
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 */
146 kiblnd_concurrent_sends_v1(void)
148 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
149 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
151 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
152 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
154 return *kiblnd_tunables.kib_concurrent_sends;
157 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
158 kiblnd_concurrent_sends_v1() : \
159 *kiblnd_tunables.kib_concurrent_sends)
160 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
161 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
162 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
164 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
165 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
166 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
167 *kiblnd_tunables.kib_map_on_demand : \
168 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
169 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
170 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
172 /************************/
173 /* derived constants... */
175 /* TX messages (shared by all connections) */
176 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
178 /* RX messages (per connection) */
179 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
180 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
181 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
183 /* WRs and CQEs (per connection) */
184 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
185 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
186 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
192 cfs_list_t ibd_list; /* chain on kib_devs */
193 cfs_list_t ibd_fail_list; /* chain on kib_failed_devs */
194 __u32 ibd_ifip; /* IPoIB interface IP */
195 char ibd_ifname[32]; /* IPoIB interface name */
196 int ibd_nnets; /* # nets extant */
198 cfs_time_t ibd_next_failover;
199 int ibd_failed_failover; /* # failover failures */
200 unsigned int ibd_failover; /* failover in progress */
201 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
203 struct kib_hca_dev *ibd_hdev;
206 typedef struct kib_hca_dev
208 struct rdma_cm_id *ibh_cmid; /* listener cmid */
209 struct ib_device *ibh_ibdev; /* IB device */
210 int ibh_page_shift; /* page shift of current HCA */
211 int ibh_page_size; /* page size of current HCA */
212 __u64 ibh_page_mask; /* page mask of current HCA */
213 int ibh_mr_shift; /* bits shift of max MR size */
214 __u64 ibh_mr_size; /* size of MR */
215 int ibh_nmrs; /* # of global MRs */
216 struct ib_mr **ibh_mrs; /* global MR */
217 struct ib_pd *ibh_pd; /* PD */
218 kib_dev_t *ibh_dev; /* owner */
219 cfs_atomic_t ibh_ref; /* refcount */
222 #define IBLND_POOL_DEADLINE 300 /* # of seconds to keep pool alive */
226 int ibp_npages; /* # pages */
227 struct page *ibp_pages[0]; /* page array */
233 cfs_list_t 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,
248 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po,
253 #define IBLND_POOL_NAME_LEN 32
255 typedef struct kib_poolset
257 cfs_spinlock_t ps_lock; /* serialize */
258 struct kib_net *ps_net; /* network it belongs to */
259 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
260 cfs_list_t ps_pool_list; /* list of pools */
261 cfs_list_t ps_failed_pool_list; /* failed pool list */
262 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
263 int ps_increasing; /* is allocating new pool */
264 int ps_pool_size; /* new pool size */
266 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
267 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
268 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
269 kib_ps_node_fini_t ps_node_fini; /* finalize node */
272 typedef struct kib_pool
274 cfs_list_t po_list; /* chain on pool list */
275 cfs_list_t po_free_list; /* pre-allocated node */
276 kib_poolset_t *po_owner; /* pool_set of this pool */
277 cfs_time_t po_deadline; /* deadline of this pool */
278 int po_allocated; /* # of elements in use */
279 int po_failed; /* pool is created on failed HCA */
280 int po_size; /* # of pre-allocated elements */
284 kib_poolset_t tps_poolset; /* pool-set */
285 __u64 tps_next_tx_cookie; /* cookie of TX */
289 kib_pool_t tpo_pool; /* pool */
290 struct kib_hca_dev *tpo_hdev; /* device for this pool */
291 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
292 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
296 kib_poolset_t pps_poolset; /* pool-set */
299 typedef struct kib_pmr_pool {
300 struct kib_hca_dev *ppo_hdev; /* device for this pool */
301 kib_pool_t ppo_pool; /* pool */
306 cfs_spinlock_t fps_lock; /* serialize */
307 struct kib_net *fps_net; /* IB network */
308 cfs_list_t fps_pool_list; /* FMR pool list */
309 cfs_list_t fps_failed_pool_list; /* FMR pool list */
310 __u64 fps_version; /* validity stamp */
311 int fps_increasing; /* is allocating new pool */
312 cfs_time_t fps_next_retry; /* time stamp for retry if failed to allocate */
317 cfs_list_t fpo_list; /* chain on pool list */
318 struct kib_hca_dev *fpo_hdev; /* device for this pool */
319 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
320 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
321 cfs_time_t fpo_deadline; /* deadline of this pool */
322 int fpo_failed; /* fmr pool is failed */
323 int fpo_map_count; /* # of mapped FMR */
327 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
328 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
331 typedef struct kib_net
333 cfs_list_t ibn_list; /* chain on kib_dev_t::ibd_nets */
334 __u64 ibn_incarnation; /* my epoch */
335 int ibn_init; /* initialisation state */
336 int ibn_shutdown; /* shutting down? */
337 unsigned int ibn_with_fmr:1; /* FMR? */
338 unsigned int ibn_with_pmr:1; /* PMR? */
340 cfs_atomic_t ibn_npeers; /* # peers extant */
341 cfs_atomic_t ibn_nconns; /* # connections extant */
343 kib_tx_poolset_t ibn_tx_ps; /* tx pool-set */
344 kib_fmr_poolset_t ibn_fmr_ps; /* fmr pool-set */
345 kib_pmr_poolset_t ibn_pmr_ps; /* pmr pool-set */
347 kib_dev_t *ibn_dev; /* underlying IB device */
352 int kib_init; /* initialisation state */
353 int kib_shutdown; /* shut down? */
354 cfs_list_t kib_devs; /* IB devices extant */
355 cfs_list_t kib_failed_devs; /* list head of failed devices */
356 cfs_atomic_t kib_nthreads; /* # live threads */
357 cfs_rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
359 cfs_list_t *kib_peers; /* hash table of all my known peers */
360 int kib_peer_hash_size;/* size of kib_peers */
362 void *kib_connd; /* the connd task (serialisation assertions) */
363 cfs_list_t kib_connd_conns; /* connections to setup/teardown */
364 cfs_list_t kib_connd_zombies;/* connections with zero refcount */
365 cfs_waitq_t kib_connd_waitq; /* connection daemon sleeps here */
366 cfs_spinlock_t kib_connd_lock; /* serialise */
368 cfs_waitq_t kib_sched_waitq; /* schedulers sleep here */
369 cfs_list_t kib_sched_conns; /* conns to check for rx completions */
370 cfs_spinlock_t kib_sched_lock; /* serialise */
371 cfs_waitq_t kib_failover_waitq; /* schedulers sleep here */
373 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
376 #define IBLND_INIT_NOTHING 0
377 #define IBLND_INIT_DATA 1
378 #define IBLND_INIT_ALL 2
380 /************************************************************************
381 * IB Wire message format.
382 * These are sent in sender's byte order (i.e. receiver flips).
385 typedef struct kib_connparams
387 __u16 ibcp_queue_depth;
388 __u16 ibcp_max_frags;
389 __u32 ibcp_max_msg_size;
390 } WIRE_ATTR kib_connparams_t;
394 lnet_hdr_t ibim_hdr; /* portals header */
395 char ibim_payload[0]; /* piggy-backed payload */
396 } WIRE_ATTR kib_immediate_msg_t;
400 __u32 rf_nob; /* # bytes this frag */
401 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
402 } WIRE_ATTR kib_rdma_frag_t;
406 __u32 rd_key; /* local/remote key */
407 __u32 rd_nfrags; /* # fragments */
408 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
409 } WIRE_ATTR kib_rdma_desc_t;
413 lnet_hdr_t ibprm_hdr; /* portals header */
414 __u64 ibprm_cookie; /* opaque completion cookie */
415 } WIRE_ATTR kib_putreq_msg_t;
419 __u64 ibpam_src_cookie; /* reflected completion cookie */
420 __u64 ibpam_dst_cookie; /* opaque completion cookie */
421 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
422 } WIRE_ATTR kib_putack_msg_t;
426 lnet_hdr_t ibgm_hdr; /* portals header */
427 __u64 ibgm_cookie; /* opaque completion cookie */
428 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
429 } WIRE_ATTR kib_get_msg_t;
433 __u64 ibcm_cookie; /* opaque completion cookie */
434 __s32 ibcm_status; /* < 0 failure: >= 0 length */
435 } WIRE_ATTR kib_completion_msg_t;
439 /* First 2 fields fixed FOR ALL TIME */
440 __u32 ibm_magic; /* I'm an ibnal message */
441 __u16 ibm_version; /* this is my version number */
443 __u8 ibm_type; /* msg type */
444 __u8 ibm_credits; /* returned credits */
445 __u32 ibm_nob; /* # bytes in whole message */
446 __u32 ibm_cksum; /* checksum (0 == no checksum) */
447 __u64 ibm_srcnid; /* sender's NID */
448 __u64 ibm_srcstamp; /* sender's incarnation */
449 __u64 ibm_dstnid; /* destination's NID */
450 __u64 ibm_dststamp; /* destination's incarnation */
453 kib_connparams_t connparams;
454 kib_immediate_msg_t immediate;
455 kib_putreq_msg_t putreq;
456 kib_putack_msg_t putack;
458 kib_completion_msg_t completion;
460 } WIRE_ATTR kib_msg_t;
462 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
464 #define IBLND_MSG_VERSION_1 0x11
465 #define IBLND_MSG_VERSION_2 0x12
466 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
468 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
469 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
470 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
471 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
472 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
473 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
474 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
475 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
476 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
477 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
480 __u32 ibr_magic; /* sender's magic */
481 __u16 ibr_version; /* sender's version */
482 __u8 ibr_why; /* reject reason */
483 __u8 ibr_padding; /* padding */
484 __u64 ibr_incarnation; /* incarnation of peer */
485 kib_connparams_t ibr_cp; /* connection parameters */
486 } WIRE_ATTR kib_rej_t;
488 /* connection rejection reasons */
489 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
490 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
491 #define IBLND_REJECT_FATAL 3 /* Anything else */
493 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
494 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
496 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
497 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
499 /***********************************************************************/
501 typedef struct kib_rx /* receive message */
503 cfs_list_t rx_list; /* queue for attention */
504 struct kib_conn *rx_conn; /* owning conn */
505 int rx_nob; /* # bytes received (-1 while posted) */
506 enum ib_wc_status rx_status; /* completion status */
507 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
508 __u64 rx_msgaddr; /* message buffer (I/O addr) */
509 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
510 struct ib_recv_wr rx_wrq; /* receive work item... */
511 struct ib_sge rx_sge; /* ...and its memory */
514 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
515 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
516 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
517 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
519 typedef struct kib_tx /* transmit message */
521 cfs_list_t tx_list; /* queue on idle_txs ibc_tx_queue etc. */
522 kib_tx_pool_t *tx_pool; /* pool I'm from */
523 struct kib_conn *tx_conn; /* owning conn */
524 short tx_sending; /* # tx callbacks outstanding */
525 short tx_queued; /* queued for sending */
526 short tx_waiting; /* waiting for peer */
527 int tx_status; /* LNET completion status */
528 unsigned long tx_deadline; /* completion deadline */
529 __u64 tx_cookie; /* completion cookie */
530 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
531 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
532 __u64 tx_msgaddr; /* message buffer (I/O addr) */
533 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
534 int tx_nwrq; /* # send work items */
535 struct ib_send_wr *tx_wrq; /* send work items... */
536 struct ib_sge *tx_sge; /* ...and their memory */
537 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
538 int tx_nfrags; /* # entries in... */
539 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
540 __u64 *tx_pages; /* rdma phys page addrs */
542 kib_phys_mr_t *pmr; /* MR for physical buffer */
543 kib_fmr_t fmr; /* FMR */
545 int tx_dmadir; /* dma direction */
548 typedef struct kib_connvars
550 /* connection-in-progress variables */
554 typedef struct kib_conn
556 struct kib_peer *ibc_peer; /* owning peer */
557 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
558 cfs_list_t ibc_list; /* stash on peer's conn list */
559 cfs_list_t ibc_sched_list; /* schedule for attention */
560 cfs_list_t ibc_connd_list; /* kiblnd_check_conns only */
561 __u16 ibc_version; /* version of connection */
562 __u64 ibc_incarnation; /* which instance of the peer */
563 cfs_atomic_t ibc_refcount; /* # users */
564 int ibc_state; /* what's happening */
565 int ibc_nsends_posted; /* # uncompleted sends */
566 int ibc_noops_posted; /* # uncompleted NOOPs */
567 int ibc_credits; /* # credits I have */
568 int ibc_outstanding_credits; /* # credits to return */
569 int ibc_reserved_credits;/* # ACK/DONE msg credits */
570 int ibc_retry_noop; /* need to retry returning credits */
571 int ibc_comms_error; /* set on comms error */
572 int ibc_nrx:16; /* receive buffers owned */
573 int ibc_scheduled:1; /* scheduled for attention */
574 int ibc_ready:1; /* CQ callback fired */
575 unsigned long ibc_last_send; /* time of last send */
576 cfs_list_t ibc_early_rxs; /* rxs completed before ESTABLISHED */
577 cfs_list_t ibc_tx_noops; /* IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
578 cfs_list_t ibc_tx_queue; /* sends that need a credit */
579 cfs_list_t ibc_tx_queue_nocred;/* sends that don't need a credit */
580 cfs_list_t ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
581 cfs_list_t ibc_active_txs; /* active tx awaiting completion */
582 cfs_spinlock_t ibc_lock; /* serialise */
583 kib_rx_t *ibc_rxs; /* the rx descs */
584 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
586 struct rdma_cm_id *ibc_cmid; /* CM id */
587 struct ib_cq *ibc_cq; /* completion queue */
589 kib_connvars_t *ibc_connvars; /* in-progress connection state */
592 #define IBLND_CONN_INIT 0 /* being initialised */
593 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
594 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
595 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
596 #define IBLND_CONN_CLOSING 4 /* being closed */
597 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
599 typedef struct kib_peer
601 cfs_list_t ibp_list; /* stash on global peer list */
602 lnet_nid_t ibp_nid; /* who's on the other end(s) */
603 lnet_ni_t *ibp_ni; /* LNet interface */
604 cfs_atomic_t ibp_refcount; /* # users */
605 cfs_list_t ibp_conns; /* all active connections */
606 cfs_list_t ibp_tx_queue; /* msgs waiting for a conn */
607 __u16 ibp_version; /* version of peer */
608 __u64 ibp_incarnation; /* incarnation of peer */
609 int ibp_connecting; /* current active connection attempts */
610 int ibp_accepting; /* current passive connection attempts */
611 int ibp_error; /* errno on closing this peer */
612 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
615 extern kib_data_t kiblnd_data;
617 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
620 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
622 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
623 cfs_atomic_inc(&hdev->ibh_ref);
627 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
629 LASSERT (cfs_atomic_read(&hdev->ibh_ref) > 0);
630 if (cfs_atomic_dec_and_test(&hdev->ibh_ref))
631 kiblnd_hdev_destroy(hdev);
635 kiblnd_dev_can_failover(kib_dev_t *dev)
637 if (!cfs_list_empty(&dev->ibd_fail_list)) /* already scheduled */
640 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
643 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
646 return dev->ibd_can_failover;
649 #define kiblnd_conn_addref(conn) \
651 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
652 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
653 cfs_atomic_inc(&(conn)->ibc_refcount); \
656 #define kiblnd_conn_decref(conn) \
658 unsigned long flags; \
660 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
661 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
662 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
663 if (cfs_atomic_dec_and_test(&(conn)->ibc_refcount)) { \
664 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
665 cfs_list_add_tail(&(conn)->ibc_list, \
666 &kiblnd_data.kib_connd_zombies); \
667 cfs_waitq_signal(&kiblnd_data.kib_connd_waitq); \
668 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
672 #define kiblnd_peer_addref(peer) \
674 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
675 (peer), libcfs_nid2str((peer)->ibp_nid), \
676 cfs_atomic_read (&(peer)->ibp_refcount)); \
677 cfs_atomic_inc(&(peer)->ibp_refcount); \
680 #define kiblnd_peer_decref(peer) \
682 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
683 (peer), libcfs_nid2str((peer)->ibp_nid), \
684 cfs_atomic_read (&(peer)->ibp_refcount)); \
685 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
686 if (cfs_atomic_dec_and_test(&(peer)->ibp_refcount)) \
687 kiblnd_destroy_peer(peer); \
690 static inline cfs_list_t *
691 kiblnd_nid2peerlist (lnet_nid_t nid)
694 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
696 return (&kiblnd_data.kib_peers [hash]);
700 kiblnd_peer_active (kib_peer_t *peer)
702 /* Am I in the peer hash table? */
703 return (!cfs_list_empty(&peer->ibp_list));
706 static inline kib_conn_t *
707 kiblnd_get_conn_locked (kib_peer_t *peer)
709 LASSERT (!cfs_list_empty(&peer->ibp_conns));
711 /* just return the first connection */
712 return cfs_list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
716 kiblnd_send_keepalive(kib_conn_t *conn)
718 return (*kiblnd_tunables.kib_keepalive > 0) &&
719 cfs_time_after(jiffies, conn->ibc_last_send +
720 *kiblnd_tunables.kib_keepalive*CFS_HZ);
724 kiblnd_send_noop(kib_conn_t *conn)
726 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
728 if (conn->ibc_outstanding_credits <
729 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
730 !kiblnd_send_keepalive(conn))
731 return 0; /* No need to send NOOP */
733 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
734 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred))
735 return 0; /* NOOP can be piggybacked */
737 /* No tx to piggyback NOOP onto or no credit to send a tx */
738 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 0);
741 if (!cfs_list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
742 !cfs_list_empty(&conn->ibc_tx_queue_nocred) || /* can be piggybacked */
743 conn->ibc_credits == 0) /* no credit */
746 if (conn->ibc_credits == 1 && /* last credit reserved for */
747 conn->ibc_outstanding_credits == 0) /* giving back credits */
750 /* No tx to piggyback NOOP onto or no credit to send a tx */
751 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
755 kiblnd_abort_receives(kib_conn_t *conn)
757 ib_modify_qp(conn->ibc_cmid->qp,
758 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
761 static inline const char *
762 kiblnd_queue2str (kib_conn_t *conn, cfs_list_t *q)
764 if (q == &conn->ibc_tx_queue)
767 if (q == &conn->ibc_tx_queue_rsrvd)
768 return "tx_queue_rsrvd";
770 if (q == &conn->ibc_tx_queue_nocred)
771 return "tx_queue_nocred";
773 if (q == &conn->ibc_active_txs)
780 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
781 * lowest bits of the work request id to stash the work item type. */
783 #define IBLND_WID_TX 0
784 #define IBLND_WID_RDMA 1
785 #define IBLND_WID_RX 2
786 #define IBLND_WID_MASK 3UL
789 kiblnd_ptr2wreqid (void *ptr, int type)
791 unsigned long lptr = (unsigned long)ptr;
793 LASSERT ((lptr & IBLND_WID_MASK) == 0);
794 LASSERT ((type & ~IBLND_WID_MASK) == 0);
795 return (__u64)(lptr | type);
799 kiblnd_wreqid2ptr (__u64 wreqid)
801 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
805 kiblnd_wreqid2type (__u64 wreqid)
807 return (wreqid & IBLND_WID_MASK);
811 kiblnd_set_conn_state (kib_conn_t *conn, int state)
813 conn->ibc_state = state;
818 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
820 msg->ibm_type = type;
821 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
825 kiblnd_rd_size (kib_rdma_desc_t *rd)
830 for (i = size = 0; i < rd->rd_nfrags; i++)
831 size += rd->rd_frags[i].rf_nob;
837 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
839 return rd->rd_frags[index].rf_addr;
843 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
845 return rd->rd_frags[index].rf_nob;
849 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
855 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
857 if (nob < rd->rd_frags[index].rf_nob) {
858 rd->rd_frags[index].rf_addr += nob;
859 rd->rd_frags[index].rf_nob -= nob;
868 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
870 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
871 msgtype == IBLND_MSG_PUT_ACK);
873 return msgtype == IBLND_MSG_GET_REQ ?
874 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
875 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
878 #ifdef HAVE_OFED_IB_DMA_MAP
881 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
883 return ib_dma_mapping_error(dev, dma_addr);
886 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
887 void *msg, size_t size,
888 enum dma_data_direction direction)
890 return ib_dma_map_single(dev, msg, size, direction);
893 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
894 __u64 addr, size_t size,
895 enum dma_data_direction direction)
897 ib_dma_unmap_single(dev, addr, size, direction);
900 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
901 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
903 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
904 struct scatterlist *sg, int nents,
905 enum dma_data_direction direction)
907 return ib_dma_map_sg(dev, sg, nents, direction);
910 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
911 struct scatterlist *sg, int nents,
912 enum dma_data_direction direction)
914 ib_dma_unmap_sg(dev, sg, nents, direction);
917 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
918 struct scatterlist *sg)
920 return ib_sg_dma_address(dev, sg);
923 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
924 struct scatterlist *sg)
926 return ib_sg_dma_len(dev, sg);
929 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
930 * right because OFED1.2 defines it as const, to use it we have to add
931 * (void *) cast to overcome "const" */
933 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
934 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
939 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
941 return dma_mapping_error(dma_addr);
944 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
945 void *msg, size_t size,
946 enum dma_data_direction direction)
948 return dma_map_single(dev->dma_device, msg, size, direction);
951 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
952 dma_addr_t addr, size_t size,
953 enum dma_data_direction direction)
955 dma_unmap_single(dev->dma_device, addr, size, direction);
958 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
959 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
961 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
962 struct scatterlist *sg, int nents,
963 enum dma_data_direction direction)
965 return dma_map_sg(dev->dma_device, sg, nents, direction);
968 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
969 struct scatterlist *sg, int nents,
970 enum dma_data_direction direction)
972 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
976 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
977 struct scatterlist *sg)
979 return sg_dma_address(sg);
983 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
984 struct scatterlist *sg)
986 return sg_dma_len(sg);
989 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
990 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
994 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
995 kib_rdma_desc_t *rd);
996 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
997 __u64 addr, __u64 size);
998 void kiblnd_map_rx_descs(kib_conn_t *conn);
999 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1000 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
1001 kib_rdma_desc_t *rd, int nfrags);
1002 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
1003 void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node);
1004 cfs_list_t *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1006 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1007 int npages, __u64 iov, kib_fmr_t *fmr);
1008 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1010 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1011 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1012 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1014 int kiblnd_startup (lnet_ni_t *ni);
1015 void kiblnd_shutdown (lnet_ni_t *ni);
1016 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
1017 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1019 int kiblnd_tunables_init(void);
1020 void kiblnd_tunables_fini(void);
1022 int kiblnd_connd (void *arg);
1023 int kiblnd_scheduler(void *arg);
1024 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
1025 int kiblnd_failover_thread (void *arg);
1027 int kiblnd_alloc_pages (kib_pages_t **pp, int npages);
1028 void kiblnd_free_pages (kib_pages_t *p);
1030 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1031 struct rdma_cm_event *event);
1032 int kiblnd_translate_mtu(int value);
1034 int kiblnd_dev_failover(kib_dev_t *dev);
1035 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1036 void kiblnd_destroy_peer (kib_peer_t *peer);
1037 void kiblnd_destroy_dev (kib_dev_t *dev);
1038 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1039 void kiblnd_peer_alive (kib_peer_t *peer);
1040 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1041 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1042 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1043 int version, __u64 incarnation);
1044 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1046 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1047 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1048 int state, int version);
1049 void kiblnd_destroy_conn (kib_conn_t *conn);
1050 void kiblnd_close_conn (kib_conn_t *conn, int error);
1051 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1053 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1054 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1056 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1057 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1058 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1059 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1060 void kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist,
1062 void kiblnd_check_sends (kib_conn_t *conn);
1064 void kiblnd_qp_event(struct ib_event *event, void *arg);
1065 void kiblnd_cq_event(struct ib_event *event, void *arg);
1066 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1068 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1069 int credits, lnet_nid_t dstnid, __u64 dststamp);
1070 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1071 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1073 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1074 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1075 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1076 unsigned int offset, unsigned int mlen, unsigned int rlen);