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, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lnet/klnds/o2iblnd/o2iblnd.h
38 * Author: Eric Barton <eric@bartonsoftware.com>
41 #include <linux/version.h>
42 #include <linux/module.h>
43 #include <linux/kernel.h>
45 #include <linux/string.h>
46 #include <linux/stat.h>
47 #include <linux/errno.h>
48 #include <linux/unistd.h>
49 #include <linux/uio.h>
51 #include <asm/uaccess.h>
54 #include <linux/init.h>
56 #include <linux/file.h>
57 #include <linux/stat.h>
58 #include <linux/list.h>
59 #include <linux/kmod.h>
60 #include <linux/sysctl.h>
61 #include <linux/pci.h>
62 #if LINUX_VERSION_CODE > KERNEL_VERSION(2,6,32)
63 #include <linux/pci-dma.h>
69 #define DEBUG_SUBSYSTEM S_LND
71 #include <libcfs/libcfs.h>
72 #include <lnet/lnet.h>
73 #include <lnet/lib-lnet.h>
75 #ifdef HAVE_COMPAT_RDMA
76 #include <linux/compat-2.6.h>
78 #include <rdma/rdma_cm.h>
79 #include <rdma/ib_cm.h>
80 #include <rdma/ib_verbs.h>
81 #include <rdma/ib_fmr_pool.h>
83 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
84 /* # scheduler loops before reschedule */
85 #define IBLND_RESCHED 100
87 #define IBLND_N_SCHED 2
88 #define IBLND_N_SCHED_HIGH 4
92 int *kib_dev_failover; /* HCA failover */
93 unsigned int *kib_service; /* IB service number */
94 int *kib_min_reconnect_interval; /* first failed connection retry... */
95 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
96 int *kib_cksum; /* checksum kib_msg_t? */
97 int *kib_timeout; /* comms timeout (seconds) */
98 int *kib_keepalive; /* keepalive timeout (seconds) */
99 int *kib_ntx; /* # tx descs */
100 int *kib_credits; /* # concurrent sends */
101 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
102 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
103 int *kib_peercredits_hiw; /* # when eagerly to return credits */
104 int *kib_peertimeout; /* seconds to consider peer dead */
105 char **kib_default_ipif; /* default IPoIB interface */
106 int *kib_retry_count;
107 int *kib_rnr_retry_count;
108 int *kib_concurrent_sends; /* send work queue sizing */
109 int *kib_ib_mtu; /* IB MTU */
110 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
111 * than this value, 0 disable map-on-demand */
112 int *kib_pmr_pool_size; /* # physical MR in pool */
113 int *kib_fmr_pool_size; /* # FMRs in pool */
114 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
115 int *kib_fmr_cache; /* enable FMR pool cache? */
116 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
117 struct ctl_table_header *kib_sysctl; /* sysctl interface */
119 int *kib_require_priv_port;/* accept only privileged ports */
120 int *kib_use_priv_port; /* use privileged port for active connect */
121 /* # threads on each CPT */
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... */
175 /* Pools (shared by connections on each CPT) */
176 /* These pools can grow at runtime, so don't need give a very large value */
177 #define IBLND_TX_POOL 256
178 #define IBLND_PMR_POOL 256
179 #define IBLND_FMR_POOL 256
180 #define IBLND_FMR_POOL_FLUSH 192
182 /* TX messages (shared by all connections) */
183 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
185 /* RX messages (per connection) */
186 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
187 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
188 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
190 /* WRs and CQEs (per connection) */
191 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
192 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
193 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
197 /* o2iblnd can run over aliased interface */
199 #define KIB_IFNAME_SIZE IFALIASZ
201 #define KIB_IFNAME_SIZE 256
206 cfs_list_t ibd_list; /* chain on kib_devs */
207 cfs_list_t ibd_fail_list; /* chain on kib_failed_devs */
208 __u32 ibd_ifip; /* IPoIB interface IP */
209 /** IPoIB interface name */
210 char ibd_ifname[KIB_IFNAME_SIZE];
211 int ibd_nnets; /* # nets extant */
213 cfs_time_t ibd_next_failover;
214 int ibd_failed_failover; /* # failover failures */
215 unsigned int ibd_failover; /* failover in progress */
216 unsigned int ibd_can_failover; /* IPoIB interface is a bonding master */
218 struct kib_hca_dev *ibd_hdev;
221 typedef struct kib_hca_dev
223 struct rdma_cm_id *ibh_cmid; /* listener cmid */
224 struct ib_device *ibh_ibdev; /* IB device */
225 int ibh_page_shift; /* page shift of current HCA */
226 int ibh_page_size; /* page size of current HCA */
227 __u64 ibh_page_mask; /* page mask of current HCA */
228 int ibh_mr_shift; /* bits shift of max MR size */
229 __u64 ibh_mr_size; /* size of MR */
230 int ibh_nmrs; /* # of global MRs */
231 struct ib_mr **ibh_mrs; /* global MR */
232 struct ib_pd *ibh_pd; /* PD */
233 kib_dev_t *ibh_dev; /* owner */
234 atomic_t ibh_ref; /* refcount */
237 /** # of seconds to keep pool alive */
238 #define IBLND_POOL_DEADLINE 300
239 /** # of seconds to retry if allocation failed */
240 #define IBLND_POOL_RETRY 1
244 int ibp_npages; /* # pages */
245 struct page *ibp_pages[0]; /* page array */
251 cfs_list_t pmr_list; /* chain node */
252 struct ib_phys_buf *pmr_ipb; /* physical buffer */
253 struct ib_mr *pmr_mr; /* IB MR */
254 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
255 __u64 pmr_iova; /* Virtual I/O address */
256 int pmr_refcount; /* reference count */
262 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps,
263 int inc, struct kib_pool **pp_po);
264 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
265 typedef void (*kib_ps_node_init_t)(struct kib_pool *po, cfs_list_t *node);
266 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po, cfs_list_t *node);
270 #define IBLND_POOL_NAME_LEN 32
272 typedef struct kib_poolset
274 spinlock_t ps_lock; /* serialize */
275 struct kib_net *ps_net; /* network it belongs to */
276 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
277 cfs_list_t ps_pool_list; /* list of pools */
278 cfs_list_t ps_failed_pool_list; /* failed pool list */
279 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
280 int ps_increasing; /* is allocating new pool */
281 int ps_pool_size; /* new pool size */
282 int ps_cpt; /* CPT id */
284 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
285 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
286 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
287 kib_ps_node_fini_t ps_node_fini; /* finalize node */
290 typedef struct kib_pool
292 cfs_list_t po_list; /* chain on pool list */
293 cfs_list_t po_free_list; /* pre-allocated node */
294 kib_poolset_t *po_owner; /* pool_set of this pool */
295 cfs_time_t po_deadline; /* deadline of this pool */
296 int po_allocated; /* # of elements in use */
297 int po_failed; /* pool is created on failed HCA */
298 int po_size; /* # of pre-allocated elements */
302 kib_poolset_t tps_poolset; /* pool-set */
303 __u64 tps_next_tx_cookie; /* cookie of TX */
307 kib_pool_t tpo_pool; /* pool */
308 struct kib_hca_dev *tpo_hdev; /* device for this pool */
309 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
310 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
314 kib_poolset_t pps_poolset; /* pool-set */
317 typedef struct kib_pmr_pool {
318 struct kib_hca_dev *ppo_hdev; /* device for this pool */
319 kib_pool_t ppo_pool; /* pool */
324 spinlock_t fps_lock; /* serialize */
325 struct kib_net *fps_net; /* IB network */
326 cfs_list_t fps_pool_list; /* FMR pool list */
327 cfs_list_t fps_failed_pool_list; /* FMR pool list */
328 __u64 fps_version; /* validity stamp */
329 int fps_cpt; /* CPT id */
331 int fps_flush_trigger;
332 /* is allocating new pool */
334 /* time stamp for retry if failed to allocate */
335 cfs_time_t fps_next_retry;
340 cfs_list_t fpo_list; /* chain on pool list */
341 struct kib_hca_dev *fpo_hdev; /* device for this pool */
342 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
343 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
344 cfs_time_t fpo_deadline; /* deadline of this pool */
345 int fpo_failed; /* fmr pool is failed */
346 int fpo_map_count; /* # of mapped FMR */
350 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
351 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
354 typedef struct kib_net
356 cfs_list_t ibn_list; /* chain on kib_dev_t::ibd_nets */
357 __u64 ibn_incarnation; /* my epoch */
358 int ibn_init; /* initialisation state */
359 int ibn_shutdown; /* shutting down? */
361 atomic_t ibn_npeers; /* # peers extant */
362 atomic_t ibn_nconns; /* # connections extant */
364 kib_tx_poolset_t **ibn_tx_ps; /* tx pool-set */
365 kib_fmr_poolset_t **ibn_fmr_ps; /* fmr pool-set */
366 kib_pmr_poolset_t **ibn_pmr_ps; /* pmr pool-set */
368 kib_dev_t *ibn_dev; /* underlying IB device */
371 #define KIB_THREAD_SHIFT 16
372 #define KIB_THREAD_ID(cpt, tid) ((cpt) << KIB_THREAD_SHIFT | (tid))
373 #define KIB_THREAD_CPT(id) ((id) >> KIB_THREAD_SHIFT)
374 #define KIB_THREAD_TID(id) ((id) & ((1UL << KIB_THREAD_SHIFT) - 1))
376 struct kib_sched_info {
379 /* schedulers sleep here */
380 wait_queue_head_t ibs_waitq;
381 /* conns to check for rx completions */
382 cfs_list_t ibs_conns;
383 /* number of scheduler threads */
385 /* max allowed scheduler threads */
386 int ibs_nthreads_max;
387 int ibs_cpt; /* CPT id */
392 int kib_init; /* initialisation state */
393 int kib_shutdown; /* shut down? */
394 cfs_list_t kib_devs; /* IB devices extant */
395 /* list head of failed devices */
396 cfs_list_t kib_failed_devs;
397 /* schedulers sleep here */
398 wait_queue_head_t kib_failover_waitq;
399 atomic_t kib_nthreads; /* # live threads */
400 /* stabilize net/dev/peer/conn ops */
401 rwlock_t kib_global_lock;
402 /* hash table of all my known peers */
403 cfs_list_t *kib_peers;
404 /* size of kib_peers */
405 int kib_peer_hash_size;
406 /* the connd task (serialisation assertions) */
408 /* connections to setup/teardown */
409 cfs_list_t kib_connd_conns;
410 /* connections with zero refcount */
411 cfs_list_t kib_connd_zombies;
412 /* connection daemon sleeps here */
413 wait_queue_head_t kib_connd_waitq;
414 spinlock_t kib_connd_lock; /* serialise */
415 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
416 /* percpt data for schedulers */
417 struct kib_sched_info **kib_scheds;
420 #define IBLND_INIT_NOTHING 0
421 #define IBLND_INIT_DATA 1
422 #define IBLND_INIT_ALL 2
424 /************************************************************************
425 * IB Wire message format.
426 * These are sent in sender's byte order (i.e. receiver flips).
429 typedef struct kib_connparams
431 __u16 ibcp_queue_depth;
432 __u16 ibcp_max_frags;
433 __u32 ibcp_max_msg_size;
434 } WIRE_ATTR kib_connparams_t;
438 lnet_hdr_t ibim_hdr; /* portals header */
439 char ibim_payload[0]; /* piggy-backed payload */
440 } WIRE_ATTR kib_immediate_msg_t;
444 __u32 rf_nob; /* # bytes this frag */
445 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
446 } WIRE_ATTR kib_rdma_frag_t;
450 __u32 rd_key; /* local/remote key */
451 __u32 rd_nfrags; /* # fragments */
452 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
453 } WIRE_ATTR kib_rdma_desc_t;
457 lnet_hdr_t ibprm_hdr; /* portals header */
458 __u64 ibprm_cookie; /* opaque completion cookie */
459 } WIRE_ATTR kib_putreq_msg_t;
463 __u64 ibpam_src_cookie; /* reflected completion cookie */
464 __u64 ibpam_dst_cookie; /* opaque completion cookie */
465 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
466 } WIRE_ATTR kib_putack_msg_t;
470 lnet_hdr_t ibgm_hdr; /* portals header */
471 __u64 ibgm_cookie; /* opaque completion cookie */
472 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
473 } WIRE_ATTR kib_get_msg_t;
477 __u64 ibcm_cookie; /* opaque completion cookie */
478 __s32 ibcm_status; /* < 0 failure: >= 0 length */
479 } WIRE_ATTR kib_completion_msg_t;
483 /* First 2 fields fixed FOR ALL TIME */
484 __u32 ibm_magic; /* I'm an ibnal message */
485 __u16 ibm_version; /* this is my version number */
487 __u8 ibm_type; /* msg type */
488 __u8 ibm_credits; /* returned credits */
489 __u32 ibm_nob; /* # bytes in whole message */
490 __u32 ibm_cksum; /* checksum (0 == no checksum) */
491 __u64 ibm_srcnid; /* sender's NID */
492 __u64 ibm_srcstamp; /* sender's incarnation */
493 __u64 ibm_dstnid; /* destination's NID */
494 __u64 ibm_dststamp; /* destination's incarnation */
497 kib_connparams_t connparams;
498 kib_immediate_msg_t immediate;
499 kib_putreq_msg_t putreq;
500 kib_putack_msg_t putack;
502 kib_completion_msg_t completion;
504 } WIRE_ATTR kib_msg_t;
506 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
508 #define IBLND_MSG_VERSION_1 0x11
509 #define IBLND_MSG_VERSION_2 0x12
510 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
512 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
513 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
514 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
515 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
516 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
517 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
518 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
519 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
520 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
521 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
524 __u32 ibr_magic; /* sender's magic */
525 __u16 ibr_version; /* sender's version */
526 __u8 ibr_why; /* reject reason */
527 __u8 ibr_padding; /* padding */
528 __u64 ibr_incarnation; /* incarnation of peer */
529 kib_connparams_t ibr_cp; /* connection parameters */
530 } WIRE_ATTR kib_rej_t;
532 /* connection rejection reasons */
533 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
534 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
535 #define IBLND_REJECT_FATAL 3 /* Anything else */
537 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
538 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
540 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
541 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
543 /***********************************************************************/
545 typedef struct kib_rx /* receive message */
547 cfs_list_t rx_list; /* queue for attention */
548 struct kib_conn *rx_conn; /* owning conn */
549 int rx_nob; /* # bytes received (-1 while posted) */
550 enum ib_wc_status rx_status; /* completion status */
551 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
552 __u64 rx_msgaddr; /* message buffer (I/O addr) */
553 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
554 struct ib_recv_wr rx_wrq; /* receive work item... */
555 struct ib_sge rx_sge; /* ...and its memory */
558 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
559 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
560 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
561 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
563 typedef struct kib_tx /* transmit message */
565 cfs_list_t tx_list; /* queue on idle_txs ibc_tx_queue etc. */
566 kib_tx_pool_t *tx_pool; /* pool I'm from */
567 struct kib_conn *tx_conn; /* owning conn */
568 short tx_sending; /* # tx callbacks outstanding */
569 short tx_queued; /* queued for sending */
570 short tx_waiting; /* waiting for peer */
571 int tx_status; /* LNET completion status */
572 unsigned long tx_deadline; /* completion deadline */
573 __u64 tx_cookie; /* completion cookie */
574 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
575 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
576 __u64 tx_msgaddr; /* message buffer (I/O addr) */
577 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
578 int tx_nwrq; /* # send work items */
579 struct ib_send_wr *tx_wrq; /* send work items... */
580 struct ib_sge *tx_sge; /* ...and their memory */
581 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
582 int tx_nfrags; /* # entries in... */
583 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
584 __u64 *tx_pages; /* rdma phys page addrs */
586 kib_phys_mr_t *pmr; /* MR for physical buffer */
587 kib_fmr_t fmr; /* FMR */
589 int tx_dmadir; /* dma direction */
592 typedef struct kib_connvars
594 /* connection-in-progress variables */
598 typedef struct kib_conn
600 struct kib_sched_info *ibc_sched; /* scheduler information */
601 struct kib_peer *ibc_peer; /* owning peer */
602 kib_hca_dev_t *ibc_hdev; /* HCA bound on */
603 cfs_list_t ibc_list; /* stash on peer's conn list */
604 cfs_list_t ibc_sched_list; /* schedule for attention */
605 __u16 ibc_version; /* version of connection */
606 __u64 ibc_incarnation; /* which instance of the peer */
607 atomic_t ibc_refcount; /* # users */
608 int ibc_state; /* what's happening */
609 int ibc_nsends_posted; /* # uncompleted sends */
610 int ibc_noops_posted; /* # uncompleted NOOPs */
611 int ibc_credits; /* # credits I have */
612 int ibc_outstanding_credits; /* # credits to return */
613 int ibc_reserved_credits;/* # ACK/DONE msg credits */
614 int ibc_comms_error; /* set on comms error */
615 unsigned int ibc_nrx:16; /* receive buffers owned */
616 unsigned int ibc_scheduled:1; /* scheduled for attention */
617 unsigned int ibc_ready:1; /* CQ callback fired */
618 /* time of last send */
619 unsigned long ibc_last_send;
620 /** link chain for kiblnd_check_conns only */
621 cfs_list_t ibc_connd_list;
622 /** rxs completed before ESTABLISHED */
623 cfs_list_t ibc_early_rxs;
624 /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
625 cfs_list_t ibc_tx_noops;
626 cfs_list_t ibc_tx_queue; /* sends that need a credit */
627 cfs_list_t ibc_tx_queue_nocred;/* sends that don't need a credit */
628 cfs_list_t ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
629 cfs_list_t ibc_active_txs; /* active tx awaiting completion */
630 spinlock_t ibc_lock; /* serialise */
631 kib_rx_t *ibc_rxs; /* the rx descs */
632 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
634 struct rdma_cm_id *ibc_cmid; /* CM id */
635 struct ib_cq *ibc_cq; /* completion queue */
637 kib_connvars_t *ibc_connvars; /* in-progress connection state */
640 #define IBLND_CONN_INIT 0 /* being initialised */
641 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
642 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
643 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
644 #define IBLND_CONN_CLOSING 4 /* being closed */
645 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
647 typedef struct kib_peer
649 cfs_list_t ibp_list; /* stash on global peer list */
650 lnet_nid_t ibp_nid; /* who's on the other end(s) */
651 lnet_ni_t *ibp_ni; /* LNet interface */
652 atomic_t ibp_refcount; /* # users */
653 cfs_list_t ibp_conns; /* all active connections */
654 cfs_list_t ibp_tx_queue; /* msgs waiting for a conn */
655 __u16 ibp_version; /* version of peer */
656 __u64 ibp_incarnation; /* incarnation of peer */
657 int ibp_connecting; /* current active connection attempts */
658 int ibp_accepting; /* current passive connection attempts */
659 int ibp_error; /* errno on closing this peer */
660 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
663 extern kib_data_t kiblnd_data;
665 extern void kiblnd_hdev_destroy(kib_hca_dev_t *hdev);
668 kiblnd_hdev_addref_locked(kib_hca_dev_t *hdev)
670 LASSERT (atomic_read(&hdev->ibh_ref) > 0);
671 atomic_inc(&hdev->ibh_ref);
675 kiblnd_hdev_decref(kib_hca_dev_t *hdev)
677 LASSERT (atomic_read(&hdev->ibh_ref) > 0);
678 if (atomic_dec_and_test(&hdev->ibh_ref))
679 kiblnd_hdev_destroy(hdev);
683 kiblnd_dev_can_failover(kib_dev_t *dev)
685 if (!cfs_list_empty(&dev->ibd_fail_list)) /* already scheduled */
688 if (*kiblnd_tunables.kib_dev_failover == 0) /* disabled */
691 if (*kiblnd_tunables.kib_dev_failover > 1) /* force failover */
694 return dev->ibd_can_failover;
697 #define kiblnd_conn_addref(conn) \
699 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
700 (conn), atomic_read(&(conn)->ibc_refcount)); \
701 atomic_inc(&(conn)->ibc_refcount); \
704 #define kiblnd_conn_decref(conn) \
706 unsigned long flags; \
708 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
709 (conn), atomic_read(&(conn)->ibc_refcount)); \
710 LASSERT_ATOMIC_POS(&(conn)->ibc_refcount); \
711 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
712 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
713 cfs_list_add_tail(&(conn)->ibc_list, \
714 &kiblnd_data.kib_connd_zombies); \
715 wake_up(&kiblnd_data.kib_connd_waitq); \
716 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
720 #define kiblnd_peer_addref(peer) \
722 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
723 (peer), libcfs_nid2str((peer)->ibp_nid), \
724 atomic_read (&(peer)->ibp_refcount)); \
725 atomic_inc(&(peer)->ibp_refcount); \
728 #define kiblnd_peer_decref(peer) \
730 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
731 (peer), libcfs_nid2str((peer)->ibp_nid), \
732 atomic_read (&(peer)->ibp_refcount)); \
733 LASSERT_ATOMIC_POS(&(peer)->ibp_refcount); \
734 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
735 kiblnd_destroy_peer(peer); \
738 static inline cfs_list_t *
739 kiblnd_nid2peerlist (lnet_nid_t nid)
742 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
744 return (&kiblnd_data.kib_peers [hash]);
748 kiblnd_peer_active (kib_peer_t *peer)
750 /* Am I in the peer hash table? */
751 return (!cfs_list_empty(&peer->ibp_list));
754 static inline kib_conn_t *
755 kiblnd_get_conn_locked (kib_peer_t *peer)
757 LASSERT (!cfs_list_empty(&peer->ibp_conns));
759 /* just return the first connection */
760 return cfs_list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
764 kiblnd_send_keepalive(kib_conn_t *conn)
766 return (*kiblnd_tunables.kib_keepalive > 0) &&
767 cfs_time_after(jiffies, conn->ibc_last_send +
768 *kiblnd_tunables.kib_keepalive*HZ);
772 kiblnd_need_noop(kib_conn_t *conn)
774 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
776 if (conn->ibc_outstanding_credits <
777 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
778 !kiblnd_send_keepalive(conn))
779 return 0; /* No need to send NOOP */
781 if (IBLND_OOB_CAPABLE(conn->ibc_version)) {
782 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred))
783 return 0; /* NOOP can be piggybacked */
785 /* No tx to piggyback NOOP onto or no credit to send a tx */
786 return (cfs_list_empty(&conn->ibc_tx_queue) ||
787 conn->ibc_credits == 0);
790 if (!cfs_list_empty(&conn->ibc_tx_noops) || /* NOOP already queued */
791 !cfs_list_empty(&conn->ibc_tx_queue_nocred) || /* piggyback NOOP */
792 conn->ibc_credits == 0) /* no credit */
795 if (conn->ibc_credits == 1 && /* last credit reserved for */
796 conn->ibc_outstanding_credits == 0) /* giving back credits */
799 /* No tx to piggyback NOOP onto or no credit to send a tx */
800 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 1);
804 kiblnd_abort_receives(kib_conn_t *conn)
806 ib_modify_qp(conn->ibc_cmid->qp,
807 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
810 static inline const char *
811 kiblnd_queue2str (kib_conn_t *conn, cfs_list_t *q)
813 if (q == &conn->ibc_tx_queue)
816 if (q == &conn->ibc_tx_queue_rsrvd)
817 return "tx_queue_rsrvd";
819 if (q == &conn->ibc_tx_queue_nocred)
820 return "tx_queue_nocred";
822 if (q == &conn->ibc_active_txs)
829 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
830 * lowest bits of the work request id to stash the work item type. */
832 #define IBLND_WID_TX 0
833 #define IBLND_WID_RDMA 1
834 #define IBLND_WID_RX 2
835 #define IBLND_WID_MASK 3UL
838 kiblnd_ptr2wreqid (void *ptr, int type)
840 unsigned long lptr = (unsigned long)ptr;
842 LASSERT ((lptr & IBLND_WID_MASK) == 0);
843 LASSERT ((type & ~IBLND_WID_MASK) == 0);
844 return (__u64)(lptr | type);
848 kiblnd_wreqid2ptr (__u64 wreqid)
850 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
854 kiblnd_wreqid2type (__u64 wreqid)
856 return (wreqid & IBLND_WID_MASK);
860 kiblnd_set_conn_state (kib_conn_t *conn, int state)
862 conn->ibc_state = state;
867 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
869 msg->ibm_type = type;
870 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
874 kiblnd_rd_size (kib_rdma_desc_t *rd)
879 for (i = size = 0; i < rd->rd_nfrags; i++)
880 size += rd->rd_frags[i].rf_nob;
886 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
888 return rd->rd_frags[index].rf_addr;
892 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
894 return rd->rd_frags[index].rf_nob;
898 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
904 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
906 if (nob < rd->rd_frags[index].rf_nob) {
907 rd->rd_frags[index].rf_addr += nob;
908 rd->rd_frags[index].rf_nob -= nob;
917 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
919 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
920 msgtype == IBLND_MSG_PUT_ACK);
922 return msgtype == IBLND_MSG_GET_REQ ?
923 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
924 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
927 #ifdef HAVE_OFED_IB_DMA_MAP
930 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
932 return ib_dma_mapping_error(dev, dma_addr);
935 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
936 void *msg, size_t size,
937 enum dma_data_direction direction)
939 return ib_dma_map_single(dev, msg, size, direction);
942 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
943 __u64 addr, size_t size,
944 enum dma_data_direction direction)
946 ib_dma_unmap_single(dev, addr, size, direction);
949 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
950 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
952 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
953 struct scatterlist *sg, int nents,
954 enum dma_data_direction direction)
956 return ib_dma_map_sg(dev, sg, nents, direction);
959 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
960 struct scatterlist *sg, int nents,
961 enum dma_data_direction direction)
963 ib_dma_unmap_sg(dev, sg, nents, direction);
966 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
967 struct scatterlist *sg)
969 return ib_sg_dma_address(dev, sg);
972 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
973 struct scatterlist *sg)
975 return ib_sg_dma_len(dev, sg);
978 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
979 * right because OFED1.2 defines it as const, to use it we have to add
980 * (void *) cast to overcome "const" */
982 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
983 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
988 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
990 return dma_mapping_error(dma_addr);
993 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
994 void *msg, size_t size,
995 enum dma_data_direction direction)
997 return dma_map_single(dev->dma_device, msg, size, direction);
1000 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
1001 dma_addr_t addr, size_t size,
1002 enum dma_data_direction direction)
1004 dma_unmap_single(dev->dma_device, addr, size, direction);
1007 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
1008 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
1010 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
1011 struct scatterlist *sg, int nents,
1012 enum dma_data_direction direction)
1014 return dma_map_sg(dev->dma_device, sg, nents, direction);
1017 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
1018 struct scatterlist *sg, int nents,
1019 enum dma_data_direction direction)
1021 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
1025 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
1026 struct scatterlist *sg)
1028 return sg_dma_address(sg);
1032 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
1033 struct scatterlist *sg)
1035 return sg_dma_len(sg);
1038 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
1039 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
1043 struct ib_mr *kiblnd_find_rd_dma_mr(kib_hca_dev_t *hdev,
1044 kib_rdma_desc_t *rd);
1045 struct ib_mr *kiblnd_find_dma_mr(kib_hca_dev_t *hdev,
1046 __u64 addr, __u64 size);
1047 void kiblnd_map_rx_descs(kib_conn_t *conn);
1048 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
1049 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
1050 kib_rdma_desc_t *rd, int nfrags);
1051 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
1052 void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node);
1053 cfs_list_t *kiblnd_pool_alloc_node(kib_poolset_t *ps);
1055 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
1056 int npages, __u64 iov, kib_fmr_t *fmr);
1057 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
1059 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_hca_dev_t *hdev,
1060 kib_rdma_desc_t *rd, __u64 *iova, kib_phys_mr_t **pp_pmr);
1061 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
1063 int kiblnd_startup (lnet_ni_t *ni);
1064 void kiblnd_shutdown (lnet_ni_t *ni);
1065 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
1066 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
1068 int kiblnd_tunables_init(void);
1069 void kiblnd_tunables_fini(void);
1071 int kiblnd_connd (void *arg);
1072 int kiblnd_scheduler(void *arg);
1073 int kiblnd_thread_start(int (*fn)(void *arg), void *arg, char *name);
1074 int kiblnd_failover_thread (void *arg);
1076 int kiblnd_alloc_pages(kib_pages_t **pp, int cpt, int npages);
1077 void kiblnd_free_pages (kib_pages_t *p);
1079 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
1080 struct rdma_cm_event *event);
1081 int kiblnd_translate_mtu(int value);
1083 int kiblnd_dev_failover(kib_dev_t *dev);
1084 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
1085 void kiblnd_destroy_peer (kib_peer_t *peer);
1086 void kiblnd_destroy_dev (kib_dev_t *dev);
1087 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
1088 void kiblnd_peer_alive (kib_peer_t *peer);
1089 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
1090 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
1091 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
1092 int version, __u64 incarnation);
1093 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
1095 void kiblnd_connreq_done(kib_conn_t *conn, int status);
1096 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
1097 int state, int version);
1098 void kiblnd_destroy_conn (kib_conn_t *conn);
1099 void kiblnd_close_conn (kib_conn_t *conn, int error);
1100 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
1102 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
1103 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
1105 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
1106 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
1107 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
1108 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
1109 void kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist,
1111 void kiblnd_check_sends (kib_conn_t *conn);
1113 void kiblnd_qp_event(struct ib_event *event, void *arg);
1114 void kiblnd_cq_event(struct ib_event *event, void *arg);
1115 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
1117 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
1118 int credits, lnet_nid_t dstnid, __u64 dststamp);
1119 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1120 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1122 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1123 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1124 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1125 unsigned int offset, unsigned int mlen, unsigned int rlen);