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 unsigned int *kib_service; /* IB service number */
102 int *kib_min_reconnect_interval; /* first failed connection retry... */
103 int *kib_max_reconnect_interval; /* ...exponentially increasing to this */
104 int *kib_cksum; /* checksum kib_msg_t? */
105 int *kib_timeout; /* comms timeout (seconds) */
106 int *kib_keepalive; /* keepalive timeout (seconds) */
107 int *kib_ntx; /* # tx descs */
108 int *kib_credits; /* # concurrent sends */
109 int *kib_peertxcredits; /* # concurrent sends to 1 peer */
110 int *kib_peerrtrcredits; /* # per-peer router buffer credits */
111 int *kib_peercredits_hiw; /* # when eagerly to return credits */
112 int *kib_peertimeout; /* seconds to consider peer dead */
113 char **kib_default_ipif; /* default IPoIB interface */
114 int *kib_retry_count;
115 int *kib_rnr_retry_count;
116 int *kib_concurrent_sends; /* send work queue sizing */
117 int *kib_ib_mtu; /* IB MTU */
118 int *kib_map_on_demand; /* map-on-demand if RD has more fragments
119 * than this value, 0 disable map-on-demand */
120 int *kib_pmr_pool_size; /* # physical MR in pool */
121 int *kib_fmr_pool_size; /* # FMRs in pool */
122 int *kib_fmr_flush_trigger; /* When to trigger FMR flush */
123 int *kib_fmr_cache; /* enable FMR pool cache? */
124 #if defined(CONFIG_SYSCTL) && !CFS_SYSFS_MODULE_PARM
125 cfs_sysctl_table_header_t *kib_sysctl; /* sysctl interface */
129 extern kib_tunables_t kiblnd_tunables;
131 #define IBLND_MSG_QUEUE_SIZE_V1 8 /* V1 only : # messages/RDMAs in-flight */
132 #define IBLND_CREDIT_HIGHWATER_V1 7 /* V1 only : when eagerly to return credits */
134 #define IBLND_CREDITS_DEFAULT 8 /* default # of peer credits */
135 #define IBLND_CREDITS_MAX ((typeof(((kib_msg_t*) 0)->ibm_credits)) - 1) /* Max # of peer credits */
137 #define IBLND_MSG_QUEUE_SIZE(v) ((v) == IBLND_MSG_VERSION_1 ? \
138 IBLND_MSG_QUEUE_SIZE_V1 : \
139 *kiblnd_tunables.kib_peertxcredits) /* # messages/RDMAs in-flight */
140 #define IBLND_CREDITS_HIGHWATER(v) ((v) == IBLND_MSG_VERSION_1 ? \
141 IBLND_CREDIT_HIGHWATER_V1 : \
142 *kiblnd_tunables.kib_peercredits_hiw) /* when eagerly to return credits */
145 kiblnd_concurrent_sends_v1(void)
147 if (*kiblnd_tunables.kib_concurrent_sends > IBLND_MSG_QUEUE_SIZE_V1 * 2)
148 return IBLND_MSG_QUEUE_SIZE_V1 * 2;
150 if (*kiblnd_tunables.kib_concurrent_sends < IBLND_MSG_QUEUE_SIZE_V1 / 2)
151 return IBLND_MSG_QUEUE_SIZE_V1 / 2;
153 return *kiblnd_tunables.kib_concurrent_sends;
156 #define IBLND_CONCURRENT_SENDS(v) ((v) == IBLND_MSG_VERSION_1 ? \
157 kiblnd_concurrent_sends_v1() : \
158 *kiblnd_tunables.kib_concurrent_sends)
159 /* 2 OOB shall suffice for 1 keepalive and 1 returning credits */
160 #define IBLND_OOB_CAPABLE(v) ((v) != IBLND_MSG_VERSION_1)
161 #define IBLND_OOB_MSGS(v) (IBLND_OOB_CAPABLE(v) ? 2 : 0)
163 #define IBLND_MSG_SIZE (4<<10) /* max size of queued messages (inc hdr) */
164 #define IBLND_MAX_RDMA_FRAGS LNET_MAX_IOV /* max # of fragments supported */
165 #define IBLND_CFG_RDMA_FRAGS (*kiblnd_tunables.kib_map_on_demand != 0 ? \
166 *kiblnd_tunables.kib_map_on_demand : \
167 IBLND_MAX_RDMA_FRAGS) /* max # of fragments configured by user */
168 #define IBLND_RDMA_FRAGS(v) ((v) == IBLND_MSG_VERSION_1 ? \
169 IBLND_MAX_RDMA_FRAGS : IBLND_CFG_RDMA_FRAGS)
171 /************************/
172 /* derived constants... */
174 /* TX messages (shared by all connections) */
175 #define IBLND_TX_MSGS() (*kiblnd_tunables.kib_ntx)
177 /* RX messages (per connection) */
178 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
179 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
180 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
182 /* WRs and CQEs (per connection) */
183 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
184 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
185 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
189 cfs_list_t ibd_list; /* chain on kib_devs */
190 __u32 ibd_ifip; /* IPoIB interface IP */
191 char ibd_ifname[32]; /* IPoIB interface name */
192 int ibd_nnets; /* # nets extant */
194 struct rdma_cm_id *ibd_cmid; /* IB listener (bound to 1 device) */
195 struct ib_pd *ibd_pd; /* PD for the device */
196 int ibd_page_shift; /* page shift of current HCA */
197 int ibd_page_size; /* page size of current HCA */
198 __u64 ibd_page_mask; /* page mask of current HCA */
199 int ibd_mr_shift; /* bits shift of max MR size */
200 __u64 ibd_mr_size; /* size of MR */
202 int ibd_nmrs; /* # of global MRs */
203 struct ib_mr **ibd_mrs; /* MR for non RDMA I/O */
206 #define IBLND_POOL_DEADLINE 300 /* # of seconds to keep pool alive */
210 struct ib_device *ibp_device; /* device for mapping */
211 int ibp_npages; /* # pages */
212 struct page *ibp_pages[0]; /* page array */
218 cfs_list_t pmr_list; /* chain node */
219 struct ib_phys_buf *pmr_ipb; /* physical buffer */
220 struct ib_mr *pmr_mr; /* IB MR */
221 struct kib_pmr_pool *pmr_pool; /* owner of this MR */
222 __u64 pmr_iova; /* Virtual I/O address */
223 int pmr_refcount; /* reference count */
229 typedef int (*kib_ps_pool_create_t)(struct kib_poolset *ps, int inc, struct kib_pool **pp_po);
230 typedef void (*kib_ps_pool_destroy_t)(struct kib_pool *po);
231 typedef void (*kib_ps_node_init_t)(struct kib_pool *po,
233 typedef void (*kib_ps_node_fini_t)(struct kib_pool *po,
238 #define IBLND_POOL_NAME_LEN 32
240 typedef struct kib_poolset
242 cfs_spinlock_t ps_lock; /* serialize */
243 struct kib_net *ps_net; /* network it belongs to */
244 char ps_name[IBLND_POOL_NAME_LEN]; /* pool set name */
245 cfs_list_t ps_pool_list; /* list of pools */
246 cfs_time_t ps_next_retry; /* time stamp for retry if failed to allocate */
247 int ps_increasing; /* is allocating new pool */
248 int ps_pool_size; /* new pool size */
250 kib_ps_pool_create_t ps_pool_create; /* create a new pool */
251 kib_ps_pool_destroy_t ps_pool_destroy; /* destroy a pool */
252 kib_ps_node_init_t ps_node_init; /* initialize new allocated node */
253 kib_ps_node_fini_t ps_node_fini; /* finalize node */
256 typedef struct kib_pool
258 cfs_list_t po_list; /* chain on pool list */
259 cfs_list_t po_free_list; /* pre-allocated node */
260 kib_poolset_t *po_owner; /* pool_set of this pool */
261 cfs_time_t po_deadline; /* deadline of this pool */
262 int po_allocated; /* # of elements in use */
263 int po_size; /* # of pre-allocated elements */
267 kib_poolset_t tps_poolset; /* pool-set */
268 __u64 tps_next_tx_cookie; /* cookie of TX */
272 kib_pool_t tpo_pool; /* pool */
273 struct kib_tx *tpo_tx_descs; /* all the tx descriptors */
274 kib_pages_t *tpo_tx_pages; /* premapped tx msg pages */
278 kib_poolset_t pps_poolset; /* pool-set */
281 typedef struct kib_pmr_pool {
282 kib_pool_t ppo_pool; /* pool */
287 cfs_spinlock_t fps_lock; /* serialize */
288 struct kib_net *fps_net; /* IB network */
289 cfs_list_t fps_pool_list; /* FMR pool list */
290 __u64 fps_version; /* validity stamp */
291 int fps_increasing; /* is allocating new pool */
292 cfs_time_t fps_next_retry; /* time stamp for retry if failed to allocate */
297 cfs_list_t fpo_list; /* chain on pool list */
298 kib_fmr_poolset_t *fpo_owner; /* owner of this pool */
299 struct ib_fmr_pool *fpo_fmr_pool; /* IB FMR pool */
300 cfs_time_t fpo_deadline; /* deadline of this pool */
301 int fpo_map_count; /* # of mapped FMR */
305 struct ib_pool_fmr *fmr_pfmr; /* IB pool fmr */
306 kib_fmr_pool_t *fmr_pool; /* pool of FMR */
309 typedef struct kib_net
311 __u64 ibn_incarnation; /* my epoch */
312 int ibn_init; /* initialisation state */
313 int ibn_shutdown; /* shutting down? */
314 unsigned int ibn_with_fmr:1; /* FMR? */
315 unsigned int ibn_with_pmr:1; /* PMR? */
317 cfs_atomic_t ibn_npeers; /* # peers extant */
318 cfs_atomic_t ibn_nconns; /* # connections extant */
320 kib_tx_poolset_t ibn_tx_ps; /* tx pool-set */
321 kib_fmr_poolset_t ibn_fmr_ps; /* fmr pool-set */
322 kib_pmr_poolset_t ibn_pmr_ps; /* pmr pool-set */
324 kib_dev_t *ibn_dev; /* underlying IB device */
329 int kib_init; /* initialisation state */
330 int kib_shutdown; /* shut down? */
331 cfs_list_t kib_devs; /* IB devices extant */
332 cfs_atomic_t kib_nthreads; /* # live threads */
333 cfs_rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
335 cfs_list_t *kib_peers; /* hash table of all my known peers */
336 int kib_peer_hash_size;/* size of kib_peers */
338 void *kib_connd; /* the connd task (serialisation assertions) */
339 cfs_list_t kib_connd_conns; /* connections to setup/teardown */
340 cfs_list_t kib_connd_zombies;/* connections with zero refcount */
341 cfs_waitq_t kib_connd_waitq; /* connection daemon sleeps here */
342 cfs_spinlock_t kib_connd_lock; /* serialise */
344 cfs_waitq_t kib_sched_waitq; /* schedulers sleep here */
345 cfs_list_t kib_sched_conns; /* conns to check for rx completions */
346 cfs_spinlock_t kib_sched_lock; /* serialise */
348 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
351 #define IBLND_INIT_NOTHING 0
352 #define IBLND_INIT_DATA 1
353 #define IBLND_INIT_ALL 2
355 /************************************************************************
356 * IB Wire message format.
357 * These are sent in sender's byte order (i.e. receiver flips).
360 typedef struct kib_connparams
362 __u16 ibcp_queue_depth;
363 __u16 ibcp_max_frags;
364 __u32 ibcp_max_msg_size;
365 } WIRE_ATTR kib_connparams_t;
369 lnet_hdr_t ibim_hdr; /* portals header */
370 char ibim_payload[0]; /* piggy-backed payload */
371 } WIRE_ATTR kib_immediate_msg_t;
375 __u32 rf_nob; /* # bytes this frag */
376 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
377 } WIRE_ATTR kib_rdma_frag_t;
381 __u32 rd_key; /* local/remote key */
382 __u32 rd_nfrags; /* # fragments */
383 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
384 } WIRE_ATTR kib_rdma_desc_t;
388 lnet_hdr_t ibprm_hdr; /* portals header */
389 __u64 ibprm_cookie; /* opaque completion cookie */
390 } WIRE_ATTR kib_putreq_msg_t;
394 __u64 ibpam_src_cookie; /* reflected completion cookie */
395 __u64 ibpam_dst_cookie; /* opaque completion cookie */
396 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
397 } WIRE_ATTR kib_putack_msg_t;
401 lnet_hdr_t ibgm_hdr; /* portals header */
402 __u64 ibgm_cookie; /* opaque completion cookie */
403 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
404 } WIRE_ATTR kib_get_msg_t;
408 __u64 ibcm_cookie; /* opaque completion cookie */
409 __s32 ibcm_status; /* < 0 failure: >= 0 length */
410 } WIRE_ATTR kib_completion_msg_t;
414 /* First 2 fields fixed FOR ALL TIME */
415 __u32 ibm_magic; /* I'm an openibnal message */
416 __u16 ibm_version; /* this is my version number */
418 __u8 ibm_type; /* msg type */
419 __u8 ibm_credits; /* returned credits */
420 __u32 ibm_nob; /* # bytes in whole message */
421 __u32 ibm_cksum; /* checksum (0 == no checksum) */
422 __u64 ibm_srcnid; /* sender's NID */
423 __u64 ibm_srcstamp; /* sender's incarnation */
424 __u64 ibm_dstnid; /* destination's NID */
425 __u64 ibm_dststamp; /* destination's incarnation */
428 kib_connparams_t connparams;
429 kib_immediate_msg_t immediate;
430 kib_putreq_msg_t putreq;
431 kib_putack_msg_t putack;
433 kib_completion_msg_t completion;
435 } WIRE_ATTR kib_msg_t;
437 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
439 #define IBLND_MSG_VERSION_1 0x11
440 #define IBLND_MSG_VERSION_2 0x12
441 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
443 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
444 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
445 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
446 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
447 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
448 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
449 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
450 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
451 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
452 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
455 __u32 ibr_magic; /* sender's magic */
456 __u16 ibr_version; /* sender's version */
457 __u8 ibr_why; /* reject reason */
458 __u8 ibr_padding; /* padding */
459 __u64 ibr_incarnation; /* incarnation of peer */
460 kib_connparams_t ibr_cp; /* connection parameters */
461 } WIRE_ATTR kib_rej_t;
463 /* connection rejection reasons */
464 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
465 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
466 #define IBLND_REJECT_FATAL 3 /* Anything else */
468 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
469 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
471 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
472 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
474 /***********************************************************************/
476 typedef struct kib_rx /* receive message */
478 cfs_list_t rx_list; /* queue for attention */
479 struct kib_conn *rx_conn; /* owning conn */
480 int rx_nob; /* # bytes received (-1 while posted) */
481 enum ib_wc_status rx_status; /* completion status */
482 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
483 __u64 rx_msgaddr; /* message buffer (I/O addr) */
484 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
485 struct ib_recv_wr rx_wrq; /* receive work item... */
486 struct ib_sge rx_sge; /* ...and its memory */
489 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
490 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
491 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
492 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
494 typedef struct kib_tx /* transmit message */
496 cfs_list_t tx_list; /* queue on idle_txs ibc_tx_queue etc. */
497 kib_tx_pool_t *tx_pool; /* pool I'm from */
498 struct kib_conn *tx_conn; /* owning conn */
499 short tx_sending; /* # tx callbacks outstanding */
500 short tx_queued; /* queued for sending */
501 short tx_waiting; /* waiting for peer */
502 int tx_status; /* LNET completion status */
503 unsigned long tx_deadline; /* completion deadline */
504 __u64 tx_cookie; /* completion cookie */
505 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
506 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
507 __u64 tx_msgaddr; /* message buffer (I/O addr) */
508 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
509 int tx_nwrq; /* # send work items */
510 struct ib_send_wr *tx_wrq; /* send work items... */
511 struct ib_sge *tx_sge; /* ...and their memory */
512 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
513 int tx_nfrags; /* # entries in... */
514 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
515 __u64 *tx_pages; /* rdma phys page addrs */
517 kib_phys_mr_t *pmr; /* MR for physical buffer */
518 kib_fmr_t fmr; /* FMR */
520 int tx_dmadir; /* dma direction */
523 typedef struct kib_connvars
525 /* connection-in-progress variables */
529 typedef struct kib_conn
531 struct kib_peer *ibc_peer; /* owning peer */
532 cfs_list_t ibc_list; /* stash on peer's conn list */
533 cfs_list_t ibc_sched_list; /* schedule for attention */
534 __u16 ibc_version; /* version of connection */
535 __u64 ibc_incarnation; /* which instance of the peer */
536 cfs_atomic_t ibc_refcount; /* # users */
537 int ibc_state; /* what's happening */
538 int ibc_nsends_posted; /* # uncompleted sends */
539 int ibc_noops_posted; /* # uncompleted NOOPs */
540 int ibc_credits; /* # credits I have */
541 int ibc_outstanding_credits; /* # credits to return */
542 int ibc_reserved_credits;/* # ACK/DONE msg credits */
543 int ibc_retry_noop; /* need to retry returning credits */
544 int ibc_comms_error; /* set on comms error */
545 int ibc_nrx:16; /* receive buffers owned */
546 int ibc_scheduled:1; /* scheduled for attention */
547 int ibc_ready:1; /* CQ callback fired */
548 unsigned long ibc_last_send; /* time of last send */
549 cfs_list_t ibc_early_rxs; /* rxs completed before ESTABLISHED */
550 cfs_list_t ibc_tx_queue; /* sends that need a credit */
551 cfs_list_t ibc_tx_queue_nocred;/* sends that don't need a credit */
552 cfs_list_t ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
553 cfs_list_t ibc_active_txs; /* active tx awaiting completion */
554 cfs_spinlock_t ibc_lock; /* serialise */
555 kib_rx_t *ibc_rxs; /* the rx descs */
556 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
558 struct rdma_cm_id *ibc_cmid; /* CM id */
559 struct ib_cq *ibc_cq; /* completion queue */
561 kib_connvars_t *ibc_connvars; /* in-progress connection state */
564 #define IBLND_CONN_INIT 0 /* being intialised */
565 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
566 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
567 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
568 #define IBLND_CONN_CLOSING 4 /* being closed */
569 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
571 typedef struct kib_peer
573 cfs_list_t ibp_list; /* stash on global peer list */
574 lnet_nid_t ibp_nid; /* who's on the other end(s) */
575 lnet_ni_t *ibp_ni; /* LNet interface */
576 cfs_atomic_t ibp_refcount; /* # users */
577 cfs_list_t ibp_conns; /* all active connections */
578 cfs_list_t ibp_tx_queue; /* msgs waiting for a conn */
579 __u16 ibp_version; /* version of peer */
580 __u64 ibp_incarnation; /* incarnation of peer */
581 int ibp_connecting; /* current active connection attempts */
582 int ibp_accepting; /* current passive connection attempts */
583 int ibp_error; /* errno on closing this peer */
584 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
587 extern kib_data_t kiblnd_data;
589 #define kiblnd_conn_addref(conn) \
591 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
592 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
593 LASSERT(cfs_atomic_read(&(conn)->ibc_refcount) > 0); \
594 cfs_atomic_inc(&(conn)->ibc_refcount); \
597 #define kiblnd_conn_decref(conn) \
599 unsigned long flags; \
601 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
602 (conn), cfs_atomic_read(&(conn)->ibc_refcount)); \
603 LASSERT(cfs_atomic_read(&(conn)->ibc_refcount) > 0); \
604 if (cfs_atomic_dec_and_test(&(conn)->ibc_refcount)) { \
605 cfs_spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
606 cfs_list_add_tail(&(conn)->ibc_list, \
607 &kiblnd_data.kib_connd_zombies); \
608 cfs_waitq_signal(&kiblnd_data.kib_connd_waitq); \
609 cfs_spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags);\
613 #define kiblnd_peer_addref(peer) \
615 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
616 (peer), libcfs_nid2str((peer)->ibp_nid), \
617 cfs_atomic_read (&(peer)->ibp_refcount)); \
618 LASSERT(cfs_atomic_read(&(peer)->ibp_refcount) > 0); \
619 cfs_atomic_inc(&(peer)->ibp_refcount); \
622 #define kiblnd_peer_decref(peer) \
624 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
625 (peer), libcfs_nid2str((peer)->ibp_nid), \
626 cfs_atomic_read (&(peer)->ibp_refcount)); \
627 LASSERT(cfs_atomic_read(&(peer)->ibp_refcount) > 0); \
628 if (cfs_atomic_dec_and_test(&(peer)->ibp_refcount)) \
629 kiblnd_destroy_peer(peer); \
632 static inline cfs_list_t *
633 kiblnd_nid2peerlist (lnet_nid_t nid)
636 ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
638 return (&kiblnd_data.kib_peers [hash]);
642 kiblnd_peer_active (kib_peer_t *peer)
644 /* Am I in the peer hash table? */
645 return (!cfs_list_empty(&peer->ibp_list));
648 static inline kib_conn_t *
649 kiblnd_get_conn_locked (kib_peer_t *peer)
651 LASSERT (!cfs_list_empty(&peer->ibp_conns));
653 /* just return the first connection */
654 return cfs_list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
658 kiblnd_send_keepalive(kib_conn_t *conn)
660 return (*kiblnd_tunables.kib_keepalive > 0) &&
661 cfs_time_after(jiffies, conn->ibc_last_send +
662 *kiblnd_tunables.kib_keepalive*CFS_HZ);
666 kiblnd_send_noop(kib_conn_t *conn)
668 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
670 if (conn->ibc_outstanding_credits <
671 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
672 !kiblnd_send_keepalive(conn))
673 return 0; /* No need to send NOOP */
675 if (!cfs_list_empty(&conn->ibc_tx_queue_nocred))
676 return 0; /* NOOP can be piggybacked */
678 if (!IBLND_OOB_CAPABLE(conn->ibc_version))
679 /* can't piggyback? */
680 return cfs_list_empty(&conn->ibc_tx_queue);
682 /* No tx to piggyback NOOP onto or no credit to send a tx */
683 return (cfs_list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 0);
687 kiblnd_abort_receives(kib_conn_t *conn)
689 ib_modify_qp(conn->ibc_cmid->qp,
690 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
693 static inline const char *
694 kiblnd_queue2str (kib_conn_t *conn, cfs_list_t *q)
696 if (q == &conn->ibc_tx_queue)
699 if (q == &conn->ibc_tx_queue_rsrvd)
700 return "tx_queue_rsrvd";
702 if (q == &conn->ibc_tx_queue_nocred)
703 return "tx_queue_nocred";
705 if (q == &conn->ibc_active_txs)
712 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
713 * lowest bits of the work request id to stash the work item type. */
715 #define IBLND_WID_TX 0
716 #define IBLND_WID_RDMA 1
717 #define IBLND_WID_RX 2
718 #define IBLND_WID_MASK 3UL
721 kiblnd_ptr2wreqid (void *ptr, int type)
723 unsigned long lptr = (unsigned long)ptr;
725 LASSERT ((lptr & IBLND_WID_MASK) == 0);
726 LASSERT ((type & ~IBLND_WID_MASK) == 0);
727 return (__u64)(lptr | type);
731 kiblnd_wreqid2ptr (__u64 wreqid)
733 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
737 kiblnd_wreqid2type (__u64 wreqid)
739 return (wreqid & IBLND_WID_MASK);
743 kiblnd_set_conn_state (kib_conn_t *conn, int state)
745 conn->ibc_state = state;
750 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
752 msg->ibm_type = type;
753 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
757 kiblnd_rd_size (kib_rdma_desc_t *rd)
762 for (i = size = 0; i < rd->rd_nfrags; i++)
763 size += rd->rd_frags[i].rf_nob;
769 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
771 return rd->rd_frags[index].rf_addr;
775 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
777 return rd->rd_frags[index].rf_nob;
781 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
787 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
789 if (nob < rd->rd_frags[index].rf_nob) {
790 rd->rd_frags[index].rf_addr += nob;
791 rd->rd_frags[index].rf_nob -= nob;
800 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
802 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
803 msgtype == IBLND_MSG_PUT_ACK);
805 return msgtype == IBLND_MSG_GET_REQ ?
806 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
807 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
810 #ifdef HAVE_OFED_IB_DMA_MAP
813 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
815 return ib_dma_mapping_error(dev, dma_addr);
818 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
819 void *msg, size_t size,
820 enum dma_data_direction direction)
822 return ib_dma_map_single(dev, msg, size, direction);
825 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
826 __u64 addr, size_t size,
827 enum dma_data_direction direction)
829 ib_dma_unmap_single(dev, addr, size, direction);
832 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
833 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
835 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
836 struct scatterlist *sg, int nents,
837 enum dma_data_direction direction)
839 return ib_dma_map_sg(dev, sg, nents, direction);
842 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
843 struct scatterlist *sg, int nents,
844 enum dma_data_direction direction)
846 ib_dma_unmap_sg(dev, sg, nents, direction);
849 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
850 struct scatterlist *sg)
852 return ib_sg_dma_address(dev, sg);
855 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
856 struct scatterlist *sg)
858 return ib_sg_dma_len(dev, sg);
861 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
862 * right because OFED1.2 defines it as const, to use it we have to add
863 * (void *) cast to overcome "const" */
865 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
866 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
871 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
873 return dma_mapping_error(dma_addr);
876 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
877 void *msg, size_t size,
878 enum dma_data_direction direction)
880 return dma_map_single(dev->dma_device, msg, size, direction);
883 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
884 dma_addr_t addr, size_t size,
885 enum dma_data_direction direction)
887 dma_unmap_single(dev->dma_device, addr, size, direction);
890 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
891 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
893 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
894 struct scatterlist *sg, int nents,
895 enum dma_data_direction direction)
897 return dma_map_sg(dev->dma_device, sg, nents, direction);
900 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
901 struct scatterlist *sg, int nents,
902 enum dma_data_direction direction)
904 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
908 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
909 struct scatterlist *sg)
911 return sg_dma_address(sg);
915 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
916 struct scatterlist *sg)
918 return sg_dma_len(sg);
921 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
922 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
926 struct ib_mr *kiblnd_find_rd_dma_mr(kib_net_t *net,
927 kib_rdma_desc_t *rd);
928 struct ib_mr *kiblnd_find_dma_mr(kib_net_t *net,
929 __u64 addr, __u64 size);
930 void kiblnd_map_rx_descs(kib_conn_t *conn);
931 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
932 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
933 kib_rdma_desc_t *rd, int nfrags);
934 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
935 void kiblnd_pool_free_node(kib_pool_t *pool, cfs_list_t *node);
936 cfs_list_t *kiblnd_pool_alloc_node(kib_poolset_t *ps);
938 int kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, __u64 *pages,
939 int npages, __u64 iov, kib_fmr_t *fmr);
940 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
942 int kiblnd_pmr_pool_map(kib_pmr_poolset_t *pps, kib_rdma_desc_t *rd,
943 __u64 *iova, kib_phys_mr_t **pp_pmr);
944 void kiblnd_pmr_pool_unmap(kib_phys_mr_t *pmr);
946 int kiblnd_startup (lnet_ni_t *ni);
947 void kiblnd_shutdown (lnet_ni_t *ni);
948 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
949 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when);
951 int kiblnd_tunables_init(void);
952 void kiblnd_tunables_fini(void);
954 int kiblnd_connd (void *arg);
955 int kiblnd_scheduler(void *arg);
956 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
958 int kiblnd_alloc_pages (kib_pages_t **pp, int npages);
959 void kiblnd_free_pages (kib_pages_t *p);
961 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
962 struct rdma_cm_event *event);
963 int kiblnd_translate_mtu(int value);
965 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
966 void kiblnd_destroy_peer (kib_peer_t *peer);
967 void kiblnd_destroy_dev (kib_dev_t *dev);
968 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
969 void kiblnd_peer_alive (kib_peer_t *peer);
970 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
971 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
972 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
973 int version, __u64 incarnation);
974 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
976 void kiblnd_connreq_done(kib_conn_t *conn, int status);
977 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
978 int state, int version);
979 void kiblnd_destroy_conn (kib_conn_t *conn);
980 void kiblnd_close_conn (kib_conn_t *conn, int error);
981 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
983 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
984 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
986 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
987 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
988 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
989 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
990 void kiblnd_txlist_done (lnet_ni_t *ni, cfs_list_t *txlist,
992 void kiblnd_check_sends (kib_conn_t *conn);
994 void kiblnd_qp_event(struct ib_event *event, void *arg);
995 void kiblnd_cq_event(struct ib_event *event, void *arg);
996 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
998 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
999 int credits, lnet_nid_t dstnid, __u64 dststamp);
1000 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
1001 int kiblnd_post_rx (kib_rx_t *rx, int credit);
1003 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
1004 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
1005 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
1006 unsigned int offset, unsigned int mlen, unsigned int rlen);