1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
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/random.h>
69 #include <linux/pci.h>
74 #define DEBUG_SUBSYSTEM S_LND
76 #include <libcfs/libcfs.h>
77 #include <lnet/lnet.h>
78 #include <lnet/lib-lnet.h>
79 #include <lnet/lnet-sysctl.h>
85 #include <rdma/rdma_cm.h>
86 #include <rdma/ib_cm.h>
87 #include <rdma/ib_verbs.h>
88 #include <rdma/ib_fmr_pool.h>
90 /* tunables fixed at compile time */
92 # define IBLND_N_SCHED num_online_cpus() /* # schedulers */
94 # define IBLND_N_SCHED 1 /* # schedulers */
97 #define IBLND_PEER_HASH_SIZE 101 /* # peer lists */
98 #define IBLND_RESCHED 100 /* # scheduler loops before reschedule */
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_peercredits; /* # concurrent sends to 1 peer */
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 4096 /* 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_peercredits) /* # 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)
176 #define IBLND_TX_MSG_BYTES() (IBLND_TX_MSGS() * IBLND_MSG_SIZE)
177 #define IBLND_TX_MSG_PAGES() ((IBLND_TX_MSG_BYTES() + PAGE_SIZE - 1) / PAGE_SIZE)
179 /* RX messages (per connection) */
180 #define IBLND_RX_MSGS(v) (IBLND_MSG_QUEUE_SIZE(v) * 2 + IBLND_OOB_MSGS(v))
181 #define IBLND_RX_MSG_BYTES(v) (IBLND_RX_MSGS(v) * IBLND_MSG_SIZE)
182 #define IBLND_RX_MSG_PAGES(v) ((IBLND_RX_MSG_BYTES(v) + PAGE_SIZE - 1) / PAGE_SIZE)
184 /* WRs and CQEs (per connection) */
185 #define IBLND_RECV_WRS(v) IBLND_RX_MSGS(v)
186 #define IBLND_SEND_WRS(v) ((IBLND_RDMA_FRAGS(v) + 1) * IBLND_CONCURRENT_SENDS(v))
187 #define IBLND_CQ_ENTRIES(v) (IBLND_RECV_WRS(v) + IBLND_SEND_WRS(v))
191 struct ib_device *ibp_device; /* device for mapping */
192 int ibp_npages; /* # pages */
193 struct page *ibp_pages[0];
197 spinlock_t ibmp_lock; /* serialize */
198 int ibmp_allocated; /* MR in use */
199 struct list_head ibmp_free_list; /* pre-allocated MR */
200 } kib_phys_mr_pool_t;
203 struct list_head ibpm_link; /* link node */
204 struct ib_mr *ibpm_mr; /* MR */
205 __u64 ibpm_iova; /* Virtual I/O address */
206 int ibpm_refcount; /* reference count */
211 struct list_head ibd_list; /* chain on kib_devs */
212 __u32 ibd_ifip; /* IPoIB interface IP */
213 char ibd_ifname[32]; /* IPoIB interface name */
214 int ibd_nnets; /* # nets extant */
216 struct rdma_cm_id *ibd_cmid; /* IB listener (bound to 1 device) */
217 struct ib_pd *ibd_pd; /* PD for the device */
218 int ibd_page_shift; /* page shift of current HCA */
219 int ibd_page_size; /* page size of current HCA */
220 __u64 ibd_page_mask; /* page mask of current HCA */
221 int ibd_mr_shift; /* bits shift of max MR size */
222 __u64 ibd_mr_size; /* size of MR */
224 int ibd_nmrs; /* # of global MRs */
225 struct ib_mr **ibd_mrs; /* MR for non RDMA I/O */
230 __u64 ibn_incarnation; /* my epoch */
231 int ibn_init; /* initialisation state */
232 int ibn_shutdown; /* shutting down? */
234 atomic_t ibn_npeers; /* # peers extant */
235 atomic_t ibn_nconns; /* # connections extant */
237 __u64 ibn_tx_next_cookie; /* RDMA completion cookie */
238 struct kib_tx *ibn_tx_descs; /* all the tx descriptors */
239 kib_pages_t *ibn_tx_pages; /* premapped tx msg pages */
240 struct list_head ibn_idle_txs; /* idle tx descriptors */
241 spinlock_t ibn_tx_lock; /* serialise */
243 struct ib_fmr_pool *ibn_fmrpool; /* FMR pool for RDMA I/O */
244 kib_phys_mr_pool_t *ibn_pmrpool; /* Physical MR pool for RDMA I/O */
246 kib_dev_t *ibn_dev; /* underlying IB device */
251 int kib_init; /* initialisation state */
252 int kib_shutdown; /* shut down? */
253 struct list_head kib_devs; /* IB devices extant */
254 atomic_t kib_nthreads; /* # live threads */
255 rwlock_t kib_global_lock; /* stabilize net/dev/peer/conn ops */
257 struct list_head *kib_peers; /* hash table of all my known peers */
258 int kib_peer_hash_size; /* size of kib_peers */
260 void *kib_connd; /* the connd task (serialisation assertions) */
261 struct list_head kib_connd_conns; /* connections to setup/teardown */
262 struct list_head kib_connd_zombies; /* connections with zero refcount */
263 wait_queue_head_t kib_connd_waitq; /* connection daemon sleeps here */
264 spinlock_t kib_connd_lock; /* serialise */
266 wait_queue_head_t kib_sched_waitq; /* schedulers sleep here */
267 struct list_head kib_sched_conns; /* conns to check for rx completions */
268 spinlock_t kib_sched_lock; /* serialise */
270 struct ib_qp_attr kib_error_qpa; /* QP->ERROR */
273 #define IBLND_INIT_NOTHING 0
274 #define IBLND_INIT_DATA 1
275 #define IBLND_INIT_ALL 2
277 /************************************************************************
278 * IB Wire message format.
279 * These are sent in sender's byte order (i.e. receiver flips).
282 typedef struct kib_connparams
284 __u16 ibcp_queue_depth;
285 __u16 ibcp_max_frags;
286 __u32 ibcp_max_msg_size;
287 } WIRE_ATTR kib_connparams_t;
291 lnet_hdr_t ibim_hdr; /* portals header */
292 char ibim_payload[0]; /* piggy-backed payload */
293 } WIRE_ATTR kib_immediate_msg_t;
297 __u32 rf_nob; /* # bytes this frag */
298 __u64 rf_addr; /* CAVEAT EMPTOR: misaligned!! */
299 } WIRE_ATTR kib_rdma_frag_t;
303 __u32 rd_key; /* local/remote key */
304 __u32 rd_nfrags; /* # fragments */
305 kib_rdma_frag_t rd_frags[0]; /* buffer frags */
306 } WIRE_ATTR kib_rdma_desc_t;
310 lnet_hdr_t ibprm_hdr; /* portals header */
311 __u64 ibprm_cookie; /* opaque completion cookie */
312 } WIRE_ATTR kib_putreq_msg_t;
316 __u64 ibpam_src_cookie; /* reflected completion cookie */
317 __u64 ibpam_dst_cookie; /* opaque completion cookie */
318 kib_rdma_desc_t ibpam_rd; /* sender's sink buffer */
319 } WIRE_ATTR kib_putack_msg_t;
323 lnet_hdr_t ibgm_hdr; /* portals header */
324 __u64 ibgm_cookie; /* opaque completion cookie */
325 kib_rdma_desc_t ibgm_rd; /* rdma descriptor */
326 } WIRE_ATTR kib_get_msg_t;
330 __u64 ibcm_cookie; /* opaque completion cookie */
331 __s32 ibcm_status; /* < 0 failure: >= 0 length */
332 } WIRE_ATTR kib_completion_msg_t;
336 /* First 2 fields fixed FOR ALL TIME */
337 __u32 ibm_magic; /* I'm an openibnal message */
338 __u16 ibm_version; /* this is my version number */
340 __u8 ibm_type; /* msg type */
341 __u8 ibm_credits; /* returned credits */
342 __u32 ibm_nob; /* # bytes in whole message */
343 __u32 ibm_cksum; /* checksum (0 == no checksum) */
344 __u64 ibm_srcnid; /* sender's NID */
345 __u64 ibm_srcstamp; /* sender's incarnation */
346 __u64 ibm_dstnid; /* destination's NID */
347 __u64 ibm_dststamp; /* destination's incarnation */
350 kib_connparams_t connparams;
351 kib_immediate_msg_t immediate;
352 kib_putreq_msg_t putreq;
353 kib_putack_msg_t putack;
355 kib_completion_msg_t completion;
357 } WIRE_ATTR kib_msg_t;
359 #define IBLND_MSG_MAGIC LNET_PROTO_IB_MAGIC /* unique magic */
361 #define IBLND_MSG_VERSION_1 0x11
362 #define IBLND_MSG_VERSION_2 0x12
363 #define IBLND_MSG_VERSION IBLND_MSG_VERSION_2
365 #define IBLND_MSG_CONNREQ 0xc0 /* connection request */
366 #define IBLND_MSG_CONNACK 0xc1 /* connection acknowledge */
367 #define IBLND_MSG_NOOP 0xd0 /* nothing (just credits) */
368 #define IBLND_MSG_IMMEDIATE 0xd1 /* immediate */
369 #define IBLND_MSG_PUT_REQ 0xd2 /* putreq (src->sink) */
370 #define IBLND_MSG_PUT_NAK 0xd3 /* completion (sink->src) */
371 #define IBLND_MSG_PUT_ACK 0xd4 /* putack (sink->src) */
372 #define IBLND_MSG_PUT_DONE 0xd5 /* completion (src->sink) */
373 #define IBLND_MSG_GET_REQ 0xd6 /* getreq (sink->src) */
374 #define IBLND_MSG_GET_DONE 0xd7 /* completion (src->sink: all OK) */
377 __u32 ibr_magic; /* sender's magic */
378 __u16 ibr_version; /* sender's version */
379 __u8 ibr_why; /* reject reason */
380 __u8 ibr_padding; /* padding */
381 __u64 ibr_incarnation; /* incarnation of peer */
382 kib_connparams_t ibr_cp; /* connection parameters */
383 } WIRE_ATTR kib_rej_t;
385 /* connection rejection reasons */
386 #define IBLND_REJECT_CONN_RACE 1 /* You lost connection race */
387 #define IBLND_REJECT_NO_RESOURCES 2 /* Out of memory/conns etc */
388 #define IBLND_REJECT_FATAL 3 /* Anything else */
390 #define IBLND_REJECT_CONN_UNCOMPAT 4 /* incompatible version peer */
391 #define IBLND_REJECT_CONN_STALE 5 /* stale peer */
393 #define IBLND_REJECT_RDMA_FRAGS 6 /* Fatal: peer's rdma frags can't match mine */
394 #define IBLND_REJECT_MSG_QUEUE_SIZE 7 /* Fatal: peer's msg queue size can't match mine */
396 /***********************************************************************/
398 typedef struct kib_rx /* receive message */
400 struct list_head rx_list; /* queue for attention */
401 struct kib_conn *rx_conn; /* owning conn */
402 int rx_nob; /* # bytes received (-1 while posted) */
403 enum ib_wc_status rx_status; /* completion status */
404 kib_msg_t *rx_msg; /* message buffer (host vaddr) */
405 __u64 rx_msgaddr; /* message buffer (I/O addr) */
406 DECLARE_PCI_UNMAP_ADDR (rx_msgunmap); /* for dma_unmap_single() */
407 struct ib_recv_wr rx_wrq; /* receive work item... */
408 struct ib_sge rx_sge; /* ...and its memory */
411 #define IBLND_POSTRX_DONT_POST 0 /* don't post */
412 #define IBLND_POSTRX_NO_CREDIT 1 /* post: no credits */
413 #define IBLND_POSTRX_PEER_CREDIT 2 /* post: give peer back 1 credit */
414 #define IBLND_POSTRX_RSRVD_CREDIT 3 /* post: give myself back 1 reserved credit */
416 typedef struct kib_tx /* transmit message */
418 struct list_head tx_list; /* queue on idle_txs ibc_tx_queue etc. */
419 struct kib_conn *tx_conn; /* owning conn */
420 int tx_sending; /* # tx callbacks outstanding */
421 int tx_queued; /* queued for sending */
422 int tx_waiting; /* waiting for peer */
423 int tx_status; /* LNET completion status */
424 unsigned long tx_deadline; /* completion deadline */
425 __u64 tx_cookie; /* completion cookie */
426 lnet_msg_t *tx_lntmsg[2]; /* lnet msgs to finalize on completion */
427 kib_msg_t *tx_msg; /* message buffer (host vaddr) */
428 __u64 tx_msgaddr; /* message buffer (I/O addr) */
429 DECLARE_PCI_UNMAP_ADDR (tx_msgunmap); /* for dma_unmap_single() */
430 int tx_nwrq; /* # send work items */
431 struct ib_send_wr *tx_wrq; /* send work items... */
432 struct ib_sge *tx_sge; /* ...and their memory */
433 kib_rdma_desc_t *tx_rd; /* rdma descriptor */
434 int tx_nfrags; /* # entries in... */
435 struct scatterlist *tx_frags; /* dma_map_sg descriptor */
436 struct ib_phys_buf *tx_ipb; /* physical buffer (for iWARP) */
437 __u64 *tx_pages; /* rdma phys page addrs */
439 kib_phys_mr_t *pmr; /* MR for physical buffer */
440 struct ib_pool_fmr *fmr; /* rdma mapping (mapped if != NULL) */
442 int tx_dmadir; /* dma direction */
445 typedef struct kib_connvars
447 /* connection-in-progress variables */
451 typedef struct kib_conn
453 struct kib_peer *ibc_peer; /* owning peer */
454 struct list_head ibc_list; /* stash on peer's conn list */
455 struct list_head ibc_sched_list; /* schedule for attention */
456 __u16 ibc_version; /* version of connection */
457 __u64 ibc_incarnation; /* which instance of the peer */
458 atomic_t ibc_refcount; /* # users */
459 int ibc_state; /* what's happening */
460 int ibc_nsends_posted; /* # uncompleted sends */
461 int ibc_noops_posted; /* # uncompleted NOOPs */
462 int ibc_credits; /* # credits I have */
463 int ibc_outstanding_credits; /* # credits to return */
464 int ibc_reserved_credits;/* # ACK/DONE msg credits */
465 int ibc_comms_error; /* set on comms error */
466 int ibc_nrx:16; /* receive buffers owned */
467 int ibc_scheduled:1; /* scheduled for attention */
468 int ibc_ready:1; /* CQ callback fired */
469 unsigned long ibc_last_send; /* time of last send */
470 struct list_head ibc_early_rxs; /* rxs completed before ESTABLISHED */
471 struct list_head ibc_tx_queue; /* sends that need a credit */
472 struct list_head ibc_tx_queue_nocred;/* sends that don't need a credit */
473 struct list_head ibc_tx_queue_rsrvd; /* sends that need to reserve an ACK/DONE msg */
474 struct list_head ibc_active_txs; /* active tx awaiting completion */
475 spinlock_t ibc_lock; /* serialise */
476 kib_rx_t *ibc_rxs; /* the rx descs */
477 kib_pages_t *ibc_rx_pages; /* premapped rx msg pages */
479 struct rdma_cm_id *ibc_cmid; /* CM id */
480 struct ib_cq *ibc_cq; /* completion queue */
482 kib_connvars_t *ibc_connvars; /* in-progress connection state */
485 #define IBLND_CONN_INIT 0 /* being intialised */
486 #define IBLND_CONN_ACTIVE_CONNECT 1 /* active sending req */
487 #define IBLND_CONN_PASSIVE_WAIT 2 /* passive waiting for rtu */
488 #define IBLND_CONN_ESTABLISHED 3 /* connection established */
489 #define IBLND_CONN_CLOSING 4 /* being closed */
490 #define IBLND_CONN_DISCONNECTED 5 /* disconnected */
492 typedef struct kib_peer
494 struct list_head ibp_list; /* stash on global peer list */
495 lnet_nid_t ibp_nid; /* who's on the other end(s) */
496 lnet_ni_t *ibp_ni; /* LNet interface */
497 atomic_t ibp_refcount; /* # users */
498 struct list_head ibp_conns; /* all active connections */
499 struct list_head ibp_tx_queue; /* msgs waiting for a conn */
500 __u16 ibp_version; /* version of peer */
501 __u64 ibp_incarnation; /* incarnation of peer */
502 int ibp_connecting; /* current active connection attempts */
503 int ibp_accepting; /* current passive connection attempts */
504 int ibp_error; /* errno on closing this peer */
505 cfs_time_t ibp_last_alive; /* when (in jiffies) I was last alive */
508 extern kib_data_t kiblnd_data;
510 #define kiblnd_conn_addref(conn) \
512 CDEBUG(D_NET, "conn[%p] (%d)++\n", \
513 (conn), atomic_read(&(conn)->ibc_refcount)); \
514 LASSERT(atomic_read(&(conn)->ibc_refcount) > 0); \
515 atomic_inc(&(conn)->ibc_refcount); \
518 #define kiblnd_conn_decref(conn) \
520 unsigned long flags; \
522 CDEBUG(D_NET, "conn[%p] (%d)--\n", \
523 (conn), atomic_read(&(conn)->ibc_refcount)); \
524 LASSERT(atomic_read(&(conn)->ibc_refcount) > 0); \
525 if (atomic_dec_and_test(&(conn)->ibc_refcount)) { \
526 spin_lock_irqsave(&kiblnd_data.kib_connd_lock, flags); \
527 list_add_tail(&(conn)->ibc_list, \
528 &kiblnd_data.kib_connd_zombies); \
529 wake_up(&kiblnd_data.kib_connd_waitq); \
530 spin_unlock_irqrestore(&kiblnd_data.kib_connd_lock, flags); \
534 #define kiblnd_peer_addref(peer) \
536 CDEBUG(D_NET, "peer[%p] -> %s (%d)++\n", \
537 (peer), libcfs_nid2str((peer)->ibp_nid), \
538 atomic_read (&(peer)->ibp_refcount)); \
539 LASSERT(atomic_read(&(peer)->ibp_refcount) > 0); \
540 atomic_inc(&(peer)->ibp_refcount); \
543 #define kiblnd_peer_decref(peer) \
545 CDEBUG(D_NET, "peer[%p] -> %s (%d)--\n", \
546 (peer), libcfs_nid2str((peer)->ibp_nid), \
547 atomic_read (&(peer)->ibp_refcount)); \
548 LASSERT(atomic_read(&(peer)->ibp_refcount) > 0); \
549 if (atomic_dec_and_test(&(peer)->ibp_refcount)) \
550 kiblnd_destroy_peer(peer); \
553 static inline struct list_head *
554 kiblnd_nid2peerlist (lnet_nid_t nid)
556 unsigned int hash = ((unsigned int)nid) % kiblnd_data.kib_peer_hash_size;
558 return (&kiblnd_data.kib_peers [hash]);
562 kiblnd_peer_active (kib_peer_t *peer)
564 /* Am I in the peer hash table? */
565 return (!list_empty(&peer->ibp_list));
568 static inline kib_conn_t *
569 kiblnd_get_conn_locked (kib_peer_t *peer)
571 LASSERT (!list_empty(&peer->ibp_conns));
573 /* just return the first connection */
574 return list_entry(peer->ibp_conns.next, kib_conn_t, ibc_list);
578 kiblnd_send_keepalive(kib_conn_t *conn)
580 return (*kiblnd_tunables.kib_keepalive > 0) &&
581 time_after(jiffies, conn->ibc_last_send +
582 *kiblnd_tunables.kib_keepalive*HZ);
586 kiblnd_send_noop(kib_conn_t *conn)
588 LASSERT (conn->ibc_state >= IBLND_CONN_ESTABLISHED);
590 if (conn->ibc_outstanding_credits <
591 IBLND_CREDITS_HIGHWATER(conn->ibc_version) &&
592 !kiblnd_send_keepalive(conn))
593 return 0; /* No need to send NOOP */
595 if (!list_empty(&conn->ibc_tx_queue_nocred))
596 return 0; /* NOOP can be piggybacked */
598 if (!IBLND_OOB_CAPABLE(conn->ibc_version))
599 return list_empty(&conn->ibc_tx_queue); /* can't piggyback? */
601 /* No tx to piggyback NOOP onto or no credit to send a tx */
602 return (list_empty(&conn->ibc_tx_queue) || conn->ibc_credits == 0);
606 kiblnd_abort_receives(kib_conn_t *conn)
608 ib_modify_qp(conn->ibc_cmid->qp,
609 &kiblnd_data.kib_error_qpa, IB_QP_STATE);
612 static inline const char *
613 kiblnd_queue2str (kib_conn_t *conn, struct list_head *q)
615 if (q == &conn->ibc_tx_queue)
618 if (q == &conn->ibc_tx_queue_rsrvd)
619 return "tx_queue_rsrvd";
621 if (q == &conn->ibc_tx_queue_nocred)
622 return "tx_queue_nocred";
624 if (q == &conn->ibc_active_txs)
631 /* CAVEAT EMPTOR: We rely on descriptor alignment to allow us to use the
632 * lowest bits of the work request id to stash the work item type. */
634 #define IBLND_WID_TX 0
635 #define IBLND_WID_RDMA 1
636 #define IBLND_WID_RX 2
637 #define IBLND_WID_MASK 3UL
640 kiblnd_ptr2wreqid (void *ptr, int type)
642 unsigned long lptr = (unsigned long)ptr;
644 LASSERT ((lptr & IBLND_WID_MASK) == 0);
645 LASSERT ((type & ~IBLND_WID_MASK) == 0);
646 return (__u64)(lptr | type);
650 kiblnd_wreqid2ptr (__u64 wreqid)
652 return (void *)(((unsigned long)wreqid) & ~IBLND_WID_MASK);
656 kiblnd_wreqid2type (__u64 wreqid)
658 return (wreqid & IBLND_WID_MASK);
662 kiblnd_set_conn_state (kib_conn_t *conn, int state)
664 conn->ibc_state = state;
669 kiblnd_init_msg (kib_msg_t *msg, int type, int body_nob)
671 msg->ibm_type = type;
672 msg->ibm_nob = offsetof(kib_msg_t, ibm_u) + body_nob;
676 kiblnd_rd_size (kib_rdma_desc_t *rd)
681 for (i = size = 0; i < rd->rd_nfrags; i++)
682 size += rd->rd_frags[i].rf_nob;
688 kiblnd_rd_frag_addr(kib_rdma_desc_t *rd, int index)
690 return rd->rd_frags[index].rf_addr;
694 kiblnd_rd_frag_size(kib_rdma_desc_t *rd, int index)
696 return rd->rd_frags[index].rf_nob;
700 kiblnd_rd_frag_key(kib_rdma_desc_t *rd, int index)
706 kiblnd_rd_consume_frag(kib_rdma_desc_t *rd, int index, __u32 nob)
708 if (nob < rd->rd_frags[index].rf_nob) {
709 rd->rd_frags[index].rf_addr += nob;
710 rd->rd_frags[index].rf_nob -= nob;
719 kiblnd_rd_msg_size(kib_rdma_desc_t *rd, int msgtype, int n)
721 LASSERT (msgtype == IBLND_MSG_GET_REQ ||
722 msgtype == IBLND_MSG_PUT_ACK);
724 return msgtype == IBLND_MSG_GET_REQ ?
725 offsetof(kib_get_msg_t, ibgm_rd.rd_frags[n]) :
726 offsetof(kib_putack_msg_t, ibpam_rd.rd_frags[n]);
729 #ifdef HAVE_OFED_IB_DMA_MAP
732 kiblnd_dma_mapping_error(struct ib_device *dev, u64 dma_addr)
734 return ib_dma_mapping_error(dev, dma_addr);
737 static inline __u64 kiblnd_dma_map_single(struct ib_device *dev,
738 void *msg, size_t size,
739 enum dma_data_direction direction)
741 return ib_dma_map_single(dev, msg, size, direction);
744 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
745 __u64 addr, size_t size,
746 enum dma_data_direction direction)
748 ib_dma_unmap_single(dev, addr, size, direction);
751 #define KIBLND_UNMAP_ADDR_SET(p, m, a) do {} while (0)
752 #define KIBLND_UNMAP_ADDR(p, m, a) (a)
754 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
755 struct scatterlist *sg, int nents,
756 enum dma_data_direction direction)
758 return ib_dma_map_sg(dev, sg, nents, direction);
761 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
762 struct scatterlist *sg, int nents,
763 enum dma_data_direction direction)
765 ib_dma_unmap_sg(dev, sg, nents, direction);
768 static inline __u64 kiblnd_sg_dma_address(struct ib_device *dev,
769 struct scatterlist *sg)
771 return ib_sg_dma_address(dev, sg);
774 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
775 struct scatterlist *sg)
777 return ib_sg_dma_len(dev, sg);
780 /* XXX We use KIBLND_CONN_PARAM(e) as writable buffer, it's not strictly
781 * right because OFED1.2 defines it as const, to use it we have to add
782 * (void *) cast to overcome "const" */
784 #define KIBLND_CONN_PARAM(e) ((e)->param.conn.private_data)
785 #define KIBLND_CONN_PARAM_LEN(e) ((e)->param.conn.private_data_len)
790 kiblnd_dma_mapping_error(struct ib_device *dev, dma_addr_t dma_addr)
792 return dma_mapping_error(dma_addr);
795 static inline dma_addr_t kiblnd_dma_map_single(struct ib_device *dev,
796 void *msg, size_t size,
797 enum dma_data_direction direction)
799 return dma_map_single(dev->dma_device, msg, size, direction);
802 static inline void kiblnd_dma_unmap_single(struct ib_device *dev,
803 dma_addr_t addr, size_t size,
804 enum dma_data_direction direction)
806 dma_unmap_single(dev->dma_device, addr, size, direction);
809 #define KIBLND_UNMAP_ADDR_SET(p, m, a) pci_unmap_addr_set(p, m, a)
810 #define KIBLND_UNMAP_ADDR(p, m, a) pci_unmap_addr(p, m)
812 static inline int kiblnd_dma_map_sg(struct ib_device *dev,
813 struct scatterlist *sg, int nents,
814 enum dma_data_direction direction)
816 return dma_map_sg(dev->dma_device, sg, nents, direction);
819 static inline void kiblnd_dma_unmap_sg(struct ib_device *dev,
820 struct scatterlist *sg, int nents,
821 enum dma_data_direction direction)
823 return dma_unmap_sg(dev->dma_device, sg, nents, direction);
827 static inline dma_addr_t kiblnd_sg_dma_address(struct ib_device *dev,
828 struct scatterlist *sg)
830 return sg_dma_address(sg);
834 static inline unsigned int kiblnd_sg_dma_len(struct ib_device *dev,
835 struct scatterlist *sg)
837 return sg_dma_len(sg);
840 #define KIBLND_CONN_PARAM(e) ((e)->private_data)
841 #define KIBLND_CONN_PARAM_LEN(e) ((e)->private_data_len)
845 struct ib_mr *kiblnd_find_rd_dma_mr(kib_net_t *net,
846 kib_rdma_desc_t *rd);
847 struct ib_mr *kiblnd_find_dma_mr(kib_net_t *net,
848 __u64 addr, __u64 size);
849 void kiblnd_map_rx_descs(kib_conn_t *conn);
850 void kiblnd_unmap_rx_descs(kib_conn_t *conn);
851 void kiblnd_map_tx_descs (lnet_ni_t *ni);
852 void kiblnd_unmap_tx_descs(lnet_ni_t *ni);
853 int kiblnd_map_tx(lnet_ni_t *ni, kib_tx_t *tx,
854 kib_rdma_desc_t *rd, int nfrags);
855 void kiblnd_unmap_tx(lnet_ni_t *ni, kib_tx_t *tx);
856 kib_phys_mr_t *kiblnd_phys_mr_map(kib_net_t *net, kib_rdma_desc_t *rd,
857 struct ib_phys_buf *ipb, __u64 *iova);
858 void kiblnd_phys_mr_unmap(kib_net_t *net, kib_phys_mr_t *pmr);
860 int kiblnd_startup (lnet_ni_t *ni);
861 void kiblnd_shutdown (lnet_ni_t *ni);
862 int kiblnd_ctl (lnet_ni_t *ni, unsigned int cmd, void *arg);
863 void kiblnd_query (struct lnet_ni *ni, lnet_nid_t nid, time_t *when);
865 int kiblnd_tunables_init(void);
866 void kiblnd_tunables_fini(void);
868 int kiblnd_connd (void *arg);
869 int kiblnd_scheduler(void *arg);
870 int kiblnd_thread_start (int (*fn)(void *arg), void *arg);
872 int kiblnd_alloc_pages (kib_pages_t **pp, int npages);
873 void kiblnd_free_pages (kib_pages_t *p);
875 int kiblnd_cm_callback(struct rdma_cm_id *cmid,
876 struct rdma_cm_event *event);
877 int kiblnd_translate_mtu(int value);
879 int kiblnd_create_peer (lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid);
880 void kiblnd_destroy_peer (kib_peer_t *peer);
881 void kiblnd_destroy_dev (kib_dev_t *dev);
882 void kiblnd_unlink_peer_locked (kib_peer_t *peer);
883 void kiblnd_peer_alive (kib_peer_t *peer);
884 kib_peer_t *kiblnd_find_peer_locked (lnet_nid_t nid);
885 void kiblnd_peer_connect_failed (kib_peer_t *peer, int active, int error);
886 int kiblnd_close_stale_conns_locked (kib_peer_t *peer,
887 int version, __u64 incarnation);
888 int kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
890 void kiblnd_connreq_done(kib_conn_t *conn, int status);
891 kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
892 int state, int version);
893 void kiblnd_destroy_conn (kib_conn_t *conn);
894 void kiblnd_close_conn (kib_conn_t *conn, int error);
895 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);
897 int kiblnd_init_rdma (kib_conn_t *conn, kib_tx_t *tx, int type,
898 int nob, kib_rdma_desc_t *dstrd, __u64 dstcookie);
900 void kiblnd_launch_tx (lnet_ni_t *ni, kib_tx_t *tx, lnet_nid_t nid);
901 void kiblnd_queue_tx_locked (kib_tx_t *tx, kib_conn_t *conn);
902 void kiblnd_queue_tx (kib_tx_t *tx, kib_conn_t *conn);
903 void kiblnd_init_tx_msg (lnet_ni_t *ni, kib_tx_t *tx, int type, int body_nob);
904 void kiblnd_txlist_done (lnet_ni_t *ni, struct list_head *txlist, int status);
905 void kiblnd_check_sends (kib_conn_t *conn);
907 void kiblnd_qp_event(struct ib_event *event, void *arg);
908 void kiblnd_cq_event(struct ib_event *event, void *arg);
909 void kiblnd_cq_completion(struct ib_cq *cq, void *arg);
911 void kiblnd_pack_msg (lnet_ni_t *ni, kib_msg_t *msg, int version,
912 int credits, lnet_nid_t dstnid, __u64 dststamp);
913 int kiblnd_unpack_msg(kib_msg_t *msg, int nob);
914 int kiblnd_post_rx (kib_rx_t *rx, int credit);
916 int kiblnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
917 int kiblnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg, int delayed,
918 unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
919 unsigned int offset, unsigned int mlen, unsigned int rlen);