1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2004 Cluster File Systems, Inc.
5 * Author: Eric Barton <eric@bartonsoftware.com>
7 * This file is part of Lustre, http://www.lustre.org.
9 * Lustre is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Lustre is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Lustre; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #include "openibnal.h"
27 * LIB functions follow
31 kibnal_schedule_tx_done (kib_tx_t *tx)
35 spin_lock_irqsave (&kibnal_data.kib_sched_lock, flags);
37 list_add_tail(&tx->tx_list, &kibnal_data.kib_sched_txq);
38 wake_up (&kibnal_data.kib_sched_waitq);
40 spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags);
44 kibnal_tx_done (kib_tx_t *tx)
46 ptl_err_t ptlrc = (tx->tx_status == 0) ? PTL_OK : PTL_FAIL;
51 LASSERT (tx->tx_sending == 0); /* mustn't be awaiting callback */
52 LASSERT (!tx->tx_passive_rdma_wait); /* mustn't be awaiting RDMA */
54 switch (tx->tx_mapped) {
63 /* can't deregister memory in IRQ context... */
64 kibnal_schedule_tx_done(tx);
67 rc = ib_memory_deregister(tx->tx_md.md_handle.mr);
69 tx->tx_mapped = KIB_TX_UNMAPPED;
73 case KIB_TX_MAPPED_FMR:
74 if (in_interrupt() && tx->tx_status != 0) {
75 /* can't flush FMRs in IRQ context... */
76 kibnal_schedule_tx_done(tx);
80 rc = ib_fmr_deregister(tx->tx_md.md_handle.fmr);
83 if (tx->tx_status != 0)
84 ib_fmr_pool_force_flush(kibnal_data.kib_fmr_pool);
85 tx->tx_mapped = KIB_TX_UNMAPPED;
90 for (i = 0; i < 2; i++) {
91 /* tx may have up to 2 libmsgs to finalise */
92 if (tx->tx_libmsg[i] == NULL)
95 lib_finalize (&kibnal_lib, NULL, tx->tx_libmsg[i], ptlrc);
96 tx->tx_libmsg[i] = NULL;
99 if (tx->tx_conn != NULL) {
100 kibnal_put_conn (tx->tx_conn);
105 tx->tx_passive_rdma = 0;
108 spin_lock_irqsave (&kibnal_data.kib_tx_lock, flags);
111 list_add_tail (&tx->tx_list, &kibnal_data.kib_idle_nblk_txs);
113 list_add_tail (&tx->tx_list, &kibnal_data.kib_idle_txs);
114 wake_up (&kibnal_data.kib_idle_tx_waitq);
117 spin_unlock_irqrestore (&kibnal_data.kib_tx_lock, flags);
121 kibnal_get_idle_tx (int may_block)
127 spin_lock_irqsave (&kibnal_data.kib_tx_lock, flags);
129 /* "normal" descriptor is free */
130 if (!list_empty (&kibnal_data.kib_idle_txs)) {
131 tx = list_entry (kibnal_data.kib_idle_txs.next,
137 /* may dip into reserve pool */
138 if (list_empty (&kibnal_data.kib_idle_nblk_txs)) {
139 CERROR ("reserved tx desc pool exhausted\n");
143 tx = list_entry (kibnal_data.kib_idle_nblk_txs.next,
148 /* block for idle tx */
149 spin_unlock_irqrestore (&kibnal_data.kib_tx_lock, flags);
151 wait_event (kibnal_data.kib_idle_tx_waitq,
152 !list_empty (&kibnal_data.kib_idle_txs) ||
153 kibnal_data.kib_shutdown);
157 list_del (&tx->tx_list);
159 /* Allocate a new passive RDMA completion cookie. It might
160 * not be needed, but we've got a lock right now and we're
161 * unlikely to wrap... */
162 tx->tx_passive_rdma_cookie = kibnal_data.kib_next_tx_cookie++;
164 LASSERT (tx->tx_mapped == KIB_TX_UNMAPPED);
165 LASSERT (tx->tx_nsp == 0);
166 LASSERT (tx->tx_sending == 0);
167 LASSERT (tx->tx_status == 0);
168 LASSERT (tx->tx_conn == NULL);
169 LASSERT (!tx->tx_passive_rdma);
170 LASSERT (!tx->tx_passive_rdma_wait);
171 LASSERT (tx->tx_libmsg[0] == NULL);
172 LASSERT (tx->tx_libmsg[1] == NULL);
175 spin_unlock_irqrestore (&kibnal_data.kib_tx_lock, flags);
181 kibnal_dist(lib_nal_t *nal, ptl_nid_t nid, unsigned long *dist)
183 /* I would guess that if kibnal_get_peer (nid) == NULL,
184 and we're not routing, then 'nid' is very distant :) */
185 if ( nal->libnal_ni.ni_pid.nid == nid ) {
195 kibnal_complete_passive_rdma(kib_conn_t *conn, __u64 cookie, int status)
197 struct list_head *ttmp;
201 spin_lock_irqsave (&conn->ibc_lock, flags);
203 list_for_each (ttmp, &conn->ibc_active_txs) {
204 kib_tx_t *tx = list_entry(ttmp, kib_tx_t, tx_list);
206 LASSERT (tx->tx_passive_rdma ||
207 !tx->tx_passive_rdma_wait);
209 LASSERT (tx->tx_passive_rdma_wait ||
210 tx->tx_sending != 0);
212 if (!tx->tx_passive_rdma_wait ||
213 tx->tx_passive_rdma_cookie != cookie)
216 CDEBUG(D_NET, "Complete %p "LPD64": %d\n", tx, cookie, status);
218 tx->tx_status = status;
219 tx->tx_passive_rdma_wait = 0;
220 idle = (tx->tx_sending == 0);
223 list_del (&tx->tx_list);
225 spin_unlock_irqrestore (&conn->ibc_lock, flags);
227 /* I could be racing with tx callbacks. It's whoever
228 * _makes_ tx idle that frees it */
234 spin_unlock_irqrestore (&conn->ibc_lock, flags);
236 CERROR ("Unmatched (late?) RDMA completion "LPX64" from "LPX64"\n",
237 cookie, conn->ibc_peer->ibp_nid);
241 kibnal_post_rx (kib_rx_t *rx, int do_credits)
243 kib_conn_t *conn = rx->rx_conn;
247 rx->rx_gl = (struct ib_gather_scatter) {
248 .address = rx->rx_vaddr,
249 .length = IBNAL_MSG_SIZE,
250 .key = conn->ibc_rx_pages->ibp_lkey,
253 rx->rx_sp = (struct ib_receive_param) {
254 .work_request_id = kibnal_ptr2wreqid(rx, 1),
255 .scatter_list = &rx->rx_gl,
256 .num_scatter_entries = 1,
257 .device_specific = NULL,
261 LASSERT (conn->ibc_state >= IBNAL_CONN_ESTABLISHED);
262 LASSERT (!rx->rx_posted);
266 if (conn->ibc_state != IBNAL_CONN_ESTABLISHED)
269 rc = ib_receive (conn->ibc_qp, &rx->rx_sp, 1);
273 spin_lock_irqsave(&conn->ibc_lock, flags);
274 conn->ibc_outstanding_credits++;
275 spin_unlock_irqrestore(&conn->ibc_lock, flags);
277 kibnal_check_sends(conn);
282 if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) {
283 CERROR ("Error posting receive -> "LPX64": %d\n",
284 conn->ibc_peer->ibp_nid, rc);
285 kibnal_close_conn (rx->rx_conn, rc);
287 CDEBUG (D_NET, "Error posting receive -> "LPX64": %d\n",
288 conn->ibc_peer->ibp_nid, rc);
292 kibnal_put_conn (conn);
296 kibnal_rx_callback (struct ib_cq_entry *e)
298 kib_rx_t *rx = (kib_rx_t *)kibnal_wreqid2ptr(e->work_request_id);
299 kib_msg_t *msg = rx->rx_msg;
300 kib_conn_t *conn = rx->rx_conn;
305 CDEBUG (D_NET, "rx %p conn %p\n", rx, conn);
306 LASSERT (rx->rx_posted);
310 /* receives complete with error in any case after we've started
312 if (conn->ibc_state >= IBNAL_CONN_DEATHROW)
315 /* We don't post receives until the conn is established */
316 LASSERT (conn->ibc_state == IBNAL_CONN_ESTABLISHED);
318 if (e->status != IB_COMPLETION_STATUS_SUCCESS) {
319 CERROR("Rx from "LPX64" failed: %d\n",
320 conn->ibc_peer->ibp_nid, e->status);
324 rc = kibnal_unpack_msg(msg, e->bytes_transferred);
326 CERROR ("Error %d unpacking rx from "LPX64"\n",
327 rc, conn->ibc_peer->ibp_nid);
331 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
332 msg->ibm_srcstamp != conn->ibc_incarnation ||
333 msg->ibm_dstnid != kibnal_lib.libnal_ni.ni_pid.nid ||
334 msg->ibm_dststamp != kibnal_data.kib_incarnation) {
335 CERROR ("Stale rx from "LPX64"\n",
336 conn->ibc_peer->ibp_nid);
340 /* Have I received credits that will let me send? */
341 credits = msg->ibm_credits;
343 spin_lock_irqsave(&conn->ibc_lock, flags);
344 conn->ibc_credits += credits;
345 spin_unlock_irqrestore(&conn->ibc_lock, flags);
347 kibnal_check_sends(conn);
350 switch (msg->ibm_type) {
352 kibnal_post_rx (rx, 1);
355 case IBNAL_MSG_IMMEDIATE:
358 case IBNAL_MSG_PUT_RDMA:
359 case IBNAL_MSG_GET_RDMA:
360 CDEBUG(D_NET, "%d RDMA: cookie "LPX64", key %x, addr "LPX64", nob %d\n",
361 msg->ibm_type, msg->ibm_u.rdma.ibrm_cookie,
362 msg->ibm_u.rdma.ibrm_desc.rd_key,
363 msg->ibm_u.rdma.ibrm_desc.rd_addr,
364 msg->ibm_u.rdma.ibrm_desc.rd_nob);
367 case IBNAL_MSG_PUT_DONE:
368 case IBNAL_MSG_GET_DONE:
369 CDEBUG(D_NET, "%d DONE: cookie "LPX64", status %d\n",
370 msg->ibm_type, msg->ibm_u.completion.ibcm_cookie,
371 msg->ibm_u.completion.ibcm_status);
373 kibnal_complete_passive_rdma (conn,
374 msg->ibm_u.completion.ibcm_cookie,
375 msg->ibm_u.completion.ibcm_status);
376 kibnal_post_rx (rx, 1);
380 CERROR ("Bad msg type %x from "LPX64"\n",
381 msg->ibm_type, conn->ibc_peer->ibp_nid);
385 /* schedule for kibnal_rx() in thread context */
386 spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags);
388 list_add_tail (&rx->rx_list, &kibnal_data.kib_sched_rxq);
389 wake_up (&kibnal_data.kib_sched_waitq);
391 spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags);
395 CDEBUG(D_NET, "rx %p conn %p\n", rx, conn);
396 kibnal_close_conn(conn, -ECONNABORTED);
398 /* Don't re-post rx & drop its ref on conn */
399 kibnal_put_conn(conn);
403 kibnal_rx (kib_rx_t *rx)
405 kib_msg_t *msg = rx->rx_msg;
407 /* Clear flag so I can detect if I've sent an RDMA completion */
410 switch (msg->ibm_type) {
411 case IBNAL_MSG_GET_RDMA:
412 lib_parse(&kibnal_lib, &msg->ibm_u.rdma.ibrm_hdr, rx);
413 /* If the incoming get was matched, I'll have initiated the
414 * RDMA and the completion message... */
418 /* Otherwise, I'll send a failed completion now to prevent
419 * the peer's GET blocking for the full timeout. */
420 CERROR ("Completing unmatched RDMA GET from "LPX64"\n",
421 rx->rx_conn->ibc_peer->ibp_nid);
422 kibnal_start_active_rdma (IBNAL_MSG_GET_DONE, -EIO,
423 rx, NULL, 0, NULL, NULL, 0, 0);
426 case IBNAL_MSG_PUT_RDMA:
427 lib_parse(&kibnal_lib, &msg->ibm_u.rdma.ibrm_hdr, rx);
430 /* This is most unusual, since even if lib_parse() didn't
431 * match anything, it should have asked us to read (and
432 * discard) the payload. The portals header must be
433 * inconsistent with this message type, so it's the
434 * sender's fault for sending garbage and she can time
436 CERROR ("Uncompleted RMDA PUT from "LPX64"\n",
437 rx->rx_conn->ibc_peer->ibp_nid);
440 case IBNAL_MSG_IMMEDIATE:
441 lib_parse(&kibnal_lib, &msg->ibm_u.immediate.ibim_hdr, rx);
442 LASSERT (!rx->rx_rdma);
450 kibnal_post_rx (rx, 1);
455 kibnal_kvaddr_to_phys (unsigned long vaddr, __u64 *physp)
459 if (vaddr >= VMALLOC_START &&
461 page = vmalloc_to_page ((void *)vaddr);
463 else if (vaddr >= PKMAP_BASE &&
464 vaddr < (PKMAP_BASE + LAST_PKMAP * PAGE_SIZE))
465 page = vmalloc_to_page ((void *)vaddr);
466 /* in 2.4 ^ just walks the page tables */
469 page = virt_to_page (vaddr);
475 *physp = kibnal_page2phys(page) + (vaddr & (PAGE_SIZE - 1));
481 kibnal_map_iov (kib_tx_t *tx, enum ib_memory_access access,
482 int niov, struct iovec *iov, int offset, int nob)
490 LASSERT (tx->tx_mapped == KIB_TX_UNMAPPED);
492 while (offset >= iov->iov_len) {
493 offset -= iov->iov_len;
499 if (nob > iov->iov_len - offset) {
500 CERROR ("Can't map multiple vaddr fragments\n");
504 vaddr = (void *)(((unsigned long)iov->iov_base) + offset);
505 tx->tx_md.md_addr = (__u64)((unsigned long)vaddr);
507 rc = ib_memory_register (kibnal_data.kib_pd,
510 &tx->tx_md.md_handle.mr,
515 CERROR ("Can't map vaddr: %d\n", rc);
519 tx->tx_mapped = KIB_TX_MAPPED;
524 kibnal_map_kiov (kib_tx_t *tx, enum ib_memory_access access,
525 int nkiov, ptl_kiov_t *kiov,
530 const int mapped = KIB_TX_MAPPED_FMR;
532 struct ib_physical_buffer *phys;
533 const int mapped = KIB_TX_MAPPED;
541 CDEBUG(D_NET, "niov %d offset %d nob %d\n", nkiov, offset, nob);
545 LASSERT (tx->tx_mapped == KIB_TX_UNMAPPED);
547 while (offset >= kiov->kiov_len) {
548 offset -= kiov->kiov_len;
554 phys_size = nkiov * sizeof (*phys);
555 PORTAL_ALLOC(phys, phys_size);
557 CERROR ("Can't allocate tmp phys\n");
561 page_offset = kiov->kiov_offset + offset;
563 phys[0] = kibnal_page2phys(kiov->kiov_page);
565 phys[0].address = kibnal_page2phys(kiov->kiov_page);
566 phys[0].size = PAGE_SIZE;
569 resid = nob - (kiov->kiov_len - offset);
576 if (kiov->kiov_offset != 0 ||
577 ((resid > PAGE_SIZE) &&
578 kiov->kiov_len < PAGE_SIZE)) {
580 /* Can't have gaps */
581 CERROR ("Can't make payload contiguous in I/O VM:"
582 "page %d, offset %d, len %d \n", nphys,
583 kiov->kiov_offset, kiov->kiov_len);
585 for (i = -nphys; i < nkiov; i++)
587 CERROR("kiov[%d] %p +%d for %d\n",
588 i, kiov[i].kiov_page, kiov[i].kiov_offset, kiov[i].kiov_len);
595 if (nphys == PTL_MD_MAX_IOV) {
596 CERROR ("payload too big (%d)\n", nphys);
601 LASSERT (nphys * sizeof (*phys) < phys_size);
603 phys[nphys] = kibnal_page2phys(kiov->kiov_page);
605 phys[nphys].address = kibnal_page2phys(kiov->kiov_page);
606 phys[nphys].size = PAGE_SIZE;
614 CWARN ("nphys %d, nob %d, page_offset %d\n", nphys, nob, page_offset);
615 for (rc = 0; rc < nphys; rc++)
616 CWARN (" [%d] "LPX64" / %d\n", rc, phys[rc].address, phys[rc].size);
618 tx->tx_md.md_addr = IBNAL_RDMA_BASE;
621 rc = ib_fmr_register_physical (kibnal_data.kib_fmr_pool,
625 &tx->tx_md.md_handle.fmr,
629 rc = ib_memory_register_physical (kibnal_data.kib_pd,
634 &tx->tx_md.md_handle.mr,
639 CDEBUG(D_NET, "Mapped %d pages %d bytes @ offset %d: lkey %x, rkey %x\n",
640 nphys, nob, page_offset, tx->tx_md.md_lkey, tx->tx_md.md_rkey);
641 tx->tx_mapped = mapped;
643 CERROR ("Can't map phys: %d\n", rc);
648 PORTAL_FREE(phys, phys_size);
653 kibnal_find_conn_locked (kib_peer_t *peer)
655 struct list_head *tmp;
657 /* just return the first connection */
658 list_for_each (tmp, &peer->ibp_conns) {
659 return (list_entry(tmp, kib_conn_t, ibc_list));
666 kibnal_check_sends (kib_conn_t *conn)
675 spin_lock_irqsave (&conn->ibc_lock, flags);
677 LASSERT (conn->ibc_nsends_posted <= IBNAL_MSG_QUEUE_SIZE);
679 if (list_empty(&conn->ibc_tx_queue) &&
680 conn->ibc_outstanding_credits >= IBNAL_CREDIT_HIGHWATER) {
681 spin_unlock_irqrestore(&conn->ibc_lock, flags);
683 tx = kibnal_get_idle_tx(0); /* don't block */
685 kibnal_init_tx_msg(tx, IBNAL_MSG_NOOP, 0);
687 spin_lock_irqsave(&conn->ibc_lock, flags);
690 atomic_inc(&conn->ibc_refcount);
691 kibnal_queue_tx_locked(tx, conn);
695 while (!list_empty (&conn->ibc_tx_queue)) {
696 tx = list_entry (conn->ibc_tx_queue.next, kib_tx_t, tx_list);
698 /* We rely on this for QP sizing */
699 LASSERT (tx->tx_nsp > 0 && tx->tx_nsp <= 2);
701 LASSERT (conn->ibc_outstanding_credits >= 0);
702 LASSERT (conn->ibc_outstanding_credits <= IBNAL_MSG_QUEUE_SIZE);
703 LASSERT (conn->ibc_credits >= 0);
704 LASSERT (conn->ibc_credits <= IBNAL_MSG_QUEUE_SIZE);
706 /* Not on ibc_rdma_queue */
707 LASSERT (!tx->tx_passive_rdma_wait);
709 if (conn->ibc_nsends_posted == IBNAL_MSG_QUEUE_SIZE)
712 if (conn->ibc_credits == 0) /* no credits */
715 if (conn->ibc_credits == 1 && /* last credit reserved for */
716 conn->ibc_outstanding_credits == 0) /* giving back credits */
719 list_del (&tx->tx_list);
721 if (tx->tx_msg->ibm_type == IBNAL_MSG_NOOP &&
722 (!list_empty(&conn->ibc_tx_queue) ||
723 conn->ibc_outstanding_credits < IBNAL_CREDIT_HIGHWATER)) {
725 spin_unlock_irqrestore(&conn->ibc_lock, flags);
727 spin_lock_irqsave(&conn->ibc_lock, flags);
731 kibnal_pack_msg(tx->tx_msg, conn->ibc_outstanding_credits,
732 conn->ibc_peer->ibp_nid, conn->ibc_incarnation);
734 conn->ibc_outstanding_credits = 0;
735 conn->ibc_nsends_posted++;
738 tx->tx_sending = tx->tx_nsp;
739 tx->tx_passive_rdma_wait = tx->tx_passive_rdma;
740 list_add (&tx->tx_list, &conn->ibc_active_txs);
742 spin_unlock_irqrestore (&conn->ibc_lock, flags);
744 /* NB the gap between removing tx from the queue and sending it
745 * allows message re-ordering to occur */
747 LASSERT (tx->tx_nsp > 0);
751 if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) {
753 /* Driver only accepts 1 item at a time */
754 for (i = 0; i < tx->tx_nsp; i++) {
755 rc = ib_send (conn->ibc_qp, &tx->tx_sp[i], 1);
762 spin_lock_irqsave (&conn->ibc_lock, flags);
764 /* NB credits are transferred in the actual
765 * message, which can only be the last work item */
766 conn->ibc_outstanding_credits += tx->tx_msg->ibm_credits;
768 conn->ibc_nsends_posted--;
771 tx->tx_passive_rdma_wait = 0;
772 tx->tx_sending -= tx->tx_nsp - nwork;
774 done = (tx->tx_sending == 0);
776 list_del (&tx->tx_list);
778 spin_unlock_irqrestore (&conn->ibc_lock, flags);
780 if (conn->ibc_state == IBNAL_CONN_ESTABLISHED)
781 CERROR ("Error %d posting transmit to "LPX64"\n",
782 rc, conn->ibc_peer->ibp_nid);
784 CDEBUG (D_NET, "Error %d posting transmit to "
785 LPX64"\n", rc, conn->ibc_peer->ibp_nid);
787 kibnal_close_conn (conn, rc);
796 spin_unlock_irqrestore (&conn->ibc_lock, flags);
800 kibnal_tx_callback (struct ib_cq_entry *e)
802 kib_tx_t *tx = (kib_tx_t *)kibnal_wreqid2ptr(e->work_request_id);
808 LASSERT (conn != NULL);
809 LASSERT (tx->tx_sending != 0);
811 spin_lock_irqsave(&conn->ibc_lock, flags);
813 CDEBUG(D_NET, "conn %p tx %p [%d/%d]: %d\n", conn, tx,
814 tx->tx_nsp - tx->tx_sending, tx->tx_nsp,
817 /* I could be racing with rdma completion. Whoever makes 'tx' idle
818 * gets to free it, which also drops its ref on 'conn'. If it's
819 * not me, then I take an extra ref on conn so it can't disappear
823 idle = (tx->tx_sending == 0) && /* This is the final callback */
824 (!tx->tx_passive_rdma_wait); /* Not waiting for RDMA completion */
826 list_del(&tx->tx_list);
828 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
829 conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
830 atomic_read (&conn->ibc_refcount));
831 atomic_inc (&conn->ibc_refcount);
833 if (tx->tx_sending == 0)
834 conn->ibc_nsends_posted--;
836 if (e->status != IB_COMPLETION_STATUS_SUCCESS &&
838 tx->tx_status = -ECONNABORTED;
840 spin_unlock_irqrestore(&conn->ibc_lock, flags);
845 if (e->status != IB_COMPLETION_STATUS_SUCCESS) {
846 CERROR ("Tx completion to "LPX64" failed: %d\n",
847 conn->ibc_peer->ibp_nid, e->status);
848 kibnal_close_conn (conn, -ENETDOWN);
850 /* can I shovel some more sends out the door? */
851 kibnal_check_sends(conn);
854 kibnal_put_conn (conn);
858 kibnal_callback (struct ib_cq *cq, struct ib_cq_entry *e, void *arg)
860 if (kibnal_wreqid_is_rx(e->work_request_id))
861 kibnal_rx_callback (e);
863 kibnal_tx_callback (e);
867 kibnal_init_tx_msg (kib_tx_t *tx, int type, int body_nob)
869 struct ib_gather_scatter *gl = &tx->tx_gl[tx->tx_nsp];
870 struct ib_send_param *sp = &tx->tx_sp[tx->tx_nsp];
872 int nob = offsetof (kib_msg_t, ibm_u) + body_nob;
874 LASSERT (tx->tx_nsp >= 0 &&
875 tx->tx_nsp < sizeof(tx->tx_sp)/sizeof(tx->tx_sp[0]));
876 LASSERT (nob <= IBNAL_MSG_SIZE);
878 kibnal_init_msg(tx->tx_msg, type, body_nob);
880 /* Fence the message if it's bundled with an RDMA read */
881 fence = (tx->tx_nsp > 0) &&
882 (type == IBNAL_MSG_PUT_DONE);
884 *gl = (struct ib_gather_scatter) {
885 .address = tx->tx_vaddr,
887 .key = kibnal_data.kib_tx_pages->ibp_lkey,
890 /* NB If this is an RDMA read, the completion message must wait for
891 * the RDMA to complete. Sends wait for previous RDMA writes
893 *sp = (struct ib_send_param) {
894 .work_request_id = kibnal_ptr2wreqid(tx, 0),
897 .num_gather_entries = 1,
898 .device_specific = NULL,
899 .solicited_event = 1,
901 .immediate_data_valid = 0,
910 kibnal_queue_tx (kib_tx_t *tx, kib_conn_t *conn)
914 spin_lock_irqsave(&conn->ibc_lock, flags);
916 kibnal_queue_tx_locked (tx, conn);
918 spin_unlock_irqrestore(&conn->ibc_lock, flags);
920 kibnal_check_sends(conn);
924 kibnal_launch_tx (kib_tx_t *tx, ptl_nid_t nid)
929 rwlock_t *g_lock = &kibnal_data.kib_global_lock;
931 /* If I get here, I've committed to send, so I complete the tx with
932 * failure on any problems */
934 LASSERT (tx->tx_conn == NULL); /* only set when assigned a conn */
935 LASSERT (tx->tx_nsp > 0); /* work items have been set up */
939 peer = kibnal_find_peer_locked (nid);
941 read_unlock (g_lock);
942 tx->tx_status = -EHOSTUNREACH;
947 conn = kibnal_find_conn_locked (peer);
949 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
950 conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
951 atomic_read (&conn->ibc_refcount));
952 atomic_inc (&conn->ibc_refcount); /* 1 ref for the tx */
953 read_unlock (g_lock);
955 kibnal_queue_tx (tx, conn);
959 /* Making one or more connections; I'll need a write lock... */
960 read_unlock (g_lock);
961 write_lock_irqsave (g_lock, flags);
963 peer = kibnal_find_peer_locked (nid);
965 write_unlock_irqrestore (g_lock, flags);
966 tx->tx_status = -EHOSTUNREACH;
971 conn = kibnal_find_conn_locked (peer);
973 /* Connection exists; queue message on it */
974 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
975 conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
976 atomic_read (&conn->ibc_refcount));
977 atomic_inc (&conn->ibc_refcount); /* 1 ref for the tx */
978 write_unlock_irqrestore (g_lock, flags);
980 kibnal_queue_tx (tx, conn);
984 if (peer->ibp_connecting == 0) {
985 if (!time_after_eq(jiffies, peer->ibp_reconnect_time)) {
986 write_unlock_irqrestore (g_lock, flags);
987 tx->tx_status = -EHOSTUNREACH;
992 peer->ibp_connecting = 1;
993 atomic_inc (&peer->ibp_refcount); /* extra ref for connd */
995 spin_lock (&kibnal_data.kib_connd_lock);
997 list_add_tail (&peer->ibp_connd_list,
998 &kibnal_data.kib_connd_peers);
999 wake_up (&kibnal_data.kib_connd_waitq);
1001 spin_unlock (&kibnal_data.kib_connd_lock);
1004 /* A connection is being established; queue the message... */
1005 list_add_tail (&tx->tx_list, &peer->ibp_tx_queue);
1007 write_unlock_irqrestore (g_lock, flags);
1011 kibnal_start_passive_rdma (int type, ptl_nid_t nid,
1012 lib_msg_t *libmsg, ptl_hdr_t *hdr)
1014 int nob = libmsg->md->length;
1020 LASSERT (type == IBNAL_MSG_PUT_RDMA ||
1021 type == IBNAL_MSG_GET_RDMA);
1023 LASSERT (!in_interrupt()); /* Mapping could block */
1025 if (type == IBNAL_MSG_PUT_RDMA) {
1026 access = IB_ACCESS_REMOTE_READ;
1028 access = IB_ACCESS_REMOTE_WRITE |
1029 IB_ACCESS_LOCAL_WRITE;
1032 tx = kibnal_get_idle_tx (1); /* May block; caller is an app thread */
1033 LASSERT (tx != NULL);
1035 if ((libmsg->md->options & PTL_MD_KIOV) == 0)
1036 rc = kibnal_map_iov (tx, access,
1037 libmsg->md->md_niov,
1038 libmsg->md->md_iov.iov,
1041 rc = kibnal_map_kiov (tx, access,
1042 libmsg->md->md_niov,
1043 libmsg->md->md_iov.kiov,
1047 CERROR ("Can't map RDMA for "LPX64": %d\n", nid, rc);
1051 if (type == IBNAL_MSG_GET_RDMA) {
1052 /* reply gets finalized when tx completes */
1053 tx->tx_libmsg[1] = lib_create_reply_msg(&kibnal_lib,
1055 if (tx->tx_libmsg[1] == NULL) {
1056 CERROR ("Can't create reply for GET -> "LPX64"\n",
1063 tx->tx_passive_rdma = 1;
1067 ibmsg->ibm_u.rdma.ibrm_hdr = *hdr;
1068 ibmsg->ibm_u.rdma.ibrm_cookie = tx->tx_passive_rdma_cookie;
1069 ibmsg->ibm_u.rdma.ibrm_desc.rd_key = tx->tx_md.md_rkey;
1070 ibmsg->ibm_u.rdma.ibrm_desc.rd_addr = tx->tx_md.md_addr;
1071 ibmsg->ibm_u.rdma.ibrm_desc.rd_nob = nob;
1073 kibnal_init_tx_msg (tx, type, sizeof (kib_rdma_msg_t));
1075 CDEBUG(D_NET, "Passive: %p cookie "LPX64", key %x, addr "
1077 tx, tx->tx_passive_rdma_cookie, tx->tx_md.md_rkey,
1078 tx->tx_md.md_addr, nob);
1080 /* libmsg gets finalized when tx completes. */
1081 tx->tx_libmsg[0] = libmsg;
1083 kibnal_launch_tx(tx, nid);
1088 kibnal_tx_done (tx);
1093 kibnal_start_active_rdma (int type, int status,
1094 kib_rx_t *rx, lib_msg_t *libmsg,
1096 struct iovec *iov, ptl_kiov_t *kiov,
1097 int offset, int nob)
1099 kib_msg_t *rxmsg = rx->rx_msg;
1106 CDEBUG(D_NET, "type %d, status %d, niov %d, offset %d, nob %d\n",
1107 type, status, niov, offset, nob);
1109 /* Called by scheduler */
1110 LASSERT (!in_interrupt ());
1112 /* Either all pages or all vaddrs */
1113 LASSERT (!(kiov != NULL && iov != NULL));
1115 /* No data if we're completing with failure */
1116 LASSERT (status == 0 || nob == 0);
1118 LASSERT (type == IBNAL_MSG_GET_DONE ||
1119 type == IBNAL_MSG_PUT_DONE);
1121 /* Flag I'm completing the RDMA. Even if I fail to send the
1122 * completion message, I will have tried my best so further
1123 * attempts shouldn't be tried. */
1124 LASSERT (!rx->rx_rdma);
1127 if (type == IBNAL_MSG_GET_DONE) {
1129 rdma_op = IB_OP_RDMA_WRITE;
1130 LASSERT (rxmsg->ibm_type == IBNAL_MSG_GET_RDMA);
1132 access = IB_ACCESS_LOCAL_WRITE;
1133 rdma_op = IB_OP_RDMA_READ;
1134 LASSERT (rxmsg->ibm_type == IBNAL_MSG_PUT_RDMA);
1137 tx = kibnal_get_idle_tx (0); /* Mustn't block */
1139 CERROR ("tx descs exhausted on RDMA from "LPX64
1140 " completing locally with failure\n",
1141 rx->rx_conn->ibc_peer->ibp_nid);
1142 lib_finalize (&kibnal_lib, NULL, libmsg, PTL_NO_SPACE);
1145 LASSERT (tx->tx_nsp == 0);
1148 /* We actually need to transfer some data (the transfer
1149 * size could get truncated to zero when the incoming
1150 * message is matched) */
1153 rc = kibnal_map_kiov (tx, access,
1154 niov, kiov, offset, nob);
1156 rc = kibnal_map_iov (tx, access,
1157 niov, iov, offset, nob);
1160 CERROR ("Can't map RDMA -> "LPX64": %d\n",
1161 rx->rx_conn->ibc_peer->ibp_nid, rc);
1162 /* We'll skip the RDMA and complete with failure. */
1166 tx->tx_gl[0] = (struct ib_gather_scatter) {
1167 .address = tx->tx_md.md_addr,
1169 .key = tx->tx_md.md_lkey,
1172 tx->tx_sp[0] = (struct ib_send_param) {
1173 .work_request_id = kibnal_ptr2wreqid(tx, 0),
1175 .gather_list = &tx->tx_gl[0],
1176 .num_gather_entries = 1,
1177 .remote_address = rxmsg->ibm_u.rdma.ibrm_desc.rd_addr,
1178 .rkey = rxmsg->ibm_u.rdma.ibrm_desc.rd_key,
1179 .device_specific = NULL,
1180 .solicited_event = 0,
1182 .immediate_data_valid = 0,
1193 txmsg->ibm_u.completion.ibcm_cookie = rxmsg->ibm_u.rdma.ibrm_cookie;
1194 txmsg->ibm_u.completion.ibcm_status = status;
1196 kibnal_init_tx_msg(tx, type, sizeof (kib_completion_msg_t));
1198 if (status == 0 && nob != 0) {
1199 LASSERT (tx->tx_nsp > 1);
1200 /* RDMA: libmsg gets finalized when the tx completes. This
1201 * is after the completion message has been sent, which in
1202 * turn is after the RDMA has finished. */
1203 tx->tx_libmsg[0] = libmsg;
1205 LASSERT (tx->tx_nsp == 1);
1206 /* No RDMA: local completion happens now! */
1207 CDEBUG(D_NET, "No data: immediate completion\n");
1208 lib_finalize (&kibnal_lib, NULL, libmsg,
1209 status == 0 ? PTL_OK : PTL_FAIL);
1212 /* +1 ref for this tx... */
1213 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
1214 rx->rx_conn, rx->rx_conn->ibc_state,
1215 rx->rx_conn->ibc_peer->ibp_nid,
1216 atomic_read (&rx->rx_conn->ibc_refcount));
1217 atomic_inc (&rx->rx_conn->ibc_refcount);
1218 /* ...and queue it up */
1219 kibnal_queue_tx(tx, rx->rx_conn);
1223 kibnal_sendmsg(lib_nal_t *nal,
1230 unsigned int payload_niov,
1231 struct iovec *payload_iov,
1232 ptl_kiov_t *payload_kiov,
1240 /* NB 'private' is different depending on what we're sending.... */
1242 CDEBUG(D_NET, "sending %d bytes in %d frags to nid:"LPX64" pid %d\n",
1243 payload_nob, payload_niov, nid , pid);
1245 LASSERT (payload_nob == 0 || payload_niov > 0);
1246 LASSERT (payload_niov <= PTL_MD_MAX_IOV);
1248 /* Thread context if we're sending payload */
1249 LASSERT (!in_interrupt() || payload_niov == 0);
1250 /* payload is either all vaddrs or all pages */
1251 LASSERT (!(payload_kiov != NULL && payload_iov != NULL));
1258 case PTL_MSG_REPLY: {
1259 /* reply's 'private' is the incoming receive */
1260 kib_rx_t *rx = private;
1262 /* RDMA reply expected? */
1263 if (rx->rx_msg->ibm_type == IBNAL_MSG_GET_RDMA) {
1264 kibnal_start_active_rdma(IBNAL_MSG_GET_DONE, 0,
1265 rx, libmsg, payload_niov,
1266 payload_iov, payload_kiov,
1267 payload_offset, payload_nob);
1271 /* Incoming message consistent with immediate reply? */
1272 if (rx->rx_msg->ibm_type != IBNAL_MSG_IMMEDIATE) {
1273 CERROR ("REPLY to "LPX64" bad opbm type %d!!!\n",
1274 nid, rx->rx_msg->ibm_type);
1278 /* Will it fit in a message? */
1279 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1280 if (nob >= IBNAL_MSG_SIZE) {
1281 CERROR("REPLY for "LPX64" too big (RDMA not requested): %d\n",
1289 /* might the REPLY message be big enough to need RDMA? */
1290 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[libmsg->md->length]);
1291 if (nob > IBNAL_MSG_SIZE)
1292 return (kibnal_start_passive_rdma(IBNAL_MSG_GET_RDMA,
1297 LASSERT (payload_nob == 0);
1301 /* Is the payload big enough to need RDMA? */
1302 nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[payload_nob]);
1303 if (nob > IBNAL_MSG_SIZE)
1304 return (kibnal_start_passive_rdma(IBNAL_MSG_PUT_RDMA,
1310 tx = kibnal_get_idle_tx(!(type == PTL_MSG_ACK ||
1311 type == PTL_MSG_REPLY ||
1314 CERROR ("Can't send %d to "LPX64": tx descs exhausted%s\n",
1315 type, nid, in_interrupt() ? " (intr)" : "");
1316 return (PTL_NO_SPACE);
1320 ibmsg->ibm_u.immediate.ibim_hdr = *hdr;
1322 if (payload_nob > 0) {
1323 if (payload_kiov != NULL)
1324 lib_copy_kiov2buf(ibmsg->ibm_u.immediate.ibim_payload,
1325 payload_niov, payload_kiov,
1326 payload_offset, payload_nob);
1328 lib_copy_iov2buf(ibmsg->ibm_u.immediate.ibim_payload,
1329 payload_niov, payload_iov,
1330 payload_offset, payload_nob);
1333 kibnal_init_tx_msg (tx, IBNAL_MSG_IMMEDIATE,
1334 offsetof(kib_immediate_msg_t,
1335 ibim_payload[payload_nob]));
1337 /* libmsg gets finalized when tx completes */
1338 tx->tx_libmsg[0] = libmsg;
1340 kibnal_launch_tx(tx, nid);
1345 kibnal_send (lib_nal_t *nal, void *private, lib_msg_t *cookie,
1346 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
1347 unsigned int payload_niov, struct iovec *payload_iov,
1348 size_t payload_offset, size_t payload_len)
1350 return (kibnal_sendmsg(nal, private, cookie,
1351 hdr, type, nid, pid,
1352 payload_niov, payload_iov, NULL,
1353 payload_offset, payload_len));
1357 kibnal_send_pages (lib_nal_t *nal, void *private, lib_msg_t *cookie,
1358 ptl_hdr_t *hdr, int type, ptl_nid_t nid, ptl_pid_t pid,
1359 unsigned int payload_niov, ptl_kiov_t *payload_kiov,
1360 size_t payload_offset, size_t payload_len)
1362 return (kibnal_sendmsg(nal, private, cookie,
1363 hdr, type, nid, pid,
1364 payload_niov, NULL, payload_kiov,
1365 payload_offset, payload_len));
1369 kibnal_recvmsg (lib_nal_t *nal, void *private, lib_msg_t *libmsg,
1370 unsigned int niov, struct iovec *iov, ptl_kiov_t *kiov,
1371 int offset, int mlen, int rlen)
1373 kib_rx_t *rx = private;
1374 kib_msg_t *rxmsg = rx->rx_msg;
1377 LASSERT (mlen <= rlen);
1378 LASSERT (!in_interrupt ());
1379 /* Either all pages or all vaddrs */
1380 LASSERT (!(kiov != NULL && iov != NULL));
1382 switch (rxmsg->ibm_type) {
1387 case IBNAL_MSG_IMMEDIATE:
1388 msg_nob = offsetof(kib_msg_t, ibm_u.immediate.ibim_payload[rlen]);
1389 if (msg_nob > IBNAL_MSG_SIZE) {
1390 CERROR ("Immediate message from "LPX64" too big: %d\n",
1391 rxmsg->ibm_u.immediate.ibim_hdr.src_nid, rlen);
1396 lib_copy_buf2kiov(niov, kiov, offset,
1397 rxmsg->ibm_u.immediate.ibim_payload,
1400 lib_copy_buf2iov(niov, iov, offset,
1401 rxmsg->ibm_u.immediate.ibim_payload,
1404 lib_finalize (nal, NULL, libmsg, PTL_OK);
1407 case IBNAL_MSG_GET_RDMA:
1408 /* We get called here just to discard any junk after the
1410 LASSERT (libmsg == NULL);
1411 lib_finalize (nal, NULL, libmsg, PTL_OK);
1414 case IBNAL_MSG_PUT_RDMA:
1415 kibnal_start_active_rdma (IBNAL_MSG_PUT_DONE, 0,
1417 niov, iov, kiov, offset, mlen);
1423 kibnal_recv (lib_nal_t *nal, void *private, lib_msg_t *msg,
1424 unsigned int niov, struct iovec *iov,
1425 size_t offset, size_t mlen, size_t rlen)
1427 return (kibnal_recvmsg (nal, private, msg, niov, iov, NULL,
1428 offset, mlen, rlen));
1432 kibnal_recv_pages (lib_nal_t *nal, void *private, lib_msg_t *msg,
1433 unsigned int niov, ptl_kiov_t *kiov,
1434 size_t offset, size_t mlen, size_t rlen)
1436 return (kibnal_recvmsg (nal, private, msg, niov, NULL, kiov,
1437 offset, mlen, rlen));
1441 kibnal_thread_start (int (*fn)(void *arg), void *arg)
1443 long pid = kernel_thread (fn, arg, 0);
1448 atomic_inc (&kibnal_data.kib_nthreads);
1453 kibnal_thread_fini (void)
1455 atomic_dec (&kibnal_data.kib_nthreads);
1459 kibnal_close_conn_locked (kib_conn_t *conn, int error)
1461 /* This just does the immmediate housekeeping, and schedules the
1462 * connection for the reaper to finish off.
1463 * Caller holds kib_global_lock exclusively in irq context */
1464 kib_peer_t *peer = conn->ibc_peer;
1466 CDEBUG (error == 0 ? D_NET : D_ERROR,
1467 "closing conn to "LPX64": error %d\n", peer->ibp_nid, error);
1469 LASSERT (conn->ibc_state == IBNAL_CONN_ESTABLISHED ||
1470 conn->ibc_state == IBNAL_CONN_CONNECTING);
1472 if (conn->ibc_state == IBNAL_CONN_ESTABLISHED) {
1473 /* kib_reaper_conns takes ibc_list's ref */
1474 list_del (&conn->ibc_list);
1476 /* new ref for kib_reaper_conns */
1477 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
1478 conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
1479 atomic_read (&conn->ibc_refcount));
1480 atomic_inc (&conn->ibc_refcount);
1483 if (list_empty (&peer->ibp_conns) &&
1484 peer->ibp_persistence == 0) {
1485 /* Non-persistent peer with no more conns... */
1486 kibnal_unlink_peer_locked (peer);
1489 conn->ibc_state = IBNAL_CONN_DEATHROW;
1491 /* Schedule conn for closing/destruction */
1492 spin_lock (&kibnal_data.kib_reaper_lock);
1494 list_add_tail (&conn->ibc_list, &kibnal_data.kib_reaper_conns);
1495 wake_up (&kibnal_data.kib_reaper_waitq);
1497 spin_unlock (&kibnal_data.kib_reaper_lock);
1501 kibnal_close_conn (kib_conn_t *conn, int why)
1503 unsigned long flags;
1506 write_lock_irqsave (&kibnal_data.kib_global_lock, flags);
1508 LASSERT (conn->ibc_state >= IBNAL_CONN_CONNECTING);
1510 if (conn->ibc_state <= IBNAL_CONN_ESTABLISHED) {
1512 kibnal_close_conn_locked (conn, why);
1515 write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
1520 kibnal_peer_connect_failed (kib_peer_t *peer, int active, int rc)
1522 LIST_HEAD (zombies);
1524 unsigned long flags;
1527 LASSERT (peer->ibp_reconnect_interval >= IBNAL_MIN_RECONNECT_INTERVAL);
1529 write_lock_irqsave (&kibnal_data.kib_global_lock, flags);
1531 LASSERT (peer->ibp_connecting != 0);
1532 peer->ibp_connecting--;
1534 if (peer->ibp_connecting != 0) {
1535 /* another connection attempt under way (loopback?)... */
1536 write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
1540 if (list_empty(&peer->ibp_conns)) {
1541 /* Say when active connection can be re-attempted */
1542 peer->ibp_reconnect_time = jiffies + peer->ibp_reconnect_interval;
1543 /* Increase reconnection interval */
1544 peer->ibp_reconnect_interval = MIN (peer->ibp_reconnect_interval * 2,
1545 IBNAL_MAX_RECONNECT_INTERVAL);
1547 /* Take peer's blocked blocked transmits; I'll complete
1548 * them with error */
1549 while (!list_empty (&peer->ibp_tx_queue)) {
1550 tx = list_entry (peer->ibp_tx_queue.next,
1553 list_del (&tx->tx_list);
1554 list_add_tail (&tx->tx_list, &zombies);
1557 if (kibnal_peer_active(peer) &&
1558 (peer->ibp_persistence == 0)) {
1559 /* failed connection attempt on non-persistent peer */
1560 kibnal_unlink_peer_locked (peer);
1563 /* Can't have blocked transmits if there are connections */
1564 LASSERT (list_empty(&peer->ibp_tx_queue));
1567 write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
1569 if (!list_empty (&zombies))
1570 CERROR ("Deleting messages for "LPX64": connection failed\n",
1573 while (!list_empty (&zombies)) {
1574 tx = list_entry (zombies.next, kib_tx_t, tx_list);
1576 list_del (&tx->tx_list);
1578 tx->tx_status = -EHOSTUNREACH;
1579 kibnal_tx_done (tx);
1584 kibnal_connreq_done (kib_conn_t *conn, int active, int status)
1586 int state = conn->ibc_state;
1587 kib_peer_t *peer = conn->ibc_peer;
1589 unsigned long flags;
1593 /* passive connection has no connreq & vice versa */
1594 LASSERT (!active == !(conn->ibc_connreq != NULL));
1596 PORTAL_FREE (conn->ibc_connreq, sizeof (*conn->ibc_connreq));
1597 conn->ibc_connreq = NULL;
1600 if (state == IBNAL_CONN_CONNECTING) {
1601 /* Install common (active/passive) callback for
1602 * disconnect/idle notification if I got as far as getting
1604 rc = tsIbCmCallbackModify(conn->ibc_comm_id,
1605 kibnal_conn_callback, conn);
1609 write_lock_irqsave (&kibnal_data.kib_global_lock, flags);
1611 LASSERT (peer->ibp_connecting != 0);
1614 /* connection established... */
1615 LASSERT (state == IBNAL_CONN_CONNECTING);
1616 conn->ibc_state = IBNAL_CONN_ESTABLISHED;
1618 if (!kibnal_peer_active(peer)) {
1619 /* ...but peer deleted meantime */
1620 status = -ECONNABORTED;
1623 LASSERT (state == IBNAL_CONN_INIT_QP ||
1624 state == IBNAL_CONN_CONNECTING);
1628 /* Everything worked! */
1630 #warning "purge old conn incarnations"
1632 peer->ibp_connecting--;
1634 /* +1 ref for ibc_list; caller(== CM)'s ref remains until
1635 * the IB_CM_IDLE callback */
1636 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
1637 conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
1638 atomic_read (&conn->ibc_refcount));
1639 atomic_inc (&conn->ibc_refcount);
1640 list_add (&conn->ibc_list, &peer->ibp_conns);
1642 /* reset reconnect interval for next attempt */
1643 peer->ibp_reconnect_interval = IBNAL_MIN_RECONNECT_INTERVAL;
1645 /* post blocked sends to the new connection */
1646 spin_lock (&conn->ibc_lock);
1648 while (!list_empty (&peer->ibp_tx_queue)) {
1649 tx = list_entry (peer->ibp_tx_queue.next,
1652 list_del (&tx->tx_list);
1654 /* +1 ref for each tx */
1655 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
1656 conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
1657 atomic_read (&conn->ibc_refcount));
1658 atomic_inc (&conn->ibc_refcount);
1659 kibnal_queue_tx_locked (tx, conn);
1662 spin_unlock (&conn->ibc_lock);
1664 /* Nuke any dangling conns from a different peer instance... */
1665 kibnal_close_stale_conns_locked (conn->ibc_peer,
1666 conn->ibc_incarnation);
1668 write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
1670 /* queue up all the receives */
1671 for (i = 0; i < IBNAL_RX_MSGS; i++) {
1672 /* +1 ref for rx desc */
1673 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
1674 conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
1675 atomic_read (&conn->ibc_refcount));
1676 atomic_inc (&conn->ibc_refcount);
1678 CDEBUG(D_NET, "RX[%d] %p->%p - "LPX64"\n",
1679 i, &conn->ibc_rxs[i], conn->ibc_rxs[i].rx_msg,
1680 conn->ibc_rxs[i].rx_vaddr);
1682 kibnal_post_rx (&conn->ibc_rxs[i], 0);
1685 kibnal_check_sends (conn);
1689 /* connection failed */
1690 if (state == IBNAL_CONN_CONNECTING) {
1691 /* schedule for reaper to close */
1692 kibnal_close_conn_locked (conn, status);
1694 /* Don't have a CM comm_id; just wait for refs to drain */
1695 conn->ibc_state = IBNAL_CONN_ZOMBIE;
1698 write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
1700 kibnal_peer_connect_failed (conn->ibc_peer, active, status);
1702 if (state != IBNAL_CONN_CONNECTING) {
1703 /* drop caller's ref if we're not waiting for the
1704 * IB_CM_IDLE callback */
1705 kibnal_put_conn (conn);
1710 kibnal_accept (kib_conn_t **connp, tTS_IB_CM_COMM_ID cid,
1711 kib_msg_t *msg, int nob)
1716 unsigned long flags;
1719 rc = kibnal_unpack_msg(msg, nob);
1721 CERROR("Can't unpack connreq msg: %d\n", rc);
1725 CDEBUG(D_NET, "connreq from "LPX64"\n", msg->ibm_srcnid);
1727 if (msg->ibm_type != IBNAL_MSG_CONNREQ) {
1728 CERROR("Unexpected connreq msg type: %x from "LPX64"\n",
1729 msg->ibm_type, msg->ibm_srcnid);
1733 if (msg->ibm_u.connparams.ibcp_queue_depth != IBNAL_MSG_QUEUE_SIZE) {
1734 CERROR("Can't accept "LPX64": bad queue depth %d (%d expected)\n",
1735 msg->ibm_srcnid, msg->ibm_u.connparams.ibcp_queue_depth,
1736 IBNAL_MSG_QUEUE_SIZE);
1740 conn = kibnal_create_conn();
1744 /* assume 'nid' is a new peer */
1745 peer = kibnal_create_peer (msg->ibm_srcnid);
1747 CDEBUG(D_NET, "--conn[%p] state %d -> "LPX64" (%d)\n",
1748 conn, conn->ibc_state, conn->ibc_peer->ibp_nid,
1749 atomic_read (&conn->ibc_refcount));
1750 atomic_dec (&conn->ibc_refcount);
1751 kibnal_destroy_conn(conn);
1755 write_lock_irqsave (&kibnal_data.kib_global_lock, flags);
1757 /* Check I'm the same instance that gave the connection parameters.
1758 * NB If my incarnation changes after this, the peer will get nuked and
1759 * we'll spot that when the connection is finally added into the peer's
1761 if (msg->ibm_dstnid != kibnal_lib.libnal_ni.ni_pid.nid ||
1762 msg->ibm_dststamp != kibnal_data.kib_incarnation) {
1763 write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
1765 CERROR("Stale connection params from "LPX64"\n",
1767 atomic_dec(&conn->ibc_refcount);
1768 kibnal_destroy_conn(conn);
1769 kibnal_put_peer(peer);
1773 peer2 = kibnal_find_peer_locked(msg->ibm_srcnid);
1774 if (peer2 == NULL) {
1775 /* peer table takes my ref on peer */
1776 list_add_tail (&peer->ibp_list,
1777 kibnal_nid2peerlist(msg->ibm_srcnid));
1779 kibnal_put_peer (peer);
1783 /* +1 ref for conn */
1784 atomic_inc (&peer->ibp_refcount);
1785 peer->ibp_connecting++;
1787 write_unlock_irqrestore (&kibnal_data.kib_global_lock, flags);
1789 conn->ibc_peer = peer;
1790 conn->ibc_state = IBNAL_CONN_CONNECTING;
1791 conn->ibc_comm_id = cid;
1792 conn->ibc_incarnation = msg->ibm_srcstamp;
1793 conn->ibc_credits = IBNAL_MSG_QUEUE_SIZE;
1799 tTS_IB_CM_CALLBACK_RETURN
1800 kibnal_idle_conn_callback (tTS_IB_CM_EVENT event,
1801 tTS_IB_CM_COMM_ID cid,
1805 /* Shouldn't ever get a callback after TS_IB_CM_IDLE */
1806 CERROR ("Unexpected event %d: conn %p\n", event, arg);
1808 return TS_IB_CM_CALLBACK_PROCEED;
1811 tTS_IB_CM_CALLBACK_RETURN
1812 kibnal_conn_callback (tTS_IB_CM_EVENT event,
1813 tTS_IB_CM_COMM_ID cid,
1817 kib_conn_t *conn = arg;
1818 LIST_HEAD (zombies);
1819 struct list_head *tmp;
1820 struct list_head *nxt;
1822 unsigned long flags;
1826 /* Established Connection Notifier */
1830 CERROR("Connection %p -> "LPX64" ERROR %d\n",
1831 conn, conn->ibc_peer->ibp_nid, event);
1832 kibnal_close_conn (conn, -ECONNABORTED);
1835 case TS_IB_CM_DISCONNECTED:
1836 CDEBUG(D_WARNING, "Connection %p -> "LPX64" DISCONNECTED.\n",
1837 conn, conn->ibc_peer->ibp_nid);
1838 kibnal_close_conn (conn, 0);
1842 CDEBUG(D_NET, "Connection %p -> "LPX64" IDLE.\n",
1843 conn, conn->ibc_peer->ibp_nid);
1844 kibnal_put_conn (conn); /* Lose CM's ref */
1846 /* LASSERT (no further callbacks) */
1847 rc = tsIbCmCallbackModify(cid,
1848 kibnal_idle_conn_callback, conn);
1851 /* NB we wait until the connection has closed before
1852 * completing outstanding passive RDMAs so we can be sure
1853 * the network can't touch the mapped memory any more. */
1855 spin_lock_irqsave (&conn->ibc_lock, flags);
1857 /* grab passive RDMAs not waiting for the tx callback */
1858 list_for_each_safe (tmp, nxt, &conn->ibc_active_txs) {
1859 tx = list_entry (tmp, kib_tx_t, tx_list);
1861 LASSERT (tx->tx_passive_rdma ||
1862 !tx->tx_passive_rdma_wait);
1864 LASSERT (tx->tx_passive_rdma_wait ||
1865 tx->tx_sending != 0);
1867 /* still waiting for tx callback? */
1868 if (!tx->tx_passive_rdma_wait)
1871 tx->tx_status = -ECONNABORTED;
1872 tx->tx_passive_rdma_wait = 0;
1873 done = (tx->tx_sending == 0);
1878 list_del (&tx->tx_list);
1879 list_add (&tx->tx_list, &zombies);
1882 /* grab all blocked transmits */
1883 list_for_each_safe (tmp, nxt, &conn->ibc_tx_queue) {
1884 tx = list_entry (tmp, kib_tx_t, tx_list);
1886 list_del (&tx->tx_list);
1887 list_add (&tx->tx_list, &zombies);
1890 spin_unlock_irqrestore (&conn->ibc_lock, flags);
1892 while (!list_empty(&zombies)) {
1893 tx = list_entry (zombies.next, kib_tx_t, tx_list);
1895 list_del(&tx->tx_list);
1896 kibnal_tx_done (tx);
1901 return TS_IB_CM_CALLBACK_PROCEED;
1904 tTS_IB_CM_CALLBACK_RETURN
1905 kibnal_passive_conn_callback (tTS_IB_CM_EVENT event,
1906 tTS_IB_CM_COMM_ID cid,
1910 kib_conn_t *conn = arg;
1916 /* no connection yet */
1917 CERROR ("Unexpected event: %d\n", event);
1918 return TS_IB_CM_CALLBACK_ABORT;
1921 CERROR ("Unexpected event %p -> "LPX64": %d\n",
1922 conn, conn->ibc_peer->ibp_nid, event);
1923 kibnal_connreq_done (conn, 0, -ECONNABORTED);
1926 case TS_IB_CM_REQ_RECEIVED: {
1927 struct ib_cm_req_received_param *req = param;
1928 kib_msg_t *msg = req->remote_private_data;
1930 LASSERT (conn == NULL);
1932 /* Don't really know srcnid until successful unpack */
1933 CDEBUG(D_NET, "REQ from ?"LPX64"?\n", msg->ibm_srcnid);
1935 rc = kibnal_accept(&conn, cid, msg,
1936 req->remote_private_data_len);
1938 CERROR ("Can't accept ?"LPX64"?: %d\n",
1939 msg->ibm_srcnid, rc);
1940 return TS_IB_CM_CALLBACK_ABORT;
1943 /* update 'arg' for next callback */
1944 rc = tsIbCmCallbackModify(cid, kibnal_passive_conn_callback, conn);
1947 msg = req->accept_param.reply_private_data;
1948 kibnal_init_msg(msg, IBNAL_MSG_CONNACK,
1949 sizeof(msg->ibm_u.connparams));
1951 msg->ibm_u.connparams.ibcp_queue_depth = IBNAL_MSG_QUEUE_SIZE;
1953 kibnal_pack_msg(msg, 0,
1954 conn->ibc_peer->ibp_nid,
1955 conn->ibc_incarnation);
1957 req->accept_param.qp = conn->ibc_qp;
1958 req->accept_param.reply_private_data_len = msg->ibm_nob;
1959 req->accept_param.responder_resources = IBNAL_RESPONDER_RESOURCES;
1960 req->accept_param.initiator_depth = IBNAL_RESPONDER_RESOURCES;
1961 req->accept_param.rnr_retry_count = IBNAL_RNR_RETRY;
1962 req->accept_param.flow_control = IBNAL_FLOW_CONTROL;
1964 CDEBUG(D_NET, "Proceeding\n");
1968 case TS_IB_CM_ESTABLISHED:
1969 LASSERT (conn != NULL);
1970 CDEBUG(D_WARNING, "Connection %p -> "LPX64" ESTABLISHED.\n",
1971 conn, conn->ibc_peer->ibp_nid);
1973 kibnal_connreq_done (conn, 0, 0);
1977 /* NB if the connreq is done, we switch to kibnal_conn_callback */
1978 return TS_IB_CM_CALLBACK_PROCEED;
1981 tTS_IB_CM_CALLBACK_RETURN
1982 kibnal_active_conn_callback (tTS_IB_CM_EVENT event,
1983 tTS_IB_CM_COMM_ID cid,
1987 kib_conn_t *conn = arg;
1990 case TS_IB_CM_REP_RECEIVED: {
1991 struct ib_cm_rep_received_param *rep = param;
1992 kib_msg_t *msg = rep->remote_private_data;
1993 int nob = rep->remote_private_data_len;
1996 rc = kibnal_unpack_msg(msg, nob);
1998 CERROR ("Error %d unpacking conn ack from "LPX64"\n",
1999 rc, conn->ibc_peer->ibp_nid);
2000 kibnal_connreq_done (conn, 1, rc);
2004 if (msg->ibm_type != IBNAL_MSG_CONNACK) {
2005 CERROR ("Unexpected conn ack type %d from "LPX64"\n",
2006 msg->ibm_type, conn->ibc_peer->ibp_nid);
2007 kibnal_connreq_done (conn, 1, -EPROTO);
2011 if (msg->ibm_srcnid != conn->ibc_peer->ibp_nid ||
2012 msg->ibm_srcstamp != conn->ibc_incarnation ||
2013 msg->ibm_dstnid != kibnal_lib.libnal_ni.ni_pid.nid ||
2014 msg->ibm_dststamp != kibnal_data.kib_incarnation) {
2015 CERROR("Stale conn ack from "LPX64"\n",
2016 conn->ibc_peer->ibp_nid);
2017 kibnal_connreq_done (conn, 1, -ESTALE);
2021 if (msg->ibm_u.connparams.ibcp_queue_depth != IBNAL_MSG_QUEUE_SIZE) {
2022 CERROR ("Bad queue depth %d from "LPX64"\n",
2023 msg->ibm_u.connparams.ibcp_queue_depth,
2024 conn->ibc_peer->ibp_nid);
2025 kibnal_connreq_done (conn, 1, -EPROTO);
2029 CDEBUG(D_NET, "Connection %p -> "LPX64" REP_RECEIVED.\n",
2030 conn, conn->ibc_peer->ibp_nid);
2032 conn->ibc_credits = IBNAL_MSG_QUEUE_SIZE;
2036 case TS_IB_CM_ESTABLISHED:
2037 CDEBUG(D_WARNING, "Connection %p -> "LPX64" Established\n",
2038 conn, conn->ibc_peer->ibp_nid);
2040 kibnal_connreq_done (conn, 1, 0);
2044 CERROR("Connection %p -> "LPX64" IDLE\n",
2045 conn, conn->ibc_peer->ibp_nid);
2046 /* Back out state change: I'm disengaged from CM */
2047 conn->ibc_state = IBNAL_CONN_INIT_QP;
2049 kibnal_connreq_done (conn, 1, -ECONNABORTED);
2053 CERROR("Connection %p -> "LPX64" ERROR %d\n",
2054 conn, conn->ibc_peer->ibp_nid, event);
2055 kibnal_connreq_done (conn, 1, -ECONNABORTED);
2059 /* NB if the connreq is done, we switch to kibnal_conn_callback */
2060 return TS_IB_CM_CALLBACK_PROCEED;
2064 kibnal_pathreq_callback (tTS_IB_CLIENT_QUERY_TID tid, int status,
2065 struct ib_path_record *resp, int remaining,
2068 kib_conn_t *conn = arg;
2069 kib_peer_t *peer = conn->ibc_peer;
2070 kib_msg_t *msg = &conn->ibc_connreq->cr_msg;
2073 CERROR ("status %d\n", status);
2074 kibnal_connreq_done (conn, 1, status);
2078 conn->ibc_connreq->cr_path = *resp;
2080 kibnal_init_msg(msg, IBNAL_MSG_CONNREQ, sizeof(msg->ibm_u.connparams));
2081 msg->ibm_u.connparams.ibcp_queue_depth = IBNAL_MSG_QUEUE_SIZE;
2082 kibnal_pack_msg(msg, 0, peer->ibp_nid, conn->ibc_incarnation);
2084 conn->ibc_connreq->cr_connparam = (struct ib_cm_active_param) {
2086 .req_private_data = msg,
2087 .req_private_data_len = msg->ibm_nob,
2088 .responder_resources = IBNAL_RESPONDER_RESOURCES,
2089 .initiator_depth = IBNAL_RESPONDER_RESOURCES,
2090 .retry_count = IBNAL_RETRY,
2091 .rnr_retry_count = IBNAL_RNR_RETRY,
2092 .cm_response_timeout = kibnal_tunables.kib_io_timeout,
2093 .max_cm_retries = IBNAL_CM_RETRY,
2094 .flow_control = IBNAL_FLOW_CONTROL,
2097 /* XXX set timeout just like SDP!!!*/
2098 conn->ibc_connreq->cr_path.packet_life = 13;
2100 /* Flag I'm getting involved with the CM... */
2101 conn->ibc_state = IBNAL_CONN_CONNECTING;
2103 CDEBUG(D_WARNING, "Connecting to, service id "LPX64", on "LPX64"\n",
2104 conn->ibc_connreq->cr_svcrsp.ibsr_svc_id, peer->ibp_nid);
2106 /* kibnal_connect_callback gets my conn ref */
2107 status = ib_cm_connect (&conn->ibc_connreq->cr_connparam,
2108 &conn->ibc_connreq->cr_path, NULL,
2109 conn->ibc_connreq->cr_svcrsp.ibsr_svc_id, 0,
2110 kibnal_active_conn_callback, conn,
2111 &conn->ibc_comm_id);
2113 CERROR ("Connect: %d\n", status);
2114 /* Back out state change: I've not got a CM comm_id yet... */
2115 conn->ibc_state = IBNAL_CONN_INIT_QP;
2116 kibnal_connreq_done (conn, 1, status);
2120 /* return non-zero to prevent further callbacks */
2125 kibnal_connect_peer (kib_peer_t *peer)
2130 conn = kibnal_create_conn();
2132 CERROR ("Can't allocate conn\n");
2133 kibnal_peer_connect_failed (peer, 1, -ENOMEM);
2137 conn->ibc_peer = peer;
2138 atomic_inc (&peer->ibp_refcount);
2140 PORTAL_ALLOC (conn->ibc_connreq, sizeof (*conn->ibc_connreq));
2141 if (conn->ibc_connreq == NULL) {
2142 CERROR ("Can't allocate connreq\n");
2143 kibnal_connreq_done (conn, 1, -ENOMEM);
2147 memset(conn->ibc_connreq, 0, sizeof (*conn->ibc_connreq));
2149 rc = kibnal_make_svcqry(conn);
2151 kibnal_connreq_done (conn, 1, rc);
2155 rc = ib_cached_gid_get(kibnal_data.kib_device,
2156 kibnal_data.kib_port, 0,
2157 conn->ibc_connreq->cr_gid);
2160 /* kibnal_pathreq_callback gets my conn ref */
2161 rc = tsIbPathRecordRequest (kibnal_data.kib_device,
2162 kibnal_data.kib_port,
2163 conn->ibc_connreq->cr_gid,
2164 conn->ibc_connreq->cr_svcrsp.ibsr_svc_gid,
2165 conn->ibc_connreq->cr_svcrsp.ibsr_svc_pkey,
2167 kibnal_tunables.kib_io_timeout * HZ,
2169 kibnal_pathreq_callback, conn,
2170 &conn->ibc_connreq->cr_tid);
2174 CERROR ("Path record request: %d\n", rc);
2175 kibnal_connreq_done (conn, 1, rc);
2179 kibnal_conn_timed_out (kib_conn_t *conn)
2182 struct list_head *ttmp;
2183 unsigned long flags;
2185 spin_lock_irqsave (&conn->ibc_lock, flags);
2187 list_for_each (ttmp, &conn->ibc_tx_queue) {
2188 tx = list_entry (ttmp, kib_tx_t, tx_list);
2190 LASSERT (!tx->tx_passive_rdma_wait);
2191 LASSERT (tx->tx_sending == 0);
2193 if (time_after_eq (jiffies, tx->tx_deadline)) {
2194 spin_unlock_irqrestore (&conn->ibc_lock, flags);
2199 list_for_each (ttmp, &conn->ibc_active_txs) {
2200 tx = list_entry (ttmp, kib_tx_t, tx_list);
2202 LASSERT (tx->tx_passive_rdma ||
2203 !tx->tx_passive_rdma_wait);
2205 LASSERT (tx->tx_passive_rdma_wait ||
2206 tx->tx_sending != 0);
2208 if (time_after_eq (jiffies, tx->tx_deadline)) {
2209 spin_unlock_irqrestore (&conn->ibc_lock, flags);
2214 spin_unlock_irqrestore (&conn->ibc_lock, flags);
2220 kibnal_check_conns (int idx)
2222 struct list_head *peers = &kibnal_data.kib_peers[idx];
2223 struct list_head *ptmp;
2226 struct list_head *ctmp;
2229 /* NB. We expect to have a look at all the peers and not find any
2230 * rdmas to time out, so we just use a shared lock while we
2232 read_lock (&kibnal_data.kib_global_lock);
2234 list_for_each (ptmp, peers) {
2235 peer = list_entry (ptmp, kib_peer_t, ibp_list);
2237 list_for_each (ctmp, &peer->ibp_conns) {
2238 conn = list_entry (ctmp, kib_conn_t, ibc_list);
2240 LASSERT (conn->ibc_state == IBNAL_CONN_ESTABLISHED);
2243 /* In case we have enough credits to return via a
2244 * NOOP, but there were no non-blocking tx descs
2245 * free to do it last time... */
2246 kibnal_check_sends(conn);
2248 if (!kibnal_conn_timed_out(conn))
2251 CDEBUG(D_NET, "++conn[%p] state %d -> "LPX64" (%d)\n",
2252 conn, conn->ibc_state, peer->ibp_nid,
2253 atomic_read (&conn->ibc_refcount));
2255 atomic_inc (&conn->ibc_refcount);
2256 read_unlock (&kibnal_data.kib_global_lock);
2258 CERROR("Timed out RDMA with "LPX64"\n",
2261 kibnal_close_conn (conn, -ETIMEDOUT);
2262 kibnal_put_conn (conn);
2264 /* start again now I've dropped the lock */
2269 read_unlock (&kibnal_data.kib_global_lock);
2273 kibnal_terminate_conn (kib_conn_t *conn)
2277 CDEBUG(D_NET, "conn %p\n", conn);
2278 LASSERT (conn->ibc_state == IBNAL_CONN_DEATHROW);
2279 conn->ibc_state = IBNAL_CONN_ZOMBIE;
2281 rc = ib_cm_disconnect (conn->ibc_comm_id);
2283 CERROR ("Error %d disconnecting conn %p -> "LPX64"\n",
2284 rc, conn, conn->ibc_peer->ibp_nid);
2288 kibnal_reaper (void *arg)
2291 unsigned long flags;
2296 unsigned long deadline = jiffies;
2298 kportal_daemonize ("kibnal_reaper");
2299 kportal_blockallsigs ();
2301 init_waitqueue_entry (&wait, current);
2303 spin_lock_irqsave (&kibnal_data.kib_reaper_lock, flags);
2305 while (!kibnal_data.kib_shutdown) {
2306 if (!list_empty (&kibnal_data.kib_reaper_conns)) {
2307 conn = list_entry (kibnal_data.kib_reaper_conns.next,
2308 kib_conn_t, ibc_list);
2309 list_del (&conn->ibc_list);
2311 spin_unlock_irqrestore (&kibnal_data.kib_reaper_lock, flags);
2313 switch (conn->ibc_state) {
2314 case IBNAL_CONN_DEATHROW:
2315 LASSERT (conn->ibc_comm_id != TS_IB_CM_COMM_ID_INVALID);
2316 /* Disconnect: conn becomes a zombie in the
2317 * callback and last ref reschedules it
2319 kibnal_terminate_conn(conn);
2320 kibnal_put_conn (conn);
2323 case IBNAL_CONN_ZOMBIE:
2324 kibnal_destroy_conn (conn);
2328 CERROR ("Bad conn %p state: %d\n",
2329 conn, conn->ibc_state);
2333 spin_lock_irqsave (&kibnal_data.kib_reaper_lock, flags);
2337 spin_unlock_irqrestore (&kibnal_data.kib_reaper_lock, flags);
2339 /* careful with the jiffy wrap... */
2340 while ((timeout = (int)(deadline - jiffies)) <= 0) {
2343 int chunk = kibnal_data.kib_peer_hash_size;
2345 /* Time to check for RDMA timeouts on a few more
2346 * peers: I do checks every 'p' seconds on a
2347 * proportion of the peer table and I need to check
2348 * every connection 'n' times within a timeout
2349 * interval, to ensure I detect a timeout on any
2350 * connection within (n+1)/n times the timeout
2353 if (kibnal_tunables.kib_io_timeout > n * p)
2354 chunk = (chunk * n * p) /
2355 kibnal_tunables.kib_io_timeout;
2359 for (i = 0; i < chunk; i++) {
2360 kibnal_check_conns (peer_index);
2361 peer_index = (peer_index + 1) %
2362 kibnal_data.kib_peer_hash_size;
2368 kibnal_data.kib_reaper_waketime = jiffies + timeout;
2370 set_current_state (TASK_INTERRUPTIBLE);
2371 add_wait_queue (&kibnal_data.kib_reaper_waitq, &wait);
2373 schedule_timeout (timeout);
2375 set_current_state (TASK_RUNNING);
2376 remove_wait_queue (&kibnal_data.kib_reaper_waitq, &wait);
2378 spin_lock_irqsave (&kibnal_data.kib_reaper_lock, flags);
2381 spin_unlock_irqrestore (&kibnal_data.kib_reaper_lock, flags);
2383 kibnal_thread_fini ();
2388 kibnal_connd (void *arg)
2390 long id = (long)arg;
2393 unsigned long flags;
2395 kib_acceptsock_t *as;
2398 snprintf(name, sizeof(name), "kibnal_connd_%02ld", id);
2399 kportal_daemonize(name);
2400 kportal_blockallsigs();
2402 init_waitqueue_entry (&wait, current);
2404 spin_lock_irqsave(&kibnal_data.kib_connd_lock, flags);
2406 while (!kibnal_data.kib_shutdown) {
2409 if (!list_empty (&kibnal_data.kib_connd_acceptq)) {
2410 as = list_entry (kibnal_data.kib_connd_acceptq.next,
2411 kib_acceptsock_t, ibas_list);
2412 list_del (&as->ibas_list);
2414 spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags);
2416 kibnal_handle_svcqry(as->ibas_sock);
2417 sock_release(as->ibas_sock);
2418 PORTAL_FREE(as, sizeof(*as));
2420 spin_lock_irqsave(&kibnal_data.kib_connd_lock, flags);
2424 if (!list_empty (&kibnal_data.kib_connd_peers)) {
2425 peer = list_entry (kibnal_data.kib_connd_peers.next,
2426 kib_peer_t, ibp_connd_list);
2428 list_del_init (&peer->ibp_connd_list);
2429 spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags);
2431 kibnal_connect_peer (peer);
2432 kibnal_put_peer (peer);
2434 spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags);
2441 set_current_state (TASK_INTERRUPTIBLE);
2442 add_wait_queue (&kibnal_data.kib_connd_waitq, &wait);
2444 spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags);
2448 set_current_state (TASK_RUNNING);
2449 remove_wait_queue (&kibnal_data.kib_connd_waitq, &wait);
2451 spin_lock_irqsave (&kibnal_data.kib_connd_lock, flags);
2454 spin_unlock_irqrestore (&kibnal_data.kib_connd_lock, flags);
2456 kibnal_thread_fini ();
2461 kibnal_scheduler(void *arg)
2463 long id = (long)arg;
2467 unsigned long flags;
2472 snprintf(name, sizeof(name), "kibnal_sd_%02ld", id);
2473 kportal_daemonize(name);
2474 kportal_blockallsigs();
2476 spin_lock_irqsave(&kibnal_data.kib_sched_lock, flags);
2478 while (!kibnal_data.kib_shutdown) {
2481 while (!list_empty(&kibnal_data.kib_sched_txq)) {
2482 tx = list_entry(kibnal_data.kib_sched_txq.next,
2484 list_del(&tx->tx_list);
2485 spin_unlock_irqrestore(&kibnal_data.kib_sched_lock,
2489 spin_lock_irqsave(&kibnal_data.kib_sched_lock,
2493 if (!list_empty(&kibnal_data.kib_sched_rxq)) {
2494 rx = list_entry(kibnal_data.kib_sched_rxq.next,
2496 list_del(&rx->rx_list);
2497 spin_unlock_irqrestore(&kibnal_data.kib_sched_lock,
2503 spin_lock_irqsave(&kibnal_data.kib_sched_lock,
2507 /* nothing to do or hogging CPU */
2508 if (!did_something || counter++ == IBNAL_RESCHED) {
2509 spin_unlock_irqrestore(&kibnal_data.kib_sched_lock,
2513 if (!did_something) {
2514 rc = wait_event_interruptible(
2515 kibnal_data.kib_sched_waitq,
2516 !list_empty(&kibnal_data.kib_sched_txq) ||
2517 !list_empty(&kibnal_data.kib_sched_rxq) ||
2518 kibnal_data.kib_shutdown);
2523 spin_lock_irqsave(&kibnal_data.kib_sched_lock,
2528 spin_unlock_irqrestore(&kibnal_data.kib_sched_lock, flags);
2530 kibnal_thread_fini();
2535 lib_nal_t kibnal_lib = {
2536 libnal_data: &kibnal_data, /* NAL private data */
2537 libnal_send: kibnal_send,
2538 libnal_send_pages: kibnal_send_pages,
2539 libnal_recv: kibnal_recv,
2540 libnal_recv_pages: kibnal_recv_pages,
2541 libnal_dist: kibnal_dist