1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 #define DEBUG_SUBSYSTEM S_RPC
25 #include <linux/obd_support.h>
26 #include <linux/obd_class.h>
27 #include <linux/lustre_lib.h>
28 #include <linux/lustre_ha.h>
29 #include <linux/lustre_import.h>
31 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
32 struct ptlrpc_client *cl)
34 cl->cli_request_portal = req_portal;
35 cl->cli_reply_portal = rep_portal;
39 __u8 *ptlrpc_req_to_uuid(struct ptlrpc_request *req)
41 return req->rq_connection->c_remote_uuid;
44 struct ptlrpc_connection *ptlrpc_uuid_to_connection(obd_uuid_t uuid)
46 struct ptlrpc_connection *c;
47 struct lustre_peer peer;
50 err = kportal_uuid_to_peer(uuid, &peer);
52 CERROR("cannot find peer %s!\n", uuid);
56 c = ptlrpc_get_connection(&peer, uuid);
58 memcpy(c->c_remote_uuid, uuid, sizeof(c->c_remote_uuid));
62 CDEBUG(D_INFO, "%s -> %p\n", uuid, c);
67 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,obd_uuid_t uuid)
69 struct lustre_peer peer;
72 err = kportal_uuid_to_peer(uuid, &peer);
74 CERROR("cannot find peer %s!\n", uuid);
78 memcpy(&conn->c_peer, &peer, sizeof(peer));
82 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk(struct ptlrpc_connection *conn)
84 struct ptlrpc_bulk_desc *desc;
86 OBD_ALLOC(desc, sizeof(*desc));
88 desc->bd_connection = ptlrpc_connection_addref(conn);
89 atomic_set(&desc->bd_refcount, 1);
90 init_waitqueue_head(&desc->bd_waitq);
91 INIT_LIST_HEAD(&desc->bd_page_list);
92 ptl_set_inv_handle(&desc->bd_md_h);
93 ptl_set_inv_handle(&desc->bd_me_h);
99 struct ptlrpc_bulk_page *ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc)
101 struct ptlrpc_bulk_page *bulk;
103 OBD_ALLOC(bulk, sizeof(*bulk));
105 bulk->bp_desc = desc;
106 list_add_tail(&bulk->bp_link, &desc->bd_page_list);
107 desc->bd_page_count++;
112 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
114 struct list_head *tmp, *next;
121 list_for_each_safe(tmp, next, &desc->bd_page_list) {
122 struct ptlrpc_bulk_page *bulk;
123 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
124 ptlrpc_free_bulk_page(bulk);
127 ptlrpc_put_connection(desc->bd_connection);
129 OBD_FREE(desc, sizeof(*desc));
133 void ptlrpc_free_bulk_page(struct ptlrpc_bulk_page *bulk)
141 list_del(&bulk->bp_link);
142 bulk->bp_desc->bd_page_count--;
143 OBD_FREE(bulk, sizeof(*bulk));
147 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
148 int count, int *lengths, char **bufs)
150 struct ptlrpc_connection *conn = imp->imp_connection;
151 struct ptlrpc_request *request;
155 OBD_ALLOC(request, sizeof(*request));
157 CERROR("request allocation out of memory\n");
161 rc = lustre_pack_msg(count, lengths, bufs,
162 &request->rq_reqlen, &request->rq_reqmsg);
164 CERROR("cannot pack request %d\n", rc);
165 OBD_FREE(request, sizeof(*request));
169 request->rq_level = LUSTRE_CONN_FULL;
170 request->rq_type = PTL_RPC_MSG_REQUEST;
171 request->rq_import = imp;
172 request->rq_connection = ptlrpc_connection_addref(conn);
174 INIT_LIST_HEAD(&request->rq_list);
176 * This will be reduced once when the sender is finished (waiting for
177 * reply, f.e.), and once when the request has been committed and is
178 * removed from the to-be-committed list.
180 * Also, the refcount will be increased in ptl_send_rpc immediately
181 * before we hand it off to portals, and there will be a corresponding
182 * decrease in request_out_cb (which is called to indicate that portals
183 * is finished with the request, and it can be safely freed).
185 * (Except in the DLM server case, where it will be dropped twice
186 * by the sender, and then the last time by request_out_callback.)
188 atomic_set(&request->rq_refcount, 2);
190 spin_lock(&conn->c_lock);
191 request->rq_xid = HTON__u32(++conn->c_xid_out);
192 spin_unlock(&conn->c_lock);
194 request->rq_reqmsg->magic = PTLRPC_MSG_MAGIC;
195 request->rq_reqmsg->version = PTLRPC_MSG_VERSION;
196 request->rq_reqmsg->opc = HTON__u32(opcode);
198 ptlrpc_hdl2req(request, &imp->imp_handle);
202 void ptlrpc_req_finished(struct ptlrpc_request *request)
207 if (atomic_dec_and_test(&request->rq_refcount))
208 ptlrpc_free_req(request);
211 void ptlrpc_free_req(struct ptlrpc_request *request)
214 if (request == NULL) {
219 if (atomic_read(&request->rq_refcount) != 0) {
220 CERROR("freeing request %p (%d->%s:%d) with refcount %d\n",
221 request, request->rq_reqmsg->opc,
222 request->rq_connection->c_remote_uuid,
223 request->rq_import->imp_client->cli_request_portal,
224 request->rq_refcount);
228 if (request->rq_repmsg != NULL) {
229 OBD_FREE(request->rq_repmsg, request->rq_replen);
230 request->rq_repmsg = NULL;
231 request->rq_reply_md.start = NULL;
233 if (request->rq_reqmsg != NULL) {
234 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
235 request->rq_reqmsg = NULL;
238 if (request->rq_connection) {
239 spin_lock(&request->rq_connection->c_lock);
240 list_del_init(&request->rq_list);
241 spin_unlock(&request->rq_connection->c_lock);
244 ptlrpc_put_connection(request->rq_connection);
245 OBD_FREE(request, sizeof(*request));
249 static int ptlrpc_check_reply(struct ptlrpc_request *req)
253 if (req->rq_repmsg != NULL) {
254 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
255 req->rq_flags |= PTL_RPC_FL_REPLIED;
259 if (req->rq_flags & PTL_RPC_FL_RESEND) {
260 CERROR("-- RESTART --\n");
265 CDEBUG(D_NET, "req = %p, rc = %d\n", req, rc);
269 int ptlrpc_check_status(struct ptlrpc_request *req, int err)
274 CERROR("err is %d\n", err);
279 CERROR("req == NULL\n");
283 if (req->rq_repmsg == NULL) {
284 CERROR("req->rq_repmsg == NULL\n");
288 err = req->rq_repmsg->status;
289 if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
290 CERROR("req->rq_repmsg->type == PTL_RPC_MSG_ERR\n");
291 RETURN(err ? err : -EINVAL);
296 CERROR("req->rq_repmsg->status is %d\n", err);
298 CDEBUG(D_INFO, "req->rq_repmsg->status is %d\n", err);
299 /* XXX: translate this error from net to host */
306 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
308 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
309 request->rq_reqmsg = NULL;
310 request->rq_reqlen = 0;
313 /* Abort this request and cleanup any resources associated with it. */
314 static int ptlrpc_abort(struct ptlrpc_request *request)
316 /* First remove the ME for the reply; in theory, this means
317 * that we can tear down the buffer safely. */
318 PtlMEUnlink(request->rq_reply_me_h);
319 OBD_FREE(request->rq_reply_md.start, request->rq_replen);
320 request->rq_repmsg = NULL;
321 request->rq_replen = 0;
325 /* caller must hold conn->c_lock */
326 void ptlrpc_free_committed(struct ptlrpc_connection *conn)
328 struct list_head *tmp, *saved;
329 struct ptlrpc_request *req;
332 list_for_each_safe(tmp, saved, &conn->c_sending_head) {
333 req = list_entry(tmp, struct ptlrpc_request, rq_list);
335 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
336 CDEBUG(D_INFO, "Keeping req %p xid "LPD64" for replay\n",
341 if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
342 CDEBUG(D_INFO, "Keeping in-flight req %p xid "LPD64
343 " for replay\n", req, req->rq_xid);
347 /* not yet committed */
348 if (req->rq_transno > conn->c_last_committed)
351 CDEBUG(D_INFO, "Marking request %p xid %Ld as committed "
352 "transno=%Lu, last_committed=%Lu\n", req,
353 (long long)req->rq_xid, (long long)req->rq_transno,
354 (long long)conn->c_last_committed);
355 if (atomic_dec_and_test(&req->rq_refcount)) {
356 /* We do this to prevent free_req deadlock.
357 * Restarting after each removal is not so bad, as we are
358 * almost always deleting the first item in the list.
360 * If we use a recursive lock here, we can skip the
361 * unlock/lock/restart sequence.
363 spin_unlock(&conn->c_lock);
364 ptlrpc_free_req(req);
365 spin_lock(&conn->c_lock);
368 list_del(&req->rq_list);
369 list_add(&req->rq_list, &conn->c_dying_head);
377 void ptlrpc_cleanup_client(struct obd_import *imp)
379 struct list_head *tmp, *saved;
380 struct ptlrpc_request *req;
381 struct ptlrpc_connection *conn = imp->imp_connection;
387 spin_lock(&conn->c_lock);
388 list_for_each_safe(tmp, saved, &conn->c_sending_head) {
389 req = list_entry(tmp, struct ptlrpc_request, rq_list);
390 if (req->rq_import != imp)
392 /* XXX we should make sure that nobody's sleeping on these! */
393 CDEBUG(D_INFO, "Cleaning req %p from sending list.\n", req);
394 list_del_init(&req->rq_list);
395 req->rq_import = NULL;
396 spin_unlock(&conn->c_lock);
397 ptlrpc_req_finished(req);
401 list_for_each_safe(tmp, saved, &conn->c_dying_head) {
402 req = list_entry(tmp, struct ptlrpc_request, rq_list);
403 if (req->rq_import != imp)
405 CERROR("Request %p is on the dying list at cleanup!\n", req);
406 list_del_init(&req->rq_list);
407 req->rq_import = NULL;
408 spin_unlock(&conn->c_lock);
409 ptlrpc_req_finished(req);
410 spin_lock(&conn->c_lock);
413 spin_unlock(&conn->c_lock);
419 void ptlrpc_continue_req(struct ptlrpc_request *req)
422 CDEBUG(D_HA, "continue delayed request "LPD64" opc %d\n",
423 req->rq_xid, req->rq_reqmsg->opc);
424 req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
425 req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
426 wake_up(&req->rq_wait_for_rep);
430 void ptlrpc_resend_req(struct ptlrpc_request *req)
433 CDEBUG(D_HA, "resend request "LPD64", opc %d\n",
434 req->rq_xid, req->rq_reqmsg->opc);
435 req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
436 req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
437 req->rq_status = -EAGAIN;
438 req->rq_level = LUSTRE_CONN_RECOVD;
439 req->rq_flags |= PTL_RPC_FL_RESEND;
440 req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
441 wake_up(&req->rq_wait_for_rep);
445 void ptlrpc_restart_req(struct ptlrpc_request *req)
448 CDEBUG(D_HA, "restart completed request "LPD64", opc %d\n",
449 req->rq_xid, req->rq_reqmsg->opc);
450 req->rq_status = -ERESTARTSYS;
451 req->rq_flags |= PTL_RPC_FL_RECOVERY;
452 req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
453 wake_up(&req->rq_wait_for_rep);
457 static int expired_request(void *data)
459 struct ptlrpc_request *req = data;
462 CERROR("req xid "LPD64" op %d: timeout on conn to %s:%d\n",
463 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
464 req->rq_connection->c_remote_uuid,
465 req->rq_import->imp_client->cli_request_portal);
466 req->rq_flags |= PTL_RPC_FL_TIMEOUT;
467 if (!req->rq_import->imp_connection->c_recovd_data.rd_recovd)
471 req->rq_connection->c_level = LUSTRE_CONN_RECOVD;
472 recovd_conn_fail(req->rq_import->imp_connection);
474 /* If this request is for recovery or other primordial tasks,
475 * don't go back to sleep.
477 if (req->rq_level < LUSTRE_CONN_FULL)
482 static int interrupted_request(void *data)
484 struct ptlrpc_request *req = data;
486 req->rq_flags |= PTL_RPC_FL_INTR;
487 RETURN(1); /* ignored, as of this writing */
490 int ptlrpc_queue_wait(struct ptlrpc_request *req)
493 struct l_wait_info lwi;
494 struct ptlrpc_client *cli = req->rq_import->imp_client;
495 struct ptlrpc_connection *conn = req->rq_import->imp_connection;
498 init_waitqueue_head(&req->rq_wait_for_rep);
499 CDEBUG(D_NET, "subsys: %s req "LPD64" opc %d level %d, conn level %d\n",
500 cli->cli_name, req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
501 req->rq_connection->c_level);
503 /* XXX probably both an import and connection level are needed */
504 if (req->rq_level > conn->c_level) {
505 spin_lock(&conn->c_lock);
506 if (conn->c_flags & CONN_INVALID) {
507 /* being torn down by "umount -f" */
508 CERROR("req xid "LPD64" op %d to %s:%d: CONN_INVALID\n",
509 (unsigned long long)req->rq_xid,
511 req->rq_connection->c_remote_uuid,
512 req->rq_import->imp_client->cli_request_portal);
513 spin_unlock(&conn->c_lock);
516 list_del(&req->rq_list);
517 list_add_tail(&req->rq_list, &conn->c_delayed_head);
518 spin_unlock(&conn->c_lock);
520 CDEBUG(D_HA, "req xid "LPD64" op %d to %s:%d: waiting for "
521 "recovery (%d < %d)\n",
522 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
523 req->rq_connection->c_remote_uuid,
524 req->rq_import->imp_client->cli_request_portal,
525 req->rq_level, conn->c_level);
527 lwi = LWI_INTR(NULL, NULL);
528 rc = l_wait_event(req->rq_wait_for_rep,
529 (req->rq_level <= conn->c_level) ||
530 (req->rq_flags & PTL_RPC_FL_ERR), &lwi);
532 spin_lock(&conn->c_lock);
533 list_del_init(&req->rq_list);
534 spin_unlock(&conn->c_lock);
536 if (req->rq_flags & PTL_RPC_FL_ERR)
542 CERROR("process %d resumed\n", current->pid);
545 req->rq_timeout = obd_timeout;
546 spin_lock(&conn->c_lock);
547 if (conn->c_flags & CONN_INVALID) {
548 CERROR("req xid "LPD64" op %d to %s:%d: CONN_INVALID\n",
549 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
550 req->rq_connection->c_remote_uuid,
551 req->rq_import->imp_client->cli_request_portal);
552 spin_unlock(&conn->c_lock); /* being torn down by "umount -f" */
556 list_del(&req->rq_list);
557 list_add_tail(&req->rq_list, &conn->c_sending_head);
558 spin_unlock(&conn->c_lock);
559 rc = ptl_send_rpc(req);
561 CDEBUG(D_HA, "error %d, opcode %d, need recovery\n", rc,
562 req->rq_reqmsg->opc);
563 /* the sleep below will time out, triggering recovery */
566 CDEBUG(D_NET, "-- sleeping on req xid "LPD64" op %d to %s:%d\n",
567 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
568 req->rq_connection->c_remote_uuid,
569 req->rq_import->imp_client->cli_request_portal);
570 lwi = LWI_TIMEOUT_INTR(req->rq_timeout * HZ, expired_request,
571 interrupted_request,req);
572 l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
573 CDEBUG(D_NET, "-- done sleeping on req xid "LPD64" op %d to %s:%d\n",
574 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
575 req->rq_connection->c_remote_uuid,
576 req->rq_import->imp_client->cli_request_portal);
578 /* Don't resend if we were interrupted. */
579 if ((req->rq_flags & (PTL_RPC_FL_RESEND | PTL_RPC_FL_INTR)) ==
581 req->rq_flags &= ~PTL_RPC_FL_RESEND;
582 CDEBUG(D_HA, "req xid "LPD64" op %d to %s:%d\n",
583 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
584 req->rq_connection->c_remote_uuid,
585 req->rq_import->imp_client->cli_request_portal);
586 /* we'll get sent again, so balance 2nd request_out_callback */
590 // up(&cli->cli_rpc_sem);
591 if (req->rq_flags & PTL_RPC_FL_INTR) {
592 if (!(req->rq_flags & PTL_RPC_FL_TIMEOUT))
593 LBUG(); /* should only be interrupted if we timed out */
594 /* Clean up the dangling reply buffers */
596 GOTO(out, rc = -EINTR);
599 if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
600 GOTO(out, rc = -ETIMEDOUT);
602 if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
603 GOTO(out, rc = req->rq_status);
605 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
607 CERROR("unpack_rep failed: %d\n", rc);
611 /* FIXME: Enable when BlueArc makes new release */
612 if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
613 req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
614 CERROR("invalid packet type received (type=%u)\n",
615 req->rq_repmsg->type);
617 GOTO(out, rc = -EINVAL);
620 CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
621 if (req->rq_repmsg->status == 0)
622 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
623 req->rq_replen, req->rq_repmsg->status);
625 spin_lock(&conn->c_lock);
626 conn->c_last_xid = req->rq_repmsg->last_xid;
627 conn->c_last_committed = req->rq_repmsg->last_committed;
628 ptlrpc_free_committed(conn);
629 spin_unlock(&conn->c_lock);
636 int ptlrpc_replay_req(struct ptlrpc_request *req)
639 // struct ptlrpc_client *cli = req->rq_import->imp_client;
640 struct l_wait_info lwi;
643 init_waitqueue_head(&req->rq_wait_for_rep);
644 CDEBUG(D_NET, "req "LPD64" opc %d level %d, conn level %d\n",
645 req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
646 req->rq_connection->c_level);
648 req->rq_timeout = obd_timeout;
649 req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
650 req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
652 rc = ptl_send_rpc(req);
654 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
655 ptlrpc_cleanup_request_buf(req);
656 // up(&cli->cli_rpc_sem);
660 CDEBUG(D_OTHER, "-- sleeping\n");
661 lwi = LWI_INTR(NULL, NULL); /* XXX needs timeout, nested recovery */
662 l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
663 CDEBUG(D_OTHER, "-- done\n");
665 // up(&cli->cli_rpc_sem);
667 if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
668 CERROR("Unknown reason for wakeup\n");
669 /* XXX Phil - I end up here when I kill obdctl */
671 GOTO(out, rc = -EINTR);
674 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
676 CERROR("unpack_rep failed: %d\n", rc);
680 CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
681 if (req->rq_repmsg->status == 0)
682 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
683 req->rq_replen, req->rq_repmsg->status);
685 CERROR("recovery failed: ");
686 CERROR("req "LPD64" opc %d level %d, conn level %d\n",
687 req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
688 req->rq_connection->c_level);
692 if (req->rq_replay_cb)
693 req->rq_replay_cb(req, &req->rq_replay_cb_handle);