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 struct ptlrpc_connection *conn = req->rq_import->imp_connection;
255 spin_lock(&conn->c_lock);
256 if (req->rq_level > conn->c_level) {
258 "rep to xid "LPD64" op %d to %s:%d: "
259 "recovery started, ignoring (%d > %d)\n",
260 (unsigned long long)req->rq_xid,
261 req->rq_reqmsg->opc, conn->c_remote_uuid,
262 req->rq_import->imp_client->cli_request_portal,
263 req->rq_level, conn->c_level);
264 req->rq_repmsg = NULL;
265 spin_unlock(&conn->c_lock);
268 spin_unlock(&conn->c_lock);
269 req->rq_transno = NTOH__u64(req->rq_repmsg->transno);
270 req->rq_flags |= PTL_RPC_FL_REPLIED;
274 if (req->rq_flags & PTL_RPC_FL_RESEND) {
275 CERROR("-- RESTART --\n");
279 if (req->rq_flags & PTL_RPC_FL_ERR) {
280 CERROR("-- ABORTED --\n");
285 CDEBUG(D_NET, "req = %p, rc = %d\n", req, rc);
289 int ptlrpc_check_status(struct ptlrpc_request *req, int err)
294 CERROR("err is %d\n", err);
299 CERROR("req == NULL\n");
303 if (req->rq_repmsg == NULL) {
304 CERROR("req->rq_repmsg == NULL\n");
308 err = req->rq_repmsg->status;
309 if (req->rq_repmsg->type == NTOH__u32(PTL_RPC_MSG_ERR)) {
310 CERROR("req->rq_repmsg->type == PTL_RPC_MSG_ERR\n");
311 RETURN(err ? err : -EINVAL);
316 CERROR("req->rq_repmsg->status is %d\n", err);
318 CDEBUG(D_INFO, "req->rq_repmsg->status is %d\n", err);
319 /* XXX: translate this error from net to host */
326 static void ptlrpc_cleanup_request_buf(struct ptlrpc_request *request)
328 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
329 request->rq_reqmsg = NULL;
330 request->rq_reqlen = 0;
333 /* Abort this request and cleanup any resources associated with it. */
334 static int ptlrpc_abort(struct ptlrpc_request *request)
336 /* First remove the ME for the reply; in theory, this means
337 * that we can tear down the buffer safely. */
338 PtlMEUnlink(request->rq_reply_me_h);
339 OBD_FREE(request->rq_reply_md.start, request->rq_replen);
340 request->rq_repmsg = NULL;
341 request->rq_replen = 0;
345 /* caller must hold conn->c_lock */
346 void ptlrpc_free_committed(struct ptlrpc_connection *conn)
348 struct list_head *tmp, *saved;
349 struct ptlrpc_request *req;
352 list_for_each_safe(tmp, saved, &conn->c_sending_head) {
353 req = list_entry(tmp, struct ptlrpc_request, rq_list);
355 if (req->rq_flags & PTL_RPC_FL_REPLAY) {
356 CDEBUG(D_INFO, "Keeping req %p xid "LPD64" for replay\n",
361 if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
362 CDEBUG(D_INFO, "Keeping in-flight req %p xid "LPD64
363 " for replay\n", req, req->rq_xid);
367 /* not yet committed */
368 if (req->rq_transno > conn->c_last_committed)
371 CDEBUG(D_INFO, "Marking request %p xid %Ld as committed "
372 "transno=%Lu, last_committed=%Lu\n", req,
373 (long long)req->rq_xid, (long long)req->rq_transno,
374 (long long)conn->c_last_committed);
375 if (atomic_dec_and_test(&req->rq_refcount)) {
376 /* We do this to prevent free_req deadlock.
377 * Restarting after each removal is not so bad, as we are
378 * almost always deleting the first item in the list.
380 * If we use a recursive lock here, we can skip the
381 * unlock/lock/restart sequence.
383 spin_unlock(&conn->c_lock);
384 ptlrpc_free_req(req);
385 spin_lock(&conn->c_lock);
388 list_del(&req->rq_list);
389 list_add(&req->rq_list, &conn->c_dying_head);
397 void ptlrpc_cleanup_client(struct obd_import *imp)
399 struct list_head *tmp, *saved;
400 struct ptlrpc_request *req;
401 struct ptlrpc_connection *conn = imp->imp_connection;
407 spin_lock(&conn->c_lock);
408 list_for_each_safe(tmp, saved, &conn->c_sending_head) {
409 req = list_entry(tmp, struct ptlrpc_request, rq_list);
410 if (req->rq_import != imp)
412 /* XXX we should make sure that nobody's sleeping on these! */
413 CDEBUG(D_INFO, "Cleaning req %p from sending list.\n", req);
414 list_del_init(&req->rq_list);
415 req->rq_import = NULL;
416 spin_unlock(&conn->c_lock);
417 ptlrpc_req_finished(req);
421 list_for_each_safe(tmp, saved, &conn->c_dying_head) {
422 req = list_entry(tmp, struct ptlrpc_request, rq_list);
423 if (req->rq_import != imp)
425 CERROR("Request %p is on the dying list at cleanup!\n", req);
426 list_del_init(&req->rq_list);
427 req->rq_import = NULL;
428 spin_unlock(&conn->c_lock);
429 ptlrpc_req_finished(req);
430 spin_lock(&conn->c_lock);
433 spin_unlock(&conn->c_lock);
439 void ptlrpc_continue_req(struct ptlrpc_request *req)
442 CDEBUG(D_HA, "continue delayed request "LPD64" opc %d\n",
443 req->rq_xid, req->rq_reqmsg->opc);
444 req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
445 req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
446 wake_up(&req->rq_wait_for_rep);
450 void ptlrpc_resend_req(struct ptlrpc_request *req)
453 CDEBUG(D_HA, "resend request "LPD64", opc %d\n",
454 req->rq_xid, req->rq_reqmsg->opc);
455 req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
456 req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
457 req->rq_status = -EAGAIN;
458 req->rq_level = LUSTRE_CONN_RECOVD;
459 req->rq_flags |= PTL_RPC_FL_RESEND;
460 req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
461 wake_up(&req->rq_wait_for_rep);
465 void ptlrpc_restart_req(struct ptlrpc_request *req)
468 CDEBUG(D_HA, "restart completed request "LPD64", opc %d\n",
469 req->rq_xid, req->rq_reqmsg->opc);
470 req->rq_status = -ERESTARTSYS;
471 req->rq_flags |= PTL_RPC_FL_RECOVERY;
472 req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
473 wake_up(&req->rq_wait_for_rep);
477 static int expired_request(void *data)
479 struct ptlrpc_request *req = data;
482 CERROR("req xid "LPD64" op %d: timeout on conn to %s:%d\n",
483 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
484 req->rq_connection->c_remote_uuid,
485 req->rq_import->imp_client->cli_request_portal);
486 req->rq_flags |= PTL_RPC_FL_TIMEOUT;
487 if (!req->rq_import->imp_connection->c_recovd_data.rd_recovd)
491 req->rq_connection->c_level = LUSTRE_CONN_RECOVD;
492 recovd_conn_fail(req->rq_import->imp_connection);
494 /* If this request is for recovery or other primordial tasks,
495 * don't go back to sleep.
497 if (req->rq_level < LUSTRE_CONN_FULL)
502 static int interrupted_request(void *data)
504 struct ptlrpc_request *req = data;
506 req->rq_flags |= PTL_RPC_FL_INTR;
507 RETURN(1); /* ignored, as of this writing */
510 /* If we're being torn down by umount -f, or the import has been
511 * invalidated (such as by an OST failure), the request must fail with
514 * Must be called with conn->c_lock held, will drop it if it returns -EIO.
516 * XXX this should just be testing the import, and umount_begin shouldn't touch
517 * XXX the connection.
519 #define EIO_IF_INVALID(conn, req) \
520 if ((conn->c_flags & CONN_INVALID) || \
521 (req->rq_import->imp_flags & IMP_INVALID)) { \
522 CERROR("req xid "LPD64" op %d to %s:%d: %s_INVALID\n", \
523 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc, \
524 req->rq_connection->c_remote_uuid, \
525 req->rq_import->imp_client->cli_request_portal, \
526 (conn->c_flags & CONN_INVALID) ? "CONN_" : "IMP_"); \
527 spin_unlock(&conn->c_lock); \
531 int ptlrpc_queue_wait(struct ptlrpc_request *req)
534 struct l_wait_info lwi;
535 struct ptlrpc_client *cli = req->rq_import->imp_client;
536 struct ptlrpc_connection *conn = req->rq_import->imp_connection;
539 init_waitqueue_head(&req->rq_wait_for_rep);
540 CDEBUG(D_NET, "subsys: %s req "LPD64" opc %d level %d, conn level %d\n",
541 cli->cli_name, req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
542 req->rq_connection->c_level);
544 /* XXX probably both an import and connection level are needed */
545 if (req->rq_level > conn->c_level) {
546 spin_lock(&conn->c_lock);
547 EIO_IF_INVALID(conn, req);
548 list_del(&req->rq_list);
549 list_add_tail(&req->rq_list, &conn->c_delayed_head);
550 spin_unlock(&conn->c_lock);
552 CDEBUG(D_HA, "req xid "LPD64" op %d to %s:%d: waiting for "
553 "recovery (%d < %d)\n",
554 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
555 req->rq_connection->c_remote_uuid,
556 req->rq_import->imp_client->cli_request_portal,
557 req->rq_level, conn->c_level);
559 lwi = LWI_INTR(NULL, NULL);
560 rc = l_wait_event(req->rq_wait_for_rep,
561 (req->rq_level <= conn->c_level) ||
562 (req->rq_flags & PTL_RPC_FL_ERR), &lwi);
564 spin_lock(&conn->c_lock);
565 list_del_init(&req->rq_list);
566 spin_unlock(&conn->c_lock);
568 if (req->rq_flags & PTL_RPC_FL_ERR)
574 CERROR("process %d resumed\n", current->pid);
577 req->rq_timeout = obd_timeout;
578 spin_lock(&conn->c_lock);
579 EIO_IF_INVALID(conn, req);
581 list_del(&req->rq_list);
582 list_add_tail(&req->rq_list, &conn->c_sending_head);
583 spin_unlock(&conn->c_lock);
584 rc = ptl_send_rpc(req);
586 CDEBUG(D_HA, "error %d, opcode %d, need recovery\n", rc,
587 req->rq_reqmsg->opc);
588 /* the sleep below will time out, triggering recovery */
591 CDEBUG(D_NET, "-- sleeping on req xid "LPD64" op %d to %s:%d\n",
592 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
593 req->rq_connection->c_remote_uuid,
594 req->rq_import->imp_client->cli_request_portal);
595 lwi = LWI_TIMEOUT_INTR(req->rq_timeout * HZ, expired_request,
596 interrupted_request,req);
597 l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
598 CDEBUG(D_NET, "-- done sleeping on req xid "LPD64" op %d to %s:%d\n",
599 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
600 req->rq_connection->c_remote_uuid,
601 req->rq_import->imp_client->cli_request_portal);
603 if (req->rq_flags & PTL_RPC_FL_ERR) {
605 GOTO(out, rc = -EIO);
608 /* Don't resend if we were interrupted. */
609 if ((req->rq_flags & (PTL_RPC_FL_RESEND | PTL_RPC_FL_INTR)) ==
611 req->rq_flags &= ~PTL_RPC_FL_RESEND;
612 CDEBUG(D_HA, "resending req xid "LPD64" op %d to %s:%d\n",
613 (unsigned long long)req->rq_xid, req->rq_reqmsg->opc,
614 req->rq_connection->c_remote_uuid,
615 req->rq_import->imp_client->cli_request_portal);
619 // up(&cli->cli_rpc_sem);
620 if (req->rq_flags & PTL_RPC_FL_INTR) {
621 if (!(req->rq_flags & PTL_RPC_FL_TIMEOUT))
622 LBUG(); /* should only be interrupted if we timed out */
623 /* Clean up the dangling reply buffers */
625 GOTO(out, rc = -EINTR);
628 if (req->rq_flags & PTL_RPC_FL_TIMEOUT)
629 GOTO(out, rc = -ETIMEDOUT);
631 if (!(req->rq_flags & PTL_RPC_FL_REPLIED))
632 GOTO(out, rc = req->rq_status);
634 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
636 CERROR("unpack_rep failed: %d\n", rc);
640 /* FIXME: Enable when BlueArc makes new release */
641 if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
642 req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
643 CERROR("invalid packet type received (type=%u)\n",
644 req->rq_repmsg->type);
646 GOTO(out, rc = -EINVAL);
649 CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
650 if (req->rq_repmsg->status == 0)
651 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
652 req->rq_replen, req->rq_repmsg->status);
654 spin_lock(&conn->c_lock);
655 conn->c_last_xid = req->rq_repmsg->last_xid;
656 conn->c_last_committed = req->rq_repmsg->last_committed;
657 ptlrpc_free_committed(conn);
658 spin_unlock(&conn->c_lock);
665 #undef EIO_IF_INVALID
667 int ptlrpc_replay_req(struct ptlrpc_request *req)
669 int rc = 0, old_level;
670 // struct ptlrpc_client *cli = req->rq_import->imp_client;
671 struct l_wait_info lwi;
674 init_waitqueue_head(&req->rq_wait_for_rep);
675 CDEBUG(D_NET, "req "LPD64" opc %d level %d, conn level %d\n",
676 req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
677 req->rq_connection->c_level);
679 req->rq_timeout = obd_timeout;
680 req->rq_reqmsg->addr = req->rq_import->imp_handle.addr;
681 req->rq_reqmsg->cookie = req->rq_import->imp_handle.cookie;
683 /* temporarily set request to RECOVD level (reset at out:) */
684 old_level = req->rq_level;
685 req->rq_level = LUSTRE_CONN_RECOVD;
686 rc = ptl_send_rpc(req);
688 CERROR("error %d, opcode %d\n", rc, req->rq_reqmsg->opc);
689 ptlrpc_cleanup_request_buf(req);
690 // up(&cli->cli_rpc_sem);
694 CDEBUG(D_OTHER, "-- sleeping\n");
695 lwi = LWI_INTR(NULL, NULL); /* XXX needs timeout, nested recovery */
696 l_wait_event(req->rq_wait_for_rep, ptlrpc_check_reply(req), &lwi);
697 CDEBUG(D_OTHER, "-- done\n");
699 // up(&cli->cli_rpc_sem);
701 if (!(req->rq_flags & PTL_RPC_FL_REPLIED)) {
702 CERROR("Unknown reason for wakeup\n");
703 /* XXX Phil - I end up here when I kill obdctl */
705 GOTO(out, rc = -EINTR);
708 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
710 CERROR("unpack_rep failed: %d\n", rc);
714 CDEBUG(D_NET, "got rep "LPD64"\n", req->rq_xid);
716 /* let the callback do fixups, possibly including in the request */
717 if (req->rq_replay_cb)
718 req->rq_replay_cb(req, req->rq_replay_cb_data);
720 if (req->rq_repmsg->status == 0) {
721 CDEBUG(D_NET, "--> buf %p len %d status %d\n", req->rq_repmsg,
722 req->rq_replen, req->rq_repmsg->status);
724 CERROR("recovery failed: ");
725 CERROR("req "LPD64" opc %d level %d, conn level %d\n",
726 req->rq_xid, req->rq_reqmsg->opc, req->rq_level,
727 req->rq_connection->c_level);
732 req->rq_level = old_level;