4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_RPC
34 #include <obd_support.h>
35 #include <lustre_net.h>
36 #include <lustre_lib.h>
38 #include <obd_class.h>
39 #include "ptlrpc_internal.h"
42 * Helper function. Sends \a len bytes from \a base at offset \a offset
43 * over \a conn connection to portal \a portal.
44 * Returns 0 on success or error code.
46 static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len,
47 enum lnet_ack_req ack, struct ptlrpc_cb_id *cbid,
48 lnet_nid_t self, struct lnet_process_id peer_id,
49 int portal, __u64 xid, unsigned int offset,
50 struct lnet_handle_md *bulk_cookie)
56 LASSERT (portal != 0);
57 CDEBUG (D_INFO, "peer_id %s\n", libcfs_id2str(peer_id));
60 md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
61 md.options = PTLRPC_MD_OPTIONS;
63 md.eq_handle = ptlrpc_eq_h;
64 LNetInvalidateMDHandle(&md.bulk_handle);
67 md.bulk_handle = *bulk_cookie;
68 md.options |= LNET_MD_BULK_HANDLE;
71 if (unlikely(ack == LNET_ACK_REQ &&
72 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
73 /* don't ask for the ack to simulate failing client */
77 rc = LNetMDBind (md, LNET_UNLINK, mdh);
78 if (unlikely(rc != 0)) {
79 CERROR ("LNetMDBind failed: %d\n", rc);
80 LASSERT (rc == -ENOMEM);
84 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
85 len, portal, xid, offset);
87 rc = LNetPut(self, *mdh, ack,
88 peer_id, portal, xid, offset, 0);
89 if (unlikely(rc != 0)) {
91 /* We're going to get an UNLINK event when I unlink below,
92 * which will complete just like any other failed send, so
93 * I fall through and return success here! */
94 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
95 libcfs_id2str(peer_id), portal, xid, rc);
96 rc2 = LNetMDUnlink(*mdh);
97 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
103 static void mdunlink_iterate_helper(struct lnet_handle_md *bd_mds, int count)
107 for (i = 0; i < count; i++)
108 LNetMDUnlink(bd_mds[i]);
111 #ifdef HAVE_SERVER_SUPPORT
113 * Prepare bulk descriptor for specified incoming request \a req that
114 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
115 * the bulk to be sent. Used on server-side after request was already
117 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
120 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
121 unsigned nfrags, unsigned max_brw,
124 const struct ptlrpc_bulk_frag_ops
127 struct obd_export *exp = req->rq_export;
128 struct ptlrpc_bulk_desc *desc;
131 LASSERT(ptlrpc_is_bulk_op_active(type));
133 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
137 desc->bd_export = class_export_get(exp);
140 desc->bd_cbid.cbid_fn = server_bulk_callback;
141 desc->bd_cbid.cbid_arg = desc;
143 /* NB we don't assign rq_bulk here; server-side requests are
144 * re-used, and the handler frees the bulk desc explicitly. */
148 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
151 * Starts bulk transfer for descriptor \a desc on the server.
152 * Returns 0 on success or error code.
154 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
156 struct obd_export *exp = desc->bd_export;
158 struct lnet_process_id peer_id;
166 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
169 /* NB no locking required until desc is on the network */
170 LASSERT(desc->bd_md_count == 0);
171 LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
173 LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
174 LASSERT(desc->bd_cbid.cbid_arg == desc);
177 * Multi-Rail: get the preferred self and peer NIDs from the
178 * request, so they are based on the route taken by the
181 self_nid = desc->bd_req->rq_self;
182 peer_id = desc->bd_req->rq_source;
184 /* NB total length may be 0 for a read past EOF, so we send 0
185 * length bulks, since the client expects bulk events.
187 * The client may not need all of the bulk mbits for the RPC. The RPC
188 * used the mbits of the highest bulk mbits needed, and the server masks
189 * off high bits to get bulk count for this RPC. LU-1431 */
190 mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
191 total_md = desc->bd_req->rq_mbits - mbits + 1;
193 desc->bd_md_count = total_md;
194 desc->bd_failure = 0;
196 md.user_ptr = &desc->bd_cbid;
197 md.eq_handle = ptlrpc_eq_h;
198 md.threshold = 2; /* SENT and ACK/REPLY */
200 for (posted_md = 0; posted_md < total_md; mbits++) {
201 md.options = PTLRPC_MD_OPTIONS;
203 /* NB it's assumed that source and sink buffer frags are
204 * page-aligned. Otherwise we'd have to send client bulk
205 * sizes over and split server buffer accordingly */
206 ptlrpc_fill_bulk_md(&md, desc, posted_md);
207 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
209 CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
210 exp->exp_obd->obd_name, posted_md, rc);
211 LASSERT(rc == -ENOMEM);
212 if (posted_md == 0) {
213 desc->bd_md_count = 0;
219 /* LU-6441: last md is not sent and desc->bd_md_count == 1 */
220 if (OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
222 total_md > 1 && posted_md == total_md - 1) {
227 /* Network is about to get at the memory */
228 if (ptlrpc_is_bulk_put_source(desc->bd_type))
229 rc = LNetPut(self_nid, desc->bd_mds[posted_md],
230 LNET_ACK_REQ, peer_id,
231 desc->bd_portal, mbits, 0, 0);
233 rc = LNetGet(self_nid, desc->bd_mds[posted_md],
234 peer_id, desc->bd_portal, mbits, 0);
238 CERROR("%s: failed bulk transfer with %s:%u x%llu: "
239 "rc = %d\n", exp->exp_obd->obd_name,
240 libcfs_id2str(peer_id), desc->bd_portal,
247 /* Can't send, so we unlink the MD bound above. The UNLINK
248 * event this creates will signal completion with failure,
249 * so we return SUCCESS here! */
250 spin_lock(&desc->bd_lock);
251 desc->bd_md_count -= total_md - posted_md;
252 spin_unlock(&desc->bd_lock);
253 LASSERT(desc->bd_md_count >= 0);
255 mdunlink_iterate_helper(desc->bd_mds, posted_md);
259 CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
260 "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
261 desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id),
262 mbits - posted_md, mbits - 1);
268 * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
269 * serialises with completion callback)
271 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
273 struct l_wait_info lwi;
276 LASSERT(!in_interrupt()); /* might sleep */
278 if (!ptlrpc_server_bulk_active(desc)) /* completed or */
279 return; /* never started */
281 /* We used to poison the pages with 0xab here because we did not want to
282 * send any meaningful data over the wire for evicted clients (bug 9297)
283 * However, this is no longer safe now that we use the page cache on the
286 /* The unlink ensures the callback happens ASAP and is the last
287 * one. If it fails, it must be because completion just happened,
288 * but we must still l_wait_event() in this case, to give liblustre
289 * a chance to run server_bulk_callback()*/
290 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
293 /* Network access will complete in finite time but the HUGE
294 * timeout lets us CWARN for visibility of sluggish NALs */
295 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
296 cfs_time_seconds(1), NULL, NULL);
297 rc = l_wait_event(desc->bd_waitq,
298 !ptlrpc_server_bulk_active(desc), &lwi);
302 LASSERT(rc == -ETIMEDOUT);
303 CWARN("Unexpectedly long timeout: desc %p\n", desc);
306 #endif /* HAVE_SERVER_SUPPORT */
309 * Register bulk at the sender for later transfer.
310 * Returns 0 on success or error code.
312 int ptlrpc_register_bulk(struct ptlrpc_request *req)
314 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
315 struct lnet_process_id peer;
321 struct lnet_handle_me me_h;
325 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
328 /* NB no locking required until desc is on the network */
329 LASSERT(desc->bd_nob > 0);
330 LASSERT(desc->bd_md_count == 0);
331 LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
332 LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
333 LASSERT(desc->bd_req != NULL);
334 LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
336 /* cleanup the state of the bulk for it will be reused */
337 if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
338 desc->bd_nob_transferred = 0;
339 else if (desc->bd_nob_transferred != 0)
340 /* If the network failed after an RPC was sent, this condition
341 * could happen. Rather than assert (was here before), return
345 desc->bd_failure = 0;
347 peer = desc->bd_import->imp_connection->c_peer;
349 LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
350 LASSERT(desc->bd_cbid.cbid_arg == desc);
352 total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
353 /* rq_mbits is matchbits of the final bulk */
354 mbits = req->rq_mbits - total_md + 1;
356 LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
357 "first mbits = x%llu, last mbits = x%llu\n",
358 mbits, req->rq_mbits);
359 LASSERTF(!(desc->bd_registered &&
360 req->rq_send_state != LUSTRE_IMP_REPLAY) ||
361 mbits != desc->bd_last_mbits,
362 "registered: %d rq_mbits: %llu bd_last_mbits: %llu\n",
363 desc->bd_registered, mbits, desc->bd_last_mbits);
365 desc->bd_registered = 1;
366 desc->bd_last_mbits = mbits;
367 desc->bd_md_count = total_md;
368 md.user_ptr = &desc->bd_cbid;
369 md.eq_handle = ptlrpc_eq_h;
370 md.threshold = 1; /* PUT or GET */
372 for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
373 md.options = PTLRPC_MD_OPTIONS |
374 (ptlrpc_is_bulk_op_get(desc->bd_type) ?
375 LNET_MD_OP_GET : LNET_MD_OP_PUT);
376 ptlrpc_fill_bulk_md(&md, desc, posted_md);
378 if (posted_md > 0 && posted_md + 1 == total_md &&
379 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
382 rc = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
383 LNET_UNLINK, LNET_INS_AFTER, &me_h);
386 CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
387 desc->bd_import->imp_obd->obd_name, mbits,
392 /* About to let the network at it... */
393 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
394 &desc->bd_mds[posted_md]);
396 CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
397 desc->bd_import->imp_obd->obd_name, mbits,
399 rc2 = LNetMEUnlink(me_h);
406 LASSERT(rc == -ENOMEM);
407 spin_lock(&desc->bd_lock);
408 desc->bd_md_count -= total_md - posted_md;
409 spin_unlock(&desc->bd_lock);
410 LASSERT(desc->bd_md_count >= 0);
411 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
412 req->rq_status = -ENOMEM;
413 desc->bd_registered = 0;
417 spin_lock(&desc->bd_lock);
418 /* Holler if peer manages to touch buffers before he knows the mbits */
419 if (desc->bd_md_count != total_md)
420 CWARN("%s: Peer %s touched %d buffers while I registered\n",
421 desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
422 total_md - desc->bd_md_count);
423 spin_unlock(&desc->bd_lock);
425 CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
426 "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count,
427 ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
428 desc->bd_iov_count, desc->bd_nob,
429 desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
435 * Disconnect a bulk desc from the network. Idempotent. Not
436 * thread-safe (i.e. only interlocks with completion callback).
437 * Returns 1 on success or 0 if network unregistration failed for whatever
440 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
442 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
443 struct l_wait_info lwi;
447 LASSERT(!in_interrupt()); /* might sleep */
449 /* Let's setup deadline for reply unlink. */
450 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
451 async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
452 req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK;
454 if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
455 RETURN(1); /* never registered */
457 LASSERT(desc->bd_req == req); /* bd_req NULL until registered */
459 /* the unlink ensures the callback happens ASAP and is the last
460 * one. If it fails, it must be because completion just happened,
461 * but we must still l_wait_event() in this case to give liblustre
462 * a chance to run client_bulk_callback() */
463 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
465 if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
466 RETURN(1); /* never registered */
468 /* Move to "Unregistering" phase as bulk was not unlinked yet. */
469 ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
471 /* Do not wait for unlink to finish. */
476 /* The wq argument is ignored by user-space wait_event macros */
477 wait_queue_head_t *wq = (req->rq_set != NULL) ?
478 &req->rq_set->set_waitq :
479 &req->rq_reply_waitq;
480 /* Network access will complete in finite time but the HUGE
481 * timeout lets us CWARN for visibility of sluggish NALs */
482 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
483 cfs_time_seconds(1), NULL, NULL);
484 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
486 ptlrpc_rqphase_move(req, req->rq_next_phase);
490 LASSERT(rc == -ETIMEDOUT);
491 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
497 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
499 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
500 struct ptlrpc_service *svc = svcpt->scp_service;
501 int service_time = max_t(int, ktime_get_real_seconds() -
502 req->rq_arrival_time.tv_sec, 1);
504 if (!(flags & PTLRPC_REPLY_EARLY) &&
505 (req->rq_type != PTL_RPC_MSG_ERR) &&
506 (req->rq_reqmsg != NULL) &&
507 !(lustre_msg_get_flags(req->rq_reqmsg) &
508 (MSG_RESENT | MSG_REPLAY |
509 MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
510 /* early replies, errors and recovery requests don't count
511 * toward our service time estimate */
512 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
515 DEBUG_REQ(D_ADAPTTO, req,
516 "svc %s changed estimate from %d to %d",
517 svc->srv_name, oldse,
518 at_get(&svcpt->scp_at_estimate));
521 /* Report actual service time for client latency calc */
522 lustre_msg_set_service_time(req->rq_repmsg, service_time);
523 /* Report service time estimate for future client reqs, but report 0
524 * (to be ignored by client) if it's an error reply during recovery.
527 if (req->rq_type == PTL_RPC_MSG_ERR &&
528 (req->rq_export == NULL ||
529 req->rq_export->exp_obd->obd_recovering)) {
530 lustre_msg_set_timeout(req->rq_repmsg, 0);
534 if (req->rq_export && req->rq_reqmsg != NULL &&
535 (flags & PTLRPC_REPLY_EARLY) &&
536 lustre_msg_get_flags(req->rq_reqmsg) &
537 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
538 struct obd_device *exp_obd = req->rq_export->exp_obd;
540 timeout = ktime_get_real_seconds() -
541 req->rq_arrival_time.tv_sec +
542 min_t(time64_t, at_extra,
543 exp_obd->obd_recovery_timeout / 4);
545 timeout = at_get(&svcpt->scp_at_estimate);
547 lustre_msg_set_timeout(req->rq_repmsg, timeout);
550 if (req->rq_reqmsg &&
551 !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
552 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
553 "req_flags=%#x magic=%x/%x len=%d\n",
554 flags, lustre_msg_get_flags(req->rq_reqmsg),
555 lustre_msg_get_magic(req->rq_reqmsg),
556 lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
561 * Send request reply from request \a req reply buffer.
562 * \a flags defines reply types
563 * Returns 0 on success or error code
565 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
567 struct ptlrpc_reply_state *rs = req->rq_reply_state;
568 struct ptlrpc_connection *conn;
571 /* We must already have a reply buffer (only ptlrpc_error() may be
572 * called without one). The reply generated by sptlrpc layer (e.g.
573 * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
574 * have a request buffer which is either the actual (swabbed) incoming
575 * request, or a saved copy if this is a req saved in
576 * target_queue_final_reply().
578 LASSERT (req->rq_no_reply == 0);
579 LASSERT (req->rq_reqbuf != NULL);
580 LASSERT (rs != NULL);
581 LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
582 LASSERT (req->rq_repmsg != NULL);
583 LASSERT (req->rq_repmsg == rs->rs_msg);
584 LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
585 LASSERT (rs->rs_cb_id.cbid_arg == rs);
587 /* There may be no rq_export during failover */
589 if (unlikely(req->rq_export && req->rq_export->exp_obd &&
590 req->rq_export->exp_obd->obd_fail)) {
591 /* Failed obd's only send ENODEV */
592 req->rq_type = PTL_RPC_MSG_ERR;
593 req->rq_status = -ENODEV;
594 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
595 req->rq_export->exp_obd->obd_minor);
598 /* In order to keep interoprability with the client (< 2.3) which
599 * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
600 * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
601 * reply buffer on client will be overflow.
603 * XXX Remove this whenver we drop the interoprability with such client.
605 req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
606 sizeof(struct ptlrpc_body_v2), 1);
608 if (req->rq_type != PTL_RPC_MSG_ERR)
609 req->rq_type = PTL_RPC_MSG_REPLY;
611 lustre_msg_set_type(req->rq_repmsg, req->rq_type);
612 lustre_msg_set_status(req->rq_repmsg,
613 ptlrpc_status_hton(req->rq_status));
614 lustre_msg_set_opc(req->rq_repmsg,
615 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
617 target_pack_pool_reply(req);
619 ptlrpc_at_set_reply(req, flags);
621 if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
622 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
624 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
626 if (unlikely(conn == NULL)) {
627 CERROR("not replying on NULL connection\n"); /* bug 9635 */
630 ptlrpc_rs_addref(rs); /* +1 ref for the network */
632 rc = sptlrpc_svc_wrap_reply(req);
636 req->rq_sent = ktime_get_real_seconds();
638 rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
639 (rs->rs_difficult && !rs->rs_no_ack) ?
640 LNET_ACK_REQ : LNET_NOACK_REQ,
641 &rs->rs_cb_id, req->rq_self, req->rq_source,
642 ptlrpc_req2svc(req)->srv_rep_portal,
643 req->rq_xid, req->rq_reply_off, NULL);
645 if (unlikely(rc != 0))
646 ptlrpc_req_drop_rs(req);
647 ptlrpc_connection_put(conn);
651 int ptlrpc_reply (struct ptlrpc_request *req)
653 if (req->rq_no_reply)
656 return (ptlrpc_send_reply(req, 0));
660 * For request \a req send an error reply back. Create empty
661 * reply buffers if necessary.
663 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
668 if (req->rq_no_reply)
671 if (!req->rq_repmsg) {
672 rc = lustre_pack_reply(req, 1, NULL, NULL);
677 if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
678 req->rq_status != -EPERM && req->rq_status != -ENOENT &&
679 req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
680 req->rq_type = PTL_RPC_MSG_ERR;
682 rc = ptlrpc_send_reply(req, may_be_difficult);
686 int ptlrpc_error(struct ptlrpc_request *req)
688 return ptlrpc_send_error(req, 0);
692 * Send request \a request.
693 * if \a noreply is set, don't expect any reply back and don't set up
695 * Returns 0 on success or error code.
697 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
702 struct lnet_handle_md bulk_cookie;
703 struct ptlrpc_connection *connection;
704 struct lnet_handle_me reply_me_h;
705 struct lnet_md reply_md;
706 struct obd_import *imp = request->rq_import;
707 struct obd_device *obd = imp->imp_obd;
710 LNetInvalidateMDHandle(&bulk_cookie);
712 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
715 LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
716 LASSERT(request->rq_wait_ctx == 0);
718 /* If this is a re-transmit, we're required to have disengaged
719 * cleanly from the previous attempt */
720 LASSERT(!request->rq_receiving_reply);
721 LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
722 (imp->imp_state == LUSTRE_IMP_FULL)));
724 if (unlikely(obd != NULL && obd->obd_fail)) {
725 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
727 /* this prevents us from waiting in ptlrpc_queue_wait */
728 spin_lock(&request->rq_lock);
730 spin_unlock(&request->rq_lock);
731 request->rq_status = -ENODEV;
735 connection = imp->imp_connection;
737 lustre_msg_set_handle(request->rq_reqmsg,
738 &imp->imp_remote_handle);
739 lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
740 lustre_msg_set_conn_cnt(request->rq_reqmsg,
742 lustre_msghdr_set_flags(request->rq_reqmsg,
743 imp->imp_msghdr_flags);
745 /* If it's the first time to resend the request for EINPROGRESS,
746 * we need to allocate a new XID (see after_reply()), it's different
747 * from the resend for reply timeout. */
748 if (request->rq_nr_resend != 0 &&
749 list_empty(&request->rq_unreplied_list)) {
751 /* resend for EINPROGRESS, allocate new xid to avoid reply
753 spin_lock(&imp->imp_lock);
754 ptlrpc_assign_next_xid_nolock(request);
755 min_xid = ptlrpc_known_replied_xid(imp);
756 spin_unlock(&imp->imp_lock);
758 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
759 DEBUG_REQ(D_RPCTRACE, request, "Allocating new xid for "
760 "resend on EINPROGRESS");
763 if (request->rq_bulk != NULL) {
764 ptlrpc_set_bulk_mbits(request);
765 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
768 if (list_empty(&request->rq_unreplied_list) ||
769 request->rq_xid <= imp->imp_known_replied_xid) {
770 DEBUG_REQ(D_ERROR, request, "xid: %llu, replied: %llu, "
771 "list_empty:%d\n", request->rq_xid,
772 imp->imp_known_replied_xid,
773 list_empty(&request->rq_unreplied_list));
777 /** For enabled AT all request should have AT_SUPPORT in the
778 * FULL import state when OBD_CONNECT_AT is set */
779 LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
780 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
781 !(imp->imp_connect_data.ocd_connect_flags &
784 if (request->rq_resend) {
785 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
786 if (request->rq_resend_cb != NULL)
787 request->rq_resend_cb(request, &request->rq_async_args);
789 if (request->rq_memalloc)
790 mpflag = cfs_memory_pressure_get_and_set();
792 rc = sptlrpc_cli_wrap_request(request);
796 /* bulk register should be done after wrap_request() */
797 if (request->rq_bulk != NULL) {
798 rc = ptlrpc_register_bulk (request);
800 GOTO(cleanup_bulk, rc);
802 * All the mds in the request will have the same cpt
803 * encoded in the cookie. So we can just get the first
806 bulk_cookie = request->rq_bulk->bd_mds[0];
810 LASSERT (request->rq_replen != 0);
811 if (request->rq_repbuf == NULL) {
812 LASSERT(request->rq_repdata == NULL);
813 LASSERT(request->rq_repmsg == NULL);
814 rc = sptlrpc_cli_alloc_repbuf(request,
817 /* this prevents us from looping in
818 * ptlrpc_queue_wait */
819 spin_lock(&request->rq_lock);
821 spin_unlock(&request->rq_lock);
822 request->rq_status = rc;
823 GOTO(cleanup_bulk, rc);
826 request->rq_repdata = NULL;
827 request->rq_repmsg = NULL;
830 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
831 connection->c_peer, request->rq_xid, 0,
832 LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
834 CERROR("LNetMEAttach failed: %d\n", rc);
835 LASSERT (rc == -ENOMEM);
836 GOTO(cleanup_bulk, rc = -ENOMEM);
840 spin_lock(&request->rq_lock);
841 /* We are responsible for unlinking the reply buffer */
842 request->rq_reply_unlinked = noreply;
843 request->rq_receiving_reply = !noreply;
844 /* Clear any flags that may be present from previous sends. */
845 request->rq_req_unlinked = 0;
846 request->rq_replied = 0;
848 request->rq_timedout = 0;
849 request->rq_net_err = 0;
850 request->rq_resend = 0;
851 request->rq_restart = 0;
852 request->rq_reply_truncated = 0;
853 spin_unlock(&request->rq_lock);
856 reply_md.start = request->rq_repbuf;
857 reply_md.length = request->rq_repbuf_len;
858 /* Allow multiple early replies */
859 reply_md.threshold = LNET_MD_THRESH_INF;
860 /* Manage remote for early replies */
861 reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
862 LNET_MD_MANAGE_REMOTE |
863 LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
864 reply_md.user_ptr = &request->rq_reply_cbid;
865 reply_md.eq_handle = ptlrpc_eq_h;
867 /* We must see the unlink callback to set rq_reply_unlinked,
868 * so we can't auto-unlink */
869 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
870 &request->rq_reply_md_h);
872 CERROR("LNetMDAttach failed: %d\n", rc);
873 LASSERT (rc == -ENOMEM);
874 spin_lock(&request->rq_lock);
875 /* ...but the MD attach didn't succeed... */
876 request->rq_receiving_reply = 0;
877 spin_unlock(&request->rq_lock);
878 GOTO(cleanup_me, rc = -ENOMEM);
881 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu"
883 request->rq_repbuf_len, request->rq_xid,
884 request->rq_reply_portal);
887 /* add references on request for request_out_callback */
888 ptlrpc_request_addref(request);
889 if (obd != NULL && obd->obd_svc_stats != NULL)
890 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
891 atomic_read(&imp->imp_inflight));
893 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
895 request->rq_sent_ns = ktime_get_real();
896 request->rq_sent = ktime_get_real_seconds();
897 /* We give the server rq_timeout secs to process the req, and
898 add the network latency for our local timeout. */
899 request->rq_deadline = request->rq_sent + request->rq_timeout +
900 ptlrpc_at_get_net_latency(request);
902 ptlrpc_pinger_sending_on_import(imp);
904 DEBUG_REQ(D_INFO, request, "send flg=%x",
905 lustre_msg_get_flags(request->rq_reqmsg));
906 rc = ptl_send_buf(&request->rq_req_md_h,
907 request->rq_reqbuf, request->rq_reqdata_len,
908 LNET_NOACK_REQ, &request->rq_req_cbid,
909 LNET_NID_ANY, connection->c_peer,
910 request->rq_request_portal,
911 request->rq_xid, 0, &bulk_cookie);
915 request->rq_req_unlinked = 1;
916 ptlrpc_req_finished(request);
921 /* MEUnlink is safe; the PUT didn't even get off the ground, and
922 * nobody apart from the PUT's target has the right nid+XID to
923 * access the reply buffer. */
924 rc2 = LNetMEUnlink(reply_me_h);
926 /* UNLINKED callback called synchronously */
927 LASSERT(!request->rq_receiving_reply);
930 /* We do sync unlink here as there was no real transfer here so
931 * the chance to have long unlink to sluggish net is smaller here. */
932 ptlrpc_unregister_bulk(request, 0);
935 /* set rq_sent so that this request is treated
936 * as a delayed send in the upper layers */
937 request->rq_sent = ktime_get_real_seconds();
940 if (request->rq_memalloc)
941 cfs_memory_pressure_restore(mpflag);
945 EXPORT_SYMBOL(ptl_send_rpc);
948 * Register request buffer descriptor for request receiving.
950 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
952 struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
953 static struct lnet_process_id match_id = {
959 struct lnet_handle_me me_h;
961 CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
962 service->srv_req_portal);
964 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
967 /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
968 * which means buffer can only be attached on local CPT, and LND
969 * threads can find it by grabbing a local lock */
970 rc = LNetMEAttach(service->srv_req_portal,
971 match_id, 0, ~0, LNET_UNLINK,
972 rqbd->rqbd_svcpt->scp_cpt >= 0 ?
973 LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
975 CERROR("LNetMEAttach failed: %d\n", rc);
979 LASSERT(rqbd->rqbd_refcount == 0);
980 rqbd->rqbd_refcount = 1;
982 md.start = rqbd->rqbd_buffer;
983 md.length = service->srv_buf_size;
984 md.max_size = service->srv_max_req_size;
985 md.threshold = LNET_MD_THRESH_INF;
986 md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
987 md.user_ptr = &rqbd->rqbd_cbid;
988 md.eq_handle = ptlrpc_eq_h;
990 rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
994 CERROR("LNetMDAttach failed: %d; \n", rc);
995 LASSERT (rc == -ENOMEM);
996 rc = LNetMEUnlink (me_h);
998 rqbd->rqbd_refcount = 0;