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"
40 #include <lnet/lib-lnet.h> /* for CFS_FAIL_PTLRPC_OST_BULK_CB2 */
43 * Helper function. Sends \a len bytes from \a base at offset \a offset
44 * over \a conn connection to portal \a portal.
45 * Returns 0 on success or error code.
47 static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len,
48 enum lnet_ack_req ack, struct ptlrpc_cb_id *cbid,
49 lnet_nid_t self, struct lnet_process_id peer_id,
50 int portal, __u64 xid, unsigned int offset,
51 struct lnet_handle_md *bulk_cookie)
57 LASSERT (portal != 0);
58 CDEBUG (D_INFO, "peer_id %s\n", libcfs_id2str(peer_id));
61 md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
62 md.options = PTLRPC_MD_OPTIONS;
64 md.eq_handle = ptlrpc_eq_h;
65 LNetInvalidateMDHandle(&md.bulk_handle);
68 md.bulk_handle = *bulk_cookie;
69 md.options |= LNET_MD_BULK_HANDLE;
72 if (unlikely(ack == LNET_ACK_REQ &&
73 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
74 /* don't ask for the ack to simulate failing client */
78 rc = LNetMDBind (md, LNET_UNLINK, mdh);
79 if (unlikely(rc != 0)) {
80 CERROR ("LNetMDBind failed: %d\n", rc);
81 LASSERT (rc == -ENOMEM);
85 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
86 len, portal, xid, offset);
88 rc = LNetPut(self, *mdh, ack,
89 peer_id, portal, xid, offset, 0);
90 if (unlikely(rc != 0)) {
92 /* We're going to get an UNLINK event when I unlink below,
93 * which will complete just like any other failed send, so
94 * I fall through and return success here! */
95 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
96 libcfs_id2str(peer_id), portal, xid, rc);
97 rc2 = LNetMDUnlink(*mdh);
98 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
104 static void mdunlink_iterate_helper(struct lnet_handle_md *bd_mds, int count)
108 for (i = 0; i < count; i++)
109 LNetMDUnlink(bd_mds[i]);
112 #ifdef HAVE_SERVER_SUPPORT
114 * Prepare bulk descriptor for specified incoming request \a req that
115 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
116 * the bulk to be sent. Used on server-side after request was already
118 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
121 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
122 unsigned nfrags, unsigned max_brw,
125 const struct ptlrpc_bulk_frag_ops
128 struct obd_export *exp = req->rq_export;
129 struct ptlrpc_bulk_desc *desc;
132 LASSERT(ptlrpc_is_bulk_op_active(type));
134 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
138 desc->bd_export = class_export_get(exp);
141 desc->bd_cbid.cbid_fn = server_bulk_callback;
142 desc->bd_cbid.cbid_arg = desc;
144 /* NB we don't assign rq_bulk here; server-side requests are
145 * re-used, and the handler frees the bulk desc explicitly. */
149 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
152 * Starts bulk transfer for descriptor \a desc on the server.
153 * Returns 0 on success or error code.
155 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
157 struct obd_export *exp = desc->bd_export;
159 struct lnet_process_id peer_id;
167 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
170 /* NB no locking required until desc is on the network */
171 LASSERT(desc->bd_md_count == 0);
172 LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
174 LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
175 LASSERT(desc->bd_cbid.cbid_arg == desc);
178 * Multi-Rail: get the preferred self and peer NIDs from the
179 * request, so they are based on the route taken by the
182 self_nid = desc->bd_req->rq_self;
183 peer_id = desc->bd_req->rq_source;
185 /* NB total length may be 0 for a read past EOF, so we send 0
186 * length bulks, since the client expects bulk events.
188 * The client may not need all of the bulk mbits for the RPC. The RPC
189 * used the mbits of the highest bulk mbits needed, and the server masks
190 * off high bits to get bulk count for this RPC. LU-1431 */
191 mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
192 total_md = desc->bd_req->rq_mbits - mbits + 1;
194 desc->bd_md_count = total_md;
195 desc->bd_failure = 0;
197 md.user_ptr = &desc->bd_cbid;
198 md.eq_handle = ptlrpc_eq_h;
199 md.threshold = 2; /* SENT and ACK/REPLY */
201 for (posted_md = 0; posted_md < total_md; mbits++) {
202 md.options = PTLRPC_MD_OPTIONS;
204 /* NB it's assumed that source and sink buffer frags are
205 * page-aligned. Otherwise we'd have to send client bulk
206 * sizes over and split server buffer accordingly */
207 ptlrpc_fill_bulk_md(&md, desc, posted_md);
208 rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
210 CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
211 exp->exp_obd->obd_name, posted_md, rc);
212 LASSERT(rc == -ENOMEM);
213 if (posted_md == 0) {
214 desc->bd_md_count = 0;
220 /* sanity.sh 224c: lets skip last md */
221 if (posted_md == desc->bd_md_max_brw - 1)
222 OBD_FAIL_CHECK_RESET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
223 CFS_FAIL_PTLRPC_OST_BULK_CB2);
225 /* Network is about to get at the memory */
226 if (ptlrpc_is_bulk_put_source(desc->bd_type))
227 rc = LNetPut(self_nid, desc->bd_mds[posted_md],
228 LNET_ACK_REQ, peer_id,
229 desc->bd_portal, mbits, 0, 0);
231 rc = LNetGet(self_nid, desc->bd_mds[posted_md],
232 peer_id, desc->bd_portal, mbits, 0, false);
236 CERROR("%s: failed bulk transfer with %s:%u x%llu: "
237 "rc = %d\n", exp->exp_obd->obd_name,
238 libcfs_id2str(peer_id), desc->bd_portal,
245 /* Can't send, so we unlink the MD bound above. The UNLINK
246 * event this creates will signal completion with failure,
247 * so we return SUCCESS here! */
248 spin_lock(&desc->bd_lock);
249 desc->bd_md_count -= total_md - posted_md;
250 spin_unlock(&desc->bd_lock);
251 LASSERT(desc->bd_md_count >= 0);
253 mdunlink_iterate_helper(desc->bd_mds, posted_md);
257 CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
258 "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
259 desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id),
260 mbits - posted_md, mbits - 1);
266 * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
267 * serialises with completion callback)
269 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
271 struct l_wait_info lwi;
274 LASSERT(!in_interrupt()); /* might sleep */
276 if (!ptlrpc_server_bulk_active(desc)) /* completed or */
277 return; /* never started */
279 /* We used to poison the pages with 0xab here because we did not want to
280 * send any meaningful data over the wire for evicted clients (bug 9297)
281 * However, this is no longer safe now that we use the page cache on the
284 /* The unlink ensures the callback happens ASAP and is the last
285 * one. If it fails, it must be because completion just happened,
286 * but we must still l_wait_event() in this case, to give liblustre
287 * a chance to run server_bulk_callback()*/
288 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
291 /* Network access will complete in finite time but the HUGE
292 * timeout lets us CWARN for visibility of sluggish NALs */
293 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
294 cfs_time_seconds(1), NULL, NULL);
295 rc = l_wait_event(desc->bd_waitq,
296 !ptlrpc_server_bulk_active(desc), &lwi);
300 LASSERT(rc == -ETIMEDOUT);
301 CWARN("Unexpectedly long timeout: desc %p\n", desc);
304 #endif /* HAVE_SERVER_SUPPORT */
307 * Register bulk at the sender for later transfer.
308 * Returns 0 on success or error code.
310 int ptlrpc_register_bulk(struct ptlrpc_request *req)
312 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
313 struct lnet_process_id peer;
322 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
325 /* NB no locking required until desc is on the network */
326 LASSERT(desc->bd_nob > 0);
327 LASSERT(desc->bd_md_count == 0);
328 LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
329 LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
330 LASSERT(desc->bd_req != NULL);
331 LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
333 /* cleanup the state of the bulk for it will be reused */
334 if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
335 desc->bd_nob_transferred = 0;
336 else if (desc->bd_nob_transferred != 0)
337 /* If the network failed after an RPC was sent, this condition
338 * could happen. Rather than assert (was here before), return
342 desc->bd_failure = 0;
344 peer = desc->bd_import->imp_connection->c_peer;
346 LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
347 LASSERT(desc->bd_cbid.cbid_arg == desc);
349 total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
350 /* rq_mbits is matchbits of the final bulk */
351 mbits = req->rq_mbits - total_md + 1;
353 LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
354 "first mbits = x%llu, last mbits = x%llu\n",
355 mbits, req->rq_mbits);
356 LASSERTF(!(desc->bd_registered &&
357 req->rq_send_state != LUSTRE_IMP_REPLAY) ||
358 mbits != desc->bd_last_mbits,
359 "registered: %d rq_mbits: %llu bd_last_mbits: %llu\n",
360 desc->bd_registered, mbits, desc->bd_last_mbits);
362 desc->bd_registered = 1;
363 desc->bd_last_mbits = mbits;
364 desc->bd_md_count = total_md;
365 md.user_ptr = &desc->bd_cbid;
366 md.eq_handle = ptlrpc_eq_h;
367 md.threshold = 1; /* PUT or GET */
369 for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) {
370 md.options = PTLRPC_MD_OPTIONS |
371 (ptlrpc_is_bulk_op_get(desc->bd_type) ?
372 LNET_MD_OP_GET : LNET_MD_OP_PUT);
373 ptlrpc_fill_bulk_md(&md, desc, posted_md);
375 if (posted_md > 0 && posted_md + 1 == total_md &&
376 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
379 me = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
380 LNET_UNLINK, LNET_INS_AFTER);
381 rc = PTR_ERR_OR_ZERO(me);
384 CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
385 desc->bd_import->imp_obd->obd_name, mbits,
390 /* About to let the network at it... */
391 rc = LNetMDAttach(me, md, LNET_UNLINK,
392 &desc->bd_mds[posted_md]);
394 CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
395 desc->bd_import->imp_obd->obd_name, mbits,
403 LASSERT(rc == -ENOMEM);
404 spin_lock(&desc->bd_lock);
405 desc->bd_md_count -= total_md - posted_md;
406 spin_unlock(&desc->bd_lock);
407 LASSERT(desc->bd_md_count >= 0);
408 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
409 req->rq_status = -ENOMEM;
410 desc->bd_registered = 0;
414 spin_lock(&desc->bd_lock);
415 /* Holler if peer manages to touch buffers before he knows the mbits */
416 if (desc->bd_md_count != total_md)
417 CWARN("%s: Peer %s touched %d buffers while I registered\n",
418 desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
419 total_md - desc->bd_md_count);
420 spin_unlock(&desc->bd_lock);
422 CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
423 "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count,
424 ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
425 desc->bd_iov_count, desc->bd_nob,
426 desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
432 * Disconnect a bulk desc from the network. Idempotent. Not
433 * thread-safe (i.e. only interlocks with completion callback).
434 * Returns 1 on success or 0 if network unregistration failed for whatever
437 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
439 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
440 struct l_wait_info lwi;
444 LASSERT(!in_interrupt()); /* might sleep */
446 /* Let's setup deadline for reply unlink. */
447 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
448 async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
449 req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK;
451 if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
452 RETURN(1); /* never registered */
454 LASSERT(desc->bd_req == req); /* bd_req NULL until registered */
456 /* the unlink ensures the callback happens ASAP and is the last
457 * one. If it fails, it must be because completion just happened,
458 * but we must still l_wait_event() in this case to give liblustre
459 * a chance to run client_bulk_callback() */
460 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
462 if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
463 RETURN(1); /* never registered */
465 /* Move to "Unregistering" phase as bulk was not unlinked yet. */
466 ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
468 /* Do not wait for unlink to finish. */
473 /* The wq argument is ignored by user-space wait_event macros */
474 wait_queue_head_t *wq = (req->rq_set != NULL) ?
475 &req->rq_set->set_waitq :
476 &req->rq_reply_waitq;
478 * Network access will complete in finite time but the HUGE
479 * timeout lets us CWARN for visibility of sluggish NALs.
481 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
482 cfs_time_seconds(1), NULL, NULL);
483 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
485 ptlrpc_rqphase_move(req, req->rq_next_phase);
489 LASSERT(rc == -ETIMEDOUT);
490 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
496 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
498 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
499 struct ptlrpc_service *svc = svcpt->scp_service;
500 int service_time = max_t(int, ktime_get_real_seconds() -
501 req->rq_arrival_time.tv_sec, 1);
503 if (!(flags & PTLRPC_REPLY_EARLY) &&
504 (req->rq_type != PTL_RPC_MSG_ERR) &&
505 (req->rq_reqmsg != NULL) &&
506 !(lustre_msg_get_flags(req->rq_reqmsg) &
507 (MSG_RESENT | MSG_REPLAY |
508 MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
509 /* early replies, errors and recovery requests don't count
510 * toward our service time estimate */
511 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
514 DEBUG_REQ(D_ADAPTTO, req,
515 "svc %s changed estimate from %d to %d",
516 svc->srv_name, oldse,
517 at_get(&svcpt->scp_at_estimate));
520 /* Report actual service time for client latency calc */
521 lustre_msg_set_service_time(req->rq_repmsg, service_time);
522 /* Report service time estimate for future client reqs, but report 0
523 * (to be ignored by client) if it's an error reply during recovery.
526 if (req->rq_type == PTL_RPC_MSG_ERR &&
527 (req->rq_export == NULL ||
528 req->rq_export->exp_obd->obd_recovering)) {
529 lustre_msg_set_timeout(req->rq_repmsg, 0);
533 if (req->rq_export && req->rq_reqmsg != NULL &&
534 (flags & PTLRPC_REPLY_EARLY) &&
535 lustre_msg_get_flags(req->rq_reqmsg) &
536 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
537 struct obd_device *exp_obd = req->rq_export->exp_obd;
539 timeout = ktime_get_real_seconds() -
540 req->rq_arrival_time.tv_sec +
541 min_t(time64_t, at_extra,
542 exp_obd->obd_recovery_timeout / 4);
544 timeout = at_get(&svcpt->scp_at_estimate);
546 lustre_msg_set_timeout(req->rq_repmsg, timeout);
549 if (req->rq_reqmsg &&
550 !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
551 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
552 "req_flags=%#x magic=%x/%x len=%d\n",
553 flags, lustre_msg_get_flags(req->rq_reqmsg),
554 lustre_msg_get_magic(req->rq_reqmsg),
555 lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
560 * Send request reply from request \a req reply buffer.
561 * \a flags defines reply types
562 * Returns 0 on success or error code
564 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
566 struct ptlrpc_reply_state *rs = req->rq_reply_state;
567 struct ptlrpc_connection *conn;
570 /* We must already have a reply buffer (only ptlrpc_error() may be
571 * called without one). The reply generated by sptlrpc layer (e.g.
572 * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
573 * have a request buffer which is either the actual (swabbed) incoming
574 * request, or a saved copy if this is a req saved in
575 * target_queue_final_reply().
577 LASSERT (req->rq_no_reply == 0);
578 LASSERT (req->rq_reqbuf != NULL);
579 LASSERT (rs != NULL);
580 LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
581 LASSERT (req->rq_repmsg != NULL);
582 LASSERT (req->rq_repmsg == rs->rs_msg);
583 LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
584 LASSERT (rs->rs_cb_id.cbid_arg == rs);
586 /* There may be no rq_export during failover */
588 if (unlikely(req->rq_export && req->rq_export->exp_obd &&
589 req->rq_export->exp_obd->obd_fail)) {
590 /* Failed obd's only send ENODEV */
591 req->rq_type = PTL_RPC_MSG_ERR;
592 req->rq_status = -ENODEV;
593 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
594 req->rq_export->exp_obd->obd_minor);
597 if (req->rq_type != PTL_RPC_MSG_ERR)
598 req->rq_type = PTL_RPC_MSG_REPLY;
600 lustre_msg_set_type(req->rq_repmsg, req->rq_type);
601 lustre_msg_set_status(req->rq_repmsg,
602 ptlrpc_status_hton(req->rq_status));
603 lustre_msg_set_opc(req->rq_repmsg,
604 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
606 target_pack_pool_reply(req);
608 ptlrpc_at_set_reply(req, flags);
610 if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
611 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
613 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
615 if (unlikely(conn == NULL)) {
616 CERROR("not replying on NULL connection\n"); /* bug 9635 */
619 ptlrpc_rs_addref(rs); /* +1 ref for the network */
621 rc = sptlrpc_svc_wrap_reply(req);
625 req->rq_sent = ktime_get_real_seconds();
627 rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
628 (rs->rs_difficult && !rs->rs_no_ack) ?
629 LNET_ACK_REQ : LNET_NOACK_REQ,
630 &rs->rs_cb_id, req->rq_self, req->rq_source,
631 ptlrpc_req2svc(req)->srv_rep_portal,
632 req->rq_xid, req->rq_reply_off, NULL);
634 if (unlikely(rc != 0))
635 ptlrpc_req_drop_rs(req);
636 ptlrpc_connection_put(conn);
640 int ptlrpc_reply (struct ptlrpc_request *req)
642 if (req->rq_no_reply)
645 return (ptlrpc_send_reply(req, 0));
649 * For request \a req send an error reply back. Create empty
650 * reply buffers if necessary.
652 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
657 if (req->rq_no_reply)
660 if (!req->rq_repmsg) {
661 rc = lustre_pack_reply(req, 1, NULL, NULL);
666 if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
667 req->rq_status != -EPERM && req->rq_status != -ENOENT &&
668 req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
669 req->rq_type = PTL_RPC_MSG_ERR;
671 rc = ptlrpc_send_reply(req, may_be_difficult);
675 int ptlrpc_error(struct ptlrpc_request *req)
677 return ptlrpc_send_error(req, 0);
681 * Send request \a request.
682 * if \a noreply is set, don't expect any reply back and don't set up
684 * Returns 0 on success or error code.
686 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
690 struct lnet_handle_md bulk_cookie;
691 struct ptlrpc_connection *connection;
692 struct lnet_me *reply_me = NULL;
693 struct lnet_md reply_md;
694 struct obd_import *imp = request->rq_import;
695 struct obd_device *obd = imp->imp_obd;
698 LNetInvalidateMDHandle(&bulk_cookie);
700 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
703 LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
704 LASSERT(request->rq_wait_ctx == 0);
706 /* If this is a re-transmit, we're required to have disengaged
707 * cleanly from the previous attempt */
708 LASSERT(!request->rq_receiving_reply);
709 LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
710 (imp->imp_state == LUSTRE_IMP_FULL)));
712 if (unlikely(obd != NULL && obd->obd_fail)) {
713 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
715 /* this prevents us from waiting in ptlrpc_queue_wait */
716 spin_lock(&request->rq_lock);
718 spin_unlock(&request->rq_lock);
719 request->rq_status = -ENODEV;
723 connection = imp->imp_connection;
725 lustre_msg_set_handle(request->rq_reqmsg,
726 &imp->imp_remote_handle);
727 lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
728 lustre_msg_set_conn_cnt(request->rq_reqmsg,
730 lustre_msghdr_set_flags(request->rq_reqmsg,
731 imp->imp_msghdr_flags);
733 /* If it's the first time to resend the request for EINPROGRESS,
734 * we need to allocate a new XID (see after_reply()), it's different
735 * from the resend for reply timeout. */
736 if (request->rq_nr_resend != 0 &&
737 list_empty(&request->rq_unreplied_list)) {
739 /* resend for EINPROGRESS, allocate new xid to avoid reply
741 spin_lock(&imp->imp_lock);
742 ptlrpc_assign_next_xid_nolock(request);
743 min_xid = ptlrpc_known_replied_xid(imp);
744 spin_unlock(&imp->imp_lock);
746 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
747 DEBUG_REQ(D_RPCTRACE, request,
748 "Allocating new XID for resend on EINPROGRESS");
751 if (request->rq_bulk != NULL) {
752 ptlrpc_set_bulk_mbits(request);
753 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
756 if (list_empty(&request->rq_unreplied_list) ||
757 request->rq_xid <= imp->imp_known_replied_xid) {
758 DEBUG_REQ(D_ERROR, request,
759 "xid=%llu, replied=%llu, list_empty=%d",
760 request->rq_xid, imp->imp_known_replied_xid,
761 list_empty(&request->rq_unreplied_list));
765 /** For enabled AT all request should have AT_SUPPORT in the
766 * FULL import state when OBD_CONNECT_AT is set */
767 LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
768 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
769 !(imp->imp_connect_data.ocd_connect_flags &
772 if (request->rq_resend) {
773 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
774 if (request->rq_resend_cb != NULL)
775 request->rq_resend_cb(request, &request->rq_async_args);
777 if (request->rq_memalloc)
778 mpflag = cfs_memory_pressure_get_and_set();
780 rc = sptlrpc_cli_wrap_request(request);
784 /* bulk register should be done after wrap_request() */
785 if (request->rq_bulk != NULL) {
786 rc = ptlrpc_register_bulk (request);
788 GOTO(cleanup_bulk, rc);
790 * All the mds in the request will have the same cpt
791 * encoded in the cookie. So we can just get the first
794 bulk_cookie = request->rq_bulk->bd_mds[0];
798 LASSERT (request->rq_replen != 0);
799 if (request->rq_repbuf == NULL) {
800 LASSERT(request->rq_repdata == NULL);
801 LASSERT(request->rq_repmsg == NULL);
802 rc = sptlrpc_cli_alloc_repbuf(request,
805 /* this prevents us from looping in
806 * ptlrpc_queue_wait */
807 spin_lock(&request->rq_lock);
809 spin_unlock(&request->rq_lock);
810 request->rq_status = rc;
811 GOTO(cleanup_bulk, rc);
814 request->rq_repdata = NULL;
815 request->rq_repmsg = NULL;
818 reply_me = LNetMEAttach(request->rq_reply_portal,
819 connection->c_peer, request->rq_xid, 0,
820 LNET_UNLINK, LNET_INS_AFTER);
821 if (IS_ERR(reply_me)) {
822 rc = PTR_ERR(reply_me);
823 CERROR("LNetMEAttach failed: %d\n", rc);
824 LASSERT(rc == -ENOMEM);
825 GOTO(cleanup_bulk, rc = -ENOMEM);
829 spin_lock(&request->rq_lock);
830 /* We are responsible for unlinking the reply buffer */
831 request->rq_reply_unlinked = noreply;
832 request->rq_receiving_reply = !noreply;
833 /* Clear any flags that may be present from previous sends. */
834 request->rq_req_unlinked = 0;
835 request->rq_replied = 0;
837 request->rq_timedout = 0;
838 request->rq_net_err = 0;
839 request->rq_resend = 0;
840 request->rq_restart = 0;
841 request->rq_reply_truncated = 0;
842 spin_unlock(&request->rq_lock);
845 reply_md.start = request->rq_repbuf;
846 reply_md.length = request->rq_repbuf_len;
847 /* Allow multiple early replies */
848 reply_md.threshold = LNET_MD_THRESH_INF;
849 /* Manage remote for early replies */
850 reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
851 LNET_MD_MANAGE_REMOTE |
852 LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
853 reply_md.user_ptr = &request->rq_reply_cbid;
854 reply_md.eq_handle = ptlrpc_eq_h;
856 /* We must see the unlink callback to set rq_reply_unlinked,
857 * so we can't auto-unlink */
858 rc = LNetMDAttach(reply_me, reply_md, LNET_RETAIN,
859 &request->rq_reply_md_h);
861 CERROR("LNetMDAttach failed: %d\n", rc);
862 LASSERT(rc == -ENOMEM);
863 spin_lock(&request->rq_lock);
864 /* ...but the MD attach didn't succeed... */
865 request->rq_receiving_reply = 0;
866 spin_unlock(&request->rq_lock);
867 GOTO(cleanup_me, rc = -ENOMEM);
870 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu"
872 request->rq_repbuf_len, request->rq_xid,
873 request->rq_reply_portal);
876 /* add references on request for request_out_callback */
877 ptlrpc_request_addref(request);
878 if (obd != NULL && obd->obd_svc_stats != NULL)
879 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
880 atomic_read(&imp->imp_inflight));
882 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
884 request->rq_sent_ns = ktime_get_real();
885 request->rq_sent = ktime_get_real_seconds();
886 /* We give the server rq_timeout secs to process the req, and
887 add the network latency for our local timeout. */
888 request->rq_deadline = request->rq_sent + request->rq_timeout +
889 ptlrpc_at_get_net_latency(request);
891 ptlrpc_pinger_sending_on_import(imp);
893 DEBUG_REQ(D_INFO, request, "send flags=%x",
894 lustre_msg_get_flags(request->rq_reqmsg));
895 rc = ptl_send_buf(&request->rq_req_md_h,
896 request->rq_reqbuf, request->rq_reqdata_len,
897 LNET_NOACK_REQ, &request->rq_req_cbid,
898 LNET_NID_ANY, connection->c_peer,
899 request->rq_request_portal,
900 request->rq_xid, 0, &bulk_cookie);
904 request->rq_req_unlinked = 1;
905 ptlrpc_req_finished(request);
910 /* MEUnlink is safe; the PUT didn't even get off the ground, and
911 * nobody apart from the PUT's target has the right nid+XID to
912 * access the reply buffer.
914 LNetMEUnlink(reply_me);
915 /* UNLINKED callback called synchronously */
916 LASSERT(!request->rq_receiving_reply);
919 /* We do sync unlink here as there was no real transfer here so
920 * the chance to have long unlink to sluggish net is smaller here. */
921 ptlrpc_unregister_bulk(request, 0);
922 if (request->rq_bulk != NULL)
923 request->rq_bulk->bd_registered = 0;
926 /* set rq_sent so that this request is treated
927 * as a delayed send in the upper layers */
928 request->rq_sent = ktime_get_real_seconds();
931 if (request->rq_memalloc)
932 cfs_memory_pressure_restore(mpflag);
936 EXPORT_SYMBOL(ptl_send_rpc);
939 * Register request buffer descriptor for request receiving.
941 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
943 struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
944 static struct lnet_process_id match_id = {
952 CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
953 service->srv_req_portal);
955 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
958 /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
959 * which means buffer can only be attached on local CPT, and LND
960 * threads can find it by grabbing a local lock */
961 me = LNetMEAttach(service->srv_req_portal,
962 match_id, 0, ~0, LNET_UNLINK,
963 rqbd->rqbd_svcpt->scp_cpt >= 0 ?
964 LNET_INS_LOCAL : LNET_INS_AFTER);
966 CERROR("LNetMEAttach failed: %ld\n", PTR_ERR(me));
970 LASSERT(rqbd->rqbd_refcount == 0);
971 rqbd->rqbd_refcount = 1;
973 md.start = rqbd->rqbd_buffer;
974 md.length = service->srv_buf_size;
975 md.max_size = service->srv_max_req_size;
976 md.threshold = LNET_MD_THRESH_INF;
977 md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
978 md.user_ptr = &rqbd->rqbd_cbid;
979 md.eq_handle = ptlrpc_eq_h;
981 rc = LNetMDAttach(me, md, LNET_UNLINK, &rqbd->rqbd_md_h);
985 CERROR("LNetMDAttach failed: %d;\n", rc);
986 LASSERT(rc == -ENOMEM);
989 rqbd->rqbd_refcount = 0;