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 <libcfs/linux/linux-mem.h>
35 #include <obd_support.h>
36 #include <lustre_net.h>
37 #include <lustre_lib.h>
39 #include <obd_class.h>
40 #include "ptlrpc_internal.h"
41 #include <lnet/lib-lnet.h> /* for CFS_FAIL_PTLRPC_OST_BULK_CB2 */
44 * Helper function. Sends \a len bytes from \a base at offset \a offset
45 * over \a conn connection to portal \a portal.
46 * Returns 0 on success or error code.
48 static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len,
49 enum lnet_ack_req ack, struct ptlrpc_cb_id *cbid,
50 lnet_nid_t self, struct lnet_process_id peer_id,
51 int portal, __u64 xid, unsigned int offset,
52 struct lnet_handle_md *bulk_cookie)
58 LASSERT (portal != 0);
59 CDEBUG (D_INFO, "peer_id %s\n", libcfs_id2str(peer_id));
62 md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
63 md.options = PTLRPC_MD_OPTIONS;
65 md.handler = ptlrpc_handler;
66 LNetInvalidateMDHandle(&md.bulk_handle);
69 md.bulk_handle = *bulk_cookie;
70 md.options |= LNET_MD_BULK_HANDLE;
73 if (unlikely(ack == LNET_ACK_REQ &&
74 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
75 /* don't ask for the ack to simulate failing client */
79 rc = LNetMDBind(&md, LNET_UNLINK, mdh);
80 if (unlikely(rc != 0)) {
81 CERROR ("LNetMDBind failed: %d\n", rc);
82 LASSERT (rc == -ENOMEM);
86 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n",
87 len, portal, xid, offset);
89 percpu_ref_get(&ptlrpc_pending);
91 rc = LNetPut(self, *mdh, ack,
92 peer_id, portal, xid, offset, 0);
93 if (unlikely(rc != 0)) {
95 /* We're going to get an UNLINK event when I unlink below,
96 * which will complete just like any other failed send, so
97 * I fall through and return success here! */
98 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
99 libcfs_id2str(peer_id), portal, xid, rc);
100 rc2 = LNetMDUnlink(*mdh);
101 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
107 static void mdunlink_iterate_helper(struct lnet_handle_md *bd_mds, int count)
111 for (i = 0; i < count; i++)
112 LNetMDUnlink(bd_mds[i]);
115 #ifdef HAVE_SERVER_SUPPORT
117 * Prepare bulk descriptor for specified incoming request \a req that
118 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
119 * the bulk to be sent. Used on server-side after request was already
121 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
124 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
125 unsigned nfrags, unsigned max_brw,
128 const struct ptlrpc_bulk_frag_ops
131 struct obd_export *exp = req->rq_export;
132 struct ptlrpc_bulk_desc *desc;
135 LASSERT(ptlrpc_is_bulk_op_active(type));
137 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
141 desc->bd_export = class_export_get(exp);
144 desc->bd_cbid.cbid_fn = server_bulk_callback;
145 desc->bd_cbid.cbid_arg = desc;
147 /* NB we don't assign rq_bulk here; server-side requests are
148 * re-used, and the handler frees the bulk desc explicitly. */
152 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
155 * Starts bulk transfer for descriptor \a desc on the server.
156 * Returns 0 on success or error code.
158 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
160 struct obd_export *exp = desc->bd_export;
162 struct lnet_process_id peer_id;
170 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
173 /* NB no locking required until desc is on the network */
174 LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
176 LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
177 LASSERT(desc->bd_cbid.cbid_arg == desc);
180 * Multi-Rail: get the preferred self and peer NIDs from the
181 * request, so they are based on the route taken by the
184 self_nid = desc->bd_req->rq_self;
185 peer_id = desc->bd_req->rq_source;
187 /* NB total length may be 0 for a read past EOF, so we send 0
188 * length bulks, since the client expects bulk events.
190 * The client may not need all of the bulk mbits for the RPC. The RPC
191 * used the mbits of the highest bulk mbits needed, and the server masks
192 * off high bits to get bulk count for this RPC. LU-1431 */
193 mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
194 total_md = desc->bd_req->rq_mbits - mbits + 1;
195 desc->bd_refs = total_md;
196 desc->bd_failure = 0;
198 md.user_ptr = &desc->bd_cbid;
199 md.handler = ptlrpc_handler;
200 md.threshold = 2; /* SENT and ACK/REPLY */
202 for (posted_md = 0; posted_md < total_md; mbits++) {
203 md.options = PTLRPC_MD_OPTIONS;
205 /* NB it's assumed that source and sink buffer frags are
206 * page-aligned. Otherwise we'd have to send client bulk
207 * sizes over and split server buffer accordingly */
208 ptlrpc_fill_bulk_md(&md, desc, posted_md);
209 rc = LNetMDBind(&md, LNET_UNLINK, &desc->bd_mds[posted_md]);
211 CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
212 exp->exp_obd->obd_name, posted_md, rc);
213 LASSERT(rc == -ENOMEM);
214 if (posted_md == 0) {
215 desc->bd_md_count = 0;
220 percpu_ref_get(&ptlrpc_pending);
222 /* sanity.sh 224c: lets skip last md */
223 if (posted_md == desc->bd_md_max_brw - 1)
224 OBD_FAIL_CHECK_RESET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
225 CFS_FAIL_PTLRPC_OST_BULK_CB2);
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, false);
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_refs -= total_md - posted_md;
252 spin_unlock(&desc->bd_lock);
253 LASSERT(desc->bd_refs >= 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 LASSERT(!in_interrupt()); /* might sleep */
275 if (!ptlrpc_server_bulk_active(desc)) /* completed or */
276 return; /* never started */
278 /* We used to poison the pages with 0xab here because we did not want to
279 * send any meaningful data over the wire for evicted clients (bug 9297)
280 * However, this is no longer safe now that we use the page cache on the
283 /* The unlink ensures the callback happens ASAP and is the last
284 * one. If it fails, it must be because completion just happened,
285 * but we must still wait_event_idle_timeout() in this case, to give
286 * us 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 int seconds = PTLRPC_REQ_LONG_UNLINK;
295 while (seconds > 0 &&
296 wait_event_idle_timeout(desc->bd_waitq,
297 !ptlrpc_server_bulk_active(desc),
298 cfs_time_seconds(1)) == 0)
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;
324 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
327 /* NB no locking required until desc is on the network */
328 LASSERT(desc->bd_nob > 0);
329 LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
330 LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
331 LASSERT(desc->bd_req != NULL);
332 LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type));
334 /* cleanup the state of the bulk for it will be reused */
335 if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
336 desc->bd_nob_transferred = 0;
337 else if (desc->bd_nob_transferred != 0)
338 /* If the network failed after an RPC was sent, this condition
339 * could happen. Rather than assert (was here before), return
343 desc->bd_failure = 0;
345 peer = desc->bd_import->imp_connection->c_peer;
347 LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
348 LASSERT(desc->bd_cbid.cbid_arg == desc);
350 total_md = desc->bd_md_count;
351 /* rq_mbits is matchbits of the final bulk */
352 mbits = req->rq_mbits - desc->bd_md_count + 1;
354 LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK),
355 "first mbits = x%llu, last mbits = x%llu\n",
356 mbits, req->rq_mbits);
357 LASSERTF(!(desc->bd_registered &&
358 req->rq_send_state != LUSTRE_IMP_REPLAY) ||
359 mbits != desc->bd_last_mbits,
360 "registered: %d rq_mbits: %llu bd_last_mbits: %llu\n",
361 desc->bd_registered, mbits, desc->bd_last_mbits);
363 desc->bd_registered = 1;
364 desc->bd_last_mbits = mbits;
365 desc->bd_refs = total_md;
366 md.user_ptr = &desc->bd_cbid;
367 md.handler = ptlrpc_handler;
368 md.threshold = 1; /* PUT or GET */
370 for (posted_md = 0; posted_md < desc->bd_md_count;
371 posted_md++, mbits++) {
372 md.options = PTLRPC_MD_OPTIONS |
373 (ptlrpc_is_bulk_op_get(desc->bd_type) ?
374 LNET_MD_OP_GET : LNET_MD_OP_PUT);
375 ptlrpc_fill_bulk_md(&md, desc, posted_md);
377 if (posted_md > 0 && posted_md + 1 == desc->bd_md_count &&
378 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
381 me = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
382 LNET_UNLINK, LNET_INS_AFTER);
383 rc = PTR_ERR_OR_ZERO(me);
386 CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
387 desc->bd_import->imp_obd->obd_name, mbits,
391 percpu_ref_get(&ptlrpc_pending);
393 /* About to let the network at it... */
394 rc = LNetMDAttach(me, &md, LNET_UNLINK,
395 &desc->bd_mds[posted_md]);
397 CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
398 desc->bd_import->imp_obd->obd_name, mbits,
405 LASSERT(rc == -ENOMEM);
406 spin_lock(&desc->bd_lock);
407 desc->bd_refs -= total_md - posted_md;
408 spin_unlock(&desc->bd_lock);
409 LASSERT(desc->bd_refs >= 0);
410 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
411 req->rq_status = -ENOMEM;
412 desc->bd_registered = 0;
416 spin_lock(&desc->bd_lock);
417 /* Holler if peer manages to touch buffers before he knows the mbits */
418 if (desc->bd_refs != total_md)
419 CWARN("%s: Peer %s touched %d buffers while I registered\n",
420 desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
421 total_md - desc->bd_refs);
422 spin_unlock(&desc->bd_lock);
425 "Setup %u bulk %s buffers: %u pages %u bytes, mbits x%#llx-%#llx, portal %u\n",
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;
445 LASSERT(!in_interrupt()); /* might sleep */
448 desc->bd_registered = 0;
450 /* Let's setup deadline for reply unlink. */
451 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
452 async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
453 req->rq_bulk_deadline = ktime_get_real_seconds() +
454 PTLRPC_REQ_LONG_UNLINK;
456 if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
457 RETURN(1); /* never registered */
459 LASSERT(desc->bd_req == req); /* bd_req NULL until registered */
461 /* the unlink ensures the callback happens ASAP and is the last
462 * one. If it fails, it must be because completion just happened,
463 * but we must still wait_event_idle_timeout() in this case to give
464 * us a chance to run client_bulk_callback()
466 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
468 if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
469 RETURN(1); /* never registered */
471 /* Move to "Unregistering" phase as bulk was not unlinked yet. */
472 ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
474 /* Do not wait for unlink to finish. */
479 /* The wq argument is ignored by user-space wait_event macros */
480 wait_queue_head_t *wq = (req->rq_set != NULL) ?
481 &req->rq_set->set_waitq :
482 &req->rq_reply_waitq;
484 * Network access will complete in finite time but the HUGE
485 * timeout lets us CWARN for visibility of sluggish NALs.
487 int seconds = PTLRPC_REQ_LONG_UNLINK;
489 while (seconds > 0 &&
490 wait_event_idle_timeout(*wq,
491 !ptlrpc_client_bulk_active(req),
492 cfs_time_seconds(1)) == 0)
495 ptlrpc_rqphase_move(req, req->rq_next_phase);
499 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
505 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
507 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
508 struct ptlrpc_service *svc = svcpt->scp_service;
509 timeout_t service_timeout;
511 service_timeout = clamp_t(timeout_t, ktime_get_real_seconds() -
512 req->rq_arrival_time.tv_sec, 1,
513 (AT_OFF ? obd_timeout * 3 / 2 : at_max));
514 if (!(flags & PTLRPC_REPLY_EARLY) &&
515 (req->rq_type != PTL_RPC_MSG_ERR) &&
516 (req->rq_reqmsg != NULL) &&
517 !(lustre_msg_get_flags(req->rq_reqmsg) &
518 (MSG_RESENT | MSG_REPLAY |
519 MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
520 /* early replies, errors and recovery requests don't count
521 * toward our service time estimate
523 timeout_t oldse = at_measured(&svcpt->scp_at_estimate,
527 DEBUG_REQ(D_ADAPTTO, req,
528 "svc %s changed estimate from %d to %d",
529 svc->srv_name, oldse,
530 at_get(&svcpt->scp_at_estimate));
533 /* Report actual service time for client latency calc */
534 lustre_msg_set_service_timeout(req->rq_repmsg, service_timeout);
535 /* Report service time estimate for future client reqs, but report 0
536 * (to be ignored by client) if it's an error reply during recovery.
539 if (req->rq_type == PTL_RPC_MSG_ERR &&
540 (req->rq_export == NULL ||
541 req->rq_export->exp_obd->obd_recovering)) {
542 lustre_msg_set_timeout(req->rq_repmsg, 0);
546 if (req->rq_export && req->rq_reqmsg != NULL &&
547 (flags & PTLRPC_REPLY_EARLY) &&
548 lustre_msg_get_flags(req->rq_reqmsg) &
549 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
550 struct obd_device *exp_obd = req->rq_export->exp_obd;
552 timeout = ktime_get_real_seconds() -
553 req->rq_arrival_time.tv_sec +
554 min_t(timeout_t, at_extra,
555 exp_obd->obd_recovery_timeout / 4);
557 timeout = at_get(&svcpt->scp_at_estimate);
559 lustre_msg_set_timeout(req->rq_repmsg, timeout);
562 if (req->rq_reqmsg &&
563 !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
564 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
565 "req_flags=%#x magic=%x/%x len=%d\n",
566 flags, lustre_msg_get_flags(req->rq_reqmsg),
567 lustre_msg_get_magic(req->rq_reqmsg),
568 lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
573 * Send request reply from request \a req reply buffer.
574 * \a flags defines reply types
575 * Returns 0 on success or error code
577 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
579 struct ptlrpc_reply_state *rs = req->rq_reply_state;
580 struct ptlrpc_connection *conn;
583 /* We must already have a reply buffer (only ptlrpc_error() may be
584 * called without one). The reply generated by sptlrpc layer (e.g.
585 * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
586 * have a request buffer which is either the actual (swabbed) incoming
587 * request, or a saved copy if this is a req saved in
588 * target_queue_final_reply().
590 LASSERT (req->rq_no_reply == 0);
591 LASSERT (req->rq_reqbuf != NULL);
592 LASSERT (rs != NULL);
593 LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
594 LASSERT (req->rq_repmsg != NULL);
595 LASSERT (req->rq_repmsg == rs->rs_msg);
596 LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
597 LASSERT (rs->rs_cb_id.cbid_arg == rs);
599 /* There may be no rq_export during failover */
601 if (unlikely(req->rq_export && req->rq_export->exp_obd &&
602 req->rq_export->exp_obd->obd_fail)) {
603 /* Failed obd's only send ENODEV */
604 req->rq_type = PTL_RPC_MSG_ERR;
605 req->rq_status = -ENODEV;
606 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
607 req->rq_export->exp_obd->obd_minor);
610 if (req->rq_type != PTL_RPC_MSG_ERR)
611 req->rq_type = PTL_RPC_MSG_REPLY;
613 lustre_msg_set_type(req->rq_repmsg, req->rq_type);
614 lustre_msg_set_status(req->rq_repmsg,
615 ptlrpc_status_hton(req->rq_status));
616 lustre_msg_set_opc(req->rq_repmsg,
617 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
619 target_pack_pool_reply(req);
621 ptlrpc_at_set_reply(req, flags);
623 if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
624 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
626 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
628 if (unlikely(conn == NULL)) {
629 CERROR("not replying on NULL connection\n"); /* bug 9635 */
632 ptlrpc_rs_addref(rs); /* +1 ref for the network */
634 rc = sptlrpc_svc_wrap_reply(req);
638 req->rq_sent = ktime_get_real_seconds();
640 rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
641 (rs->rs_difficult && !rs->rs_no_ack) ?
642 LNET_ACK_REQ : LNET_NOACK_REQ,
643 &rs->rs_cb_id, req->rq_self, req->rq_source,
644 ptlrpc_req2svc(req)->srv_rep_portal,
645 req->rq_xid, req->rq_reply_off, NULL);
647 if (unlikely(rc != 0))
648 ptlrpc_req_drop_rs(req);
649 ptlrpc_connection_put(conn);
653 int ptlrpc_reply (struct ptlrpc_request *req)
655 if (req->rq_no_reply)
658 return (ptlrpc_send_reply(req, 0));
662 * For request \a req send an error reply back. Create empty
663 * reply buffers if necessary.
665 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
670 if (req->rq_no_reply)
673 if (!req->rq_repmsg) {
674 rc = lustre_pack_reply(req, 1, NULL, NULL);
679 if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
680 req->rq_status != -EPERM && req->rq_status != -ENOENT &&
681 req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
682 req->rq_type = PTL_RPC_MSG_ERR;
684 rc = ptlrpc_send_reply(req, may_be_difficult);
688 int ptlrpc_error(struct ptlrpc_request *req)
690 return ptlrpc_send_error(req, 0);
694 * Send request \a request.
695 * if \a noreply is set, don't expect any reply back and don't set up
697 * Returns 0 on success or error code.
699 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
703 struct lnet_handle_md bulk_cookie;
704 struct ptlrpc_connection *connection;
705 struct lnet_me *reply_me = NULL;
706 struct lnet_md reply_md;
707 struct obd_import *imp = request->rq_import;
708 struct obd_device *obd = imp->imp_obd;
711 LNetInvalidateMDHandle(&bulk_cookie);
713 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
716 LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
717 LASSERT(request->rq_wait_ctx == 0);
719 /* If this is a re-transmit, we're required to have disengaged
720 * cleanly from the previous attempt */
721 LASSERT(!request->rq_receiving_reply);
722 LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
723 (imp->imp_state == LUSTRE_IMP_FULL)));
725 if (unlikely(obd != NULL && obd->obd_fail)) {
726 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
728 /* this prevents us from waiting in ptlrpc_queue_wait */
729 spin_lock(&request->rq_lock);
731 spin_unlock(&request->rq_lock);
732 request->rq_status = -ENODEV;
736 connection = imp->imp_connection;
738 lustre_msg_set_handle(request->rq_reqmsg,
739 &imp->imp_remote_handle);
740 lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
741 lustre_msg_set_conn_cnt(request->rq_reqmsg,
743 lustre_msghdr_set_flags(request->rq_reqmsg,
744 imp->imp_msghdr_flags);
746 /* If it's the first time to resend the request for EINPROGRESS,
747 * we need to allocate a new XID (see after_reply()), it's different
748 * from the resend for reply timeout. */
749 if (request->rq_nr_resend != 0 &&
750 list_empty(&request->rq_unreplied_list)) {
752 /* resend for EINPROGRESS, allocate new xid to avoid reply
754 spin_lock(&imp->imp_lock);
755 ptlrpc_assign_next_xid_nolock(request);
756 min_xid = ptlrpc_known_replied_xid(imp);
757 spin_unlock(&imp->imp_lock);
759 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
760 DEBUG_REQ(D_RPCTRACE, request,
761 "Allocating new XID for resend on EINPROGRESS");
764 if (request->rq_bulk != NULL) {
765 ptlrpc_set_bulk_mbits(request);
766 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
769 if (list_empty(&request->rq_unreplied_list) ||
770 request->rq_xid <= imp->imp_known_replied_xid) {
771 DEBUG_REQ(D_ERROR, request,
772 "xid=%llu, replied=%llu, list_empty=%d",
773 request->rq_xid, imp->imp_known_replied_xid,
774 list_empty(&request->rq_unreplied_list));
778 /** For enabled AT all request should have AT_SUPPORT in the
779 * FULL import state when OBD_CONNECT_AT is set */
780 LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
781 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
782 !(imp->imp_connect_data.ocd_connect_flags &
785 if (request->rq_resend) {
786 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
787 if (request->rq_resend_cb != NULL)
788 request->rq_resend_cb(request, &request->rq_async_args);
790 if (request->rq_memalloc)
791 mpflag = memalloc_noreclaim_save();
793 rc = sptlrpc_cli_wrap_request(request);
797 /* bulk register should be done after wrap_request() */
798 if (request->rq_bulk != NULL) {
799 rc = ptlrpc_register_bulk (request);
801 GOTO(cleanup_bulk, rc);
803 * All the mds in the request will have the same cpt
804 * encoded in the cookie. So we can just get the first
807 bulk_cookie = request->rq_bulk->bd_mds[0];
811 LASSERT (request->rq_replen != 0);
812 if (request->rq_repbuf == NULL) {
813 LASSERT(request->rq_repdata == NULL);
814 LASSERT(request->rq_repmsg == NULL);
815 rc = sptlrpc_cli_alloc_repbuf(request,
818 /* this prevents us from looping in
819 * ptlrpc_queue_wait */
820 spin_lock(&request->rq_lock);
822 spin_unlock(&request->rq_lock);
823 request->rq_status = rc;
824 GOTO(cleanup_bulk, rc);
827 request->rq_repdata = NULL;
828 request->rq_repmsg = NULL;
831 if (request->rq_bulk &&
832 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH)) {
833 reply_me = ERR_PTR(-ENOMEM);
835 reply_me = LNetMEAttach(request->rq_reply_portal,
838 LNET_UNLINK, LNET_INS_AFTER);
841 if (IS_ERR(reply_me)) {
842 rc = PTR_ERR(reply_me);
843 CERROR("LNetMEAttach failed: %d\n", rc);
844 LASSERT(rc == -ENOMEM);
845 GOTO(cleanup_bulk, rc = -ENOMEM);
849 spin_lock(&request->rq_lock);
850 /* We are responsible for unlinking the reply buffer */
851 request->rq_reply_unlinked = noreply;
852 request->rq_receiving_reply = !noreply;
853 /* Clear any flags that may be present from previous sends. */
854 request->rq_req_unlinked = 0;
855 request->rq_replied = 0;
857 request->rq_timedout = 0;
858 request->rq_net_err = 0;
859 request->rq_resend = 0;
860 request->rq_restart = 0;
861 request->rq_reply_truncated = 0;
862 spin_unlock(&request->rq_lock);
865 reply_md.start = request->rq_repbuf;
866 reply_md.length = request->rq_repbuf_len;
867 /* Allow multiple early replies */
868 reply_md.threshold = LNET_MD_THRESH_INF;
869 /* Manage remote for early replies */
870 reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
871 LNET_MD_MANAGE_REMOTE |
872 LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
873 reply_md.user_ptr = &request->rq_reply_cbid;
874 reply_md.handler = ptlrpc_handler;
876 /* We must see the unlink callback to set rq_reply_unlinked,
877 * so we can't auto-unlink */
878 rc = LNetMDAttach(reply_me, &reply_md, LNET_RETAIN,
879 &request->rq_reply_md_h);
881 CERROR("LNetMDAttach failed: %d\n", rc);
882 LASSERT(rc == -ENOMEM);
883 spin_lock(&request->rq_lock);
884 /* ...but the MD attach didn't succeed... */
885 request->rq_receiving_reply = 0;
886 spin_unlock(&request->rq_lock);
887 GOTO(cleanup_bulk, rc = -ENOMEM);
889 percpu_ref_get(&ptlrpc_pending);
892 "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
893 request->rq_repbuf_len, request->rq_xid,
894 request->rq_reply_portal);
897 /* add references on request for request_out_callback */
898 ptlrpc_request_addref(request);
899 if (obd != NULL && obd->obd_svc_stats != NULL)
900 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
901 atomic_read(&imp->imp_inflight));
903 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
905 request->rq_sent_ns = ktime_get_real();
906 request->rq_sent = ktime_get_real_seconds();
907 /* We give the server rq_timeout secs to process the req, and
908 * add the network latency for our local timeout.
910 request->rq_deadline = request->rq_sent + request->rq_timeout +
911 ptlrpc_at_get_net_latency(request);
913 ptlrpc_pinger_sending_on_import(imp);
915 DEBUG_REQ(D_INFO, request, "send flags=%x",
916 lustre_msg_get_flags(request->rq_reqmsg));
917 rc = ptl_send_buf(&request->rq_req_md_h,
918 request->rq_reqbuf, request->rq_reqdata_len,
919 LNET_NOACK_REQ, &request->rq_req_cbid,
920 LNET_NID_ANY, connection->c_peer,
921 request->rq_request_portal,
922 request->rq_xid, 0, &bulk_cookie);
926 request->rq_req_unlinked = 1;
927 ptlrpc_req_finished(request);
931 LNetMDUnlink(request->rq_reply_md_h);
933 /* UNLINKED callback called synchronously */
934 LASSERT(!request->rq_receiving_reply);
937 /* We do sync unlink here as there was no real transfer here so
938 * the chance to have long unlink to sluggish net is smaller here. */
939 ptlrpc_unregister_bulk(request, 0);
942 /* set rq_sent so that this request is treated
943 * as a delayed send in the upper layers */
944 request->rq_sent = ktime_get_real_seconds();
947 if (request->rq_memalloc)
948 memalloc_noreclaim_restore(mpflag);
952 EXPORT_SYMBOL(ptl_send_rpc);
955 * Register request buffer descriptor for request receiving.
957 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
959 struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
960 static struct lnet_process_id match_id = {
968 CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
969 service->srv_req_portal);
971 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
974 /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
975 * which means buffer can only be attached on local CPT, and LND
976 * threads can find it by grabbing a local lock */
977 me = LNetMEAttach(service->srv_req_portal,
978 match_id, 0, ~0, LNET_UNLINK,
979 rqbd->rqbd_svcpt->scp_cpt >= 0 ?
980 LNET_INS_LOCAL : LNET_INS_AFTER);
982 CERROR("LNetMEAttach failed: %ld\n", PTR_ERR(me));
986 LASSERT(rqbd->rqbd_refcount == 0);
987 rqbd->rqbd_refcount = 1;
989 md.start = rqbd->rqbd_buffer;
990 md.length = service->srv_buf_size;
991 md.max_size = service->srv_max_req_size;
992 md.threshold = LNET_MD_THRESH_INF;
993 md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
994 md.user_ptr = &rqbd->rqbd_cbid;
995 md.handler = ptlrpc_handler;
997 rc = LNetMDAttach(me, &md, LNET_UNLINK, &rqbd->rqbd_md_h);
999 percpu_ref_get(&ptlrpc_pending);
1003 CERROR("ptlrpc: LNetMDAttach failed: rc = %d\n", rc);
1004 LASSERT(rc == -ENOMEM);
1006 rqbd->rqbd_refcount = 0;