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/
32 #define DEBUG_SUBSYSTEM S_RPC
33 #include <libcfs/linux/linux-mem.h>
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.handler = ptlrpc_handler;
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 percpu_ref_get(&ptlrpc_pending);
90 rc = LNetPut(self, *mdh, ack,
91 peer_id, portal, xid, offset, 0);
92 if (unlikely(rc != 0)) {
94 /* We're going to get an UNLINK event when I unlink below,
95 * which will complete just like any other failed send, so
96 * I fall through and return success here! */
97 CERROR("LNetPut(%s, %d, %lld) failed: %d\n",
98 libcfs_id2str(peer_id), portal, xid, rc);
99 rc2 = LNetMDUnlink(*mdh);
100 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
106 #define mdunlink_iterate_helper(mds, count) \
107 __mdunlink_iterate_helper(mds, count, false)
108 static void __mdunlink_iterate_helper(struct lnet_handle_md *bd_mds,
109 int count, bool discard)
113 for (i = 0; i < count; i++)
114 __LNetMDUnlink(bd_mds[i], discard);
117 #ifdef HAVE_SERVER_SUPPORT
119 * Prepare bulk descriptor for specified incoming request \a req that
120 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
121 * the bulk to be sent. Used on server-side after request was already
123 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
126 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
127 unsigned nfrags, unsigned max_brw,
130 const struct ptlrpc_bulk_frag_ops
133 struct obd_export *exp = req->rq_export;
134 struct ptlrpc_bulk_desc *desc;
137 LASSERT(ptlrpc_is_bulk_op_active(type));
139 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
143 desc->bd_export = class_export_get(exp);
146 desc->bd_cbid.cbid_fn = server_bulk_callback;
147 desc->bd_cbid.cbid_arg = desc;
149 /* NB we don't assign rq_bulk here; server-side requests are
150 * re-used, and the handler frees the bulk desc explicitly. */
154 EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
157 * Starts bulk transfer for descriptor \a desc on the server.
158 * Returns 0 on success or error code.
160 int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
162 struct obd_export *exp = desc->bd_export;
164 struct lnet_process_id peer_id;
172 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
175 /* NB no locking required until desc is on the network */
176 LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type));
178 LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
179 LASSERT(desc->bd_cbid.cbid_arg == desc);
182 * Multi-Rail: get the preferred self and peer NIDs from the
183 * request, so they are based on the route taken by the
186 self_nid = desc->bd_req->rq_self;
187 peer_id = desc->bd_req->rq_source;
189 /* NB total length may be 0 for a read past EOF, so we send 0
190 * length bulks, since the client expects bulk events.
192 * The client may not need all of the bulk mbits for the RPC. The RPC
193 * used the mbits of the highest bulk mbits needed, and the server masks
194 * off high bits to get bulk count for this RPC. LU-1431 */
195 mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1);
196 total_md = desc->bd_req->rq_mbits - mbits + 1;
197 desc->bd_refs = total_md;
198 desc->bd_failure = 0;
200 md.user_ptr = &desc->bd_cbid;
201 md.handler = ptlrpc_handler;
202 md.threshold = 2; /* SENT and ACK/REPLY */
204 for (posted_md = 0; posted_md < total_md; mbits++) {
205 md.options = PTLRPC_MD_OPTIONS;
207 /* NB it's assumed that source and sink buffer frags are
208 * page-aligned. Otherwise we'd have to send client bulk
209 * sizes over and split server buffer accordingly */
210 ptlrpc_fill_bulk_md(&md, desc, posted_md);
211 rc = LNetMDBind(&md, LNET_UNLINK, &desc->bd_mds[posted_md]);
213 CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
214 exp->exp_obd->obd_name, posted_md, rc);
215 LASSERT(rc == -ENOMEM);
216 if (posted_md == 0) {
217 desc->bd_md_count = 0;
222 percpu_ref_get(&ptlrpc_pending);
224 /* sanity.sh 224c: lets skip last md */
225 if (posted_md == desc->bd_md_max_brw - 1)
226 OBD_FAIL_CHECK_RESET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3,
227 CFS_FAIL_PTLRPC_OST_BULK_CB2);
229 /* Network is about to get at the memory */
230 if (ptlrpc_is_bulk_put_source(desc->bd_type))
231 rc = LNetPut(self_nid, desc->bd_mds[posted_md],
232 LNET_ACK_REQ, peer_id,
233 desc->bd_portal, mbits, 0, 0);
235 rc = LNetGet(self_nid, desc->bd_mds[posted_md],
236 peer_id, desc->bd_portal, mbits, 0, false);
240 CERROR("%s: failed bulk transfer with %s:%u x%llu: "
241 "rc = %d\n", exp->exp_obd->obd_name,
242 libcfs_id2str(peer_id), desc->bd_portal,
249 /* Can't send, so we unlink the MD bound above. The UNLINK
250 * event this creates will signal completion with failure,
251 * so we return SUCCESS here! */
252 spin_lock(&desc->bd_lock);
253 desc->bd_refs -= total_md - posted_md;
254 spin_unlock(&desc->bd_lock);
255 LASSERT(desc->bd_refs >= 0);
257 mdunlink_iterate_helper(desc->bd_mds, posted_md);
261 CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
262 "id %s mbits %#llx-%#llx\n", desc->bd_iov_count,
263 desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id),
264 mbits - posted_md, mbits - 1);
270 * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
271 * serialises with completion callback)
273 void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
275 LASSERT(!in_interrupt()); /* might sleep */
277 if (!ptlrpc_server_bulk_active(desc)) /* completed or */
278 return; /* never started */
280 /* We used to poison the pages with 0xab here because we did not want to
281 * send any meaningful data over the wire for evicted clients (bug 9297)
282 * However, this is no longer safe now that we use the page cache on the
285 /* The unlink ensures the callback happens ASAP and is the last
286 * one. If it fails, it must be because completion just happened,
287 * but we must still wait_event_idle_timeout() in this case, to give
288 * us a chance to run server_bulk_callback()
290 __mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw, true);
293 /* Network access will complete in finite time but the HUGE
294 * timeout lets us CWARN for visibility of sluggish NALs */
295 int seconds = PTLRPC_REQ_LONG_UNLINK;
297 while (seconds > 0 &&
298 wait_event_idle_timeout(desc->bd_waitq,
299 !ptlrpc_server_bulk_active(desc),
300 cfs_time_seconds(1)) == 0)
305 CWARN("Unexpectedly long timeout: desc %p\n", desc);
308 #endif /* HAVE_SERVER_SUPPORT */
311 * Register bulk at the sender for later transfer.
312 * Returns 0 on success or error code.
314 int ptlrpc_register_bulk(struct ptlrpc_request *req)
316 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
317 struct lnet_process_id peer;
326 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
329 /* NB no locking required until desc is on the network */
330 LASSERT(desc->bd_nob > 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_md_count;
353 /* rq_mbits is matchbits of the final bulk */
354 mbits = req->rq_mbits - desc->bd_md_count + 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_refs = total_md;
368 md.user_ptr = &desc->bd_cbid;
369 md.handler = ptlrpc_handler;
370 md.threshold = 1; /* PUT or GET */
372 for (posted_md = 0; posted_md < desc->bd_md_count;
373 posted_md++, mbits++) {
374 md.options = PTLRPC_MD_OPTIONS |
375 (ptlrpc_is_bulk_op_get(desc->bd_type) ?
376 LNET_MD_OP_GET : LNET_MD_OP_PUT);
377 ptlrpc_fill_bulk_md(&md, desc, posted_md);
379 if (posted_md > 0 && posted_md + 1 == desc->bd_md_count &&
380 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) {
383 me = LNetMEAttach(desc->bd_portal, peer, mbits, 0,
384 LNET_UNLINK, LNET_INS_AFTER);
385 rc = PTR_ERR_OR_ZERO(me);
388 CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n",
389 desc->bd_import->imp_obd->obd_name, mbits,
393 percpu_ref_get(&ptlrpc_pending);
395 /* About to let the network at it... */
396 rc = LNetMDAttach(me, &md, LNET_UNLINK,
397 &desc->bd_mds[posted_md]);
399 CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n",
400 desc->bd_import->imp_obd->obd_name, mbits,
407 LASSERT(rc == -ENOMEM);
408 spin_lock(&desc->bd_lock);
409 desc->bd_refs -= total_md - posted_md;
410 spin_unlock(&desc->bd_lock);
411 LASSERT(desc->bd_refs >= 0);
412 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
413 req->rq_status = -ENOMEM;
414 desc->bd_registered = 0;
418 spin_lock(&desc->bd_lock);
419 /* Holler if peer manages to touch buffers before he knows the mbits */
420 if (desc->bd_refs != total_md)
421 CWARN("%s: Peer %s touched %d buffers while I registered\n",
422 desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
423 total_md - desc->bd_refs);
424 spin_unlock(&desc->bd_lock);
427 "Setup %u bulk %s buffers: %u pages %u bytes, mbits x%#llx-%#llx, portal %u\n",
429 ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink",
430 desc->bd_iov_count, desc->bd_nob,
431 desc->bd_last_mbits, req->rq_mbits, desc->bd_portal);
437 * Disconnect a bulk desc from the network. Idempotent. Not
438 * thread-safe (i.e. only interlocks with completion callback).
439 * Returns 1 on success or 0 if network unregistration failed for whatever
442 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
444 struct ptlrpc_bulk_desc *desc = req->rq_bulk;
447 LASSERT(!in_interrupt()); /* might sleep */
450 desc->bd_registered = 0;
452 /* Let's setup deadline for reply unlink. */
453 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
454 async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
455 req->rq_bulk_deadline = ktime_get_real_seconds() +
456 PTLRPC_REQ_LONG_UNLINK;
458 if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
459 RETURN(1); /* never registered */
461 LASSERT(desc->bd_req == req); /* bd_req NULL until registered */
463 /* the unlink ensures the callback happens ASAP and is the last
464 * one. If it fails, it must be because completion just happened,
465 * but we must still wait_event_idle_timeout() in this case to give
466 * us a chance to run client_bulk_callback()
468 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
470 if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
471 RETURN(1); /* never registered */
473 /* Move to "Unregistering" phase as bulk was not unlinked yet. */
474 ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK);
476 /* Do not wait for unlink to finish. */
481 /* The wq argument is ignored by user-space wait_event macros */
482 wait_queue_head_t *wq = (req->rq_set != NULL) ?
483 &req->rq_set->set_waitq :
484 &req->rq_reply_waitq;
486 * Network access will complete in finite time but the HUGE
487 * timeout lets us CWARN for visibility of sluggish NALs.
489 int seconds = PTLRPC_REQ_LONG_UNLINK;
491 while (seconds > 0 &&
492 wait_event_idle_timeout(*wq,
493 !ptlrpc_client_bulk_active(req),
494 cfs_time_seconds(1)) == 0)
497 ptlrpc_rqphase_move(req, req->rq_next_phase);
501 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
507 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
509 struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
510 struct ptlrpc_service *svc = svcpt->scp_service;
511 timeout_t service_timeout;
513 service_timeout = clamp_t(timeout_t, ktime_get_real_seconds() -
514 req->rq_arrival_time.tv_sec, 1,
515 (AT_OFF ? obd_timeout * 3 / 2 : at_max));
516 if (!(flags & PTLRPC_REPLY_EARLY) &&
517 (req->rq_type != PTL_RPC_MSG_ERR) &&
518 (req->rq_reqmsg != NULL) &&
519 !(lustre_msg_get_flags(req->rq_reqmsg) &
520 (MSG_RESENT | MSG_REPLAY |
521 MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
522 /* early replies, errors and recovery requests don't count
523 * toward our service time estimate
525 timeout_t oldse = at_measured(&svcpt->scp_at_estimate,
529 DEBUG_REQ(D_ADAPTTO, req,
530 "svc %s changed estimate from %d to %d",
531 svc->srv_name, oldse,
532 at_get(&svcpt->scp_at_estimate));
535 /* Report actual service time for client latency calc */
536 lustre_msg_set_service_timeout(req->rq_repmsg, service_timeout);
537 /* Report service time estimate for future client reqs, but report 0
538 * (to be ignored by client) if it's an error reply during recovery.
541 if (req->rq_type == PTL_RPC_MSG_ERR &&
542 (req->rq_export == NULL ||
543 req->rq_export->exp_obd->obd_recovering)) {
544 lustre_msg_set_timeout(req->rq_repmsg, 0);
548 if (req->rq_export && req->rq_reqmsg != NULL &&
549 (flags & PTLRPC_REPLY_EARLY) &&
550 lustre_msg_get_flags(req->rq_reqmsg) &
551 (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) {
552 struct obd_device *exp_obd = req->rq_export->exp_obd;
554 timeout = ktime_get_real_seconds() -
555 req->rq_arrival_time.tv_sec +
556 min_t(timeout_t, at_extra,
557 exp_obd->obd_recovery_timeout / 4);
559 timeout = at_get(&svcpt->scp_at_estimate);
561 lustre_msg_set_timeout(req->rq_repmsg, timeout);
564 if (req->rq_reqmsg &&
565 !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
566 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
567 "req_flags=%#x magic=%x/%x len=%d\n",
568 flags, lustre_msg_get_flags(req->rq_reqmsg),
569 lustre_msg_get_magic(req->rq_reqmsg),
570 lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
575 * Send request reply from request \a req reply buffer.
576 * \a flags defines reply types
577 * Returns 0 on success or error code
579 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
581 struct ptlrpc_reply_state *rs = req->rq_reply_state;
582 struct ptlrpc_connection *conn;
585 /* We must already have a reply buffer (only ptlrpc_error() may be
586 * called without one). The reply generated by sptlrpc layer (e.g.
587 * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
588 * have a request buffer which is either the actual (swabbed) incoming
589 * request, or a saved copy if this is a req saved in
590 * target_queue_final_reply().
592 LASSERT (req->rq_no_reply == 0);
593 LASSERT (req->rq_reqbuf != NULL);
594 LASSERT (rs != NULL);
595 LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
596 LASSERT (req->rq_repmsg != NULL);
597 LASSERT (req->rq_repmsg == rs->rs_msg);
598 LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
599 LASSERT (rs->rs_cb_id.cbid_arg == rs);
601 /* There may be no rq_export during failover */
603 if (unlikely(req->rq_export && req->rq_export->exp_obd &&
604 req->rq_export->exp_obd->obd_fail)) {
605 /* Failed obd's only send ENODEV */
606 req->rq_type = PTL_RPC_MSG_ERR;
607 req->rq_status = -ENODEV;
608 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
609 req->rq_export->exp_obd->obd_minor);
612 if (req->rq_type != PTL_RPC_MSG_ERR)
613 req->rq_type = PTL_RPC_MSG_REPLY;
615 lustre_msg_set_type(req->rq_repmsg, req->rq_type);
616 lustre_msg_set_status(req->rq_repmsg,
617 ptlrpc_status_hton(req->rq_status));
618 lustre_msg_set_opc(req->rq_repmsg,
619 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
621 target_pack_pool_reply(req);
623 ptlrpc_at_set_reply(req, flags);
625 if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
626 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
628 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
630 if (unlikely(conn == NULL)) {
631 CERROR("not replying on NULL connection\n"); /* bug 9635 */
634 ptlrpc_rs_addref(rs); /* +1 ref for the network */
636 rc = sptlrpc_svc_wrap_reply(req);
640 req->rq_sent = ktime_get_real_seconds();
642 rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
643 (rs->rs_difficult && !rs->rs_no_ack) ?
644 LNET_ACK_REQ : LNET_NOACK_REQ,
645 &rs->rs_cb_id, req->rq_self, req->rq_source,
646 ptlrpc_req2svc(req)->srv_rep_portal,
647 req->rq_rep_mbits ? req->rq_rep_mbits : req->rq_xid,
648 req->rq_reply_off, NULL);
650 if (unlikely(rc != 0))
651 ptlrpc_req_drop_rs(req);
652 ptlrpc_connection_put(conn);
656 int ptlrpc_reply (struct ptlrpc_request *req)
658 if (req->rq_no_reply)
661 return (ptlrpc_send_reply(req, 0));
665 * For request \a req send an error reply back. Create empty
666 * reply buffers if necessary.
668 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
673 if (req->rq_no_reply)
676 if (!req->rq_repmsg) {
677 rc = lustre_pack_reply(req, 1, NULL, NULL);
682 if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
683 req->rq_status != -EPERM && req->rq_status != -ENOENT &&
684 req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
685 req->rq_type = PTL_RPC_MSG_ERR;
687 rc = ptlrpc_send_reply(req, may_be_difficult);
691 int ptlrpc_error(struct ptlrpc_request *req)
693 return ptlrpc_send_error(req, 0);
697 * Send request \a request.
698 * if \a noreply is set, don't expect any reply back and don't set up
700 * Returns 0 on success or error code.
702 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
707 bool rep_mbits = false;
708 struct lnet_handle_md bulk_cookie;
709 struct ptlrpc_connection *connection;
710 struct lnet_me *reply_me = NULL;
711 struct lnet_md reply_md;
712 struct obd_import *imp = request->rq_import;
713 struct obd_device *obd = imp->imp_obd;
716 LNetInvalidateMDHandle(&bulk_cookie);
718 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
721 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DELAY_RECOV) &&
722 lustre_msg_get_opc(request->rq_reqmsg) == MDS_CONNECT &&
723 strcmp(obd->obd_type->typ_name, LUSTRE_OSP_NAME) == 0)) {
727 LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
728 LASSERT(request->rq_wait_ctx == 0);
730 /* If this is a re-transmit, we're required to have disengaged
731 * cleanly from the previous attempt */
732 LASSERT(!request->rq_receiving_reply);
733 LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
734 (imp->imp_state == LUSTRE_IMP_FULL)));
736 if (unlikely(obd != NULL && obd->obd_fail)) {
737 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
739 /* this prevents us from waiting in ptlrpc_queue_wait */
740 spin_lock(&request->rq_lock);
742 spin_unlock(&request->rq_lock);
743 request->rq_status = -ENODEV;
747 connection = imp->imp_connection;
749 lustre_msg_set_handle(request->rq_reqmsg,
750 &imp->imp_remote_handle);
751 lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
752 lustre_msg_set_conn_cnt(request->rq_reqmsg,
754 lustre_msghdr_set_flags(request->rq_reqmsg,
755 imp->imp_msghdr_flags);
757 /* If it's the first time to resend the request for EINPROGRESS,
758 * we need to allocate a new XID (see after_reply()), it's different
759 * from the resend for reply timeout. */
760 if (request->rq_nr_resend != 0 &&
761 list_empty(&request->rq_unreplied_list)) {
763 /* resend for EINPROGRESS, allocate new xid to avoid reply
765 spin_lock(&imp->imp_lock);
766 ptlrpc_assign_next_xid_nolock(request);
767 min_xid = ptlrpc_known_replied_xid(imp);
768 spin_unlock(&imp->imp_lock);
770 lustre_msg_set_last_xid(request->rq_reqmsg, min_xid);
771 DEBUG_REQ(D_RPCTRACE, request,
772 "Allocating new XID for resend on EINPROGRESS");
775 opc = lustre_msg_get_opc(request->rq_reqmsg);
776 if (opc != OST_CONNECT && opc != MDS_CONNECT &&
777 opc != MGS_CONNECT && OCD_HAS_FLAG(&imp->imp_connect_data, FLAGS2))
778 rep_mbits = imp->imp_connect_data.ocd_connect_flags2 &
779 OBD_CONNECT2_REP_MBITS;
781 if ((request->rq_bulk != NULL) || rep_mbits) {
782 ptlrpc_set_mbits(request);
783 lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits);
786 if (list_empty(&request->rq_unreplied_list) ||
787 request->rq_xid <= imp->imp_known_replied_xid) {
788 DEBUG_REQ(D_ERROR, request,
789 "xid=%llu, replied=%llu, list_empty=%d",
790 request->rq_xid, imp->imp_known_replied_xid,
791 list_empty(&request->rq_unreplied_list));
795 /** For enabled AT all request should have AT_SUPPORT in the
796 * FULL import state when OBD_CONNECT_AT is set */
797 LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
798 (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
799 !(imp->imp_connect_data.ocd_connect_flags &
802 if (request->rq_resend) {
803 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
804 if (request->rq_resend_cb != NULL)
805 request->rq_resend_cb(request, &request->rq_async_args);
807 if (request->rq_memalloc)
808 mpflag = memalloc_noreclaim_save();
810 rc = sptlrpc_cli_wrap_request(request);
814 /* bulk register should be done after wrap_request() */
815 if (request->rq_bulk != NULL) {
816 rc = ptlrpc_register_bulk (request);
818 GOTO(cleanup_bulk, rc);
820 * All the mds in the request will have the same cpt
821 * encoded in the cookie. So we can just get the first
824 bulk_cookie = request->rq_bulk->bd_mds[0];
828 LASSERT (request->rq_replen != 0);
829 if (request->rq_repbuf == NULL) {
830 LASSERT(request->rq_repdata == NULL);
831 LASSERT(request->rq_repmsg == NULL);
832 rc = sptlrpc_cli_alloc_repbuf(request,
835 /* this prevents us from looping in
836 * ptlrpc_queue_wait */
837 spin_lock(&request->rq_lock);
839 spin_unlock(&request->rq_lock);
840 request->rq_status = rc;
841 GOTO(cleanup_bulk, rc);
844 request->rq_repdata = NULL;
845 request->rq_repmsg = NULL;
848 if (request->rq_bulk &&
849 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH)) {
850 reply_me = ERR_PTR(-ENOMEM);
852 reply_me = LNetMEAttach(request->rq_reply_portal,
854 rep_mbits ? request->rq_mbits :
856 0, LNET_UNLINK, LNET_INS_AFTER);
859 if (IS_ERR(reply_me)) {
860 rc = PTR_ERR(reply_me);
861 CERROR("LNetMEAttach failed: %d\n", rc);
862 LASSERT(rc == -ENOMEM);
863 GOTO(cleanup_bulk, rc = -ENOMEM);
867 spin_lock(&request->rq_lock);
868 /* We are responsible for unlinking the reply buffer */
869 request->rq_reply_unlinked = noreply;
870 request->rq_receiving_reply = !noreply;
871 /* Clear any flags that may be present from previous sends. */
872 request->rq_req_unlinked = 0;
873 request->rq_replied = 0;
875 request->rq_timedout = 0;
876 request->rq_net_err = 0;
877 request->rq_resend = 0;
878 request->rq_restart = 0;
879 request->rq_reply_truncated = 0;
880 spin_unlock(&request->rq_lock);
883 reply_md.start = request->rq_repbuf;
884 reply_md.length = request->rq_repbuf_len;
885 /* Allow multiple early replies */
886 reply_md.threshold = LNET_MD_THRESH_INF;
887 /* Manage remote for early replies */
888 reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
889 LNET_MD_MANAGE_REMOTE |
890 LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
891 reply_md.user_ptr = &request->rq_reply_cbid;
892 reply_md.handler = ptlrpc_handler;
894 /* We must see the unlink callback to set rq_reply_unlinked,
895 * so we can't auto-unlink */
896 rc = LNetMDAttach(reply_me, &reply_md, LNET_RETAIN,
897 &request->rq_reply_md_h);
899 CERROR("LNetMDAttach failed: %d\n", rc);
900 LASSERT(rc == -ENOMEM);
901 spin_lock(&request->rq_lock);
902 /* ...but the MD attach didn't succeed... */
903 request->rq_receiving_reply = 0;
904 spin_unlock(&request->rq_lock);
905 GOTO(cleanup_bulk, rc = -ENOMEM);
907 percpu_ref_get(&ptlrpc_pending);
910 "Setup reply buffer: %u bytes, xid %llu, portal %u\n",
911 request->rq_repbuf_len, request->rq_xid,
912 request->rq_reply_portal);
915 /* add references on request for request_out_callback */
916 ptlrpc_request_addref(request);
917 if (obd != NULL && obd->obd_svc_stats != NULL)
918 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
919 atomic_read(&imp->imp_inflight));
921 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
923 request->rq_sent_ns = ktime_get_real();
924 request->rq_sent = ktime_get_real_seconds();
925 /* We give the server rq_timeout secs to process the req, and
926 * add the network latency for our local timeout.
928 request->rq_deadline = request->rq_sent + request->rq_timeout +
929 ptlrpc_at_get_net_latency(request);
931 DEBUG_REQ(D_INFO, request, "send flags=%x",
932 lustre_msg_get_flags(request->rq_reqmsg));
933 rc = ptl_send_buf(&request->rq_req_md_h,
934 request->rq_reqbuf, request->rq_reqdata_len,
935 LNET_NOACK_REQ, &request->rq_req_cbid,
936 LNET_NID_ANY, connection->c_peer,
937 request->rq_request_portal,
938 request->rq_xid, 0, &bulk_cookie);
942 request->rq_req_unlinked = 1;
943 ptlrpc_req_finished(request);
947 LNetMDUnlink(request->rq_reply_md_h);
949 /* UNLINKED callback called synchronously */
950 LASSERT(!request->rq_receiving_reply);
953 /* We do sync unlink here as there was no real transfer here so
954 * the chance to have long unlink to sluggish net is smaller here. */
955 ptlrpc_unregister_bulk(request, 0);
958 /* set rq_sent so that this request is treated
959 * as a delayed send in the upper layers */
960 request->rq_sent = ktime_get_real_seconds();
963 if (request->rq_memalloc)
964 memalloc_noreclaim_restore(mpflag);
968 EXPORT_SYMBOL(ptl_send_rpc);
971 * Register request buffer descriptor for request receiving.
973 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
975 struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
976 static struct lnet_process_id match_id = {
984 CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
985 service->srv_req_portal);
987 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
990 /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
991 * which means buffer can only be attached on local CPT, and LND
992 * threads can find it by grabbing a local lock */
993 me = LNetMEAttach(service->srv_req_portal,
994 match_id, 0, ~0, LNET_UNLINK,
995 rqbd->rqbd_svcpt->scp_cpt >= 0 ?
996 LNET_INS_LOCAL : LNET_INS_AFTER);
998 CERROR("LNetMEAttach failed: %ld\n", PTR_ERR(me));
1002 LASSERT(rqbd->rqbd_refcount == 0);
1003 rqbd->rqbd_refcount = 1;
1005 md.start = rqbd->rqbd_buffer;
1006 md.length = service->srv_buf_size;
1007 md.max_size = service->srv_max_req_size;
1008 md.threshold = LNET_MD_THRESH_INF;
1009 md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
1010 md.user_ptr = &rqbd->rqbd_cbid;
1011 md.handler = ptlrpc_handler;
1013 rc = LNetMDAttach(me, &md, LNET_UNLINK, &rqbd->rqbd_md_h);
1015 percpu_ref_get(&ptlrpc_pending);
1019 CERROR("ptlrpc: LNetMDAttach failed: rc = %d\n", rc);
1020 LASSERT(rc == -ENOMEM);
1022 rqbd->rqbd_refcount = 0;