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, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
24 * Copyright (c) 2013, 2017, Intel Corporation.
27 * lustre/target/tgt_handler.c
29 * Lustre Unified Target request handler code
31 * Author: Brian Behlendorf <behlendorf1@llnl.gov>
32 * Author: Mikhail Pershin <mike.pershin@intel.com>
35 #define DEBUG_SUBSYSTEM S_CLASS
37 #include <linux/user_namespace.h>
38 #ifdef HAVE_UIDGID_HEADER
39 # include <linux/uidgid.h>
43 #include <obd_class.h>
44 #include <obd_cksum.h>
45 #include <lustre_lfsck.h>
46 #include <lustre_nodemap.h>
47 #include <lustre_acl.h>
49 #include "tgt_internal.h"
51 char *tgt_name(struct lu_target *tgt)
53 LASSERT(tgt->lut_obd != NULL);
54 return tgt->lut_obd->obd_name;
56 EXPORT_SYMBOL(tgt_name);
59 * Generic code handling requests that have struct mdt_body passed in:
61 * - extract mdt_body from request and save it in @tsi, if present;
63 * - create lu_object, corresponding to the fid in mdt_body, and save it in
66 * - if HABEO_CORPUS flag is set for this request type check whether object
67 * actually exists on storage (lu_object_exists()).
70 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
72 const struct mdt_body *body;
73 struct lu_object *obj;
74 struct req_capsule *pill = tsi->tsi_pill;
79 body = req_capsule_client_get(pill, &RMF_MDT_BODY);
83 tsi->tsi_mdt_body = body;
85 if (!(body->mbo_valid & OBD_MD_FLID))
88 /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
89 * in any case in pre-2.5 clients. Fix that here if needed */
90 if (unlikely(fid_is_zero(&body->mbo_fid1)))
93 if (!fid_is_sane(&body->mbo_fid1)) {
94 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
95 PFID(&body->mbo_fid1));
99 obj = lu_object_find(tsi->tsi_env,
100 &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
101 &body->mbo_fid1, NULL);
103 if ((flags & HABEO_CORPUS) && !lu_object_exists(obj)) {
104 lu_object_put(tsi->tsi_env, obj);
107 tsi->tsi_corpus = obj;
114 tsi->tsi_fid = body->mbo_fid1;
120 * Validate oa from client.
121 * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
123 * a. objects in Single MDT FS seq = FID_SEQ_OST_MDT0, oi_id != 0
124 * b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
125 * pack ost_id. Because non-zero oi_seq will make it diffcult to tell
126 * whether this is oi_fid or real ostid. So it will check
127 * OBD_CONNECT_FID, then convert the ostid to FID for old client.
128 * c. Old FID-disable osc will send IDIF.
129 * d. new FID-enable osc/osp will send normal FID.
131 * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
132 * be used for LAST_ID file, and only being accessed inside OST now.
134 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
136 struct ost_id *oi = &oa->o_oi;
137 u64 seq = ostid_seq(oi);
138 u64 id = ostid_id(oi);
142 if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
143 fid_seq_is_echo(seq))) {
144 /* Sigh 2.[123] client still sends echo req with oi_id = 0
145 * during create, and we will reset this to 1, since this
146 * oi_id is basically useless in the following create process,
147 * but oi_id == 0 will make it difficult to tell whether it is
148 * real FID or ost_id. */
149 oi->oi_fid.f_seq = FID_SEQ_ECHO;
150 oi->oi_fid.f_oid = id ?: 1;
151 oi->oi_fid.f_ver = 0;
153 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
155 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
156 GOTO(out, rc = -EPROTO);
158 /* Note: this check might be forced in 2.5 or 2.6, i.e.
159 * all of the requests are required to setup FLGROUP */
160 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
161 ostid_set_seq_mdt0(oi);
162 oa->o_valid |= OBD_MD_FLGROUP;
166 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
167 fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
168 GOTO(out, rc = -EPROTO);
170 rc = ostid_to_fid(&tti->tti_fid1, oi,
171 tsi->tsi_tgt->lut_lsd.lsd_osd_index);
172 if (unlikely(rc != 0))
175 oi->oi_fid = tti->tti_fid1;
181 CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
182 tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
186 EXPORT_SYMBOL(tgt_validate_obdo);
188 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
191 struct niobuf_remote *rnb;
192 struct obd_ioobj *ioo;
197 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
201 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
205 max_brw = ioobj_max_brw_get(ioo);
206 if (unlikely((max_brw & (max_brw - 1)) != 0)) {
207 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
208 ": rc = %d\n", tgt_name(tsi->tsi_tgt),
209 obd_export_nid2str(tsi->tsi_exp), max_brw,
210 POSTID(oi), -EPROTO);
215 obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
216 RCL_CLIENT) / sizeof(*ioo);
217 if (obj_count == 0) {
218 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
220 } else if (obj_count > 1) {
221 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
226 if (ioo->ioo_bufcnt == 0) {
227 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
231 if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
232 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
233 "bulk has too many pages (%d)",
241 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
243 struct ost_body *body;
244 struct req_capsule *pill = tsi->tsi_pill;
245 struct lu_nodemap *nodemap;
250 body = req_capsule_client_get(pill, &RMF_OST_BODY);
254 rc = tgt_validate_obdo(tsi, &body->oa);
258 nodemap = nodemap_get_from_exp(tsi->tsi_exp);
260 RETURN(PTR_ERR(nodemap));
262 body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
263 NODEMAP_CLIENT_TO_FS,
265 body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
266 NODEMAP_CLIENT_TO_FS,
268 nodemap_putref(nodemap);
270 tsi->tsi_ost_body = body;
271 tsi->tsi_fid = body->oa.o_oi.oi_fid;
273 if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
274 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
279 if (!(body->oa.o_valid & OBD_MD_FLID)) {
280 if (flags & HABEO_CORPUS) {
281 CERROR("%s: OBD_MD_FLID flag is not set in ost_body "
282 "but OID/FID is mandatory with HABEO_CORPUS\n",
283 tgt_name(tsi->tsi_tgt));
290 ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
293 * OST doesn't get object in advance for further use to prevent
294 * situations with nested object_find which is potential deadlock.
296 tsi->tsi_corpus = NULL;
301 * Do necessary preprocessing according to handler ->th_flags.
303 static int tgt_request_preprocess(struct tgt_session_info *tsi,
304 struct tgt_handler *h,
305 struct ptlrpc_request *req)
307 struct req_capsule *pill = tsi->tsi_pill;
308 __u32 flags = h->th_flags;
313 if (tsi->tsi_preprocessed)
316 LASSERT(h->th_act != NULL);
317 LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
318 LASSERT(current->journal_info == NULL);
320 LASSERT(ergo(flags & (HABEO_CORPUS | HABEO_REFERO),
322 if (h->th_fmt != NULL) {
323 req_capsule_set(pill, h->th_fmt);
324 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
325 rc = tgt_mdt_body_unpack(tsi, flags);
328 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
330 rc = tgt_ost_body_unpack(tsi, flags);
336 if (flags & MUTABOR && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
339 if (flags & HABEO_CLAVIS) {
340 struct ldlm_request *dlm_req;
342 LASSERT(h->th_fmt != NULL);
344 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
345 if (dlm_req != NULL) {
346 union ldlm_wire_policy_data *policy =
347 &dlm_req->lock_desc.l_policy_data;
349 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
351 (policy->l_inodebits.bits |
352 policy->l_inodebits.try_bits) == 0)) {
354 * Lock without inodebits makes no sense and
355 * will oops later in ldlm. If client miss to
356 * set such bits, do not trigger ASSERTION.
358 * For liblustre flock case, it maybe zero.
362 tsi->tsi_dlm_req = dlm_req;
368 tsi->tsi_preprocessed = 1;
373 * Invoke handler for this request opc. Also do necessary preprocessing
374 * (according to handler ->th_flags), and post-processing (setting of
375 * ->last_{xid,committed}).
377 static int tgt_handle_request0(struct tgt_session_info *tsi,
378 struct tgt_handler *h,
379 struct ptlrpc_request *req)
383 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
388 /* When dealing with sec context requests, no export is associated yet,
389 * because these requests are sent before *_CONNECT requests.
390 * A NULL req->rq_export means the normal *_common_slice handlers will
391 * not be called, because there is no reference to the target.
392 * So deal with them by hand and jump directly to target_send_reply().
396 case SEC_CTX_INIT_CONT:
398 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
403 * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
404 * to put same checks into handlers like mdt_close(), mdt_reint(),
405 * etc., without talking to mdt authors first. Checking same thing
406 * there again is useless and returning 0 error without packing reply
407 * is buggy! Handlers either pack reply or return error.
409 * We return 0 here and do not send any reply in order to emulate
410 * network failure. Do not send any reply in case any of NET related
411 * fail_id has occured.
413 if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
415 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
416 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
419 rc = tgt_request_preprocess(tsi, h, req);
420 /* pack reply if reply format is fixed */
421 if (rc == 0 && h->th_flags & HABEO_REFERO) {
423 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
425 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
427 tsi->tsi_mdt_body->mbo_eadatasize);
428 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
430 req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
432 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
433 req_capsule_set_size(tsi->tsi_pill,
434 &RMF_ACL, RCL_SERVER,
435 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
437 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
439 struct niobuf_remote *remote_nb =
440 req_capsule_client_get(tsi->tsi_pill,
442 struct ost_body *body = tsi->tsi_ost_body;
444 req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
446 (body->oa.o_flags & OBD_FL_SHORT_IO) ?
447 remote_nb[0].rnb_len : 0);
450 rc = req_capsule_server_pack(tsi->tsi_pill);
453 if (likely(rc == 0)) {
455 * Process request, there can be two types of rc:
456 * 1) errors with msg unpack/pack, other failures outside the
457 * operation itself. This is counted as serious errors;
458 * 2) errors during fs operation, should be placed in rq_status
462 if (!is_serious(rc) &&
463 !req->rq_no_reply && req->rq_reply_state == NULL) {
464 DEBUG_REQ(D_ERROR, req, "%s \"handler\" %s did not "
465 "pack reply and returned 0 error\n",
466 tgt_name(tsi->tsi_tgt), h->th_name);
469 serious = is_serious(rc);
470 rc = clear_serious(rc);
478 * ELDLM_* codes which > 0 should be in rq_status only as well as
479 * all non-serious errors.
481 if (rc > 0 || !serious)
484 LASSERT(current->journal_info == NULL);
486 if (likely(rc == 0 && req->rq_export))
487 target_committed_to_req(req);
490 target_send_reply(req, rc, tsi->tsi_reply_fail_id);
494 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
495 struct obd_device *obd, int *process)
497 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
504 case MDS_SYNC: /* used in unmounting */
518 case MDS_HSM_PROGRESS:
519 case MDS_HSM_STATE_SET:
520 case MDS_HSM_REQUEST:
521 *process = target_queue_recovery_request(req, obd);
525 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
532 * Handle recovery. Return:
533 * +1: continue request processing;
534 * -ve: abort immediately with the given error code;
535 * 0: send reply with error code in req->rq_status;
537 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
541 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
546 case SEC_CTX_INIT_CONT:
551 if (!req->rq_export->exp_obd->obd_replayable)
554 /* sanity check: if the xid matches, the request must be marked as a
555 * resent or replayed */
556 if (req_can_reconstruct(req, NULL)) {
557 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
558 (MSG_RESENT | MSG_REPLAY))) {
559 DEBUG_REQ(D_WARNING, req, "rq_xid %llu matches "
560 "saved xid, expected REPLAY or RESENT flag "
562 lustre_msg_get_flags(req->rq_reqmsg));
563 req->rq_status = -ENOTCONN;
567 /* else: note the opposite is not always true; a RESENT req after a
568 * failover will usually not match the last_xid, since it was likely
569 * never committed. A REPLAYed request will almost never match the
570 * last xid, however it could for a committed, but still retained,
573 /* Check for aborted recovery... */
574 if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
578 DEBUG_REQ(D_INFO, req, "Got new replay");
579 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
581 if (rc != 0 || !should_process)
583 else if (should_process < 0) {
584 req->rq_status = should_process;
585 rc = ptlrpc_error(req);
592 /* Initial check for request, it is validation mostly */
593 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
595 struct tgt_handler *h;
596 struct tgt_opc_slice *s;
597 struct lu_target *tgt;
598 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
602 tgt = class_exp2tgt(req->rq_export);
603 if (unlikely(tgt == NULL)) {
604 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
605 class_exp2obd(req->rq_export)->obd_name);
606 RETURN(ERR_PTR(-EINVAL));
609 for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
610 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
613 /* opcode was not found in slice */
614 if (unlikely(s->tos_hs == NULL)) {
615 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
617 RETURN(ERR_PTR(-ENOTSUPP));
620 LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
621 h = s->tos_hs + (opc - s->tos_opc_start);
622 if (unlikely(h->th_opc == 0)) {
623 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
624 RETURN(ERR_PTR(-ENOTSUPP));
630 static int process_req_last_xid(struct ptlrpc_request *req)
635 /* check request's xid is consistent with export's last_xid */
636 last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
637 if (last_xid > req->rq_export->exp_last_xid)
638 req->rq_export->exp_last_xid = last_xid;
640 if (req->rq_xid == 0 ||
641 (req->rq_xid <= req->rq_export->exp_last_xid)) {
642 DEBUG_REQ(D_ERROR, req, "Unexpected xid %llx vs. "
643 "last_xid %llx\n", req->rq_xid,
644 req->rq_export->exp_last_xid);
645 /* Some request is allowed to be sent during replay,
646 * such as OUT update requests, FLD requests, so it
647 * is possible that replay requests has smaller XID
648 * than the exp_last_xid.
650 * Some non-replay requests may have smaller XID as
653 * - Client send a no_resend RPC, like statfs;
654 * - The RPC timedout (or some other error) on client,
655 * then it's removed from the unreplied list;
656 * - Client send some other request to bump the
657 * exp_last_xid on server;
658 * - The former RPC got chance to be processed;
660 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
664 /* The "last_xid" is the minimum xid among unreplied requests,
665 * if the request is from the previous connection, its xid can
666 * still be larger than "exp_last_xid", then the above check of
667 * xid is not enough to determine whether the request is delayed.
669 * For example, if some replay request was delayed and caused
670 * timeout at client and the replay is restarted, the delayed
671 * replay request will have the larger xid than "exp_last_xid"
673 if (req->rq_export->exp_conn_cnt >
674 lustre_msg_get_conn_cnt(req->rq_reqmsg))
677 /* try to release in-memory reply data */
678 if (tgt_is_multimodrpcs_client(req->rq_export)) {
679 tgt_handle_received_xid(req->rq_export,
680 lustre_msg_get_last_xid(req->rq_reqmsg));
681 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
682 (MSG_RESENT | MSG_REPLAY)))
683 tgt_handle_tag(req->rq_export,
684 lustre_msg_get_tag(req->rq_reqmsg));
689 int tgt_request_handle(struct ptlrpc_request *req)
691 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
693 struct lustre_msg *msg = req->rq_reqmsg;
694 struct tgt_handler *h;
695 struct lu_target *tgt;
696 int request_fail_id = 0;
697 __u32 opc = lustre_msg_get_opc(msg);
698 struct obd_device *obd;
700 bool is_connect = false;
703 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
704 if (cfs_fail_val == 0 &&
705 lustre_msg_get_opc(msg) != OBD_PING &&
706 lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
707 struct l_wait_info lwi = { 0 };
711 l_wait_event(cfs_race_waitq, (cfs_race_state == 1),
716 /* Refill the context, to make sure all thread keys are allocated */
717 lu_env_refill(req->rq_svc_thread->t_env);
719 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
720 tsi->tsi_pill = &req->rq_pill;
721 tsi->tsi_env = req->rq_svc_thread->t_env;
723 /* if request has export then get handlers slice from corresponding
724 * target, otherwise that should be connect operation */
725 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
726 opc == MGS_CONNECT) {
728 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
729 rc = target_handle_connect(req);
731 rc = ptlrpc_error(req);
734 /* recovery-small test 18c asks to drop connect reply */
735 if (unlikely(opc == OST_CONNECT &&
736 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
740 if (unlikely(!class_connected_export(req->rq_export))) {
741 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
742 opc == SEC_CTX_FINI) {
743 /* sec context initialization has to be handled
744 * by hand in tgt_handle_request0() */
745 tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
747 GOTO(handle_recov, rc = 0);
749 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
750 opc, libcfs_id2str(req->rq_peer));
751 req->rq_status = -ENOTCONN;
752 rc = ptlrpc_error(req);
756 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
757 tsi->tsi_exp = req->rq_export;
758 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
759 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
761 tsi->tsi_jobid = NULL;
764 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
765 class_exp2obd(req->rq_export)->obd_name);
766 req->rq_status = -EINVAL;
767 rc = ptlrpc_error(req);
771 /* Skip last_xid processing for the recovery thread, otherwise, the
772 * last_xid on same request could be processed twice: first time when
773 * processing the incoming request, second time when the request is
774 * being processed by recovery thread. */
775 obd = class_exp2obd(req->rq_export);
777 /* reset the exp_last_xid on each connection. */
778 req->rq_export->exp_last_xid = 0;
779 } else if (obd->obd_recovery_data.trd_processing_task !=
781 rc = process_req_last_xid(req);
784 rc = ptlrpc_error(req);
789 request_fail_id = tgt->lut_request_fail_id;
790 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
792 h = tgt_handler_find_check(req);
794 req->rq_status = PTR_ERR(h);
795 rc = ptlrpc_error(req);
799 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
802 if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
803 CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
806 rc = lustre_msg_check_version(msg, h->th_version);
808 DEBUG_REQ(D_ERROR, req, "%s: drop mal-formed request, version"
809 " %08x, expecting %08x\n", tgt_name(tgt),
810 lustre_msg_get_version(msg), h->th_version);
811 req->rq_status = -EINVAL;
812 rc = ptlrpc_error(req);
817 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
818 if (likely(rc == 1)) {
819 rc = tgt_handle_request0(tsi, h, req);
825 req_capsule_fini(tsi->tsi_pill);
826 if (tsi->tsi_corpus != NULL) {
827 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
828 tsi->tsi_corpus = NULL;
832 EXPORT_SYMBOL(tgt_request_handle);
834 /** Assign high priority operations to the request if needed. */
835 int tgt_hpreq_handler(struct ptlrpc_request *req)
837 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
838 struct tgt_handler *h;
843 if (req->rq_export == NULL)
846 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
847 tsi->tsi_pill = &req->rq_pill;
848 tsi->tsi_env = req->rq_svc_thread->t_env;
849 tsi->tsi_tgt = class_exp2tgt(req->rq_export);
850 tsi->tsi_exp = req->rq_export;
852 h = tgt_handler_find_check(req);
858 rc = tgt_request_preprocess(tsi, h, req);
859 if (unlikely(rc != 0))
862 if (h->th_hp != NULL)
866 EXPORT_SYMBOL(tgt_hpreq_handler);
868 void tgt_counter_incr(struct obd_export *exp, int opcode)
870 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
871 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
872 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
874 EXPORT_SYMBOL(tgt_counter_incr);
877 * Unified target generic handlers.
880 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
882 struct lu_target *tgt = class_exp2tgt(exp);
883 struct sptlrpc_flavor flvr;
887 LASSERT(tgt->lut_obd);
888 LASSERT(tgt->lut_slice);
890 /* always allow ECHO client */
891 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
892 LUSTRE_ECHO_NAME) == 0)) {
893 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
897 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
898 read_lock(&tgt->lut_sptlrpc_lock);
899 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
903 read_unlock(&tgt->lut_sptlrpc_lock);
905 spin_lock(&exp->exp_lock);
906 exp->exp_sp_peer = req->rq_sp_from;
907 exp->exp_flvr = flvr;
909 /* when on mgs, if no restriction is set, or if client
910 * is loopback, allow any flavor */
911 if ((strcmp(exp->exp_obd->obd_type->typ_name,
912 LUSTRE_MGS_NAME) == 0) &&
913 (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
914 LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
916 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
918 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
919 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
920 CERROR("%s: unauthorized rpc flavor %x from %s, "
921 "expect %x\n", tgt_name(tgt),
923 libcfs_nid2str(req->rq_peer.nid),
924 exp->exp_flvr.sf_rpc);
927 spin_unlock(&exp->exp_lock);
929 if (exp->exp_sp_peer != req->rq_sp_from) {
930 CERROR("%s: RPC source %s doesn't match %s\n",
932 sptlrpc_part2name(req->rq_sp_from),
933 sptlrpc_part2name(exp->exp_sp_peer));
936 rc = sptlrpc_target_export_check(exp, req);
943 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
945 struct sptlrpc_rule_set tmp_rset;
948 if (unlikely(tgt == NULL)) {
949 CERROR("No target passed");
953 sptlrpc_rule_set_init(&tmp_rset);
954 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
956 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
961 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
963 write_lock(&tgt->lut_sptlrpc_lock);
964 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
965 tgt->lut_sptlrpc_rset = tmp_rset;
966 write_unlock(&tgt->lut_sptlrpc_lock);
970 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
972 int tgt_connect(struct tgt_session_info *tsi)
974 struct ptlrpc_request *req = tgt_ses_req(tsi);
975 struct obd_connect_data *reply;
980 /* XXX: better to call this check right after getting new export but
981 * before last_rcvd slot allocation to avoid server load upon insecure
982 * connects. This is to be fixed after unifiyng all targets.
984 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
988 /* To avoid exposing partially initialized connection flags, changes up
989 * to this point have been staged in reply->ocd_connect_flags. Now that
990 * connection handling has completed successfully, atomically update
991 * the connect flags in the shared export data structure. LU-1623 */
992 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
993 spin_lock(&tsi->tsi_exp->exp_lock);
994 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
995 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
996 *exp_connect_flags2_ptr(tsi->tsi_exp) =
997 reply->ocd_connect_flags2;
998 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
999 spin_unlock(&tsi->tsi_exp->exp_lock);
1003 obd_disconnect(class_export_get(tsi->tsi_exp));
1006 EXPORT_SYMBOL(tgt_connect);
1008 int tgt_disconnect(struct tgt_session_info *tsi)
1014 rc = target_handle_disconnect(tgt_ses_req(tsi));
1016 RETURN(err_serious(rc));
1020 EXPORT_SYMBOL(tgt_disconnect);
1023 * Unified target OBD handlers
1025 int tgt_obd_ping(struct tgt_session_info *tsi)
1031 rc = target_handle_ping(tgt_ses_req(tsi));
1033 RETURN(err_serious(rc));
1037 EXPORT_SYMBOL(tgt_obd_ping);
1039 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1041 return err_serious(-EOPNOTSUPP);
1044 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1046 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1047 struct ptlrpc_request *req = tgt_ses_req(tsi);
1048 struct obd_export *exp = req->rq_export;
1049 struct ptlrpc_bulk_desc *desc;
1050 struct l_wait_info *lwi = &tti->tti_u.update.tti_wait_info;
1056 desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1057 PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1058 MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1062 for (i = 0; i < rdbuf->rb_nbufs; i++)
1063 desc->bd_frag_ops->add_iov_frag(desc,
1064 rdbuf->rb_bufs[i].lb_buf,
1065 rdbuf->rb_bufs[i].lb_len);
1067 rc = target_bulk_io(exp, desc, lwi);
1068 ptlrpc_free_bulk(desc);
1071 EXPORT_SYMBOL(tgt_send_buffer);
1073 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1075 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1076 struct ptlrpc_request *req = tgt_ses_req(tsi);
1077 struct obd_export *exp = req->rq_export;
1078 struct ptlrpc_bulk_desc *desc;
1079 struct l_wait_info *lwi = &tti->tti_u.rdpg.tti_wait_info;
1087 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1088 PTLRPC_BULK_PUT_SOURCE |
1089 PTLRPC_BULK_BUF_KIOV,
1091 &ptlrpc_bulk_kiov_pin_ops);
1095 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1096 /* old client requires reply size in it's PAGE_SIZE,
1097 * which is rdpg->rp_count */
1098 nob = rdpg->rp_count;
1100 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1101 i++, tmpcount -= tmpsize) {
1102 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1103 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1107 LASSERT(desc->bd_nob == nob);
1108 rc = target_bulk_io(exp, desc, lwi);
1109 ptlrpc_free_bulk(desc);
1112 EXPORT_SYMBOL(tgt_sendpage);
1115 * OBD_IDX_READ handler
1117 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1119 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1120 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1121 struct idx_info *req_ii, *rep_ii;
1126 memset(rdpg, 0, sizeof(*rdpg));
1127 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1129 /* extract idx_info buffer from request & reply */
1130 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1131 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1132 RETURN(err_serious(-EPROTO));
1134 rc = req_capsule_server_pack(tsi->tsi_pill);
1136 RETURN(err_serious(rc));
1138 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1140 RETURN(err_serious(-EFAULT));
1141 rep_ii->ii_magic = IDX_INFO_MAGIC;
1143 /* extract hash to start with */
1144 rdpg->rp_hash = req_ii->ii_hash_start;
1146 /* extract requested attributes */
1147 rdpg->rp_attrs = req_ii->ii_attrs;
1149 /* check that fid packed in request is valid and supported */
1150 if (!fid_is_sane(&req_ii->ii_fid))
1152 rep_ii->ii_fid = req_ii->ii_fid;
1155 rep_ii->ii_flags = req_ii->ii_flags;
1157 /* compute number of pages to allocate, ii_count is the number of 4KB
1159 if (req_ii->ii_count <= 0)
1160 GOTO(out, rc = -EFAULT);
1161 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1162 exp_max_brw_size(tsi->tsi_exp));
1163 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1165 /* allocate pages to store the containers */
1166 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1167 if (rdpg->rp_pages == NULL)
1168 GOTO(out, rc = -ENOMEM);
1169 for (i = 0; i < rdpg->rp_npages; i++) {
1170 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1171 if (rdpg->rp_pages[i] == NULL)
1172 GOTO(out, rc = -ENOMEM);
1175 /* populate pages with key/record pairs */
1176 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1180 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1181 "asked %d > %d\n", rc, rdpg->rp_count);
1183 /* send pages to client */
1184 rc = tgt_sendpage(tsi, rdpg, rc);
1189 if (rdpg->rp_pages) {
1190 for (i = 0; i < rdpg->rp_npages; i++)
1191 if (rdpg->rp_pages[i])
1192 __free_page(rdpg->rp_pages[i]);
1193 OBD_FREE(rdpg->rp_pages,
1194 rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1199 struct tgt_handler tgt_obd_handlers[] = {
1200 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1201 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1203 EXPORT_SYMBOL(tgt_obd_handlers);
1205 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1206 struct dt_object *obj, __u64 start, __u64 end)
1212 /* if no objid is specified, it means "sync whole filesystem" */
1214 rc = dt_sync(env, tgt->lut_bottom);
1215 } else if (dt_version_get(env, obj) >
1216 tgt->lut_obd->obd_last_committed) {
1217 rc = dt_object_sync(env, obj, start, end);
1219 atomic_inc(&tgt->lut_sync_count);
1223 EXPORT_SYMBOL(tgt_sync);
1225 * Unified target DLM handlers.
1229 * Unified target BAST
1231 * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1232 * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1233 * but if it's cross-MDT lock, because remote object version is not set, a
1234 * filesystem sync is needed.
1236 * \param lock server side lock
1237 * \param desc lock desc
1238 * \param data ldlm_cb_set_arg
1239 * \param flag indicates whether this cancelling or blocking callback
1240 * \retval 0 on success
1241 * \retval negative number on error
1243 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1244 void *data, int flag)
1247 struct lu_target *tgt;
1248 struct dt_object *obj = NULL;
1254 tgt = class_exp2tgt(lock->l_export);
1256 if (unlikely(tgt == NULL)) {
1257 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1258 class_exp2obd(lock->l_export)->obd_name);
1262 if (flag == LDLM_CB_CANCELING &&
1263 (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1264 (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1265 (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1266 ldlm_is_cbpending(lock))) &&
1267 ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1268 lock->l_resource->lr_type == LDLM_EXTENT)) {
1270 __u64 end = OBD_OBJECT_EOF;
1272 rc = lu_env_init(&env, LCT_DT_THREAD);
1273 if (unlikely(rc != 0))
1276 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1277 tgt->lut_lsd.lsd_osd_index);
1279 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1280 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1282 GOTO(err_env, rc = PTR_ERR(obj));
1284 if (!dt_object_exists(obj))
1285 GOTO(err_put, rc = -ENOENT);
1287 start = lock->l_policy_data.l_extent.start;
1288 end = lock->l_policy_data.l_extent.end;
1291 rc = tgt_sync(&env, tgt, obj, start, end);
1293 CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1294 "cancel: rc = %d\n",
1295 tgt_name(tgt), PFID(&fid),
1296 lock->l_policy_data.l_extent.start,
1297 lock->l_policy_data.l_extent.end, rc);
1301 dt_object_put(&env, obj);
1306 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1310 static struct ldlm_callback_suite tgt_dlm_cbs = {
1311 .lcs_completion = ldlm_server_completion_ast,
1312 .lcs_blocking = tgt_blocking_ast,
1313 .lcs_glimpse = ldlm_server_glimpse_ast
1316 int tgt_enqueue(struct tgt_session_info *tsi)
1318 struct ptlrpc_request *req = tgt_ses_req(tsi);
1323 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1324 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1326 LASSERT(tsi->tsi_dlm_req != NULL);
1327 rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1328 tsi->tsi_dlm_req, &tgt_dlm_cbs);
1330 RETURN(err_serious(rc));
1332 switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1334 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1337 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1340 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1343 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1346 RETURN(req->rq_status);
1348 EXPORT_SYMBOL(tgt_enqueue);
1350 int tgt_convert(struct tgt_session_info *tsi)
1352 struct ptlrpc_request *req = tgt_ses_req(tsi);
1356 LASSERT(tsi->tsi_dlm_req);
1357 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1359 RETURN(err_serious(rc));
1361 RETURN(req->rq_status);
1364 int tgt_bl_callback(struct tgt_session_info *tsi)
1366 return err_serious(-EOPNOTSUPP);
1369 int tgt_cp_callback(struct tgt_session_info *tsi)
1371 return err_serious(-EOPNOTSUPP);
1374 /* generic LDLM target handler */
1375 struct tgt_handler tgt_dlm_handlers[] = {
1376 TGT_DLM_HDL (HABEO_CLAVIS, LDLM_ENQUEUE, tgt_enqueue),
1377 TGT_DLM_HDL (HABEO_CLAVIS, LDLM_CONVERT, tgt_convert),
1378 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1379 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1381 EXPORT_SYMBOL(tgt_dlm_handlers);
1384 * Unified target LLOG handlers.
1386 int tgt_llog_open(struct tgt_session_info *tsi)
1392 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1396 EXPORT_SYMBOL(tgt_llog_open);
1398 int tgt_llog_read_header(struct tgt_session_info *tsi)
1404 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1408 EXPORT_SYMBOL(tgt_llog_read_header);
1410 int tgt_llog_next_block(struct tgt_session_info *tsi)
1416 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1420 EXPORT_SYMBOL(tgt_llog_next_block);
1422 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1428 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1432 EXPORT_SYMBOL(tgt_llog_prev_block);
1434 /* generic llog target handler */
1435 struct tgt_handler tgt_llog_handlers[] = {
1436 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1437 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1438 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1439 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1441 EXPORT_SYMBOL(tgt_llog_handlers);
1444 * sec context handlers
1446 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1447 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1452 struct tgt_handler tgt_sec_ctx_handlers[] = {
1453 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle),
1454 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle),
1455 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle),
1457 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1459 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1460 struct dt_device *key,
1461 struct lfsck_req_local *lrl,
1462 struct thandle *th) = NULL;
1464 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1466 struct lfsck_req_local *,
1469 tgt_lfsck_in_notify_local = notify;
1471 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1473 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1474 struct dt_device *key,
1475 struct lfsck_request *lr) = NULL;
1477 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1479 struct lfsck_request *))
1481 tgt_lfsck_in_notify = notify;
1483 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1485 static int (*tgt_lfsck_query)(const struct lu_env *env,
1486 struct dt_device *key,
1487 struct lfsck_request *req,
1488 struct lfsck_reply *rep,
1489 struct lfsck_query *que) = NULL;
1491 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1493 struct lfsck_request *,
1494 struct lfsck_reply *,
1495 struct lfsck_query *))
1497 tgt_lfsck_query = query;
1499 EXPORT_SYMBOL(tgt_register_lfsck_query);
1501 /* LFSCK request handlers */
1502 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1504 const struct lu_env *env = tsi->tsi_env;
1505 struct dt_device *key = tsi->tsi_tgt->lut_bottom;
1506 struct lfsck_request *lr;
1510 lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1514 rc = tgt_lfsck_in_notify(env, key, lr);
1519 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1521 struct lfsck_request *request;
1522 struct lfsck_reply *reply;
1526 request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1527 if (request == NULL)
1530 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1534 rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1535 request, reply, NULL);
1537 RETURN(rc < 0 ? rc : 0);
1540 struct tgt_handler tgt_lfsck_handlers[] = {
1541 TGT_LFSCK_HDL(HABEO_REFERO, LFSCK_NOTIFY, tgt_handle_lfsck_notify),
1542 TGT_LFSCK_HDL(HABEO_REFERO, LFSCK_QUERY, tgt_handle_lfsck_query),
1544 EXPORT_SYMBOL(tgt_lfsck_handlers);
1547 * initialize per-thread page pool (bug 5137).
1549 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1551 struct tgt_thread_big_cache *tbc;
1555 LASSERT(thread != NULL);
1556 LASSERT(thread->t_data == NULL);
1558 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1561 thread->t_data = tbc;
1564 EXPORT_SYMBOL(tgt_io_thread_init);
1567 * free per-thread pool created by tgt_thread_init().
1569 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1571 struct tgt_thread_big_cache *tbc;
1575 LASSERT(thread != NULL);
1578 * be prepared to handle partially-initialized pools (because this is
1579 * called from ost_io_thread_init() for cleanup.
1581 tbc = thread->t_data;
1583 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1584 thread->t_data = NULL;
1588 EXPORT_SYMBOL(tgt_io_thread_done);
1591 * Helper function for getting Data-on-MDT file server DLM lock
1592 * if asked by client.
1594 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1595 struct lustre_handle *lh, int mode, __u64 *flags)
1597 union ldlm_policy_data policy = {
1598 .l_inodebits.bits = MDS_INODELOCK_DOM,
1604 LASSERT(lh != NULL);
1605 LASSERT(ns != NULL);
1606 LASSERT(!lustre_handle_is_used(lh));
1608 rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1609 flags, ldlm_blocking_ast,
1610 ldlm_completion_ast, ldlm_glimpse_ast,
1611 NULL, 0, LVB_T_NONE, NULL, lh);
1613 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1615 EXPORT_SYMBOL(tgt_mdt_data_lock);
1617 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1619 LASSERT(lustre_handle_is_used(lh));
1620 ldlm_lock_decref(lh, mode);
1622 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1625 * Helper function for getting server side [start, start+count] DLM lock
1626 * if asked by client.
1628 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1629 struct ldlm_res_id *res_id, __u64 start, __u64 end,
1630 struct lustre_handle *lh, int mode, __u64 *flags)
1632 union ldlm_policy_data policy;
1637 LASSERT(lh != NULL);
1638 LASSERT(ns != NULL);
1639 LASSERT(!lustre_handle_is_used(lh));
1641 policy.l_extent.gid = 0;
1642 policy.l_extent.start = start & PAGE_MASK;
1645 * If ->o_blocks is EOF it means "lock till the end of the file".
1646 * Otherwise, it's size of an extent or hole being punched (in bytes).
1648 if (end == OBD_OBJECT_EOF || end < start)
1649 policy.l_extent.end = OBD_OBJECT_EOF;
1651 policy.l_extent.end = end | ~PAGE_MASK;
1653 rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1654 mode, flags, ldlm_blocking_ast,
1655 ldlm_completion_ast, ldlm_glimpse_ast,
1656 NULL, 0, LVB_T_NONE, NULL, lh);
1657 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1659 EXPORT_SYMBOL(tgt_extent_lock);
1661 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1663 LASSERT(lustre_handle_is_used(lh));
1664 ldlm_lock_decref(lh, mode);
1666 EXPORT_SYMBOL(tgt_extent_unlock);
1668 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1669 struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1670 struct niobuf_remote *nb, struct lustre_handle *lh,
1671 enum ldlm_mode mode)
1673 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1675 int nrbufs = obj->ioo_bufcnt;
1681 LASSERT(mode == LCK_PR || mode == LCK_PW);
1682 LASSERT(!lustre_handle_is_used(lh));
1684 if (ns->ns_obd->obd_recovering)
1687 if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1690 for (i = 1; i < nrbufs; i++)
1691 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1694 /* MDT IO for data-on-mdt */
1695 if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1696 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1698 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1699 nb[nrbufs - 1].rnb_offset +
1700 nb[nrbufs - 1].rnb_len - 1,
1705 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1706 struct lustre_handle *lh, enum ldlm_mode mode)
1710 LASSERT(mode == LCK_PR || mode == LCK_PW);
1711 LASSERT((obj->ioo_bufcnt > 0 &&
1712 (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1713 lustre_handle_is_used(lh));
1715 if (lustre_handle_is_used(lh))
1716 tgt_extent_unlock(lh, mode);
1719 static int tgt_checksum_niobuf(struct lu_target *tgt,
1720 struct niobuf_local *local_nb, int npages,
1721 int opc, enum cksum_types cksum_type,
1724 struct ahash_request *req;
1725 unsigned int bufsize;
1727 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1729 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1731 CERROR("%s: unable to initialize checksum hash %s\n",
1732 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1733 return PTR_ERR(req);
1736 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1737 for (i = 0; i < npages; i++) {
1738 /* corrupt the data before we compute the checksum, to
1739 * simulate a client->OST data error */
1740 if (i == 0 && opc == OST_WRITE &&
1741 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1742 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1743 int len = local_nb[i].lnb_len;
1744 struct page *np = tgt_page_to_corrupt;
1747 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1749 char *ptr2 = page_address(np);
1751 memcpy(ptr2 + off, ptr + off, len);
1752 memcpy(ptr2 + off, "bad3", min(4, len));
1753 ll_kunmap_atomic(ptr, KM_USER0);
1755 /* LU-8376 to preserve original index for
1756 * display in dump_all_bulk_pages() */
1759 cfs_crypto_hash_update_page(req, np, off,
1763 CERROR("%s: can't alloc page for corruption\n",
1767 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1768 local_nb[i].lnb_page_offset & ~PAGE_MASK,
1769 local_nb[i].lnb_len);
1771 /* corrupt the data after we compute the checksum, to
1772 * simulate an OST->client data error */
1773 if (i == 0 && opc == OST_READ &&
1774 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1775 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1776 int len = local_nb[i].lnb_len;
1777 struct page *np = tgt_page_to_corrupt;
1780 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1782 char *ptr2 = page_address(np);
1784 memcpy(ptr2 + off, ptr + off, len);
1785 memcpy(ptr2 + off, "bad4", min(4, len));
1786 ll_kunmap_atomic(ptr, KM_USER0);
1788 /* LU-8376 to preserve original index for
1789 * display in dump_all_bulk_pages() */
1792 cfs_crypto_hash_update_page(req, np, off,
1796 CERROR("%s: can't alloc page for corruption\n",
1802 bufsize = sizeof(*cksum);
1803 err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1808 char dbgcksum_file_name[PATH_MAX];
1810 static void dump_all_bulk_pages(struct obdo *oa, int count,
1811 struct niobuf_local *local_nb,
1812 __u32 server_cksum, __u32 client_cksum)
1819 /* will only keep dump of pages on first error for the same range in
1820 * file/fid, not during the resends/retries. */
1821 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1822 "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1823 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1824 libcfs_debug_file_path_arr :
1825 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1826 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1827 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1828 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1829 local_nb[0].lnb_file_offset,
1830 local_nb[count-1].lnb_file_offset +
1831 local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1832 filp = filp_open(dbgcksum_file_name,
1833 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1837 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1838 "checksum error: rc = %d\n", dbgcksum_file_name,
1841 CERROR("%s: can't open to dump pages with checksum "
1842 "error: rc = %d\n", dbgcksum_file_name, rc);
1846 for (i = 0; i < count; i++) {
1847 len = local_nb[i].lnb_len;
1848 buf = kmap(local_nb[i].lnb_page);
1850 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1852 CERROR("%s: wanted to write %u but got %d "
1853 "error\n", dbgcksum_file_name, len, rc);
1858 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1859 dbgcksum_file_name, rc);
1861 kunmap(local_nb[i].lnb_page);
1864 rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1866 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1867 filp_close(filp, NULL);
1871 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1872 struct obd_export *exp, struct obdo *oa,
1873 const struct lnet_process_id *peer,
1874 __u32 client_cksum, __u32 server_cksum,
1875 enum cksum_types server_cksum_type)
1878 enum cksum_types cksum_type;
1881 /* unlikely to happen and only if resend does not occur due to cksum
1882 * control failure on Client */
1883 if (unlikely(server_cksum == client_cksum)) {
1884 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1889 if (exp->exp_obd->obd_checksum_dump)
1890 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1893 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1896 if (cksum_type != server_cksum_type)
1897 msg = "the server may have not used the checksum type specified"
1898 " in the original request - likely a protocol problem";
1900 msg = "should have changed on the client or in transit";
1902 start = local_nb[0].lnb_file_offset;
1903 end = local_nb[npages-1].lnb_file_offset +
1904 local_nb[npages-1].lnb_len - 1;
1906 LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1907 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1908 " %x (type %x), server csum %x (type %x)\n",
1909 exp->exp_obd->obd_name,
1910 msg, libcfs_nid2str(peer->nid),
1911 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1912 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1913 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1915 start, end, client_cksum, cksum_type, server_cksum,
1921 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1922 unsigned char *buf, int size)
1924 int i, off, len, copied = size;
1927 for (i = 0; i < npages; i++) {
1928 off = local[i].lnb_page_offset & ~PAGE_MASK;
1929 len = local[i].lnb_len;
1931 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1936 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
1937 memcpy(buf + off, ptr, len);
1938 ll_kunmap_atomic(ptr, KM_USER0);
1942 return copied - size;
1945 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1946 struct niobuf_local *local_nb,
1947 int npages, int opc,
1948 obd_dif_csum_fn *fn,
1952 enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1953 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1954 const char *obd_name = tgt->lut_obd->obd_name;
1955 struct ahash_request *req;
1956 unsigned int bufsize;
1957 unsigned char *buffer;
1958 struct page *__page;
1961 int used_number = 0;
1967 __page = alloc_page(GFP_KERNEL);
1971 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1973 CERROR("%s: unable to initialize checksum hash %s\n",
1974 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1975 return PTR_ERR(req);
1978 buffer = kmap(__page);
1979 guard_start = (__u16 *)buffer;
1980 guard_number = PAGE_SIZE / sizeof(*guard_start);
1981 for (i = 0; i < npages; i++) {
1982 /* corrupt the data before we compute the checksum, to
1983 * simulate a client->OST data error */
1984 if (i == 0 && opc == OST_WRITE &&
1985 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1986 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1987 int len = local_nb[i].lnb_len;
1988 struct page *np = tgt_page_to_corrupt;
1991 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1993 char *ptr2 = page_address(np);
1995 memcpy(ptr2 + off, ptr + off, len);
1996 memcpy(ptr2 + off, "bad3", min(4, len));
1997 ll_kunmap_atomic(ptr, KM_USER0);
1999 /* LU-8376 to preserve original index for
2000 * display in dump_all_bulk_pages() */
2003 cfs_crypto_hash_update_page(req, np, off,
2007 CERROR("%s: can't alloc page for corruption\n",
2013 * The left guard number should be able to hold checksums of a
2016 if (t10_cksum_type && opc == OST_READ &&
2017 local_nb[i].lnb_guard_disk) {
2018 used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2019 if (used > (guard_number - used_number)) {
2023 memcpy(guard_start + used_number,
2024 local_nb[i].lnb_guards,
2025 used * sizeof(*local_nb[i].lnb_guards));
2027 rc = obd_page_dif_generate_buffer(obd_name,
2028 local_nb[i].lnb_page,
2029 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2030 local_nb[i].lnb_len, guard_start + used_number,
2031 guard_number - used_number, &used, sector_size,
2037 LASSERT(used <= MAX_GUARD_NUMBER);
2039 * If disk support T10PI checksum, copy guards to local_nb.
2040 * If the write is partial page, do not use the guards for bio
2041 * submission since the data might not be full-sector. The bio
2042 * guards will be generated later based on the full sectors. If
2043 * the sector size is 512B rather than 4 KB, or the page size
2044 * is larger than 4KB, this might drop some useful guards for
2045 * partial page write, but it will only add minimal extra time
2046 * of checksum calculation.
2048 if (t10_cksum_type && opc == OST_WRITE &&
2049 local_nb[i].lnb_len == PAGE_SIZE) {
2050 local_nb[i].lnb_guard_rpc = 1;
2051 memcpy(local_nb[i].lnb_guards,
2052 guard_start + used_number,
2053 used * sizeof(*local_nb[i].lnb_guards));
2056 used_number += used;
2057 if (used_number == guard_number) {
2058 cfs_crypto_hash_update_page(req, __page, 0,
2059 used_number * sizeof(*guard_start));
2063 /* corrupt the data after we compute the checksum, to
2064 * simulate an OST->client data error */
2065 if (unlikely(i == 0 && opc == OST_READ &&
2066 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2067 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2068 int len = local_nb[i].lnb_len;
2069 struct page *np = tgt_page_to_corrupt;
2072 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2074 char *ptr2 = page_address(np);
2076 memcpy(ptr2 + off, ptr + off, len);
2077 memcpy(ptr2 + off, "bad4", min(4, len));
2078 ll_kunmap_atomic(ptr, KM_USER0);
2080 /* LU-8376 to preserve original index for
2081 * display in dump_all_bulk_pages() */
2084 cfs_crypto_hash_update_page(req, np, off,
2088 CERROR("%s: can't alloc page for corruption\n",
2097 if (used_number != 0)
2098 cfs_crypto_hash_update_page(req, __page, 0,
2099 used_number * sizeof(*guard_start));
2101 bufsize = sizeof(cksum);
2102 rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2107 __free_page(__page);
2111 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2112 enum cksum_types cksum_type,
2113 struct niobuf_local *local_nb,
2114 int npages, int opc, u32 *check_sum)
2116 obd_dif_csum_fn *fn = NULL;
2117 int sector_size = 0;
2121 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
2124 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2125 opc, fn, sector_size,
2128 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2129 cksum_type, check_sum);
2133 int tgt_brw_read(struct tgt_session_info *tsi)
2135 struct ptlrpc_request *req = tgt_ses_req(tsi);
2136 struct ptlrpc_bulk_desc *desc = NULL;
2137 struct obd_export *exp = tsi->tsi_exp;
2138 struct niobuf_remote *remote_nb;
2139 struct niobuf_local *local_nb;
2140 struct obd_ioobj *ioo;
2141 struct ost_body *body, *repbody;
2142 struct l_wait_info lwi;
2143 struct lustre_handle lockh = { 0 };
2144 int npages, nob = 0, rc, i, no_reply = 0,
2146 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2147 const char *obd_name = exp->exp_obd->obd_name;
2151 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2152 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2153 CERROR("%s: deny read request from %s to portal %u\n",
2154 tgt_name(tsi->tsi_tgt),
2155 obd_export_nid2str(req->rq_export),
2156 ptlrpc_req2svc(req)->srv_req_portal);
2160 req->rq_bulk_read = 1;
2162 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2165 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2166 cfs_fail_val : (obd_timeout + 1) / 4);
2168 /* Check if there is eviction in progress, and if so, wait for it to
2170 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2171 /* We do not care how long it takes */
2172 lwi = LWI_INTR(NULL, NULL);
2173 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2174 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
2178 /* There must be big cache in current thread to process this request
2179 * if it is NULL then something went wrong and it wasn't allocated,
2180 * report -ENOMEM in that case */
2184 body = tsi->tsi_ost_body;
2185 LASSERT(body != NULL);
2187 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2188 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2190 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2191 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2193 local_nb = tbc->local;
2195 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2201 * If getting the lock took more time than
2202 * client was willing to wait, drop it. b=11330
2204 if (ktime_get_real_seconds() > req->rq_deadline ||
2205 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2207 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2208 libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2209 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2210 req->rq_deadline - req->rq_arrival_time.tv_sec);
2211 GOTO(out_lock, rc = -ETIMEDOUT);
2214 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2215 repbody->oa = body->oa;
2217 npages = PTLRPC_MAX_BRW_PAGES;
2218 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2219 ioo, remote_nb, &npages, local_nb);
2223 if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2226 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2227 PTLRPC_BULK_PUT_SOURCE |
2228 PTLRPC_BULK_BUF_KIOV,
2230 &ptlrpc_bulk_kiov_nopin_ops);
2232 GOTO(out_commitrw, rc = -ENOMEM);
2236 npages_read = npages;
2237 for (i = 0; i < npages; i++) {
2238 int page_rc = local_nb[i].lnb_rc;
2247 if (page_rc != 0 && desc != NULL) { /* some data! */
2248 LASSERT(local_nb[i].lnb_page != NULL);
2249 desc->bd_frag_ops->add_kiov_frag
2250 (desc, local_nb[i].lnb_page,
2251 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2255 if (page_rc != local_nb[i].lnb_len) { /* short read */
2256 local_nb[i].lnb_len = page_rc;
2257 npages_read = i + (page_rc != 0 ? 1 : 0);
2258 /* All subsequent pages should be 0 */
2259 while (++i < npages)
2260 LASSERT(local_nb[i].lnb_rc == 0);
2264 if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
2265 nob != cfs_fail_val)
2268 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2269 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2270 body->oa.o_flags : 0;
2271 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2273 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2275 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2277 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2278 local_nb, npages_read, OST_READ,
2279 &repbody->oa.o_cksum);
2281 GOTO(out_commitrw, rc);
2282 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2283 repbody->oa.o_cksum);
2285 /* if a resend it could be for a cksum error, so check Server
2286 * cksum with returned Client cksum (this should even cover
2287 * zero-cksum case) */
2288 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2289 (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2290 check_read_checksum(local_nb, npages_read, exp,
2291 &body->oa, &req->rq_peer,
2293 repbody->oa.o_cksum, cksum_type);
2295 repbody->oa.o_valid = 0;
2297 /* We're finishing using body->oa as an input variable */
2299 /* Check if client was evicted while we were doing i/o before touching
2302 if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2303 unsigned char *short_io_buf;
2306 short_io_buf = req_capsule_server_get(&req->rq_pill,
2308 short_io_size = req_capsule_get_size(&req->rq_pill,
2311 rc = tgt_pages2shortio(local_nb, npages_read,
2312 short_io_buf, short_io_size);
2314 req_capsule_shrink(&req->rq_pill,
2317 rc = rc > 0 ? 0 : rc;
2318 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2319 rc = target_bulk_io(exp, desc, &lwi);
2323 if (body->oa.o_flags & OBD_FL_SHORT_IO)
2324 req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2329 /* Must commit after prep above in all cases */
2330 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2331 remote_nb, npages, local_nb, rc);
2333 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2335 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2336 ptlrpc_free_bulk(desc);
2341 ptlrpc_lprocfs_brw(req, nob);
2342 } else if (no_reply) {
2343 req->rq_no_reply = 1;
2344 /* reply out callback would free */
2345 ptlrpc_req_drop_rs(req);
2346 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2347 "client will retry: rc %d\n",
2349 obd_uuid2str(&exp->exp_client_uuid),
2350 obd_export_nid2str(exp), rc);
2352 /* send a bulk after reply to simulate a network delay or reordering
2353 * by a router - Note that !desc implies short io, so there is no bulk
2355 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2357 wait_queue_head_t waitq;
2358 struct l_wait_info lwi1;
2360 CDEBUG(D_INFO, "reorder BULK\n");
2361 init_waitqueue_head(&waitq);
2363 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
2364 l_wait_event(waitq, 0, &lwi1);
2365 target_bulk_io(exp, desc, &lwi);
2366 ptlrpc_free_bulk(desc);
2371 EXPORT_SYMBOL(tgt_brw_read);
2373 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2374 unsigned char *buf, int size)
2379 for (i = 0; i < npages; i++) {
2380 off = local[i].lnb_page_offset & ~PAGE_MASK;
2381 len = local[i].lnb_len;
2386 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2388 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
2391 memcpy(ptr + off, buf, len < size ? len : size);
2392 ll_kunmap_atomic(ptr, KM_USER0);
2399 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2400 struct ptlrpc_bulk_desc *desc,
2401 struct niobuf_local *local_nb, int npages,
2402 u32 client_cksum, u32 server_cksum,
2405 struct obd_export *exp = req->rq_export;
2406 struct ost_body *body;
2410 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2411 LASSERT(body != NULL);
2413 if (desc && req->rq_peer.nid != desc->bd_sender) {
2415 router = libcfs_nid2str(desc->bd_sender);
2418 if (exp->exp_obd->obd_checksum_dump)
2419 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2423 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2424 client_cksum, server_cksum);
2428 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2429 DFID" object "DOSTID" extent [%llu-%llu"
2430 "]: client csum %x, server csum %x\n",
2431 exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2433 body->oa.o_valid & OBD_MD_FLFID ?
2434 body->oa.o_parent_seq : (__u64)0,
2435 body->oa.o_valid & OBD_MD_FLFID ?
2436 body->oa.o_parent_oid : 0,
2437 body->oa.o_valid & OBD_MD_FLFID ?
2438 body->oa.o_parent_ver : 0,
2439 POSTID(&body->oa.o_oi),
2440 local_nb[0].lnb_file_offset,
2441 local_nb[npages-1].lnb_file_offset +
2442 local_nb[npages - 1].lnb_len - 1,
2443 client_cksum, server_cksum);
2446 int tgt_brw_write(struct tgt_session_info *tsi)
2448 struct ptlrpc_request *req = tgt_ses_req(tsi);
2449 struct ptlrpc_bulk_desc *desc = NULL;
2450 struct obd_export *exp = req->rq_export;
2451 struct niobuf_remote *remote_nb;
2452 struct niobuf_local *local_nb;
2453 struct obd_ioobj *ioo;
2454 struct ost_body *body, *repbody;
2455 struct l_wait_info lwi;
2456 struct lustre_handle lockh = {0};
2458 int objcount, niocount, npages;
2460 enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2461 bool no_reply = false, mmap;
2462 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2463 bool wait_sync = false;
2464 const char *obd_name = exp->exp_obd->obd_name;
2468 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2469 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2470 CERROR("%s: deny write request from %s to portal %u\n",
2471 tgt_name(tsi->tsi_tgt),
2472 obd_export_nid2str(req->rq_export),
2473 ptlrpc_req2svc(req)->srv_req_portal);
2474 RETURN(err_serious(-EPROTO));
2477 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2478 RETURN(err_serious(-ENOSPC));
2479 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2480 RETURN(err_serious(-EROFS));
2482 req->rq_bulk_write = 1;
2484 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2485 RETURN(err_serious(-EIO));
2486 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2487 RETURN(err_serious(-EFAULT));
2489 /* pause before transaction has been started */
2490 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2491 cfs_fail_val : (obd_timeout + 1) / 4);
2493 /* There must be big cache in current thread to process this request
2494 * if it is NULL then something went wrong and it wasn't allocated,
2495 * report -ENOMEM in that case */
2499 body = tsi->tsi_ost_body;
2500 LASSERT(body != NULL);
2502 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2503 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2505 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2506 RCL_CLIENT) / sizeof(*ioo);
2508 for (niocount = i = 0; i < objcount; i++)
2509 niocount += ioo[i].ioo_bufcnt;
2511 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2512 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2513 if (niocount != req_capsule_get_size(&req->rq_pill,
2514 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2516 RETURN(err_serious(-EPROTO));
2518 if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2519 ptlrpc_connection_is_local(exp->exp_connection))
2520 memory_pressure_set();
2522 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2523 niocount * sizeof(*rcs));
2524 rc = req_capsule_server_pack(&req->rq_pill);
2526 GOTO(out, rc = err_serious(rc));
2528 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2529 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2531 local_nb = tbc->local;
2533 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2539 * If getting the lock took more time than
2540 * client was willing to wait, drop it. b=11330
2542 if (ktime_get_real_seconds() > req->rq_deadline ||
2543 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2545 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2546 tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2547 POSTID(&ioo->ioo_oid),
2548 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2549 req->rq_deadline - req->rq_arrival_time.tv_sec);
2550 GOTO(out_lock, rc = -ETIMEDOUT);
2553 /* Because we already sync grant info with client when reconnect,
2554 * grant info will be cleared for resent req, then fed_grant and
2555 * total_grant will not be modified in following preprw_write */
2556 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2557 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2558 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2561 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2562 if (repbody == NULL)
2563 GOTO(out_lock, rc = -ENOMEM);
2564 repbody->oa = body->oa;
2566 npages = PTLRPC_MAX_BRW_PAGES;
2567 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2568 objcount, ioo, remote_nb, &npages, local_nb);
2571 if (body->oa.o_flags & OBD_FL_SHORT_IO) {
2573 unsigned char *short_io_buf;
2575 short_io_size = req_capsule_get_size(&req->rq_pill,
2578 short_io_buf = req_capsule_client_get(&req->rq_pill,
2580 CDEBUG(D_INFO, "Client use short io for data transfer,"
2581 " size = %d\n", short_io_size);
2583 /* Copy short io buf to pages */
2584 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2588 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2589 PTLRPC_BULK_GET_SINK |
2590 PTLRPC_BULK_BUF_KIOV,
2592 &ptlrpc_bulk_kiov_nopin_ops);
2594 GOTO(skip_transfer, rc = -ENOMEM);
2596 /* NB Having prepped, we must commit... */
2597 for (i = 0; i < npages; i++)
2598 desc->bd_frag_ops->add_kiov_frag(desc,
2599 local_nb[i].lnb_page,
2600 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2601 local_nb[i].lnb_len);
2603 rc = sptlrpc_svc_prep_bulk(req, desc);
2605 GOTO(skip_transfer, rc);
2607 rc = target_bulk_io(exp, desc, &lwi);
2613 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2614 static int cksum_counter;
2616 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2617 cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2619 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2620 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2621 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2624 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2625 local_nb, npages, OST_WRITE,
2626 &repbody->oa.o_cksum);
2628 GOTO(out_commitrw, rc);
2632 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2633 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2634 body->oa.o_flags & OBD_FL_MMAP);
2636 tgt_warn_on_cksum(req, desc, local_nb, npages,
2638 repbody->oa.o_cksum, mmap);
2640 } else if ((cksum_counter & (-cksum_counter)) ==
2642 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2643 cksum_counter, libcfs_id2str(req->rq_peer),
2644 repbody->oa.o_cksum);
2649 /* Must commit after prep above in all cases */
2650 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2651 objcount, ioo, remote_nb, npages, local_nb, rc);
2652 if (rc == -ENOTCONN)
2653 /* quota acquire process has been given up because
2654 * either the client has been evicted or the client
2655 * has timed out the request already */
2658 for (i = 0; i < niocount; i++) {
2659 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2665 * Disable sending mtime back to the client. If the client locked the
2666 * whole object, then it has already updated the mtime on its side,
2667 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2669 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2674 /* set per-requested niobuf return codes */
2675 for (i = j = 0; i < niocount; i++) {
2676 int len = remote_nb[i].rnb_len;
2681 LASSERT(j < npages);
2682 if (local_nb[j].lnb_rc < 0)
2683 rcs[i] = local_nb[j].lnb_rc;
2684 len -= local_nb[j].lnb_len;
2689 LASSERT(j == npages);
2690 ptlrpc_lprocfs_brw(req, nob);
2693 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2695 ptlrpc_free_bulk(desc);
2697 if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2698 req->rq_no_reply = 1;
2699 /* reply out callback would free */
2700 ptlrpc_req_drop_rs(req);
2701 if (!exp->exp_obd->obd_no_transno)
2702 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2703 " client will retry: rc = %d\n",
2705 obd_uuid2str(&exp->exp_client_uuid),
2706 obd_export_nid2str(exp), rc);
2708 memory_pressure_clr();
2711 EXPORT_SYMBOL(tgt_brw_write);
2713 /* Check if request can be reconstructed from saved reply data
2714 * A copy of the reply data is returned in @trd if the pointer is not NULL
2716 bool req_can_reconstruct(struct ptlrpc_request *req,
2717 struct tg_reply_data *trd)
2719 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2720 struct lsd_client_data *lcd = ted->ted_lcd;
2723 if (tgt_is_multimodrpcs_client(req->rq_export))
2724 return tgt_lookup_reply(req, trd);
2726 mutex_lock(&ted->ted_lcd_lock);
2727 found = req->rq_xid == lcd->lcd_last_xid ||
2728 req->rq_xid == lcd->lcd_last_close_xid;
2730 if (found && trd != NULL) {
2731 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2732 trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2733 trd->trd_reply.lrd_transno =
2734 lcd->lcd_last_close_transno;
2735 trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2737 trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2738 trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2739 trd->trd_reply.lrd_result = lcd->lcd_last_result;
2740 trd->trd_reply.lrd_data = lcd->lcd_last_data;
2741 trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2742 trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2743 trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2744 trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2747 mutex_unlock(&ted->ted_lcd_lock);
2751 EXPORT_SYMBOL(req_can_reconstruct);