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 #include <linux/delay.h>
39 #include <linux/uidgid.h>
41 #include <libcfs/linux/linux-mem.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 HAS_BODY 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 & HAS_BODY) && !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 & HAS_BODY) {
281 CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\n",
282 tgt_name(tsi->tsi_tgt));
289 ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
292 * OST doesn't get object in advance for further use to prevent
293 * situations with nested object_find which is potential deadlock.
295 tsi->tsi_corpus = NULL;
300 * Do necessary preprocessing according to handler ->th_flags.
302 static int tgt_request_preprocess(struct tgt_session_info *tsi,
303 struct tgt_handler *h,
304 struct ptlrpc_request *req)
306 struct req_capsule *pill = tsi->tsi_pill;
307 __u32 flags = h->th_flags;
312 if (tsi->tsi_preprocessed)
315 LASSERT(h->th_act != NULL);
316 LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
317 LASSERT(current->journal_info == NULL);
319 LASSERT(ergo(flags & (HAS_BODY | HAS_REPLY),
321 if (h->th_fmt != NULL) {
322 req_capsule_set(pill, h->th_fmt);
323 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT) &&
324 req_capsule_field_present(pill, &RMF_MDT_BODY,
326 rc = tgt_mdt_body_unpack(tsi, flags);
329 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
331 req_capsule_field_present(pill, &RMF_OST_BODY,
333 rc = tgt_ost_body_unpack(tsi, flags);
339 if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
342 if (flags & HAS_KEY) {
343 struct ldlm_request *dlm_req;
345 LASSERT(h->th_fmt != NULL);
347 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
348 if (dlm_req != NULL) {
349 union ldlm_wire_policy_data *policy =
350 &dlm_req->lock_desc.l_policy_data;
352 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
354 (policy->l_inodebits.bits |
355 policy->l_inodebits.try_bits) == 0)) {
357 * Lock without inodebits makes no sense and
358 * will oops later in ldlm. If client miss to
359 * set such bits, do not trigger ASSERTION.
361 * For liblustre flock case, it maybe zero.
365 tsi->tsi_dlm_req = dlm_req;
371 tsi->tsi_preprocessed = 1;
376 * Invoke handler for this request opc. Also do necessary preprocessing
377 * (according to handler ->th_flags), and post-processing (setting of
378 * ->last_{xid,committed}).
380 static int tgt_handle_request0(struct tgt_session_info *tsi,
381 struct tgt_handler *h,
382 struct ptlrpc_request *req)
386 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
391 /* When dealing with sec context requests, no export is associated yet,
392 * because these requests are sent before *_CONNECT requests.
393 * A NULL req->rq_export means the normal *_common_slice handlers will
394 * not be called, because there is no reference to the target.
395 * So deal with them by hand and jump directly to target_send_reply().
399 case SEC_CTX_INIT_CONT:
401 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
406 * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
407 * to put same checks into handlers like mdt_close(), mdt_reint(),
408 * etc., without talking to mdt authors first. Checking same thing
409 * there again is useless and returning 0 error without packing reply
410 * is buggy! Handlers either pack reply or return error.
412 * We return 0 here and do not send any reply in order to emulate
413 * network failure. Do not send any reply in case any of NET related
414 * fail_id has occured.
416 if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
418 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
419 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
422 /* drop OUT_UPDATE rpc */
423 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == OUT_UPDATE &&
424 OBD_FAIL_CHECK(OBD_FAIL_OUT_UPDATE_DROP)))
427 rc = tgt_request_preprocess(tsi, h, req);
428 /* pack reply if reply format is fixed */
429 if (rc == 0 && h->th_flags & HAS_REPLY) {
431 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
433 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
435 tsi->tsi_mdt_body->mbo_eadatasize);
436 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
438 req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
440 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
441 req_capsule_set_size(tsi->tsi_pill,
442 &RMF_ACL, RCL_SERVER,
443 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
445 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
447 struct niobuf_remote *remote_nb =
448 req_capsule_client_get(tsi->tsi_pill,
450 struct ost_body *body = tsi->tsi_ost_body;
452 req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
454 (body->oa.o_valid & OBD_MD_FLFLAGS &&
455 body->oa.o_flags & OBD_FL_SHORT_IO) ?
456 remote_nb[0].rnb_len : 0);
459 rc = req_capsule_server_pack(tsi->tsi_pill);
462 if (likely(rc == 0)) {
464 * Process request, there can be two types of rc:
465 * 1) errors with msg unpack/pack, other failures outside the
466 * operation itself. This is counted as serious errors;
467 * 2) errors during fs operation, should be placed in rq_status
471 if (!is_serious(rc) &&
472 !req->rq_no_reply && req->rq_reply_state == NULL) {
473 DEBUG_REQ(D_ERROR, req,
474 "%s: %s handler did not pack reply but returned no error",
475 tgt_name(tsi->tsi_tgt), h->th_name);
478 serious = is_serious(rc);
479 rc = clear_serious(rc);
487 * ELDLM_* codes which > 0 should be in rq_status only as well as
488 * all non-serious errors.
490 if (rc > 0 || !serious)
493 LASSERT(current->journal_info == NULL);
495 if (likely(rc == 0 && req->rq_export))
496 target_committed_to_req(req);
499 target_send_reply(req, rc, tsi->tsi_reply_fail_id);
503 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
504 struct obd_device *obd, int *process)
506 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
513 case MDS_SYNC: /* used in unmounting */
527 case MDS_HSM_PROGRESS:
528 case MDS_HSM_STATE_SET:
529 case MDS_HSM_REQUEST:
531 *process = target_queue_recovery_request(req, obd);
535 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
542 * Handle recovery. Return:
543 * +1: continue request processing;
544 * -ve: abort immediately with the given error code;
545 * 0: send reply with error code in req->rq_status;
547 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
551 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
556 case SEC_CTX_INIT_CONT:
561 if (!req->rq_export->exp_obd->obd_replayable)
564 /* sanity check: if the xid matches, the request must be marked as a
565 * resent or replayed */
566 if (req_can_reconstruct(req, NULL) == 1) {
567 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
568 (MSG_RESENT | MSG_REPLAY))) {
569 DEBUG_REQ(D_WARNING, req,
570 "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
572 lustre_msg_get_flags(req->rq_reqmsg));
573 req->rq_status = -ENOTCONN;
577 /* else: note the opposite is not always true; a RESENT req after a
578 * failover will usually not match the last_xid, since it was likely
579 * never committed. A REPLAYed request will almost never match the
580 * last xid, however it could for a committed, but still retained,
583 /* Check for aborted recovery... */
584 if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
588 DEBUG_REQ(D_INFO, req, "Got new replay");
589 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
591 if (rc != 0 || !should_process)
593 else if (should_process < 0) {
594 req->rq_status = should_process;
595 rc = ptlrpc_error(req);
602 /* Initial check for request, it is validation mostly */
603 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
605 struct tgt_handler *h;
606 struct tgt_opc_slice *s;
607 struct lu_target *tgt;
608 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
612 tgt = class_exp2tgt(req->rq_export);
613 if (unlikely(tgt == NULL)) {
614 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
615 class_exp2obd(req->rq_export)->obd_name);
616 RETURN(ERR_PTR(-EINVAL));
619 for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
620 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
623 /* opcode was not found in slice */
624 if (unlikely(s->tos_hs == NULL)) {
627 /* don't spew error messages for unhandled RPCs */
629 CERROR("%s: no handler for opcode 0x%x from %s\n",
630 tgt_name(tgt), opc, libcfs_id2str(req->rq_peer));
633 RETURN(ERR_PTR(-ENOTSUPP));
636 LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
637 h = s->tos_hs + (opc - s->tos_opc_start);
638 if (unlikely(h->th_opc == 0)) {
639 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
640 RETURN(ERR_PTR(-ENOTSUPP));
646 static int process_req_last_xid(struct ptlrpc_request *req)
650 struct obd_export *exp = req->rq_export;
651 struct tg_export_data *ted = &exp->exp_target_data;
652 bool need_lock = tgt_is_multimodrpcs_client(exp);
656 mutex_lock(&ted->ted_lcd_lock);
657 /* check request's xid is consistent with export's last_xid */
658 last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
659 if (last_xid > exp->exp_last_xid)
660 exp->exp_last_xid = last_xid;
662 if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
663 /* Some request is allowed to be sent during replay,
664 * such as OUT update requests, FLD requests, so it
665 * is possible that replay requests has smaller XID
666 * than the exp_last_xid.
668 * Some non-replay requests may have smaller XID as
671 * - Client send a no_resend RPC, like statfs;
672 * - The RPC timedout (or some other error) on client,
673 * then it's removed from the unreplied list;
674 * - Client send some other request to bump the
675 * exp_last_xid on server;
676 * - The former RPC got chance to be processed;
678 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
681 DEBUG_REQ(D_WARNING, req,
682 "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
683 req->rq_xid, exp->exp_last_xid, rc);
688 /* The "last_xid" is the minimum xid among unreplied requests,
689 * if the request is from the previous connection, its xid can
690 * still be larger than "exp_last_xid", then the above check of
691 * xid is not enough to determine whether the request is delayed.
693 * For example, if some replay request was delayed and caused
694 * timeout at client and the replay is restarted, the delayed
695 * replay request will have the larger xid than "exp_last_xid"
697 if (req->rq_export->exp_conn_cnt >
698 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
700 "Dropping request %llu from an old epoch %u/%u\n",
702 lustre_msg_get_conn_cnt(req->rq_reqmsg),
703 req->rq_export->exp_conn_cnt);
704 req->rq_no_reply = 1;
705 GOTO(out, rc = -ESTALE);
708 /* try to release in-memory reply data */
709 if (tgt_is_multimodrpcs_client(exp)) {
710 tgt_handle_received_xid(exp, last_xid);
711 rc = tgt_handle_tag(req);
716 mutex_unlock(&ted->ted_lcd_lock);
721 int tgt_request_handle(struct ptlrpc_request *req)
723 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
725 struct lustre_msg *msg = req->rq_reqmsg;
726 struct tgt_handler *h;
727 struct lu_target *tgt;
728 int request_fail_id = 0;
729 __u32 opc = lustre_msg_get_opc(msg);
730 struct obd_device *obd;
732 bool is_connect = false;
735 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
736 if (cfs_fail_val == 0 &&
737 lustre_msg_get_opc(msg) != OBD_PING &&
738 lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
741 wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
745 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
746 tsi->tsi_pill = &req->rq_pill;
747 tsi->tsi_env = req->rq_svc_thread->t_env;
749 /* if request has export then get handlers slice from corresponding
750 * target, otherwise that should be connect operation */
751 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
752 opc == MGS_CONNECT) {
754 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
755 rc = target_handle_connect(req);
757 rc = ptlrpc_error(req);
760 /* recovery-small test 18c asks to drop connect reply */
761 if (unlikely(opc == OST_CONNECT &&
762 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
766 if (unlikely(!class_connected_export(req->rq_export))) {
767 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
768 opc == SEC_CTX_FINI) {
769 /* sec context initialization has to be handled
770 * by hand in tgt_handle_request0() */
771 tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
773 GOTO(handle_recov, rc = 0);
775 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
776 opc, libcfs_id2str(req->rq_peer));
777 req->rq_status = -ENOTCONN;
778 rc = ptlrpc_error(req);
782 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
783 tsi->tsi_exp = req->rq_export;
784 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
785 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
787 tsi->tsi_jobid = NULL;
790 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
791 class_exp2obd(req->rq_export)->obd_name);
792 req->rq_status = -EINVAL;
793 rc = ptlrpc_error(req);
797 /* Skip last_xid processing for the recovery thread, otherwise, the
798 * last_xid on same request could be processed twice: first time when
799 * processing the incoming request, second time when the request is
800 * being processed by recovery thread. */
801 obd = class_exp2obd(req->rq_export);
803 /* reset the exp_last_xid on each connection. */
804 req->rq_export->exp_last_xid = 0;
805 } else if (obd->obd_recovery_data.trd_processing_task !=
807 rc = process_req_last_xid(req);
810 rc = ptlrpc_error(req);
815 request_fail_id = tgt->lut_request_fail_id;
816 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
818 h = tgt_handler_find_check(req);
820 req->rq_status = PTR_ERR(h);
821 rc = ptlrpc_error(req);
825 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
828 if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
829 CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
832 rc = lustre_msg_check_version(msg, h->th_version);
834 DEBUG_REQ(D_ERROR, req,
835 "%s: drop malformed request version=%08x expect=%08x",
836 tgt_name(tgt), lustre_msg_get_version(msg),
838 req->rq_status = -EINVAL;
839 rc = ptlrpc_error(req);
844 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
845 if (likely(rc == 1)) {
846 rc = tgt_handle_request0(tsi, h, req);
852 req_capsule_fini(tsi->tsi_pill);
853 if (tsi->tsi_corpus != NULL) {
854 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
855 tsi->tsi_corpus = NULL;
859 EXPORT_SYMBOL(tgt_request_handle);
861 /** Assign high priority operations to the request if needed. */
862 int tgt_hpreq_handler(struct ptlrpc_request *req)
864 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
865 struct tgt_handler *h;
870 if (req->rq_export == NULL)
873 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
874 tsi->tsi_pill = &req->rq_pill;
875 tsi->tsi_env = req->rq_svc_thread->t_env;
876 tsi->tsi_tgt = class_exp2tgt(req->rq_export);
877 tsi->tsi_exp = req->rq_export;
879 h = tgt_handler_find_check(req);
885 rc = tgt_request_preprocess(tsi, h, req);
886 if (unlikely(rc != 0))
889 if (h->th_hp != NULL)
893 EXPORT_SYMBOL(tgt_hpreq_handler);
895 void tgt_counter_incr(struct obd_export *exp, int opcode)
897 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
898 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
899 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
901 EXPORT_SYMBOL(tgt_counter_incr);
904 * Unified target generic handlers.
907 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
909 struct lu_target *tgt = class_exp2tgt(exp);
910 struct sptlrpc_flavor flvr;
914 LASSERT(tgt->lut_obd);
915 LASSERT(tgt->lut_slice);
917 /* always allow ECHO client */
918 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
919 LUSTRE_ECHO_NAME) == 0)) {
920 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
924 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
925 read_lock(&tgt->lut_sptlrpc_lock);
926 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
930 read_unlock(&tgt->lut_sptlrpc_lock);
932 spin_lock(&exp->exp_lock);
933 exp->exp_sp_peer = req->rq_sp_from;
934 exp->exp_flvr = flvr;
936 /* when on mgs, if no restriction is set, or if the client
937 * NID is on the local node, allow any flavor
939 if ((strcmp(exp->exp_obd->obd_type->typ_name,
940 LUSTRE_MGS_NAME) == 0) &&
941 (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
942 LNetIsPeerLocal(exp->exp_connection->c_peer.nid)))
943 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
945 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
946 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
947 CERROR("%s: unauthorized rpc flavor %x from %s, "
948 "expect %x\n", tgt_name(tgt),
950 libcfs_nid2str(req->rq_peer.nid),
951 exp->exp_flvr.sf_rpc);
954 spin_unlock(&exp->exp_lock);
956 if (exp->exp_sp_peer != req->rq_sp_from) {
957 CERROR("%s: RPC source %s doesn't match %s\n",
959 sptlrpc_part2name(req->rq_sp_from),
960 sptlrpc_part2name(exp->exp_sp_peer));
963 rc = sptlrpc_target_export_check(exp, req);
970 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
972 struct sptlrpc_rule_set tmp_rset;
975 if (unlikely(tgt == NULL)) {
976 CERROR("No target passed\n");
980 sptlrpc_rule_set_init(&tmp_rset);
981 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
983 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
988 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
990 write_lock(&tgt->lut_sptlrpc_lock);
991 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
992 tgt->lut_sptlrpc_rset = tmp_rset;
993 write_unlock(&tgt->lut_sptlrpc_lock);
997 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
999 int tgt_connect(struct tgt_session_info *tsi)
1001 struct ptlrpc_request *req = tgt_ses_req(tsi);
1002 struct obd_connect_data *reply;
1007 /* XXX: better to call this check right after getting new export but
1008 * before last_rcvd slot allocation to avoid server load upon insecure
1009 * connects. This is to be fixed after unifiyng all targets.
1011 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
1015 /* To avoid exposing partially initialized connection flags, changes up
1016 * to this point have been staged in reply->ocd_connect_flags. Now that
1017 * connection handling has completed successfully, atomically update
1018 * the connect flags in the shared export data structure. LU-1623 */
1019 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1020 spin_lock(&tsi->tsi_exp->exp_lock);
1021 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1022 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1023 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1024 reply->ocd_connect_flags2;
1025 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1026 spin_unlock(&tsi->tsi_exp->exp_lock);
1028 if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1029 LUSTRE_MDT_NAME) == 0) {
1030 rc = req_check_sepol(tsi->tsi_pill);
1034 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2 &&
1035 reply->ocd_connect_flags2 & OBD_CONNECT2_ENCRYPT &&
1036 tsi->tsi_pill->rc_req->rq_export) {
1037 bool forbid_encrypt = true;
1038 struct lu_nodemap *nm =
1039 nodemap_get_from_exp(tsi->tsi_pill->rc_req->rq_export);
1042 /* nodemap_get_from_exp returns NULL in case
1043 * nodemap is not active, so we do not forbid
1045 forbid_encrypt = false;
1046 } else if (!IS_ERR(nm)) {
1047 forbid_encrypt = nm->nmf_forbid_encryption;
1052 GOTO(out, rc = -EACCES);
1058 obd_disconnect(class_export_get(tsi->tsi_exp));
1061 EXPORT_SYMBOL(tgt_connect);
1063 int tgt_disconnect(struct tgt_session_info *tsi)
1069 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1071 rc = target_handle_disconnect(tgt_ses_req(tsi));
1073 RETURN(err_serious(rc));
1077 EXPORT_SYMBOL(tgt_disconnect);
1080 * Unified target OBD handlers
1082 int tgt_obd_ping(struct tgt_session_info *tsi)
1088 /* The target-specific part of OBD_PING request handling.
1089 * It controls Filter Modification Data (FMD) expiration each time
1092 * Valid only for replayable targets, e.g. MDT and OFD
1094 if (tsi->tsi_exp->exp_obd->obd_replayable)
1095 tgt_fmd_expire(tsi->tsi_exp);
1097 rc = req_capsule_server_pack(tsi->tsi_pill);
1099 RETURN(err_serious(rc));
1103 EXPORT_SYMBOL(tgt_obd_ping);
1105 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1107 return err_serious(-EOPNOTSUPP);
1110 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1112 struct ptlrpc_request *req = tgt_ses_req(tsi);
1113 struct obd_export *exp = req->rq_export;
1114 struct ptlrpc_bulk_desc *desc;
1121 for (i = 0; i < rdbuf->rb_nbufs; i++) {
1122 unsigned int offset;
1124 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1125 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1129 desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1130 PTLRPC_BULK_PUT_SOURCE,
1132 &ptlrpc_bulk_kiov_nopin_ops);
1136 for (i = 0; i < rdbuf->rb_nbufs; i++)
1137 desc->bd_frag_ops->add_iov_frag(desc,
1138 rdbuf->rb_bufs[i].lb_buf,
1139 rdbuf->rb_bufs[i].lb_len);
1141 rc = target_bulk_io(exp, desc);
1142 ptlrpc_free_bulk(desc);
1145 EXPORT_SYMBOL(tgt_send_buffer);
1147 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1149 struct ptlrpc_request *req = tgt_ses_req(tsi);
1150 struct obd_export *exp = req->rq_export;
1151 struct ptlrpc_bulk_desc *desc;
1159 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1160 PTLRPC_BULK_PUT_SOURCE,
1162 &ptlrpc_bulk_kiov_pin_ops);
1166 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1167 /* old client requires reply size in it's PAGE_SIZE,
1168 * which is rdpg->rp_count */
1169 nob = rdpg->rp_count;
1171 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1172 i++, tmpcount -= tmpsize) {
1173 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1174 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1178 LASSERT(desc->bd_nob == nob);
1179 rc = target_bulk_io(exp, desc);
1180 ptlrpc_free_bulk(desc);
1183 EXPORT_SYMBOL(tgt_sendpage);
1186 * OBD_IDX_READ handler
1188 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1190 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1191 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1192 struct idx_info *req_ii, *rep_ii;
1197 memset(rdpg, 0, sizeof(*rdpg));
1198 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1200 /* extract idx_info buffer from request & reply */
1201 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1202 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1203 RETURN(err_serious(-EPROTO));
1205 rc = req_capsule_server_pack(tsi->tsi_pill);
1207 RETURN(err_serious(rc));
1209 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1211 RETURN(err_serious(-EFAULT));
1212 rep_ii->ii_magic = IDX_INFO_MAGIC;
1214 /* extract hash to start with */
1215 rdpg->rp_hash = req_ii->ii_hash_start;
1217 /* extract requested attributes */
1218 rdpg->rp_attrs = req_ii->ii_attrs;
1220 /* check that fid packed in request is valid and supported */
1221 if (!fid_is_sane(&req_ii->ii_fid))
1223 rep_ii->ii_fid = req_ii->ii_fid;
1226 rep_ii->ii_flags = req_ii->ii_flags;
1228 /* compute number of pages to allocate, ii_count is the number of 4KB
1230 if (req_ii->ii_count <= 0)
1231 GOTO(out, rc = -EFAULT);
1232 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1233 exp_max_brw_size(tsi->tsi_exp));
1234 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1236 /* allocate pages to store the containers */
1237 OBD_ALLOC_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1238 if (rdpg->rp_pages == NULL)
1239 GOTO(out, rc = -ENOMEM);
1240 for (i = 0; i < rdpg->rp_npages; i++) {
1241 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1242 if (rdpg->rp_pages[i] == NULL)
1243 GOTO(out, rc = -ENOMEM);
1246 /* populate pages with key/record pairs */
1247 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1251 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1252 "asked %d > %d\n", rc, rdpg->rp_count);
1254 /* send pages to client */
1255 rc = tgt_sendpage(tsi, rdpg, rc);
1260 if (rdpg->rp_pages) {
1261 for (i = 0; i < rdpg->rp_npages; i++)
1262 if (rdpg->rp_pages[i])
1263 __free_page(rdpg->rp_pages[i]);
1264 OBD_FREE_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1269 struct tgt_handler tgt_obd_handlers[] = {
1270 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1271 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1273 EXPORT_SYMBOL(tgt_obd_handlers);
1275 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1276 struct dt_object *obj, __u64 start, __u64 end)
1282 /* if no objid is specified, it means "sync whole filesystem" */
1284 rc = dt_sync(env, tgt->lut_bottom);
1285 } else if (dt_version_get(env, obj) >
1286 tgt->lut_obd->obd_last_committed) {
1287 rc = dt_object_sync(env, obj, start, end);
1289 atomic_inc(&tgt->lut_sync_count);
1293 EXPORT_SYMBOL(tgt_sync);
1295 * Unified target DLM handlers.
1299 * Unified target BAST
1301 * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1302 * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1303 * but if it's cross-MDT lock, because remote object version is not set, a
1304 * filesystem sync is needed.
1306 * \param lock server side lock
1307 * \param desc lock desc
1308 * \param data ldlm_cb_set_arg
1309 * \param flag indicates whether this cancelling or blocking callback
1310 * \retval 0 on success
1311 * \retval negative number on error
1313 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1314 void *data, int flag)
1317 struct lu_target *tgt;
1318 struct dt_object *obj = NULL;
1324 tgt = class_exp2tgt(lock->l_export);
1326 if (unlikely(tgt == NULL)) {
1327 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1328 class_exp2obd(lock->l_export)->obd_name);
1332 if (flag == LDLM_CB_CANCELING &&
1333 (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1334 (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1335 (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1336 ldlm_is_cbpending(lock))) &&
1337 ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1338 lock->l_resource->lr_type == LDLM_EXTENT)) {
1340 __u64 end = OBD_OBJECT_EOF;
1342 rc = lu_env_init(&env, LCT_DT_THREAD);
1343 if (unlikely(rc != 0))
1346 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1347 tgt->lut_lsd.lsd_osd_index);
1349 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1350 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1352 GOTO(err_env, rc = PTR_ERR(obj));
1354 if (!dt_object_exists(obj))
1355 GOTO(err_put, rc = -ENOENT);
1357 start = lock->l_policy_data.l_extent.start;
1358 end = lock->l_policy_data.l_extent.end;
1361 rc = tgt_sync(&env, tgt, obj, start, end);
1363 CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1364 "cancel: rc = %d\n",
1365 tgt_name(tgt), PFID(&fid),
1366 lock->l_policy_data.l_extent.start,
1367 lock->l_policy_data.l_extent.end, rc);
1371 dt_object_put(&env, obj);
1376 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1380 static struct ldlm_callback_suite tgt_dlm_cbs = {
1381 .lcs_completion = ldlm_server_completion_ast,
1382 .lcs_blocking = tgt_blocking_ast,
1383 .lcs_glimpse = ldlm_server_glimpse_ast
1386 int tgt_enqueue(struct tgt_session_info *tsi)
1388 struct ptlrpc_request *req = tgt_ses_req(tsi);
1393 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1394 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1396 LASSERT(tsi->tsi_dlm_req != NULL);
1397 rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1398 tsi->tsi_dlm_req, &tgt_dlm_cbs);
1400 RETURN(err_serious(rc));
1402 switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1404 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1407 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1410 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1413 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1416 RETURN(req->rq_status);
1418 EXPORT_SYMBOL(tgt_enqueue);
1420 int tgt_convert(struct tgt_session_info *tsi)
1422 struct ptlrpc_request *req = tgt_ses_req(tsi);
1426 LASSERT(tsi->tsi_dlm_req);
1427 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1429 RETURN(err_serious(rc));
1431 RETURN(req->rq_status);
1434 int tgt_bl_callback(struct tgt_session_info *tsi)
1436 return err_serious(-EOPNOTSUPP);
1439 int tgt_cp_callback(struct tgt_session_info *tsi)
1441 return err_serious(-EOPNOTSUPP);
1444 /* generic LDLM target handler */
1445 struct tgt_handler tgt_dlm_handlers[] = {
1446 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1447 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1448 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1449 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1451 EXPORT_SYMBOL(tgt_dlm_handlers);
1454 * Unified target LLOG handlers.
1456 int tgt_llog_open(struct tgt_session_info *tsi)
1462 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1466 EXPORT_SYMBOL(tgt_llog_open);
1468 int tgt_llog_read_header(struct tgt_session_info *tsi)
1474 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1478 EXPORT_SYMBOL(tgt_llog_read_header);
1480 int tgt_llog_next_block(struct tgt_session_info *tsi)
1486 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1490 EXPORT_SYMBOL(tgt_llog_next_block);
1492 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1498 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1502 EXPORT_SYMBOL(tgt_llog_prev_block);
1504 /* generic llog target handler */
1505 struct tgt_handler tgt_llog_handlers[] = {
1506 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1507 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1508 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1509 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1511 EXPORT_SYMBOL(tgt_llog_handlers);
1514 * sec context handlers
1516 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1517 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1522 struct tgt_handler tgt_sec_ctx_handlers[] = {
1523 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle),
1524 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle),
1525 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle),
1527 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1529 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1530 struct dt_device *key,
1531 struct lfsck_req_local *lrl,
1532 struct thandle *th) = NULL;
1534 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1536 struct lfsck_req_local *,
1539 tgt_lfsck_in_notify_local = notify;
1541 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1543 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1544 struct dt_device *key,
1545 struct lfsck_request *lr) = NULL;
1547 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1549 struct lfsck_request *))
1551 tgt_lfsck_in_notify = notify;
1553 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1555 static int (*tgt_lfsck_query)(const struct lu_env *env,
1556 struct dt_device *key,
1557 struct lfsck_request *req,
1558 struct lfsck_reply *rep,
1559 struct lfsck_query *que) = NULL;
1561 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1563 struct lfsck_request *,
1564 struct lfsck_reply *,
1565 struct lfsck_query *))
1567 tgt_lfsck_query = query;
1569 EXPORT_SYMBOL(tgt_register_lfsck_query);
1571 /* LFSCK request handlers */
1572 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1574 const struct lu_env *env = tsi->tsi_env;
1575 struct dt_device *key = tsi->tsi_tgt->lut_bottom;
1576 struct lfsck_request *lr;
1580 lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1584 rc = tgt_lfsck_in_notify(env, key, lr);
1589 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1591 struct lfsck_request *request;
1592 struct lfsck_reply *reply;
1596 request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1597 if (request == NULL)
1600 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1604 rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1605 request, reply, NULL);
1607 RETURN(rc < 0 ? rc : 0);
1610 struct tgt_handler tgt_lfsck_handlers[] = {
1611 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_NOTIFY, tgt_handle_lfsck_notify),
1612 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_QUERY, tgt_handle_lfsck_query),
1614 EXPORT_SYMBOL(tgt_lfsck_handlers);
1617 * initialize per-thread page pool (bug 5137).
1619 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1621 struct tgt_thread_big_cache *tbc;
1625 LASSERT(thread != NULL);
1626 LASSERT(thread->t_data == NULL);
1628 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1631 thread->t_data = tbc;
1634 EXPORT_SYMBOL(tgt_io_thread_init);
1637 * free per-thread pool created by tgt_thread_init().
1639 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1641 struct tgt_thread_big_cache *tbc;
1645 LASSERT(thread != NULL);
1648 * be prepared to handle partially-initialized pools (because this is
1649 * called from ost_io_thread_init() for cleanup.
1651 tbc = thread->t_data;
1653 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1654 thread->t_data = NULL;
1658 EXPORT_SYMBOL(tgt_io_thread_done);
1661 * Helper function for getting Data-on-MDT file server DLM lock
1662 * if asked by client.
1664 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1665 struct lustre_handle *lh, int mode, __u64 *flags)
1667 union ldlm_policy_data policy = {
1668 .l_inodebits.bits = MDS_INODELOCK_DOM,
1674 LASSERT(lh != NULL);
1675 LASSERT(ns != NULL);
1676 LASSERT(!lustre_handle_is_used(lh));
1678 rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1679 flags, ldlm_blocking_ast,
1680 ldlm_completion_ast, ldlm_glimpse_ast,
1681 NULL, 0, LVB_T_NONE, NULL, lh);
1683 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1685 EXPORT_SYMBOL(tgt_mdt_data_lock);
1688 * Helper function for getting server side [start, start+count] DLM lock
1689 * if asked by client.
1691 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1692 struct ldlm_res_id *res_id, __u64 start, __u64 end,
1693 struct lustre_handle *lh, int mode, __u64 *flags)
1695 union ldlm_policy_data policy;
1700 LASSERT(lh != NULL);
1701 LASSERT(ns != NULL);
1702 LASSERT(!lustre_handle_is_used(lh));
1704 policy.l_extent.gid = 0;
1705 policy.l_extent.start = start & PAGE_MASK;
1708 * If ->o_blocks is EOF it means "lock till the end of the file".
1709 * Otherwise, it's size of an extent or hole being punched (in bytes).
1711 if (end == OBD_OBJECT_EOF || end < start)
1712 policy.l_extent.end = OBD_OBJECT_EOF;
1714 policy.l_extent.end = end | ~PAGE_MASK;
1716 rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1717 mode, flags, ldlm_blocking_ast,
1718 ldlm_completion_ast, ldlm_glimpse_ast,
1719 NULL, 0, LVB_T_NONE, NULL, lh);
1720 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1722 EXPORT_SYMBOL(tgt_extent_lock);
1724 static int tgt_data_lock(const struct lu_env *env, struct obd_export *exp,
1725 struct ldlm_res_id *res_id, __u64 start, __u64 end,
1726 struct lustre_handle *lh, enum ldlm_mode mode)
1728 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1731 /* MDT IO for data-on-mdt */
1732 if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1733 return tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1735 return tgt_extent_lock(env, ns, res_id, start, end, lh, mode, &flags);
1738 void tgt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1740 LASSERT(lustre_handle_is_used(lh));
1741 ldlm_lock_decref(lh, mode);
1743 EXPORT_SYMBOL(tgt_data_unlock);
1745 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1746 struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1747 struct niobuf_remote *nb, struct lustre_handle *lh,
1748 enum ldlm_mode mode)
1750 int nrbufs = obj->ioo_bufcnt;
1755 LASSERT(mode == LCK_PR || mode == LCK_PW);
1756 LASSERT(!lustre_handle_is_used(lh));
1758 if (exp->exp_obd->obd_recovering)
1761 if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1764 for (i = 1; i < nrbufs; i++)
1765 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1768 return tgt_data_lock(env, exp, res_id, nb[0].rnb_offset,
1769 nb[nrbufs - 1].rnb_offset +
1770 nb[nrbufs - 1].rnb_len - 1, lh, mode);
1773 static void tgt_brw_unlock(struct obd_export *exp, struct obd_ioobj *obj,
1774 struct niobuf_remote *niob,
1775 struct lustre_handle *lh, enum ldlm_mode mode)
1779 LASSERT(mode == LCK_PR || mode == LCK_PW);
1780 LASSERT((!exp->exp_obd->obd_recovering && obj->ioo_bufcnt &&
1781 niob[0].rnb_flags & OBD_BRW_SRVLOCK) ==
1782 lustre_handle_is_used(lh));
1784 if (lustre_handle_is_used(lh))
1785 tgt_data_unlock(lh, mode);
1789 static int tgt_checksum_niobuf(struct lu_target *tgt,
1790 struct niobuf_local *local_nb, int npages,
1791 int opc, enum cksum_types cksum_type,
1794 struct ahash_request *req;
1795 unsigned int bufsize;
1797 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1799 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1801 CERROR("%s: unable to initialize checksum hash %s\n",
1802 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1803 return PTR_ERR(req);
1806 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1807 for (i = 0; i < npages; i++) {
1808 /* corrupt the data before we compute the checksum, to
1809 * simulate a client->OST data error */
1810 if (i == 0 && opc == OST_WRITE &&
1811 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1812 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1813 int len = local_nb[i].lnb_len;
1814 struct page *np = tgt_page_to_corrupt;
1817 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1818 char *ptr2 = page_address(np);
1820 memcpy(ptr2 + off, ptr + off, len);
1821 memcpy(ptr2 + off, "bad3", min(4, len));
1824 /* LU-8376 to preserve original index for
1825 * display in dump_all_bulk_pages() */
1828 cfs_crypto_hash_update_page(req, np, off,
1832 CERROR("%s: can't alloc page for corruption\n",
1836 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1837 local_nb[i].lnb_page_offset & ~PAGE_MASK,
1838 local_nb[i].lnb_len);
1840 /* corrupt the data after we compute the checksum, to
1841 * simulate an OST->client data error */
1842 if (i == 0 && opc == OST_READ &&
1843 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1844 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1845 int len = local_nb[i].lnb_len;
1846 struct page *np = tgt_page_to_corrupt;
1849 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1850 char *ptr2 = page_address(np);
1852 memcpy(ptr2 + off, ptr + off, len);
1853 memcpy(ptr2 + off, "bad4", min(4, len));
1856 /* LU-8376 to preserve original index for
1857 * display in dump_all_bulk_pages() */
1860 cfs_crypto_hash_update_page(req, np, off,
1864 CERROR("%s: can't alloc page for corruption\n",
1870 bufsize = sizeof(*cksum);
1871 err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1876 char dbgcksum_file_name[PATH_MAX];
1878 static void dump_all_bulk_pages(struct obdo *oa, int count,
1879 struct niobuf_local *local_nb,
1880 __u32 server_cksum, __u32 client_cksum)
1887 /* will only keep dump of pages on first error for the same range in
1888 * file/fid, not during the resends/retries. */
1889 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1890 "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1891 (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1892 libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1893 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1894 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1895 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1896 local_nb[0].lnb_file_offset,
1897 local_nb[count-1].lnb_file_offset +
1898 local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1899 filp = filp_open(dbgcksum_file_name,
1900 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1904 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1905 "checksum error: rc = %d\n", dbgcksum_file_name,
1908 CERROR("%s: can't open to dump pages with checksum "
1909 "error: rc = %d\n", dbgcksum_file_name, rc);
1913 for (i = 0; i < count; i++) {
1914 len = local_nb[i].lnb_len;
1915 buf = kmap(local_nb[i].lnb_page);
1917 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1919 CERROR("%s: wanted to write %u but got %d "
1920 "error\n", dbgcksum_file_name, len, rc);
1925 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1926 dbgcksum_file_name, rc);
1928 kunmap(local_nb[i].lnb_page);
1931 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1933 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1934 filp_close(filp, NULL);
1937 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1938 struct obd_export *exp, struct obdo *oa,
1939 const struct lnet_process_id *peer,
1940 __u32 client_cksum, __u32 server_cksum,
1941 enum cksum_types server_cksum_type)
1944 enum cksum_types cksum_type;
1947 /* unlikely to happen and only if resend does not occur due to cksum
1948 * control failure on Client */
1949 if (unlikely(server_cksum == client_cksum)) {
1950 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1955 if (exp->exp_obd->obd_checksum_dump)
1956 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1959 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1962 if (cksum_type != server_cksum_type)
1963 msg = "the server may have not used the checksum type specified"
1964 " in the original request - likely a protocol problem";
1966 msg = "should have changed on the client or in transit";
1968 start = local_nb[0].lnb_file_offset;
1969 end = local_nb[npages-1].lnb_file_offset +
1970 local_nb[npages-1].lnb_len - 1;
1972 LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1973 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1974 " %x (type %x), server csum %x (type %x)\n",
1975 exp->exp_obd->obd_name,
1976 msg, libcfs_nid2str(peer->nid),
1977 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1978 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1979 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1981 start, end, client_cksum, cksum_type, server_cksum,
1987 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1988 unsigned char *buf, int size)
1990 int i, off, len, copied = size;
1993 for (i = 0; i < npages; i++) {
1994 off = local[i].lnb_page_offset & ~PAGE_MASK;
1995 len = local[i].lnb_len;
1997 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2002 ptr = kmap_atomic(local[i].lnb_page);
2003 memcpy(buf, ptr + off, len);
2008 return copied - size;
2011 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
2012 struct niobuf_local *local_nb,
2013 int npages, int opc,
2014 obd_dif_csum_fn *fn,
2018 enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
2019 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
2020 const char *obd_name = tgt->lut_obd->obd_name;
2021 struct ahash_request *req;
2022 unsigned int bufsize;
2023 unsigned char *buffer;
2024 struct page *__page;
2027 int used_number = 0;
2033 __page = alloc_page(GFP_KERNEL);
2037 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
2039 CERROR("%s: unable to initialize checksum hash %s\n",
2040 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
2041 return PTR_ERR(req);
2044 buffer = kmap(__page);
2045 guard_start = (__u16 *)buffer;
2046 guard_number = PAGE_SIZE / sizeof(*guard_start);
2047 for (i = 0; i < npages; i++) {
2048 /* corrupt the data before we compute the checksum, to
2049 * simulate a client->OST data error */
2050 if (i == 0 && opc == OST_WRITE &&
2051 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2052 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2053 int len = local_nb[i].lnb_len;
2054 struct page *np = tgt_page_to_corrupt;
2057 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2058 char *ptr2 = page_address(np);
2060 memcpy(ptr2 + off, ptr + off, len);
2061 memcpy(ptr2 + off, "bad3", min(4, len));
2064 /* LU-8376 to preserve original index for
2065 * display in dump_all_bulk_pages() */
2068 cfs_crypto_hash_update_page(req, np, off,
2072 CERROR("%s: can't alloc page for corruption\n",
2078 * The left guard number should be able to hold checksums of a
2081 if (t10_cksum_type && opc == OST_READ &&
2082 local_nb[i].lnb_len == PAGE_SIZE &&
2083 local_nb[i].lnb_guard_disk) {
2084 used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2085 if (used > (guard_number - used_number)) {
2089 memcpy(guard_start + used_number,
2090 local_nb[i].lnb_guards,
2091 used * sizeof(*local_nb[i].lnb_guards));
2093 rc = obd_page_dif_generate_buffer(obd_name,
2094 local_nb[i].lnb_page,
2095 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2096 local_nb[i].lnb_len, guard_start + used_number,
2097 guard_number - used_number, &used, sector_size,
2103 LASSERT(used <= MAX_GUARD_NUMBER);
2105 * If disk support T10PI checksum, copy guards to local_nb.
2106 * If the write is partial page, do not use the guards for bio
2107 * submission since the data might not be full-sector. The bio
2108 * guards will be generated later based on the full sectors. If
2109 * the sector size is 512B rather than 4 KB, or the page size
2110 * is larger than 4KB, this might drop some useful guards for
2111 * partial page write, but it will only add minimal extra time
2112 * of checksum calculation.
2114 if (t10_cksum_type && opc == OST_WRITE &&
2115 local_nb[i].lnb_len == PAGE_SIZE) {
2116 local_nb[i].lnb_guard_rpc = 1;
2117 memcpy(local_nb[i].lnb_guards,
2118 guard_start + used_number,
2119 used * sizeof(*local_nb[i].lnb_guards));
2122 used_number += used;
2123 if (used_number == guard_number) {
2124 cfs_crypto_hash_update_page(req, __page, 0,
2125 used_number * sizeof(*guard_start));
2129 /* corrupt the data after we compute the checksum, to
2130 * simulate an OST->client data error */
2131 if (unlikely(i == 0 && opc == OST_READ &&
2132 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2133 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2134 int len = local_nb[i].lnb_len;
2135 struct page *np = tgt_page_to_corrupt;
2138 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2139 char *ptr2 = page_address(np);
2141 memcpy(ptr2 + off, ptr + off, len);
2142 memcpy(ptr2 + off, "bad4", min(4, len));
2145 /* LU-8376 to preserve original index for
2146 * display in dump_all_bulk_pages() */
2149 cfs_crypto_hash_update_page(req, np, off,
2153 CERROR("%s: can't alloc page for corruption\n",
2162 if (used_number != 0)
2163 cfs_crypto_hash_update_page(req, __page, 0,
2164 used_number * sizeof(*guard_start));
2166 bufsize = sizeof(cksum);
2167 rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2172 __free_page(__page);
2176 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2177 enum cksum_types cksum_type,
2178 struct niobuf_local *local_nb,
2179 int npages, int opc, u32 *check_sum)
2181 obd_dif_csum_fn *fn = NULL;
2182 int sector_size = 0;
2186 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
2189 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2190 opc, fn, sector_size,
2193 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2194 cksum_type, check_sum);
2198 int tgt_brw_read(struct tgt_session_info *tsi)
2200 struct ptlrpc_request *req = tgt_ses_req(tsi);
2201 struct ptlrpc_bulk_desc *desc = NULL;
2202 struct obd_export *exp = tsi->tsi_exp;
2203 struct niobuf_remote *remote_nb;
2204 struct niobuf_local *local_nb;
2205 struct obd_ioobj *ioo;
2206 struct ost_body *body, *repbody;
2207 struct lustre_handle lockh = { 0 };
2208 int npages, nob = 0, rc, i, no_reply = 0,
2210 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2211 const char *obd_name = exp->exp_obd->obd_name;
2215 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2216 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2217 CERROR("%s: deny read request from %s to portal %u\n",
2218 tgt_name(tsi->tsi_tgt),
2219 obd_export_nid2str(req->rq_export),
2220 ptlrpc_req2svc(req)->srv_req_portal);
2224 req->rq_bulk_read = 1;
2226 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2229 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2230 cfs_fail_val : (obd_timeout + 1) / 4);
2232 /* Check if there is eviction in progress, and if so, wait for it to
2234 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2235 /* We do not care how long it takes */
2237 exp->exp_obd->obd_evict_inprogress_waitq,
2238 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2241 /* There must be big cache in current thread to process this request
2242 * if it is NULL then something went wrong and it wasn't allocated,
2243 * report -ENOMEM in that case */
2247 body = tsi->tsi_ost_body;
2248 LASSERT(body != NULL);
2250 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2251 body->oa.o_flags & OBD_FL_NORPC)
2254 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2255 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2257 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2258 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2260 local_nb = tbc->local;
2262 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2268 * If getting the lock took more time than
2269 * client was willing to wait, drop it. b=11330
2271 if (ktime_get_real_seconds() > req->rq_deadline ||
2272 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2274 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2275 libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2276 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2277 req->rq_deadline - req->rq_arrival_time.tv_sec);
2278 GOTO(out_lock, rc = -ETIMEDOUT);
2281 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2282 repbody->oa = body->oa;
2284 npages = PTLRPC_MAX_BRW_PAGES;
2285 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2286 ioo, remote_nb, &npages, local_nb);
2290 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2291 body->oa.o_flags & OBD_FL_SHORT_IO) {
2294 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2295 PTLRPC_BULK_PUT_SOURCE,
2297 &ptlrpc_bulk_kiov_nopin_ops);
2299 GOTO(out_commitrw, rc = -ENOMEM);
2303 npages_read = npages;
2304 for (i = 0; i < npages; i++) {
2305 int page_rc = local_nb[i].lnb_rc;
2314 if (page_rc != 0 && desc != NULL) { /* some data! */
2315 LASSERT(local_nb[i].lnb_page != NULL);
2316 desc->bd_frag_ops->add_kiov_frag
2317 (desc, local_nb[i].lnb_page,
2318 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2322 if (page_rc != local_nb[i].lnb_len) { /* short read */
2323 local_nb[i].lnb_len = page_rc;
2324 npages_read = i + (page_rc != 0 ? 1 : 0);
2325 /* All subsequent pages should be 0 */
2326 while (++i < npages)
2327 LASSERT(local_nb[i].lnb_rc == 0);
2332 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2333 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2334 body->oa.o_flags : 0;
2335 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2337 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2339 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2341 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2342 local_nb, npages_read, OST_READ,
2343 &repbody->oa.o_cksum);
2345 GOTO(out_commitrw, rc);
2346 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2347 repbody->oa.o_cksum);
2349 /* if a resend it could be for a cksum error, so check Server
2350 * cksum with returned Client cksum (this should even cover
2351 * zero-cksum case) */
2352 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2353 (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2354 check_read_checksum(local_nb, npages_read, exp,
2355 &body->oa, &req->rq_peer,
2357 repbody->oa.o_cksum, cksum_type);
2359 repbody->oa.o_valid = 0;
2361 /* We're finishing using body->oa as an input variable */
2363 /* Check if client was evicted while we were doing i/o before touching
2366 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2367 body->oa.o_flags & OBD_FL_SHORT_IO) {
2368 unsigned char *short_io_buf;
2371 short_io_buf = req_capsule_server_get(&req->rq_pill,
2373 short_io_size = req_capsule_get_size(&req->rq_pill,
2376 rc = tgt_pages2shortio(local_nb, npages_read,
2377 short_io_buf, short_io_size);
2379 req_capsule_shrink(&req->rq_pill,
2382 rc = rc > 0 ? 0 : rc;
2383 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2384 rc = target_bulk_io(exp, desc);
2388 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2389 body->oa.o_flags & OBD_FL_SHORT_IO)
2390 req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2395 /* Must commit after prep above in all cases */
2396 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2397 remote_nb, npages, local_nb, rc);
2399 tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PR);
2401 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2402 ptlrpc_free_bulk(desc);
2407 ptlrpc_lprocfs_brw(req, nob);
2408 } else if (no_reply) {
2409 req->rq_no_reply = 1;
2410 /* reply out callback would free */
2411 ptlrpc_req_drop_rs(req);
2412 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2413 "client will retry: rc %d\n",
2415 obd_uuid2str(&exp->exp_client_uuid),
2416 obd_export_nid2str(exp), rc);
2418 /* send a bulk after reply to simulate a network delay or reordering
2419 * by a router - Note that !desc implies short io, so there is no bulk
2421 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2423 CDEBUG(D_INFO, "reorder BULK\n");
2426 target_bulk_io(exp, desc);
2427 ptlrpc_free_bulk(desc);
2432 EXPORT_SYMBOL(tgt_brw_read);
2434 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2435 unsigned char *buf, unsigned int size)
2440 for (i = 0; i < npages; i++) {
2441 off = local[i].lnb_page_offset & ~PAGE_MASK;
2442 len = local[i].lnb_len;
2447 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2449 ptr = kmap_atomic(local[i].lnb_page);
2452 memcpy(ptr + off, buf, len < size ? len : size);
2460 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2461 struct ptlrpc_bulk_desc *desc,
2462 struct niobuf_local *local_nb, int npages,
2463 u32 client_cksum, u32 server_cksum,
2466 struct obd_export *exp = req->rq_export;
2467 struct ost_body *body;
2471 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2472 LASSERT(body != NULL);
2474 if (desc && req->rq_peer.nid != desc->bd_sender) {
2476 router = libcfs_nid2str(desc->bd_sender);
2479 if (exp->exp_obd->obd_checksum_dump)
2480 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2484 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2485 client_cksum, server_cksum);
2489 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2490 DFID" object "DOSTID" extent [%llu-%llu"
2491 "]: client csum %x, server csum %x\n",
2492 exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2494 body->oa.o_valid & OBD_MD_FLFID ?
2495 body->oa.o_parent_seq : (__u64)0,
2496 body->oa.o_valid & OBD_MD_FLFID ?
2497 body->oa.o_parent_oid : 0,
2498 body->oa.o_valid & OBD_MD_FLFID ?
2499 body->oa.o_parent_ver : 0,
2500 POSTID(&body->oa.o_oi),
2501 local_nb[0].lnb_file_offset,
2502 local_nb[npages-1].lnb_file_offset +
2503 local_nb[npages - 1].lnb_len - 1,
2504 client_cksum, server_cksum);
2507 int tgt_brw_write(struct tgt_session_info *tsi)
2509 struct ptlrpc_request *req = tgt_ses_req(tsi);
2510 struct ptlrpc_bulk_desc *desc = NULL;
2511 struct obd_export *exp = req->rq_export;
2512 struct niobuf_remote *remote_nb;
2513 struct niobuf_local *local_nb;
2514 struct obd_ioobj *ioo;
2515 struct ost_body *body, *repbody;
2516 struct lustre_handle lockh = {0};
2518 int objcount, niocount, npages;
2520 enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2521 bool no_reply = false, mmap;
2522 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2523 bool wait_sync = false;
2524 const char *obd_name = exp->exp_obd->obd_name;
2525 /* '1' for consistency with code that checks !mpflag to restore */
2526 unsigned int mpflags = 1;
2530 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2531 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2532 CERROR("%s: deny write request from %s to portal %u\n",
2533 tgt_name(tsi->tsi_tgt),
2534 obd_export_nid2str(req->rq_export),
2535 ptlrpc_req2svc(req)->srv_req_portal);
2536 RETURN(err_serious(-EPROTO));
2539 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2540 RETURN(err_serious(-ENOSPC));
2541 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2542 RETURN(err_serious(-EROFS));
2544 req->rq_bulk_write = 1;
2546 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2547 RETURN(err_serious(-EIO));
2548 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2549 RETURN(err_serious(-EFAULT));
2551 /* pause before transaction has been started */
2552 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2553 cfs_fail_val : (obd_timeout + 1) / 4);
2555 /* Delay write commit to show stale size information */
2556 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2558 /* There must be big cache in current thread to process this request
2559 * if it is NULL then something went wrong and it wasn't allocated,
2560 * report -ENOMEM in that case */
2564 body = tsi->tsi_ost_body;
2565 LASSERT(body != NULL);
2567 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2568 body->oa.o_flags & OBD_FL_NORPC)
2572 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2573 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2575 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2576 RCL_CLIENT) / sizeof(*ioo);
2578 for (niocount = i = 0; i < objcount; i++)
2579 niocount += ioo[i].ioo_bufcnt;
2581 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2582 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2583 if (niocount != req_capsule_get_size(&req->rq_pill,
2584 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2586 RETURN(err_serious(-EPROTO));
2588 if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2589 ptlrpc_connection_is_local(exp->exp_connection))
2590 mpflags = memalloc_noreclaim_save();
2592 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2593 niocount * sizeof(*rcs));
2594 rc = req_capsule_server_pack(&req->rq_pill);
2596 GOTO(out, rc = err_serious(rc));
2598 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2599 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2601 local_nb = tbc->local;
2603 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2609 * If getting the lock took more time than
2610 * client was willing to wait, drop it. b=11330
2612 if (ktime_get_real_seconds() > req->rq_deadline ||
2613 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2615 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2616 tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2617 POSTID(&ioo->ioo_oid),
2618 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2619 req->rq_deadline - req->rq_arrival_time.tv_sec);
2620 GOTO(out_lock, rc = -ETIMEDOUT);
2623 /* Because we already sync grant info with client when reconnect,
2624 * grant info will be cleared for resent req, then fed_grant and
2625 * total_grant will not be modified in following preprw_write */
2626 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2627 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2628 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2631 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2632 if (repbody == NULL)
2633 GOTO(out_lock, rc = -ENOMEM);
2634 repbody->oa = body->oa;
2636 npages = PTLRPC_MAX_BRW_PAGES;
2637 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2638 objcount, ioo, remote_nb, &npages, local_nb);
2641 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2642 body->oa.o_flags & OBD_FL_SHORT_IO) {
2643 unsigned int short_io_size;
2644 unsigned char *short_io_buf;
2646 short_io_size = req_capsule_get_size(&req->rq_pill,
2649 short_io_buf = req_capsule_client_get(&req->rq_pill,
2651 CDEBUG(D_INFO, "Client use short io for data transfer,"
2652 " size = %d\n", short_io_size);
2654 /* Copy short io buf to pages */
2655 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2659 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2660 PTLRPC_BULK_GET_SINK,
2662 &ptlrpc_bulk_kiov_nopin_ops);
2664 GOTO(skip_transfer, rc = -ENOMEM);
2666 /* NB Having prepped, we must commit... */
2667 for (i = 0; i < npages; i++)
2668 desc->bd_frag_ops->add_kiov_frag(desc,
2669 local_nb[i].lnb_page,
2670 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2671 local_nb[i].lnb_len);
2673 rc = sptlrpc_svc_prep_bulk(req, desc);
2675 GOTO(skip_transfer, rc);
2677 rc = target_bulk_io(exp, desc);
2683 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2684 static int cksum_counter;
2686 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2687 cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2689 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2690 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2691 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2694 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2695 local_nb, npages, OST_WRITE,
2696 &repbody->oa.o_cksum);
2698 GOTO(out_commitrw, rc);
2702 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2703 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2704 body->oa.o_flags & OBD_FL_MMAP);
2706 tgt_warn_on_cksum(req, desc, local_nb, npages,
2708 repbody->oa.o_cksum, mmap);
2710 } else if ((cksum_counter & (-cksum_counter)) ==
2712 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2713 cksum_counter, libcfs_id2str(req->rq_peer),
2714 repbody->oa.o_cksum);
2718 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, cfs_fail_val);
2721 /* Must commit after prep above in all cases */
2722 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2723 objcount, ioo, remote_nb, npages, local_nb, rc);
2724 if (rc == -ENOTCONN)
2725 /* quota acquire process has been given up because
2726 * either the client has been evicted or the client
2727 * has timed out the request already */
2730 for (i = 0; i < niocount; i++) {
2731 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2737 * Disable sending mtime back to the client. If the client locked the
2738 * whole object, then it has already updated the mtime on its side,
2739 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2741 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2746 /* set per-requested niobuf return codes */
2747 for (i = j = 0; i < niocount; i++) {
2748 int len = remote_nb[i].rnb_len;
2753 LASSERT(j < npages);
2754 if (local_nb[j].lnb_rc < 0)
2755 rcs[i] = local_nb[j].lnb_rc;
2756 len -= local_nb[j].lnb_len;
2761 LASSERT(j == npages);
2762 ptlrpc_lprocfs_brw(req, nob);
2765 tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PW);
2767 ptlrpc_free_bulk(desc);
2769 if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2770 req->rq_no_reply = 1;
2771 /* reply out callback would free */
2772 ptlrpc_req_drop_rs(req);
2773 if (!exp->exp_obd->obd_no_transno)
2774 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2775 " client will retry: rc = %d\n",
2777 obd_uuid2str(&exp->exp_client_uuid),
2778 obd_export_nid2str(exp), rc);
2782 memalloc_noreclaim_restore(mpflags);
2786 EXPORT_SYMBOL(tgt_brw_write);
2789 * Common request handler for OST_SEEK RPC.
2791 * Unified request handling for OST_SEEK RPC.
2792 * It takes object by its FID, does needed lseek and packs result
2793 * into reply. Only SEEK_HOLE and SEEK_DATA are supported.
2795 * \param[in] tsi target session environment for this request
2797 * \retval 0 if successful
2798 * \retval negative value on error
2800 int tgt_lseek(struct tgt_session_info *tsi)
2802 struct lustre_handle lh = { 0 };
2803 struct dt_object *dob;
2804 struct ost_body *repbody;
2805 loff_t offset = tsi->tsi_ost_body->oa.o_size;
2806 int whence = tsi->tsi_ost_body->oa.o_mode;
2812 if (whence != SEEK_HOLE && whence != SEEK_DATA)
2815 /* Negative offset is prohibited on wire and must be handled on client
2816 * prior sending RPC.
2821 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
2822 if (repbody == NULL)
2824 repbody->oa = tsi->tsi_ost_body->oa;
2826 srvlock = tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
2827 tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK;
2829 rc = tgt_data_lock(tsi->tsi_env, tsi->tsi_exp, &tsi->tsi_resid,
2830 offset, OBD_OBJECT_EOF, &lh, LCK_PR);
2835 dob = dt_locate(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, &tsi->tsi_fid);
2837 GOTO(out, rc = PTR_ERR(dob));
2839 if (!dt_object_exists(dob))
2840 GOTO(obj_put, rc = -ENOENT);
2842 repbody->oa.o_size = dt_lseek(tsi->tsi_env, dob, offset, whence);
2845 dt_object_put(tsi->tsi_env, dob);
2848 tgt_data_unlock(&lh, LCK_PR);
2852 EXPORT_SYMBOL(tgt_lseek);
2854 /* Check if request can be reconstructed from saved reply data
2855 * A copy of the reply data is returned in @trd if the pointer is not NULL
2857 int req_can_reconstruct(struct ptlrpc_request *req,
2858 struct tg_reply_data *trd)
2860 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2861 struct lsd_client_data *lcd = ted->ted_lcd;
2864 if (tgt_is_multimodrpcs_client(req->rq_export))
2865 return tgt_lookup_reply(req, trd);
2867 mutex_lock(&ted->ted_lcd_lock);
2868 found = req->rq_xid == lcd->lcd_last_xid ||
2869 req->rq_xid == lcd->lcd_last_close_xid;
2871 if (found && trd != NULL) {
2872 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2873 trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2874 trd->trd_reply.lrd_transno =
2875 lcd->lcd_last_close_transno;
2876 trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2878 trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2879 trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2880 trd->trd_reply.lrd_result = lcd->lcd_last_result;
2881 trd->trd_reply.lrd_data = lcd->lcd_last_data;
2882 trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2883 trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2884 trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2885 trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2888 mutex_unlock(&ted->ted_lcd_lock);
2892 EXPORT_SYMBOL(req_can_reconstruct);