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 rc = tgt_request_preprocess(tsi, h, req);
423 /* pack reply if reply format is fixed */
424 if (rc == 0 && h->th_flags & HAS_REPLY) {
426 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
428 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
430 tsi->tsi_mdt_body->mbo_eadatasize);
431 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
433 req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
435 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
436 req_capsule_set_size(tsi->tsi_pill,
437 &RMF_ACL, RCL_SERVER,
438 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
440 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
442 struct niobuf_remote *remote_nb =
443 req_capsule_client_get(tsi->tsi_pill,
445 struct ost_body *body = tsi->tsi_ost_body;
447 req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
449 (body->oa.o_valid & OBD_MD_FLFLAGS &&
450 body->oa.o_flags & OBD_FL_SHORT_IO) ?
451 remote_nb[0].rnb_len : 0);
454 rc = req_capsule_server_pack(tsi->tsi_pill);
457 if (likely(rc == 0)) {
459 * Process request, there can be two types of rc:
460 * 1) errors with msg unpack/pack, other failures outside the
461 * operation itself. This is counted as serious errors;
462 * 2) errors during fs operation, should be placed in rq_status
466 if (!is_serious(rc) &&
467 !req->rq_no_reply && req->rq_reply_state == NULL) {
468 DEBUG_REQ(D_ERROR, req,
469 "%s: %s handler did not pack reply but returned no error",
470 tgt_name(tsi->tsi_tgt), h->th_name);
473 serious = is_serious(rc);
474 rc = clear_serious(rc);
482 * ELDLM_* codes which > 0 should be in rq_status only as well as
483 * all non-serious errors.
485 if (rc > 0 || !serious)
488 LASSERT(current->journal_info == NULL);
490 if (likely(rc == 0 && req->rq_export))
491 target_committed_to_req(req);
494 target_send_reply(req, rc, tsi->tsi_reply_fail_id);
498 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
499 struct obd_device *obd, int *process)
501 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
508 case MDS_SYNC: /* used in unmounting */
522 case MDS_HSM_PROGRESS:
523 case MDS_HSM_STATE_SET:
524 case MDS_HSM_REQUEST:
526 *process = target_queue_recovery_request(req, obd);
530 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
537 * Handle recovery. Return:
538 * +1: continue request processing;
539 * -ve: abort immediately with the given error code;
540 * 0: send reply with error code in req->rq_status;
542 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
546 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
551 case SEC_CTX_INIT_CONT:
556 if (!req->rq_export->exp_obd->obd_replayable)
559 /* sanity check: if the xid matches, the request must be marked as a
560 * resent or replayed */
561 if (req_can_reconstruct(req, NULL) == 1) {
562 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
563 (MSG_RESENT | MSG_REPLAY))) {
564 DEBUG_REQ(D_WARNING, req,
565 "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
567 lustre_msg_get_flags(req->rq_reqmsg));
568 req->rq_status = -ENOTCONN;
572 /* else: note the opposite is not always true; a RESENT req after a
573 * failover will usually not match the last_xid, since it was likely
574 * never committed. A REPLAYed request will almost never match the
575 * last xid, however it could for a committed, but still retained,
578 /* Check for aborted recovery... */
579 if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
583 DEBUG_REQ(D_INFO, req, "Got new replay");
584 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
586 if (rc != 0 || !should_process)
588 else if (should_process < 0) {
589 req->rq_status = should_process;
590 rc = ptlrpc_error(req);
597 /* Initial check for request, it is validation mostly */
598 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
600 struct tgt_handler *h;
601 struct tgt_opc_slice *s;
602 struct lu_target *tgt;
603 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
607 tgt = class_exp2tgt(req->rq_export);
608 if (unlikely(tgt == NULL)) {
609 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
610 class_exp2obd(req->rq_export)->obd_name);
611 RETURN(ERR_PTR(-EINVAL));
614 for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
615 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
618 /* opcode was not found in slice */
619 if (unlikely(s->tos_hs == NULL)) {
620 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
622 RETURN(ERR_PTR(-ENOTSUPP));
625 LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
626 h = s->tos_hs + (opc - s->tos_opc_start);
627 if (unlikely(h->th_opc == 0)) {
628 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
629 RETURN(ERR_PTR(-ENOTSUPP));
635 static int process_req_last_xid(struct ptlrpc_request *req)
639 struct obd_export *exp = req->rq_export;
640 struct tg_export_data *ted = &exp->exp_target_data;
641 bool need_lock = tgt_is_multimodrpcs_client(exp);
645 mutex_lock(&ted->ted_lcd_lock);
646 /* check request's xid is consistent with export's last_xid */
647 last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
648 if (last_xid > exp->exp_last_xid)
649 exp->exp_last_xid = last_xid;
651 if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
652 /* Some request is allowed to be sent during replay,
653 * such as OUT update requests, FLD requests, so it
654 * is possible that replay requests has smaller XID
655 * than the exp_last_xid.
657 * Some non-replay requests may have smaller XID as
660 * - Client send a no_resend RPC, like statfs;
661 * - The RPC timedout (or some other error) on client,
662 * then it's removed from the unreplied list;
663 * - Client send some other request to bump the
664 * exp_last_xid on server;
665 * - The former RPC got chance to be processed;
667 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
670 DEBUG_REQ(D_WARNING, req,
671 "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
672 req->rq_xid, exp->exp_last_xid, rc);
677 /* The "last_xid" is the minimum xid among unreplied requests,
678 * if the request is from the previous connection, its xid can
679 * still be larger than "exp_last_xid", then the above check of
680 * xid is not enough to determine whether the request is delayed.
682 * For example, if some replay request was delayed and caused
683 * timeout at client and the replay is restarted, the delayed
684 * replay request will have the larger xid than "exp_last_xid"
686 if (req->rq_export->exp_conn_cnt >
687 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
689 "Dropping request %llu from an old epoch %u/%u\n",
691 lustre_msg_get_conn_cnt(req->rq_reqmsg),
692 req->rq_export->exp_conn_cnt);
693 req->rq_no_reply = 1;
694 GOTO(out, rc = -ESTALE);
697 /* try to release in-memory reply data */
698 if (tgt_is_multimodrpcs_client(exp)) {
699 tgt_handle_received_xid(exp, last_xid);
700 rc = tgt_handle_tag(req);
705 mutex_unlock(&ted->ted_lcd_lock);
710 int tgt_request_handle(struct ptlrpc_request *req)
712 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
714 struct lustre_msg *msg = req->rq_reqmsg;
715 struct tgt_handler *h;
716 struct lu_target *tgt;
717 int request_fail_id = 0;
718 __u32 opc = lustre_msg_get_opc(msg);
719 struct obd_device *obd;
721 bool is_connect = false;
724 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
725 if (cfs_fail_val == 0 &&
726 lustre_msg_get_opc(msg) != OBD_PING &&
727 lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
730 wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
734 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
735 tsi->tsi_pill = &req->rq_pill;
736 tsi->tsi_env = req->rq_svc_thread->t_env;
738 /* if request has export then get handlers slice from corresponding
739 * target, otherwise that should be connect operation */
740 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
741 opc == MGS_CONNECT) {
743 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
744 rc = target_handle_connect(req);
746 rc = ptlrpc_error(req);
749 /* recovery-small test 18c asks to drop connect reply */
750 if (unlikely(opc == OST_CONNECT &&
751 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
755 if (unlikely(!class_connected_export(req->rq_export))) {
756 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
757 opc == SEC_CTX_FINI) {
758 /* sec context initialization has to be handled
759 * by hand in tgt_handle_request0() */
760 tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
762 GOTO(handle_recov, rc = 0);
764 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
765 opc, libcfs_id2str(req->rq_peer));
766 req->rq_status = -ENOTCONN;
767 rc = ptlrpc_error(req);
771 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
772 tsi->tsi_exp = req->rq_export;
773 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
774 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
776 tsi->tsi_jobid = NULL;
779 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
780 class_exp2obd(req->rq_export)->obd_name);
781 req->rq_status = -EINVAL;
782 rc = ptlrpc_error(req);
786 /* Skip last_xid processing for the recovery thread, otherwise, the
787 * last_xid on same request could be processed twice: first time when
788 * processing the incoming request, second time when the request is
789 * being processed by recovery thread. */
790 obd = class_exp2obd(req->rq_export);
792 /* reset the exp_last_xid on each connection. */
793 req->rq_export->exp_last_xid = 0;
794 } else if (obd->obd_recovery_data.trd_processing_task !=
796 rc = process_req_last_xid(req);
799 rc = ptlrpc_error(req);
804 request_fail_id = tgt->lut_request_fail_id;
805 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
807 h = tgt_handler_find_check(req);
809 req->rq_status = PTR_ERR(h);
810 rc = ptlrpc_error(req);
814 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
817 if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
818 CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
821 rc = lustre_msg_check_version(msg, h->th_version);
823 DEBUG_REQ(D_ERROR, req,
824 "%s: drop malformed request version=%08x expect=%08x",
825 tgt_name(tgt), lustre_msg_get_version(msg),
827 req->rq_status = -EINVAL;
828 rc = ptlrpc_error(req);
833 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
834 if (likely(rc == 1)) {
835 rc = tgt_handle_request0(tsi, h, req);
841 req_capsule_fini(tsi->tsi_pill);
842 if (tsi->tsi_corpus != NULL) {
843 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
844 tsi->tsi_corpus = NULL;
848 EXPORT_SYMBOL(tgt_request_handle);
850 /** Assign high priority operations to the request if needed. */
851 int tgt_hpreq_handler(struct ptlrpc_request *req)
853 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
854 struct tgt_handler *h;
859 if (req->rq_export == NULL)
862 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
863 tsi->tsi_pill = &req->rq_pill;
864 tsi->tsi_env = req->rq_svc_thread->t_env;
865 tsi->tsi_tgt = class_exp2tgt(req->rq_export);
866 tsi->tsi_exp = req->rq_export;
868 h = tgt_handler_find_check(req);
874 rc = tgt_request_preprocess(tsi, h, req);
875 if (unlikely(rc != 0))
878 if (h->th_hp != NULL)
882 EXPORT_SYMBOL(tgt_hpreq_handler);
884 void tgt_counter_incr(struct obd_export *exp, int opcode)
886 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
887 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
888 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
890 EXPORT_SYMBOL(tgt_counter_incr);
893 * Unified target generic handlers.
896 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
898 struct lu_target *tgt = class_exp2tgt(exp);
899 struct sptlrpc_flavor flvr;
903 LASSERT(tgt->lut_obd);
904 LASSERT(tgt->lut_slice);
906 /* always allow ECHO client */
907 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
908 LUSTRE_ECHO_NAME) == 0)) {
909 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
913 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
914 read_lock(&tgt->lut_sptlrpc_lock);
915 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
919 read_unlock(&tgt->lut_sptlrpc_lock);
921 spin_lock(&exp->exp_lock);
922 exp->exp_sp_peer = req->rq_sp_from;
923 exp->exp_flvr = flvr;
925 /* when on mgs, if no restriction is set, or if the client
926 * NID is on the local node, allow any flavor
928 if ((strcmp(exp->exp_obd->obd_type->typ_name,
929 LUSTRE_MGS_NAME) == 0) &&
930 (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
931 LNetIsPeerLocal(exp->exp_connection->c_peer.nid)))
932 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
934 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
935 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
936 CERROR("%s: unauthorized rpc flavor %x from %s, "
937 "expect %x\n", tgt_name(tgt),
939 libcfs_nid2str(req->rq_peer.nid),
940 exp->exp_flvr.sf_rpc);
943 spin_unlock(&exp->exp_lock);
945 if (exp->exp_sp_peer != req->rq_sp_from) {
946 CERROR("%s: RPC source %s doesn't match %s\n",
948 sptlrpc_part2name(req->rq_sp_from),
949 sptlrpc_part2name(exp->exp_sp_peer));
952 rc = sptlrpc_target_export_check(exp, req);
959 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
961 struct sptlrpc_rule_set tmp_rset;
964 if (unlikely(tgt == NULL)) {
965 CERROR("No target passed");
969 sptlrpc_rule_set_init(&tmp_rset);
970 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
972 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
977 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
979 write_lock(&tgt->lut_sptlrpc_lock);
980 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
981 tgt->lut_sptlrpc_rset = tmp_rset;
982 write_unlock(&tgt->lut_sptlrpc_lock);
986 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
988 int tgt_connect(struct tgt_session_info *tsi)
990 struct ptlrpc_request *req = tgt_ses_req(tsi);
991 struct obd_connect_data *reply;
996 /* XXX: better to call this check right after getting new export but
997 * before last_rcvd slot allocation to avoid server load upon insecure
998 * connects. This is to be fixed after unifiyng all targets.
1000 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
1004 /* To avoid exposing partially initialized connection flags, changes up
1005 * to this point have been staged in reply->ocd_connect_flags. Now that
1006 * connection handling has completed successfully, atomically update
1007 * the connect flags in the shared export data structure. LU-1623 */
1008 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1009 spin_lock(&tsi->tsi_exp->exp_lock);
1010 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1011 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1012 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1013 reply->ocd_connect_flags2;
1014 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1015 spin_unlock(&tsi->tsi_exp->exp_lock);
1017 if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1018 LUSTRE_MDT_NAME) == 0) {
1019 rc = req_check_sepol(tsi->tsi_pill);
1023 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2 &&
1024 reply->ocd_connect_flags2 & OBD_CONNECT2_ENCRYPT &&
1025 tsi->tsi_pill->rc_req->rq_export) {
1026 bool forbid_encrypt = true;
1027 struct lu_nodemap *nm =
1028 nodemap_get_from_exp(tsi->tsi_pill->rc_req->rq_export);
1031 /* nodemap_get_from_exp returns NULL in case
1032 * nodemap is not active, so we do not forbid
1034 forbid_encrypt = false;
1035 } else if (!IS_ERR(nm)) {
1036 forbid_encrypt = nm->nmf_forbid_encryption;
1041 GOTO(out, rc = -EACCES);
1047 obd_disconnect(class_export_get(tsi->tsi_exp));
1050 EXPORT_SYMBOL(tgt_connect);
1052 int tgt_disconnect(struct tgt_session_info *tsi)
1058 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1060 rc = target_handle_disconnect(tgt_ses_req(tsi));
1062 RETURN(err_serious(rc));
1066 EXPORT_SYMBOL(tgt_disconnect);
1069 * Unified target OBD handlers
1071 int tgt_obd_ping(struct tgt_session_info *tsi)
1077 /* The target-specific part of OBD_PING request handling.
1078 * It controls Filter Modification Data (FMD) expiration each time
1081 * Valid only for replayable targets, e.g. MDT and OFD
1083 if (tsi->tsi_exp->exp_obd->obd_replayable)
1084 tgt_fmd_expire(tsi->tsi_exp);
1086 rc = req_capsule_server_pack(tsi->tsi_pill);
1088 RETURN(err_serious(rc));
1092 EXPORT_SYMBOL(tgt_obd_ping);
1094 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1096 return err_serious(-EOPNOTSUPP);
1099 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1101 struct ptlrpc_request *req = tgt_ses_req(tsi);
1102 struct obd_export *exp = req->rq_export;
1103 struct ptlrpc_bulk_desc *desc;
1110 for (i = 0; i < rdbuf->rb_nbufs; i++) {
1111 unsigned int offset;
1113 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1114 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1118 desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1119 PTLRPC_BULK_PUT_SOURCE,
1121 &ptlrpc_bulk_kiov_nopin_ops);
1125 for (i = 0; i < rdbuf->rb_nbufs; i++)
1126 desc->bd_frag_ops->add_iov_frag(desc,
1127 rdbuf->rb_bufs[i].lb_buf,
1128 rdbuf->rb_bufs[i].lb_len);
1130 rc = target_bulk_io(exp, desc);
1131 ptlrpc_free_bulk(desc);
1134 EXPORT_SYMBOL(tgt_send_buffer);
1136 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1138 struct ptlrpc_request *req = tgt_ses_req(tsi);
1139 struct obd_export *exp = req->rq_export;
1140 struct ptlrpc_bulk_desc *desc;
1148 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1149 PTLRPC_BULK_PUT_SOURCE,
1151 &ptlrpc_bulk_kiov_pin_ops);
1155 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1156 /* old client requires reply size in it's PAGE_SIZE,
1157 * which is rdpg->rp_count */
1158 nob = rdpg->rp_count;
1160 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1161 i++, tmpcount -= tmpsize) {
1162 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1163 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1167 LASSERT(desc->bd_nob == nob);
1168 rc = target_bulk_io(exp, desc);
1169 ptlrpc_free_bulk(desc);
1172 EXPORT_SYMBOL(tgt_sendpage);
1175 * OBD_IDX_READ handler
1177 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1179 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1180 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1181 struct idx_info *req_ii, *rep_ii;
1186 memset(rdpg, 0, sizeof(*rdpg));
1187 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1189 /* extract idx_info buffer from request & reply */
1190 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1191 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1192 RETURN(err_serious(-EPROTO));
1194 rc = req_capsule_server_pack(tsi->tsi_pill);
1196 RETURN(err_serious(rc));
1198 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1200 RETURN(err_serious(-EFAULT));
1201 rep_ii->ii_magic = IDX_INFO_MAGIC;
1203 /* extract hash to start with */
1204 rdpg->rp_hash = req_ii->ii_hash_start;
1206 /* extract requested attributes */
1207 rdpg->rp_attrs = req_ii->ii_attrs;
1209 /* check that fid packed in request is valid and supported */
1210 if (!fid_is_sane(&req_ii->ii_fid))
1212 rep_ii->ii_fid = req_ii->ii_fid;
1215 rep_ii->ii_flags = req_ii->ii_flags;
1217 /* compute number of pages to allocate, ii_count is the number of 4KB
1219 if (req_ii->ii_count <= 0)
1220 GOTO(out, rc = -EFAULT);
1221 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1222 exp_max_brw_size(tsi->tsi_exp));
1223 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1225 /* allocate pages to store the containers */
1226 OBD_ALLOC_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1227 if (rdpg->rp_pages == NULL)
1228 GOTO(out, rc = -ENOMEM);
1229 for (i = 0; i < rdpg->rp_npages; i++) {
1230 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1231 if (rdpg->rp_pages[i] == NULL)
1232 GOTO(out, rc = -ENOMEM);
1235 /* populate pages with key/record pairs */
1236 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1240 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1241 "asked %d > %d\n", rc, rdpg->rp_count);
1243 /* send pages to client */
1244 rc = tgt_sendpage(tsi, rdpg, rc);
1249 if (rdpg->rp_pages) {
1250 for (i = 0; i < rdpg->rp_npages; i++)
1251 if (rdpg->rp_pages[i])
1252 __free_page(rdpg->rp_pages[i]);
1253 OBD_FREE_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1258 struct tgt_handler tgt_obd_handlers[] = {
1259 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1260 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1262 EXPORT_SYMBOL(tgt_obd_handlers);
1264 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1265 struct dt_object *obj, __u64 start, __u64 end)
1271 /* if no objid is specified, it means "sync whole filesystem" */
1273 rc = dt_sync(env, tgt->lut_bottom);
1274 } else if (dt_version_get(env, obj) >
1275 tgt->lut_obd->obd_last_committed) {
1276 rc = dt_object_sync(env, obj, start, end);
1278 atomic_inc(&tgt->lut_sync_count);
1282 EXPORT_SYMBOL(tgt_sync);
1284 * Unified target DLM handlers.
1288 * Unified target BAST
1290 * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1291 * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1292 * but if it's cross-MDT lock, because remote object version is not set, a
1293 * filesystem sync is needed.
1295 * \param lock server side lock
1296 * \param desc lock desc
1297 * \param data ldlm_cb_set_arg
1298 * \param flag indicates whether this cancelling or blocking callback
1299 * \retval 0 on success
1300 * \retval negative number on error
1302 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1303 void *data, int flag)
1306 struct lu_target *tgt;
1307 struct dt_object *obj = NULL;
1313 tgt = class_exp2tgt(lock->l_export);
1315 if (unlikely(tgt == NULL)) {
1316 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1317 class_exp2obd(lock->l_export)->obd_name);
1321 if (flag == LDLM_CB_CANCELING &&
1322 (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1323 (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1324 (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1325 ldlm_is_cbpending(lock))) &&
1326 ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1327 lock->l_resource->lr_type == LDLM_EXTENT)) {
1329 __u64 end = OBD_OBJECT_EOF;
1331 rc = lu_env_init(&env, LCT_DT_THREAD);
1332 if (unlikely(rc != 0))
1335 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1336 tgt->lut_lsd.lsd_osd_index);
1338 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1339 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1341 GOTO(err_env, rc = PTR_ERR(obj));
1343 if (!dt_object_exists(obj))
1344 GOTO(err_put, rc = -ENOENT);
1346 start = lock->l_policy_data.l_extent.start;
1347 end = lock->l_policy_data.l_extent.end;
1350 rc = tgt_sync(&env, tgt, obj, start, end);
1352 CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1353 "cancel: rc = %d\n",
1354 tgt_name(tgt), PFID(&fid),
1355 lock->l_policy_data.l_extent.start,
1356 lock->l_policy_data.l_extent.end, rc);
1360 dt_object_put(&env, obj);
1365 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1369 static struct ldlm_callback_suite tgt_dlm_cbs = {
1370 .lcs_completion = ldlm_server_completion_ast,
1371 .lcs_blocking = tgt_blocking_ast,
1372 .lcs_glimpse = ldlm_server_glimpse_ast
1375 int tgt_enqueue(struct tgt_session_info *tsi)
1377 struct ptlrpc_request *req = tgt_ses_req(tsi);
1382 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1383 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1385 LASSERT(tsi->tsi_dlm_req != NULL);
1386 rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1387 tsi->tsi_dlm_req, &tgt_dlm_cbs);
1389 RETURN(err_serious(rc));
1391 switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1393 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1396 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1399 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1402 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1405 RETURN(req->rq_status);
1407 EXPORT_SYMBOL(tgt_enqueue);
1409 int tgt_convert(struct tgt_session_info *tsi)
1411 struct ptlrpc_request *req = tgt_ses_req(tsi);
1415 LASSERT(tsi->tsi_dlm_req);
1416 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1418 RETURN(err_serious(rc));
1420 RETURN(req->rq_status);
1423 int tgt_bl_callback(struct tgt_session_info *tsi)
1425 return err_serious(-EOPNOTSUPP);
1428 int tgt_cp_callback(struct tgt_session_info *tsi)
1430 return err_serious(-EOPNOTSUPP);
1433 /* generic LDLM target handler */
1434 struct tgt_handler tgt_dlm_handlers[] = {
1435 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1436 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1437 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1438 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1440 EXPORT_SYMBOL(tgt_dlm_handlers);
1443 * Unified target LLOG handlers.
1445 int tgt_llog_open(struct tgt_session_info *tsi)
1451 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1455 EXPORT_SYMBOL(tgt_llog_open);
1457 int tgt_llog_read_header(struct tgt_session_info *tsi)
1463 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1467 EXPORT_SYMBOL(tgt_llog_read_header);
1469 int tgt_llog_next_block(struct tgt_session_info *tsi)
1475 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1479 EXPORT_SYMBOL(tgt_llog_next_block);
1481 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1487 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1491 EXPORT_SYMBOL(tgt_llog_prev_block);
1493 /* generic llog target handler */
1494 struct tgt_handler tgt_llog_handlers[] = {
1495 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1496 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1497 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1498 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1500 EXPORT_SYMBOL(tgt_llog_handlers);
1503 * sec context handlers
1505 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1506 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1511 struct tgt_handler tgt_sec_ctx_handlers[] = {
1512 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle),
1513 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle),
1514 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle),
1516 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1518 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1519 struct dt_device *key,
1520 struct lfsck_req_local *lrl,
1521 struct thandle *th) = NULL;
1523 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1525 struct lfsck_req_local *,
1528 tgt_lfsck_in_notify_local = notify;
1530 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1532 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1533 struct dt_device *key,
1534 struct lfsck_request *lr) = NULL;
1536 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1538 struct lfsck_request *))
1540 tgt_lfsck_in_notify = notify;
1542 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1544 static int (*tgt_lfsck_query)(const struct lu_env *env,
1545 struct dt_device *key,
1546 struct lfsck_request *req,
1547 struct lfsck_reply *rep,
1548 struct lfsck_query *que) = NULL;
1550 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1552 struct lfsck_request *,
1553 struct lfsck_reply *,
1554 struct lfsck_query *))
1556 tgt_lfsck_query = query;
1558 EXPORT_SYMBOL(tgt_register_lfsck_query);
1560 /* LFSCK request handlers */
1561 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1563 const struct lu_env *env = tsi->tsi_env;
1564 struct dt_device *key = tsi->tsi_tgt->lut_bottom;
1565 struct lfsck_request *lr;
1569 lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1573 rc = tgt_lfsck_in_notify(env, key, lr);
1578 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1580 struct lfsck_request *request;
1581 struct lfsck_reply *reply;
1585 request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1586 if (request == NULL)
1589 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1593 rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1594 request, reply, NULL);
1596 RETURN(rc < 0 ? rc : 0);
1599 struct tgt_handler tgt_lfsck_handlers[] = {
1600 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_NOTIFY, tgt_handle_lfsck_notify),
1601 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_QUERY, tgt_handle_lfsck_query),
1603 EXPORT_SYMBOL(tgt_lfsck_handlers);
1606 * initialize per-thread page pool (bug 5137).
1608 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1610 struct tgt_thread_big_cache *tbc;
1614 LASSERT(thread != NULL);
1615 LASSERT(thread->t_data == NULL);
1617 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1620 thread->t_data = tbc;
1623 EXPORT_SYMBOL(tgt_io_thread_init);
1626 * free per-thread pool created by tgt_thread_init().
1628 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1630 struct tgt_thread_big_cache *tbc;
1634 LASSERT(thread != NULL);
1637 * be prepared to handle partially-initialized pools (because this is
1638 * called from ost_io_thread_init() for cleanup.
1640 tbc = thread->t_data;
1642 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1643 thread->t_data = NULL;
1647 EXPORT_SYMBOL(tgt_io_thread_done);
1650 * Helper function for getting Data-on-MDT file server DLM lock
1651 * if asked by client.
1653 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1654 struct lustre_handle *lh, int mode, __u64 *flags)
1656 union ldlm_policy_data policy = {
1657 .l_inodebits.bits = MDS_INODELOCK_DOM,
1663 LASSERT(lh != NULL);
1664 LASSERT(ns != NULL);
1665 LASSERT(!lustre_handle_is_used(lh));
1667 rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1668 flags, ldlm_blocking_ast,
1669 ldlm_completion_ast, ldlm_glimpse_ast,
1670 NULL, 0, LVB_T_NONE, NULL, lh);
1672 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1674 EXPORT_SYMBOL(tgt_mdt_data_lock);
1676 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1678 LASSERT(lustre_handle_is_used(lh));
1679 ldlm_lock_decref(lh, mode);
1681 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1684 * Helper function for getting server side [start, start+count] DLM lock
1685 * if asked by client.
1687 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1688 struct ldlm_res_id *res_id, __u64 start, __u64 end,
1689 struct lustre_handle *lh, int mode, __u64 *flags)
1691 union ldlm_policy_data policy;
1696 LASSERT(lh != NULL);
1697 LASSERT(ns != NULL);
1698 LASSERT(!lustre_handle_is_used(lh));
1700 policy.l_extent.gid = 0;
1701 policy.l_extent.start = start & PAGE_MASK;
1704 * If ->o_blocks is EOF it means "lock till the end of the file".
1705 * Otherwise, it's size of an extent or hole being punched (in bytes).
1707 if (end == OBD_OBJECT_EOF || end < start)
1708 policy.l_extent.end = OBD_OBJECT_EOF;
1710 policy.l_extent.end = end | ~PAGE_MASK;
1712 rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1713 mode, flags, ldlm_blocking_ast,
1714 ldlm_completion_ast, ldlm_glimpse_ast,
1715 NULL, 0, LVB_T_NONE, NULL, lh);
1716 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1718 EXPORT_SYMBOL(tgt_extent_lock);
1720 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1722 LASSERT(lustre_handle_is_used(lh));
1723 ldlm_lock_decref(lh, mode);
1725 EXPORT_SYMBOL(tgt_extent_unlock);
1727 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1728 struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1729 struct niobuf_remote *nb, struct lustre_handle *lh,
1730 enum ldlm_mode mode)
1732 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1734 int nrbufs = obj->ioo_bufcnt;
1740 LASSERT(mode == LCK_PR || mode == LCK_PW);
1741 LASSERT(!lustre_handle_is_used(lh));
1743 if (ns->ns_obd->obd_recovering)
1746 if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1749 for (i = 1; i < nrbufs; i++)
1750 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1753 /* MDT IO for data-on-mdt */
1754 if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1755 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1757 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1758 nb[nrbufs - 1].rnb_offset +
1759 nb[nrbufs - 1].rnb_len - 1,
1764 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1765 struct lustre_handle *lh, enum ldlm_mode mode)
1769 LASSERT(mode == LCK_PR || mode == LCK_PW);
1770 LASSERT((obj->ioo_bufcnt > 0 &&
1771 (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1772 lustre_handle_is_used(lh));
1774 if (lustre_handle_is_used(lh))
1775 tgt_extent_unlock(lh, mode);
1778 static int tgt_checksum_niobuf(struct lu_target *tgt,
1779 struct niobuf_local *local_nb, int npages,
1780 int opc, enum cksum_types cksum_type,
1783 struct ahash_request *req;
1784 unsigned int bufsize;
1786 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1788 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1790 CERROR("%s: unable to initialize checksum hash %s\n",
1791 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1792 return PTR_ERR(req);
1795 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1796 for (i = 0; i < npages; i++) {
1797 /* corrupt the data before we compute the checksum, to
1798 * simulate a client->OST data error */
1799 if (i == 0 && opc == OST_WRITE &&
1800 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1801 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1802 int len = local_nb[i].lnb_len;
1803 struct page *np = tgt_page_to_corrupt;
1806 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1807 char *ptr2 = page_address(np);
1809 memcpy(ptr2 + off, ptr + off, len);
1810 memcpy(ptr2 + off, "bad3", min(4, len));
1813 /* LU-8376 to preserve original index for
1814 * display in dump_all_bulk_pages() */
1817 cfs_crypto_hash_update_page(req, np, off,
1821 CERROR("%s: can't alloc page for corruption\n",
1825 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1826 local_nb[i].lnb_page_offset & ~PAGE_MASK,
1827 local_nb[i].lnb_len);
1829 /* corrupt the data after we compute the checksum, to
1830 * simulate an OST->client data error */
1831 if (i == 0 && opc == OST_READ &&
1832 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1833 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1834 int len = local_nb[i].lnb_len;
1835 struct page *np = tgt_page_to_corrupt;
1838 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1839 char *ptr2 = page_address(np);
1841 memcpy(ptr2 + off, ptr + off, len);
1842 memcpy(ptr2 + off, "bad4", min(4, len));
1845 /* LU-8376 to preserve original index for
1846 * display in dump_all_bulk_pages() */
1849 cfs_crypto_hash_update_page(req, np, off,
1853 CERROR("%s: can't alloc page for corruption\n",
1859 bufsize = sizeof(*cksum);
1860 err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1865 char dbgcksum_file_name[PATH_MAX];
1867 static void dump_all_bulk_pages(struct obdo *oa, int count,
1868 struct niobuf_local *local_nb,
1869 __u32 server_cksum, __u32 client_cksum)
1876 /* will only keep dump of pages on first error for the same range in
1877 * file/fid, not during the resends/retries. */
1878 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1879 "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1880 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1881 libcfs_debug_file_path_arr :
1882 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1883 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1884 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1885 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1886 local_nb[0].lnb_file_offset,
1887 local_nb[count-1].lnb_file_offset +
1888 local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1889 filp = filp_open(dbgcksum_file_name,
1890 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1894 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1895 "checksum error: rc = %d\n", dbgcksum_file_name,
1898 CERROR("%s: can't open to dump pages with checksum "
1899 "error: rc = %d\n", dbgcksum_file_name, rc);
1903 for (i = 0; i < count; i++) {
1904 len = local_nb[i].lnb_len;
1905 buf = kmap(local_nb[i].lnb_page);
1907 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1909 CERROR("%s: wanted to write %u but got %d "
1910 "error\n", dbgcksum_file_name, len, rc);
1915 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1916 dbgcksum_file_name, rc);
1918 kunmap(local_nb[i].lnb_page);
1921 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1923 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1924 filp_close(filp, NULL);
1927 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1928 struct obd_export *exp, struct obdo *oa,
1929 const struct lnet_process_id *peer,
1930 __u32 client_cksum, __u32 server_cksum,
1931 enum cksum_types server_cksum_type)
1934 enum cksum_types cksum_type;
1937 /* unlikely to happen and only if resend does not occur due to cksum
1938 * control failure on Client */
1939 if (unlikely(server_cksum == client_cksum)) {
1940 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1945 if (exp->exp_obd->obd_checksum_dump)
1946 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1949 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1952 if (cksum_type != server_cksum_type)
1953 msg = "the server may have not used the checksum type specified"
1954 " in the original request - likely a protocol problem";
1956 msg = "should have changed on the client or in transit";
1958 start = local_nb[0].lnb_file_offset;
1959 end = local_nb[npages-1].lnb_file_offset +
1960 local_nb[npages-1].lnb_len - 1;
1962 LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1963 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1964 " %x (type %x), server csum %x (type %x)\n",
1965 exp->exp_obd->obd_name,
1966 msg, libcfs_nid2str(peer->nid),
1967 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1968 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1969 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1971 start, end, client_cksum, cksum_type, server_cksum,
1977 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1978 unsigned char *buf, int size)
1980 int i, off, len, copied = size;
1983 for (i = 0; i < npages; i++) {
1984 off = local[i].lnb_page_offset & ~PAGE_MASK;
1985 len = local[i].lnb_len;
1987 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1992 ptr = kmap_atomic(local[i].lnb_page);
1993 memcpy(buf, ptr + off, len);
1998 return copied - size;
2001 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
2002 struct niobuf_local *local_nb,
2003 int npages, int opc,
2004 obd_dif_csum_fn *fn,
2008 enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
2009 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
2010 const char *obd_name = tgt->lut_obd->obd_name;
2011 struct ahash_request *req;
2012 unsigned int bufsize;
2013 unsigned char *buffer;
2014 struct page *__page;
2017 int used_number = 0;
2023 __page = alloc_page(GFP_KERNEL);
2027 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
2029 CERROR("%s: unable to initialize checksum hash %s\n",
2030 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
2031 return PTR_ERR(req);
2034 buffer = kmap(__page);
2035 guard_start = (__u16 *)buffer;
2036 guard_number = PAGE_SIZE / sizeof(*guard_start);
2037 for (i = 0; i < npages; i++) {
2038 /* corrupt the data before we compute the checksum, to
2039 * simulate a client->OST data error */
2040 if (i == 0 && opc == OST_WRITE &&
2041 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2042 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2043 int len = local_nb[i].lnb_len;
2044 struct page *np = tgt_page_to_corrupt;
2047 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2048 char *ptr2 = page_address(np);
2050 memcpy(ptr2 + off, ptr + off, len);
2051 memcpy(ptr2 + off, "bad3", min(4, len));
2054 /* LU-8376 to preserve original index for
2055 * display in dump_all_bulk_pages() */
2058 cfs_crypto_hash_update_page(req, np, off,
2062 CERROR("%s: can't alloc page for corruption\n",
2068 * The left guard number should be able to hold checksums of a
2071 if (t10_cksum_type && opc == OST_READ &&
2072 local_nb[i].lnb_guard_disk) {
2073 used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2074 if (used > (guard_number - used_number)) {
2078 memcpy(guard_start + used_number,
2079 local_nb[i].lnb_guards,
2080 used * sizeof(*local_nb[i].lnb_guards));
2082 rc = obd_page_dif_generate_buffer(obd_name,
2083 local_nb[i].lnb_page,
2084 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2085 local_nb[i].lnb_len, guard_start + used_number,
2086 guard_number - used_number, &used, sector_size,
2092 LASSERT(used <= MAX_GUARD_NUMBER);
2094 * If disk support T10PI checksum, copy guards to local_nb.
2095 * If the write is partial page, do not use the guards for bio
2096 * submission since the data might not be full-sector. The bio
2097 * guards will be generated later based on the full sectors. If
2098 * the sector size is 512B rather than 4 KB, or the page size
2099 * is larger than 4KB, this might drop some useful guards for
2100 * partial page write, but it will only add minimal extra time
2101 * of checksum calculation.
2103 if (t10_cksum_type && opc == OST_WRITE &&
2104 local_nb[i].lnb_len == PAGE_SIZE) {
2105 local_nb[i].lnb_guard_rpc = 1;
2106 memcpy(local_nb[i].lnb_guards,
2107 guard_start + used_number,
2108 used * sizeof(*local_nb[i].lnb_guards));
2111 used_number += used;
2112 if (used_number == guard_number) {
2113 cfs_crypto_hash_update_page(req, __page, 0,
2114 used_number * sizeof(*guard_start));
2118 /* corrupt the data after we compute the checksum, to
2119 * simulate an OST->client data error */
2120 if (unlikely(i == 0 && opc == OST_READ &&
2121 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2122 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2123 int len = local_nb[i].lnb_len;
2124 struct page *np = tgt_page_to_corrupt;
2127 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2128 char *ptr2 = page_address(np);
2130 memcpy(ptr2 + off, ptr + off, len);
2131 memcpy(ptr2 + off, "bad4", min(4, len));
2134 /* LU-8376 to preserve original index for
2135 * display in dump_all_bulk_pages() */
2138 cfs_crypto_hash_update_page(req, np, off,
2142 CERROR("%s: can't alloc page for corruption\n",
2151 if (used_number != 0)
2152 cfs_crypto_hash_update_page(req, __page, 0,
2153 used_number * sizeof(*guard_start));
2155 bufsize = sizeof(cksum);
2156 rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2161 __free_page(__page);
2165 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2166 enum cksum_types cksum_type,
2167 struct niobuf_local *local_nb,
2168 int npages, int opc, u32 *check_sum)
2170 obd_dif_csum_fn *fn = NULL;
2171 int sector_size = 0;
2175 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
2178 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2179 opc, fn, sector_size,
2182 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2183 cksum_type, check_sum);
2187 int tgt_brw_read(struct tgt_session_info *tsi)
2189 struct ptlrpc_request *req = tgt_ses_req(tsi);
2190 struct ptlrpc_bulk_desc *desc = NULL;
2191 struct obd_export *exp = tsi->tsi_exp;
2192 struct niobuf_remote *remote_nb;
2193 struct niobuf_local *local_nb;
2194 struct obd_ioobj *ioo;
2195 struct ost_body *body, *repbody;
2196 struct lustre_handle lockh = { 0 };
2197 int npages, nob = 0, rc, i, no_reply = 0,
2199 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2200 const char *obd_name = exp->exp_obd->obd_name;
2204 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2205 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2206 CERROR("%s: deny read request from %s to portal %u\n",
2207 tgt_name(tsi->tsi_tgt),
2208 obd_export_nid2str(req->rq_export),
2209 ptlrpc_req2svc(req)->srv_req_portal);
2213 req->rq_bulk_read = 1;
2215 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2218 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2219 cfs_fail_val : (obd_timeout + 1) / 4);
2221 /* Check if there is eviction in progress, and if so, wait for it to
2223 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2224 /* We do not care how long it takes */
2226 exp->exp_obd->obd_evict_inprogress_waitq,
2227 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2230 /* There must be big cache in current thread to process this request
2231 * if it is NULL then something went wrong and it wasn't allocated,
2232 * report -ENOMEM in that case */
2236 body = tsi->tsi_ost_body;
2237 LASSERT(body != NULL);
2239 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2240 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2242 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2243 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2245 local_nb = tbc->local;
2247 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2253 * If getting the lock took more time than
2254 * client was willing to wait, drop it. b=11330
2256 if (ktime_get_real_seconds() > req->rq_deadline ||
2257 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2259 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2260 libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2261 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2262 req->rq_deadline - req->rq_arrival_time.tv_sec);
2263 GOTO(out_lock, rc = -ETIMEDOUT);
2266 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2267 repbody->oa = body->oa;
2269 npages = PTLRPC_MAX_BRW_PAGES;
2270 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2271 ioo, remote_nb, &npages, local_nb);
2275 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2276 body->oa.o_flags & OBD_FL_SHORT_IO) {
2279 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2280 PTLRPC_BULK_PUT_SOURCE,
2282 &ptlrpc_bulk_kiov_nopin_ops);
2284 GOTO(out_commitrw, rc = -ENOMEM);
2288 npages_read = npages;
2289 for (i = 0; i < npages; i++) {
2290 int page_rc = local_nb[i].lnb_rc;
2299 if (page_rc != 0 && desc != NULL) { /* some data! */
2300 LASSERT(local_nb[i].lnb_page != NULL);
2301 desc->bd_frag_ops->add_kiov_frag
2302 (desc, local_nb[i].lnb_page,
2303 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2307 if (page_rc != local_nb[i].lnb_len) { /* short read */
2308 local_nb[i].lnb_len = page_rc;
2309 npages_read = i + (page_rc != 0 ? 1 : 0);
2310 /* All subsequent pages should be 0 */
2311 while (++i < npages)
2312 LASSERT(local_nb[i].lnb_rc == 0);
2317 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2318 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2319 body->oa.o_flags : 0;
2320 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2322 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2324 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2326 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2327 local_nb, npages_read, OST_READ,
2328 &repbody->oa.o_cksum);
2330 GOTO(out_commitrw, rc);
2331 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2332 repbody->oa.o_cksum);
2334 /* if a resend it could be for a cksum error, so check Server
2335 * cksum with returned Client cksum (this should even cover
2336 * zero-cksum case) */
2337 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2338 (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2339 check_read_checksum(local_nb, npages_read, exp,
2340 &body->oa, &req->rq_peer,
2342 repbody->oa.o_cksum, cksum_type);
2344 repbody->oa.o_valid = 0;
2346 /* We're finishing using body->oa as an input variable */
2348 /* Check if client was evicted while we were doing i/o before touching
2351 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2352 body->oa.o_flags & OBD_FL_SHORT_IO) {
2353 unsigned char *short_io_buf;
2356 short_io_buf = req_capsule_server_get(&req->rq_pill,
2358 short_io_size = req_capsule_get_size(&req->rq_pill,
2361 rc = tgt_pages2shortio(local_nb, npages_read,
2362 short_io_buf, short_io_size);
2364 req_capsule_shrink(&req->rq_pill,
2367 rc = rc > 0 ? 0 : rc;
2368 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2369 rc = target_bulk_io(exp, desc);
2373 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2374 body->oa.o_flags & OBD_FL_SHORT_IO)
2375 req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2380 /* Must commit after prep above in all cases */
2381 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2382 remote_nb, npages, local_nb, rc);
2384 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2386 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2387 ptlrpc_free_bulk(desc);
2392 ptlrpc_lprocfs_brw(req, nob);
2393 } else if (no_reply) {
2394 req->rq_no_reply = 1;
2395 /* reply out callback would free */
2396 ptlrpc_req_drop_rs(req);
2397 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2398 "client will retry: rc %d\n",
2400 obd_uuid2str(&exp->exp_client_uuid),
2401 obd_export_nid2str(exp), rc);
2403 /* send a bulk after reply to simulate a network delay or reordering
2404 * by a router - Note that !desc implies short io, so there is no bulk
2406 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2408 CDEBUG(D_INFO, "reorder BULK\n");
2411 target_bulk_io(exp, desc);
2412 ptlrpc_free_bulk(desc);
2417 EXPORT_SYMBOL(tgt_brw_read);
2419 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2420 unsigned char *buf, unsigned int size)
2425 for (i = 0; i < npages; i++) {
2426 off = local[i].lnb_page_offset & ~PAGE_MASK;
2427 len = local[i].lnb_len;
2432 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2434 ptr = kmap_atomic(local[i].lnb_page);
2437 memcpy(ptr + off, buf, len < size ? len : size);
2445 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2446 struct ptlrpc_bulk_desc *desc,
2447 struct niobuf_local *local_nb, int npages,
2448 u32 client_cksum, u32 server_cksum,
2451 struct obd_export *exp = req->rq_export;
2452 struct ost_body *body;
2456 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2457 LASSERT(body != NULL);
2459 if (desc && req->rq_peer.nid != desc->bd_sender) {
2461 router = libcfs_nid2str(desc->bd_sender);
2464 if (exp->exp_obd->obd_checksum_dump)
2465 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2469 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2470 client_cksum, server_cksum);
2474 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2475 DFID" object "DOSTID" extent [%llu-%llu"
2476 "]: client csum %x, server csum %x\n",
2477 exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2479 body->oa.o_valid & OBD_MD_FLFID ?
2480 body->oa.o_parent_seq : (__u64)0,
2481 body->oa.o_valid & OBD_MD_FLFID ?
2482 body->oa.o_parent_oid : 0,
2483 body->oa.o_valid & OBD_MD_FLFID ?
2484 body->oa.o_parent_ver : 0,
2485 POSTID(&body->oa.o_oi),
2486 local_nb[0].lnb_file_offset,
2487 local_nb[npages-1].lnb_file_offset +
2488 local_nb[npages - 1].lnb_len - 1,
2489 client_cksum, server_cksum);
2492 int tgt_brw_write(struct tgt_session_info *tsi)
2494 struct ptlrpc_request *req = tgt_ses_req(tsi);
2495 struct ptlrpc_bulk_desc *desc = NULL;
2496 struct obd_export *exp = req->rq_export;
2497 struct niobuf_remote *remote_nb;
2498 struct niobuf_local *local_nb;
2499 struct obd_ioobj *ioo;
2500 struct ost_body *body, *repbody;
2501 struct lustre_handle lockh = {0};
2503 int objcount, niocount, npages;
2505 enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2506 bool no_reply = false, mmap;
2507 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2508 bool wait_sync = false;
2509 const char *obd_name = exp->exp_obd->obd_name;
2510 /* '1' for consistency with code that checks !mpflag to restore */
2511 unsigned int mpflags = 1;
2515 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2516 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2517 CERROR("%s: deny write request from %s to portal %u\n",
2518 tgt_name(tsi->tsi_tgt),
2519 obd_export_nid2str(req->rq_export),
2520 ptlrpc_req2svc(req)->srv_req_portal);
2521 RETURN(err_serious(-EPROTO));
2524 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2525 RETURN(err_serious(-ENOSPC));
2526 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2527 RETURN(err_serious(-EROFS));
2529 req->rq_bulk_write = 1;
2531 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2532 RETURN(err_serious(-EIO));
2533 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2534 RETURN(err_serious(-EFAULT));
2536 /* pause before transaction has been started */
2537 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2538 cfs_fail_val : (obd_timeout + 1) / 4);
2540 /* Delay write commit to show stale size information */
2541 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2543 /* There must be big cache in current thread to process this request
2544 * if it is NULL then something went wrong and it wasn't allocated,
2545 * report -ENOMEM in that case */
2549 body = tsi->tsi_ost_body;
2550 LASSERT(body != NULL);
2552 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2553 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2555 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2556 RCL_CLIENT) / sizeof(*ioo);
2558 for (niocount = i = 0; i < objcount; i++)
2559 niocount += ioo[i].ioo_bufcnt;
2561 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2562 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2563 if (niocount != req_capsule_get_size(&req->rq_pill,
2564 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2566 RETURN(err_serious(-EPROTO));
2568 if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2569 ptlrpc_connection_is_local(exp->exp_connection))
2570 mpflags = memalloc_noreclaim_save();
2572 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2573 niocount * sizeof(*rcs));
2574 rc = req_capsule_server_pack(&req->rq_pill);
2576 GOTO(out, rc = err_serious(rc));
2578 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2579 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2581 local_nb = tbc->local;
2583 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2589 * If getting the lock took more time than
2590 * client was willing to wait, drop it. b=11330
2592 if (ktime_get_real_seconds() > req->rq_deadline ||
2593 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2595 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2596 tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2597 POSTID(&ioo->ioo_oid),
2598 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2599 req->rq_deadline - req->rq_arrival_time.tv_sec);
2600 GOTO(out_lock, rc = -ETIMEDOUT);
2603 /* Because we already sync grant info with client when reconnect,
2604 * grant info will be cleared for resent req, then fed_grant and
2605 * total_grant will not be modified in following preprw_write */
2606 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2607 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2608 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2611 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2612 if (repbody == NULL)
2613 GOTO(out_lock, rc = -ENOMEM);
2614 repbody->oa = body->oa;
2616 npages = PTLRPC_MAX_BRW_PAGES;
2617 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2618 objcount, ioo, remote_nb, &npages, local_nb);
2621 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2622 body->oa.o_flags & OBD_FL_SHORT_IO) {
2623 unsigned int short_io_size;
2624 unsigned char *short_io_buf;
2626 short_io_size = req_capsule_get_size(&req->rq_pill,
2629 short_io_buf = req_capsule_client_get(&req->rq_pill,
2631 CDEBUG(D_INFO, "Client use short io for data transfer,"
2632 " size = %d\n", short_io_size);
2634 /* Copy short io buf to pages */
2635 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2639 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2640 PTLRPC_BULK_GET_SINK,
2642 &ptlrpc_bulk_kiov_nopin_ops);
2644 GOTO(skip_transfer, rc = -ENOMEM);
2646 /* NB Having prepped, we must commit... */
2647 for (i = 0; i < npages; i++)
2648 desc->bd_frag_ops->add_kiov_frag(desc,
2649 local_nb[i].lnb_page,
2650 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2651 local_nb[i].lnb_len);
2653 rc = sptlrpc_svc_prep_bulk(req, desc);
2655 GOTO(skip_transfer, rc);
2657 rc = target_bulk_io(exp, desc);
2663 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2664 static int cksum_counter;
2666 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2667 cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2669 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2670 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2671 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2674 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2675 local_nb, npages, OST_WRITE,
2676 &repbody->oa.o_cksum);
2678 GOTO(out_commitrw, rc);
2682 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2683 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2684 body->oa.o_flags & OBD_FL_MMAP);
2686 tgt_warn_on_cksum(req, desc, local_nb, npages,
2688 repbody->oa.o_cksum, mmap);
2690 } else if ((cksum_counter & (-cksum_counter)) ==
2692 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2693 cksum_counter, libcfs_id2str(req->rq_peer),
2694 repbody->oa.o_cksum);
2699 /* Must commit after prep above in all cases */
2700 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2701 objcount, ioo, remote_nb, npages, local_nb, rc);
2702 if (rc == -ENOTCONN)
2703 /* quota acquire process has been given up because
2704 * either the client has been evicted or the client
2705 * has timed out the request already */
2708 for (i = 0; i < niocount; i++) {
2709 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2715 * Disable sending mtime back to the client. If the client locked the
2716 * whole object, then it has already updated the mtime on its side,
2717 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2719 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2724 /* set per-requested niobuf return codes */
2725 for (i = j = 0; i < niocount; i++) {
2726 int len = remote_nb[i].rnb_len;
2731 LASSERT(j < npages);
2732 if (local_nb[j].lnb_rc < 0)
2733 rcs[i] = local_nb[j].lnb_rc;
2734 len -= local_nb[j].lnb_len;
2739 LASSERT(j == npages);
2740 ptlrpc_lprocfs_brw(req, nob);
2743 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2745 ptlrpc_free_bulk(desc);
2747 if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2748 req->rq_no_reply = 1;
2749 /* reply out callback would free */
2750 ptlrpc_req_drop_rs(req);
2751 if (!exp->exp_obd->obd_no_transno)
2752 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2753 " client will retry: rc = %d\n",
2755 obd_uuid2str(&exp->exp_client_uuid),
2756 obd_export_nid2str(exp), rc);
2760 memalloc_noreclaim_restore(mpflags);
2764 EXPORT_SYMBOL(tgt_brw_write);
2766 /* Check if request can be reconstructed from saved reply data
2767 * A copy of the reply data is returned in @trd if the pointer is not NULL
2769 int req_can_reconstruct(struct ptlrpc_request *req,
2770 struct tg_reply_data *trd)
2772 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2773 struct lsd_client_data *lcd = ted->ted_lcd;
2776 if (tgt_is_multimodrpcs_client(req->rq_export))
2777 return tgt_lookup_reply(req, trd);
2779 mutex_lock(&ted->ted_lcd_lock);
2780 found = req->rq_xid == lcd->lcd_last_xid ||
2781 req->rq_xid == lcd->lcd_last_close_xid;
2783 if (found && trd != NULL) {
2784 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2785 trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2786 trd->trd_reply.lrd_transno =
2787 lcd->lcd_last_close_transno;
2788 trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2790 trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2791 trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2792 trd->trd_reply.lrd_result = lcd->lcd_last_result;
2793 trd->trd_reply.lrd_data = lcd->lcd_last_data;
2794 trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2795 trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2796 trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2797 trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2800 mutex_unlock(&ted->ted_lcd_lock);
2804 EXPORT_SYMBOL(req_can_reconstruct);