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>
42 #include <obd_class.h>
43 #include <obd_cksum.h>
44 #include <lustre_lfsck.h>
45 #include <lustre_nodemap.h>
46 #include <lustre_acl.h>
48 #include "tgt_internal.h"
50 char *tgt_name(struct lu_target *tgt)
52 LASSERT(tgt->lut_obd != NULL);
53 return tgt->lut_obd->obd_name;
55 EXPORT_SYMBOL(tgt_name);
58 * Generic code handling requests that have struct mdt_body passed in:
60 * - extract mdt_body from request and save it in @tsi, if present;
62 * - create lu_object, corresponding to the fid in mdt_body, and save it in
65 * - if HAS_BODY flag is set for this request type check whether object
66 * actually exists on storage (lu_object_exists()).
69 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
71 const struct mdt_body *body;
72 struct lu_object *obj;
73 struct req_capsule *pill = tsi->tsi_pill;
78 body = req_capsule_client_get(pill, &RMF_MDT_BODY);
82 tsi->tsi_mdt_body = body;
84 if (!(body->mbo_valid & OBD_MD_FLID))
87 /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
88 * in any case in pre-2.5 clients. Fix that here if needed */
89 if (unlikely(fid_is_zero(&body->mbo_fid1)))
92 if (!fid_is_sane(&body->mbo_fid1)) {
93 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
94 PFID(&body->mbo_fid1));
98 obj = lu_object_find(tsi->tsi_env,
99 &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
100 &body->mbo_fid1, NULL);
102 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
103 lu_object_put(tsi->tsi_env, obj);
106 tsi->tsi_corpus = obj;
113 tsi->tsi_fid = body->mbo_fid1;
119 * Validate oa from client.
120 * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
122 * a. objects in Single MDT FS seq = FID_SEQ_OST_MDT0, oi_id != 0
123 * b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
124 * pack ost_id. Because non-zero oi_seq will make it diffcult to tell
125 * whether this is oi_fid or real ostid. So it will check
126 * OBD_CONNECT_FID, then convert the ostid to FID for old client.
127 * c. Old FID-disable osc will send IDIF.
128 * d. new FID-enable osc/osp will send normal FID.
130 * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
131 * be used for LAST_ID file, and only being accessed inside OST now.
133 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
135 struct ost_id *oi = &oa->o_oi;
136 u64 seq = ostid_seq(oi);
137 u64 id = ostid_id(oi);
141 if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
142 fid_seq_is_echo(seq))) {
143 /* Sigh 2.[123] client still sends echo req with oi_id = 0
144 * during create, and we will reset this to 1, since this
145 * oi_id is basically useless in the following create process,
146 * but oi_id == 0 will make it difficult to tell whether it is
147 * real FID or ost_id. */
148 oi->oi_fid.f_seq = FID_SEQ_ECHO;
149 oi->oi_fid.f_oid = id ?: 1;
150 oi->oi_fid.f_ver = 0;
152 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
154 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
155 GOTO(out, rc = -EPROTO);
157 /* Note: this check might be forced in 2.5 or 2.6, i.e.
158 * all of the requests are required to setup FLGROUP */
159 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
160 ostid_set_seq_mdt0(oi);
161 oa->o_valid |= OBD_MD_FLGROUP;
165 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
166 fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
167 GOTO(out, rc = -EPROTO);
169 rc = ostid_to_fid(&tti->tti_fid1, oi,
170 tsi->tsi_tgt->lut_lsd.lsd_osd_index);
171 if (unlikely(rc != 0))
174 oi->oi_fid = tti->tti_fid1;
180 CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
181 tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
185 EXPORT_SYMBOL(tgt_validate_obdo);
187 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
190 struct niobuf_remote *rnb;
191 struct obd_ioobj *ioo;
196 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
200 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
204 max_brw = ioobj_max_brw_get(ioo);
205 if (unlikely((max_brw & (max_brw - 1)) != 0)) {
206 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
207 ": rc = %d\n", tgt_name(tsi->tsi_tgt),
208 obd_export_nid2str(tsi->tsi_exp), max_brw,
209 POSTID(oi), -EPROTO);
214 obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
215 RCL_CLIENT) / sizeof(*ioo);
216 if (obj_count == 0) {
217 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
219 } else if (obj_count > 1) {
220 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
225 if (ioo->ioo_bufcnt == 0) {
226 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
230 if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
231 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
232 "bulk has too many pages (%d)",
240 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
242 struct ost_body *body;
243 struct req_capsule *pill = tsi->tsi_pill;
244 struct lu_nodemap *nodemap;
249 body = req_capsule_client_get(pill, &RMF_OST_BODY);
253 rc = tgt_validate_obdo(tsi, &body->oa);
257 nodemap = nodemap_get_from_exp(tsi->tsi_exp);
259 RETURN(PTR_ERR(nodemap));
261 body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
262 NODEMAP_CLIENT_TO_FS,
264 body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
265 NODEMAP_CLIENT_TO_FS,
267 nodemap_putref(nodemap);
269 tsi->tsi_ost_body = body;
270 tsi->tsi_fid = body->oa.o_oi.oi_fid;
272 if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
273 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
278 if (!(body->oa.o_valid & OBD_MD_FLID)) {
279 if (flags & HAS_BODY) {
280 CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\n",
281 tgt_name(tsi->tsi_tgt));
288 ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
291 * OST doesn't get object in advance for further use to prevent
292 * situations with nested object_find which is potential deadlock.
294 tsi->tsi_corpus = NULL;
299 * Do necessary preprocessing according to handler ->th_flags.
301 static int tgt_request_preprocess(struct tgt_session_info *tsi,
302 struct tgt_handler *h,
303 struct ptlrpc_request *req)
305 struct req_capsule *pill = tsi->tsi_pill;
306 __u32 flags = h->th_flags;
311 if (tsi->tsi_preprocessed)
314 LASSERT(h->th_act != NULL);
315 LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
316 LASSERT(current->journal_info == NULL);
318 LASSERT(ergo(flags & (HAS_BODY | HAS_REPLY),
320 if (h->th_fmt != NULL) {
321 req_capsule_set(pill, h->th_fmt);
322 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
323 rc = tgt_mdt_body_unpack(tsi, flags);
326 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
328 rc = tgt_ost_body_unpack(tsi, flags);
334 if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
337 if (flags & HAS_KEY) {
338 struct ldlm_request *dlm_req;
340 LASSERT(h->th_fmt != NULL);
342 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
343 if (dlm_req != NULL) {
344 union ldlm_wire_policy_data *policy =
345 &dlm_req->lock_desc.l_policy_data;
347 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
349 (policy->l_inodebits.bits |
350 policy->l_inodebits.try_bits) == 0)) {
352 * Lock without inodebits makes no sense and
353 * will oops later in ldlm. If client miss to
354 * set such bits, do not trigger ASSERTION.
356 * For liblustre flock case, it maybe zero.
360 tsi->tsi_dlm_req = dlm_req;
366 tsi->tsi_preprocessed = 1;
371 * Invoke handler for this request opc. Also do necessary preprocessing
372 * (according to handler ->th_flags), and post-processing (setting of
373 * ->last_{xid,committed}).
375 static int tgt_handle_request0(struct tgt_session_info *tsi,
376 struct tgt_handler *h,
377 struct ptlrpc_request *req)
381 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
386 /* When dealing with sec context requests, no export is associated yet,
387 * because these requests are sent before *_CONNECT requests.
388 * A NULL req->rq_export means the normal *_common_slice handlers will
389 * not be called, because there is no reference to the target.
390 * So deal with them by hand and jump directly to target_send_reply().
394 case SEC_CTX_INIT_CONT:
396 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
401 * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
402 * to put same checks into handlers like mdt_close(), mdt_reint(),
403 * etc., without talking to mdt authors first. Checking same thing
404 * there again is useless and returning 0 error without packing reply
405 * is buggy! Handlers either pack reply or return error.
407 * We return 0 here and do not send any reply in order to emulate
408 * network failure. Do not send any reply in case any of NET related
409 * fail_id has occured.
411 if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
413 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
414 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
417 rc = tgt_request_preprocess(tsi, h, req);
418 /* pack reply if reply format is fixed */
419 if (rc == 0 && h->th_flags & HAS_REPLY) {
421 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
423 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
425 tsi->tsi_mdt_body->mbo_eadatasize);
426 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
428 req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
430 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
431 req_capsule_set_size(tsi->tsi_pill,
432 &RMF_ACL, RCL_SERVER,
433 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
435 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
437 struct niobuf_remote *remote_nb =
438 req_capsule_client_get(tsi->tsi_pill,
440 struct ost_body *body = tsi->tsi_ost_body;
442 req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
444 (body->oa.o_valid & OBD_MD_FLFLAGS &&
445 body->oa.o_flags & OBD_FL_SHORT_IO) ?
446 remote_nb[0].rnb_len : 0);
449 rc = req_capsule_server_pack(tsi->tsi_pill);
452 if (likely(rc == 0)) {
454 * Process request, there can be two types of rc:
455 * 1) errors with msg unpack/pack, other failures outside the
456 * operation itself. This is counted as serious errors;
457 * 2) errors during fs operation, should be placed in rq_status
461 if (!is_serious(rc) &&
462 !req->rq_no_reply && req->rq_reply_state == NULL) {
463 DEBUG_REQ(D_ERROR, req,
464 "%s: %s handler did not pack reply but returned no error",
465 tgt_name(tsi->tsi_tgt), h->th_name);
468 serious = is_serious(rc);
469 rc = clear_serious(rc);
477 * ELDLM_* codes which > 0 should be in rq_status only as well as
478 * all non-serious errors.
480 if (rc > 0 || !serious)
483 LASSERT(current->journal_info == NULL);
485 if (likely(rc == 0 && req->rq_export))
486 target_committed_to_req(req);
489 target_send_reply(req, rc, tsi->tsi_reply_fail_id);
493 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
494 struct obd_device *obd, int *process)
496 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
503 case MDS_SYNC: /* used in unmounting */
517 case MDS_HSM_PROGRESS:
518 case MDS_HSM_STATE_SET:
519 case MDS_HSM_REQUEST:
520 *process = target_queue_recovery_request(req, obd);
524 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
531 * Handle recovery. Return:
532 * +1: continue request processing;
533 * -ve: abort immediately with the given error code;
534 * 0: send reply with error code in req->rq_status;
536 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
540 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
545 case SEC_CTX_INIT_CONT:
550 if (!req->rq_export->exp_obd->obd_replayable)
553 /* sanity check: if the xid matches, the request must be marked as a
554 * resent or replayed */
555 if (req_can_reconstruct(req, NULL) == 1) {
556 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
557 (MSG_RESENT | MSG_REPLAY))) {
558 DEBUG_REQ(D_WARNING, req,
559 "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
561 lustre_msg_get_flags(req->rq_reqmsg));
562 req->rq_status = -ENOTCONN;
566 /* else: note the opposite is not always true; a RESENT req after a
567 * failover will usually not match the last_xid, since it was likely
568 * never committed. A REPLAYed request will almost never match the
569 * last xid, however it could for a committed, but still retained,
572 /* Check for aborted recovery... */
573 if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
577 DEBUG_REQ(D_INFO, req, "Got new replay");
578 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
580 if (rc != 0 || !should_process)
582 else if (should_process < 0) {
583 req->rq_status = should_process;
584 rc = ptlrpc_error(req);
591 /* Initial check for request, it is validation mostly */
592 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
594 struct tgt_handler *h;
595 struct tgt_opc_slice *s;
596 struct lu_target *tgt;
597 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
601 tgt = class_exp2tgt(req->rq_export);
602 if (unlikely(tgt == NULL)) {
603 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
604 class_exp2obd(req->rq_export)->obd_name);
605 RETURN(ERR_PTR(-EINVAL));
608 for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
609 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
612 /* opcode was not found in slice */
613 if (unlikely(s->tos_hs == NULL)) {
614 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
616 RETURN(ERR_PTR(-ENOTSUPP));
619 LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
620 h = s->tos_hs + (opc - s->tos_opc_start);
621 if (unlikely(h->th_opc == 0)) {
622 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
623 RETURN(ERR_PTR(-ENOTSUPP));
629 static int process_req_last_xid(struct ptlrpc_request *req)
633 struct obd_export *exp = req->rq_export;
634 struct tg_export_data *ted = &exp->exp_target_data;
635 bool need_lock = tgt_is_multimodrpcs_client(exp);
639 mutex_lock(&ted->ted_lcd_lock);
640 /* check request's xid is consistent with export's last_xid */
641 last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
642 if (last_xid > exp->exp_last_xid)
643 exp->exp_last_xid = last_xid;
645 if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
646 /* Some request is allowed to be sent during replay,
647 * such as OUT update requests, FLD requests, so it
648 * is possible that replay requests has smaller XID
649 * than the exp_last_xid.
651 * Some non-replay requests may have smaller XID as
654 * - Client send a no_resend RPC, like statfs;
655 * - The RPC timedout (or some other error) on client,
656 * then it's removed from the unreplied list;
657 * - Client send some other request to bump the
658 * exp_last_xid on server;
659 * - The former RPC got chance to be processed;
661 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
664 DEBUG_REQ(D_WARNING, req,
665 "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
666 req->rq_xid, exp->exp_last_xid, rc);
671 /* The "last_xid" is the minimum xid among unreplied requests,
672 * if the request is from the previous connection, its xid can
673 * still be larger than "exp_last_xid", then the above check of
674 * xid is not enough to determine whether the request is delayed.
676 * For example, if some replay request was delayed and caused
677 * timeout at client and the replay is restarted, the delayed
678 * replay request will have the larger xid than "exp_last_xid"
680 if (req->rq_export->exp_conn_cnt >
681 lustre_msg_get_conn_cnt(req->rq_reqmsg))
682 GOTO(out, rc = -ESTALE);
684 /* try to release in-memory reply data */
685 if (tgt_is_multimodrpcs_client(exp)) {
686 tgt_handle_received_xid(exp, last_xid);
687 rc = tgt_handle_tag(req);
692 mutex_unlock(&ted->ted_lcd_lock);
697 int tgt_request_handle(struct ptlrpc_request *req)
699 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
701 struct lustre_msg *msg = req->rq_reqmsg;
702 struct tgt_handler *h;
703 struct lu_target *tgt;
704 int request_fail_id = 0;
705 __u32 opc = lustre_msg_get_opc(msg);
706 struct obd_device *obd;
708 bool is_connect = false;
711 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
712 if (cfs_fail_val == 0 &&
713 lustre_msg_get_opc(msg) != OBD_PING &&
714 lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
717 wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
721 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
722 tsi->tsi_pill = &req->rq_pill;
723 tsi->tsi_env = req->rq_svc_thread->t_env;
725 /* if request has export then get handlers slice from corresponding
726 * target, otherwise that should be connect operation */
727 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
728 opc == MGS_CONNECT) {
730 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
731 rc = target_handle_connect(req);
733 rc = ptlrpc_error(req);
736 /* recovery-small test 18c asks to drop connect reply */
737 if (unlikely(opc == OST_CONNECT &&
738 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
742 if (unlikely(!class_connected_export(req->rq_export))) {
743 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
744 opc == SEC_CTX_FINI) {
745 /* sec context initialization has to be handled
746 * by hand in tgt_handle_request0() */
747 tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
749 GOTO(handle_recov, rc = 0);
751 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
752 opc, libcfs_id2str(req->rq_peer));
753 req->rq_status = -ENOTCONN;
754 rc = ptlrpc_error(req);
758 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
759 tsi->tsi_exp = req->rq_export;
760 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
761 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
763 tsi->tsi_jobid = NULL;
766 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
767 class_exp2obd(req->rq_export)->obd_name);
768 req->rq_status = -EINVAL;
769 rc = ptlrpc_error(req);
773 /* Skip last_xid processing for the recovery thread, otherwise, the
774 * last_xid on same request could be processed twice: first time when
775 * processing the incoming request, second time when the request is
776 * being processed by recovery thread. */
777 obd = class_exp2obd(req->rq_export);
779 /* reset the exp_last_xid on each connection. */
780 req->rq_export->exp_last_xid = 0;
781 } else if (obd->obd_recovery_data.trd_processing_task !=
783 rc = process_req_last_xid(req);
786 rc = ptlrpc_error(req);
791 request_fail_id = tgt->lut_request_fail_id;
792 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
794 h = tgt_handler_find_check(req);
796 req->rq_status = PTR_ERR(h);
797 rc = ptlrpc_error(req);
801 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
804 if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
805 CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
808 rc = lustre_msg_check_version(msg, h->th_version);
810 DEBUG_REQ(D_ERROR, req,
811 "%s: drop malformed request version=%08x expect=%08x",
812 tgt_name(tgt), lustre_msg_get_version(msg),
814 req->rq_status = -EINVAL;
815 rc = ptlrpc_error(req);
820 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
821 if (likely(rc == 1)) {
822 rc = tgt_handle_request0(tsi, h, req);
828 req_capsule_fini(tsi->tsi_pill);
829 if (tsi->tsi_corpus != NULL) {
830 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
831 tsi->tsi_corpus = NULL;
835 EXPORT_SYMBOL(tgt_request_handle);
837 /** Assign high priority operations to the request if needed. */
838 int tgt_hpreq_handler(struct ptlrpc_request *req)
840 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
841 struct tgt_handler *h;
846 if (req->rq_export == NULL)
849 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
850 tsi->tsi_pill = &req->rq_pill;
851 tsi->tsi_env = req->rq_svc_thread->t_env;
852 tsi->tsi_tgt = class_exp2tgt(req->rq_export);
853 tsi->tsi_exp = req->rq_export;
855 h = tgt_handler_find_check(req);
861 rc = tgt_request_preprocess(tsi, h, req);
862 if (unlikely(rc != 0))
865 if (h->th_hp != NULL)
869 EXPORT_SYMBOL(tgt_hpreq_handler);
871 void tgt_counter_incr(struct obd_export *exp, int opcode)
873 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
874 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
875 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
877 EXPORT_SYMBOL(tgt_counter_incr);
880 * Unified target generic handlers.
883 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
885 struct lu_target *tgt = class_exp2tgt(exp);
886 struct sptlrpc_flavor flvr;
890 LASSERT(tgt->lut_obd);
891 LASSERT(tgt->lut_slice);
893 /* always allow ECHO client */
894 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
895 LUSTRE_ECHO_NAME) == 0)) {
896 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
900 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
901 read_lock(&tgt->lut_sptlrpc_lock);
902 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
906 read_unlock(&tgt->lut_sptlrpc_lock);
908 spin_lock(&exp->exp_lock);
909 exp->exp_sp_peer = req->rq_sp_from;
910 exp->exp_flvr = flvr;
912 /* when on mgs, if no restriction is set, or if client
913 * is loopback, allow any flavor */
914 if ((strcmp(exp->exp_obd->obd_type->typ_name,
915 LUSTRE_MGS_NAME) == 0) &&
916 (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
917 LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
919 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
921 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
922 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
923 CERROR("%s: unauthorized rpc flavor %x from %s, "
924 "expect %x\n", tgt_name(tgt),
926 libcfs_nid2str(req->rq_peer.nid),
927 exp->exp_flvr.sf_rpc);
930 spin_unlock(&exp->exp_lock);
932 if (exp->exp_sp_peer != req->rq_sp_from) {
933 CERROR("%s: RPC source %s doesn't match %s\n",
935 sptlrpc_part2name(req->rq_sp_from),
936 sptlrpc_part2name(exp->exp_sp_peer));
939 rc = sptlrpc_target_export_check(exp, req);
946 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
948 struct sptlrpc_rule_set tmp_rset;
951 if (unlikely(tgt == NULL)) {
952 CERROR("No target passed");
956 sptlrpc_rule_set_init(&tmp_rset);
957 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
959 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
964 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
966 write_lock(&tgt->lut_sptlrpc_lock);
967 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
968 tgt->lut_sptlrpc_rset = tmp_rset;
969 write_unlock(&tgt->lut_sptlrpc_lock);
973 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
975 int tgt_connect(struct tgt_session_info *tsi)
977 struct ptlrpc_request *req = tgt_ses_req(tsi);
978 struct obd_connect_data *reply;
983 /* XXX: better to call this check right after getting new export but
984 * before last_rcvd slot allocation to avoid server load upon insecure
985 * connects. This is to be fixed after unifiyng all targets.
987 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
991 /* To avoid exposing partially initialized connection flags, changes up
992 * to this point have been staged in reply->ocd_connect_flags. Now that
993 * connection handling has completed successfully, atomically update
994 * the connect flags in the shared export data structure. LU-1623 */
995 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
996 spin_lock(&tsi->tsi_exp->exp_lock);
997 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
998 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
999 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1000 reply->ocd_connect_flags2;
1001 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1002 spin_unlock(&tsi->tsi_exp->exp_lock);
1004 if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1005 LUSTRE_MDT_NAME) == 0) {
1006 rc = req_check_sepol(tsi->tsi_pill);
1013 obd_disconnect(class_export_get(tsi->tsi_exp));
1016 EXPORT_SYMBOL(tgt_connect);
1018 int tgt_disconnect(struct tgt_session_info *tsi)
1024 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1026 rc = target_handle_disconnect(tgt_ses_req(tsi));
1028 RETURN(err_serious(rc));
1032 EXPORT_SYMBOL(tgt_disconnect);
1035 * Unified target OBD handlers
1037 int tgt_obd_ping(struct tgt_session_info *tsi)
1043 /* The target-specific part of OBD_PING request handling.
1044 * It controls Filter Modification Data (FMD) expiration each time
1047 * Valid only for replayable targets, e.g. MDT and OFD
1049 if (tsi->tsi_exp->exp_obd->obd_replayable)
1050 tgt_fmd_expire(tsi->tsi_exp);
1052 rc = req_capsule_server_pack(tsi->tsi_pill);
1054 RETURN(err_serious(rc));
1058 EXPORT_SYMBOL(tgt_obd_ping);
1060 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1062 return err_serious(-EOPNOTSUPP);
1065 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1067 struct ptlrpc_request *req = tgt_ses_req(tsi);
1068 struct obd_export *exp = req->rq_export;
1069 struct ptlrpc_bulk_desc *desc;
1076 for (i = 0; i < rdbuf->rb_nbufs; i++) {
1077 unsigned int offset;
1079 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1080 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1084 desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1085 PTLRPC_BULK_PUT_SOURCE,
1087 &ptlrpc_bulk_kiov_nopin_ops);
1091 for (i = 0; i < rdbuf->rb_nbufs; i++)
1092 desc->bd_frag_ops->add_iov_frag(desc,
1093 rdbuf->rb_bufs[i].lb_buf,
1094 rdbuf->rb_bufs[i].lb_len);
1096 rc = target_bulk_io(exp, desc);
1097 ptlrpc_free_bulk(desc);
1100 EXPORT_SYMBOL(tgt_send_buffer);
1102 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1104 struct ptlrpc_request *req = tgt_ses_req(tsi);
1105 struct obd_export *exp = req->rq_export;
1106 struct ptlrpc_bulk_desc *desc;
1114 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1115 PTLRPC_BULK_PUT_SOURCE,
1117 &ptlrpc_bulk_kiov_pin_ops);
1121 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1122 /* old client requires reply size in it's PAGE_SIZE,
1123 * which is rdpg->rp_count */
1124 nob = rdpg->rp_count;
1126 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1127 i++, tmpcount -= tmpsize) {
1128 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1129 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1133 LASSERT(desc->bd_nob == nob);
1134 rc = target_bulk_io(exp, desc);
1135 ptlrpc_free_bulk(desc);
1138 EXPORT_SYMBOL(tgt_sendpage);
1141 * OBD_IDX_READ handler
1143 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1145 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1146 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1147 struct idx_info *req_ii, *rep_ii;
1152 memset(rdpg, 0, sizeof(*rdpg));
1153 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1155 /* extract idx_info buffer from request & reply */
1156 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1157 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1158 RETURN(err_serious(-EPROTO));
1160 rc = req_capsule_server_pack(tsi->tsi_pill);
1162 RETURN(err_serious(rc));
1164 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1166 RETURN(err_serious(-EFAULT));
1167 rep_ii->ii_magic = IDX_INFO_MAGIC;
1169 /* extract hash to start with */
1170 rdpg->rp_hash = req_ii->ii_hash_start;
1172 /* extract requested attributes */
1173 rdpg->rp_attrs = req_ii->ii_attrs;
1175 /* check that fid packed in request is valid and supported */
1176 if (!fid_is_sane(&req_ii->ii_fid))
1178 rep_ii->ii_fid = req_ii->ii_fid;
1181 rep_ii->ii_flags = req_ii->ii_flags;
1183 /* compute number of pages to allocate, ii_count is the number of 4KB
1185 if (req_ii->ii_count <= 0)
1186 GOTO(out, rc = -EFAULT);
1187 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1188 exp_max_brw_size(tsi->tsi_exp));
1189 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1191 /* allocate pages to store the containers */
1192 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1193 if (rdpg->rp_pages == NULL)
1194 GOTO(out, rc = -ENOMEM);
1195 for (i = 0; i < rdpg->rp_npages; i++) {
1196 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1197 if (rdpg->rp_pages[i] == NULL)
1198 GOTO(out, rc = -ENOMEM);
1201 /* populate pages with key/record pairs */
1202 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1206 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1207 "asked %d > %d\n", rc, rdpg->rp_count);
1209 /* send pages to client */
1210 rc = tgt_sendpage(tsi, rdpg, rc);
1215 if (rdpg->rp_pages) {
1216 for (i = 0; i < rdpg->rp_npages; i++)
1217 if (rdpg->rp_pages[i])
1218 __free_page(rdpg->rp_pages[i]);
1219 OBD_FREE(rdpg->rp_pages,
1220 rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1225 struct tgt_handler tgt_obd_handlers[] = {
1226 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1227 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1229 EXPORT_SYMBOL(tgt_obd_handlers);
1231 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1232 struct dt_object *obj, __u64 start, __u64 end)
1238 /* if no objid is specified, it means "sync whole filesystem" */
1240 rc = dt_sync(env, tgt->lut_bottom);
1241 } else if (dt_version_get(env, obj) >
1242 tgt->lut_obd->obd_last_committed) {
1243 rc = dt_object_sync(env, obj, start, end);
1245 atomic_inc(&tgt->lut_sync_count);
1249 EXPORT_SYMBOL(tgt_sync);
1251 * Unified target DLM handlers.
1255 * Unified target BAST
1257 * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1258 * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1259 * but if it's cross-MDT lock, because remote object version is not set, a
1260 * filesystem sync is needed.
1262 * \param lock server side lock
1263 * \param desc lock desc
1264 * \param data ldlm_cb_set_arg
1265 * \param flag indicates whether this cancelling or blocking callback
1266 * \retval 0 on success
1267 * \retval negative number on error
1269 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1270 void *data, int flag)
1273 struct lu_target *tgt;
1274 struct dt_object *obj = NULL;
1280 tgt = class_exp2tgt(lock->l_export);
1282 if (unlikely(tgt == NULL)) {
1283 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1284 class_exp2obd(lock->l_export)->obd_name);
1288 if (flag == LDLM_CB_CANCELING &&
1289 (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1290 (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1291 (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1292 ldlm_is_cbpending(lock))) &&
1293 ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1294 lock->l_resource->lr_type == LDLM_EXTENT)) {
1296 __u64 end = OBD_OBJECT_EOF;
1298 rc = lu_env_init(&env, LCT_DT_THREAD);
1299 if (unlikely(rc != 0))
1302 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1303 tgt->lut_lsd.lsd_osd_index);
1305 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1306 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1308 GOTO(err_env, rc = PTR_ERR(obj));
1310 if (!dt_object_exists(obj))
1311 GOTO(err_put, rc = -ENOENT);
1313 start = lock->l_policy_data.l_extent.start;
1314 end = lock->l_policy_data.l_extent.end;
1317 rc = tgt_sync(&env, tgt, obj, start, end);
1319 CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1320 "cancel: rc = %d\n",
1321 tgt_name(tgt), PFID(&fid),
1322 lock->l_policy_data.l_extent.start,
1323 lock->l_policy_data.l_extent.end, rc);
1327 dt_object_put(&env, obj);
1332 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1336 static struct ldlm_callback_suite tgt_dlm_cbs = {
1337 .lcs_completion = ldlm_server_completion_ast,
1338 .lcs_blocking = tgt_blocking_ast,
1339 .lcs_glimpse = ldlm_server_glimpse_ast
1342 int tgt_enqueue(struct tgt_session_info *tsi)
1344 struct ptlrpc_request *req = tgt_ses_req(tsi);
1349 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1350 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1352 LASSERT(tsi->tsi_dlm_req != NULL);
1353 rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1354 tsi->tsi_dlm_req, &tgt_dlm_cbs);
1356 RETURN(err_serious(rc));
1358 switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1360 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1363 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1366 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1369 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1372 RETURN(req->rq_status);
1374 EXPORT_SYMBOL(tgt_enqueue);
1376 int tgt_convert(struct tgt_session_info *tsi)
1378 struct ptlrpc_request *req = tgt_ses_req(tsi);
1382 LASSERT(tsi->tsi_dlm_req);
1383 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1385 RETURN(err_serious(rc));
1387 RETURN(req->rq_status);
1390 int tgt_bl_callback(struct tgt_session_info *tsi)
1392 return err_serious(-EOPNOTSUPP);
1395 int tgt_cp_callback(struct tgt_session_info *tsi)
1397 return err_serious(-EOPNOTSUPP);
1400 /* generic LDLM target handler */
1401 struct tgt_handler tgt_dlm_handlers[] = {
1402 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1403 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1404 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1405 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1407 EXPORT_SYMBOL(tgt_dlm_handlers);
1410 * Unified target LLOG handlers.
1412 int tgt_llog_open(struct tgt_session_info *tsi)
1418 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1422 EXPORT_SYMBOL(tgt_llog_open);
1424 int tgt_llog_read_header(struct tgt_session_info *tsi)
1430 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1434 EXPORT_SYMBOL(tgt_llog_read_header);
1436 int tgt_llog_next_block(struct tgt_session_info *tsi)
1442 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1446 EXPORT_SYMBOL(tgt_llog_next_block);
1448 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1454 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1458 EXPORT_SYMBOL(tgt_llog_prev_block);
1460 /* generic llog target handler */
1461 struct tgt_handler tgt_llog_handlers[] = {
1462 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1463 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1464 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1465 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1467 EXPORT_SYMBOL(tgt_llog_handlers);
1470 * sec context handlers
1472 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1473 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1478 struct tgt_handler tgt_sec_ctx_handlers[] = {
1479 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle),
1480 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle),
1481 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle),
1483 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1485 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1486 struct dt_device *key,
1487 struct lfsck_req_local *lrl,
1488 struct thandle *th) = NULL;
1490 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1492 struct lfsck_req_local *,
1495 tgt_lfsck_in_notify_local = notify;
1497 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1499 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1500 struct dt_device *key,
1501 struct lfsck_request *lr) = NULL;
1503 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1505 struct lfsck_request *))
1507 tgt_lfsck_in_notify = notify;
1509 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1511 static int (*tgt_lfsck_query)(const struct lu_env *env,
1512 struct dt_device *key,
1513 struct lfsck_request *req,
1514 struct lfsck_reply *rep,
1515 struct lfsck_query *que) = NULL;
1517 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1519 struct lfsck_request *,
1520 struct lfsck_reply *,
1521 struct lfsck_query *))
1523 tgt_lfsck_query = query;
1525 EXPORT_SYMBOL(tgt_register_lfsck_query);
1527 /* LFSCK request handlers */
1528 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1530 const struct lu_env *env = tsi->tsi_env;
1531 struct dt_device *key = tsi->tsi_tgt->lut_bottom;
1532 struct lfsck_request *lr;
1536 lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1540 rc = tgt_lfsck_in_notify(env, key, lr);
1545 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1547 struct lfsck_request *request;
1548 struct lfsck_reply *reply;
1552 request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1553 if (request == NULL)
1556 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1560 rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1561 request, reply, NULL);
1563 RETURN(rc < 0 ? rc : 0);
1566 struct tgt_handler tgt_lfsck_handlers[] = {
1567 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_NOTIFY, tgt_handle_lfsck_notify),
1568 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_QUERY, tgt_handle_lfsck_query),
1570 EXPORT_SYMBOL(tgt_lfsck_handlers);
1573 * initialize per-thread page pool (bug 5137).
1575 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1577 struct tgt_thread_big_cache *tbc;
1581 LASSERT(thread != NULL);
1582 LASSERT(thread->t_data == NULL);
1584 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1587 thread->t_data = tbc;
1590 EXPORT_SYMBOL(tgt_io_thread_init);
1593 * free per-thread pool created by tgt_thread_init().
1595 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1597 struct tgt_thread_big_cache *tbc;
1601 LASSERT(thread != NULL);
1604 * be prepared to handle partially-initialized pools (because this is
1605 * called from ost_io_thread_init() for cleanup.
1607 tbc = thread->t_data;
1609 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1610 thread->t_data = NULL;
1614 EXPORT_SYMBOL(tgt_io_thread_done);
1617 * Helper function for getting Data-on-MDT file server DLM lock
1618 * if asked by client.
1620 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1621 struct lustre_handle *lh, int mode, __u64 *flags)
1623 union ldlm_policy_data policy = {
1624 .l_inodebits.bits = MDS_INODELOCK_DOM,
1630 LASSERT(lh != NULL);
1631 LASSERT(ns != NULL);
1632 LASSERT(!lustre_handle_is_used(lh));
1634 rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1635 flags, ldlm_blocking_ast,
1636 ldlm_completion_ast, ldlm_glimpse_ast,
1637 NULL, 0, LVB_T_NONE, NULL, lh);
1639 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1641 EXPORT_SYMBOL(tgt_mdt_data_lock);
1643 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1645 LASSERT(lustre_handle_is_used(lh));
1646 ldlm_lock_decref(lh, mode);
1648 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1651 * Helper function for getting server side [start, start+count] DLM lock
1652 * if asked by client.
1654 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1655 struct ldlm_res_id *res_id, __u64 start, __u64 end,
1656 struct lustre_handle *lh, int mode, __u64 *flags)
1658 union ldlm_policy_data policy;
1663 LASSERT(lh != NULL);
1664 LASSERT(ns != NULL);
1665 LASSERT(!lustre_handle_is_used(lh));
1667 policy.l_extent.gid = 0;
1668 policy.l_extent.start = start & PAGE_MASK;
1671 * If ->o_blocks is EOF it means "lock till the end of the file".
1672 * Otherwise, it's size of an extent or hole being punched (in bytes).
1674 if (end == OBD_OBJECT_EOF || end < start)
1675 policy.l_extent.end = OBD_OBJECT_EOF;
1677 policy.l_extent.end = end | ~PAGE_MASK;
1679 rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1680 mode, flags, ldlm_blocking_ast,
1681 ldlm_completion_ast, ldlm_glimpse_ast,
1682 NULL, 0, LVB_T_NONE, NULL, lh);
1683 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1685 EXPORT_SYMBOL(tgt_extent_lock);
1687 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1689 LASSERT(lustre_handle_is_used(lh));
1690 ldlm_lock_decref(lh, mode);
1692 EXPORT_SYMBOL(tgt_extent_unlock);
1694 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1695 struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1696 struct niobuf_remote *nb, struct lustre_handle *lh,
1697 enum ldlm_mode mode)
1699 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1701 int nrbufs = obj->ioo_bufcnt;
1707 LASSERT(mode == LCK_PR || mode == LCK_PW);
1708 LASSERT(!lustre_handle_is_used(lh));
1710 if (ns->ns_obd->obd_recovering)
1713 if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1716 for (i = 1; i < nrbufs; i++)
1717 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1720 /* MDT IO for data-on-mdt */
1721 if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1722 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1724 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1725 nb[nrbufs - 1].rnb_offset +
1726 nb[nrbufs - 1].rnb_len - 1,
1731 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1732 struct lustre_handle *lh, enum ldlm_mode mode)
1736 LASSERT(mode == LCK_PR || mode == LCK_PW);
1737 LASSERT((obj->ioo_bufcnt > 0 &&
1738 (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1739 lustre_handle_is_used(lh));
1741 if (lustre_handle_is_used(lh))
1742 tgt_extent_unlock(lh, mode);
1745 static int tgt_checksum_niobuf(struct lu_target *tgt,
1746 struct niobuf_local *local_nb, int npages,
1747 int opc, enum cksum_types cksum_type,
1750 struct ahash_request *req;
1751 unsigned int bufsize;
1753 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1755 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1757 CERROR("%s: unable to initialize checksum hash %s\n",
1758 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1759 return PTR_ERR(req);
1762 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1763 for (i = 0; i < npages; i++) {
1764 /* corrupt the data before we compute the checksum, to
1765 * simulate a client->OST data error */
1766 if (i == 0 && opc == OST_WRITE &&
1767 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1768 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1769 int len = local_nb[i].lnb_len;
1770 struct page *np = tgt_page_to_corrupt;
1773 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1774 char *ptr2 = page_address(np);
1776 memcpy(ptr2 + off, ptr + off, len);
1777 memcpy(ptr2 + off, "bad3", min(4, len));
1780 /* LU-8376 to preserve original index for
1781 * display in dump_all_bulk_pages() */
1784 cfs_crypto_hash_update_page(req, np, off,
1788 CERROR("%s: can't alloc page for corruption\n",
1792 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1793 local_nb[i].lnb_page_offset & ~PAGE_MASK,
1794 local_nb[i].lnb_len);
1796 /* corrupt the data after we compute the checksum, to
1797 * simulate an OST->client data error */
1798 if (i == 0 && opc == OST_READ &&
1799 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1800 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1801 int len = local_nb[i].lnb_len;
1802 struct page *np = tgt_page_to_corrupt;
1805 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1806 char *ptr2 = page_address(np);
1808 memcpy(ptr2 + off, ptr + off, len);
1809 memcpy(ptr2 + off, "bad4", min(4, len));
1812 /* LU-8376 to preserve original index for
1813 * display in dump_all_bulk_pages() */
1816 cfs_crypto_hash_update_page(req, np, off,
1820 CERROR("%s: can't alloc page for corruption\n",
1826 bufsize = sizeof(*cksum);
1827 err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1832 char dbgcksum_file_name[PATH_MAX];
1834 static void dump_all_bulk_pages(struct obdo *oa, int count,
1835 struct niobuf_local *local_nb,
1836 __u32 server_cksum, __u32 client_cksum)
1843 /* will only keep dump of pages on first error for the same range in
1844 * file/fid, not during the resends/retries. */
1845 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1846 "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1847 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1848 libcfs_debug_file_path_arr :
1849 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1850 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1851 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1852 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1853 local_nb[0].lnb_file_offset,
1854 local_nb[count-1].lnb_file_offset +
1855 local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1856 filp = filp_open(dbgcksum_file_name,
1857 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1861 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1862 "checksum error: rc = %d\n", dbgcksum_file_name,
1865 CERROR("%s: can't open to dump pages with checksum "
1866 "error: rc = %d\n", dbgcksum_file_name, rc);
1870 for (i = 0; i < count; i++) {
1871 len = local_nb[i].lnb_len;
1872 buf = kmap(local_nb[i].lnb_page);
1874 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1876 CERROR("%s: wanted to write %u but got %d "
1877 "error\n", dbgcksum_file_name, len, rc);
1882 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1883 dbgcksum_file_name, rc);
1885 kunmap(local_nb[i].lnb_page);
1888 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1890 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1891 filp_close(filp, NULL);
1894 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1895 struct obd_export *exp, struct obdo *oa,
1896 const struct lnet_process_id *peer,
1897 __u32 client_cksum, __u32 server_cksum,
1898 enum cksum_types server_cksum_type)
1901 enum cksum_types cksum_type;
1904 /* unlikely to happen and only if resend does not occur due to cksum
1905 * control failure on Client */
1906 if (unlikely(server_cksum == client_cksum)) {
1907 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1912 if (exp->exp_obd->obd_checksum_dump)
1913 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1916 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1919 if (cksum_type != server_cksum_type)
1920 msg = "the server may have not used the checksum type specified"
1921 " in the original request - likely a protocol problem";
1923 msg = "should have changed on the client or in transit";
1925 start = local_nb[0].lnb_file_offset;
1926 end = local_nb[npages-1].lnb_file_offset +
1927 local_nb[npages-1].lnb_len - 1;
1929 LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1930 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1931 " %x (type %x), server csum %x (type %x)\n",
1932 exp->exp_obd->obd_name,
1933 msg, libcfs_nid2str(peer->nid),
1934 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1935 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1936 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1938 start, end, client_cksum, cksum_type, server_cksum,
1944 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1945 unsigned char *buf, int size)
1947 int i, off, len, copied = size;
1950 for (i = 0; i < npages; i++) {
1951 off = local[i].lnb_page_offset & ~PAGE_MASK;
1952 len = local[i].lnb_len;
1954 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1959 ptr = kmap_atomic(local[i].lnb_page);
1960 memcpy(buf + off, ptr, len);
1965 return copied - size;
1968 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1969 struct niobuf_local *local_nb,
1970 int npages, int opc,
1971 obd_dif_csum_fn *fn,
1975 enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1976 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1977 const char *obd_name = tgt->lut_obd->obd_name;
1978 struct ahash_request *req;
1979 unsigned int bufsize;
1980 unsigned char *buffer;
1981 struct page *__page;
1984 int used_number = 0;
1990 __page = alloc_page(GFP_KERNEL);
1994 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1996 CERROR("%s: unable to initialize checksum hash %s\n",
1997 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1998 return PTR_ERR(req);
2001 buffer = kmap(__page);
2002 guard_start = (__u16 *)buffer;
2003 guard_number = PAGE_SIZE / sizeof(*guard_start);
2004 for (i = 0; i < npages; i++) {
2005 /* corrupt the data before we compute the checksum, to
2006 * simulate a client->OST data error */
2007 if (i == 0 && opc == OST_WRITE &&
2008 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2009 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2010 int len = local_nb[i].lnb_len;
2011 struct page *np = tgt_page_to_corrupt;
2014 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2015 char *ptr2 = page_address(np);
2017 memcpy(ptr2 + off, ptr + off, len);
2018 memcpy(ptr2 + off, "bad3", min(4, len));
2021 /* LU-8376 to preserve original index for
2022 * display in dump_all_bulk_pages() */
2025 cfs_crypto_hash_update_page(req, np, off,
2029 CERROR("%s: can't alloc page for corruption\n",
2035 * The left guard number should be able to hold checksums of a
2038 if (t10_cksum_type && opc == OST_READ &&
2039 local_nb[i].lnb_guard_disk) {
2040 used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2041 if (used > (guard_number - used_number)) {
2045 memcpy(guard_start + used_number,
2046 local_nb[i].lnb_guards,
2047 used * sizeof(*local_nb[i].lnb_guards));
2049 rc = obd_page_dif_generate_buffer(obd_name,
2050 local_nb[i].lnb_page,
2051 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2052 local_nb[i].lnb_len, guard_start + used_number,
2053 guard_number - used_number, &used, sector_size,
2059 LASSERT(used <= MAX_GUARD_NUMBER);
2061 * If disk support T10PI checksum, copy guards to local_nb.
2062 * If the write is partial page, do not use the guards for bio
2063 * submission since the data might not be full-sector. The bio
2064 * guards will be generated later based on the full sectors. If
2065 * the sector size is 512B rather than 4 KB, or the page size
2066 * is larger than 4KB, this might drop some useful guards for
2067 * partial page write, but it will only add minimal extra time
2068 * of checksum calculation.
2070 if (t10_cksum_type && opc == OST_WRITE &&
2071 local_nb[i].lnb_len == PAGE_SIZE) {
2072 local_nb[i].lnb_guard_rpc = 1;
2073 memcpy(local_nb[i].lnb_guards,
2074 guard_start + used_number,
2075 used * sizeof(*local_nb[i].lnb_guards));
2078 used_number += used;
2079 if (used_number == guard_number) {
2080 cfs_crypto_hash_update_page(req, __page, 0,
2081 used_number * sizeof(*guard_start));
2085 /* corrupt the data after we compute the checksum, to
2086 * simulate an OST->client data error */
2087 if (unlikely(i == 0 && opc == OST_READ &&
2088 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2089 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2090 int len = local_nb[i].lnb_len;
2091 struct page *np = tgt_page_to_corrupt;
2094 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2095 char *ptr2 = page_address(np);
2097 memcpy(ptr2 + off, ptr + off, len);
2098 memcpy(ptr2 + off, "bad4", min(4, len));
2101 /* LU-8376 to preserve original index for
2102 * display in dump_all_bulk_pages() */
2105 cfs_crypto_hash_update_page(req, np, off,
2109 CERROR("%s: can't alloc page for corruption\n",
2118 if (used_number != 0)
2119 cfs_crypto_hash_update_page(req, __page, 0,
2120 used_number * sizeof(*guard_start));
2122 bufsize = sizeof(cksum);
2123 rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2128 __free_page(__page);
2132 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2133 enum cksum_types cksum_type,
2134 struct niobuf_local *local_nb,
2135 int npages, int opc, u32 *check_sum)
2137 obd_dif_csum_fn *fn = NULL;
2138 int sector_size = 0;
2142 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
2145 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2146 opc, fn, sector_size,
2149 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2150 cksum_type, check_sum);
2154 int tgt_brw_read(struct tgt_session_info *tsi)
2156 struct ptlrpc_request *req = tgt_ses_req(tsi);
2157 struct ptlrpc_bulk_desc *desc = NULL;
2158 struct obd_export *exp = tsi->tsi_exp;
2159 struct niobuf_remote *remote_nb;
2160 struct niobuf_local *local_nb;
2161 struct obd_ioobj *ioo;
2162 struct ost_body *body, *repbody;
2163 struct lustre_handle lockh = { 0 };
2164 int npages, nob = 0, rc, i, no_reply = 0,
2166 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2167 const char *obd_name = exp->exp_obd->obd_name;
2171 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2172 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2173 CERROR("%s: deny read request from %s to portal %u\n",
2174 tgt_name(tsi->tsi_tgt),
2175 obd_export_nid2str(req->rq_export),
2176 ptlrpc_req2svc(req)->srv_req_portal);
2180 req->rq_bulk_read = 1;
2182 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2185 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2186 cfs_fail_val : (obd_timeout + 1) / 4);
2188 /* Check if there is eviction in progress, and if so, wait for it to
2190 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2191 /* We do not care how long it takes */
2193 exp->exp_obd->obd_evict_inprogress_waitq,
2194 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2197 /* There must be big cache in current thread to process this request
2198 * if it is NULL then something went wrong and it wasn't allocated,
2199 * report -ENOMEM in that case */
2203 body = tsi->tsi_ost_body;
2204 LASSERT(body != NULL);
2206 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2207 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2209 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2210 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2212 local_nb = tbc->local;
2214 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2220 * If getting the lock took more time than
2221 * client was willing to wait, drop it. b=11330
2223 if (ktime_get_real_seconds() > req->rq_deadline ||
2224 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2226 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2227 libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2228 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2229 req->rq_deadline - req->rq_arrival_time.tv_sec);
2230 GOTO(out_lock, rc = -ETIMEDOUT);
2233 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2234 repbody->oa = body->oa;
2236 npages = PTLRPC_MAX_BRW_PAGES;
2237 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2238 ioo, remote_nb, &npages, local_nb);
2242 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2243 body->oa.o_flags & OBD_FL_SHORT_IO) {
2246 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2247 PTLRPC_BULK_PUT_SOURCE,
2249 &ptlrpc_bulk_kiov_nopin_ops);
2251 GOTO(out_commitrw, rc = -ENOMEM);
2255 npages_read = npages;
2256 for (i = 0; i < npages; i++) {
2257 int page_rc = local_nb[i].lnb_rc;
2266 if (page_rc != 0 && desc != NULL) { /* some data! */
2267 LASSERT(local_nb[i].lnb_page != NULL);
2268 desc->bd_frag_ops->add_kiov_frag
2269 (desc, local_nb[i].lnb_page,
2270 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2274 if (page_rc != local_nb[i].lnb_len) { /* short read */
2275 local_nb[i].lnb_len = page_rc;
2276 npages_read = i + (page_rc != 0 ? 1 : 0);
2277 /* All subsequent pages should be 0 */
2278 while (++i < npages)
2279 LASSERT(local_nb[i].lnb_rc == 0);
2284 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2285 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2286 body->oa.o_flags : 0;
2287 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2289 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2291 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2293 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2294 local_nb, npages_read, OST_READ,
2295 &repbody->oa.o_cksum);
2297 GOTO(out_commitrw, rc);
2298 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2299 repbody->oa.o_cksum);
2301 /* if a resend it could be for a cksum error, so check Server
2302 * cksum with returned Client cksum (this should even cover
2303 * zero-cksum case) */
2304 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2305 (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2306 check_read_checksum(local_nb, npages_read, exp,
2307 &body->oa, &req->rq_peer,
2309 repbody->oa.o_cksum, cksum_type);
2311 repbody->oa.o_valid = 0;
2313 /* We're finishing using body->oa as an input variable */
2315 /* Check if client was evicted while we were doing i/o before touching
2318 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2319 body->oa.o_flags & OBD_FL_SHORT_IO) {
2320 unsigned char *short_io_buf;
2323 short_io_buf = req_capsule_server_get(&req->rq_pill,
2325 short_io_size = req_capsule_get_size(&req->rq_pill,
2328 rc = tgt_pages2shortio(local_nb, npages_read,
2329 short_io_buf, short_io_size);
2331 req_capsule_shrink(&req->rq_pill,
2334 rc = rc > 0 ? 0 : rc;
2335 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2336 rc = target_bulk_io(exp, desc);
2340 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2341 body->oa.o_flags & OBD_FL_SHORT_IO)
2342 req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2347 /* Must commit after prep above in all cases */
2348 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2349 remote_nb, npages, local_nb, rc);
2351 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2353 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2354 ptlrpc_free_bulk(desc);
2359 ptlrpc_lprocfs_brw(req, nob);
2360 } else if (no_reply) {
2361 req->rq_no_reply = 1;
2362 /* reply out callback would free */
2363 ptlrpc_req_drop_rs(req);
2364 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2365 "client will retry: rc %d\n",
2367 obd_uuid2str(&exp->exp_client_uuid),
2368 obd_export_nid2str(exp), rc);
2370 /* send a bulk after reply to simulate a network delay or reordering
2371 * by a router - Note that !desc implies short io, so there is no bulk
2373 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2375 CDEBUG(D_INFO, "reorder BULK\n");
2378 target_bulk_io(exp, desc);
2379 ptlrpc_free_bulk(desc);
2384 EXPORT_SYMBOL(tgt_brw_read);
2386 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2387 unsigned char *buf, unsigned int size)
2392 for (i = 0; i < npages; i++) {
2393 off = local[i].lnb_page_offset & ~PAGE_MASK;
2394 len = local[i].lnb_len;
2399 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2401 ptr = kmap_atomic(local[i].lnb_page);
2404 memcpy(ptr + off, buf, len < size ? len : size);
2412 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2413 struct ptlrpc_bulk_desc *desc,
2414 struct niobuf_local *local_nb, int npages,
2415 u32 client_cksum, u32 server_cksum,
2418 struct obd_export *exp = req->rq_export;
2419 struct ost_body *body;
2423 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2424 LASSERT(body != NULL);
2426 if (desc && req->rq_peer.nid != desc->bd_sender) {
2428 router = libcfs_nid2str(desc->bd_sender);
2431 if (exp->exp_obd->obd_checksum_dump)
2432 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2436 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2437 client_cksum, server_cksum);
2441 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2442 DFID" object "DOSTID" extent [%llu-%llu"
2443 "]: client csum %x, server csum %x\n",
2444 exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2446 body->oa.o_valid & OBD_MD_FLFID ?
2447 body->oa.o_parent_seq : (__u64)0,
2448 body->oa.o_valid & OBD_MD_FLFID ?
2449 body->oa.o_parent_oid : 0,
2450 body->oa.o_valid & OBD_MD_FLFID ?
2451 body->oa.o_parent_ver : 0,
2452 POSTID(&body->oa.o_oi),
2453 local_nb[0].lnb_file_offset,
2454 local_nb[npages-1].lnb_file_offset +
2455 local_nb[npages - 1].lnb_len - 1,
2456 client_cksum, server_cksum);
2459 int tgt_brw_write(struct tgt_session_info *tsi)
2461 struct ptlrpc_request *req = tgt_ses_req(tsi);
2462 struct ptlrpc_bulk_desc *desc = NULL;
2463 struct obd_export *exp = req->rq_export;
2464 struct niobuf_remote *remote_nb;
2465 struct niobuf_local *local_nb;
2466 struct obd_ioobj *ioo;
2467 struct ost_body *body, *repbody;
2468 struct lustre_handle lockh = {0};
2470 int objcount, niocount, npages;
2472 enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2473 bool no_reply = false, mmap;
2474 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2475 bool wait_sync = false;
2476 const char *obd_name = exp->exp_obd->obd_name;
2480 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2481 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2482 CERROR("%s: deny write request from %s to portal %u\n",
2483 tgt_name(tsi->tsi_tgt),
2484 obd_export_nid2str(req->rq_export),
2485 ptlrpc_req2svc(req)->srv_req_portal);
2486 RETURN(err_serious(-EPROTO));
2489 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2490 RETURN(err_serious(-ENOSPC));
2491 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2492 RETURN(err_serious(-EROFS));
2494 req->rq_bulk_write = 1;
2496 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2497 RETURN(err_serious(-EIO));
2498 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2499 RETURN(err_serious(-EFAULT));
2501 /* pause before transaction has been started */
2502 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2503 cfs_fail_val : (obd_timeout + 1) / 4);
2505 /* Delay write commit to show stale size information */
2506 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2508 /* There must be big cache in current thread to process this request
2509 * if it is NULL then something went wrong and it wasn't allocated,
2510 * report -ENOMEM in that case */
2514 body = tsi->tsi_ost_body;
2515 LASSERT(body != NULL);
2517 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2518 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2520 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2521 RCL_CLIENT) / sizeof(*ioo);
2523 for (niocount = i = 0; i < objcount; i++)
2524 niocount += ioo[i].ioo_bufcnt;
2526 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2527 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2528 if (niocount != req_capsule_get_size(&req->rq_pill,
2529 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2531 RETURN(err_serious(-EPROTO));
2533 if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2534 ptlrpc_connection_is_local(exp->exp_connection))
2535 memory_pressure_set();
2537 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2538 niocount * sizeof(*rcs));
2539 rc = req_capsule_server_pack(&req->rq_pill);
2541 GOTO(out, rc = err_serious(rc));
2543 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2544 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2546 local_nb = tbc->local;
2548 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2554 * If getting the lock took more time than
2555 * client was willing to wait, drop it. b=11330
2557 if (ktime_get_real_seconds() > req->rq_deadline ||
2558 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2560 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2561 tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2562 POSTID(&ioo->ioo_oid),
2563 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2564 req->rq_deadline - req->rq_arrival_time.tv_sec);
2565 GOTO(out_lock, rc = -ETIMEDOUT);
2568 /* Because we already sync grant info with client when reconnect,
2569 * grant info will be cleared for resent req, then fed_grant and
2570 * total_grant will not be modified in following preprw_write */
2571 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2572 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2573 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2576 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2577 if (repbody == NULL)
2578 GOTO(out_lock, rc = -ENOMEM);
2579 repbody->oa = body->oa;
2581 npages = PTLRPC_MAX_BRW_PAGES;
2582 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2583 objcount, ioo, remote_nb, &npages, local_nb);
2586 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2587 body->oa.o_flags & OBD_FL_SHORT_IO) {
2588 unsigned int short_io_size;
2589 unsigned char *short_io_buf;
2591 short_io_size = req_capsule_get_size(&req->rq_pill,
2594 short_io_buf = req_capsule_client_get(&req->rq_pill,
2596 CDEBUG(D_INFO, "Client use short io for data transfer,"
2597 " size = %d\n", short_io_size);
2599 /* Copy short io buf to pages */
2600 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2604 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2605 PTLRPC_BULK_GET_SINK,
2607 &ptlrpc_bulk_kiov_nopin_ops);
2609 GOTO(skip_transfer, rc = -ENOMEM);
2611 /* NB Having prepped, we must commit... */
2612 for (i = 0; i < npages; i++)
2613 desc->bd_frag_ops->add_kiov_frag(desc,
2614 local_nb[i].lnb_page,
2615 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2616 local_nb[i].lnb_len);
2618 rc = sptlrpc_svc_prep_bulk(req, desc);
2620 GOTO(skip_transfer, rc);
2622 rc = target_bulk_io(exp, desc);
2628 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2629 static int cksum_counter;
2631 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2632 cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2634 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2635 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2636 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2639 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2640 local_nb, npages, OST_WRITE,
2641 &repbody->oa.o_cksum);
2643 GOTO(out_commitrw, rc);
2647 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2648 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2649 body->oa.o_flags & OBD_FL_MMAP);
2651 tgt_warn_on_cksum(req, desc, local_nb, npages,
2653 repbody->oa.o_cksum, mmap);
2655 } else if ((cksum_counter & (-cksum_counter)) ==
2657 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2658 cksum_counter, libcfs_id2str(req->rq_peer),
2659 repbody->oa.o_cksum);
2664 /* Must commit after prep above in all cases */
2665 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2666 objcount, ioo, remote_nb, npages, local_nb, rc);
2667 if (rc == -ENOTCONN)
2668 /* quota acquire process has been given up because
2669 * either the client has been evicted or the client
2670 * has timed out the request already */
2673 for (i = 0; i < niocount; i++) {
2674 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2680 * Disable sending mtime back to the client. If the client locked the
2681 * whole object, then it has already updated the mtime on its side,
2682 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2684 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2689 /* set per-requested niobuf return codes */
2690 for (i = j = 0; i < niocount; i++) {
2691 int len = remote_nb[i].rnb_len;
2696 LASSERT(j < npages);
2697 if (local_nb[j].lnb_rc < 0)
2698 rcs[i] = local_nb[j].lnb_rc;
2699 len -= local_nb[j].lnb_len;
2704 LASSERT(j == npages);
2705 ptlrpc_lprocfs_brw(req, nob);
2708 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2710 ptlrpc_free_bulk(desc);
2712 if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2713 req->rq_no_reply = 1;
2714 /* reply out callback would free */
2715 ptlrpc_req_drop_rs(req);
2716 if (!exp->exp_obd->obd_no_transno)
2717 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2718 " client will retry: rc = %d\n",
2720 obd_uuid2str(&exp->exp_client_uuid),
2721 obd_export_nid2str(exp), rc);
2723 memory_pressure_clr();
2726 EXPORT_SYMBOL(tgt_brw_write);
2728 /* Check if request can be reconstructed from saved reply data
2729 * A copy of the reply data is returned in @trd if the pointer is not NULL
2731 int req_can_reconstruct(struct ptlrpc_request *req,
2732 struct tg_reply_data *trd)
2734 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2735 struct lsd_client_data *lcd = ted->ted_lcd;
2738 if (tgt_is_multimodrpcs_client(req->rq_export))
2739 return tgt_lookup_reply(req, trd);
2741 mutex_lock(&ted->ted_lcd_lock);
2742 found = req->rq_xid == lcd->lcd_last_xid ||
2743 req->rq_xid == lcd->lcd_last_close_xid;
2745 if (found && trd != NULL) {
2746 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2747 trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2748 trd->trd_reply.lrd_transno =
2749 lcd->lcd_last_close_transno;
2750 trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2752 trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2753 trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2754 trd->trd_reply.lrd_result = lcd->lcd_last_result;
2755 trd->trd_reply.lrd_data = lcd->lcd_last_data;
2756 trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2757 trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2758 trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2759 trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2762 mutex_unlock(&ted->ted_lcd_lock);
2766 EXPORT_SYMBOL(req_can_reconstruct);