4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; if not, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
24 * Copyright (c) 2013, 2017, Intel Corporation.
27 * lustre/target/tgt_handler.c
29 * Lustre Unified Target request handler code
31 * Author: Brian Behlendorf <behlendorf1@llnl.gov>
32 * Author: Mikhail Pershin <mike.pershin@intel.com>
35 #define DEBUG_SUBSYSTEM S_CLASS
37 #include <linux/user_namespace.h>
38 #ifdef HAVE_UIDGID_HEADER
39 # include <linux/uidgid.h>
43 #include <obd_class.h>
44 #include <obd_cksum.h>
45 #include <lustre_lfsck.h>
46 #include <lustre_nodemap.h>
47 #include <lustre_acl.h>
49 #include "tgt_internal.h"
51 char *tgt_name(struct lu_target *tgt)
53 LASSERT(tgt->lut_obd != NULL);
54 return tgt->lut_obd->obd_name;
56 EXPORT_SYMBOL(tgt_name);
59 * Generic code handling requests that have struct mdt_body passed in:
61 * - extract mdt_body from request and save it in @tsi, if present;
63 * - create lu_object, corresponding to the fid in mdt_body, and save it in
66 * - if 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 rc = tgt_mdt_body_unpack(tsi, flags);
327 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
329 rc = tgt_ost_body_unpack(tsi, flags);
335 if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
338 if (flags & HAS_KEY) {
339 struct ldlm_request *dlm_req;
341 LASSERT(h->th_fmt != NULL);
343 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
344 if (dlm_req != NULL) {
345 union ldlm_wire_policy_data *policy =
346 &dlm_req->lock_desc.l_policy_data;
348 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
350 (policy->l_inodebits.bits |
351 policy->l_inodebits.try_bits) == 0)) {
353 * Lock without inodebits makes no sense and
354 * will oops later in ldlm. If client miss to
355 * set such bits, do not trigger ASSERTION.
357 * For liblustre flock case, it maybe zero.
361 tsi->tsi_dlm_req = dlm_req;
367 tsi->tsi_preprocessed = 1;
372 * Invoke handler for this request opc. Also do necessary preprocessing
373 * (according to handler ->th_flags), and post-processing (setting of
374 * ->last_{xid,committed}).
376 static int tgt_handle_request0(struct tgt_session_info *tsi,
377 struct tgt_handler *h,
378 struct ptlrpc_request *req)
382 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
387 /* When dealing with sec context requests, no export is associated yet,
388 * because these requests are sent before *_CONNECT requests.
389 * A NULL req->rq_export means the normal *_common_slice handlers will
390 * not be called, because there is no reference to the target.
391 * So deal with them by hand and jump directly to target_send_reply().
395 case SEC_CTX_INIT_CONT:
397 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
402 * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
403 * to put same checks into handlers like mdt_close(), mdt_reint(),
404 * etc., without talking to mdt authors first. Checking same thing
405 * there again is useless and returning 0 error without packing reply
406 * is buggy! Handlers either pack reply or return error.
408 * We return 0 here and do not send any reply in order to emulate
409 * network failure. Do not send any reply in case any of NET related
410 * fail_id has occured.
412 if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
414 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
415 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
418 rc = tgt_request_preprocess(tsi, h, req);
419 /* pack reply if reply format is fixed */
420 if (rc == 0 && h->th_flags & HAS_REPLY) {
422 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
424 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
426 tsi->tsi_mdt_body->mbo_eadatasize);
427 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
429 req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
431 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
432 req_capsule_set_size(tsi->tsi_pill,
433 &RMF_ACL, RCL_SERVER,
434 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
436 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
438 struct niobuf_remote *remote_nb =
439 req_capsule_client_get(tsi->tsi_pill,
441 struct ost_body *body = tsi->tsi_ost_body;
443 req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
445 (body->oa.o_valid & OBD_MD_FLFLAGS &&
446 body->oa.o_flags & OBD_FL_SHORT_IO) ?
447 remote_nb[0].rnb_len : 0);
450 rc = req_capsule_server_pack(tsi->tsi_pill);
453 if (likely(rc == 0)) {
455 * Process request, there can be two types of rc:
456 * 1) errors with msg unpack/pack, other failures outside the
457 * operation itself. This is counted as serious errors;
458 * 2) errors during fs operation, should be placed in rq_status
462 if (!is_serious(rc) &&
463 !req->rq_no_reply && req->rq_reply_state == NULL) {
464 DEBUG_REQ(D_ERROR, req,
465 "%s: %s handler did not pack reply but returned no error",
466 tgt_name(tsi->tsi_tgt), h->th_name);
469 serious = is_serious(rc);
470 rc = clear_serious(rc);
478 * ELDLM_* codes which > 0 should be in rq_status only as well as
479 * all non-serious errors.
481 if (rc > 0 || !serious)
484 LASSERT(current->journal_info == NULL);
486 if (likely(rc == 0 && req->rq_export))
487 target_committed_to_req(req);
490 target_send_reply(req, rc, tsi->tsi_reply_fail_id);
494 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
495 struct obd_device *obd, int *process)
497 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
504 case MDS_SYNC: /* used in unmounting */
518 case MDS_HSM_PROGRESS:
519 case MDS_HSM_STATE_SET:
520 case MDS_HSM_REQUEST:
521 *process = target_queue_recovery_request(req, obd);
525 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
532 * Handle recovery. Return:
533 * +1: continue request processing;
534 * -ve: abort immediately with the given error code;
535 * 0: send reply with error code in req->rq_status;
537 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
541 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
546 case SEC_CTX_INIT_CONT:
551 if (!req->rq_export->exp_obd->obd_replayable)
554 /* sanity check: if the xid matches, the request must be marked as a
555 * resent or replayed */
556 if (req_can_reconstruct(req, NULL) == 1) {
557 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
558 (MSG_RESENT | MSG_REPLAY))) {
559 DEBUG_REQ(D_WARNING, req,
560 "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
562 lustre_msg_get_flags(req->rq_reqmsg));
563 req->rq_status = -ENOTCONN;
567 /* else: note the opposite is not always true; a RESENT req after a
568 * failover will usually not match the last_xid, since it was likely
569 * never committed. A REPLAYed request will almost never match the
570 * last xid, however it could for a committed, but still retained,
573 /* Check for aborted recovery... */
574 if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
578 DEBUG_REQ(D_INFO, req, "Got new replay");
579 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
581 if (rc != 0 || !should_process)
583 else if (should_process < 0) {
584 req->rq_status = should_process;
585 rc = ptlrpc_error(req);
592 /* Initial check for request, it is validation mostly */
593 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
595 struct tgt_handler *h;
596 struct tgt_opc_slice *s;
597 struct lu_target *tgt;
598 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
602 tgt = class_exp2tgt(req->rq_export);
603 if (unlikely(tgt == NULL)) {
604 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
605 class_exp2obd(req->rq_export)->obd_name);
606 RETURN(ERR_PTR(-EINVAL));
609 for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
610 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
613 /* opcode was not found in slice */
614 if (unlikely(s->tos_hs == NULL)) {
615 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
617 RETURN(ERR_PTR(-ENOTSUPP));
620 LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
621 h = s->tos_hs + (opc - s->tos_opc_start);
622 if (unlikely(h->th_opc == 0)) {
623 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
624 RETURN(ERR_PTR(-ENOTSUPP));
630 static int process_req_last_xid(struct ptlrpc_request *req)
634 struct obd_export *exp = req->rq_export;
635 struct tg_export_data *ted = &exp->exp_target_data;
636 bool need_lock = tgt_is_multimodrpcs_client(exp);
640 mutex_lock(&ted->ted_lcd_lock);
641 /* check request's xid is consistent with export's last_xid */
642 last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
643 if (last_xid > exp->exp_last_xid)
644 exp->exp_last_xid = last_xid;
646 if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
647 /* Some request is allowed to be sent during replay,
648 * such as OUT update requests, FLD requests, so it
649 * is possible that replay requests has smaller XID
650 * than the exp_last_xid.
652 * Some non-replay requests may have smaller XID as
655 * - Client send a no_resend RPC, like statfs;
656 * - The RPC timedout (or some other error) on client,
657 * then it's removed from the unreplied list;
658 * - Client send some other request to bump the
659 * exp_last_xid on server;
660 * - The former RPC got chance to be processed;
662 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
665 DEBUG_REQ(D_WARNING, req,
666 "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
667 req->rq_xid, exp->exp_last_xid, rc);
672 /* The "last_xid" is the minimum xid among unreplied requests,
673 * if the request is from the previous connection, its xid can
674 * still be larger than "exp_last_xid", then the above check of
675 * xid is not enough to determine whether the request is delayed.
677 * For example, if some replay request was delayed and caused
678 * timeout at client and the replay is restarted, the delayed
679 * replay request will have the larger xid than "exp_last_xid"
681 if (req->rq_export->exp_conn_cnt >
682 lustre_msg_get_conn_cnt(req->rq_reqmsg))
683 GOTO(out, rc = -ESTALE);
685 /* try to release in-memory reply data */
686 if (tgt_is_multimodrpcs_client(exp)) {
687 tgt_handle_received_xid(exp, last_xid);
688 rc = tgt_handle_tag(req);
693 mutex_unlock(&ted->ted_lcd_lock);
698 int tgt_request_handle(struct ptlrpc_request *req)
700 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
702 struct lustre_msg *msg = req->rq_reqmsg;
703 struct tgt_handler *h;
704 struct lu_target *tgt;
705 int request_fail_id = 0;
706 __u32 opc = lustre_msg_get_opc(msg);
707 struct obd_device *obd;
709 bool is_connect = false;
712 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
713 if (cfs_fail_val == 0 &&
714 lustre_msg_get_opc(msg) != OBD_PING &&
715 lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
716 struct l_wait_info lwi = { 0 };
720 l_wait_event(cfs_race_waitq, (cfs_race_state == 1),
725 /* Refill the context, to make sure all thread keys are allocated */
726 lu_env_refill(req->rq_svc_thread->t_env);
728 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
729 tsi->tsi_pill = &req->rq_pill;
730 tsi->tsi_env = req->rq_svc_thread->t_env;
732 /* if request has export then get handlers slice from corresponding
733 * target, otherwise that should be connect operation */
734 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
735 opc == MGS_CONNECT) {
737 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
738 rc = target_handle_connect(req);
740 rc = ptlrpc_error(req);
743 /* recovery-small test 18c asks to drop connect reply */
744 if (unlikely(opc == OST_CONNECT &&
745 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
749 if (unlikely(!class_connected_export(req->rq_export))) {
750 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
751 opc == SEC_CTX_FINI) {
752 /* sec context initialization has to be handled
753 * by hand in tgt_handle_request0() */
754 tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
756 GOTO(handle_recov, rc = 0);
758 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
759 opc, libcfs_id2str(req->rq_peer));
760 req->rq_status = -ENOTCONN;
761 rc = ptlrpc_error(req);
765 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
766 tsi->tsi_exp = req->rq_export;
767 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
768 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
770 tsi->tsi_jobid = NULL;
773 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
774 class_exp2obd(req->rq_export)->obd_name);
775 req->rq_status = -EINVAL;
776 rc = ptlrpc_error(req);
780 /* Skip last_xid processing for the recovery thread, otherwise, the
781 * last_xid on same request could be processed twice: first time when
782 * processing the incoming request, second time when the request is
783 * being processed by recovery thread. */
784 obd = class_exp2obd(req->rq_export);
786 /* reset the exp_last_xid on each connection. */
787 req->rq_export->exp_last_xid = 0;
788 } else if (obd->obd_recovery_data.trd_processing_task !=
790 rc = process_req_last_xid(req);
793 rc = ptlrpc_error(req);
798 request_fail_id = tgt->lut_request_fail_id;
799 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
801 h = tgt_handler_find_check(req);
803 req->rq_status = PTR_ERR(h);
804 rc = ptlrpc_error(req);
808 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
811 if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
812 CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
815 rc = lustre_msg_check_version(msg, h->th_version);
817 DEBUG_REQ(D_ERROR, req,
818 "%s: drop malformed request version=%08x expect=%08x",
819 tgt_name(tgt), lustre_msg_get_version(msg),
821 req->rq_status = -EINVAL;
822 rc = ptlrpc_error(req);
827 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
828 if (likely(rc == 1)) {
829 rc = tgt_handle_request0(tsi, h, req);
835 req_capsule_fini(tsi->tsi_pill);
836 if (tsi->tsi_corpus != NULL) {
837 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
838 tsi->tsi_corpus = NULL;
842 EXPORT_SYMBOL(tgt_request_handle);
844 /** Assign high priority operations to the request if needed. */
845 int tgt_hpreq_handler(struct ptlrpc_request *req)
847 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
848 struct tgt_handler *h;
853 if (req->rq_export == NULL)
856 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
857 tsi->tsi_pill = &req->rq_pill;
858 tsi->tsi_env = req->rq_svc_thread->t_env;
859 tsi->tsi_tgt = class_exp2tgt(req->rq_export);
860 tsi->tsi_exp = req->rq_export;
862 h = tgt_handler_find_check(req);
868 rc = tgt_request_preprocess(tsi, h, req);
869 if (unlikely(rc != 0))
872 if (h->th_hp != NULL)
876 EXPORT_SYMBOL(tgt_hpreq_handler);
878 void tgt_counter_incr(struct obd_export *exp, int opcode)
880 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
881 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
882 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
884 EXPORT_SYMBOL(tgt_counter_incr);
887 * Unified target generic handlers.
890 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
892 struct lu_target *tgt = class_exp2tgt(exp);
893 struct sptlrpc_flavor flvr;
897 LASSERT(tgt->lut_obd);
898 LASSERT(tgt->lut_slice);
900 /* always allow ECHO client */
901 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
902 LUSTRE_ECHO_NAME) == 0)) {
903 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
907 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
908 read_lock(&tgt->lut_sptlrpc_lock);
909 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
913 read_unlock(&tgt->lut_sptlrpc_lock);
915 spin_lock(&exp->exp_lock);
916 exp->exp_sp_peer = req->rq_sp_from;
917 exp->exp_flvr = flvr;
919 /* when on mgs, if no restriction is set, or if client
920 * is loopback, allow any flavor */
921 if ((strcmp(exp->exp_obd->obd_type->typ_name,
922 LUSTRE_MGS_NAME) == 0) &&
923 (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
924 LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
926 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
928 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
929 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
930 CERROR("%s: unauthorized rpc flavor %x from %s, "
931 "expect %x\n", tgt_name(tgt),
933 libcfs_nid2str(req->rq_peer.nid),
934 exp->exp_flvr.sf_rpc);
937 spin_unlock(&exp->exp_lock);
939 if (exp->exp_sp_peer != req->rq_sp_from) {
940 CERROR("%s: RPC source %s doesn't match %s\n",
942 sptlrpc_part2name(req->rq_sp_from),
943 sptlrpc_part2name(exp->exp_sp_peer));
946 rc = sptlrpc_target_export_check(exp, req);
953 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
955 struct sptlrpc_rule_set tmp_rset;
958 if (unlikely(tgt == NULL)) {
959 CERROR("No target passed");
963 sptlrpc_rule_set_init(&tmp_rset);
964 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
966 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
971 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
973 write_lock(&tgt->lut_sptlrpc_lock);
974 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
975 tgt->lut_sptlrpc_rset = tmp_rset;
976 write_unlock(&tgt->lut_sptlrpc_lock);
980 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
982 int tgt_connect(struct tgt_session_info *tsi)
984 struct ptlrpc_request *req = tgt_ses_req(tsi);
985 struct obd_connect_data *reply;
990 /* XXX: better to call this check right after getting new export but
991 * before last_rcvd slot allocation to avoid server load upon insecure
992 * connects. This is to be fixed after unifiyng all targets.
994 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
998 /* To avoid exposing partially initialized connection flags, changes up
999 * to this point have been staged in reply->ocd_connect_flags. Now that
1000 * connection handling has completed successfully, atomically update
1001 * the connect flags in the shared export data structure. LU-1623 */
1002 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1003 spin_lock(&tsi->tsi_exp->exp_lock);
1004 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1005 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1006 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1007 reply->ocd_connect_flags2;
1008 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1009 spin_unlock(&tsi->tsi_exp->exp_lock);
1011 if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1012 LUSTRE_MDT_NAME) == 0) {
1013 rc = req_check_sepol(tsi->tsi_pill);
1020 obd_disconnect(class_export_get(tsi->tsi_exp));
1023 EXPORT_SYMBOL(tgt_connect);
1025 int tgt_disconnect(struct tgt_session_info *tsi)
1031 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1033 rc = target_handle_disconnect(tgt_ses_req(tsi));
1035 RETURN(err_serious(rc));
1039 EXPORT_SYMBOL(tgt_disconnect);
1042 * Unified target OBD handlers
1044 int tgt_obd_ping(struct tgt_session_info *tsi)
1050 /* The target-specific part of OBD_PING request handling.
1051 * It controls Filter Modification Data (FMD) expiration each time
1054 * Valid only for replayable targets, e.g. MDT and OFD
1056 if (tsi->tsi_exp->exp_obd->obd_replayable)
1057 tgt_fmd_expire(tsi->tsi_exp);
1059 rc = req_capsule_server_pack(tsi->tsi_pill);
1061 RETURN(err_serious(rc));
1065 EXPORT_SYMBOL(tgt_obd_ping);
1067 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1069 return err_serious(-EOPNOTSUPP);
1072 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1074 struct ptlrpc_request *req = tgt_ses_req(tsi);
1075 struct obd_export *exp = req->rq_export;
1076 struct ptlrpc_bulk_desc *desc;
1082 desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1083 PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1084 MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1088 for (i = 0; i < rdbuf->rb_nbufs; i++)
1089 desc->bd_frag_ops->add_iov_frag(desc,
1090 rdbuf->rb_bufs[i].lb_buf,
1091 rdbuf->rb_bufs[i].lb_len);
1093 rc = target_bulk_io(exp, desc);
1094 ptlrpc_free_bulk(desc);
1097 EXPORT_SYMBOL(tgt_send_buffer);
1099 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1101 struct ptlrpc_request *req = tgt_ses_req(tsi);
1102 struct obd_export *exp = req->rq_export;
1103 struct ptlrpc_bulk_desc *desc;
1111 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1112 PTLRPC_BULK_PUT_SOURCE |
1113 PTLRPC_BULK_BUF_KIOV,
1115 &ptlrpc_bulk_kiov_pin_ops);
1119 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1120 /* old client requires reply size in it's PAGE_SIZE,
1121 * which is rdpg->rp_count */
1122 nob = rdpg->rp_count;
1124 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1125 i++, tmpcount -= tmpsize) {
1126 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1127 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1131 LASSERT(desc->bd_nob == nob);
1132 rc = target_bulk_io(exp, desc);
1133 ptlrpc_free_bulk(desc);
1136 EXPORT_SYMBOL(tgt_sendpage);
1139 * OBD_IDX_READ handler
1141 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1143 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1144 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1145 struct idx_info *req_ii, *rep_ii;
1150 memset(rdpg, 0, sizeof(*rdpg));
1151 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1153 /* extract idx_info buffer from request & reply */
1154 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1155 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1156 RETURN(err_serious(-EPROTO));
1158 rc = req_capsule_server_pack(tsi->tsi_pill);
1160 RETURN(err_serious(rc));
1162 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1164 RETURN(err_serious(-EFAULT));
1165 rep_ii->ii_magic = IDX_INFO_MAGIC;
1167 /* extract hash to start with */
1168 rdpg->rp_hash = req_ii->ii_hash_start;
1170 /* extract requested attributes */
1171 rdpg->rp_attrs = req_ii->ii_attrs;
1173 /* check that fid packed in request is valid and supported */
1174 if (!fid_is_sane(&req_ii->ii_fid))
1176 rep_ii->ii_fid = req_ii->ii_fid;
1179 rep_ii->ii_flags = req_ii->ii_flags;
1181 /* compute number of pages to allocate, ii_count is the number of 4KB
1183 if (req_ii->ii_count <= 0)
1184 GOTO(out, rc = -EFAULT);
1185 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1186 exp_max_brw_size(tsi->tsi_exp));
1187 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1189 /* allocate pages to store the containers */
1190 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1191 if (rdpg->rp_pages == NULL)
1192 GOTO(out, rc = -ENOMEM);
1193 for (i = 0; i < rdpg->rp_npages; i++) {
1194 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1195 if (rdpg->rp_pages[i] == NULL)
1196 GOTO(out, rc = -ENOMEM);
1199 /* populate pages with key/record pairs */
1200 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1204 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1205 "asked %d > %d\n", rc, rdpg->rp_count);
1207 /* send pages to client */
1208 rc = tgt_sendpage(tsi, rdpg, rc);
1213 if (rdpg->rp_pages) {
1214 for (i = 0; i < rdpg->rp_npages; i++)
1215 if (rdpg->rp_pages[i])
1216 __free_page(rdpg->rp_pages[i]);
1217 OBD_FREE(rdpg->rp_pages,
1218 rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1223 struct tgt_handler tgt_obd_handlers[] = {
1224 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1225 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1227 EXPORT_SYMBOL(tgt_obd_handlers);
1229 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1230 struct dt_object *obj, __u64 start, __u64 end)
1236 /* if no objid is specified, it means "sync whole filesystem" */
1238 rc = dt_sync(env, tgt->lut_bottom);
1239 } else if (dt_version_get(env, obj) >
1240 tgt->lut_obd->obd_last_committed) {
1241 rc = dt_object_sync(env, obj, start, end);
1243 atomic_inc(&tgt->lut_sync_count);
1247 EXPORT_SYMBOL(tgt_sync);
1249 * Unified target DLM handlers.
1253 * Unified target BAST
1255 * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1256 * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1257 * but if it's cross-MDT lock, because remote object version is not set, a
1258 * filesystem sync is needed.
1260 * \param lock server side lock
1261 * \param desc lock desc
1262 * \param data ldlm_cb_set_arg
1263 * \param flag indicates whether this cancelling or blocking callback
1264 * \retval 0 on success
1265 * \retval negative number on error
1267 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1268 void *data, int flag)
1271 struct lu_target *tgt;
1272 struct dt_object *obj = NULL;
1278 tgt = class_exp2tgt(lock->l_export);
1280 if (unlikely(tgt == NULL)) {
1281 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1282 class_exp2obd(lock->l_export)->obd_name);
1286 if (flag == LDLM_CB_CANCELING &&
1287 (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1288 (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1289 (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1290 ldlm_is_cbpending(lock))) &&
1291 ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1292 lock->l_resource->lr_type == LDLM_EXTENT)) {
1294 __u64 end = OBD_OBJECT_EOF;
1296 rc = lu_env_init(&env, LCT_DT_THREAD);
1297 if (unlikely(rc != 0))
1300 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1301 tgt->lut_lsd.lsd_osd_index);
1303 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1304 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1306 GOTO(err_env, rc = PTR_ERR(obj));
1308 if (!dt_object_exists(obj))
1309 GOTO(err_put, rc = -ENOENT);
1311 start = lock->l_policy_data.l_extent.start;
1312 end = lock->l_policy_data.l_extent.end;
1315 rc = tgt_sync(&env, tgt, obj, start, end);
1317 CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1318 "cancel: rc = %d\n",
1319 tgt_name(tgt), PFID(&fid),
1320 lock->l_policy_data.l_extent.start,
1321 lock->l_policy_data.l_extent.end, rc);
1325 dt_object_put(&env, obj);
1330 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1334 static struct ldlm_callback_suite tgt_dlm_cbs = {
1335 .lcs_completion = ldlm_server_completion_ast,
1336 .lcs_blocking = tgt_blocking_ast,
1337 .lcs_glimpse = ldlm_server_glimpse_ast
1340 int tgt_enqueue(struct tgt_session_info *tsi)
1342 struct ptlrpc_request *req = tgt_ses_req(tsi);
1347 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1348 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1350 LASSERT(tsi->tsi_dlm_req != NULL);
1351 rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1352 tsi->tsi_dlm_req, &tgt_dlm_cbs);
1354 RETURN(err_serious(rc));
1356 switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1358 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1361 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1364 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1367 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1370 RETURN(req->rq_status);
1372 EXPORT_SYMBOL(tgt_enqueue);
1374 int tgt_convert(struct tgt_session_info *tsi)
1376 struct ptlrpc_request *req = tgt_ses_req(tsi);
1380 LASSERT(tsi->tsi_dlm_req);
1381 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1383 RETURN(err_serious(rc));
1385 RETURN(req->rq_status);
1388 int tgt_bl_callback(struct tgt_session_info *tsi)
1390 return err_serious(-EOPNOTSUPP);
1393 int tgt_cp_callback(struct tgt_session_info *tsi)
1395 return err_serious(-EOPNOTSUPP);
1398 /* generic LDLM target handler */
1399 struct tgt_handler tgt_dlm_handlers[] = {
1400 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1401 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1402 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1403 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1405 EXPORT_SYMBOL(tgt_dlm_handlers);
1408 * Unified target LLOG handlers.
1410 int tgt_llog_open(struct tgt_session_info *tsi)
1416 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1420 EXPORT_SYMBOL(tgt_llog_open);
1422 int tgt_llog_read_header(struct tgt_session_info *tsi)
1428 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1432 EXPORT_SYMBOL(tgt_llog_read_header);
1434 int tgt_llog_next_block(struct tgt_session_info *tsi)
1440 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1444 EXPORT_SYMBOL(tgt_llog_next_block);
1446 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1452 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1456 EXPORT_SYMBOL(tgt_llog_prev_block);
1458 /* generic llog target handler */
1459 struct tgt_handler tgt_llog_handlers[] = {
1460 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1461 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1462 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1463 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1465 EXPORT_SYMBOL(tgt_llog_handlers);
1468 * sec context handlers
1470 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1471 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1476 struct tgt_handler tgt_sec_ctx_handlers[] = {
1477 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle),
1478 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle),
1479 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle),
1481 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1483 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1484 struct dt_device *key,
1485 struct lfsck_req_local *lrl,
1486 struct thandle *th) = NULL;
1488 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1490 struct lfsck_req_local *,
1493 tgt_lfsck_in_notify_local = notify;
1495 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1497 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1498 struct dt_device *key,
1499 struct lfsck_request *lr) = NULL;
1501 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1503 struct lfsck_request *))
1505 tgt_lfsck_in_notify = notify;
1507 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1509 static int (*tgt_lfsck_query)(const struct lu_env *env,
1510 struct dt_device *key,
1511 struct lfsck_request *req,
1512 struct lfsck_reply *rep,
1513 struct lfsck_query *que) = NULL;
1515 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1517 struct lfsck_request *,
1518 struct lfsck_reply *,
1519 struct lfsck_query *))
1521 tgt_lfsck_query = query;
1523 EXPORT_SYMBOL(tgt_register_lfsck_query);
1525 /* LFSCK request handlers */
1526 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1528 const struct lu_env *env = tsi->tsi_env;
1529 struct dt_device *key = tsi->tsi_tgt->lut_bottom;
1530 struct lfsck_request *lr;
1534 lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1538 rc = tgt_lfsck_in_notify(env, key, lr);
1543 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1545 struct lfsck_request *request;
1546 struct lfsck_reply *reply;
1550 request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1551 if (request == NULL)
1554 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1558 rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1559 request, reply, NULL);
1561 RETURN(rc < 0 ? rc : 0);
1564 struct tgt_handler tgt_lfsck_handlers[] = {
1565 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_NOTIFY, tgt_handle_lfsck_notify),
1566 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_QUERY, tgt_handle_lfsck_query),
1568 EXPORT_SYMBOL(tgt_lfsck_handlers);
1571 * initialize per-thread page pool (bug 5137).
1573 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1575 struct tgt_thread_big_cache *tbc;
1579 LASSERT(thread != NULL);
1580 LASSERT(thread->t_data == NULL);
1582 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1585 thread->t_data = tbc;
1588 EXPORT_SYMBOL(tgt_io_thread_init);
1591 * free per-thread pool created by tgt_thread_init().
1593 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1595 struct tgt_thread_big_cache *tbc;
1599 LASSERT(thread != NULL);
1602 * be prepared to handle partially-initialized pools (because this is
1603 * called from ost_io_thread_init() for cleanup.
1605 tbc = thread->t_data;
1607 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1608 thread->t_data = NULL;
1612 EXPORT_SYMBOL(tgt_io_thread_done);
1615 * Helper function for getting Data-on-MDT file server DLM lock
1616 * if asked by client.
1618 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1619 struct lustre_handle *lh, int mode, __u64 *flags)
1621 union ldlm_policy_data policy = {
1622 .l_inodebits.bits = MDS_INODELOCK_DOM,
1628 LASSERT(lh != NULL);
1629 LASSERT(ns != NULL);
1630 LASSERT(!lustre_handle_is_used(lh));
1632 rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1633 flags, ldlm_blocking_ast,
1634 ldlm_completion_ast, ldlm_glimpse_ast,
1635 NULL, 0, LVB_T_NONE, NULL, lh);
1637 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1639 EXPORT_SYMBOL(tgt_mdt_data_lock);
1641 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1643 LASSERT(lustre_handle_is_used(lh));
1644 ldlm_lock_decref(lh, mode);
1646 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1649 * Helper function for getting server side [start, start+count] DLM lock
1650 * if asked by client.
1652 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1653 struct ldlm_res_id *res_id, __u64 start, __u64 end,
1654 struct lustre_handle *lh, int mode, __u64 *flags)
1656 union ldlm_policy_data policy;
1661 LASSERT(lh != NULL);
1662 LASSERT(ns != NULL);
1663 LASSERT(!lustre_handle_is_used(lh));
1665 policy.l_extent.gid = 0;
1666 policy.l_extent.start = start & PAGE_MASK;
1669 * If ->o_blocks is EOF it means "lock till the end of the file".
1670 * Otherwise, it's size of an extent or hole being punched (in bytes).
1672 if (end == OBD_OBJECT_EOF || end < start)
1673 policy.l_extent.end = OBD_OBJECT_EOF;
1675 policy.l_extent.end = end | ~PAGE_MASK;
1677 rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1678 mode, flags, ldlm_blocking_ast,
1679 ldlm_completion_ast, ldlm_glimpse_ast,
1680 NULL, 0, LVB_T_NONE, NULL, lh);
1681 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1683 EXPORT_SYMBOL(tgt_extent_lock);
1685 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1687 LASSERT(lustre_handle_is_used(lh));
1688 ldlm_lock_decref(lh, mode);
1690 EXPORT_SYMBOL(tgt_extent_unlock);
1692 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1693 struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1694 struct niobuf_remote *nb, struct lustre_handle *lh,
1695 enum ldlm_mode mode)
1697 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1699 int nrbufs = obj->ioo_bufcnt;
1705 LASSERT(mode == LCK_PR || mode == LCK_PW);
1706 LASSERT(!lustre_handle_is_used(lh));
1708 if (ns->ns_obd->obd_recovering)
1711 if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1714 for (i = 1; i < nrbufs; i++)
1715 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1718 /* MDT IO for data-on-mdt */
1719 if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1720 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1722 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1723 nb[nrbufs - 1].rnb_offset +
1724 nb[nrbufs - 1].rnb_len - 1,
1729 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1730 struct lustre_handle *lh, enum ldlm_mode mode)
1734 LASSERT(mode == LCK_PR || mode == LCK_PW);
1735 LASSERT((obj->ioo_bufcnt > 0 &&
1736 (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1737 lustre_handle_is_used(lh));
1739 if (lustre_handle_is_used(lh))
1740 tgt_extent_unlock(lh, mode);
1743 static int tgt_checksum_niobuf(struct lu_target *tgt,
1744 struct niobuf_local *local_nb, int npages,
1745 int opc, enum cksum_types cksum_type,
1748 struct ahash_request *req;
1749 unsigned int bufsize;
1751 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1753 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1755 CERROR("%s: unable to initialize checksum hash %s\n",
1756 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1757 return PTR_ERR(req);
1760 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1761 for (i = 0; i < npages; i++) {
1762 /* corrupt the data before we compute the checksum, to
1763 * simulate a client->OST data error */
1764 if (i == 0 && opc == OST_WRITE &&
1765 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1766 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1767 int len = local_nb[i].lnb_len;
1768 struct page *np = tgt_page_to_corrupt;
1771 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1773 char *ptr2 = page_address(np);
1775 memcpy(ptr2 + off, ptr + off, len);
1776 memcpy(ptr2 + off, "bad3", min(4, len));
1777 ll_kunmap_atomic(ptr, KM_USER0);
1779 /* LU-8376 to preserve original index for
1780 * display in dump_all_bulk_pages() */
1783 cfs_crypto_hash_update_page(req, np, off,
1787 CERROR("%s: can't alloc page for corruption\n",
1791 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1792 local_nb[i].lnb_page_offset & ~PAGE_MASK,
1793 local_nb[i].lnb_len);
1795 /* corrupt the data after we compute the checksum, to
1796 * simulate an OST->client data error */
1797 if (i == 0 && opc == OST_READ &&
1798 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1799 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1800 int len = local_nb[i].lnb_len;
1801 struct page *np = tgt_page_to_corrupt;
1804 char *ptr = ll_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));
1810 ll_kunmap_atomic(ptr, KM_USER0);
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);
1895 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1896 struct obd_export *exp, struct obdo *oa,
1897 const struct lnet_process_id *peer,
1898 __u32 client_cksum, __u32 server_cksum,
1899 enum cksum_types server_cksum_type)
1902 enum cksum_types cksum_type;
1905 /* unlikely to happen and only if resend does not occur due to cksum
1906 * control failure on Client */
1907 if (unlikely(server_cksum == client_cksum)) {
1908 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1913 if (exp->exp_obd->obd_checksum_dump)
1914 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1917 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1920 if (cksum_type != server_cksum_type)
1921 msg = "the server may have not used the checksum type specified"
1922 " in the original request - likely a protocol problem";
1924 msg = "should have changed on the client or in transit";
1926 start = local_nb[0].lnb_file_offset;
1927 end = local_nb[npages-1].lnb_file_offset +
1928 local_nb[npages-1].lnb_len - 1;
1930 LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1931 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1932 " %x (type %x), server csum %x (type %x)\n",
1933 exp->exp_obd->obd_name,
1934 msg, libcfs_nid2str(peer->nid),
1935 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1936 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1937 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1939 start, end, client_cksum, cksum_type, server_cksum,
1945 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1946 unsigned char *buf, int size)
1948 int i, off, len, copied = size;
1951 for (i = 0; i < npages; i++) {
1952 off = local[i].lnb_page_offset & ~PAGE_MASK;
1953 len = local[i].lnb_len;
1955 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1960 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
1961 memcpy(buf + off, ptr, len);
1962 ll_kunmap_atomic(ptr, KM_USER0);
1966 return copied - size;
1969 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1970 struct niobuf_local *local_nb,
1971 int npages, int opc,
1972 obd_dif_csum_fn *fn,
1976 enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1977 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1978 const char *obd_name = tgt->lut_obd->obd_name;
1979 struct ahash_request *req;
1980 unsigned int bufsize;
1981 unsigned char *buffer;
1982 struct page *__page;
1985 int used_number = 0;
1991 __page = alloc_page(GFP_KERNEL);
1995 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1997 CERROR("%s: unable to initialize checksum hash %s\n",
1998 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1999 return PTR_ERR(req);
2002 buffer = kmap(__page);
2003 guard_start = (__u16 *)buffer;
2004 guard_number = PAGE_SIZE / sizeof(*guard_start);
2005 for (i = 0; i < npages; i++) {
2006 /* corrupt the data before we compute the checksum, to
2007 * simulate a client->OST data error */
2008 if (i == 0 && opc == OST_WRITE &&
2009 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2010 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2011 int len = local_nb[i].lnb_len;
2012 struct page *np = tgt_page_to_corrupt;
2015 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2017 char *ptr2 = page_address(np);
2019 memcpy(ptr2 + off, ptr + off, len);
2020 memcpy(ptr2 + off, "bad3", min(4, len));
2021 ll_kunmap_atomic(ptr, KM_USER0);
2023 /* LU-8376 to preserve original index for
2024 * display in dump_all_bulk_pages() */
2027 cfs_crypto_hash_update_page(req, np, off,
2031 CERROR("%s: can't alloc page for corruption\n",
2037 * The left guard number should be able to hold checksums of a
2040 if (t10_cksum_type && opc == OST_READ &&
2041 local_nb[i].lnb_guard_disk) {
2042 used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2043 if (used > (guard_number - used_number)) {
2047 memcpy(guard_start + used_number,
2048 local_nb[i].lnb_guards,
2049 used * sizeof(*local_nb[i].lnb_guards));
2051 rc = obd_page_dif_generate_buffer(obd_name,
2052 local_nb[i].lnb_page,
2053 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2054 local_nb[i].lnb_len, guard_start + used_number,
2055 guard_number - used_number, &used, sector_size,
2061 LASSERT(used <= MAX_GUARD_NUMBER);
2063 * If disk support T10PI checksum, copy guards to local_nb.
2064 * If the write is partial page, do not use the guards for bio
2065 * submission since the data might not be full-sector. The bio
2066 * guards will be generated later based on the full sectors. If
2067 * the sector size is 512B rather than 4 KB, or the page size
2068 * is larger than 4KB, this might drop some useful guards for
2069 * partial page write, but it will only add minimal extra time
2070 * of checksum calculation.
2072 if (t10_cksum_type && opc == OST_WRITE &&
2073 local_nb[i].lnb_len == PAGE_SIZE) {
2074 local_nb[i].lnb_guard_rpc = 1;
2075 memcpy(local_nb[i].lnb_guards,
2076 guard_start + used_number,
2077 used * sizeof(*local_nb[i].lnb_guards));
2080 used_number += used;
2081 if (used_number == guard_number) {
2082 cfs_crypto_hash_update_page(req, __page, 0,
2083 used_number * sizeof(*guard_start));
2087 /* corrupt the data after we compute the checksum, to
2088 * simulate an OST->client data error */
2089 if (unlikely(i == 0 && opc == OST_READ &&
2090 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2091 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2092 int len = local_nb[i].lnb_len;
2093 struct page *np = tgt_page_to_corrupt;
2096 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2098 char *ptr2 = page_address(np);
2100 memcpy(ptr2 + off, ptr + off, len);
2101 memcpy(ptr2 + off, "bad4", min(4, len));
2102 ll_kunmap_atomic(ptr, KM_USER0);
2104 /* LU-8376 to preserve original index for
2105 * display in dump_all_bulk_pages() */
2108 cfs_crypto_hash_update_page(req, np, off,
2112 CERROR("%s: can't alloc page for corruption\n",
2121 if (used_number != 0)
2122 cfs_crypto_hash_update_page(req, __page, 0,
2123 used_number * sizeof(*guard_start));
2125 bufsize = sizeof(cksum);
2126 rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2131 __free_page(__page);
2135 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2136 enum cksum_types cksum_type,
2137 struct niobuf_local *local_nb,
2138 int npages, int opc, u32 *check_sum)
2140 obd_dif_csum_fn *fn = NULL;
2141 int sector_size = 0;
2145 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
2148 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2149 opc, fn, sector_size,
2152 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2153 cksum_type, check_sum);
2157 int tgt_brw_read(struct tgt_session_info *tsi)
2159 struct ptlrpc_request *req = tgt_ses_req(tsi);
2160 struct ptlrpc_bulk_desc *desc = NULL;
2161 struct obd_export *exp = tsi->tsi_exp;
2162 struct niobuf_remote *remote_nb;
2163 struct niobuf_local *local_nb;
2164 struct obd_ioobj *ioo;
2165 struct ost_body *body, *repbody;
2166 struct l_wait_info lwi;
2167 struct lustre_handle lockh = { 0 };
2168 int npages, nob = 0, rc, i, no_reply = 0,
2170 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2171 const char *obd_name = exp->exp_obd->obd_name;
2175 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2176 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2177 CERROR("%s: deny read request from %s to portal %u\n",
2178 tgt_name(tsi->tsi_tgt),
2179 obd_export_nid2str(req->rq_export),
2180 ptlrpc_req2svc(req)->srv_req_portal);
2184 req->rq_bulk_read = 1;
2186 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2189 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2190 cfs_fail_val : (obd_timeout + 1) / 4);
2192 /* Check if there is eviction in progress, and if so, wait for it to
2194 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2195 /* We do not care how long it takes */
2196 lwi = LWI_INTR(NULL, NULL);
2197 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2198 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
2202 /* There must be big cache in current thread to process this request
2203 * if it is NULL then something went wrong and it wasn't allocated,
2204 * report -ENOMEM in that case */
2208 body = tsi->tsi_ost_body;
2209 LASSERT(body != NULL);
2211 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2212 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2214 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2215 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2217 local_nb = tbc->local;
2219 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2225 * If getting the lock took more time than
2226 * client was willing to wait, drop it. b=11330
2228 if (ktime_get_real_seconds() > req->rq_deadline ||
2229 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2231 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2232 libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2233 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2234 req->rq_deadline - req->rq_arrival_time.tv_sec);
2235 GOTO(out_lock, rc = -ETIMEDOUT);
2238 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2239 repbody->oa = body->oa;
2241 npages = PTLRPC_MAX_BRW_PAGES;
2242 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2243 ioo, remote_nb, &npages, local_nb);
2247 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2248 body->oa.o_flags & OBD_FL_SHORT_IO) {
2251 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2252 PTLRPC_BULK_PUT_SOURCE |
2253 PTLRPC_BULK_BUF_KIOV,
2255 &ptlrpc_bulk_kiov_nopin_ops);
2257 GOTO(out_commitrw, rc = -ENOMEM);
2261 npages_read = npages;
2262 for (i = 0; i < npages; i++) {
2263 int page_rc = local_nb[i].lnb_rc;
2272 if (page_rc != 0 && desc != NULL) { /* some data! */
2273 LASSERT(local_nb[i].lnb_page != NULL);
2274 desc->bd_frag_ops->add_kiov_frag
2275 (desc, local_nb[i].lnb_page,
2276 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2280 if (page_rc != local_nb[i].lnb_len) { /* short read */
2281 local_nb[i].lnb_len = page_rc;
2282 npages_read = i + (page_rc != 0 ? 1 : 0);
2283 /* All subsequent pages should be 0 */
2284 while (++i < npages)
2285 LASSERT(local_nb[i].lnb_rc == 0);
2289 if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
2290 nob != cfs_fail_val)
2293 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2294 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2295 body->oa.o_flags : 0;
2296 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2298 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2300 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2302 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2303 local_nb, npages_read, OST_READ,
2304 &repbody->oa.o_cksum);
2306 GOTO(out_commitrw, rc);
2307 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2308 repbody->oa.o_cksum);
2310 /* if a resend it could be for a cksum error, so check Server
2311 * cksum with returned Client cksum (this should even cover
2312 * zero-cksum case) */
2313 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2314 (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2315 check_read_checksum(local_nb, npages_read, exp,
2316 &body->oa, &req->rq_peer,
2318 repbody->oa.o_cksum, cksum_type);
2320 repbody->oa.o_valid = 0;
2322 /* We're finishing using body->oa as an input variable */
2324 /* Check if client was evicted while we were doing i/o before touching
2327 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2328 body->oa.o_flags & OBD_FL_SHORT_IO) {
2329 unsigned char *short_io_buf;
2332 short_io_buf = req_capsule_server_get(&req->rq_pill,
2334 short_io_size = req_capsule_get_size(&req->rq_pill,
2337 rc = tgt_pages2shortio(local_nb, npages_read,
2338 short_io_buf, short_io_size);
2340 req_capsule_shrink(&req->rq_pill,
2343 rc = rc > 0 ? 0 : rc;
2344 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2345 rc = target_bulk_io(exp, desc);
2349 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2350 body->oa.o_flags & OBD_FL_SHORT_IO)
2351 req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2356 /* Must commit after prep above in all cases */
2357 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2358 remote_nb, npages, local_nb, rc);
2360 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2362 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2363 ptlrpc_free_bulk(desc);
2368 ptlrpc_lprocfs_brw(req, nob);
2369 } else if (no_reply) {
2370 req->rq_no_reply = 1;
2371 /* reply out callback would free */
2372 ptlrpc_req_drop_rs(req);
2373 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2374 "client will retry: rc %d\n",
2376 obd_uuid2str(&exp->exp_client_uuid),
2377 obd_export_nid2str(exp), rc);
2379 /* send a bulk after reply to simulate a network delay or reordering
2380 * by a router - Note that !desc implies short io, so there is no bulk
2382 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2384 wait_queue_head_t waitq;
2385 struct l_wait_info lwi1;
2387 CDEBUG(D_INFO, "reorder BULK\n");
2388 init_waitqueue_head(&waitq);
2390 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
2391 l_wait_event(waitq, 0, &lwi1);
2392 target_bulk_io(exp, desc);
2393 ptlrpc_free_bulk(desc);
2398 EXPORT_SYMBOL(tgt_brw_read);
2400 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2401 unsigned char *buf, unsigned int size)
2406 for (i = 0; i < npages; i++) {
2407 off = local[i].lnb_page_offset & ~PAGE_MASK;
2408 len = local[i].lnb_len;
2413 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2415 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
2418 memcpy(ptr + off, buf, len < size ? len : size);
2419 ll_kunmap_atomic(ptr, KM_USER0);
2426 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2427 struct ptlrpc_bulk_desc *desc,
2428 struct niobuf_local *local_nb, int npages,
2429 u32 client_cksum, u32 server_cksum,
2432 struct obd_export *exp = req->rq_export;
2433 struct ost_body *body;
2437 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2438 LASSERT(body != NULL);
2440 if (desc && req->rq_peer.nid != desc->bd_sender) {
2442 router = libcfs_nid2str(desc->bd_sender);
2445 if (exp->exp_obd->obd_checksum_dump)
2446 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2450 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2451 client_cksum, server_cksum);
2455 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2456 DFID" object "DOSTID" extent [%llu-%llu"
2457 "]: client csum %x, server csum %x\n",
2458 exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2460 body->oa.o_valid & OBD_MD_FLFID ?
2461 body->oa.o_parent_seq : (__u64)0,
2462 body->oa.o_valid & OBD_MD_FLFID ?
2463 body->oa.o_parent_oid : 0,
2464 body->oa.o_valid & OBD_MD_FLFID ?
2465 body->oa.o_parent_ver : 0,
2466 POSTID(&body->oa.o_oi),
2467 local_nb[0].lnb_file_offset,
2468 local_nb[npages-1].lnb_file_offset +
2469 local_nb[npages - 1].lnb_len - 1,
2470 client_cksum, server_cksum);
2473 int tgt_brw_write(struct tgt_session_info *tsi)
2475 struct ptlrpc_request *req = tgt_ses_req(tsi);
2476 struct ptlrpc_bulk_desc *desc = NULL;
2477 struct obd_export *exp = req->rq_export;
2478 struct niobuf_remote *remote_nb;
2479 struct niobuf_local *local_nb;
2480 struct obd_ioobj *ioo;
2481 struct ost_body *body, *repbody;
2482 struct lustre_handle lockh = {0};
2484 int objcount, niocount, npages;
2486 enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2487 bool no_reply = false, mmap;
2488 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2489 bool wait_sync = false;
2490 const char *obd_name = exp->exp_obd->obd_name;
2494 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2495 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2496 CERROR("%s: deny write request from %s to portal %u\n",
2497 tgt_name(tsi->tsi_tgt),
2498 obd_export_nid2str(req->rq_export),
2499 ptlrpc_req2svc(req)->srv_req_portal);
2500 RETURN(err_serious(-EPROTO));
2503 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2504 RETURN(err_serious(-ENOSPC));
2505 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2506 RETURN(err_serious(-EROFS));
2508 req->rq_bulk_write = 1;
2510 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2511 RETURN(err_serious(-EIO));
2512 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2513 RETURN(err_serious(-EFAULT));
2515 /* pause before transaction has been started */
2516 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2517 cfs_fail_val : (obd_timeout + 1) / 4);
2519 /* Delay write commit to show stale size information */
2520 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2522 /* There must be big cache in current thread to process this request
2523 * if it is NULL then something went wrong and it wasn't allocated,
2524 * report -ENOMEM in that case */
2528 body = tsi->tsi_ost_body;
2529 LASSERT(body != NULL);
2531 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2532 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2534 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2535 RCL_CLIENT) / sizeof(*ioo);
2537 for (niocount = i = 0; i < objcount; i++)
2538 niocount += ioo[i].ioo_bufcnt;
2540 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2541 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2542 if (niocount != req_capsule_get_size(&req->rq_pill,
2543 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2545 RETURN(err_serious(-EPROTO));
2547 if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2548 ptlrpc_connection_is_local(exp->exp_connection))
2549 memory_pressure_set();
2551 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2552 niocount * sizeof(*rcs));
2553 rc = req_capsule_server_pack(&req->rq_pill);
2555 GOTO(out, rc = err_serious(rc));
2557 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2558 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2560 local_nb = tbc->local;
2562 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2568 * If getting the lock took more time than
2569 * client was willing to wait, drop it. b=11330
2571 if (ktime_get_real_seconds() > req->rq_deadline ||
2572 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2574 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2575 tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2576 POSTID(&ioo->ioo_oid),
2577 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2578 req->rq_deadline - req->rq_arrival_time.tv_sec);
2579 GOTO(out_lock, rc = -ETIMEDOUT);
2582 /* Because we already sync grant info with client when reconnect,
2583 * grant info will be cleared for resent req, then fed_grant and
2584 * total_grant will not be modified in following preprw_write */
2585 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2586 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2587 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2590 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2591 if (repbody == NULL)
2592 GOTO(out_lock, rc = -ENOMEM);
2593 repbody->oa = body->oa;
2595 npages = PTLRPC_MAX_BRW_PAGES;
2596 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2597 objcount, ioo, remote_nb, &npages, local_nb);
2600 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2601 body->oa.o_flags & OBD_FL_SHORT_IO) {
2602 unsigned int short_io_size;
2603 unsigned char *short_io_buf;
2605 short_io_size = req_capsule_get_size(&req->rq_pill,
2608 short_io_buf = req_capsule_client_get(&req->rq_pill,
2610 CDEBUG(D_INFO, "Client use short io for data transfer,"
2611 " size = %d\n", short_io_size);
2613 /* Copy short io buf to pages */
2614 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2618 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2619 PTLRPC_BULK_GET_SINK |
2620 PTLRPC_BULK_BUF_KIOV,
2622 &ptlrpc_bulk_kiov_nopin_ops);
2624 GOTO(skip_transfer, rc = -ENOMEM);
2626 /* NB Having prepped, we must commit... */
2627 for (i = 0; i < npages; i++)
2628 desc->bd_frag_ops->add_kiov_frag(desc,
2629 local_nb[i].lnb_page,
2630 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2631 local_nb[i].lnb_len);
2633 rc = sptlrpc_svc_prep_bulk(req, desc);
2635 GOTO(skip_transfer, rc);
2637 rc = target_bulk_io(exp, desc);
2643 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2644 static int cksum_counter;
2646 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2647 cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2649 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2650 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2651 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2654 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2655 local_nb, npages, OST_WRITE,
2656 &repbody->oa.o_cksum);
2658 GOTO(out_commitrw, rc);
2662 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2663 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2664 body->oa.o_flags & OBD_FL_MMAP);
2666 tgt_warn_on_cksum(req, desc, local_nb, npages,
2668 repbody->oa.o_cksum, mmap);
2670 } else if ((cksum_counter & (-cksum_counter)) ==
2672 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2673 cksum_counter, libcfs_id2str(req->rq_peer),
2674 repbody->oa.o_cksum);
2679 /* Must commit after prep above in all cases */
2680 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2681 objcount, ioo, remote_nb, npages, local_nb, rc);
2682 if (rc == -ENOTCONN)
2683 /* quota acquire process has been given up because
2684 * either the client has been evicted or the client
2685 * has timed out the request already */
2688 for (i = 0; i < niocount; i++) {
2689 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2695 * Disable sending mtime back to the client. If the client locked the
2696 * whole object, then it has already updated the mtime on its side,
2697 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2699 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2704 /* set per-requested niobuf return codes */
2705 for (i = j = 0; i < niocount; i++) {
2706 int len = remote_nb[i].rnb_len;
2711 LASSERT(j < npages);
2712 if (local_nb[j].lnb_rc < 0)
2713 rcs[i] = local_nb[j].lnb_rc;
2714 len -= local_nb[j].lnb_len;
2719 LASSERT(j == npages);
2720 ptlrpc_lprocfs_brw(req, nob);
2723 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2725 ptlrpc_free_bulk(desc);
2727 if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2728 req->rq_no_reply = 1;
2729 /* reply out callback would free */
2730 ptlrpc_req_drop_rs(req);
2731 if (!exp->exp_obd->obd_no_transno)
2732 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2733 " client will retry: rc = %d\n",
2735 obd_uuid2str(&exp->exp_client_uuid),
2736 obd_export_nid2str(exp), rc);
2738 memory_pressure_clr();
2741 EXPORT_SYMBOL(tgt_brw_write);
2743 /* Check if request can be reconstructed from saved reply data
2744 * A copy of the reply data is returned in @trd if the pointer is not NULL
2746 int req_can_reconstruct(struct ptlrpc_request *req,
2747 struct tg_reply_data *trd)
2749 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2750 struct lsd_client_data *lcd = ted->ted_lcd;
2753 if (tgt_is_multimodrpcs_client(req->rq_export))
2754 return tgt_lookup_reply(req, trd);
2756 mutex_lock(&ted->ted_lcd_lock);
2757 found = req->rq_xid == lcd->lcd_last_xid ||
2758 req->rq_xid == lcd->lcd_last_close_xid;
2760 if (found && trd != NULL) {
2761 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2762 trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2763 trd->trd_reply.lrd_transno =
2764 lcd->lcd_last_close_transno;
2765 trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2767 trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2768 trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2769 trd->trd_reply.lrd_result = lcd->lcd_last_result;
2770 trd->trd_reply.lrd_data = lcd->lcd_last_data;
2771 trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2772 trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2773 trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2774 trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2777 mutex_unlock(&ted->ted_lcd_lock);
2781 EXPORT_SYMBOL(req_can_reconstruct);