4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; if not, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
24 * Copyright (c) 2013, 2017, Intel Corporation.
27 * lustre/target/tgt_handler.c
29 * Lustre Unified Target request handler code
31 * Author: Brian Behlendorf <behlendorf1@llnl.gov>
32 * Author: Mikhail Pershin <mike.pershin@intel.com>
35 #define DEBUG_SUBSYSTEM S_CLASS
37 #include <linux/user_namespace.h>
38 #ifdef HAVE_UIDGID_HEADER
39 # include <linux/uidgid.h>
43 #include <obd_class.h>
44 #include <obd_cksum.h>
45 #include <lustre_lfsck.h>
46 #include <lustre_nodemap.h>
47 #include <lustre_acl.h>
49 #include "tgt_internal.h"
51 char *tgt_name(struct lu_target *tgt)
53 LASSERT(tgt->lut_obd != NULL);
54 return tgt->lut_obd->obd_name;
56 EXPORT_SYMBOL(tgt_name);
59 * Generic code handling requests that have struct mdt_body passed in:
61 * - extract mdt_body from request and save it in @tsi, if present;
63 * - create lu_object, corresponding to the fid in mdt_body, and save it in
66 * - if HABEO_CORPUS flag is set for this request type check whether object
67 * actually exists on storage (lu_object_exists()).
70 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
72 const struct mdt_body *body;
73 struct lu_object *obj;
74 struct req_capsule *pill = tsi->tsi_pill;
79 body = req_capsule_client_get(pill, &RMF_MDT_BODY);
83 tsi->tsi_mdt_body = body;
85 if (!(body->mbo_valid & OBD_MD_FLID))
88 /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
89 * in any case in pre-2.5 clients. Fix that here if needed */
90 if (unlikely(fid_is_zero(&body->mbo_fid1)))
93 if (!fid_is_sane(&body->mbo_fid1)) {
94 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
95 PFID(&body->mbo_fid1));
99 obj = lu_object_find(tsi->tsi_env,
100 &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
101 &body->mbo_fid1, NULL);
103 if ((flags & HABEO_CORPUS) && !lu_object_exists(obj)) {
104 lu_object_put(tsi->tsi_env, obj);
107 tsi->tsi_corpus = obj;
114 tsi->tsi_fid = body->mbo_fid1;
120 * Validate oa from client.
121 * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
123 * a. objects in Single MDT FS seq = FID_SEQ_OST_MDT0, oi_id != 0
124 * b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
125 * pack ost_id. Because non-zero oi_seq will make it diffcult to tell
126 * whether this is oi_fid or real ostid. So it will check
127 * OBD_CONNECT_FID, then convert the ostid to FID for old client.
128 * c. Old FID-disable osc will send IDIF.
129 * d. new FID-enable osc/osp will send normal FID.
131 * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
132 * be used for LAST_ID file, and only being accessed inside OST now.
134 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
136 struct ost_id *oi = &oa->o_oi;
137 u64 seq = ostid_seq(oi);
138 u64 id = ostid_id(oi);
142 if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
143 fid_seq_is_echo(seq))) {
144 /* Sigh 2.[123] client still sends echo req with oi_id = 0
145 * during create, and we will reset this to 1, since this
146 * oi_id is basically useless in the following create process,
147 * but oi_id == 0 will make it difficult to tell whether it is
148 * real FID or ost_id. */
149 oi->oi_fid.f_seq = FID_SEQ_ECHO;
150 oi->oi_fid.f_oid = id ?: 1;
151 oi->oi_fid.f_ver = 0;
153 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
155 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
156 GOTO(out, rc = -EPROTO);
158 /* Note: this check might be forced in 2.5 or 2.6, i.e.
159 * all of the requests are required to setup FLGROUP */
160 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
161 ostid_set_seq_mdt0(oi);
162 oa->o_valid |= OBD_MD_FLGROUP;
166 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
167 fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
168 GOTO(out, rc = -EPROTO);
170 rc = ostid_to_fid(&tti->tti_fid1, oi,
171 tsi->tsi_tgt->lut_lsd.lsd_osd_index);
172 if (unlikely(rc != 0))
175 oi->oi_fid = tti->tti_fid1;
181 CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
182 tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
186 EXPORT_SYMBOL(tgt_validate_obdo);
188 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
191 struct niobuf_remote *rnb;
192 struct obd_ioobj *ioo;
197 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
201 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
205 max_brw = ioobj_max_brw_get(ioo);
206 if (unlikely((max_brw & (max_brw - 1)) != 0)) {
207 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
208 ": rc = %d\n", tgt_name(tsi->tsi_tgt),
209 obd_export_nid2str(tsi->tsi_exp), max_brw,
210 POSTID(oi), -EPROTO);
215 obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
216 RCL_CLIENT) / sizeof(*ioo);
217 if (obj_count == 0) {
218 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
220 } else if (obj_count > 1) {
221 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
226 if (ioo->ioo_bufcnt == 0) {
227 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
231 if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
232 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
233 "bulk has too many pages (%d)",
241 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
243 struct ost_body *body;
244 struct req_capsule *pill = tsi->tsi_pill;
245 struct lu_nodemap *nodemap;
250 body = req_capsule_client_get(pill, &RMF_OST_BODY);
254 rc = tgt_validate_obdo(tsi, &body->oa);
258 nodemap = nodemap_get_from_exp(tsi->tsi_exp);
260 RETURN(PTR_ERR(nodemap));
262 body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
263 NODEMAP_CLIENT_TO_FS,
265 body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
266 NODEMAP_CLIENT_TO_FS,
268 nodemap_putref(nodemap);
270 tsi->tsi_ost_body = body;
271 tsi->tsi_fid = body->oa.o_oi.oi_fid;
273 if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
274 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
279 if (!(body->oa.o_valid & OBD_MD_FLID)) {
280 if (flags & HABEO_CORPUS) {
281 CERROR("%s: OBD_MD_FLID flag is not set in ost_body "
282 "but OID/FID is mandatory with HABEO_CORPUS\n",
283 tgt_name(tsi->tsi_tgt));
290 ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
293 * OST doesn't get object in advance for further use to prevent
294 * situations with nested object_find which is potential deadlock.
296 tsi->tsi_corpus = NULL;
301 * Do necessary preprocessing according to handler ->th_flags.
303 static int tgt_request_preprocess(struct tgt_session_info *tsi,
304 struct tgt_handler *h,
305 struct ptlrpc_request *req)
307 struct req_capsule *pill = tsi->tsi_pill;
308 __u32 flags = h->th_flags;
313 if (tsi->tsi_preprocessed)
316 LASSERT(h->th_act != NULL);
317 LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
318 LASSERT(current->journal_info == NULL);
320 LASSERT(ergo(flags & (HABEO_CORPUS | HABEO_REFERO),
322 if (h->th_fmt != NULL) {
323 req_capsule_set(pill, h->th_fmt);
324 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
325 rc = tgt_mdt_body_unpack(tsi, flags);
328 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
330 rc = tgt_ost_body_unpack(tsi, flags);
336 if (flags & MUTABOR && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
339 if (flags & HABEO_CLAVIS) {
340 struct ldlm_request *dlm_req;
342 LASSERT(h->th_fmt != NULL);
344 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
345 if (dlm_req != NULL) {
346 union ldlm_wire_policy_data *policy =
347 &dlm_req->lock_desc.l_policy_data;
349 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
351 (policy->l_inodebits.bits |
352 policy->l_inodebits.try_bits) == 0)) {
354 * Lock without inodebits makes no sense and
355 * will oops later in ldlm. If client miss to
356 * set such bits, do not trigger ASSERTION.
358 * For liblustre flock case, it maybe zero.
362 tsi->tsi_dlm_req = dlm_req;
368 tsi->tsi_preprocessed = 1;
373 * Invoke handler for this request opc. Also do necessary preprocessing
374 * (according to handler ->th_flags), and post-processing (setting of
375 * ->last_{xid,committed}).
377 static int tgt_handle_request0(struct tgt_session_info *tsi,
378 struct tgt_handler *h,
379 struct ptlrpc_request *req)
383 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
388 /* When dealing with sec context requests, no export is associated yet,
389 * because these requests are sent before *_CONNECT requests.
390 * A NULL req->rq_export means the normal *_common_slice handlers will
391 * not be called, because there is no reference to the target.
392 * So deal with them by hand and jump directly to target_send_reply().
396 case SEC_CTX_INIT_CONT:
398 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
403 * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
404 * to put same checks into handlers like mdt_close(), mdt_reint(),
405 * etc., without talking to mdt authors first. Checking same thing
406 * there again is useless and returning 0 error without packing reply
407 * is buggy! Handlers either pack reply or return error.
409 * We return 0 here and do not send any reply in order to emulate
410 * network failure. Do not send any reply in case any of NET related
411 * fail_id has occured.
413 if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
415 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
416 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
419 rc = tgt_request_preprocess(tsi, h, req);
420 /* pack reply if reply format is fixed */
421 if (rc == 0 && h->th_flags & HABEO_REFERO) {
423 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
425 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
427 tsi->tsi_mdt_body->mbo_eadatasize);
428 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
430 req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
432 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
433 req_capsule_set_size(tsi->tsi_pill,
434 &RMF_ACL, RCL_SERVER,
435 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
437 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
439 struct niobuf_remote *remote_nb =
440 req_capsule_client_get(tsi->tsi_pill,
442 struct ost_body *body = tsi->tsi_ost_body;
444 req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
446 (body->oa.o_valid & OBD_MD_FLFLAGS &&
447 body->oa.o_flags & OBD_FL_SHORT_IO) ?
448 remote_nb[0].rnb_len : 0);
451 rc = req_capsule_server_pack(tsi->tsi_pill);
454 if (likely(rc == 0)) {
456 * Process request, there can be two types of rc:
457 * 1) errors with msg unpack/pack, other failures outside the
458 * operation itself. This is counted as serious errors;
459 * 2) errors during fs operation, should be placed in rq_status
463 if (!is_serious(rc) &&
464 !req->rq_no_reply && req->rq_reply_state == NULL) {
465 DEBUG_REQ(D_ERROR, req, "%s \"handler\" %s did not "
466 "pack reply and returned 0 error\n",
467 tgt_name(tsi->tsi_tgt), h->th_name);
470 serious = is_serious(rc);
471 rc = clear_serious(rc);
479 * ELDLM_* codes which > 0 should be in rq_status only as well as
480 * all non-serious errors.
482 if (rc > 0 || !serious)
485 LASSERT(current->journal_info == NULL);
487 if (likely(rc == 0 && req->rq_export))
488 target_committed_to_req(req);
491 target_send_reply(req, rc, tsi->tsi_reply_fail_id);
495 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
496 struct obd_device *obd, int *process)
498 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
505 case MDS_SYNC: /* used in unmounting */
519 case MDS_HSM_PROGRESS:
520 case MDS_HSM_STATE_SET:
521 case MDS_HSM_REQUEST:
522 *process = target_queue_recovery_request(req, obd);
526 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
533 * Handle recovery. Return:
534 * +1: continue request processing;
535 * -ve: abort immediately with the given error code;
536 * 0: send reply with error code in req->rq_status;
538 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
542 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
547 case SEC_CTX_INIT_CONT:
552 if (!req->rq_export->exp_obd->obd_replayable)
555 /* sanity check: if the xid matches, the request must be marked as a
556 * resent or replayed */
557 if (req_can_reconstruct(req, NULL)) {
558 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
559 (MSG_RESENT | MSG_REPLAY))) {
560 DEBUG_REQ(D_WARNING, req, "rq_xid %llu matches "
561 "saved xid, expected REPLAY or RESENT flag "
563 lustre_msg_get_flags(req->rq_reqmsg));
564 req->rq_status = -ENOTCONN;
568 /* else: note the opposite is not always true; a RESENT req after a
569 * failover will usually not match the last_xid, since it was likely
570 * never committed. A REPLAYed request will almost never match the
571 * last xid, however it could for a committed, but still retained,
574 /* Check for aborted recovery... */
575 if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
579 DEBUG_REQ(D_INFO, req, "Got new replay");
580 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
582 if (rc != 0 || !should_process)
584 else if (should_process < 0) {
585 req->rq_status = should_process;
586 rc = ptlrpc_error(req);
593 /* Initial check for request, it is validation mostly */
594 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
596 struct tgt_handler *h;
597 struct tgt_opc_slice *s;
598 struct lu_target *tgt;
599 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
603 tgt = class_exp2tgt(req->rq_export);
604 if (unlikely(tgt == NULL)) {
605 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
606 class_exp2obd(req->rq_export)->obd_name);
607 RETURN(ERR_PTR(-EINVAL));
610 for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
611 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
614 /* opcode was not found in slice */
615 if (unlikely(s->tos_hs == NULL)) {
616 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
618 RETURN(ERR_PTR(-ENOTSUPP));
621 LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
622 h = s->tos_hs + (opc - s->tos_opc_start);
623 if (unlikely(h->th_opc == 0)) {
624 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
625 RETURN(ERR_PTR(-ENOTSUPP));
631 static int process_req_last_xid(struct ptlrpc_request *req)
636 /* check request's xid is consistent with export's last_xid */
637 last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
638 if (last_xid > req->rq_export->exp_last_xid)
639 req->rq_export->exp_last_xid = last_xid;
641 if (req->rq_xid == 0 ||
642 (req->rq_xid <= req->rq_export->exp_last_xid)) {
643 DEBUG_REQ(D_ERROR, req, "Unexpected xid %llx vs. "
644 "last_xid %llx\n", req->rq_xid,
645 req->rq_export->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))
665 /* The "last_xid" is the minimum xid among unreplied requests,
666 * if the request is from the previous connection, its xid can
667 * still be larger than "exp_last_xid", then the above check of
668 * xid is not enough to determine whether the request is delayed.
670 * For example, if some replay request was delayed and caused
671 * timeout at client and the replay is restarted, the delayed
672 * replay request will have the larger xid than "exp_last_xid"
674 if (req->rq_export->exp_conn_cnt >
675 lustre_msg_get_conn_cnt(req->rq_reqmsg))
678 /* try to release in-memory reply data */
679 if (tgt_is_multimodrpcs_client(req->rq_export)) {
680 tgt_handle_received_xid(req->rq_export,
681 lustre_msg_get_last_xid(req->rq_reqmsg));
682 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
683 (MSG_RESENT | MSG_REPLAY)))
684 tgt_handle_tag(req->rq_export,
685 lustre_msg_get_tag(req->rq_reqmsg));
690 int tgt_request_handle(struct ptlrpc_request *req)
692 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
694 struct lustre_msg *msg = req->rq_reqmsg;
695 struct tgt_handler *h;
696 struct lu_target *tgt;
697 int request_fail_id = 0;
698 __u32 opc = lustre_msg_get_opc(msg);
699 struct obd_device *obd;
701 bool is_connect = false;
704 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
705 if (cfs_fail_val == 0 &&
706 lustre_msg_get_opc(msg) != OBD_PING &&
707 lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
708 struct l_wait_info lwi = { 0 };
712 l_wait_event(cfs_race_waitq, (cfs_race_state == 1),
717 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
718 tsi->tsi_pill = &req->rq_pill;
719 tsi->tsi_env = req->rq_svc_thread->t_env;
721 /* if request has export then get handlers slice from corresponding
722 * target, otherwise that should be connect operation */
723 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
724 opc == MGS_CONNECT) {
726 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
727 rc = target_handle_connect(req);
729 rc = ptlrpc_error(req);
732 /* recovery-small test 18c asks to drop connect reply */
733 if (unlikely(opc == OST_CONNECT &&
734 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
738 if (unlikely(!class_connected_export(req->rq_export))) {
739 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
740 opc == SEC_CTX_FINI) {
741 /* sec context initialization has to be handled
742 * by hand in tgt_handle_request0() */
743 tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
745 GOTO(handle_recov, rc = 0);
747 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
748 opc, libcfs_id2str(req->rq_peer));
749 req->rq_status = -ENOTCONN;
750 rc = ptlrpc_error(req);
754 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
755 tsi->tsi_exp = req->rq_export;
756 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
757 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
759 tsi->tsi_jobid = NULL;
762 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
763 class_exp2obd(req->rq_export)->obd_name);
764 req->rq_status = -EINVAL;
765 rc = ptlrpc_error(req);
769 /* Skip last_xid processing for the recovery thread, otherwise, the
770 * last_xid on same request could be processed twice: first time when
771 * processing the incoming request, second time when the request is
772 * being processed by recovery thread. */
773 obd = class_exp2obd(req->rq_export);
775 /* reset the exp_last_xid on each connection. */
776 req->rq_export->exp_last_xid = 0;
777 } else if (obd->obd_recovery_data.trd_processing_task !=
779 rc = process_req_last_xid(req);
782 rc = ptlrpc_error(req);
787 request_fail_id = tgt->lut_request_fail_id;
788 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
790 h = tgt_handler_find_check(req);
792 req->rq_status = PTR_ERR(h);
793 rc = ptlrpc_error(req);
797 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
800 if (CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
803 rc = lustre_msg_check_version(msg, h->th_version);
805 DEBUG_REQ(D_ERROR, req, "%s: drop mal-formed request, version"
806 " %08x, expecting %08x\n", tgt_name(tgt),
807 lustre_msg_get_version(msg), h->th_version);
808 req->rq_status = -EINVAL;
809 rc = ptlrpc_error(req);
814 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
815 if (likely(rc == 1)) {
816 rc = tgt_handle_request0(tsi, h, req);
822 req_capsule_fini(tsi->tsi_pill);
823 if (tsi->tsi_corpus != NULL) {
824 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
825 tsi->tsi_corpus = NULL;
829 EXPORT_SYMBOL(tgt_request_handle);
831 /** Assign high priority operations to the request if needed. */
832 int tgt_hpreq_handler(struct ptlrpc_request *req)
834 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
835 struct tgt_handler *h;
840 if (req->rq_export == NULL)
843 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
844 tsi->tsi_pill = &req->rq_pill;
845 tsi->tsi_env = req->rq_svc_thread->t_env;
846 tsi->tsi_tgt = class_exp2tgt(req->rq_export);
847 tsi->tsi_exp = req->rq_export;
849 h = tgt_handler_find_check(req);
855 rc = tgt_request_preprocess(tsi, h, req);
856 if (unlikely(rc != 0))
859 if (h->th_hp != NULL)
863 EXPORT_SYMBOL(tgt_hpreq_handler);
865 void tgt_counter_incr(struct obd_export *exp, int opcode)
867 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
868 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
869 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
871 EXPORT_SYMBOL(tgt_counter_incr);
874 * Unified target generic handlers.
877 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
879 struct lu_target *tgt = class_exp2tgt(exp);
880 struct sptlrpc_flavor flvr;
884 LASSERT(tgt->lut_obd);
885 LASSERT(tgt->lut_slice);
887 /* always allow ECHO client */
888 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
889 LUSTRE_ECHO_NAME) == 0)) {
890 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
894 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
895 read_lock(&tgt->lut_sptlrpc_lock);
896 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
900 read_unlock(&tgt->lut_sptlrpc_lock);
902 spin_lock(&exp->exp_lock);
903 exp->exp_sp_peer = req->rq_sp_from;
904 exp->exp_flvr = flvr;
906 /* when on mgs, if no restriction is set, or if client
907 * is loopback, allow any flavor */
908 if ((strcmp(exp->exp_obd->obd_type->typ_name,
909 LUSTRE_MGS_NAME) == 0) &&
910 (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
911 exp->exp_connection->c_peer.nid == LNET_NID_LO_0))
912 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
914 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
915 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
916 CERROR("%s: unauthorized rpc flavor %x from %s, "
917 "expect %x\n", tgt_name(tgt),
919 libcfs_nid2str(req->rq_peer.nid),
920 exp->exp_flvr.sf_rpc);
923 spin_unlock(&exp->exp_lock);
925 if (exp->exp_sp_peer != req->rq_sp_from) {
926 CERROR("%s: RPC source %s doesn't match %s\n",
928 sptlrpc_part2name(req->rq_sp_from),
929 sptlrpc_part2name(exp->exp_sp_peer));
932 rc = sptlrpc_target_export_check(exp, req);
939 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
941 struct sptlrpc_rule_set tmp_rset;
944 if (unlikely(tgt == NULL)) {
945 CERROR("No target passed");
949 sptlrpc_rule_set_init(&tmp_rset);
950 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
952 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
957 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
959 write_lock(&tgt->lut_sptlrpc_lock);
960 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
961 tgt->lut_sptlrpc_rset = tmp_rset;
962 write_unlock(&tgt->lut_sptlrpc_lock);
966 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
968 int tgt_connect(struct tgt_session_info *tsi)
970 struct ptlrpc_request *req = tgt_ses_req(tsi);
971 struct obd_connect_data *reply;
976 /* XXX: better to call this check right after getting new export but
977 * before last_rcvd slot allocation to avoid server load upon insecure
978 * connects. This is to be fixed after unifiyng all targets.
980 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
984 /* To avoid exposing partially initialized connection flags, changes up
985 * to this point have been staged in reply->ocd_connect_flags. Now that
986 * connection handling has completed successfully, atomically update
987 * the connect flags in the shared export data structure. LU-1623 */
988 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
989 spin_lock(&tsi->tsi_exp->exp_lock);
990 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
991 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
992 *exp_connect_flags2_ptr(tsi->tsi_exp) =
993 reply->ocd_connect_flags2;
994 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
995 spin_unlock(&tsi->tsi_exp->exp_lock);
997 if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
998 LUSTRE_MDT_NAME) == 0) {
999 rc = req_check_sepol(tsi->tsi_pill);
1006 obd_disconnect(class_export_get(tsi->tsi_exp));
1009 EXPORT_SYMBOL(tgt_connect);
1011 int tgt_disconnect(struct tgt_session_info *tsi)
1017 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1019 rc = target_handle_disconnect(tgt_ses_req(tsi));
1021 RETURN(err_serious(rc));
1025 EXPORT_SYMBOL(tgt_disconnect);
1028 * Unified target OBD handlers
1030 int tgt_obd_ping(struct tgt_session_info *tsi)
1036 /* The target-specific part of OBD_PING request handling.
1037 * It controls Filter Modification Data (FMD) expiration each time
1040 * Valid only for replayable targets, e.g. MDT and OFD
1042 if (tsi->tsi_exp->exp_obd->obd_replayable)
1043 tgt_fmd_expire(tsi->tsi_exp);
1045 rc = req_capsule_server_pack(tsi->tsi_pill);
1047 RETURN(err_serious(rc));
1051 EXPORT_SYMBOL(tgt_obd_ping);
1053 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1055 return err_serious(-EOPNOTSUPP);
1058 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1060 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1061 struct ptlrpc_request *req = tgt_ses_req(tsi);
1062 struct obd_export *exp = req->rq_export;
1063 struct ptlrpc_bulk_desc *desc;
1064 struct l_wait_info *lwi = &tti->tti_u.update.tti_wait_info;
1070 desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1071 PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1072 MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1076 for (i = 0; i < rdbuf->rb_nbufs; i++)
1077 desc->bd_frag_ops->add_iov_frag(desc,
1078 rdbuf->rb_bufs[i].lb_buf,
1079 rdbuf->rb_bufs[i].lb_len);
1081 rc = target_bulk_io(exp, desc, lwi);
1082 ptlrpc_free_bulk(desc);
1085 EXPORT_SYMBOL(tgt_send_buffer);
1087 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1089 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1090 struct ptlrpc_request *req = tgt_ses_req(tsi);
1091 struct obd_export *exp = req->rq_export;
1092 struct ptlrpc_bulk_desc *desc;
1093 struct l_wait_info *lwi = &tti->tti_u.rdpg.tti_wait_info;
1101 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1102 PTLRPC_BULK_PUT_SOURCE |
1103 PTLRPC_BULK_BUF_KIOV,
1105 &ptlrpc_bulk_kiov_pin_ops);
1109 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1110 /* old client requires reply size in it's PAGE_SIZE,
1111 * which is rdpg->rp_count */
1112 nob = rdpg->rp_count;
1114 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1115 i++, tmpcount -= tmpsize) {
1116 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1117 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1121 LASSERT(desc->bd_nob == nob);
1122 rc = target_bulk_io(exp, desc, lwi);
1123 ptlrpc_free_bulk(desc);
1126 EXPORT_SYMBOL(tgt_sendpage);
1129 * OBD_IDX_READ handler
1131 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1133 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1134 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1135 struct idx_info *req_ii, *rep_ii;
1140 memset(rdpg, 0, sizeof(*rdpg));
1141 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1143 /* extract idx_info buffer from request & reply */
1144 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1145 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1146 RETURN(err_serious(-EPROTO));
1148 rc = req_capsule_server_pack(tsi->tsi_pill);
1150 RETURN(err_serious(rc));
1152 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1154 RETURN(err_serious(-EFAULT));
1155 rep_ii->ii_magic = IDX_INFO_MAGIC;
1157 /* extract hash to start with */
1158 rdpg->rp_hash = req_ii->ii_hash_start;
1160 /* extract requested attributes */
1161 rdpg->rp_attrs = req_ii->ii_attrs;
1163 /* check that fid packed in request is valid and supported */
1164 if (!fid_is_sane(&req_ii->ii_fid))
1166 rep_ii->ii_fid = req_ii->ii_fid;
1169 rep_ii->ii_flags = req_ii->ii_flags;
1171 /* compute number of pages to allocate, ii_count is the number of 4KB
1173 if (req_ii->ii_count <= 0)
1174 GOTO(out, rc = -EFAULT);
1175 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1176 exp_max_brw_size(tsi->tsi_exp));
1177 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1179 /* allocate pages to store the containers */
1180 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1181 if (rdpg->rp_pages == NULL)
1182 GOTO(out, rc = -ENOMEM);
1183 for (i = 0; i < rdpg->rp_npages; i++) {
1184 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1185 if (rdpg->rp_pages[i] == NULL)
1186 GOTO(out, rc = -ENOMEM);
1189 /* populate pages with key/record pairs */
1190 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1194 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1195 "asked %d > %d\n", rc, rdpg->rp_count);
1197 /* send pages to client */
1198 rc = tgt_sendpage(tsi, rdpg, rc);
1203 if (rdpg->rp_pages) {
1204 for (i = 0; i < rdpg->rp_npages; i++)
1205 if (rdpg->rp_pages[i])
1206 __free_page(rdpg->rp_pages[i]);
1207 OBD_FREE(rdpg->rp_pages,
1208 rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1213 struct tgt_handler tgt_obd_handlers[] = {
1214 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1215 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1217 EXPORT_SYMBOL(tgt_obd_handlers);
1219 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1220 struct dt_object *obj, __u64 start, __u64 end)
1226 /* if no objid is specified, it means "sync whole filesystem" */
1228 rc = dt_sync(env, tgt->lut_bottom);
1229 } else if (dt_version_get(env, obj) >
1230 tgt->lut_obd->obd_last_committed) {
1231 rc = dt_object_sync(env, obj, start, end);
1233 atomic_inc(&tgt->lut_sync_count);
1237 EXPORT_SYMBOL(tgt_sync);
1239 * Unified target DLM handlers.
1243 * Unified target BAST
1245 * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1246 * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1247 * but if it's cross-MDT lock, because remote object version is not set, a
1248 * filesystem sync is needed.
1250 * \param lock server side lock
1251 * \param desc lock desc
1252 * \param data ldlm_cb_set_arg
1253 * \param flag indicates whether this cancelling or blocking callback
1254 * \retval 0 on success
1255 * \retval negative number on error
1257 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1258 void *data, int flag)
1261 struct lu_target *tgt;
1262 struct dt_object *obj = NULL;
1268 tgt = class_exp2tgt(lock->l_export);
1270 if (unlikely(tgt == NULL)) {
1271 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1272 class_exp2obd(lock->l_export)->obd_name);
1276 if (flag == LDLM_CB_CANCELING &&
1277 (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1278 (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1279 (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1280 ldlm_is_cbpending(lock))) &&
1281 ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1282 lock->l_resource->lr_type == LDLM_EXTENT)) {
1284 __u64 end = OBD_OBJECT_EOF;
1286 rc = lu_env_init(&env, LCT_DT_THREAD);
1287 if (unlikely(rc != 0))
1290 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1291 tgt->lut_lsd.lsd_osd_index);
1293 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1294 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1296 GOTO(err_env, rc = PTR_ERR(obj));
1298 if (!dt_object_exists(obj))
1299 GOTO(err_put, rc = -ENOENT);
1301 start = lock->l_policy_data.l_extent.start;
1302 end = lock->l_policy_data.l_extent.end;
1305 rc = tgt_sync(&env, tgt, obj, start, end);
1307 CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1308 "cancel: rc = %d\n",
1309 tgt_name(tgt), PFID(&fid),
1310 lock->l_policy_data.l_extent.start,
1311 lock->l_policy_data.l_extent.end, rc);
1315 dt_object_put(&env, obj);
1320 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1324 static struct ldlm_callback_suite tgt_dlm_cbs = {
1325 .lcs_completion = ldlm_server_completion_ast,
1326 .lcs_blocking = tgt_blocking_ast,
1327 .lcs_glimpse = ldlm_server_glimpse_ast
1330 int tgt_enqueue(struct tgt_session_info *tsi)
1332 struct ptlrpc_request *req = tgt_ses_req(tsi);
1337 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1338 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1340 LASSERT(tsi->tsi_dlm_req != NULL);
1341 rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1342 tsi->tsi_dlm_req, &tgt_dlm_cbs);
1344 RETURN(err_serious(rc));
1346 switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1348 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1351 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1354 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1357 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1360 RETURN(req->rq_status);
1362 EXPORT_SYMBOL(tgt_enqueue);
1364 int tgt_convert(struct tgt_session_info *tsi)
1366 struct ptlrpc_request *req = tgt_ses_req(tsi);
1370 LASSERT(tsi->tsi_dlm_req);
1371 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1373 RETURN(err_serious(rc));
1375 RETURN(req->rq_status);
1378 int tgt_bl_callback(struct tgt_session_info *tsi)
1380 return err_serious(-EOPNOTSUPP);
1383 int tgt_cp_callback(struct tgt_session_info *tsi)
1385 return err_serious(-EOPNOTSUPP);
1388 /* generic LDLM target handler */
1389 struct tgt_handler tgt_dlm_handlers[] = {
1390 TGT_DLM_HDL (HABEO_CLAVIS, LDLM_ENQUEUE, tgt_enqueue),
1391 TGT_DLM_HDL (HABEO_CLAVIS, LDLM_CONVERT, tgt_convert),
1392 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1393 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1395 EXPORT_SYMBOL(tgt_dlm_handlers);
1398 * Unified target LLOG handlers.
1400 int tgt_llog_open(struct tgt_session_info *tsi)
1406 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1410 EXPORT_SYMBOL(tgt_llog_open);
1412 int tgt_llog_read_header(struct tgt_session_info *tsi)
1418 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1422 EXPORT_SYMBOL(tgt_llog_read_header);
1424 int tgt_llog_next_block(struct tgt_session_info *tsi)
1430 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1434 EXPORT_SYMBOL(tgt_llog_next_block);
1436 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1442 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1446 EXPORT_SYMBOL(tgt_llog_prev_block);
1448 /* generic llog target handler */
1449 struct tgt_handler tgt_llog_handlers[] = {
1450 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1451 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1452 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1453 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1455 EXPORT_SYMBOL(tgt_llog_handlers);
1458 * sec context handlers
1460 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1461 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1466 struct tgt_handler tgt_sec_ctx_handlers[] = {
1467 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle),
1468 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle),
1469 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle),
1471 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1473 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1474 struct dt_device *key,
1475 struct lfsck_req_local *lrl,
1476 struct thandle *th) = NULL;
1478 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1480 struct lfsck_req_local *,
1483 tgt_lfsck_in_notify_local = notify;
1485 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1487 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1488 struct dt_device *key,
1489 struct lfsck_request *lr) = NULL;
1491 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1493 struct lfsck_request *))
1495 tgt_lfsck_in_notify = notify;
1497 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1499 static int (*tgt_lfsck_query)(const struct lu_env *env,
1500 struct dt_device *key,
1501 struct lfsck_request *req,
1502 struct lfsck_reply *rep,
1503 struct lfsck_query *que) = NULL;
1505 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1507 struct lfsck_request *,
1508 struct lfsck_reply *,
1509 struct lfsck_query *))
1511 tgt_lfsck_query = query;
1513 EXPORT_SYMBOL(tgt_register_lfsck_query);
1515 /* LFSCK request handlers */
1516 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1518 const struct lu_env *env = tsi->tsi_env;
1519 struct dt_device *key = tsi->tsi_tgt->lut_bottom;
1520 struct lfsck_request *lr;
1524 lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1528 rc = tgt_lfsck_in_notify(env, key, lr);
1533 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1535 struct lfsck_request *request;
1536 struct lfsck_reply *reply;
1540 request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1541 if (request == NULL)
1544 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1548 rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1549 request, reply, NULL);
1551 RETURN(rc < 0 ? rc : 0);
1554 struct tgt_handler tgt_lfsck_handlers[] = {
1555 TGT_LFSCK_HDL(HABEO_REFERO, LFSCK_NOTIFY, tgt_handle_lfsck_notify),
1556 TGT_LFSCK_HDL(HABEO_REFERO, LFSCK_QUERY, tgt_handle_lfsck_query),
1558 EXPORT_SYMBOL(tgt_lfsck_handlers);
1561 * initialize per-thread page pool (bug 5137).
1563 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1565 struct tgt_thread_big_cache *tbc;
1569 LASSERT(thread != NULL);
1570 LASSERT(thread->t_data == NULL);
1572 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1575 thread->t_data = tbc;
1578 EXPORT_SYMBOL(tgt_io_thread_init);
1581 * free per-thread pool created by tgt_thread_init().
1583 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1585 struct tgt_thread_big_cache *tbc;
1589 LASSERT(thread != NULL);
1592 * be prepared to handle partially-initialized pools (because this is
1593 * called from ost_io_thread_init() for cleanup.
1595 tbc = thread->t_data;
1597 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1598 thread->t_data = NULL;
1602 EXPORT_SYMBOL(tgt_io_thread_done);
1605 * Helper function for getting Data-on-MDT file server DLM lock
1606 * if asked by client.
1608 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1609 struct lustre_handle *lh, int mode, __u64 *flags)
1611 union ldlm_policy_data policy = {
1612 .l_inodebits.bits = MDS_INODELOCK_DOM,
1618 LASSERT(lh != NULL);
1619 LASSERT(ns != NULL);
1620 LASSERT(!lustre_handle_is_used(lh));
1622 rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1623 flags, ldlm_blocking_ast,
1624 ldlm_completion_ast, ldlm_glimpse_ast,
1625 NULL, 0, LVB_T_NONE, NULL, lh);
1627 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1629 EXPORT_SYMBOL(tgt_mdt_data_lock);
1631 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1633 LASSERT(lustre_handle_is_used(lh));
1634 ldlm_lock_decref(lh, mode);
1636 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1639 * Helper function for getting server side [start, start+count] DLM lock
1640 * if asked by client.
1642 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1643 struct ldlm_res_id *res_id, __u64 start, __u64 end,
1644 struct lustre_handle *lh, int mode, __u64 *flags)
1646 union ldlm_policy_data policy;
1651 LASSERT(lh != NULL);
1652 LASSERT(ns != NULL);
1653 LASSERT(!lustre_handle_is_used(lh));
1655 policy.l_extent.gid = 0;
1656 policy.l_extent.start = start & PAGE_MASK;
1659 * If ->o_blocks is EOF it means "lock till the end of the file".
1660 * Otherwise, it's size of an extent or hole being punched (in bytes).
1662 if (end == OBD_OBJECT_EOF || end < start)
1663 policy.l_extent.end = OBD_OBJECT_EOF;
1665 policy.l_extent.end = end | ~PAGE_MASK;
1667 rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1668 mode, flags, ldlm_blocking_ast,
1669 ldlm_completion_ast, ldlm_glimpse_ast,
1670 NULL, 0, LVB_T_NONE, NULL, lh);
1671 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1673 EXPORT_SYMBOL(tgt_extent_lock);
1675 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1677 LASSERT(lustre_handle_is_used(lh));
1678 ldlm_lock_decref(lh, mode);
1680 EXPORT_SYMBOL(tgt_extent_unlock);
1682 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1683 struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1684 struct niobuf_remote *nb, struct lustre_handle *lh,
1685 enum ldlm_mode mode)
1687 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1689 int nrbufs = obj->ioo_bufcnt;
1695 LASSERT(mode == LCK_PR || mode == LCK_PW);
1696 LASSERT(!lustre_handle_is_used(lh));
1698 if (ns->ns_obd->obd_recovering)
1701 if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1704 for (i = 1; i < nrbufs; i++)
1705 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1708 /* MDT IO for data-on-mdt */
1709 if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1710 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1712 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1713 nb[nrbufs - 1].rnb_offset +
1714 nb[nrbufs - 1].rnb_len - 1,
1719 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1720 struct lustre_handle *lh, enum ldlm_mode mode)
1724 LASSERT(mode == LCK_PR || mode == LCK_PW);
1725 LASSERT((obj->ioo_bufcnt > 0 &&
1726 (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1727 lustre_handle_is_used(lh));
1729 if (lustre_handle_is_used(lh))
1730 tgt_extent_unlock(lh, mode);
1733 static int tgt_checksum_niobuf(struct lu_target *tgt,
1734 struct niobuf_local *local_nb, int npages,
1735 int opc, enum cksum_types cksum_type,
1738 struct ahash_request *req;
1739 unsigned int bufsize;
1741 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1743 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1745 CERROR("%s: unable to initialize checksum hash %s\n",
1746 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1747 return PTR_ERR(req);
1750 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1751 for (i = 0; i < npages; i++) {
1752 /* corrupt the data before we compute the checksum, to
1753 * simulate a client->OST data error */
1754 if (i == 0 && opc == OST_WRITE &&
1755 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1756 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1757 int len = local_nb[i].lnb_len;
1758 struct page *np = tgt_page_to_corrupt;
1761 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1763 char *ptr2 = page_address(np);
1765 memcpy(ptr2 + off, ptr + off, len);
1766 memcpy(ptr2 + off, "bad3", min(4, len));
1767 ll_kunmap_atomic(ptr, KM_USER0);
1769 /* LU-8376 to preserve original index for
1770 * display in dump_all_bulk_pages() */
1773 cfs_crypto_hash_update_page(req, np, off,
1777 CERROR("%s: can't alloc page for corruption\n",
1781 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1782 local_nb[i].lnb_page_offset & ~PAGE_MASK,
1783 local_nb[i].lnb_len);
1785 /* corrupt the data after we compute the checksum, to
1786 * simulate an OST->client data error */
1787 if (i == 0 && opc == OST_READ &&
1788 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1789 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1790 int len = local_nb[i].lnb_len;
1791 struct page *np = tgt_page_to_corrupt;
1794 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1796 char *ptr2 = page_address(np);
1798 memcpy(ptr2 + off, ptr + off, len);
1799 memcpy(ptr2 + off, "bad4", min(4, len));
1800 ll_kunmap_atomic(ptr, KM_USER0);
1802 /* LU-8376 to preserve original index for
1803 * display in dump_all_bulk_pages() */
1806 cfs_crypto_hash_update_page(req, np, off,
1810 CERROR("%s: can't alloc page for corruption\n",
1816 bufsize = sizeof(*cksum);
1817 err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1822 char dbgcksum_file_name[PATH_MAX];
1824 static void dump_all_bulk_pages(struct obdo *oa, int count,
1825 struct niobuf_local *local_nb,
1826 __u32 server_cksum, __u32 client_cksum)
1833 /* will only keep dump of pages on first error for the same range in
1834 * file/fid, not during the resends/retries. */
1835 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1836 "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1837 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1838 libcfs_debug_file_path_arr :
1839 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1840 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1841 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1842 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1843 local_nb[0].lnb_file_offset,
1844 local_nb[count-1].lnb_file_offset +
1845 local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1846 filp = filp_open(dbgcksum_file_name,
1847 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1851 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1852 "checksum error: rc = %d\n", dbgcksum_file_name,
1855 CERROR("%s: can't open to dump pages with checksum "
1856 "error: rc = %d\n", dbgcksum_file_name, rc);
1860 for (i = 0; i < count; i++) {
1861 len = local_nb[i].lnb_len;
1862 buf = kmap(local_nb[i].lnb_page);
1864 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1866 CERROR("%s: wanted to write %u but got %d "
1867 "error\n", dbgcksum_file_name, len, rc);
1872 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1873 dbgcksum_file_name, rc);
1875 kunmap(local_nb[i].lnb_page);
1878 rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1880 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1881 filp_close(filp, NULL);
1885 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1886 struct obd_export *exp, struct obdo *oa,
1887 const struct lnet_process_id *peer,
1888 __u32 client_cksum, __u32 server_cksum,
1889 enum cksum_types server_cksum_type)
1892 enum cksum_types cksum_type;
1895 /* unlikely to happen and only if resend does not occur due to cksum
1896 * control failure on Client */
1897 if (unlikely(server_cksum == client_cksum)) {
1898 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1903 if (exp->exp_obd->obd_checksum_dump)
1904 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1907 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1910 if (cksum_type != server_cksum_type)
1911 msg = "the server may have not used the checksum type specified"
1912 " in the original request - likely a protocol problem";
1914 msg = "should have changed on the client or in transit";
1916 start = local_nb[0].lnb_file_offset;
1917 end = local_nb[npages-1].lnb_file_offset +
1918 local_nb[npages-1].lnb_len - 1;
1920 LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1921 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1922 " %x (type %x), server csum %x (type %x)\n",
1923 exp->exp_obd->obd_name,
1924 msg, libcfs_nid2str(peer->nid),
1925 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1926 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1927 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1929 start, end, client_cksum, cksum_type, server_cksum,
1935 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1936 unsigned char *buf, int size)
1938 int i, off, len, copied = size;
1941 for (i = 0; i < npages; i++) {
1942 off = local[i].lnb_page_offset & ~PAGE_MASK;
1943 len = local[i].lnb_len;
1945 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1950 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
1951 memcpy(buf + off, ptr, len);
1952 ll_kunmap_atomic(ptr, KM_USER0);
1956 return copied - size;
1959 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1960 struct niobuf_local *local_nb,
1961 int npages, int opc,
1962 obd_dif_csum_fn *fn,
1966 enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1967 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1968 const char *obd_name = tgt->lut_obd->obd_name;
1969 struct ahash_request *req;
1970 unsigned int bufsize;
1971 unsigned char *buffer;
1972 struct page *__page;
1975 int used_number = 0;
1981 __page = alloc_page(GFP_KERNEL);
1985 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1987 CERROR("%s: unable to initialize checksum hash %s\n",
1988 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1989 return PTR_ERR(req);
1992 buffer = kmap(__page);
1993 guard_start = (__u16 *)buffer;
1994 guard_number = PAGE_SIZE / sizeof(*guard_start);
1995 for (i = 0; i < npages; i++) {
1996 /* corrupt the data before we compute the checksum, to
1997 * simulate a client->OST data error */
1998 if (i == 0 && opc == OST_WRITE &&
1999 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2000 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2001 int len = local_nb[i].lnb_len;
2002 struct page *np = tgt_page_to_corrupt;
2005 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2007 char *ptr2 = page_address(np);
2009 memcpy(ptr2 + off, ptr + off, len);
2010 memcpy(ptr2 + off, "bad3", min(4, len));
2011 ll_kunmap_atomic(ptr, KM_USER0);
2013 /* LU-8376 to preserve original index for
2014 * display in dump_all_bulk_pages() */
2017 cfs_crypto_hash_update_page(req, np, off,
2021 CERROR("%s: can't alloc page for corruption\n",
2027 * The left guard number should be able to hold checksums of a
2030 if (t10_cksum_type && opc == OST_READ &&
2031 local_nb[i].lnb_guard_disk) {
2032 used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2033 if (used > (guard_number - used_number)) {
2037 memcpy(guard_start + used_number,
2038 local_nb[i].lnb_guards,
2039 used * sizeof(*local_nb[i].lnb_guards));
2041 rc = obd_page_dif_generate_buffer(obd_name,
2042 local_nb[i].lnb_page,
2043 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2044 local_nb[i].lnb_len, guard_start + used_number,
2045 guard_number - used_number, &used, sector_size,
2051 LASSERT(used <= MAX_GUARD_NUMBER);
2053 * If disk support T10PI checksum, copy guards to local_nb.
2054 * If the write is partial page, do not use the guards for bio
2055 * submission since the data might not be full-sector. The bio
2056 * guards will be generated later based on the full sectors. If
2057 * the sector size is 512B rather than 4 KB, or the page size
2058 * is larger than 4KB, this might drop some useful guards for
2059 * partial page write, but it will only add minimal extra time
2060 * of checksum calculation.
2062 if (t10_cksum_type && opc == OST_WRITE &&
2063 local_nb[i].lnb_len == PAGE_SIZE) {
2064 local_nb[i].lnb_guard_rpc = 1;
2065 memcpy(local_nb[i].lnb_guards,
2066 guard_start + used_number,
2067 used * sizeof(*local_nb[i].lnb_guards));
2070 used_number += used;
2071 if (used_number == guard_number) {
2072 cfs_crypto_hash_update_page(req, __page, 0,
2073 used_number * sizeof(*guard_start));
2077 /* corrupt the data after we compute the checksum, to
2078 * simulate an OST->client data error */
2079 if (unlikely(i == 0 && opc == OST_READ &&
2080 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2081 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2082 int len = local_nb[i].lnb_len;
2083 struct page *np = tgt_page_to_corrupt;
2086 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2088 char *ptr2 = page_address(np);
2090 memcpy(ptr2 + off, ptr + off, len);
2091 memcpy(ptr2 + off, "bad4", min(4, len));
2092 ll_kunmap_atomic(ptr, KM_USER0);
2094 /* LU-8376 to preserve original index for
2095 * display in dump_all_bulk_pages() */
2098 cfs_crypto_hash_update_page(req, np, off,
2102 CERROR("%s: can't alloc page for corruption\n",
2111 if (used_number != 0)
2112 cfs_crypto_hash_update_page(req, __page, 0,
2113 used_number * sizeof(*guard_start));
2115 bufsize = sizeof(cksum);
2116 rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2121 __free_page(__page);
2125 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2126 enum cksum_types cksum_type,
2127 struct niobuf_local *local_nb,
2128 int npages, int opc, u32 *check_sum)
2130 obd_dif_csum_fn *fn = NULL;
2131 int sector_size = 0;
2135 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
2138 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2139 opc, fn, sector_size,
2142 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2143 cksum_type, check_sum);
2147 int tgt_brw_read(struct tgt_session_info *tsi)
2149 struct ptlrpc_request *req = tgt_ses_req(tsi);
2150 struct ptlrpc_bulk_desc *desc = NULL;
2151 struct obd_export *exp = tsi->tsi_exp;
2152 struct niobuf_remote *remote_nb;
2153 struct niobuf_local *local_nb;
2154 struct obd_ioobj *ioo;
2155 struct ost_body *body, *repbody;
2156 struct l_wait_info lwi;
2157 struct lustre_handle lockh = { 0 };
2158 int npages, nob = 0, rc, i, no_reply = 0,
2160 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2161 const char *obd_name = exp->exp_obd->obd_name;
2165 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2166 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2167 CERROR("%s: deny read request from %s to portal %u\n",
2168 tgt_name(tsi->tsi_tgt),
2169 obd_export_nid2str(req->rq_export),
2170 ptlrpc_req2svc(req)->srv_req_portal);
2174 req->rq_bulk_read = 1;
2176 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2179 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2180 cfs_fail_val : (obd_timeout + 1) / 4);
2182 /* Check if there is eviction in progress, and if so, wait for it to
2184 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2185 /* We do not care how long it takes */
2186 lwi = LWI_INTR(NULL, NULL);
2187 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2188 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
2192 /* There must be big cache in current thread to process this request
2193 * if it is NULL then something went wrong and it wasn't allocated,
2194 * report -ENOMEM in that case */
2198 body = tsi->tsi_ost_body;
2199 LASSERT(body != NULL);
2201 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2202 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2204 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2205 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2207 local_nb = tbc->local;
2209 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2215 * If getting the lock took more time than
2216 * client was willing to wait, drop it. b=11330
2218 if (ktime_get_real_seconds() > req->rq_deadline ||
2219 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2221 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2222 libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2223 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2224 req->rq_deadline - req->rq_arrival_time.tv_sec);
2225 GOTO(out_lock, rc = -ETIMEDOUT);
2228 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2229 repbody->oa = body->oa;
2231 npages = PTLRPC_MAX_BRW_PAGES;
2232 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2233 ioo, remote_nb, &npages, local_nb);
2237 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2238 body->oa.o_flags & OBD_FL_SHORT_IO) {
2241 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2242 PTLRPC_BULK_PUT_SOURCE |
2243 PTLRPC_BULK_BUF_KIOV,
2245 &ptlrpc_bulk_kiov_nopin_ops);
2247 GOTO(out_commitrw, rc = -ENOMEM);
2251 npages_read = npages;
2252 for (i = 0; i < npages; i++) {
2253 int page_rc = local_nb[i].lnb_rc;
2262 if (page_rc != 0 && desc != NULL) { /* some data! */
2263 LASSERT(local_nb[i].lnb_page != NULL);
2264 desc->bd_frag_ops->add_kiov_frag
2265 (desc, local_nb[i].lnb_page,
2266 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2270 if (page_rc != local_nb[i].lnb_len) { /* short read */
2271 local_nb[i].lnb_len = page_rc;
2272 npages_read = i + (page_rc != 0 ? 1 : 0);
2273 /* All subsequent pages should be 0 */
2274 while (++i < npages)
2275 LASSERT(local_nb[i].lnb_rc == 0);
2279 if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
2280 nob != cfs_fail_val)
2283 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2284 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2285 body->oa.o_flags : 0;
2286 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2288 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2290 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2292 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2293 local_nb, npages_read, OST_READ,
2294 &repbody->oa.o_cksum);
2296 GOTO(out_commitrw, rc);
2297 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2298 repbody->oa.o_cksum);
2300 /* if a resend it could be for a cksum error, so check Server
2301 * cksum with returned Client cksum (this should even cover
2302 * zero-cksum case) */
2303 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2304 (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2305 check_read_checksum(local_nb, npages_read, exp,
2306 &body->oa, &req->rq_peer,
2308 repbody->oa.o_cksum, cksum_type);
2310 repbody->oa.o_valid = 0;
2312 /* We're finishing using body->oa as an input variable */
2314 /* Check if client was evicted while we were doing i/o before touching
2317 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2318 body->oa.o_flags & OBD_FL_SHORT_IO) {
2319 unsigned char *short_io_buf;
2322 short_io_buf = req_capsule_server_get(&req->rq_pill,
2324 short_io_size = req_capsule_get_size(&req->rq_pill,
2327 rc = tgt_pages2shortio(local_nb, npages_read,
2328 short_io_buf, short_io_size);
2330 req_capsule_shrink(&req->rq_pill,
2333 rc = rc > 0 ? 0 : rc;
2334 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2335 rc = target_bulk_io(exp, desc, &lwi);
2339 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2340 body->oa.o_flags & OBD_FL_SHORT_IO)
2341 req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2346 /* Must commit after prep above in all cases */
2347 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2348 remote_nb, npages, local_nb, rc);
2350 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2352 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2353 ptlrpc_free_bulk(desc);
2358 ptlrpc_lprocfs_brw(req, nob);
2359 } else if (no_reply) {
2360 req->rq_no_reply = 1;
2361 /* reply out callback would free */
2362 ptlrpc_req_drop_rs(req);
2363 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2364 "client will retry: rc %d\n",
2366 obd_uuid2str(&exp->exp_client_uuid),
2367 obd_export_nid2str(exp), rc);
2369 /* send a bulk after reply to simulate a network delay or reordering
2370 * by a router - Note that !desc implies short io, so there is no bulk
2372 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2374 wait_queue_head_t waitq;
2375 struct l_wait_info lwi1;
2377 CDEBUG(D_INFO, "reorder BULK\n");
2378 init_waitqueue_head(&waitq);
2380 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
2381 l_wait_event(waitq, 0, &lwi1);
2382 target_bulk_io(exp, desc, &lwi);
2383 ptlrpc_free_bulk(desc);
2388 EXPORT_SYMBOL(tgt_brw_read);
2390 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2391 unsigned char *buf, unsigned int size)
2396 for (i = 0; i < npages; i++) {
2397 off = local[i].lnb_page_offset & ~PAGE_MASK;
2398 len = local[i].lnb_len;
2403 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2405 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
2408 memcpy(ptr + off, buf, len < size ? len : size);
2409 ll_kunmap_atomic(ptr, KM_USER0);
2416 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2417 struct ptlrpc_bulk_desc *desc,
2418 struct niobuf_local *local_nb, int npages,
2419 u32 client_cksum, u32 server_cksum,
2422 struct obd_export *exp = req->rq_export;
2423 struct ost_body *body;
2427 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2428 LASSERT(body != NULL);
2430 if (desc && req->rq_peer.nid != desc->bd_sender) {
2432 router = libcfs_nid2str(desc->bd_sender);
2435 if (exp->exp_obd->obd_checksum_dump)
2436 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2440 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2441 client_cksum, server_cksum);
2445 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2446 DFID" object "DOSTID" extent [%llu-%llu"
2447 "]: client csum %x, server csum %x\n",
2448 exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2450 body->oa.o_valid & OBD_MD_FLFID ?
2451 body->oa.o_parent_seq : (__u64)0,
2452 body->oa.o_valid & OBD_MD_FLFID ?
2453 body->oa.o_parent_oid : 0,
2454 body->oa.o_valid & OBD_MD_FLFID ?
2455 body->oa.o_parent_ver : 0,
2456 POSTID(&body->oa.o_oi),
2457 local_nb[0].lnb_file_offset,
2458 local_nb[npages-1].lnb_file_offset +
2459 local_nb[npages - 1].lnb_len - 1,
2460 client_cksum, server_cksum);
2463 int tgt_brw_write(struct tgt_session_info *tsi)
2465 struct ptlrpc_request *req = tgt_ses_req(tsi);
2466 struct ptlrpc_bulk_desc *desc = NULL;
2467 struct obd_export *exp = req->rq_export;
2468 struct niobuf_remote *remote_nb;
2469 struct niobuf_local *local_nb;
2470 struct obd_ioobj *ioo;
2471 struct ost_body *body, *repbody;
2472 struct l_wait_info lwi;
2473 struct lustre_handle lockh = {0};
2475 int objcount, niocount, npages;
2477 enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2478 bool no_reply = false, mmap;
2479 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2480 bool wait_sync = false;
2481 const char *obd_name = exp->exp_obd->obd_name;
2485 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2486 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2487 CERROR("%s: deny write request from %s to portal %u\n",
2488 tgt_name(tsi->tsi_tgt),
2489 obd_export_nid2str(req->rq_export),
2490 ptlrpc_req2svc(req)->srv_req_portal);
2491 RETURN(err_serious(-EPROTO));
2494 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2495 RETURN(err_serious(-ENOSPC));
2496 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2497 RETURN(err_serious(-EROFS));
2499 req->rq_bulk_write = 1;
2501 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2502 RETURN(err_serious(-EIO));
2503 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2504 RETURN(err_serious(-EFAULT));
2506 /* pause before transaction has been started */
2507 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2508 cfs_fail_val : (obd_timeout + 1) / 4);
2510 /* Delay write commit to show stale size information */
2511 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2513 /* There must be big cache in current thread to process this request
2514 * if it is NULL then something went wrong and it wasn't allocated,
2515 * report -ENOMEM in that case */
2519 body = tsi->tsi_ost_body;
2520 LASSERT(body != NULL);
2522 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2523 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2525 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2526 RCL_CLIENT) / sizeof(*ioo);
2528 for (niocount = i = 0; i < objcount; i++)
2529 niocount += ioo[i].ioo_bufcnt;
2531 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2532 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2533 if (niocount != req_capsule_get_size(&req->rq_pill,
2534 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2536 RETURN(err_serious(-EPROTO));
2538 if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2539 ptlrpc_connection_is_local(exp->exp_connection))
2540 memory_pressure_set();
2542 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2543 niocount * sizeof(*rcs));
2544 rc = req_capsule_server_pack(&req->rq_pill);
2546 GOTO(out, rc = err_serious(rc));
2548 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2549 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2551 local_nb = tbc->local;
2553 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2559 * If getting the lock took more time than
2560 * client was willing to wait, drop it. b=11330
2562 if (ktime_get_real_seconds() > req->rq_deadline ||
2563 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2565 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2566 tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2567 POSTID(&ioo->ioo_oid),
2568 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2569 req->rq_deadline - req->rq_arrival_time.tv_sec);
2570 GOTO(out_lock, rc = -ETIMEDOUT);
2573 /* Because we already sync grant info with client when reconnect,
2574 * grant info will be cleared for resent req, then fed_grant and
2575 * total_grant will not be modified in following preprw_write */
2576 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2577 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2578 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2581 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2582 if (repbody == NULL)
2583 GOTO(out_lock, rc = -ENOMEM);
2584 repbody->oa = body->oa;
2586 npages = PTLRPC_MAX_BRW_PAGES;
2587 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2588 objcount, ioo, remote_nb, &npages, local_nb);
2591 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2592 body->oa.o_flags & OBD_FL_SHORT_IO) {
2593 unsigned int short_io_size;
2594 unsigned char *short_io_buf;
2596 short_io_size = req_capsule_get_size(&req->rq_pill,
2599 short_io_buf = req_capsule_client_get(&req->rq_pill,
2601 CDEBUG(D_INFO, "Client use short io for data transfer,"
2602 " size = %d\n", short_io_size);
2604 /* Copy short io buf to pages */
2605 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2609 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2610 PTLRPC_BULK_GET_SINK |
2611 PTLRPC_BULK_BUF_KIOV,
2613 &ptlrpc_bulk_kiov_nopin_ops);
2615 GOTO(skip_transfer, rc = -ENOMEM);
2617 /* NB Having prepped, we must commit... */
2618 for (i = 0; i < npages; i++)
2619 desc->bd_frag_ops->add_kiov_frag(desc,
2620 local_nb[i].lnb_page,
2621 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2622 local_nb[i].lnb_len);
2624 rc = sptlrpc_svc_prep_bulk(req, desc);
2626 GOTO(skip_transfer, rc);
2628 rc = target_bulk_io(exp, desc, &lwi);
2634 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2635 static int cksum_counter;
2637 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2638 cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2640 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2641 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2642 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2645 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2646 local_nb, npages, OST_WRITE,
2647 &repbody->oa.o_cksum);
2649 GOTO(out_commitrw, rc);
2653 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2654 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2655 body->oa.o_flags & OBD_FL_MMAP);
2657 tgt_warn_on_cksum(req, desc, local_nb, npages,
2659 repbody->oa.o_cksum, mmap);
2661 } else if ((cksum_counter & (-cksum_counter)) ==
2663 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2664 cksum_counter, libcfs_id2str(req->rq_peer),
2665 repbody->oa.o_cksum);
2670 /* Must commit after prep above in all cases */
2671 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2672 objcount, ioo, remote_nb, npages, local_nb, rc);
2673 if (rc == -ENOTCONN)
2674 /* quota acquire process has been given up because
2675 * either the client has been evicted or the client
2676 * has timed out the request already */
2679 for (i = 0; i < niocount; i++) {
2680 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2686 * Disable sending mtime back to the client. If the client locked the
2687 * whole object, then it has already updated the mtime on its side,
2688 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2690 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2695 /* set per-requested niobuf return codes */
2696 for (i = j = 0; i < niocount; i++) {
2697 int len = remote_nb[i].rnb_len;
2702 LASSERT(j < npages);
2703 if (local_nb[j].lnb_rc < 0)
2704 rcs[i] = local_nb[j].lnb_rc;
2705 len -= local_nb[j].lnb_len;
2710 LASSERT(j == npages);
2711 ptlrpc_lprocfs_brw(req, nob);
2714 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2716 ptlrpc_free_bulk(desc);
2718 if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2719 req->rq_no_reply = 1;
2720 /* reply out callback would free */
2721 ptlrpc_req_drop_rs(req);
2722 if (!exp->exp_obd->obd_no_transno)
2723 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2724 " client will retry: rc = %d\n",
2726 obd_uuid2str(&exp->exp_client_uuid),
2727 obd_export_nid2str(exp), rc);
2729 memory_pressure_clr();
2732 EXPORT_SYMBOL(tgt_brw_write);
2734 /* Check if request can be reconstructed from saved reply data
2735 * A copy of the reply data is returned in @trd if the pointer is not NULL
2737 bool req_can_reconstruct(struct ptlrpc_request *req,
2738 struct tg_reply_data *trd)
2740 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2741 struct lsd_client_data *lcd = ted->ted_lcd;
2744 if (tgt_is_multimodrpcs_client(req->rq_export))
2745 return tgt_lookup_reply(req, trd);
2747 mutex_lock(&ted->ted_lcd_lock);
2748 found = req->rq_xid == lcd->lcd_last_xid ||
2749 req->rq_xid == lcd->lcd_last_close_xid;
2751 if (found && trd != NULL) {
2752 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2753 trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2754 trd->trd_reply.lrd_transno =
2755 lcd->lcd_last_close_transno;
2756 trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2758 trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2759 trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2760 trd->trd_reply.lrd_result = lcd->lcd_last_result;
2761 trd->trd_reply.lrd_data = lcd->lcd_last_data;
2762 trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2763 trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2764 trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2765 trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2768 mutex_unlock(&ted->ted_lcd_lock);
2772 EXPORT_SYMBOL(req_can_reconstruct);