4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; if not, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
24 * Copyright (c) 2013, 2017, Intel Corporation.
27 * lustre/target/tgt_handler.c
29 * Lustre Unified Target request handler code
31 * Author: Brian Behlendorf <behlendorf1@llnl.gov>
32 * Author: Mikhail Pershin <mike.pershin@intel.com>
35 #define DEBUG_SUBSYSTEM S_CLASS
37 #include <linux/user_namespace.h>
38 #include <linux/delay.h>
39 #include <linux/uidgid.h>
41 #include <libcfs/linux/linux-mem.h>
43 #include <obd_class.h>
44 #include <obd_cksum.h>
45 #include <lustre_lfsck.h>
46 #include <lustre_nodemap.h>
47 #include <lustre_acl.h>
49 #include "tgt_internal.h"
51 char *tgt_name(struct lu_target *tgt)
53 LASSERT(tgt->lut_obd != NULL);
54 return tgt->lut_obd->obd_name;
56 EXPORT_SYMBOL(tgt_name);
59 * Generic code handling requests that have struct mdt_body passed in:
61 * - extract mdt_body from request and save it in @tsi, if present;
63 * - create lu_object, corresponding to the fid in mdt_body, and save it in
66 * - if HAS_BODY flag is set for this request type check whether object
67 * actually exists on storage (lu_object_exists()).
70 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
72 const struct mdt_body *body;
73 struct lu_object *obj;
74 struct req_capsule *pill = tsi->tsi_pill;
79 body = req_capsule_client_get(pill, &RMF_MDT_BODY);
83 tsi->tsi_mdt_body = body;
85 if (!(body->mbo_valid & OBD_MD_FLID))
88 /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
89 * in any case in pre-2.5 clients. Fix that here if needed */
90 if (unlikely(fid_is_zero(&body->mbo_fid1)))
93 if (!fid_is_sane(&body->mbo_fid1)) {
94 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
95 PFID(&body->mbo_fid1));
99 obj = lu_object_find(tsi->tsi_env,
100 &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
101 &body->mbo_fid1, NULL);
103 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
104 lu_object_put(tsi->tsi_env, obj);
107 tsi->tsi_corpus = obj;
114 tsi->tsi_fid = body->mbo_fid1;
120 * Validate oa from client.
121 * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
123 * a. objects in Single MDT FS seq = FID_SEQ_OST_MDT0, oi_id != 0
124 * b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
125 * pack ost_id. Because non-zero oi_seq will make it diffcult to tell
126 * whether this is oi_fid or real ostid. So it will check
127 * OBD_CONNECT_FID, then convert the ostid to FID for old client.
128 * c. Old FID-disable osc will send IDIF.
129 * d. new FID-enable osc/osp will send normal FID.
131 * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
132 * be used for LAST_ID file, and only being accessed inside OST now.
134 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
136 struct ost_id *oi = &oa->o_oi;
137 u64 seq = ostid_seq(oi);
138 u64 id = ostid_id(oi);
142 if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
143 fid_seq_is_echo(seq))) {
144 /* Sigh 2.[123] client still sends echo req with oi_id = 0
145 * during create, and we will reset this to 1, since this
146 * oi_id is basically useless in the following create process,
147 * but oi_id == 0 will make it difficult to tell whether it is
148 * real FID or ost_id. */
149 oi->oi_fid.f_seq = FID_SEQ_ECHO;
150 oi->oi_fid.f_oid = id ?: 1;
151 oi->oi_fid.f_ver = 0;
153 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
155 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
156 GOTO(out, rc = -EPROTO);
158 /* Note: this check might be forced in 2.5 or 2.6, i.e.
159 * all of the requests are required to setup FLGROUP */
160 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP)))
161 oa->o_valid |= OBD_MD_FLGROUP;
163 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
164 fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
165 GOTO(out, rc = -EPROTO);
167 rc = ostid_to_fid(&tti->tti_fid1, oi,
168 tsi->tsi_tgt->lut_lsd.lsd_osd_index);
169 if (unlikely(rc != 0))
172 oi->oi_fid = tti->tti_fid1;
178 CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
179 tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
183 EXPORT_SYMBOL(tgt_validate_obdo);
185 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
188 struct niobuf_remote *rnb;
189 struct obd_ioobj *ioo;
194 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
198 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
202 max_brw = ioobj_max_brw_get(ioo);
203 if (unlikely((max_brw & (max_brw - 1)) != 0)) {
204 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
205 ": rc = %d\n", tgt_name(tsi->tsi_tgt),
206 obd_export_nid2str(tsi->tsi_exp), max_brw,
207 POSTID(oi), -EPROTO);
212 obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
213 RCL_CLIENT) / sizeof(*ioo);
214 if (obj_count == 0) {
215 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
217 } else if (obj_count > 1) {
218 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
223 if (ioo->ioo_bufcnt == 0) {
224 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
228 if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
229 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
230 "bulk has too many pages (%d)",
238 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
240 struct ost_body *body;
241 struct req_capsule *pill = tsi->tsi_pill;
242 struct lu_nodemap *nodemap;
247 body = req_capsule_client_get(pill, &RMF_OST_BODY);
251 rc = tgt_validate_obdo(tsi, &body->oa);
255 nodemap = nodemap_get_from_exp(tsi->tsi_exp);
257 RETURN(PTR_ERR(nodemap));
259 body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
260 NODEMAP_CLIENT_TO_FS,
262 body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
263 NODEMAP_CLIENT_TO_FS,
265 body->oa.o_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
266 NODEMAP_CLIENT_TO_FS,
268 nodemap_putref(nodemap);
270 tsi->tsi_ost_body = body;
271 tsi->tsi_fid = body->oa.o_oi.oi_fid;
273 if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
274 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
279 if (!(body->oa.o_valid & OBD_MD_FLID)) {
280 if (flags & HAS_BODY) {
281 CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\n",
282 tgt_name(tsi->tsi_tgt));
289 ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
292 * OST doesn't get object in advance for further use to prevent
293 * situations with nested object_find which is potential deadlock.
295 tsi->tsi_corpus = NULL;
300 * Do necessary preprocessing according to handler ->th_flags.
302 static int tgt_request_preprocess(struct tgt_session_info *tsi,
303 struct tgt_handler *h,
304 struct ptlrpc_request *req)
306 struct req_capsule *pill = tsi->tsi_pill;
307 __u32 flags = h->th_flags;
312 if (tsi->tsi_preprocessed)
315 LASSERT(h->th_act != NULL);
316 LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
317 LASSERT(current->journal_info == NULL);
319 LASSERT(ergo(flags & (HAS_BODY | HAS_REPLY),
321 if (h->th_fmt != NULL) {
322 req_capsule_set(pill, h->th_fmt);
323 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT) &&
324 req_capsule_field_present(pill, &RMF_MDT_BODY,
326 rc = tgt_mdt_body_unpack(tsi, flags);
329 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
331 req_capsule_field_present(pill, &RMF_OST_BODY,
333 rc = tgt_ost_body_unpack(tsi, flags);
339 if (flags & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
342 if (flags & HAS_KEY) {
343 struct ldlm_request *dlm_req;
345 LASSERT(h->th_fmt != NULL);
347 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
348 if (dlm_req != NULL) {
349 union ldlm_wire_policy_data *policy =
350 &dlm_req->lock_desc.l_policy_data;
352 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
354 (policy->l_inodebits.bits |
355 policy->l_inodebits.try_bits) == 0)) {
357 * Lock without inodebits makes no sense and
358 * will oops later in ldlm. If client miss to
359 * set such bits, do not trigger ASSERTION.
361 * For liblustre flock case, it maybe zero.
365 tsi->tsi_dlm_req = dlm_req;
371 tsi->tsi_preprocessed = 1;
376 * Invoke handler for this request opc. Also do necessary preprocessing
377 * (according to handler ->th_flags), and post-processing (setting of
378 * ->last_{xid,committed}).
380 static int tgt_handle_request0(struct tgt_session_info *tsi,
381 struct tgt_handler *h,
382 struct ptlrpc_request *req)
386 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
391 /* When dealing with sec context requests, no export is associated yet,
392 * because these requests are sent before *_CONNECT requests.
393 * A NULL req->rq_export means the normal *_common_slice handlers will
394 * not be called, because there is no reference to the target.
395 * So deal with them by hand and jump directly to target_send_reply().
399 case SEC_CTX_INIT_CONT:
401 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
406 * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
407 * to put same checks into handlers like mdt_close(), mdt_reint(),
408 * etc., without talking to mdt authors first. Checking same thing
409 * there again is useless and returning 0 error without packing reply
410 * is buggy! Handlers either pack reply or return error.
412 * We return 0 here and do not send any reply in order to emulate
413 * network failure. Do not send any reply in case any of NET related
414 * fail_id has occured.
416 if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
418 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
419 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
422 /* drop OUT_UPDATE rpc */
423 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == OUT_UPDATE &&
424 OBD_FAIL_CHECK(OBD_FAIL_OUT_UPDATE_DROP)))
427 rc = tgt_request_preprocess(tsi, h, req);
428 /* pack reply if reply format is fixed */
429 if (rc == 0 && h->th_flags & HAS_REPLY) {
431 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
433 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
435 tsi->tsi_mdt_body->mbo_eadatasize);
436 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
438 req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
440 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
441 req_capsule_set_size(tsi->tsi_pill,
442 &RMF_ACL, RCL_SERVER,
443 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
445 if (req_capsule_has_field(tsi->tsi_pill, &RMF_SHORT_IO,
447 struct niobuf_remote *remote_nb =
448 req_capsule_client_get(tsi->tsi_pill,
450 struct ost_body *body = tsi->tsi_ost_body;
452 req_capsule_set_size(tsi->tsi_pill, &RMF_SHORT_IO,
454 (body->oa.o_valid & OBD_MD_FLFLAGS &&
455 body->oa.o_flags & OBD_FL_SHORT_IO) ?
456 remote_nb[0].rnb_len : 0);
458 if (req_capsule_has_field(tsi->tsi_pill, &RMF_FILE_ENCCTX,
460 req_capsule_set_size(tsi->tsi_pill, &RMF_FILE_ENCCTX,
463 rc = req_capsule_server_pack(tsi->tsi_pill);
466 if (likely(rc == 0)) {
468 * Process request, there can be two types of rc:
469 * 1) errors with msg unpack/pack, other failures outside the
470 * operation itself. This is counted as serious errors;
471 * 2) errors during fs operation, should be placed in rq_status
475 if (!is_serious(rc) &&
476 !req->rq_no_reply && req->rq_reply_state == NULL) {
477 DEBUG_REQ(D_ERROR, req,
478 "%s: %s handler did not pack reply but returned no error",
479 tgt_name(tsi->tsi_tgt), h->th_name);
482 serious = is_serious(rc);
483 rc = clear_serious(rc);
491 * ELDLM_* codes which > 0 should be in rq_status only as well as
492 * all non-serious errors.
494 if (rc > 0 || !serious)
497 LASSERT(current->journal_info == NULL);
499 if (likely(rc == 0 && req->rq_export))
500 target_committed_to_req(req);
503 target_send_reply(req, rc, tsi->tsi_reply_fail_id);
507 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
508 struct obd_device *obd, int *process)
510 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
517 case MDS_SYNC: /* used in unmounting */
532 case MDS_HSM_PROGRESS:
533 case MDS_HSM_STATE_SET:
534 case MDS_HSM_REQUEST:
536 *process = target_queue_recovery_request(req, obd);
540 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
547 * Handle recovery. Return:
548 * +1: continue request processing;
549 * -ve: abort immediately with the given error code;
550 * 0: send reply with error code in req->rq_status;
552 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
556 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
561 case SEC_CTX_INIT_CONT:
566 if (!req->rq_export->exp_obd->obd_replayable)
569 /* sanity check: if the xid matches, the request must be marked as a
570 * resent or replayed */
571 if (req_can_reconstruct(req, NULL) == 1) {
572 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
573 (MSG_RESENT | MSG_REPLAY))) {
574 DEBUG_REQ(D_WARNING, req,
575 "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
577 lustre_msg_get_flags(req->rq_reqmsg));
578 req->rq_status = -ENOTCONN;
582 /* else: note the opposite is not always true; a RESENT req after a
583 * failover will usually not match the last_xid, since it was likely
584 * never committed. A REPLAYed request will almost never match the
585 * last xid, however it could for a committed, but still retained,
588 /* Check for aborted recovery... */
589 if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
593 DEBUG_REQ(D_INFO, req, "Got new replay");
594 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
596 if (rc != 0 || !should_process)
598 else if (should_process < 0) {
599 req->rq_status = should_process;
600 rc = ptlrpc_error(req);
607 /* Initial check for request, it is validation mostly */
608 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
610 struct tgt_handler *h;
611 struct tgt_opc_slice *s;
612 struct lu_target *tgt;
613 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
614 /* don't spew error messages for unhandled RPCs */
619 tgt = class_exp2tgt(req->rq_export);
620 if (unlikely(tgt == NULL)) {
621 DEBUG_REQ(D_ERROR, req, "%s: no target for connected export",
622 class_exp2obd(req->rq_export)->obd_name);
623 RETURN(ERR_PTR(-EINVAL));
626 for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
627 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
630 /* opcode was not found in slice */
631 if (unlikely(s->tos_hs == NULL)) {
633 CERROR("%s: no handler for opcode 0x%x from %s\n",
634 tgt_name(tgt), opc, libcfs_idstr(&req->rq_peer));
637 goto err_unsupported;
640 LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
641 h = s->tos_hs + (opc - s->tos_opc_start);
642 if (unlikely(h->th_opc == 0)) {
644 CERROR("%s: unsupported opcode 0x%x\n",
648 goto err_unsupported;
651 if (OBD_FAIL_CHECK(OBD_FAIL_OST_OPCODE) && opc == cfs_fail_val)
652 goto err_unsupported;
657 * Unknown opcode does not necessarily means insane client. A new
658 * client might send RPCs with new opcodes to an old server. The
659 * client might desperately stuck there waiting for a reply. So,
660 * send an error back here.
662 * An old client might also send RPCs with deprecated opcodes (e.g.
665 * Error in ptlrpc_send_error() is ignored.
667 req->rq_status = -EOPNOTSUPP;
668 ptlrpc_send_error(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
669 RETURN(ERR_PTR(-EOPNOTSUPP));
672 static int process_req_last_xid(struct ptlrpc_request *req)
676 struct obd_export *exp = req->rq_export;
677 struct tg_export_data *ted = &exp->exp_target_data;
678 bool need_lock = tgt_is_multimodrpcs_client(exp);
682 mutex_lock(&ted->ted_lcd_lock);
683 /* check request's xid is consistent with export's last_xid */
684 last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
685 if (last_xid > exp->exp_last_xid)
686 exp->exp_last_xid = last_xid;
688 if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
689 /* Some request is allowed to be sent during replay,
690 * such as OUT update requests, FLD requests, so it
691 * is possible that replay requests has smaller XID
692 * than the exp_last_xid.
694 * Some non-replay requests may have smaller XID as
697 * - Client send a no_resend RPC, like statfs;
698 * - The RPC timedout (or some other error) on client,
699 * then it's removed from the unreplied list;
700 * - Client send some other request to bump the
701 * exp_last_xid on server;
702 * - The former RPC got chance to be processed;
704 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
707 DEBUG_REQ(D_WARNING, req,
708 "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
709 req->rq_xid, exp->exp_last_xid, rc);
714 /* The "last_xid" is the minimum xid among unreplied requests,
715 * if the request is from the previous connection, its xid can
716 * still be larger than "exp_last_xid", then the above check of
717 * xid is not enough to determine whether the request is delayed.
719 * For example, if some replay request was delayed and caused
720 * timeout at client and the replay is restarted, the delayed
721 * replay request will have the larger xid than "exp_last_xid"
723 if (req->rq_export->exp_conn_cnt >
724 lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
726 "Dropping request %llu from an old epoch %u/%u\n",
728 lustre_msg_get_conn_cnt(req->rq_reqmsg),
729 req->rq_export->exp_conn_cnt);
730 req->rq_no_reply = 1;
731 GOTO(out, rc = -ESTALE);
734 /* try to release in-memory reply data */
735 if (tgt_is_multimodrpcs_client(exp)) {
736 tgt_handle_received_xid(exp, last_xid);
737 rc = tgt_handle_tag(req);
742 mutex_unlock(&ted->ted_lcd_lock);
747 int tgt_request_handle(struct ptlrpc_request *req)
749 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
751 struct lustre_msg *msg = req->rq_reqmsg;
752 struct tgt_handler *h;
753 struct lu_target *tgt;
754 int request_fail_id = 0;
755 __u32 opc = lustre_msg_get_opc(msg);
756 struct obd_device *obd;
758 bool is_connect = false;
761 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
762 if (cfs_fail_val == 0 &&
763 lustre_msg_get_opc(msg) != OBD_PING &&
764 lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
767 wait_event_idle(cfs_race_waitq, (cfs_race_state == 1));
771 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
772 tsi->tsi_pill = &req->rq_pill;
773 tsi->tsi_env = req->rq_svc_thread->t_env;
775 /* if request has export then get handlers slice from corresponding
776 * target, otherwise that should be connect operation */
777 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
778 opc == MGS_CONNECT) {
780 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
781 rc = target_handle_connect(req);
783 rc = ptlrpc_error(req);
786 /* recovery-small test 18c asks to drop connect reply */
787 if (unlikely(opc == OST_CONNECT &&
788 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
792 if (unlikely(!class_connected_export(req->rq_export))) {
793 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
794 opc == SEC_CTX_FINI) {
795 /* sec context initialization has to be handled
796 * by hand in tgt_handle_request0() */
797 tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
799 GOTO(handle_recov, rc = 0);
801 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
802 opc, libcfs_idstr(&req->rq_peer));
803 req->rq_status = -ENOTCONN;
804 rc = ptlrpc_error(req);
808 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
809 tsi->tsi_exp = req->rq_export;
810 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
811 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
813 tsi->tsi_jobid = NULL;
816 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
817 class_exp2obd(req->rq_export)->obd_name);
818 req->rq_status = -EINVAL;
819 rc = ptlrpc_error(req);
823 /* Skip last_xid processing for the recovery thread, otherwise, the
824 * last_xid on same request could be processed twice: first time when
825 * processing the incoming request, second time when the request is
826 * being processed by recovery thread. */
827 obd = class_exp2obd(req->rq_export);
829 /* reset the exp_last_xid on each connection. */
830 req->rq_export->exp_last_xid = 0;
831 } else if (obd->obd_recovery_data.trd_processing_task !=
833 rc = process_req_last_xid(req);
836 rc = ptlrpc_error(req);
841 request_fail_id = tgt->lut_request_fail_id;
842 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
844 h = tgt_handler_find_check(req);
846 req->rq_status = PTR_ERR(h);
847 rc = ptlrpc_error(req);
851 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
854 if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
855 CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
858 rc = lustre_msg_check_version(msg, h->th_version);
860 DEBUG_REQ(D_ERROR, req,
861 "%s: drop malformed request version=%08x expect=%08x",
862 tgt_name(tgt), lustre_msg_get_version(msg),
864 req->rq_status = -EINVAL;
865 rc = ptlrpc_error(req);
870 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
871 if (likely(rc == 1)) {
872 rc = tgt_handle_request0(tsi, h, req);
878 req_capsule_fini(tsi->tsi_pill);
879 if (tsi->tsi_corpus != NULL) {
880 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
881 tsi->tsi_corpus = NULL;
885 EXPORT_SYMBOL(tgt_request_handle);
887 /** Assign high priority operations to the request if needed. */
888 int tgt_hpreq_handler(struct ptlrpc_request *req)
890 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
891 struct tgt_handler *h;
896 if (req->rq_export == NULL)
899 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
900 tsi->tsi_pill = &req->rq_pill;
901 tsi->tsi_env = req->rq_svc_thread->t_env;
902 tsi->tsi_tgt = class_exp2tgt(req->rq_export);
903 tsi->tsi_exp = req->rq_export;
905 h = tgt_handler_find_check(req);
911 rc = tgt_request_preprocess(tsi, h, req);
912 if (unlikely(rc != 0))
915 if (h->th_hp != NULL)
919 EXPORT_SYMBOL(tgt_hpreq_handler);
921 void tgt_counter_incr(struct obd_export *exp, int opcode)
923 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
924 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
925 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
927 EXPORT_SYMBOL(tgt_counter_incr);
930 * Unified target generic handlers.
933 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
935 struct lu_target *tgt = class_exp2tgt(exp);
936 struct sptlrpc_flavor flvr;
940 LASSERT(tgt->lut_obd);
941 LASSERT(tgt->lut_slice);
943 /* always allow ECHO client */
944 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
945 LUSTRE_ECHO_NAME) == 0)) {
946 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
950 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
951 read_lock(&tgt->lut_sptlrpc_lock);
952 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
954 lnet_nid_to_nid4(&req->rq_peer.nid),
956 read_unlock(&tgt->lut_sptlrpc_lock);
958 spin_lock(&exp->exp_lock);
959 exp->exp_sp_peer = req->rq_sp_from;
960 exp->exp_flvr = flvr;
962 /* when on mgs, if no restriction is set, or if the client
963 * NID is on the local node, allow any flavor
965 if ((strcmp(exp->exp_obd->obd_type->typ_name,
966 LUSTRE_MGS_NAME) == 0) &&
967 (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
968 LNetIsPeerLocal(&exp->exp_connection->c_peer.nid)))
969 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
971 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
972 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
973 CERROR("%s: unauthorized rpc flavor %x from %s, "
974 "expect %x\n", tgt_name(tgt),
976 libcfs_nidstr(&req->rq_peer.nid),
977 exp->exp_flvr.sf_rpc);
980 spin_unlock(&exp->exp_lock);
982 if (exp->exp_sp_peer != req->rq_sp_from) {
983 CERROR("%s: RPC source %s doesn't match %s\n",
985 sptlrpc_part2name(req->rq_sp_from),
986 sptlrpc_part2name(exp->exp_sp_peer));
989 rc = sptlrpc_target_export_check(exp, req);
996 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
998 struct sptlrpc_rule_set tmp_rset;
1001 if (unlikely(tgt == NULL)) {
1002 CERROR("No target passed\n");
1006 sptlrpc_rule_set_init(&tmp_rset);
1007 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
1009 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
1014 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
1016 write_lock(&tgt->lut_sptlrpc_lock);
1017 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
1018 tgt->lut_sptlrpc_rset = tmp_rset;
1019 write_unlock(&tgt->lut_sptlrpc_lock);
1023 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
1025 int tgt_connect(struct tgt_session_info *tsi)
1027 struct ptlrpc_request *req = tgt_ses_req(tsi);
1028 struct obd_connect_data *reply;
1033 /* XXX: better to call this check right after getting new export but
1034 * before last_rcvd slot allocation to avoid server load upon insecure
1035 * connects. This is to be fixed after unifiyng all targets.
1037 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
1041 /* To avoid exposing partially initialized connection flags, changes up
1042 * to this point have been staged in reply->ocd_connect_flags. Now that
1043 * connection handling has completed successfully, atomically update
1044 * the connect flags in the shared export data structure. LU-1623 */
1045 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1046 spin_lock(&tsi->tsi_exp->exp_lock);
1047 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1048 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1049 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1050 reply->ocd_connect_flags2;
1051 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1052 spin_unlock(&tsi->tsi_exp->exp_lock);
1054 if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1055 LUSTRE_MDT_NAME) == 0) {
1056 struct lu_nodemap *nm = NULL;
1058 rc = req_check_sepol(tsi->tsi_pill);
1062 if (tsi->tsi_pill->rc_req->rq_export)
1064 nodemap_get_from_exp(tsi->tsi_pill->rc_req->rq_export);
1066 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2 &&
1067 reply->ocd_connect_flags2 & OBD_CONNECT2_ENCRYPT) {
1068 bool forbid_encrypt = true;
1071 /* nodemap_get_from_exp returns NULL in case
1072 * nodemap is not active, so we do not forbid
1074 forbid_encrypt = false;
1075 else if (!IS_ERR(nm))
1076 forbid_encrypt = nm->nmf_forbid_encryption;
1078 GOTO(put_nm, rc = -EACCES);
1081 if (!(reply->ocd_connect_flags & OBD_CONNECT_RDONLY)) {
1082 bool readonly = false;
1084 if (!IS_ERR_OR_NULL(nm))
1085 readonly = nm->nmf_readonly_mount;
1086 if (unlikely(readonly))
1087 GOTO(put_nm, rc = -EROFS);
1091 if (!IS_ERR_OR_NULL(nm))
1099 obd_disconnect(class_export_get(tsi->tsi_exp));
1102 EXPORT_SYMBOL(tgt_connect);
1104 int tgt_disconnect(struct tgt_session_info *tsi)
1110 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1112 rc = target_handle_disconnect(tgt_ses_req(tsi));
1114 RETURN(err_serious(rc));
1118 EXPORT_SYMBOL(tgt_disconnect);
1121 * Unified target OBD handlers
1123 int tgt_obd_ping(struct tgt_session_info *tsi)
1129 /* The target-specific part of OBD_PING request handling.
1130 * It controls Filter Modification Data (FMD) expiration each time
1133 * Valid only for replayable targets, e.g. MDT and OFD
1135 if (tsi->tsi_exp->exp_obd->obd_replayable)
1136 tgt_fmd_expire(tsi->tsi_exp);
1138 rc = req_capsule_server_pack(tsi->tsi_pill);
1140 RETURN(err_serious(rc));
1142 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_CONNECT_VS_EVICT)) {
1143 if (strstr(tsi->tsi_exp->exp_obd->obd_name, "MDT0000") &&
1144 (exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_MDS_MDS))
1145 tsi->tsi_pill->rc_req->rq_no_reply = 1;
1150 EXPORT_SYMBOL(tgt_obd_ping);
1152 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1154 return err_serious(-EOPNOTSUPP);
1157 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1159 struct ptlrpc_request *req = tgt_ses_req(tsi);
1160 struct obd_export *exp = req->rq_export;
1161 struct ptlrpc_bulk_desc *desc;
1168 for (i = 0; i < rdbuf->rb_nbufs; i++) {
1169 unsigned int offset;
1171 offset = (unsigned long)rdbuf->rb_bufs[i].lb_buf & ~PAGE_MASK;
1172 pages += DIV_ROUND_UP(rdbuf->rb_bufs[i].lb_len + offset,
1176 desc = ptlrpc_prep_bulk_exp(req, pages, 1,
1177 PTLRPC_BULK_PUT_SOURCE,
1179 &ptlrpc_bulk_kiov_nopin_ops);
1183 for (i = 0; i < rdbuf->rb_nbufs; i++)
1184 desc->bd_frag_ops->add_iov_frag(desc,
1185 rdbuf->rb_bufs[i].lb_buf,
1186 rdbuf->rb_bufs[i].lb_len);
1188 rc = target_bulk_io(exp, desc);
1189 ptlrpc_free_bulk(desc);
1192 EXPORT_SYMBOL(tgt_send_buffer);
1194 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1196 struct ptlrpc_request *req = tgt_ses_req(tsi);
1197 struct obd_export *exp = req->rq_export;
1198 struct ptlrpc_bulk_desc *desc;
1206 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1207 PTLRPC_BULK_PUT_SOURCE,
1209 &ptlrpc_bulk_kiov_pin_ops);
1213 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1214 /* old client requires reply size in it's PAGE_SIZE,
1215 * which is rdpg->rp_count */
1216 nob = rdpg->rp_count;
1218 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1219 i++, tmpcount -= tmpsize) {
1220 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1221 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1225 LASSERT(desc->bd_nob == nob);
1226 rc = target_bulk_io(exp, desc);
1227 ptlrpc_free_bulk(desc);
1230 EXPORT_SYMBOL(tgt_sendpage);
1233 * OBD_IDX_READ handler
1235 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1237 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1238 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1239 struct idx_info *req_ii, *rep_ii;
1244 memset(rdpg, 0, sizeof(*rdpg));
1245 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1247 /* extract idx_info buffer from request & reply */
1248 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1249 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1250 RETURN(err_serious(-EPROTO));
1252 rc = req_capsule_server_pack(tsi->tsi_pill);
1254 RETURN(err_serious(rc));
1256 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1258 RETURN(err_serious(-EFAULT));
1259 rep_ii->ii_magic = IDX_INFO_MAGIC;
1261 /* extract hash to start with */
1262 rdpg->rp_hash = req_ii->ii_hash_start;
1264 /* extract requested attributes */
1265 rdpg->rp_attrs = req_ii->ii_attrs;
1267 /* check that fid packed in request is valid and supported */
1268 if (!fid_is_sane(&req_ii->ii_fid))
1270 rep_ii->ii_fid = req_ii->ii_fid;
1273 rep_ii->ii_flags = req_ii->ii_flags;
1275 /* compute number of pages to allocate, ii_count is the number of 4KB
1277 if (req_ii->ii_count <= 0)
1278 GOTO(out, rc = -EFAULT);
1279 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1280 exp_max_brw_size(tsi->tsi_exp));
1281 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1283 /* allocate pages to store the containers */
1284 OBD_ALLOC_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1285 if (rdpg->rp_pages == NULL)
1286 GOTO(out, rc = -ENOMEM);
1287 for (i = 0; i < rdpg->rp_npages; i++) {
1288 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1289 if (rdpg->rp_pages[i] == NULL)
1290 GOTO(out, rc = -ENOMEM);
1293 /* populate pages with key/record pairs */
1294 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1298 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1299 "asked %d > %d\n", rc, rdpg->rp_count);
1301 /* send pages to client */
1302 rc = tgt_sendpage(tsi, rdpg, rc);
1307 if (rdpg->rp_pages) {
1308 for (i = 0; i < rdpg->rp_npages; i++)
1309 if (rdpg->rp_pages[i])
1310 __free_page(rdpg->rp_pages[i]);
1311 OBD_FREE_PTR_ARRAY(rdpg->rp_pages, rdpg->rp_npages);
1316 struct tgt_handler tgt_obd_handlers[] = {
1317 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1318 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1320 EXPORT_SYMBOL(tgt_obd_handlers);
1322 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1323 struct dt_object *obj, __u64 start, __u64 end)
1329 /* if no objid is specified, it means "sync whole filesystem" */
1331 rc = dt_sync(env, tgt->lut_bottom);
1332 } else if (dt_version_get(env, obj) >
1333 tgt->lut_obd->obd_last_committed) {
1334 rc = dt_object_sync(env, obj, start, end);
1336 atomic_inc(&tgt->lut_sync_count);
1340 EXPORT_SYMBOL(tgt_sync);
1342 * Unified target DLM handlers.
1346 * Unified target BAST
1348 * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1349 * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1350 * but if it's cross-MDT lock, because remote object version is not set, a
1351 * filesystem sync is needed.
1353 * \param lock server side lock
1354 * \param desc lock desc
1355 * \param data ldlm_cb_set_arg
1356 * \param flag indicates whether this cancelling or blocking callback
1357 * \retval 0 on success
1358 * \retval negative number on error
1360 int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1361 void *data, int flag)
1364 struct lu_target *tgt;
1365 struct dt_object *obj = NULL;
1371 tgt = class_exp2tgt(lock->l_export);
1373 if (unlikely(tgt == NULL)) {
1374 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1375 class_exp2obd(lock->l_export)->obd_name);
1379 if (flag == LDLM_CB_CANCELING &&
1380 (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1381 (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_ALWAYS ||
1382 (tgt->lut_sync_lock_cancel == SYNC_LOCK_CANCEL_BLOCKING &&
1383 ldlm_is_cbpending(lock))) &&
1384 ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1385 lock->l_resource->lr_type == LDLM_EXTENT)) {
1387 __u64 end = OBD_OBJECT_EOF;
1389 rc = lu_env_init(&env, LCT_DT_THREAD);
1390 if (unlikely(rc != 0))
1393 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1394 tgt->lut_lsd.lsd_osd_index);
1396 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1397 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1399 GOTO(err_env, rc = PTR_ERR(obj));
1401 if (!dt_object_exists(obj))
1402 GOTO(err_put, rc = -ENOENT);
1404 start = lock->l_policy_data.l_extent.start;
1405 end = lock->l_policy_data.l_extent.end;
1408 rc = tgt_sync(&env, tgt, obj, start, end);
1410 CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1411 "cancel: rc = %d\n",
1412 tgt_name(tgt), PFID(&fid),
1413 lock->l_policy_data.l_extent.start,
1414 lock->l_policy_data.l_extent.end, rc);
1418 dt_object_put(&env, obj);
1423 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1426 EXPORT_SYMBOL(tgt_blocking_ast);
1428 static struct ldlm_callback_suite tgt_dlm_cbs = {
1429 .lcs_completion = ldlm_server_completion_ast,
1430 .lcs_blocking = tgt_blocking_ast,
1431 .lcs_glimpse = ldlm_server_glimpse_ast
1434 int tgt_enqueue(struct tgt_session_info *tsi)
1436 struct ptlrpc_request *req = tgt_ses_req(tsi);
1441 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1442 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1444 LASSERT(tsi->tsi_dlm_req != NULL);
1445 rc = ldlm_handle_enqueue(tsi->tsi_exp->exp_obd->obd_namespace,
1446 &req->rq_pill, tsi->tsi_dlm_req, &tgt_dlm_cbs);
1448 RETURN(err_serious(rc));
1450 switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1452 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1455 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1458 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1461 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1464 RETURN(req->rq_status);
1466 EXPORT_SYMBOL(tgt_enqueue);
1468 int tgt_convert(struct tgt_session_info *tsi)
1470 struct ptlrpc_request *req = tgt_ses_req(tsi);
1474 LASSERT(tsi->tsi_dlm_req);
1475 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1477 RETURN(err_serious(rc));
1479 RETURN(req->rq_status);
1482 int tgt_bl_callback(struct tgt_session_info *tsi)
1484 return err_serious(-EOPNOTSUPP);
1487 int tgt_cp_callback(struct tgt_session_info *tsi)
1489 return err_serious(-EOPNOTSUPP);
1492 /* generic LDLM target handler */
1493 struct tgt_handler tgt_dlm_handlers[] = {
1494 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1495 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1496 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1497 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1499 EXPORT_SYMBOL(tgt_dlm_handlers);
1502 * Unified target LLOG handlers.
1504 int tgt_llog_open(struct tgt_session_info *tsi)
1510 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1514 EXPORT_SYMBOL(tgt_llog_open);
1516 int tgt_llog_read_header(struct tgt_session_info *tsi)
1522 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1526 EXPORT_SYMBOL(tgt_llog_read_header);
1528 int tgt_llog_next_block(struct tgt_session_info *tsi)
1534 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1538 EXPORT_SYMBOL(tgt_llog_next_block);
1540 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1546 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1550 EXPORT_SYMBOL(tgt_llog_prev_block);
1552 /* generic llog target handler */
1553 struct tgt_handler tgt_llog_handlers[] = {
1554 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1555 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1556 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1557 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1559 EXPORT_SYMBOL(tgt_llog_handlers);
1562 * sec context handlers
1564 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1565 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1570 struct tgt_handler tgt_sec_ctx_handlers[] = {
1571 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle),
1572 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle),
1573 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle),
1575 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1577 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1578 struct dt_device *key,
1579 struct lfsck_req_local *lrl,
1580 struct thandle *th) = NULL;
1582 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1584 struct lfsck_req_local *,
1587 tgt_lfsck_in_notify_local = notify;
1589 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1591 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1592 struct dt_device *key,
1593 struct lfsck_request *lr) = NULL;
1595 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1597 struct lfsck_request *))
1599 tgt_lfsck_in_notify = notify;
1601 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1603 static int (*tgt_lfsck_query)(const struct lu_env *env,
1604 struct dt_device *key,
1605 struct lfsck_request *req,
1606 struct lfsck_reply *rep,
1607 struct lfsck_query *que) = NULL;
1609 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1611 struct lfsck_request *,
1612 struct lfsck_reply *,
1613 struct lfsck_query *))
1615 tgt_lfsck_query = query;
1617 EXPORT_SYMBOL(tgt_register_lfsck_query);
1619 /* LFSCK request handlers */
1620 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1622 const struct lu_env *env = tsi->tsi_env;
1623 struct dt_device *key = tsi->tsi_tgt->lut_bottom;
1624 struct lfsck_request *lr;
1628 lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1632 rc = tgt_lfsck_in_notify(env, key, lr);
1637 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1639 struct lfsck_request *request;
1640 struct lfsck_reply *reply;
1644 request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1645 if (request == NULL)
1648 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1652 rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1653 request, reply, NULL);
1655 RETURN(rc < 0 ? rc : 0);
1658 struct tgt_handler tgt_lfsck_handlers[] = {
1659 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_NOTIFY, tgt_handle_lfsck_notify),
1660 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_QUERY, tgt_handle_lfsck_query),
1662 EXPORT_SYMBOL(tgt_lfsck_handlers);
1665 * initialize per-thread page pool (bug 5137).
1667 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1669 struct tgt_thread_big_cache *tbc;
1673 LASSERT(thread != NULL);
1674 LASSERT(thread->t_data == NULL);
1676 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1679 thread->t_data = tbc;
1682 EXPORT_SYMBOL(tgt_io_thread_init);
1685 * free per-thread pool created by tgt_thread_init().
1687 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1689 struct tgt_thread_big_cache *tbc;
1693 LASSERT(thread != NULL);
1696 * be prepared to handle partially-initialized pools (because this is
1697 * called from ost_io_thread_init() for cleanup.
1699 tbc = thread->t_data;
1701 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1702 thread->t_data = NULL;
1706 EXPORT_SYMBOL(tgt_io_thread_done);
1709 * Helper function for getting Data-on-MDT file server DLM lock
1710 * if asked by client.
1712 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1713 struct lustre_handle *lh, int mode, __u64 *flags)
1715 union ldlm_policy_data policy = {
1716 .l_inodebits.bits = MDS_INODELOCK_DOM,
1722 LASSERT(lh != NULL);
1723 LASSERT(ns != NULL);
1724 LASSERT(!lustre_handle_is_used(lh));
1726 rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1727 flags, ldlm_blocking_ast,
1728 ldlm_completion_ast, ldlm_glimpse_ast,
1729 NULL, 0, LVB_T_NONE, NULL, lh);
1731 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1733 EXPORT_SYMBOL(tgt_mdt_data_lock);
1736 * Helper function for getting server side [start, start+count] DLM lock
1737 * if asked by client.
1739 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1740 struct ldlm_res_id *res_id, __u64 start, __u64 end,
1741 struct lustre_handle *lh, int mode, __u64 *flags)
1743 union ldlm_policy_data policy;
1748 LASSERT(lh != NULL);
1749 LASSERT(ns != NULL);
1750 LASSERT(!lustre_handle_is_used(lh));
1752 policy.l_extent.gid = 0;
1753 policy.l_extent.start = start & PAGE_MASK;
1756 * If ->o_blocks is EOF it means "lock till the end of the file".
1757 * Otherwise, it's size of an extent or hole being punched (in bytes).
1759 if (end == OBD_OBJECT_EOF || end < start)
1760 policy.l_extent.end = OBD_OBJECT_EOF;
1762 policy.l_extent.end = end | ~PAGE_MASK;
1764 rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1765 mode, flags, ldlm_blocking_ast,
1766 ldlm_completion_ast, ldlm_glimpse_ast,
1767 NULL, 0, LVB_T_NONE, NULL, lh);
1768 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1770 EXPORT_SYMBOL(tgt_extent_lock);
1772 static int tgt_data_lock(const struct lu_env *env, struct obd_export *exp,
1773 struct ldlm_res_id *res_id, __u64 start, __u64 end,
1774 struct lustre_handle *lh, enum ldlm_mode mode)
1776 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1779 /* MDT IO for data-on-mdt */
1780 if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1781 return tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1783 return tgt_extent_lock(env, ns, res_id, start, end, lh, mode, &flags);
1786 void tgt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1788 LASSERT(lustre_handle_is_used(lh));
1789 ldlm_lock_decref(lh, mode);
1791 EXPORT_SYMBOL(tgt_data_unlock);
1793 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1794 struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1795 struct niobuf_remote *nb, struct lustre_handle *lh,
1796 enum ldlm_mode mode)
1798 int nrbufs = obj->ioo_bufcnt;
1803 LASSERT(mode == LCK_PR || mode == LCK_PW);
1804 LASSERT(!lustre_handle_is_used(lh));
1806 if (exp->exp_obd->obd_recovering)
1809 if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1812 for (i = 1; i < nrbufs; i++)
1813 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1816 return tgt_data_lock(env, exp, res_id, nb[0].rnb_offset,
1817 nb[nrbufs - 1].rnb_offset +
1818 nb[nrbufs - 1].rnb_len - 1, lh, mode);
1821 static void tgt_brw_unlock(struct obd_export *exp, struct obd_ioobj *obj,
1822 struct niobuf_remote *niob,
1823 struct lustre_handle *lh, enum ldlm_mode mode)
1827 LASSERT(mode == LCK_PR || mode == LCK_PW);
1828 LASSERT((!exp->exp_obd->obd_recovering && obj->ioo_bufcnt &&
1829 niob[0].rnb_flags & OBD_BRW_SRVLOCK) ==
1830 lustre_handle_is_used(lh));
1832 if (lustre_handle_is_used(lh))
1833 tgt_data_unlock(lh, mode);
1837 static int tgt_checksum_niobuf(struct lu_target *tgt,
1838 struct niobuf_local *local_nb, int npages,
1839 int opc, enum cksum_types cksum_type,
1842 struct ahash_request *req;
1843 unsigned int bufsize;
1845 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1847 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1849 CERROR("%s: unable to initialize checksum hash %s\n",
1850 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1851 return PTR_ERR(req);
1854 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1855 for (i = 0; i < npages; i++) {
1856 /* corrupt the data before we compute the checksum, to
1857 * simulate a client->OST data error */
1858 if (i == 0 && opc == OST_WRITE &&
1859 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1860 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1861 int len = local_nb[i].lnb_len;
1862 struct page *np = tgt_page_to_corrupt;
1865 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1866 char *ptr2 = page_address(np);
1868 memcpy(ptr2 + off, ptr + off, len);
1869 memcpy(ptr2 + off, "bad3", min(4, len));
1872 /* LU-8376 to preserve original index for
1873 * display in dump_all_bulk_pages() */
1876 cfs_crypto_hash_update_page(req, np, off,
1880 CERROR("%s: can't alloc page for corruption\n",
1884 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1885 local_nb[i].lnb_page_offset & ~PAGE_MASK,
1886 local_nb[i].lnb_len);
1888 /* corrupt the data after we compute the checksum, to
1889 * simulate an OST->client data error */
1890 if (i == 0 && opc == OST_READ &&
1891 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1892 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1893 int len = local_nb[i].lnb_len;
1894 struct page *np = tgt_page_to_corrupt;
1897 char *ptr = kmap_atomic(local_nb[i].lnb_page);
1898 char *ptr2 = page_address(np);
1900 memcpy(ptr2 + off, ptr + off, len);
1901 memcpy(ptr2 + off, "bad4", min(4, len));
1904 /* LU-8376 to preserve original index for
1905 * display in dump_all_bulk_pages() */
1908 cfs_crypto_hash_update_page(req, np, off,
1912 CERROR("%s: can't alloc page for corruption\n",
1918 bufsize = sizeof(*cksum);
1919 err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1924 char dbgcksum_file_name[PATH_MAX];
1926 static void dump_all_bulk_pages(struct obdo *oa, int count,
1927 struct niobuf_local *local_nb,
1928 __u32 server_cksum, __u32 client_cksum)
1935 /* will only keep dump of pages on first error for the same range in
1936 * file/fid, not during the resends/retries. */
1937 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1938 "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1939 (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1940 libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1941 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1942 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1943 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1944 local_nb[0].lnb_file_offset,
1945 local_nb[count-1].lnb_file_offset +
1946 local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1947 CWARN("dumping checksum data to %s\n", dbgcksum_file_name);
1948 filp = filp_open(dbgcksum_file_name,
1949 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1953 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1954 "checksum error: rc = %d\n", dbgcksum_file_name,
1957 CERROR("%s: can't open to dump pages with checksum "
1958 "error: rc = %d\n", dbgcksum_file_name, rc);
1962 for (i = 0; i < count; i++) {
1963 len = local_nb[i].lnb_len;
1964 buf = kmap(local_nb[i].lnb_page);
1966 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1968 CERROR("%s: wanted to write %u but got %d "
1969 "error\n", dbgcksum_file_name, len, rc);
1975 kunmap(local_nb[i].lnb_page);
1978 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1980 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1981 filp_close(filp, NULL);
1983 libcfs_debug_dumplog();
1986 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1987 struct obd_export *exp, struct obdo *oa,
1988 const struct lnet_processid *peer,
1989 __u32 client_cksum, __u32 server_cksum,
1990 enum cksum_types server_cksum_type)
1993 enum cksum_types cksum_type;
1996 if (unlikely(npages <= 0))
1999 /* unlikely to happen and only if resend does not occur due to cksum
2000 * control failure on Client */
2001 if (unlikely(server_cksum == client_cksum)) {
2002 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
2007 if (exp->exp_obd->obd_checksum_dump)
2008 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
2011 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
2014 if (cksum_type != server_cksum_type)
2015 msg = "the server may have not used the checksum type specified"
2016 " in the original request - likely a protocol problem";
2018 msg = "should have changed on the client or in transit";
2020 start = local_nb[0].lnb_file_offset;
2021 end = local_nb[npages-1].lnb_file_offset +
2022 local_nb[npages-1].lnb_len - 1;
2024 LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
2025 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
2026 " %x (type %x), server csum %x (type %x)\n",
2027 exp->exp_obd->obd_name,
2028 msg, libcfs_nidstr(&peer->nid),
2029 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
2030 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
2031 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
2033 start, end, client_cksum, cksum_type, server_cksum,
2039 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
2040 unsigned char *buf, int size)
2042 int i, off, len, copied = size;
2045 for (i = 0; i < npages; i++) {
2046 off = local[i].lnb_page_offset & ~PAGE_MASK;
2047 len = local[i].lnb_len;
2049 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2054 ptr = kmap_atomic(local[i].lnb_page);
2055 memcpy(buf, ptr + off, len);
2060 return copied - size;
2063 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
2064 enum cksum_types cksum_type,
2065 struct niobuf_local *local_nb, int npages,
2066 int opc, obd_dif_csum_fn *fn,
2067 int sector_size, u32 *check_sum,
2070 enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
2071 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
2072 const char *obd_name = tgt->lut_obd->obd_name;
2073 struct ahash_request *req;
2074 unsigned char *buffer;
2075 struct page *__page;
2076 __be16 *guard_start;
2078 int used_number = 0;
2080 unsigned int bufsize = sizeof(cksum);
2085 __page = alloc_page(GFP_KERNEL);
2089 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
2092 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
2093 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg), rc);
2097 buffer = kmap(__page);
2098 guard_start = (__be16 *)buffer;
2099 guard_number = PAGE_SIZE / sizeof(*guard_start);
2100 if (unlikely(resend))
2101 CDEBUG(D_PAGE | D_HA, "GRD tags per page = %u\n", guard_number);
2102 for (i = 0; i < npages; i++) {
2105 /* corrupt the data before we compute the checksum, to
2106 * simulate a client->OST data error */
2107 if (i == 0 && opc == OST_WRITE &&
2108 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2109 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2110 int len = local_nb[i].lnb_len;
2111 struct page *np = tgt_page_to_corrupt;
2114 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2115 char *ptr2 = page_address(np);
2117 memcpy(ptr2 + off, ptr + off, len);
2118 memcpy(ptr2 + off, "bad3", min(4, len));
2121 /* LU-8376 to preserve original index for
2122 * display in dump_all_bulk_pages() */
2125 cfs_crypto_hash_update_page(req, np, off,
2129 CERROR("%s: can't alloc page for corruption\n",
2135 * The left guard number should be able to hold checksums of a
2138 use_t10_grd = t10_cksum_type && t10_cksum_type == cksum_type &&
2140 local_nb[i].lnb_len == PAGE_SIZE &&
2141 local_nb[i].lnb_guard_disk;
2143 used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2144 if (used > (guard_number - used_number)) {
2146 CDEBUG(D_PAGE | D_HA,
2147 "%s: used %u, guard %u/%u, data size %u+%u, sector_size %u: rc = %d\n",
2148 obd_name, used, guard_number,
2149 used_number, local_nb[i].lnb_page_offset,
2150 local_nb[i].lnb_len, sector_size, rc);
2153 memcpy(guard_start + used_number,
2154 local_nb[i].lnb_guards,
2155 used * sizeof(*guard_start));
2156 if (unlikely(resend))
2157 CDEBUG(D_PAGE | D_HA,
2158 "lnb[%u]: used %u off %u+%u lnb checksum: %*phN\n",
2160 local_nb[i].lnb_page_offset,
2161 local_nb[i].lnb_len,
2162 (int)(used * sizeof(*guard_start)),
2163 guard_start + used_number);
2165 if (!use_t10_grd || unlikely(resend)) {
2166 __be16 guard_tmp[MAX_GUARD_NUMBER];
2167 __be16 *guards = guard_start + used_number;
2168 int used_tmp = -1, *usedp = &used;
2170 if (unlikely(use_t10_grd)) {
2174 rc = obd_page_dif_generate_buffer(obd_name,
2175 local_nb[i].lnb_page,
2176 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2177 local_nb[i].lnb_len, guards,
2178 guard_number - used_number, usedp, sector_size,
2180 if (unlikely(resend)) {
2181 bool bad = use_t10_grd &&
2183 local_nb[i].lnb_guards,
2184 used_tmp * sizeof(*guard_tmp));
2187 CERROR("lnb[%u]: used %u/%u off %u+%u tmp checksum: %*phN\n",
2189 local_nb[i].lnb_page_offset,
2190 local_nb[i].lnb_len,
2191 (int)(used_tmp * sizeof(*guard_start)),
2193 CDEBUG_LIMIT(D_PAGE | D_HA | (bad ? D_ERROR : 0),
2194 "lnb[%u]: used %u/%u off %u+%u gen checksum: %*phN\n",
2196 local_nb[i].lnb_page_offset,
2197 local_nb[i].lnb_len,
2198 (int)(used * sizeof(*guard_start)),
2199 guard_start + used_number);
2205 LASSERT(used <= MAX_GUARD_NUMBER);
2207 * If disk support T10PI checksum, copy guards to local_nb.
2208 * If the write is partial page, do not use the guards for bio
2209 * submission since the data might not be full-sector. The bio
2210 * guards will be generated later based on the full sectors. If
2211 * the sector size is 512B rather than 4 KB, or the page size
2212 * is larger than 4KB, this might drop some useful guards for
2213 * partial page write, but it will only add minimal extra time
2214 * of checksum calculation.
2216 if (t10_cksum_type && t10_cksum_type == cksum_type &&
2218 local_nb[i].lnb_len == PAGE_SIZE) {
2219 local_nb[i].lnb_guard_rpc = 1;
2220 memcpy(local_nb[i].lnb_guards,
2221 guard_start + used_number,
2222 used * sizeof(*local_nb[i].lnb_guards));
2225 used_number += used;
2226 if (used_number == guard_number) {
2227 cfs_crypto_hash_update_page(req, __page, 0,
2228 used_number * sizeof(*guard_start));
2232 /* corrupt the data after we compute the checksum, to
2233 * simulate an OST->client data error */
2234 if (unlikely(i == 0 && opc == OST_READ &&
2235 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2236 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2237 int len = local_nb[i].lnb_len;
2238 struct page *np = tgt_page_to_corrupt;
2241 char *ptr = kmap_atomic(local_nb[i].lnb_page);
2242 char *ptr2 = page_address(np);
2244 memcpy(ptr2 + off, ptr + off, len);
2245 memcpy(ptr2 + off, "bad4", min(4, len));
2248 /* LU-8376 to preserve original index for
2249 * display in dump_all_bulk_pages() */
2252 cfs_crypto_hash_update_page(req, np, off,
2256 CERROR("%s: can't alloc page for corruption\n",
2265 if (used_number != 0)
2266 cfs_crypto_hash_update_page(req, __page, 0,
2267 used_number * sizeof(*guard_start));
2269 rc2 = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2275 __free_page(__page);
2279 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2280 enum cksum_types cksum_type,
2281 struct niobuf_local *local_nb,
2282 int npages, int opc, u32 *check_sum,
2285 obd_dif_csum_fn *fn = NULL;
2286 int sector_size = 0;
2290 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
2293 rc = tgt_checksum_niobuf_t10pi(tgt, cksum_type,
2295 opc, fn, sector_size,
2298 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2299 cksum_type, check_sum);
2304 int tgt_brw_read(struct tgt_session_info *tsi)
2306 struct ptlrpc_request *req = tgt_ses_req(tsi);
2307 struct ptlrpc_bulk_desc *desc = NULL;
2308 struct obd_export *exp = tsi->tsi_exp;
2309 struct niobuf_remote *remote_nb;
2310 struct niobuf_local *local_nb;
2311 struct obd_ioobj *ioo;
2312 struct ost_body *body, *repbody;
2313 struct lustre_handle lockh = { 0 };
2314 int npages, nob = 0, rc, i, no_reply = 0,
2316 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2317 const char *obd_name = exp->exp_obd->obd_name;
2322 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2323 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2324 CERROR("%s: deny read request from %s to portal %u\n",
2325 tgt_name(tsi->tsi_tgt),
2326 obd_export_nid2str(req->rq_export),
2327 ptlrpc_req2svc(req)->srv_req_portal);
2331 req->rq_bulk_read = 1;
2333 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK)) {
2334 /* optionally use cfs_fail_val - 1 to select a specific OST on
2335 * this server to fail requests.
2337 char fail_ost_name[MAX_OBD_NAME];
2339 if (cfs_fail_val > 0) {
2340 snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2343 if (strstr(obd_name, fail_ost_name))
2344 RETURN(err_serious(-EIO));
2346 RETURN(err_serious(-EIO));
2350 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2351 cfs_fail_val : (obd_timeout + 1) / 4);
2353 /* Check if there is eviction in progress, and if so, wait for it to
2355 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2356 /* We do not care how long it takes */
2358 exp->exp_obd->obd_evict_inprogress_waitq,
2359 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
2362 /* There must be big cache in current thread to process this request
2363 * if it is NULL then something went wrong and it wasn't allocated,
2364 * report -ENOMEM in that case */
2368 body = tsi->tsi_ost_body;
2369 LASSERT(body != NULL);
2371 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2372 body->oa.o_flags & OBD_FL_NORPC)
2375 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2376 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2378 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2379 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2381 local_nb = tbc->local;
2383 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2389 * If getting the lock took more time than
2390 * client was willing to wait, drop it. b=11330
2392 if (ktime_get_real_seconds() > req->rq_deadline ||
2393 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2395 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2396 libcfs_idstr(&req->rq_peer), POSTID(&ioo->ioo_oid),
2397 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2398 req->rq_deadline - req->rq_arrival_time.tv_sec);
2399 GOTO(out_lock, rc = -ETIMEDOUT);
2403 * Because we already sync grant info with client when
2404 * reconnect, grant info will be cleared for resent req,
2405 * otherwise, outdated grant count in the rpc would de-sync
2406 * grant counters in case of shrink
2408 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2409 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2410 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2413 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2414 repbody->oa = body->oa;
2416 npages = PTLRPC_MAX_BRW_PAGES;
2417 kstart = ktime_get();
2418 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2419 ioo, remote_nb, &npages, local_nb);
2423 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2424 body->oa.o_flags & OBD_FL_SHORT_IO) {
2427 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2428 PTLRPC_BULK_PUT_SOURCE,
2430 &ptlrpc_bulk_kiov_nopin_ops);
2432 GOTO(out_commitrw, rc = -ENOMEM);
2435 npages_read = npages;
2436 for (i = 0; i < npages; i++) {
2437 int page_rc = local_nb[i].lnb_rc;
2446 if (page_rc != 0 && desc != NULL) { /* some data! */
2447 LASSERT(local_nb[i].lnb_page != NULL);
2448 desc->bd_frag_ops->add_kiov_frag
2449 (desc, local_nb[i].lnb_page,
2450 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2454 if (page_rc != local_nb[i].lnb_len) { /* short read */
2455 local_nb[i].lnb_len = page_rc;
2456 npages_read = i + (page_rc != 0 ? 1 : 0);
2457 /* All subsequent pages should be 0 */
2458 while (++i < npages)
2459 LASSERT(local_nb[i].lnb_rc == 0);
2464 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2465 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2466 body->oa.o_flags : 0;
2467 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2468 bool resend = (body->oa.o_valid & OBD_MD_FLFLAGS) &&
2469 (body->oa.o_flags & OBD_FL_RECOV_RESEND);
2471 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2473 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2475 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2476 local_nb, npages_read, OST_READ,
2477 &repbody->oa.o_cksum, resend);
2479 GOTO(out_commitrw, rc);
2480 CDEBUG(D_PAGE | (resend ? D_HA : 0),
2481 "checksum at read origin: %x (%x)\n",
2482 repbody->oa.o_cksum, cksum_type);
2484 /* if a resend it could be for a cksum error, so check Server
2485 * cksum with returned Client cksum (this should even cover
2486 * zero-cksum case) */
2488 check_read_checksum(local_nb, npages_read, exp,
2489 &body->oa, &req->rq_peer,
2491 repbody->oa.o_cksum, cksum_type);
2493 repbody->oa.o_valid = 0;
2495 if (body->oa.o_valid & OBD_MD_FLGRANT)
2496 repbody->oa.o_valid |= OBD_MD_FLGRANT;
2497 /* We're finishing using body->oa as an input variable */
2499 /* Check if client was evicted while we were doing i/o before touching
2502 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2503 body->oa.o_flags & OBD_FL_SHORT_IO) {
2504 unsigned char *short_io_buf;
2507 short_io_buf = req_capsule_server_get(&req->rq_pill,
2509 short_io_size = req_capsule_get_size(&req->rq_pill,
2512 rc = tgt_pages2shortio(local_nb, npages_read,
2513 short_io_buf, short_io_size);
2515 req_capsule_shrink(&req->rq_pill,
2518 rc = rc > 0 ? 0 : rc;
2519 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2520 rc = target_bulk_io(exp, desc);
2524 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2525 body->oa.o_flags & OBD_FL_SHORT_IO)
2526 req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2531 /* Must commit after prep above in all cases */
2532 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2533 remote_nb, npages, local_nb, rc, nob, kstart);
2535 tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PR);
2537 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2538 ptlrpc_free_bulk(desc);
2543 ptlrpc_lprocfs_brw(req, nob);
2544 } else if (no_reply) {
2545 req->rq_no_reply = 1;
2546 /* reply out callback would free */
2547 ptlrpc_req_drop_rs(req);
2548 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2549 "client will retry: rc %d\n",
2551 obd_uuid2str(&exp->exp_client_uuid),
2552 obd_export_nid2str(exp), rc);
2554 /* send a bulk after reply to simulate a network delay or reordering
2555 * by a router - Note that !desc implies short io, so there is no bulk
2557 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2559 /* Calculate checksum before request transfer, original
2560 * it is done by target_bulk_io() */
2561 rc = sptlrpc_svc_wrap_bulk(req, desc);
2562 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
2563 req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
2564 else /* old version, bulk matchbits is rq_xid */
2565 req->rq_mbits = req->rq_xid;
2567 req->rq_status = rc;
2568 target_committed_to_req(req);
2569 target_send_reply(req, 0, 0);
2571 CDEBUG(D_INFO, "reorder BULK\n");
2572 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,
2573 cfs_fail_val ? : 3);
2575 target_bulk_io(exp, desc);
2576 ptlrpc_free_bulk(desc);
2581 EXPORT_SYMBOL(tgt_brw_read);
2583 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2584 unsigned char *buf, unsigned int size)
2589 for (i = 0; i < npages; i++) {
2590 off = local[i].lnb_page_offset & ~PAGE_MASK;
2591 len = local[i].lnb_len;
2596 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2598 ptr = kmap_atomic(local[i].lnb_page);
2601 memcpy(ptr + off, buf, len < size ? len : size);
2609 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2610 struct ptlrpc_bulk_desc *desc,
2611 struct niobuf_local *local_nb, int npages,
2612 u32 client_cksum, u32 server_cksum,
2615 struct obd_export *exp = req->rq_export;
2616 struct ost_body *body;
2620 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2621 LASSERT(body != NULL);
2623 if (desc && !nid_same(&req->rq_peer.nid, &desc->bd_sender)) {
2625 router = libcfs_nidstr(&desc->bd_sender);
2628 if (exp->exp_obd->obd_checksum_dump)
2629 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2633 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2634 client_cksum, server_cksum);
2638 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2639 DFID" object "DOSTID" extent [%llu-%llu"
2640 "]: client csum %x, server csum %x\n",
2641 exp->exp_obd->obd_name, libcfs_idstr(&req->rq_peer),
2643 body->oa.o_valid & OBD_MD_FLFID ?
2644 body->oa.o_parent_seq : (__u64)0,
2645 body->oa.o_valid & OBD_MD_FLFID ?
2646 body->oa.o_parent_oid : 0,
2647 body->oa.o_valid & OBD_MD_FLFID ?
2648 body->oa.o_parent_ver : 0,
2649 POSTID(&body->oa.o_oi),
2650 local_nb[0].lnb_file_offset,
2651 local_nb[npages-1].lnb_file_offset +
2652 local_nb[npages - 1].lnb_len - 1,
2653 client_cksum, server_cksum);
2656 int tgt_brw_write(struct tgt_session_info *tsi)
2658 struct ptlrpc_request *req = tgt_ses_req(tsi);
2659 struct ptlrpc_bulk_desc *desc = NULL;
2660 struct obd_export *exp = req->rq_export;
2661 struct niobuf_remote *remote_nb;
2662 struct niobuf_local *local_nb;
2663 struct obd_ioobj *ioo;
2664 struct ost_body *body, *repbody;
2665 struct lustre_handle lockh = {0};
2667 int objcount, niocount, npages;
2670 enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2671 bool no_reply = false, mmap;
2672 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2673 bool wait_sync = false;
2674 const char *obd_name = exp->exp_obd->obd_name;
2675 /* '1' for consistency with code that checks !mpflag to restore */
2676 unsigned int mpflags = 1;
2682 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2683 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2684 CERROR("%s: deny write request from %s to portal %u\n",
2685 tgt_name(tsi->tsi_tgt),
2686 obd_export_nid2str(req->rq_export),
2687 ptlrpc_req2svc(req)->srv_req_portal);
2688 RETURN(err_serious(-EPROTO));
2691 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2692 RETURN(err_serious(-ENOSPC));
2693 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2694 RETURN(err_serious(-EROFS));
2696 req->rq_bulk_write = 1;
2698 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2700 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2703 /* optionally use cfs_fail_val - 1 to select a specific OST on
2704 * this server to fail requests.
2706 char fail_ost_name[MAX_OBD_NAME];
2708 if (cfs_fail_val > 0) {
2709 snprintf(fail_ost_name, MAX_OBD_NAME, "OST%04X",
2712 if (strstr(obd_name, fail_ost_name))
2713 RETURN(err_serious(rc));
2715 RETURN(err_serious(rc));
2719 /* pause before transaction has been started */
2720 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2721 cfs_fail_val : (obd_timeout + 1) / 4);
2723 /* Delay write commit to show stale size information */
2724 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2726 /* There must be big cache in current thread to process this request
2727 * if it is NULL then something went wrong and it wasn't allocated,
2728 * report -ENOMEM in that case */
2732 body = tsi->tsi_ost_body;
2733 LASSERT(body != NULL);
2735 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2736 body->oa.o_flags & OBD_FL_NORPC)
2740 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2741 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2743 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2744 RCL_CLIENT) / sizeof(*ioo);
2746 for (niocount = i = 0; i < objcount; i++)
2747 niocount += ioo[i].ioo_bufcnt;
2749 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2750 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2751 if (niocount != req_capsule_get_size(&req->rq_pill,
2752 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2754 RETURN(err_serious(-EPROTO));
2756 if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2757 ptlrpc_connection_is_local(exp->exp_connection))
2758 mpflags = memalloc_noreclaim_save();
2760 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2761 niocount * sizeof(*rcs));
2762 rc = req_capsule_server_pack(&req->rq_pill);
2764 GOTO(out, rc = err_serious(rc));
2766 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2767 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2769 local_nb = tbc->local;
2771 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2777 * If getting the lock took more time than
2778 * client was willing to wait, drop it. b=11330
2780 if (ktime_get_real_seconds() > req->rq_deadline ||
2781 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2783 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2784 tgt_name(tsi->tsi_tgt), libcfs_idstr(&req->rq_peer),
2785 POSTID(&ioo->ioo_oid),
2786 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2787 req->rq_deadline - req->rq_arrival_time.tv_sec);
2788 GOTO(out_lock, rc = -ETIMEDOUT);
2791 /* Because we already sync grant info with client when reconnect,
2792 * grant info will be cleared for resent req, then fed_grant and
2793 * total_grant will not be modified in following preprw_write */
2794 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2795 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2796 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2799 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2800 if (repbody == NULL)
2801 GOTO(out_lock, rc = -ENOMEM);
2802 repbody->oa = body->oa;
2804 npages = PTLRPC_MAX_BRW_PAGES;
2805 kstart = ktime_get();
2806 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2807 objcount, ioo, remote_nb, &npages, local_nb);
2810 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2811 body->oa.o_flags & OBD_FL_SHORT_IO) {
2812 unsigned int short_io_size;
2813 unsigned char *short_io_buf;
2815 short_io_size = req_capsule_get_size(&req->rq_pill,
2818 short_io_buf = req_capsule_client_get(&req->rq_pill,
2820 CDEBUG(D_INFO, "Client use short io for data transfer,"
2821 " size = %d\n", short_io_size);
2823 /* Copy short io buf to pages */
2824 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2828 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2829 PTLRPC_BULK_GET_SINK,
2831 &ptlrpc_bulk_kiov_nopin_ops);
2833 GOTO(skip_transfer, rc = -ENOMEM);
2835 /* NB Having prepped, we must commit... */
2836 for (i = 0; i < npages; i++)
2837 desc->bd_frag_ops->add_kiov_frag(desc,
2838 local_nb[i].lnb_page,
2839 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2840 local_nb[i].lnb_len);
2842 rc = sptlrpc_svc_prep_bulk(req, desc);
2844 GOTO(skip_transfer, rc);
2846 rc = target_bulk_io(exp, desc);
2852 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2853 static int cksum_counter;
2855 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2856 cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2858 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2859 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2860 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2863 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2864 local_nb, npages, OST_WRITE,
2865 &repbody->oa.o_cksum, false);
2867 GOTO(out_commitrw, rc);
2871 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2872 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2873 body->oa.o_flags & OBD_FL_MMAP);
2875 tgt_warn_on_cksum(req, desc, local_nb, npages,
2877 repbody->oa.o_cksum, mmap);
2879 } else if ((cksum_counter & (-cksum_counter)) ==
2881 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2882 cksum_counter, libcfs_idstr(&req->rq_peer),
2883 repbody->oa.o_cksum);
2887 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK2, cfs_fail_val);
2890 /* calculate the expected actual write bytes (nob) for OFD stats.
2891 * Technically, if commit fails this would be wrong, but that should be
2894 for (i = 0; i < niocount; i++) {
2895 int len = remote_nb[i].rnb_len;
2900 /* multiple transactions can be assigned during write commit */
2901 tsi->tsi_mult_trans = 1;
2903 /* Must commit after prep above in all cases */
2904 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2905 objcount, ioo, remote_nb, npages, local_nb, rc, nob,
2907 if (rc == -ENOTCONN)
2908 /* quota acquire process has been given up because
2909 * either the client has been evicted or the client
2910 * has timed out the request already
2914 for (i = 0; i < niocount; i++) {
2915 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2921 * Disable sending mtime back to the client. If the client locked the
2922 * whole object, then it has already updated the mtime on its side,
2923 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2925 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2928 /* set per-requested niobuf return codes */
2929 for (i = j = 0; i < niocount; i++) {
2930 int len = remote_nb[i].rnb_len;
2934 LASSERT(j < npages);
2935 if (local_nb[j].lnb_rc < 0)
2936 rcs[i] = local_nb[j].lnb_rc;
2937 len -= local_nb[j].lnb_len;
2942 LASSERT(j == npages);
2943 ptlrpc_lprocfs_brw(req, nob);
2946 tgt_brw_unlock(exp, ioo, remote_nb, &lockh, LCK_PW);
2948 ptlrpc_free_bulk(desc);
2950 if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2951 req->rq_no_reply = 1;
2952 /* reply out callback would free */
2953 ptlrpc_req_drop_rs(req);
2954 if (!exp->exp_obd->obd_no_transno)
2955 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2956 " client will retry: rc = %d\n",
2958 obd_uuid2str(&exp->exp_client_uuid),
2959 obd_export_nid2str(exp), rc);
2963 memalloc_noreclaim_restore(mpflags);
2967 EXPORT_SYMBOL(tgt_brw_write);
2970 * Common request handler for OST_SEEK RPC.
2972 * Unified request handling for OST_SEEK RPC.
2973 * It takes object by its FID, does needed lseek and packs result
2974 * into reply. Only SEEK_HOLE and SEEK_DATA are supported.
2976 * \param[in] tsi target session environment for this request
2978 * \retval 0 if successful
2979 * \retval negative value on error
2981 int tgt_lseek(struct tgt_session_info *tsi)
2983 struct lustre_handle lh = { 0 };
2984 struct dt_object *dob;
2985 struct ost_body *repbody;
2986 loff_t offset = tsi->tsi_ost_body->oa.o_size;
2987 int whence = tsi->tsi_ost_body->oa.o_mode;
2993 if (whence != SEEK_HOLE && whence != SEEK_DATA)
2996 /* Negative offset is prohibited on wire and must be handled on client
2997 * prior sending RPC.
3002 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
3003 if (repbody == NULL)
3005 repbody->oa = tsi->tsi_ost_body->oa;
3007 srvlock = tsi->tsi_ost_body->oa.o_valid & OBD_MD_FLFLAGS &&
3008 tsi->tsi_ost_body->oa.o_flags & OBD_FL_SRVLOCK;
3010 rc = tgt_data_lock(tsi->tsi_env, tsi->tsi_exp, &tsi->tsi_resid,
3011 offset, OBD_OBJECT_EOF, &lh, LCK_PR);
3016 dob = dt_locate(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, &tsi->tsi_fid);
3018 GOTO(out, rc = PTR_ERR(dob));
3020 if (!dt_object_exists(dob))
3021 GOTO(obj_put, rc = -ENOENT);
3023 repbody->oa.o_size = dt_lseek(tsi->tsi_env, dob, offset, whence);
3026 dt_object_put(tsi->tsi_env, dob);
3029 tgt_data_unlock(&lh, LCK_PR);
3033 EXPORT_SYMBOL(tgt_lseek);
3035 /* Check if request can be reconstructed from saved reply data
3036 * A copy of the reply data is returned in @trd if the pointer is not NULL
3038 int req_can_reconstruct(struct ptlrpc_request *req,
3039 struct tg_reply_data *trd)
3041 struct tg_export_data *ted = &req->rq_export->exp_target_data;
3042 struct lsd_client_data *lcd = ted->ted_lcd;
3045 if (tgt_is_multimodrpcs_client(req->rq_export))
3046 return tgt_lookup_reply(req, trd);
3048 mutex_lock(&ted->ted_lcd_lock);
3049 found = req->rq_xid == lcd->lcd_last_xid ||
3050 req->rq_xid == lcd->lcd_last_close_xid;
3052 if (found && trd != NULL) {
3053 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
3054 trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
3055 trd->trd_reply.lrd_transno =
3056 lcd->lcd_last_close_transno;
3057 trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
3059 trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
3060 trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
3061 trd->trd_reply.lrd_result = lcd->lcd_last_result;
3062 trd->trd_reply.lrd_data = lcd->lcd_last_data;
3063 trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
3064 trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
3065 trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
3066 trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
3069 mutex_unlock(&ted->ted_lcd_lock);
3073 EXPORT_SYMBOL(req_can_reconstruct);
3075 bool tgt_check_resent(struct ptlrpc_request *req, struct tg_reply_data *trd)
3077 struct lsd_reply_data *lrd;
3078 bool need_reconstruct = false;
3082 if (likely(!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT)))
3085 if (req_can_reconstruct(req, trd)) {
3086 lrd = &trd->trd_reply;
3087 req->rq_transno = lrd->lrd_transno;
3088 req->rq_status = lrd->lrd_result;
3089 if (req->rq_status != 0)
3090 req->rq_transno = 0;
3091 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
3092 lustre_msg_set_status(req->rq_repmsg, req->rq_status);
3094 DEBUG_REQ(D_HA, req, "reconstruct resent RPC");
3095 need_reconstruct = true;
3097 DEBUG_REQ(D_HA, req, "no reply for RESENT req");
3100 return need_reconstruct;
3102 EXPORT_SYMBOL(tgt_check_resent);