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 #ifdef HAVE_UIDGID_HEADER
40 # include <linux/uidgid.h>
44 #include <obd_class.h>
45 #include <obd_cksum.h>
46 #include <lustre_lfsck.h>
47 #include <lustre_nodemap.h>
48 #include <lustre_acl.h>
50 #include "tgt_internal.h"
52 char *tgt_name(struct lu_target *tgt)
54 LASSERT(tgt->lut_obd != NULL);
55 return tgt->lut_obd->obd_name;
57 EXPORT_SYMBOL(tgt_name);
60 * Generic code handling requests that have struct mdt_body passed in:
62 * - extract mdt_body from request and save it in @tsi, if present;
64 * - create lu_object, corresponding to the fid in mdt_body, and save it in
67 * - if HAS_BODY flag is set for this request type check whether object
68 * actually exists on storage (lu_object_exists()).
71 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
73 const struct mdt_body *body;
74 struct lu_object *obj;
75 struct req_capsule *pill = tsi->tsi_pill;
80 body = req_capsule_client_get(pill, &RMF_MDT_BODY);
84 tsi->tsi_mdt_body = body;
86 if (!(body->mbo_valid & OBD_MD_FLID))
89 /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
90 * in any case in pre-2.5 clients. Fix that here if needed */
91 if (unlikely(fid_is_zero(&body->mbo_fid1)))
94 if (!fid_is_sane(&body->mbo_fid1)) {
95 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
96 PFID(&body->mbo_fid1));
100 obj = lu_object_find(tsi->tsi_env,
101 &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
102 &body->mbo_fid1, NULL);
104 if ((flags & HAS_BODY) && !lu_object_exists(obj)) {
105 lu_object_put(tsi->tsi_env, obj);
108 tsi->tsi_corpus = obj;
115 tsi->tsi_fid = body->mbo_fid1;
121 * Validate oa from client.
122 * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
124 * a. objects in Single MDT FS seq = FID_SEQ_OST_MDT0, oi_id != 0
125 * b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
126 * pack ost_id. Because non-zero oi_seq will make it diffcult to tell
127 * whether this is oi_fid or real ostid. So it will check
128 * OBD_CONNECT_FID, then convert the ostid to FID for old client.
129 * c. Old FID-disable osc will send IDIF.
130 * d. new FID-enable osc/osp will send normal FID.
132 * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
133 * be used for LAST_ID file, and only being accessed inside OST now.
135 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
137 struct ost_id *oi = &oa->o_oi;
138 u64 seq = ostid_seq(oi);
139 u64 id = ostid_id(oi);
143 if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
144 fid_seq_is_echo(seq))) {
145 /* Sigh 2.[123] client still sends echo req with oi_id = 0
146 * during create, and we will reset this to 1, since this
147 * oi_id is basically useless in the following create process,
148 * but oi_id == 0 will make it difficult to tell whether it is
149 * real FID or ost_id. */
150 oi->oi_fid.f_seq = FID_SEQ_ECHO;
151 oi->oi_fid.f_oid = id ?: 1;
152 oi->oi_fid.f_ver = 0;
154 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
156 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
157 GOTO(out, rc = -EPROTO);
159 /* Note: this check might be forced in 2.5 or 2.6, i.e.
160 * all of the requests are required to setup FLGROUP */
161 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
162 ostid_set_seq_mdt0(oi);
163 oa->o_valid |= OBD_MD_FLGROUP;
167 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
168 fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
169 GOTO(out, rc = -EPROTO);
171 rc = ostid_to_fid(&tti->tti_fid1, oi,
172 tsi->tsi_tgt->lut_lsd.lsd_osd_index);
173 if (unlikely(rc != 0))
176 oi->oi_fid = tti->tti_fid1;
182 CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
183 tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
187 EXPORT_SYMBOL(tgt_validate_obdo);
189 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
192 struct niobuf_remote *rnb;
193 struct obd_ioobj *ioo;
198 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
202 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
206 max_brw = ioobj_max_brw_get(ioo);
207 if (unlikely((max_brw & (max_brw - 1)) != 0)) {
208 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
209 ": rc = %d\n", tgt_name(tsi->tsi_tgt),
210 obd_export_nid2str(tsi->tsi_exp), max_brw,
211 POSTID(oi), -EPROTO);
216 obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
217 RCL_CLIENT) / sizeof(*ioo);
218 if (obj_count == 0) {
219 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
221 } else if (obj_count > 1) {
222 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
227 if (ioo->ioo_bufcnt == 0) {
228 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
232 if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
233 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
234 "bulk has too many pages (%d)",
242 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
244 struct ost_body *body;
245 struct req_capsule *pill = tsi->tsi_pill;
246 struct lu_nodemap *nodemap;
251 body = req_capsule_client_get(pill, &RMF_OST_BODY);
255 rc = tgt_validate_obdo(tsi, &body->oa);
259 nodemap = nodemap_get_from_exp(tsi->tsi_exp);
261 RETURN(PTR_ERR(nodemap));
263 body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
264 NODEMAP_CLIENT_TO_FS,
266 body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
267 NODEMAP_CLIENT_TO_FS,
269 nodemap_putref(nodemap);
271 tsi->tsi_ost_body = body;
272 tsi->tsi_fid = body->oa.o_oi.oi_fid;
274 if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
275 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
280 if (!(body->oa.o_valid & OBD_MD_FLID)) {
281 if (flags & HAS_BODY) {
282 CERROR("%s: OBD_MD_FLID flag is not set in ost_body but OID/FID is mandatory with HAS_BODY\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 & (HAS_BODY | HAS_REPLY),
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 & IS_MUTABLE && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
339 if (flags & HAS_KEY) {
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 & HAS_REPLY) {
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,
466 "%s: %s handler did not pack reply but returned no error",
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) == 1) {
558 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
559 (MSG_RESENT | MSG_REPLAY))) {
560 DEBUG_REQ(D_WARNING, req,
561 "rq_xid=%llu matches saved XID, expected REPLAY or RESENT flag (%x)",
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",
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)
635 struct obd_export *exp = req->rq_export;
636 struct tg_export_data *ted = &exp->exp_target_data;
637 bool need_lock = tgt_is_multimodrpcs_client(exp);
641 mutex_lock(&ted->ted_lcd_lock);
642 /* check request's xid is consistent with export's last_xid */
643 last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
644 if (last_xid > exp->exp_last_xid)
645 exp->exp_last_xid = last_xid;
647 if (req->rq_xid == 0 || req->rq_xid <= exp->exp_last_xid) {
648 /* Some request is allowed to be sent during replay,
649 * such as OUT update requests, FLD requests, so it
650 * is possible that replay requests has smaller XID
651 * than the exp_last_xid.
653 * Some non-replay requests may have smaller XID as
656 * - Client send a no_resend RPC, like statfs;
657 * - The RPC timedout (or some other error) on client,
658 * then it's removed from the unreplied list;
659 * - Client send some other request to bump the
660 * exp_last_xid on server;
661 * - The former RPC got chance to be processed;
663 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
666 DEBUG_REQ(D_WARNING, req,
667 "unexpected xid=%llx != exp_last_xid=%llx, rc = %d",
668 req->rq_xid, exp->exp_last_xid, rc);
673 /* The "last_xid" is the minimum xid among unreplied requests,
674 * if the request is from the previous connection, its xid can
675 * still be larger than "exp_last_xid", then the above check of
676 * xid is not enough to determine whether the request is delayed.
678 * For example, if some replay request was delayed and caused
679 * timeout at client and the replay is restarted, the delayed
680 * replay request will have the larger xid than "exp_last_xid"
682 if (req->rq_export->exp_conn_cnt >
683 lustre_msg_get_conn_cnt(req->rq_reqmsg))
684 GOTO(out, rc = -ESTALE);
686 /* try to release in-memory reply data */
687 if (tgt_is_multimodrpcs_client(exp)) {
688 tgt_handle_received_xid(exp, last_xid);
689 rc = tgt_handle_tag(req);
694 mutex_unlock(&ted->ted_lcd_lock);
699 int tgt_request_handle(struct ptlrpc_request *req)
701 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
703 struct lustre_msg *msg = req->rq_reqmsg;
704 struct tgt_handler *h;
705 struct lu_target *tgt;
706 int request_fail_id = 0;
707 __u32 opc = lustre_msg_get_opc(msg);
708 struct obd_device *obd;
710 bool is_connect = false;
713 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
714 if (cfs_fail_val == 0 &&
715 lustre_msg_get_opc(msg) != OBD_PING &&
716 lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
717 struct l_wait_info lwi = { 0 };
721 l_wait_event(cfs_race_waitq, (cfs_race_state == 1),
726 /* Refill the context, to make sure all thread keys are allocated */
727 lu_env_refill(req->rq_svc_thread->t_env);
729 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
730 tsi->tsi_pill = &req->rq_pill;
731 tsi->tsi_env = req->rq_svc_thread->t_env;
733 /* if request has export then get handlers slice from corresponding
734 * target, otherwise that should be connect operation */
735 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
736 opc == MGS_CONNECT) {
738 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
739 rc = target_handle_connect(req);
741 rc = ptlrpc_error(req);
744 /* recovery-small test 18c asks to drop connect reply */
745 if (unlikely(opc == OST_CONNECT &&
746 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
750 if (unlikely(!class_connected_export(req->rq_export))) {
751 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
752 opc == SEC_CTX_FINI) {
753 /* sec context initialization has to be handled
754 * by hand in tgt_handle_request0() */
755 tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
757 GOTO(handle_recov, rc = 0);
759 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
760 opc, libcfs_id2str(req->rq_peer));
761 req->rq_status = -ENOTCONN;
762 rc = ptlrpc_error(req);
766 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
767 tsi->tsi_exp = req->rq_export;
768 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
769 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
771 tsi->tsi_jobid = NULL;
774 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export",
775 class_exp2obd(req->rq_export)->obd_name);
776 req->rq_status = -EINVAL;
777 rc = ptlrpc_error(req);
781 /* Skip last_xid processing for the recovery thread, otherwise, the
782 * last_xid on same request could be processed twice: first time when
783 * processing the incoming request, second time when the request is
784 * being processed by recovery thread. */
785 obd = class_exp2obd(req->rq_export);
787 /* reset the exp_last_xid on each connection. */
788 req->rq_export->exp_last_xid = 0;
789 } else if (obd->obd_recovery_data.trd_processing_task !=
791 rc = process_req_last_xid(req);
794 rc = ptlrpc_error(req);
799 request_fail_id = tgt->lut_request_fail_id;
800 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
802 h = tgt_handler_find_check(req);
804 req->rq_status = PTR_ERR(h);
805 rc = ptlrpc_error(req);
809 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
812 if ((cfs_fail_val == 0 || cfs_fail_val == opc) &&
813 CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
816 rc = lustre_msg_check_version(msg, h->th_version);
818 DEBUG_REQ(D_ERROR, req,
819 "%s: drop malformed request version=%08x expect=%08x",
820 tgt_name(tgt), lustre_msg_get_version(msg),
822 req->rq_status = -EINVAL;
823 rc = ptlrpc_error(req);
828 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
829 if (likely(rc == 1)) {
830 rc = tgt_handle_request0(tsi, h, req);
836 req_capsule_fini(tsi->tsi_pill);
837 if (tsi->tsi_corpus != NULL) {
838 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
839 tsi->tsi_corpus = NULL;
843 EXPORT_SYMBOL(tgt_request_handle);
845 /** Assign high priority operations to the request if needed. */
846 int tgt_hpreq_handler(struct ptlrpc_request *req)
848 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
849 struct tgt_handler *h;
854 if (req->rq_export == NULL)
857 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
858 tsi->tsi_pill = &req->rq_pill;
859 tsi->tsi_env = req->rq_svc_thread->t_env;
860 tsi->tsi_tgt = class_exp2tgt(req->rq_export);
861 tsi->tsi_exp = req->rq_export;
863 h = tgt_handler_find_check(req);
869 rc = tgt_request_preprocess(tsi, h, req);
870 if (unlikely(rc != 0))
873 if (h->th_hp != NULL)
877 EXPORT_SYMBOL(tgt_hpreq_handler);
879 void tgt_counter_incr(struct obd_export *exp, int opcode)
881 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
882 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
883 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
885 EXPORT_SYMBOL(tgt_counter_incr);
888 * Unified target generic handlers.
891 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
893 struct lu_target *tgt = class_exp2tgt(exp);
894 struct sptlrpc_flavor flvr;
898 LASSERT(tgt->lut_obd);
899 LASSERT(tgt->lut_slice);
901 /* always allow ECHO client */
902 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
903 LUSTRE_ECHO_NAME) == 0)) {
904 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
908 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
909 read_lock(&tgt->lut_sptlrpc_lock);
910 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
914 read_unlock(&tgt->lut_sptlrpc_lock);
916 spin_lock(&exp->exp_lock);
917 exp->exp_sp_peer = req->rq_sp_from;
918 exp->exp_flvr = flvr;
920 /* when on mgs, if no restriction is set, or if client
921 * is loopback, allow any flavor */
922 if ((strcmp(exp->exp_obd->obd_type->typ_name,
923 LUSTRE_MGS_NAME) == 0) &&
924 (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
925 LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
927 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
929 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
930 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
931 CERROR("%s: unauthorized rpc flavor %x from %s, "
932 "expect %x\n", tgt_name(tgt),
934 libcfs_nid2str(req->rq_peer.nid),
935 exp->exp_flvr.sf_rpc);
938 spin_unlock(&exp->exp_lock);
940 if (exp->exp_sp_peer != req->rq_sp_from) {
941 CERROR("%s: RPC source %s doesn't match %s\n",
943 sptlrpc_part2name(req->rq_sp_from),
944 sptlrpc_part2name(exp->exp_sp_peer));
947 rc = sptlrpc_target_export_check(exp, req);
954 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
956 struct sptlrpc_rule_set tmp_rset;
959 if (unlikely(tgt == NULL)) {
960 CERROR("No target passed");
964 sptlrpc_rule_set_init(&tmp_rset);
965 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
967 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
972 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
974 write_lock(&tgt->lut_sptlrpc_lock);
975 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
976 tgt->lut_sptlrpc_rset = tmp_rset;
977 write_unlock(&tgt->lut_sptlrpc_lock);
981 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
983 int tgt_connect(struct tgt_session_info *tsi)
985 struct ptlrpc_request *req = tgt_ses_req(tsi);
986 struct obd_connect_data *reply;
991 /* XXX: better to call this check right after getting new export but
992 * before last_rcvd slot allocation to avoid server load upon insecure
993 * connects. This is to be fixed after unifiyng all targets.
995 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
999 /* To avoid exposing partially initialized connection flags, changes up
1000 * to this point have been staged in reply->ocd_connect_flags. Now that
1001 * connection handling has completed successfully, atomically update
1002 * the connect flags in the shared export data structure. LU-1623 */
1003 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
1004 spin_lock(&tsi->tsi_exp->exp_lock);
1005 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
1006 if (reply->ocd_connect_flags & OBD_CONNECT_FLAGS2)
1007 *exp_connect_flags2_ptr(tsi->tsi_exp) =
1008 reply->ocd_connect_flags2;
1009 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
1010 spin_unlock(&tsi->tsi_exp->exp_lock);
1012 if (strcmp(tsi->tsi_exp->exp_obd->obd_type->typ_name,
1013 LUSTRE_MDT_NAME) == 0) {
1014 rc = req_check_sepol(tsi->tsi_pill);
1021 obd_disconnect(class_export_get(tsi->tsi_exp));
1024 EXPORT_SYMBOL(tgt_connect);
1026 int tgt_disconnect(struct tgt_session_info *tsi)
1032 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_DISCONNECT_DELAY, cfs_fail_val);
1034 rc = target_handle_disconnect(tgt_ses_req(tsi));
1036 RETURN(err_serious(rc));
1040 EXPORT_SYMBOL(tgt_disconnect);
1043 * Unified target OBD handlers
1045 int tgt_obd_ping(struct tgt_session_info *tsi)
1051 /* The target-specific part of OBD_PING request handling.
1052 * It controls Filter Modification Data (FMD) expiration each time
1055 * Valid only for replayable targets, e.g. MDT and OFD
1057 if (tsi->tsi_exp->exp_obd->obd_replayable)
1058 tgt_fmd_expire(tsi->tsi_exp);
1060 rc = req_capsule_server_pack(tsi->tsi_pill);
1062 RETURN(err_serious(rc));
1066 EXPORT_SYMBOL(tgt_obd_ping);
1068 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1070 return err_serious(-EOPNOTSUPP);
1073 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1075 struct ptlrpc_request *req = tgt_ses_req(tsi);
1076 struct obd_export *exp = req->rq_export;
1077 struct ptlrpc_bulk_desc *desc;
1083 desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1084 PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1085 MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1089 for (i = 0; i < rdbuf->rb_nbufs; i++)
1090 desc->bd_frag_ops->add_iov_frag(desc,
1091 rdbuf->rb_bufs[i].lb_buf,
1092 rdbuf->rb_bufs[i].lb_len);
1094 rc = target_bulk_io(exp, desc);
1095 ptlrpc_free_bulk(desc);
1098 EXPORT_SYMBOL(tgt_send_buffer);
1100 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1102 struct ptlrpc_request *req = tgt_ses_req(tsi);
1103 struct obd_export *exp = req->rq_export;
1104 struct ptlrpc_bulk_desc *desc;
1112 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1113 PTLRPC_BULK_PUT_SOURCE |
1114 PTLRPC_BULK_BUF_KIOV,
1116 &ptlrpc_bulk_kiov_pin_ops);
1120 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1121 /* old client requires reply size in it's PAGE_SIZE,
1122 * which is rdpg->rp_count */
1123 nob = rdpg->rp_count;
1125 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1126 i++, tmpcount -= tmpsize) {
1127 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1128 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1132 LASSERT(desc->bd_nob == nob);
1133 rc = target_bulk_io(exp, desc);
1134 ptlrpc_free_bulk(desc);
1137 EXPORT_SYMBOL(tgt_sendpage);
1140 * OBD_IDX_READ handler
1142 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1144 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1145 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1146 struct idx_info *req_ii, *rep_ii;
1151 memset(rdpg, 0, sizeof(*rdpg));
1152 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1154 /* extract idx_info buffer from request & reply */
1155 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1156 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1157 RETURN(err_serious(-EPROTO));
1159 rc = req_capsule_server_pack(tsi->tsi_pill);
1161 RETURN(err_serious(rc));
1163 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1165 RETURN(err_serious(-EFAULT));
1166 rep_ii->ii_magic = IDX_INFO_MAGIC;
1168 /* extract hash to start with */
1169 rdpg->rp_hash = req_ii->ii_hash_start;
1171 /* extract requested attributes */
1172 rdpg->rp_attrs = req_ii->ii_attrs;
1174 /* check that fid packed in request is valid and supported */
1175 if (!fid_is_sane(&req_ii->ii_fid))
1177 rep_ii->ii_fid = req_ii->ii_fid;
1180 rep_ii->ii_flags = req_ii->ii_flags;
1182 /* compute number of pages to allocate, ii_count is the number of 4KB
1184 if (req_ii->ii_count <= 0)
1185 GOTO(out, rc = -EFAULT);
1186 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1187 exp_max_brw_size(tsi->tsi_exp));
1188 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1190 /* allocate pages to store the containers */
1191 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1192 if (rdpg->rp_pages == NULL)
1193 GOTO(out, rc = -ENOMEM);
1194 for (i = 0; i < rdpg->rp_npages; i++) {
1195 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1196 if (rdpg->rp_pages[i] == NULL)
1197 GOTO(out, rc = -ENOMEM);
1200 /* populate pages with key/record pairs */
1201 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1205 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1206 "asked %d > %d\n", rc, rdpg->rp_count);
1208 /* send pages to client */
1209 rc = tgt_sendpage(tsi, rdpg, rc);
1214 if (rdpg->rp_pages) {
1215 for (i = 0; i < rdpg->rp_npages; i++)
1216 if (rdpg->rp_pages[i])
1217 __free_page(rdpg->rp_pages[i]);
1218 OBD_FREE(rdpg->rp_pages,
1219 rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1224 struct tgt_handler tgt_obd_handlers[] = {
1225 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1226 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1228 EXPORT_SYMBOL(tgt_obd_handlers);
1230 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1231 struct dt_object *obj, __u64 start, __u64 end)
1237 /* if no objid is specified, it means "sync whole filesystem" */
1239 rc = dt_sync(env, tgt->lut_bottom);
1240 } else if (dt_version_get(env, obj) >
1241 tgt->lut_obd->obd_last_committed) {
1242 rc = dt_object_sync(env, obj, start, end);
1244 atomic_inc(&tgt->lut_sync_count);
1248 EXPORT_SYMBOL(tgt_sync);
1250 * Unified target DLM handlers.
1254 * Unified target BAST
1256 * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1257 * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1258 * but if it's cross-MDT lock, because remote object version is not set, a
1259 * filesystem sync is needed.
1261 * \param lock server side lock
1262 * \param desc lock desc
1263 * \param data ldlm_cb_set_arg
1264 * \param flag indicates whether this cancelling or blocking callback
1265 * \retval 0 on success
1266 * \retval negative number on error
1268 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1269 void *data, int flag)
1272 struct lu_target *tgt;
1273 struct dt_object *obj = NULL;
1279 tgt = class_exp2tgt(lock->l_export);
1281 if (unlikely(tgt == NULL)) {
1282 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1283 class_exp2obd(lock->l_export)->obd_name);
1287 if (flag == LDLM_CB_CANCELING &&
1288 (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1289 (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1290 (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1291 ldlm_is_cbpending(lock))) &&
1292 ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1293 lock->l_resource->lr_type == LDLM_EXTENT)) {
1295 __u64 end = OBD_OBJECT_EOF;
1297 rc = lu_env_init(&env, LCT_DT_THREAD);
1298 if (unlikely(rc != 0))
1301 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1302 tgt->lut_lsd.lsd_osd_index);
1304 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1305 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1307 GOTO(err_env, rc = PTR_ERR(obj));
1309 if (!dt_object_exists(obj))
1310 GOTO(err_put, rc = -ENOENT);
1312 start = lock->l_policy_data.l_extent.start;
1313 end = lock->l_policy_data.l_extent.end;
1316 rc = tgt_sync(&env, tgt, obj, start, end);
1318 CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1319 "cancel: rc = %d\n",
1320 tgt_name(tgt), PFID(&fid),
1321 lock->l_policy_data.l_extent.start,
1322 lock->l_policy_data.l_extent.end, rc);
1326 dt_object_put(&env, obj);
1331 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1335 static struct ldlm_callback_suite tgt_dlm_cbs = {
1336 .lcs_completion = ldlm_server_completion_ast,
1337 .lcs_blocking = tgt_blocking_ast,
1338 .lcs_glimpse = ldlm_server_glimpse_ast
1341 int tgt_enqueue(struct tgt_session_info *tsi)
1343 struct ptlrpc_request *req = tgt_ses_req(tsi);
1348 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1349 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1351 LASSERT(tsi->tsi_dlm_req != NULL);
1352 rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1353 tsi->tsi_dlm_req, &tgt_dlm_cbs);
1355 RETURN(err_serious(rc));
1357 switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1359 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1362 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1365 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1368 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1371 RETURN(req->rq_status);
1373 EXPORT_SYMBOL(tgt_enqueue);
1375 int tgt_convert(struct tgt_session_info *tsi)
1377 struct ptlrpc_request *req = tgt_ses_req(tsi);
1381 LASSERT(tsi->tsi_dlm_req);
1382 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1384 RETURN(err_serious(rc));
1386 RETURN(req->rq_status);
1389 int tgt_bl_callback(struct tgt_session_info *tsi)
1391 return err_serious(-EOPNOTSUPP);
1394 int tgt_cp_callback(struct tgt_session_info *tsi)
1396 return err_serious(-EOPNOTSUPP);
1399 /* generic LDLM target handler */
1400 struct tgt_handler tgt_dlm_handlers[] = {
1401 TGT_DLM_HDL(HAS_KEY, LDLM_ENQUEUE, tgt_enqueue),
1402 TGT_DLM_HDL(HAS_KEY, LDLM_CONVERT, tgt_convert),
1403 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1404 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1406 EXPORT_SYMBOL(tgt_dlm_handlers);
1409 * Unified target LLOG handlers.
1411 int tgt_llog_open(struct tgt_session_info *tsi)
1417 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1421 EXPORT_SYMBOL(tgt_llog_open);
1423 int tgt_llog_read_header(struct tgt_session_info *tsi)
1429 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1433 EXPORT_SYMBOL(tgt_llog_read_header);
1435 int tgt_llog_next_block(struct tgt_session_info *tsi)
1441 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1445 EXPORT_SYMBOL(tgt_llog_next_block);
1447 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1453 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1457 EXPORT_SYMBOL(tgt_llog_prev_block);
1459 /* generic llog target handler */
1460 struct tgt_handler tgt_llog_handlers[] = {
1461 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1462 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1463 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1464 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1466 EXPORT_SYMBOL(tgt_llog_handlers);
1469 * sec context handlers
1471 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1472 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1477 struct tgt_handler tgt_sec_ctx_handlers[] = {
1478 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle),
1479 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle),
1480 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle),
1482 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1484 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1485 struct dt_device *key,
1486 struct lfsck_req_local *lrl,
1487 struct thandle *th) = NULL;
1489 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1491 struct lfsck_req_local *,
1494 tgt_lfsck_in_notify_local = notify;
1496 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1498 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1499 struct dt_device *key,
1500 struct lfsck_request *lr) = NULL;
1502 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1504 struct lfsck_request *))
1506 tgt_lfsck_in_notify = notify;
1508 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1510 static int (*tgt_lfsck_query)(const struct lu_env *env,
1511 struct dt_device *key,
1512 struct lfsck_request *req,
1513 struct lfsck_reply *rep,
1514 struct lfsck_query *que) = NULL;
1516 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1518 struct lfsck_request *,
1519 struct lfsck_reply *,
1520 struct lfsck_query *))
1522 tgt_lfsck_query = query;
1524 EXPORT_SYMBOL(tgt_register_lfsck_query);
1526 /* LFSCK request handlers */
1527 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1529 const struct lu_env *env = tsi->tsi_env;
1530 struct dt_device *key = tsi->tsi_tgt->lut_bottom;
1531 struct lfsck_request *lr;
1535 lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1539 rc = tgt_lfsck_in_notify(env, key, lr);
1544 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1546 struct lfsck_request *request;
1547 struct lfsck_reply *reply;
1551 request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1552 if (request == NULL)
1555 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1559 rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1560 request, reply, NULL);
1562 RETURN(rc < 0 ? rc : 0);
1565 struct tgt_handler tgt_lfsck_handlers[] = {
1566 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_NOTIFY, tgt_handle_lfsck_notify),
1567 TGT_LFSCK_HDL(HAS_REPLY, LFSCK_QUERY, tgt_handle_lfsck_query),
1569 EXPORT_SYMBOL(tgt_lfsck_handlers);
1572 * initialize per-thread page pool (bug 5137).
1574 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1576 struct tgt_thread_big_cache *tbc;
1580 LASSERT(thread != NULL);
1581 LASSERT(thread->t_data == NULL);
1583 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1586 thread->t_data = tbc;
1589 EXPORT_SYMBOL(tgt_io_thread_init);
1592 * free per-thread pool created by tgt_thread_init().
1594 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1596 struct tgt_thread_big_cache *tbc;
1600 LASSERT(thread != NULL);
1603 * be prepared to handle partially-initialized pools (because this is
1604 * called from ost_io_thread_init() for cleanup.
1606 tbc = thread->t_data;
1608 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1609 thread->t_data = NULL;
1613 EXPORT_SYMBOL(tgt_io_thread_done);
1616 * Helper function for getting Data-on-MDT file server DLM lock
1617 * if asked by client.
1619 int tgt_mdt_data_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1620 struct lustre_handle *lh, int mode, __u64 *flags)
1622 union ldlm_policy_data policy = {
1623 .l_inodebits.bits = MDS_INODELOCK_DOM,
1629 LASSERT(lh != NULL);
1630 LASSERT(ns != NULL);
1631 LASSERT(!lustre_handle_is_used(lh));
1633 rc = ldlm_cli_enqueue_local(NULL, ns, res_id, LDLM_IBITS, &policy, mode,
1634 flags, ldlm_blocking_ast,
1635 ldlm_completion_ast, ldlm_glimpse_ast,
1636 NULL, 0, LVB_T_NONE, NULL, lh);
1638 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1640 EXPORT_SYMBOL(tgt_mdt_data_lock);
1642 void tgt_mdt_data_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1644 LASSERT(lustre_handle_is_used(lh));
1645 ldlm_lock_decref(lh, mode);
1647 EXPORT_SYMBOL(tgt_mdt_data_unlock);
1650 * Helper function for getting server side [start, start+count] DLM lock
1651 * if asked by client.
1653 int tgt_extent_lock(const struct lu_env *env, struct ldlm_namespace *ns,
1654 struct ldlm_res_id *res_id, __u64 start, __u64 end,
1655 struct lustre_handle *lh, int mode, __u64 *flags)
1657 union ldlm_policy_data policy;
1662 LASSERT(lh != NULL);
1663 LASSERT(ns != NULL);
1664 LASSERT(!lustre_handle_is_used(lh));
1666 policy.l_extent.gid = 0;
1667 policy.l_extent.start = start & PAGE_MASK;
1670 * If ->o_blocks is EOF it means "lock till the end of the file".
1671 * Otherwise, it's size of an extent or hole being punched (in bytes).
1673 if (end == OBD_OBJECT_EOF || end < start)
1674 policy.l_extent.end = OBD_OBJECT_EOF;
1676 policy.l_extent.end = end | ~PAGE_MASK;
1678 rc = ldlm_cli_enqueue_local(env, ns, res_id, LDLM_EXTENT, &policy,
1679 mode, flags, ldlm_blocking_ast,
1680 ldlm_completion_ast, ldlm_glimpse_ast,
1681 NULL, 0, LVB_T_NONE, NULL, lh);
1682 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1684 EXPORT_SYMBOL(tgt_extent_lock);
1686 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1688 LASSERT(lustre_handle_is_used(lh));
1689 ldlm_lock_decref(lh, mode);
1691 EXPORT_SYMBOL(tgt_extent_unlock);
1693 static int tgt_brw_lock(const struct lu_env *env, struct obd_export *exp,
1694 struct ldlm_res_id *res_id, struct obd_ioobj *obj,
1695 struct niobuf_remote *nb, struct lustre_handle *lh,
1696 enum ldlm_mode mode)
1698 struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
1700 int nrbufs = obj->ioo_bufcnt;
1706 LASSERT(mode == LCK_PR || mode == LCK_PW);
1707 LASSERT(!lustre_handle_is_used(lh));
1709 if (ns->ns_obd->obd_recovering)
1712 if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1715 for (i = 1; i < nrbufs; i++)
1716 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1719 /* MDT IO for data-on-mdt */
1720 if (exp->exp_connect_data.ocd_connect_flags & OBD_CONNECT_IBITS)
1721 rc = tgt_mdt_data_lock(ns, res_id, lh, mode, &flags);
1723 rc = tgt_extent_lock(env, ns, res_id, nb[0].rnb_offset,
1724 nb[nrbufs - 1].rnb_offset +
1725 nb[nrbufs - 1].rnb_len - 1,
1730 static void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1731 struct lustre_handle *lh, enum ldlm_mode mode)
1735 LASSERT(mode == LCK_PR || mode == LCK_PW);
1736 LASSERT((obj->ioo_bufcnt > 0 &&
1737 (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1738 lustre_handle_is_used(lh));
1740 if (lustre_handle_is_used(lh))
1741 tgt_extent_unlock(lh, mode);
1744 static int tgt_checksum_niobuf(struct lu_target *tgt,
1745 struct niobuf_local *local_nb, int npages,
1746 int opc, enum cksum_types cksum_type,
1749 struct ahash_request *req;
1750 unsigned int bufsize;
1752 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1754 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1756 CERROR("%s: unable to initialize checksum hash %s\n",
1757 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1758 return PTR_ERR(req);
1761 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1762 for (i = 0; i < npages; i++) {
1763 /* corrupt the data before we compute the checksum, to
1764 * simulate a client->OST data error */
1765 if (i == 0 && opc == OST_WRITE &&
1766 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1767 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1768 int len = local_nb[i].lnb_len;
1769 struct page *np = tgt_page_to_corrupt;
1772 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1774 char *ptr2 = page_address(np);
1776 memcpy(ptr2 + off, ptr + off, len);
1777 memcpy(ptr2 + off, "bad3", min(4, len));
1778 ll_kunmap_atomic(ptr, KM_USER0);
1780 /* LU-8376 to preserve original index for
1781 * display in dump_all_bulk_pages() */
1784 cfs_crypto_hash_update_page(req, np, off,
1788 CERROR("%s: can't alloc page for corruption\n",
1792 cfs_crypto_hash_update_page(req, local_nb[i].lnb_page,
1793 local_nb[i].lnb_page_offset & ~PAGE_MASK,
1794 local_nb[i].lnb_len);
1796 /* corrupt the data after we compute the checksum, to
1797 * simulate an OST->client data error */
1798 if (i == 0 && opc == OST_READ &&
1799 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1800 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
1801 int len = local_nb[i].lnb_len;
1802 struct page *np = tgt_page_to_corrupt;
1805 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
1807 char *ptr2 = page_address(np);
1809 memcpy(ptr2 + off, ptr + off, len);
1810 memcpy(ptr2 + off, "bad4", min(4, len));
1811 ll_kunmap_atomic(ptr, KM_USER0);
1813 /* LU-8376 to preserve original index for
1814 * display in dump_all_bulk_pages() */
1817 cfs_crypto_hash_update_page(req, np, off,
1821 CERROR("%s: can't alloc page for corruption\n",
1827 bufsize = sizeof(*cksum);
1828 err = cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1833 char dbgcksum_file_name[PATH_MAX];
1835 static void dump_all_bulk_pages(struct obdo *oa, int count,
1836 struct niobuf_local *local_nb,
1837 __u32 server_cksum, __u32 client_cksum)
1844 /* will only keep dump of pages on first error for the same range in
1845 * file/fid, not during the resends/retries. */
1846 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1847 "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1848 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1849 libcfs_debug_file_path_arr :
1850 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1851 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1852 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1853 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1854 local_nb[0].lnb_file_offset,
1855 local_nb[count-1].lnb_file_offset +
1856 local_nb[count-1].lnb_len - 1, client_cksum, server_cksum);
1857 filp = filp_open(dbgcksum_file_name,
1858 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1862 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1863 "checksum error: rc = %d\n", dbgcksum_file_name,
1866 CERROR("%s: can't open to dump pages with checksum "
1867 "error: rc = %d\n", dbgcksum_file_name, rc);
1871 for (i = 0; i < count; i++) {
1872 len = local_nb[i].lnb_len;
1873 buf = kmap(local_nb[i].lnb_page);
1875 rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1877 CERROR("%s: wanted to write %u but got %d "
1878 "error\n", dbgcksum_file_name, len, rc);
1883 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1884 dbgcksum_file_name, rc);
1886 kunmap(local_nb[i].lnb_page);
1889 rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1891 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1892 filp_close(filp, NULL);
1896 static int check_read_checksum(struct niobuf_local *local_nb, int npages,
1897 struct obd_export *exp, struct obdo *oa,
1898 const struct lnet_process_id *peer,
1899 __u32 client_cksum, __u32 server_cksum,
1900 enum cksum_types server_cksum_type)
1903 enum cksum_types cksum_type;
1906 /* unlikely to happen and only if resend does not occur due to cksum
1907 * control failure on Client */
1908 if (unlikely(server_cksum == client_cksum)) {
1909 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1914 if (exp->exp_obd->obd_checksum_dump)
1915 dump_all_bulk_pages(oa, npages, local_nb, server_cksum,
1918 cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1921 if (cksum_type != server_cksum_type)
1922 msg = "the server may have not used the checksum type specified"
1923 " in the original request - likely a protocol problem";
1925 msg = "should have changed on the client or in transit";
1927 start = local_nb[0].lnb_file_offset;
1928 end = local_nb[npages-1].lnb_file_offset +
1929 local_nb[npages-1].lnb_len - 1;
1931 LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1932 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1933 " %x (type %x), server csum %x (type %x)\n",
1934 exp->exp_obd->obd_name,
1935 msg, libcfs_nid2str(peer->nid),
1936 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1937 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1938 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1940 start, end, client_cksum, cksum_type, server_cksum,
1946 static int tgt_pages2shortio(struct niobuf_local *local, int npages,
1947 unsigned char *buf, int size)
1949 int i, off, len, copied = size;
1952 for (i = 0; i < npages; i++) {
1953 off = local[i].lnb_page_offset & ~PAGE_MASK;
1954 len = local[i].lnb_len;
1956 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
1961 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
1962 memcpy(buf + off, ptr, len);
1963 ll_kunmap_atomic(ptr, KM_USER0);
1967 return copied - size;
1970 static int tgt_checksum_niobuf_t10pi(struct lu_target *tgt,
1971 struct niobuf_local *local_nb,
1972 int npages, int opc,
1973 obd_dif_csum_fn *fn,
1977 enum cksum_types t10_cksum_type = tgt->lut_dt_conf.ddp_t10_cksum_type;
1978 unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1979 const char *obd_name = tgt->lut_obd->obd_name;
1980 struct ahash_request *req;
1981 unsigned int bufsize;
1982 unsigned char *buffer;
1983 struct page *__page;
1986 int used_number = 0;
1992 __page = alloc_page(GFP_KERNEL);
1996 req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1998 CERROR("%s: unable to initialize checksum hash %s\n",
1999 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
2000 return PTR_ERR(req);
2003 buffer = kmap(__page);
2004 guard_start = (__u16 *)buffer;
2005 guard_number = PAGE_SIZE / sizeof(*guard_start);
2006 for (i = 0; i < npages; i++) {
2007 /* corrupt the data before we compute the checksum, to
2008 * simulate a client->OST data error */
2009 if (i == 0 && opc == OST_WRITE &&
2010 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
2011 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2012 int len = local_nb[i].lnb_len;
2013 struct page *np = tgt_page_to_corrupt;
2016 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2018 char *ptr2 = page_address(np);
2020 memcpy(ptr2 + off, ptr + off, len);
2021 memcpy(ptr2 + off, "bad3", min(4, len));
2022 ll_kunmap_atomic(ptr, KM_USER0);
2024 /* LU-8376 to preserve original index for
2025 * display in dump_all_bulk_pages() */
2028 cfs_crypto_hash_update_page(req, np, off,
2032 CERROR("%s: can't alloc page for corruption\n",
2038 * The left guard number should be able to hold checksums of a
2041 if (t10_cksum_type && opc == OST_READ &&
2042 local_nb[i].lnb_guard_disk) {
2043 used = DIV_ROUND_UP(local_nb[i].lnb_len, sector_size);
2044 if (used > (guard_number - used_number)) {
2048 memcpy(guard_start + used_number,
2049 local_nb[i].lnb_guards,
2050 used * sizeof(*local_nb[i].lnb_guards));
2052 rc = obd_page_dif_generate_buffer(obd_name,
2053 local_nb[i].lnb_page,
2054 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2055 local_nb[i].lnb_len, guard_start + used_number,
2056 guard_number - used_number, &used, sector_size,
2062 LASSERT(used <= MAX_GUARD_NUMBER);
2064 * If disk support T10PI checksum, copy guards to local_nb.
2065 * If the write is partial page, do not use the guards for bio
2066 * submission since the data might not be full-sector. The bio
2067 * guards will be generated later based on the full sectors. If
2068 * the sector size is 512B rather than 4 KB, or the page size
2069 * is larger than 4KB, this might drop some useful guards for
2070 * partial page write, but it will only add minimal extra time
2071 * of checksum calculation.
2073 if (t10_cksum_type && opc == OST_WRITE &&
2074 local_nb[i].lnb_len == PAGE_SIZE) {
2075 local_nb[i].lnb_guard_rpc = 1;
2076 memcpy(local_nb[i].lnb_guards,
2077 guard_start + used_number,
2078 used * sizeof(*local_nb[i].lnb_guards));
2081 used_number += used;
2082 if (used_number == guard_number) {
2083 cfs_crypto_hash_update_page(req, __page, 0,
2084 used_number * sizeof(*guard_start));
2088 /* corrupt the data after we compute the checksum, to
2089 * simulate an OST->client data error */
2090 if (unlikely(i == 0 && opc == OST_READ &&
2091 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))) {
2092 int off = local_nb[i].lnb_page_offset & ~PAGE_MASK;
2093 int len = local_nb[i].lnb_len;
2094 struct page *np = tgt_page_to_corrupt;
2097 char *ptr = ll_kmap_atomic(local_nb[i].lnb_page,
2099 char *ptr2 = page_address(np);
2101 memcpy(ptr2 + off, ptr + off, len);
2102 memcpy(ptr2 + off, "bad4", min(4, len));
2103 ll_kunmap_atomic(ptr, KM_USER0);
2105 /* LU-8376 to preserve original index for
2106 * display in dump_all_bulk_pages() */
2109 cfs_crypto_hash_update_page(req, np, off,
2113 CERROR("%s: can't alloc page for corruption\n",
2122 if (used_number != 0)
2123 cfs_crypto_hash_update_page(req, __page, 0,
2124 used_number * sizeof(*guard_start));
2126 bufsize = sizeof(cksum);
2127 rc = cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
2132 __free_page(__page);
2136 static int tgt_checksum_niobuf_rw(struct lu_target *tgt,
2137 enum cksum_types cksum_type,
2138 struct niobuf_local *local_nb,
2139 int npages, int opc, u32 *check_sum)
2141 obd_dif_csum_fn *fn = NULL;
2142 int sector_size = 0;
2146 obd_t10_cksum2dif(cksum_type, &fn, §or_size);
2149 rc = tgt_checksum_niobuf_t10pi(tgt, local_nb, npages,
2150 opc, fn, sector_size,
2153 rc = tgt_checksum_niobuf(tgt, local_nb, npages, opc,
2154 cksum_type, check_sum);
2158 int tgt_brw_read(struct tgt_session_info *tsi)
2160 struct ptlrpc_request *req = tgt_ses_req(tsi);
2161 struct ptlrpc_bulk_desc *desc = NULL;
2162 struct obd_export *exp = tsi->tsi_exp;
2163 struct niobuf_remote *remote_nb;
2164 struct niobuf_local *local_nb;
2165 struct obd_ioobj *ioo;
2166 struct ost_body *body, *repbody;
2167 struct l_wait_info lwi;
2168 struct lustre_handle lockh = { 0 };
2169 int npages, nob = 0, rc, i, no_reply = 0,
2171 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2172 const char *obd_name = exp->exp_obd->obd_name;
2176 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2177 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2178 CERROR("%s: deny read request from %s to portal %u\n",
2179 tgt_name(tsi->tsi_tgt),
2180 obd_export_nid2str(req->rq_export),
2181 ptlrpc_req2svc(req)->srv_req_portal);
2185 req->rq_bulk_read = 1;
2187 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
2190 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2191 cfs_fail_val : (obd_timeout + 1) / 4);
2193 /* Check if there is eviction in progress, and if so, wait for it to
2195 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2196 /* We do not care how long it takes */
2197 lwi = LWI_INTR(NULL, NULL);
2198 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2199 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
2203 /* There must be big cache in current thread to process this request
2204 * if it is NULL then something went wrong and it wasn't allocated,
2205 * report -ENOMEM in that case */
2209 body = tsi->tsi_ost_body;
2210 LASSERT(body != NULL);
2212 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
2213 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2215 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2216 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2218 local_nb = tbc->local;
2220 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2226 * If getting the lock took more time than
2227 * client was willing to wait, drop it. b=11330
2229 if (ktime_get_real_seconds() > req->rq_deadline ||
2230 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2232 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2233 libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
2234 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2235 req->rq_deadline - req->rq_arrival_time.tv_sec);
2236 GOTO(out_lock, rc = -ETIMEDOUT);
2239 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2240 repbody->oa = body->oa;
2242 npages = PTLRPC_MAX_BRW_PAGES;
2243 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
2244 ioo, remote_nb, &npages, local_nb);
2248 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2249 body->oa.o_flags & OBD_FL_SHORT_IO) {
2252 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2253 PTLRPC_BULK_PUT_SOURCE |
2254 PTLRPC_BULK_BUF_KIOV,
2256 &ptlrpc_bulk_kiov_nopin_ops);
2258 GOTO(out_commitrw, rc = -ENOMEM);
2262 npages_read = npages;
2263 for (i = 0; i < npages; i++) {
2264 int page_rc = local_nb[i].lnb_rc;
2273 if (page_rc != 0 && desc != NULL) { /* some data! */
2274 LASSERT(local_nb[i].lnb_page != NULL);
2275 desc->bd_frag_ops->add_kiov_frag
2276 (desc, local_nb[i].lnb_page,
2277 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2281 if (page_rc != local_nb[i].lnb_len) { /* short read */
2282 local_nb[i].lnb_len = page_rc;
2283 npages_read = i + (page_rc != 0 ? 1 : 0);
2284 /* All subsequent pages should be 0 */
2285 while (++i < npages)
2286 LASSERT(local_nb[i].lnb_rc == 0);
2290 if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
2291 nob != cfs_fail_val)
2294 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2295 u32 flag = body->oa.o_valid & OBD_MD_FLFLAGS ?
2296 body->oa.o_flags : 0;
2297 enum cksum_types cksum_type = obd_cksum_type_unpack(flag);
2299 repbody->oa.o_flags = obd_cksum_type_pack(obd_name,
2301 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2303 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2304 local_nb, npages_read, OST_READ,
2305 &repbody->oa.o_cksum);
2307 GOTO(out_commitrw, rc);
2308 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2309 repbody->oa.o_cksum);
2311 /* if a resend it could be for a cksum error, so check Server
2312 * cksum with returned Client cksum (this should even cover
2313 * zero-cksum case) */
2314 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2315 (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2316 check_read_checksum(local_nb, npages_read, exp,
2317 &body->oa, &req->rq_peer,
2319 repbody->oa.o_cksum, cksum_type);
2321 repbody->oa.o_valid = 0;
2323 /* We're finishing using body->oa as an input variable */
2325 /* Check if client was evicted while we were doing i/o before touching
2328 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2329 body->oa.o_flags & OBD_FL_SHORT_IO) {
2330 unsigned char *short_io_buf;
2333 short_io_buf = req_capsule_server_get(&req->rq_pill,
2335 short_io_size = req_capsule_get_size(&req->rq_pill,
2338 rc = tgt_pages2shortio(local_nb, npages_read,
2339 short_io_buf, short_io_size);
2341 req_capsule_shrink(&req->rq_pill,
2344 rc = rc > 0 ? 0 : rc;
2345 } else if (!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) {
2346 rc = target_bulk_io(exp, desc);
2350 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2351 body->oa.o_flags & OBD_FL_SHORT_IO)
2352 req_capsule_shrink(&req->rq_pill, &RMF_SHORT_IO, 0,
2357 /* Must commit after prep above in all cases */
2358 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2359 remote_nb, npages, local_nb, rc);
2361 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2363 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2364 ptlrpc_free_bulk(desc);
2369 ptlrpc_lprocfs_brw(req, nob);
2370 } else if (no_reply) {
2371 req->rq_no_reply = 1;
2372 /* reply out callback would free */
2373 ptlrpc_req_drop_rs(req);
2374 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2375 "client will retry: rc %d\n",
2377 obd_uuid2str(&exp->exp_client_uuid),
2378 obd_export_nid2str(exp), rc);
2380 /* send a bulk after reply to simulate a network delay or reordering
2381 * by a router - Note that !desc implies short io, so there is no bulk
2383 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) &&
2385 CDEBUG(D_INFO, "reorder BULK\n");
2388 target_bulk_io(exp, desc);
2389 ptlrpc_free_bulk(desc);
2394 EXPORT_SYMBOL(tgt_brw_read);
2396 static int tgt_shortio2pages(struct niobuf_local *local, int npages,
2397 unsigned char *buf, unsigned int size)
2402 for (i = 0; i < npages; i++) {
2403 off = local[i].lnb_page_offset & ~PAGE_MASK;
2404 len = local[i].lnb_len;
2409 CDEBUG(D_PAGE, "index %d offset = %d len = %d left = %d\n",
2411 ptr = ll_kmap_atomic(local[i].lnb_page, KM_USER0);
2414 memcpy(ptr + off, buf, len < size ? len : size);
2415 ll_kunmap_atomic(ptr, KM_USER0);
2422 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2423 struct ptlrpc_bulk_desc *desc,
2424 struct niobuf_local *local_nb, int npages,
2425 u32 client_cksum, u32 server_cksum,
2428 struct obd_export *exp = req->rq_export;
2429 struct ost_body *body;
2433 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2434 LASSERT(body != NULL);
2436 if (desc && req->rq_peer.nid != desc->bd_sender) {
2438 router = libcfs_nid2str(desc->bd_sender);
2441 if (exp->exp_obd->obd_checksum_dump)
2442 dump_all_bulk_pages(&body->oa, npages, local_nb, server_cksum,
2446 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2447 client_cksum, server_cksum);
2451 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2452 DFID" object "DOSTID" extent [%llu-%llu"
2453 "]: client csum %x, server csum %x\n",
2454 exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2456 body->oa.o_valid & OBD_MD_FLFID ?
2457 body->oa.o_parent_seq : (__u64)0,
2458 body->oa.o_valid & OBD_MD_FLFID ?
2459 body->oa.o_parent_oid : 0,
2460 body->oa.o_valid & OBD_MD_FLFID ?
2461 body->oa.o_parent_ver : 0,
2462 POSTID(&body->oa.o_oi),
2463 local_nb[0].lnb_file_offset,
2464 local_nb[npages-1].lnb_file_offset +
2465 local_nb[npages - 1].lnb_len - 1,
2466 client_cksum, server_cksum);
2469 int tgt_brw_write(struct tgt_session_info *tsi)
2471 struct ptlrpc_request *req = tgt_ses_req(tsi);
2472 struct ptlrpc_bulk_desc *desc = NULL;
2473 struct obd_export *exp = req->rq_export;
2474 struct niobuf_remote *remote_nb;
2475 struct niobuf_local *local_nb;
2476 struct obd_ioobj *ioo;
2477 struct ost_body *body, *repbody;
2478 struct lustre_handle lockh = {0};
2480 int objcount, niocount, npages;
2482 enum cksum_types cksum_type = OBD_CKSUM_CRC32;
2483 bool no_reply = false, mmap;
2484 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2485 bool wait_sync = false;
2486 const char *obd_name = exp->exp_obd->obd_name;
2490 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL &&
2491 ptlrpc_req2svc(req)->srv_req_portal != MDS_IO_PORTAL) {
2492 CERROR("%s: deny write request from %s to portal %u\n",
2493 tgt_name(tsi->tsi_tgt),
2494 obd_export_nid2str(req->rq_export),
2495 ptlrpc_req2svc(req)->srv_req_portal);
2496 RETURN(err_serious(-EPROTO));
2499 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2500 RETURN(err_serious(-ENOSPC));
2501 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2502 RETURN(err_serious(-EROFS));
2504 req->rq_bulk_write = 1;
2506 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2507 RETURN(err_serious(-EIO));
2508 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2509 RETURN(err_serious(-EFAULT));
2511 /* pause before transaction has been started */
2512 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2513 cfs_fail_val : (obd_timeout + 1) / 4);
2515 /* Delay write commit to show stale size information */
2516 CFS_FAIL_TIMEOUT(OBD_FAIL_OSC_NO_SIZE_DATA, cfs_fail_val);
2518 /* There must be big cache in current thread to process this request
2519 * if it is NULL then something went wrong and it wasn't allocated,
2520 * report -ENOMEM in that case */
2524 body = tsi->tsi_ost_body;
2525 LASSERT(body != NULL);
2527 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2528 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2530 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2531 RCL_CLIENT) / sizeof(*ioo);
2533 for (niocount = i = 0; i < objcount; i++)
2534 niocount += ioo[i].ioo_bufcnt;
2536 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2537 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2538 if (niocount != req_capsule_get_size(&req->rq_pill,
2539 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2541 RETURN(err_serious(-EPROTO));
2543 if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2544 ptlrpc_connection_is_local(exp->exp_connection))
2545 memory_pressure_set();
2547 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2548 niocount * sizeof(*rcs));
2549 rc = req_capsule_server_pack(&req->rq_pill);
2551 GOTO(out, rc = err_serious(rc));
2553 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2554 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2556 local_nb = tbc->local;
2558 rc = tgt_brw_lock(tsi->tsi_env, exp, &tsi->tsi_resid, ioo, remote_nb,
2564 * If getting the lock took more time than
2565 * client was willing to wait, drop it. b=11330
2567 if (ktime_get_real_seconds() > req->rq_deadline ||
2568 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2570 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2571 tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2572 POSTID(&ioo->ioo_oid),
2573 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2574 req->rq_deadline - req->rq_arrival_time.tv_sec);
2575 GOTO(out_lock, rc = -ETIMEDOUT);
2578 /* Because we already sync grant info with client when reconnect,
2579 * grant info will be cleared for resent req, then fed_grant and
2580 * total_grant will not be modified in following preprw_write */
2581 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2582 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2583 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2586 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2587 if (repbody == NULL)
2588 GOTO(out_lock, rc = -ENOMEM);
2589 repbody->oa = body->oa;
2591 npages = PTLRPC_MAX_BRW_PAGES;
2592 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2593 objcount, ioo, remote_nb, &npages, local_nb);
2596 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
2597 body->oa.o_flags & OBD_FL_SHORT_IO) {
2598 unsigned int short_io_size;
2599 unsigned char *short_io_buf;
2601 short_io_size = req_capsule_get_size(&req->rq_pill,
2604 short_io_buf = req_capsule_client_get(&req->rq_pill,
2606 CDEBUG(D_INFO, "Client use short io for data transfer,"
2607 " size = %d\n", short_io_size);
2609 /* Copy short io buf to pages */
2610 rc = tgt_shortio2pages(local_nb, npages, short_io_buf,
2614 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2615 PTLRPC_BULK_GET_SINK |
2616 PTLRPC_BULK_BUF_KIOV,
2618 &ptlrpc_bulk_kiov_nopin_ops);
2620 GOTO(skip_transfer, rc = -ENOMEM);
2622 /* NB Having prepped, we must commit... */
2623 for (i = 0; i < npages; i++)
2624 desc->bd_frag_ops->add_kiov_frag(desc,
2625 local_nb[i].lnb_page,
2626 local_nb[i].lnb_page_offset & ~PAGE_MASK,
2627 local_nb[i].lnb_len);
2629 rc = sptlrpc_svc_prep_bulk(req, desc);
2631 GOTO(skip_transfer, rc);
2633 rc = target_bulk_io(exp, desc);
2639 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2640 static int cksum_counter;
2642 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2643 cksum_type = obd_cksum_type_unpack(body->oa.o_flags);
2645 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2646 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2647 repbody->oa.o_flags |= obd_cksum_type_pack(obd_name,
2650 rc = tgt_checksum_niobuf_rw(tsi->tsi_tgt, cksum_type,
2651 local_nb, npages, OST_WRITE,
2652 &repbody->oa.o_cksum);
2654 GOTO(out_commitrw, rc);
2658 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2659 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2660 body->oa.o_flags & OBD_FL_MMAP);
2662 tgt_warn_on_cksum(req, desc, local_nb, npages,
2664 repbody->oa.o_cksum, mmap);
2666 } else if ((cksum_counter & (-cksum_counter)) ==
2668 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2669 cksum_counter, libcfs_id2str(req->rq_peer),
2670 repbody->oa.o_cksum);
2675 /* Must commit after prep above in all cases */
2676 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2677 objcount, ioo, remote_nb, npages, local_nb, rc);
2678 if (rc == -ENOTCONN)
2679 /* quota acquire process has been given up because
2680 * either the client has been evicted or the client
2681 * has timed out the request already */
2684 for (i = 0; i < niocount; i++) {
2685 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2691 * Disable sending mtime back to the client. If the client locked the
2692 * whole object, then it has already updated the mtime on its side,
2693 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2695 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2700 /* set per-requested niobuf return codes */
2701 for (i = j = 0; i < niocount; i++) {
2702 int len = remote_nb[i].rnb_len;
2707 LASSERT(j < npages);
2708 if (local_nb[j].lnb_rc < 0)
2709 rcs[i] = local_nb[j].lnb_rc;
2710 len -= local_nb[j].lnb_len;
2715 LASSERT(j == npages);
2716 ptlrpc_lprocfs_brw(req, nob);
2719 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2721 ptlrpc_free_bulk(desc);
2723 if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2724 req->rq_no_reply = 1;
2725 /* reply out callback would free */
2726 ptlrpc_req_drop_rs(req);
2727 if (!exp->exp_obd->obd_no_transno)
2728 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2729 " client will retry: rc = %d\n",
2731 obd_uuid2str(&exp->exp_client_uuid),
2732 obd_export_nid2str(exp), rc);
2734 memory_pressure_clr();
2737 EXPORT_SYMBOL(tgt_brw_write);
2739 /* Check if request can be reconstructed from saved reply data
2740 * A copy of the reply data is returned in @trd if the pointer is not NULL
2742 int req_can_reconstruct(struct ptlrpc_request *req,
2743 struct tg_reply_data *trd)
2745 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2746 struct lsd_client_data *lcd = ted->ted_lcd;
2749 if (tgt_is_multimodrpcs_client(req->rq_export))
2750 return tgt_lookup_reply(req, trd);
2752 mutex_lock(&ted->ted_lcd_lock);
2753 found = req->rq_xid == lcd->lcd_last_xid ||
2754 req->rq_xid == lcd->lcd_last_close_xid;
2756 if (found && trd != NULL) {
2757 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2758 trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2759 trd->trd_reply.lrd_transno =
2760 lcd->lcd_last_close_transno;
2761 trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2763 trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2764 trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2765 trd->trd_reply.lrd_result = lcd->lcd_last_result;
2766 trd->trd_reply.lrd_data = lcd->lcd_last_data;
2767 trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2768 trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2769 trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2770 trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2773 mutex_unlock(&ted->ted_lcd_lock);
2777 EXPORT_SYMBOL(req_can_reconstruct);