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, 2016, Intel Corporation.
27 * lustre/target/tgt_handler.c
29 * Lustre Unified Target request handler code
31 * Author: Brian Behlendorf <behlendorf1@llnl.gov>
32 * Author: Mikhail Pershin <mike.pershin@intel.com>
35 #define DEBUG_SUBSYSTEM S_CLASS
37 #include <linux/user_namespace.h>
38 #ifdef HAVE_UIDGID_HEADER
39 # include <linux/uidgid.h>
43 #include <obd_class.h>
44 #include <obd_cksum.h>
45 #include <lustre_lfsck.h>
46 #include <lustre_nodemap.h>
47 #include <lustre_acl.h>
49 #include "tgt_internal.h"
51 char *tgt_name(struct lu_target *tgt)
53 LASSERT(tgt->lut_obd != NULL);
54 return tgt->lut_obd->obd_name;
56 EXPORT_SYMBOL(tgt_name);
59 * Generic code handling requests that have struct mdt_body passed in:
61 * - extract mdt_body from request and save it in @tsi, if present;
63 * - create lu_object, corresponding to the fid in mdt_body, and save it in
66 * - if HABEO_CORPUS flag is set for this request type check whether object
67 * actually exists on storage (lu_object_exists()).
70 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
72 const struct mdt_body *body;
73 struct lu_object *obj;
74 struct req_capsule *pill = tsi->tsi_pill;
79 body = req_capsule_client_get(pill, &RMF_MDT_BODY);
83 tsi->tsi_mdt_body = body;
85 if (!(body->mbo_valid & OBD_MD_FLID))
88 /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
89 * in any case in pre-2.5 clients. Fix that here if needed */
90 if (unlikely(fid_is_zero(&body->mbo_fid1)))
93 if (!fid_is_sane(&body->mbo_fid1)) {
94 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
95 PFID(&body->mbo_fid1));
99 obj = lu_object_find(tsi->tsi_env,
100 &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
101 &body->mbo_fid1, NULL);
103 if ((flags & HABEO_CORPUS) && !lu_object_exists(obj)) {
104 lu_object_put(tsi->tsi_env, obj);
107 tsi->tsi_corpus = obj;
114 tsi->tsi_fid = body->mbo_fid1;
120 * Validate oa from client.
121 * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
123 * a. objects in Single MDT FS seq = FID_SEQ_OST_MDT0, oi_id != 0
124 * b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
125 * pack ost_id. Because non-zero oi_seq will make it diffcult to tell
126 * whether this is oi_fid or real ostid. So it will check
127 * OBD_CONNECT_FID, then convert the ostid to FID for old client.
128 * c. Old FID-disable osc will send IDIF.
129 * d. new FID-enable osc/osp will send normal FID.
131 * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
132 * be used for LAST_ID file, and only being accessed inside OST now.
134 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
136 struct ost_id *oi = &oa->o_oi;
137 u64 seq = ostid_seq(oi);
138 u64 id = ostid_id(oi);
142 if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
143 fid_seq_is_echo(seq))) {
144 /* Sigh 2.[123] client still sends echo req with oi_id = 0
145 * during create, and we will reset this to 1, since this
146 * oi_id is basically useless in the following create process,
147 * but oi_id == 0 will make it difficult to tell whether it is
148 * real FID or ost_id. */
149 oi->oi_fid.f_seq = FID_SEQ_ECHO;
150 oi->oi_fid.f_oid = id ?: 1;
151 oi->oi_fid.f_ver = 0;
153 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
155 if (unlikely((oa->o_valid & OBD_MD_FLID) && id == 0))
156 GOTO(out, rc = -EPROTO);
158 /* Note: this check might be forced in 2.5 or 2.6, i.e.
159 * all of the requests are required to setup FLGROUP */
160 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
161 ostid_set_seq_mdt0(oi);
162 oa->o_valid |= OBD_MD_FLGROUP;
166 if (unlikely(!(fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq) ||
167 fid_seq_is_norm(seq) || fid_seq_is_echo(seq))))
168 GOTO(out, rc = -EPROTO);
170 rc = ostid_to_fid(&tti->tti_fid1, oi,
171 tsi->tsi_tgt->lut_lsd.lsd_osd_index);
172 if (unlikely(rc != 0))
175 oi->oi_fid = tti->tti_fid1;
181 CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
182 tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
186 EXPORT_SYMBOL(tgt_validate_obdo);
188 static int tgt_io_data_unpack(struct tgt_session_info *tsi, struct ost_id *oi)
191 struct niobuf_remote *rnb;
192 struct obd_ioobj *ioo;
197 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
201 rnb = req_capsule_client_get(tsi->tsi_pill, &RMF_NIOBUF_REMOTE);
205 max_brw = ioobj_max_brw_get(ioo);
206 if (unlikely((max_brw & (max_brw - 1)) != 0)) {
207 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
208 ": rc = %d\n", tgt_name(tsi->tsi_tgt),
209 obd_export_nid2str(tsi->tsi_exp), max_brw,
210 POSTID(oi), -EPROTO);
215 obj_count = req_capsule_get_size(tsi->tsi_pill, &RMF_OBD_IOOBJ,
216 RCL_CLIENT) / sizeof(*ioo);
217 if (obj_count == 0) {
218 CERROR("%s: short ioobj\n", tgt_name(tsi->tsi_tgt));
220 } else if (obj_count > 1) {
221 CERROR("%s: too many ioobjs (%d)\n", tgt_name(tsi->tsi_tgt),
226 if (ioo->ioo_bufcnt == 0) {
227 CERROR("%s: ioo has zero bufcnt\n", tgt_name(tsi->tsi_tgt));
231 if (ioo->ioo_bufcnt > PTLRPC_MAX_BRW_PAGES) {
232 DEBUG_REQ(D_RPCTRACE, tgt_ses_req(tsi),
233 "bulk has too many pages (%d)",
241 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
243 struct ost_body *body;
244 struct req_capsule *pill = tsi->tsi_pill;
245 struct lu_nodemap *nodemap;
250 body = req_capsule_client_get(pill, &RMF_OST_BODY);
254 rc = tgt_validate_obdo(tsi, &body->oa);
258 nodemap = nodemap_get_from_exp(tsi->tsi_exp);
260 RETURN(PTR_ERR(nodemap));
262 body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
263 NODEMAP_CLIENT_TO_FS,
265 body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
266 NODEMAP_CLIENT_TO_FS,
268 nodemap_putref(nodemap);
270 tsi->tsi_ost_body = body;
271 tsi->tsi_fid = body->oa.o_oi.oi_fid;
273 if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
274 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
279 if (!(body->oa.o_valid & OBD_MD_FLID)) {
280 if (flags & HABEO_CORPUS) {
281 CERROR("%s: OBD_MD_FLID flag is not set in ost_body "
282 "but OID/FID is mandatory with HABEO_CORPUS\n",
283 tgt_name(tsi->tsi_tgt));
290 ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
293 * OST doesn't get object in advance for further use to prevent
294 * situations with nested object_find which is potential deadlock.
296 tsi->tsi_corpus = NULL;
301 * Do necessary preprocessing according to handler ->th_flags.
303 static int tgt_request_preprocess(struct tgt_session_info *tsi,
304 struct tgt_handler *h,
305 struct ptlrpc_request *req)
307 struct req_capsule *pill = tsi->tsi_pill;
308 __u32 flags = h->th_flags;
313 if (tsi->tsi_preprocessed)
316 LASSERT(h->th_act != NULL);
317 LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
318 LASSERT(current->journal_info == NULL);
320 LASSERT(ergo(flags & (HABEO_CORPUS | HABEO_REFERO),
322 if (h->th_fmt != NULL) {
323 req_capsule_set(pill, h->th_fmt);
324 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
325 rc = tgt_mdt_body_unpack(tsi, flags);
328 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
330 rc = tgt_ost_body_unpack(tsi, flags);
336 if (flags & MUTABOR && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
339 if (flags & HABEO_CLAVIS) {
340 struct ldlm_request *dlm_req;
342 LASSERT(h->th_fmt != NULL);
344 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
345 if (dlm_req != NULL) {
346 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
348 dlm_req->lock_desc.l_policy_data.\
349 l_inodebits.bits == 0)) {
351 * Lock without inodebits makes no sense and
352 * will oops later in ldlm. If client miss to
353 * set such bits, do not trigger ASSERTION.
355 * For liblustre flock case, it maybe zero.
359 tsi->tsi_dlm_req = dlm_req;
365 tsi->tsi_preprocessed = 1;
370 * Invoke handler for this request opc. Also do necessary preprocessing
371 * (according to handler ->th_flags), and post-processing (setting of
372 * ->last_{xid,committed}).
374 static int tgt_handle_request0(struct tgt_session_info *tsi,
375 struct tgt_handler *h,
376 struct ptlrpc_request *req)
380 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
385 /* When dealing with sec context requests, no export is associated yet,
386 * because these requests are sent before *_CONNECT requests.
387 * A NULL req->rq_export means the normal *_common_slice handlers will
388 * not be called, because there is no reference to the target.
389 * So deal with them by hand and jump directly to target_send_reply().
393 case SEC_CTX_INIT_CONT:
395 CFS_FAIL_TIMEOUT(OBD_FAIL_SEC_CTX_HDL_PAUSE, cfs_fail_val);
400 * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
401 * to put same checks into handlers like mdt_close(), mdt_reint(),
402 * etc., without talking to mdt authors first. Checking same thing
403 * there again is useless and returning 0 error without packing reply
404 * is buggy! Handlers either pack reply or return error.
406 * We return 0 here and do not send any reply in order to emulate
407 * network failure. Do not send any reply in case any of NET related
408 * fail_id has occured.
410 if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
412 if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
413 OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET)))
416 rc = tgt_request_preprocess(tsi, h, req);
417 /* pack reply if reply format is fixed */
418 if (rc == 0 && h->th_flags & HABEO_REFERO) {
420 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
422 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
424 tsi->tsi_mdt_body->mbo_eadatasize);
425 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
427 req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
429 if (req_capsule_has_field(tsi->tsi_pill, &RMF_ACL, RCL_SERVER))
430 req_capsule_set_size(tsi->tsi_pill,
431 &RMF_ACL, RCL_SERVER,
432 LUSTRE_POSIX_ACL_MAX_SIZE_OLD);
434 rc = req_capsule_server_pack(tsi->tsi_pill);
437 if (likely(rc == 0)) {
439 * Process request, there can be two types of rc:
440 * 1) errors with msg unpack/pack, other failures outside the
441 * operation itself. This is counted as serious errors;
442 * 2) errors during fs operation, should be placed in rq_status
446 if (!is_serious(rc) &&
447 !req->rq_no_reply && req->rq_reply_state == NULL) {
448 DEBUG_REQ(D_ERROR, req, "%s \"handler\" %s did not "
449 "pack reply and returned 0 error\n",
450 tgt_name(tsi->tsi_tgt), h->th_name);
453 serious = is_serious(rc);
454 rc = clear_serious(rc);
462 * ELDLM_* codes which > 0 should be in rq_status only as well as
463 * all non-serious errors.
465 if (rc > 0 || !serious)
468 LASSERT(current->journal_info == NULL);
470 if (likely(rc == 0 && req->rq_export))
471 target_committed_to_req(req);
474 target_send_reply(req, rc, tsi->tsi_reply_fail_id);
478 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
479 struct obd_device *obd, int *process)
481 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
488 case MDS_SYNC: /* used in unmounting */
502 case MDS_HSM_PROGRESS:
503 case MDS_HSM_STATE_SET:
504 case MDS_HSM_REQUEST:
505 *process = target_queue_recovery_request(req, obd);
509 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
516 * Handle recovery. Return:
517 * +1: continue request processing;
518 * -ve: abort immediately with the given error code;
519 * 0: send reply with error code in req->rq_status;
521 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
525 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
530 case SEC_CTX_INIT_CONT:
535 if (!req->rq_export->exp_obd->obd_replayable)
538 /* sanity check: if the xid matches, the request must be marked as a
539 * resent or replayed */
540 if (req_can_reconstruct(req, NULL)) {
541 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
542 (MSG_RESENT | MSG_REPLAY))) {
543 DEBUG_REQ(D_WARNING, req, "rq_xid %llu matches "
544 "saved xid, expected REPLAY or RESENT flag "
546 lustre_msg_get_flags(req->rq_reqmsg));
547 req->rq_status = -ENOTCONN;
551 /* else: note the opposite is not always true; a RESENT req after a
552 * failover will usually not match the last_xid, since it was likely
553 * never committed. A REPLAYed request will almost never match the
554 * last xid, however it could for a committed, but still retained,
557 /* Check for aborted recovery... */
558 if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
562 DEBUG_REQ(D_INFO, req, "Got new replay");
563 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
565 if (rc != 0 || !should_process)
567 else if (should_process < 0) {
568 req->rq_status = should_process;
569 rc = ptlrpc_error(req);
576 /* Initial check for request, it is validation mostly */
577 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
579 struct tgt_handler *h;
580 struct tgt_opc_slice *s;
581 struct lu_target *tgt;
582 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
586 tgt = class_exp2tgt(req->rq_export);
587 if (unlikely(tgt == NULL)) {
588 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
589 class_exp2obd(req->rq_export)->obd_name);
590 RETURN(ERR_PTR(-EINVAL));
593 for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
594 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
597 /* opcode was not found in slice */
598 if (unlikely(s->tos_hs == NULL)) {
599 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
601 RETURN(ERR_PTR(-ENOTSUPP));
604 LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
605 h = s->tos_hs + (opc - s->tos_opc_start);
606 if (unlikely(h->th_opc == 0)) {
607 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
608 RETURN(ERR_PTR(-ENOTSUPP));
614 static int process_req_last_xid(struct ptlrpc_request *req)
619 /* check request's xid is consistent with export's last_xid */
620 last_xid = lustre_msg_get_last_xid(req->rq_reqmsg);
621 if (last_xid > req->rq_export->exp_last_xid)
622 req->rq_export->exp_last_xid = last_xid;
624 if (req->rq_xid == 0 ||
625 (req->rq_xid <= req->rq_export->exp_last_xid)) {
626 DEBUG_REQ(D_ERROR, req, "Unexpected xid %llx vs. "
627 "last_xid %llx\n", req->rq_xid,
628 req->rq_export->exp_last_xid);
629 /* Some request is allowed to be sent during replay,
630 * such as OUT update requests, FLD requests, so it
631 * is possible that replay requests has smaller XID
632 * than the exp_last_xid.
634 * Some non-replay requests may have smaller XID as
637 * - Client send a no_resend RPC, like statfs;
638 * - The RPC timedout (or some other error) on client,
639 * then it's removed from the unreplied list;
640 * - Client send some other request to bump the
641 * exp_last_xid on server;
642 * - The former RPC got chance to be processed;
644 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
648 /* try to release in-memory reply data */
649 if (tgt_is_multimodrpcs_client(req->rq_export)) {
650 tgt_handle_received_xid(req->rq_export,
651 lustre_msg_get_last_xid(req->rq_reqmsg));
652 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
653 (MSG_RESENT | MSG_REPLAY)))
654 tgt_handle_tag(req->rq_export,
655 lustre_msg_get_tag(req->rq_reqmsg));
660 int tgt_request_handle(struct ptlrpc_request *req)
662 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
664 struct lustre_msg *msg = req->rq_reqmsg;
665 struct tgt_handler *h;
666 struct lu_target *tgt;
667 int request_fail_id = 0;
668 __u32 opc = lustre_msg_get_opc(msg);
669 struct obd_device *obd;
671 bool is_connect = false;
674 /* Refill the context, to make sure all thread keys are allocated */
675 lu_env_refill(req->rq_svc_thread->t_env);
677 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
678 tsi->tsi_pill = &req->rq_pill;
679 tsi->tsi_env = req->rq_svc_thread->t_env;
681 /* if request has export then get handlers slice from corresponding
682 * target, otherwise that should be connect operation */
683 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
684 opc == MGS_CONNECT) {
686 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
687 rc = target_handle_connect(req);
689 rc = ptlrpc_error(req);
692 /* recovery-small test 18c asks to drop connect reply */
693 if (unlikely(opc == OST_CONNECT &&
694 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
698 if (unlikely(!class_connected_export(req->rq_export))) {
699 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
700 opc == SEC_CTX_FINI) {
701 /* sec context initialization has to be handled
702 * by hand in tgt_handle_request0() */
703 tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
705 GOTO(handle_recov, rc = 0);
707 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
708 opc, libcfs_id2str(req->rq_peer));
709 req->rq_status = -ENOTCONN;
710 rc = ptlrpc_error(req);
714 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
715 tsi->tsi_exp = req->rq_export;
716 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
717 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
719 tsi->tsi_jobid = NULL;
722 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
723 class_exp2obd(req->rq_export)->obd_name);
724 req->rq_status = -EINVAL;
725 rc = ptlrpc_error(req);
729 /* Skip last_xid processing for the recovery thread, otherwise, the
730 * last_xid on same request could be processed twice: first time when
731 * processing the incoming request, second time when the request is
732 * being processed by recovery thread. */
733 obd = class_exp2obd(req->rq_export);
735 /* reset the exp_last_xid on each connection. */
736 req->rq_export->exp_last_xid = 0;
737 } else if (obd->obd_recovery_data.trd_processing_task !=
739 rc = process_req_last_xid(req);
742 rc = ptlrpc_error(req);
747 request_fail_id = tgt->lut_request_fail_id;
748 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
750 h = tgt_handler_find_check(req);
752 req->rq_status = PTR_ERR(h);
753 rc = ptlrpc_error(req);
757 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
760 if (CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
763 rc = lustre_msg_check_version(msg, h->th_version);
765 DEBUG_REQ(D_ERROR, req, "%s: drop mal-formed request, version"
766 " %08x, expecting %08x\n", tgt_name(tgt),
767 lustre_msg_get_version(msg), h->th_version);
768 req->rq_status = -EINVAL;
769 rc = ptlrpc_error(req);
774 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
775 if (likely(rc == 1)) {
776 rc = tgt_handle_request0(tsi, h, req);
782 req_capsule_fini(tsi->tsi_pill);
783 if (tsi->tsi_corpus != NULL) {
784 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
785 tsi->tsi_corpus = NULL;
789 EXPORT_SYMBOL(tgt_request_handle);
791 /** Assign high priority operations to the request if needed. */
792 int tgt_hpreq_handler(struct ptlrpc_request *req)
794 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
795 struct tgt_handler *h;
800 if (req->rq_export == NULL)
803 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
804 tsi->tsi_pill = &req->rq_pill;
805 tsi->tsi_env = req->rq_svc_thread->t_env;
806 tsi->tsi_tgt = class_exp2tgt(req->rq_export);
807 tsi->tsi_exp = req->rq_export;
809 h = tgt_handler_find_check(req);
815 rc = tgt_request_preprocess(tsi, h, req);
816 if (unlikely(rc != 0))
819 if (h->th_hp != NULL)
823 EXPORT_SYMBOL(tgt_hpreq_handler);
825 void tgt_counter_incr(struct obd_export *exp, int opcode)
827 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
828 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
829 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
831 EXPORT_SYMBOL(tgt_counter_incr);
834 * Unified target generic handlers.
837 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
839 struct lu_target *tgt = class_exp2tgt(exp);
840 struct sptlrpc_flavor flvr;
844 LASSERT(tgt->lut_obd);
845 LASSERT(tgt->lut_slice);
847 /* always allow ECHO client */
848 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
849 LUSTRE_ECHO_NAME) == 0)) {
850 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
854 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
855 read_lock(&tgt->lut_sptlrpc_lock);
856 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
860 read_unlock(&tgt->lut_sptlrpc_lock);
862 spin_lock(&exp->exp_lock);
863 exp->exp_sp_peer = req->rq_sp_from;
864 exp->exp_flvr = flvr;
866 /* when on mgs, if no restriction is set, or if client
867 * is loopback, allow any flavor */
868 if ((strcmp(exp->exp_obd->obd_type->typ_name,
869 LUSTRE_MGS_NAME) == 0) &&
870 (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_NULL ||
871 LNET_NETTYP(LNET_NIDNET(exp->exp_connection->c_peer.nid))
873 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
875 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
876 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
877 CERROR("%s: unauthorized rpc flavor %x from %s, "
878 "expect %x\n", tgt_name(tgt),
880 libcfs_nid2str(req->rq_peer.nid),
881 exp->exp_flvr.sf_rpc);
884 spin_unlock(&exp->exp_lock);
886 if (exp->exp_sp_peer != req->rq_sp_from) {
887 CERROR("%s: RPC source %s doesn't match %s\n",
889 sptlrpc_part2name(req->rq_sp_from),
890 sptlrpc_part2name(exp->exp_sp_peer));
893 rc = sptlrpc_target_export_check(exp, req);
900 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt)
902 struct sptlrpc_rule_set tmp_rset;
905 if (unlikely(tgt == NULL)) {
906 CERROR("No target passed");
910 sptlrpc_rule_set_init(&tmp_rset);
911 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset);
913 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
918 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
920 write_lock(&tgt->lut_sptlrpc_lock);
921 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
922 tgt->lut_sptlrpc_rset = tmp_rset;
923 write_unlock(&tgt->lut_sptlrpc_lock);
927 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
929 int tgt_connect(struct tgt_session_info *tsi)
931 struct ptlrpc_request *req = tgt_ses_req(tsi);
932 struct obd_connect_data *reply;
937 /* XXX: better to call this check right after getting new export but
938 * before last_rcvd slot allocation to avoid server load upon insecure
939 * connects. This is to be fixed after unifiyng all targets.
941 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
945 /* To avoid exposing partially initialized connection flags, changes up
946 * to this point have been staged in reply->ocd_connect_flags. Now that
947 * connection handling has completed successfully, atomically update
948 * the connect flags in the shared export data structure. LU-1623 */
949 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
950 spin_lock(&tsi->tsi_exp->exp_lock);
951 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
952 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
953 spin_unlock(&tsi->tsi_exp->exp_lock);
957 obd_disconnect(class_export_get(tsi->tsi_exp));
960 EXPORT_SYMBOL(tgt_connect);
962 int tgt_disconnect(struct tgt_session_info *tsi)
968 rc = target_handle_disconnect(tgt_ses_req(tsi));
970 RETURN(err_serious(rc));
974 EXPORT_SYMBOL(tgt_disconnect);
977 * Unified target OBD handlers
979 int tgt_obd_ping(struct tgt_session_info *tsi)
985 rc = target_handle_ping(tgt_ses_req(tsi));
987 RETURN(err_serious(rc));
991 EXPORT_SYMBOL(tgt_obd_ping);
993 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
995 return err_serious(-EOPNOTSUPP);
998 int tgt_send_buffer(struct tgt_session_info *tsi, struct lu_rdbuf *rdbuf)
1000 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1001 struct ptlrpc_request *req = tgt_ses_req(tsi);
1002 struct obd_export *exp = req->rq_export;
1003 struct ptlrpc_bulk_desc *desc;
1004 struct l_wait_info *lwi = &tti->tti_u.update.tti_wait_info;
1010 desc = ptlrpc_prep_bulk_exp(req, rdbuf->rb_nbufs, 1,
1011 PTLRPC_BULK_PUT_SOURCE | PTLRPC_BULK_BUF_KVEC,
1012 MDS_BULK_PORTAL, &ptlrpc_bulk_kvec_ops);
1016 for (i = 0; i < rdbuf->rb_nbufs; i++)
1017 desc->bd_frag_ops->add_iov_frag(desc,
1018 rdbuf->rb_bufs[i].lb_buf,
1019 rdbuf->rb_bufs[i].lb_len);
1021 rc = target_bulk_io(exp, desc, lwi);
1022 ptlrpc_free_bulk(desc);
1025 EXPORT_SYMBOL(tgt_send_buffer);
1027 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1029 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1030 struct ptlrpc_request *req = tgt_ses_req(tsi);
1031 struct obd_export *exp = req->rq_export;
1032 struct ptlrpc_bulk_desc *desc;
1033 struct l_wait_info *lwi = &tti->tti_u.rdpg.tti_wait_info;
1041 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1042 PTLRPC_BULK_PUT_SOURCE |
1043 PTLRPC_BULK_BUF_KIOV,
1045 &ptlrpc_bulk_kiov_pin_ops);
1049 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1050 /* old client requires reply size in it's PAGE_SIZE,
1051 * which is rdpg->rp_count */
1052 nob = rdpg->rp_count;
1054 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1055 i++, tmpcount -= tmpsize) {
1056 tmpsize = min_t(int, tmpcount, PAGE_SIZE);
1057 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1061 LASSERT(desc->bd_nob == nob);
1062 rc = target_bulk_io(exp, desc, lwi);
1063 ptlrpc_free_bulk(desc);
1066 EXPORT_SYMBOL(tgt_sendpage);
1069 * OBD_IDX_READ handler
1071 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1073 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1074 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1075 struct idx_info *req_ii, *rep_ii;
1080 memset(rdpg, 0, sizeof(*rdpg));
1081 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1083 /* extract idx_info buffer from request & reply */
1084 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1085 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1086 RETURN(err_serious(-EPROTO));
1088 rc = req_capsule_server_pack(tsi->tsi_pill);
1090 RETURN(err_serious(rc));
1092 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1094 RETURN(err_serious(-EFAULT));
1095 rep_ii->ii_magic = IDX_INFO_MAGIC;
1097 /* extract hash to start with */
1098 rdpg->rp_hash = req_ii->ii_hash_start;
1100 /* extract requested attributes */
1101 rdpg->rp_attrs = req_ii->ii_attrs;
1103 /* check that fid packed in request is valid and supported */
1104 if (!fid_is_sane(&req_ii->ii_fid))
1106 rep_ii->ii_fid = req_ii->ii_fid;
1109 rep_ii->ii_flags = req_ii->ii_flags;
1111 /* compute number of pages to allocate, ii_count is the number of 4KB
1113 if (req_ii->ii_count <= 0)
1114 GOTO(out, rc = -EFAULT);
1115 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1116 exp_max_brw_size(tsi->tsi_exp));
1117 rdpg->rp_npages = (rdpg->rp_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1119 /* allocate pages to store the containers */
1120 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1121 if (rdpg->rp_pages == NULL)
1122 GOTO(out, rc = -ENOMEM);
1123 for (i = 0; i < rdpg->rp_npages; i++) {
1124 rdpg->rp_pages[i] = alloc_page(GFP_NOFS);
1125 if (rdpg->rp_pages[i] == NULL)
1126 GOTO(out, rc = -ENOMEM);
1129 /* populate pages with key/record pairs */
1130 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1134 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1135 "asked %d > %d\n", rc, rdpg->rp_count);
1137 /* send pages to client */
1138 rc = tgt_sendpage(tsi, rdpg, rc);
1143 if (rdpg->rp_pages) {
1144 for (i = 0; i < rdpg->rp_npages; i++)
1145 if (rdpg->rp_pages[i])
1146 __free_page(rdpg->rp_pages[i]);
1147 OBD_FREE(rdpg->rp_pages,
1148 rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1153 struct tgt_handler tgt_obd_handlers[] = {
1154 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1155 TGT_OBD_HDL_VAR(0, OBD_LOG_CANCEL, tgt_obd_log_cancel),
1156 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1158 EXPORT_SYMBOL(tgt_obd_handlers);
1160 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1161 struct dt_object *obj, __u64 start, __u64 end)
1167 /* if no objid is specified, it means "sync whole filesystem" */
1169 rc = dt_sync(env, tgt->lut_bottom);
1170 } else if (dt_version_get(env, obj) >
1171 tgt->lut_obd->obd_last_committed) {
1172 rc = dt_object_sync(env, obj, start, end);
1174 atomic_inc(&tgt->lut_sync_count);
1178 EXPORT_SYMBOL(tgt_sync);
1180 * Unified target DLM handlers.
1184 * Unified target BAST
1186 * Ensure data and metadata are synced to disk when lock is canceled if Sync on
1187 * Cancel (SOC) is enabled. If it's extent lock, normally sync obj is enough,
1188 * but if it's cross-MDT lock, because remote object version is not set, a
1189 * filesystem sync is needed.
1191 * \param lock server side lock
1192 * \param desc lock desc
1193 * \param data ldlm_cb_set_arg
1194 * \param flag indicates whether this cancelling or blocking callback
1195 * \retval 0 on success
1196 * \retval negative number on error
1198 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1199 void *data, int flag)
1202 struct lu_target *tgt;
1203 struct dt_object *obj = NULL;
1209 tgt = class_exp2tgt(lock->l_export);
1211 if (unlikely(tgt == NULL)) {
1212 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1213 class_exp2obd(lock->l_export)->obd_name);
1217 if (flag == LDLM_CB_CANCELING &&
1218 (lock->l_granted_mode & (LCK_EX | LCK_PW | LCK_GROUP)) &&
1219 (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1220 (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1221 ldlm_is_cbpending(lock))) &&
1222 ((exp_connect_flags(lock->l_export) & OBD_CONNECT_MDS_MDS) ||
1223 lock->l_resource->lr_type == LDLM_EXTENT)) {
1225 __u64 end = OBD_OBJECT_EOF;
1227 rc = lu_env_init(&env, LCT_DT_THREAD);
1228 if (unlikely(rc != 0))
1231 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1232 tgt->lut_lsd.lsd_osd_index);
1234 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1235 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1237 GOTO(err_env, rc = PTR_ERR(obj));
1239 if (!dt_object_exists(obj))
1240 GOTO(err_put, rc = -ENOENT);
1242 start = lock->l_policy_data.l_extent.start;
1243 end = lock->l_policy_data.l_extent.end;
1246 rc = tgt_sync(&env, tgt, obj, start, end);
1248 CERROR("%s: syncing "DFID" (%llu-%llu) on lock "
1249 "cancel: rc = %d\n",
1250 tgt_name(tgt), PFID(&fid),
1251 lock->l_policy_data.l_extent.start,
1252 lock->l_policy_data.l_extent.end, rc);
1256 dt_object_put(&env, obj);
1261 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1265 static struct ldlm_callback_suite tgt_dlm_cbs = {
1266 .lcs_completion = ldlm_server_completion_ast,
1267 .lcs_blocking = tgt_blocking_ast,
1268 .lcs_glimpse = ldlm_server_glimpse_ast
1271 int tgt_enqueue(struct tgt_session_info *tsi)
1273 struct ptlrpc_request *req = tgt_ses_req(tsi);
1278 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1279 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1281 LASSERT(tsi->tsi_dlm_req != NULL);
1282 rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1283 tsi->tsi_dlm_req, &tgt_dlm_cbs);
1285 RETURN(err_serious(rc));
1287 switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1289 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1292 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1295 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1298 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1301 RETURN(req->rq_status);
1303 EXPORT_SYMBOL(tgt_enqueue);
1305 int tgt_convert(struct tgt_session_info *tsi)
1307 struct ptlrpc_request *req = tgt_ses_req(tsi);
1311 LASSERT(tsi->tsi_dlm_req);
1312 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1314 RETURN(err_serious(rc));
1316 RETURN(req->rq_status);
1319 int tgt_bl_callback(struct tgt_session_info *tsi)
1321 return err_serious(-EOPNOTSUPP);
1324 int tgt_cp_callback(struct tgt_session_info *tsi)
1326 return err_serious(-EOPNOTSUPP);
1329 /* generic LDLM target handler */
1330 struct tgt_handler tgt_dlm_handlers[] = {
1331 TGT_DLM_HDL (HABEO_CLAVIS, LDLM_ENQUEUE, tgt_enqueue),
1332 TGT_DLM_HDL_VAR(HABEO_CLAVIS, LDLM_CONVERT, tgt_convert),
1333 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1334 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1336 EXPORT_SYMBOL(tgt_dlm_handlers);
1339 * Unified target LLOG handlers.
1341 int tgt_llog_open(struct tgt_session_info *tsi)
1347 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1351 EXPORT_SYMBOL(tgt_llog_open);
1353 int tgt_llog_close(struct tgt_session_info *tsi)
1359 rc = llog_origin_handle_close(tgt_ses_req(tsi));
1363 EXPORT_SYMBOL(tgt_llog_close);
1366 int tgt_llog_destroy(struct tgt_session_info *tsi)
1372 rc = llog_origin_handle_destroy(tgt_ses_req(tsi));
1377 int tgt_llog_read_header(struct tgt_session_info *tsi)
1383 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1387 EXPORT_SYMBOL(tgt_llog_read_header);
1389 int tgt_llog_next_block(struct tgt_session_info *tsi)
1395 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1399 EXPORT_SYMBOL(tgt_llog_next_block);
1401 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1407 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1411 EXPORT_SYMBOL(tgt_llog_prev_block);
1413 /* generic llog target handler */
1414 struct tgt_handler tgt_llog_handlers[] = {
1415 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1416 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1417 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1418 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1419 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_DESTROY, tgt_llog_destroy),
1420 TGT_LLOG_HDL_VAR(0, LLOG_ORIGIN_HANDLE_CLOSE, tgt_llog_close),
1422 EXPORT_SYMBOL(tgt_llog_handlers);
1425 * sec context handlers
1427 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1428 static int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1433 struct tgt_handler tgt_sec_ctx_handlers[] = {
1434 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle),
1435 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle),
1436 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle),
1438 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1440 int (*tgt_lfsck_in_notify_local)(const struct lu_env *env,
1441 struct dt_device *key,
1442 struct lfsck_req_local *lrl,
1443 struct thandle *th) = NULL;
1445 void tgt_register_lfsck_in_notify_local(int (*notify)(const struct lu_env *,
1447 struct lfsck_req_local *,
1450 tgt_lfsck_in_notify_local = notify;
1452 EXPORT_SYMBOL(tgt_register_lfsck_in_notify_local);
1454 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1455 struct dt_device *key,
1456 struct lfsck_request *lr) = NULL;
1458 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1460 struct lfsck_request *))
1462 tgt_lfsck_in_notify = notify;
1464 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1466 static int (*tgt_lfsck_query)(const struct lu_env *env,
1467 struct dt_device *key,
1468 struct lfsck_request *req,
1469 struct lfsck_reply *rep,
1470 struct lfsck_query *que) = NULL;
1472 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1474 struct lfsck_request *,
1475 struct lfsck_reply *,
1476 struct lfsck_query *))
1478 tgt_lfsck_query = query;
1480 EXPORT_SYMBOL(tgt_register_lfsck_query);
1482 /* LFSCK request handlers */
1483 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1485 const struct lu_env *env = tsi->tsi_env;
1486 struct dt_device *key = tsi->tsi_tgt->lut_bottom;
1487 struct lfsck_request *lr;
1491 lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1495 rc = tgt_lfsck_in_notify(env, key, lr);
1500 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1502 struct lfsck_request *request;
1503 struct lfsck_reply *reply;
1507 request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1508 if (request == NULL)
1511 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1515 rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom,
1516 request, reply, NULL);
1518 RETURN(rc < 0 ? rc : 0);
1521 struct tgt_handler tgt_lfsck_handlers[] = {
1522 TGT_LFSCK_HDL(HABEO_REFERO, LFSCK_NOTIFY, tgt_handle_lfsck_notify),
1523 TGT_LFSCK_HDL(HABEO_REFERO, LFSCK_QUERY, tgt_handle_lfsck_query),
1525 EXPORT_SYMBOL(tgt_lfsck_handlers);
1528 * initialize per-thread page pool (bug 5137).
1530 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1532 struct tgt_thread_big_cache *tbc;
1536 LASSERT(thread != NULL);
1537 LASSERT(thread->t_data == NULL);
1539 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1542 thread->t_data = tbc;
1545 EXPORT_SYMBOL(tgt_io_thread_init);
1548 * free per-thread pool created by tgt_thread_init().
1550 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1552 struct tgt_thread_big_cache *tbc;
1556 LASSERT(thread != NULL);
1559 * be prepared to handle partially-initialized pools (because this is
1560 * called from ost_io_thread_init() for cleanup.
1562 tbc = thread->t_data;
1564 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1565 thread->t_data = NULL;
1569 EXPORT_SYMBOL(tgt_io_thread_done);
1571 * Helper function for getting server side [start, start+count] DLM lock
1572 * if asked by client.
1574 int tgt_extent_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1575 __u64 start, __u64 end, struct lustre_handle *lh,
1576 int mode, __u64 *flags)
1578 union ldlm_policy_data policy;
1583 LASSERT(lh != NULL);
1584 LASSERT(ns != NULL);
1585 LASSERT(!lustre_handle_is_used(lh));
1587 policy.l_extent.gid = 0;
1588 policy.l_extent.start = start & PAGE_MASK;
1591 * If ->o_blocks is EOF it means "lock till the end of the file".
1592 * Otherwise, it's size of an extent or hole being punched (in bytes).
1594 if (end == OBD_OBJECT_EOF || end < start)
1595 policy.l_extent.end = OBD_OBJECT_EOF;
1597 policy.l_extent.end = end | ~PAGE_MASK;
1599 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_EXTENT, &policy, mode,
1600 flags, ldlm_blocking_ast,
1601 ldlm_completion_ast, ldlm_glimpse_ast,
1602 NULL, 0, LVB_T_NONE, NULL, lh);
1603 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1605 EXPORT_SYMBOL(tgt_extent_lock);
1607 void tgt_extent_unlock(struct lustre_handle *lh, enum ldlm_mode mode)
1609 LASSERT(lustre_handle_is_used(lh));
1610 ldlm_lock_decref(lh, mode);
1612 EXPORT_SYMBOL(tgt_extent_unlock);
1614 int tgt_brw_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1615 struct obd_ioobj *obj, struct niobuf_remote *nb,
1616 struct lustre_handle *lh, enum ldlm_mode mode)
1619 int nrbufs = obj->ioo_bufcnt;
1624 LASSERT(mode == LCK_PR || mode == LCK_PW);
1625 LASSERT(!lustre_handle_is_used(lh));
1627 if (ns->ns_obd->obd_recovering)
1630 if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1633 for (i = 1; i < nrbufs; i++)
1634 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1637 RETURN(tgt_extent_lock(ns, res_id, nb[0].rnb_offset,
1638 nb[nrbufs - 1].rnb_offset +
1639 nb[nrbufs - 1].rnb_len - 1,
1643 void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1644 struct lustre_handle *lh, enum ldlm_mode mode)
1648 LASSERT(mode == LCK_PR || mode == LCK_PW);
1649 LASSERT((obj->ioo_bufcnt > 0 &&
1650 (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1651 lustre_handle_is_used(lh));
1653 if (lustre_handle_is_used(lh))
1654 tgt_extent_unlock(lh, mode);
1658 static __u32 tgt_checksum_bulk(struct lu_target *tgt,
1659 struct ptlrpc_bulk_desc *desc, int opc,
1660 cksum_type_t cksum_type)
1662 struct cfs_crypto_hash_desc *hdesc;
1663 unsigned int bufsize;
1665 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1668 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
1670 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1671 if (IS_ERR(hdesc)) {
1672 CERROR("%s: unable to initialize checksum hash %s\n",
1673 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1674 return PTR_ERR(hdesc);
1677 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1678 for (i = 0; i < desc->bd_iov_count; i++) {
1679 /* corrupt the data before we compute the checksum, to
1680 * simulate a client->OST data error */
1681 if (i == 0 && opc == OST_WRITE &&
1682 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1683 int off = BD_GET_KIOV(desc, i).kiov_offset &
1685 int len = BD_GET_KIOV(desc, i).kiov_len;
1686 struct page *np = tgt_page_to_corrupt;
1687 char *ptr = kmap(BD_GET_KIOV(desc, i).kiov_page) + off;
1690 char *ptr2 = kmap(np) + off;
1692 memcpy(ptr2, ptr, len);
1693 memcpy(ptr2, "bad3", min(4, len));
1696 /* LU-8376 to preserve original index for
1697 * display in dump_all_bulk_pages() */
1698 np->index = BD_GET_KIOV(desc,
1699 i).kiov_page->index;
1701 BD_GET_KIOV(desc, i).kiov_page = np;
1703 CERROR("%s: can't alloc page for corruption\n",
1707 cfs_crypto_hash_update_page(hdesc,
1708 BD_GET_KIOV(desc, i).kiov_page,
1709 BD_GET_KIOV(desc, i).kiov_offset &
1711 BD_GET_KIOV(desc, i).kiov_len);
1713 /* corrupt the data after we compute the checksum, to
1714 * simulate an OST->client data error */
1715 if (i == 0 && opc == OST_READ &&
1716 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1717 int off = BD_GET_KIOV(desc, i).kiov_offset
1719 int len = BD_GET_KIOV(desc, i).kiov_len;
1720 struct page *np = tgt_page_to_corrupt;
1722 kmap(BD_GET_KIOV(desc, i).kiov_page) + off;
1725 char *ptr2 = kmap(np) + off;
1727 memcpy(ptr2, ptr, len);
1728 memcpy(ptr2, "bad4", min(4, len));
1731 /* LU-8376 to preserve original index for
1732 * display in dump_all_bulk_pages() */
1733 np->index = BD_GET_KIOV(desc,
1734 i).kiov_page->index;
1736 BD_GET_KIOV(desc, i).kiov_page = np;
1738 CERROR("%s: can't alloc page for corruption\n",
1744 bufsize = sizeof(cksum);
1745 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1750 char dbgcksum_file_name[PATH_MAX];
1752 static void dump_all_bulk_pages(struct obdo *oa, int count,
1753 lnet_kiov_t *iov, __u32 server_cksum,
1762 /* will only keep dump of pages on first error for the same range in
1763 * file/fid, not during the resends/retries. */
1764 snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1765 "%s-checksum_dump-ost-"DFID":[%llu-%llu]-%x-%x",
1766 (strncmp(libcfs_debug_file_path_arr, "NONE", 4) != 0 ?
1767 libcfs_debug_file_path_arr :
1768 LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1769 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1770 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1771 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1772 (__u64)iov[0].kiov_page->index << PAGE_SHIFT,
1773 ((__u64)iov[count - 1].kiov_page->index << PAGE_SHIFT) +
1774 iov[count - 1].kiov_len - 1, client_cksum, server_cksum);
1775 filp = filp_open(dbgcksum_file_name,
1776 O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1780 CDEBUG(D_INFO, "%s: can't open to dump pages with "
1781 "checksum error: rc = %d\n", dbgcksum_file_name,
1784 CERROR("%s: can't open to dump pages with checksum "
1785 "error: rc = %d\n", dbgcksum_file_name, rc);
1791 for (i = 0; i < count; i++) {
1792 len = iov[i].kiov_len;
1793 buf = kmap(iov[i].kiov_page);
1795 rc = vfs_write(filp, (__force const char __user *)buf,
1798 CERROR("%s: wanted to write %u but got %d "
1799 "error\n", dbgcksum_file_name, len, rc);
1804 CDEBUG(D_INFO, "%s: wrote %d bytes\n",
1805 dbgcksum_file_name, rc);
1807 kunmap(iov[i].kiov_page);
1811 rc = ll_vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1813 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1814 filp_close(filp, NULL);
1818 static int check_read_checksum(struct ptlrpc_bulk_desc *desc, struct obdo *oa,
1819 const lnet_process_id_t *peer,
1820 __u32 client_cksum, __u32 server_cksum,
1821 cksum_type_t server_cksum_type)
1824 cksum_type_t cksum_type;
1826 /* unlikely to happen and only if resend does not occur due to cksum
1827 * control failure on Client */
1828 if (unlikely(server_cksum == client_cksum)) {
1829 CDEBUG(D_PAGE, "checksum %x confirmed upon retry\n",
1834 if (desc->bd_export->exp_obd->obd_checksum_dump)
1835 dump_all_bulk_pages(oa, desc->bd_iov_count,
1836 &BD_GET_KIOV(desc, 0), server_cksum,
1839 cksum_type = cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1842 if (cksum_type != server_cksum_type)
1843 msg = "the server may have not used the checksum type specified"
1844 " in the original request - likely a protocol problem";
1846 msg = "should have changed on the client or in transit";
1848 LCONSOLE_ERROR_MSG(0x132, "%s: BAD READ CHECKSUM: %s: from %s inode "
1849 DFID " object "DOSTID" extent [%llu-%llu], client returned csum"
1850 " %x (type %x), server csum %x (type %x)\n",
1851 desc->bd_export->exp_obd->obd_name,
1852 msg, libcfs_nid2str(peer->nid),
1853 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1854 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1855 oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1857 (__u64)BD_GET_KIOV(desc, 0).kiov_page->index << PAGE_SHIFT,
1858 ((__u64)BD_GET_KIOV(desc,
1859 desc->bd_iov_count - 1).kiov_page->index
1861 BD_GET_KIOV(desc, desc->bd_iov_count - 1).kiov_len - 1,
1862 client_cksum, cksum_type, server_cksum, server_cksum_type);
1866 int tgt_brw_read(struct tgt_session_info *tsi)
1868 struct ptlrpc_request *req = tgt_ses_req(tsi);
1869 struct ptlrpc_bulk_desc *desc = NULL;
1870 struct obd_export *exp = tsi->tsi_exp;
1871 struct niobuf_remote *remote_nb;
1872 struct niobuf_local *local_nb;
1873 struct obd_ioobj *ioo;
1874 struct ost_body *body, *repbody;
1875 struct l_wait_info lwi;
1876 struct lustre_handle lockh = { 0 };
1877 int npages, nob = 0, rc, i, no_reply = 0;
1878 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
1882 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) {
1883 CERROR("%s: deny read request from %s to portal %u\n",
1884 tgt_name(tsi->tsi_tgt),
1885 obd_export_nid2str(req->rq_export),
1886 ptlrpc_req2svc(req)->srv_req_portal);
1890 req->rq_bulk_read = 1;
1892 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
1895 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
1896 cfs_fail_val : (obd_timeout + 1) / 4);
1898 /* Check if there is eviction in progress, and if so, wait for it to
1900 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
1901 /* We do not care how long it takes */
1902 lwi = LWI_INTR(NULL, NULL);
1903 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
1904 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
1908 /* There must be big cache in current thread to process this request
1909 * if it is NULL then something went wrong and it wasn't allocated,
1910 * report -ENOMEM in that case */
1914 body = tsi->tsi_ost_body;
1915 LASSERT(body != NULL);
1917 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
1918 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
1920 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1921 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
1923 local_nb = tbc->local;
1925 rc = tgt_brw_lock(exp->exp_obd->obd_namespace, &tsi->tsi_resid, ioo,
1926 remote_nb, &lockh, LCK_PR);
1931 * If getting the lock took more time than
1932 * client was willing to wait, drop it. b=11330
1934 if (ktime_get_real_seconds() > req->rq_deadline ||
1935 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1937 CERROR("Dropping timed-out read from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
1938 libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
1939 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
1940 req->rq_deadline - req->rq_arrival_time.tv_sec);
1941 GOTO(out_lock, rc = -ETIMEDOUT);
1944 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1945 repbody->oa = body->oa;
1947 npages = PTLRPC_MAX_BRW_PAGES;
1948 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
1949 ioo, remote_nb, &npages, local_nb);
1953 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
1954 PTLRPC_BULK_PUT_SOURCE |
1955 PTLRPC_BULK_BUF_KIOV,
1957 &ptlrpc_bulk_kiov_nopin_ops);
1959 GOTO(out_commitrw, rc = -ENOMEM);
1962 for (i = 0; i < npages; i++) {
1963 int page_rc = local_nb[i].lnb_rc;
1971 if (page_rc != 0) { /* some data! */
1972 LASSERT(local_nb[i].lnb_page != NULL);
1973 desc->bd_frag_ops->add_kiov_frag
1974 (desc, local_nb[i].lnb_page,
1975 local_nb[i].lnb_page_offset,
1979 if (page_rc != local_nb[i].lnb_len) { /* short read */
1980 /* All subsequent pages should be 0 */
1981 while (++i < npages)
1982 LASSERT(local_nb[i].lnb_rc == 0);
1986 if (OBD_FAIL_CHECK(OBD_FAIL_OST_READ_SIZE) &&
1987 nob != cfs_fail_val)
1990 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1991 cksum_type_t cksum_type =
1992 cksum_type_unpack(body->oa.o_valid & OBD_MD_FLFLAGS ?
1993 body->oa.o_flags : 0);
1995 repbody->oa.o_flags = cksum_type_pack(cksum_type);
1996 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1997 repbody->oa.o_cksum = tgt_checksum_bulk(tsi->tsi_tgt, desc,
1998 OST_READ, cksum_type);
1999 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
2000 repbody->oa.o_cksum);
2002 /* if a resend it could be for a cksum error, so check Server
2003 * cksum with returned Client cksum (this should even cover
2004 * zero-cksum case) */
2005 if ((body->oa.o_valid & OBD_MD_FLFLAGS) &&
2006 (body->oa.o_flags & OBD_FL_RECOV_RESEND))
2007 check_read_checksum(desc, &body->oa, &req->rq_peer,
2009 repbody->oa.o_cksum, cksum_type);
2011 repbody->oa.o_valid = 0;
2013 /* We're finishing using body->oa as an input variable */
2015 /* Check if client was evicted while we were doing i/o before touching
2017 if (likely(rc == 0 &&
2018 !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2) &&
2019 !CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_BULK))) {
2020 rc = target_bulk_io(exp, desc, &lwi);
2025 /* Must commit after prep above in all cases */
2026 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
2027 remote_nb, npages, local_nb, rc);
2029 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
2031 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
2032 ptlrpc_free_bulk(desc);
2037 ptlrpc_lprocfs_brw(req, nob);
2038 } else if (no_reply) {
2039 req->rq_no_reply = 1;
2040 /* reply out callback would free */
2041 ptlrpc_req_drop_rs(req);
2042 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
2043 "client will retry: rc %d\n",
2044 exp->exp_obd->obd_name,
2045 obd_uuid2str(&exp->exp_client_uuid),
2046 obd_export_nid2str(exp), rc);
2048 /* send a bulk after reply to simulate a network delay or reordering
2050 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) {
2051 wait_queue_head_t waitq;
2052 struct l_wait_info lwi1;
2054 CDEBUG(D_INFO, "reorder BULK\n");
2055 init_waitqueue_head(&waitq);
2057 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
2058 l_wait_event(waitq, 0, &lwi1);
2059 target_bulk_io(exp, desc, &lwi);
2060 ptlrpc_free_bulk(desc);
2065 EXPORT_SYMBOL(tgt_brw_read);
2067 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
2068 struct ptlrpc_bulk_desc *desc,
2069 struct niobuf_local *local_nb, int npages,
2070 u32 client_cksum, u32 server_cksum,
2073 struct obd_export *exp = req->rq_export;
2074 struct ost_body *body;
2078 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2079 LASSERT(body != NULL);
2081 if (req->rq_peer.nid != desc->bd_sender) {
2083 router = libcfs_nid2str(desc->bd_sender);
2086 if (exp->exp_obd->obd_checksum_dump)
2087 dump_all_bulk_pages(&body->oa, desc->bd_iov_count,
2088 &BD_GET_KIOV(desc, 0), server_cksum,
2092 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
2093 client_cksum, server_cksum);
2097 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: from %s%s%s inode "
2098 DFID" object "DOSTID" extent [%llu-%llu"
2099 "]: client csum %x, server csum %x\n",
2100 exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
2102 body->oa.o_valid & OBD_MD_FLFID ?
2103 body->oa.o_parent_seq : (__u64)0,
2104 body->oa.o_valid & OBD_MD_FLFID ?
2105 body->oa.o_parent_oid : 0,
2106 body->oa.o_valid & OBD_MD_FLFID ?
2107 body->oa.o_parent_ver : 0,
2108 POSTID(&body->oa.o_oi),
2109 local_nb[0].lnb_file_offset,
2110 local_nb[npages-1].lnb_file_offset +
2111 local_nb[npages - 1].lnb_len - 1,
2112 client_cksum, server_cksum);
2115 int tgt_brw_write(struct tgt_session_info *tsi)
2117 struct ptlrpc_request *req = tgt_ses_req(tsi);
2118 struct ptlrpc_bulk_desc *desc = NULL;
2119 struct obd_export *exp = req->rq_export;
2120 struct niobuf_remote *remote_nb;
2121 struct niobuf_local *local_nb;
2122 struct obd_ioobj *ioo;
2123 struct ost_body *body, *repbody;
2124 struct l_wait_info lwi;
2125 struct lustre_handle lockh = {0};
2127 int objcount, niocount, npages;
2129 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
2130 bool no_reply = false, mmap;
2131 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
2132 bool wait_sync = false;
2136 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) {
2137 CERROR("%s: deny write request from %s to portal %u\n",
2138 tgt_name(tsi->tsi_tgt),
2139 obd_export_nid2str(req->rq_export),
2140 ptlrpc_req2svc(req)->srv_req_portal);
2141 RETURN(err_serious(-EPROTO));
2144 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2145 RETURN(err_serious(-ENOSPC));
2146 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
2147 RETURN(err_serious(-EROFS));
2149 req->rq_bulk_write = 1;
2151 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
2152 RETURN(err_serious(-EIO));
2153 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
2154 RETURN(err_serious(-EFAULT));
2156 /* pause before transaction has been started */
2157 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
2158 cfs_fail_val : (obd_timeout + 1) / 4);
2160 /* There must be big cache in current thread to process this request
2161 * if it is NULL then something went wrong and it wasn't allocated,
2162 * report -ENOMEM in that case */
2166 body = tsi->tsi_ost_body;
2167 LASSERT(body != NULL);
2169 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2170 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
2172 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2173 RCL_CLIENT) / sizeof(*ioo);
2175 for (niocount = i = 0; i < objcount; i++)
2176 niocount += ioo[i].ioo_bufcnt;
2178 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2179 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
2180 if (niocount != req_capsule_get_size(&req->rq_pill,
2181 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
2183 RETURN(err_serious(-EPROTO));
2185 if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
2186 (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self))
2187 memory_pressure_set();
2189 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
2190 niocount * sizeof(*rcs));
2191 rc = req_capsule_server_pack(&req->rq_pill);
2193 GOTO(out, rc = err_serious(rc));
2195 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2196 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2198 local_nb = tbc->local;
2200 rc = tgt_brw_lock(exp->exp_obd->obd_namespace, &tsi->tsi_resid, ioo,
2201 remote_nb, &lockh, LCK_PW);
2206 * If getting the lock took more time than
2207 * client was willing to wait, drop it. b=11330
2209 if (ktime_get_real_seconds() > req->rq_deadline ||
2210 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2212 CERROR("%s: Dropping timed-out write from %s because locking object " DOSTID " took %lld seconds (limit was %lld).\n",
2213 tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2214 POSTID(&ioo->ioo_oid),
2215 ktime_get_real_seconds() - req->rq_arrival_time.tv_sec,
2216 req->rq_deadline - req->rq_arrival_time.tv_sec);
2217 GOTO(out_lock, rc = -ETIMEDOUT);
2220 /* Because we already sync grant info with client when reconnect,
2221 * grant info will be cleared for resent req, then fed_grant and
2222 * total_grant will not be modified in following preprw_write */
2223 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2224 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2225 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2228 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2229 if (repbody == NULL)
2230 GOTO(out_lock, rc = -ENOMEM);
2231 repbody->oa = body->oa;
2233 npages = PTLRPC_MAX_BRW_PAGES;
2234 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2235 objcount, ioo, remote_nb, &npages, local_nb);
2239 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2240 PTLRPC_BULK_GET_SINK | PTLRPC_BULK_BUF_KIOV,
2242 &ptlrpc_bulk_kiov_nopin_ops);
2244 GOTO(skip_transfer, rc = -ENOMEM);
2246 /* NB Having prepped, we must commit... */
2247 for (i = 0; i < npages; i++)
2248 desc->bd_frag_ops->add_kiov_frag(desc,
2249 local_nb[i].lnb_page,
2250 local_nb[i].lnb_page_offset,
2251 local_nb[i].lnb_len);
2253 rc = sptlrpc_svc_prep_bulk(req, desc);
2255 GOTO(skip_transfer, rc);
2257 rc = target_bulk_io(exp, desc, &lwi);
2261 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2262 static int cksum_counter;
2264 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2265 cksum_type = cksum_type_unpack(body->oa.o_flags);
2267 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2268 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2269 repbody->oa.o_flags |= cksum_type_pack(cksum_type);
2270 repbody->oa.o_cksum = tgt_checksum_bulk(tsi->tsi_tgt, desc,
2271 OST_WRITE, cksum_type);
2274 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2275 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2276 body->oa.o_flags & OBD_FL_MMAP);
2278 tgt_warn_on_cksum(req, desc, local_nb, npages,
2280 repbody->oa.o_cksum, mmap);
2282 } else if ((cksum_counter & (-cksum_counter)) ==
2284 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2285 cksum_counter, libcfs_id2str(req->rq_peer),
2286 repbody->oa.o_cksum);
2290 /* Must commit after prep above in all cases */
2291 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2292 objcount, ioo, remote_nb, npages, local_nb, rc);
2293 if (rc == -ENOTCONN)
2294 /* quota acquire process has been given up because
2295 * either the client has been evicted or the client
2296 * has timed out the request already */
2299 for (i = 0; i < niocount; i++) {
2300 if (!(local_nb[i].lnb_flags & OBD_BRW_ASYNC)) {
2306 * Disable sending mtime back to the client. If the client locked the
2307 * whole object, then it has already updated the mtime on its side,
2308 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2310 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2315 /* set per-requested niobuf return codes */
2316 for (i = j = 0; i < niocount; i++) {
2317 int len = remote_nb[i].rnb_len;
2322 LASSERT(j < npages);
2323 if (local_nb[j].lnb_rc < 0)
2324 rcs[i] = local_nb[j].lnb_rc;
2325 len -= local_nb[j].lnb_len;
2330 LASSERT(j == npages);
2331 ptlrpc_lprocfs_brw(req, nob);
2334 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2336 ptlrpc_free_bulk(desc);
2338 if (unlikely(no_reply || (exp->exp_obd->obd_no_transno && wait_sync))) {
2339 req->rq_no_reply = 1;
2340 /* reply out callback would free */
2341 ptlrpc_req_drop_rs(req);
2342 if (!exp->exp_obd->obd_no_transno)
2343 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s),"
2344 " client will retry: rc = %d\n",
2345 exp->exp_obd->obd_name,
2346 obd_uuid2str(&exp->exp_client_uuid),
2347 obd_export_nid2str(exp), rc);
2349 memory_pressure_clr();
2352 EXPORT_SYMBOL(tgt_brw_write);
2354 /* Check if request can be reconstructed from saved reply data
2355 * A copy of the reply data is returned in @trd if the pointer is not NULL
2357 bool req_can_reconstruct(struct ptlrpc_request *req,
2358 struct tg_reply_data *trd)
2360 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2361 struct lsd_client_data *lcd = ted->ted_lcd;
2364 if (tgt_is_multimodrpcs_client(req->rq_export))
2365 return tgt_lookup_reply(req, trd);
2367 mutex_lock(&ted->ted_lcd_lock);
2368 found = req->rq_xid == lcd->lcd_last_xid ||
2369 req->rq_xid == lcd->lcd_last_close_xid;
2371 if (found && trd != NULL) {
2372 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
2373 trd->trd_reply.lrd_xid = lcd->lcd_last_close_xid;
2374 trd->trd_reply.lrd_transno =
2375 lcd->lcd_last_close_transno;
2376 trd->trd_reply.lrd_result = lcd->lcd_last_close_result;
2378 trd->trd_reply.lrd_xid = lcd->lcd_last_xid;
2379 trd->trd_reply.lrd_transno = lcd->lcd_last_transno;
2380 trd->trd_reply.lrd_result = lcd->lcd_last_result;
2381 trd->trd_reply.lrd_data = lcd->lcd_last_data;
2382 trd->trd_pre_versions[0] = lcd->lcd_pre_versions[0];
2383 trd->trd_pre_versions[1] = lcd->lcd_pre_versions[1];
2384 trd->trd_pre_versions[2] = lcd->lcd_pre_versions[2];
2385 trd->trd_pre_versions[3] = lcd->lcd_pre_versions[3];
2388 mutex_unlock(&ted->ted_lcd_lock);
2392 EXPORT_SYMBOL(req_can_reconstruct);