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) 2011, 2014, 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 <md_object.h>
46 #include <lustre_lfsck.h>
47 #include <lustre_nodemap.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 = tsi->tsi_exp->exp_target_data.ted_nodemap;
260 body->oa.o_uid = nodemap_map_id(nodemap, NODEMAP_UID,
261 NODEMAP_CLIENT_TO_FS,
263 body->oa.o_gid = nodemap_map_id(nodemap, NODEMAP_GID,
264 NODEMAP_CLIENT_TO_FS,
267 tsi->tsi_ost_body = body;
268 tsi->tsi_fid = body->oa.o_oi.oi_fid;
270 if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
271 rc = tgt_io_data_unpack(tsi, &body->oa.o_oi);
276 if (!(body->oa.o_valid & OBD_MD_FLID)) {
277 if (flags & HABEO_CORPUS) {
278 CERROR("%s: OBD_MD_FLID flag is not set in ost_body "
279 "but OID/FID is mandatory with HABEO_CORPUS\n",
280 tgt_name(tsi->tsi_tgt));
287 ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
290 * OST doesn't get object in advance for further use to prevent
291 * situations with nested object_find which is potential deadlock.
293 tsi->tsi_corpus = NULL;
298 * Do necessary preprocessing according to handler ->th_flags.
300 static int tgt_request_preprocess(struct tgt_session_info *tsi,
301 struct tgt_handler *h,
302 struct ptlrpc_request *req)
304 struct req_capsule *pill = tsi->tsi_pill;
305 __u32 flags = h->th_flags;
310 if (tsi->tsi_preprocessed)
313 LASSERT(h->th_act != NULL);
314 LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
315 LASSERT(current->journal_info == NULL);
317 LASSERT(ergo(flags & (HABEO_CORPUS | HABEO_REFERO),
319 if (h->th_fmt != NULL) {
320 req_capsule_set(pill, h->th_fmt);
321 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
322 rc = tgt_mdt_body_unpack(tsi, flags);
325 } else if (req_capsule_has_field(pill, &RMF_OST_BODY,
327 rc = tgt_ost_body_unpack(tsi, flags);
333 if (flags & MUTABOR && tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
336 if (flags & HABEO_CLAVIS) {
337 struct ldlm_request *dlm_req;
339 LASSERT(h->th_fmt != NULL);
341 dlm_req = req_capsule_client_get(pill, &RMF_DLM_REQ);
342 if (dlm_req != NULL) {
343 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
345 dlm_req->lock_desc.l_policy_data.\
346 l_inodebits.bits == 0)) {
348 * Lock without inodebits makes no sense and
349 * will oops later in ldlm. If client miss to
350 * set such bits, do not trigger ASSERTION.
352 * For liblustre flock case, it maybe zero.
356 tsi->tsi_dlm_req = dlm_req;
362 tsi->tsi_preprocessed = 1;
367 * Invoke handler for this request opc. Also do necessary preprocessing
368 * (according to handler ->th_flags), and post-processing (setting of
369 * ->last_{xid,committed}).
371 static int tgt_handle_request0(struct tgt_session_info *tsi,
372 struct tgt_handler *h,
373 struct ptlrpc_request *req)
377 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
382 /* When dealing with sec context requests, no export is associated yet,
383 * because these requests are sent before *_CONNECT requests.
384 * A NULL req->rq_export means the normal *_common_slice handlers will
385 * not be called, because there is no reference to the target.
386 * So deal with them by hand and jump directly to target_send_reply().
390 case SEC_CTX_INIT_CONT:
396 * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
397 * to put same checks into handlers like mdt_close(), mdt_reint(),
398 * etc., without talking to mdt authors first. Checking same thing
399 * there again is useless and returning 0 error without packing reply
400 * is buggy! Handlers either pack reply or return error.
402 * We return 0 here and do not send any reply in order to emulate
403 * network failure. Do not send any reply in case any of NET related
404 * fail_id has occured.
406 if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
409 rc = tgt_request_preprocess(tsi, h, req);
410 /* pack reply if reply format is fixed */
411 if (rc == 0 && h->th_flags & HABEO_REFERO) {
413 if (req_capsule_has_field(tsi->tsi_pill, &RMF_MDT_MD,
415 req_capsule_set_size(tsi->tsi_pill, &RMF_MDT_MD,
417 tsi->tsi_mdt_body->mbo_eadatasize);
418 if (req_capsule_has_field(tsi->tsi_pill, &RMF_LOGCOOKIES,
420 req_capsule_set_size(tsi->tsi_pill, &RMF_LOGCOOKIES,
423 rc = req_capsule_server_pack(tsi->tsi_pill);
426 if (likely(rc == 0)) {
428 * Process request, there can be two types of rc:
429 * 1) errors with msg unpack/pack, other failures outside the
430 * operation itself. This is counted as serious errors;
431 * 2) errors during fs operation, should be placed in rq_status
435 if (!is_serious(rc) &&
436 !req->rq_no_reply && req->rq_reply_state == NULL) {
437 DEBUG_REQ(D_ERROR, req, "%s \"handler\" %s did not "
438 "pack reply and returned 0 error\n",
439 tgt_name(tsi->tsi_tgt), h->th_name);
442 serious = is_serious(rc);
443 rc = clear_serious(rc);
451 * ELDLM_* codes which > 0 should be in rq_status only as well as
452 * all non-serious errors.
454 if (rc > 0 || !serious)
457 LASSERT(current->journal_info == NULL);
459 if (likely(rc == 0 && req->rq_export))
460 target_committed_to_req(req);
463 target_send_reply(req, rc, tsi->tsi_reply_fail_id);
467 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
468 struct obd_device *obd, int *process)
470 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
477 case MDS_SYNC: /* used in unmounting */
491 case MDS_HSM_PROGRESS:
492 case MDS_HSM_STATE_SET:
493 case MDS_HSM_REQUEST:
494 *process = target_queue_recovery_request(req, obd);
498 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
505 * Handle recovery. Return:
506 * +1: continue request processing;
507 * -ve: abort immediately with the given error code;
508 * 0: send reply with error code in req->rq_status;
510 static int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
514 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
519 case SEC_CTX_INIT_CONT:
524 if (!req->rq_export->exp_obd->obd_replayable)
527 /* sanity check: if the xid matches, the request must be marked as a
528 * resent or replayed */
529 if (req_xid_is_last(req)) {
530 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
531 (MSG_RESENT | MSG_REPLAY))) {
532 DEBUG_REQ(D_WARNING, req, "rq_xid "LPU64" matches "
533 "last_xid, expected REPLAY or RESENT flag "
535 lustre_msg_get_flags(req->rq_reqmsg));
536 req->rq_status = -ENOTCONN;
540 /* else: note the opposite is not always true; a RESENT req after a
541 * failover will usually not match the last_xid, since it was likely
542 * never committed. A REPLAYed request will almost never match the
543 * last xid, however it could for a committed, but still retained,
546 /* Check for aborted recovery... */
547 if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
551 DEBUG_REQ(D_INFO, req, "Got new replay");
552 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
554 if (rc != 0 || !should_process)
556 else if (should_process < 0) {
557 req->rq_status = should_process;
558 rc = ptlrpc_error(req);
565 /* Initial check for request, it is validation mostly */
566 static struct tgt_handler *tgt_handler_find_check(struct ptlrpc_request *req)
568 struct tgt_handler *h;
569 struct tgt_opc_slice *s;
570 struct lu_target *tgt;
571 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
575 tgt = class_exp2tgt(req->rq_export);
576 if (unlikely(tgt == NULL)) {
577 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
578 class_exp2obd(req->rq_export)->obd_name);
579 RETURN(ERR_PTR(-EINVAL));
582 for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
583 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
586 /* opcode was not found in slice */
587 if (unlikely(s->tos_hs == NULL)) {
588 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt),
590 RETURN(ERR_PTR(-ENOTSUPP));
593 LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
594 h = s->tos_hs + (opc - s->tos_opc_start);
595 if (unlikely(h->th_opc == 0)) {
596 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
597 RETURN(ERR_PTR(-ENOTSUPP));
603 int tgt_request_handle(struct ptlrpc_request *req)
605 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
607 struct lustre_msg *msg = req->rq_reqmsg;
608 struct tgt_handler *h;
609 struct lu_target *tgt;
610 int request_fail_id = 0;
611 __u32 opc = lustre_msg_get_opc(msg);
616 /* Refill the context, to make sure all thread keys are allocated */
617 lu_env_refill(req->rq_svc_thread->t_env);
619 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
620 tsi->tsi_pill = &req->rq_pill;
621 tsi->tsi_env = req->rq_svc_thread->t_env;
623 /* if request has export then get handlers slice from corresponding
624 * target, otherwise that should be connect operation */
625 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
626 opc == MGS_CONNECT) {
627 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
628 rc = target_handle_connect(req);
630 rc = ptlrpc_error(req);
633 /* recovery-small test 18c asks to drop connect reply */
634 if (unlikely(opc == OST_CONNECT &&
635 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
639 if (unlikely(!class_connected_export(req->rq_export))) {
640 if (opc == SEC_CTX_INIT || opc == SEC_CTX_INIT_CONT ||
641 opc == SEC_CTX_FINI) {
642 /* sec context initialization has to be handled
643 * by hand in tgt_handle_request0() */
644 tsi->tsi_reply_fail_id = OBD_FAIL_SEC_CTX_INIT_NET;
646 GOTO(handle_recov, rc = 0);
648 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
649 opc, libcfs_id2str(req->rq_peer));
650 req->rq_status = -ENOTCONN;
651 rc = ptlrpc_error(req);
655 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
656 tsi->tsi_exp = req->rq_export;
657 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
658 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
660 tsi->tsi_jobid = NULL;
663 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
664 class_exp2obd(req->rq_export)->obd_name);
665 req->rq_status = -EINVAL;
666 rc = ptlrpc_error(req);
670 request_fail_id = tgt->lut_request_fail_id;
671 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
673 h = tgt_handler_find_check(req);
675 req->rq_status = PTR_ERR(h);
676 rc = ptlrpc_error(req);
680 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
683 if (CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
686 rc = lustre_msg_check_version(msg, h->th_version);
688 DEBUG_REQ(D_ERROR, req, "%s: drop mal-formed request, version"
689 " %08x, expecting %08x\n", tgt_name(tgt),
690 lustre_msg_get_version(msg), h->th_version);
691 req->rq_status = -EINVAL;
692 rc = ptlrpc_error(req);
697 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
698 if (likely(rc == 1)) {
699 rc = tgt_handle_request0(tsi, h, req);
705 req_capsule_fini(tsi->tsi_pill);
706 if (tsi->tsi_corpus != NULL) {
707 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
708 tsi->tsi_corpus = NULL;
712 EXPORT_SYMBOL(tgt_request_handle);
714 /** Assign high priority operations to the request if needed. */
715 int tgt_hpreq_handler(struct ptlrpc_request *req)
717 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
718 struct tgt_handler *h;
723 if (req->rq_export == NULL)
726 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
727 tsi->tsi_pill = &req->rq_pill;
728 tsi->tsi_env = req->rq_svc_thread->t_env;
729 tsi->tsi_tgt = class_exp2tgt(req->rq_export);
730 tsi->tsi_exp = req->rq_export;
732 h = tgt_handler_find_check(req);
738 rc = tgt_request_preprocess(tsi, h, req);
739 if (unlikely(rc != 0))
742 if (h->th_hp != NULL)
746 EXPORT_SYMBOL(tgt_hpreq_handler);
748 void tgt_counter_incr(struct obd_export *exp, int opcode)
750 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
751 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
752 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
754 EXPORT_SYMBOL(tgt_counter_incr);
757 * Unified target generic handlers.
763 static inline void tgt_init_sec_none(struct obd_connect_data *reply)
765 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT |
766 OBD_CONNECT_RMT_CLIENT_FORCE);
769 static int tgt_init_sec_level(struct ptlrpc_request *req)
771 struct lu_target *tgt = class_exp2tgt(req->rq_export);
773 struct obd_connect_data *data, *reply;
778 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
779 reply = req_capsule_server_get(&req->rq_pill, &RMF_CONNECT_DATA);
780 if (data == NULL || reply == NULL)
783 /* connection from MDT is always trusted */
784 if (req->rq_auth_usr_mdt) {
785 tgt_init_sec_none(reply);
789 if (unlikely(tgt == NULL)) {
790 DEBUG_REQ(D_ERROR, req, "%s: No target for connected export\n",
791 class_exp2obd(req->rq_export)->obd_name);
795 client = libcfs_nid2str(req->rq_peer.nid);
796 /* no GSS support case */
797 if (!req->rq_auth_gss) {
798 if (tgt->lut_sec_level > LUSTRE_SEC_NONE) {
799 CWARN("client %s -> target %s does not use GSS, "
800 "can not run under security level %d.\n",
801 client, tgt_name(tgt), tgt->lut_sec_level);
804 tgt_init_sec_none(reply);
809 /* old version case */
810 if (unlikely(!(data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT) ||
811 !(data->ocd_connect_flags & OBD_CONNECT_MDS_CAPA) ||
812 !(data->ocd_connect_flags & OBD_CONNECT_OSS_CAPA))) {
813 if (tgt->lut_sec_level > LUSTRE_SEC_NONE) {
814 CWARN("client %s -> target %s uses old version, "
815 "can not run under security level %d.\n",
816 client, tgt_name(tgt), tgt->lut_sec_level);
819 CWARN("client %s -> target %s uses old version, "
820 "run under security level %d.\n",
821 client, tgt_name(tgt), tgt->lut_sec_level);
822 tgt_init_sec_none(reply);
827 remote = data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT_FORCE;
829 if (!req->rq_auth_remote)
830 CDEBUG(D_SEC, "client (local realm) %s -> target %s "
831 "asked to be remote.\n", client, tgt_name(tgt));
832 } else if (req->rq_auth_remote) {
834 CDEBUG(D_SEC, "client (remote realm) %s -> target %s is set "
835 "as remote by default.\n", client, tgt_name(tgt));
839 if (!uid_valid(make_kuid(&init_user_ns, req->rq_auth_uid))) {
840 CDEBUG(D_SEC, "client %s -> target %s: user is not "
841 "authenticated!\n", client, tgt_name(tgt));
847 switch (tgt->lut_sec_level) {
848 case LUSTRE_SEC_NONE:
851 "client %s -> target %s is set as remote, "
852 "can not run under security level %d.\n",
853 client, tgt_name(tgt), tgt->lut_sec_level);
856 tgt_init_sec_none(reply);
858 case LUSTRE_SEC_REMOTE:
860 tgt_init_sec_none(reply);
865 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT |
866 OBD_CONNECT_RMT_CLIENT_FORCE);
867 reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA;
868 reply->ocd_connect_flags &= ~OBD_CONNECT_MDS_CAPA;
877 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
879 struct lu_target *tgt = class_exp2tgt(exp);
880 struct sptlrpc_flavor flvr;
884 LASSERT(tgt->lut_obd);
885 LASSERT(tgt->lut_slice);
887 /* always allow ECHO client */
888 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
889 LUSTRE_ECHO_NAME) == 0)) {
890 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
894 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
895 read_lock(&tgt->lut_sptlrpc_lock);
896 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
900 read_unlock(&tgt->lut_sptlrpc_lock);
902 spin_lock(&exp->exp_lock);
903 exp->exp_sp_peer = req->rq_sp_from;
904 exp->exp_flvr = flvr;
905 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
906 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
907 CERROR("%s: unauthorized rpc flavor %x from %s, "
908 "expect %x\n", tgt_name(tgt),
910 libcfs_nid2str(req->rq_peer.nid),
911 exp->exp_flvr.sf_rpc);
914 spin_unlock(&exp->exp_lock);
916 if (exp->exp_sp_peer != req->rq_sp_from) {
917 CERROR("%s: RPC source %s doesn't match %s\n",
919 sptlrpc_part2name(req->rq_sp_from),
920 sptlrpc_part2name(exp->exp_sp_peer));
923 rc = sptlrpc_target_export_check(exp, req);
930 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt, int initial)
932 struct sptlrpc_rule_set tmp_rset;
935 if (unlikely(tgt == NULL)) {
936 CERROR("No target passed");
940 sptlrpc_rule_set_init(&tmp_rset);
941 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset, initial);
943 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
948 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
950 write_lock(&tgt->lut_sptlrpc_lock);
951 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
952 tgt->lut_sptlrpc_rset = tmp_rset;
953 write_unlock(&tgt->lut_sptlrpc_lock);
957 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
959 int tgt_connect(struct tgt_session_info *tsi)
961 struct ptlrpc_request *req = tgt_ses_req(tsi);
962 struct obd_connect_data *reply;
967 rc = tgt_init_sec_level(req);
971 /* XXX: better to call this check right after getting new export but
972 * before last_rcvd slot allocation to avoid server load upon insecure
973 * connects. This is to be fixed after unifiyng all targets.
975 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
979 /* To avoid exposing partially initialized connection flags, changes up
980 * to this point have been staged in reply->ocd_connect_flags. Now that
981 * connection handling has completed successfully, atomically update
982 * the connect flags in the shared export data structure. LU-1623 */
983 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
984 spin_lock(&tsi->tsi_exp->exp_lock);
985 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
986 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
987 spin_unlock(&tsi->tsi_exp->exp_lock);
991 obd_disconnect(class_export_get(tsi->tsi_exp));
994 EXPORT_SYMBOL(tgt_connect);
996 int tgt_disconnect(struct tgt_session_info *tsi)
1002 rc = target_handle_disconnect(tgt_ses_req(tsi));
1004 RETURN(err_serious(rc));
1008 EXPORT_SYMBOL(tgt_disconnect);
1011 * Unified target OBD handlers
1013 int tgt_obd_ping(struct tgt_session_info *tsi)
1019 rc = target_handle_ping(tgt_ses_req(tsi));
1021 RETURN(err_serious(rc));
1025 EXPORT_SYMBOL(tgt_obd_ping);
1027 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
1029 return err_serious(-EOPNOTSUPP);
1032 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
1034 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1035 struct ptlrpc_request *req = tgt_ses_req(tsi);
1036 struct obd_export *exp = req->rq_export;
1037 struct ptlrpc_bulk_desc *desc;
1038 struct l_wait_info *lwi = &tti->tti_u.rdpg.tti_wait_info;
1046 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1,
1047 PTLRPC_BULK_PUT_SOURCE |
1048 PTLRPC_BULK_BUF_KIOV,
1050 &ptlrpc_bulk_kiov_pin_ops);
1054 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
1055 /* old client requires reply size in it's PAGE_CACHE_SIZE,
1056 * which is rdpg->rp_count */
1057 nob = rdpg->rp_count;
1059 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
1060 i++, tmpcount -= tmpsize) {
1061 tmpsize = min_t(int, tmpcount, PAGE_CACHE_SIZE);
1062 desc->bd_frag_ops->add_kiov_frag(desc, rdpg->rp_pages[i], 0,
1066 LASSERT(desc->bd_nob == nob);
1067 rc = target_bulk_io(exp, desc, lwi);
1068 ptlrpc_free_bulk(desc);
1071 EXPORT_SYMBOL(tgt_sendpage);
1074 * OBD_IDX_READ handler
1076 static int tgt_obd_idx_read(struct tgt_session_info *tsi)
1078 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
1079 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
1080 struct idx_info *req_ii, *rep_ii;
1085 memset(rdpg, 0, sizeof(*rdpg));
1086 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
1088 /* extract idx_info buffer from request & reply */
1089 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
1090 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
1091 RETURN(err_serious(-EPROTO));
1093 rc = req_capsule_server_pack(tsi->tsi_pill);
1095 RETURN(err_serious(rc));
1097 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
1099 RETURN(err_serious(-EFAULT));
1100 rep_ii->ii_magic = IDX_INFO_MAGIC;
1102 /* extract hash to start with */
1103 rdpg->rp_hash = req_ii->ii_hash_start;
1105 /* extract requested attributes */
1106 rdpg->rp_attrs = req_ii->ii_attrs;
1108 /* check that fid packed in request is valid and supported */
1109 if (!fid_is_sane(&req_ii->ii_fid))
1111 rep_ii->ii_fid = req_ii->ii_fid;
1114 rep_ii->ii_flags = req_ii->ii_flags;
1116 /* compute number of pages to allocate, ii_count is the number of 4KB
1118 if (req_ii->ii_count <= 0)
1119 GOTO(out, rc = -EFAULT);
1120 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
1121 exp_max_brw_size(tsi->tsi_exp));
1122 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE -1) >> PAGE_CACHE_SHIFT;
1124 /* allocate pages to store the containers */
1125 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1126 if (rdpg->rp_pages == NULL)
1127 GOTO(out, rc = -ENOMEM);
1128 for (i = 0; i < rdpg->rp_npages; i++) {
1129 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
1130 if (rdpg->rp_pages[i] == NULL)
1131 GOTO(out, rc = -ENOMEM);
1134 /* populate pages with key/record pairs */
1135 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1139 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1140 "asked %d > %d\n", rc, rdpg->rp_count);
1142 /* send pages to client */
1143 rc = tgt_sendpage(tsi, rdpg, rc);
1148 if (rdpg->rp_pages) {
1149 for (i = 0; i < rdpg->rp_npages; i++)
1150 if (rdpg->rp_pages[i])
1151 __free_page(rdpg->rp_pages[i]);
1152 OBD_FREE(rdpg->rp_pages,
1153 rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1158 struct tgt_handler tgt_obd_handlers[] = {
1159 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1160 TGT_OBD_HDL_VAR(0, OBD_LOG_CANCEL, tgt_obd_log_cancel),
1161 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1163 EXPORT_SYMBOL(tgt_obd_handlers);
1165 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1166 struct dt_object *obj, __u64 start, __u64 end)
1172 /* if no objid is specified, it means "sync whole filesystem" */
1174 rc = dt_sync(env, tgt->lut_bottom);
1175 } else if (dt_version_get(env, obj) >
1176 tgt->lut_obd->obd_last_committed) {
1177 rc = dt_object_sync(env, obj, start, end);
1182 EXPORT_SYMBOL(tgt_sync);
1184 * Unified target DLM handlers.
1187 /* Ensure that data and metadata are synced to the disk when lock is cancelled
1189 static int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1190 void *data, int flag)
1193 struct lu_target *tgt;
1194 struct dt_object *obj;
1200 tgt = class_exp2tgt(lock->l_export);
1202 if (unlikely(tgt == NULL)) {
1203 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1204 class_exp2obd(lock->l_export)->obd_name);
1208 if (flag == LDLM_CB_CANCELING &&
1209 (lock->l_granted_mode & (LCK_PW | LCK_GROUP)) &&
1210 (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1211 (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1212 lock->l_flags & LDLM_FL_CBPENDING))) {
1214 __u64 end = OBD_OBJECT_EOF;
1216 rc = lu_env_init(&env, LCT_DT_THREAD);
1217 if (unlikely(rc != 0))
1220 ost_fid_from_resid(&fid, &lock->l_resource->lr_name,
1221 tgt->lut_lsd.lsd_osd_index);
1222 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1224 GOTO(err_env, rc = PTR_ERR(obj));
1226 if (!dt_object_exists(obj))
1227 GOTO(err_put, rc = -ENOENT);
1229 if (lock->l_resource->lr_type == LDLM_EXTENT) {
1230 start = lock->l_policy_data.l_extent.start;
1231 end = lock->l_policy_data.l_extent.end;
1234 rc = tgt_sync(&env, tgt, obj, start, end);
1236 CERROR("%s: syncing "DFID" ("LPU64"-"LPU64") on lock "
1237 "cancel: rc = %d\n",
1238 tgt_name(tgt), PFID(&fid),
1239 lock->l_policy_data.l_extent.start,
1240 lock->l_policy_data.l_extent.end, rc);
1243 lu_object_put(&env, &obj->do_lu);
1248 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1252 static struct ldlm_callback_suite tgt_dlm_cbs = {
1253 .lcs_completion = ldlm_server_completion_ast,
1254 .lcs_blocking = tgt_blocking_ast,
1255 .lcs_glimpse = ldlm_server_glimpse_ast
1258 int tgt_enqueue(struct tgt_session_info *tsi)
1260 struct ptlrpc_request *req = tgt_ses_req(tsi);
1265 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1266 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1268 LASSERT(tsi->tsi_dlm_req != NULL);
1269 rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1270 tsi->tsi_dlm_req, &tgt_dlm_cbs);
1272 RETURN(err_serious(rc));
1274 switch (LUT_FAIL_CLASS(tsi->tsi_reply_fail_id)) {
1276 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_LDLM_REPLY_NET;
1279 tsi->tsi_reply_fail_id = OBD_FAIL_OST_LDLM_REPLY_NET;
1282 tsi->tsi_reply_fail_id = OBD_FAIL_MGS_LDLM_REPLY_NET;
1285 tsi->tsi_reply_fail_id = OBD_FAIL_LDLM_REPLY;
1288 RETURN(req->rq_status);
1290 EXPORT_SYMBOL(tgt_enqueue);
1292 int tgt_convert(struct tgt_session_info *tsi)
1294 struct ptlrpc_request *req = tgt_ses_req(tsi);
1298 LASSERT(tsi->tsi_dlm_req);
1299 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1301 RETURN(err_serious(rc));
1303 RETURN(req->rq_status);
1306 int tgt_bl_callback(struct tgt_session_info *tsi)
1308 return err_serious(-EOPNOTSUPP);
1311 int tgt_cp_callback(struct tgt_session_info *tsi)
1313 return err_serious(-EOPNOTSUPP);
1316 /* generic LDLM target handler */
1317 struct tgt_handler tgt_dlm_handlers[] = {
1318 TGT_DLM_HDL (HABEO_CLAVIS, LDLM_ENQUEUE, tgt_enqueue),
1319 TGT_DLM_HDL_VAR(HABEO_CLAVIS, LDLM_CONVERT, tgt_convert),
1320 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1321 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1323 EXPORT_SYMBOL(tgt_dlm_handlers);
1326 * Unified target LLOG handlers.
1328 int tgt_llog_open(struct tgt_session_info *tsi)
1334 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1338 EXPORT_SYMBOL(tgt_llog_open);
1340 int tgt_llog_close(struct tgt_session_info *tsi)
1346 rc = llog_origin_handle_close(tgt_ses_req(tsi));
1350 EXPORT_SYMBOL(tgt_llog_close);
1353 int tgt_llog_destroy(struct tgt_session_info *tsi)
1359 rc = llog_origin_handle_destroy(tgt_ses_req(tsi));
1364 int tgt_llog_read_header(struct tgt_session_info *tsi)
1370 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1374 EXPORT_SYMBOL(tgt_llog_read_header);
1376 int tgt_llog_next_block(struct tgt_session_info *tsi)
1382 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1386 EXPORT_SYMBOL(tgt_llog_next_block);
1388 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1394 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1398 EXPORT_SYMBOL(tgt_llog_prev_block);
1400 /* generic llog target handler */
1401 struct tgt_handler tgt_llog_handlers[] = {
1402 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1403 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1404 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1405 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1406 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_DESTROY, tgt_llog_destroy),
1407 TGT_LLOG_HDL_VAR(0, LLOG_ORIGIN_HANDLE_CLOSE, tgt_llog_close),
1409 EXPORT_SYMBOL(tgt_llog_handlers);
1411 int (*tgt_lfsck_in_notify)(const struct lu_env *env,
1412 struct dt_device *key,
1413 struct lfsck_request *lr,
1414 struct thandle *th) = NULL;
1416 void tgt_register_lfsck_in_notify(int (*notify)(const struct lu_env *,
1418 struct lfsck_request *,
1421 tgt_lfsck_in_notify = notify;
1423 EXPORT_SYMBOL(tgt_register_lfsck_in_notify);
1425 static int (*tgt_lfsck_query)(const struct lu_env *env,
1426 struct dt_device *key,
1427 struct lfsck_request *lr) = NULL;
1429 void tgt_register_lfsck_query(int (*query)(const struct lu_env *,
1431 struct lfsck_request *))
1433 tgt_lfsck_query = query;
1435 EXPORT_SYMBOL(tgt_register_lfsck_query);
1437 /* LFSCK request handlers */
1438 static int tgt_handle_lfsck_notify(struct tgt_session_info *tsi)
1440 const struct lu_env *env = tsi->tsi_env;
1441 struct dt_device *key = tsi->tsi_tgt->lut_bottom;
1442 struct lfsck_request *lr;
1446 lr = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1450 rc = tgt_lfsck_in_notify(env, key, lr, NULL);
1455 static int tgt_handle_lfsck_query(struct tgt_session_info *tsi)
1457 struct lfsck_request *request;
1458 struct lfsck_reply *reply;
1462 request = req_capsule_client_get(tsi->tsi_pill, &RMF_LFSCK_REQUEST);
1463 if (request == NULL)
1466 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_LFSCK_REPLY);
1470 rc = tgt_lfsck_query(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, request);
1471 reply->lr_status = rc;
1473 RETURN(rc < 0 ? rc : 0);
1476 struct tgt_handler tgt_lfsck_handlers[] = {
1477 TGT_LFSCK_HDL(HABEO_REFERO, LFSCK_NOTIFY, tgt_handle_lfsck_notify),
1478 TGT_LFSCK_HDL(HABEO_REFERO, LFSCK_QUERY, tgt_handle_lfsck_query),
1480 EXPORT_SYMBOL(tgt_lfsck_handlers);
1483 * initialize per-thread page pool (bug 5137).
1485 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1487 struct tgt_thread_big_cache *tbc;
1491 LASSERT(thread != NULL);
1492 LASSERT(thread->t_data == NULL);
1494 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1497 thread->t_data = tbc;
1500 EXPORT_SYMBOL(tgt_io_thread_init);
1503 * free per-thread pool created by tgt_thread_init().
1505 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1507 struct tgt_thread_big_cache *tbc;
1511 LASSERT(thread != NULL);
1514 * be prepared to handle partially-initialized pools (because this is
1515 * called from ost_io_thread_init() for cleanup.
1517 tbc = thread->t_data;
1519 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1520 thread->t_data = NULL;
1524 EXPORT_SYMBOL(tgt_io_thread_done);
1526 * Helper function for getting server side [start, start+count] DLM lock
1527 * if asked by client.
1529 int tgt_extent_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1530 __u64 start, __u64 end, struct lustre_handle *lh,
1531 int mode, __u64 *flags)
1533 ldlm_policy_data_t policy;
1538 LASSERT(lh != NULL);
1539 LASSERT(ns != NULL);
1540 LASSERT(!lustre_handle_is_used(lh));
1542 policy.l_extent.gid = 0;
1543 policy.l_extent.start = start & PAGE_MASK;
1546 * If ->o_blocks is EOF it means "lock till the end of the file".
1547 * Otherwise, it's size of an extent or hole being punched (in bytes).
1549 if (end == OBD_OBJECT_EOF || end < start)
1550 policy.l_extent.end = OBD_OBJECT_EOF;
1552 policy.l_extent.end = end | ~PAGE_MASK;
1554 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_EXTENT, &policy, mode,
1555 flags, ldlm_blocking_ast,
1556 ldlm_completion_ast, ldlm_glimpse_ast,
1557 NULL, 0, LVB_T_NONE, NULL, lh);
1558 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1560 EXPORT_SYMBOL(tgt_extent_lock);
1562 void tgt_extent_unlock(struct lustre_handle *lh, ldlm_mode_t mode)
1564 LASSERT(lustre_handle_is_used(lh));
1565 ldlm_lock_decref(lh, mode);
1567 EXPORT_SYMBOL(tgt_extent_unlock);
1569 int tgt_brw_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1570 struct obd_ioobj *obj, struct niobuf_remote *nb,
1571 struct lustre_handle *lh, int mode)
1574 int nrbufs = obj->ioo_bufcnt;
1579 LASSERT(mode == LCK_PR || mode == LCK_PW);
1580 LASSERT(!lustre_handle_is_used(lh));
1582 if (nrbufs == 0 || !(nb[0].rnb_flags & OBD_BRW_SRVLOCK))
1585 for (i = 1; i < nrbufs; i++)
1586 if (!(nb[i].rnb_flags & OBD_BRW_SRVLOCK))
1589 RETURN(tgt_extent_lock(ns, res_id, nb[0].rnb_offset,
1590 nb[nrbufs - 1].rnb_offset +
1591 nb[nrbufs - 1].rnb_len - 1,
1595 void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1596 struct lustre_handle *lh, int mode)
1600 LASSERT(mode == LCK_PR || mode == LCK_PW);
1601 LASSERT((obj->ioo_bufcnt > 0 &&
1602 (niob[0].rnb_flags & OBD_BRW_SRVLOCK)) ==
1603 lustre_handle_is_used(lh));
1605 if (lustre_handle_is_used(lh))
1606 tgt_extent_unlock(lh, mode);
1610 static __u32 tgt_checksum_bulk(struct lu_target *tgt,
1611 struct ptlrpc_bulk_desc *desc, int opc,
1612 cksum_type_t cksum_type)
1614 struct cfs_crypto_hash_desc *hdesc;
1615 unsigned int bufsize;
1617 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1620 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
1622 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1623 if (IS_ERR(hdesc)) {
1624 CERROR("%s: unable to initialize checksum hash %s\n",
1625 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1626 return PTR_ERR(hdesc);
1629 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1630 for (i = 0; i < desc->bd_iov_count; i++) {
1631 /* corrupt the data before we compute the checksum, to
1632 * simulate a client->OST data error */
1633 if (i == 0 && opc == OST_WRITE &&
1634 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1635 int off = BD_GET_KIOV(desc, i).kiov_offset &
1637 int len = BD_GET_KIOV(desc, i).kiov_len;
1638 struct page *np = tgt_page_to_corrupt;
1639 char *ptr = kmap(BD_GET_KIOV(desc, i).kiov_page) + off;
1642 char *ptr2 = kmap(np) + off;
1644 memcpy(ptr2, ptr, len);
1645 memcpy(ptr2, "bad3", min(4, len));
1647 BD_GET_KIOV(desc, i).kiov_page = np;
1649 CERROR("%s: can't alloc page for corruption\n",
1653 cfs_crypto_hash_update_page(hdesc,
1654 BD_GET_KIOV(desc, i).kiov_page,
1655 BD_GET_KIOV(desc, i).kiov_offset &
1657 BD_GET_KIOV(desc, i).kiov_len);
1659 /* corrupt the data after we compute the checksum, to
1660 * simulate an OST->client data error */
1661 if (i == 0 && opc == OST_READ &&
1662 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1663 int off = BD_GET_KIOV(desc, i).kiov_offset
1665 int len = BD_GET_KIOV(desc, i).kiov_len;
1666 struct page *np = tgt_page_to_corrupt;
1668 kmap(BD_GET_KIOV(desc, i).kiov_page) + off;
1671 char *ptr2 = kmap(np) + off;
1673 memcpy(ptr2, ptr, len);
1674 memcpy(ptr2, "bad4", min(4, len));
1676 BD_GET_KIOV(desc, i).kiov_page = np;
1678 CERROR("%s: can't alloc page for corruption\n",
1684 bufsize = sizeof(cksum);
1685 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1690 int tgt_brw_read(struct tgt_session_info *tsi)
1692 struct ptlrpc_request *req = tgt_ses_req(tsi);
1693 struct ptlrpc_bulk_desc *desc = NULL;
1694 struct obd_export *exp = tsi->tsi_exp;
1695 struct niobuf_remote *remote_nb;
1696 struct niobuf_local *local_nb;
1697 struct obd_ioobj *ioo;
1698 struct ost_body *body, *repbody;
1699 struct l_wait_info lwi;
1700 struct lustre_handle lockh = { 0 };
1701 int npages, nob = 0, rc, i, no_reply = 0;
1702 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
1706 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) {
1707 CERROR("%s: deny read request from %s to portal %u\n",
1708 tgt_name(tsi->tsi_tgt),
1709 obd_export_nid2str(req->rq_export),
1710 ptlrpc_req2svc(req)->srv_req_portal);
1714 req->rq_bulk_read = 1;
1716 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
1719 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
1720 cfs_fail_val : (obd_timeout + 1) / 4);
1722 /* Check if there is eviction in progress, and if so, wait for it to
1724 if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
1725 /* We do not care how long it takes */
1726 lwi = LWI_INTR(NULL, NULL);
1727 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
1728 !atomic_read(&exp->exp_obd->obd_evict_inprogress),
1732 /* There must be big cache in current thread to process this request
1733 * if it is NULL then something went wrong and it wasn't allocated,
1734 * report -ENOMEM in that case */
1738 body = tsi->tsi_ost_body;
1739 LASSERT(body != NULL);
1741 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
1742 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
1744 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1745 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
1747 local_nb = tbc->local;
1749 rc = tgt_brw_lock(exp->exp_obd->obd_namespace, &tsi->tsi_resid, ioo,
1750 remote_nb, &lockh, LCK_PR);
1755 * If getting the lock took more time than
1756 * client was willing to wait, drop it. b=11330
1758 if (cfs_time_current_sec() > req->rq_deadline ||
1759 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1761 CERROR("Dropping timed-out read from %s because locking"
1762 "object "DOSTID" took %ld seconds (limit was %ld).\n",
1763 libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
1764 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
1765 req->rq_deadline - req->rq_arrival_time.tv_sec);
1766 GOTO(out_lock, rc = -ETIMEDOUT);
1769 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1770 repbody->oa = body->oa;
1772 npages = PTLRPC_MAX_BRW_PAGES;
1773 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
1774 ioo, remote_nb, &npages, local_nb);
1778 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
1779 PTLRPC_BULK_PUT_SOURCE |
1780 PTLRPC_BULK_BUF_KIOV,
1782 &ptlrpc_bulk_kiov_nopin_ops);
1784 GOTO(out_commitrw, rc = -ENOMEM);
1787 for (i = 0; i < npages; i++) {
1788 int page_rc = local_nb[i].lnb_rc;
1796 if (page_rc != 0) { /* some data! */
1797 LASSERT(local_nb[i].lnb_page != NULL);
1798 desc->bd_frag_ops->add_kiov_frag
1799 (desc, local_nb[i].lnb_page,
1800 local_nb[i].lnb_page_offset,
1804 if (page_rc != local_nb[i].lnb_len) { /* short read */
1805 /* All subsequent pages should be 0 */
1806 while (++i < npages)
1807 LASSERT(local_nb[i].lnb_rc == 0);
1812 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1813 cksum_type_t cksum_type =
1814 cksum_type_unpack(body->oa.o_valid & OBD_MD_FLFLAGS ?
1815 body->oa.o_flags : 0);
1816 repbody->oa.o_flags = cksum_type_pack(cksum_type);
1817 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1818 repbody->oa.o_cksum = tgt_checksum_bulk(tsi->tsi_tgt, desc,
1819 OST_READ, cksum_type);
1820 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
1821 repbody->oa.o_cksum);
1823 repbody->oa.o_valid = 0;
1825 /* We're finishing using body->oa as an input variable */
1827 /* Check if client was evicted while we were doing i/o before touching
1829 if (likely(rc == 0 &&
1830 !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) {
1831 rc = target_bulk_io(exp, desc, &lwi);
1836 /* Must commit after prep above in all cases */
1837 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1, ioo,
1838 remote_nb, npages, local_nb, rc);
1840 tgt_drop_id(exp, &repbody->oa);
1842 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
1844 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
1845 ptlrpc_free_bulk(desc);
1850 ptlrpc_lprocfs_brw(req, nob);
1851 } else if (no_reply) {
1852 req->rq_no_reply = 1;
1853 /* reply out callback would free */
1854 ptlrpc_req_drop_rs(req);
1855 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
1856 "client will retry: rc %d\n",
1857 exp->exp_obd->obd_name,
1858 obd_uuid2str(&exp->exp_client_uuid),
1859 obd_export_nid2str(exp), rc);
1861 /* send a bulk after reply to simulate a network delay or reordering
1863 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) {
1864 wait_queue_head_t waitq;
1865 struct l_wait_info lwi1;
1867 CDEBUG(D_INFO, "reorder BULK\n");
1868 init_waitqueue_head(&waitq);
1870 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
1871 l_wait_event(waitq, 0, &lwi1);
1872 target_bulk_io(exp, desc, &lwi);
1873 ptlrpc_free_bulk(desc);
1878 EXPORT_SYMBOL(tgt_brw_read);
1880 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
1881 struct ptlrpc_bulk_desc *desc,
1882 struct niobuf_local *local_nb, int npages,
1883 u32 client_cksum, u32 server_cksum,
1886 struct obd_export *exp = req->rq_export;
1887 struct ost_body *body;
1891 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1892 LASSERT(body != NULL);
1894 if (req->rq_peer.nid != desc->bd_sender) {
1896 router = libcfs_nid2str(desc->bd_sender);
1900 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
1901 client_cksum, server_cksum);
1905 LCONSOLE_ERROR_MSG(0x168, "BAD WRITE CHECKSUM: %s from %s%s%s inode "
1906 DFID" object "DOSTID" extent ["LPU64"-"LPU64
1907 "]: client csum %x, server csum %x\n",
1908 exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
1910 body->oa.o_valid & OBD_MD_FLFID ?
1911 body->oa.o_parent_seq : (__u64)0,
1912 body->oa.o_valid & OBD_MD_FLFID ?
1913 body->oa.o_parent_oid : 0,
1914 body->oa.o_valid & OBD_MD_FLFID ?
1915 body->oa.o_parent_ver : 0,
1916 POSTID(&body->oa.o_oi),
1917 local_nb[0].lnb_file_offset,
1918 local_nb[npages-1].lnb_file_offset +
1919 local_nb[npages - 1].lnb_len - 1,
1920 client_cksum, server_cksum);
1923 int tgt_brw_write(struct tgt_session_info *tsi)
1925 struct ptlrpc_request *req = tgt_ses_req(tsi);
1926 struct ptlrpc_bulk_desc *desc = NULL;
1927 struct obd_export *exp = req->rq_export;
1928 struct niobuf_remote *remote_nb;
1929 struct niobuf_local *local_nb;
1930 struct obd_ioobj *ioo;
1931 struct ost_body *body, *repbody;
1932 struct l_wait_info lwi;
1933 struct lustre_handle lockh = {0};
1935 int objcount, niocount, npages;
1937 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
1938 bool no_reply = false, mmap;
1939 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
1943 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) {
1944 CERROR("%s: deny write request from %s to portal %u\n",
1945 tgt_name(tsi->tsi_tgt),
1946 obd_export_nid2str(req->rq_export),
1947 ptlrpc_req2svc(req)->srv_req_portal);
1948 RETURN(err_serious(-EPROTO));
1951 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
1952 RETURN(err_serious(-ENOSPC));
1953 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
1954 RETURN(err_serious(-EROFS));
1956 req->rq_bulk_write = 1;
1958 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
1959 RETURN(err_serious(-EIO));
1960 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
1961 RETURN(err_serious(-EFAULT));
1963 /* pause before transaction has been started */
1964 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, cfs_fail_val > 0 ?
1965 cfs_fail_val : (obd_timeout + 1) / 4);
1967 /* There must be big cache in current thread to process this request
1968 * if it is NULL then something went wrong and it wasn't allocated,
1969 * report -ENOMEM in that case */
1973 body = tsi->tsi_ost_body;
1974 LASSERT(body != NULL);
1976 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1977 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
1979 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
1980 RCL_CLIENT) / sizeof(*ioo);
1982 for (niocount = i = 0; i < objcount; i++)
1983 niocount += ioo[i].ioo_bufcnt;
1985 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1986 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
1987 if (niocount != req_capsule_get_size(&req->rq_pill,
1988 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
1990 RETURN(err_serious(-EPROTO));
1992 if ((remote_nb[0].rnb_flags & OBD_BRW_MEMALLOC) &&
1993 (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self))
1994 memory_pressure_set();
1996 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
1997 niocount * sizeof(*rcs));
1998 rc = req_capsule_server_pack(&req->rq_pill);
2000 GOTO(out, rc = err_serious(rc));
2002 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
2003 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2005 local_nb = tbc->local;
2007 rc = tgt_brw_lock(exp->exp_obd->obd_namespace, &tsi->tsi_resid, ioo,
2008 remote_nb, &lockh, LCK_PW);
2013 * If getting the lock took more time than
2014 * client was willing to wait, drop it. b=11330
2016 if (cfs_time_current_sec() > req->rq_deadline ||
2017 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
2019 CERROR("%s: Dropping timed-out write from %s because locking "
2020 "object "DOSTID" took %ld seconds (limit was %ld).\n",
2021 tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
2022 POSTID(&ioo->ioo_oid),
2023 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
2024 req->rq_deadline - req->rq_arrival_time.tv_sec);
2025 GOTO(out_lock, rc = -ETIMEDOUT);
2028 /* Because we already sync grant info with client when reconnect,
2029 * grant info will be cleared for resent req, then fed_grant and
2030 * total_grant will not be modified in following preprw_write */
2031 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
2032 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
2033 body->oa.o_valid &= ~OBD_MD_FLGRANT;
2036 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
2037 if (repbody == NULL)
2038 GOTO(out_lock, rc = -ENOMEM);
2039 repbody->oa = body->oa;
2041 npages = PTLRPC_MAX_BRW_PAGES;
2042 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2043 objcount, ioo, remote_nb, &npages, local_nb);
2047 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
2048 PTLRPC_BULK_GET_SINK | PTLRPC_BULK_BUF_KIOV,
2050 &ptlrpc_bulk_kiov_nopin_ops);
2052 GOTO(skip_transfer, rc = -ENOMEM);
2054 /* NB Having prepped, we must commit... */
2055 for (i = 0; i < npages; i++)
2056 desc->bd_frag_ops->add_kiov_frag(desc,
2057 local_nb[i].lnb_page,
2058 local_nb[i].lnb_page_offset,
2059 local_nb[i].lnb_len);
2061 rc = sptlrpc_svc_prep_bulk(req, desc);
2063 GOTO(skip_transfer, rc);
2065 rc = target_bulk_io(exp, desc, &lwi);
2069 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
2070 static int cksum_counter;
2072 if (body->oa.o_valid & OBD_MD_FLFLAGS)
2073 cksum_type = cksum_type_unpack(body->oa.o_flags);
2075 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
2076 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
2077 repbody->oa.o_flags |= cksum_type_pack(cksum_type);
2078 repbody->oa.o_cksum = tgt_checksum_bulk(tsi->tsi_tgt, desc,
2079 OST_WRITE, cksum_type);
2082 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
2083 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
2084 body->oa.o_flags & OBD_FL_MMAP);
2086 tgt_warn_on_cksum(req, desc, local_nb, npages,
2088 repbody->oa.o_cksum, mmap);
2090 } else if ((cksum_counter & (-cksum_counter)) ==
2092 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
2093 cksum_counter, libcfs_id2str(req->rq_peer),
2094 repbody->oa.o_cksum);
2098 /* Must commit after prep above in all cases */
2099 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
2100 objcount, ioo, remote_nb, npages, local_nb, rc);
2101 if (rc == -ENOTCONN)
2102 /* quota acquire process has been given up because
2103 * either the client has been evicted or the client
2104 * has timed out the request already */
2108 * Disable sending mtime back to the client. If the client locked the
2109 * whole object, then it has already updated the mtime on its side,
2110 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
2112 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
2117 /* set per-requested niobuf return codes */
2118 for (i = j = 0; i < niocount; i++) {
2119 int len = remote_nb[i].rnb_len;
2124 LASSERT(j < npages);
2125 if (local_nb[j].lnb_rc < 0)
2126 rcs[i] = local_nb[j].lnb_rc;
2127 len -= local_nb[j].lnb_len;
2132 LASSERT(j == npages);
2133 ptlrpc_lprocfs_brw(req, nob);
2135 tgt_drop_id(exp, &repbody->oa);
2138 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
2140 ptlrpc_free_bulk(desc);
2143 req->rq_no_reply = 1;
2144 /* reply out callback would free */
2145 ptlrpc_req_drop_rs(req);
2146 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s), "
2147 "client will retry: rc %d\n",
2148 exp->exp_obd->obd_name,
2149 obd_uuid2str(&exp->exp_client_uuid),
2150 obd_export_nid2str(exp), rc);
2152 memory_pressure_clr();
2155 EXPORT_SYMBOL(tgt_brw_write);