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, 2012, 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
38 #include <obd_class.h>
39 #include <obd_cksum.h>
41 #include "tgt_internal.h"
43 char *tgt_name(struct lu_target *tgt)
45 LASSERT(tgt->lut_obd != NULL);
46 return tgt->lut_obd->obd_name;
48 EXPORT_SYMBOL(tgt_name);
51 * Generic code handling requests that have struct mdt_body passed in:
53 * - extract mdt_body from request and save it in @tsi, if present;
55 * - create lu_object, corresponding to the fid in mdt_body, and save it in
58 * - if HABEO_CORPUS flag is set for this request type check whether object
59 * actually exists on storage (lu_object_exists()).
62 static int tgt_mdt_body_unpack(struct tgt_session_info *tsi, __u32 flags)
64 const struct mdt_body *body;
65 struct lu_object *obj;
66 struct req_capsule *pill = tsi->tsi_pill;
71 body = req_capsule_client_get(pill, &RMF_MDT_BODY);
75 tsi->tsi_mdt_body = body;
77 if (!(body->valid & OBD_MD_FLID))
80 /* mdc_pack_body() doesn't check if fid is zero and set OBD_ML_FID
81 * in any case in pre-2.5 clients. Fix that here if needed */
82 if (unlikely(fid_is_zero(&body->fid1)))
85 if (!fid_is_sane(&body->fid1)) {
86 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
91 obj = lu_object_find(tsi->tsi_env,
92 &tsi->tsi_tgt->lut_bottom->dd_lu_dev,
95 if ((flags & HABEO_CORPUS) && !lu_object_exists(obj)) {
96 lu_object_put(tsi->tsi_env, obj);
97 /* for capability renew ENOENT will be handled in
99 if (body->valid & OBD_MD_FLOSSCAPA)
104 tsi->tsi_corpus = obj;
114 * Validate oa from client.
115 * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
117 * a. objects in Single MDT FS seq = FID_SEQ_OST_MDT0, oi_id != 0
118 * b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to
119 * pack ost_id. Because non-zero oi_seq will make it diffcult to tell
120 * whether this is oi_fid or real ostid. So it will check
121 * OBD_CONNECT_FID, then convert the ostid to FID for old client.
122 * c. Old FID-disable osc will send IDIF.
123 * d. new FID-enable osc/osp will send normal FID.
125 * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will
126 * be used for LAST_ID file, and only being accessed inside OST now.
128 int tgt_validate_obdo(struct tgt_session_info *tsi, struct obdo *oa)
134 if (unlikely(!(exp_connect_flags(tsi->tsi_exp) & OBD_CONNECT_FID) &&
135 fid_seq_is_echo(oa->o_oi.oi.oi_seq))) {
136 /* Sigh 2.[123] client still sends echo req with oi_id = 0
137 * during create, and we will reset this to 1, since this
138 * oi_id is basically useless in the following create process,
139 * but oi_id == 0 will make it difficult to tell whether it is
140 * real FID or ost_id. */
141 oa->o_oi.oi_fid.f_oid = oa->o_oi.oi.oi_id ?: 1;
142 oa->o_oi.oi_fid.f_seq = FID_SEQ_ECHO;
143 oa->o_oi.oi_fid.f_ver = 0;
145 if (unlikely((oa->o_valid & OBD_MD_FLID &&
146 ostid_id(&oa->o_oi) == 0)))
147 GOTO(out, rc = -EPROTO);
149 /* Note: this check might be forced in 2.5 or 2.6, i.e.
150 * all of the requests are required to setup FLGROUP */
151 if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) {
152 ostid_set_seq_mdt0(&oa->o_oi);
153 oa->o_valid |= OBD_MD_FLGROUP;
156 if (unlikely(!(fid_seq_is_idif(ostid_seq(&oa->o_oi)) ||
157 fid_seq_is_mdt0(ostid_seq(&oa->o_oi)) ||
158 fid_seq_is_norm(ostid_seq(&oa->o_oi)) ||
159 fid_seq_is_echo(ostid_seq(&oa->o_oi)))))
160 GOTO(out, rc = -EPROTO);
164 CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n",
165 tgt_name(tsi->tsi_tgt), obd_export_nid2str(tsi->tsi_exp),
166 ostid_seq(&oa->o_oi), ostid_id(&oa->o_oi), rc);
169 EXPORT_SYMBOL(tgt_validate_obdo);
171 static int tgt_ost_body_unpack(struct tgt_session_info *tsi, __u32 flags)
173 struct ost_body *body;
174 struct req_capsule *pill = tsi->tsi_pill;
175 struct lustre_capa *capa;
176 struct obd_ioobj *ioo;
181 body = req_capsule_client_get(pill, &RMF_OST_BODY);
185 rc = tgt_validate_obdo(tsi, &body->oa);
189 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
190 capa = req_capsule_client_get(tsi->tsi_pill, &RMF_CAPA1);
192 CERROR("%s: OSSCAPA flag is set without capability\n",
193 tgt_name(tsi->tsi_tgt));
198 tsi->tsi_ost_body = body;
200 if (req_capsule_has_field(pill, &RMF_OBD_IOOBJ, RCL_CLIENT)) {
202 struct niobuf_remote *rnb;
204 ioo = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
208 rnb = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
212 max_brw = ioobj_max_brw_get(ioo);
213 if (unlikely((max_brw & (max_brw - 1)) != 0)) {
214 CERROR("%s: client %s sent bad ioobj max %u for "DOSTID
215 ": rc = %d\n", tgt_name(tsi->tsi_tgt),
216 obd_export_nid2str(tsi->tsi_exp), max_brw,
217 POSTID(&body->oa.o_oi), -EPROTO);
220 ioo->ioo_oid = body->oa.o_oi;
223 if (!(body->oa.o_valid & OBD_MD_FLID)) {
224 if (flags & HABEO_CORPUS) {
225 CERROR("%s: OBD_MD_FLID flag is not set in ost_body "
226 "but OID/FID is mandatory with HABEO_CORPUS\n",
227 tgt_name(tsi->tsi_tgt));
234 rc = ostid_to_fid(&tsi->tsi_fid, &body->oa.o_oi, 0);
238 if (!fid_is_sane(&tsi->tsi_fid)) {
239 CERROR("%s: invalid FID: "DFID"\n", tgt_name(tsi->tsi_tgt),
240 PFID(&tsi->tsi_fid));
244 ost_fid_build_resid(&tsi->tsi_fid, &tsi->tsi_resid);
247 * OST doesn't get object in advance for further use to prevent
248 * situations with nested object_find which is potential deadlock.
250 tsi->tsi_corpus = NULL;
254 static int tgt_unpack_req_pack_rep(struct tgt_session_info *tsi, __u32 flags)
256 struct req_capsule *pill = tsi->tsi_pill;
261 if (req_capsule_has_field(pill, &RMF_MDT_BODY, RCL_CLIENT)) {
262 rc = tgt_mdt_body_unpack(tsi, flags);
263 } else if (req_capsule_has_field(pill, &RMF_OST_BODY, RCL_CLIENT)) {
264 rc = tgt_ost_body_unpack(tsi, flags);
269 if (rc == 0 && flags & HABEO_REFERO) {
271 if (req_capsule_has_field(pill, &RMF_MDT_MD, RCL_SERVER))
272 req_capsule_set_size(pill, &RMF_MDT_MD, RCL_SERVER,
273 tsi->tsi_mdt_body->eadatasize);
274 if (req_capsule_has_field(pill, &RMF_LOGCOOKIES, RCL_SERVER))
275 req_capsule_set_size(pill, &RMF_LOGCOOKIES,
278 rc = req_capsule_server_pack(pill);
284 * Invoke handler for this request opc. Also do necessary preprocessing
285 * (according to handler ->th_flags), and post-processing (setting of
286 * ->last_{xid,committed}).
288 static int tgt_handle_request0(struct tgt_session_info *tsi,
289 struct tgt_handler *h,
290 struct ptlrpc_request *req)
298 LASSERT(h->th_act != NULL);
299 LASSERT(h->th_opc == lustre_msg_get_opc(req->rq_reqmsg));
300 LASSERT(current->journal_info == NULL);
303 * Checking for various OBD_FAIL_$PREF_$OPC_NET codes. _Do_ not try
304 * to put same checks into handlers like mdt_close(), mdt_reint(),
305 * etc., without talking to mdt authors first. Checking same thing
306 * there again is useless and returning 0 error without packing reply
307 * is buggy! Handlers either pack reply or return error.
309 * We return 0 here and do not send any reply in order to emulate
310 * network failure. Do not send any reply in case any of NET related
311 * fail_id has occured.
313 if (OBD_FAIL_CHECK_ORSET(h->th_fail_id, OBD_FAIL_ONCE))
318 LASSERT(ergo(flags & (HABEO_CORPUS | HABEO_REFERO),
320 if (h->th_fmt != NULL) {
321 req_capsule_set(tsi->tsi_pill, h->th_fmt);
322 rc = tgt_unpack_req_pack_rep(tsi, flags);
325 if (rc == 0 && flags & MUTABOR &&
326 tgt_conn_flags(tsi) & OBD_CONNECT_RDONLY)
329 if (rc == 0 && flags & HABEO_CLAVIS) {
330 struct ldlm_request *dlm_req;
332 LASSERT(h->th_fmt != NULL);
334 dlm_req = req_capsule_client_get(tsi->tsi_pill, &RMF_DLM_REQ);
335 if (dlm_req != NULL) {
336 if (unlikely(dlm_req->lock_desc.l_resource.lr_type ==
338 dlm_req->lock_desc.l_policy_data.\
339 l_inodebits.bits == 0)) {
341 * Lock without inodebits makes no sense and
342 * will oops later in ldlm. If client miss to
343 * set such bits, do not trigger ASSERTION.
345 * For liblustre flock case, it maybe zero.
349 tsi->tsi_dlm_req = dlm_req;
356 if (likely(rc == 0)) {
358 * Process request, there can be two types of rc:
359 * 1) errors with msg unpack/pack, other failures outside the
360 * operation itself. This is counted as serious errors;
361 * 2) errors during fs operation, should be placed in rq_status
365 if (!is_serious(rc) &&
366 !req->rq_no_reply && req->rq_reply_state == NULL) {
367 DEBUG_REQ(D_ERROR, req, "%s \"handler\" %s did not "
368 "pack reply and returned 0 error\n",
369 tgt_name(tsi->tsi_tgt), h->th_name);
372 serious = is_serious(rc);
373 rc = clear_serious(rc);
381 * ELDLM_* codes which > 0 should be in rq_status only as well as
382 * all non-serious errors.
384 if (rc > 0 || !serious)
387 LASSERT(current->journal_info == NULL);
389 if (likely(rc == 0 && req->rq_export))
390 target_committed_to_req(req);
392 target_send_reply(req, rc, tsi->tsi_reply_fail_id);
396 static int tgt_filter_recovery_request(struct ptlrpc_request *req,
397 struct obd_device *obd, int *process)
399 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
406 case MDS_DONE_WRITING:
407 case MDS_SYNC: /* used in unmounting */
420 *process = target_queue_recovery_request(req, obd);
424 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
431 * Handle recovery. Return:
432 * +1: continue request processing;
433 * -ve: abort immediately with the given error code;
434 * 0: send reply with error code in req->rq_status;
436 int tgt_handle_recovery(struct ptlrpc_request *req, int reply_fail_id)
440 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
445 case SEC_CTX_INIT_CONT:
450 if (!req->rq_export->exp_obd->obd_replayable)
453 /* sanity check: if the xid matches, the request must be marked as a
454 * resent or replayed */
455 if (req_xid_is_last(req)) {
456 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
457 (MSG_RESENT | MSG_REPLAY))) {
458 DEBUG_REQ(D_WARNING, req, "rq_xid "LPU64" matches "
459 "last_xid, expected REPLAY or RESENT flag "
461 lustre_msg_get_flags(req->rq_reqmsg));
462 req->rq_status = -ENOTCONN;
466 /* else: note the opposite is not always true; a RESENT req after a
467 * failover will usually not match the last_xid, since it was likely
468 * never committed. A REPLAYed request will almost never match the
469 * last xid, however it could for a committed, but still retained,
472 /* Check for aborted recovery... */
473 if (unlikely(req->rq_export->exp_obd->obd_recovering)) {
477 DEBUG_REQ(D_INFO, req, "Got new replay");
478 rc = tgt_filter_recovery_request(req, req->rq_export->exp_obd,
480 if (rc != 0 || !should_process)
482 else if (should_process < 0) {
483 req->rq_status = should_process;
484 rc = ptlrpc_error(req);
491 int tgt_request_handle(struct ptlrpc_request *req)
493 struct tgt_session_info *tsi = tgt_ses_info(req->rq_svc_thread->t_env);
495 struct lustre_msg *msg = req->rq_reqmsg;
496 struct tgt_handler *h;
497 struct tgt_opc_slice *s;
498 struct lu_target *tgt;
499 int request_fail_id = 0;
500 __u32 opc = lustre_msg_get_opc(msg);
505 /* Refill the context, to make sure all thread keys are allocated */
506 lu_env_refill(req->rq_svc_thread->t_env);
508 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
509 tsi->tsi_pill = &req->rq_pill;
510 tsi->tsi_env = req->rq_svc_thread->t_env;
512 /* if request has export then get handlers slice from corresponding
513 * target, otherwise that should be connect operation */
514 if (opc == MDS_CONNECT || opc == OST_CONNECT ||
515 opc == MGS_CONNECT) {
516 req_capsule_set(&req->rq_pill, &RQF_CONNECT);
517 rc = target_handle_connect(req);
519 rc = ptlrpc_error(req);
522 /* recovery-small test 18c asks to drop connect reply */
523 if (unlikely(opc == OST_CONNECT &&
524 OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2)))
528 if (unlikely(!class_connected_export(req->rq_export))) {
529 CDEBUG(D_HA, "operation %d on unconnected OST from %s\n",
530 opc, libcfs_id2str(req->rq_peer));
531 req->rq_status = -ENOTCONN;
532 rc = ptlrpc_error(req);
536 tsi->tsi_tgt = tgt = class_exp2tgt(req->rq_export);
537 tsi->tsi_exp = req->rq_export;
538 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)
539 tsi->tsi_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
541 tsi->tsi_jobid = NULL;
543 request_fail_id = tgt->lut_request_fail_id;
544 tsi->tsi_reply_fail_id = tgt->lut_reply_fail_id;
546 for (s = tgt->lut_slice; s->tos_hs != NULL; s++)
547 if (s->tos_opc_start <= opc && opc < s->tos_opc_end)
550 /* opcode was not found in slice */
551 if (unlikely(s->tos_hs == NULL)) {
552 CERROR("%s: no handlers for opcode 0x%x\n", tgt_name(tgt), opc);
553 req->rq_status = -ENOTSUPP;
554 rc = ptlrpc_error(req);
558 if (CFS_FAIL_CHECK_ORSET(request_fail_id, CFS_FAIL_ONCE))
561 LASSERT(current->journal_info == NULL);
563 LASSERT(opc >= s->tos_opc_start && opc < s->tos_opc_end);
564 h = s->tos_hs + (opc - s->tos_opc_start);
565 if (unlikely(h->th_opc == 0)) {
566 CERROR("%s: unsupported opcode 0x%x\n", tgt_name(tgt), opc);
567 req->rq_status = -ENOTSUPP;
568 rc = ptlrpc_error(req);
572 rc = lustre_msg_check_version(msg, h->th_version);
574 DEBUG_REQ(D_ERROR, req, "%s: drop mal-formed request, version"
575 " %08x, expecting %08x\n", tgt_name(tgt),
576 lustre_msg_get_version(msg), h->th_version);
577 req->rq_status = -EINVAL;
578 rc = ptlrpc_error(req);
582 rc = tgt_handle_recovery(req, tsi->tsi_reply_fail_id);
583 if (likely(rc == 1)) {
584 LASSERTF(h->th_opc == opc, "opcode mismatch %d != %d\n",
586 rc = tgt_handle_request0(tsi, h, req);
592 req_capsule_fini(tsi->tsi_pill);
593 tsi->tsi_pill = NULL;
594 if (tsi->tsi_corpus != NULL) {
595 lu_object_put(tsi->tsi_env, tsi->tsi_corpus);
596 tsi->tsi_corpus = NULL;
599 tsi->tsi_mdt_body = NULL;
600 tsi->tsi_dlm_req = NULL;
601 fid_zero(&tsi->tsi_fid);
602 memset(&tsi->tsi_resid, 0, sizeof tsi->tsi_resid);
605 EXPORT_SYMBOL(tgt_request_handle);
607 void tgt_counter_incr(struct obd_export *exp, int opcode)
609 lprocfs_counter_incr(exp->exp_obd->obd_stats, opcode);
610 if (exp->exp_nid_stats && exp->exp_nid_stats->nid_stats != NULL)
611 lprocfs_counter_incr(exp->exp_nid_stats->nid_stats, opcode);
613 EXPORT_SYMBOL(tgt_counter_incr);
616 * Unified target generic handlers.
622 static inline void tgt_init_sec_none(struct obd_connect_data *reply)
624 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT |
625 OBD_CONNECT_RMT_CLIENT_FORCE |
626 OBD_CONNECT_MDS_CAPA |
627 OBD_CONNECT_OSS_CAPA);
630 static int tgt_init_sec_level(struct ptlrpc_request *req)
632 struct lu_target *tgt = class_exp2tgt(req->rq_export);
633 char *client = libcfs_nid2str(req->rq_peer.nid);
634 struct obd_connect_data *data, *reply;
640 data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
641 reply = req_capsule_server_get(&req->rq_pill, &RMF_CONNECT_DATA);
642 if (data == NULL || reply == NULL)
645 /* connection from MDT is always trusted */
646 if (req->rq_auth_usr_mdt) {
647 tgt_init_sec_none(reply);
651 /* no GSS support case */
652 if (!req->rq_auth_gss) {
653 if (tgt->lut_sec_level > LUSTRE_SEC_NONE) {
654 CWARN("client %s -> target %s does not use GSS, "
655 "can not run under security level %d.\n",
656 client, tgt_name(tgt), tgt->lut_sec_level);
659 tgt_init_sec_none(reply);
664 /* old version case */
665 if (unlikely(!(data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT) ||
666 !(data->ocd_connect_flags & OBD_CONNECT_MDS_CAPA) ||
667 !(data->ocd_connect_flags & OBD_CONNECT_OSS_CAPA))) {
668 if (tgt->lut_sec_level > LUSTRE_SEC_NONE) {
669 CWARN("client %s -> target %s uses old version, "
670 "can not run under security level %d.\n",
671 client, tgt_name(tgt), tgt->lut_sec_level);
674 CWARN("client %s -> target %s uses old version, "
675 "run under security level %d.\n",
676 client, tgt_name(tgt), tgt->lut_sec_level);
677 tgt_init_sec_none(reply);
682 remote = data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT_FORCE;
684 if (!req->rq_auth_remote)
685 CDEBUG(D_SEC, "client (local realm) %s -> target %s "
686 "asked to be remote.\n", client, tgt_name(tgt));
687 } else if (req->rq_auth_remote) {
689 CDEBUG(D_SEC, "client (remote realm) %s -> target %s is set "
690 "as remote by default.\n", client, tgt_name(tgt));
694 if (!tgt->lut_oss_capa) {
696 "client %s -> target %s is set as remote,"
697 " but OSS capabilities are not enabled: %d.\n",
698 client, tgt_name(tgt), tgt->lut_oss_capa);
702 if (req->rq_auth_uid == INVALID_UID) {
703 CDEBUG(D_SEC, "client %s -> target %s: user is not "
704 "authenticated!\n", client, tgt_name(tgt));
710 switch (tgt->lut_sec_level) {
711 case LUSTRE_SEC_NONE:
714 "client %s -> target %s is set as remote, "
715 "can not run under security level %d.\n",
716 client, tgt_name(tgt), tgt->lut_sec_level);
719 tgt_init_sec_none(reply);
721 case LUSTRE_SEC_REMOTE:
723 tgt_init_sec_none(reply);
728 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT |
729 OBD_CONNECT_RMT_CLIENT_FORCE);
730 if (!tgt->lut_oss_capa)
731 reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA;
732 if (!tgt->lut_mds_capa)
733 reply->ocd_connect_flags &= ~OBD_CONNECT_MDS_CAPA;
742 int tgt_connect_check_sptlrpc(struct ptlrpc_request *req, struct obd_export *exp)
744 struct lu_target *tgt = class_exp2tgt(exp);
745 struct sptlrpc_flavor flvr;
749 LASSERT(tgt->lut_obd);
750 LASSERT(tgt->lut_slice);
752 /* always allow ECHO client */
753 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
754 LUSTRE_ECHO_NAME) == 0)) {
755 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
759 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
760 read_lock(&tgt->lut_sptlrpc_lock);
761 sptlrpc_target_choose_flavor(&tgt->lut_sptlrpc_rset,
765 read_unlock(&tgt->lut_sptlrpc_lock);
767 spin_lock(&exp->exp_lock);
768 exp->exp_sp_peer = req->rq_sp_from;
769 exp->exp_flvr = flvr;
770 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
771 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
772 CERROR("%s: unauthorized rpc flavor %x from %s, "
773 "expect %x\n", tgt_name(tgt),
775 libcfs_nid2str(req->rq_peer.nid),
776 exp->exp_flvr.sf_rpc);
779 spin_unlock(&exp->exp_lock);
781 if (exp->exp_sp_peer != req->rq_sp_from) {
782 CERROR("%s: RPC source %s doesn't match %s\n",
784 sptlrpc_part2name(req->rq_sp_from),
785 sptlrpc_part2name(exp->exp_sp_peer));
788 rc = sptlrpc_target_export_check(exp, req);
795 int tgt_adapt_sptlrpc_conf(struct lu_target *tgt, int initial)
797 struct sptlrpc_rule_set tmp_rset;
800 sptlrpc_rule_set_init(&tmp_rset);
801 rc = sptlrpc_conf_target_get_rules(tgt->lut_obd, &tmp_rset, initial);
803 CERROR("%s: failed get sptlrpc rules: rc = %d\n",
808 sptlrpc_target_update_exp_flavor(tgt->lut_obd, &tmp_rset);
810 write_lock(&tgt->lut_sptlrpc_lock);
811 sptlrpc_rule_set_free(&tgt->lut_sptlrpc_rset);
812 tgt->lut_sptlrpc_rset = tmp_rset;
813 write_unlock(&tgt->lut_sptlrpc_lock);
817 EXPORT_SYMBOL(tgt_adapt_sptlrpc_conf);
819 int tgt_connect(struct tgt_session_info *tsi)
821 struct ptlrpc_request *req = tgt_ses_req(tsi);
822 struct obd_connect_data *reply;
827 rc = tgt_init_sec_level(req);
831 /* XXX: better to call this check right after getting new export but
832 * before last_rcvd slot allocation to avoid server load upon insecure
833 * connects. This is to be fixed after unifiyng all targets.
835 rc = tgt_connect_check_sptlrpc(req, tsi->tsi_exp);
839 /* To avoid exposing partially initialized connection flags, changes up
840 * to this point have been staged in reply->ocd_connect_flags. Now that
841 * connection handling has completed successfully, atomically update
842 * the connect flags in the shared export data structure. LU-1623 */
843 reply = req_capsule_server_get(tsi->tsi_pill, &RMF_CONNECT_DATA);
844 spin_lock(&tsi->tsi_exp->exp_lock);
845 *exp_connect_flags_ptr(tsi->tsi_exp) = reply->ocd_connect_flags;
846 tsi->tsi_exp->exp_connect_data.ocd_brw_size = reply->ocd_brw_size;
847 spin_unlock(&tsi->tsi_exp->exp_lock);
851 obd_disconnect(class_export_get(tsi->tsi_exp));
854 EXPORT_SYMBOL(tgt_connect);
856 int tgt_disconnect(struct tgt_session_info *tsi)
862 rc = target_handle_disconnect(tgt_ses_req(tsi));
864 RETURN(err_serious(rc));
868 EXPORT_SYMBOL(tgt_disconnect);
871 * Unified target OBD handlers
873 int tgt_obd_ping(struct tgt_session_info *tsi)
879 rc = target_handle_ping(tgt_ses_req(tsi));
881 RETURN(err_serious(rc));
885 EXPORT_SYMBOL(tgt_obd_ping);
887 int tgt_obd_log_cancel(struct tgt_session_info *tsi)
889 return err_serious(-EOPNOTSUPP);
891 EXPORT_SYMBOL(tgt_obd_log_cancel);
893 int tgt_obd_qc_callback(struct tgt_session_info *tsi)
895 return err_serious(-EOPNOTSUPP);
897 EXPORT_SYMBOL(tgt_obd_qc_callback);
899 int tgt_sendpage(struct tgt_session_info *tsi, struct lu_rdpg *rdpg, int nob)
901 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
902 struct ptlrpc_request *req = tgt_ses_req(tsi);
903 struct obd_export *exp = req->rq_export;
904 struct ptlrpc_bulk_desc *desc;
905 struct l_wait_info *lwi = &tti->tti_u.rdpg.tti_wait_info;
913 desc = ptlrpc_prep_bulk_exp(req, rdpg->rp_npages, 1, BULK_PUT_SOURCE,
918 if (!(exp_connect_flags(exp) & OBD_CONNECT_BRW_SIZE))
919 /* old client requires reply size in it's PAGE_CACHE_SIZE,
920 * which is rdpg->rp_count */
921 nob = rdpg->rp_count;
923 for (i = 0, tmpcount = nob; i < rdpg->rp_npages && tmpcount > 0;
924 i++, tmpcount -= tmpsize) {
925 tmpsize = min_t(int, tmpcount, PAGE_CACHE_SIZE);
926 ptlrpc_prep_bulk_page_pin(desc, rdpg->rp_pages[i], 0, tmpsize);
929 LASSERT(desc->bd_nob == nob);
930 rc = target_bulk_io(exp, desc, lwi);
931 ptlrpc_free_bulk_pin(desc);
934 EXPORT_SYMBOL(tgt_sendpage);
937 * OBD_IDX_READ handler
939 int tgt_obd_idx_read(struct tgt_session_info *tsi)
941 struct tgt_thread_info *tti = tgt_th_info(tsi->tsi_env);
942 struct lu_rdpg *rdpg = &tti->tti_u.rdpg.tti_rdpg;
943 struct idx_info *req_ii, *rep_ii;
948 memset(rdpg, 0, sizeof(*rdpg));
949 req_capsule_set(tsi->tsi_pill, &RQF_OBD_IDX_READ);
951 /* extract idx_info buffer from request & reply */
952 req_ii = req_capsule_client_get(tsi->tsi_pill, &RMF_IDX_INFO);
953 if (req_ii == NULL || req_ii->ii_magic != IDX_INFO_MAGIC)
954 RETURN(err_serious(-EPROTO));
956 rc = req_capsule_server_pack(tsi->tsi_pill);
958 RETURN(err_serious(rc));
960 rep_ii = req_capsule_server_get(tsi->tsi_pill, &RMF_IDX_INFO);
962 RETURN(err_serious(-EFAULT));
963 rep_ii->ii_magic = IDX_INFO_MAGIC;
965 /* extract hash to start with */
966 rdpg->rp_hash = req_ii->ii_hash_start;
968 /* extract requested attributes */
969 rdpg->rp_attrs = req_ii->ii_attrs;
971 /* check that fid packed in request is valid and supported */
972 if (!fid_is_sane(&req_ii->ii_fid))
974 rep_ii->ii_fid = req_ii->ii_fid;
977 rep_ii->ii_flags = req_ii->ii_flags;
979 /* compute number of pages to allocate, ii_count is the number of 4KB
981 if (req_ii->ii_count <= 0)
982 GOTO(out, rc = -EFAULT);
983 rdpg->rp_count = min_t(unsigned int, req_ii->ii_count << LU_PAGE_SHIFT,
984 exp_max_brw_size(tsi->tsi_exp));
985 rdpg->rp_npages = (rdpg->rp_count + PAGE_CACHE_SIZE -1) >> PAGE_CACHE_SHIFT;
987 /* allocate pages to store the containers */
988 OBD_ALLOC(rdpg->rp_pages, rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
989 if (rdpg->rp_pages == NULL)
990 GOTO(out, rc = -ENOMEM);
991 for (i = 0; i < rdpg->rp_npages; i++) {
992 rdpg->rp_pages[i] = alloc_page(GFP_IOFS);
993 if (rdpg->rp_pages[i] == NULL)
994 GOTO(out, rc = -ENOMEM);
997 /* populate pages with key/record pairs */
998 rc = dt_index_read(tsi->tsi_env, tsi->tsi_tgt->lut_bottom, rep_ii, rdpg);
1002 LASSERTF(rc <= rdpg->rp_count, "dt_index_read() returned more than "
1003 "asked %d > %d\n", rc, rdpg->rp_count);
1005 /* send pages to client */
1006 rc = tgt_sendpage(tsi, rdpg, rc);
1011 if (rdpg->rp_pages) {
1012 for (i = 0; i < rdpg->rp_npages; i++)
1013 if (rdpg->rp_pages[i])
1014 __free_page(rdpg->rp_pages[i]);
1015 OBD_FREE(rdpg->rp_pages,
1016 rdpg->rp_npages * sizeof(rdpg->rp_pages[0]));
1020 EXPORT_SYMBOL(tgt_obd_idx_read);
1022 struct tgt_handler tgt_obd_handlers[] = {
1023 TGT_OBD_HDL (0, OBD_PING, tgt_obd_ping),
1024 TGT_OBD_HDL_VAR(0, OBD_LOG_CANCEL, tgt_obd_log_cancel),
1025 TGT_OBD_HDL_VAR(0, OBD_QC_CALLBACK, tgt_obd_qc_callback),
1026 TGT_OBD_HDL (0, OBD_IDX_READ, tgt_obd_idx_read)
1028 EXPORT_SYMBOL(tgt_obd_handlers);
1030 int tgt_sync(const struct lu_env *env, struct lu_target *tgt,
1031 struct dt_object *obj)
1037 /* if no objid is specified, it means "sync whole filesystem" */
1039 rc = dt_sync(env, tgt->lut_bottom);
1040 } else if (dt_version_get(env, obj) >
1041 tgt->lut_obd->obd_last_committed) {
1042 rc = dt_object_sync(env, obj);
1047 EXPORT_SYMBOL(tgt_sync);
1049 * Unified target DLM handlers.
1052 /* Ensure that data and metadata are synced to the disk when lock is cancelled
1054 int tgt_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1055 void *data, int flag)
1058 struct lu_target *tgt;
1059 struct dt_object *obj;
1065 tgt = class_exp2tgt(lock->l_export);
1067 if (flag == LDLM_CB_CANCELING &&
1068 (lock->l_granted_mode & (LCK_PW | LCK_GROUP)) &&
1069 (tgt->lut_sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1070 (tgt->lut_sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1071 lock->l_flags & LDLM_FL_CBPENDING))) {
1072 rc = lu_env_init(&env, LCT_DT_THREAD);
1073 if (unlikely(rc != 0))
1076 ost_fid_from_resid(&fid, &lock->l_resource->lr_name);
1077 obj = dt_locate(&env, tgt->lut_bottom, &fid);
1079 GOTO(err_env, rc = PTR_ERR(obj));
1081 if (!dt_object_exists(obj))
1082 GOTO(err_put, rc = -ENOENT);
1084 rc = tgt_sync(&env, tgt, obj);
1086 CERROR("%s: sync failed on lock cancel: rc = %d\n",
1090 lu_object_put(&env, &obj->do_lu);
1095 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1099 struct ldlm_callback_suite tgt_dlm_cbs = {
1100 .lcs_completion = ldlm_server_completion_ast,
1101 .lcs_blocking = tgt_blocking_ast,
1102 .lcs_glimpse = ldlm_server_glimpse_ast
1105 int tgt_enqueue(struct tgt_session_info *tsi)
1107 struct ptlrpc_request *req = tgt_ses_req(tsi);
1112 * tsi->tsi_dlm_req was already swapped and (if necessary) converted,
1113 * tsi->tsi_dlm_cbs was set by the *_req_handle() function.
1115 LASSERT(tsi->tsi_dlm_req != NULL);
1116 rc = ldlm_handle_enqueue0(tsi->tsi_exp->exp_obd->obd_namespace, req,
1117 tsi->tsi_dlm_req, &tgt_dlm_cbs);
1119 RETURN(err_serious(rc));
1121 RETURN(req->rq_status);
1123 EXPORT_SYMBOL(tgt_enqueue);
1125 int tgt_convert(struct tgt_session_info *tsi)
1127 struct ptlrpc_request *req = tgt_ses_req(tsi);
1131 LASSERT(tsi->tsi_dlm_req);
1132 rc = ldlm_handle_convert0(req, tsi->tsi_dlm_req);
1134 RETURN(err_serious(rc));
1136 RETURN(req->rq_status);
1138 EXPORT_SYMBOL(tgt_convert);
1140 int tgt_bl_callback(struct tgt_session_info *tsi)
1142 return err_serious(-EOPNOTSUPP);
1144 EXPORT_SYMBOL(tgt_bl_callback);
1146 int tgt_cp_callback(struct tgt_session_info *tsi)
1148 return err_serious(-EOPNOTSUPP);
1150 EXPORT_SYMBOL(tgt_cp_callback);
1152 /* generic LDLM target handler */
1153 struct tgt_handler tgt_dlm_handlers[] = {
1154 TGT_DLM_HDL (HABEO_CLAVIS, LDLM_ENQUEUE, tgt_enqueue),
1155 TGT_DLM_HDL_VAR(HABEO_CLAVIS, LDLM_CONVERT, tgt_convert),
1156 TGT_DLM_HDL_VAR(0, LDLM_BL_CALLBACK, tgt_bl_callback),
1157 TGT_DLM_HDL_VAR(0, LDLM_CP_CALLBACK, tgt_cp_callback)
1159 EXPORT_SYMBOL(tgt_dlm_handlers);
1162 * Unified target LLOG handlers.
1164 int tgt_llog_open(struct tgt_session_info *tsi)
1170 rc = llog_origin_handle_open(tgt_ses_req(tsi));
1174 EXPORT_SYMBOL(tgt_llog_open);
1176 int tgt_llog_close(struct tgt_session_info *tsi)
1182 rc = llog_origin_handle_close(tgt_ses_req(tsi));
1186 EXPORT_SYMBOL(tgt_llog_close);
1189 int tgt_llog_destroy(struct tgt_session_info *tsi)
1195 rc = llog_origin_handle_destroy(tgt_ses_req(tsi));
1199 EXPORT_SYMBOL(tgt_llog_destroy);
1201 int tgt_llog_read_header(struct tgt_session_info *tsi)
1207 rc = llog_origin_handle_read_header(tgt_ses_req(tsi));
1211 EXPORT_SYMBOL(tgt_llog_read_header);
1213 int tgt_llog_next_block(struct tgt_session_info *tsi)
1219 rc = llog_origin_handle_next_block(tgt_ses_req(tsi));
1223 EXPORT_SYMBOL(tgt_llog_next_block);
1225 int tgt_llog_prev_block(struct tgt_session_info *tsi)
1231 rc = llog_origin_handle_prev_block(tgt_ses_req(tsi));
1235 EXPORT_SYMBOL(tgt_llog_prev_block);
1237 /* generic llog target handler */
1238 struct tgt_handler tgt_llog_handlers[] = {
1239 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_CREATE, tgt_llog_open),
1240 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_NEXT_BLOCK, tgt_llog_next_block),
1241 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_READ_HEADER, tgt_llog_read_header),
1242 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_PREV_BLOCK, tgt_llog_prev_block),
1243 TGT_LLOG_HDL (0, LLOG_ORIGIN_HANDLE_DESTROY, tgt_llog_destroy),
1244 TGT_LLOG_HDL_VAR(0, LLOG_ORIGIN_HANDLE_CLOSE, tgt_llog_close),
1246 EXPORT_SYMBOL(tgt_llog_handlers);
1249 * sec context handlers
1251 /* XXX: Implement based on mdt_sec_ctx_handle()? */
1252 int tgt_sec_ctx_handle(struct tgt_session_info *tsi)
1257 struct tgt_handler tgt_sec_ctx_handlers[] = {
1258 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT, tgt_sec_ctx_handle),
1259 TGT_SEC_HDL_VAR(0, SEC_CTX_INIT_CONT, tgt_sec_ctx_handle),
1260 TGT_SEC_HDL_VAR(0, SEC_CTX_FINI, tgt_sec_ctx_handle),
1262 EXPORT_SYMBOL(tgt_sec_ctx_handlers);
1265 * initialize per-thread page pool (bug 5137).
1267 int tgt_io_thread_init(struct ptlrpc_thread *thread)
1269 struct tgt_thread_big_cache *tbc;
1273 LASSERT(thread != NULL);
1274 LASSERT(thread->t_data == NULL);
1276 OBD_ALLOC_LARGE(tbc, sizeof(*tbc));
1279 thread->t_data = tbc;
1282 EXPORT_SYMBOL(tgt_io_thread_init);
1285 * free per-thread pool created by tgt_thread_init().
1287 void tgt_io_thread_done(struct ptlrpc_thread *thread)
1289 struct tgt_thread_big_cache *tbc;
1293 LASSERT(thread != NULL);
1296 * be prepared to handle partially-initialized pools (because this is
1297 * called from ost_io_thread_init() for cleanup.
1299 tbc = thread->t_data;
1301 OBD_FREE_LARGE(tbc, sizeof(*tbc));
1302 thread->t_data = NULL;
1306 EXPORT_SYMBOL(tgt_io_thread_done);
1308 * Helper function for getting server side [start, start+count] DLM lock
1309 * if asked by client.
1311 int tgt_extent_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1312 __u64 start, __u64 end, struct lustre_handle *lh,
1313 int mode, __u64 *flags)
1315 ldlm_policy_data_t policy;
1320 LASSERT(lh != NULL);
1321 LASSERT(ns != NULL);
1322 LASSERT(!lustre_handle_is_used(lh));
1324 policy.l_extent.gid = 0;
1325 policy.l_extent.start = start & CFS_PAGE_MASK;
1328 * If ->o_blocks is EOF it means "lock till the end of the file".
1329 * Otherwise, it's size of an extent or hole being punched (in bytes).
1331 if (end == OBD_OBJECT_EOF || end < start)
1332 policy.l_extent.end = OBD_OBJECT_EOF;
1334 policy.l_extent.end = end | ~CFS_PAGE_MASK;
1336 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_EXTENT, &policy, mode,
1337 flags, ldlm_blocking_ast,
1338 ldlm_completion_ast, ldlm_glimpse_ast,
1339 NULL, 0, LVB_T_NONE, NULL, lh);
1340 RETURN(rc == ELDLM_OK ? 0 : -EIO);
1342 EXPORT_SYMBOL(tgt_extent_lock);
1344 void tgt_extent_unlock(struct lustre_handle *lh, ldlm_mode_t mode)
1346 LASSERT(lustre_handle_is_used(lh));
1347 ldlm_lock_decref(lh, mode);
1349 EXPORT_SYMBOL(tgt_extent_unlock);
1351 int tgt_brw_lock(struct ldlm_namespace *ns, struct ldlm_res_id *res_id,
1352 struct obd_ioobj *obj, struct niobuf_remote *nb,
1353 struct lustre_handle *lh, int mode)
1356 int nrbufs = obj->ioo_bufcnt;
1361 LASSERT(mode == LCK_PR || mode == LCK_PW);
1362 LASSERT(!lustre_handle_is_used(lh));
1364 if (nrbufs == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
1367 for (i = 1; i < nrbufs; i++)
1368 if (!(nb[i].flags & OBD_BRW_SRVLOCK))
1371 RETURN(tgt_extent_lock(ns, res_id, nb[0].offset,
1372 nb[nrbufs - 1].offset + nb[nrbufs - 1].len - 1,
1375 EXPORT_SYMBOL(tgt_brw_lock);
1377 void tgt_brw_unlock(struct obd_ioobj *obj, struct niobuf_remote *niob,
1378 struct lustre_handle *lh, int mode)
1382 LASSERT(mode == LCK_PR || mode == LCK_PW);
1383 LASSERT((obj->ioo_bufcnt > 0 && (niob[0].flags & OBD_BRW_SRVLOCK)) ==
1384 lustre_handle_is_used(lh));
1385 if (lustre_handle_is_used(lh))
1386 tgt_extent_unlock(lh, mode);
1389 EXPORT_SYMBOL(tgt_brw_unlock);
1391 static __u32 tgt_checksum_bulk(struct lu_target *tgt,
1392 struct ptlrpc_bulk_desc *desc, int opc,
1393 cksum_type_t cksum_type)
1395 struct cfs_crypto_hash_desc *hdesc;
1396 unsigned int bufsize;
1398 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
1401 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1402 if (IS_ERR(hdesc)) {
1403 CERROR("%s: unable to initialize checksum hash %s\n",
1404 tgt_name(tgt), cfs_crypto_hash_name(cfs_alg));
1405 return PTR_ERR(hdesc);
1408 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
1409 for (i = 0; i < desc->bd_iov_count; i++) {
1410 /* corrupt the data before we compute the checksum, to
1411 * simulate a client->OST data error */
1412 if (i == 0 && opc == OST_WRITE &&
1413 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
1414 int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
1415 int len = desc->bd_iov[i].kiov_len;
1416 struct page *np = tgt_page_to_corrupt;
1417 char *ptr = kmap(desc->bd_iov[i].kiov_page) + off;
1420 char *ptr2 = kmap(np) + off;
1422 memcpy(ptr2, ptr, len);
1423 memcpy(ptr2, "bad3", min(4, len));
1425 desc->bd_iov[i].kiov_page = np;
1427 CERROR("%s: can't alloc page for corruption\n",
1431 cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page,
1432 desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK,
1433 desc->bd_iov[i].kiov_len);
1435 /* corrupt the data after we compute the checksum, to
1436 * simulate an OST->client data error */
1437 if (i == 0 && opc == OST_READ &&
1438 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
1439 int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
1440 int len = desc->bd_iov[i].kiov_len;
1441 struct page *np = tgt_page_to_corrupt;
1442 char *ptr = kmap(desc->bd_iov[i].kiov_page) + off;
1445 char *ptr2 = kmap(np) + off;
1447 memcpy(ptr2, ptr, len);
1448 memcpy(ptr2, "bad4", min(4, len));
1450 desc->bd_iov[i].kiov_page = np;
1452 CERROR("%s: can't alloc page for corruption\n",
1459 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
1461 cfs_crypto_hash_final(hdesc, NULL, NULL);
1466 int tgt_brw_read(struct tgt_session_info *tsi)
1468 struct ptlrpc_request *req = tgt_ses_req(tsi);
1469 struct ptlrpc_bulk_desc *desc = NULL;
1470 struct obd_export *exp = tsi->tsi_exp;
1471 struct niobuf_remote *remote_nb;
1472 struct niobuf_local *local_nb;
1473 struct obd_ioobj *ioo;
1474 struct ost_body *body, *repbody;
1475 struct l_wait_info lwi;
1476 struct lustre_handle lockh = { 0 };
1477 int niocount, npages, nob = 0, rc, i;
1479 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
1483 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) {
1484 CERROR("%s: deny read request from %s to portal %u\n",
1485 tgt_name(tsi->tsi_tgt),
1486 obd_export_nid2str(req->rq_export),
1487 ptlrpc_req2svc(req)->srv_req_portal);
1491 req->rq_bulk_read = 1;
1493 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
1496 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
1498 /* Check if there is eviction in progress, and if so, wait for it to
1500 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
1501 /* We do not care how long it takes */
1502 lwi = LWI_INTR(NULL, NULL);
1503 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
1504 !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress),
1508 /* There must be big cache in current thread to process this request
1509 * if it is NULL then something went wrong and it wasn't allocated,
1510 * report -ENOMEM in that case */
1514 body = tsi->tsi_ost_body;
1515 LASSERT(body != NULL);
1517 ioo = req_capsule_client_get(tsi->tsi_pill, &RMF_OBD_IOOBJ);
1518 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
1520 niocount = ioo->ioo_bufcnt;
1521 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1522 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
1524 local_nb = tbc->local;
1526 rc = tgt_brw_lock(exp->exp_obd->obd_namespace, &tsi->tsi_resid, ioo,
1527 remote_nb, &lockh, LCK_PR);
1532 * If getting the lock took more time than
1533 * client was willing to wait, drop it. b=11330
1535 if (cfs_time_current_sec() > req->rq_deadline ||
1536 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1538 CERROR("Dropping timed-out read from %s because locking"
1539 "object "DOSTID" took %ld seconds (limit was %ld).\n",
1540 libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid),
1541 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
1542 req->rq_deadline - req->rq_arrival_time.tv_sec);
1543 GOTO(out_lock, rc = -ETIMEDOUT);
1546 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1547 repbody->oa = body->oa;
1549 npages = PTLRPC_MAX_BRW_PAGES;
1550 rc = obd_preprw(tsi->tsi_env, OBD_BRW_READ, exp, &repbody->oa, 1,
1551 ioo, remote_nb, &npages, local_nb, NULL, BYPASS_CAPA);
1555 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
1556 BULK_PUT_SOURCE, OST_BULK_PORTAL);
1558 GOTO(out_commitrw, rc = -ENOMEM);
1561 for (i = 0; i < npages; i++) {
1562 int page_rc = local_nb[i].rc;
1570 if (page_rc != 0) { /* some data! */
1571 LASSERT(local_nb[i].page != NULL);
1572 ptlrpc_prep_bulk_page_nopin(desc, local_nb[i].page,
1573 local_nb[i].lnb_page_offset,
1577 if (page_rc != local_nb[i].len) { /* short read */
1578 /* All subsequent pages should be 0 */
1579 while (++i < npages)
1580 LASSERT(local_nb[i].rc == 0);
1585 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1586 cksum_type_t cksum_type =
1587 cksum_type_unpack(body->oa.o_valid & OBD_MD_FLFLAGS ?
1588 body->oa.o_flags : 0);
1589 repbody->oa.o_flags = cksum_type_pack(cksum_type);
1590 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1591 repbody->oa.o_cksum = tgt_checksum_bulk(tsi->tsi_tgt, desc,
1592 OST_READ, cksum_type);
1593 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
1594 repbody->oa.o_cksum);
1596 repbody->oa.o_valid = 0;
1598 /* We're finishing using body->oa as an input variable */
1600 /* Check if client was evicted while we were doing i/o before touching
1602 if (likely(rc == 0 &&
1603 !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) {
1604 rc = target_bulk_io(exp, desc, &lwi);
1609 /* Must commit after prep above in all cases */
1610 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_READ, exp,
1611 &repbody->oa, 1, ioo, remote_nb, npages, local_nb,
1614 tgt_drop_id(exp, &repbody->oa);
1616 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PR);
1618 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
1619 ptlrpc_free_bulk_nopin(desc);
1624 ptlrpc_lprocfs_brw(req, nob);
1625 } else if (no_reply) {
1626 req->rq_no_reply = 1;
1627 /* reply out callback would free */
1628 ptlrpc_req_drop_rs(req);
1629 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
1630 "client will retry: rc %d\n",
1631 exp->exp_obd->obd_name,
1632 obd_uuid2str(&exp->exp_client_uuid),
1633 obd_export_nid2str(exp), rc);
1635 /* send a bulk after reply to simulate a network delay or reordering
1637 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) {
1638 wait_queue_head_t waitq;
1639 struct l_wait_info lwi1;
1641 CDEBUG(D_INFO, "reorder BULK\n");
1642 init_waitqueue_head(&waitq);
1644 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
1645 l_wait_event(waitq, 0, &lwi1);
1646 target_bulk_io(exp, desc, &lwi);
1647 ptlrpc_free_bulk_nopin(desc);
1652 EXPORT_SYMBOL(tgt_brw_read);
1654 static void tgt_warn_on_cksum(struct ptlrpc_request *req,
1655 struct ptlrpc_bulk_desc *desc,
1656 struct niobuf_local *local_nb, int npages,
1657 obd_count client_cksum, obd_count server_cksum,
1660 struct obd_export *exp = req->rq_export;
1661 struct ost_body *body;
1665 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1666 LASSERT(body != NULL);
1668 if (req->rq_peer.nid == desc->bd_sender) {
1672 router = libcfs_nid2str(desc->bd_sender);
1676 CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n",
1677 client_cksum, server_cksum);
1681 LCONSOLE_ERROR_MSG(0x168, "BAD WRITE CHECKSUM: %s from %s%s%s inode "
1682 DFID" object "DOSTID" extent ["LPU64"-"LPU64
1683 "]: client csum %x, server csum %x\n",
1684 exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer),
1686 body->oa.o_valid & OBD_MD_FLFID ?
1687 body->oa.o_parent_seq : (__u64)0,
1688 body->oa.o_valid & OBD_MD_FLFID ?
1689 body->oa.o_parent_oid : 0,
1690 body->oa.o_valid & OBD_MD_FLFID ?
1691 body->oa.o_parent_ver : 0,
1692 POSTID(&body->oa.o_oi),
1693 local_nb[0].lnb_file_offset,
1694 local_nb[npages-1].lnb_file_offset +
1695 local_nb[npages-1].len - 1,
1696 client_cksum, server_cksum);
1699 int tgt_brw_write(struct tgt_session_info *tsi)
1701 struct ptlrpc_request *req = tgt_ses_req(tsi);
1702 struct ptlrpc_bulk_desc *desc = NULL;
1703 struct obd_export *exp = req->rq_export;
1704 struct niobuf_remote *remote_nb;
1705 struct niobuf_local *local_nb;
1706 struct obd_ioobj *ioo;
1707 struct ost_body *body, *repbody;
1708 struct l_wait_info lwi;
1709 struct lustre_handle lockh = {0};
1711 int objcount, niocount, npages;
1713 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
1714 bool no_reply = false, mmap;
1715 struct tgt_thread_big_cache *tbc = req->rq_svc_thread->t_data;
1719 if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) {
1720 CERROR("%s: deny write request from %s to portal %u\n",
1721 tgt_name(tsi->tsi_tgt),
1722 obd_export_nid2str(req->rq_export),
1723 ptlrpc_req2svc(req)->srv_req_portal);
1724 RETURN(err_serious(-EPROTO));
1727 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
1728 RETURN(err_serious(-ENOSPC));
1729 if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1))
1730 RETURN(err_serious(-EROFS));
1732 req->rq_bulk_write = 1;
1734 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
1735 RETURN(err_serious(-EIO));
1736 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
1737 RETURN(err_serious(-EFAULT));
1739 /* pause before transaction has been started */
1740 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
1742 /* There must be big cache in current thread to process this request
1743 * if it is NULL then something went wrong and it wasn't allocated,
1744 * report -ENOMEM in that case */
1748 body = tsi->tsi_ost_body;
1749 LASSERT(body != NULL);
1751 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1752 LASSERT(ioo != NULL); /* must exists after tgt_ost_body_unpack */
1754 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
1755 RCL_CLIENT) / sizeof(*ioo);
1757 for (niocount = i = 0; i < objcount; i++)
1758 niocount += ioo[i].ioo_bufcnt;
1760 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1761 LASSERT(remote_nb != NULL); /* must exists after tgt_ost_body_unpack */
1762 if (niocount != req_capsule_get_size(&req->rq_pill,
1763 &RMF_NIOBUF_REMOTE, RCL_CLIENT) /
1765 RETURN(err_serious(-EPROTO));
1767 if ((remote_nb[0].flags & OBD_BRW_MEMALLOC) &&
1768 (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self))
1769 memory_pressure_set();
1771 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
1772 niocount * sizeof(*rcs));
1773 rc = req_capsule_server_pack(&req->rq_pill);
1775 GOTO(out, rc = err_serious(rc));
1777 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
1778 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
1780 local_nb = tbc->local;
1782 rc = tgt_brw_lock(exp->exp_obd->obd_namespace, &tsi->tsi_resid, ioo,
1783 remote_nb, &lockh, LCK_PW);
1788 * If getting the lock took more time than
1789 * client was willing to wait, drop it. b=11330
1791 if (cfs_time_current_sec() > req->rq_deadline ||
1792 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1794 CERROR("%s: Dropping timed-out write from %s because locking "
1795 "object "DOSTID" took %ld seconds (limit was %ld).\n",
1796 tgt_name(tsi->tsi_tgt), libcfs_id2str(req->rq_peer),
1797 POSTID(&ioo->ioo_oid),
1798 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
1799 req->rq_deadline - req->rq_arrival_time.tv_sec);
1800 GOTO(out_lock, rc = -ETIMEDOUT);
1803 /* Because we already sync grant info with client when reconnect,
1804 * grant info will be cleared for resent req, then fed_grant and
1805 * total_grant will not be modified in following preprw_write */
1806 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
1807 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
1808 body->oa.o_valid &= ~OBD_MD_FLGRANT;
1811 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1812 if (repbody == NULL)
1813 GOTO(out_lock, rc = -ENOMEM);
1814 repbody->oa = body->oa;
1816 npages = PTLRPC_MAX_BRW_PAGES;
1817 rc = obd_preprw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
1818 objcount, ioo, remote_nb, &npages, local_nb, NULL,
1823 desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo),
1824 BULK_GET_SINK, OST_BULK_PORTAL);
1826 GOTO(skip_transfer, rc = -ENOMEM);
1828 /* NB Having prepped, we must commit... */
1829 for (i = 0; i < npages; i++)
1830 ptlrpc_prep_bulk_page_nopin(desc, local_nb[i].page,
1831 local_nb[i].lnb_page_offset,
1834 rc = sptlrpc_svc_prep_bulk(req, desc);
1836 GOTO(skip_transfer, rc);
1838 rc = target_bulk_io(exp, desc, &lwi);
1842 if (body->oa.o_valid & OBD_MD_FLCKSUM && rc == 0) {
1843 static int cksum_counter;
1845 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1846 cksum_type = cksum_type_unpack(body->oa.o_flags);
1848 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1849 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
1850 repbody->oa.o_flags |= cksum_type_pack(cksum_type);
1851 repbody->oa.o_cksum = tgt_checksum_bulk(tsi->tsi_tgt, desc,
1852 OST_WRITE, cksum_type);
1855 if (unlikely(body->oa.o_cksum != repbody->oa.o_cksum)) {
1856 mmap = (body->oa.o_valid & OBD_MD_FLFLAGS &&
1857 body->oa.o_flags & OBD_FL_MMAP);
1859 tgt_warn_on_cksum(req, desc, local_nb, npages,
1861 repbody->oa.o_cksum, mmap);
1863 } else if ((cksum_counter & (-cksum_counter)) ==
1865 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
1866 cksum_counter, libcfs_id2str(req->rq_peer),
1867 repbody->oa.o_cksum);
1871 /* Must commit after prep above in all cases */
1872 rc = obd_commitrw(tsi->tsi_env, OBD_BRW_WRITE, exp, &repbody->oa,
1873 objcount, ioo, remote_nb, npages, local_nb, NULL,
1875 if (rc == -ENOTCONN)
1876 /* quota acquire process has been given up because
1877 * either the client has been evicted or the client
1878 * has timed out the request already */
1882 * Disable sending mtime back to the client. If the client locked the
1883 * whole object, then it has already updated the mtime on its side,
1884 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
1886 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
1891 /* set per-requested niobuf return codes */
1892 for (i = j = 0; i < niocount; i++) {
1893 int len = remote_nb[i].len;
1898 LASSERT(j < npages);
1899 if (local_nb[j].rc < 0)
1900 rcs[i] = local_nb[j].rc;
1901 len -= local_nb[j].len;
1906 LASSERT(j == npages);
1907 ptlrpc_lprocfs_brw(req, nob);
1909 tgt_drop_id(exp, &repbody->oa);
1912 tgt_brw_unlock(ioo, remote_nb, &lockh, LCK_PW);
1914 ptlrpc_free_bulk_nopin(desc);
1917 req->rq_no_reply = 1;
1918 /* reply out callback would free */
1919 ptlrpc_req_drop_rs(req);
1920 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s), "
1921 "client will retry: rc %d\n",
1922 exp->exp_obd->obd_name,
1923 obd_uuid2str(&exp->exp_client_uuid),
1924 obd_export_nid2str(exp), rc);
1926 memory_pressure_clr();
1929 EXPORT_SYMBOL(tgt_brw_write);