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, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #define DEBUG_SUBSYSTEM S_MDC
35 #include <linux/init.h>
36 #include <linux/kthread.h>
37 #include <linux/module.h>
38 #include <linux/pagemap.h>
39 #include <linux/user_namespace.h>
40 #include <linux/utsname.h>
41 #ifdef HAVE_UIDGID_HEADER
42 # include <linux/uidgid.h>
45 #include <lustre/lustre_errno.h>
47 #include <cl_object.h>
48 #include <llog_swab.h>
49 #include <lprocfs_status.h>
50 #include <lustre_acl.h>
51 #include <lustre_fid.h>
52 #include <uapi/linux/lustre_ioctl.h>
53 #include <lustre_kernelcomm.h>
54 #include <lustre_lmv.h>
55 #include <lustre_log.h>
56 #include <lustre_param.h>
57 #include <lustre_swab.h>
58 #include <obd_class.h>
60 #include "mdc_internal.h"
62 #define REQUEST_MINOR 244
64 static int mdc_cleanup(struct obd_device *obd);
66 static inline int mdc_queue_wait(struct ptlrpc_request *req)
68 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
71 /* obd_get_request_slot() ensures that this client has no more
72 * than cl_max_rpcs_in_flight RPCs simultaneously inf light
74 rc = obd_get_request_slot(cli);
78 rc = ptlrpc_queue_wait(req);
79 obd_put_request_slot(cli);
85 * Send MDS_GET_ROOT RPC to fetch root FID.
87 * If \a fileset is not NULL it should contain a subdirectory off
88 * the ROOT/ directory to be mounted on the client. Return the FID
89 * of the subdirectory to the client to mount onto its mountpoint.
91 * \param[in] imp MDC import
92 * \param[in] fileset fileset name, which could be NULL
93 * \param[out] rootfid root FID of this mountpoint
94 * \param[out] pc root capa will be unpacked and saved in this pointer
96 * \retval 0 on success, negative errno on failure
98 static int mdc_get_root(struct obd_export *exp, const char *fileset,
99 struct lu_fid *rootfid)
101 struct ptlrpc_request *req;
102 struct mdt_body *body;
107 if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
110 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
116 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
117 strlen(fileset) + 1);
118 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
120 ptlrpc_request_free(req);
123 mdc_pack_body(req, NULL, 0, 0, -1, 0);
124 if (fileset != NULL) {
125 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
127 memcpy(name, fileset, strlen(fileset));
129 lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
130 req->rq_send_state = LUSTRE_IMP_FULL;
132 ptlrpc_request_set_replen(req);
134 rc = ptlrpc_queue_wait(req);
138 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
140 GOTO(out, rc = -EPROTO);
142 *rootfid = body->mbo_fid1;
143 CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
144 PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
147 ptlrpc_req_finished(req);
153 * This function now is known to always saying that it will receive 4 buffers
154 * from server. Even for cases when acl_size and md_size is zero, RPC header
155 * will contain 4 fields and RPC itself will contain zero size fields. This is
156 * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
157 * and thus zero, it shrinks it, making zero size. The same story about
158 * md_size. And this is course of problem when client waits for smaller number
159 * of fields. This issue will be fixed later when client gets aware of RPC
162 static int mdc_getattr_common(struct obd_export *exp,
163 struct ptlrpc_request *req)
165 struct req_capsule *pill = &req->rq_pill;
166 struct mdt_body *body;
171 /* Request message already built. */
172 rc = ptlrpc_queue_wait(req);
176 /* sanity check for the reply */
177 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
181 CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
183 mdc_update_max_ea_from_body(exp, body);
184 if (body->mbo_eadatasize != 0) {
185 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
186 body->mbo_eadatasize);
194 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
195 struct ptlrpc_request **request)
197 struct ptlrpc_request *req;
201 /* Single MDS without an LMV case */
202 if (op_data->op_flags & MF_GET_MDT_IDX) {
207 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_GETATTR);
211 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
213 ptlrpc_request_free(req);
217 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
218 op_data->op_mode, -1, 0);
220 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
221 req->rq_import->imp_connect_data.ocd_max_easize);
222 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
224 ptlrpc_request_set_replen(req);
226 rc = mdc_getattr_common(exp, req);
228 ptlrpc_req_finished(req);
234 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
235 struct ptlrpc_request **request)
237 struct ptlrpc_request *req;
242 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
243 &RQF_MDS_GETATTR_NAME);
247 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
248 op_data->op_namelen + 1);
250 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
252 ptlrpc_request_free(req);
256 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
257 op_data->op_mode, op_data->op_suppgids[0], 0);
259 if (op_data->op_name) {
260 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
261 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
262 op_data->op_namelen);
263 memcpy(name, op_data->op_name, op_data->op_namelen);
266 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
268 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
269 req->rq_import->imp_connect_data.ocd_max_easize);
270 ptlrpc_request_set_replen(req);
272 rc = mdc_getattr_common(exp, req);
274 ptlrpc_req_finished(req);
280 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
281 const struct lu_fid *fid, int opcode, u64 valid,
282 const char *xattr_name, const char *input,
283 int input_size, int output_size, int flags,
284 __u32 suppgid, struct ptlrpc_request **request)
286 struct ptlrpc_request *req;
287 int xattr_namelen = 0;
293 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
298 xattr_namelen = strlen(xattr_name) + 1;
299 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
304 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
308 /* Flush local XATTR locks to get rid of a possible cancel RPC */
309 if (opcode == MDS_REINT && fid_is_sane(fid) &&
310 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
311 struct list_head cancels = LIST_HEAD_INIT(cancels);
314 /* Without that packing would fail */
316 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
319 count = mdc_resource_get_unused(exp, fid,
321 MDS_INODELOCK_XATTR);
323 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
325 ptlrpc_request_free(req);
329 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
331 ptlrpc_request_free(req);
336 if (opcode == MDS_REINT) {
337 struct mdt_rec_setxattr *rec;
339 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
340 sizeof(struct mdt_rec_reint));
341 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
342 rec->sx_opcode = REINT_SETXATTR;
343 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
344 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
345 rec->sx_cap = cfs_curproc_cap_pack();
346 rec->sx_suppgid1 = suppgid;
347 rec->sx_suppgid2 = -1;
349 rec->sx_valid = valid | OBD_MD_FLCTIME;
350 rec->sx_time = ktime_get_real_seconds();
351 rec->sx_size = output_size;
352 rec->sx_flags = flags;
354 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
358 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
359 memcpy(tmp, xattr_name, xattr_namelen);
362 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
363 memcpy(tmp, input, input_size);
366 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
367 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
368 RCL_SERVER, output_size);
369 ptlrpc_request_set_replen(req);
372 if (opcode == MDS_REINT)
373 mdc_get_mod_rpc_slot(req, NULL);
375 rc = ptlrpc_queue_wait(req);
377 if (opcode == MDS_REINT)
378 mdc_put_mod_rpc_slot(req, NULL);
381 ptlrpc_req_finished(req);
387 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
388 u64 valid, const char *xattr_name,
389 const char *input, int input_size, int output_size,
390 int flags, __u32 suppgid,
391 struct ptlrpc_request **request)
393 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
394 fid, MDS_REINT, valid, xattr_name,
395 input, input_size, output_size, flags,
399 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
400 u64 valid, const char *xattr_name,
401 const char *input, int input_size, int output_size,
402 int flags, struct ptlrpc_request **request)
404 return mdc_xattr_common(exp, &RQF_MDS_GETXATTR,
405 fid, MDS_GETXATTR, valid, xattr_name,
406 input, input_size, output_size, flags,
410 #ifdef CONFIG_FS_POSIX_ACL
411 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
413 struct req_capsule *pill = &req->rq_pill;
414 struct mdt_body *body = md->body;
415 struct posix_acl *acl;
420 if (!body->mbo_aclsize)
423 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
428 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
433 CERROR("convert xattr to acl: %d\n", rc);
437 rc = posix_acl_valid(&init_user_ns, acl);
439 CERROR("validate acl: %d\n", rc);
440 posix_acl_release(acl);
448 #define mdc_unpack_acl(req, md) 0
451 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
452 struct obd_export *dt_exp, struct obd_export *md_exp,
453 struct lustre_md *md)
455 struct req_capsule *pill = &req->rq_pill;
460 memset(md, 0, sizeof(*md));
462 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
463 LASSERT(md->body != NULL);
465 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
466 if (!S_ISREG(md->body->mbo_mode)) {
467 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
468 "regular file, but is not\n");
469 GOTO(out, rc = -EPROTO);
472 if (md->body->mbo_eadatasize == 0) {
473 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
474 "but eadatasize 0\n");
475 GOTO(out, rc = -EPROTO);
478 md->layout.lb_len = md->body->mbo_eadatasize;
479 md->layout.lb_buf = req_capsule_server_sized_get(pill,
482 if (md->layout.lb_buf == NULL)
483 GOTO(out, rc = -EPROTO);
484 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
485 const union lmv_mds_md *lmv;
488 if (!S_ISDIR(md->body->mbo_mode)) {
489 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
490 "directory, but is not\n");
491 GOTO(out, rc = -EPROTO);
494 lmv_size = md->body->mbo_eadatasize;
496 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
497 "but eadatasize 0\n");
501 if (md->body->mbo_valid & OBD_MD_MEA) {
502 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
505 GOTO(out, rc = -EPROTO);
507 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
511 if (rc < (typeof(rc))sizeof(*md->lmv)) {
512 CDEBUG(D_INFO, "size too small: "
513 "rc < sizeof(*md->lmv) (%d < %d)\n",
514 rc, (int)sizeof(*md->lmv));
515 GOTO(out, rc = -EPROTO);
521 if (md->body->mbo_valid & OBD_MD_FLACL) {
522 /* for ACL, it's possible that FLACL is set but aclsize is zero.
523 * only when aclsize != 0 there's an actual segment for ACL
526 if (md->body->mbo_aclsize) {
527 rc = mdc_unpack_acl(req, md);
530 #ifdef CONFIG_FS_POSIX_ACL
532 md->posix_acl = NULL;
540 #ifdef CONFIG_FS_POSIX_ACL
541 posix_acl_release(md->posix_acl);
547 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
553 void mdc_replay_open(struct ptlrpc_request *req)
555 struct md_open_data *mod = req->rq_cb_data;
556 struct ptlrpc_request *close_req;
557 struct obd_client_handle *och;
558 struct lustre_handle old;
559 struct mdt_body *body;
563 DEBUG_REQ(D_ERROR, req,
564 "Can't properly replay without open data.");
569 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
570 LASSERT(body != NULL);
574 struct lustre_handle *file_fh;
576 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
578 file_fh = &och->och_fh;
579 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
580 file_fh->cookie, body->mbo_handle.cookie);
582 *file_fh = body->mbo_handle;
584 close_req = mod->mod_close_req;
585 if (close_req != NULL) {
586 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
587 struct mdt_ioepoch *epoch;
589 LASSERT(opc == MDS_CLOSE);
590 epoch = req_capsule_client_get(&close_req->rq_pill,
595 LASSERT(!memcmp(&old, &epoch->mio_handle, sizeof(old)));
597 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
598 epoch->mio_handle = body->mbo_handle;
603 void mdc_commit_open(struct ptlrpc_request *req)
605 struct md_open_data *mod = req->rq_cb_data;
610 * No need to touch md_open_data::mod_och, it holds a reference on
611 * \var mod and will zero references to each other, \var mod will be
612 * freed after that when md_open_data::mod_och will put the reference.
616 * Do not let open request to disappear as it still may be needed
617 * for close rpc to happen (it may happen on evict only, otherwise
618 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
619 * called), just mark this rpc as committed to distinguish these 2
620 * cases, see mdc_close() for details. The open request reference will
621 * be put along with freeing \var mod.
623 ptlrpc_request_addref(req);
624 spin_lock(&req->rq_lock);
625 req->rq_committed = 1;
626 spin_unlock(&req->rq_lock);
627 req->rq_cb_data = NULL;
631 int mdc_set_open_replay_data(struct obd_export *exp,
632 struct obd_client_handle *och,
633 struct lookup_intent *it)
635 struct md_open_data *mod;
636 struct mdt_rec_create *rec;
637 struct mdt_body *body;
638 struct ptlrpc_request *open_req = it->it_request;
639 struct obd_import *imp = open_req->rq_import;
642 if (!open_req->rq_replay)
645 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
646 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
647 LASSERT(rec != NULL);
648 /* Incoming message in my byte order (it's been swabbed). */
649 /* Outgoing messages always in my byte order. */
650 LASSERT(body != NULL);
652 /* Only if the import is replayable, we set replay_open data */
653 if (och && imp->imp_replayable) {
654 mod = obd_mod_alloc();
656 DEBUG_REQ(D_ERROR, open_req,
657 "Can't allocate md_open_data");
662 * Take a reference on \var mod, to be freed on mdc_close().
663 * It protects \var mod from being freed on eviction (commit
664 * callback is called despite rq_replay flag).
665 * Another reference for \var och.
670 spin_lock(&open_req->rq_lock);
673 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
674 it_disposition(it, DISP_OPEN_STRIPE);
675 mod->mod_open_req = open_req;
676 open_req->rq_cb_data = mod;
677 open_req->rq_commit_cb = mdc_commit_open;
678 spin_unlock(&open_req->rq_lock);
681 rec->cr_fid2 = body->mbo_fid1;
682 rec->cr_ioepoch = body->mbo_ioepoch;
683 rec->cr_old_handle.cookie = body->mbo_handle.cookie;
684 open_req->rq_replay_cb = mdc_replay_open;
685 if (!fid_is_sane(&body->mbo_fid1)) {
686 DEBUG_REQ(D_ERROR, open_req, "Saving replay request with "
691 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
695 static void mdc_free_open(struct md_open_data *mod)
699 if (mod->mod_is_create == 0 &&
700 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
704 * No reason to asssert here if the open request has
705 * rq_replay == 1. It means that mdc_close failed, and
706 * close request wasn`t sent. It is not fatal to client.
707 * The worst thing is eviction if the client gets open lock
710 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
711 "= %d\n", mod->mod_open_req->rq_replay);
713 ptlrpc_request_committed(mod->mod_open_req, committed);
714 if (mod->mod_close_req)
715 ptlrpc_request_committed(mod->mod_close_req, committed);
718 int mdc_clear_open_replay_data(struct obd_export *exp,
719 struct obd_client_handle *och)
721 struct md_open_data *mod = och->och_mod;
725 * It is possible to not have \var mod in a case of eviction between
726 * lookup and ll_file_open().
731 LASSERT(mod != LP_POISON);
732 LASSERT(mod->mod_open_req != NULL);
742 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
743 struct md_open_data *mod, struct ptlrpc_request **request)
745 struct obd_device *obd = class_exp2obd(exp);
746 struct ptlrpc_request *req;
747 struct req_format *req_fmt;
752 if (op_data->op_bias & MDS_HSM_RELEASE) {
753 req_fmt = &RQF_MDS_INTENT_CLOSE;
755 /* allocate a FID for volatile file */
756 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
758 CERROR("%s: "DFID" failed to allocate FID: %d\n",
759 obd->obd_name, PFID(&op_data->op_fid1), rc);
760 /* save the errcode and proceed to close */
763 } else if (op_data->op_bias & MDS_CLOSE_LAYOUT_SWAP) {
764 req_fmt = &RQF_MDS_INTENT_CLOSE;
766 req_fmt = &RQF_MDS_CLOSE;
770 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
773 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
775 /* Ensure that this close's handle is fixed up during replay. */
776 if (likely(mod != NULL)) {
777 LASSERTF(mod->mod_open_req != NULL &&
778 mod->mod_open_req->rq_type != LI_POISON,
779 "POISONED open %p!\n", mod->mod_open_req);
781 mod->mod_close_req = req;
783 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
784 /* We no longer want to preserve this open for replay even
785 * though the open was committed. b=3632, b=3633 */
786 spin_lock(&mod->mod_open_req->rq_lock);
787 mod->mod_open_req->rq_replay = 0;
788 spin_unlock(&mod->mod_open_req->rq_lock);
790 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
794 * TODO: repeat close after errors
796 CWARN("%s: close of FID "DFID" failed, file reference will be "
797 "dropped when this client unmounts or is evicted\n",
798 obd->obd_name, PFID(&op_data->op_fid1));
799 GOTO(out, rc = -ENOMEM);
802 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
804 ptlrpc_request_free(req);
809 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
810 * portal whose threads are not taking any DLM locks and are therefore
811 * always progressing */
812 req->rq_request_portal = MDS_READPAGE_PORTAL;
813 ptlrpc_at_set_req_timeout(req);
816 mdc_close_pack(req, op_data);
818 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
819 obd->u.cli.cl_default_mds_easize);
821 ptlrpc_request_set_replen(req);
823 mdc_get_mod_rpc_slot(req, NULL);
824 rc = ptlrpc_queue_wait(req);
825 mdc_put_mod_rpc_slot(req, NULL);
827 if (req->rq_repmsg == NULL) {
828 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
831 rc = req->rq_status ?: -EIO;
832 } else if (rc == 0 || rc == -EAGAIN) {
833 struct mdt_body *body;
835 rc = lustre_msg_get_status(req->rq_repmsg);
836 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
837 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
842 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
845 } else if (rc == -ESTALE) {
847 * it can be allowed error after 3633 if open was committed and
848 * server failed before close was sent. Let's check if mod
849 * exists and return no error in that case
852 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
853 LASSERT(mod->mod_open_req != NULL);
854 if (mod->mod_open_req->rq_committed)
862 mod->mod_close_req = NULL;
863 /* Since now, mod is accessed through open_req only,
864 * thus close req does not keep a reference on mod anymore. */
869 RETURN(rc < 0 ? rc : saved_rc);
872 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
873 u64 offset, struct page **pages, int npages,
874 struct ptlrpc_request **request)
876 struct ptlrpc_request *req;
877 struct ptlrpc_bulk_desc *desc;
879 wait_queue_head_t waitq;
881 struct l_wait_info lwi;
886 init_waitqueue_head(&waitq);
889 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
893 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
895 ptlrpc_request_free(req);
899 req->rq_request_portal = MDS_READPAGE_PORTAL;
900 ptlrpc_at_set_req_timeout(req);
902 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
903 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
905 &ptlrpc_bulk_kiov_pin_ops);
907 ptlrpc_req_finished(req);
911 /* NB req now owns desc and will free it when it gets freed */
912 for (i = 0; i < npages; i++)
913 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
916 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
918 ptlrpc_request_set_replen(req);
919 rc = ptlrpc_queue_wait(req);
921 ptlrpc_req_finished(req);
922 if (rc != -ETIMEDOUT)
926 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
927 CERROR("%s: too many resend retries: rc = %d\n",
928 exp->exp_obd->obd_name, -EIO);
931 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
933 l_wait_event(waitq, 0, &lwi);
938 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
939 req->rq_bulk->bd_nob_transferred);
941 ptlrpc_req_finished(req);
945 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
946 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
947 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
949 ptlrpc_req_finished(req);
957 static void mdc_release_page(struct page *page, int remove)
961 if (likely(page->mapping != NULL))
962 truncate_complete_page(page->mapping, page);
968 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
969 __u64 *start, __u64 *end, int hash64)
972 * Complement of hash is used as an index so that
973 * radix_tree_gang_lookup() can be used to find a page with starting
974 * hash _smaller_ than one we are looking for.
976 unsigned long offset = hash_x_index(*hash, hash64);
980 spin_lock_irq(&mapping->tree_lock);
981 found = radix_tree_gang_lookup(&mapping->page_tree,
982 (void **)&page, offset, 1);
983 if (found > 0 && !radix_tree_exceptional_entry(page)) {
984 struct lu_dirpage *dp;
987 spin_unlock_irq(&mapping->tree_lock);
989 * In contrast to find_lock_page() we are sure that directory
990 * page cannot be truncated (while DLM lock is held) and,
991 * hence, can avoid restart.
993 * In fact, page cannot be locked here at all, because
994 * mdc_read_page_remote does synchronous io.
996 wait_on_page_locked(page);
997 if (PageUptodate(page)) {
999 if (BITS_PER_LONG == 32 && hash64) {
1000 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1001 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1002 *hash = *hash >> 32;
1004 *start = le64_to_cpu(dp->ldp_hash_start);
1005 *end = le64_to_cpu(dp->ldp_hash_end);
1007 if (unlikely(*start == 1 && *hash == 0))
1010 LASSERTF(*start <= *hash, "start = %#llx"
1011 ",end = %#llx,hash = %#llx\n",
1012 *start, *end, *hash);
1013 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1014 " hash %#llx\n", offset, *start, *end, *hash);
1017 mdc_release_page(page, 0);
1019 } else if (*end != *start && *hash == *end) {
1021 * upon hash collision, remove this page,
1022 * otherwise put page reference, and
1023 * mdc_read_page_remote() will issue RPC to
1024 * fetch the page we want.
1027 mdc_release_page(page,
1028 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1033 page = ERR_PTR(-EIO);
1036 spin_unlock_irq(&mapping->tree_lock);
1043 * Adjust a set of pages, each page containing an array of lu_dirpages,
1044 * so that each page can be used as a single logical lu_dirpage.
1046 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1047 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1048 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1049 * value is used as a cookie to request the next lu_dirpage in a
1050 * directory listing that spans multiple pages (two in this example):
1053 * .|--------v------- -----.
1054 * |s|e|f|p|ent|ent| ... |ent|
1055 * '--|-------------- -----' Each PAGE contains a single
1056 * '------. lu_dirpage.
1057 * .---------v------- -----.
1058 * |s|e|f|p|ent| 0 | ... | 0 |
1059 * '----------------- -----'
1061 * However, on hosts where the native VM page size (PAGE_SIZE) is
1062 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1063 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1064 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1065 * after it in the same PAGE (arrows simplified for brevity, but
1066 * in general e0==s1, e1==s2, etc.):
1068 * .-------------------- -----.
1069 * |s0|e0|f0|p|ent|ent| ... |ent|
1070 * |---v---------------- -----|
1071 * |s1|e1|f1|p|ent|ent| ... |ent|
1072 * |---v---------------- -----| Here, each PAGE contains
1073 * ... multiple lu_dirpages.
1074 * |---v---------------- -----|
1075 * |s'|e'|f'|p|ent|ent| ... |ent|
1076 * '---|---------------- -----'
1078 * .----------------------------.
1081 * This structure is transformed into a single logical lu_dirpage as follows:
1083 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1084 * labeled 'next PAGE'.
1086 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1087 * a hash collision with the next page exists.
1089 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1090 * to the first entry of the next lu_dirpage.
1092 #if PAGE_SIZE > LU_PAGE_SIZE
1093 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1097 for (i = 0; i < cfs_pgs; i++) {
1098 struct lu_dirpage *dp = kmap(pages[i]);
1099 struct lu_dirpage *first = dp;
1100 struct lu_dirent *end_dirent = NULL;
1101 struct lu_dirent *ent;
1102 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1103 __u32 flags = le32_to_cpu(dp->ldp_flags);
1105 while (--lu_pgs > 0) {
1106 ent = lu_dirent_start(dp);
1107 for (end_dirent = ent; ent != NULL;
1108 end_dirent = ent, ent = lu_dirent_next(ent));
1110 /* Advance dp to next lu_dirpage. */
1111 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1113 /* Check if we've reached the end of the PAGE. */
1114 if (!((unsigned long)dp & ~PAGE_MASK))
1117 /* Save the hash and flags of this lu_dirpage. */
1118 hash_end = le64_to_cpu(dp->ldp_hash_end);
1119 flags = le32_to_cpu(dp->ldp_flags);
1121 /* Check if lu_dirpage contains no entries. */
1122 if (end_dirent == NULL)
1125 /* Enlarge the end entry lde_reclen from 0 to
1126 * first entry of next lu_dirpage. */
1127 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1128 end_dirent->lde_reclen =
1129 cpu_to_le16((char *)(dp->ldp_entries) -
1130 (char *)end_dirent);
1133 first->ldp_hash_end = hash_end;
1134 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1135 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1139 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1142 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1143 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1145 /* parameters for readdir page */
1146 struct readpage_param {
1147 struct md_op_data *rp_mod;
1150 struct obd_export *rp_exp;
1151 struct md_callback *rp_cb;
1154 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1155 static inline void delete_from_page_cache(struct page *page)
1157 remove_from_page_cache(page);
1163 * Read pages from server.
1165 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1166 * a header lu_dirpage which describes the start/end hash, and whether this
1167 * page is empty (contains no dir entry) or hash collide with next page.
1168 * After client receives reply, several pages will be integrated into dir page
1169 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1170 * lu_dirpage for this integrated page will be adjusted.
1172 static int mdc_read_page_remote(void *data, struct page *page0)
1174 struct readpage_param *rp = data;
1175 struct page **page_pool;
1177 struct lu_dirpage *dp;
1178 struct md_op_data *op_data = rp->rp_mod;
1179 struct ptlrpc_request *req;
1181 struct inode *inode;
1183 int rd_pgs = 0; /* number of pages actually read */
1189 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1190 inode = op_data->op_data;
1191 fid = &op_data->op_fid1;
1192 LASSERT(inode != NULL);
1194 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1195 if (page_pool != NULL) {
1196 page_pool[0] = page0;
1202 for (npages = 1; npages < max_pages; npages++) {
1203 page = page_cache_alloc_cold(inode->i_mapping);
1206 page_pool[npages] = page;
1209 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1211 /* page0 is special, which was added into page cache early */
1212 delete_from_page_cache(page0);
1216 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1218 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1219 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1221 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1223 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1225 SetPageUptodate(page0);
1229 ptlrpc_req_finished(req);
1230 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1231 for (i = 1; i < npages; i++) {
1232 unsigned long offset;
1236 page = page_pool[i];
1238 if (rc < 0 || i >= rd_pgs) {
1243 SetPageUptodate(page);
1246 hash = le64_to_cpu(dp->ldp_hash_start);
1249 offset = hash_x_index(hash, rp->rp_hash64);
1251 prefetchw(&page->flags);
1252 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1257 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1258 " rc = %d\n", offset, ret);
1262 if (page_pool != &page0)
1263 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1269 * Read dir page from cache first, if it can not find it, read it from
1270 * server and add into the cache.
1272 * \param[in] exp MDC export
1273 * \param[in] op_data client MD stack parameters, transfering parameters
1274 * between different layers on client MD stack.
1275 * \param[in] cb_op callback required for ldlm lock enqueue during
1277 * \param[in] hash_offset the hash offset of the page to be read
1278 * \param[in] ppage the page to be read
1280 * retval = 0 get the page successfully
1281 * errno(<0) get the page failed
1283 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1284 struct md_callback *cb_op, __u64 hash_offset,
1285 struct page **ppage)
1287 struct lookup_intent it = { .it_op = IT_READDIR };
1289 struct inode *dir = op_data->op_data;
1290 struct address_space *mapping;
1291 struct lu_dirpage *dp;
1294 struct lustre_handle lockh;
1295 struct ptlrpc_request *enq_req = NULL;
1296 struct readpage_param rp_param;
1303 LASSERT(dir != NULL);
1304 mapping = dir->i_mapping;
1306 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1307 cb_op->md_blocking_ast, 0);
1308 if (enq_req != NULL)
1309 ptlrpc_req_finished(enq_req);
1312 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1313 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1318 lockh.cookie = it.it_lock_handle;
1319 mdc_set_lock_data(exp, &lockh, dir, NULL);
1321 rp_param.rp_off = hash_offset;
1322 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1323 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1324 rp_param.rp_hash64);
1326 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1327 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1328 rp_param.rp_off, PTR_ERR(page));
1329 GOTO(out_unlock, rc = PTR_ERR(page));
1330 } else if (page != NULL) {
1332 * XXX nikita: not entirely correct handling of a corner case:
1333 * suppose hash chain of entries with hash value HASH crosses
1334 * border between pages P0 and P1. First both P0 and P1 are
1335 * cached, seekdir() is called for some entry from the P0 part
1336 * of the chain. Later P0 goes out of cache. telldir(HASH)
1337 * happens and finds P1, as it starts with matching hash
1338 * value. Remaining entries from P0 part of the chain are
1339 * skipped. (Is that really a bug?)
1341 * Possible solutions: 0. don't cache P1 is such case, handle
1342 * it as an "overflow" page. 1. invalidate all pages at
1343 * once. 2. use HASH|1 as an index for P1.
1345 GOTO(hash_collision, page);
1348 rp_param.rp_exp = exp;
1349 rp_param.rp_mod = op_data;
1350 page = read_cache_page(mapping,
1351 hash_x_index(rp_param.rp_off,
1352 rp_param.rp_hash64),
1353 mdc_read_page_remote, &rp_param);
1355 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1356 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1357 rp_param.rp_off, PTR_ERR(page));
1358 GOTO(out_unlock, rc = PTR_ERR(page));
1361 wait_on_page_locked(page);
1363 if (!PageUptodate(page)) {
1364 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1365 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1366 rp_param.rp_off, -5);
1369 if (!PageChecked(page))
1370 SetPageChecked(page);
1371 if (PageError(page)) {
1372 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1373 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1374 rp_param.rp_off, -5);
1379 dp = page_address(page);
1380 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1381 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1382 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1383 rp_param.rp_off = hash_offset >> 32;
1385 start = le64_to_cpu(dp->ldp_hash_start);
1386 end = le64_to_cpu(dp->ldp_hash_end);
1387 rp_param.rp_off = hash_offset;
1390 LASSERT(start == rp_param.rp_off);
1391 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1392 #if BITS_PER_LONG == 32
1393 CWARN("Real page-wide hash collision at [%llu %llu] with "
1394 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1395 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1399 * Fetch whole overflow chain...
1407 ldlm_lock_decref(&lockh, it.it_lock_mode);
1411 mdc_release_page(page, 1);
1417 static int mdc_statfs(const struct lu_env *env,
1418 struct obd_export *exp, struct obd_statfs *osfs,
1419 __u64 max_age, __u32 flags)
1421 struct obd_device *obd = class_exp2obd(exp);
1422 struct ptlrpc_request *req;
1423 struct obd_statfs *msfs;
1424 struct obd_import *imp = NULL;
1429 * Since the request might also come from lprocfs, so we need
1430 * sync this with client_disconnect_export Bug15684
1432 down_read(&obd->u.cli.cl_sem);
1433 if (obd->u.cli.cl_import)
1434 imp = class_import_get(obd->u.cli.cl_import);
1435 up_read(&obd->u.cli.cl_sem);
1439 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_STATFS,
1440 LUSTRE_MDS_VERSION, MDS_STATFS);
1442 GOTO(output, rc = -ENOMEM);
1444 ptlrpc_request_set_replen(req);
1446 if (flags & OBD_STATFS_NODELAY) {
1447 /* procfs requests not want stay in wait for avoid deadlock */
1448 req->rq_no_resend = 1;
1449 req->rq_no_delay = 1;
1452 rc = ptlrpc_queue_wait(req);
1454 /* check connection error first */
1455 if (imp->imp_connect_error)
1456 rc = imp->imp_connect_error;
1460 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1462 GOTO(out, rc = -EPROTO);
1467 ptlrpc_req_finished(req);
1469 class_import_put(imp);
1473 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1475 __u32 keylen, vallen;
1479 if (gf->gf_pathlen > PATH_MAX)
1480 RETURN(-ENAMETOOLONG);
1481 if (gf->gf_pathlen < 2)
1484 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1485 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1486 sizeof(struct lu_fid));
1487 OBD_ALLOC(key, keylen);
1490 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1491 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1492 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1493 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1494 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1495 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1497 if (!fid_is_sane(&gf->gf_fid))
1498 GOTO(out, rc = -EINVAL);
1500 /* Val is struct getinfo_fid2path result plus path */
1501 vallen = sizeof(*gf) + gf->gf_pathlen;
1503 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1504 if (rc != 0 && rc != -EREMOTE)
1507 if (vallen <= sizeof(*gf))
1508 GOTO(out, rc = -EPROTO);
1509 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1510 GOTO(out, rc = -EOVERFLOW);
1512 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1513 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1514 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1515 /* only log the last 512 characters of the path */
1516 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1519 OBD_FREE(key, keylen);
1523 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1524 struct hsm_progress_kernel *hpk)
1526 struct obd_import *imp = class_exp2cliimp(exp);
1527 struct hsm_progress_kernel *req_hpk;
1528 struct ptlrpc_request *req;
1532 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1533 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1535 GOTO(out, rc = -ENOMEM);
1537 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1539 /* Copy hsm_progress struct */
1540 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1541 if (req_hpk == NULL)
1542 GOTO(out, rc = -EPROTO);
1545 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1547 ptlrpc_request_set_replen(req);
1549 mdc_get_mod_rpc_slot(req, NULL);
1550 rc = ptlrpc_queue_wait(req);
1551 mdc_put_mod_rpc_slot(req, NULL);
1555 ptlrpc_req_finished(req);
1559 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archives)
1561 __u32 *archive_mask;
1562 struct ptlrpc_request *req;
1566 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_REGISTER,
1568 MDS_HSM_CT_REGISTER);
1570 GOTO(out, rc = -ENOMEM);
1572 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1574 /* Copy hsm_progress struct */
1575 archive_mask = req_capsule_client_get(&req->rq_pill,
1576 &RMF_MDS_HSM_ARCHIVE);
1577 if (archive_mask == NULL)
1578 GOTO(out, rc = -EPROTO);
1580 *archive_mask = archives;
1582 ptlrpc_request_set_replen(req);
1584 rc = mdc_queue_wait(req);
1587 ptlrpc_req_finished(req);
1591 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1592 struct md_op_data *op_data)
1594 struct hsm_current_action *hca = op_data->op_data;
1595 struct hsm_current_action *req_hca;
1596 struct ptlrpc_request *req;
1600 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1601 &RQF_MDS_HSM_ACTION);
1605 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1607 ptlrpc_request_free(req);
1611 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1612 op_data->op_suppgids[0], 0);
1614 ptlrpc_request_set_replen(req);
1616 rc = mdc_queue_wait(req);
1620 req_hca = req_capsule_server_get(&req->rq_pill,
1621 &RMF_MDS_HSM_CURRENT_ACTION);
1622 if (req_hca == NULL)
1623 GOTO(out, rc = -EPROTO);
1629 ptlrpc_req_finished(req);
1633 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1635 struct ptlrpc_request *req;
1639 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1641 MDS_HSM_CT_UNREGISTER);
1643 GOTO(out, rc = -ENOMEM);
1645 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1647 ptlrpc_request_set_replen(req);
1649 rc = mdc_queue_wait(req);
1652 ptlrpc_req_finished(req);
1656 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1657 struct md_op_data *op_data)
1659 struct hsm_user_state *hus = op_data->op_data;
1660 struct hsm_user_state *req_hus;
1661 struct ptlrpc_request *req;
1665 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1666 &RQF_MDS_HSM_STATE_GET);
1670 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1672 ptlrpc_request_free(req);
1676 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1677 op_data->op_suppgids[0], 0);
1679 ptlrpc_request_set_replen(req);
1681 rc = mdc_queue_wait(req);
1685 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1686 if (req_hus == NULL)
1687 GOTO(out, rc = -EPROTO);
1693 ptlrpc_req_finished(req);
1697 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1698 struct md_op_data *op_data)
1700 struct hsm_state_set *hss = op_data->op_data;
1701 struct hsm_state_set *req_hss;
1702 struct ptlrpc_request *req;
1706 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1707 &RQF_MDS_HSM_STATE_SET);
1711 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1713 ptlrpc_request_free(req);
1717 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1718 op_data->op_suppgids[0], 0);
1721 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1722 if (req_hss == NULL)
1723 GOTO(out, rc = -EPROTO);
1726 ptlrpc_request_set_replen(req);
1728 mdc_get_mod_rpc_slot(req, NULL);
1729 rc = ptlrpc_queue_wait(req);
1730 mdc_put_mod_rpc_slot(req, NULL);
1734 ptlrpc_req_finished(req);
1738 static int mdc_ioc_hsm_request(struct obd_export *exp,
1739 struct hsm_user_request *hur)
1741 struct obd_import *imp = class_exp2cliimp(exp);
1742 struct ptlrpc_request *req;
1743 struct hsm_request *req_hr;
1744 struct hsm_user_item *req_hui;
1749 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1751 GOTO(out, rc = -ENOMEM);
1753 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1754 hur->hur_request.hr_itemcount
1755 * sizeof(struct hsm_user_item));
1756 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1757 hur->hur_request.hr_data_len);
1759 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1761 ptlrpc_request_free(req);
1765 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1767 /* Copy hsm_request struct */
1768 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1770 GOTO(out, rc = -EPROTO);
1771 *req_hr = hur->hur_request;
1773 /* Copy hsm_user_item structs */
1774 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1775 if (req_hui == NULL)
1776 GOTO(out, rc = -EPROTO);
1777 memcpy(req_hui, hur->hur_user_item,
1778 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1780 /* Copy opaque field */
1781 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1782 if (req_opaque == NULL)
1783 GOTO(out, rc = -EPROTO);
1784 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1786 ptlrpc_request_set_replen(req);
1788 mdc_get_mod_rpc_slot(req, NULL);
1789 rc = ptlrpc_queue_wait(req);
1790 mdc_put_mod_rpc_slot(req, NULL);
1795 ptlrpc_req_finished(req);
1799 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1800 struct lustre_kernelcomm *lk);
1802 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1803 struct obd_quotactl *oqctl)
1805 struct ptlrpc_request *req;
1806 struct obd_quotactl *oqc;
1810 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1811 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1816 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1819 ptlrpc_request_set_replen(req);
1820 ptlrpc_at_set_req_timeout(req);
1821 req->rq_no_resend = 1;
1823 rc = ptlrpc_queue_wait(req);
1825 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1827 if (req->rq_repmsg &&
1828 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1831 CERROR ("Can't unpack obd_quotactl\n");
1834 ptlrpc_req_finished(req);
1839 static int mdc_ioc_swap_layouts(struct obd_export *exp,
1840 struct md_op_data *op_data)
1842 struct list_head cancels = LIST_HEAD_INIT(cancels);
1843 struct ptlrpc_request *req;
1845 struct mdc_swap_layouts *msl, *payload;
1848 msl = op_data->op_data;
1850 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
1851 * first thing it will do is to cancel the 2 layout
1852 * locks held by this client.
1853 * So the client must cancel its layout locks on the 2 fids
1854 * with the request RPC to avoid extra RPC round trips.
1856 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
1857 LCK_EX, MDS_INODELOCK_LAYOUT |
1858 MDS_INODELOCK_XATTR);
1859 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
1860 LCK_EX, MDS_INODELOCK_LAYOUT |
1861 MDS_INODELOCK_XATTR);
1863 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1864 &RQF_MDS_SWAP_LAYOUTS);
1866 ldlm_lock_list_put(&cancels, l_bl_ast, count);
1870 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
1872 ptlrpc_request_free(req);
1876 mdc_swap_layouts_pack(req, op_data);
1878 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
1883 ptlrpc_request_set_replen(req);
1885 rc = ptlrpc_queue_wait(req);
1891 ptlrpc_req_finished(req);
1895 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1896 void *karg, void __user *uarg)
1898 struct obd_device *obd = exp->exp_obd;
1899 struct obd_ioctl_data *data = karg;
1900 struct obd_import *imp = obd->u.cli.cl_import;
1904 if (!try_module_get(THIS_MODULE)) {
1905 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1906 module_name(THIS_MODULE));
1910 case OBD_IOC_FID2PATH:
1911 rc = mdc_ioc_fid2path(exp, karg);
1913 case LL_IOC_HSM_CT_START:
1914 rc = mdc_ioc_hsm_ct_start(exp, karg);
1915 /* ignore if it was already registered on this MDS. */
1919 case LL_IOC_HSM_PROGRESS:
1920 rc = mdc_ioc_hsm_progress(exp, karg);
1922 case LL_IOC_HSM_STATE_GET:
1923 rc = mdc_ioc_hsm_state_get(exp, karg);
1925 case LL_IOC_HSM_STATE_SET:
1926 rc = mdc_ioc_hsm_state_set(exp, karg);
1928 case LL_IOC_HSM_ACTION:
1929 rc = mdc_ioc_hsm_current_action(exp, karg);
1931 case LL_IOC_HSM_REQUEST:
1932 rc = mdc_ioc_hsm_request(exp, karg);
1934 case OBD_IOC_CLIENT_RECOVER:
1935 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
1939 case IOC_OSC_SET_ACTIVE:
1940 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
1942 case OBD_IOC_PING_TARGET:
1943 rc = ptlrpc_obd_ping(obd);
1946 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
1947 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
1948 * there'd be no LMV layer thus we might be called here. Eventually
1949 * this code should be removed.
1952 case IOC_OBD_STATFS: {
1953 struct obd_statfs stat_buf = {0};
1955 if (*((__u32 *) data->ioc_inlbuf2) != 0)
1956 GOTO(out, rc = -ENODEV);
1959 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
1960 min((int)data->ioc_plen2,
1961 (int)sizeof(struct obd_uuid))))
1962 GOTO(out, rc = -EFAULT);
1964 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
1965 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
1970 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
1971 min((int) data->ioc_plen1,
1972 (int) sizeof(stat_buf))))
1973 GOTO(out, rc = -EFAULT);
1977 case OBD_IOC_QUOTACTL: {
1978 struct if_quotactl *qctl = karg;
1979 struct obd_quotactl *oqctl;
1981 OBD_ALLOC_PTR(oqctl);
1983 GOTO(out, rc = -ENOMEM);
1985 QCTL_COPY(oqctl, qctl);
1986 rc = obd_quotactl(exp, oqctl);
1988 QCTL_COPY(qctl, oqctl);
1989 qctl->qc_valid = QC_MDTIDX;
1990 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
1993 OBD_FREE_PTR(oqctl);
1996 case LL_IOC_GET_CONNECT_FLAGS:
1997 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
1998 sizeof(*exp_connect_flags_ptr(exp))))
1999 GOTO(out, rc = -EFAULT);
2002 case LL_IOC_LOV_SWAP_LAYOUTS:
2003 rc = mdc_ioc_swap_layouts(exp, karg);
2006 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2007 GOTO(out, rc = -ENOTTY);
2010 module_put(THIS_MODULE);
2015 static int mdc_get_info_rpc(struct obd_export *exp,
2016 u32 keylen, void *key,
2017 u32 vallen, void *val)
2019 struct obd_import *imp = class_exp2cliimp(exp);
2020 struct ptlrpc_request *req;
2025 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2029 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2030 RCL_CLIENT, keylen);
2031 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2032 RCL_CLIENT, sizeof(vallen));
2034 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2036 ptlrpc_request_free(req);
2040 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2041 memcpy(tmp, key, keylen);
2042 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2043 memcpy(tmp, &vallen, sizeof(vallen));
2045 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2046 RCL_SERVER, vallen);
2047 ptlrpc_request_set_replen(req);
2049 rc = ptlrpc_queue_wait(req);
2050 /* -EREMOTE means the get_info result is partial, and it needs to
2051 * continue on another MDT, see fid2path part in lmv_iocontrol */
2052 if (rc == 0 || rc == -EREMOTE) {
2053 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2054 memcpy(val, tmp, vallen);
2055 if (ptlrpc_rep_need_swab(req)) {
2056 if (KEY_IS(KEY_FID2PATH))
2057 lustre_swab_fid2path(val);
2060 ptlrpc_req_finished(req);
2065 static void lustre_swab_hai(struct hsm_action_item *h)
2067 __swab32s(&h->hai_len);
2068 __swab32s(&h->hai_action);
2069 lustre_swab_lu_fid(&h->hai_fid);
2070 lustre_swab_lu_fid(&h->hai_dfid);
2071 __swab64s(&h->hai_cookie);
2072 __swab64s(&h->hai_extent.offset);
2073 __swab64s(&h->hai_extent.length);
2074 __swab64s(&h->hai_gid);
2077 static void lustre_swab_hal(struct hsm_action_list *h)
2079 struct hsm_action_item *hai;
2082 __swab32s(&h->hal_version);
2083 __swab32s(&h->hal_count);
2084 __swab32s(&h->hal_archive_id);
2085 __swab64s(&h->hal_flags);
2087 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2088 lustre_swab_hai(hai);
2091 static void lustre_swab_kuch(struct kuc_hdr *l)
2093 __swab16s(&l->kuc_magic);
2094 /* __u8 l->kuc_transport */
2095 __swab16s(&l->kuc_msgtype);
2096 __swab16s(&l->kuc_msglen);
2099 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2100 struct lustre_kernelcomm *lk)
2102 struct obd_import *imp = class_exp2cliimp(exp);
2103 __u32 archive = lk->lk_data;
2106 if (lk->lk_group != KUC_GRP_HSM) {
2107 CERROR("Bad copytool group %d\n", lk->lk_group);
2111 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2112 lk->lk_uid, lk->lk_group, lk->lk_flags);
2114 if (lk->lk_flags & LK_FLG_STOP) {
2115 /* Unregister with the coordinator */
2116 rc = mdc_ioc_hsm_ct_unregister(imp);
2118 rc = mdc_ioc_hsm_ct_register(imp, archive);
2125 * Send a message to any listening copytools
2126 * @param val KUC message (kuc_hdr + hsm_action_list)
2127 * @param len total length of message
2129 static int mdc_hsm_copytool_send(size_t len, void *val)
2131 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2132 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2136 if (len < sizeof(*lh) + sizeof(*hal)) {
2137 CERROR("Short HSM message %zu < %zu\n", len,
2138 sizeof(*lh) + sizeof(*hal));
2141 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2142 lustre_swab_kuch(lh);
2143 lustre_swab_hal(hal);
2144 } else if (lh->kuc_magic != KUC_MAGIC) {
2145 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2149 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2151 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2152 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2154 /* Broadcast to HSM listeners */
2155 rc = libcfs_kkuc_group_put(KUC_GRP_HSM, lh);
2161 * callback function passed to kuc for re-registering each HSM copytool
2162 * running on MDC, after MDT shutdown/recovery.
2163 * @param data copytool registration data
2164 * @param cb_arg callback argument (obd_import)
2166 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2168 struct kkuc_ct_data *kcd = data;
2169 struct obd_import *imp = (struct obd_import *)cb_arg;
2172 if (kcd == NULL || kcd->kcd_magic != KKUC_CT_DATA_MAGIC)
2175 if (!obd_uuid_equals(&kcd->kcd_uuid, &imp->imp_obd->obd_uuid))
2178 CDEBUG(D_HA, "%s: recover copytool registration to MDT (archive=%#x)\n",
2179 imp->imp_obd->obd_name, kcd->kcd_archive);
2180 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_archive);
2182 /* ignore error if the copytool is already registered */
2183 return (rc == -EEXIST) ? 0 : rc;
2187 * Re-establish all kuc contexts with MDT
2188 * after MDT shutdown/recovery.
2190 static int mdc_kuc_reregister(struct obd_import *imp)
2192 /* re-register HSM agents */
2193 return libcfs_kkuc_group_foreach(KUC_GRP_HSM, mdc_hsm_ct_reregister,
2197 static int mdc_set_info_async(const struct lu_env *env,
2198 struct obd_export *exp,
2199 u32 keylen, void *key,
2200 u32 vallen, void *val,
2201 struct ptlrpc_request_set *set)
2203 struct obd_import *imp = class_exp2cliimp(exp);
2207 if (KEY_IS(KEY_READ_ONLY)) {
2208 if (vallen != sizeof(int))
2211 spin_lock(&imp->imp_lock);
2212 if (*((int *)val)) {
2213 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2214 imp->imp_connect_data.ocd_connect_flags |=
2217 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2218 imp->imp_connect_data.ocd_connect_flags &=
2219 ~OBD_CONNECT_RDONLY;
2221 spin_unlock(&imp->imp_lock);
2223 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2224 keylen, key, vallen, val, set);
2227 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2228 sptlrpc_conf_client_adapt(exp->exp_obd);
2231 if (KEY_IS(KEY_FLUSH_CTX)) {
2232 sptlrpc_import_flush_my_ctx(imp);
2235 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2236 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2237 keylen, key, vallen, val, set);
2240 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2241 rc = mdc_hsm_copytool_send(vallen, val);
2245 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2246 __u32 *default_easize = val;
2248 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2252 CERROR("Unknown key %s\n", (char *)key);
2256 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2257 __u32 keylen, void *key, __u32 *vallen, void *val)
2261 if (KEY_IS(KEY_MAX_EASIZE)) {
2262 __u32 mdsize, *max_easize;
2264 if (*vallen != sizeof(int))
2266 mdsize = *(__u32 *)val;
2267 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2268 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2270 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2272 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2273 __u32 *default_easize;
2275 if (*vallen != sizeof(int))
2277 default_easize = val;
2278 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2280 } else if (KEY_IS(KEY_CONN_DATA)) {
2281 struct obd_import *imp = class_exp2cliimp(exp);
2282 struct obd_connect_data *data = val;
2284 if (*vallen != sizeof(*data))
2287 *data = imp->imp_connect_data;
2289 } else if (KEY_IS(KEY_TGT_COUNT)) {
2290 *((__u32 *)val) = 1;
2294 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2299 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2300 struct ptlrpc_request **request)
2302 struct ptlrpc_request *req;
2307 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2311 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2313 ptlrpc_request_free(req);
2317 mdc_pack_body(req, fid, 0, 0, -1, 0);
2319 ptlrpc_request_set_replen(req);
2321 rc = ptlrpc_queue_wait(req);
2323 ptlrpc_req_finished(req);
2329 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2330 enum obd_import_event event)
2334 LASSERT(imp->imp_obd == obd);
2338 case IMP_EVENT_INACTIVE: {
2339 struct client_obd *cli = &obd->u.cli;
2341 * Flush current sequence to make client obtain new one
2342 * from server in case of disconnect/reconnect.
2344 down_read(&cli->cl_seq_rwsem);
2346 seq_client_flush(cli->cl_seq);
2347 up_read(&cli->cl_seq_rwsem);
2349 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2352 case IMP_EVENT_INVALIDATE: {
2353 struct ldlm_namespace *ns = obd->obd_namespace;
2355 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2359 case IMP_EVENT_ACTIVE:
2360 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2361 /* redo the kuc registration after reconnecting */
2363 rc = mdc_kuc_reregister(imp);
2366 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2368 case IMP_EVENT_DISCON:
2369 case IMP_EVENT_DEACTIVATE:
2370 case IMP_EVENT_ACTIVATE:
2373 CERROR("Unknown import event %x\n", event);
2379 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2380 struct lu_fid *fid, struct md_op_data *op_data)
2382 struct client_obd *cli = &exp->exp_obd->u.cli;
2387 down_read(&cli->cl_seq_rwsem);
2389 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2390 up_read(&cli->cl_seq_rwsem);
2395 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2397 struct client_obd *cli = &exp->exp_obd->u.cli;
2398 return &cli->cl_target_uuid;
2402 * Determine whether the lock can be canceled before replaying it during
2403 * recovery, non zero value will be return if the lock can be canceled,
2404 * or zero returned for not
2406 static int mdc_cancel_weight(struct ldlm_lock *lock)
2408 if (lock->l_resource->lr_type != LDLM_IBITS)
2411 /* FIXME: if we ever get into a situation where there are too many
2412 * opened files with open locks on a single node, then we really
2413 * should replay these open locks to reget it */
2414 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2420 static int mdc_resource_inode_free(struct ldlm_resource *res)
2422 if (res->lr_lvb_inode)
2423 res->lr_lvb_inode = NULL;
2428 static struct ldlm_valblock_ops inode_lvbo = {
2429 .lvbo_free = mdc_resource_inode_free
2432 static int mdc_llog_init(struct obd_device *obd)
2434 struct obd_llog_group *olg = &obd->obd_olg;
2435 struct llog_ctxt *ctxt;
2440 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2445 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2446 llog_initiator_connect(ctxt);
2447 llog_ctxt_put(ctxt);
2452 static void mdc_llog_finish(struct obd_device *obd)
2454 struct llog_ctxt *ctxt;
2458 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2460 llog_cleanup(NULL, ctxt);
2465 static int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2470 rc = ptlrpcd_addref();
2474 rc = client_obd_setup(obd, cfg);
2476 GOTO(err_ptlrpcd_decref, rc);
2477 #ifdef CONFIG_PROC_FS
2478 obd->obd_vars = lprocfs_mdc_obd_vars;
2479 lprocfs_obd_setup(obd);
2480 lprocfs_alloc_md_stats(obd, 0);
2482 sptlrpc_lprocfs_cliobd_attach(obd);
2483 ptlrpc_lprocfs_register_obd(obd);
2485 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2487 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2489 rc = mdc_llog_init(obd);
2491 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2493 GOTO(err_mdc_cleanup, rc);
2496 rc = mdc_changelog_cdev_init(obd);
2498 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2500 GOTO(err_mdc_cleanup, rc);
2506 client_obd_cleanup(obd);
2515 /* Initialize the default and maximum LOV EA sizes. This allows
2516 * us to make MDS RPCs with large enough reply buffers to hold a default
2517 * sized EA without having to calculate this (via a call into the
2518 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2519 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2520 * a large number of stripes is possible. If a larger reply buffer is
2521 * required it will be reallocated in the ptlrpc layer due to overflow.
2523 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2526 struct obd_device *obd = exp->exp_obd;
2527 struct client_obd *cli = &obd->u.cli;
2530 if (cli->cl_max_mds_easize < easize)
2531 cli->cl_max_mds_easize = easize;
2533 if (cli->cl_default_mds_easize < def_easize)
2534 cli->cl_default_mds_easize = def_easize;
2539 static int mdc_precleanup(struct obd_device *obd)
2543 /* Failsafe, ok if racy */
2544 if (obd->obd_type->typ_refcnt <= 1)
2545 libcfs_kkuc_group_rem(0, KUC_GRP_HSM);
2547 mdc_changelog_cdev_finish(obd);
2549 obd_cleanup_client_import(obd);
2550 ptlrpc_lprocfs_unregister_obd(obd);
2551 lprocfs_obd_cleanup(obd);
2552 lprocfs_free_md_stats(obd);
2553 mdc_llog_finish(obd);
2557 static int mdc_cleanup(struct obd_device *obd)
2561 return client_obd_cleanup(obd);
2564 static int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2566 struct lustre_cfg *lcfg = buf;
2567 int rc = class_process_proc_param(PARAM_MDC, obd->obd_vars, lcfg, obd);
2568 return (rc > 0 ? 0: rc);
2571 static struct obd_ops mdc_obd_ops = {
2572 .o_owner = THIS_MODULE,
2573 .o_setup = mdc_setup,
2574 .o_precleanup = mdc_precleanup,
2575 .o_cleanup = mdc_cleanup,
2576 .o_add_conn = client_import_add_conn,
2577 .o_del_conn = client_import_del_conn,
2578 .o_connect = client_connect_import,
2579 .o_disconnect = client_disconnect_export,
2580 .o_iocontrol = mdc_iocontrol,
2581 .o_set_info_async = mdc_set_info_async,
2582 .o_statfs = mdc_statfs,
2583 .o_fid_init = client_fid_init,
2584 .o_fid_fini = client_fid_fini,
2585 .o_fid_alloc = mdc_fid_alloc,
2586 .o_import_event = mdc_import_event,
2587 .o_get_info = mdc_get_info,
2588 .o_process_config = mdc_process_config,
2589 .o_get_uuid = mdc_get_uuid,
2590 .o_quotactl = mdc_quotactl,
2593 static struct md_ops mdc_md_ops = {
2594 .m_get_root = mdc_get_root,
2595 .m_null_inode = mdc_null_inode,
2596 .m_close = mdc_close,
2597 .m_create = mdc_create,
2598 .m_enqueue = mdc_enqueue,
2599 .m_getattr = mdc_getattr,
2600 .m_getattr_name = mdc_getattr_name,
2601 .m_intent_lock = mdc_intent_lock,
2603 .m_rename = mdc_rename,
2604 .m_setattr = mdc_setattr,
2605 .m_setxattr = mdc_setxattr,
2606 .m_getxattr = mdc_getxattr,
2607 .m_fsync = mdc_fsync,
2608 .m_read_page = mdc_read_page,
2609 .m_unlink = mdc_unlink,
2610 .m_cancel_unused = mdc_cancel_unused,
2611 .m_init_ea_size = mdc_init_ea_size,
2612 .m_set_lock_data = mdc_set_lock_data,
2613 .m_lock_match = mdc_lock_match,
2614 .m_get_lustre_md = mdc_get_lustre_md,
2615 .m_free_lustre_md = mdc_free_lustre_md,
2616 .m_set_open_replay_data = mdc_set_open_replay_data,
2617 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2618 .m_intent_getattr_async = mdc_intent_getattr_async,
2619 .m_revalidate_lock = mdc_revalidate_lock
2622 static int __init mdc_init(void)
2624 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2625 LUSTRE_MDC_NAME, NULL);
2628 static void __exit mdc_exit(void)
2630 class_unregister_type(LUSTRE_MDC_NAME);
2633 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2634 MODULE_DESCRIPTION("Lustre Metadata Client");
2635 MODULE_VERSION(LUSTRE_VERSION_STRING);
2636 MODULE_LICENSE("GPL");
2638 module_init(mdc_init);
2639 module_exit(mdc_exit);