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, 2017, 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_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/lustre_ioctl.h>
53 #include <lustre_kernelcomm.h>
54 #include <lustre_lmv.h>
55 #include <lustre_log.h>
56 #include <lustre_swab.h>
57 #include <obd_class.h>
58 #include <lustre_osc.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 void mdc_reset_acl_req(struct ptlrpc_request *req)
196 spin_lock(&req->rq_early_free_lock);
197 sptlrpc_cli_free_repbuf(req);
198 req->rq_repbuf = NULL;
199 req->rq_repbuf_len = 0;
200 req->rq_repdata = NULL;
201 req->rq_reqdata_len = 0;
202 spin_unlock(&req->rq_early_free_lock);
205 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
206 struct ptlrpc_request **request)
208 struct ptlrpc_request *req;
209 struct obd_import *imp = class_exp2cliimp(exp);
210 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
214 /* Single MDS without an LMV case */
215 if (op_data->op_flags & MF_GET_MDT_IDX) {
221 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
225 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
227 ptlrpc_request_free(req);
232 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
233 op_data->op_mode, -1, 0);
234 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
235 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
237 ptlrpc_request_set_replen(req);
239 rc = mdc_getattr_common(exp, req);
242 acl_bufsize = MIN(imp->imp_connect_data.ocd_max_easize,
244 mdc_reset_acl_req(req);
248 ptlrpc_req_finished(req);
256 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
257 struct ptlrpc_request **request)
259 struct ptlrpc_request *req;
260 struct obd_import *imp = class_exp2cliimp(exp);
261 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
266 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
270 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
271 op_data->op_namelen + 1);
273 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
275 ptlrpc_request_free(req);
279 if (op_data->op_name) {
280 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
281 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
282 op_data->op_namelen);
283 memcpy(name, op_data->op_name, op_data->op_namelen);
287 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
288 op_data->op_mode, op_data->op_suppgids[0], 0);
289 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
291 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
292 ptlrpc_request_set_replen(req);
294 rc = mdc_getattr_common(exp, req);
297 acl_bufsize = MIN(imp->imp_connect_data.ocd_max_easize,
299 mdc_reset_acl_req(req);
303 ptlrpc_req_finished(req);
311 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
312 const struct lu_fid *fid, int opcode, u64 valid,
313 const char *xattr_name, const char *input,
314 int input_size, int output_size, int flags,
315 __u32 suppgid, struct ptlrpc_request **request)
317 struct ptlrpc_request *req;
318 int xattr_namelen = 0;
324 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
329 xattr_namelen = strlen(xattr_name) + 1;
330 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
335 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
338 /* get SELinux policy info if any */
339 rc = sptlrpc_get_sepol(req);
341 ptlrpc_request_free(req);
344 req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
345 strlen(req->rq_sepol) ?
346 strlen(req->rq_sepol) + 1 : 0);
348 /* Flush local XATTR locks to get rid of a possible cancel RPC */
349 if (opcode == MDS_REINT && fid_is_sane(fid) &&
350 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
351 struct list_head cancels = LIST_HEAD_INIT(cancels);
354 /* Without that packing would fail */
356 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
359 count = mdc_resource_get_unused(exp, fid,
361 MDS_INODELOCK_XATTR);
363 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
365 ptlrpc_request_free(req);
369 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
371 ptlrpc_request_free(req);
376 if (opcode == MDS_REINT) {
377 struct mdt_rec_setxattr *rec;
379 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
380 sizeof(struct mdt_rec_reint));
381 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
382 rec->sx_opcode = REINT_SETXATTR;
383 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
384 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
385 rec->sx_cap = cfs_curproc_cap_pack();
386 rec->sx_suppgid1 = suppgid;
387 rec->sx_suppgid2 = -1;
389 rec->sx_valid = valid | OBD_MD_FLCTIME;
390 rec->sx_time = ktime_get_real_seconds();
391 rec->sx_size = output_size;
392 rec->sx_flags = flags;
394 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
398 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
399 memcpy(tmp, xattr_name, xattr_namelen);
402 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
403 memcpy(tmp, input, input_size);
406 mdc_file_sepol_pack(req);
408 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
409 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
410 RCL_SERVER, output_size);
411 ptlrpc_request_set_replen(req);
414 if (opcode == MDS_REINT)
415 mdc_get_mod_rpc_slot(req, NULL);
417 rc = ptlrpc_queue_wait(req);
419 if (opcode == MDS_REINT)
420 mdc_put_mod_rpc_slot(req, NULL);
423 ptlrpc_req_finished(req);
429 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
430 u64 obd_md_valid, const char *name,
431 const void *value, size_t value_size,
432 unsigned int xattr_flags, u32 suppgid,
433 struct ptlrpc_request **req)
435 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
436 obd_md_valid == OBD_MD_FLXATTRRM);
438 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
439 fid, MDS_REINT, obd_md_valid, name,
440 value, value_size, 0, xattr_flags, suppgid,
444 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
445 u64 obd_md_valid, const char *name, size_t buf_size,
446 struct ptlrpc_request **req)
448 struct mdt_body *body;
451 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
452 obd_md_valid == OBD_MD_FLXATTRLS);
454 CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
455 exp->exp_obd->obd_name, name, PFID(fid));
456 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
457 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
462 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
464 GOTO(out, rc = -EPROTO);
466 /* only detect the xattr size */
468 /* LU-11109: Older MDTs do not distinguish
469 * between nonexistent xattrs and zero length
470 * values in this case. Newer MDTs will return
471 * -ENODATA or set OBD_MD_FLXATTR. */
472 GOTO(out, rc = body->mbo_eadatasize);
475 if (body->mbo_eadatasize == 0) {
476 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
477 * success so that we can distinguish between
478 * zero length value and nonexistent xattr.
480 * If OBD_MD_FLXATTR is not set then we keep
481 * the old behavior and return -ENODATA for
482 * getxattr() when mbo_eadatasize is 0. But
483 * -ENODATA only makes sense for getxattr()
484 * and not for listxattr(). */
485 if (body->mbo_valid & OBD_MD_FLXATTR)
487 else if (obd_md_valid == OBD_MD_FLXATTR)
488 GOTO(out, rc = -ENODATA);
493 GOTO(out, rc = body->mbo_eadatasize);
496 ptlrpc_req_finished(*req);
503 #ifdef CONFIG_FS_POSIX_ACL
504 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
506 struct req_capsule *pill = &req->rq_pill;
507 struct mdt_body *body = md->body;
508 struct posix_acl *acl;
513 if (!body->mbo_aclsize)
516 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
521 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
526 CERROR("convert xattr to acl: %d\n", rc);
530 rc = posix_acl_valid(&init_user_ns, acl);
532 CERROR("validate acl: %d\n", rc);
533 posix_acl_release(acl);
541 #define mdc_unpack_acl(req, md) 0
544 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
545 struct obd_export *dt_exp, struct obd_export *md_exp,
546 struct lustre_md *md)
548 struct req_capsule *pill = &req->rq_pill;
553 memset(md, 0, sizeof(*md));
555 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
556 LASSERT(md->body != NULL);
558 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
559 if (!S_ISREG(md->body->mbo_mode)) {
560 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
561 "regular file, but is not\n");
562 GOTO(out, rc = -EPROTO);
565 if (md->body->mbo_eadatasize == 0) {
566 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
567 "but eadatasize 0\n");
568 GOTO(out, rc = -EPROTO);
571 md->layout.lb_len = md->body->mbo_eadatasize;
572 md->layout.lb_buf = req_capsule_server_sized_get(pill,
575 if (md->layout.lb_buf == NULL)
576 GOTO(out, rc = -EPROTO);
577 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
578 const union lmv_mds_md *lmv;
581 if (!S_ISDIR(md->body->mbo_mode)) {
582 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
583 "directory, but is not\n");
584 GOTO(out, rc = -EPROTO);
587 lmv_size = md->body->mbo_eadatasize;
589 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
590 "but eadatasize 0\n");
594 if (md->body->mbo_valid & OBD_MD_MEA) {
595 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
598 GOTO(out, rc = -EPROTO);
600 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
604 if (rc < (typeof(rc))sizeof(*md->lmv)) {
605 CDEBUG(D_INFO, "size too small: "
606 "rc < sizeof(*md->lmv) (%d < %d)\n",
607 rc, (int)sizeof(*md->lmv));
608 GOTO(out, rc = -EPROTO);
614 if (md->body->mbo_valid & OBD_MD_FLACL) {
615 /* for ACL, it's possible that FLACL is set but aclsize is zero.
616 * only when aclsize != 0 there's an actual segment for ACL
619 if (md->body->mbo_aclsize) {
620 rc = mdc_unpack_acl(req, md);
623 #ifdef CONFIG_FS_POSIX_ACL
625 md->posix_acl = NULL;
633 #ifdef CONFIG_FS_POSIX_ACL
634 posix_acl_release(md->posix_acl);
640 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
646 void mdc_replay_open(struct ptlrpc_request *req)
648 struct md_open_data *mod = req->rq_cb_data;
649 struct ptlrpc_request *close_req;
650 struct obd_client_handle *och;
651 struct lustre_handle old_open_handle = { };
652 struct mdt_body *body;
656 DEBUG_REQ(D_ERROR, req,
657 "Can't properly replay without open data.");
662 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
663 LASSERT(body != NULL);
665 spin_lock(&req->rq_lock);
667 if (och && och->och_open_handle.cookie)
668 req->rq_early_free_repbuf = 1;
670 req->rq_early_free_repbuf = 0;
671 spin_unlock(&req->rq_lock);
673 if (req->rq_early_free_repbuf) {
674 struct lustre_handle *file_open_handle;
676 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
678 file_open_handle = &och->och_open_handle;
679 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
680 file_open_handle->cookie, body->mbo_open_handle.cookie);
681 old_open_handle = *file_open_handle;
682 *file_open_handle = body->mbo_open_handle;
685 close_req = mod->mod_close_req;
687 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
688 struct mdt_ioepoch *epoch;
690 LASSERT(opc == MDS_CLOSE);
691 epoch = req_capsule_client_get(&close_req->rq_pill,
695 if (req->rq_early_free_repbuf)
696 LASSERT(old_open_handle.cookie ==
697 epoch->mio_open_handle.cookie);
699 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
700 epoch->mio_open_handle = body->mbo_open_handle;
705 void mdc_commit_open(struct ptlrpc_request *req)
707 struct md_open_data *mod = req->rq_cb_data;
712 * No need to touch md_open_data::mod_och, it holds a reference on
713 * \var mod and will zero references to each other, \var mod will be
714 * freed after that when md_open_data::mod_och will put the reference.
718 * Do not let open request to disappear as it still may be needed
719 * for close rpc to happen (it may happen on evict only, otherwise
720 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
721 * called), just mark this rpc as committed to distinguish these 2
722 * cases, see mdc_close() for details. The open request reference will
723 * be put along with freeing \var mod.
725 ptlrpc_request_addref(req);
726 spin_lock(&req->rq_lock);
727 req->rq_committed = 1;
728 spin_unlock(&req->rq_lock);
729 req->rq_cb_data = NULL;
733 int mdc_set_open_replay_data(struct obd_export *exp,
734 struct obd_client_handle *och,
735 struct lookup_intent *it)
737 struct md_open_data *mod;
738 struct mdt_rec_create *rec;
739 struct mdt_body *body;
740 struct ptlrpc_request *open_req = it->it_request;
741 struct obd_import *imp = open_req->rq_import;
744 if (!open_req->rq_replay)
747 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
748 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
749 LASSERT(rec != NULL);
750 /* Incoming message in my byte order (it's been swabbed). */
751 /* Outgoing messages always in my byte order. */
752 LASSERT(body != NULL);
754 /* Only if the import is replayable, we set replay_open data */
755 if (och && imp->imp_replayable) {
756 mod = obd_mod_alloc();
758 DEBUG_REQ(D_ERROR, open_req,
759 "Can't allocate md_open_data");
764 * Take a reference on \var mod, to be freed on mdc_close().
765 * It protects \var mod from being freed on eviction (commit
766 * callback is called despite rq_replay flag).
767 * Another reference for \var och.
772 spin_lock(&open_req->rq_lock);
775 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
776 it_disposition(it, DISP_OPEN_STRIPE);
777 mod->mod_open_req = open_req;
778 open_req->rq_cb_data = mod;
779 open_req->rq_commit_cb = mdc_commit_open;
780 open_req->rq_early_free_repbuf = 1;
781 spin_unlock(&open_req->rq_lock);
784 rec->cr_fid2 = body->mbo_fid1;
785 rec->cr_open_handle_old = body->mbo_open_handle;
786 open_req->rq_replay_cb = mdc_replay_open;
787 if (!fid_is_sane(&body->mbo_fid1)) {
788 DEBUG_REQ(D_ERROR, open_req,
789 "saving replay request with insane FID " DFID,
790 PFID(&body->mbo_fid1));
794 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
798 static void mdc_free_open(struct md_open_data *mod)
802 if (mod->mod_is_create == 0 &&
803 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
807 * No reason to asssert here if the open request has
808 * rq_replay == 1. It means that mdc_close failed, and
809 * close request wasn`t sent. It is not fatal to client.
810 * The worst thing is eviction if the client gets open lock
813 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
814 "= %d\n", mod->mod_open_req->rq_replay);
816 ptlrpc_request_committed(mod->mod_open_req, committed);
817 if (mod->mod_close_req)
818 ptlrpc_request_committed(mod->mod_close_req, committed);
821 int mdc_clear_open_replay_data(struct obd_export *exp,
822 struct obd_client_handle *och)
824 struct md_open_data *mod = och->och_mod;
828 * It is possible to not have \var mod in a case of eviction between
829 * lookup and ll_file_open().
834 LASSERT(mod != LP_POISON);
835 LASSERT(mod->mod_open_req != NULL);
837 spin_lock(&mod->mod_open_req->rq_lock);
839 mod->mod_och->och_open_handle.cookie = 0;
840 mod->mod_open_req->rq_early_free_repbuf = 0;
841 spin_unlock(&mod->mod_open_req->rq_lock);
851 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
852 struct md_open_data *mod, struct ptlrpc_request **request)
854 struct obd_device *obd = class_exp2obd(exp);
855 struct ptlrpc_request *req;
856 struct req_format *req_fmt;
857 size_t u32_count = 0;
862 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
863 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
866 if (op_data->op_bias & MDS_CLOSE_INTENT) {
867 req_fmt = &RQF_MDS_CLOSE_INTENT;
868 if (op_data->op_bias & MDS_HSM_RELEASE) {
869 /* allocate a FID for volatile file */
870 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
873 CERROR("%s: "DFID" allocating FID: rc = %d\n",
874 obd->obd_name, PFID(&op_data->op_fid1),
876 /* save the errcode and proceed to close */
880 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
881 size_t count = op_data->op_data_size / sizeof(__u32);
883 if (count > INLINE_RESYNC_ARRAY_SIZE)
887 req_fmt = &RQF_MDS_CLOSE;
891 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
894 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
896 /* Ensure that this close's handle is fixed up during replay. */
897 if (likely(mod != NULL)) {
898 LASSERTF(mod->mod_open_req != NULL &&
899 mod->mod_open_req->rq_type != LI_POISON,
900 "POISONED open %p!\n", mod->mod_open_req);
902 mod->mod_close_req = req;
904 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
905 /* We no longer want to preserve this open for replay even
906 * though the open was committed. b=3632, b=3633 */
907 spin_lock(&mod->mod_open_req->rq_lock);
908 mod->mod_open_req->rq_replay = 0;
909 spin_unlock(&mod->mod_open_req->rq_lock);
911 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
915 * TODO: repeat close after errors
917 CWARN("%s: close of FID "DFID" failed, file reference will be "
918 "dropped when this client unmounts or is evicted\n",
919 obd->obd_name, PFID(&op_data->op_fid1));
920 GOTO(out, rc = -ENOMEM);
924 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
925 u32_count * sizeof(__u32));
927 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
929 ptlrpc_request_free(req);
934 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
935 * portal whose threads are not taking any DLM locks and are therefore
936 * always progressing */
937 req->rq_request_portal = MDS_READPAGE_PORTAL;
938 ptlrpc_at_set_req_timeout(req);
940 if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
941 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
942 OP_XVALID_LAZYBLOCKS);
944 mdc_close_pack(req, op_data);
946 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
947 obd->u.cli.cl_default_mds_easize);
949 ptlrpc_request_set_replen(req);
951 mdc_get_mod_rpc_slot(req, NULL);
952 rc = ptlrpc_queue_wait(req);
953 mdc_put_mod_rpc_slot(req, NULL);
955 if (req->rq_repmsg == NULL) {
956 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
959 rc = req->rq_status ?: -EIO;
960 } else if (rc == 0 || rc == -EAGAIN) {
961 struct mdt_body *body;
963 rc = lustre_msg_get_status(req->rq_repmsg);
964 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
965 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
970 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
973 } else if (rc == -ESTALE) {
975 * it can be allowed error after 3633 if open was committed and
976 * server failed before close was sent. Let's check if mod
977 * exists and return no error in that case
980 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
981 LASSERT(mod->mod_open_req != NULL);
982 if (mod->mod_open_req->rq_committed)
990 mod->mod_close_req = NULL;
991 /* Since now, mod is accessed through open_req only,
992 * thus close req does not keep a reference on mod anymore. */
997 RETURN(rc < 0 ? rc : saved_rc);
1000 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1001 u64 offset, struct page **pages, int npages,
1002 struct ptlrpc_request **request)
1004 struct ptlrpc_request *req;
1005 struct ptlrpc_bulk_desc *desc;
1007 wait_queue_head_t waitq;
1009 struct l_wait_info lwi;
1014 init_waitqueue_head(&waitq);
1017 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1021 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1023 ptlrpc_request_free(req);
1027 req->rq_request_portal = MDS_READPAGE_PORTAL;
1028 ptlrpc_at_set_req_timeout(req);
1030 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1031 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1033 &ptlrpc_bulk_kiov_pin_ops);
1035 ptlrpc_req_finished(req);
1039 /* NB req now owns desc and will free it when it gets freed */
1040 for (i = 0; i < npages; i++)
1041 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1044 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1046 ptlrpc_request_set_replen(req);
1047 rc = ptlrpc_queue_wait(req);
1049 ptlrpc_req_finished(req);
1050 if (rc != -ETIMEDOUT)
1054 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1055 CERROR("%s: too many resend retries: rc = %d\n",
1056 exp->exp_obd->obd_name, -EIO);
1059 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1061 l_wait_event(waitq, 0, &lwi);
1066 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1067 req->rq_bulk->bd_nob_transferred);
1069 ptlrpc_req_finished(req);
1073 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1074 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1075 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1076 PAGE_SIZE * npages);
1077 ptlrpc_req_finished(req);
1085 static void mdc_release_page(struct page *page, int remove)
1089 if (likely(page->mapping != NULL))
1090 truncate_complete_page(page->mapping, page);
1096 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1097 __u64 *start, __u64 *end, int hash64)
1100 * Complement of hash is used as an index so that
1101 * radix_tree_gang_lookup() can be used to find a page with starting
1102 * hash _smaller_ than one we are looking for.
1104 unsigned long offset = hash_x_index(*hash, hash64);
1108 spin_lock_irq(&mapping->tree_lock);
1109 found = radix_tree_gang_lookup(&mapping->page_tree,
1110 (void **)&page, offset, 1);
1111 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1112 struct lu_dirpage *dp;
1115 spin_unlock_irq(&mapping->tree_lock);
1117 * In contrast to find_lock_page() we are sure that directory
1118 * page cannot be truncated (while DLM lock is held) and,
1119 * hence, can avoid restart.
1121 * In fact, page cannot be locked here at all, because
1122 * mdc_read_page_remote does synchronous io.
1124 wait_on_page_locked(page);
1125 if (PageUptodate(page)) {
1127 if (BITS_PER_LONG == 32 && hash64) {
1128 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1129 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1130 *hash = *hash >> 32;
1132 *start = le64_to_cpu(dp->ldp_hash_start);
1133 *end = le64_to_cpu(dp->ldp_hash_end);
1135 if (unlikely(*start == 1 && *hash == 0))
1138 LASSERTF(*start <= *hash, "start = %#llx"
1139 ",end = %#llx,hash = %#llx\n",
1140 *start, *end, *hash);
1141 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1142 " hash %#llx\n", offset, *start, *end, *hash);
1145 mdc_release_page(page, 0);
1147 } else if (*end != *start && *hash == *end) {
1149 * upon hash collision, remove this page,
1150 * otherwise put page reference, and
1151 * mdc_read_page_remote() will issue RPC to
1152 * fetch the page we want.
1155 mdc_release_page(page,
1156 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1161 page = ERR_PTR(-EIO);
1164 spin_unlock_irq(&mapping->tree_lock);
1171 * Adjust a set of pages, each page containing an array of lu_dirpages,
1172 * so that each page can be used as a single logical lu_dirpage.
1174 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1175 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1176 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1177 * value is used as a cookie to request the next lu_dirpage in a
1178 * directory listing that spans multiple pages (two in this example):
1181 * .|--------v------- -----.
1182 * |s|e|f|p|ent|ent| ... |ent|
1183 * '--|-------------- -----' Each PAGE contains a single
1184 * '------. lu_dirpage.
1185 * .---------v------- -----.
1186 * |s|e|f|p|ent| 0 | ... | 0 |
1187 * '----------------- -----'
1189 * However, on hosts where the native VM page size (PAGE_SIZE) is
1190 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1191 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1192 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1193 * after it in the same PAGE (arrows simplified for brevity, but
1194 * in general e0==s1, e1==s2, etc.):
1196 * .-------------------- -----.
1197 * |s0|e0|f0|p|ent|ent| ... |ent|
1198 * |---v---------------- -----|
1199 * |s1|e1|f1|p|ent|ent| ... |ent|
1200 * |---v---------------- -----| Here, each PAGE contains
1201 * ... multiple lu_dirpages.
1202 * |---v---------------- -----|
1203 * |s'|e'|f'|p|ent|ent| ... |ent|
1204 * '---|---------------- -----'
1206 * .----------------------------.
1209 * This structure is transformed into a single logical lu_dirpage as follows:
1211 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1212 * labeled 'next PAGE'.
1214 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1215 * a hash collision with the next page exists.
1217 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1218 * to the first entry of the next lu_dirpage.
1220 #if PAGE_SIZE > LU_PAGE_SIZE
1221 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1225 for (i = 0; i < cfs_pgs; i++) {
1226 struct lu_dirpage *dp = kmap(pages[i]);
1227 struct lu_dirpage *first = dp;
1228 struct lu_dirent *end_dirent = NULL;
1229 struct lu_dirent *ent;
1230 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1231 __u32 flags = le32_to_cpu(dp->ldp_flags);
1233 while (--lu_pgs > 0) {
1234 ent = lu_dirent_start(dp);
1235 for (end_dirent = ent; ent != NULL;
1236 end_dirent = ent, ent = lu_dirent_next(ent));
1238 /* Advance dp to next lu_dirpage. */
1239 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1241 /* Check if we've reached the end of the PAGE. */
1242 if (!((unsigned long)dp & ~PAGE_MASK))
1245 /* Save the hash and flags of this lu_dirpage. */
1246 hash_end = le64_to_cpu(dp->ldp_hash_end);
1247 flags = le32_to_cpu(dp->ldp_flags);
1249 /* Check if lu_dirpage contains no entries. */
1250 if (end_dirent == NULL)
1253 /* Enlarge the end entry lde_reclen from 0 to
1254 * first entry of next lu_dirpage. */
1255 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1256 end_dirent->lde_reclen =
1257 cpu_to_le16((char *)(dp->ldp_entries) -
1258 (char *)end_dirent);
1261 first->ldp_hash_end = hash_end;
1262 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1263 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1267 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1270 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1271 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1273 /* parameters for readdir page */
1274 struct readpage_param {
1275 struct md_op_data *rp_mod;
1278 struct obd_export *rp_exp;
1279 struct md_callback *rp_cb;
1282 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1283 static inline void delete_from_page_cache(struct page *page)
1285 remove_from_page_cache(page);
1291 * Read pages from server.
1293 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1294 * a header lu_dirpage which describes the start/end hash, and whether this
1295 * page is empty (contains no dir entry) or hash collide with next page.
1296 * After client receives reply, several pages will be integrated into dir page
1297 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1298 * lu_dirpage for this integrated page will be adjusted.
1300 static int mdc_read_page_remote(void *data, struct page *page0)
1302 struct readpage_param *rp = data;
1303 struct page **page_pool;
1305 struct lu_dirpage *dp;
1306 struct md_op_data *op_data = rp->rp_mod;
1307 struct ptlrpc_request *req;
1309 struct inode *inode;
1311 int rd_pgs = 0; /* number of pages actually read */
1317 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1318 inode = op_data->op_data;
1319 fid = &op_data->op_fid1;
1320 LASSERT(inode != NULL);
1322 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1323 if (page_pool != NULL) {
1324 page_pool[0] = page0;
1330 for (npages = 1; npages < max_pages; npages++) {
1331 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1335 page_pool[npages] = page;
1338 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1340 /* page0 is special, which was added into page cache early */
1341 delete_from_page_cache(page0);
1345 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1347 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1348 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1350 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1352 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1354 SetPageUptodate(page0);
1358 ptlrpc_req_finished(req);
1359 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1360 for (i = 1; i < npages; i++) {
1361 unsigned long offset;
1365 page = page_pool[i];
1367 if (rc < 0 || i >= rd_pgs) {
1372 SetPageUptodate(page);
1375 hash = le64_to_cpu(dp->ldp_hash_start);
1378 offset = hash_x_index(hash, rp->rp_hash64);
1380 prefetchw(&page->flags);
1381 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1386 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1387 " rc = %d\n", offset, ret);
1391 if (page_pool != &page0)
1392 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1398 * Read dir page from cache first, if it can not find it, read it from
1399 * server and add into the cache.
1401 * \param[in] exp MDC export
1402 * \param[in] op_data client MD stack parameters, transfering parameters
1403 * between different layers on client MD stack.
1404 * \param[in] cb_op callback required for ldlm lock enqueue during
1406 * \param[in] hash_offset the hash offset of the page to be read
1407 * \param[in] ppage the page to be read
1409 * retval = 0 get the page successfully
1410 * errno(<0) get the page failed
1412 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1413 struct md_callback *cb_op, __u64 hash_offset,
1414 struct page **ppage)
1416 struct lookup_intent it = { .it_op = IT_READDIR };
1418 struct inode *dir = op_data->op_data;
1419 struct address_space *mapping;
1420 struct lu_dirpage *dp;
1423 struct lustre_handle lockh;
1424 struct ptlrpc_request *enq_req = NULL;
1425 struct readpage_param rp_param;
1432 LASSERT(dir != NULL);
1433 mapping = dir->i_mapping;
1435 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1436 cb_op->md_blocking_ast, 0);
1437 if (enq_req != NULL)
1438 ptlrpc_req_finished(enq_req);
1441 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1442 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1447 lockh.cookie = it.it_lock_handle;
1448 mdc_set_lock_data(exp, &lockh, dir, NULL);
1450 rp_param.rp_off = hash_offset;
1451 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1452 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1453 rp_param.rp_hash64);
1455 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1456 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1457 rp_param.rp_off, PTR_ERR(page));
1458 GOTO(out_unlock, rc = PTR_ERR(page));
1459 } else if (page != NULL) {
1461 * XXX nikita: not entirely correct handling of a corner case:
1462 * suppose hash chain of entries with hash value HASH crosses
1463 * border between pages P0 and P1. First both P0 and P1 are
1464 * cached, seekdir() is called for some entry from the P0 part
1465 * of the chain. Later P0 goes out of cache. telldir(HASH)
1466 * happens and finds P1, as it starts with matching hash
1467 * value. Remaining entries from P0 part of the chain are
1468 * skipped. (Is that really a bug?)
1470 * Possible solutions: 0. don't cache P1 is such case, handle
1471 * it as an "overflow" page. 1. invalidate all pages at
1472 * once. 2. use HASH|1 as an index for P1.
1474 GOTO(hash_collision, page);
1477 rp_param.rp_exp = exp;
1478 rp_param.rp_mod = op_data;
1479 page = read_cache_page(mapping,
1480 hash_x_index(rp_param.rp_off,
1481 rp_param.rp_hash64),
1482 mdc_read_page_remote, &rp_param);
1484 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1485 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1486 rp_param.rp_off, PTR_ERR(page));
1487 GOTO(out_unlock, rc = PTR_ERR(page));
1490 wait_on_page_locked(page);
1492 if (!PageUptodate(page)) {
1493 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1494 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1495 rp_param.rp_off, -5);
1498 if (!PageChecked(page))
1499 SetPageChecked(page);
1500 if (PageError(page)) {
1501 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1502 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1503 rp_param.rp_off, -5);
1508 dp = page_address(page);
1509 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1510 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1511 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1512 rp_param.rp_off = hash_offset >> 32;
1514 start = le64_to_cpu(dp->ldp_hash_start);
1515 end = le64_to_cpu(dp->ldp_hash_end);
1516 rp_param.rp_off = hash_offset;
1519 LASSERT(start == rp_param.rp_off);
1520 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1521 #if BITS_PER_LONG == 32
1522 CWARN("Real page-wide hash collision at [%llu %llu] with "
1523 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1524 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1528 * Fetch whole overflow chain...
1536 ldlm_lock_decref(&lockh, it.it_lock_mode);
1540 mdc_release_page(page, 1);
1545 static int mdc_statfs(const struct lu_env *env,
1546 struct obd_export *exp, struct obd_statfs *osfs,
1547 time64_t max_age, __u32 flags)
1549 struct obd_device *obd = class_exp2obd(exp);
1550 struct req_format *fmt;
1551 struct ptlrpc_request *req;
1552 struct obd_statfs *msfs;
1553 struct obd_import *imp = NULL;
1558 * Since the request might also come from lprocfs, so we need
1559 * sync this with client_disconnect_export Bug15684
1561 down_read(&obd->u.cli.cl_sem);
1562 if (obd->u.cli.cl_import)
1563 imp = class_import_get(obd->u.cli.cl_import);
1564 up_read(&obd->u.cli.cl_sem);
1568 fmt = &RQF_MDS_STATFS;
1569 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1570 (flags & OBD_STATFS_SUM))
1571 fmt = &RQF_MDS_STATFS_NEW;
1572 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1575 GOTO(output, rc = -ENOMEM);
1577 if ((flags & OBD_STATFS_SUM) &&
1578 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1579 /* request aggregated states */
1580 struct mdt_body *body;
1582 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1584 GOTO(out, rc = -EPROTO);
1585 body->mbo_valid = OBD_MD_FLAGSTATFS;
1588 ptlrpc_request_set_replen(req);
1590 if (flags & OBD_STATFS_NODELAY) {
1591 /* procfs requests not want stay in wait for avoid deadlock */
1592 req->rq_no_resend = 1;
1593 req->rq_no_delay = 1;
1596 rc = ptlrpc_queue_wait(req);
1598 /* check connection error first */
1599 if (imp->imp_connect_error)
1600 rc = imp->imp_connect_error;
1604 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1606 GOTO(out, rc = -EPROTO);
1611 ptlrpc_req_finished(req);
1613 class_import_put(imp);
1617 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1619 __u32 keylen, vallen;
1623 if (gf->gf_pathlen > PATH_MAX)
1624 RETURN(-ENAMETOOLONG);
1625 if (gf->gf_pathlen < 2)
1628 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1629 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1630 sizeof(struct lu_fid));
1631 OBD_ALLOC(key, keylen);
1634 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1635 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1636 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1637 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1638 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1639 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1641 if (!fid_is_sane(&gf->gf_fid))
1642 GOTO(out, rc = -EINVAL);
1644 /* Val is struct getinfo_fid2path result plus path */
1645 vallen = sizeof(*gf) + gf->gf_pathlen;
1647 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1648 if (rc != 0 && rc != -EREMOTE)
1651 if (vallen <= sizeof(*gf))
1652 GOTO(out, rc = -EPROTO);
1653 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1654 GOTO(out, rc = -EOVERFLOW);
1656 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1657 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1658 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1659 /* only log the last 512 characters of the path */
1660 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1663 OBD_FREE(key, keylen);
1667 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1668 struct hsm_progress_kernel *hpk)
1670 struct obd_import *imp = class_exp2cliimp(exp);
1671 struct hsm_progress_kernel *req_hpk;
1672 struct ptlrpc_request *req;
1676 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1677 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1679 GOTO(out, rc = -ENOMEM);
1681 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1683 /* Copy hsm_progress struct */
1684 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1685 if (req_hpk == NULL)
1686 GOTO(out, rc = -EPROTO);
1689 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1691 ptlrpc_request_set_replen(req);
1693 mdc_get_mod_rpc_slot(req, NULL);
1694 rc = ptlrpc_queue_wait(req);
1695 mdc_put_mod_rpc_slot(req, NULL);
1699 ptlrpc_req_finished(req);
1703 * Send hsm_ct_register to MDS
1705 * \param[in] imp import
1706 * \param[in] archive_count if in bitmap format, it is the bitmap,
1707 * else it is the count of archive_ids
1708 * \param[in] archives if in bitmap format, it is NULL,
1709 * else it is archive_id lists
1711 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1714 struct ptlrpc_request *req;
1715 __u32 *archive_array;
1716 size_t archives_size;
1720 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1724 if (archives != NULL)
1725 archives_size = sizeof(*archive_array) * archive_count;
1727 archives_size = sizeof(archive_count);
1729 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1730 RCL_CLIENT, archives_size);
1732 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1734 ptlrpc_request_free(req);
1738 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1740 archive_array = req_capsule_client_get(&req->rq_pill,
1741 &RMF_MDS_HSM_ARCHIVE);
1742 if (archive_array == NULL)
1743 GOTO(out, rc = -EPROTO);
1745 if (archives != NULL)
1746 memcpy(archive_array, archives, archives_size);
1748 *archive_array = archive_count;
1750 ptlrpc_request_set_replen(req);
1752 rc = mdc_queue_wait(req);
1755 ptlrpc_req_finished(req);
1759 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1760 struct md_op_data *op_data)
1762 struct hsm_current_action *hca = op_data->op_data;
1763 struct hsm_current_action *req_hca;
1764 struct ptlrpc_request *req;
1768 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1769 &RQF_MDS_HSM_ACTION);
1773 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1775 ptlrpc_request_free(req);
1779 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1780 op_data->op_suppgids[0], 0);
1782 ptlrpc_request_set_replen(req);
1784 rc = mdc_queue_wait(req);
1788 req_hca = req_capsule_server_get(&req->rq_pill,
1789 &RMF_MDS_HSM_CURRENT_ACTION);
1790 if (req_hca == NULL)
1791 GOTO(out, rc = -EPROTO);
1797 ptlrpc_req_finished(req);
1801 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1803 struct ptlrpc_request *req;
1807 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1809 MDS_HSM_CT_UNREGISTER);
1811 GOTO(out, rc = -ENOMEM);
1813 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1815 ptlrpc_request_set_replen(req);
1817 rc = mdc_queue_wait(req);
1820 ptlrpc_req_finished(req);
1824 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1825 struct md_op_data *op_data)
1827 struct hsm_user_state *hus = op_data->op_data;
1828 struct hsm_user_state *req_hus;
1829 struct ptlrpc_request *req;
1833 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1834 &RQF_MDS_HSM_STATE_GET);
1838 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1840 ptlrpc_request_free(req);
1844 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1845 op_data->op_suppgids[0], 0);
1847 ptlrpc_request_set_replen(req);
1849 rc = mdc_queue_wait(req);
1853 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1854 if (req_hus == NULL)
1855 GOTO(out, rc = -EPROTO);
1861 ptlrpc_req_finished(req);
1865 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1866 struct md_op_data *op_data)
1868 struct hsm_state_set *hss = op_data->op_data;
1869 struct hsm_state_set *req_hss;
1870 struct ptlrpc_request *req;
1874 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1875 &RQF_MDS_HSM_STATE_SET);
1879 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1881 ptlrpc_request_free(req);
1885 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1886 op_data->op_suppgids[0], 0);
1889 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1890 if (req_hss == NULL)
1891 GOTO(out, rc = -EPROTO);
1894 ptlrpc_request_set_replen(req);
1896 mdc_get_mod_rpc_slot(req, NULL);
1897 rc = ptlrpc_queue_wait(req);
1898 mdc_put_mod_rpc_slot(req, NULL);
1902 ptlrpc_req_finished(req);
1906 static int mdc_ioc_hsm_request(struct obd_export *exp,
1907 struct hsm_user_request *hur)
1909 struct obd_import *imp = class_exp2cliimp(exp);
1910 struct ptlrpc_request *req;
1911 struct hsm_request *req_hr;
1912 struct hsm_user_item *req_hui;
1917 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1919 GOTO(out, rc = -ENOMEM);
1921 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1922 hur->hur_request.hr_itemcount
1923 * sizeof(struct hsm_user_item));
1924 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1925 hur->hur_request.hr_data_len);
1927 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1929 ptlrpc_request_free(req);
1933 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1935 /* Copy hsm_request struct */
1936 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1938 GOTO(out, rc = -EPROTO);
1939 *req_hr = hur->hur_request;
1941 /* Copy hsm_user_item structs */
1942 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1943 if (req_hui == NULL)
1944 GOTO(out, rc = -EPROTO);
1945 memcpy(req_hui, hur->hur_user_item,
1946 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1948 /* Copy opaque field */
1949 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1950 if (req_opaque == NULL)
1951 GOTO(out, rc = -EPROTO);
1952 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1954 ptlrpc_request_set_replen(req);
1956 mdc_get_mod_rpc_slot(req, NULL);
1957 rc = ptlrpc_queue_wait(req);
1958 mdc_put_mod_rpc_slot(req, NULL);
1963 ptlrpc_req_finished(req);
1967 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1968 struct lustre_kernelcomm *lk);
1970 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1971 struct obd_quotactl *oqctl)
1973 struct ptlrpc_request *req;
1974 struct obd_quotactl *oqc;
1978 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1979 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1984 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1987 ptlrpc_request_set_replen(req);
1988 ptlrpc_at_set_req_timeout(req);
1990 rc = ptlrpc_queue_wait(req);
1992 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1994 if (req->rq_repmsg &&
1995 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1998 CERROR ("Can't unpack obd_quotactl\n");
2001 ptlrpc_req_finished(req);
2006 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2007 struct md_op_data *op_data)
2009 struct list_head cancels = LIST_HEAD_INIT(cancels);
2010 struct ptlrpc_request *req;
2012 struct mdc_swap_layouts *msl, *payload;
2015 msl = op_data->op_data;
2017 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2018 * first thing it will do is to cancel the 2 layout
2019 * locks held by this client.
2020 * So the client must cancel its layout locks on the 2 fids
2021 * with the request RPC to avoid extra RPC round trips.
2023 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2024 LCK_EX, MDS_INODELOCK_LAYOUT |
2025 MDS_INODELOCK_XATTR);
2026 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2027 LCK_EX, MDS_INODELOCK_LAYOUT |
2028 MDS_INODELOCK_XATTR);
2030 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2031 &RQF_MDS_SWAP_LAYOUTS);
2033 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2037 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2039 ptlrpc_request_free(req);
2043 mdc_swap_layouts_pack(req, op_data);
2045 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2050 ptlrpc_request_set_replen(req);
2052 rc = ptlrpc_queue_wait(req);
2058 ptlrpc_req_finished(req);
2062 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2063 void *karg, void __user *uarg)
2065 struct obd_device *obd = exp->exp_obd;
2066 struct obd_ioctl_data *data = karg;
2067 struct obd_import *imp = obd->u.cli.cl_import;
2071 if (!try_module_get(THIS_MODULE)) {
2072 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2073 module_name(THIS_MODULE));
2077 case OBD_IOC_FID2PATH:
2078 rc = mdc_ioc_fid2path(exp, karg);
2080 case LL_IOC_HSM_CT_START:
2081 rc = mdc_ioc_hsm_ct_start(exp, karg);
2082 /* ignore if it was already registered on this MDS. */
2086 case LL_IOC_HSM_PROGRESS:
2087 rc = mdc_ioc_hsm_progress(exp, karg);
2089 case LL_IOC_HSM_STATE_GET:
2090 rc = mdc_ioc_hsm_state_get(exp, karg);
2092 case LL_IOC_HSM_STATE_SET:
2093 rc = mdc_ioc_hsm_state_set(exp, karg);
2095 case LL_IOC_HSM_ACTION:
2096 rc = mdc_ioc_hsm_current_action(exp, karg);
2098 case LL_IOC_HSM_REQUEST:
2099 rc = mdc_ioc_hsm_request(exp, karg);
2101 case OBD_IOC_CLIENT_RECOVER:
2102 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2106 case IOC_OSC_SET_ACTIVE:
2107 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2110 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2111 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2112 * there'd be no LMV layer thus we might be called here. Eventually
2113 * this code should be removed.
2116 case IOC_OBD_STATFS: {
2117 struct obd_statfs stat_buf = {0};
2119 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2120 GOTO(out, rc = -ENODEV);
2123 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2124 min((int)data->ioc_plen2,
2125 (int)sizeof(struct obd_uuid))))
2126 GOTO(out, rc = -EFAULT);
2128 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2129 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2134 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2135 min((int) data->ioc_plen1,
2136 (int) sizeof(stat_buf))))
2137 GOTO(out, rc = -EFAULT);
2141 case OBD_IOC_QUOTACTL: {
2142 struct if_quotactl *qctl = karg;
2143 struct obd_quotactl *oqctl;
2145 OBD_ALLOC_PTR(oqctl);
2147 GOTO(out, rc = -ENOMEM);
2149 QCTL_COPY(oqctl, qctl);
2150 rc = obd_quotactl(exp, oqctl);
2152 QCTL_COPY(qctl, oqctl);
2153 qctl->qc_valid = QC_MDTIDX;
2154 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2157 OBD_FREE_PTR(oqctl);
2160 case LL_IOC_GET_CONNECT_FLAGS:
2161 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2162 sizeof(*exp_connect_flags_ptr(exp))))
2163 GOTO(out, rc = -EFAULT);
2166 case LL_IOC_LOV_SWAP_LAYOUTS:
2167 rc = mdc_ioc_swap_layouts(exp, karg);
2170 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2171 GOTO(out, rc = -ENOTTY);
2174 module_put(THIS_MODULE);
2179 static int mdc_get_info_rpc(struct obd_export *exp,
2180 u32 keylen, void *key,
2181 u32 vallen, void *val)
2183 struct obd_import *imp = class_exp2cliimp(exp);
2184 struct ptlrpc_request *req;
2189 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2193 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2194 RCL_CLIENT, keylen);
2195 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2196 RCL_CLIENT, sizeof(vallen));
2198 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2200 ptlrpc_request_free(req);
2204 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2205 memcpy(tmp, key, keylen);
2206 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2207 memcpy(tmp, &vallen, sizeof(vallen));
2209 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2210 RCL_SERVER, vallen);
2211 ptlrpc_request_set_replen(req);
2213 rc = ptlrpc_queue_wait(req);
2214 /* -EREMOTE means the get_info result is partial, and it needs to
2215 * continue on another MDT, see fid2path part in lmv_iocontrol */
2216 if (rc == 0 || rc == -EREMOTE) {
2217 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2218 memcpy(val, tmp, vallen);
2219 if (ptlrpc_rep_need_swab(req)) {
2220 if (KEY_IS(KEY_FID2PATH))
2221 lustre_swab_fid2path(val);
2224 ptlrpc_req_finished(req);
2229 static void lustre_swab_hai(struct hsm_action_item *h)
2231 __swab32s(&h->hai_len);
2232 __swab32s(&h->hai_action);
2233 lustre_swab_lu_fid(&h->hai_fid);
2234 lustre_swab_lu_fid(&h->hai_dfid);
2235 __swab64s(&h->hai_cookie);
2236 __swab64s(&h->hai_extent.offset);
2237 __swab64s(&h->hai_extent.length);
2238 __swab64s(&h->hai_gid);
2241 static void lustre_swab_hal(struct hsm_action_list *h)
2243 struct hsm_action_item *hai;
2246 __swab32s(&h->hal_version);
2247 __swab32s(&h->hal_count);
2248 __swab32s(&h->hal_archive_id);
2249 __swab64s(&h->hal_flags);
2251 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2252 lustre_swab_hai(hai);
2255 static void lustre_swab_kuch(struct kuc_hdr *l)
2257 __swab16s(&l->kuc_magic);
2258 /* __u8 l->kuc_transport */
2259 __swab16s(&l->kuc_msgtype);
2260 __swab16s(&l->kuc_msglen);
2263 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2264 struct lustre_kernelcomm *lk)
2266 struct obd_import *imp = class_exp2cliimp(exp);
2269 if (lk->lk_group != KUC_GRP_HSM) {
2270 CERROR("Bad copytool group %d\n", lk->lk_group);
2274 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2275 lk->lk_uid, lk->lk_group, lk->lk_flags);
2277 if (lk->lk_flags & LK_FLG_STOP) {
2278 /* Unregister with the coordinator */
2279 rc = mdc_ioc_hsm_ct_unregister(imp);
2281 __u32 *archives = NULL;
2283 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2284 archives = lk->lk_data;
2286 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2293 * Send a message to any listening copytools
2294 * @param val KUC message (kuc_hdr + hsm_action_list)
2295 * @param len total length of message
2297 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2298 size_t len, void *val)
2300 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2301 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2305 if (len < sizeof(*lh) + sizeof(*hal)) {
2306 CERROR("Short HSM message %zu < %zu\n", len,
2307 sizeof(*lh) + sizeof(*hal));
2310 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2311 lustre_swab_kuch(lh);
2312 lustre_swab_hal(hal);
2313 } else if (lh->kuc_magic != KUC_MAGIC) {
2314 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2318 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2320 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2321 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2323 /* Broadcast to HSM listeners */
2324 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2330 * callback function passed to kuc for re-registering each HSM copytool
2331 * running on MDC, after MDT shutdown/recovery.
2332 * @param data copytool registration data
2333 * @param cb_arg callback argument (obd_import)
2335 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2337 struct obd_import *imp = (struct obd_import *)cb_arg;
2338 struct kkuc_ct_data *kcd = data;
2339 __u32 *archives = NULL;
2343 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2344 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2347 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2348 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2349 "(archive=%#x)\n", imp->imp_obd->obd_name,
2350 kcd->kcd_nr_archives);
2352 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2353 "(archive nr = %u)\n",
2354 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2355 if (kcd->kcd_nr_archives != 0)
2356 archives = kcd->kcd_archives;
2359 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2360 /* ignore error if the copytool is already registered */
2361 return (rc == -EEXIST) ? 0 : rc;
2365 * Re-establish all kuc contexts with MDT
2366 * after MDT shutdown/recovery.
2368 static int mdc_kuc_reregister(struct obd_import *imp)
2370 /* re-register HSM agents */
2371 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2372 mdc_hsm_ct_reregister, imp);
2375 static int mdc_set_info_async(const struct lu_env *env,
2376 struct obd_export *exp,
2377 u32 keylen, void *key,
2378 u32 vallen, void *val,
2379 struct ptlrpc_request_set *set)
2381 struct obd_import *imp = class_exp2cliimp(exp);
2385 if (KEY_IS(KEY_READ_ONLY)) {
2386 if (vallen != sizeof(int))
2389 spin_lock(&imp->imp_lock);
2390 if (*((int *)val)) {
2391 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2392 imp->imp_connect_data.ocd_connect_flags |=
2395 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2396 imp->imp_connect_data.ocd_connect_flags &=
2397 ~OBD_CONNECT_RDONLY;
2399 spin_unlock(&imp->imp_lock);
2401 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2402 keylen, key, vallen, val, set);
2405 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2406 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2407 keylen, key, vallen, val, set);
2410 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2411 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2416 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2417 __u32 *default_easize = val;
2419 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2423 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2427 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2428 __u32 keylen, void *key, __u32 *vallen, void *val)
2432 if (KEY_IS(KEY_MAX_EASIZE)) {
2433 __u32 mdsize, *max_easize;
2435 if (*vallen != sizeof(int))
2437 mdsize = *(__u32 *)val;
2438 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2439 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2441 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2443 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2444 __u32 *default_easize;
2446 if (*vallen != sizeof(int))
2448 default_easize = val;
2449 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2451 } else if (KEY_IS(KEY_CONN_DATA)) {
2452 struct obd_import *imp = class_exp2cliimp(exp);
2453 struct obd_connect_data *data = val;
2455 if (*vallen != sizeof(*data))
2458 *data = imp->imp_connect_data;
2460 } else if (KEY_IS(KEY_TGT_COUNT)) {
2461 *((__u32 *)val) = 1;
2465 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2470 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2471 struct ptlrpc_request **request)
2473 struct ptlrpc_request *req;
2478 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2482 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2484 ptlrpc_request_free(req);
2488 mdc_pack_body(req, fid, 0, 0, -1, 0);
2490 ptlrpc_request_set_replen(req);
2492 rc = ptlrpc_queue_wait(req);
2494 ptlrpc_req_finished(req);
2500 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2501 enum obd_import_event event)
2503 struct client_obd *cli = &obd->u.cli;
2506 LASSERT(imp->imp_obd == obd);
2509 case IMP_EVENT_DISCON:
2510 spin_lock(&cli->cl_loi_list_lock);
2511 cli->cl_avail_grant = 0;
2512 cli->cl_lost_grant = 0;
2513 spin_unlock(&cli->cl_loi_list_lock);
2515 case IMP_EVENT_INACTIVE:
2517 * Flush current sequence to make client obtain new one
2518 * from server in case of disconnect/reconnect.
2520 down_read(&cli->cl_seq_rwsem);
2522 seq_client_flush(cli->cl_seq);
2523 up_read(&cli->cl_seq_rwsem);
2525 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2527 case IMP_EVENT_INVALIDATE: {
2528 struct ldlm_namespace *ns = obd->obd_namespace;
2532 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2534 env = cl_env_get(&refcheck);
2536 /* Reset grants. All pages go to failing rpcs due to
2537 * the invalid import.
2539 osc_io_unplug(env, cli, NULL);
2541 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2542 osc_ldlm_resource_invalidate,
2544 cl_env_put(env, &refcheck);
2545 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2551 case IMP_EVENT_ACTIVE:
2552 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2553 /* redo the kuc registration after reconnecting */
2555 rc = mdc_kuc_reregister(imp);
2557 case IMP_EVENT_OCD: {
2558 struct obd_connect_data *ocd = &imp->imp_connect_data;
2560 if (OCD_HAS_FLAG(ocd, GRANT))
2561 osc_init_grant(cli, ocd);
2563 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2566 case IMP_EVENT_DEACTIVATE:
2567 case IMP_EVENT_ACTIVATE:
2570 CERROR("Unknown import event %x\n", event);
2576 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2577 struct lu_fid *fid, struct md_op_data *op_data)
2579 struct client_obd *cli = &exp->exp_obd->u.cli;
2584 down_read(&cli->cl_seq_rwsem);
2586 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2587 up_read(&cli->cl_seq_rwsem);
2592 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2594 struct client_obd *cli = &exp->exp_obd->u.cli;
2595 return &cli->cl_target_uuid;
2599 * Determine whether the lock can be canceled before replaying it during
2600 * recovery, non zero value will be return if the lock can be canceled,
2601 * or zero returned for not
2603 static int mdc_cancel_weight(struct ldlm_lock *lock)
2605 if (lock->l_resource->lr_type != LDLM_IBITS)
2608 /* FIXME: if we ever get into a situation where there are too many
2609 * opened files with open locks on a single node, then we really
2610 * should replay these open locks to reget it */
2611 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2614 /* Special case for DoM locks, cancel only unused and granted locks */
2615 if (ldlm_has_dom(lock) &&
2616 (lock->l_granted_mode != lock->l_req_mode ||
2617 osc_ldlm_weigh_ast(lock) != 0))
2623 static int mdc_resource_inode_free(struct ldlm_resource *res)
2625 if (res->lr_lvb_inode)
2626 res->lr_lvb_inode = NULL;
2631 static struct ldlm_valblock_ops inode_lvbo = {
2632 .lvbo_free = mdc_resource_inode_free
2635 static int mdc_llog_init(struct obd_device *obd)
2637 struct obd_llog_group *olg = &obd->obd_olg;
2638 struct llog_ctxt *ctxt;
2643 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2648 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2649 llog_initiator_connect(ctxt);
2650 llog_ctxt_put(ctxt);
2655 static void mdc_llog_finish(struct obd_device *obd)
2657 struct llog_ctxt *ctxt;
2661 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2663 llog_cleanup(NULL, ctxt);
2668 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2674 rc = osc_setup_common(obd, cfg);
2678 rc = mdc_tunables_init(obd);
2680 GOTO(err_osc_cleanup, rc);
2682 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2684 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2686 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2688 rc = mdc_llog_init(obd);
2690 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2692 GOTO(err_llog_cleanup, rc);
2695 rc = mdc_changelog_cdev_init(obd);
2697 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2699 GOTO(err_changelog_cleanup, rc);
2704 err_changelog_cleanup:
2705 mdc_llog_finish(obd);
2707 lprocfs_free_md_stats(obd);
2708 ptlrpc_lprocfs_unregister_obd(obd);
2710 osc_cleanup_common(obd);
2714 /* Initialize the default and maximum LOV EA sizes. This allows
2715 * us to make MDS RPCs with large enough reply buffers to hold a default
2716 * sized EA without having to calculate this (via a call into the
2717 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2718 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2719 * a large number of stripes is possible. If a larger reply buffer is
2720 * required it will be reallocated in the ptlrpc layer due to overflow.
2722 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2725 struct obd_device *obd = exp->exp_obd;
2726 struct client_obd *cli = &obd->u.cli;
2729 if (cli->cl_max_mds_easize < easize)
2730 cli->cl_max_mds_easize = easize;
2732 if (cli->cl_default_mds_easize < def_easize)
2733 cli->cl_default_mds_easize = def_easize;
2738 static int mdc_precleanup(struct obd_device *obd)
2742 osc_precleanup_common(obd);
2743 mdc_changelog_cdev_finish(obd);
2745 obd_cleanup_client_import(obd);
2746 ptlrpc_lprocfs_unregister_obd(obd);
2747 lprocfs_free_md_stats(obd);
2748 mdc_llog_finish(obd);
2752 static int mdc_cleanup(struct obd_device *obd)
2754 return osc_cleanup_common(obd);
2757 static struct obd_ops mdc_obd_ops = {
2758 .o_owner = THIS_MODULE,
2759 .o_setup = mdc_setup,
2760 .o_precleanup = mdc_precleanup,
2761 .o_cleanup = mdc_cleanup,
2762 .o_add_conn = client_import_add_conn,
2763 .o_del_conn = client_import_del_conn,
2764 .o_connect = client_connect_import,
2765 .o_reconnect = osc_reconnect,
2766 .o_disconnect = osc_disconnect,
2767 .o_iocontrol = mdc_iocontrol,
2768 .o_set_info_async = mdc_set_info_async,
2769 .o_statfs = mdc_statfs,
2770 .o_fid_init = client_fid_init,
2771 .o_fid_fini = client_fid_fini,
2772 .o_fid_alloc = mdc_fid_alloc,
2773 .o_import_event = mdc_import_event,
2774 .o_get_info = mdc_get_info,
2775 .o_get_uuid = mdc_get_uuid,
2776 .o_quotactl = mdc_quotactl,
2779 static struct md_ops mdc_md_ops = {
2780 .m_get_root = mdc_get_root,
2781 .m_null_inode = mdc_null_inode,
2782 .m_close = mdc_close,
2783 .m_create = mdc_create,
2784 .m_enqueue = mdc_enqueue,
2785 .m_getattr = mdc_getattr,
2786 .m_getattr_name = mdc_getattr_name,
2787 .m_intent_lock = mdc_intent_lock,
2789 .m_rename = mdc_rename,
2790 .m_setattr = mdc_setattr,
2791 .m_setxattr = mdc_setxattr,
2792 .m_getxattr = mdc_getxattr,
2793 .m_fsync = mdc_fsync,
2794 .m_file_resync = mdc_file_resync,
2795 .m_read_page = mdc_read_page,
2796 .m_unlink = mdc_unlink,
2797 .m_cancel_unused = mdc_cancel_unused,
2798 .m_init_ea_size = mdc_init_ea_size,
2799 .m_set_lock_data = mdc_set_lock_data,
2800 .m_lock_match = mdc_lock_match,
2801 .m_get_lustre_md = mdc_get_lustre_md,
2802 .m_free_lustre_md = mdc_free_lustre_md,
2803 .m_set_open_replay_data = mdc_set_open_replay_data,
2804 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2805 .m_intent_getattr_async = mdc_intent_getattr_async,
2806 .m_revalidate_lock = mdc_revalidate_lock
2809 static int __init mdc_init(void)
2811 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2812 LUSTRE_MDC_NAME, &mdc_device_type);
2815 static void __exit mdc_exit(void)
2817 class_unregister_type(LUSTRE_MDC_NAME);
2820 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2821 MODULE_DESCRIPTION("Lustre Metadata Client");
2822 MODULE_VERSION(LUSTRE_VERSION_STRING);
2823 MODULE_LICENSE("GPL");
2825 module_init(mdc_init);
2826 module_exit(mdc_exit);