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 struct lmv_foreign_md *lfm = md->lfm;
607 /* short (< sizeof(struct lmv_stripe_md))
610 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
612 "size too small: rc < sizeof(*md->lmv) (%d < %d)\n",
613 rc, (int)sizeof(*md->lmv));
614 GOTO(out, rc = -EPROTO);
621 if (md->body->mbo_valid & OBD_MD_FLACL) {
622 /* for ACL, it's possible that FLACL is set but aclsize is zero.
623 * only when aclsize != 0 there's an actual segment for ACL
626 if (md->body->mbo_aclsize) {
627 rc = mdc_unpack_acl(req, md);
630 #ifdef CONFIG_FS_POSIX_ACL
632 md->posix_acl = NULL;
640 #ifdef CONFIG_FS_POSIX_ACL
641 posix_acl_release(md->posix_acl);
647 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
653 void mdc_replay_open(struct ptlrpc_request *req)
655 struct md_open_data *mod = req->rq_cb_data;
656 struct ptlrpc_request *close_req;
657 struct obd_client_handle *och;
658 struct lustre_handle old_open_handle = { };
659 struct mdt_body *body;
663 DEBUG_REQ(D_ERROR, req,
664 "Can't properly replay without open data.");
669 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
670 LASSERT(body != NULL);
672 spin_lock(&req->rq_lock);
674 if (och && och->och_open_handle.cookie)
675 req->rq_early_free_repbuf = 1;
677 req->rq_early_free_repbuf = 0;
678 spin_unlock(&req->rq_lock);
680 if (req->rq_early_free_repbuf) {
681 struct lustre_handle *file_open_handle;
683 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
685 file_open_handle = &och->och_open_handle;
686 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
687 file_open_handle->cookie, body->mbo_open_handle.cookie);
688 old_open_handle = *file_open_handle;
689 *file_open_handle = body->mbo_open_handle;
692 close_req = mod->mod_close_req;
694 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
695 struct mdt_ioepoch *epoch;
697 LASSERT(opc == MDS_CLOSE);
698 epoch = req_capsule_client_get(&close_req->rq_pill,
702 if (req->rq_early_free_repbuf)
703 LASSERT(old_open_handle.cookie ==
704 epoch->mio_open_handle.cookie);
706 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
707 epoch->mio_open_handle = body->mbo_open_handle;
712 void mdc_commit_open(struct ptlrpc_request *req)
714 struct md_open_data *mod = req->rq_cb_data;
719 * No need to touch md_open_data::mod_och, it holds a reference on
720 * \var mod and will zero references to each other, \var mod will be
721 * freed after that when md_open_data::mod_och will put the reference.
725 * Do not let open request to disappear as it still may be needed
726 * for close rpc to happen (it may happen on evict only, otherwise
727 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
728 * called), just mark this rpc as committed to distinguish these 2
729 * cases, see mdc_close() for details. The open request reference will
730 * be put along with freeing \var mod.
732 ptlrpc_request_addref(req);
733 spin_lock(&req->rq_lock);
734 req->rq_committed = 1;
735 spin_unlock(&req->rq_lock);
736 req->rq_cb_data = NULL;
740 int mdc_set_open_replay_data(struct obd_export *exp,
741 struct obd_client_handle *och,
742 struct lookup_intent *it)
744 struct md_open_data *mod;
745 struct mdt_rec_create *rec;
746 struct mdt_body *body;
747 struct ptlrpc_request *open_req = it->it_request;
748 struct obd_import *imp = open_req->rq_import;
751 if (!open_req->rq_replay)
754 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
755 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
756 LASSERT(rec != NULL);
757 /* Incoming message in my byte order (it's been swabbed). */
758 /* Outgoing messages always in my byte order. */
759 LASSERT(body != NULL);
761 /* Only if the import is replayable, we set replay_open data */
762 if (och && imp->imp_replayable) {
763 mod = obd_mod_alloc();
765 DEBUG_REQ(D_ERROR, open_req,
766 "Can't allocate md_open_data");
771 * Take a reference on \var mod, to be freed on mdc_close().
772 * It protects \var mod from being freed on eviction (commit
773 * callback is called despite rq_replay flag).
774 * Another reference for \var och.
779 spin_lock(&open_req->rq_lock);
782 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
783 it_disposition(it, DISP_OPEN_STRIPE);
784 mod->mod_open_req = open_req;
785 open_req->rq_cb_data = mod;
786 open_req->rq_commit_cb = mdc_commit_open;
787 open_req->rq_early_free_repbuf = 1;
788 spin_unlock(&open_req->rq_lock);
791 rec->cr_fid2 = body->mbo_fid1;
792 rec->cr_open_handle_old = body->mbo_open_handle;
793 open_req->rq_replay_cb = mdc_replay_open;
794 if (!fid_is_sane(&body->mbo_fid1)) {
795 DEBUG_REQ(D_ERROR, open_req,
796 "saving replay request with insane FID " DFID,
797 PFID(&body->mbo_fid1));
801 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
805 static void mdc_free_open(struct md_open_data *mod)
809 if (mod->mod_is_create == 0 &&
810 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
814 * No reason to asssert here if the open request has
815 * rq_replay == 1. It means that mdc_close failed, and
816 * close request wasn`t sent. It is not fatal to client.
817 * The worst thing is eviction if the client gets open lock
820 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
821 "= %d\n", mod->mod_open_req->rq_replay);
823 ptlrpc_request_committed(mod->mod_open_req, committed);
824 if (mod->mod_close_req)
825 ptlrpc_request_committed(mod->mod_close_req, committed);
828 int mdc_clear_open_replay_data(struct obd_export *exp,
829 struct obd_client_handle *och)
831 struct md_open_data *mod = och->och_mod;
835 * It is possible to not have \var mod in a case of eviction between
836 * lookup and ll_file_open().
841 LASSERT(mod != LP_POISON);
842 LASSERT(mod->mod_open_req != NULL);
844 spin_lock(&mod->mod_open_req->rq_lock);
846 mod->mod_och->och_open_handle.cookie = 0;
847 mod->mod_open_req->rq_early_free_repbuf = 0;
848 spin_unlock(&mod->mod_open_req->rq_lock);
858 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
859 struct md_open_data *mod, struct ptlrpc_request **request)
861 struct obd_device *obd = class_exp2obd(exp);
862 struct ptlrpc_request *req;
863 struct req_format *req_fmt;
864 size_t u32_count = 0;
869 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
870 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
873 if (op_data->op_bias & MDS_CLOSE_INTENT) {
874 req_fmt = &RQF_MDS_CLOSE_INTENT;
875 if (op_data->op_bias & MDS_HSM_RELEASE) {
876 /* allocate a FID for volatile file */
877 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
880 CERROR("%s: "DFID" allocating FID: rc = %d\n",
881 obd->obd_name, PFID(&op_data->op_fid1),
883 /* save the errcode and proceed to close */
887 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
888 size_t count = op_data->op_data_size / sizeof(__u32);
890 if (count > INLINE_RESYNC_ARRAY_SIZE)
894 req_fmt = &RQF_MDS_CLOSE;
898 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
901 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
903 /* Ensure that this close's handle is fixed up during replay. */
904 if (likely(mod != NULL)) {
905 LASSERTF(mod->mod_open_req != NULL &&
906 mod->mod_open_req->rq_type != LI_POISON,
907 "POISONED open %p!\n", mod->mod_open_req);
909 mod->mod_close_req = req;
911 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
912 /* We no longer want to preserve this open for replay even
913 * though the open was committed. b=3632, b=3633 */
914 spin_lock(&mod->mod_open_req->rq_lock);
915 mod->mod_open_req->rq_replay = 0;
916 spin_unlock(&mod->mod_open_req->rq_lock);
918 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
922 * TODO: repeat close after errors
924 CWARN("%s: close of FID "DFID" failed, file reference will be "
925 "dropped when this client unmounts or is evicted\n",
926 obd->obd_name, PFID(&op_data->op_fid1));
927 GOTO(out, rc = -ENOMEM);
931 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
932 u32_count * sizeof(__u32));
934 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
936 ptlrpc_request_free(req);
941 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
942 * portal whose threads are not taking any DLM locks and are therefore
943 * always progressing */
944 req->rq_request_portal = MDS_READPAGE_PORTAL;
945 ptlrpc_at_set_req_timeout(req);
947 if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
948 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
949 OP_XVALID_LAZYBLOCKS);
951 mdc_close_pack(req, op_data);
953 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
954 obd->u.cli.cl_default_mds_easize);
956 ptlrpc_request_set_replen(req);
958 mdc_get_mod_rpc_slot(req, NULL);
959 rc = ptlrpc_queue_wait(req);
960 mdc_put_mod_rpc_slot(req, NULL);
962 if (req->rq_repmsg == NULL) {
963 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
966 rc = req->rq_status ?: -EIO;
967 } else if (rc == 0 || rc == -EAGAIN) {
968 struct mdt_body *body;
970 rc = lustre_msg_get_status(req->rq_repmsg);
971 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
972 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
977 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
980 } else if (rc == -ESTALE) {
982 * it can be allowed error after 3633 if open was committed and
983 * server failed before close was sent. Let's check if mod
984 * exists and return no error in that case
987 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
988 LASSERT(mod->mod_open_req != NULL);
989 if (mod->mod_open_req->rq_committed)
997 mod->mod_close_req = NULL;
998 /* Since now, mod is accessed through open_req only,
999 * thus close req does not keep a reference on mod anymore. */
1004 RETURN(rc < 0 ? rc : saved_rc);
1007 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1008 u64 offset, struct page **pages, int npages,
1009 struct ptlrpc_request **request)
1011 struct ptlrpc_request *req;
1012 struct ptlrpc_bulk_desc *desc;
1014 wait_queue_head_t waitq;
1016 struct l_wait_info lwi;
1021 init_waitqueue_head(&waitq);
1024 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1028 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1030 ptlrpc_request_free(req);
1034 req->rq_request_portal = MDS_READPAGE_PORTAL;
1035 ptlrpc_at_set_req_timeout(req);
1037 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1038 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1040 &ptlrpc_bulk_kiov_pin_ops);
1042 ptlrpc_req_finished(req);
1046 /* NB req now owns desc and will free it when it gets freed */
1047 for (i = 0; i < npages; i++)
1048 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1051 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1053 ptlrpc_request_set_replen(req);
1054 rc = ptlrpc_queue_wait(req);
1056 ptlrpc_req_finished(req);
1057 if (rc != -ETIMEDOUT)
1061 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1062 CERROR("%s: too many resend retries: rc = %d\n",
1063 exp->exp_obd->obd_name, -EIO);
1066 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1068 l_wait_event(waitq, 0, &lwi);
1073 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1074 req->rq_bulk->bd_nob_transferred);
1076 ptlrpc_req_finished(req);
1080 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1081 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1082 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1083 PAGE_SIZE * npages);
1084 ptlrpc_req_finished(req);
1092 static void mdc_release_page(struct page *page, int remove)
1096 if (likely(page->mapping != NULL))
1097 truncate_complete_page(page->mapping, page);
1103 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1104 __u64 *start, __u64 *end, int hash64)
1107 * Complement of hash is used as an index so that
1108 * radix_tree_gang_lookup() can be used to find a page with starting
1109 * hash _smaller_ than one we are looking for.
1111 unsigned long offset = hash_x_index(*hash, hash64);
1115 xa_lock_irq(&mapping->i_pages);
1116 found = radix_tree_gang_lookup(&mapping->page_tree,
1117 (void **)&page, offset, 1);
1118 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1119 struct lu_dirpage *dp;
1122 xa_unlock_irq(&mapping->i_pages);
1124 * In contrast to find_lock_page() we are sure that directory
1125 * page cannot be truncated (while DLM lock is held) and,
1126 * hence, can avoid restart.
1128 * In fact, page cannot be locked here at all, because
1129 * mdc_read_page_remote does synchronous io.
1131 wait_on_page_locked(page);
1132 if (PageUptodate(page)) {
1134 if (BITS_PER_LONG == 32 && hash64) {
1135 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1136 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1137 *hash = *hash >> 32;
1139 *start = le64_to_cpu(dp->ldp_hash_start);
1140 *end = le64_to_cpu(dp->ldp_hash_end);
1142 if (unlikely(*start == 1 && *hash == 0))
1145 LASSERTF(*start <= *hash, "start = %#llx"
1146 ",end = %#llx,hash = %#llx\n",
1147 *start, *end, *hash);
1148 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1149 " hash %#llx\n", offset, *start, *end, *hash);
1152 mdc_release_page(page, 0);
1154 } else if (*end != *start && *hash == *end) {
1156 * upon hash collision, remove this page,
1157 * otherwise put page reference, and
1158 * mdc_read_page_remote() will issue RPC to
1159 * fetch the page we want.
1162 mdc_release_page(page,
1163 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1168 page = ERR_PTR(-EIO);
1171 xa_unlock_irq(&mapping->i_pages);
1178 * Adjust a set of pages, each page containing an array of lu_dirpages,
1179 * so that each page can be used as a single logical lu_dirpage.
1181 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1182 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1183 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1184 * value is used as a cookie to request the next lu_dirpage in a
1185 * directory listing that spans multiple pages (two in this example):
1188 * .|--------v------- -----.
1189 * |s|e|f|p|ent|ent| ... |ent|
1190 * '--|-------------- -----' Each PAGE contains a single
1191 * '------. lu_dirpage.
1192 * .---------v------- -----.
1193 * |s|e|f|p|ent| 0 | ... | 0 |
1194 * '----------------- -----'
1196 * However, on hosts where the native VM page size (PAGE_SIZE) is
1197 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1198 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1199 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1200 * after it in the same PAGE (arrows simplified for brevity, but
1201 * in general e0==s1, e1==s2, etc.):
1203 * .-------------------- -----.
1204 * |s0|e0|f0|p|ent|ent| ... |ent|
1205 * |---v---------------- -----|
1206 * |s1|e1|f1|p|ent|ent| ... |ent|
1207 * |---v---------------- -----| Here, each PAGE contains
1208 * ... multiple lu_dirpages.
1209 * |---v---------------- -----|
1210 * |s'|e'|f'|p|ent|ent| ... |ent|
1211 * '---|---------------- -----'
1213 * .----------------------------.
1216 * This structure is transformed into a single logical lu_dirpage as follows:
1218 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1219 * labeled 'next PAGE'.
1221 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1222 * a hash collision with the next page exists.
1224 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1225 * to the first entry of the next lu_dirpage.
1227 #if PAGE_SIZE > LU_PAGE_SIZE
1228 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1232 for (i = 0; i < cfs_pgs; i++) {
1233 struct lu_dirpage *dp = kmap(pages[i]);
1234 struct lu_dirpage *first = dp;
1235 struct lu_dirent *end_dirent = NULL;
1236 struct lu_dirent *ent;
1237 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1238 __u32 flags = le32_to_cpu(dp->ldp_flags);
1240 while (--lu_pgs > 0) {
1241 ent = lu_dirent_start(dp);
1242 for (end_dirent = ent; ent != NULL;
1243 end_dirent = ent, ent = lu_dirent_next(ent));
1245 /* Advance dp to next lu_dirpage. */
1246 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1248 /* Check if we've reached the end of the PAGE. */
1249 if (!((unsigned long)dp & ~PAGE_MASK))
1252 /* Save the hash and flags of this lu_dirpage. */
1253 hash_end = le64_to_cpu(dp->ldp_hash_end);
1254 flags = le32_to_cpu(dp->ldp_flags);
1256 /* Check if lu_dirpage contains no entries. */
1257 if (end_dirent == NULL)
1260 /* Enlarge the end entry lde_reclen from 0 to
1261 * first entry of next lu_dirpage. */
1262 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1263 end_dirent->lde_reclen =
1264 cpu_to_le16((char *)(dp->ldp_entries) -
1265 (char *)end_dirent);
1268 first->ldp_hash_end = hash_end;
1269 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1270 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1274 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1277 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1278 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1280 /* parameters for readdir page */
1281 struct readpage_param {
1282 struct md_op_data *rp_mod;
1285 struct obd_export *rp_exp;
1286 struct md_callback *rp_cb;
1289 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1290 static inline void delete_from_page_cache(struct page *page)
1292 remove_from_page_cache(page);
1298 * Read pages from server.
1300 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1301 * a header lu_dirpage which describes the start/end hash, and whether this
1302 * page is empty (contains no dir entry) or hash collide with next page.
1303 * After client receives reply, several pages will be integrated into dir page
1304 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1305 * lu_dirpage for this integrated page will be adjusted.
1307 static int mdc_read_page_remote(void *data, struct page *page0)
1309 struct readpage_param *rp = data;
1310 struct page **page_pool;
1312 struct lu_dirpage *dp;
1313 struct md_op_data *op_data = rp->rp_mod;
1314 struct ptlrpc_request *req;
1316 struct inode *inode;
1318 int rd_pgs = 0; /* number of pages actually read */
1324 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1325 inode = op_data->op_data;
1326 fid = &op_data->op_fid1;
1327 LASSERT(inode != NULL);
1329 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1330 if (page_pool != NULL) {
1331 page_pool[0] = page0;
1337 for (npages = 1; npages < max_pages; npages++) {
1338 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1342 page_pool[npages] = page;
1345 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1347 /* page0 is special, which was added into page cache early */
1348 delete_from_page_cache(page0);
1352 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1354 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1355 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1357 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1359 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1361 SetPageUptodate(page0);
1365 ptlrpc_req_finished(req);
1366 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1367 for (i = 1; i < npages; i++) {
1368 unsigned long offset;
1372 page = page_pool[i];
1374 if (rc < 0 || i >= rd_pgs) {
1379 SetPageUptodate(page);
1382 hash = le64_to_cpu(dp->ldp_hash_start);
1385 offset = hash_x_index(hash, rp->rp_hash64);
1387 prefetchw(&page->flags);
1388 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1393 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1394 " rc = %d\n", offset, ret);
1398 if (page_pool != &page0)
1399 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1405 * Read dir page from cache first, if it can not find it, read it from
1406 * server and add into the cache.
1408 * \param[in] exp MDC export
1409 * \param[in] op_data client MD stack parameters, transfering parameters
1410 * between different layers on client MD stack.
1411 * \param[in] cb_op callback required for ldlm lock enqueue during
1413 * \param[in] hash_offset the hash offset of the page to be read
1414 * \param[in] ppage the page to be read
1416 * retval = 0 get the page successfully
1417 * errno(<0) get the page failed
1419 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1420 struct md_callback *cb_op, __u64 hash_offset,
1421 struct page **ppage)
1423 struct lookup_intent it = { .it_op = IT_READDIR };
1425 struct inode *dir = op_data->op_data;
1426 struct address_space *mapping;
1427 struct lu_dirpage *dp;
1430 struct lustre_handle lockh;
1431 struct ptlrpc_request *enq_req = NULL;
1432 struct readpage_param rp_param;
1439 LASSERT(dir != NULL);
1440 mapping = dir->i_mapping;
1442 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1443 cb_op->md_blocking_ast, 0);
1444 if (enq_req != NULL)
1445 ptlrpc_req_finished(enq_req);
1448 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1449 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1454 lockh.cookie = it.it_lock_handle;
1455 mdc_set_lock_data(exp, &lockh, dir, NULL);
1457 rp_param.rp_off = hash_offset;
1458 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1459 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1460 rp_param.rp_hash64);
1462 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1463 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1464 rp_param.rp_off, PTR_ERR(page));
1465 GOTO(out_unlock, rc = PTR_ERR(page));
1466 } else if (page != NULL) {
1468 * XXX nikita: not entirely correct handling of a corner case:
1469 * suppose hash chain of entries with hash value HASH crosses
1470 * border between pages P0 and P1. First both P0 and P1 are
1471 * cached, seekdir() is called for some entry from the P0 part
1472 * of the chain. Later P0 goes out of cache. telldir(HASH)
1473 * happens and finds P1, as it starts with matching hash
1474 * value. Remaining entries from P0 part of the chain are
1475 * skipped. (Is that really a bug?)
1477 * Possible solutions: 0. don't cache P1 is such case, handle
1478 * it as an "overflow" page. 1. invalidate all pages at
1479 * once. 2. use HASH|1 as an index for P1.
1481 GOTO(hash_collision, page);
1484 rp_param.rp_exp = exp;
1485 rp_param.rp_mod = op_data;
1486 page = read_cache_page(mapping,
1487 hash_x_index(rp_param.rp_off,
1488 rp_param.rp_hash64),
1489 mdc_read_page_remote, &rp_param);
1491 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1492 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1493 rp_param.rp_off, PTR_ERR(page));
1494 GOTO(out_unlock, rc = PTR_ERR(page));
1497 wait_on_page_locked(page);
1499 if (!PageUptodate(page)) {
1500 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1501 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1502 rp_param.rp_off, -5);
1505 if (!PageChecked(page))
1506 SetPageChecked(page);
1507 if (PageError(page)) {
1508 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1509 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1510 rp_param.rp_off, -5);
1515 dp = page_address(page);
1516 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1517 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1518 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1519 rp_param.rp_off = hash_offset >> 32;
1521 start = le64_to_cpu(dp->ldp_hash_start);
1522 end = le64_to_cpu(dp->ldp_hash_end);
1523 rp_param.rp_off = hash_offset;
1526 LASSERT(start == rp_param.rp_off);
1527 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1528 #if BITS_PER_LONG == 32
1529 CWARN("Real page-wide hash collision at [%llu %llu] with "
1530 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1531 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1535 * Fetch whole overflow chain...
1543 ldlm_lock_decref(&lockh, it.it_lock_mode);
1547 mdc_release_page(page, 1);
1552 static int mdc_statfs(const struct lu_env *env,
1553 struct obd_export *exp, struct obd_statfs *osfs,
1554 time64_t max_age, __u32 flags)
1556 struct obd_device *obd = class_exp2obd(exp);
1557 struct req_format *fmt;
1558 struct ptlrpc_request *req;
1559 struct obd_statfs *msfs;
1560 struct obd_import *imp = NULL;
1565 * Since the request might also come from lprocfs, so we need
1566 * sync this with client_disconnect_export Bug15684
1568 down_read(&obd->u.cli.cl_sem);
1569 if (obd->u.cli.cl_import)
1570 imp = class_import_get(obd->u.cli.cl_import);
1571 up_read(&obd->u.cli.cl_sem);
1575 fmt = &RQF_MDS_STATFS;
1576 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1577 (flags & OBD_STATFS_SUM))
1578 fmt = &RQF_MDS_STATFS_NEW;
1579 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1582 GOTO(output, rc = -ENOMEM);
1584 if ((flags & OBD_STATFS_SUM) &&
1585 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1586 /* request aggregated states */
1587 struct mdt_body *body;
1589 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1591 GOTO(out, rc = -EPROTO);
1592 body->mbo_valid = OBD_MD_FLAGSTATFS;
1595 ptlrpc_request_set_replen(req);
1597 if (flags & OBD_STATFS_NODELAY) {
1598 /* procfs requests not want stay in wait for avoid deadlock */
1599 req->rq_no_resend = 1;
1600 req->rq_no_delay = 1;
1603 rc = ptlrpc_queue_wait(req);
1605 /* check connection error first */
1606 if (imp->imp_connect_error)
1607 rc = imp->imp_connect_error;
1611 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1613 GOTO(out, rc = -EPROTO);
1618 ptlrpc_req_finished(req);
1620 class_import_put(imp);
1624 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1626 __u32 keylen, vallen;
1630 if (gf->gf_pathlen > PATH_MAX)
1631 RETURN(-ENAMETOOLONG);
1632 if (gf->gf_pathlen < 2)
1635 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1636 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1637 sizeof(struct lu_fid));
1638 OBD_ALLOC(key, keylen);
1641 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1642 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1643 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1644 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1645 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1646 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1648 if (!fid_is_sane(&gf->gf_fid))
1649 GOTO(out, rc = -EINVAL);
1651 /* Val is struct getinfo_fid2path result plus path */
1652 vallen = sizeof(*gf) + gf->gf_pathlen;
1654 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1655 if (rc != 0 && rc != -EREMOTE)
1658 if (vallen <= sizeof(*gf))
1659 GOTO(out, rc = -EPROTO);
1660 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1661 GOTO(out, rc = -EOVERFLOW);
1663 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1664 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1665 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1666 /* only log the last 512 characters of the path */
1667 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1670 OBD_FREE(key, keylen);
1674 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1675 struct hsm_progress_kernel *hpk)
1677 struct obd_import *imp = class_exp2cliimp(exp);
1678 struct hsm_progress_kernel *req_hpk;
1679 struct ptlrpc_request *req;
1683 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1684 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1686 GOTO(out, rc = -ENOMEM);
1688 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1690 /* Copy hsm_progress struct */
1691 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1692 if (req_hpk == NULL)
1693 GOTO(out, rc = -EPROTO);
1696 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1698 ptlrpc_request_set_replen(req);
1700 mdc_get_mod_rpc_slot(req, NULL);
1701 rc = ptlrpc_queue_wait(req);
1702 mdc_put_mod_rpc_slot(req, NULL);
1706 ptlrpc_req_finished(req);
1710 * Send hsm_ct_register to MDS
1712 * \param[in] imp import
1713 * \param[in] archive_count if in bitmap format, it is the bitmap,
1714 * else it is the count of archive_ids
1715 * \param[in] archives if in bitmap format, it is NULL,
1716 * else it is archive_id lists
1718 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1721 struct ptlrpc_request *req;
1722 __u32 *archive_array;
1723 size_t archives_size;
1727 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1731 if (archives != NULL)
1732 archives_size = sizeof(*archive_array) * archive_count;
1734 archives_size = sizeof(archive_count);
1736 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1737 RCL_CLIENT, archives_size);
1739 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1741 ptlrpc_request_free(req);
1745 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1747 archive_array = req_capsule_client_get(&req->rq_pill,
1748 &RMF_MDS_HSM_ARCHIVE);
1749 if (archive_array == NULL)
1750 GOTO(out, rc = -EPROTO);
1752 if (archives != NULL)
1753 memcpy(archive_array, archives, archives_size);
1755 *archive_array = archive_count;
1757 ptlrpc_request_set_replen(req);
1759 rc = mdc_queue_wait(req);
1762 ptlrpc_req_finished(req);
1766 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1767 struct md_op_data *op_data)
1769 struct hsm_current_action *hca = op_data->op_data;
1770 struct hsm_current_action *req_hca;
1771 struct ptlrpc_request *req;
1775 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1776 &RQF_MDS_HSM_ACTION);
1780 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1782 ptlrpc_request_free(req);
1786 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1787 op_data->op_suppgids[0], 0);
1789 ptlrpc_request_set_replen(req);
1791 rc = mdc_queue_wait(req);
1795 req_hca = req_capsule_server_get(&req->rq_pill,
1796 &RMF_MDS_HSM_CURRENT_ACTION);
1797 if (req_hca == NULL)
1798 GOTO(out, rc = -EPROTO);
1804 ptlrpc_req_finished(req);
1808 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1810 struct ptlrpc_request *req;
1814 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1816 MDS_HSM_CT_UNREGISTER);
1818 GOTO(out, rc = -ENOMEM);
1820 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1822 ptlrpc_request_set_replen(req);
1824 rc = mdc_queue_wait(req);
1827 ptlrpc_req_finished(req);
1831 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1832 struct md_op_data *op_data)
1834 struct hsm_user_state *hus = op_data->op_data;
1835 struct hsm_user_state *req_hus;
1836 struct ptlrpc_request *req;
1840 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1841 &RQF_MDS_HSM_STATE_GET);
1845 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1847 ptlrpc_request_free(req);
1851 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1852 op_data->op_suppgids[0], 0);
1854 ptlrpc_request_set_replen(req);
1856 rc = mdc_queue_wait(req);
1860 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1861 if (req_hus == NULL)
1862 GOTO(out, rc = -EPROTO);
1868 ptlrpc_req_finished(req);
1872 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1873 struct md_op_data *op_data)
1875 struct hsm_state_set *hss = op_data->op_data;
1876 struct hsm_state_set *req_hss;
1877 struct ptlrpc_request *req;
1881 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1882 &RQF_MDS_HSM_STATE_SET);
1886 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1888 ptlrpc_request_free(req);
1892 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1893 op_data->op_suppgids[0], 0);
1896 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1897 if (req_hss == NULL)
1898 GOTO(out, rc = -EPROTO);
1901 ptlrpc_request_set_replen(req);
1903 mdc_get_mod_rpc_slot(req, NULL);
1904 rc = ptlrpc_queue_wait(req);
1905 mdc_put_mod_rpc_slot(req, NULL);
1909 ptlrpc_req_finished(req);
1913 static int mdc_ioc_hsm_request(struct obd_export *exp,
1914 struct hsm_user_request *hur)
1916 struct obd_import *imp = class_exp2cliimp(exp);
1917 struct ptlrpc_request *req;
1918 struct hsm_request *req_hr;
1919 struct hsm_user_item *req_hui;
1924 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1926 GOTO(out, rc = -ENOMEM);
1928 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1929 hur->hur_request.hr_itemcount
1930 * sizeof(struct hsm_user_item));
1931 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1932 hur->hur_request.hr_data_len);
1934 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1936 ptlrpc_request_free(req);
1940 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1942 /* Copy hsm_request struct */
1943 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1945 GOTO(out, rc = -EPROTO);
1946 *req_hr = hur->hur_request;
1948 /* Copy hsm_user_item structs */
1949 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1950 if (req_hui == NULL)
1951 GOTO(out, rc = -EPROTO);
1952 memcpy(req_hui, hur->hur_user_item,
1953 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1955 /* Copy opaque field */
1956 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1957 if (req_opaque == NULL)
1958 GOTO(out, rc = -EPROTO);
1959 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1961 ptlrpc_request_set_replen(req);
1963 mdc_get_mod_rpc_slot(req, NULL);
1964 rc = ptlrpc_queue_wait(req);
1965 mdc_put_mod_rpc_slot(req, NULL);
1970 ptlrpc_req_finished(req);
1974 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1975 struct lustre_kernelcomm *lk);
1977 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1978 struct obd_quotactl *oqctl)
1980 struct ptlrpc_request *req;
1981 struct obd_quotactl *oqc;
1985 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1986 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1991 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1994 ptlrpc_request_set_replen(req);
1995 ptlrpc_at_set_req_timeout(req);
1997 rc = ptlrpc_queue_wait(req);
1999 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
2001 if (req->rq_repmsg &&
2002 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2005 CERROR ("Can't unpack obd_quotactl\n");
2008 ptlrpc_req_finished(req);
2013 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2014 struct md_op_data *op_data)
2016 struct list_head cancels = LIST_HEAD_INIT(cancels);
2017 struct ptlrpc_request *req;
2019 struct mdc_swap_layouts *msl, *payload;
2022 msl = op_data->op_data;
2024 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2025 * first thing it will do is to cancel the 2 layout
2026 * locks held by this client.
2027 * So the client must cancel its layout locks on the 2 fids
2028 * with the request RPC to avoid extra RPC round trips.
2030 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2031 LCK_EX, MDS_INODELOCK_LAYOUT |
2032 MDS_INODELOCK_XATTR);
2033 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2034 LCK_EX, MDS_INODELOCK_LAYOUT |
2035 MDS_INODELOCK_XATTR);
2037 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2038 &RQF_MDS_SWAP_LAYOUTS);
2040 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2044 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2046 ptlrpc_request_free(req);
2050 mdc_swap_layouts_pack(req, op_data);
2052 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2057 ptlrpc_request_set_replen(req);
2059 rc = ptlrpc_queue_wait(req);
2065 ptlrpc_req_finished(req);
2069 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2070 void *karg, void __user *uarg)
2072 struct obd_device *obd = exp->exp_obd;
2073 struct obd_ioctl_data *data = karg;
2074 struct obd_import *imp = obd->u.cli.cl_import;
2078 if (!try_module_get(THIS_MODULE)) {
2079 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2080 module_name(THIS_MODULE));
2084 case OBD_IOC_FID2PATH:
2085 rc = mdc_ioc_fid2path(exp, karg);
2087 case LL_IOC_HSM_CT_START:
2088 rc = mdc_ioc_hsm_ct_start(exp, karg);
2089 /* ignore if it was already registered on this MDS. */
2093 case LL_IOC_HSM_PROGRESS:
2094 rc = mdc_ioc_hsm_progress(exp, karg);
2096 case LL_IOC_HSM_STATE_GET:
2097 rc = mdc_ioc_hsm_state_get(exp, karg);
2099 case LL_IOC_HSM_STATE_SET:
2100 rc = mdc_ioc_hsm_state_set(exp, karg);
2102 case LL_IOC_HSM_ACTION:
2103 rc = mdc_ioc_hsm_current_action(exp, karg);
2105 case LL_IOC_HSM_REQUEST:
2106 rc = mdc_ioc_hsm_request(exp, karg);
2108 case OBD_IOC_CLIENT_RECOVER:
2109 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2113 case IOC_OSC_SET_ACTIVE:
2114 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2117 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2118 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2119 * there'd be no LMV layer thus we might be called here. Eventually
2120 * this code should be removed.
2123 case IOC_OBD_STATFS: {
2124 struct obd_statfs stat_buf = {0};
2126 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2127 GOTO(out, rc = -ENODEV);
2130 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2131 min((int)data->ioc_plen2,
2132 (int)sizeof(struct obd_uuid))))
2133 GOTO(out, rc = -EFAULT);
2135 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2136 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2141 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2142 min((int) data->ioc_plen1,
2143 (int) sizeof(stat_buf))))
2144 GOTO(out, rc = -EFAULT);
2148 case OBD_IOC_QUOTACTL: {
2149 struct if_quotactl *qctl = karg;
2150 struct obd_quotactl *oqctl;
2152 OBD_ALLOC_PTR(oqctl);
2154 GOTO(out, rc = -ENOMEM);
2156 QCTL_COPY(oqctl, qctl);
2157 rc = obd_quotactl(exp, oqctl);
2159 QCTL_COPY(qctl, oqctl);
2160 qctl->qc_valid = QC_MDTIDX;
2161 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2164 OBD_FREE_PTR(oqctl);
2167 case LL_IOC_GET_CONNECT_FLAGS:
2168 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2169 sizeof(*exp_connect_flags_ptr(exp))))
2170 GOTO(out, rc = -EFAULT);
2173 case LL_IOC_LOV_SWAP_LAYOUTS:
2174 rc = mdc_ioc_swap_layouts(exp, karg);
2177 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2178 GOTO(out, rc = -ENOTTY);
2181 module_put(THIS_MODULE);
2186 static int mdc_get_info_rpc(struct obd_export *exp,
2187 u32 keylen, void *key,
2188 u32 vallen, void *val)
2190 struct obd_import *imp = class_exp2cliimp(exp);
2191 struct ptlrpc_request *req;
2196 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2200 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2201 RCL_CLIENT, keylen);
2202 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2203 RCL_CLIENT, sizeof(vallen));
2205 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2207 ptlrpc_request_free(req);
2211 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2212 memcpy(tmp, key, keylen);
2213 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2214 memcpy(tmp, &vallen, sizeof(vallen));
2216 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2217 RCL_SERVER, vallen);
2218 ptlrpc_request_set_replen(req);
2220 rc = ptlrpc_queue_wait(req);
2221 /* -EREMOTE means the get_info result is partial, and it needs to
2222 * continue on another MDT, see fid2path part in lmv_iocontrol */
2223 if (rc == 0 || rc == -EREMOTE) {
2224 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2225 memcpy(val, tmp, vallen);
2226 if (ptlrpc_rep_need_swab(req)) {
2227 if (KEY_IS(KEY_FID2PATH))
2228 lustre_swab_fid2path(val);
2231 ptlrpc_req_finished(req);
2236 static void lustre_swab_hai(struct hsm_action_item *h)
2238 __swab32s(&h->hai_len);
2239 __swab32s(&h->hai_action);
2240 lustre_swab_lu_fid(&h->hai_fid);
2241 lustre_swab_lu_fid(&h->hai_dfid);
2242 __swab64s(&h->hai_cookie);
2243 __swab64s(&h->hai_extent.offset);
2244 __swab64s(&h->hai_extent.length);
2245 __swab64s(&h->hai_gid);
2248 static void lustre_swab_hal(struct hsm_action_list *h)
2250 struct hsm_action_item *hai;
2253 __swab32s(&h->hal_version);
2254 __swab32s(&h->hal_count);
2255 __swab32s(&h->hal_archive_id);
2256 __swab64s(&h->hal_flags);
2258 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2259 lustre_swab_hai(hai);
2262 static void lustre_swab_kuch(struct kuc_hdr *l)
2264 __swab16s(&l->kuc_magic);
2265 /* __u8 l->kuc_transport */
2266 __swab16s(&l->kuc_msgtype);
2267 __swab16s(&l->kuc_msglen);
2270 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2271 struct lustre_kernelcomm *lk)
2273 struct obd_import *imp = class_exp2cliimp(exp);
2276 if (lk->lk_group != KUC_GRP_HSM) {
2277 CERROR("Bad copytool group %d\n", lk->lk_group);
2281 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2282 lk->lk_uid, lk->lk_group, lk->lk_flags);
2284 if (lk->lk_flags & LK_FLG_STOP) {
2285 /* Unregister with the coordinator */
2286 rc = mdc_ioc_hsm_ct_unregister(imp);
2288 __u32 *archives = NULL;
2290 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2291 archives = lk->lk_data;
2293 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2300 * Send a message to any listening copytools
2301 * @param val KUC message (kuc_hdr + hsm_action_list)
2302 * @param len total length of message
2304 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2305 size_t len, void *val)
2307 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2308 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2312 if (len < sizeof(*lh) + sizeof(*hal)) {
2313 CERROR("Short HSM message %zu < %zu\n", len,
2314 sizeof(*lh) + sizeof(*hal));
2317 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2318 lustre_swab_kuch(lh);
2319 lustre_swab_hal(hal);
2320 } else if (lh->kuc_magic != KUC_MAGIC) {
2321 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2325 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2327 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2328 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2330 /* Broadcast to HSM listeners */
2331 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2337 * callback function passed to kuc for re-registering each HSM copytool
2338 * running on MDC, after MDT shutdown/recovery.
2339 * @param data copytool registration data
2340 * @param cb_arg callback argument (obd_import)
2342 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2344 struct obd_import *imp = (struct obd_import *)cb_arg;
2345 struct kkuc_ct_data *kcd = data;
2346 __u32 *archives = NULL;
2350 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2351 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2354 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2355 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2356 "(archive=%#x)\n", imp->imp_obd->obd_name,
2357 kcd->kcd_nr_archives);
2359 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2360 "(archive nr = %u)\n",
2361 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2362 if (kcd->kcd_nr_archives != 0)
2363 archives = kcd->kcd_archives;
2366 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2367 /* ignore error if the copytool is already registered */
2368 return (rc == -EEXIST) ? 0 : rc;
2372 * Re-establish all kuc contexts with MDT
2373 * after MDT shutdown/recovery.
2375 static int mdc_kuc_reregister(struct obd_import *imp)
2377 /* re-register HSM agents */
2378 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2379 mdc_hsm_ct_reregister, imp);
2382 static int mdc_set_info_async(const struct lu_env *env,
2383 struct obd_export *exp,
2384 u32 keylen, void *key,
2385 u32 vallen, void *val,
2386 struct ptlrpc_request_set *set)
2388 struct obd_import *imp = class_exp2cliimp(exp);
2392 if (KEY_IS(KEY_READ_ONLY)) {
2393 if (vallen != sizeof(int))
2396 spin_lock(&imp->imp_lock);
2397 if (*((int *)val)) {
2398 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2399 imp->imp_connect_data.ocd_connect_flags |=
2402 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2403 imp->imp_connect_data.ocd_connect_flags &=
2404 ~OBD_CONNECT_RDONLY;
2406 spin_unlock(&imp->imp_lock);
2408 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2409 keylen, key, vallen, val, set);
2412 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2413 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2414 keylen, key, vallen, val, set);
2417 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2418 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2423 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2424 __u32 *default_easize = val;
2426 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2430 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2434 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2435 __u32 keylen, void *key, __u32 *vallen, void *val)
2439 if (KEY_IS(KEY_MAX_EASIZE)) {
2440 __u32 mdsize, *max_easize;
2442 if (*vallen != sizeof(int))
2444 mdsize = *(__u32 *)val;
2445 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2446 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2448 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2450 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2451 __u32 *default_easize;
2453 if (*vallen != sizeof(int))
2455 default_easize = val;
2456 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2458 } else if (KEY_IS(KEY_CONN_DATA)) {
2459 struct obd_import *imp = class_exp2cliimp(exp);
2460 struct obd_connect_data *data = val;
2462 if (*vallen != sizeof(*data))
2465 *data = imp->imp_connect_data;
2467 } else if (KEY_IS(KEY_TGT_COUNT)) {
2468 *((__u32 *)val) = 1;
2472 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2477 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2478 struct ptlrpc_request **request)
2480 struct ptlrpc_request *req;
2485 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2489 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2491 ptlrpc_request_free(req);
2495 mdc_pack_body(req, fid, 0, 0, -1, 0);
2497 ptlrpc_request_set_replen(req);
2499 rc = ptlrpc_queue_wait(req);
2501 ptlrpc_req_finished(req);
2507 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2508 enum obd_import_event event)
2510 struct client_obd *cli = &obd->u.cli;
2513 LASSERT(imp->imp_obd == obd);
2516 case IMP_EVENT_DISCON:
2517 spin_lock(&cli->cl_loi_list_lock);
2518 cli->cl_avail_grant = 0;
2519 cli->cl_lost_grant = 0;
2520 spin_unlock(&cli->cl_loi_list_lock);
2522 case IMP_EVENT_INACTIVE:
2524 * Flush current sequence to make client obtain new one
2525 * from server in case of disconnect/reconnect.
2527 down_read(&cli->cl_seq_rwsem);
2529 seq_client_flush(cli->cl_seq);
2530 up_read(&cli->cl_seq_rwsem);
2532 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2534 case IMP_EVENT_INVALIDATE: {
2535 struct ldlm_namespace *ns = obd->obd_namespace;
2539 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2541 env = cl_env_get(&refcheck);
2543 /* Reset grants. All pages go to failing rpcs due to
2544 * the invalid import.
2546 osc_io_unplug(env, cli, NULL);
2548 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2549 osc_ldlm_resource_invalidate,
2551 cl_env_put(env, &refcheck);
2552 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2558 case IMP_EVENT_ACTIVE:
2559 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2560 /* redo the kuc registration after reconnecting */
2562 rc = mdc_kuc_reregister(imp);
2564 case IMP_EVENT_OCD: {
2565 struct obd_connect_data *ocd = &imp->imp_connect_data;
2567 if (OCD_HAS_FLAG(ocd, GRANT))
2568 osc_init_grant(cli, ocd);
2570 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2573 case IMP_EVENT_DEACTIVATE:
2574 case IMP_EVENT_ACTIVATE:
2577 CERROR("Unknown import event %x\n", event);
2583 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2584 struct lu_fid *fid, struct md_op_data *op_data)
2586 struct client_obd *cli = &exp->exp_obd->u.cli;
2591 down_read(&cli->cl_seq_rwsem);
2593 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2594 up_read(&cli->cl_seq_rwsem);
2599 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2601 struct client_obd *cli = &exp->exp_obd->u.cli;
2602 return &cli->cl_target_uuid;
2606 * Determine whether the lock can be canceled before replaying it during
2607 * recovery, non zero value will be return if the lock can be canceled,
2608 * or zero returned for not
2610 static int mdc_cancel_weight(struct ldlm_lock *lock)
2612 if (lock->l_resource->lr_type != LDLM_IBITS)
2615 /* FIXME: if we ever get into a situation where there are too many
2616 * opened files with open locks on a single node, then we really
2617 * should replay these open locks to reget it */
2618 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2621 /* Special case for DoM locks, cancel only unused and granted locks */
2622 if (ldlm_has_dom(lock) &&
2623 (lock->l_granted_mode != lock->l_req_mode ||
2624 osc_ldlm_weigh_ast(lock) != 0))
2630 static int mdc_resource_inode_free(struct ldlm_resource *res)
2632 if (res->lr_lvb_inode)
2633 res->lr_lvb_inode = NULL;
2638 static struct ldlm_valblock_ops inode_lvbo = {
2639 .lvbo_free = mdc_resource_inode_free
2642 static int mdc_llog_init(struct obd_device *obd)
2644 struct obd_llog_group *olg = &obd->obd_olg;
2645 struct llog_ctxt *ctxt;
2650 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2655 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2656 llog_initiator_connect(ctxt);
2657 llog_ctxt_put(ctxt);
2662 static void mdc_llog_finish(struct obd_device *obd)
2664 struct llog_ctxt *ctxt;
2668 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2670 llog_cleanup(NULL, ctxt);
2675 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2681 rc = osc_setup_common(obd, cfg);
2685 rc = mdc_tunables_init(obd);
2687 GOTO(err_osc_cleanup, rc);
2689 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2691 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2693 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2695 rc = mdc_llog_init(obd);
2697 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2699 GOTO(err_llog_cleanup, rc);
2702 rc = mdc_changelog_cdev_init(obd);
2704 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2706 GOTO(err_changelog_cleanup, rc);
2711 err_changelog_cleanup:
2712 mdc_llog_finish(obd);
2714 lprocfs_free_md_stats(obd);
2715 ptlrpc_lprocfs_unregister_obd(obd);
2717 osc_cleanup_common(obd);
2721 /* Initialize the default and maximum LOV EA sizes. This allows
2722 * us to make MDS RPCs with large enough reply buffers to hold a default
2723 * sized EA without having to calculate this (via a call into the
2724 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2725 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2726 * a large number of stripes is possible. If a larger reply buffer is
2727 * required it will be reallocated in the ptlrpc layer due to overflow.
2729 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2732 struct obd_device *obd = exp->exp_obd;
2733 struct client_obd *cli = &obd->u.cli;
2736 if (cli->cl_max_mds_easize < easize)
2737 cli->cl_max_mds_easize = easize;
2739 if (cli->cl_default_mds_easize < def_easize)
2740 cli->cl_default_mds_easize = def_easize;
2745 static int mdc_precleanup(struct obd_device *obd)
2749 osc_precleanup_common(obd);
2750 mdc_changelog_cdev_finish(obd);
2752 obd_cleanup_client_import(obd);
2753 ptlrpc_lprocfs_unregister_obd(obd);
2754 lprocfs_free_md_stats(obd);
2755 mdc_llog_finish(obd);
2759 static int mdc_cleanup(struct obd_device *obd)
2761 return osc_cleanup_common(obd);
2764 static struct obd_ops mdc_obd_ops = {
2765 .o_owner = THIS_MODULE,
2766 .o_setup = mdc_setup,
2767 .o_precleanup = mdc_precleanup,
2768 .o_cleanup = mdc_cleanup,
2769 .o_add_conn = client_import_add_conn,
2770 .o_del_conn = client_import_del_conn,
2771 .o_connect = client_connect_import,
2772 .o_reconnect = osc_reconnect,
2773 .o_disconnect = osc_disconnect,
2774 .o_iocontrol = mdc_iocontrol,
2775 .o_set_info_async = mdc_set_info_async,
2776 .o_statfs = mdc_statfs,
2777 .o_fid_init = client_fid_init,
2778 .o_fid_fini = client_fid_fini,
2779 .o_fid_alloc = mdc_fid_alloc,
2780 .o_import_event = mdc_import_event,
2781 .o_get_info = mdc_get_info,
2782 .o_get_uuid = mdc_get_uuid,
2783 .o_quotactl = mdc_quotactl,
2786 static struct md_ops mdc_md_ops = {
2787 .m_get_root = mdc_get_root,
2788 .m_null_inode = mdc_null_inode,
2789 .m_close = mdc_close,
2790 .m_create = mdc_create,
2791 .m_enqueue = mdc_enqueue,
2792 .m_getattr = mdc_getattr,
2793 .m_getattr_name = mdc_getattr_name,
2794 .m_intent_lock = mdc_intent_lock,
2796 .m_rename = mdc_rename,
2797 .m_setattr = mdc_setattr,
2798 .m_setxattr = mdc_setxattr,
2799 .m_getxattr = mdc_getxattr,
2800 .m_fsync = mdc_fsync,
2801 .m_file_resync = mdc_file_resync,
2802 .m_read_page = mdc_read_page,
2803 .m_unlink = mdc_unlink,
2804 .m_cancel_unused = mdc_cancel_unused,
2805 .m_init_ea_size = mdc_init_ea_size,
2806 .m_set_lock_data = mdc_set_lock_data,
2807 .m_lock_match = mdc_lock_match,
2808 .m_get_lustre_md = mdc_get_lustre_md,
2809 .m_free_lustre_md = mdc_free_lustre_md,
2810 .m_set_open_replay_data = mdc_set_open_replay_data,
2811 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2812 .m_intent_getattr_async = mdc_intent_getattr_async,
2813 .m_revalidate_lock = mdc_revalidate_lock
2816 static int __init mdc_init(void)
2818 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2819 LUSTRE_MDC_NAME, &mdc_device_type);
2822 static void __exit mdc_exit(void)
2824 class_unregister_type(LUSTRE_MDC_NAME);
2827 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2828 MODULE_DESCRIPTION("Lustre Metadata Client");
2829 MODULE_VERSION(LUSTRE_VERSION_STRING);
2830 MODULE_LICENSE("GPL");
2832 module_init(mdc_init);
2833 module_exit(mdc_exit);