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,
339 /* Flush local XATTR locks to get rid of a possible cancel RPC */
340 if (opcode == MDS_REINT && fid_is_sane(fid) &&
341 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
342 struct list_head cancels = LIST_HEAD_INIT(cancels);
345 /* Without that packing would fail */
347 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
350 count = mdc_resource_get_unused(exp, fid,
352 MDS_INODELOCK_XATTR);
354 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
356 ptlrpc_request_free(req);
360 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
362 ptlrpc_request_free(req);
367 if (opcode == MDS_REINT) {
368 struct mdt_rec_setxattr *rec;
370 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
371 sizeof(struct mdt_rec_reint));
372 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
373 rec->sx_opcode = REINT_SETXATTR;
374 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
375 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
376 rec->sx_cap = cfs_curproc_cap_pack();
377 rec->sx_suppgid1 = suppgid;
378 rec->sx_suppgid2 = -1;
380 rec->sx_valid = valid | OBD_MD_FLCTIME;
381 rec->sx_time = ktime_get_real_seconds();
382 rec->sx_size = output_size;
383 rec->sx_flags = flags;
385 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
389 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
390 memcpy(tmp, xattr_name, xattr_namelen);
393 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
394 memcpy(tmp, input, input_size);
397 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
398 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
399 RCL_SERVER, output_size);
400 ptlrpc_request_set_replen(req);
403 if (opcode == MDS_REINT)
404 mdc_get_mod_rpc_slot(req, NULL);
406 rc = ptlrpc_queue_wait(req);
408 if (opcode == MDS_REINT)
409 mdc_put_mod_rpc_slot(req, NULL);
412 ptlrpc_req_finished(req);
418 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
419 u64 obd_md_valid, const char *name,
420 const void *value, size_t value_size,
421 unsigned int xattr_flags, u32 suppgid,
422 struct ptlrpc_request **req)
424 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
425 obd_md_valid == OBD_MD_FLXATTRRM);
427 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
428 fid, MDS_REINT, obd_md_valid, name,
429 value, value_size, 0, xattr_flags, suppgid,
433 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
434 u64 obd_md_valid, const char *name, size_t buf_size,
435 struct ptlrpc_request **req)
437 struct mdt_body *body;
440 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
441 obd_md_valid == OBD_MD_FLXATTRLS);
443 CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
444 exp->exp_obd->obd_name, name, PFID(fid));
445 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
446 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
451 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
453 GOTO(out, rc = -EPROTO);
455 /* only detect the xattr size */
457 /* LU-11109: Older MDTs do not distinguish
458 * between nonexistent xattrs and zero length
459 * values in this case. Newer MDTs will return
460 * -ENODATA or set OBD_MD_FLXATTR. */
461 GOTO(out, rc = body->mbo_eadatasize);
464 if (body->mbo_eadatasize == 0) {
465 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
466 * success so that we can distinguish between
467 * zero length value and nonexistent xattr.
469 * If OBD_MD_FLXATTR is not set then we keep
470 * the old behavior and return -ENODATA for
471 * getxattr() when mbo_eadatasize is 0. But
472 * -ENODATA only makes sense for getxattr()
473 * and not for listxattr(). */
474 if (body->mbo_valid & OBD_MD_FLXATTR)
476 else if (obd_md_valid == OBD_MD_FLXATTR)
477 GOTO(out, rc = -ENODATA);
482 GOTO(out, rc = body->mbo_eadatasize);
485 ptlrpc_req_finished(*req);
492 #ifdef CONFIG_FS_POSIX_ACL
493 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
495 struct req_capsule *pill = &req->rq_pill;
496 struct mdt_body *body = md->body;
497 struct posix_acl *acl;
502 if (!body->mbo_aclsize)
505 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
510 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
515 CERROR("convert xattr to acl: %d\n", rc);
519 rc = posix_acl_valid(&init_user_ns, acl);
521 CERROR("validate acl: %d\n", rc);
522 posix_acl_release(acl);
530 #define mdc_unpack_acl(req, md) 0
533 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
534 struct obd_export *dt_exp, struct obd_export *md_exp,
535 struct lustre_md *md)
537 struct req_capsule *pill = &req->rq_pill;
542 memset(md, 0, sizeof(*md));
544 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
545 LASSERT(md->body != NULL);
547 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
548 if (!S_ISREG(md->body->mbo_mode)) {
549 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
550 "regular file, but is not\n");
551 GOTO(out, rc = -EPROTO);
554 if (md->body->mbo_eadatasize == 0) {
555 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
556 "but eadatasize 0\n");
557 GOTO(out, rc = -EPROTO);
560 md->layout.lb_len = md->body->mbo_eadatasize;
561 md->layout.lb_buf = req_capsule_server_sized_get(pill,
564 if (md->layout.lb_buf == NULL)
565 GOTO(out, rc = -EPROTO);
566 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
567 const union lmv_mds_md *lmv;
570 if (!S_ISDIR(md->body->mbo_mode)) {
571 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
572 "directory, but is not\n");
573 GOTO(out, rc = -EPROTO);
576 lmv_size = md->body->mbo_eadatasize;
578 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
579 "but eadatasize 0\n");
583 if (md->body->mbo_valid & OBD_MD_MEA) {
584 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
587 GOTO(out, rc = -EPROTO);
589 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
593 if (rc < (typeof(rc))sizeof(*md->lmv)) {
594 CDEBUG(D_INFO, "size too small: "
595 "rc < sizeof(*md->lmv) (%d < %d)\n",
596 rc, (int)sizeof(*md->lmv));
597 GOTO(out, rc = -EPROTO);
603 if (md->body->mbo_valid & OBD_MD_FLACL) {
604 /* for ACL, it's possible that FLACL is set but aclsize is zero.
605 * only when aclsize != 0 there's an actual segment for ACL
608 if (md->body->mbo_aclsize) {
609 rc = mdc_unpack_acl(req, md);
612 #ifdef CONFIG_FS_POSIX_ACL
614 md->posix_acl = NULL;
622 #ifdef CONFIG_FS_POSIX_ACL
623 posix_acl_release(md->posix_acl);
629 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
635 void mdc_replay_open(struct ptlrpc_request *req)
637 struct md_open_data *mod = req->rq_cb_data;
638 struct ptlrpc_request *close_req;
639 struct obd_client_handle *och;
640 struct lustre_handle old_open_handle = { };
641 struct mdt_body *body;
645 DEBUG_REQ(D_ERROR, req,
646 "Can't properly replay without open data.");
651 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
652 LASSERT(body != NULL);
654 spin_lock(&req->rq_lock);
656 if (och && och->och_open_handle.cookie)
657 req->rq_early_free_repbuf = 1;
659 req->rq_early_free_repbuf = 0;
660 spin_unlock(&req->rq_lock);
662 if (req->rq_early_free_repbuf) {
663 struct lustre_handle *file_open_handle;
665 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
667 file_open_handle = &och->och_open_handle;
668 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
669 file_open_handle->cookie, body->mbo_open_handle.cookie);
670 old_open_handle = *file_open_handle;
671 *file_open_handle = body->mbo_open_handle;
674 close_req = mod->mod_close_req;
676 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
677 struct mdt_ioepoch *epoch;
679 LASSERT(opc == MDS_CLOSE);
680 epoch = req_capsule_client_get(&close_req->rq_pill,
684 if (req->rq_early_free_repbuf)
685 LASSERT(old_open_handle.cookie ==
686 epoch->mio_open_handle.cookie);
688 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
689 epoch->mio_open_handle = body->mbo_open_handle;
694 void mdc_commit_open(struct ptlrpc_request *req)
696 struct md_open_data *mod = req->rq_cb_data;
701 * No need to touch md_open_data::mod_och, it holds a reference on
702 * \var mod and will zero references to each other, \var mod will be
703 * freed after that when md_open_data::mod_och will put the reference.
707 * Do not let open request to disappear as it still may be needed
708 * for close rpc to happen (it may happen on evict only, otherwise
709 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
710 * called), just mark this rpc as committed to distinguish these 2
711 * cases, see mdc_close() for details. The open request reference will
712 * be put along with freeing \var mod.
714 ptlrpc_request_addref(req);
715 spin_lock(&req->rq_lock);
716 req->rq_committed = 1;
717 spin_unlock(&req->rq_lock);
718 req->rq_cb_data = NULL;
722 int mdc_set_open_replay_data(struct obd_export *exp,
723 struct obd_client_handle *och,
724 struct lookup_intent *it)
726 struct md_open_data *mod;
727 struct mdt_rec_create *rec;
728 struct mdt_body *body;
729 struct ptlrpc_request *open_req = it->it_request;
730 struct obd_import *imp = open_req->rq_import;
733 if (!open_req->rq_replay)
736 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
737 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
738 LASSERT(rec != NULL);
739 /* Incoming message in my byte order (it's been swabbed). */
740 /* Outgoing messages always in my byte order. */
741 LASSERT(body != NULL);
743 /* Only if the import is replayable, we set replay_open data */
744 if (och && imp->imp_replayable) {
745 mod = obd_mod_alloc();
747 DEBUG_REQ(D_ERROR, open_req,
748 "Can't allocate md_open_data");
753 * Take a reference on \var mod, to be freed on mdc_close().
754 * It protects \var mod from being freed on eviction (commit
755 * callback is called despite rq_replay flag).
756 * Another reference for \var och.
761 spin_lock(&open_req->rq_lock);
764 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
765 it_disposition(it, DISP_OPEN_STRIPE);
766 mod->mod_open_req = open_req;
767 open_req->rq_cb_data = mod;
768 open_req->rq_commit_cb = mdc_commit_open;
769 open_req->rq_early_free_repbuf = 1;
770 spin_unlock(&open_req->rq_lock);
773 rec->cr_fid2 = body->mbo_fid1;
774 rec->cr_open_handle_old = body->mbo_open_handle;
775 open_req->rq_replay_cb = mdc_replay_open;
776 if (!fid_is_sane(&body->mbo_fid1)) {
777 DEBUG_REQ(D_ERROR, open_req,
778 "saving replay request with insane FID " DFID,
779 PFID(&body->mbo_fid1));
783 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
787 static void mdc_free_open(struct md_open_data *mod)
791 if (mod->mod_is_create == 0 &&
792 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
796 * No reason to asssert here if the open request has
797 * rq_replay == 1. It means that mdc_close failed, and
798 * close request wasn`t sent. It is not fatal to client.
799 * The worst thing is eviction if the client gets open lock
802 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
803 "= %d\n", mod->mod_open_req->rq_replay);
805 ptlrpc_request_committed(mod->mod_open_req, committed);
806 if (mod->mod_close_req)
807 ptlrpc_request_committed(mod->mod_close_req, committed);
810 int mdc_clear_open_replay_data(struct obd_export *exp,
811 struct obd_client_handle *och)
813 struct md_open_data *mod = och->och_mod;
817 * It is possible to not have \var mod in a case of eviction between
818 * lookup and ll_file_open().
823 LASSERT(mod != LP_POISON);
824 LASSERT(mod->mod_open_req != NULL);
826 spin_lock(&mod->mod_open_req->rq_lock);
828 mod->mod_och->och_open_handle.cookie = 0;
829 mod->mod_open_req->rq_early_free_repbuf = 0;
830 spin_unlock(&mod->mod_open_req->rq_lock);
840 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
841 struct md_open_data *mod, struct ptlrpc_request **request)
843 struct obd_device *obd = class_exp2obd(exp);
844 struct ptlrpc_request *req;
845 struct req_format *req_fmt;
846 size_t u32_count = 0;
851 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
852 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
855 if (op_data->op_bias & MDS_CLOSE_INTENT) {
856 req_fmt = &RQF_MDS_CLOSE_INTENT;
857 if (op_data->op_bias & MDS_HSM_RELEASE) {
858 /* allocate a FID for volatile file */
859 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
862 CERROR("%s: "DFID" allocating FID: rc = %d\n",
863 obd->obd_name, PFID(&op_data->op_fid1),
865 /* save the errcode and proceed to close */
869 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
870 size_t count = op_data->op_data_size / sizeof(__u32);
872 if (count > INLINE_RESYNC_ARRAY_SIZE)
876 req_fmt = &RQF_MDS_CLOSE;
880 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
883 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
885 /* Ensure that this close's handle is fixed up during replay. */
886 if (likely(mod != NULL)) {
887 LASSERTF(mod->mod_open_req != NULL &&
888 mod->mod_open_req->rq_type != LI_POISON,
889 "POISONED open %p!\n", mod->mod_open_req);
891 mod->mod_close_req = req;
893 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
894 /* We no longer want to preserve this open for replay even
895 * though the open was committed. b=3632, b=3633 */
896 spin_lock(&mod->mod_open_req->rq_lock);
897 mod->mod_open_req->rq_replay = 0;
898 spin_unlock(&mod->mod_open_req->rq_lock);
900 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
904 * TODO: repeat close after errors
906 CWARN("%s: close of FID "DFID" failed, file reference will be "
907 "dropped when this client unmounts or is evicted\n",
908 obd->obd_name, PFID(&op_data->op_fid1));
909 GOTO(out, rc = -ENOMEM);
913 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
914 u32_count * sizeof(__u32));
916 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
918 ptlrpc_request_free(req);
923 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
924 * portal whose threads are not taking any DLM locks and are therefore
925 * always progressing */
926 req->rq_request_portal = MDS_READPAGE_PORTAL;
927 ptlrpc_at_set_req_timeout(req);
930 mdc_close_pack(req, op_data);
932 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
933 obd->u.cli.cl_default_mds_easize);
935 ptlrpc_request_set_replen(req);
937 mdc_get_mod_rpc_slot(req, NULL);
938 rc = ptlrpc_queue_wait(req);
939 mdc_put_mod_rpc_slot(req, NULL);
941 if (req->rq_repmsg == NULL) {
942 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
945 rc = req->rq_status ?: -EIO;
946 } else if (rc == 0 || rc == -EAGAIN) {
947 struct mdt_body *body;
949 rc = lustre_msg_get_status(req->rq_repmsg);
950 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
951 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
956 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
959 } else if (rc == -ESTALE) {
961 * it can be allowed error after 3633 if open was committed and
962 * server failed before close was sent. Let's check if mod
963 * exists and return no error in that case
966 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
967 LASSERT(mod->mod_open_req != NULL);
968 if (mod->mod_open_req->rq_committed)
976 mod->mod_close_req = NULL;
977 /* Since now, mod is accessed through open_req only,
978 * thus close req does not keep a reference on mod anymore. */
983 RETURN(rc < 0 ? rc : saved_rc);
986 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
987 u64 offset, struct page **pages, int npages,
988 struct ptlrpc_request **request)
990 struct ptlrpc_request *req;
991 struct ptlrpc_bulk_desc *desc;
993 wait_queue_head_t waitq;
995 struct l_wait_info lwi;
1000 init_waitqueue_head(&waitq);
1003 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1007 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1009 ptlrpc_request_free(req);
1013 req->rq_request_portal = MDS_READPAGE_PORTAL;
1014 ptlrpc_at_set_req_timeout(req);
1016 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1017 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1019 &ptlrpc_bulk_kiov_pin_ops);
1021 ptlrpc_req_finished(req);
1025 /* NB req now owns desc and will free it when it gets freed */
1026 for (i = 0; i < npages; i++)
1027 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1030 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1032 ptlrpc_request_set_replen(req);
1033 rc = ptlrpc_queue_wait(req);
1035 ptlrpc_req_finished(req);
1036 if (rc != -ETIMEDOUT)
1040 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1041 CERROR("%s: too many resend retries: rc = %d\n",
1042 exp->exp_obd->obd_name, -EIO);
1045 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1047 l_wait_event(waitq, 0, &lwi);
1052 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1053 req->rq_bulk->bd_nob_transferred);
1055 ptlrpc_req_finished(req);
1059 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1060 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1061 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1062 PAGE_SIZE * npages);
1063 ptlrpc_req_finished(req);
1071 static void mdc_release_page(struct page *page, int remove)
1075 if (likely(page->mapping != NULL))
1076 truncate_complete_page(page->mapping, page);
1082 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1083 __u64 *start, __u64 *end, int hash64)
1086 * Complement of hash is used as an index so that
1087 * radix_tree_gang_lookup() can be used to find a page with starting
1088 * hash _smaller_ than one we are looking for.
1090 unsigned long offset = hash_x_index(*hash, hash64);
1094 spin_lock_irq(&mapping->tree_lock);
1095 found = radix_tree_gang_lookup(&mapping->page_tree,
1096 (void **)&page, offset, 1);
1097 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1098 struct lu_dirpage *dp;
1101 spin_unlock_irq(&mapping->tree_lock);
1103 * In contrast to find_lock_page() we are sure that directory
1104 * page cannot be truncated (while DLM lock is held) and,
1105 * hence, can avoid restart.
1107 * In fact, page cannot be locked here at all, because
1108 * mdc_read_page_remote does synchronous io.
1110 wait_on_page_locked(page);
1111 if (PageUptodate(page)) {
1113 if (BITS_PER_LONG == 32 && hash64) {
1114 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1115 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1116 *hash = *hash >> 32;
1118 *start = le64_to_cpu(dp->ldp_hash_start);
1119 *end = le64_to_cpu(dp->ldp_hash_end);
1121 if (unlikely(*start == 1 && *hash == 0))
1124 LASSERTF(*start <= *hash, "start = %#llx"
1125 ",end = %#llx,hash = %#llx\n",
1126 *start, *end, *hash);
1127 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1128 " hash %#llx\n", offset, *start, *end, *hash);
1131 mdc_release_page(page, 0);
1133 } else if (*end != *start && *hash == *end) {
1135 * upon hash collision, remove this page,
1136 * otherwise put page reference, and
1137 * mdc_read_page_remote() will issue RPC to
1138 * fetch the page we want.
1141 mdc_release_page(page,
1142 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1147 page = ERR_PTR(-EIO);
1150 spin_unlock_irq(&mapping->tree_lock);
1157 * Adjust a set of pages, each page containing an array of lu_dirpages,
1158 * so that each page can be used as a single logical lu_dirpage.
1160 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1161 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1162 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1163 * value is used as a cookie to request the next lu_dirpage in a
1164 * directory listing that spans multiple pages (two in this example):
1167 * .|--------v------- -----.
1168 * |s|e|f|p|ent|ent| ... |ent|
1169 * '--|-------------- -----' Each PAGE contains a single
1170 * '------. lu_dirpage.
1171 * .---------v------- -----.
1172 * |s|e|f|p|ent| 0 | ... | 0 |
1173 * '----------------- -----'
1175 * However, on hosts where the native VM page size (PAGE_SIZE) is
1176 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1177 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1178 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1179 * after it in the same PAGE (arrows simplified for brevity, but
1180 * in general e0==s1, e1==s2, etc.):
1182 * .-------------------- -----.
1183 * |s0|e0|f0|p|ent|ent| ... |ent|
1184 * |---v---------------- -----|
1185 * |s1|e1|f1|p|ent|ent| ... |ent|
1186 * |---v---------------- -----| Here, each PAGE contains
1187 * ... multiple lu_dirpages.
1188 * |---v---------------- -----|
1189 * |s'|e'|f'|p|ent|ent| ... |ent|
1190 * '---|---------------- -----'
1192 * .----------------------------.
1195 * This structure is transformed into a single logical lu_dirpage as follows:
1197 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1198 * labeled 'next PAGE'.
1200 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1201 * a hash collision with the next page exists.
1203 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1204 * to the first entry of the next lu_dirpage.
1206 #if PAGE_SIZE > LU_PAGE_SIZE
1207 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1211 for (i = 0; i < cfs_pgs; i++) {
1212 struct lu_dirpage *dp = kmap(pages[i]);
1213 struct lu_dirpage *first = dp;
1214 struct lu_dirent *end_dirent = NULL;
1215 struct lu_dirent *ent;
1216 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1217 __u32 flags = le32_to_cpu(dp->ldp_flags);
1219 while (--lu_pgs > 0) {
1220 ent = lu_dirent_start(dp);
1221 for (end_dirent = ent; ent != NULL;
1222 end_dirent = ent, ent = lu_dirent_next(ent));
1224 /* Advance dp to next lu_dirpage. */
1225 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1227 /* Check if we've reached the end of the PAGE. */
1228 if (!((unsigned long)dp & ~PAGE_MASK))
1231 /* Save the hash and flags of this lu_dirpage. */
1232 hash_end = le64_to_cpu(dp->ldp_hash_end);
1233 flags = le32_to_cpu(dp->ldp_flags);
1235 /* Check if lu_dirpage contains no entries. */
1236 if (end_dirent == NULL)
1239 /* Enlarge the end entry lde_reclen from 0 to
1240 * first entry of next lu_dirpage. */
1241 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1242 end_dirent->lde_reclen =
1243 cpu_to_le16((char *)(dp->ldp_entries) -
1244 (char *)end_dirent);
1247 first->ldp_hash_end = hash_end;
1248 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1249 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1253 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1256 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1257 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1259 /* parameters for readdir page */
1260 struct readpage_param {
1261 struct md_op_data *rp_mod;
1264 struct obd_export *rp_exp;
1265 struct md_callback *rp_cb;
1268 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1269 static inline void delete_from_page_cache(struct page *page)
1271 remove_from_page_cache(page);
1277 * Read pages from server.
1279 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1280 * a header lu_dirpage which describes the start/end hash, and whether this
1281 * page is empty (contains no dir entry) or hash collide with next page.
1282 * After client receives reply, several pages will be integrated into dir page
1283 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1284 * lu_dirpage for this integrated page will be adjusted.
1286 static int mdc_read_page_remote(void *data, struct page *page0)
1288 struct readpage_param *rp = data;
1289 struct page **page_pool;
1291 struct lu_dirpage *dp;
1292 struct md_op_data *op_data = rp->rp_mod;
1293 struct ptlrpc_request *req;
1295 struct inode *inode;
1297 int rd_pgs = 0; /* number of pages actually read */
1303 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1304 inode = op_data->op_data;
1305 fid = &op_data->op_fid1;
1306 LASSERT(inode != NULL);
1308 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1309 if (page_pool != NULL) {
1310 page_pool[0] = page0;
1316 for (npages = 1; npages < max_pages; npages++) {
1317 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1321 page_pool[npages] = page;
1324 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1326 /* page0 is special, which was added into page cache early */
1327 delete_from_page_cache(page0);
1331 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1333 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1334 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1336 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1338 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1340 SetPageUptodate(page0);
1344 ptlrpc_req_finished(req);
1345 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1346 for (i = 1; i < npages; i++) {
1347 unsigned long offset;
1351 page = page_pool[i];
1353 if (rc < 0 || i >= rd_pgs) {
1358 SetPageUptodate(page);
1361 hash = le64_to_cpu(dp->ldp_hash_start);
1364 offset = hash_x_index(hash, rp->rp_hash64);
1366 prefetchw(&page->flags);
1367 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1372 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1373 " rc = %d\n", offset, ret);
1377 if (page_pool != &page0)
1378 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1384 * Read dir page from cache first, if it can not find it, read it from
1385 * server and add into the cache.
1387 * \param[in] exp MDC export
1388 * \param[in] op_data client MD stack parameters, transfering parameters
1389 * between different layers on client MD stack.
1390 * \param[in] cb_op callback required for ldlm lock enqueue during
1392 * \param[in] hash_offset the hash offset of the page to be read
1393 * \param[in] ppage the page to be read
1395 * retval = 0 get the page successfully
1396 * errno(<0) get the page failed
1398 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1399 struct md_callback *cb_op, __u64 hash_offset,
1400 struct page **ppage)
1402 struct lookup_intent it = { .it_op = IT_READDIR };
1404 struct inode *dir = op_data->op_data;
1405 struct address_space *mapping;
1406 struct lu_dirpage *dp;
1409 struct lustre_handle lockh;
1410 struct ptlrpc_request *enq_req = NULL;
1411 struct readpage_param rp_param;
1418 LASSERT(dir != NULL);
1419 mapping = dir->i_mapping;
1421 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1422 cb_op->md_blocking_ast, 0);
1423 if (enq_req != NULL)
1424 ptlrpc_req_finished(enq_req);
1427 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1428 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1433 lockh.cookie = it.it_lock_handle;
1434 mdc_set_lock_data(exp, &lockh, dir, NULL);
1436 rp_param.rp_off = hash_offset;
1437 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1438 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1439 rp_param.rp_hash64);
1441 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1442 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1443 rp_param.rp_off, PTR_ERR(page));
1444 GOTO(out_unlock, rc = PTR_ERR(page));
1445 } else if (page != NULL) {
1447 * XXX nikita: not entirely correct handling of a corner case:
1448 * suppose hash chain of entries with hash value HASH crosses
1449 * border between pages P0 and P1. First both P0 and P1 are
1450 * cached, seekdir() is called for some entry from the P0 part
1451 * of the chain. Later P0 goes out of cache. telldir(HASH)
1452 * happens and finds P1, as it starts with matching hash
1453 * value. Remaining entries from P0 part of the chain are
1454 * skipped. (Is that really a bug?)
1456 * Possible solutions: 0. don't cache P1 is such case, handle
1457 * it as an "overflow" page. 1. invalidate all pages at
1458 * once. 2. use HASH|1 as an index for P1.
1460 GOTO(hash_collision, page);
1463 rp_param.rp_exp = exp;
1464 rp_param.rp_mod = op_data;
1465 page = read_cache_page(mapping,
1466 hash_x_index(rp_param.rp_off,
1467 rp_param.rp_hash64),
1468 mdc_read_page_remote, &rp_param);
1470 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1471 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1472 rp_param.rp_off, PTR_ERR(page));
1473 GOTO(out_unlock, rc = PTR_ERR(page));
1476 wait_on_page_locked(page);
1478 if (!PageUptodate(page)) {
1479 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1480 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1481 rp_param.rp_off, -5);
1484 if (!PageChecked(page))
1485 SetPageChecked(page);
1486 if (PageError(page)) {
1487 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1488 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1489 rp_param.rp_off, -5);
1494 dp = page_address(page);
1495 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1496 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1497 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1498 rp_param.rp_off = hash_offset >> 32;
1500 start = le64_to_cpu(dp->ldp_hash_start);
1501 end = le64_to_cpu(dp->ldp_hash_end);
1502 rp_param.rp_off = hash_offset;
1505 LASSERT(start == rp_param.rp_off);
1506 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1507 #if BITS_PER_LONG == 32
1508 CWARN("Real page-wide hash collision at [%llu %llu] with "
1509 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1510 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1514 * Fetch whole overflow chain...
1522 ldlm_lock_decref(&lockh, it.it_lock_mode);
1526 mdc_release_page(page, 1);
1531 static int mdc_statfs(const struct lu_env *env,
1532 struct obd_export *exp, struct obd_statfs *osfs,
1533 time64_t max_age, __u32 flags)
1535 struct obd_device *obd = class_exp2obd(exp);
1536 struct req_format *fmt;
1537 struct ptlrpc_request *req;
1538 struct obd_statfs *msfs;
1539 struct obd_import *imp = NULL;
1544 * Since the request might also come from lprocfs, so we need
1545 * sync this with client_disconnect_export Bug15684
1547 down_read(&obd->u.cli.cl_sem);
1548 if (obd->u.cli.cl_import)
1549 imp = class_import_get(obd->u.cli.cl_import);
1550 up_read(&obd->u.cli.cl_sem);
1554 fmt = &RQF_MDS_STATFS;
1555 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1556 (flags & OBD_STATFS_SUM))
1557 fmt = &RQF_MDS_STATFS_NEW;
1558 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1561 GOTO(output, rc = -ENOMEM);
1563 if ((flags & OBD_STATFS_SUM) &&
1564 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1565 /* request aggregated states */
1566 struct mdt_body *body;
1568 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1570 GOTO(out, rc = -EPROTO);
1571 body->mbo_valid = OBD_MD_FLAGSTATFS;
1574 ptlrpc_request_set_replen(req);
1576 if (flags & OBD_STATFS_NODELAY) {
1577 /* procfs requests not want stay in wait for avoid deadlock */
1578 req->rq_no_resend = 1;
1579 req->rq_no_delay = 1;
1582 rc = ptlrpc_queue_wait(req);
1584 /* check connection error first */
1585 if (imp->imp_connect_error)
1586 rc = imp->imp_connect_error;
1590 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1592 GOTO(out, rc = -EPROTO);
1597 ptlrpc_req_finished(req);
1599 class_import_put(imp);
1603 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1605 __u32 keylen, vallen;
1609 if (gf->gf_pathlen > PATH_MAX)
1610 RETURN(-ENAMETOOLONG);
1611 if (gf->gf_pathlen < 2)
1614 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1615 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1616 sizeof(struct lu_fid));
1617 OBD_ALLOC(key, keylen);
1620 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1621 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1622 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1623 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1624 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1625 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1627 if (!fid_is_sane(&gf->gf_fid))
1628 GOTO(out, rc = -EINVAL);
1630 /* Val is struct getinfo_fid2path result plus path */
1631 vallen = sizeof(*gf) + gf->gf_pathlen;
1633 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1634 if (rc != 0 && rc != -EREMOTE)
1637 if (vallen <= sizeof(*gf))
1638 GOTO(out, rc = -EPROTO);
1639 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1640 GOTO(out, rc = -EOVERFLOW);
1642 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1643 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1644 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1645 /* only log the last 512 characters of the path */
1646 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1649 OBD_FREE(key, keylen);
1653 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1654 struct hsm_progress_kernel *hpk)
1656 struct obd_import *imp = class_exp2cliimp(exp);
1657 struct hsm_progress_kernel *req_hpk;
1658 struct ptlrpc_request *req;
1662 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1663 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1665 GOTO(out, rc = -ENOMEM);
1667 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1669 /* Copy hsm_progress struct */
1670 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1671 if (req_hpk == NULL)
1672 GOTO(out, rc = -EPROTO);
1675 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1677 ptlrpc_request_set_replen(req);
1679 mdc_get_mod_rpc_slot(req, NULL);
1680 rc = ptlrpc_queue_wait(req);
1681 mdc_put_mod_rpc_slot(req, NULL);
1685 ptlrpc_req_finished(req);
1689 * Send hsm_ct_register to MDS
1691 * \param[in] imp import
1692 * \param[in] archive_count if in bitmap format, it is the bitmap,
1693 * else it is the count of archive_ids
1694 * \param[in] archives if in bitmap format, it is NULL,
1695 * else it is archive_id lists
1697 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1700 struct ptlrpc_request *req;
1701 __u32 *archive_array;
1702 size_t archives_size;
1706 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1710 if (archives != NULL)
1711 archives_size = sizeof(*archive_array) * archive_count;
1713 archives_size = sizeof(archive_count);
1715 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1716 RCL_CLIENT, archives_size);
1718 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1720 ptlrpc_request_free(req);
1724 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1726 archive_array = req_capsule_client_get(&req->rq_pill,
1727 &RMF_MDS_HSM_ARCHIVE);
1728 if (archive_array == NULL)
1729 GOTO(out, rc = -EPROTO);
1731 if (archives != NULL)
1732 memcpy(archive_array, archives, archives_size);
1734 *archive_array = archive_count;
1736 ptlrpc_request_set_replen(req);
1738 rc = mdc_queue_wait(req);
1741 ptlrpc_req_finished(req);
1745 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1746 struct md_op_data *op_data)
1748 struct hsm_current_action *hca = op_data->op_data;
1749 struct hsm_current_action *req_hca;
1750 struct ptlrpc_request *req;
1754 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1755 &RQF_MDS_HSM_ACTION);
1759 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1761 ptlrpc_request_free(req);
1765 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1766 op_data->op_suppgids[0], 0);
1768 ptlrpc_request_set_replen(req);
1770 rc = mdc_queue_wait(req);
1774 req_hca = req_capsule_server_get(&req->rq_pill,
1775 &RMF_MDS_HSM_CURRENT_ACTION);
1776 if (req_hca == NULL)
1777 GOTO(out, rc = -EPROTO);
1783 ptlrpc_req_finished(req);
1787 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1789 struct ptlrpc_request *req;
1793 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1795 MDS_HSM_CT_UNREGISTER);
1797 GOTO(out, rc = -ENOMEM);
1799 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1801 ptlrpc_request_set_replen(req);
1803 rc = mdc_queue_wait(req);
1806 ptlrpc_req_finished(req);
1810 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1811 struct md_op_data *op_data)
1813 struct hsm_user_state *hus = op_data->op_data;
1814 struct hsm_user_state *req_hus;
1815 struct ptlrpc_request *req;
1819 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1820 &RQF_MDS_HSM_STATE_GET);
1824 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1826 ptlrpc_request_free(req);
1830 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1831 op_data->op_suppgids[0], 0);
1833 ptlrpc_request_set_replen(req);
1835 rc = mdc_queue_wait(req);
1839 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1840 if (req_hus == NULL)
1841 GOTO(out, rc = -EPROTO);
1847 ptlrpc_req_finished(req);
1851 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1852 struct md_op_data *op_data)
1854 struct hsm_state_set *hss = op_data->op_data;
1855 struct hsm_state_set *req_hss;
1856 struct ptlrpc_request *req;
1860 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1861 &RQF_MDS_HSM_STATE_SET);
1865 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1867 ptlrpc_request_free(req);
1871 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1872 op_data->op_suppgids[0], 0);
1875 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1876 if (req_hss == NULL)
1877 GOTO(out, rc = -EPROTO);
1880 ptlrpc_request_set_replen(req);
1882 mdc_get_mod_rpc_slot(req, NULL);
1883 rc = ptlrpc_queue_wait(req);
1884 mdc_put_mod_rpc_slot(req, NULL);
1888 ptlrpc_req_finished(req);
1892 static int mdc_ioc_hsm_request(struct obd_export *exp,
1893 struct hsm_user_request *hur)
1895 struct obd_import *imp = class_exp2cliimp(exp);
1896 struct ptlrpc_request *req;
1897 struct hsm_request *req_hr;
1898 struct hsm_user_item *req_hui;
1903 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1905 GOTO(out, rc = -ENOMEM);
1907 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1908 hur->hur_request.hr_itemcount
1909 * sizeof(struct hsm_user_item));
1910 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1911 hur->hur_request.hr_data_len);
1913 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1915 ptlrpc_request_free(req);
1919 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1921 /* Copy hsm_request struct */
1922 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1924 GOTO(out, rc = -EPROTO);
1925 *req_hr = hur->hur_request;
1927 /* Copy hsm_user_item structs */
1928 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1929 if (req_hui == NULL)
1930 GOTO(out, rc = -EPROTO);
1931 memcpy(req_hui, hur->hur_user_item,
1932 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1934 /* Copy opaque field */
1935 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1936 if (req_opaque == NULL)
1937 GOTO(out, rc = -EPROTO);
1938 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1940 ptlrpc_request_set_replen(req);
1942 mdc_get_mod_rpc_slot(req, NULL);
1943 rc = ptlrpc_queue_wait(req);
1944 mdc_put_mod_rpc_slot(req, NULL);
1949 ptlrpc_req_finished(req);
1953 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1954 struct lustre_kernelcomm *lk);
1956 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1957 struct obd_quotactl *oqctl)
1959 struct ptlrpc_request *req;
1960 struct obd_quotactl *oqc;
1964 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1965 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1970 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1973 ptlrpc_request_set_replen(req);
1974 ptlrpc_at_set_req_timeout(req);
1976 rc = ptlrpc_queue_wait(req);
1978 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1980 if (req->rq_repmsg &&
1981 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1984 CERROR ("Can't unpack obd_quotactl\n");
1987 ptlrpc_req_finished(req);
1992 static int mdc_ioc_swap_layouts(struct obd_export *exp,
1993 struct md_op_data *op_data)
1995 struct list_head cancels = LIST_HEAD_INIT(cancels);
1996 struct ptlrpc_request *req;
1998 struct mdc_swap_layouts *msl, *payload;
2001 msl = op_data->op_data;
2003 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2004 * first thing it will do is to cancel the 2 layout
2005 * locks held by this client.
2006 * So the client must cancel its layout locks on the 2 fids
2007 * with the request RPC to avoid extra RPC round trips.
2009 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2010 LCK_EX, MDS_INODELOCK_LAYOUT |
2011 MDS_INODELOCK_XATTR);
2012 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2013 LCK_EX, MDS_INODELOCK_LAYOUT |
2014 MDS_INODELOCK_XATTR);
2016 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2017 &RQF_MDS_SWAP_LAYOUTS);
2019 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2023 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2025 ptlrpc_request_free(req);
2029 mdc_swap_layouts_pack(req, op_data);
2031 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2036 ptlrpc_request_set_replen(req);
2038 rc = ptlrpc_queue_wait(req);
2044 ptlrpc_req_finished(req);
2048 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2049 void *karg, void __user *uarg)
2051 struct obd_device *obd = exp->exp_obd;
2052 struct obd_ioctl_data *data = karg;
2053 struct obd_import *imp = obd->u.cli.cl_import;
2057 if (!try_module_get(THIS_MODULE)) {
2058 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2059 module_name(THIS_MODULE));
2063 case OBD_IOC_FID2PATH:
2064 rc = mdc_ioc_fid2path(exp, karg);
2066 case LL_IOC_HSM_CT_START:
2067 rc = mdc_ioc_hsm_ct_start(exp, karg);
2068 /* ignore if it was already registered on this MDS. */
2072 case LL_IOC_HSM_PROGRESS:
2073 rc = mdc_ioc_hsm_progress(exp, karg);
2075 case LL_IOC_HSM_STATE_GET:
2076 rc = mdc_ioc_hsm_state_get(exp, karg);
2078 case LL_IOC_HSM_STATE_SET:
2079 rc = mdc_ioc_hsm_state_set(exp, karg);
2081 case LL_IOC_HSM_ACTION:
2082 rc = mdc_ioc_hsm_current_action(exp, karg);
2084 case LL_IOC_HSM_REQUEST:
2085 rc = mdc_ioc_hsm_request(exp, karg);
2087 case OBD_IOC_CLIENT_RECOVER:
2088 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2092 case IOC_OSC_SET_ACTIVE:
2093 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2095 case OBD_IOC_PING_TARGET:
2096 rc = ptlrpc_obd_ping(obd);
2099 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2100 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2101 * there'd be no LMV layer thus we might be called here. Eventually
2102 * this code should be removed.
2105 case IOC_OBD_STATFS: {
2106 struct obd_statfs stat_buf = {0};
2108 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2109 GOTO(out, rc = -ENODEV);
2112 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2113 min((int)data->ioc_plen2,
2114 (int)sizeof(struct obd_uuid))))
2115 GOTO(out, rc = -EFAULT);
2117 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2118 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2123 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2124 min((int) data->ioc_plen1,
2125 (int) sizeof(stat_buf))))
2126 GOTO(out, rc = -EFAULT);
2130 case OBD_IOC_QUOTACTL: {
2131 struct if_quotactl *qctl = karg;
2132 struct obd_quotactl *oqctl;
2134 OBD_ALLOC_PTR(oqctl);
2136 GOTO(out, rc = -ENOMEM);
2138 QCTL_COPY(oqctl, qctl);
2139 rc = obd_quotactl(exp, oqctl);
2141 QCTL_COPY(qctl, oqctl);
2142 qctl->qc_valid = QC_MDTIDX;
2143 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2146 OBD_FREE_PTR(oqctl);
2149 case LL_IOC_GET_CONNECT_FLAGS:
2150 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2151 sizeof(*exp_connect_flags_ptr(exp))))
2152 GOTO(out, rc = -EFAULT);
2155 case LL_IOC_LOV_SWAP_LAYOUTS:
2156 rc = mdc_ioc_swap_layouts(exp, karg);
2159 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2160 GOTO(out, rc = -ENOTTY);
2163 module_put(THIS_MODULE);
2168 static int mdc_get_info_rpc(struct obd_export *exp,
2169 u32 keylen, void *key,
2170 u32 vallen, void *val)
2172 struct obd_import *imp = class_exp2cliimp(exp);
2173 struct ptlrpc_request *req;
2178 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2182 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2183 RCL_CLIENT, keylen);
2184 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2185 RCL_CLIENT, sizeof(vallen));
2187 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2189 ptlrpc_request_free(req);
2193 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2194 memcpy(tmp, key, keylen);
2195 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2196 memcpy(tmp, &vallen, sizeof(vallen));
2198 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2199 RCL_SERVER, vallen);
2200 ptlrpc_request_set_replen(req);
2202 rc = ptlrpc_queue_wait(req);
2203 /* -EREMOTE means the get_info result is partial, and it needs to
2204 * continue on another MDT, see fid2path part in lmv_iocontrol */
2205 if (rc == 0 || rc == -EREMOTE) {
2206 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2207 memcpy(val, tmp, vallen);
2208 if (ptlrpc_rep_need_swab(req)) {
2209 if (KEY_IS(KEY_FID2PATH))
2210 lustre_swab_fid2path(val);
2213 ptlrpc_req_finished(req);
2218 static void lustre_swab_hai(struct hsm_action_item *h)
2220 __swab32s(&h->hai_len);
2221 __swab32s(&h->hai_action);
2222 lustre_swab_lu_fid(&h->hai_fid);
2223 lustre_swab_lu_fid(&h->hai_dfid);
2224 __swab64s(&h->hai_cookie);
2225 __swab64s(&h->hai_extent.offset);
2226 __swab64s(&h->hai_extent.length);
2227 __swab64s(&h->hai_gid);
2230 static void lustre_swab_hal(struct hsm_action_list *h)
2232 struct hsm_action_item *hai;
2235 __swab32s(&h->hal_version);
2236 __swab32s(&h->hal_count);
2237 __swab32s(&h->hal_archive_id);
2238 __swab64s(&h->hal_flags);
2240 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2241 lustre_swab_hai(hai);
2244 static void lustre_swab_kuch(struct kuc_hdr *l)
2246 __swab16s(&l->kuc_magic);
2247 /* __u8 l->kuc_transport */
2248 __swab16s(&l->kuc_msgtype);
2249 __swab16s(&l->kuc_msglen);
2252 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2253 struct lustre_kernelcomm *lk)
2255 struct obd_import *imp = class_exp2cliimp(exp);
2258 if (lk->lk_group != KUC_GRP_HSM) {
2259 CERROR("Bad copytool group %d\n", lk->lk_group);
2263 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2264 lk->lk_uid, lk->lk_group, lk->lk_flags);
2266 if (lk->lk_flags & LK_FLG_STOP) {
2267 /* Unregister with the coordinator */
2268 rc = mdc_ioc_hsm_ct_unregister(imp);
2270 __u32 *archives = NULL;
2272 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2273 archives = lk->lk_data;
2275 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2282 * Send a message to any listening copytools
2283 * @param val KUC message (kuc_hdr + hsm_action_list)
2284 * @param len total length of message
2286 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2287 size_t len, void *val)
2289 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2290 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2294 if (len < sizeof(*lh) + sizeof(*hal)) {
2295 CERROR("Short HSM message %zu < %zu\n", len,
2296 sizeof(*lh) + sizeof(*hal));
2299 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2300 lustre_swab_kuch(lh);
2301 lustre_swab_hal(hal);
2302 } else if (lh->kuc_magic != KUC_MAGIC) {
2303 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2307 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2309 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2310 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2312 /* Broadcast to HSM listeners */
2313 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2319 * callback function passed to kuc for re-registering each HSM copytool
2320 * running on MDC, after MDT shutdown/recovery.
2321 * @param data copytool registration data
2322 * @param cb_arg callback argument (obd_import)
2324 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2326 struct obd_import *imp = (struct obd_import *)cb_arg;
2327 struct kkuc_ct_data *kcd = data;
2328 __u32 *archives = NULL;
2332 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2333 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2336 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2337 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2338 "(archive=%#x)\n", imp->imp_obd->obd_name,
2339 kcd->kcd_nr_archives);
2341 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2342 "(archive nr = %u)\n",
2343 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2344 if (kcd->kcd_nr_archives != 0)
2345 archives = kcd->kcd_archives;
2348 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2349 /* ignore error if the copytool is already registered */
2350 return (rc == -EEXIST) ? 0 : rc;
2354 * Re-establish all kuc contexts with MDT
2355 * after MDT shutdown/recovery.
2357 static int mdc_kuc_reregister(struct obd_import *imp)
2359 /* re-register HSM agents */
2360 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2361 mdc_hsm_ct_reregister, imp);
2364 static int mdc_set_info_async(const struct lu_env *env,
2365 struct obd_export *exp,
2366 u32 keylen, void *key,
2367 u32 vallen, void *val,
2368 struct ptlrpc_request_set *set)
2370 struct obd_import *imp = class_exp2cliimp(exp);
2374 if (KEY_IS(KEY_READ_ONLY)) {
2375 if (vallen != sizeof(int))
2378 spin_lock(&imp->imp_lock);
2379 if (*((int *)val)) {
2380 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2381 imp->imp_connect_data.ocd_connect_flags |=
2384 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2385 imp->imp_connect_data.ocd_connect_flags &=
2386 ~OBD_CONNECT_RDONLY;
2388 spin_unlock(&imp->imp_lock);
2390 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2391 keylen, key, vallen, val, set);
2394 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2395 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2396 keylen, key, vallen, val, set);
2399 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2400 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2405 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2406 __u32 *default_easize = val;
2408 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2412 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2416 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2417 __u32 keylen, void *key, __u32 *vallen, void *val)
2421 if (KEY_IS(KEY_MAX_EASIZE)) {
2422 __u32 mdsize, *max_easize;
2424 if (*vallen != sizeof(int))
2426 mdsize = *(__u32 *)val;
2427 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2428 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2430 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2432 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2433 __u32 *default_easize;
2435 if (*vallen != sizeof(int))
2437 default_easize = val;
2438 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2440 } else if (KEY_IS(KEY_CONN_DATA)) {
2441 struct obd_import *imp = class_exp2cliimp(exp);
2442 struct obd_connect_data *data = val;
2444 if (*vallen != sizeof(*data))
2447 *data = imp->imp_connect_data;
2449 } else if (KEY_IS(KEY_TGT_COUNT)) {
2450 *((__u32 *)val) = 1;
2454 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2459 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2460 struct ptlrpc_request **request)
2462 struct ptlrpc_request *req;
2467 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2471 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2473 ptlrpc_request_free(req);
2477 mdc_pack_body(req, fid, 0, 0, -1, 0);
2479 ptlrpc_request_set_replen(req);
2481 rc = ptlrpc_queue_wait(req);
2483 ptlrpc_req_finished(req);
2489 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2490 enum obd_import_event event)
2492 struct client_obd *cli = &obd->u.cli;
2495 LASSERT(imp->imp_obd == obd);
2498 case IMP_EVENT_DISCON:
2499 spin_lock(&cli->cl_loi_list_lock);
2500 cli->cl_avail_grant = 0;
2501 cli->cl_lost_grant = 0;
2502 spin_unlock(&cli->cl_loi_list_lock);
2504 case IMP_EVENT_INACTIVE:
2506 * Flush current sequence to make client obtain new one
2507 * from server in case of disconnect/reconnect.
2509 down_read(&cli->cl_seq_rwsem);
2511 seq_client_flush(cli->cl_seq);
2512 up_read(&cli->cl_seq_rwsem);
2514 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2516 case IMP_EVENT_INVALIDATE: {
2517 struct ldlm_namespace *ns = obd->obd_namespace;
2521 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2523 env = cl_env_get(&refcheck);
2525 /* Reset grants. All pages go to failing rpcs due to
2526 * the invalid import.
2528 osc_io_unplug(env, cli, NULL);
2530 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2531 osc_ldlm_resource_invalidate,
2533 cl_env_put(env, &refcheck);
2534 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2540 case IMP_EVENT_ACTIVE:
2541 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2542 /* redo the kuc registration after reconnecting */
2544 rc = mdc_kuc_reregister(imp);
2546 case IMP_EVENT_OCD: {
2547 struct obd_connect_data *ocd = &imp->imp_connect_data;
2549 if (OCD_HAS_FLAG(ocd, GRANT))
2550 osc_init_grant(cli, ocd);
2552 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2555 case IMP_EVENT_DEACTIVATE:
2556 case IMP_EVENT_ACTIVATE:
2559 CERROR("Unknown import event %x\n", event);
2565 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2566 struct lu_fid *fid, struct md_op_data *op_data)
2568 struct client_obd *cli = &exp->exp_obd->u.cli;
2573 down_read(&cli->cl_seq_rwsem);
2575 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2576 up_read(&cli->cl_seq_rwsem);
2581 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2583 struct client_obd *cli = &exp->exp_obd->u.cli;
2584 return &cli->cl_target_uuid;
2588 * Determine whether the lock can be canceled before replaying it during
2589 * recovery, non zero value will be return if the lock can be canceled,
2590 * or zero returned for not
2592 static int mdc_cancel_weight(struct ldlm_lock *lock)
2594 if (lock->l_resource->lr_type != LDLM_IBITS)
2597 /* FIXME: if we ever get into a situation where there are too many
2598 * opened files with open locks on a single node, then we really
2599 * should replay these open locks to reget it */
2600 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2603 /* Special case for DoM locks, cancel only unused and granted locks */
2604 if (ldlm_has_dom(lock) &&
2605 (lock->l_granted_mode != lock->l_req_mode ||
2606 osc_ldlm_weigh_ast(lock) != 0))
2612 static int mdc_resource_inode_free(struct ldlm_resource *res)
2614 if (res->lr_lvb_inode)
2615 res->lr_lvb_inode = NULL;
2620 static struct ldlm_valblock_ops inode_lvbo = {
2621 .lvbo_free = mdc_resource_inode_free
2624 static int mdc_llog_init(struct obd_device *obd)
2626 struct obd_llog_group *olg = &obd->obd_olg;
2627 struct llog_ctxt *ctxt;
2632 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2637 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2638 llog_initiator_connect(ctxt);
2639 llog_ctxt_put(ctxt);
2644 static void mdc_llog_finish(struct obd_device *obd)
2646 struct llog_ctxt *ctxt;
2650 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2652 llog_cleanup(NULL, ctxt);
2657 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2663 rc = osc_setup_common(obd, cfg);
2667 rc = mdc_tunables_init(obd);
2669 GOTO(err_osc_cleanup, rc);
2671 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2673 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2675 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2677 rc = mdc_llog_init(obd);
2679 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2681 GOTO(err_llog_cleanup, rc);
2684 rc = mdc_changelog_cdev_init(obd);
2686 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2688 GOTO(err_changelog_cleanup, rc);
2693 err_changelog_cleanup:
2694 mdc_llog_finish(obd);
2696 lprocfs_free_md_stats(obd);
2697 ptlrpc_lprocfs_unregister_obd(obd);
2699 osc_cleanup_common(obd);
2703 /* Initialize the default and maximum LOV EA sizes. This allows
2704 * us to make MDS RPCs with large enough reply buffers to hold a default
2705 * sized EA without having to calculate this (via a call into the
2706 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2707 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2708 * a large number of stripes is possible. If a larger reply buffer is
2709 * required it will be reallocated in the ptlrpc layer due to overflow.
2711 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2714 struct obd_device *obd = exp->exp_obd;
2715 struct client_obd *cli = &obd->u.cli;
2718 if (cli->cl_max_mds_easize < easize)
2719 cli->cl_max_mds_easize = easize;
2721 if (cli->cl_default_mds_easize < def_easize)
2722 cli->cl_default_mds_easize = def_easize;
2727 static int mdc_precleanup(struct obd_device *obd)
2731 osc_precleanup_common(obd);
2732 mdc_changelog_cdev_finish(obd);
2734 obd_cleanup_client_import(obd);
2735 ptlrpc_lprocfs_unregister_obd(obd);
2736 lprocfs_free_md_stats(obd);
2737 mdc_llog_finish(obd);
2741 static int mdc_cleanup(struct obd_device *obd)
2743 return osc_cleanup_common(obd);
2746 static struct obd_ops mdc_obd_ops = {
2747 .o_owner = THIS_MODULE,
2748 .o_setup = mdc_setup,
2749 .o_precleanup = mdc_precleanup,
2750 .o_cleanup = mdc_cleanup,
2751 .o_add_conn = client_import_add_conn,
2752 .o_del_conn = client_import_del_conn,
2753 .o_connect = client_connect_import,
2754 .o_reconnect = osc_reconnect,
2755 .o_disconnect = osc_disconnect,
2756 .o_iocontrol = mdc_iocontrol,
2757 .o_set_info_async = mdc_set_info_async,
2758 .o_statfs = mdc_statfs,
2759 .o_fid_init = client_fid_init,
2760 .o_fid_fini = client_fid_fini,
2761 .o_fid_alloc = mdc_fid_alloc,
2762 .o_import_event = mdc_import_event,
2763 .o_get_info = mdc_get_info,
2764 .o_get_uuid = mdc_get_uuid,
2765 .o_quotactl = mdc_quotactl,
2768 static struct md_ops mdc_md_ops = {
2769 .m_get_root = mdc_get_root,
2770 .m_null_inode = mdc_null_inode,
2771 .m_close = mdc_close,
2772 .m_create = mdc_create,
2773 .m_enqueue = mdc_enqueue,
2774 .m_getattr = mdc_getattr,
2775 .m_getattr_name = mdc_getattr_name,
2776 .m_intent_lock = mdc_intent_lock,
2778 .m_rename = mdc_rename,
2779 .m_setattr = mdc_setattr,
2780 .m_setxattr = mdc_setxattr,
2781 .m_getxattr = mdc_getxattr,
2782 .m_fsync = mdc_fsync,
2783 .m_file_resync = mdc_file_resync,
2784 .m_read_page = mdc_read_page,
2785 .m_unlink = mdc_unlink,
2786 .m_cancel_unused = mdc_cancel_unused,
2787 .m_init_ea_size = mdc_init_ea_size,
2788 .m_set_lock_data = mdc_set_lock_data,
2789 .m_lock_match = mdc_lock_match,
2790 .m_get_lustre_md = mdc_get_lustre_md,
2791 .m_free_lustre_md = mdc_free_lustre_md,
2792 .m_set_open_replay_data = mdc_set_open_replay_data,
2793 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2794 .m_intent_getattr_async = mdc_intent_getattr_async,
2795 .m_revalidate_lock = mdc_revalidate_lock
2798 static int __init mdc_init(void)
2800 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2801 LUSTRE_MDC_NAME, &mdc_device_type);
2804 static void __exit mdc_exit(void)
2806 class_unregister_type(LUSTRE_MDC_NAME);
2809 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2810 MODULE_DESCRIPTION("Lustre Metadata Client");
2811 MODULE_VERSION(LUSTRE_VERSION_STRING);
2812 MODULE_LICENSE("GPL");
2814 module_init(mdc_init);
2815 module_exit(mdc_exit);