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 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
444 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
449 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
451 GOTO(out, rc = -EPROTO);
453 /* only detect the xattr size */
455 /* LU-11109: Older MDTs do not distinguish
456 * between nonexistent xattrs and zero length
457 * values in this case. Newer MDTs will return
458 * -ENODATA or set OBD_MD_FLXATTR. */
459 GOTO(out, rc = body->mbo_eadatasize);
462 if (body->mbo_eadatasize == 0) {
463 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
464 * success so that we can distinguish between
465 * zero length value and nonexistent xattr.
467 * If OBD_MD_FLXATTR is not set then we keep
468 * the old behavior and return -ENODATA for
469 * getxattr() when mbo_eadatasize is 0. But
470 * -ENODATA only makes sense for getxattr()
471 * and not for listxattr(). */
472 if (body->mbo_valid & OBD_MD_FLXATTR)
474 else if (obd_md_valid == OBD_MD_FLXATTR)
475 GOTO(out, rc = -ENODATA);
480 GOTO(out, rc = body->mbo_eadatasize);
483 ptlrpc_req_finished(*req);
490 #ifdef CONFIG_FS_POSIX_ACL
491 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
493 struct req_capsule *pill = &req->rq_pill;
494 struct mdt_body *body = md->body;
495 struct posix_acl *acl;
500 if (!body->mbo_aclsize)
503 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
508 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
513 CERROR("convert xattr to acl: %d\n", rc);
517 rc = posix_acl_valid(&init_user_ns, acl);
519 CERROR("validate acl: %d\n", rc);
520 posix_acl_release(acl);
528 #define mdc_unpack_acl(req, md) 0
531 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
532 struct obd_export *dt_exp, struct obd_export *md_exp,
533 struct lustre_md *md)
535 struct req_capsule *pill = &req->rq_pill;
540 memset(md, 0, sizeof(*md));
542 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
543 LASSERT(md->body != NULL);
545 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
546 if (!S_ISREG(md->body->mbo_mode)) {
547 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
548 "regular file, but is not\n");
549 GOTO(out, rc = -EPROTO);
552 if (md->body->mbo_eadatasize == 0) {
553 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
554 "but eadatasize 0\n");
555 GOTO(out, rc = -EPROTO);
558 md->layout.lb_len = md->body->mbo_eadatasize;
559 md->layout.lb_buf = req_capsule_server_sized_get(pill,
562 if (md->layout.lb_buf == NULL)
563 GOTO(out, rc = -EPROTO);
564 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
565 const union lmv_mds_md *lmv;
568 if (!S_ISDIR(md->body->mbo_mode)) {
569 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
570 "directory, but is not\n");
571 GOTO(out, rc = -EPROTO);
574 lmv_size = md->body->mbo_eadatasize;
576 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
577 "but eadatasize 0\n");
581 if (md->body->mbo_valid & OBD_MD_MEA) {
582 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
585 GOTO(out, rc = -EPROTO);
587 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
591 if (rc < (typeof(rc))sizeof(*md->lmv)) {
592 CDEBUG(D_INFO, "size too small: "
593 "rc < sizeof(*md->lmv) (%d < %d)\n",
594 rc, (int)sizeof(*md->lmv));
595 GOTO(out, rc = -EPROTO);
601 if (md->body->mbo_valid & OBD_MD_FLACL) {
602 /* for ACL, it's possible that FLACL is set but aclsize is zero.
603 * only when aclsize != 0 there's an actual segment for ACL
606 if (md->body->mbo_aclsize) {
607 rc = mdc_unpack_acl(req, md);
610 #ifdef CONFIG_FS_POSIX_ACL
612 md->posix_acl = NULL;
620 #ifdef CONFIG_FS_POSIX_ACL
621 posix_acl_release(md->posix_acl);
627 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
633 void mdc_replay_open(struct ptlrpc_request *req)
635 struct md_open_data *mod = req->rq_cb_data;
636 struct ptlrpc_request *close_req;
637 struct obd_client_handle *och;
638 struct lustre_handle old_open_handle = { };
639 struct mdt_body *body;
643 DEBUG_REQ(D_ERROR, req,
644 "Can't properly replay without open data.");
649 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
650 LASSERT(body != NULL);
652 spin_lock(&req->rq_lock);
654 if (och && och->och_open_handle.cookie)
655 req->rq_early_free_repbuf = 1;
657 req->rq_early_free_repbuf = 0;
658 spin_unlock(&req->rq_lock);
660 if (req->rq_early_free_repbuf) {
661 struct lustre_handle *file_open_handle;
663 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
665 file_open_handle = &och->och_open_handle;
666 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
667 file_open_handle->cookie, body->mbo_open_handle.cookie);
668 old_open_handle = *file_open_handle;
669 *file_open_handle = body->mbo_open_handle;
672 close_req = mod->mod_close_req;
674 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
675 struct mdt_ioepoch *epoch;
677 LASSERT(opc == MDS_CLOSE);
678 epoch = req_capsule_client_get(&close_req->rq_pill,
682 if (req->rq_early_free_repbuf)
683 LASSERT(old_open_handle.cookie ==
684 epoch->mio_open_handle.cookie);
686 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
687 epoch->mio_open_handle = body->mbo_open_handle;
692 void mdc_commit_open(struct ptlrpc_request *req)
694 struct md_open_data *mod = req->rq_cb_data;
699 * No need to touch md_open_data::mod_och, it holds a reference on
700 * \var mod and will zero references to each other, \var mod will be
701 * freed after that when md_open_data::mod_och will put the reference.
705 * Do not let open request to disappear as it still may be needed
706 * for close rpc to happen (it may happen on evict only, otherwise
707 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
708 * called), just mark this rpc as committed to distinguish these 2
709 * cases, see mdc_close() for details. The open request reference will
710 * be put along with freeing \var mod.
712 ptlrpc_request_addref(req);
713 spin_lock(&req->rq_lock);
714 req->rq_committed = 1;
715 spin_unlock(&req->rq_lock);
716 req->rq_cb_data = NULL;
720 int mdc_set_open_replay_data(struct obd_export *exp,
721 struct obd_client_handle *och,
722 struct lookup_intent *it)
724 struct md_open_data *mod;
725 struct mdt_rec_create *rec;
726 struct mdt_body *body;
727 struct ptlrpc_request *open_req = it->it_request;
728 struct obd_import *imp = open_req->rq_import;
731 if (!open_req->rq_replay)
734 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
735 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
736 LASSERT(rec != NULL);
737 /* Incoming message in my byte order (it's been swabbed). */
738 /* Outgoing messages always in my byte order. */
739 LASSERT(body != NULL);
741 /* Only if the import is replayable, we set replay_open data */
742 if (och && imp->imp_replayable) {
743 mod = obd_mod_alloc();
745 DEBUG_REQ(D_ERROR, open_req,
746 "Can't allocate md_open_data");
751 * Take a reference on \var mod, to be freed on mdc_close().
752 * It protects \var mod from being freed on eviction (commit
753 * callback is called despite rq_replay flag).
754 * Another reference for \var och.
759 spin_lock(&open_req->rq_lock);
762 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
763 it_disposition(it, DISP_OPEN_STRIPE);
764 mod->mod_open_req = open_req;
765 open_req->rq_cb_data = mod;
766 open_req->rq_commit_cb = mdc_commit_open;
767 open_req->rq_early_free_repbuf = 1;
768 spin_unlock(&open_req->rq_lock);
771 rec->cr_fid2 = body->mbo_fid1;
772 rec->cr_open_handle_old = body->mbo_open_handle;
773 open_req->rq_replay_cb = mdc_replay_open;
774 if (!fid_is_sane(&body->mbo_fid1)) {
775 DEBUG_REQ(D_ERROR, open_req,
776 "saving replay request with insane FID " DFID,
777 PFID(&body->mbo_fid1));
781 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
785 static void mdc_free_open(struct md_open_data *mod)
789 if (mod->mod_is_create == 0 &&
790 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
794 * No reason to asssert here if the open request has
795 * rq_replay == 1. It means that mdc_close failed, and
796 * close request wasn`t sent. It is not fatal to client.
797 * The worst thing is eviction if the client gets open lock
800 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
801 "= %d\n", mod->mod_open_req->rq_replay);
803 ptlrpc_request_committed(mod->mod_open_req, committed);
804 if (mod->mod_close_req)
805 ptlrpc_request_committed(mod->mod_close_req, committed);
808 int mdc_clear_open_replay_data(struct obd_export *exp,
809 struct obd_client_handle *och)
811 struct md_open_data *mod = och->och_mod;
815 * It is possible to not have \var mod in a case of eviction between
816 * lookup and ll_file_open().
821 LASSERT(mod != LP_POISON);
822 LASSERT(mod->mod_open_req != NULL);
824 spin_lock(&mod->mod_open_req->rq_lock);
826 mod->mod_och->och_open_handle.cookie = 0;
827 mod->mod_open_req->rq_early_free_repbuf = 0;
828 spin_unlock(&mod->mod_open_req->rq_lock);
838 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
839 struct md_open_data *mod, struct ptlrpc_request **request)
841 struct obd_device *obd = class_exp2obd(exp);
842 struct ptlrpc_request *req;
843 struct req_format *req_fmt;
844 size_t u32_count = 0;
849 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
850 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
853 if (op_data->op_bias & MDS_CLOSE_INTENT) {
854 req_fmt = &RQF_MDS_CLOSE_INTENT;
855 if (op_data->op_bias & MDS_HSM_RELEASE) {
856 /* allocate a FID for volatile file */
857 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
860 CERROR("%s: "DFID" allocating FID: rc = %d\n",
861 obd->obd_name, PFID(&op_data->op_fid1),
863 /* save the errcode and proceed to close */
867 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
868 size_t count = op_data->op_data_size / sizeof(__u32);
870 if (count > INLINE_RESYNC_ARRAY_SIZE)
874 req_fmt = &RQF_MDS_CLOSE;
878 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
881 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
883 /* Ensure that this close's handle is fixed up during replay. */
884 if (likely(mod != NULL)) {
885 LASSERTF(mod->mod_open_req != NULL &&
886 mod->mod_open_req->rq_type != LI_POISON,
887 "POISONED open %p!\n", mod->mod_open_req);
889 mod->mod_close_req = req;
891 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
892 /* We no longer want to preserve this open for replay even
893 * though the open was committed. b=3632, b=3633 */
894 spin_lock(&mod->mod_open_req->rq_lock);
895 mod->mod_open_req->rq_replay = 0;
896 spin_unlock(&mod->mod_open_req->rq_lock);
898 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
902 * TODO: repeat close after errors
904 CWARN("%s: close of FID "DFID" failed, file reference will be "
905 "dropped when this client unmounts or is evicted\n",
906 obd->obd_name, PFID(&op_data->op_fid1));
907 GOTO(out, rc = -ENOMEM);
911 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
912 u32_count * sizeof(__u32));
914 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
916 ptlrpc_request_free(req);
921 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
922 * portal whose threads are not taking any DLM locks and are therefore
923 * always progressing */
924 req->rq_request_portal = MDS_READPAGE_PORTAL;
925 ptlrpc_at_set_req_timeout(req);
928 mdc_close_pack(req, op_data);
930 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
931 obd->u.cli.cl_default_mds_easize);
933 ptlrpc_request_set_replen(req);
935 mdc_get_mod_rpc_slot(req, NULL);
936 rc = ptlrpc_queue_wait(req);
937 mdc_put_mod_rpc_slot(req, NULL);
939 if (req->rq_repmsg == NULL) {
940 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
943 rc = req->rq_status ?: -EIO;
944 } else if (rc == 0 || rc == -EAGAIN) {
945 struct mdt_body *body;
947 rc = lustre_msg_get_status(req->rq_repmsg);
948 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
949 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
954 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
957 } else if (rc == -ESTALE) {
959 * it can be allowed error after 3633 if open was committed and
960 * server failed before close was sent. Let's check if mod
961 * exists and return no error in that case
964 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
965 LASSERT(mod->mod_open_req != NULL);
966 if (mod->mod_open_req->rq_committed)
974 mod->mod_close_req = NULL;
975 /* Since now, mod is accessed through open_req only,
976 * thus close req does not keep a reference on mod anymore. */
981 RETURN(rc < 0 ? rc : saved_rc);
984 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
985 u64 offset, struct page **pages, int npages,
986 struct ptlrpc_request **request)
988 struct ptlrpc_request *req;
989 struct ptlrpc_bulk_desc *desc;
991 wait_queue_head_t waitq;
993 struct l_wait_info lwi;
998 init_waitqueue_head(&waitq);
1001 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1005 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1007 ptlrpc_request_free(req);
1011 req->rq_request_portal = MDS_READPAGE_PORTAL;
1012 ptlrpc_at_set_req_timeout(req);
1014 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1015 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1017 &ptlrpc_bulk_kiov_pin_ops);
1019 ptlrpc_req_finished(req);
1023 /* NB req now owns desc and will free it when it gets freed */
1024 for (i = 0; i < npages; i++)
1025 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1028 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1030 ptlrpc_request_set_replen(req);
1031 rc = ptlrpc_queue_wait(req);
1033 ptlrpc_req_finished(req);
1034 if (rc != -ETIMEDOUT)
1038 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1039 CERROR("%s: too many resend retries: rc = %d\n",
1040 exp->exp_obd->obd_name, -EIO);
1043 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1045 l_wait_event(waitq, 0, &lwi);
1050 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1051 req->rq_bulk->bd_nob_transferred);
1053 ptlrpc_req_finished(req);
1057 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1058 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1059 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1060 PAGE_SIZE * npages);
1061 ptlrpc_req_finished(req);
1069 static void mdc_release_page(struct page *page, int remove)
1073 if (likely(page->mapping != NULL))
1074 truncate_complete_page(page->mapping, page);
1080 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1081 __u64 *start, __u64 *end, int hash64)
1084 * Complement of hash is used as an index so that
1085 * radix_tree_gang_lookup() can be used to find a page with starting
1086 * hash _smaller_ than one we are looking for.
1088 unsigned long offset = hash_x_index(*hash, hash64);
1092 spin_lock_irq(&mapping->tree_lock);
1093 found = radix_tree_gang_lookup(&mapping->page_tree,
1094 (void **)&page, offset, 1);
1095 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1096 struct lu_dirpage *dp;
1099 spin_unlock_irq(&mapping->tree_lock);
1101 * In contrast to find_lock_page() we are sure that directory
1102 * page cannot be truncated (while DLM lock is held) and,
1103 * hence, can avoid restart.
1105 * In fact, page cannot be locked here at all, because
1106 * mdc_read_page_remote does synchronous io.
1108 wait_on_page_locked(page);
1109 if (PageUptodate(page)) {
1111 if (BITS_PER_LONG == 32 && hash64) {
1112 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1113 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1114 *hash = *hash >> 32;
1116 *start = le64_to_cpu(dp->ldp_hash_start);
1117 *end = le64_to_cpu(dp->ldp_hash_end);
1119 if (unlikely(*start == 1 && *hash == 0))
1122 LASSERTF(*start <= *hash, "start = %#llx"
1123 ",end = %#llx,hash = %#llx\n",
1124 *start, *end, *hash);
1125 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1126 " hash %#llx\n", offset, *start, *end, *hash);
1129 mdc_release_page(page, 0);
1131 } else if (*end != *start && *hash == *end) {
1133 * upon hash collision, remove this page,
1134 * otherwise put page reference, and
1135 * mdc_read_page_remote() will issue RPC to
1136 * fetch the page we want.
1139 mdc_release_page(page,
1140 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1145 page = ERR_PTR(-EIO);
1148 spin_unlock_irq(&mapping->tree_lock);
1155 * Adjust a set of pages, each page containing an array of lu_dirpages,
1156 * so that each page can be used as a single logical lu_dirpage.
1158 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1159 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1160 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1161 * value is used as a cookie to request the next lu_dirpage in a
1162 * directory listing that spans multiple pages (two in this example):
1165 * .|--------v------- -----.
1166 * |s|e|f|p|ent|ent| ... |ent|
1167 * '--|-------------- -----' Each PAGE contains a single
1168 * '------. lu_dirpage.
1169 * .---------v------- -----.
1170 * |s|e|f|p|ent| 0 | ... | 0 |
1171 * '----------------- -----'
1173 * However, on hosts where the native VM page size (PAGE_SIZE) is
1174 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1175 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1176 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1177 * after it in the same PAGE (arrows simplified for brevity, but
1178 * in general e0==s1, e1==s2, etc.):
1180 * .-------------------- -----.
1181 * |s0|e0|f0|p|ent|ent| ... |ent|
1182 * |---v---------------- -----|
1183 * |s1|e1|f1|p|ent|ent| ... |ent|
1184 * |---v---------------- -----| Here, each PAGE contains
1185 * ... multiple lu_dirpages.
1186 * |---v---------------- -----|
1187 * |s'|e'|f'|p|ent|ent| ... |ent|
1188 * '---|---------------- -----'
1190 * .----------------------------.
1193 * This structure is transformed into a single logical lu_dirpage as follows:
1195 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1196 * labeled 'next PAGE'.
1198 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1199 * a hash collision with the next page exists.
1201 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1202 * to the first entry of the next lu_dirpage.
1204 #if PAGE_SIZE > LU_PAGE_SIZE
1205 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1209 for (i = 0; i < cfs_pgs; i++) {
1210 struct lu_dirpage *dp = kmap(pages[i]);
1211 struct lu_dirpage *first = dp;
1212 struct lu_dirent *end_dirent = NULL;
1213 struct lu_dirent *ent;
1214 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1215 __u32 flags = le32_to_cpu(dp->ldp_flags);
1217 while (--lu_pgs > 0) {
1218 ent = lu_dirent_start(dp);
1219 for (end_dirent = ent; ent != NULL;
1220 end_dirent = ent, ent = lu_dirent_next(ent));
1222 /* Advance dp to next lu_dirpage. */
1223 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1225 /* Check if we've reached the end of the PAGE. */
1226 if (!((unsigned long)dp & ~PAGE_MASK))
1229 /* Save the hash and flags of this lu_dirpage. */
1230 hash_end = le64_to_cpu(dp->ldp_hash_end);
1231 flags = le32_to_cpu(dp->ldp_flags);
1233 /* Check if lu_dirpage contains no entries. */
1234 if (end_dirent == NULL)
1237 /* Enlarge the end entry lde_reclen from 0 to
1238 * first entry of next lu_dirpage. */
1239 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1240 end_dirent->lde_reclen =
1241 cpu_to_le16((char *)(dp->ldp_entries) -
1242 (char *)end_dirent);
1245 first->ldp_hash_end = hash_end;
1246 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1247 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1251 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1254 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1255 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1257 /* parameters for readdir page */
1258 struct readpage_param {
1259 struct md_op_data *rp_mod;
1262 struct obd_export *rp_exp;
1263 struct md_callback *rp_cb;
1266 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1267 static inline void delete_from_page_cache(struct page *page)
1269 remove_from_page_cache(page);
1275 * Read pages from server.
1277 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1278 * a header lu_dirpage which describes the start/end hash, and whether this
1279 * page is empty (contains no dir entry) or hash collide with next page.
1280 * After client receives reply, several pages will be integrated into dir page
1281 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1282 * lu_dirpage for this integrated page will be adjusted.
1284 static int mdc_read_page_remote(void *data, struct page *page0)
1286 struct readpage_param *rp = data;
1287 struct page **page_pool;
1289 struct lu_dirpage *dp;
1290 struct md_op_data *op_data = rp->rp_mod;
1291 struct ptlrpc_request *req;
1293 struct inode *inode;
1295 int rd_pgs = 0; /* number of pages actually read */
1301 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1302 inode = op_data->op_data;
1303 fid = &op_data->op_fid1;
1304 LASSERT(inode != NULL);
1306 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1307 if (page_pool != NULL) {
1308 page_pool[0] = page0;
1314 for (npages = 1; npages < max_pages; npages++) {
1315 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1319 page_pool[npages] = page;
1322 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1324 /* page0 is special, which was added into page cache early */
1325 delete_from_page_cache(page0);
1329 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1331 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1332 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1334 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1336 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1338 SetPageUptodate(page0);
1342 ptlrpc_req_finished(req);
1343 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1344 for (i = 1; i < npages; i++) {
1345 unsigned long offset;
1349 page = page_pool[i];
1351 if (rc < 0 || i >= rd_pgs) {
1356 SetPageUptodate(page);
1359 hash = le64_to_cpu(dp->ldp_hash_start);
1362 offset = hash_x_index(hash, rp->rp_hash64);
1364 prefetchw(&page->flags);
1365 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1370 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1371 " rc = %d\n", offset, ret);
1375 if (page_pool != &page0)
1376 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1382 * Read dir page from cache first, if it can not find it, read it from
1383 * server and add into the cache.
1385 * \param[in] exp MDC export
1386 * \param[in] op_data client MD stack parameters, transfering parameters
1387 * between different layers on client MD stack.
1388 * \param[in] cb_op callback required for ldlm lock enqueue during
1390 * \param[in] hash_offset the hash offset of the page to be read
1391 * \param[in] ppage the page to be read
1393 * retval = 0 get the page successfully
1394 * errno(<0) get the page failed
1396 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1397 struct md_callback *cb_op, __u64 hash_offset,
1398 struct page **ppage)
1400 struct lookup_intent it = { .it_op = IT_READDIR };
1402 struct inode *dir = op_data->op_data;
1403 struct address_space *mapping;
1404 struct lu_dirpage *dp;
1407 struct lustre_handle lockh;
1408 struct ptlrpc_request *enq_req = NULL;
1409 struct readpage_param rp_param;
1416 LASSERT(dir != NULL);
1417 mapping = dir->i_mapping;
1419 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1420 cb_op->md_blocking_ast, 0);
1421 if (enq_req != NULL)
1422 ptlrpc_req_finished(enq_req);
1425 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1426 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1431 lockh.cookie = it.it_lock_handle;
1432 mdc_set_lock_data(exp, &lockh, dir, NULL);
1434 rp_param.rp_off = hash_offset;
1435 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1436 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1437 rp_param.rp_hash64);
1439 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1440 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1441 rp_param.rp_off, PTR_ERR(page));
1442 GOTO(out_unlock, rc = PTR_ERR(page));
1443 } else if (page != NULL) {
1445 * XXX nikita: not entirely correct handling of a corner case:
1446 * suppose hash chain of entries with hash value HASH crosses
1447 * border between pages P0 and P1. First both P0 and P1 are
1448 * cached, seekdir() is called for some entry from the P0 part
1449 * of the chain. Later P0 goes out of cache. telldir(HASH)
1450 * happens and finds P1, as it starts with matching hash
1451 * value. Remaining entries from P0 part of the chain are
1452 * skipped. (Is that really a bug?)
1454 * Possible solutions: 0. don't cache P1 is such case, handle
1455 * it as an "overflow" page. 1. invalidate all pages at
1456 * once. 2. use HASH|1 as an index for P1.
1458 GOTO(hash_collision, page);
1461 rp_param.rp_exp = exp;
1462 rp_param.rp_mod = op_data;
1463 page = read_cache_page(mapping,
1464 hash_x_index(rp_param.rp_off,
1465 rp_param.rp_hash64),
1466 mdc_read_page_remote, &rp_param);
1468 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1469 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1470 rp_param.rp_off, PTR_ERR(page));
1471 GOTO(out_unlock, rc = PTR_ERR(page));
1474 wait_on_page_locked(page);
1476 if (!PageUptodate(page)) {
1477 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1478 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1479 rp_param.rp_off, -5);
1482 if (!PageChecked(page))
1483 SetPageChecked(page);
1484 if (PageError(page)) {
1485 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1486 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1487 rp_param.rp_off, -5);
1492 dp = page_address(page);
1493 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1494 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1495 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1496 rp_param.rp_off = hash_offset >> 32;
1498 start = le64_to_cpu(dp->ldp_hash_start);
1499 end = le64_to_cpu(dp->ldp_hash_end);
1500 rp_param.rp_off = hash_offset;
1503 LASSERT(start == rp_param.rp_off);
1504 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1505 #if BITS_PER_LONG == 32
1506 CWARN("Real page-wide hash collision at [%llu %llu] with "
1507 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1508 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1512 * Fetch whole overflow chain...
1520 ldlm_lock_decref(&lockh, it.it_lock_mode);
1524 mdc_release_page(page, 1);
1529 static int mdc_statfs(const struct lu_env *env,
1530 struct obd_export *exp, struct obd_statfs *osfs,
1531 time64_t max_age, __u32 flags)
1533 struct obd_device *obd = class_exp2obd(exp);
1534 struct req_format *fmt;
1535 struct ptlrpc_request *req;
1536 struct obd_statfs *msfs;
1537 struct obd_import *imp = NULL;
1542 * Since the request might also come from lprocfs, so we need
1543 * sync this with client_disconnect_export Bug15684
1545 down_read(&obd->u.cli.cl_sem);
1546 if (obd->u.cli.cl_import)
1547 imp = class_import_get(obd->u.cli.cl_import);
1548 up_read(&obd->u.cli.cl_sem);
1552 fmt = &RQF_MDS_STATFS;
1553 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1554 (flags & OBD_STATFS_SUM))
1555 fmt = &RQF_MDS_STATFS_NEW;
1556 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1559 GOTO(output, rc = -ENOMEM);
1561 if ((flags & OBD_STATFS_SUM) &&
1562 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1563 /* request aggregated states */
1564 struct mdt_body *body;
1566 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1568 GOTO(out, rc = -EPROTO);
1569 body->mbo_valid = OBD_MD_FLAGSTATFS;
1572 ptlrpc_request_set_replen(req);
1574 if (flags & OBD_STATFS_NODELAY) {
1575 /* procfs requests not want stay in wait for avoid deadlock */
1576 req->rq_no_resend = 1;
1577 req->rq_no_delay = 1;
1580 rc = ptlrpc_queue_wait(req);
1582 /* check connection error first */
1583 if (imp->imp_connect_error)
1584 rc = imp->imp_connect_error;
1588 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1590 GOTO(out, rc = -EPROTO);
1595 ptlrpc_req_finished(req);
1597 class_import_put(imp);
1601 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1603 __u32 keylen, vallen;
1607 if (gf->gf_pathlen > PATH_MAX)
1608 RETURN(-ENAMETOOLONG);
1609 if (gf->gf_pathlen < 2)
1612 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1613 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1614 sizeof(struct lu_fid));
1615 OBD_ALLOC(key, keylen);
1618 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1619 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1620 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1621 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1622 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1623 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1625 if (!fid_is_sane(&gf->gf_fid))
1626 GOTO(out, rc = -EINVAL);
1628 /* Val is struct getinfo_fid2path result plus path */
1629 vallen = sizeof(*gf) + gf->gf_pathlen;
1631 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1632 if (rc != 0 && rc != -EREMOTE)
1635 if (vallen <= sizeof(*gf))
1636 GOTO(out, rc = -EPROTO);
1637 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1638 GOTO(out, rc = -EOVERFLOW);
1640 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1641 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1642 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1643 /* only log the last 512 characters of the path */
1644 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1647 OBD_FREE(key, keylen);
1651 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1652 struct hsm_progress_kernel *hpk)
1654 struct obd_import *imp = class_exp2cliimp(exp);
1655 struct hsm_progress_kernel *req_hpk;
1656 struct ptlrpc_request *req;
1660 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1661 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1663 GOTO(out, rc = -ENOMEM);
1665 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1667 /* Copy hsm_progress struct */
1668 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1669 if (req_hpk == NULL)
1670 GOTO(out, rc = -EPROTO);
1673 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1675 ptlrpc_request_set_replen(req);
1677 mdc_get_mod_rpc_slot(req, NULL);
1678 rc = ptlrpc_queue_wait(req);
1679 mdc_put_mod_rpc_slot(req, NULL);
1683 ptlrpc_req_finished(req);
1687 * Send hsm_ct_register to MDS
1689 * \param[in] imp import
1690 * \param[in] archive_count if in bitmap format, it is the bitmap,
1691 * else it is the count of archive_ids
1692 * \param[in] archives if in bitmap format, it is NULL,
1693 * else it is archive_id lists
1695 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1698 struct ptlrpc_request *req;
1699 __u32 *archive_array;
1700 size_t archives_size;
1704 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1708 if (archives != NULL)
1709 archives_size = sizeof(*archive_array) * archive_count;
1711 archives_size = sizeof(archive_count);
1713 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1714 RCL_CLIENT, archives_size);
1716 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1718 ptlrpc_request_free(req);
1722 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1724 archive_array = req_capsule_client_get(&req->rq_pill,
1725 &RMF_MDS_HSM_ARCHIVE);
1726 if (archive_array == NULL)
1727 GOTO(out, rc = -EPROTO);
1729 if (archives != NULL)
1730 memcpy(archive_array, archives, archives_size);
1732 *archive_array = archive_count;
1734 ptlrpc_request_set_replen(req);
1736 rc = mdc_queue_wait(req);
1739 ptlrpc_req_finished(req);
1743 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1744 struct md_op_data *op_data)
1746 struct hsm_current_action *hca = op_data->op_data;
1747 struct hsm_current_action *req_hca;
1748 struct ptlrpc_request *req;
1752 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1753 &RQF_MDS_HSM_ACTION);
1757 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1759 ptlrpc_request_free(req);
1763 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1764 op_data->op_suppgids[0], 0);
1766 ptlrpc_request_set_replen(req);
1768 rc = mdc_queue_wait(req);
1772 req_hca = req_capsule_server_get(&req->rq_pill,
1773 &RMF_MDS_HSM_CURRENT_ACTION);
1774 if (req_hca == NULL)
1775 GOTO(out, rc = -EPROTO);
1781 ptlrpc_req_finished(req);
1785 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1787 struct ptlrpc_request *req;
1791 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1793 MDS_HSM_CT_UNREGISTER);
1795 GOTO(out, rc = -ENOMEM);
1797 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1799 ptlrpc_request_set_replen(req);
1801 rc = mdc_queue_wait(req);
1804 ptlrpc_req_finished(req);
1808 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1809 struct md_op_data *op_data)
1811 struct hsm_user_state *hus = op_data->op_data;
1812 struct hsm_user_state *req_hus;
1813 struct ptlrpc_request *req;
1817 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1818 &RQF_MDS_HSM_STATE_GET);
1822 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1824 ptlrpc_request_free(req);
1828 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1829 op_data->op_suppgids[0], 0);
1831 ptlrpc_request_set_replen(req);
1833 rc = mdc_queue_wait(req);
1837 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1838 if (req_hus == NULL)
1839 GOTO(out, rc = -EPROTO);
1845 ptlrpc_req_finished(req);
1849 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1850 struct md_op_data *op_data)
1852 struct hsm_state_set *hss = op_data->op_data;
1853 struct hsm_state_set *req_hss;
1854 struct ptlrpc_request *req;
1858 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1859 &RQF_MDS_HSM_STATE_SET);
1863 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1865 ptlrpc_request_free(req);
1869 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1870 op_data->op_suppgids[0], 0);
1873 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1874 if (req_hss == NULL)
1875 GOTO(out, rc = -EPROTO);
1878 ptlrpc_request_set_replen(req);
1880 mdc_get_mod_rpc_slot(req, NULL);
1881 rc = ptlrpc_queue_wait(req);
1882 mdc_put_mod_rpc_slot(req, NULL);
1886 ptlrpc_req_finished(req);
1890 static int mdc_ioc_hsm_request(struct obd_export *exp,
1891 struct hsm_user_request *hur)
1893 struct obd_import *imp = class_exp2cliimp(exp);
1894 struct ptlrpc_request *req;
1895 struct hsm_request *req_hr;
1896 struct hsm_user_item *req_hui;
1901 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1903 GOTO(out, rc = -ENOMEM);
1905 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1906 hur->hur_request.hr_itemcount
1907 * sizeof(struct hsm_user_item));
1908 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1909 hur->hur_request.hr_data_len);
1911 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1913 ptlrpc_request_free(req);
1917 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1919 /* Copy hsm_request struct */
1920 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1922 GOTO(out, rc = -EPROTO);
1923 *req_hr = hur->hur_request;
1925 /* Copy hsm_user_item structs */
1926 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1927 if (req_hui == NULL)
1928 GOTO(out, rc = -EPROTO);
1929 memcpy(req_hui, hur->hur_user_item,
1930 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1932 /* Copy opaque field */
1933 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1934 if (req_opaque == NULL)
1935 GOTO(out, rc = -EPROTO);
1936 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1938 ptlrpc_request_set_replen(req);
1940 mdc_get_mod_rpc_slot(req, NULL);
1941 rc = ptlrpc_queue_wait(req);
1942 mdc_put_mod_rpc_slot(req, NULL);
1947 ptlrpc_req_finished(req);
1951 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1952 struct lustre_kernelcomm *lk);
1954 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1955 struct obd_quotactl *oqctl)
1957 struct ptlrpc_request *req;
1958 struct obd_quotactl *oqc;
1962 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1963 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1968 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1971 ptlrpc_request_set_replen(req);
1972 ptlrpc_at_set_req_timeout(req);
1974 rc = ptlrpc_queue_wait(req);
1976 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1978 if (req->rq_repmsg &&
1979 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1982 CERROR ("Can't unpack obd_quotactl\n");
1985 ptlrpc_req_finished(req);
1990 static int mdc_ioc_swap_layouts(struct obd_export *exp,
1991 struct md_op_data *op_data)
1993 struct list_head cancels = LIST_HEAD_INIT(cancels);
1994 struct ptlrpc_request *req;
1996 struct mdc_swap_layouts *msl, *payload;
1999 msl = op_data->op_data;
2001 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2002 * first thing it will do is to cancel the 2 layout
2003 * locks held by this client.
2004 * So the client must cancel its layout locks on the 2 fids
2005 * with the request RPC to avoid extra RPC round trips.
2007 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2008 LCK_EX, MDS_INODELOCK_LAYOUT |
2009 MDS_INODELOCK_XATTR);
2010 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2011 LCK_EX, MDS_INODELOCK_LAYOUT |
2012 MDS_INODELOCK_XATTR);
2014 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2015 &RQF_MDS_SWAP_LAYOUTS);
2017 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2021 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2023 ptlrpc_request_free(req);
2027 mdc_swap_layouts_pack(req, op_data);
2029 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2034 ptlrpc_request_set_replen(req);
2036 rc = ptlrpc_queue_wait(req);
2042 ptlrpc_req_finished(req);
2046 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2047 void *karg, void __user *uarg)
2049 struct obd_device *obd = exp->exp_obd;
2050 struct obd_ioctl_data *data = karg;
2051 struct obd_import *imp = obd->u.cli.cl_import;
2055 if (!try_module_get(THIS_MODULE)) {
2056 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2057 module_name(THIS_MODULE));
2061 case OBD_IOC_FID2PATH:
2062 rc = mdc_ioc_fid2path(exp, karg);
2064 case LL_IOC_HSM_CT_START:
2065 rc = mdc_ioc_hsm_ct_start(exp, karg);
2066 /* ignore if it was already registered on this MDS. */
2070 case LL_IOC_HSM_PROGRESS:
2071 rc = mdc_ioc_hsm_progress(exp, karg);
2073 case LL_IOC_HSM_STATE_GET:
2074 rc = mdc_ioc_hsm_state_get(exp, karg);
2076 case LL_IOC_HSM_STATE_SET:
2077 rc = mdc_ioc_hsm_state_set(exp, karg);
2079 case LL_IOC_HSM_ACTION:
2080 rc = mdc_ioc_hsm_current_action(exp, karg);
2082 case LL_IOC_HSM_REQUEST:
2083 rc = mdc_ioc_hsm_request(exp, karg);
2085 case OBD_IOC_CLIENT_RECOVER:
2086 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2090 case IOC_OSC_SET_ACTIVE:
2091 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2093 case OBD_IOC_PING_TARGET:
2094 rc = ptlrpc_obd_ping(obd);
2097 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2098 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2099 * there'd be no LMV layer thus we might be called here. Eventually
2100 * this code should be removed.
2103 case IOC_OBD_STATFS: {
2104 struct obd_statfs stat_buf = {0};
2106 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2107 GOTO(out, rc = -ENODEV);
2110 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2111 min((int)data->ioc_plen2,
2112 (int)sizeof(struct obd_uuid))))
2113 GOTO(out, rc = -EFAULT);
2115 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2116 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2121 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2122 min((int) data->ioc_plen1,
2123 (int) sizeof(stat_buf))))
2124 GOTO(out, rc = -EFAULT);
2128 case OBD_IOC_QUOTACTL: {
2129 struct if_quotactl *qctl = karg;
2130 struct obd_quotactl *oqctl;
2132 OBD_ALLOC_PTR(oqctl);
2134 GOTO(out, rc = -ENOMEM);
2136 QCTL_COPY(oqctl, qctl);
2137 rc = obd_quotactl(exp, oqctl);
2139 QCTL_COPY(qctl, oqctl);
2140 qctl->qc_valid = QC_MDTIDX;
2141 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2144 OBD_FREE_PTR(oqctl);
2147 case LL_IOC_GET_CONNECT_FLAGS:
2148 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2149 sizeof(*exp_connect_flags_ptr(exp))))
2150 GOTO(out, rc = -EFAULT);
2153 case LL_IOC_LOV_SWAP_LAYOUTS:
2154 rc = mdc_ioc_swap_layouts(exp, karg);
2157 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2158 GOTO(out, rc = -ENOTTY);
2161 module_put(THIS_MODULE);
2166 static int mdc_get_info_rpc(struct obd_export *exp,
2167 u32 keylen, void *key,
2168 u32 vallen, void *val)
2170 struct obd_import *imp = class_exp2cliimp(exp);
2171 struct ptlrpc_request *req;
2176 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2180 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2181 RCL_CLIENT, keylen);
2182 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2183 RCL_CLIENT, sizeof(vallen));
2185 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2187 ptlrpc_request_free(req);
2191 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2192 memcpy(tmp, key, keylen);
2193 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2194 memcpy(tmp, &vallen, sizeof(vallen));
2196 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2197 RCL_SERVER, vallen);
2198 ptlrpc_request_set_replen(req);
2200 rc = ptlrpc_queue_wait(req);
2201 /* -EREMOTE means the get_info result is partial, and it needs to
2202 * continue on another MDT, see fid2path part in lmv_iocontrol */
2203 if (rc == 0 || rc == -EREMOTE) {
2204 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2205 memcpy(val, tmp, vallen);
2206 if (ptlrpc_rep_need_swab(req)) {
2207 if (KEY_IS(KEY_FID2PATH))
2208 lustre_swab_fid2path(val);
2211 ptlrpc_req_finished(req);
2216 static void lustre_swab_hai(struct hsm_action_item *h)
2218 __swab32s(&h->hai_len);
2219 __swab32s(&h->hai_action);
2220 lustre_swab_lu_fid(&h->hai_fid);
2221 lustre_swab_lu_fid(&h->hai_dfid);
2222 __swab64s(&h->hai_cookie);
2223 __swab64s(&h->hai_extent.offset);
2224 __swab64s(&h->hai_extent.length);
2225 __swab64s(&h->hai_gid);
2228 static void lustre_swab_hal(struct hsm_action_list *h)
2230 struct hsm_action_item *hai;
2233 __swab32s(&h->hal_version);
2234 __swab32s(&h->hal_count);
2235 __swab32s(&h->hal_archive_id);
2236 __swab64s(&h->hal_flags);
2238 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2239 lustre_swab_hai(hai);
2242 static void lustre_swab_kuch(struct kuc_hdr *l)
2244 __swab16s(&l->kuc_magic);
2245 /* __u8 l->kuc_transport */
2246 __swab16s(&l->kuc_msgtype);
2247 __swab16s(&l->kuc_msglen);
2250 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2251 struct lustre_kernelcomm *lk)
2253 struct obd_import *imp = class_exp2cliimp(exp);
2256 if (lk->lk_group != KUC_GRP_HSM) {
2257 CERROR("Bad copytool group %d\n", lk->lk_group);
2261 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2262 lk->lk_uid, lk->lk_group, lk->lk_flags);
2264 if (lk->lk_flags & LK_FLG_STOP) {
2265 /* Unregister with the coordinator */
2266 rc = mdc_ioc_hsm_ct_unregister(imp);
2268 __u32 *archives = NULL;
2270 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2271 archives = lk->lk_data;
2273 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2280 * Send a message to any listening copytools
2281 * @param val KUC message (kuc_hdr + hsm_action_list)
2282 * @param len total length of message
2284 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2285 size_t len, void *val)
2287 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2288 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2292 if (len < sizeof(*lh) + sizeof(*hal)) {
2293 CERROR("Short HSM message %zu < %zu\n", len,
2294 sizeof(*lh) + sizeof(*hal));
2297 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2298 lustre_swab_kuch(lh);
2299 lustre_swab_hal(hal);
2300 } else if (lh->kuc_magic != KUC_MAGIC) {
2301 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2305 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2307 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2308 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2310 /* Broadcast to HSM listeners */
2311 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2317 * callback function passed to kuc for re-registering each HSM copytool
2318 * running on MDC, after MDT shutdown/recovery.
2319 * @param data copytool registration data
2320 * @param cb_arg callback argument (obd_import)
2322 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2324 struct obd_import *imp = (struct obd_import *)cb_arg;
2325 struct kkuc_ct_data *kcd = data;
2326 __u32 *archives = NULL;
2330 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2331 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2334 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2335 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2336 "(archive=%#x)\n", imp->imp_obd->obd_name,
2337 kcd->kcd_nr_archives);
2339 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2340 "(archive nr = %u)\n",
2341 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2342 if (kcd->kcd_nr_archives != 0)
2343 archives = kcd->kcd_archives;
2346 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2347 /* ignore error if the copytool is already registered */
2348 return (rc == -EEXIST) ? 0 : rc;
2352 * Re-establish all kuc contexts with MDT
2353 * after MDT shutdown/recovery.
2355 static int mdc_kuc_reregister(struct obd_import *imp)
2357 /* re-register HSM agents */
2358 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2359 mdc_hsm_ct_reregister, imp);
2362 static int mdc_set_info_async(const struct lu_env *env,
2363 struct obd_export *exp,
2364 u32 keylen, void *key,
2365 u32 vallen, void *val,
2366 struct ptlrpc_request_set *set)
2368 struct obd_import *imp = class_exp2cliimp(exp);
2372 if (KEY_IS(KEY_READ_ONLY)) {
2373 if (vallen != sizeof(int))
2376 spin_lock(&imp->imp_lock);
2377 if (*((int *)val)) {
2378 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2379 imp->imp_connect_data.ocd_connect_flags |=
2382 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2383 imp->imp_connect_data.ocd_connect_flags &=
2384 ~OBD_CONNECT_RDONLY;
2386 spin_unlock(&imp->imp_lock);
2388 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2389 keylen, key, vallen, val, set);
2392 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2393 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2394 keylen, key, vallen, val, set);
2397 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2398 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2403 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2404 __u32 *default_easize = val;
2406 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2410 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2414 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2415 __u32 keylen, void *key, __u32 *vallen, void *val)
2419 if (KEY_IS(KEY_MAX_EASIZE)) {
2420 __u32 mdsize, *max_easize;
2422 if (*vallen != sizeof(int))
2424 mdsize = *(__u32 *)val;
2425 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2426 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2428 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2430 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2431 __u32 *default_easize;
2433 if (*vallen != sizeof(int))
2435 default_easize = val;
2436 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2438 } else if (KEY_IS(KEY_CONN_DATA)) {
2439 struct obd_import *imp = class_exp2cliimp(exp);
2440 struct obd_connect_data *data = val;
2442 if (*vallen != sizeof(*data))
2445 *data = imp->imp_connect_data;
2447 } else if (KEY_IS(KEY_TGT_COUNT)) {
2448 *((__u32 *)val) = 1;
2452 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2457 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2458 struct ptlrpc_request **request)
2460 struct ptlrpc_request *req;
2465 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2469 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2471 ptlrpc_request_free(req);
2475 mdc_pack_body(req, fid, 0, 0, -1, 0);
2477 ptlrpc_request_set_replen(req);
2479 rc = ptlrpc_queue_wait(req);
2481 ptlrpc_req_finished(req);
2487 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2488 enum obd_import_event event)
2490 struct client_obd *cli = &obd->u.cli;
2493 LASSERT(imp->imp_obd == obd);
2496 case IMP_EVENT_DISCON:
2497 spin_lock(&cli->cl_loi_list_lock);
2498 cli->cl_avail_grant = 0;
2499 cli->cl_lost_grant = 0;
2500 spin_unlock(&cli->cl_loi_list_lock);
2502 case IMP_EVENT_INACTIVE:
2504 * Flush current sequence to make client obtain new one
2505 * from server in case of disconnect/reconnect.
2507 down_read(&cli->cl_seq_rwsem);
2509 seq_client_flush(cli->cl_seq);
2510 up_read(&cli->cl_seq_rwsem);
2512 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2514 case IMP_EVENT_INVALIDATE: {
2515 struct ldlm_namespace *ns = obd->obd_namespace;
2519 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2521 env = cl_env_get(&refcheck);
2523 /* Reset grants. All pages go to failing rpcs due to
2524 * the invalid import.
2526 osc_io_unplug(env, cli, NULL);
2528 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2529 osc_ldlm_resource_invalidate,
2531 cl_env_put(env, &refcheck);
2532 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2538 case IMP_EVENT_ACTIVE:
2539 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2540 /* redo the kuc registration after reconnecting */
2542 rc = mdc_kuc_reregister(imp);
2544 case IMP_EVENT_OCD: {
2545 struct obd_connect_data *ocd = &imp->imp_connect_data;
2547 if (OCD_HAS_FLAG(ocd, GRANT))
2548 osc_init_grant(cli, ocd);
2550 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2553 case IMP_EVENT_DEACTIVATE:
2554 case IMP_EVENT_ACTIVATE:
2557 CERROR("Unknown import event %x\n", event);
2563 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2564 struct lu_fid *fid, struct md_op_data *op_data)
2566 struct client_obd *cli = &exp->exp_obd->u.cli;
2571 down_read(&cli->cl_seq_rwsem);
2573 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2574 up_read(&cli->cl_seq_rwsem);
2579 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2581 struct client_obd *cli = &exp->exp_obd->u.cli;
2582 return &cli->cl_target_uuid;
2586 * Determine whether the lock can be canceled before replaying it during
2587 * recovery, non zero value will be return if the lock can be canceled,
2588 * or zero returned for not
2590 static int mdc_cancel_weight(struct ldlm_lock *lock)
2592 if (lock->l_resource->lr_type != LDLM_IBITS)
2595 /* FIXME: if we ever get into a situation where there are too many
2596 * opened files with open locks on a single node, then we really
2597 * should replay these open locks to reget it */
2598 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2601 /* Special case for DoM locks, cancel only unused and granted locks */
2602 if (ldlm_has_dom(lock) &&
2603 (lock->l_granted_mode != lock->l_req_mode ||
2604 osc_ldlm_weigh_ast(lock) != 0))
2610 static int mdc_resource_inode_free(struct ldlm_resource *res)
2612 if (res->lr_lvb_inode)
2613 res->lr_lvb_inode = NULL;
2618 static struct ldlm_valblock_ops inode_lvbo = {
2619 .lvbo_free = mdc_resource_inode_free
2622 static int mdc_llog_init(struct obd_device *obd)
2624 struct obd_llog_group *olg = &obd->obd_olg;
2625 struct llog_ctxt *ctxt;
2630 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2635 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2636 llog_initiator_connect(ctxt);
2637 llog_ctxt_put(ctxt);
2642 static void mdc_llog_finish(struct obd_device *obd)
2644 struct llog_ctxt *ctxt;
2648 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2650 llog_cleanup(NULL, ctxt);
2655 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2661 rc = osc_setup_common(obd, cfg);
2665 rc = mdc_tunables_init(obd);
2667 GOTO(err_osc_cleanup, rc);
2669 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2671 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2673 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2675 rc = mdc_llog_init(obd);
2677 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2679 GOTO(err_llog_cleanup, rc);
2682 rc = mdc_changelog_cdev_init(obd);
2684 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2686 GOTO(err_changelog_cleanup, rc);
2691 err_changelog_cleanup:
2692 mdc_llog_finish(obd);
2694 lprocfs_free_md_stats(obd);
2695 ptlrpc_lprocfs_unregister_obd(obd);
2697 osc_cleanup_common(obd);
2701 /* Initialize the default and maximum LOV EA sizes. This allows
2702 * us to make MDS RPCs with large enough reply buffers to hold a default
2703 * sized EA without having to calculate this (via a call into the
2704 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2705 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2706 * a large number of stripes is possible. If a larger reply buffer is
2707 * required it will be reallocated in the ptlrpc layer due to overflow.
2709 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2712 struct obd_device *obd = exp->exp_obd;
2713 struct client_obd *cli = &obd->u.cli;
2716 if (cli->cl_max_mds_easize < easize)
2717 cli->cl_max_mds_easize = easize;
2719 if (cli->cl_default_mds_easize < def_easize)
2720 cli->cl_default_mds_easize = def_easize;
2725 static int mdc_precleanup(struct obd_device *obd)
2729 osc_precleanup_common(obd);
2730 mdc_changelog_cdev_finish(obd);
2732 obd_cleanup_client_import(obd);
2733 ptlrpc_lprocfs_unregister_obd(obd);
2734 lprocfs_free_md_stats(obd);
2735 mdc_llog_finish(obd);
2739 static int mdc_cleanup(struct obd_device *obd)
2741 return osc_cleanup_common(obd);
2744 static struct obd_ops mdc_obd_ops = {
2745 .o_owner = THIS_MODULE,
2746 .o_setup = mdc_setup,
2747 .o_precleanup = mdc_precleanup,
2748 .o_cleanup = mdc_cleanup,
2749 .o_add_conn = client_import_add_conn,
2750 .o_del_conn = client_import_del_conn,
2751 .o_connect = client_connect_import,
2752 .o_reconnect = osc_reconnect,
2753 .o_disconnect = osc_disconnect,
2754 .o_iocontrol = mdc_iocontrol,
2755 .o_set_info_async = mdc_set_info_async,
2756 .o_statfs = mdc_statfs,
2757 .o_fid_init = client_fid_init,
2758 .o_fid_fini = client_fid_fini,
2759 .o_fid_alloc = mdc_fid_alloc,
2760 .o_import_event = mdc_import_event,
2761 .o_get_info = mdc_get_info,
2762 .o_get_uuid = mdc_get_uuid,
2763 .o_quotactl = mdc_quotactl,
2766 static struct md_ops mdc_md_ops = {
2767 .m_get_root = mdc_get_root,
2768 .m_null_inode = mdc_null_inode,
2769 .m_close = mdc_close,
2770 .m_create = mdc_create,
2771 .m_enqueue = mdc_enqueue,
2772 .m_getattr = mdc_getattr,
2773 .m_getattr_name = mdc_getattr_name,
2774 .m_intent_lock = mdc_intent_lock,
2776 .m_rename = mdc_rename,
2777 .m_setattr = mdc_setattr,
2778 .m_setxattr = mdc_setxattr,
2779 .m_getxattr = mdc_getxattr,
2780 .m_fsync = mdc_fsync,
2781 .m_file_resync = mdc_file_resync,
2782 .m_read_page = mdc_read_page,
2783 .m_unlink = mdc_unlink,
2784 .m_cancel_unused = mdc_cancel_unused,
2785 .m_init_ea_size = mdc_init_ea_size,
2786 .m_set_lock_data = mdc_set_lock_data,
2787 .m_lock_match = mdc_lock_match,
2788 .m_get_lustre_md = mdc_get_lustre_md,
2789 .m_free_lustre_md = mdc_free_lustre_md,
2790 .m_set_open_replay_data = mdc_set_open_replay_data,
2791 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2792 .m_intent_getattr_async = mdc_intent_getattr_async,
2793 .m_revalidate_lock = mdc_revalidate_lock
2796 static int __init mdc_init(void)
2798 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2799 LUSTRE_MDC_NAME, &mdc_device_type);
2802 static void __exit mdc_exit(void)
2804 class_unregister_type(LUSTRE_MDC_NAME);
2807 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2808 MODULE_DESCRIPTION("Lustre Metadata Client");
2809 MODULE_VERSION(LUSTRE_VERSION_STRING);
2810 MODULE_LICENSE("GPL");
2812 module_init(mdc_init);
2813 module_exit(mdc_exit);