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 <uapi/linux/lustre/lustre_param.h>
57 #include <lustre_swab.h>
58 #include <obd_class.h>
59 #include <lustre_osc.h>
61 #include "mdc_internal.h"
63 #define REQUEST_MINOR 244
65 static int mdc_cleanup(struct obd_device *obd);
67 static inline int mdc_queue_wait(struct ptlrpc_request *req)
69 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
72 /* obd_get_request_slot() ensures that this client has no more
73 * than cl_max_rpcs_in_flight RPCs simultaneously inf light
75 rc = obd_get_request_slot(cli);
79 rc = ptlrpc_queue_wait(req);
80 obd_put_request_slot(cli);
86 * Send MDS_GET_ROOT RPC to fetch root FID.
88 * If \a fileset is not NULL it should contain a subdirectory off
89 * the ROOT/ directory to be mounted on the client. Return the FID
90 * of the subdirectory to the client to mount onto its mountpoint.
92 * \param[in] imp MDC import
93 * \param[in] fileset fileset name, which could be NULL
94 * \param[out] rootfid root FID of this mountpoint
95 * \param[out] pc root capa will be unpacked and saved in this pointer
97 * \retval 0 on success, negative errno on failure
99 static int mdc_get_root(struct obd_export *exp, const char *fileset,
100 struct lu_fid *rootfid)
102 struct ptlrpc_request *req;
103 struct mdt_body *body;
108 if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
111 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
117 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
118 strlen(fileset) + 1);
119 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
121 ptlrpc_request_free(req);
124 mdc_pack_body(req, NULL, 0, 0, -1, 0);
125 if (fileset != NULL) {
126 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
128 memcpy(name, fileset, strlen(fileset));
130 lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
131 req->rq_send_state = LUSTRE_IMP_FULL;
133 ptlrpc_request_set_replen(req);
135 rc = ptlrpc_queue_wait(req);
139 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
141 GOTO(out, rc = -EPROTO);
143 *rootfid = body->mbo_fid1;
144 CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
145 PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
148 ptlrpc_req_finished(req);
154 * This function now is known to always saying that it will receive 4 buffers
155 * from server. Even for cases when acl_size and md_size is zero, RPC header
156 * will contain 4 fields and RPC itself will contain zero size fields. This is
157 * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
158 * and thus zero, it shrinks it, making zero size. The same story about
159 * md_size. And this is course of problem when client waits for smaller number
160 * of fields. This issue will be fixed later when client gets aware of RPC
163 static int mdc_getattr_common(struct obd_export *exp,
164 struct ptlrpc_request *req)
166 struct req_capsule *pill = &req->rq_pill;
167 struct mdt_body *body;
172 /* Request message already built. */
173 rc = ptlrpc_queue_wait(req);
177 /* sanity check for the reply */
178 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
182 CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
184 mdc_update_max_ea_from_body(exp, body);
185 if (body->mbo_eadatasize != 0) {
186 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
187 body->mbo_eadatasize);
195 static void mdc_reset_acl_req(struct ptlrpc_request *req)
197 spin_lock(&req->rq_early_free_lock);
198 sptlrpc_cli_free_repbuf(req);
199 req->rq_repbuf = NULL;
200 req->rq_repbuf_len = 0;
201 req->rq_repdata = NULL;
202 req->rq_reqdata_len = 0;
203 spin_unlock(&req->rq_early_free_lock);
206 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
207 struct ptlrpc_request **request)
209 struct ptlrpc_request *req;
210 struct obd_import *imp = class_exp2cliimp(exp);
211 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
215 /* Single MDS without an LMV case */
216 if (op_data->op_flags & MF_GET_MDT_IDX) {
222 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
226 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
228 ptlrpc_request_free(req);
233 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
234 op_data->op_mode, -1, 0);
235 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
236 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
238 ptlrpc_request_set_replen(req);
240 rc = mdc_getattr_common(exp, req);
243 acl_bufsize != imp->imp_connect_data.ocd_max_easize) {
244 acl_bufsize = imp->imp_connect_data.ocd_max_easize;
245 mdc_reset_acl_req(req);
249 ptlrpc_req_finished(req);
257 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
258 struct ptlrpc_request **request)
260 struct ptlrpc_request *req;
261 struct obd_import *imp = class_exp2cliimp(exp);
262 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
267 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
271 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
272 op_data->op_namelen + 1);
274 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
276 ptlrpc_request_free(req);
280 if (op_data->op_name) {
281 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
282 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
283 op_data->op_namelen);
284 memcpy(name, op_data->op_name, op_data->op_namelen);
288 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
289 op_data->op_mode, op_data->op_suppgids[0], 0);
290 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
292 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
293 ptlrpc_request_set_replen(req);
295 rc = mdc_getattr_common(exp, req);
298 acl_bufsize != imp->imp_connect_data.ocd_max_easize) {
299 acl_bufsize = imp->imp_connect_data.ocd_max_easize;
300 mdc_reset_acl_req(req);
304 ptlrpc_req_finished(req);
312 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
313 const struct lu_fid *fid, int opcode, u64 valid,
314 const char *xattr_name, const char *input,
315 int input_size, int output_size, int flags,
316 __u32 suppgid, struct ptlrpc_request **request)
318 struct ptlrpc_request *req;
319 int xattr_namelen = 0;
325 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
330 xattr_namelen = strlen(xattr_name) + 1;
331 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
336 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
340 /* Flush local XATTR locks to get rid of a possible cancel RPC */
341 if (opcode == MDS_REINT && fid_is_sane(fid) &&
342 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
343 struct list_head cancels = LIST_HEAD_INIT(cancels);
346 /* Without that packing would fail */
348 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
351 count = mdc_resource_get_unused(exp, fid,
353 MDS_INODELOCK_XATTR);
355 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
357 ptlrpc_request_free(req);
361 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
363 ptlrpc_request_free(req);
368 if (opcode == MDS_REINT) {
369 struct mdt_rec_setxattr *rec;
371 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
372 sizeof(struct mdt_rec_reint));
373 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
374 rec->sx_opcode = REINT_SETXATTR;
375 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
376 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
377 rec->sx_cap = cfs_curproc_cap_pack();
378 rec->sx_suppgid1 = suppgid;
379 rec->sx_suppgid2 = -1;
381 rec->sx_valid = valid | OBD_MD_FLCTIME;
382 rec->sx_time = ktime_get_real_seconds();
383 rec->sx_size = output_size;
384 rec->sx_flags = flags;
386 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
390 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
391 memcpy(tmp, xattr_name, xattr_namelen);
394 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
395 memcpy(tmp, input, input_size);
398 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
399 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
400 RCL_SERVER, output_size);
401 ptlrpc_request_set_replen(req);
404 if (opcode == MDS_REINT)
405 mdc_get_mod_rpc_slot(req, NULL);
407 rc = ptlrpc_queue_wait(req);
409 if (opcode == MDS_REINT)
410 mdc_put_mod_rpc_slot(req, NULL);
413 ptlrpc_req_finished(req);
419 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
420 u64 obd_md_valid, const char *name,
421 const void *value, size_t value_size,
422 unsigned int xattr_flags, u32 suppgid,
423 struct ptlrpc_request **req)
425 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
426 obd_md_valid == OBD_MD_FLXATTRRM);
428 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
429 fid, MDS_REINT, obd_md_valid, name,
430 value, value_size, 0, xattr_flags, suppgid,
434 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
435 u64 obd_md_valid, const char *name, size_t buf_size,
436 struct ptlrpc_request **req)
438 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
439 obd_md_valid == OBD_MD_FLXATTRLS);
441 return mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
442 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
446 #ifdef CONFIG_FS_POSIX_ACL
447 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
449 struct req_capsule *pill = &req->rq_pill;
450 struct mdt_body *body = md->body;
451 struct posix_acl *acl;
456 if (!body->mbo_aclsize)
459 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
464 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
469 CERROR("convert xattr to acl: %d\n", rc);
473 rc = posix_acl_valid(&init_user_ns, acl);
475 CERROR("validate acl: %d\n", rc);
476 posix_acl_release(acl);
484 #define mdc_unpack_acl(req, md) 0
487 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
488 struct obd_export *dt_exp, struct obd_export *md_exp,
489 struct lustre_md *md)
491 struct req_capsule *pill = &req->rq_pill;
496 memset(md, 0, sizeof(*md));
498 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
499 LASSERT(md->body != NULL);
501 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
502 if (!S_ISREG(md->body->mbo_mode)) {
503 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
504 "regular file, but is not\n");
505 GOTO(out, rc = -EPROTO);
508 if (md->body->mbo_eadatasize == 0) {
509 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
510 "but eadatasize 0\n");
511 GOTO(out, rc = -EPROTO);
514 md->layout.lb_len = md->body->mbo_eadatasize;
515 md->layout.lb_buf = req_capsule_server_sized_get(pill,
518 if (md->layout.lb_buf == NULL)
519 GOTO(out, rc = -EPROTO);
520 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
521 const union lmv_mds_md *lmv;
524 if (!S_ISDIR(md->body->mbo_mode)) {
525 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
526 "directory, but is not\n");
527 GOTO(out, rc = -EPROTO);
530 lmv_size = md->body->mbo_eadatasize;
532 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
533 "but eadatasize 0\n");
537 if (md->body->mbo_valid & OBD_MD_MEA) {
538 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
541 GOTO(out, rc = -EPROTO);
543 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
547 if (rc < (typeof(rc))sizeof(*md->lmv)) {
548 CDEBUG(D_INFO, "size too small: "
549 "rc < sizeof(*md->lmv) (%d < %d)\n",
550 rc, (int)sizeof(*md->lmv));
551 GOTO(out, rc = -EPROTO);
557 if (md->body->mbo_valid & OBD_MD_FLACL) {
558 /* for ACL, it's possible that FLACL is set but aclsize is zero.
559 * only when aclsize != 0 there's an actual segment for ACL
562 if (md->body->mbo_aclsize) {
563 rc = mdc_unpack_acl(req, md);
566 #ifdef CONFIG_FS_POSIX_ACL
568 md->posix_acl = NULL;
576 #ifdef CONFIG_FS_POSIX_ACL
577 posix_acl_release(md->posix_acl);
583 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
589 void mdc_replay_open(struct ptlrpc_request *req)
591 struct md_open_data *mod = req->rq_cb_data;
592 struct ptlrpc_request *close_req;
593 struct obd_client_handle *och;
594 struct lustre_handle old_open_handle = { };
595 struct mdt_body *body;
599 DEBUG_REQ(D_ERROR, req,
600 "Can't properly replay without open data.");
605 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
606 LASSERT(body != NULL);
608 spin_lock(&req->rq_lock);
610 if (och && och->och_open_handle.cookie)
611 req->rq_early_free_repbuf = 1;
613 req->rq_early_free_repbuf = 0;
614 spin_unlock(&req->rq_lock);
616 if (req->rq_early_free_repbuf) {
617 struct lustre_handle *file_open_handle;
619 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
621 file_open_handle = &och->och_open_handle;
622 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
623 file_open_handle->cookie, body->mbo_open_handle.cookie);
624 old_open_handle = *file_open_handle;
625 *file_open_handle = body->mbo_open_handle;
628 close_req = mod->mod_close_req;
630 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
631 struct mdt_ioepoch *epoch;
633 LASSERT(opc == MDS_CLOSE);
634 epoch = req_capsule_client_get(&close_req->rq_pill,
638 if (req->rq_early_free_repbuf)
639 LASSERT(old_open_handle.cookie ==
640 epoch->mio_open_handle.cookie);
642 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
643 epoch->mio_open_handle = body->mbo_open_handle;
648 void mdc_commit_open(struct ptlrpc_request *req)
650 struct md_open_data *mod = req->rq_cb_data;
655 * No need to touch md_open_data::mod_och, it holds a reference on
656 * \var mod and will zero references to each other, \var mod will be
657 * freed after that when md_open_data::mod_och will put the reference.
661 * Do not let open request to disappear as it still may be needed
662 * for close rpc to happen (it may happen on evict only, otherwise
663 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
664 * called), just mark this rpc as committed to distinguish these 2
665 * cases, see mdc_close() for details. The open request reference will
666 * be put along with freeing \var mod.
668 ptlrpc_request_addref(req);
669 spin_lock(&req->rq_lock);
670 req->rq_committed = 1;
671 spin_unlock(&req->rq_lock);
672 req->rq_cb_data = NULL;
676 int mdc_set_open_replay_data(struct obd_export *exp,
677 struct obd_client_handle *och,
678 struct lookup_intent *it)
680 struct md_open_data *mod;
681 struct mdt_rec_create *rec;
682 struct mdt_body *body;
683 struct ptlrpc_request *open_req = it->it_request;
684 struct obd_import *imp = open_req->rq_import;
687 if (!open_req->rq_replay)
690 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
691 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
692 LASSERT(rec != NULL);
693 /* Incoming message in my byte order (it's been swabbed). */
694 /* Outgoing messages always in my byte order. */
695 LASSERT(body != NULL);
697 /* Only if the import is replayable, we set replay_open data */
698 if (och && imp->imp_replayable) {
699 mod = obd_mod_alloc();
701 DEBUG_REQ(D_ERROR, open_req,
702 "Can't allocate md_open_data");
707 * Take a reference on \var mod, to be freed on mdc_close().
708 * It protects \var mod from being freed on eviction (commit
709 * callback is called despite rq_replay flag).
710 * Another reference for \var och.
715 spin_lock(&open_req->rq_lock);
718 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
719 it_disposition(it, DISP_OPEN_STRIPE);
720 mod->mod_open_req = open_req;
721 open_req->rq_cb_data = mod;
722 open_req->rq_commit_cb = mdc_commit_open;
723 open_req->rq_early_free_repbuf = 1;
724 spin_unlock(&open_req->rq_lock);
727 rec->cr_fid2 = body->mbo_fid1;
728 rec->cr_open_handle_old = body->mbo_open_handle;
729 open_req->rq_replay_cb = mdc_replay_open;
730 if (!fid_is_sane(&body->mbo_fid1)) {
731 DEBUG_REQ(D_ERROR, open_req,
732 "saving replay request with insane FID " DFID,
733 PFID(&body->mbo_fid1));
737 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
741 static void mdc_free_open(struct md_open_data *mod)
745 if (mod->mod_is_create == 0 &&
746 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
750 * No reason to asssert here if the open request has
751 * rq_replay == 1. It means that mdc_close failed, and
752 * close request wasn`t sent. It is not fatal to client.
753 * The worst thing is eviction if the client gets open lock
756 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
757 "= %d\n", mod->mod_open_req->rq_replay);
759 ptlrpc_request_committed(mod->mod_open_req, committed);
760 if (mod->mod_close_req)
761 ptlrpc_request_committed(mod->mod_close_req, committed);
764 int mdc_clear_open_replay_data(struct obd_export *exp,
765 struct obd_client_handle *och)
767 struct md_open_data *mod = och->och_mod;
771 * It is possible to not have \var mod in a case of eviction between
772 * lookup and ll_file_open().
777 LASSERT(mod != LP_POISON);
778 LASSERT(mod->mod_open_req != NULL);
780 spin_lock(&mod->mod_open_req->rq_lock);
782 mod->mod_och->och_open_handle.cookie = 0;
783 mod->mod_open_req->rq_early_free_repbuf = 0;
784 spin_unlock(&mod->mod_open_req->rq_lock);
794 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
795 struct md_open_data *mod, struct ptlrpc_request **request)
797 struct obd_device *obd = class_exp2obd(exp);
798 struct ptlrpc_request *req;
799 struct req_format *req_fmt;
800 size_t u32_count = 0;
805 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
806 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
809 if (op_data->op_bias & MDS_CLOSE_INTENT) {
810 req_fmt = &RQF_MDS_CLOSE_INTENT;
811 if (op_data->op_bias & MDS_HSM_RELEASE) {
812 /* allocate a FID for volatile file */
813 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
816 CERROR("%s: "DFID" allocating FID: rc = %d\n",
817 obd->obd_name, PFID(&op_data->op_fid1),
819 /* save the errcode and proceed to close */
823 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
824 size_t count = op_data->op_data_size / sizeof(__u32);
826 if (count > INLINE_RESYNC_ARRAY_SIZE)
830 req_fmt = &RQF_MDS_CLOSE;
834 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
837 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
839 /* Ensure that this close's handle is fixed up during replay. */
840 if (likely(mod != NULL)) {
841 LASSERTF(mod->mod_open_req != NULL &&
842 mod->mod_open_req->rq_type != LI_POISON,
843 "POISONED open %p!\n", mod->mod_open_req);
845 mod->mod_close_req = req;
847 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
848 /* We no longer want to preserve this open for replay even
849 * though the open was committed. b=3632, b=3633 */
850 spin_lock(&mod->mod_open_req->rq_lock);
851 mod->mod_open_req->rq_replay = 0;
852 spin_unlock(&mod->mod_open_req->rq_lock);
854 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
858 * TODO: repeat close after errors
860 CWARN("%s: close of FID "DFID" failed, file reference will be "
861 "dropped when this client unmounts or is evicted\n",
862 obd->obd_name, PFID(&op_data->op_fid1));
863 GOTO(out, rc = -ENOMEM);
867 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
868 u32_count * sizeof(__u32));
870 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
872 ptlrpc_request_free(req);
877 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
878 * portal whose threads are not taking any DLM locks and are therefore
879 * always progressing */
880 req->rq_request_portal = MDS_READPAGE_PORTAL;
881 ptlrpc_at_set_req_timeout(req);
884 mdc_close_pack(req, op_data);
886 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
887 obd->u.cli.cl_default_mds_easize);
889 ptlrpc_request_set_replen(req);
891 mdc_get_mod_rpc_slot(req, NULL);
892 rc = ptlrpc_queue_wait(req);
893 mdc_put_mod_rpc_slot(req, NULL);
895 if (req->rq_repmsg == NULL) {
896 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
899 rc = req->rq_status ?: -EIO;
900 } else if (rc == 0 || rc == -EAGAIN) {
901 struct mdt_body *body;
903 rc = lustre_msg_get_status(req->rq_repmsg);
904 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
905 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
910 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
913 } else if (rc == -ESTALE) {
915 * it can be allowed error after 3633 if open was committed and
916 * server failed before close was sent. Let's check if mod
917 * exists and return no error in that case
920 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
921 LASSERT(mod->mod_open_req != NULL);
922 if (mod->mod_open_req->rq_committed)
930 mod->mod_close_req = NULL;
931 /* Since now, mod is accessed through open_req only,
932 * thus close req does not keep a reference on mod anymore. */
937 RETURN(rc < 0 ? rc : saved_rc);
940 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
941 u64 offset, struct page **pages, int npages,
942 struct ptlrpc_request **request)
944 struct ptlrpc_request *req;
945 struct ptlrpc_bulk_desc *desc;
947 wait_queue_head_t waitq;
949 struct l_wait_info lwi;
954 init_waitqueue_head(&waitq);
957 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
961 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
963 ptlrpc_request_free(req);
967 req->rq_request_portal = MDS_READPAGE_PORTAL;
968 ptlrpc_at_set_req_timeout(req);
970 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
971 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
973 &ptlrpc_bulk_kiov_pin_ops);
975 ptlrpc_req_finished(req);
979 /* NB req now owns desc and will free it when it gets freed */
980 for (i = 0; i < npages; i++)
981 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
984 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
986 ptlrpc_request_set_replen(req);
987 rc = ptlrpc_queue_wait(req);
989 ptlrpc_req_finished(req);
990 if (rc != -ETIMEDOUT)
994 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
995 CERROR("%s: too many resend retries: rc = %d\n",
996 exp->exp_obd->obd_name, -EIO);
999 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1001 l_wait_event(waitq, 0, &lwi);
1006 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1007 req->rq_bulk->bd_nob_transferred);
1009 ptlrpc_req_finished(req);
1013 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1014 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1015 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1016 PAGE_SIZE * npages);
1017 ptlrpc_req_finished(req);
1025 static void mdc_release_page(struct page *page, int remove)
1029 if (likely(page->mapping != NULL))
1030 truncate_complete_page(page->mapping, page);
1036 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1037 __u64 *start, __u64 *end, int hash64)
1040 * Complement of hash is used as an index so that
1041 * radix_tree_gang_lookup() can be used to find a page with starting
1042 * hash _smaller_ than one we are looking for.
1044 unsigned long offset = hash_x_index(*hash, hash64);
1048 spin_lock_irq(&mapping->tree_lock);
1049 found = radix_tree_gang_lookup(&mapping->page_tree,
1050 (void **)&page, offset, 1);
1051 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1052 struct lu_dirpage *dp;
1055 spin_unlock_irq(&mapping->tree_lock);
1057 * In contrast to find_lock_page() we are sure that directory
1058 * page cannot be truncated (while DLM lock is held) and,
1059 * hence, can avoid restart.
1061 * In fact, page cannot be locked here at all, because
1062 * mdc_read_page_remote does synchronous io.
1064 wait_on_page_locked(page);
1065 if (PageUptodate(page)) {
1067 if (BITS_PER_LONG == 32 && hash64) {
1068 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1069 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1070 *hash = *hash >> 32;
1072 *start = le64_to_cpu(dp->ldp_hash_start);
1073 *end = le64_to_cpu(dp->ldp_hash_end);
1075 if (unlikely(*start == 1 && *hash == 0))
1078 LASSERTF(*start <= *hash, "start = %#llx"
1079 ",end = %#llx,hash = %#llx\n",
1080 *start, *end, *hash);
1081 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1082 " hash %#llx\n", offset, *start, *end, *hash);
1085 mdc_release_page(page, 0);
1087 } else if (*end != *start && *hash == *end) {
1089 * upon hash collision, remove this page,
1090 * otherwise put page reference, and
1091 * mdc_read_page_remote() will issue RPC to
1092 * fetch the page we want.
1095 mdc_release_page(page,
1096 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1101 page = ERR_PTR(-EIO);
1104 spin_unlock_irq(&mapping->tree_lock);
1111 * Adjust a set of pages, each page containing an array of lu_dirpages,
1112 * so that each page can be used as a single logical lu_dirpage.
1114 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1115 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1116 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1117 * value is used as a cookie to request the next lu_dirpage in a
1118 * directory listing that spans multiple pages (two in this example):
1121 * .|--------v------- -----.
1122 * |s|e|f|p|ent|ent| ... |ent|
1123 * '--|-------------- -----' Each PAGE contains a single
1124 * '------. lu_dirpage.
1125 * .---------v------- -----.
1126 * |s|e|f|p|ent| 0 | ... | 0 |
1127 * '----------------- -----'
1129 * However, on hosts where the native VM page size (PAGE_SIZE) is
1130 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1131 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1132 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1133 * after it in the same PAGE (arrows simplified for brevity, but
1134 * in general e0==s1, e1==s2, etc.):
1136 * .-------------------- -----.
1137 * |s0|e0|f0|p|ent|ent| ... |ent|
1138 * |---v---------------- -----|
1139 * |s1|e1|f1|p|ent|ent| ... |ent|
1140 * |---v---------------- -----| Here, each PAGE contains
1141 * ... multiple lu_dirpages.
1142 * |---v---------------- -----|
1143 * |s'|e'|f'|p|ent|ent| ... |ent|
1144 * '---|---------------- -----'
1146 * .----------------------------.
1149 * This structure is transformed into a single logical lu_dirpage as follows:
1151 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1152 * labeled 'next PAGE'.
1154 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1155 * a hash collision with the next page exists.
1157 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1158 * to the first entry of the next lu_dirpage.
1160 #if PAGE_SIZE > LU_PAGE_SIZE
1161 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1165 for (i = 0; i < cfs_pgs; i++) {
1166 struct lu_dirpage *dp = kmap(pages[i]);
1167 struct lu_dirpage *first = dp;
1168 struct lu_dirent *end_dirent = NULL;
1169 struct lu_dirent *ent;
1170 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1171 __u32 flags = le32_to_cpu(dp->ldp_flags);
1173 while (--lu_pgs > 0) {
1174 ent = lu_dirent_start(dp);
1175 for (end_dirent = ent; ent != NULL;
1176 end_dirent = ent, ent = lu_dirent_next(ent));
1178 /* Advance dp to next lu_dirpage. */
1179 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1181 /* Check if we've reached the end of the PAGE. */
1182 if (!((unsigned long)dp & ~PAGE_MASK))
1185 /* Save the hash and flags of this lu_dirpage. */
1186 hash_end = le64_to_cpu(dp->ldp_hash_end);
1187 flags = le32_to_cpu(dp->ldp_flags);
1189 /* Check if lu_dirpage contains no entries. */
1190 if (end_dirent == NULL)
1193 /* Enlarge the end entry lde_reclen from 0 to
1194 * first entry of next lu_dirpage. */
1195 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1196 end_dirent->lde_reclen =
1197 cpu_to_le16((char *)(dp->ldp_entries) -
1198 (char *)end_dirent);
1201 first->ldp_hash_end = hash_end;
1202 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1203 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1207 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1210 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1211 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1213 /* parameters for readdir page */
1214 struct readpage_param {
1215 struct md_op_data *rp_mod;
1218 struct obd_export *rp_exp;
1219 struct md_callback *rp_cb;
1222 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1223 static inline void delete_from_page_cache(struct page *page)
1225 remove_from_page_cache(page);
1231 * Read pages from server.
1233 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1234 * a header lu_dirpage which describes the start/end hash, and whether this
1235 * page is empty (contains no dir entry) or hash collide with next page.
1236 * After client receives reply, several pages will be integrated into dir page
1237 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1238 * lu_dirpage for this integrated page will be adjusted.
1240 static int mdc_read_page_remote(void *data, struct page *page0)
1242 struct readpage_param *rp = data;
1243 struct page **page_pool;
1245 struct lu_dirpage *dp;
1246 struct md_op_data *op_data = rp->rp_mod;
1247 struct ptlrpc_request *req;
1249 struct inode *inode;
1251 int rd_pgs = 0; /* number of pages actually read */
1257 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1258 inode = op_data->op_data;
1259 fid = &op_data->op_fid1;
1260 LASSERT(inode != NULL);
1262 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1263 if (page_pool != NULL) {
1264 page_pool[0] = page0;
1270 for (npages = 1; npages < max_pages; npages++) {
1271 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1275 page_pool[npages] = page;
1278 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1280 /* page0 is special, which was added into page cache early */
1281 delete_from_page_cache(page0);
1285 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1287 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1288 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1290 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1292 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1294 SetPageUptodate(page0);
1298 ptlrpc_req_finished(req);
1299 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1300 for (i = 1; i < npages; i++) {
1301 unsigned long offset;
1305 page = page_pool[i];
1307 if (rc < 0 || i >= rd_pgs) {
1312 SetPageUptodate(page);
1315 hash = le64_to_cpu(dp->ldp_hash_start);
1318 offset = hash_x_index(hash, rp->rp_hash64);
1320 prefetchw(&page->flags);
1321 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1326 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1327 " rc = %d\n", offset, ret);
1331 if (page_pool != &page0)
1332 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1338 * Read dir page from cache first, if it can not find it, read it from
1339 * server and add into the cache.
1341 * \param[in] exp MDC export
1342 * \param[in] op_data client MD stack parameters, transfering parameters
1343 * between different layers on client MD stack.
1344 * \param[in] cb_op callback required for ldlm lock enqueue during
1346 * \param[in] hash_offset the hash offset of the page to be read
1347 * \param[in] ppage the page to be read
1349 * retval = 0 get the page successfully
1350 * errno(<0) get the page failed
1352 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1353 struct md_callback *cb_op, __u64 hash_offset,
1354 struct page **ppage)
1356 struct lookup_intent it = { .it_op = IT_READDIR };
1358 struct inode *dir = op_data->op_data;
1359 struct address_space *mapping;
1360 struct lu_dirpage *dp;
1363 struct lustre_handle lockh;
1364 struct ptlrpc_request *enq_req = NULL;
1365 struct readpage_param rp_param;
1372 LASSERT(dir != NULL);
1373 mapping = dir->i_mapping;
1375 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1376 cb_op->md_blocking_ast, 0);
1377 if (enq_req != NULL)
1378 ptlrpc_req_finished(enq_req);
1381 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1382 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1387 lockh.cookie = it.it_lock_handle;
1388 mdc_set_lock_data(exp, &lockh, dir, NULL);
1390 rp_param.rp_off = hash_offset;
1391 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1392 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1393 rp_param.rp_hash64);
1395 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1396 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1397 rp_param.rp_off, PTR_ERR(page));
1398 GOTO(out_unlock, rc = PTR_ERR(page));
1399 } else if (page != NULL) {
1401 * XXX nikita: not entirely correct handling of a corner case:
1402 * suppose hash chain of entries with hash value HASH crosses
1403 * border between pages P0 and P1. First both P0 and P1 are
1404 * cached, seekdir() is called for some entry from the P0 part
1405 * of the chain. Later P0 goes out of cache. telldir(HASH)
1406 * happens and finds P1, as it starts with matching hash
1407 * value. Remaining entries from P0 part of the chain are
1408 * skipped. (Is that really a bug?)
1410 * Possible solutions: 0. don't cache P1 is such case, handle
1411 * it as an "overflow" page. 1. invalidate all pages at
1412 * once. 2. use HASH|1 as an index for P1.
1414 GOTO(hash_collision, page);
1417 rp_param.rp_exp = exp;
1418 rp_param.rp_mod = op_data;
1419 page = read_cache_page(mapping,
1420 hash_x_index(rp_param.rp_off,
1421 rp_param.rp_hash64),
1422 mdc_read_page_remote, &rp_param);
1424 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1425 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1426 rp_param.rp_off, PTR_ERR(page));
1427 GOTO(out_unlock, rc = PTR_ERR(page));
1430 wait_on_page_locked(page);
1432 if (!PageUptodate(page)) {
1433 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1434 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1435 rp_param.rp_off, -5);
1438 if (!PageChecked(page))
1439 SetPageChecked(page);
1440 if (PageError(page)) {
1441 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1442 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1443 rp_param.rp_off, -5);
1448 dp = page_address(page);
1449 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1450 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1451 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1452 rp_param.rp_off = hash_offset >> 32;
1454 start = le64_to_cpu(dp->ldp_hash_start);
1455 end = le64_to_cpu(dp->ldp_hash_end);
1456 rp_param.rp_off = hash_offset;
1459 LASSERT(start == rp_param.rp_off);
1460 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1461 #if BITS_PER_LONG == 32
1462 CWARN("Real page-wide hash collision at [%llu %llu] with "
1463 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1464 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1468 * Fetch whole overflow chain...
1476 ldlm_lock_decref(&lockh, it.it_lock_mode);
1480 mdc_release_page(page, 1);
1485 static int mdc_statfs(const struct lu_env *env,
1486 struct obd_export *exp, struct obd_statfs *osfs,
1487 time64_t max_age, __u32 flags)
1489 struct obd_device *obd = class_exp2obd(exp);
1490 struct ptlrpc_request *req;
1491 struct obd_statfs *msfs;
1492 struct obd_import *imp = NULL;
1497 * Since the request might also come from lprocfs, so we need
1498 * sync this with client_disconnect_export Bug15684
1500 down_read(&obd->u.cli.cl_sem);
1501 if (obd->u.cli.cl_import)
1502 imp = class_import_get(obd->u.cli.cl_import);
1503 up_read(&obd->u.cli.cl_sem);
1507 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_STATFS,
1508 LUSTRE_MDS_VERSION, MDS_STATFS);
1510 GOTO(output, rc = -ENOMEM);
1512 if ((flags & OBD_STATFS_SUM) &&
1513 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1514 /* request aggregated states */
1515 struct mdt_body *body;
1517 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1519 GOTO(out, rc = -EPROTO);
1520 body->mbo_valid = OBD_MD_FLAGSTATFS;
1523 ptlrpc_request_set_replen(req);
1525 if (flags & OBD_STATFS_NODELAY) {
1526 /* procfs requests not want stay in wait for avoid deadlock */
1527 req->rq_no_resend = 1;
1528 req->rq_no_delay = 1;
1531 rc = ptlrpc_queue_wait(req);
1533 /* check connection error first */
1534 if (imp->imp_connect_error)
1535 rc = imp->imp_connect_error;
1539 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1541 GOTO(out, rc = -EPROTO);
1546 ptlrpc_req_finished(req);
1548 class_import_put(imp);
1552 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1554 __u32 keylen, vallen;
1558 if (gf->gf_pathlen > PATH_MAX)
1559 RETURN(-ENAMETOOLONG);
1560 if (gf->gf_pathlen < 2)
1563 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1564 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1565 sizeof(struct lu_fid));
1566 OBD_ALLOC(key, keylen);
1569 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1570 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1571 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1572 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1573 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1574 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1576 if (!fid_is_sane(&gf->gf_fid))
1577 GOTO(out, rc = -EINVAL);
1579 /* Val is struct getinfo_fid2path result plus path */
1580 vallen = sizeof(*gf) + gf->gf_pathlen;
1582 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1583 if (rc != 0 && rc != -EREMOTE)
1586 if (vallen <= sizeof(*gf))
1587 GOTO(out, rc = -EPROTO);
1588 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1589 GOTO(out, rc = -EOVERFLOW);
1591 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1592 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1593 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1594 /* only log the last 512 characters of the path */
1595 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1598 OBD_FREE(key, keylen);
1602 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1603 struct hsm_progress_kernel *hpk)
1605 struct obd_import *imp = class_exp2cliimp(exp);
1606 struct hsm_progress_kernel *req_hpk;
1607 struct ptlrpc_request *req;
1611 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1612 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1614 GOTO(out, rc = -ENOMEM);
1616 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1618 /* Copy hsm_progress struct */
1619 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1620 if (req_hpk == NULL)
1621 GOTO(out, rc = -EPROTO);
1624 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1626 ptlrpc_request_set_replen(req);
1628 mdc_get_mod_rpc_slot(req, NULL);
1629 rc = ptlrpc_queue_wait(req);
1630 mdc_put_mod_rpc_slot(req, NULL);
1634 ptlrpc_req_finished(req);
1638 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archives)
1640 __u32 *archive_mask;
1641 struct ptlrpc_request *req;
1645 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_REGISTER,
1647 MDS_HSM_CT_REGISTER);
1649 GOTO(out, rc = -ENOMEM);
1651 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1653 /* Copy hsm_progress struct */
1654 archive_mask = req_capsule_client_get(&req->rq_pill,
1655 &RMF_MDS_HSM_ARCHIVE);
1656 if (archive_mask == NULL)
1657 GOTO(out, rc = -EPROTO);
1659 *archive_mask = archives;
1661 ptlrpc_request_set_replen(req);
1663 rc = mdc_queue_wait(req);
1666 ptlrpc_req_finished(req);
1670 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1671 struct md_op_data *op_data)
1673 struct hsm_current_action *hca = op_data->op_data;
1674 struct hsm_current_action *req_hca;
1675 struct ptlrpc_request *req;
1679 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1680 &RQF_MDS_HSM_ACTION);
1684 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1686 ptlrpc_request_free(req);
1690 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1691 op_data->op_suppgids[0], 0);
1693 ptlrpc_request_set_replen(req);
1695 rc = mdc_queue_wait(req);
1699 req_hca = req_capsule_server_get(&req->rq_pill,
1700 &RMF_MDS_HSM_CURRENT_ACTION);
1701 if (req_hca == NULL)
1702 GOTO(out, rc = -EPROTO);
1708 ptlrpc_req_finished(req);
1712 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1714 struct ptlrpc_request *req;
1718 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1720 MDS_HSM_CT_UNREGISTER);
1722 GOTO(out, rc = -ENOMEM);
1724 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1726 ptlrpc_request_set_replen(req);
1728 rc = mdc_queue_wait(req);
1731 ptlrpc_req_finished(req);
1735 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1736 struct md_op_data *op_data)
1738 struct hsm_user_state *hus = op_data->op_data;
1739 struct hsm_user_state *req_hus;
1740 struct ptlrpc_request *req;
1744 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1745 &RQF_MDS_HSM_STATE_GET);
1749 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1751 ptlrpc_request_free(req);
1755 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1756 op_data->op_suppgids[0], 0);
1758 ptlrpc_request_set_replen(req);
1760 rc = mdc_queue_wait(req);
1764 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1765 if (req_hus == NULL)
1766 GOTO(out, rc = -EPROTO);
1772 ptlrpc_req_finished(req);
1776 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1777 struct md_op_data *op_data)
1779 struct hsm_state_set *hss = op_data->op_data;
1780 struct hsm_state_set *req_hss;
1781 struct ptlrpc_request *req;
1785 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1786 &RQF_MDS_HSM_STATE_SET);
1790 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1792 ptlrpc_request_free(req);
1796 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1797 op_data->op_suppgids[0], 0);
1800 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1801 if (req_hss == NULL)
1802 GOTO(out, rc = -EPROTO);
1805 ptlrpc_request_set_replen(req);
1807 mdc_get_mod_rpc_slot(req, NULL);
1808 rc = ptlrpc_queue_wait(req);
1809 mdc_put_mod_rpc_slot(req, NULL);
1813 ptlrpc_req_finished(req);
1817 static int mdc_ioc_hsm_request(struct obd_export *exp,
1818 struct hsm_user_request *hur)
1820 struct obd_import *imp = class_exp2cliimp(exp);
1821 struct ptlrpc_request *req;
1822 struct hsm_request *req_hr;
1823 struct hsm_user_item *req_hui;
1828 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1830 GOTO(out, rc = -ENOMEM);
1832 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1833 hur->hur_request.hr_itemcount
1834 * sizeof(struct hsm_user_item));
1835 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1836 hur->hur_request.hr_data_len);
1838 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1840 ptlrpc_request_free(req);
1844 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1846 /* Copy hsm_request struct */
1847 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1849 GOTO(out, rc = -EPROTO);
1850 *req_hr = hur->hur_request;
1852 /* Copy hsm_user_item structs */
1853 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1854 if (req_hui == NULL)
1855 GOTO(out, rc = -EPROTO);
1856 memcpy(req_hui, hur->hur_user_item,
1857 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1859 /* Copy opaque field */
1860 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1861 if (req_opaque == NULL)
1862 GOTO(out, rc = -EPROTO);
1863 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1865 ptlrpc_request_set_replen(req);
1867 mdc_get_mod_rpc_slot(req, NULL);
1868 rc = ptlrpc_queue_wait(req);
1869 mdc_put_mod_rpc_slot(req, NULL);
1874 ptlrpc_req_finished(req);
1878 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1879 struct lustre_kernelcomm *lk);
1881 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1882 struct obd_quotactl *oqctl)
1884 struct ptlrpc_request *req;
1885 struct obd_quotactl *oqc;
1889 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1890 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1895 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1898 ptlrpc_request_set_replen(req);
1899 ptlrpc_at_set_req_timeout(req);
1901 rc = ptlrpc_queue_wait(req);
1903 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1905 if (req->rq_repmsg &&
1906 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1909 CERROR ("Can't unpack obd_quotactl\n");
1912 ptlrpc_req_finished(req);
1917 static int mdc_ioc_swap_layouts(struct obd_export *exp,
1918 struct md_op_data *op_data)
1920 struct list_head cancels = LIST_HEAD_INIT(cancels);
1921 struct ptlrpc_request *req;
1923 struct mdc_swap_layouts *msl, *payload;
1926 msl = op_data->op_data;
1928 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
1929 * first thing it will do is to cancel the 2 layout
1930 * locks held by this client.
1931 * So the client must cancel its layout locks on the 2 fids
1932 * with the request RPC to avoid extra RPC round trips.
1934 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
1935 LCK_EX, MDS_INODELOCK_LAYOUT |
1936 MDS_INODELOCK_XATTR);
1937 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
1938 LCK_EX, MDS_INODELOCK_LAYOUT |
1939 MDS_INODELOCK_XATTR);
1941 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1942 &RQF_MDS_SWAP_LAYOUTS);
1944 ldlm_lock_list_put(&cancels, l_bl_ast, count);
1948 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
1950 ptlrpc_request_free(req);
1954 mdc_swap_layouts_pack(req, op_data);
1956 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
1961 ptlrpc_request_set_replen(req);
1963 rc = ptlrpc_queue_wait(req);
1969 ptlrpc_req_finished(req);
1973 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1974 void *karg, void __user *uarg)
1976 struct obd_device *obd = exp->exp_obd;
1977 struct obd_ioctl_data *data = karg;
1978 struct obd_import *imp = obd->u.cli.cl_import;
1982 if (!try_module_get(THIS_MODULE)) {
1983 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1984 module_name(THIS_MODULE));
1988 case OBD_IOC_FID2PATH:
1989 rc = mdc_ioc_fid2path(exp, karg);
1991 case LL_IOC_HSM_CT_START:
1992 rc = mdc_ioc_hsm_ct_start(exp, karg);
1993 /* ignore if it was already registered on this MDS. */
1997 case LL_IOC_HSM_PROGRESS:
1998 rc = mdc_ioc_hsm_progress(exp, karg);
2000 case LL_IOC_HSM_STATE_GET:
2001 rc = mdc_ioc_hsm_state_get(exp, karg);
2003 case LL_IOC_HSM_STATE_SET:
2004 rc = mdc_ioc_hsm_state_set(exp, karg);
2006 case LL_IOC_HSM_ACTION:
2007 rc = mdc_ioc_hsm_current_action(exp, karg);
2009 case LL_IOC_HSM_REQUEST:
2010 rc = mdc_ioc_hsm_request(exp, karg);
2012 case OBD_IOC_CLIENT_RECOVER:
2013 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2017 case IOC_OSC_SET_ACTIVE:
2018 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2020 case OBD_IOC_PING_TARGET:
2021 rc = ptlrpc_obd_ping(obd);
2024 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2025 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2026 * there'd be no LMV layer thus we might be called here. Eventually
2027 * this code should be removed.
2030 case IOC_OBD_STATFS: {
2031 struct obd_statfs stat_buf = {0};
2033 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2034 GOTO(out, rc = -ENODEV);
2037 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2038 min((int)data->ioc_plen2,
2039 (int)sizeof(struct obd_uuid))))
2040 GOTO(out, rc = -EFAULT);
2042 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2043 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2048 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2049 min((int) data->ioc_plen1,
2050 (int) sizeof(stat_buf))))
2051 GOTO(out, rc = -EFAULT);
2055 case OBD_IOC_QUOTACTL: {
2056 struct if_quotactl *qctl = karg;
2057 struct obd_quotactl *oqctl;
2059 OBD_ALLOC_PTR(oqctl);
2061 GOTO(out, rc = -ENOMEM);
2063 QCTL_COPY(oqctl, qctl);
2064 rc = obd_quotactl(exp, oqctl);
2066 QCTL_COPY(qctl, oqctl);
2067 qctl->qc_valid = QC_MDTIDX;
2068 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2071 OBD_FREE_PTR(oqctl);
2074 case LL_IOC_GET_CONNECT_FLAGS:
2075 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2076 sizeof(*exp_connect_flags_ptr(exp))))
2077 GOTO(out, rc = -EFAULT);
2080 case LL_IOC_LOV_SWAP_LAYOUTS:
2081 rc = mdc_ioc_swap_layouts(exp, karg);
2084 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2085 GOTO(out, rc = -ENOTTY);
2088 module_put(THIS_MODULE);
2093 static int mdc_get_info_rpc(struct obd_export *exp,
2094 u32 keylen, void *key,
2095 u32 vallen, void *val)
2097 struct obd_import *imp = class_exp2cliimp(exp);
2098 struct ptlrpc_request *req;
2103 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2107 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2108 RCL_CLIENT, keylen);
2109 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2110 RCL_CLIENT, sizeof(vallen));
2112 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2114 ptlrpc_request_free(req);
2118 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2119 memcpy(tmp, key, keylen);
2120 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2121 memcpy(tmp, &vallen, sizeof(vallen));
2123 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2124 RCL_SERVER, vallen);
2125 ptlrpc_request_set_replen(req);
2127 rc = ptlrpc_queue_wait(req);
2128 /* -EREMOTE means the get_info result is partial, and it needs to
2129 * continue on another MDT, see fid2path part in lmv_iocontrol */
2130 if (rc == 0 || rc == -EREMOTE) {
2131 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2132 memcpy(val, tmp, vallen);
2133 if (ptlrpc_rep_need_swab(req)) {
2134 if (KEY_IS(KEY_FID2PATH))
2135 lustre_swab_fid2path(val);
2138 ptlrpc_req_finished(req);
2143 static void lustre_swab_hai(struct hsm_action_item *h)
2145 __swab32s(&h->hai_len);
2146 __swab32s(&h->hai_action);
2147 lustre_swab_lu_fid(&h->hai_fid);
2148 lustre_swab_lu_fid(&h->hai_dfid);
2149 __swab64s(&h->hai_cookie);
2150 __swab64s(&h->hai_extent.offset);
2151 __swab64s(&h->hai_extent.length);
2152 __swab64s(&h->hai_gid);
2155 static void lustre_swab_hal(struct hsm_action_list *h)
2157 struct hsm_action_item *hai;
2160 __swab32s(&h->hal_version);
2161 __swab32s(&h->hal_count);
2162 __swab32s(&h->hal_archive_id);
2163 __swab64s(&h->hal_flags);
2165 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2166 lustre_swab_hai(hai);
2169 static void lustre_swab_kuch(struct kuc_hdr *l)
2171 __swab16s(&l->kuc_magic);
2172 /* __u8 l->kuc_transport */
2173 __swab16s(&l->kuc_msgtype);
2174 __swab16s(&l->kuc_msglen);
2177 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2178 struct lustre_kernelcomm *lk)
2180 struct obd_import *imp = class_exp2cliimp(exp);
2181 __u32 archive = lk->lk_data;
2184 if (lk->lk_group != KUC_GRP_HSM) {
2185 CERROR("Bad copytool group %d\n", lk->lk_group);
2189 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2190 lk->lk_uid, lk->lk_group, lk->lk_flags);
2192 if (lk->lk_flags & LK_FLG_STOP) {
2193 /* Unregister with the coordinator */
2194 rc = mdc_ioc_hsm_ct_unregister(imp);
2196 rc = mdc_ioc_hsm_ct_register(imp, archive);
2203 * Send a message to any listening copytools
2204 * @param val KUC message (kuc_hdr + hsm_action_list)
2205 * @param len total length of message
2207 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2208 size_t len, void *val)
2210 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2211 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2215 if (len < sizeof(*lh) + sizeof(*hal)) {
2216 CERROR("Short HSM message %zu < %zu\n", len,
2217 sizeof(*lh) + sizeof(*hal));
2220 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2221 lustre_swab_kuch(lh);
2222 lustre_swab_hal(hal);
2223 } else if (lh->kuc_magic != KUC_MAGIC) {
2224 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2228 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2230 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2231 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2233 /* Broadcast to HSM listeners */
2234 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2240 * callback function passed to kuc for re-registering each HSM copytool
2241 * running on MDC, after MDT shutdown/recovery.
2242 * @param data copytool registration data
2243 * @param cb_arg callback argument (obd_import)
2245 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2247 struct kkuc_ct_data *kcd = data;
2248 struct obd_import *imp = (struct obd_import *)cb_arg;
2251 if (kcd == NULL || kcd->kcd_magic != KKUC_CT_DATA_MAGIC)
2254 CDEBUG(D_HA, "%s: recover copytool registration to MDT (archive=%#x)\n",
2255 imp->imp_obd->obd_name, kcd->kcd_archive);
2256 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_archive);
2258 /* ignore error if the copytool is already registered */
2259 return (rc == -EEXIST) ? 0 : rc;
2263 * Re-establish all kuc contexts with MDT
2264 * after MDT shutdown/recovery.
2266 static int mdc_kuc_reregister(struct obd_import *imp)
2268 /* re-register HSM agents */
2269 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2270 mdc_hsm_ct_reregister, imp);
2273 static int mdc_set_info_async(const struct lu_env *env,
2274 struct obd_export *exp,
2275 u32 keylen, void *key,
2276 u32 vallen, void *val,
2277 struct ptlrpc_request_set *set)
2279 struct obd_import *imp = class_exp2cliimp(exp);
2283 if (KEY_IS(KEY_READ_ONLY)) {
2284 if (vallen != sizeof(int))
2287 spin_lock(&imp->imp_lock);
2288 if (*((int *)val)) {
2289 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2290 imp->imp_connect_data.ocd_connect_flags |=
2293 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2294 imp->imp_connect_data.ocd_connect_flags &=
2295 ~OBD_CONNECT_RDONLY;
2297 spin_unlock(&imp->imp_lock);
2299 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2300 keylen, key, vallen, val, set);
2303 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2304 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2305 keylen, key, vallen, val, set);
2308 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2309 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2314 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2315 __u32 *default_easize = val;
2317 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2321 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2325 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2326 __u32 keylen, void *key, __u32 *vallen, void *val)
2330 if (KEY_IS(KEY_MAX_EASIZE)) {
2331 __u32 mdsize, *max_easize;
2333 if (*vallen != sizeof(int))
2335 mdsize = *(__u32 *)val;
2336 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2337 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2339 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2341 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2342 __u32 *default_easize;
2344 if (*vallen != sizeof(int))
2346 default_easize = val;
2347 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2349 } else if (KEY_IS(KEY_CONN_DATA)) {
2350 struct obd_import *imp = class_exp2cliimp(exp);
2351 struct obd_connect_data *data = val;
2353 if (*vallen != sizeof(*data))
2356 *data = imp->imp_connect_data;
2358 } else if (KEY_IS(KEY_TGT_COUNT)) {
2359 *((__u32 *)val) = 1;
2363 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2368 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2369 struct ptlrpc_request **request)
2371 struct ptlrpc_request *req;
2376 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2380 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2382 ptlrpc_request_free(req);
2386 mdc_pack_body(req, fid, 0, 0, -1, 0);
2388 ptlrpc_request_set_replen(req);
2390 rc = ptlrpc_queue_wait(req);
2392 ptlrpc_req_finished(req);
2398 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2399 enum obd_import_event event)
2401 struct client_obd *cli = &obd->u.cli;
2404 LASSERT(imp->imp_obd == obd);
2407 case IMP_EVENT_DISCON:
2408 spin_lock(&cli->cl_loi_list_lock);
2409 cli->cl_avail_grant = 0;
2410 cli->cl_lost_grant = 0;
2411 spin_unlock(&cli->cl_loi_list_lock);
2413 case IMP_EVENT_INACTIVE:
2415 * Flush current sequence to make client obtain new one
2416 * from server in case of disconnect/reconnect.
2418 down_read(&cli->cl_seq_rwsem);
2420 seq_client_flush(cli->cl_seq);
2421 up_read(&cli->cl_seq_rwsem);
2423 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2425 case IMP_EVENT_INVALIDATE: {
2426 struct ldlm_namespace *ns = obd->obd_namespace;
2430 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2432 env = cl_env_get(&refcheck);
2434 /* Reset grants. All pages go to failing rpcs due to
2435 * the invalid import.
2437 osc_io_unplug(env, cli, NULL);
2439 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2440 osc_ldlm_resource_invalidate,
2442 cl_env_put(env, &refcheck);
2443 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2449 case IMP_EVENT_ACTIVE:
2450 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2451 /* redo the kuc registration after reconnecting */
2453 rc = mdc_kuc_reregister(imp);
2455 case IMP_EVENT_OCD: {
2456 struct obd_connect_data *ocd = &imp->imp_connect_data;
2458 if (OCD_HAS_FLAG(ocd, GRANT))
2459 osc_init_grant(cli, ocd);
2461 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2464 case IMP_EVENT_DEACTIVATE:
2465 case IMP_EVENT_ACTIVATE:
2468 CERROR("Unknown import event %x\n", event);
2474 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2475 struct lu_fid *fid, struct md_op_data *op_data)
2477 struct client_obd *cli = &exp->exp_obd->u.cli;
2482 down_read(&cli->cl_seq_rwsem);
2484 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2485 up_read(&cli->cl_seq_rwsem);
2490 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2492 struct client_obd *cli = &exp->exp_obd->u.cli;
2493 return &cli->cl_target_uuid;
2497 * Determine whether the lock can be canceled before replaying it during
2498 * recovery, non zero value will be return if the lock can be canceled,
2499 * or zero returned for not
2501 static int mdc_cancel_weight(struct ldlm_lock *lock)
2503 if (lock->l_resource->lr_type != LDLM_IBITS)
2506 /* FIXME: if we ever get into a situation where there are too many
2507 * opened files with open locks on a single node, then we really
2508 * should replay these open locks to reget it */
2509 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2512 /* Special case for DoM locks, cancel only unused and granted locks */
2513 if (ldlm_has_dom(lock) &&
2514 (lock->l_granted_mode != lock->l_req_mode ||
2515 osc_ldlm_weigh_ast(lock) != 0))
2521 static int mdc_resource_inode_free(struct ldlm_resource *res)
2523 if (res->lr_lvb_inode)
2524 res->lr_lvb_inode = NULL;
2529 static struct ldlm_valblock_ops inode_lvbo = {
2530 .lvbo_free = mdc_resource_inode_free
2533 static int mdc_llog_init(struct obd_device *obd)
2535 struct obd_llog_group *olg = &obd->obd_olg;
2536 struct llog_ctxt *ctxt;
2541 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2546 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2547 llog_initiator_connect(ctxt);
2548 llog_ctxt_put(ctxt);
2553 static void mdc_llog_finish(struct obd_device *obd)
2555 struct llog_ctxt *ctxt;
2559 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2561 llog_cleanup(NULL, ctxt);
2566 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2572 rc = osc_setup_common(obd, cfg);
2576 rc = mdc_tunables_init(obd);
2578 GOTO(err_osc_cleanup, rc);
2580 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2582 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2584 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2586 rc = mdc_llog_init(obd);
2588 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2590 GOTO(err_llog_cleanup, rc);
2593 rc = mdc_changelog_cdev_init(obd);
2595 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2597 GOTO(err_changelog_cleanup, rc);
2602 err_changelog_cleanup:
2603 mdc_llog_finish(obd);
2605 lprocfs_free_md_stats(obd);
2606 ptlrpc_lprocfs_unregister_obd(obd);
2608 osc_cleanup_common(obd);
2612 /* Initialize the default and maximum LOV EA sizes. This allows
2613 * us to make MDS RPCs with large enough reply buffers to hold a default
2614 * sized EA without having to calculate this (via a call into the
2615 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2616 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2617 * a large number of stripes is possible. If a larger reply buffer is
2618 * required it will be reallocated in the ptlrpc layer due to overflow.
2620 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2623 struct obd_device *obd = exp->exp_obd;
2624 struct client_obd *cli = &obd->u.cli;
2627 if (cli->cl_max_mds_easize < easize)
2628 cli->cl_max_mds_easize = easize;
2630 if (cli->cl_default_mds_easize < def_easize)
2631 cli->cl_default_mds_easize = def_easize;
2636 static int mdc_precleanup(struct obd_device *obd)
2640 osc_precleanup_common(obd);
2641 mdc_changelog_cdev_finish(obd);
2643 obd_cleanup_client_import(obd);
2644 ptlrpc_lprocfs_unregister_obd(obd);
2645 lprocfs_free_md_stats(obd);
2646 mdc_llog_finish(obd);
2650 static int mdc_cleanup(struct obd_device *obd)
2652 return osc_cleanup_common(obd);
2655 int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2657 struct lustre_cfg *lcfg = buf;
2658 size_t count = class_modify_config(lcfg, PARAM_MDC,
2659 &obd->obd_kset.kobj);
2661 return count > 0 ? 0 : count;
2664 static struct obd_ops mdc_obd_ops = {
2665 .o_owner = THIS_MODULE,
2666 .o_setup = mdc_setup,
2667 .o_precleanup = mdc_precleanup,
2668 .o_cleanup = mdc_cleanup,
2669 .o_add_conn = client_import_add_conn,
2670 .o_del_conn = client_import_del_conn,
2671 .o_connect = client_connect_import,
2672 .o_reconnect = osc_reconnect,
2673 .o_disconnect = osc_disconnect,
2674 .o_iocontrol = mdc_iocontrol,
2675 .o_set_info_async = mdc_set_info_async,
2676 .o_statfs = mdc_statfs,
2677 .o_fid_init = client_fid_init,
2678 .o_fid_fini = client_fid_fini,
2679 .o_fid_alloc = mdc_fid_alloc,
2680 .o_import_event = mdc_import_event,
2681 .o_get_info = mdc_get_info,
2682 .o_process_config = mdc_process_config,
2683 .o_get_uuid = mdc_get_uuid,
2684 .o_quotactl = mdc_quotactl,
2687 static struct md_ops mdc_md_ops = {
2688 .m_get_root = mdc_get_root,
2689 .m_null_inode = mdc_null_inode,
2690 .m_close = mdc_close,
2691 .m_create = mdc_create,
2692 .m_enqueue = mdc_enqueue,
2693 .m_getattr = mdc_getattr,
2694 .m_getattr_name = mdc_getattr_name,
2695 .m_intent_lock = mdc_intent_lock,
2697 .m_rename = mdc_rename,
2698 .m_setattr = mdc_setattr,
2699 .m_setxattr = mdc_setxattr,
2700 .m_getxattr = mdc_getxattr,
2701 .m_fsync = mdc_fsync,
2702 .m_file_resync = mdc_file_resync,
2703 .m_read_page = mdc_read_page,
2704 .m_unlink = mdc_unlink,
2705 .m_cancel_unused = mdc_cancel_unused,
2706 .m_init_ea_size = mdc_init_ea_size,
2707 .m_set_lock_data = mdc_set_lock_data,
2708 .m_lock_match = mdc_lock_match,
2709 .m_get_lustre_md = mdc_get_lustre_md,
2710 .m_free_lustre_md = mdc_free_lustre_md,
2711 .m_set_open_replay_data = mdc_set_open_replay_data,
2712 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2713 .m_intent_getattr_async = mdc_intent_getattr_async,
2714 .m_revalidate_lock = mdc_revalidate_lock
2717 static int __init mdc_init(void)
2719 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2720 LUSTRE_MDC_NAME, &mdc_device_type);
2723 static void __exit mdc_exit(void)
2725 class_unregister_type(LUSTRE_MDC_NAME);
2728 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2729 MODULE_DESCRIPTION("Lustre Metadata Client");
2730 MODULE_VERSION(LUSTRE_VERSION_STRING);
2731 MODULE_LICENSE("GPL");
2733 module_init(mdc_init);
2734 module_exit(mdc_exit);