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 valid, const char *xattr_name,
421 const char *input, int input_size, int output_size,
422 int flags, __u32 suppgid,
423 struct ptlrpc_request **request)
425 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
426 fid, MDS_REINT, valid, xattr_name,
427 input, input_size, output_size, flags,
431 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
432 u64 valid, const char *xattr_name,
433 const char *input, int input_size, int output_size,
434 int flags, struct ptlrpc_request **request)
436 return mdc_xattr_common(exp, &RQF_MDS_GETXATTR,
437 fid, MDS_GETXATTR, valid, xattr_name,
438 input, input_size, output_size, flags,
442 #ifdef CONFIG_FS_POSIX_ACL
443 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
445 struct req_capsule *pill = &req->rq_pill;
446 struct mdt_body *body = md->body;
447 struct posix_acl *acl;
452 if (!body->mbo_aclsize)
455 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
460 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
465 CERROR("convert xattr to acl: %d\n", rc);
469 rc = posix_acl_valid(&init_user_ns, acl);
471 CERROR("validate acl: %d\n", rc);
472 posix_acl_release(acl);
480 #define mdc_unpack_acl(req, md) 0
483 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
484 struct obd_export *dt_exp, struct obd_export *md_exp,
485 struct lustre_md *md)
487 struct req_capsule *pill = &req->rq_pill;
492 memset(md, 0, sizeof(*md));
494 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
495 LASSERT(md->body != NULL);
497 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
498 if (!S_ISREG(md->body->mbo_mode)) {
499 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
500 "regular file, but is not\n");
501 GOTO(out, rc = -EPROTO);
504 if (md->body->mbo_eadatasize == 0) {
505 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
506 "but eadatasize 0\n");
507 GOTO(out, rc = -EPROTO);
510 md->layout.lb_len = md->body->mbo_eadatasize;
511 md->layout.lb_buf = req_capsule_server_sized_get(pill,
514 if (md->layout.lb_buf == NULL)
515 GOTO(out, rc = -EPROTO);
516 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
517 const union lmv_mds_md *lmv;
520 if (!S_ISDIR(md->body->mbo_mode)) {
521 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
522 "directory, but is not\n");
523 GOTO(out, rc = -EPROTO);
526 lmv_size = md->body->mbo_eadatasize;
528 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
529 "but eadatasize 0\n");
533 if (md->body->mbo_valid & OBD_MD_MEA) {
534 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
537 GOTO(out, rc = -EPROTO);
539 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
543 if (rc < (typeof(rc))sizeof(*md->lmv)) {
544 CDEBUG(D_INFO, "size too small: "
545 "rc < sizeof(*md->lmv) (%d < %d)\n",
546 rc, (int)sizeof(*md->lmv));
547 GOTO(out, rc = -EPROTO);
553 if (md->body->mbo_valid & OBD_MD_FLACL) {
554 /* for ACL, it's possible that FLACL is set but aclsize is zero.
555 * only when aclsize != 0 there's an actual segment for ACL
558 if (md->body->mbo_aclsize) {
559 rc = mdc_unpack_acl(req, md);
562 #ifdef CONFIG_FS_POSIX_ACL
564 md->posix_acl = NULL;
572 #ifdef CONFIG_FS_POSIX_ACL
573 posix_acl_release(md->posix_acl);
579 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
585 void mdc_replay_open(struct ptlrpc_request *req)
587 struct md_open_data *mod = req->rq_cb_data;
588 struct ptlrpc_request *close_req;
589 struct obd_client_handle *och;
590 struct lustre_handle old;
591 struct mdt_body *body;
595 DEBUG_REQ(D_ERROR, req,
596 "Can't properly replay without open data.");
601 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
602 LASSERT(body != NULL);
604 spin_lock(&req->rq_lock);
606 if (och && och->och_fh.cookie)
607 req->rq_early_free_repbuf = 1;
609 req->rq_early_free_repbuf = 0;
610 spin_unlock(&req->rq_lock);
612 if (req->rq_early_free_repbuf) {
613 struct lustre_handle *file_fh;
615 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
617 file_fh = &och->och_fh;
618 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
619 file_fh->cookie, body->mbo_handle.cookie);
621 *file_fh = body->mbo_handle;
624 close_req = mod->mod_close_req;
626 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
627 struct mdt_ioepoch *epoch;
629 LASSERT(opc == MDS_CLOSE);
630 epoch = req_capsule_client_get(&close_req->rq_pill,
634 if (req->rq_early_free_repbuf)
635 LASSERT(!memcmp(&old, &epoch->mio_handle, sizeof(old)));
637 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
638 epoch->mio_handle = body->mbo_handle;
643 void mdc_commit_open(struct ptlrpc_request *req)
645 struct md_open_data *mod = req->rq_cb_data;
650 * No need to touch md_open_data::mod_och, it holds a reference on
651 * \var mod and will zero references to each other, \var mod will be
652 * freed after that when md_open_data::mod_och will put the reference.
656 * Do not let open request to disappear as it still may be needed
657 * for close rpc to happen (it may happen on evict only, otherwise
658 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
659 * called), just mark this rpc as committed to distinguish these 2
660 * cases, see mdc_close() for details. The open request reference will
661 * be put along with freeing \var mod.
663 ptlrpc_request_addref(req);
664 spin_lock(&req->rq_lock);
665 req->rq_committed = 1;
666 spin_unlock(&req->rq_lock);
667 req->rq_cb_data = NULL;
671 int mdc_set_open_replay_data(struct obd_export *exp,
672 struct obd_client_handle *och,
673 struct lookup_intent *it)
675 struct md_open_data *mod;
676 struct mdt_rec_create *rec;
677 struct mdt_body *body;
678 struct ptlrpc_request *open_req = it->it_request;
679 struct obd_import *imp = open_req->rq_import;
682 if (!open_req->rq_replay)
685 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
686 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
687 LASSERT(rec != NULL);
688 /* Incoming message in my byte order (it's been swabbed). */
689 /* Outgoing messages always in my byte order. */
690 LASSERT(body != NULL);
692 /* Only if the import is replayable, we set replay_open data */
693 if (och && imp->imp_replayable) {
694 mod = obd_mod_alloc();
696 DEBUG_REQ(D_ERROR, open_req,
697 "Can't allocate md_open_data");
702 * Take a reference on \var mod, to be freed on mdc_close().
703 * It protects \var mod from being freed on eviction (commit
704 * callback is called despite rq_replay flag).
705 * Another reference for \var och.
710 spin_lock(&open_req->rq_lock);
713 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
714 it_disposition(it, DISP_OPEN_STRIPE);
715 mod->mod_open_req = open_req;
716 open_req->rq_cb_data = mod;
717 open_req->rq_commit_cb = mdc_commit_open;
718 open_req->rq_early_free_repbuf = 1;
719 spin_unlock(&open_req->rq_lock);
722 rec->cr_fid2 = body->mbo_fid1;
723 rec->cr_old_handle.cookie = body->mbo_handle.cookie;
724 open_req->rq_replay_cb = mdc_replay_open;
725 if (!fid_is_sane(&body->mbo_fid1)) {
726 DEBUG_REQ(D_ERROR, open_req, "Saving replay request with "
731 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
735 static void mdc_free_open(struct md_open_data *mod)
739 if (mod->mod_is_create == 0 &&
740 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
744 * No reason to asssert here if the open request has
745 * rq_replay == 1. It means that mdc_close failed, and
746 * close request wasn`t sent. It is not fatal to client.
747 * The worst thing is eviction if the client gets open lock
750 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
751 "= %d\n", mod->mod_open_req->rq_replay);
753 ptlrpc_request_committed(mod->mod_open_req, committed);
754 if (mod->mod_close_req)
755 ptlrpc_request_committed(mod->mod_close_req, committed);
758 int mdc_clear_open_replay_data(struct obd_export *exp,
759 struct obd_client_handle *och)
761 struct md_open_data *mod = och->och_mod;
765 * It is possible to not have \var mod in a case of eviction between
766 * lookup and ll_file_open().
771 LASSERT(mod != LP_POISON);
772 LASSERT(mod->mod_open_req != NULL);
774 spin_lock(&mod->mod_open_req->rq_lock);
776 mod->mod_och->och_fh.cookie = 0;
777 mod->mod_open_req->rq_early_free_repbuf = 0;
778 spin_unlock(&mod->mod_open_req->rq_lock);
788 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
789 struct md_open_data *mod, struct ptlrpc_request **request)
791 struct obd_device *obd = class_exp2obd(exp);
792 struct ptlrpc_request *req;
793 struct req_format *req_fmt;
794 size_t u32_count = 0;
799 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
800 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
803 if (op_data->op_bias & MDS_CLOSE_INTENT) {
804 req_fmt = &RQF_MDS_CLOSE_INTENT;
805 if (op_data->op_bias & MDS_HSM_RELEASE) {
806 /* allocate a FID for volatile file */
807 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
810 CERROR("%s: "DFID" allocating FID: rc = %d\n",
811 obd->obd_name, PFID(&op_data->op_fid1),
813 /* save the errcode and proceed to close */
817 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
818 size_t count = op_data->op_data_size / sizeof(__u32);
820 if (count > INLINE_RESYNC_ARRAY_SIZE)
824 req_fmt = &RQF_MDS_CLOSE;
828 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
831 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
833 /* Ensure that this close's handle is fixed up during replay. */
834 if (likely(mod != NULL)) {
835 LASSERTF(mod->mod_open_req != NULL &&
836 mod->mod_open_req->rq_type != LI_POISON,
837 "POISONED open %p!\n", mod->mod_open_req);
839 mod->mod_close_req = req;
841 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
842 /* We no longer want to preserve this open for replay even
843 * though the open was committed. b=3632, b=3633 */
844 spin_lock(&mod->mod_open_req->rq_lock);
845 mod->mod_open_req->rq_replay = 0;
846 spin_unlock(&mod->mod_open_req->rq_lock);
848 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
852 * TODO: repeat close after errors
854 CWARN("%s: close of FID "DFID" failed, file reference will be "
855 "dropped when this client unmounts or is evicted\n",
856 obd->obd_name, PFID(&op_data->op_fid1));
857 GOTO(out, rc = -ENOMEM);
861 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
862 u32_count * sizeof(__u32));
864 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
866 ptlrpc_request_free(req);
871 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
872 * portal whose threads are not taking any DLM locks and are therefore
873 * always progressing */
874 req->rq_request_portal = MDS_READPAGE_PORTAL;
875 ptlrpc_at_set_req_timeout(req);
878 mdc_close_pack(req, op_data);
880 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
881 obd->u.cli.cl_default_mds_easize);
883 ptlrpc_request_set_replen(req);
885 mdc_get_mod_rpc_slot(req, NULL);
886 rc = ptlrpc_queue_wait(req);
887 mdc_put_mod_rpc_slot(req, NULL);
889 if (req->rq_repmsg == NULL) {
890 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
893 rc = req->rq_status ?: -EIO;
894 } else if (rc == 0 || rc == -EAGAIN) {
895 struct mdt_body *body;
897 rc = lustre_msg_get_status(req->rq_repmsg);
898 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
899 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
904 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
907 } else if (rc == -ESTALE) {
909 * it can be allowed error after 3633 if open was committed and
910 * server failed before close was sent. Let's check if mod
911 * exists and return no error in that case
914 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
915 LASSERT(mod->mod_open_req != NULL);
916 if (mod->mod_open_req->rq_committed)
924 mod->mod_close_req = NULL;
925 /* Since now, mod is accessed through open_req only,
926 * thus close req does not keep a reference on mod anymore. */
931 RETURN(rc < 0 ? rc : saved_rc);
934 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
935 u64 offset, struct page **pages, int npages,
936 struct ptlrpc_request **request)
938 struct ptlrpc_request *req;
939 struct ptlrpc_bulk_desc *desc;
941 wait_queue_head_t waitq;
943 struct l_wait_info lwi;
948 init_waitqueue_head(&waitq);
951 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
955 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
957 ptlrpc_request_free(req);
961 req->rq_request_portal = MDS_READPAGE_PORTAL;
962 ptlrpc_at_set_req_timeout(req);
964 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
965 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
967 &ptlrpc_bulk_kiov_pin_ops);
969 ptlrpc_req_finished(req);
973 /* NB req now owns desc and will free it when it gets freed */
974 for (i = 0; i < npages; i++)
975 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
978 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
980 ptlrpc_request_set_replen(req);
981 rc = ptlrpc_queue_wait(req);
983 ptlrpc_req_finished(req);
984 if (rc != -ETIMEDOUT)
988 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
989 CERROR("%s: too many resend retries: rc = %d\n",
990 exp->exp_obd->obd_name, -EIO);
993 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
995 l_wait_event(waitq, 0, &lwi);
1000 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1001 req->rq_bulk->bd_nob_transferred);
1003 ptlrpc_req_finished(req);
1007 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1008 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1009 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1010 PAGE_SIZE * npages);
1011 ptlrpc_req_finished(req);
1019 static void mdc_release_page(struct page *page, int remove)
1023 if (likely(page->mapping != NULL))
1024 truncate_complete_page(page->mapping, page);
1030 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1031 __u64 *start, __u64 *end, int hash64)
1034 * Complement of hash is used as an index so that
1035 * radix_tree_gang_lookup() can be used to find a page with starting
1036 * hash _smaller_ than one we are looking for.
1038 unsigned long offset = hash_x_index(*hash, hash64);
1042 spin_lock_irq(&mapping->tree_lock);
1043 found = radix_tree_gang_lookup(&mapping->page_tree,
1044 (void **)&page, offset, 1);
1045 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1046 struct lu_dirpage *dp;
1049 spin_unlock_irq(&mapping->tree_lock);
1051 * In contrast to find_lock_page() we are sure that directory
1052 * page cannot be truncated (while DLM lock is held) and,
1053 * hence, can avoid restart.
1055 * In fact, page cannot be locked here at all, because
1056 * mdc_read_page_remote does synchronous io.
1058 wait_on_page_locked(page);
1059 if (PageUptodate(page)) {
1061 if (BITS_PER_LONG == 32 && hash64) {
1062 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1063 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1064 *hash = *hash >> 32;
1066 *start = le64_to_cpu(dp->ldp_hash_start);
1067 *end = le64_to_cpu(dp->ldp_hash_end);
1069 if (unlikely(*start == 1 && *hash == 0))
1072 LASSERTF(*start <= *hash, "start = %#llx"
1073 ",end = %#llx,hash = %#llx\n",
1074 *start, *end, *hash);
1075 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1076 " hash %#llx\n", offset, *start, *end, *hash);
1079 mdc_release_page(page, 0);
1081 } else if (*end != *start && *hash == *end) {
1083 * upon hash collision, remove this page,
1084 * otherwise put page reference, and
1085 * mdc_read_page_remote() will issue RPC to
1086 * fetch the page we want.
1089 mdc_release_page(page,
1090 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1095 page = ERR_PTR(-EIO);
1098 spin_unlock_irq(&mapping->tree_lock);
1105 * Adjust a set of pages, each page containing an array of lu_dirpages,
1106 * so that each page can be used as a single logical lu_dirpage.
1108 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1109 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1110 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1111 * value is used as a cookie to request the next lu_dirpage in a
1112 * directory listing that spans multiple pages (two in this example):
1115 * .|--------v------- -----.
1116 * |s|e|f|p|ent|ent| ... |ent|
1117 * '--|-------------- -----' Each PAGE contains a single
1118 * '------. lu_dirpage.
1119 * .---------v------- -----.
1120 * |s|e|f|p|ent| 0 | ... | 0 |
1121 * '----------------- -----'
1123 * However, on hosts where the native VM page size (PAGE_SIZE) is
1124 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1125 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1126 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1127 * after it in the same PAGE (arrows simplified for brevity, but
1128 * in general e0==s1, e1==s2, etc.):
1130 * .-------------------- -----.
1131 * |s0|e0|f0|p|ent|ent| ... |ent|
1132 * |---v---------------- -----|
1133 * |s1|e1|f1|p|ent|ent| ... |ent|
1134 * |---v---------------- -----| Here, each PAGE contains
1135 * ... multiple lu_dirpages.
1136 * |---v---------------- -----|
1137 * |s'|e'|f'|p|ent|ent| ... |ent|
1138 * '---|---------------- -----'
1140 * .----------------------------.
1143 * This structure is transformed into a single logical lu_dirpage as follows:
1145 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1146 * labeled 'next PAGE'.
1148 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1149 * a hash collision with the next page exists.
1151 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1152 * to the first entry of the next lu_dirpage.
1154 #if PAGE_SIZE > LU_PAGE_SIZE
1155 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1159 for (i = 0; i < cfs_pgs; i++) {
1160 struct lu_dirpage *dp = kmap(pages[i]);
1161 struct lu_dirpage *first = dp;
1162 struct lu_dirent *end_dirent = NULL;
1163 struct lu_dirent *ent;
1164 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1165 __u32 flags = le32_to_cpu(dp->ldp_flags);
1167 while (--lu_pgs > 0) {
1168 ent = lu_dirent_start(dp);
1169 for (end_dirent = ent; ent != NULL;
1170 end_dirent = ent, ent = lu_dirent_next(ent));
1172 /* Advance dp to next lu_dirpage. */
1173 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1175 /* Check if we've reached the end of the PAGE. */
1176 if (!((unsigned long)dp & ~PAGE_MASK))
1179 /* Save the hash and flags of this lu_dirpage. */
1180 hash_end = le64_to_cpu(dp->ldp_hash_end);
1181 flags = le32_to_cpu(dp->ldp_flags);
1183 /* Check if lu_dirpage contains no entries. */
1184 if (end_dirent == NULL)
1187 /* Enlarge the end entry lde_reclen from 0 to
1188 * first entry of next lu_dirpage. */
1189 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1190 end_dirent->lde_reclen =
1191 cpu_to_le16((char *)(dp->ldp_entries) -
1192 (char *)end_dirent);
1195 first->ldp_hash_end = hash_end;
1196 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1197 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1201 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1204 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1205 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1207 /* parameters for readdir page */
1208 struct readpage_param {
1209 struct md_op_data *rp_mod;
1212 struct obd_export *rp_exp;
1213 struct md_callback *rp_cb;
1216 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1217 static inline void delete_from_page_cache(struct page *page)
1219 remove_from_page_cache(page);
1225 * Read pages from server.
1227 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1228 * a header lu_dirpage which describes the start/end hash, and whether this
1229 * page is empty (contains no dir entry) or hash collide with next page.
1230 * After client receives reply, several pages will be integrated into dir page
1231 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1232 * lu_dirpage for this integrated page will be adjusted.
1234 static int mdc_read_page_remote(void *data, struct page *page0)
1236 struct readpage_param *rp = data;
1237 struct page **page_pool;
1239 struct lu_dirpage *dp;
1240 struct md_op_data *op_data = rp->rp_mod;
1241 struct ptlrpc_request *req;
1243 struct inode *inode;
1245 int rd_pgs = 0; /* number of pages actually read */
1251 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1252 inode = op_data->op_data;
1253 fid = &op_data->op_fid1;
1254 LASSERT(inode != NULL);
1256 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1257 if (page_pool != NULL) {
1258 page_pool[0] = page0;
1264 for (npages = 1; npages < max_pages; npages++) {
1265 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1269 page_pool[npages] = page;
1272 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1274 /* page0 is special, which was added into page cache early */
1275 delete_from_page_cache(page0);
1279 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1281 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1282 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1284 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1286 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1288 SetPageUptodate(page0);
1292 ptlrpc_req_finished(req);
1293 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1294 for (i = 1; i < npages; i++) {
1295 unsigned long offset;
1299 page = page_pool[i];
1301 if (rc < 0 || i >= rd_pgs) {
1306 SetPageUptodate(page);
1309 hash = le64_to_cpu(dp->ldp_hash_start);
1312 offset = hash_x_index(hash, rp->rp_hash64);
1314 prefetchw(&page->flags);
1315 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1320 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1321 " rc = %d\n", offset, ret);
1325 if (page_pool != &page0)
1326 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1332 * Read dir page from cache first, if it can not find it, read it from
1333 * server and add into the cache.
1335 * \param[in] exp MDC export
1336 * \param[in] op_data client MD stack parameters, transfering parameters
1337 * between different layers on client MD stack.
1338 * \param[in] cb_op callback required for ldlm lock enqueue during
1340 * \param[in] hash_offset the hash offset of the page to be read
1341 * \param[in] ppage the page to be read
1343 * retval = 0 get the page successfully
1344 * errno(<0) get the page failed
1346 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1347 struct md_callback *cb_op, __u64 hash_offset,
1348 struct page **ppage)
1350 struct lookup_intent it = { .it_op = IT_READDIR };
1352 struct inode *dir = op_data->op_data;
1353 struct address_space *mapping;
1354 struct lu_dirpage *dp;
1357 struct lustre_handle lockh;
1358 struct ptlrpc_request *enq_req = NULL;
1359 struct readpage_param rp_param;
1366 LASSERT(dir != NULL);
1367 mapping = dir->i_mapping;
1369 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1370 cb_op->md_blocking_ast, 0);
1371 if (enq_req != NULL)
1372 ptlrpc_req_finished(enq_req);
1375 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1376 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1381 lockh.cookie = it.it_lock_handle;
1382 mdc_set_lock_data(exp, &lockh, dir, NULL);
1384 rp_param.rp_off = hash_offset;
1385 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1386 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1387 rp_param.rp_hash64);
1389 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1390 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1391 rp_param.rp_off, PTR_ERR(page));
1392 GOTO(out_unlock, rc = PTR_ERR(page));
1393 } else if (page != NULL) {
1395 * XXX nikita: not entirely correct handling of a corner case:
1396 * suppose hash chain of entries with hash value HASH crosses
1397 * border between pages P0 and P1. First both P0 and P1 are
1398 * cached, seekdir() is called for some entry from the P0 part
1399 * of the chain. Later P0 goes out of cache. telldir(HASH)
1400 * happens and finds P1, as it starts with matching hash
1401 * value. Remaining entries from P0 part of the chain are
1402 * skipped. (Is that really a bug?)
1404 * Possible solutions: 0. don't cache P1 is such case, handle
1405 * it as an "overflow" page. 1. invalidate all pages at
1406 * once. 2. use HASH|1 as an index for P1.
1408 GOTO(hash_collision, page);
1411 rp_param.rp_exp = exp;
1412 rp_param.rp_mod = op_data;
1413 page = read_cache_page(mapping,
1414 hash_x_index(rp_param.rp_off,
1415 rp_param.rp_hash64),
1416 mdc_read_page_remote, &rp_param);
1418 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1419 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1420 rp_param.rp_off, PTR_ERR(page));
1421 GOTO(out_unlock, rc = PTR_ERR(page));
1424 wait_on_page_locked(page);
1426 if (!PageUptodate(page)) {
1427 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1428 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1429 rp_param.rp_off, -5);
1432 if (!PageChecked(page))
1433 SetPageChecked(page);
1434 if (PageError(page)) {
1435 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1436 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1437 rp_param.rp_off, -5);
1442 dp = page_address(page);
1443 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1444 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1445 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1446 rp_param.rp_off = hash_offset >> 32;
1448 start = le64_to_cpu(dp->ldp_hash_start);
1449 end = le64_to_cpu(dp->ldp_hash_end);
1450 rp_param.rp_off = hash_offset;
1453 LASSERT(start == rp_param.rp_off);
1454 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1455 #if BITS_PER_LONG == 32
1456 CWARN("Real page-wide hash collision at [%llu %llu] with "
1457 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1458 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1462 * Fetch whole overflow chain...
1470 ldlm_lock_decref(&lockh, it.it_lock_mode);
1474 mdc_release_page(page, 1);
1479 static int mdc_statfs(const struct lu_env *env,
1480 struct obd_export *exp, struct obd_statfs *osfs,
1481 time64_t max_age, __u32 flags)
1483 struct obd_device *obd = class_exp2obd(exp);
1484 struct ptlrpc_request *req;
1485 struct obd_statfs *msfs;
1486 struct obd_import *imp = NULL;
1491 * Since the request might also come from lprocfs, so we need
1492 * sync this with client_disconnect_export Bug15684
1494 down_read(&obd->u.cli.cl_sem);
1495 if (obd->u.cli.cl_import)
1496 imp = class_import_get(obd->u.cli.cl_import);
1497 up_read(&obd->u.cli.cl_sem);
1501 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_STATFS,
1502 LUSTRE_MDS_VERSION, MDS_STATFS);
1504 GOTO(output, rc = -ENOMEM);
1506 ptlrpc_request_set_replen(req);
1508 if (flags & OBD_STATFS_NODELAY) {
1509 /* procfs requests not want stay in wait for avoid deadlock */
1510 req->rq_no_resend = 1;
1511 req->rq_no_delay = 1;
1514 rc = ptlrpc_queue_wait(req);
1516 /* check connection error first */
1517 if (imp->imp_connect_error)
1518 rc = imp->imp_connect_error;
1522 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1524 GOTO(out, rc = -EPROTO);
1529 ptlrpc_req_finished(req);
1531 class_import_put(imp);
1535 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1537 __u32 keylen, vallen;
1541 if (gf->gf_pathlen > PATH_MAX)
1542 RETURN(-ENAMETOOLONG);
1543 if (gf->gf_pathlen < 2)
1546 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1547 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1548 sizeof(struct lu_fid));
1549 OBD_ALLOC(key, keylen);
1552 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1553 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1554 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1555 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1556 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1557 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1559 if (!fid_is_sane(&gf->gf_fid))
1560 GOTO(out, rc = -EINVAL);
1562 /* Val is struct getinfo_fid2path result plus path */
1563 vallen = sizeof(*gf) + gf->gf_pathlen;
1565 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1566 if (rc != 0 && rc != -EREMOTE)
1569 if (vallen <= sizeof(*gf))
1570 GOTO(out, rc = -EPROTO);
1571 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1572 GOTO(out, rc = -EOVERFLOW);
1574 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1575 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1576 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1577 /* only log the last 512 characters of the path */
1578 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1581 OBD_FREE(key, keylen);
1585 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1586 struct hsm_progress_kernel *hpk)
1588 struct obd_import *imp = class_exp2cliimp(exp);
1589 struct hsm_progress_kernel *req_hpk;
1590 struct ptlrpc_request *req;
1594 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1595 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1597 GOTO(out, rc = -ENOMEM);
1599 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1601 /* Copy hsm_progress struct */
1602 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1603 if (req_hpk == NULL)
1604 GOTO(out, rc = -EPROTO);
1607 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1609 ptlrpc_request_set_replen(req);
1611 mdc_get_mod_rpc_slot(req, NULL);
1612 rc = ptlrpc_queue_wait(req);
1613 mdc_put_mod_rpc_slot(req, NULL);
1617 ptlrpc_req_finished(req);
1621 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archives)
1623 __u32 *archive_mask;
1624 struct ptlrpc_request *req;
1628 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_REGISTER,
1630 MDS_HSM_CT_REGISTER);
1632 GOTO(out, rc = -ENOMEM);
1634 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1636 /* Copy hsm_progress struct */
1637 archive_mask = req_capsule_client_get(&req->rq_pill,
1638 &RMF_MDS_HSM_ARCHIVE);
1639 if (archive_mask == NULL)
1640 GOTO(out, rc = -EPROTO);
1642 *archive_mask = archives;
1644 ptlrpc_request_set_replen(req);
1646 rc = mdc_queue_wait(req);
1649 ptlrpc_req_finished(req);
1653 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1654 struct md_op_data *op_data)
1656 struct hsm_current_action *hca = op_data->op_data;
1657 struct hsm_current_action *req_hca;
1658 struct ptlrpc_request *req;
1662 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1663 &RQF_MDS_HSM_ACTION);
1667 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1669 ptlrpc_request_free(req);
1673 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1674 op_data->op_suppgids[0], 0);
1676 ptlrpc_request_set_replen(req);
1678 rc = mdc_queue_wait(req);
1682 req_hca = req_capsule_server_get(&req->rq_pill,
1683 &RMF_MDS_HSM_CURRENT_ACTION);
1684 if (req_hca == NULL)
1685 GOTO(out, rc = -EPROTO);
1691 ptlrpc_req_finished(req);
1695 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1697 struct ptlrpc_request *req;
1701 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1703 MDS_HSM_CT_UNREGISTER);
1705 GOTO(out, rc = -ENOMEM);
1707 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1709 ptlrpc_request_set_replen(req);
1711 rc = mdc_queue_wait(req);
1714 ptlrpc_req_finished(req);
1718 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1719 struct md_op_data *op_data)
1721 struct hsm_user_state *hus = op_data->op_data;
1722 struct hsm_user_state *req_hus;
1723 struct ptlrpc_request *req;
1727 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1728 &RQF_MDS_HSM_STATE_GET);
1732 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1734 ptlrpc_request_free(req);
1738 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1739 op_data->op_suppgids[0], 0);
1741 ptlrpc_request_set_replen(req);
1743 rc = mdc_queue_wait(req);
1747 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1748 if (req_hus == NULL)
1749 GOTO(out, rc = -EPROTO);
1755 ptlrpc_req_finished(req);
1759 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1760 struct md_op_data *op_data)
1762 struct hsm_state_set *hss = op_data->op_data;
1763 struct hsm_state_set *req_hss;
1764 struct ptlrpc_request *req;
1768 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1769 &RQF_MDS_HSM_STATE_SET);
1773 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1775 ptlrpc_request_free(req);
1779 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1780 op_data->op_suppgids[0], 0);
1783 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1784 if (req_hss == NULL)
1785 GOTO(out, rc = -EPROTO);
1788 ptlrpc_request_set_replen(req);
1790 mdc_get_mod_rpc_slot(req, NULL);
1791 rc = ptlrpc_queue_wait(req);
1792 mdc_put_mod_rpc_slot(req, NULL);
1796 ptlrpc_req_finished(req);
1800 static int mdc_ioc_hsm_request(struct obd_export *exp,
1801 struct hsm_user_request *hur)
1803 struct obd_import *imp = class_exp2cliimp(exp);
1804 struct ptlrpc_request *req;
1805 struct hsm_request *req_hr;
1806 struct hsm_user_item *req_hui;
1811 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1813 GOTO(out, rc = -ENOMEM);
1815 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1816 hur->hur_request.hr_itemcount
1817 * sizeof(struct hsm_user_item));
1818 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1819 hur->hur_request.hr_data_len);
1821 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1823 ptlrpc_request_free(req);
1827 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1829 /* Copy hsm_request struct */
1830 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1832 GOTO(out, rc = -EPROTO);
1833 *req_hr = hur->hur_request;
1835 /* Copy hsm_user_item structs */
1836 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1837 if (req_hui == NULL)
1838 GOTO(out, rc = -EPROTO);
1839 memcpy(req_hui, hur->hur_user_item,
1840 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1842 /* Copy opaque field */
1843 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1844 if (req_opaque == NULL)
1845 GOTO(out, rc = -EPROTO);
1846 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1848 ptlrpc_request_set_replen(req);
1850 mdc_get_mod_rpc_slot(req, NULL);
1851 rc = ptlrpc_queue_wait(req);
1852 mdc_put_mod_rpc_slot(req, NULL);
1857 ptlrpc_req_finished(req);
1861 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1862 struct lustre_kernelcomm *lk);
1864 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1865 struct obd_quotactl *oqctl)
1867 struct ptlrpc_request *req;
1868 struct obd_quotactl *oqc;
1872 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1873 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1878 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1881 ptlrpc_request_set_replen(req);
1882 ptlrpc_at_set_req_timeout(req);
1883 req->rq_no_resend = 1;
1885 rc = ptlrpc_queue_wait(req);
1887 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1889 if (req->rq_repmsg &&
1890 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1893 CERROR ("Can't unpack obd_quotactl\n");
1896 ptlrpc_req_finished(req);
1901 static int mdc_ioc_swap_layouts(struct obd_export *exp,
1902 struct md_op_data *op_data)
1904 struct list_head cancels = LIST_HEAD_INIT(cancels);
1905 struct ptlrpc_request *req;
1907 struct mdc_swap_layouts *msl, *payload;
1910 msl = op_data->op_data;
1912 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
1913 * first thing it will do is to cancel the 2 layout
1914 * locks held by this client.
1915 * So the client must cancel its layout locks on the 2 fids
1916 * with the request RPC to avoid extra RPC round trips.
1918 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
1919 LCK_EX, MDS_INODELOCK_LAYOUT |
1920 MDS_INODELOCK_XATTR);
1921 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
1922 LCK_EX, MDS_INODELOCK_LAYOUT |
1923 MDS_INODELOCK_XATTR);
1925 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1926 &RQF_MDS_SWAP_LAYOUTS);
1928 ldlm_lock_list_put(&cancels, l_bl_ast, count);
1932 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
1934 ptlrpc_request_free(req);
1938 mdc_swap_layouts_pack(req, op_data);
1940 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
1945 ptlrpc_request_set_replen(req);
1947 rc = ptlrpc_queue_wait(req);
1953 ptlrpc_req_finished(req);
1957 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1958 void *karg, void __user *uarg)
1960 struct obd_device *obd = exp->exp_obd;
1961 struct obd_ioctl_data *data = karg;
1962 struct obd_import *imp = obd->u.cli.cl_import;
1966 if (!try_module_get(THIS_MODULE)) {
1967 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
1968 module_name(THIS_MODULE));
1972 case OBD_IOC_FID2PATH:
1973 rc = mdc_ioc_fid2path(exp, karg);
1975 case LL_IOC_HSM_CT_START:
1976 rc = mdc_ioc_hsm_ct_start(exp, karg);
1977 /* ignore if it was already registered on this MDS. */
1981 case LL_IOC_HSM_PROGRESS:
1982 rc = mdc_ioc_hsm_progress(exp, karg);
1984 case LL_IOC_HSM_STATE_GET:
1985 rc = mdc_ioc_hsm_state_get(exp, karg);
1987 case LL_IOC_HSM_STATE_SET:
1988 rc = mdc_ioc_hsm_state_set(exp, karg);
1990 case LL_IOC_HSM_ACTION:
1991 rc = mdc_ioc_hsm_current_action(exp, karg);
1993 case LL_IOC_HSM_REQUEST:
1994 rc = mdc_ioc_hsm_request(exp, karg);
1996 case OBD_IOC_CLIENT_RECOVER:
1997 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2001 case IOC_OSC_SET_ACTIVE:
2002 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2004 case OBD_IOC_PING_TARGET:
2005 rc = ptlrpc_obd_ping(obd);
2008 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2009 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2010 * there'd be no LMV layer thus we might be called here. Eventually
2011 * this code should be removed.
2014 case IOC_OBD_STATFS: {
2015 struct obd_statfs stat_buf = {0};
2017 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2018 GOTO(out, rc = -ENODEV);
2021 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2022 min((int)data->ioc_plen2,
2023 (int)sizeof(struct obd_uuid))))
2024 GOTO(out, rc = -EFAULT);
2026 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2027 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2032 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2033 min((int) data->ioc_plen1,
2034 (int) sizeof(stat_buf))))
2035 GOTO(out, rc = -EFAULT);
2039 case OBD_IOC_QUOTACTL: {
2040 struct if_quotactl *qctl = karg;
2041 struct obd_quotactl *oqctl;
2043 OBD_ALLOC_PTR(oqctl);
2045 GOTO(out, rc = -ENOMEM);
2047 QCTL_COPY(oqctl, qctl);
2048 rc = obd_quotactl(exp, oqctl);
2050 QCTL_COPY(qctl, oqctl);
2051 qctl->qc_valid = QC_MDTIDX;
2052 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2055 OBD_FREE_PTR(oqctl);
2058 case LL_IOC_GET_CONNECT_FLAGS:
2059 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2060 sizeof(*exp_connect_flags_ptr(exp))))
2061 GOTO(out, rc = -EFAULT);
2064 case LL_IOC_LOV_SWAP_LAYOUTS:
2065 rc = mdc_ioc_swap_layouts(exp, karg);
2068 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2069 GOTO(out, rc = -ENOTTY);
2072 module_put(THIS_MODULE);
2077 static int mdc_get_info_rpc(struct obd_export *exp,
2078 u32 keylen, void *key,
2079 u32 vallen, void *val)
2081 struct obd_import *imp = class_exp2cliimp(exp);
2082 struct ptlrpc_request *req;
2087 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2091 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2092 RCL_CLIENT, keylen);
2093 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2094 RCL_CLIENT, sizeof(vallen));
2096 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2098 ptlrpc_request_free(req);
2102 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2103 memcpy(tmp, key, keylen);
2104 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2105 memcpy(tmp, &vallen, sizeof(vallen));
2107 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2108 RCL_SERVER, vallen);
2109 ptlrpc_request_set_replen(req);
2111 rc = ptlrpc_queue_wait(req);
2112 /* -EREMOTE means the get_info result is partial, and it needs to
2113 * continue on another MDT, see fid2path part in lmv_iocontrol */
2114 if (rc == 0 || rc == -EREMOTE) {
2115 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2116 memcpy(val, tmp, vallen);
2117 if (ptlrpc_rep_need_swab(req)) {
2118 if (KEY_IS(KEY_FID2PATH))
2119 lustre_swab_fid2path(val);
2122 ptlrpc_req_finished(req);
2127 static void lustre_swab_hai(struct hsm_action_item *h)
2129 __swab32s(&h->hai_len);
2130 __swab32s(&h->hai_action);
2131 lustre_swab_lu_fid(&h->hai_fid);
2132 lustre_swab_lu_fid(&h->hai_dfid);
2133 __swab64s(&h->hai_cookie);
2134 __swab64s(&h->hai_extent.offset);
2135 __swab64s(&h->hai_extent.length);
2136 __swab64s(&h->hai_gid);
2139 static void lustre_swab_hal(struct hsm_action_list *h)
2141 struct hsm_action_item *hai;
2144 __swab32s(&h->hal_version);
2145 __swab32s(&h->hal_count);
2146 __swab32s(&h->hal_archive_id);
2147 __swab64s(&h->hal_flags);
2149 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2150 lustre_swab_hai(hai);
2153 static void lustre_swab_kuch(struct kuc_hdr *l)
2155 __swab16s(&l->kuc_magic);
2156 /* __u8 l->kuc_transport */
2157 __swab16s(&l->kuc_msgtype);
2158 __swab16s(&l->kuc_msglen);
2161 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2162 struct lustre_kernelcomm *lk)
2164 struct obd_import *imp = class_exp2cliimp(exp);
2165 __u32 archive = lk->lk_data;
2168 if (lk->lk_group != KUC_GRP_HSM) {
2169 CERROR("Bad copytool group %d\n", lk->lk_group);
2173 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2174 lk->lk_uid, lk->lk_group, lk->lk_flags);
2176 if (lk->lk_flags & LK_FLG_STOP) {
2177 /* Unregister with the coordinator */
2178 rc = mdc_ioc_hsm_ct_unregister(imp);
2180 rc = mdc_ioc_hsm_ct_register(imp, archive);
2187 * Send a message to any listening copytools
2188 * @param val KUC message (kuc_hdr + hsm_action_list)
2189 * @param len total length of message
2191 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2192 size_t len, void *val)
2194 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2195 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2199 if (len < sizeof(*lh) + sizeof(*hal)) {
2200 CERROR("Short HSM message %zu < %zu\n", len,
2201 sizeof(*lh) + sizeof(*hal));
2204 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2205 lustre_swab_kuch(lh);
2206 lustre_swab_hal(hal);
2207 } else if (lh->kuc_magic != KUC_MAGIC) {
2208 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2212 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2214 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2215 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2217 /* Broadcast to HSM listeners */
2218 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2224 * callback function passed to kuc for re-registering each HSM copytool
2225 * running on MDC, after MDT shutdown/recovery.
2226 * @param data copytool registration data
2227 * @param cb_arg callback argument (obd_import)
2229 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2231 struct kkuc_ct_data *kcd = data;
2232 struct obd_import *imp = (struct obd_import *)cb_arg;
2235 if (kcd == NULL || kcd->kcd_magic != KKUC_CT_DATA_MAGIC)
2238 CDEBUG(D_HA, "%s: recover copytool registration to MDT (archive=%#x)\n",
2239 imp->imp_obd->obd_name, kcd->kcd_archive);
2240 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_archive);
2242 /* ignore error if the copytool is already registered */
2243 return (rc == -EEXIST) ? 0 : rc;
2247 * Re-establish all kuc contexts with MDT
2248 * after MDT shutdown/recovery.
2250 static int mdc_kuc_reregister(struct obd_import *imp)
2252 /* re-register HSM agents */
2253 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2254 mdc_hsm_ct_reregister, imp);
2257 static int mdc_set_info_async(const struct lu_env *env,
2258 struct obd_export *exp,
2259 u32 keylen, void *key,
2260 u32 vallen, void *val,
2261 struct ptlrpc_request_set *set)
2263 struct obd_import *imp = class_exp2cliimp(exp);
2267 if (KEY_IS(KEY_READ_ONLY)) {
2268 if (vallen != sizeof(int))
2271 spin_lock(&imp->imp_lock);
2272 if (*((int *)val)) {
2273 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2274 imp->imp_connect_data.ocd_connect_flags |=
2277 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2278 imp->imp_connect_data.ocd_connect_flags &=
2279 ~OBD_CONNECT_RDONLY;
2281 spin_unlock(&imp->imp_lock);
2283 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2284 keylen, key, vallen, val, set);
2287 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2288 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2289 keylen, key, vallen, val, set);
2292 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2293 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2298 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2299 __u32 *default_easize = val;
2301 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2305 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2309 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2310 __u32 keylen, void *key, __u32 *vallen, void *val)
2314 if (KEY_IS(KEY_MAX_EASIZE)) {
2315 __u32 mdsize, *max_easize;
2317 if (*vallen != sizeof(int))
2319 mdsize = *(__u32 *)val;
2320 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2321 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2323 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2325 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2326 __u32 *default_easize;
2328 if (*vallen != sizeof(int))
2330 default_easize = val;
2331 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2333 } else if (KEY_IS(KEY_CONN_DATA)) {
2334 struct obd_import *imp = class_exp2cliimp(exp);
2335 struct obd_connect_data *data = val;
2337 if (*vallen != sizeof(*data))
2340 *data = imp->imp_connect_data;
2342 } else if (KEY_IS(KEY_TGT_COUNT)) {
2343 *((__u32 *)val) = 1;
2347 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2352 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2353 struct ptlrpc_request **request)
2355 struct ptlrpc_request *req;
2360 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2364 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2366 ptlrpc_request_free(req);
2370 mdc_pack_body(req, fid, 0, 0, -1, 0);
2372 ptlrpc_request_set_replen(req);
2374 rc = ptlrpc_queue_wait(req);
2376 ptlrpc_req_finished(req);
2382 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2383 enum obd_import_event event)
2385 struct client_obd *cli = &obd->u.cli;
2388 LASSERT(imp->imp_obd == obd);
2391 case IMP_EVENT_DISCON:
2392 spin_lock(&cli->cl_loi_list_lock);
2393 cli->cl_avail_grant = 0;
2394 cli->cl_lost_grant = 0;
2395 spin_unlock(&cli->cl_loi_list_lock);
2397 case IMP_EVENT_INACTIVE:
2399 * Flush current sequence to make client obtain new one
2400 * from server in case of disconnect/reconnect.
2402 down_read(&cli->cl_seq_rwsem);
2404 seq_client_flush(cli->cl_seq);
2405 up_read(&cli->cl_seq_rwsem);
2407 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2409 case IMP_EVENT_INVALIDATE: {
2410 struct ldlm_namespace *ns = obd->obd_namespace;
2414 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2416 env = cl_env_get(&refcheck);
2418 /* Reset grants. All pages go to failing rpcs due to
2419 * the invalid import.
2421 osc_io_unplug(env, cli, NULL);
2423 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2424 osc_ldlm_resource_invalidate,
2426 cl_env_put(env, &refcheck);
2427 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2433 case IMP_EVENT_ACTIVE:
2434 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2435 /* redo the kuc registration after reconnecting */
2437 rc = mdc_kuc_reregister(imp);
2439 case IMP_EVENT_OCD: {
2440 struct obd_connect_data *ocd = &imp->imp_connect_data;
2442 if (OCD_HAS_FLAG(ocd, GRANT))
2443 osc_init_grant(cli, ocd);
2445 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2448 case IMP_EVENT_DEACTIVATE:
2449 case IMP_EVENT_ACTIVATE:
2452 CERROR("Unknown import event %x\n", event);
2458 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2459 struct lu_fid *fid, struct md_op_data *op_data)
2461 struct client_obd *cli = &exp->exp_obd->u.cli;
2466 down_read(&cli->cl_seq_rwsem);
2468 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2469 up_read(&cli->cl_seq_rwsem);
2474 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2476 struct client_obd *cli = &exp->exp_obd->u.cli;
2477 return &cli->cl_target_uuid;
2481 * Determine whether the lock can be canceled before replaying it during
2482 * recovery, non zero value will be return if the lock can be canceled,
2483 * or zero returned for not
2485 static int mdc_cancel_weight(struct ldlm_lock *lock)
2487 if (lock->l_resource->lr_type != LDLM_IBITS)
2490 /* FIXME: if we ever get into a situation where there are too many
2491 * opened files with open locks on a single node, then we really
2492 * should replay these open locks to reget it */
2493 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2499 static int mdc_resource_inode_free(struct ldlm_resource *res)
2501 if (res->lr_lvb_inode)
2502 res->lr_lvb_inode = NULL;
2507 static struct ldlm_valblock_ops inode_lvbo = {
2508 .lvbo_free = mdc_resource_inode_free
2511 static int mdc_llog_init(struct obd_device *obd)
2513 struct obd_llog_group *olg = &obd->obd_olg;
2514 struct llog_ctxt *ctxt;
2519 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2524 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2525 llog_initiator_connect(ctxt);
2526 llog_ctxt_put(ctxt);
2531 static void mdc_llog_finish(struct obd_device *obd)
2533 struct llog_ctxt *ctxt;
2537 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2539 llog_cleanup(NULL, ctxt);
2544 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2550 rc = osc_setup_common(obd, cfg);
2554 #ifdef CONFIG_PROC_FS
2555 obd->obd_vars = lprocfs_mdc_obd_vars;
2556 lprocfs_obd_setup(obd, false);
2557 lprocfs_alloc_md_stats(obd, 0);
2560 sptlrpc_lprocfs_cliobd_attach(obd);
2561 ptlrpc_lprocfs_register_obd(obd);
2563 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2565 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2567 rc = mdc_llog_init(obd);
2569 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2571 GOTO(err_llog_cleanup, rc);
2574 rc = mdc_changelog_cdev_init(obd);
2576 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2578 GOTO(err_changelog_cleanup, rc);
2583 err_changelog_cleanup:
2584 mdc_llog_finish(obd);
2586 ptlrpc_lprocfs_unregister_obd(obd);
2587 lprocfs_free_md_stats(obd);
2589 osc_cleanup_common(obd);
2593 /* Initialize the default and maximum LOV EA sizes. This allows
2594 * us to make MDS RPCs with large enough reply buffers to hold a default
2595 * sized EA without having to calculate this (via a call into the
2596 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2597 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2598 * a large number of stripes is possible. If a larger reply buffer is
2599 * required it will be reallocated in the ptlrpc layer due to overflow.
2601 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2604 struct obd_device *obd = exp->exp_obd;
2605 struct client_obd *cli = &obd->u.cli;
2608 if (cli->cl_max_mds_easize < easize)
2609 cli->cl_max_mds_easize = easize;
2611 if (cli->cl_default_mds_easize < def_easize)
2612 cli->cl_default_mds_easize = def_easize;
2617 static int mdc_precleanup(struct obd_device *obd)
2621 osc_precleanup_common(obd);
2622 mdc_changelog_cdev_finish(obd);
2624 obd_cleanup_client_import(obd);
2625 ptlrpc_lprocfs_unregister_obd(obd);
2626 lprocfs_free_md_stats(obd);
2627 mdc_llog_finish(obd);
2631 static int mdc_cleanup(struct obd_device *obd)
2633 return osc_cleanup_common(obd);
2636 int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2638 struct lustre_cfg *lcfg = buf;
2641 rc = class_process_proc_param(PARAM_MDC, obd->obd_vars, lcfg, obd);
2642 return (rc > 0 ? 0: rc);
2645 static struct obd_ops mdc_obd_ops = {
2646 .o_owner = THIS_MODULE,
2647 .o_setup = mdc_setup,
2648 .o_precleanup = mdc_precleanup,
2649 .o_cleanup = mdc_cleanup,
2650 .o_add_conn = client_import_add_conn,
2651 .o_del_conn = client_import_del_conn,
2652 .o_connect = client_connect_import,
2653 .o_reconnect = osc_reconnect,
2654 .o_disconnect = osc_disconnect,
2655 .o_iocontrol = mdc_iocontrol,
2656 .o_set_info_async = mdc_set_info_async,
2657 .o_statfs = mdc_statfs,
2658 .o_fid_init = client_fid_init,
2659 .o_fid_fini = client_fid_fini,
2660 .o_fid_alloc = mdc_fid_alloc,
2661 .o_import_event = mdc_import_event,
2662 .o_get_info = mdc_get_info,
2663 .o_process_config = mdc_process_config,
2664 .o_get_uuid = mdc_get_uuid,
2665 .o_quotactl = mdc_quotactl,
2668 static struct md_ops mdc_md_ops = {
2669 .m_get_root = mdc_get_root,
2670 .m_null_inode = mdc_null_inode,
2671 .m_close = mdc_close,
2672 .m_create = mdc_create,
2673 .m_enqueue = mdc_enqueue,
2674 .m_getattr = mdc_getattr,
2675 .m_getattr_name = mdc_getattr_name,
2676 .m_intent_lock = mdc_intent_lock,
2678 .m_rename = mdc_rename,
2679 .m_setattr = mdc_setattr,
2680 .m_setxattr = mdc_setxattr,
2681 .m_getxattr = mdc_getxattr,
2682 .m_fsync = mdc_fsync,
2683 .m_file_resync = mdc_file_resync,
2684 .m_read_page = mdc_read_page,
2685 .m_unlink = mdc_unlink,
2686 .m_cancel_unused = mdc_cancel_unused,
2687 .m_init_ea_size = mdc_init_ea_size,
2688 .m_set_lock_data = mdc_set_lock_data,
2689 .m_lock_match = mdc_lock_match,
2690 .m_get_lustre_md = mdc_get_lustre_md,
2691 .m_free_lustre_md = mdc_free_lustre_md,
2692 .m_set_open_replay_data = mdc_set_open_replay_data,
2693 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2694 .m_intent_getattr_async = mdc_intent_getattr_async,
2695 .m_revalidate_lock = mdc_revalidate_lock
2698 static int __init mdc_init(void)
2700 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2701 LUSTRE_MDC_NAME, &mdc_device_type);
2704 static void __exit mdc_exit(void)
2706 class_unregister_type(LUSTRE_MDC_NAME);
2709 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2710 MODULE_DESCRIPTION("Lustre Metadata Client");
2711 MODULE_VERSION(LUSTRE_VERSION_STRING);
2712 MODULE_LICENSE("GPL");
2714 module_init(mdc_init);
2715 module_exit(mdc_exit);