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,
339 /* get SELinux policy info if any */
340 rc = sptlrpc_get_sepol(req);
342 ptlrpc_request_free(req);
345 req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
346 strlen(req->rq_sepol) ?
347 strlen(req->rq_sepol) + 1 : 0);
349 /* Flush local XATTR locks to get rid of a possible cancel RPC */
350 if (opcode == MDS_REINT && fid_is_sane(fid) &&
351 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
352 struct list_head cancels = LIST_HEAD_INIT(cancels);
355 /* Without that packing would fail */
357 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
360 count = mdc_resource_get_unused(exp, fid,
362 MDS_INODELOCK_XATTR);
364 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
366 ptlrpc_request_free(req);
370 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
372 ptlrpc_request_free(req);
377 if (opcode == MDS_REINT) {
378 struct mdt_rec_setxattr *rec;
380 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
381 sizeof(struct mdt_rec_reint));
382 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
383 rec->sx_opcode = REINT_SETXATTR;
384 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
385 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
386 rec->sx_cap = cfs_curproc_cap_pack();
387 rec->sx_suppgid1 = suppgid;
388 rec->sx_suppgid2 = -1;
390 rec->sx_valid = valid | OBD_MD_FLCTIME;
391 rec->sx_time = ktime_get_real_seconds();
392 rec->sx_size = output_size;
393 rec->sx_flags = flags;
395 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
399 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
400 memcpy(tmp, xattr_name, xattr_namelen);
403 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
404 memcpy(tmp, input, input_size);
407 mdc_file_sepol_pack(req);
409 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
410 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
411 RCL_SERVER, output_size);
412 ptlrpc_request_set_replen(req);
415 if (opcode == MDS_REINT)
416 mdc_get_mod_rpc_slot(req, NULL);
418 rc = ptlrpc_queue_wait(req);
420 if (opcode == MDS_REINT)
421 mdc_put_mod_rpc_slot(req, NULL);
424 ptlrpc_req_finished(req);
430 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
431 u64 obd_md_valid, const char *name,
432 const void *value, size_t value_size,
433 unsigned int xattr_flags, u32 suppgid,
434 struct ptlrpc_request **req)
436 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
437 obd_md_valid == OBD_MD_FLXATTRRM);
439 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
440 fid, MDS_REINT, obd_md_valid, name,
441 value, value_size, 0, xattr_flags, suppgid,
445 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
446 u64 obd_md_valid, const char *name, size_t buf_size,
447 struct ptlrpc_request **req)
449 struct mdt_body *body;
452 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
453 obd_md_valid == OBD_MD_FLXATTRLS);
455 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
456 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
461 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
463 GOTO(out, rc = -EPROTO);
465 /* only detect the xattr size */
467 /* LU-11109: Older MDTs do not distinguish
468 * between nonexistent xattrs and zero length
469 * values in this case. Newer MDTs will return
470 * -ENODATA or set OBD_MD_FLXATTR. */
471 GOTO(out, rc = body->mbo_eadatasize);
474 if (body->mbo_eadatasize == 0) {
475 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
476 * success so that we can distinguish between
477 * zero length value and nonexistent xattr.
479 * If OBD_MD_FLXATTR is not set then we keep
480 * the old behavior and return -ENODATA for
481 * getxattr() when mbo_eadatasize is 0. But
482 * -ENODATA only makes sense for getxattr()
483 * and not for listxattr(). */
484 if (body->mbo_valid & OBD_MD_FLXATTR)
486 else if (obd_md_valid == OBD_MD_FLXATTR)
487 GOTO(out, rc = -ENODATA);
492 GOTO(out, rc = body->mbo_eadatasize);
495 ptlrpc_req_finished(*req);
502 #ifdef CONFIG_FS_POSIX_ACL
503 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
505 struct req_capsule *pill = &req->rq_pill;
506 struct mdt_body *body = md->body;
507 struct posix_acl *acl;
512 if (!body->mbo_aclsize)
515 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
520 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
525 CERROR("convert xattr to acl: %d\n", rc);
529 rc = posix_acl_valid(&init_user_ns, acl);
531 CERROR("validate acl: %d\n", rc);
532 posix_acl_release(acl);
540 #define mdc_unpack_acl(req, md) 0
543 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
544 struct obd_export *dt_exp, struct obd_export *md_exp,
545 struct lustre_md *md)
547 struct req_capsule *pill = &req->rq_pill;
552 memset(md, 0, sizeof(*md));
554 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
555 LASSERT(md->body != NULL);
557 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
558 if (!S_ISREG(md->body->mbo_mode)) {
559 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
560 "regular file, but is not\n");
561 GOTO(out, rc = -EPROTO);
564 if (md->body->mbo_eadatasize == 0) {
565 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
566 "but eadatasize 0\n");
567 GOTO(out, rc = -EPROTO);
570 md->layout.lb_len = md->body->mbo_eadatasize;
571 md->layout.lb_buf = req_capsule_server_sized_get(pill,
574 if (md->layout.lb_buf == NULL)
575 GOTO(out, rc = -EPROTO);
576 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
577 const union lmv_mds_md *lmv;
580 if (!S_ISDIR(md->body->mbo_mode)) {
581 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
582 "directory, but is not\n");
583 GOTO(out, rc = -EPROTO);
586 lmv_size = md->body->mbo_eadatasize;
588 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
589 "but eadatasize 0\n");
593 if (md->body->mbo_valid & OBD_MD_MEA) {
594 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
597 GOTO(out, rc = -EPROTO);
599 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
603 if (rc < (typeof(rc))sizeof(*md->lmv)) {
604 CDEBUG(D_INFO, "size too small: "
605 "rc < sizeof(*md->lmv) (%d < %d)\n",
606 rc, (int)sizeof(*md->lmv));
607 GOTO(out, rc = -EPROTO);
613 if (md->body->mbo_valid & OBD_MD_FLACL) {
614 /* for ACL, it's possible that FLACL is set but aclsize is zero.
615 * only when aclsize != 0 there's an actual segment for ACL
618 if (md->body->mbo_aclsize) {
619 rc = mdc_unpack_acl(req, md);
622 #ifdef CONFIG_FS_POSIX_ACL
624 md->posix_acl = NULL;
632 #ifdef CONFIG_FS_POSIX_ACL
633 posix_acl_release(md->posix_acl);
639 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
645 void mdc_replay_open(struct ptlrpc_request *req)
647 struct md_open_data *mod = req->rq_cb_data;
648 struct ptlrpc_request *close_req;
649 struct obd_client_handle *och;
650 struct lustre_handle old_open_handle = { };
651 struct mdt_body *body;
655 DEBUG_REQ(D_ERROR, req,
656 "Can't properly replay without open data.");
661 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
662 LASSERT(body != NULL);
664 spin_lock(&req->rq_lock);
666 if (och && och->och_open_handle.cookie)
667 req->rq_early_free_repbuf = 1;
669 req->rq_early_free_repbuf = 0;
670 spin_unlock(&req->rq_lock);
672 if (req->rq_early_free_repbuf) {
673 struct lustre_handle *file_open_handle;
675 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
677 file_open_handle = &och->och_open_handle;
678 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
679 file_open_handle->cookie, body->mbo_open_handle.cookie);
680 old_open_handle = *file_open_handle;
681 *file_open_handle = body->mbo_open_handle;
684 close_req = mod->mod_close_req;
686 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
687 struct mdt_ioepoch *epoch;
689 LASSERT(opc == MDS_CLOSE);
690 epoch = req_capsule_client_get(&close_req->rq_pill,
694 if (req->rq_early_free_repbuf)
695 LASSERT(old_open_handle.cookie ==
696 epoch->mio_open_handle.cookie);
698 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
699 epoch->mio_open_handle = body->mbo_open_handle;
704 void mdc_commit_open(struct ptlrpc_request *req)
706 struct md_open_data *mod = req->rq_cb_data;
711 * No need to touch md_open_data::mod_och, it holds a reference on
712 * \var mod and will zero references to each other, \var mod will be
713 * freed after that when md_open_data::mod_och will put the reference.
717 * Do not let open request to disappear as it still may be needed
718 * for close rpc to happen (it may happen on evict only, otherwise
719 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
720 * called), just mark this rpc as committed to distinguish these 2
721 * cases, see mdc_close() for details. The open request reference will
722 * be put along with freeing \var mod.
724 ptlrpc_request_addref(req);
725 spin_lock(&req->rq_lock);
726 req->rq_committed = 1;
727 spin_unlock(&req->rq_lock);
728 req->rq_cb_data = NULL;
732 int mdc_set_open_replay_data(struct obd_export *exp,
733 struct obd_client_handle *och,
734 struct lookup_intent *it)
736 struct md_open_data *mod;
737 struct mdt_rec_create *rec;
738 struct mdt_body *body;
739 struct ptlrpc_request *open_req = it->it_request;
740 struct obd_import *imp = open_req->rq_import;
743 if (!open_req->rq_replay)
746 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
747 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
748 LASSERT(rec != NULL);
749 /* Incoming message in my byte order (it's been swabbed). */
750 /* Outgoing messages always in my byte order. */
751 LASSERT(body != NULL);
753 /* Only if the import is replayable, we set replay_open data */
754 if (och && imp->imp_replayable) {
755 mod = obd_mod_alloc();
757 DEBUG_REQ(D_ERROR, open_req,
758 "Can't allocate md_open_data");
763 * Take a reference on \var mod, to be freed on mdc_close().
764 * It protects \var mod from being freed on eviction (commit
765 * callback is called despite rq_replay flag).
766 * Another reference for \var och.
771 spin_lock(&open_req->rq_lock);
774 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
775 it_disposition(it, DISP_OPEN_STRIPE);
776 mod->mod_open_req = open_req;
777 open_req->rq_cb_data = mod;
778 open_req->rq_commit_cb = mdc_commit_open;
779 open_req->rq_early_free_repbuf = 1;
780 spin_unlock(&open_req->rq_lock);
783 rec->cr_fid2 = body->mbo_fid1;
784 rec->cr_open_handle_old = body->mbo_open_handle;
785 open_req->rq_replay_cb = mdc_replay_open;
786 if (!fid_is_sane(&body->mbo_fid1)) {
787 DEBUG_REQ(D_ERROR, open_req,
788 "saving replay request with insane FID " DFID,
789 PFID(&body->mbo_fid1));
793 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
797 static void mdc_free_open(struct md_open_data *mod)
801 if (mod->mod_is_create == 0 &&
802 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
806 * No reason to asssert here if the open request has
807 * rq_replay == 1. It means that mdc_close failed, and
808 * close request wasn`t sent. It is not fatal to client.
809 * The worst thing is eviction if the client gets open lock
812 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
813 "= %d\n", mod->mod_open_req->rq_replay);
815 ptlrpc_request_committed(mod->mod_open_req, committed);
816 if (mod->mod_close_req)
817 ptlrpc_request_committed(mod->mod_close_req, committed);
820 int mdc_clear_open_replay_data(struct obd_export *exp,
821 struct obd_client_handle *och)
823 struct md_open_data *mod = och->och_mod;
827 * It is possible to not have \var mod in a case of eviction between
828 * lookup and ll_file_open().
833 LASSERT(mod != LP_POISON);
834 LASSERT(mod->mod_open_req != NULL);
836 spin_lock(&mod->mod_open_req->rq_lock);
838 mod->mod_och->och_open_handle.cookie = 0;
839 mod->mod_open_req->rq_early_free_repbuf = 0;
840 spin_unlock(&mod->mod_open_req->rq_lock);
850 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
851 struct md_open_data *mod, struct ptlrpc_request **request)
853 struct obd_device *obd = class_exp2obd(exp);
854 struct ptlrpc_request *req;
855 struct req_format *req_fmt;
856 size_t u32_count = 0;
861 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
862 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
865 if (op_data->op_bias & MDS_CLOSE_INTENT) {
866 req_fmt = &RQF_MDS_CLOSE_INTENT;
867 if (op_data->op_bias & MDS_HSM_RELEASE) {
868 /* allocate a FID for volatile file */
869 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
872 CERROR("%s: "DFID" allocating FID: rc = %d\n",
873 obd->obd_name, PFID(&op_data->op_fid1),
875 /* save the errcode and proceed to close */
879 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
880 size_t count = op_data->op_data_size / sizeof(__u32);
882 if (count > INLINE_RESYNC_ARRAY_SIZE)
886 req_fmt = &RQF_MDS_CLOSE;
890 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
893 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
895 /* Ensure that this close's handle is fixed up during replay. */
896 if (likely(mod != NULL)) {
897 LASSERTF(mod->mod_open_req != NULL &&
898 mod->mod_open_req->rq_type != LI_POISON,
899 "POISONED open %p!\n", mod->mod_open_req);
901 mod->mod_close_req = req;
903 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
904 /* We no longer want to preserve this open for replay even
905 * though the open was committed. b=3632, b=3633 */
906 spin_lock(&mod->mod_open_req->rq_lock);
907 mod->mod_open_req->rq_replay = 0;
908 spin_unlock(&mod->mod_open_req->rq_lock);
910 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
914 * TODO: repeat close after errors
916 CWARN("%s: close of FID "DFID" failed, file reference will be "
917 "dropped when this client unmounts or is evicted\n",
918 obd->obd_name, PFID(&op_data->op_fid1));
919 GOTO(out, rc = -ENOMEM);
923 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
924 u32_count * sizeof(__u32));
926 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
928 ptlrpc_request_free(req);
933 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
934 * portal whose threads are not taking any DLM locks and are therefore
935 * always progressing */
936 req->rq_request_portal = MDS_READPAGE_PORTAL;
937 ptlrpc_at_set_req_timeout(req);
940 mdc_close_pack(req, op_data);
942 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
943 obd->u.cli.cl_default_mds_easize);
945 ptlrpc_request_set_replen(req);
947 mdc_get_mod_rpc_slot(req, NULL);
948 rc = ptlrpc_queue_wait(req);
949 mdc_put_mod_rpc_slot(req, NULL);
951 if (req->rq_repmsg == NULL) {
952 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
955 rc = req->rq_status ?: -EIO;
956 } else if (rc == 0 || rc == -EAGAIN) {
957 struct mdt_body *body;
959 rc = lustre_msg_get_status(req->rq_repmsg);
960 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
961 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
966 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
969 } else if (rc == -ESTALE) {
971 * it can be allowed error after 3633 if open was committed and
972 * server failed before close was sent. Let's check if mod
973 * exists and return no error in that case
976 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
977 LASSERT(mod->mod_open_req != NULL);
978 if (mod->mod_open_req->rq_committed)
986 mod->mod_close_req = NULL;
987 /* Since now, mod is accessed through open_req only,
988 * thus close req does not keep a reference on mod anymore. */
993 RETURN(rc < 0 ? rc : saved_rc);
996 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
997 u64 offset, struct page **pages, int npages,
998 struct ptlrpc_request **request)
1000 struct ptlrpc_request *req;
1001 struct ptlrpc_bulk_desc *desc;
1003 wait_queue_head_t waitq;
1005 struct l_wait_info lwi;
1010 init_waitqueue_head(&waitq);
1013 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1017 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1019 ptlrpc_request_free(req);
1023 req->rq_request_portal = MDS_READPAGE_PORTAL;
1024 ptlrpc_at_set_req_timeout(req);
1026 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1027 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1029 &ptlrpc_bulk_kiov_pin_ops);
1031 ptlrpc_req_finished(req);
1035 /* NB req now owns desc and will free it when it gets freed */
1036 for (i = 0; i < npages; i++)
1037 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1040 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1042 ptlrpc_request_set_replen(req);
1043 rc = ptlrpc_queue_wait(req);
1045 ptlrpc_req_finished(req);
1046 if (rc != -ETIMEDOUT)
1050 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1051 CERROR("%s: too many resend retries: rc = %d\n",
1052 exp->exp_obd->obd_name, -EIO);
1055 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1057 l_wait_event(waitq, 0, &lwi);
1062 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1063 req->rq_bulk->bd_nob_transferred);
1065 ptlrpc_req_finished(req);
1069 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1070 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1071 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1072 PAGE_SIZE * npages);
1073 ptlrpc_req_finished(req);
1081 static void mdc_release_page(struct page *page, int remove)
1085 if (likely(page->mapping != NULL))
1086 truncate_complete_page(page->mapping, page);
1092 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1093 __u64 *start, __u64 *end, int hash64)
1096 * Complement of hash is used as an index so that
1097 * radix_tree_gang_lookup() can be used to find a page with starting
1098 * hash _smaller_ than one we are looking for.
1100 unsigned long offset = hash_x_index(*hash, hash64);
1104 spin_lock_irq(&mapping->tree_lock);
1105 found = radix_tree_gang_lookup(&mapping->page_tree,
1106 (void **)&page, offset, 1);
1107 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1108 struct lu_dirpage *dp;
1111 spin_unlock_irq(&mapping->tree_lock);
1113 * In contrast to find_lock_page() we are sure that directory
1114 * page cannot be truncated (while DLM lock is held) and,
1115 * hence, can avoid restart.
1117 * In fact, page cannot be locked here at all, because
1118 * mdc_read_page_remote does synchronous io.
1120 wait_on_page_locked(page);
1121 if (PageUptodate(page)) {
1123 if (BITS_PER_LONG == 32 && hash64) {
1124 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1125 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1126 *hash = *hash >> 32;
1128 *start = le64_to_cpu(dp->ldp_hash_start);
1129 *end = le64_to_cpu(dp->ldp_hash_end);
1131 if (unlikely(*start == 1 && *hash == 0))
1134 LASSERTF(*start <= *hash, "start = %#llx"
1135 ",end = %#llx,hash = %#llx\n",
1136 *start, *end, *hash);
1137 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1138 " hash %#llx\n", offset, *start, *end, *hash);
1141 mdc_release_page(page, 0);
1143 } else if (*end != *start && *hash == *end) {
1145 * upon hash collision, remove this page,
1146 * otherwise put page reference, and
1147 * mdc_read_page_remote() will issue RPC to
1148 * fetch the page we want.
1151 mdc_release_page(page,
1152 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1157 page = ERR_PTR(-EIO);
1160 spin_unlock_irq(&mapping->tree_lock);
1167 * Adjust a set of pages, each page containing an array of lu_dirpages,
1168 * so that each page can be used as a single logical lu_dirpage.
1170 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1171 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1172 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1173 * value is used as a cookie to request the next lu_dirpage in a
1174 * directory listing that spans multiple pages (two in this example):
1177 * .|--------v------- -----.
1178 * |s|e|f|p|ent|ent| ... |ent|
1179 * '--|-------------- -----' Each PAGE contains a single
1180 * '------. lu_dirpage.
1181 * .---------v------- -----.
1182 * |s|e|f|p|ent| 0 | ... | 0 |
1183 * '----------------- -----'
1185 * However, on hosts where the native VM page size (PAGE_SIZE) is
1186 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1187 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1188 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1189 * after it in the same PAGE (arrows simplified for brevity, but
1190 * in general e0==s1, e1==s2, etc.):
1192 * .-------------------- -----.
1193 * |s0|e0|f0|p|ent|ent| ... |ent|
1194 * |---v---------------- -----|
1195 * |s1|e1|f1|p|ent|ent| ... |ent|
1196 * |---v---------------- -----| Here, each PAGE contains
1197 * ... multiple lu_dirpages.
1198 * |---v---------------- -----|
1199 * |s'|e'|f'|p|ent|ent| ... |ent|
1200 * '---|---------------- -----'
1202 * .----------------------------.
1205 * This structure is transformed into a single logical lu_dirpage as follows:
1207 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1208 * labeled 'next PAGE'.
1210 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1211 * a hash collision with the next page exists.
1213 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1214 * to the first entry of the next lu_dirpage.
1216 #if PAGE_SIZE > LU_PAGE_SIZE
1217 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1221 for (i = 0; i < cfs_pgs; i++) {
1222 struct lu_dirpage *dp = kmap(pages[i]);
1223 struct lu_dirpage *first = dp;
1224 struct lu_dirent *end_dirent = NULL;
1225 struct lu_dirent *ent;
1226 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1227 __u32 flags = le32_to_cpu(dp->ldp_flags);
1229 while (--lu_pgs > 0) {
1230 ent = lu_dirent_start(dp);
1231 for (end_dirent = ent; ent != NULL;
1232 end_dirent = ent, ent = lu_dirent_next(ent));
1234 /* Advance dp to next lu_dirpage. */
1235 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1237 /* Check if we've reached the end of the PAGE. */
1238 if (!((unsigned long)dp & ~PAGE_MASK))
1241 /* Save the hash and flags of this lu_dirpage. */
1242 hash_end = le64_to_cpu(dp->ldp_hash_end);
1243 flags = le32_to_cpu(dp->ldp_flags);
1245 /* Check if lu_dirpage contains no entries. */
1246 if (end_dirent == NULL)
1249 /* Enlarge the end entry lde_reclen from 0 to
1250 * first entry of next lu_dirpage. */
1251 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1252 end_dirent->lde_reclen =
1253 cpu_to_le16((char *)(dp->ldp_entries) -
1254 (char *)end_dirent);
1257 first->ldp_hash_end = hash_end;
1258 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1259 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1263 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1266 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1267 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1269 /* parameters for readdir page */
1270 struct readpage_param {
1271 struct md_op_data *rp_mod;
1274 struct obd_export *rp_exp;
1275 struct md_callback *rp_cb;
1278 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1279 static inline void delete_from_page_cache(struct page *page)
1281 remove_from_page_cache(page);
1287 * Read pages from server.
1289 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1290 * a header lu_dirpage which describes the start/end hash, and whether this
1291 * page is empty (contains no dir entry) or hash collide with next page.
1292 * After client receives reply, several pages will be integrated into dir page
1293 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1294 * lu_dirpage for this integrated page will be adjusted.
1296 static int mdc_read_page_remote(void *data, struct page *page0)
1298 struct readpage_param *rp = data;
1299 struct page **page_pool;
1301 struct lu_dirpage *dp;
1302 struct md_op_data *op_data = rp->rp_mod;
1303 struct ptlrpc_request *req;
1305 struct inode *inode;
1307 int rd_pgs = 0; /* number of pages actually read */
1313 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1314 inode = op_data->op_data;
1315 fid = &op_data->op_fid1;
1316 LASSERT(inode != NULL);
1318 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1319 if (page_pool != NULL) {
1320 page_pool[0] = page0;
1326 for (npages = 1; npages < max_pages; npages++) {
1327 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1331 page_pool[npages] = page;
1334 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1336 /* page0 is special, which was added into page cache early */
1337 delete_from_page_cache(page0);
1341 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1343 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1344 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1346 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1348 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1350 SetPageUptodate(page0);
1354 ptlrpc_req_finished(req);
1355 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1356 for (i = 1; i < npages; i++) {
1357 unsigned long offset;
1361 page = page_pool[i];
1363 if (rc < 0 || i >= rd_pgs) {
1368 SetPageUptodate(page);
1371 hash = le64_to_cpu(dp->ldp_hash_start);
1374 offset = hash_x_index(hash, rp->rp_hash64);
1376 prefetchw(&page->flags);
1377 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1382 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1383 " rc = %d\n", offset, ret);
1387 if (page_pool != &page0)
1388 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1394 * Read dir page from cache first, if it can not find it, read it from
1395 * server and add into the cache.
1397 * \param[in] exp MDC export
1398 * \param[in] op_data client MD stack parameters, transfering parameters
1399 * between different layers on client MD stack.
1400 * \param[in] cb_op callback required for ldlm lock enqueue during
1402 * \param[in] hash_offset the hash offset of the page to be read
1403 * \param[in] ppage the page to be read
1405 * retval = 0 get the page successfully
1406 * errno(<0) get the page failed
1408 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1409 struct md_callback *cb_op, __u64 hash_offset,
1410 struct page **ppage)
1412 struct lookup_intent it = { .it_op = IT_READDIR };
1414 struct inode *dir = op_data->op_data;
1415 struct address_space *mapping;
1416 struct lu_dirpage *dp;
1419 struct lustre_handle lockh;
1420 struct ptlrpc_request *enq_req = NULL;
1421 struct readpage_param rp_param;
1428 LASSERT(dir != NULL);
1429 mapping = dir->i_mapping;
1431 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1432 cb_op->md_blocking_ast, 0);
1433 if (enq_req != NULL)
1434 ptlrpc_req_finished(enq_req);
1437 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1438 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1443 lockh.cookie = it.it_lock_handle;
1444 mdc_set_lock_data(exp, &lockh, dir, NULL);
1446 rp_param.rp_off = hash_offset;
1447 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1448 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1449 rp_param.rp_hash64);
1451 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1452 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1453 rp_param.rp_off, PTR_ERR(page));
1454 GOTO(out_unlock, rc = PTR_ERR(page));
1455 } else if (page != NULL) {
1457 * XXX nikita: not entirely correct handling of a corner case:
1458 * suppose hash chain of entries with hash value HASH crosses
1459 * border between pages P0 and P1. First both P0 and P1 are
1460 * cached, seekdir() is called for some entry from the P0 part
1461 * of the chain. Later P0 goes out of cache. telldir(HASH)
1462 * happens and finds P1, as it starts with matching hash
1463 * value. Remaining entries from P0 part of the chain are
1464 * skipped. (Is that really a bug?)
1466 * Possible solutions: 0. don't cache P1 is such case, handle
1467 * it as an "overflow" page. 1. invalidate all pages at
1468 * once. 2. use HASH|1 as an index for P1.
1470 GOTO(hash_collision, page);
1473 rp_param.rp_exp = exp;
1474 rp_param.rp_mod = op_data;
1475 page = read_cache_page(mapping,
1476 hash_x_index(rp_param.rp_off,
1477 rp_param.rp_hash64),
1478 mdc_read_page_remote, &rp_param);
1480 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1481 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1482 rp_param.rp_off, PTR_ERR(page));
1483 GOTO(out_unlock, rc = PTR_ERR(page));
1486 wait_on_page_locked(page);
1488 if (!PageUptodate(page)) {
1489 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1490 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1491 rp_param.rp_off, -5);
1494 if (!PageChecked(page))
1495 SetPageChecked(page);
1496 if (PageError(page)) {
1497 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1498 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1499 rp_param.rp_off, -5);
1504 dp = page_address(page);
1505 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1506 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1507 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1508 rp_param.rp_off = hash_offset >> 32;
1510 start = le64_to_cpu(dp->ldp_hash_start);
1511 end = le64_to_cpu(dp->ldp_hash_end);
1512 rp_param.rp_off = hash_offset;
1515 LASSERT(start == rp_param.rp_off);
1516 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1517 #if BITS_PER_LONG == 32
1518 CWARN("Real page-wide hash collision at [%llu %llu] with "
1519 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1520 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1524 * Fetch whole overflow chain...
1532 ldlm_lock_decref(&lockh, it.it_lock_mode);
1536 mdc_release_page(page, 1);
1541 static int mdc_statfs(const struct lu_env *env,
1542 struct obd_export *exp, struct obd_statfs *osfs,
1543 time64_t max_age, __u32 flags)
1545 struct obd_device *obd = class_exp2obd(exp);
1546 struct req_format *fmt;
1547 struct ptlrpc_request *req;
1548 struct obd_statfs *msfs;
1549 struct obd_import *imp = NULL;
1554 * Since the request might also come from lprocfs, so we need
1555 * sync this with client_disconnect_export Bug15684
1557 down_read(&obd->u.cli.cl_sem);
1558 if (obd->u.cli.cl_import)
1559 imp = class_import_get(obd->u.cli.cl_import);
1560 up_read(&obd->u.cli.cl_sem);
1564 fmt = &RQF_MDS_STATFS;
1565 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1566 (flags & OBD_STATFS_SUM))
1567 fmt = &RQF_MDS_STATFS_NEW;
1568 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1571 GOTO(output, rc = -ENOMEM);
1573 if ((flags & OBD_STATFS_SUM) &&
1574 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1575 /* request aggregated states */
1576 struct mdt_body *body;
1578 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1580 GOTO(out, rc = -EPROTO);
1581 body->mbo_valid = OBD_MD_FLAGSTATFS;
1584 ptlrpc_request_set_replen(req);
1586 if (flags & OBD_STATFS_NODELAY) {
1587 /* procfs requests not want stay in wait for avoid deadlock */
1588 req->rq_no_resend = 1;
1589 req->rq_no_delay = 1;
1592 rc = ptlrpc_queue_wait(req);
1594 /* check connection error first */
1595 if (imp->imp_connect_error)
1596 rc = imp->imp_connect_error;
1600 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1602 GOTO(out, rc = -EPROTO);
1607 ptlrpc_req_finished(req);
1609 class_import_put(imp);
1613 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1615 __u32 keylen, vallen;
1619 if (gf->gf_pathlen > PATH_MAX)
1620 RETURN(-ENAMETOOLONG);
1621 if (gf->gf_pathlen < 2)
1624 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1625 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1626 sizeof(struct lu_fid));
1627 OBD_ALLOC(key, keylen);
1630 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1631 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1632 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1633 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1634 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1635 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1637 if (!fid_is_sane(&gf->gf_fid))
1638 GOTO(out, rc = -EINVAL);
1640 /* Val is struct getinfo_fid2path result plus path */
1641 vallen = sizeof(*gf) + gf->gf_pathlen;
1643 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1644 if (rc != 0 && rc != -EREMOTE)
1647 if (vallen <= sizeof(*gf))
1648 GOTO(out, rc = -EPROTO);
1649 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1650 GOTO(out, rc = -EOVERFLOW);
1652 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1653 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1654 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1655 /* only log the last 512 characters of the path */
1656 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1659 OBD_FREE(key, keylen);
1663 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1664 struct hsm_progress_kernel *hpk)
1666 struct obd_import *imp = class_exp2cliimp(exp);
1667 struct hsm_progress_kernel *req_hpk;
1668 struct ptlrpc_request *req;
1672 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1673 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1675 GOTO(out, rc = -ENOMEM);
1677 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1679 /* Copy hsm_progress struct */
1680 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1681 if (req_hpk == NULL)
1682 GOTO(out, rc = -EPROTO);
1685 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1687 ptlrpc_request_set_replen(req);
1689 mdc_get_mod_rpc_slot(req, NULL);
1690 rc = ptlrpc_queue_wait(req);
1691 mdc_put_mod_rpc_slot(req, NULL);
1695 ptlrpc_req_finished(req);
1699 * Send hsm_ct_register to MDS
1701 * \param[in] imp import
1702 * \param[in] archive_count if in bitmap format, it is the bitmap,
1703 * else it is the count of archive_ids
1704 * \param[in] archives if in bitmap format, it is NULL,
1705 * else it is archive_id lists
1707 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1710 struct ptlrpc_request *req;
1711 __u32 *archive_array;
1712 size_t archives_size;
1716 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1720 if (archives != NULL)
1721 archives_size = sizeof(*archive_array) * archive_count;
1723 archives_size = sizeof(archive_count);
1725 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1726 RCL_CLIENT, archives_size);
1728 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1730 ptlrpc_request_free(req);
1734 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1736 archive_array = req_capsule_client_get(&req->rq_pill,
1737 &RMF_MDS_HSM_ARCHIVE);
1738 if (archive_array == NULL)
1739 GOTO(out, rc = -EPROTO);
1741 if (archives != NULL)
1742 memcpy(archive_array, archives, archives_size);
1744 *archive_array = archive_count;
1746 ptlrpc_request_set_replen(req);
1748 rc = mdc_queue_wait(req);
1751 ptlrpc_req_finished(req);
1755 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1756 struct md_op_data *op_data)
1758 struct hsm_current_action *hca = op_data->op_data;
1759 struct hsm_current_action *req_hca;
1760 struct ptlrpc_request *req;
1764 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1765 &RQF_MDS_HSM_ACTION);
1769 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1771 ptlrpc_request_free(req);
1775 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1776 op_data->op_suppgids[0], 0);
1778 ptlrpc_request_set_replen(req);
1780 rc = mdc_queue_wait(req);
1784 req_hca = req_capsule_server_get(&req->rq_pill,
1785 &RMF_MDS_HSM_CURRENT_ACTION);
1786 if (req_hca == NULL)
1787 GOTO(out, rc = -EPROTO);
1793 ptlrpc_req_finished(req);
1797 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1799 struct ptlrpc_request *req;
1803 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1805 MDS_HSM_CT_UNREGISTER);
1807 GOTO(out, rc = -ENOMEM);
1809 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1811 ptlrpc_request_set_replen(req);
1813 rc = mdc_queue_wait(req);
1816 ptlrpc_req_finished(req);
1820 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1821 struct md_op_data *op_data)
1823 struct hsm_user_state *hus = op_data->op_data;
1824 struct hsm_user_state *req_hus;
1825 struct ptlrpc_request *req;
1829 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1830 &RQF_MDS_HSM_STATE_GET);
1834 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1836 ptlrpc_request_free(req);
1840 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1841 op_data->op_suppgids[0], 0);
1843 ptlrpc_request_set_replen(req);
1845 rc = mdc_queue_wait(req);
1849 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1850 if (req_hus == NULL)
1851 GOTO(out, rc = -EPROTO);
1857 ptlrpc_req_finished(req);
1861 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1862 struct md_op_data *op_data)
1864 struct hsm_state_set *hss = op_data->op_data;
1865 struct hsm_state_set *req_hss;
1866 struct ptlrpc_request *req;
1870 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1871 &RQF_MDS_HSM_STATE_SET);
1875 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1877 ptlrpc_request_free(req);
1881 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1882 op_data->op_suppgids[0], 0);
1885 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1886 if (req_hss == NULL)
1887 GOTO(out, rc = -EPROTO);
1890 ptlrpc_request_set_replen(req);
1892 mdc_get_mod_rpc_slot(req, NULL);
1893 rc = ptlrpc_queue_wait(req);
1894 mdc_put_mod_rpc_slot(req, NULL);
1898 ptlrpc_req_finished(req);
1902 static int mdc_ioc_hsm_request(struct obd_export *exp,
1903 struct hsm_user_request *hur)
1905 struct obd_import *imp = class_exp2cliimp(exp);
1906 struct ptlrpc_request *req;
1907 struct hsm_request *req_hr;
1908 struct hsm_user_item *req_hui;
1913 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1915 GOTO(out, rc = -ENOMEM);
1917 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1918 hur->hur_request.hr_itemcount
1919 * sizeof(struct hsm_user_item));
1920 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1921 hur->hur_request.hr_data_len);
1923 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1925 ptlrpc_request_free(req);
1929 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1931 /* Copy hsm_request struct */
1932 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1934 GOTO(out, rc = -EPROTO);
1935 *req_hr = hur->hur_request;
1937 /* Copy hsm_user_item structs */
1938 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1939 if (req_hui == NULL)
1940 GOTO(out, rc = -EPROTO);
1941 memcpy(req_hui, hur->hur_user_item,
1942 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1944 /* Copy opaque field */
1945 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1946 if (req_opaque == NULL)
1947 GOTO(out, rc = -EPROTO);
1948 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1950 ptlrpc_request_set_replen(req);
1952 mdc_get_mod_rpc_slot(req, NULL);
1953 rc = ptlrpc_queue_wait(req);
1954 mdc_put_mod_rpc_slot(req, NULL);
1959 ptlrpc_req_finished(req);
1963 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1964 struct lustre_kernelcomm *lk);
1966 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1967 struct obd_quotactl *oqctl)
1969 struct ptlrpc_request *req;
1970 struct obd_quotactl *oqc;
1974 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1975 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1980 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1983 ptlrpc_request_set_replen(req);
1984 ptlrpc_at_set_req_timeout(req);
1986 rc = ptlrpc_queue_wait(req);
1988 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1990 if (req->rq_repmsg &&
1991 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1994 CERROR ("Can't unpack obd_quotactl\n");
1997 ptlrpc_req_finished(req);
2002 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2003 struct md_op_data *op_data)
2005 struct list_head cancels = LIST_HEAD_INIT(cancels);
2006 struct ptlrpc_request *req;
2008 struct mdc_swap_layouts *msl, *payload;
2011 msl = op_data->op_data;
2013 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2014 * first thing it will do is to cancel the 2 layout
2015 * locks held by this client.
2016 * So the client must cancel its layout locks on the 2 fids
2017 * with the request RPC to avoid extra RPC round trips.
2019 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2020 LCK_EX, MDS_INODELOCK_LAYOUT |
2021 MDS_INODELOCK_XATTR);
2022 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2023 LCK_EX, MDS_INODELOCK_LAYOUT |
2024 MDS_INODELOCK_XATTR);
2026 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2027 &RQF_MDS_SWAP_LAYOUTS);
2029 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2033 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2035 ptlrpc_request_free(req);
2039 mdc_swap_layouts_pack(req, op_data);
2041 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2046 ptlrpc_request_set_replen(req);
2048 rc = ptlrpc_queue_wait(req);
2054 ptlrpc_req_finished(req);
2058 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2059 void *karg, void __user *uarg)
2061 struct obd_device *obd = exp->exp_obd;
2062 struct obd_ioctl_data *data = karg;
2063 struct obd_import *imp = obd->u.cli.cl_import;
2067 if (!try_module_get(THIS_MODULE)) {
2068 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2069 module_name(THIS_MODULE));
2073 case OBD_IOC_FID2PATH:
2074 rc = mdc_ioc_fid2path(exp, karg);
2076 case LL_IOC_HSM_CT_START:
2077 rc = mdc_ioc_hsm_ct_start(exp, karg);
2078 /* ignore if it was already registered on this MDS. */
2082 case LL_IOC_HSM_PROGRESS:
2083 rc = mdc_ioc_hsm_progress(exp, karg);
2085 case LL_IOC_HSM_STATE_GET:
2086 rc = mdc_ioc_hsm_state_get(exp, karg);
2088 case LL_IOC_HSM_STATE_SET:
2089 rc = mdc_ioc_hsm_state_set(exp, karg);
2091 case LL_IOC_HSM_ACTION:
2092 rc = mdc_ioc_hsm_current_action(exp, karg);
2094 case LL_IOC_HSM_REQUEST:
2095 rc = mdc_ioc_hsm_request(exp, karg);
2097 case OBD_IOC_CLIENT_RECOVER:
2098 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2102 case IOC_OSC_SET_ACTIVE:
2103 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2105 case OBD_IOC_PING_TARGET:
2106 rc = ptlrpc_obd_ping(obd);
2109 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2110 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2111 * there'd be no LMV layer thus we might be called here. Eventually
2112 * this code should be removed.
2115 case IOC_OBD_STATFS: {
2116 struct obd_statfs stat_buf = {0};
2118 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2119 GOTO(out, rc = -ENODEV);
2122 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2123 min((int)data->ioc_plen2,
2124 (int)sizeof(struct obd_uuid))))
2125 GOTO(out, rc = -EFAULT);
2127 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2128 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2133 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2134 min((int) data->ioc_plen1,
2135 (int) sizeof(stat_buf))))
2136 GOTO(out, rc = -EFAULT);
2140 case OBD_IOC_QUOTACTL: {
2141 struct if_quotactl *qctl = karg;
2142 struct obd_quotactl *oqctl;
2144 OBD_ALLOC_PTR(oqctl);
2146 GOTO(out, rc = -ENOMEM);
2148 QCTL_COPY(oqctl, qctl);
2149 rc = obd_quotactl(exp, oqctl);
2151 QCTL_COPY(qctl, oqctl);
2152 qctl->qc_valid = QC_MDTIDX;
2153 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2156 OBD_FREE_PTR(oqctl);
2159 case LL_IOC_GET_CONNECT_FLAGS:
2160 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2161 sizeof(*exp_connect_flags_ptr(exp))))
2162 GOTO(out, rc = -EFAULT);
2165 case LL_IOC_LOV_SWAP_LAYOUTS:
2166 rc = mdc_ioc_swap_layouts(exp, karg);
2169 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2170 GOTO(out, rc = -ENOTTY);
2173 module_put(THIS_MODULE);
2178 static int mdc_get_info_rpc(struct obd_export *exp,
2179 u32 keylen, void *key,
2180 u32 vallen, void *val)
2182 struct obd_import *imp = class_exp2cliimp(exp);
2183 struct ptlrpc_request *req;
2188 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2192 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2193 RCL_CLIENT, keylen);
2194 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2195 RCL_CLIENT, sizeof(vallen));
2197 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2199 ptlrpc_request_free(req);
2203 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2204 memcpy(tmp, key, keylen);
2205 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2206 memcpy(tmp, &vallen, sizeof(vallen));
2208 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2209 RCL_SERVER, vallen);
2210 ptlrpc_request_set_replen(req);
2212 rc = ptlrpc_queue_wait(req);
2213 /* -EREMOTE means the get_info result is partial, and it needs to
2214 * continue on another MDT, see fid2path part in lmv_iocontrol */
2215 if (rc == 0 || rc == -EREMOTE) {
2216 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2217 memcpy(val, tmp, vallen);
2218 if (ptlrpc_rep_need_swab(req)) {
2219 if (KEY_IS(KEY_FID2PATH))
2220 lustre_swab_fid2path(val);
2223 ptlrpc_req_finished(req);
2228 static void lustre_swab_hai(struct hsm_action_item *h)
2230 __swab32s(&h->hai_len);
2231 __swab32s(&h->hai_action);
2232 lustre_swab_lu_fid(&h->hai_fid);
2233 lustre_swab_lu_fid(&h->hai_dfid);
2234 __swab64s(&h->hai_cookie);
2235 __swab64s(&h->hai_extent.offset);
2236 __swab64s(&h->hai_extent.length);
2237 __swab64s(&h->hai_gid);
2240 static void lustre_swab_hal(struct hsm_action_list *h)
2242 struct hsm_action_item *hai;
2245 __swab32s(&h->hal_version);
2246 __swab32s(&h->hal_count);
2247 __swab32s(&h->hal_archive_id);
2248 __swab64s(&h->hal_flags);
2250 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2251 lustre_swab_hai(hai);
2254 static void lustre_swab_kuch(struct kuc_hdr *l)
2256 __swab16s(&l->kuc_magic);
2257 /* __u8 l->kuc_transport */
2258 __swab16s(&l->kuc_msgtype);
2259 __swab16s(&l->kuc_msglen);
2262 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2263 struct lustre_kernelcomm *lk)
2265 struct obd_import *imp = class_exp2cliimp(exp);
2268 if (lk->lk_group != KUC_GRP_HSM) {
2269 CERROR("Bad copytool group %d\n", lk->lk_group);
2273 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2274 lk->lk_uid, lk->lk_group, lk->lk_flags);
2276 if (lk->lk_flags & LK_FLG_STOP) {
2277 /* Unregister with the coordinator */
2278 rc = mdc_ioc_hsm_ct_unregister(imp);
2280 __u32 *archives = NULL;
2282 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2283 archives = lk->lk_data;
2285 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2292 * Send a message to any listening copytools
2293 * @param val KUC message (kuc_hdr + hsm_action_list)
2294 * @param len total length of message
2296 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2297 size_t len, void *val)
2299 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2300 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2304 if (len < sizeof(*lh) + sizeof(*hal)) {
2305 CERROR("Short HSM message %zu < %zu\n", len,
2306 sizeof(*lh) + sizeof(*hal));
2309 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2310 lustre_swab_kuch(lh);
2311 lustre_swab_hal(hal);
2312 } else if (lh->kuc_magic != KUC_MAGIC) {
2313 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2317 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2319 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2320 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2322 /* Broadcast to HSM listeners */
2323 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2329 * callback function passed to kuc for re-registering each HSM copytool
2330 * running on MDC, after MDT shutdown/recovery.
2331 * @param data copytool registration data
2332 * @param cb_arg callback argument (obd_import)
2334 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2336 struct obd_import *imp = (struct obd_import *)cb_arg;
2337 struct kkuc_ct_data *kcd = data;
2338 __u32 *archives = NULL;
2342 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2343 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2346 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2347 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2348 "(archive=%#x)\n", imp->imp_obd->obd_name,
2349 kcd->kcd_nr_archives);
2351 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2352 "(archive nr = %u)\n",
2353 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2354 if (kcd->kcd_nr_archives != 0)
2355 archives = kcd->kcd_archives;
2358 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2359 /* ignore error if the copytool is already registered */
2360 return (rc == -EEXIST) ? 0 : rc;
2364 * Re-establish all kuc contexts with MDT
2365 * after MDT shutdown/recovery.
2367 static int mdc_kuc_reregister(struct obd_import *imp)
2369 /* re-register HSM agents */
2370 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2371 mdc_hsm_ct_reregister, imp);
2374 static int mdc_set_info_async(const struct lu_env *env,
2375 struct obd_export *exp,
2376 u32 keylen, void *key,
2377 u32 vallen, void *val,
2378 struct ptlrpc_request_set *set)
2380 struct obd_import *imp = class_exp2cliimp(exp);
2384 if (KEY_IS(KEY_READ_ONLY)) {
2385 if (vallen != sizeof(int))
2388 spin_lock(&imp->imp_lock);
2389 if (*((int *)val)) {
2390 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2391 imp->imp_connect_data.ocd_connect_flags |=
2394 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2395 imp->imp_connect_data.ocd_connect_flags &=
2396 ~OBD_CONNECT_RDONLY;
2398 spin_unlock(&imp->imp_lock);
2400 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2401 keylen, key, vallen, val, set);
2404 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2405 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2406 keylen, key, vallen, val, set);
2409 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2410 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2415 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2416 __u32 *default_easize = val;
2418 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2422 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2426 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2427 __u32 keylen, void *key, __u32 *vallen, void *val)
2431 if (KEY_IS(KEY_MAX_EASIZE)) {
2432 __u32 mdsize, *max_easize;
2434 if (*vallen != sizeof(int))
2436 mdsize = *(__u32 *)val;
2437 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2438 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2440 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2442 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2443 __u32 *default_easize;
2445 if (*vallen != sizeof(int))
2447 default_easize = val;
2448 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2450 } else if (KEY_IS(KEY_CONN_DATA)) {
2451 struct obd_import *imp = class_exp2cliimp(exp);
2452 struct obd_connect_data *data = val;
2454 if (*vallen != sizeof(*data))
2457 *data = imp->imp_connect_data;
2459 } else if (KEY_IS(KEY_TGT_COUNT)) {
2460 *((__u32 *)val) = 1;
2464 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2469 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2470 struct ptlrpc_request **request)
2472 struct ptlrpc_request *req;
2477 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2481 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2483 ptlrpc_request_free(req);
2487 mdc_pack_body(req, fid, 0, 0, -1, 0);
2489 ptlrpc_request_set_replen(req);
2491 rc = ptlrpc_queue_wait(req);
2493 ptlrpc_req_finished(req);
2499 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2500 enum obd_import_event event)
2502 struct client_obd *cli = &obd->u.cli;
2505 LASSERT(imp->imp_obd == obd);
2508 case IMP_EVENT_DISCON:
2509 spin_lock(&cli->cl_loi_list_lock);
2510 cli->cl_avail_grant = 0;
2511 cli->cl_lost_grant = 0;
2512 spin_unlock(&cli->cl_loi_list_lock);
2514 case IMP_EVENT_INACTIVE:
2516 * Flush current sequence to make client obtain new one
2517 * from server in case of disconnect/reconnect.
2519 down_read(&cli->cl_seq_rwsem);
2521 seq_client_flush(cli->cl_seq);
2522 up_read(&cli->cl_seq_rwsem);
2524 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2526 case IMP_EVENT_INVALIDATE: {
2527 struct ldlm_namespace *ns = obd->obd_namespace;
2531 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2533 env = cl_env_get(&refcheck);
2535 /* Reset grants. All pages go to failing rpcs due to
2536 * the invalid import.
2538 osc_io_unplug(env, cli, NULL);
2540 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2541 osc_ldlm_resource_invalidate,
2543 cl_env_put(env, &refcheck);
2544 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2550 case IMP_EVENT_ACTIVE:
2551 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2552 /* redo the kuc registration after reconnecting */
2554 rc = mdc_kuc_reregister(imp);
2556 case IMP_EVENT_OCD: {
2557 struct obd_connect_data *ocd = &imp->imp_connect_data;
2559 if (OCD_HAS_FLAG(ocd, GRANT))
2560 osc_init_grant(cli, ocd);
2562 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2565 case IMP_EVENT_DEACTIVATE:
2566 case IMP_EVENT_ACTIVATE:
2569 CERROR("Unknown import event %x\n", event);
2575 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2576 struct lu_fid *fid, struct md_op_data *op_data)
2578 struct client_obd *cli = &exp->exp_obd->u.cli;
2583 down_read(&cli->cl_seq_rwsem);
2585 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2586 up_read(&cli->cl_seq_rwsem);
2591 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2593 struct client_obd *cli = &exp->exp_obd->u.cli;
2594 return &cli->cl_target_uuid;
2598 * Determine whether the lock can be canceled before replaying it during
2599 * recovery, non zero value will be return if the lock can be canceled,
2600 * or zero returned for not
2602 static int mdc_cancel_weight(struct ldlm_lock *lock)
2604 if (lock->l_resource->lr_type != LDLM_IBITS)
2607 /* FIXME: if we ever get into a situation where there are too many
2608 * opened files with open locks on a single node, then we really
2609 * should replay these open locks to reget it */
2610 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2613 /* Special case for DoM locks, cancel only unused and granted locks */
2614 if (ldlm_has_dom(lock) &&
2615 (lock->l_granted_mode != lock->l_req_mode ||
2616 osc_ldlm_weigh_ast(lock) != 0))
2622 static int mdc_resource_inode_free(struct ldlm_resource *res)
2624 if (res->lr_lvb_inode)
2625 res->lr_lvb_inode = NULL;
2630 static struct ldlm_valblock_ops inode_lvbo = {
2631 .lvbo_free = mdc_resource_inode_free
2634 static int mdc_llog_init(struct obd_device *obd)
2636 struct obd_llog_group *olg = &obd->obd_olg;
2637 struct llog_ctxt *ctxt;
2642 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2647 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2648 llog_initiator_connect(ctxt);
2649 llog_ctxt_put(ctxt);
2654 static void mdc_llog_finish(struct obd_device *obd)
2656 struct llog_ctxt *ctxt;
2660 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2662 llog_cleanup(NULL, ctxt);
2667 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2673 rc = osc_setup_common(obd, cfg);
2677 rc = mdc_tunables_init(obd);
2679 GOTO(err_osc_cleanup, rc);
2681 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2683 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2685 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2687 rc = mdc_llog_init(obd);
2689 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2691 GOTO(err_llog_cleanup, rc);
2694 rc = mdc_changelog_cdev_init(obd);
2696 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2698 GOTO(err_changelog_cleanup, rc);
2703 err_changelog_cleanup:
2704 mdc_llog_finish(obd);
2706 lprocfs_free_md_stats(obd);
2707 ptlrpc_lprocfs_unregister_obd(obd);
2709 osc_cleanup_common(obd);
2713 /* Initialize the default and maximum LOV EA sizes. This allows
2714 * us to make MDS RPCs with large enough reply buffers to hold a default
2715 * sized EA without having to calculate this (via a call into the
2716 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2717 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2718 * a large number of stripes is possible. If a larger reply buffer is
2719 * required it will be reallocated in the ptlrpc layer due to overflow.
2721 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2724 struct obd_device *obd = exp->exp_obd;
2725 struct client_obd *cli = &obd->u.cli;
2728 if (cli->cl_max_mds_easize < easize)
2729 cli->cl_max_mds_easize = easize;
2731 if (cli->cl_default_mds_easize < def_easize)
2732 cli->cl_default_mds_easize = def_easize;
2737 static int mdc_precleanup(struct obd_device *obd)
2741 osc_precleanup_common(obd);
2742 mdc_changelog_cdev_finish(obd);
2744 obd_cleanup_client_import(obd);
2745 ptlrpc_lprocfs_unregister_obd(obd);
2746 lprocfs_free_md_stats(obd);
2747 mdc_llog_finish(obd);
2751 static int mdc_cleanup(struct obd_device *obd)
2753 return osc_cleanup_common(obd);
2756 int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2758 struct lustre_cfg *lcfg = buf;
2759 size_t count = class_modify_config(lcfg, PARAM_MDC,
2760 &obd->obd_kset.kobj);
2762 return count > 0 ? 0 : count;
2765 static struct obd_ops mdc_obd_ops = {
2766 .o_owner = THIS_MODULE,
2767 .o_setup = mdc_setup,
2768 .o_precleanup = mdc_precleanup,
2769 .o_cleanup = mdc_cleanup,
2770 .o_add_conn = client_import_add_conn,
2771 .o_del_conn = client_import_del_conn,
2772 .o_connect = client_connect_import,
2773 .o_reconnect = osc_reconnect,
2774 .o_disconnect = osc_disconnect,
2775 .o_iocontrol = mdc_iocontrol,
2776 .o_set_info_async = mdc_set_info_async,
2777 .o_statfs = mdc_statfs,
2778 .o_fid_init = client_fid_init,
2779 .o_fid_fini = client_fid_fini,
2780 .o_fid_alloc = mdc_fid_alloc,
2781 .o_import_event = mdc_import_event,
2782 .o_get_info = mdc_get_info,
2783 .o_process_config = mdc_process_config,
2784 .o_get_uuid = mdc_get_uuid,
2785 .o_quotactl = mdc_quotactl,
2788 static struct md_ops mdc_md_ops = {
2789 .m_get_root = mdc_get_root,
2790 .m_null_inode = mdc_null_inode,
2791 .m_close = mdc_close,
2792 .m_create = mdc_create,
2793 .m_enqueue = mdc_enqueue,
2794 .m_getattr = mdc_getattr,
2795 .m_getattr_name = mdc_getattr_name,
2796 .m_intent_lock = mdc_intent_lock,
2798 .m_rename = mdc_rename,
2799 .m_setattr = mdc_setattr,
2800 .m_setxattr = mdc_setxattr,
2801 .m_getxattr = mdc_getxattr,
2802 .m_fsync = mdc_fsync,
2803 .m_file_resync = mdc_file_resync,
2804 .m_read_page = mdc_read_page,
2805 .m_unlink = mdc_unlink,
2806 .m_cancel_unused = mdc_cancel_unused,
2807 .m_init_ea_size = mdc_init_ea_size,
2808 .m_set_lock_data = mdc_set_lock_data,
2809 .m_lock_match = mdc_lock_match,
2810 .m_get_lustre_md = mdc_get_lustre_md,
2811 .m_free_lustre_md = mdc_free_lustre_md,
2812 .m_set_open_replay_data = mdc_set_open_replay_data,
2813 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2814 .m_intent_getattr_async = mdc_intent_getattr_async,
2815 .m_revalidate_lock = mdc_revalidate_lock
2818 static int __init mdc_init(void)
2820 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2821 LUSTRE_MDC_NAME, &mdc_device_type);
2824 static void __exit mdc_exit(void)
2826 class_unregister_type(LUSTRE_MDC_NAME);
2829 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2830 MODULE_DESCRIPTION("Lustre Metadata Client");
2831 MODULE_VERSION(LUSTRE_VERSION_STRING);
2832 MODULE_LICENSE("GPL");
2834 module_init(mdc_init);
2835 module_exit(mdc_exit);