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 = MIN(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 = MIN(imp->imp_connect_data.ocd_max_easize,
300 mdc_reset_acl_req(req);
304 ptlrpc_req_finished(req);
312 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
313 const struct lu_fid *fid, int opcode, u64 valid,
314 const char *xattr_name, const char *input,
315 int input_size, int output_size, int flags,
316 __u32 suppgid, struct ptlrpc_request **request)
318 struct ptlrpc_request *req;
319 int xattr_namelen = 0;
325 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
330 xattr_namelen = strlen(xattr_name) + 1;
331 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
336 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
340 /* Flush local XATTR locks to get rid of a possible cancel RPC */
341 if (opcode == MDS_REINT && fid_is_sane(fid) &&
342 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
343 struct list_head cancels = LIST_HEAD_INIT(cancels);
346 /* Without that packing would fail */
348 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
351 count = mdc_resource_get_unused(exp, fid,
353 MDS_INODELOCK_XATTR);
355 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
357 ptlrpc_request_free(req);
361 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
363 ptlrpc_request_free(req);
368 if (opcode == MDS_REINT) {
369 struct mdt_rec_setxattr *rec;
371 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
372 sizeof(struct mdt_rec_reint));
373 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
374 rec->sx_opcode = REINT_SETXATTR;
375 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
376 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
377 rec->sx_cap = cfs_curproc_cap_pack();
378 rec->sx_suppgid1 = suppgid;
379 rec->sx_suppgid2 = -1;
381 rec->sx_valid = valid | OBD_MD_FLCTIME;
382 rec->sx_time = ktime_get_real_seconds();
383 rec->sx_size = output_size;
384 rec->sx_flags = flags;
386 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
390 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
391 memcpy(tmp, xattr_name, xattr_namelen);
394 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
395 memcpy(tmp, input, input_size);
398 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
399 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
400 RCL_SERVER, output_size);
401 ptlrpc_request_set_replen(req);
404 if (opcode == MDS_REINT)
405 mdc_get_mod_rpc_slot(req, NULL);
407 rc = ptlrpc_queue_wait(req);
409 if (opcode == MDS_REINT)
410 mdc_put_mod_rpc_slot(req, NULL);
413 ptlrpc_req_finished(req);
419 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
420 u64 obd_md_valid, const char *name,
421 const void *value, size_t value_size,
422 unsigned int xattr_flags, u32 suppgid,
423 struct ptlrpc_request **req)
425 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
426 obd_md_valid == OBD_MD_FLXATTRRM);
428 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
429 fid, MDS_REINT, obd_md_valid, name,
430 value, value_size, 0, xattr_flags, suppgid,
434 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
435 u64 obd_md_valid, const char *name, size_t buf_size,
436 struct ptlrpc_request **req)
438 struct mdt_body *body;
441 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
442 obd_md_valid == OBD_MD_FLXATTRLS);
444 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
445 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
450 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
452 GOTO(out, rc = -EPROTO);
454 /* only detect the xattr size */
456 /* LU-11109: Older MDTs do not distinguish
457 * between nonexistent xattrs and zero length
458 * values in this case. Newer MDTs will return
459 * -ENODATA or set OBD_MD_FLXATTR. */
460 GOTO(out, rc = body->mbo_eadatasize);
463 if (body->mbo_eadatasize == 0) {
464 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
465 * success so that we can distinguish between
466 * zero length value and nonexistent xattr.
468 * If OBD_MD_FLXATTR is not set then we keep
469 * the old behavior and return -ENODATA for
470 * getxattr() when mbo_eadatasize is 0. But
471 * -ENODATA only makes sense for getxattr()
472 * and not for listxattr(). */
473 if (body->mbo_valid & OBD_MD_FLXATTR)
475 else if (obd_md_valid == OBD_MD_FLXATTR)
476 GOTO(out, rc = -ENODATA);
481 GOTO(out, rc = body->mbo_eadatasize);
484 ptlrpc_req_finished(*req);
491 #ifdef CONFIG_FS_POSIX_ACL
492 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
494 struct req_capsule *pill = &req->rq_pill;
495 struct mdt_body *body = md->body;
496 struct posix_acl *acl;
501 if (!body->mbo_aclsize)
504 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
509 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
514 CERROR("convert xattr to acl: %d\n", rc);
518 rc = posix_acl_valid(&init_user_ns, acl);
520 CERROR("validate acl: %d\n", rc);
521 posix_acl_release(acl);
529 #define mdc_unpack_acl(req, md) 0
532 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
533 struct obd_export *dt_exp, struct obd_export *md_exp,
534 struct lustre_md *md)
536 struct req_capsule *pill = &req->rq_pill;
541 memset(md, 0, sizeof(*md));
543 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
544 LASSERT(md->body != NULL);
546 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
547 if (!S_ISREG(md->body->mbo_mode)) {
548 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
549 "regular file, but is not\n");
550 GOTO(out, rc = -EPROTO);
553 if (md->body->mbo_eadatasize == 0) {
554 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
555 "but eadatasize 0\n");
556 GOTO(out, rc = -EPROTO);
559 md->layout.lb_len = md->body->mbo_eadatasize;
560 md->layout.lb_buf = req_capsule_server_sized_get(pill,
563 if (md->layout.lb_buf == NULL)
564 GOTO(out, rc = -EPROTO);
565 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
566 const union lmv_mds_md *lmv;
569 if (!S_ISDIR(md->body->mbo_mode)) {
570 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
571 "directory, but is not\n");
572 GOTO(out, rc = -EPROTO);
575 lmv_size = md->body->mbo_eadatasize;
577 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
578 "but eadatasize 0\n");
582 if (md->body->mbo_valid & OBD_MD_MEA) {
583 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
586 GOTO(out, rc = -EPROTO);
588 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
592 if (rc < (typeof(rc))sizeof(*md->lmv)) {
593 CDEBUG(D_INFO, "size too small: "
594 "rc < sizeof(*md->lmv) (%d < %d)\n",
595 rc, (int)sizeof(*md->lmv));
596 GOTO(out, rc = -EPROTO);
602 if (md->body->mbo_valid & OBD_MD_FLACL) {
603 /* for ACL, it's possible that FLACL is set but aclsize is zero.
604 * only when aclsize != 0 there's an actual segment for ACL
607 if (md->body->mbo_aclsize) {
608 rc = mdc_unpack_acl(req, md);
611 #ifdef CONFIG_FS_POSIX_ACL
613 md->posix_acl = NULL;
621 #ifdef CONFIG_FS_POSIX_ACL
622 posix_acl_release(md->posix_acl);
628 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
634 void mdc_replay_open(struct ptlrpc_request *req)
636 struct md_open_data *mod = req->rq_cb_data;
637 struct ptlrpc_request *close_req;
638 struct obd_client_handle *och;
639 struct lustre_handle old_open_handle = { };
640 struct mdt_body *body;
644 DEBUG_REQ(D_ERROR, req,
645 "Can't properly replay without open data.");
650 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
651 LASSERT(body != NULL);
653 spin_lock(&req->rq_lock);
655 if (och && och->och_open_handle.cookie)
656 req->rq_early_free_repbuf = 1;
658 req->rq_early_free_repbuf = 0;
659 spin_unlock(&req->rq_lock);
661 if (req->rq_early_free_repbuf) {
662 struct lustre_handle *file_open_handle;
664 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
666 file_open_handle = &och->och_open_handle;
667 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
668 file_open_handle->cookie, body->mbo_open_handle.cookie);
669 old_open_handle = *file_open_handle;
670 *file_open_handle = body->mbo_open_handle;
673 close_req = mod->mod_close_req;
675 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
676 struct mdt_ioepoch *epoch;
678 LASSERT(opc == MDS_CLOSE);
679 epoch = req_capsule_client_get(&close_req->rq_pill,
683 if (req->rq_early_free_repbuf)
684 LASSERT(old_open_handle.cookie ==
685 epoch->mio_open_handle.cookie);
687 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
688 epoch->mio_open_handle = body->mbo_open_handle;
693 void mdc_commit_open(struct ptlrpc_request *req)
695 struct md_open_data *mod = req->rq_cb_data;
700 * No need to touch md_open_data::mod_och, it holds a reference on
701 * \var mod and will zero references to each other, \var mod will be
702 * freed after that when md_open_data::mod_och will put the reference.
706 * Do not let open request to disappear as it still may be needed
707 * for close rpc to happen (it may happen on evict only, otherwise
708 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
709 * called), just mark this rpc as committed to distinguish these 2
710 * cases, see mdc_close() for details. The open request reference will
711 * be put along with freeing \var mod.
713 ptlrpc_request_addref(req);
714 spin_lock(&req->rq_lock);
715 req->rq_committed = 1;
716 spin_unlock(&req->rq_lock);
717 req->rq_cb_data = NULL;
721 int mdc_set_open_replay_data(struct obd_export *exp,
722 struct obd_client_handle *och,
723 struct lookup_intent *it)
725 struct md_open_data *mod;
726 struct mdt_rec_create *rec;
727 struct mdt_body *body;
728 struct ptlrpc_request *open_req = it->it_request;
729 struct obd_import *imp = open_req->rq_import;
732 if (!open_req->rq_replay)
735 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
736 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
737 LASSERT(rec != NULL);
738 /* Incoming message in my byte order (it's been swabbed). */
739 /* Outgoing messages always in my byte order. */
740 LASSERT(body != NULL);
742 /* Only if the import is replayable, we set replay_open data */
743 if (och && imp->imp_replayable) {
744 mod = obd_mod_alloc();
746 DEBUG_REQ(D_ERROR, open_req,
747 "Can't allocate md_open_data");
752 * Take a reference on \var mod, to be freed on mdc_close().
753 * It protects \var mod from being freed on eviction (commit
754 * callback is called despite rq_replay flag).
755 * Another reference for \var och.
760 spin_lock(&open_req->rq_lock);
763 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
764 it_disposition(it, DISP_OPEN_STRIPE);
765 mod->mod_open_req = open_req;
766 open_req->rq_cb_data = mod;
767 open_req->rq_commit_cb = mdc_commit_open;
768 open_req->rq_early_free_repbuf = 1;
769 spin_unlock(&open_req->rq_lock);
772 rec->cr_fid2 = body->mbo_fid1;
773 rec->cr_open_handle_old = body->mbo_open_handle;
774 open_req->rq_replay_cb = mdc_replay_open;
775 if (!fid_is_sane(&body->mbo_fid1)) {
776 DEBUG_REQ(D_ERROR, open_req,
777 "saving replay request with insane FID " DFID,
778 PFID(&body->mbo_fid1));
782 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
786 static void mdc_free_open(struct md_open_data *mod)
790 if (mod->mod_is_create == 0 &&
791 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
795 * No reason to asssert here if the open request has
796 * rq_replay == 1. It means that mdc_close failed, and
797 * close request wasn`t sent. It is not fatal to client.
798 * The worst thing is eviction if the client gets open lock
801 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
802 "= %d\n", mod->mod_open_req->rq_replay);
804 ptlrpc_request_committed(mod->mod_open_req, committed);
805 if (mod->mod_close_req)
806 ptlrpc_request_committed(mod->mod_close_req, committed);
809 int mdc_clear_open_replay_data(struct obd_export *exp,
810 struct obd_client_handle *och)
812 struct md_open_data *mod = och->och_mod;
816 * It is possible to not have \var mod in a case of eviction between
817 * lookup and ll_file_open().
822 LASSERT(mod != LP_POISON);
823 LASSERT(mod->mod_open_req != NULL);
825 spin_lock(&mod->mod_open_req->rq_lock);
827 mod->mod_och->och_open_handle.cookie = 0;
828 mod->mod_open_req->rq_early_free_repbuf = 0;
829 spin_unlock(&mod->mod_open_req->rq_lock);
839 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
840 struct md_open_data *mod, struct ptlrpc_request **request)
842 struct obd_device *obd = class_exp2obd(exp);
843 struct ptlrpc_request *req;
844 struct req_format *req_fmt;
845 size_t u32_count = 0;
850 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
851 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
854 if (op_data->op_bias & MDS_CLOSE_INTENT) {
855 req_fmt = &RQF_MDS_CLOSE_INTENT;
856 if (op_data->op_bias & MDS_HSM_RELEASE) {
857 /* allocate a FID for volatile file */
858 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
861 CERROR("%s: "DFID" allocating FID: rc = %d\n",
862 obd->obd_name, PFID(&op_data->op_fid1),
864 /* save the errcode and proceed to close */
868 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
869 size_t count = op_data->op_data_size / sizeof(__u32);
871 if (count > INLINE_RESYNC_ARRAY_SIZE)
875 req_fmt = &RQF_MDS_CLOSE;
879 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
882 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
884 /* Ensure that this close's handle is fixed up during replay. */
885 if (likely(mod != NULL)) {
886 LASSERTF(mod->mod_open_req != NULL &&
887 mod->mod_open_req->rq_type != LI_POISON,
888 "POISONED open %p!\n", mod->mod_open_req);
890 mod->mod_close_req = req;
892 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
893 /* We no longer want to preserve this open for replay even
894 * though the open was committed. b=3632, b=3633 */
895 spin_lock(&mod->mod_open_req->rq_lock);
896 mod->mod_open_req->rq_replay = 0;
897 spin_unlock(&mod->mod_open_req->rq_lock);
899 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
903 * TODO: repeat close after errors
905 CWARN("%s: close of FID "DFID" failed, file reference will be "
906 "dropped when this client unmounts or is evicted\n",
907 obd->obd_name, PFID(&op_data->op_fid1));
908 GOTO(out, rc = -ENOMEM);
912 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
913 u32_count * sizeof(__u32));
915 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
917 ptlrpc_request_free(req);
922 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
923 * portal whose threads are not taking any DLM locks and are therefore
924 * always progressing */
925 req->rq_request_portal = MDS_READPAGE_PORTAL;
926 ptlrpc_at_set_req_timeout(req);
929 mdc_close_pack(req, op_data);
931 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
932 obd->u.cli.cl_default_mds_easize);
934 ptlrpc_request_set_replen(req);
936 mdc_get_mod_rpc_slot(req, NULL);
937 rc = ptlrpc_queue_wait(req);
938 mdc_put_mod_rpc_slot(req, NULL);
940 if (req->rq_repmsg == NULL) {
941 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
944 rc = req->rq_status ?: -EIO;
945 } else if (rc == 0 || rc == -EAGAIN) {
946 struct mdt_body *body;
948 rc = lustre_msg_get_status(req->rq_repmsg);
949 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
950 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
955 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
958 } else if (rc == -ESTALE) {
960 * it can be allowed error after 3633 if open was committed and
961 * server failed before close was sent. Let's check if mod
962 * exists and return no error in that case
965 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
966 LASSERT(mod->mod_open_req != NULL);
967 if (mod->mod_open_req->rq_committed)
975 mod->mod_close_req = NULL;
976 /* Since now, mod is accessed through open_req only,
977 * thus close req does not keep a reference on mod anymore. */
982 RETURN(rc < 0 ? rc : saved_rc);
985 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
986 u64 offset, struct page **pages, int npages,
987 struct ptlrpc_request **request)
989 struct ptlrpc_request *req;
990 struct ptlrpc_bulk_desc *desc;
992 wait_queue_head_t waitq;
994 struct l_wait_info lwi;
999 init_waitqueue_head(&waitq);
1002 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1006 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1008 ptlrpc_request_free(req);
1012 req->rq_request_portal = MDS_READPAGE_PORTAL;
1013 ptlrpc_at_set_req_timeout(req);
1015 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1016 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1018 &ptlrpc_bulk_kiov_pin_ops);
1020 ptlrpc_req_finished(req);
1024 /* NB req now owns desc and will free it when it gets freed */
1025 for (i = 0; i < npages; i++)
1026 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1029 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1031 ptlrpc_request_set_replen(req);
1032 rc = ptlrpc_queue_wait(req);
1034 ptlrpc_req_finished(req);
1035 if (rc != -ETIMEDOUT)
1039 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1040 CERROR("%s: too many resend retries: rc = %d\n",
1041 exp->exp_obd->obd_name, -EIO);
1044 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1046 l_wait_event(waitq, 0, &lwi);
1051 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1052 req->rq_bulk->bd_nob_transferred);
1054 ptlrpc_req_finished(req);
1058 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1059 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1060 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1061 PAGE_SIZE * npages);
1062 ptlrpc_req_finished(req);
1070 static void mdc_release_page(struct page *page, int remove)
1074 if (likely(page->mapping != NULL))
1075 truncate_complete_page(page->mapping, page);
1081 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1082 __u64 *start, __u64 *end, int hash64)
1085 * Complement of hash is used as an index so that
1086 * radix_tree_gang_lookup() can be used to find a page with starting
1087 * hash _smaller_ than one we are looking for.
1089 unsigned long offset = hash_x_index(*hash, hash64);
1093 spin_lock_irq(&mapping->tree_lock);
1094 found = radix_tree_gang_lookup(&mapping->page_tree,
1095 (void **)&page, offset, 1);
1096 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1097 struct lu_dirpage *dp;
1100 spin_unlock_irq(&mapping->tree_lock);
1102 * In contrast to find_lock_page() we are sure that directory
1103 * page cannot be truncated (while DLM lock is held) and,
1104 * hence, can avoid restart.
1106 * In fact, page cannot be locked here at all, because
1107 * mdc_read_page_remote does synchronous io.
1109 wait_on_page_locked(page);
1110 if (PageUptodate(page)) {
1112 if (BITS_PER_LONG == 32 && hash64) {
1113 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1114 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1115 *hash = *hash >> 32;
1117 *start = le64_to_cpu(dp->ldp_hash_start);
1118 *end = le64_to_cpu(dp->ldp_hash_end);
1120 if (unlikely(*start == 1 && *hash == 0))
1123 LASSERTF(*start <= *hash, "start = %#llx"
1124 ",end = %#llx,hash = %#llx\n",
1125 *start, *end, *hash);
1126 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1127 " hash %#llx\n", offset, *start, *end, *hash);
1130 mdc_release_page(page, 0);
1132 } else if (*end != *start && *hash == *end) {
1134 * upon hash collision, remove this page,
1135 * otherwise put page reference, and
1136 * mdc_read_page_remote() will issue RPC to
1137 * fetch the page we want.
1140 mdc_release_page(page,
1141 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1146 page = ERR_PTR(-EIO);
1149 spin_unlock_irq(&mapping->tree_lock);
1156 * Adjust a set of pages, each page containing an array of lu_dirpages,
1157 * so that each page can be used as a single logical lu_dirpage.
1159 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1160 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1161 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1162 * value is used as a cookie to request the next lu_dirpage in a
1163 * directory listing that spans multiple pages (two in this example):
1166 * .|--------v------- -----.
1167 * |s|e|f|p|ent|ent| ... |ent|
1168 * '--|-------------- -----' Each PAGE contains a single
1169 * '------. lu_dirpage.
1170 * .---------v------- -----.
1171 * |s|e|f|p|ent| 0 | ... | 0 |
1172 * '----------------- -----'
1174 * However, on hosts where the native VM page size (PAGE_SIZE) is
1175 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1176 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1177 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1178 * after it in the same PAGE (arrows simplified for brevity, but
1179 * in general e0==s1, e1==s2, etc.):
1181 * .-------------------- -----.
1182 * |s0|e0|f0|p|ent|ent| ... |ent|
1183 * |---v---------------- -----|
1184 * |s1|e1|f1|p|ent|ent| ... |ent|
1185 * |---v---------------- -----| Here, each PAGE contains
1186 * ... multiple lu_dirpages.
1187 * |---v---------------- -----|
1188 * |s'|e'|f'|p|ent|ent| ... |ent|
1189 * '---|---------------- -----'
1191 * .----------------------------.
1194 * This structure is transformed into a single logical lu_dirpage as follows:
1196 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1197 * labeled 'next PAGE'.
1199 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1200 * a hash collision with the next page exists.
1202 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1203 * to the first entry of the next lu_dirpage.
1205 #if PAGE_SIZE > LU_PAGE_SIZE
1206 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1210 for (i = 0; i < cfs_pgs; i++) {
1211 struct lu_dirpage *dp = kmap(pages[i]);
1212 struct lu_dirpage *first = dp;
1213 struct lu_dirent *end_dirent = NULL;
1214 struct lu_dirent *ent;
1215 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1216 __u32 flags = le32_to_cpu(dp->ldp_flags);
1218 while (--lu_pgs > 0) {
1219 ent = lu_dirent_start(dp);
1220 for (end_dirent = ent; ent != NULL;
1221 end_dirent = ent, ent = lu_dirent_next(ent));
1223 /* Advance dp to next lu_dirpage. */
1224 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1226 /* Check if we've reached the end of the PAGE. */
1227 if (!((unsigned long)dp & ~PAGE_MASK))
1230 /* Save the hash and flags of this lu_dirpage. */
1231 hash_end = le64_to_cpu(dp->ldp_hash_end);
1232 flags = le32_to_cpu(dp->ldp_flags);
1234 /* Check if lu_dirpage contains no entries. */
1235 if (end_dirent == NULL)
1238 /* Enlarge the end entry lde_reclen from 0 to
1239 * first entry of next lu_dirpage. */
1240 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1241 end_dirent->lde_reclen =
1242 cpu_to_le16((char *)(dp->ldp_entries) -
1243 (char *)end_dirent);
1246 first->ldp_hash_end = hash_end;
1247 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1248 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1252 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1255 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1256 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1258 /* parameters for readdir page */
1259 struct readpage_param {
1260 struct md_op_data *rp_mod;
1263 struct obd_export *rp_exp;
1264 struct md_callback *rp_cb;
1267 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1268 static inline void delete_from_page_cache(struct page *page)
1270 remove_from_page_cache(page);
1276 * Read pages from server.
1278 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1279 * a header lu_dirpage which describes the start/end hash, and whether this
1280 * page is empty (contains no dir entry) or hash collide with next page.
1281 * After client receives reply, several pages will be integrated into dir page
1282 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1283 * lu_dirpage for this integrated page will be adjusted.
1285 static int mdc_read_page_remote(void *data, struct page *page0)
1287 struct readpage_param *rp = data;
1288 struct page **page_pool;
1290 struct lu_dirpage *dp;
1291 struct md_op_data *op_data = rp->rp_mod;
1292 struct ptlrpc_request *req;
1294 struct inode *inode;
1296 int rd_pgs = 0; /* number of pages actually read */
1302 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1303 inode = op_data->op_data;
1304 fid = &op_data->op_fid1;
1305 LASSERT(inode != NULL);
1307 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1308 if (page_pool != NULL) {
1309 page_pool[0] = page0;
1315 for (npages = 1; npages < max_pages; npages++) {
1316 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1320 page_pool[npages] = page;
1323 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1325 /* page0 is special, which was added into page cache early */
1326 delete_from_page_cache(page0);
1330 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1332 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1333 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1335 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1337 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1339 SetPageUptodate(page0);
1343 ptlrpc_req_finished(req);
1344 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1345 for (i = 1; i < npages; i++) {
1346 unsigned long offset;
1350 page = page_pool[i];
1352 if (rc < 0 || i >= rd_pgs) {
1357 SetPageUptodate(page);
1360 hash = le64_to_cpu(dp->ldp_hash_start);
1363 offset = hash_x_index(hash, rp->rp_hash64);
1365 prefetchw(&page->flags);
1366 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1371 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1372 " rc = %d\n", offset, ret);
1376 if (page_pool != &page0)
1377 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1383 * Read dir page from cache first, if it can not find it, read it from
1384 * server and add into the cache.
1386 * \param[in] exp MDC export
1387 * \param[in] op_data client MD stack parameters, transfering parameters
1388 * between different layers on client MD stack.
1389 * \param[in] cb_op callback required for ldlm lock enqueue during
1391 * \param[in] hash_offset the hash offset of the page to be read
1392 * \param[in] ppage the page to be read
1394 * retval = 0 get the page successfully
1395 * errno(<0) get the page failed
1397 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1398 struct md_callback *cb_op, __u64 hash_offset,
1399 struct page **ppage)
1401 struct lookup_intent it = { .it_op = IT_READDIR };
1403 struct inode *dir = op_data->op_data;
1404 struct address_space *mapping;
1405 struct lu_dirpage *dp;
1408 struct lustre_handle lockh;
1409 struct ptlrpc_request *enq_req = NULL;
1410 struct readpage_param rp_param;
1417 LASSERT(dir != NULL);
1418 mapping = dir->i_mapping;
1420 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1421 cb_op->md_blocking_ast, 0);
1422 if (enq_req != NULL)
1423 ptlrpc_req_finished(enq_req);
1426 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1427 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1432 lockh.cookie = it.it_lock_handle;
1433 mdc_set_lock_data(exp, &lockh, dir, NULL);
1435 rp_param.rp_off = hash_offset;
1436 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1437 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1438 rp_param.rp_hash64);
1440 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1441 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1442 rp_param.rp_off, PTR_ERR(page));
1443 GOTO(out_unlock, rc = PTR_ERR(page));
1444 } else if (page != NULL) {
1446 * XXX nikita: not entirely correct handling of a corner case:
1447 * suppose hash chain of entries with hash value HASH crosses
1448 * border between pages P0 and P1. First both P0 and P1 are
1449 * cached, seekdir() is called for some entry from the P0 part
1450 * of the chain. Later P0 goes out of cache. telldir(HASH)
1451 * happens and finds P1, as it starts with matching hash
1452 * value. Remaining entries from P0 part of the chain are
1453 * skipped. (Is that really a bug?)
1455 * Possible solutions: 0. don't cache P1 is such case, handle
1456 * it as an "overflow" page. 1. invalidate all pages at
1457 * once. 2. use HASH|1 as an index for P1.
1459 GOTO(hash_collision, page);
1462 rp_param.rp_exp = exp;
1463 rp_param.rp_mod = op_data;
1464 page = read_cache_page(mapping,
1465 hash_x_index(rp_param.rp_off,
1466 rp_param.rp_hash64),
1467 mdc_read_page_remote, &rp_param);
1469 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1470 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1471 rp_param.rp_off, PTR_ERR(page));
1472 GOTO(out_unlock, rc = PTR_ERR(page));
1475 wait_on_page_locked(page);
1477 if (!PageUptodate(page)) {
1478 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1479 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1480 rp_param.rp_off, -5);
1483 if (!PageChecked(page))
1484 SetPageChecked(page);
1485 if (PageError(page)) {
1486 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1487 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1488 rp_param.rp_off, -5);
1493 dp = page_address(page);
1494 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1495 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1496 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1497 rp_param.rp_off = hash_offset >> 32;
1499 start = le64_to_cpu(dp->ldp_hash_start);
1500 end = le64_to_cpu(dp->ldp_hash_end);
1501 rp_param.rp_off = hash_offset;
1504 LASSERT(start == rp_param.rp_off);
1505 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1506 #if BITS_PER_LONG == 32
1507 CWARN("Real page-wide hash collision at [%llu %llu] with "
1508 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1509 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1513 * Fetch whole overflow chain...
1521 ldlm_lock_decref(&lockh, it.it_lock_mode);
1525 mdc_release_page(page, 1);
1530 static int mdc_statfs(const struct lu_env *env,
1531 struct obd_export *exp, struct obd_statfs *osfs,
1532 time64_t max_age, __u32 flags)
1534 struct obd_device *obd = class_exp2obd(exp);
1535 struct req_format *fmt;
1536 struct ptlrpc_request *req;
1537 struct obd_statfs *msfs;
1538 struct obd_import *imp = NULL;
1543 * Since the request might also come from lprocfs, so we need
1544 * sync this with client_disconnect_export Bug15684
1546 down_read(&obd->u.cli.cl_sem);
1547 if (obd->u.cli.cl_import)
1548 imp = class_import_get(obd->u.cli.cl_import);
1549 up_read(&obd->u.cli.cl_sem);
1553 fmt = &RQF_MDS_STATFS;
1554 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1555 (flags & OBD_STATFS_SUM))
1556 fmt = &RQF_MDS_STATFS_NEW;
1557 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1560 GOTO(output, rc = -ENOMEM);
1562 if ((flags & OBD_STATFS_SUM) &&
1563 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1564 /* request aggregated states */
1565 struct mdt_body *body;
1567 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1569 GOTO(out, rc = -EPROTO);
1570 body->mbo_valid = OBD_MD_FLAGSTATFS;
1573 ptlrpc_request_set_replen(req);
1575 if (flags & OBD_STATFS_NODELAY) {
1576 /* procfs requests not want stay in wait for avoid deadlock */
1577 req->rq_no_resend = 1;
1578 req->rq_no_delay = 1;
1581 rc = ptlrpc_queue_wait(req);
1583 /* check connection error first */
1584 if (imp->imp_connect_error)
1585 rc = imp->imp_connect_error;
1589 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1591 GOTO(out, rc = -EPROTO);
1596 ptlrpc_req_finished(req);
1598 class_import_put(imp);
1602 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1604 __u32 keylen, vallen;
1608 if (gf->gf_pathlen > PATH_MAX)
1609 RETURN(-ENAMETOOLONG);
1610 if (gf->gf_pathlen < 2)
1613 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1614 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1615 sizeof(struct lu_fid));
1616 OBD_ALLOC(key, keylen);
1619 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1620 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1621 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1622 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1623 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1624 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1626 if (!fid_is_sane(&gf->gf_fid))
1627 GOTO(out, rc = -EINVAL);
1629 /* Val is struct getinfo_fid2path result plus path */
1630 vallen = sizeof(*gf) + gf->gf_pathlen;
1632 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1633 if (rc != 0 && rc != -EREMOTE)
1636 if (vallen <= sizeof(*gf))
1637 GOTO(out, rc = -EPROTO);
1638 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1639 GOTO(out, rc = -EOVERFLOW);
1641 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1642 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1643 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1644 /* only log the last 512 characters of the path */
1645 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1648 OBD_FREE(key, keylen);
1652 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1653 struct hsm_progress_kernel *hpk)
1655 struct obd_import *imp = class_exp2cliimp(exp);
1656 struct hsm_progress_kernel *req_hpk;
1657 struct ptlrpc_request *req;
1661 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1662 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1664 GOTO(out, rc = -ENOMEM);
1666 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1668 /* Copy hsm_progress struct */
1669 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1670 if (req_hpk == NULL)
1671 GOTO(out, rc = -EPROTO);
1674 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1676 ptlrpc_request_set_replen(req);
1678 mdc_get_mod_rpc_slot(req, NULL);
1679 rc = ptlrpc_queue_wait(req);
1680 mdc_put_mod_rpc_slot(req, NULL);
1684 ptlrpc_req_finished(req);
1688 * Send hsm_ct_register to MDS
1690 * \param[in] imp import
1691 * \param[in] archive_count if in bitmap format, it is the bitmap,
1692 * else it is the count of archive_ids
1693 * \param[in] archives if in bitmap format, it is NULL,
1694 * else it is archive_id lists
1696 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1699 struct ptlrpc_request *req;
1700 __u32 *archive_array;
1701 size_t archives_size;
1705 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1709 if (archives != NULL)
1710 archives_size = sizeof(*archive_array) * archive_count;
1712 archives_size = sizeof(archive_count);
1714 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1715 RCL_CLIENT, archives_size);
1717 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1719 ptlrpc_request_free(req);
1723 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1725 archive_array = req_capsule_client_get(&req->rq_pill,
1726 &RMF_MDS_HSM_ARCHIVE);
1727 if (archive_array == NULL)
1728 GOTO(out, rc = -EPROTO);
1730 if (archives != NULL)
1731 memcpy(archive_array, archives, archives_size);
1733 *archive_array = archive_count;
1735 ptlrpc_request_set_replen(req);
1737 rc = mdc_queue_wait(req);
1740 ptlrpc_req_finished(req);
1744 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1745 struct md_op_data *op_data)
1747 struct hsm_current_action *hca = op_data->op_data;
1748 struct hsm_current_action *req_hca;
1749 struct ptlrpc_request *req;
1753 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1754 &RQF_MDS_HSM_ACTION);
1758 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1760 ptlrpc_request_free(req);
1764 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1765 op_data->op_suppgids[0], 0);
1767 ptlrpc_request_set_replen(req);
1769 rc = mdc_queue_wait(req);
1773 req_hca = req_capsule_server_get(&req->rq_pill,
1774 &RMF_MDS_HSM_CURRENT_ACTION);
1775 if (req_hca == NULL)
1776 GOTO(out, rc = -EPROTO);
1782 ptlrpc_req_finished(req);
1786 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1788 struct ptlrpc_request *req;
1792 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1794 MDS_HSM_CT_UNREGISTER);
1796 GOTO(out, rc = -ENOMEM);
1798 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1800 ptlrpc_request_set_replen(req);
1802 rc = mdc_queue_wait(req);
1805 ptlrpc_req_finished(req);
1809 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1810 struct md_op_data *op_data)
1812 struct hsm_user_state *hus = op_data->op_data;
1813 struct hsm_user_state *req_hus;
1814 struct ptlrpc_request *req;
1818 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1819 &RQF_MDS_HSM_STATE_GET);
1823 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1825 ptlrpc_request_free(req);
1829 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1830 op_data->op_suppgids[0], 0);
1832 ptlrpc_request_set_replen(req);
1834 rc = mdc_queue_wait(req);
1838 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1839 if (req_hus == NULL)
1840 GOTO(out, rc = -EPROTO);
1846 ptlrpc_req_finished(req);
1850 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1851 struct md_op_data *op_data)
1853 struct hsm_state_set *hss = op_data->op_data;
1854 struct hsm_state_set *req_hss;
1855 struct ptlrpc_request *req;
1859 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1860 &RQF_MDS_HSM_STATE_SET);
1864 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1866 ptlrpc_request_free(req);
1870 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1871 op_data->op_suppgids[0], 0);
1874 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1875 if (req_hss == NULL)
1876 GOTO(out, rc = -EPROTO);
1879 ptlrpc_request_set_replen(req);
1881 mdc_get_mod_rpc_slot(req, NULL);
1882 rc = ptlrpc_queue_wait(req);
1883 mdc_put_mod_rpc_slot(req, NULL);
1887 ptlrpc_req_finished(req);
1891 static int mdc_ioc_hsm_request(struct obd_export *exp,
1892 struct hsm_user_request *hur)
1894 struct obd_import *imp = class_exp2cliimp(exp);
1895 struct ptlrpc_request *req;
1896 struct hsm_request *req_hr;
1897 struct hsm_user_item *req_hui;
1902 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1904 GOTO(out, rc = -ENOMEM);
1906 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1907 hur->hur_request.hr_itemcount
1908 * sizeof(struct hsm_user_item));
1909 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1910 hur->hur_request.hr_data_len);
1912 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1914 ptlrpc_request_free(req);
1918 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1920 /* Copy hsm_request struct */
1921 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1923 GOTO(out, rc = -EPROTO);
1924 *req_hr = hur->hur_request;
1926 /* Copy hsm_user_item structs */
1927 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1928 if (req_hui == NULL)
1929 GOTO(out, rc = -EPROTO);
1930 memcpy(req_hui, hur->hur_user_item,
1931 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1933 /* Copy opaque field */
1934 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1935 if (req_opaque == NULL)
1936 GOTO(out, rc = -EPROTO);
1937 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1939 ptlrpc_request_set_replen(req);
1941 mdc_get_mod_rpc_slot(req, NULL);
1942 rc = ptlrpc_queue_wait(req);
1943 mdc_put_mod_rpc_slot(req, NULL);
1948 ptlrpc_req_finished(req);
1952 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1953 struct lustre_kernelcomm *lk);
1955 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1956 struct obd_quotactl *oqctl)
1958 struct ptlrpc_request *req;
1959 struct obd_quotactl *oqc;
1963 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1964 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1969 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1972 ptlrpc_request_set_replen(req);
1973 ptlrpc_at_set_req_timeout(req);
1975 rc = ptlrpc_queue_wait(req);
1977 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1979 if (req->rq_repmsg &&
1980 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
1983 CERROR ("Can't unpack obd_quotactl\n");
1986 ptlrpc_req_finished(req);
1991 static int mdc_ioc_swap_layouts(struct obd_export *exp,
1992 struct md_op_data *op_data)
1994 struct list_head cancels = LIST_HEAD_INIT(cancels);
1995 struct ptlrpc_request *req;
1997 struct mdc_swap_layouts *msl, *payload;
2000 msl = op_data->op_data;
2002 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2003 * first thing it will do is to cancel the 2 layout
2004 * locks held by this client.
2005 * So the client must cancel its layout locks on the 2 fids
2006 * with the request RPC to avoid extra RPC round trips.
2008 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2009 LCK_EX, MDS_INODELOCK_LAYOUT |
2010 MDS_INODELOCK_XATTR);
2011 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2012 LCK_EX, MDS_INODELOCK_LAYOUT |
2013 MDS_INODELOCK_XATTR);
2015 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2016 &RQF_MDS_SWAP_LAYOUTS);
2018 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2022 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2024 ptlrpc_request_free(req);
2028 mdc_swap_layouts_pack(req, op_data);
2030 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2035 ptlrpc_request_set_replen(req);
2037 rc = ptlrpc_queue_wait(req);
2043 ptlrpc_req_finished(req);
2047 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2048 void *karg, void __user *uarg)
2050 struct obd_device *obd = exp->exp_obd;
2051 struct obd_ioctl_data *data = karg;
2052 struct obd_import *imp = obd->u.cli.cl_import;
2056 if (!try_module_get(THIS_MODULE)) {
2057 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2058 module_name(THIS_MODULE));
2062 case OBD_IOC_FID2PATH:
2063 rc = mdc_ioc_fid2path(exp, karg);
2065 case LL_IOC_HSM_CT_START:
2066 rc = mdc_ioc_hsm_ct_start(exp, karg);
2067 /* ignore if it was already registered on this MDS. */
2071 case LL_IOC_HSM_PROGRESS:
2072 rc = mdc_ioc_hsm_progress(exp, karg);
2074 case LL_IOC_HSM_STATE_GET:
2075 rc = mdc_ioc_hsm_state_get(exp, karg);
2077 case LL_IOC_HSM_STATE_SET:
2078 rc = mdc_ioc_hsm_state_set(exp, karg);
2080 case LL_IOC_HSM_ACTION:
2081 rc = mdc_ioc_hsm_current_action(exp, karg);
2083 case LL_IOC_HSM_REQUEST:
2084 rc = mdc_ioc_hsm_request(exp, karg);
2086 case OBD_IOC_CLIENT_RECOVER:
2087 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2091 case IOC_OSC_SET_ACTIVE:
2092 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2094 case OBD_IOC_PING_TARGET:
2095 rc = ptlrpc_obd_ping(obd);
2098 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2099 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2100 * there'd be no LMV layer thus we might be called here. Eventually
2101 * this code should be removed.
2104 case IOC_OBD_STATFS: {
2105 struct obd_statfs stat_buf = {0};
2107 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2108 GOTO(out, rc = -ENODEV);
2111 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2112 min((int)data->ioc_plen2,
2113 (int)sizeof(struct obd_uuid))))
2114 GOTO(out, rc = -EFAULT);
2116 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2117 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2122 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2123 min((int) data->ioc_plen1,
2124 (int) sizeof(stat_buf))))
2125 GOTO(out, rc = -EFAULT);
2129 case OBD_IOC_QUOTACTL: {
2130 struct if_quotactl *qctl = karg;
2131 struct obd_quotactl *oqctl;
2133 OBD_ALLOC_PTR(oqctl);
2135 GOTO(out, rc = -ENOMEM);
2137 QCTL_COPY(oqctl, qctl);
2138 rc = obd_quotactl(exp, oqctl);
2140 QCTL_COPY(qctl, oqctl);
2141 qctl->qc_valid = QC_MDTIDX;
2142 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2145 OBD_FREE_PTR(oqctl);
2148 case LL_IOC_GET_CONNECT_FLAGS:
2149 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2150 sizeof(*exp_connect_flags_ptr(exp))))
2151 GOTO(out, rc = -EFAULT);
2154 case LL_IOC_LOV_SWAP_LAYOUTS:
2155 rc = mdc_ioc_swap_layouts(exp, karg);
2158 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2159 GOTO(out, rc = -ENOTTY);
2162 module_put(THIS_MODULE);
2167 static int mdc_get_info_rpc(struct obd_export *exp,
2168 u32 keylen, void *key,
2169 u32 vallen, void *val)
2171 struct obd_import *imp = class_exp2cliimp(exp);
2172 struct ptlrpc_request *req;
2177 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2181 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2182 RCL_CLIENT, keylen);
2183 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2184 RCL_CLIENT, sizeof(vallen));
2186 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2188 ptlrpc_request_free(req);
2192 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2193 memcpy(tmp, key, keylen);
2194 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2195 memcpy(tmp, &vallen, sizeof(vallen));
2197 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2198 RCL_SERVER, vallen);
2199 ptlrpc_request_set_replen(req);
2201 rc = ptlrpc_queue_wait(req);
2202 /* -EREMOTE means the get_info result is partial, and it needs to
2203 * continue on another MDT, see fid2path part in lmv_iocontrol */
2204 if (rc == 0 || rc == -EREMOTE) {
2205 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2206 memcpy(val, tmp, vallen);
2207 if (ptlrpc_rep_need_swab(req)) {
2208 if (KEY_IS(KEY_FID2PATH))
2209 lustre_swab_fid2path(val);
2212 ptlrpc_req_finished(req);
2217 static void lustre_swab_hai(struct hsm_action_item *h)
2219 __swab32s(&h->hai_len);
2220 __swab32s(&h->hai_action);
2221 lustre_swab_lu_fid(&h->hai_fid);
2222 lustre_swab_lu_fid(&h->hai_dfid);
2223 __swab64s(&h->hai_cookie);
2224 __swab64s(&h->hai_extent.offset);
2225 __swab64s(&h->hai_extent.length);
2226 __swab64s(&h->hai_gid);
2229 static void lustre_swab_hal(struct hsm_action_list *h)
2231 struct hsm_action_item *hai;
2234 __swab32s(&h->hal_version);
2235 __swab32s(&h->hal_count);
2236 __swab32s(&h->hal_archive_id);
2237 __swab64s(&h->hal_flags);
2239 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2240 lustre_swab_hai(hai);
2243 static void lustre_swab_kuch(struct kuc_hdr *l)
2245 __swab16s(&l->kuc_magic);
2246 /* __u8 l->kuc_transport */
2247 __swab16s(&l->kuc_msgtype);
2248 __swab16s(&l->kuc_msglen);
2251 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2252 struct lustre_kernelcomm *lk)
2254 struct obd_import *imp = class_exp2cliimp(exp);
2257 if (lk->lk_group != KUC_GRP_HSM) {
2258 CERROR("Bad copytool group %d\n", lk->lk_group);
2262 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2263 lk->lk_uid, lk->lk_group, lk->lk_flags);
2265 if (lk->lk_flags & LK_FLG_STOP) {
2266 /* Unregister with the coordinator */
2267 rc = mdc_ioc_hsm_ct_unregister(imp);
2269 __u32 *archives = NULL;
2271 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2272 archives = lk->lk_data;
2274 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2281 * Send a message to any listening copytools
2282 * @param val KUC message (kuc_hdr + hsm_action_list)
2283 * @param len total length of message
2285 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2286 size_t len, void *val)
2288 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2289 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2293 if (len < sizeof(*lh) + sizeof(*hal)) {
2294 CERROR("Short HSM message %zu < %zu\n", len,
2295 sizeof(*lh) + sizeof(*hal));
2298 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2299 lustre_swab_kuch(lh);
2300 lustre_swab_hal(hal);
2301 } else if (lh->kuc_magic != KUC_MAGIC) {
2302 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2306 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2308 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2309 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2311 /* Broadcast to HSM listeners */
2312 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2318 * callback function passed to kuc for re-registering each HSM copytool
2319 * running on MDC, after MDT shutdown/recovery.
2320 * @param data copytool registration data
2321 * @param cb_arg callback argument (obd_import)
2323 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2325 struct obd_import *imp = (struct obd_import *)cb_arg;
2326 struct kkuc_ct_data *kcd = data;
2327 __u32 *archives = NULL;
2331 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2332 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2335 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2336 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2337 "(archive=%#x)\n", imp->imp_obd->obd_name,
2338 kcd->kcd_nr_archives);
2340 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2341 "(archive nr = %u)\n",
2342 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2343 if (kcd->kcd_nr_archives != 0)
2344 archives = kcd->kcd_archives;
2347 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2348 /* ignore error if the copytool is already registered */
2349 return (rc == -EEXIST) ? 0 : rc;
2353 * Re-establish all kuc contexts with MDT
2354 * after MDT shutdown/recovery.
2356 static int mdc_kuc_reregister(struct obd_import *imp)
2358 /* re-register HSM agents */
2359 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2360 mdc_hsm_ct_reregister, imp);
2363 static int mdc_set_info_async(const struct lu_env *env,
2364 struct obd_export *exp,
2365 u32 keylen, void *key,
2366 u32 vallen, void *val,
2367 struct ptlrpc_request_set *set)
2369 struct obd_import *imp = class_exp2cliimp(exp);
2373 if (KEY_IS(KEY_READ_ONLY)) {
2374 if (vallen != sizeof(int))
2377 spin_lock(&imp->imp_lock);
2378 if (*((int *)val)) {
2379 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2380 imp->imp_connect_data.ocd_connect_flags |=
2383 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2384 imp->imp_connect_data.ocd_connect_flags &=
2385 ~OBD_CONNECT_RDONLY;
2387 spin_unlock(&imp->imp_lock);
2389 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2390 keylen, key, vallen, val, set);
2393 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2394 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2395 keylen, key, vallen, val, set);
2398 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2399 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2404 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2405 __u32 *default_easize = val;
2407 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2411 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2415 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2416 __u32 keylen, void *key, __u32 *vallen, void *val)
2420 if (KEY_IS(KEY_MAX_EASIZE)) {
2421 __u32 mdsize, *max_easize;
2423 if (*vallen != sizeof(int))
2425 mdsize = *(__u32 *)val;
2426 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2427 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2429 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2431 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2432 __u32 *default_easize;
2434 if (*vallen != sizeof(int))
2436 default_easize = val;
2437 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2439 } else if (KEY_IS(KEY_CONN_DATA)) {
2440 struct obd_import *imp = class_exp2cliimp(exp);
2441 struct obd_connect_data *data = val;
2443 if (*vallen != sizeof(*data))
2446 *data = imp->imp_connect_data;
2448 } else if (KEY_IS(KEY_TGT_COUNT)) {
2449 *((__u32 *)val) = 1;
2453 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2458 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2459 struct ptlrpc_request **request)
2461 struct ptlrpc_request *req;
2466 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2470 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2472 ptlrpc_request_free(req);
2476 mdc_pack_body(req, fid, 0, 0, -1, 0);
2478 ptlrpc_request_set_replen(req);
2480 rc = ptlrpc_queue_wait(req);
2482 ptlrpc_req_finished(req);
2488 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2489 enum obd_import_event event)
2491 struct client_obd *cli = &obd->u.cli;
2494 LASSERT(imp->imp_obd == obd);
2497 case IMP_EVENT_DISCON:
2498 spin_lock(&cli->cl_loi_list_lock);
2499 cli->cl_avail_grant = 0;
2500 cli->cl_lost_grant = 0;
2501 spin_unlock(&cli->cl_loi_list_lock);
2503 case IMP_EVENT_INACTIVE:
2505 * Flush current sequence to make client obtain new one
2506 * from server in case of disconnect/reconnect.
2508 down_read(&cli->cl_seq_rwsem);
2510 seq_client_flush(cli->cl_seq);
2511 up_read(&cli->cl_seq_rwsem);
2513 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2515 case IMP_EVENT_INVALIDATE: {
2516 struct ldlm_namespace *ns = obd->obd_namespace;
2520 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2522 env = cl_env_get(&refcheck);
2524 /* Reset grants. All pages go to failing rpcs due to
2525 * the invalid import.
2527 osc_io_unplug(env, cli, NULL);
2529 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2530 osc_ldlm_resource_invalidate,
2532 cl_env_put(env, &refcheck);
2533 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2539 case IMP_EVENT_ACTIVE:
2540 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2541 /* redo the kuc registration after reconnecting */
2543 rc = mdc_kuc_reregister(imp);
2545 case IMP_EVENT_OCD: {
2546 struct obd_connect_data *ocd = &imp->imp_connect_data;
2548 if (OCD_HAS_FLAG(ocd, GRANT))
2549 osc_init_grant(cli, ocd);
2551 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2554 case IMP_EVENT_DEACTIVATE:
2555 case IMP_EVENT_ACTIVATE:
2558 CERROR("Unknown import event %x\n", event);
2564 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2565 struct lu_fid *fid, struct md_op_data *op_data)
2567 struct client_obd *cli = &exp->exp_obd->u.cli;
2572 down_read(&cli->cl_seq_rwsem);
2574 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2575 up_read(&cli->cl_seq_rwsem);
2580 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2582 struct client_obd *cli = &exp->exp_obd->u.cli;
2583 return &cli->cl_target_uuid;
2587 * Determine whether the lock can be canceled before replaying it during
2588 * recovery, non zero value will be return if the lock can be canceled,
2589 * or zero returned for not
2591 static int mdc_cancel_weight(struct ldlm_lock *lock)
2593 if (lock->l_resource->lr_type != LDLM_IBITS)
2596 /* FIXME: if we ever get into a situation where there are too many
2597 * opened files with open locks on a single node, then we really
2598 * should replay these open locks to reget it */
2599 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2602 /* Special case for DoM locks, cancel only unused and granted locks */
2603 if (ldlm_has_dom(lock) &&
2604 (lock->l_granted_mode != lock->l_req_mode ||
2605 osc_ldlm_weigh_ast(lock) != 0))
2611 static int mdc_resource_inode_free(struct ldlm_resource *res)
2613 if (res->lr_lvb_inode)
2614 res->lr_lvb_inode = NULL;
2619 static struct ldlm_valblock_ops inode_lvbo = {
2620 .lvbo_free = mdc_resource_inode_free
2623 static int mdc_llog_init(struct obd_device *obd)
2625 struct obd_llog_group *olg = &obd->obd_olg;
2626 struct llog_ctxt *ctxt;
2631 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2636 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2637 llog_initiator_connect(ctxt);
2638 llog_ctxt_put(ctxt);
2643 static void mdc_llog_finish(struct obd_device *obd)
2645 struct llog_ctxt *ctxt;
2649 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2651 llog_cleanup(NULL, ctxt);
2656 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2662 rc = osc_setup_common(obd, cfg);
2666 rc = mdc_tunables_init(obd);
2668 GOTO(err_osc_cleanup, rc);
2670 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2672 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2674 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2676 rc = mdc_llog_init(obd);
2678 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2680 GOTO(err_llog_cleanup, rc);
2683 rc = mdc_changelog_cdev_init(obd);
2685 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2687 GOTO(err_changelog_cleanup, rc);
2692 err_changelog_cleanup:
2693 mdc_llog_finish(obd);
2695 lprocfs_free_md_stats(obd);
2696 ptlrpc_lprocfs_unregister_obd(obd);
2698 osc_cleanup_common(obd);
2702 /* Initialize the default and maximum LOV EA sizes. This allows
2703 * us to make MDS RPCs with large enough reply buffers to hold a default
2704 * sized EA without having to calculate this (via a call into the
2705 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2706 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2707 * a large number of stripes is possible. If a larger reply buffer is
2708 * required it will be reallocated in the ptlrpc layer due to overflow.
2710 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2713 struct obd_device *obd = exp->exp_obd;
2714 struct client_obd *cli = &obd->u.cli;
2717 if (cli->cl_max_mds_easize < easize)
2718 cli->cl_max_mds_easize = easize;
2720 if (cli->cl_default_mds_easize < def_easize)
2721 cli->cl_default_mds_easize = def_easize;
2726 static int mdc_precleanup(struct obd_device *obd)
2730 osc_precleanup_common(obd);
2731 mdc_changelog_cdev_finish(obd);
2733 obd_cleanup_client_import(obd);
2734 ptlrpc_lprocfs_unregister_obd(obd);
2735 lprocfs_free_md_stats(obd);
2736 mdc_llog_finish(obd);
2740 static int mdc_cleanup(struct obd_device *obd)
2742 return osc_cleanup_common(obd);
2745 int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2747 struct lustre_cfg *lcfg = buf;
2748 size_t count = class_modify_config(lcfg, PARAM_MDC,
2749 &obd->obd_kset.kobj);
2751 return count > 0 ? 0 : count;
2754 static struct obd_ops mdc_obd_ops = {
2755 .o_owner = THIS_MODULE,
2756 .o_setup = mdc_setup,
2757 .o_precleanup = mdc_precleanup,
2758 .o_cleanup = mdc_cleanup,
2759 .o_add_conn = client_import_add_conn,
2760 .o_del_conn = client_import_del_conn,
2761 .o_connect = client_connect_import,
2762 .o_reconnect = osc_reconnect,
2763 .o_disconnect = osc_disconnect,
2764 .o_iocontrol = mdc_iocontrol,
2765 .o_set_info_async = mdc_set_info_async,
2766 .o_statfs = mdc_statfs,
2767 .o_fid_init = client_fid_init,
2768 .o_fid_fini = client_fid_fini,
2769 .o_fid_alloc = mdc_fid_alloc,
2770 .o_import_event = mdc_import_event,
2771 .o_get_info = mdc_get_info,
2772 .o_process_config = mdc_process_config,
2773 .o_get_uuid = mdc_get_uuid,
2774 .o_quotactl = mdc_quotactl,
2777 static struct md_ops mdc_md_ops = {
2778 .m_get_root = mdc_get_root,
2779 .m_null_inode = mdc_null_inode,
2780 .m_close = mdc_close,
2781 .m_create = mdc_create,
2782 .m_enqueue = mdc_enqueue,
2783 .m_getattr = mdc_getattr,
2784 .m_getattr_name = mdc_getattr_name,
2785 .m_intent_lock = mdc_intent_lock,
2787 .m_rename = mdc_rename,
2788 .m_setattr = mdc_setattr,
2789 .m_setxattr = mdc_setxattr,
2790 .m_getxattr = mdc_getxattr,
2791 .m_fsync = mdc_fsync,
2792 .m_file_resync = mdc_file_resync,
2793 .m_read_page = mdc_read_page,
2794 .m_unlink = mdc_unlink,
2795 .m_cancel_unused = mdc_cancel_unused,
2796 .m_init_ea_size = mdc_init_ea_size,
2797 .m_set_lock_data = mdc_set_lock_data,
2798 .m_lock_match = mdc_lock_match,
2799 .m_get_lustre_md = mdc_get_lustre_md,
2800 .m_free_lustre_md = mdc_free_lustre_md,
2801 .m_set_open_replay_data = mdc_set_open_replay_data,
2802 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2803 .m_intent_getattr_async = mdc_intent_getattr_async,
2804 .m_revalidate_lock = mdc_revalidate_lock
2807 static int __init mdc_init(void)
2809 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2810 LUSTRE_MDC_NAME, &mdc_device_type);
2813 static void __exit mdc_exit(void)
2815 class_unregister_type(LUSTRE_MDC_NAME);
2818 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2819 MODULE_DESCRIPTION("Lustre Metadata Client");
2820 MODULE_VERSION(LUSTRE_VERSION_STRING);
2821 MODULE_LICENSE("GPL");
2823 module_init(mdc_init);
2824 module_exit(mdc_exit);