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>
44 #include <linux/device.h>
46 #include <lustre_errno.h>
48 #include <cl_object.h>
49 #include <llog_swab.h>
50 #include <lprocfs_status.h>
51 #include <lustre_acl.h>
52 #include <lustre_fid.h>
53 #include <uapi/linux/lustre/lustre_ioctl.h>
54 #include <lustre_kernelcomm.h>
55 #include <lustre_lmv.h>
56 #include <lustre_log.h>
57 #include <uapi/linux/lustre/lustre_param.h>
58 #include <lustre_swab.h>
59 #include <obd_class.h>
60 #include <lustre_osc.h>
62 #include "mdc_internal.h"
64 #define REQUEST_MINOR 244
66 static int mdc_cleanup(struct obd_device *obd);
68 static inline int mdc_queue_wait(struct ptlrpc_request *req)
70 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
73 /* obd_get_request_slot() ensures that this client has no more
74 * than cl_max_rpcs_in_flight RPCs simultaneously inf light
76 rc = obd_get_request_slot(cli);
80 rc = ptlrpc_queue_wait(req);
81 obd_put_request_slot(cli);
87 * Send MDS_GET_ROOT RPC to fetch root FID.
89 * If \a fileset is not NULL it should contain a subdirectory off
90 * the ROOT/ directory to be mounted on the client. Return the FID
91 * of the subdirectory to the client to mount onto its mountpoint.
93 * \param[in] imp MDC import
94 * \param[in] fileset fileset name, which could be NULL
95 * \param[out] rootfid root FID of this mountpoint
96 * \param[out] pc root capa will be unpacked and saved in this pointer
98 * \retval 0 on success, negative errno on failure
100 static int mdc_get_root(struct obd_export *exp, const char *fileset,
101 struct lu_fid *rootfid)
103 struct ptlrpc_request *req;
104 struct mdt_body *body;
109 if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
112 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
118 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
119 strlen(fileset) + 1);
120 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
122 ptlrpc_request_free(req);
125 mdc_pack_body(req, NULL, 0, 0, -1, 0);
126 if (fileset != NULL) {
127 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
129 memcpy(name, fileset, strlen(fileset));
131 lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
132 req->rq_send_state = LUSTRE_IMP_FULL;
134 ptlrpc_request_set_replen(req);
136 rc = ptlrpc_queue_wait(req);
140 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
142 GOTO(out, rc = -EPROTO);
144 *rootfid = body->mbo_fid1;
145 CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
146 PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
149 ptlrpc_req_finished(req);
155 * This function now is known to always saying that it will receive 4 buffers
156 * from server. Even for cases when acl_size and md_size is zero, RPC header
157 * will contain 4 fields and RPC itself will contain zero size fields. This is
158 * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
159 * and thus zero, it shrinks it, making zero size. The same story about
160 * md_size. And this is course of problem when client waits for smaller number
161 * of fields. This issue will be fixed later when client gets aware of RPC
164 static int mdc_getattr_common(struct obd_export *exp,
165 struct ptlrpc_request *req)
167 struct req_capsule *pill = &req->rq_pill;
168 struct mdt_body *body;
173 /* Request message already built. */
174 rc = ptlrpc_queue_wait(req);
178 /* sanity check for the reply */
179 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
183 CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
185 mdc_update_max_ea_from_body(exp, body);
186 if (body->mbo_eadatasize != 0) {
187 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
188 body->mbo_eadatasize);
196 static void mdc_reset_acl_req(struct ptlrpc_request *req)
198 spin_lock(&req->rq_early_free_lock);
199 sptlrpc_cli_free_repbuf(req);
200 req->rq_repbuf = NULL;
201 req->rq_repbuf_len = 0;
202 req->rq_repdata = NULL;
203 req->rq_reqdata_len = 0;
204 spin_unlock(&req->rq_early_free_lock);
207 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
208 struct ptlrpc_request **request)
210 struct ptlrpc_request *req;
211 struct obd_import *imp = class_exp2cliimp(exp);
212 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
216 /* Single MDS without an LMV case */
217 if (op_data->op_flags & MF_GET_MDT_IDX) {
223 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
227 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
229 ptlrpc_request_free(req);
234 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
235 op_data->op_mode, -1, 0);
236 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
237 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
239 ptlrpc_request_set_replen(req);
241 rc = mdc_getattr_common(exp, req);
244 acl_bufsize != imp->imp_connect_data.ocd_max_easize) {
245 acl_bufsize = imp->imp_connect_data.ocd_max_easize;
246 mdc_reset_acl_req(req);
250 ptlrpc_req_finished(req);
258 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
259 struct ptlrpc_request **request)
261 struct ptlrpc_request *req;
262 struct obd_import *imp = class_exp2cliimp(exp);
263 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
268 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
272 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
273 op_data->op_namelen + 1);
275 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
277 ptlrpc_request_free(req);
281 if (op_data->op_name) {
282 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
283 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
284 op_data->op_namelen);
285 memcpy(name, op_data->op_name, op_data->op_namelen);
289 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
290 op_data->op_mode, op_data->op_suppgids[0], 0);
291 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
293 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
294 ptlrpc_request_set_replen(req);
296 rc = mdc_getattr_common(exp, req);
299 acl_bufsize != imp->imp_connect_data.ocd_max_easize) {
300 acl_bufsize = imp->imp_connect_data.ocd_max_easize;
301 mdc_reset_acl_req(req);
305 ptlrpc_req_finished(req);
313 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
314 const struct lu_fid *fid, int opcode, u64 valid,
315 const char *xattr_name, const char *input,
316 int input_size, int output_size, int flags,
317 __u32 suppgid, struct ptlrpc_request **request)
319 struct ptlrpc_request *req;
320 int xattr_namelen = 0;
326 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
331 xattr_namelen = strlen(xattr_name) + 1;
332 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
337 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
340 /* get SELinux policy info if any */
341 rc = sptlrpc_get_sepol(req);
343 ptlrpc_request_free(req);
346 req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
347 strlen(req->rq_sepol) ?
348 strlen(req->rq_sepol) + 1 : 0);
350 /* Flush local XATTR locks to get rid of a possible cancel RPC */
351 if (opcode == MDS_REINT && fid_is_sane(fid) &&
352 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
353 struct list_head cancels = LIST_HEAD_INIT(cancels);
356 /* Without that packing would fail */
358 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
361 count = mdc_resource_get_unused(exp, fid,
363 MDS_INODELOCK_XATTR);
365 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
367 ptlrpc_request_free(req);
371 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
373 ptlrpc_request_free(req);
378 if (opcode == MDS_REINT) {
379 struct mdt_rec_setxattr *rec;
381 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
382 sizeof(struct mdt_rec_reint));
383 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
384 rec->sx_opcode = REINT_SETXATTR;
385 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
386 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
387 rec->sx_cap = cfs_curproc_cap_pack();
388 rec->sx_suppgid1 = suppgid;
389 rec->sx_suppgid2 = -1;
391 rec->sx_valid = valid | OBD_MD_FLCTIME;
392 rec->sx_time = ktime_get_real_seconds();
393 rec->sx_size = output_size;
394 rec->sx_flags = flags;
396 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
400 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
401 memcpy(tmp, xattr_name, xattr_namelen);
404 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
405 memcpy(tmp, input, input_size);
408 mdc_file_sepol_pack(req);
410 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
411 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
412 RCL_SERVER, output_size);
413 ptlrpc_request_set_replen(req);
416 if (opcode == MDS_REINT)
417 mdc_get_mod_rpc_slot(req, NULL);
419 rc = ptlrpc_queue_wait(req);
421 if (opcode == MDS_REINT)
422 mdc_put_mod_rpc_slot(req, NULL);
425 ptlrpc_req_finished(req);
431 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
432 u64 obd_md_valid, const char *name,
433 const void *value, size_t value_size,
434 unsigned int xattr_flags, u32 suppgid,
435 struct ptlrpc_request **req)
437 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
438 obd_md_valid == OBD_MD_FLXATTRRM);
440 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
441 fid, MDS_REINT, obd_md_valid, name,
442 value, value_size, 0, xattr_flags, suppgid,
446 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
447 u64 obd_md_valid, const char *name, size_t buf_size,
448 struct ptlrpc_request **req)
450 struct mdt_body *body;
453 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
454 obd_md_valid == OBD_MD_FLXATTRLS);
456 CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
457 exp->exp_obd->obd_name, name, PFID(fid));
458 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
459 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
464 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
466 GOTO(out, rc = -EPROTO);
468 /* only detect the xattr size */
470 /* LU-11109: Older MDTs do not distinguish
471 * between nonexistent xattrs and zero length
472 * values in this case. Newer MDTs will return
473 * -ENODATA or set OBD_MD_FLXATTR. */
474 GOTO(out, rc = body->mbo_eadatasize);
477 if (body->mbo_eadatasize == 0) {
478 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
479 * success so that we can distinguish between
480 * zero length value and nonexistent xattr.
482 * If OBD_MD_FLXATTR is not set then we keep
483 * the old behavior and return -ENODATA for
484 * getxattr() when mbo_eadatasize is 0. But
485 * -ENODATA only makes sense for getxattr()
486 * and not for listxattr(). */
487 if (body->mbo_valid & OBD_MD_FLXATTR)
489 else if (obd_md_valid == OBD_MD_FLXATTR)
490 GOTO(out, rc = -ENODATA);
495 GOTO(out, rc = body->mbo_eadatasize);
498 ptlrpc_req_finished(*req);
505 #ifdef CONFIG_FS_POSIX_ACL
506 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
508 struct req_capsule *pill = &req->rq_pill;
509 struct mdt_body *body = md->body;
510 struct posix_acl *acl;
515 if (!body->mbo_aclsize)
518 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
523 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
528 CERROR("convert xattr to acl: %d\n", rc);
532 rc = posix_acl_valid(&init_user_ns, acl);
534 CERROR("validate acl: %d\n", rc);
535 posix_acl_release(acl);
543 #define mdc_unpack_acl(req, md) 0
546 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
547 struct obd_export *dt_exp, struct obd_export *md_exp,
548 struct lustre_md *md)
550 struct req_capsule *pill = &req->rq_pill;
555 memset(md, 0, sizeof(*md));
557 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
558 LASSERT(md->body != NULL);
560 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
561 if (!S_ISREG(md->body->mbo_mode)) {
562 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
563 "regular file, but is not\n");
564 GOTO(out, rc = -EPROTO);
567 if (md->body->mbo_eadatasize == 0) {
568 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
569 "but eadatasize 0\n");
570 GOTO(out, rc = -EPROTO);
573 md->layout.lb_len = md->body->mbo_eadatasize;
574 md->layout.lb_buf = req_capsule_server_sized_get(pill,
577 if (md->layout.lb_buf == NULL)
578 GOTO(out, rc = -EPROTO);
579 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
580 const union lmv_mds_md *lmv;
583 if (!S_ISDIR(md->body->mbo_mode)) {
584 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
585 "directory, but is not\n");
586 GOTO(out, rc = -EPROTO);
589 lmv_size = md->body->mbo_eadatasize;
591 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
592 "but eadatasize 0\n");
596 if (md->body->mbo_valid & OBD_MD_MEA) {
597 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
600 GOTO(out, rc = -EPROTO);
602 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
606 if (rc < (typeof(rc))sizeof(*md->lmv)) {
607 CDEBUG(D_INFO, "size too small: "
608 "rc < sizeof(*md->lmv) (%d < %d)\n",
609 rc, (int)sizeof(*md->lmv));
610 GOTO(out, rc = -EPROTO);
616 if (md->body->mbo_valid & OBD_MD_FLACL) {
617 /* for ACL, it's possible that FLACL is set but aclsize is zero.
618 * only when aclsize != 0 there's an actual segment for ACL
621 if (md->body->mbo_aclsize) {
622 rc = mdc_unpack_acl(req, md);
625 #ifdef CONFIG_FS_POSIX_ACL
627 md->posix_acl = NULL;
635 #ifdef CONFIG_FS_POSIX_ACL
636 posix_acl_release(md->posix_acl);
642 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
648 void mdc_replay_open(struct ptlrpc_request *req)
650 struct md_open_data *mod = req->rq_cb_data;
651 struct ptlrpc_request *close_req;
652 struct obd_client_handle *och;
653 struct lustre_handle old_open_handle = { };
654 struct mdt_body *body;
658 DEBUG_REQ(D_ERROR, req,
659 "Can't properly replay without open data.");
664 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
665 LASSERT(body != NULL);
667 spin_lock(&req->rq_lock);
669 if (och && och->och_open_handle.cookie)
670 req->rq_early_free_repbuf = 1;
672 req->rq_early_free_repbuf = 0;
673 spin_unlock(&req->rq_lock);
675 if (req->rq_early_free_repbuf) {
676 struct lustre_handle *file_open_handle;
678 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
680 file_open_handle = &och->och_open_handle;
681 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
682 file_open_handle->cookie, body->mbo_open_handle.cookie);
683 old_open_handle = *file_open_handle;
684 *file_open_handle = body->mbo_open_handle;
687 close_req = mod->mod_close_req;
689 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
690 struct mdt_ioepoch *epoch;
692 LASSERT(opc == MDS_CLOSE);
693 epoch = req_capsule_client_get(&close_req->rq_pill,
697 if (req->rq_early_free_repbuf)
698 LASSERT(old_open_handle.cookie ==
699 epoch->mio_open_handle.cookie);
701 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
702 epoch->mio_open_handle = body->mbo_open_handle;
707 void mdc_commit_open(struct ptlrpc_request *req)
709 struct md_open_data *mod = req->rq_cb_data;
714 * No need to touch md_open_data::mod_och, it holds a reference on
715 * \var mod and will zero references to each other, \var mod will be
716 * freed after that when md_open_data::mod_och will put the reference.
720 * Do not let open request to disappear as it still may be needed
721 * for close rpc to happen (it may happen on evict only, otherwise
722 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
723 * called), just mark this rpc as committed to distinguish these 2
724 * cases, see mdc_close() for details. The open request reference will
725 * be put along with freeing \var mod.
727 ptlrpc_request_addref(req);
728 spin_lock(&req->rq_lock);
729 req->rq_committed = 1;
730 spin_unlock(&req->rq_lock);
731 req->rq_cb_data = NULL;
735 int mdc_set_open_replay_data(struct obd_export *exp,
736 struct obd_client_handle *och,
737 struct lookup_intent *it)
739 struct md_open_data *mod;
740 struct mdt_rec_create *rec;
741 struct mdt_body *body;
742 struct ptlrpc_request *open_req = it->it_request;
743 struct obd_import *imp = open_req->rq_import;
746 if (!open_req->rq_replay)
749 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
750 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
751 LASSERT(rec != NULL);
752 /* Incoming message in my byte order (it's been swabbed). */
753 /* Outgoing messages always in my byte order. */
754 LASSERT(body != NULL);
756 /* Only if the import is replayable, we set replay_open data */
757 if (och && imp->imp_replayable) {
758 mod = obd_mod_alloc();
760 DEBUG_REQ(D_ERROR, open_req,
761 "Can't allocate md_open_data");
766 * Take a reference on \var mod, to be freed on mdc_close().
767 * It protects \var mod from being freed on eviction (commit
768 * callback is called despite rq_replay flag).
769 * Another reference for \var och.
774 spin_lock(&open_req->rq_lock);
777 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
778 it_disposition(it, DISP_OPEN_STRIPE);
779 mod->mod_open_req = open_req;
780 open_req->rq_cb_data = mod;
781 open_req->rq_commit_cb = mdc_commit_open;
782 open_req->rq_early_free_repbuf = 1;
783 spin_unlock(&open_req->rq_lock);
786 rec->cr_fid2 = body->mbo_fid1;
787 rec->cr_open_handle_old = body->mbo_open_handle;
788 open_req->rq_replay_cb = mdc_replay_open;
789 if (!fid_is_sane(&body->mbo_fid1)) {
790 DEBUG_REQ(D_ERROR, open_req,
791 "saving replay request with insane FID " DFID,
792 PFID(&body->mbo_fid1));
796 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
800 static void mdc_free_open(struct md_open_data *mod)
804 if (mod->mod_is_create == 0 &&
805 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
809 * No reason to asssert here if the open request has
810 * rq_replay == 1. It means that mdc_close failed, and
811 * close request wasn`t sent. It is not fatal to client.
812 * The worst thing is eviction if the client gets open lock
815 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
816 "= %d\n", mod->mod_open_req->rq_replay);
818 ptlrpc_request_committed(mod->mod_open_req, committed);
819 if (mod->mod_close_req)
820 ptlrpc_request_committed(mod->mod_close_req, committed);
823 int mdc_clear_open_replay_data(struct obd_export *exp,
824 struct obd_client_handle *och)
826 struct md_open_data *mod = och->och_mod;
830 * It is possible to not have \var mod in a case of eviction between
831 * lookup and ll_file_open().
836 LASSERT(mod != LP_POISON);
837 LASSERT(mod->mod_open_req != NULL);
839 spin_lock(&mod->mod_open_req->rq_lock);
841 mod->mod_och->och_open_handle.cookie = 0;
842 mod->mod_open_req->rq_early_free_repbuf = 0;
843 spin_unlock(&mod->mod_open_req->rq_lock);
853 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
854 struct md_open_data *mod, struct ptlrpc_request **request)
856 struct obd_device *obd = class_exp2obd(exp);
857 struct ptlrpc_request *req;
858 struct req_format *req_fmt;
859 size_t u32_count = 0;
864 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
865 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
868 if (op_data->op_bias & MDS_CLOSE_INTENT) {
869 req_fmt = &RQF_MDS_CLOSE_INTENT;
870 if (op_data->op_bias & MDS_HSM_RELEASE) {
871 /* allocate a FID for volatile file */
872 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
875 CERROR("%s: "DFID" allocating FID: rc = %d\n",
876 obd->obd_name, PFID(&op_data->op_fid1),
878 /* save the errcode and proceed to close */
882 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
883 size_t count = op_data->op_data_size / sizeof(__u32);
885 if (count > INLINE_RESYNC_ARRAY_SIZE)
889 req_fmt = &RQF_MDS_CLOSE;
893 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
896 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
898 /* Ensure that this close's handle is fixed up during replay. */
899 if (likely(mod != NULL)) {
900 LASSERTF(mod->mod_open_req != NULL &&
901 mod->mod_open_req->rq_type != LI_POISON,
902 "POISONED open %p!\n", mod->mod_open_req);
904 mod->mod_close_req = req;
906 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
907 /* We no longer want to preserve this open for replay even
908 * though the open was committed. b=3632, b=3633 */
909 spin_lock(&mod->mod_open_req->rq_lock);
910 mod->mod_open_req->rq_replay = 0;
911 spin_unlock(&mod->mod_open_req->rq_lock);
913 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
917 * TODO: repeat close after errors
919 CWARN("%s: close of FID "DFID" failed, file reference will be "
920 "dropped when this client unmounts or is evicted\n",
921 obd->obd_name, PFID(&op_data->op_fid1));
922 GOTO(out, rc = -ENOMEM);
926 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
927 u32_count * sizeof(__u32));
929 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
931 ptlrpc_request_free(req);
936 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
937 * portal whose threads are not taking any DLM locks and are therefore
938 * always progressing */
939 req->rq_request_portal = MDS_READPAGE_PORTAL;
940 ptlrpc_at_set_req_timeout(req);
942 if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
943 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
944 OP_XVALID_LAZYBLOCKS);
946 mdc_close_pack(req, op_data);
948 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
949 obd->u.cli.cl_default_mds_easize);
951 ptlrpc_request_set_replen(req);
953 mdc_get_mod_rpc_slot(req, NULL);
954 rc = ptlrpc_queue_wait(req);
955 mdc_put_mod_rpc_slot(req, NULL);
957 if (req->rq_repmsg == NULL) {
958 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
961 rc = req->rq_status ?: -EIO;
962 } else if (rc == 0 || rc == -EAGAIN) {
963 struct mdt_body *body;
965 rc = lustre_msg_get_status(req->rq_repmsg);
966 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
967 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
972 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
975 } else if (rc == -ESTALE) {
977 * it can be allowed error after 3633 if open was committed and
978 * server failed before close was sent. Let's check if mod
979 * exists and return no error in that case
982 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
983 LASSERT(mod->mod_open_req != NULL);
984 if (mod->mod_open_req->rq_committed)
992 mod->mod_close_req = NULL;
993 /* Since now, mod is accessed through open_req only,
994 * thus close req does not keep a reference on mod anymore. */
999 RETURN(rc < 0 ? rc : saved_rc);
1002 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1003 u64 offset, struct page **pages, int npages,
1004 struct ptlrpc_request **request)
1006 struct ptlrpc_request *req;
1007 struct ptlrpc_bulk_desc *desc;
1009 wait_queue_head_t waitq;
1011 struct l_wait_info lwi;
1016 init_waitqueue_head(&waitq);
1019 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1023 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1025 ptlrpc_request_free(req);
1029 req->rq_request_portal = MDS_READPAGE_PORTAL;
1030 ptlrpc_at_set_req_timeout(req);
1032 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1033 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1035 &ptlrpc_bulk_kiov_pin_ops);
1037 ptlrpc_req_finished(req);
1041 /* NB req now owns desc and will free it when it gets freed */
1042 for (i = 0; i < npages; i++)
1043 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1046 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1048 ptlrpc_request_set_replen(req);
1049 rc = ptlrpc_queue_wait(req);
1051 ptlrpc_req_finished(req);
1052 if (rc != -ETIMEDOUT)
1056 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1057 CERROR("%s: too many resend retries: rc = %d\n",
1058 exp->exp_obd->obd_name, -EIO);
1061 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1063 l_wait_event(waitq, 0, &lwi);
1068 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1069 req->rq_bulk->bd_nob_transferred);
1071 ptlrpc_req_finished(req);
1075 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1076 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1077 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1078 PAGE_SIZE * npages);
1079 ptlrpc_req_finished(req);
1087 static void mdc_release_page(struct page *page, int remove)
1091 if (likely(page->mapping != NULL))
1092 truncate_complete_page(page->mapping, page);
1098 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1099 __u64 *start, __u64 *end, int hash64)
1102 * Complement of hash is used as an index so that
1103 * radix_tree_gang_lookup() can be used to find a page with starting
1104 * hash _smaller_ than one we are looking for.
1106 unsigned long offset = hash_x_index(*hash, hash64);
1110 xa_lock_irq(&mapping->i_pages);
1111 found = radix_tree_gang_lookup(&mapping->page_tree,
1112 (void **)&page, offset, 1);
1113 if (found > 0 && !xa_is_value(page)) {
1114 struct lu_dirpage *dp;
1117 xa_unlock_irq(&mapping->i_pages);
1119 * In contrast to find_lock_page() we are sure that directory
1120 * page cannot be truncated (while DLM lock is held) and,
1121 * hence, can avoid restart.
1123 * In fact, page cannot be locked here at all, because
1124 * mdc_read_page_remote does synchronous io.
1126 wait_on_page_locked(page);
1127 if (PageUptodate(page)) {
1129 if (BITS_PER_LONG == 32 && hash64) {
1130 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1131 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1132 *hash = *hash >> 32;
1134 *start = le64_to_cpu(dp->ldp_hash_start);
1135 *end = le64_to_cpu(dp->ldp_hash_end);
1137 if (unlikely(*start == 1 && *hash == 0))
1140 LASSERTF(*start <= *hash, "start = %#llx"
1141 ",end = %#llx,hash = %#llx\n",
1142 *start, *end, *hash);
1143 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1144 " hash %#llx\n", offset, *start, *end, *hash);
1147 mdc_release_page(page, 0);
1149 } else if (*end != *start && *hash == *end) {
1151 * upon hash collision, remove this page,
1152 * otherwise put page reference, and
1153 * mdc_read_page_remote() will issue RPC to
1154 * fetch the page we want.
1157 mdc_release_page(page,
1158 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1163 page = ERR_PTR(-EIO);
1166 xa_unlock_irq(&mapping->i_pages);
1173 * Adjust a set of pages, each page containing an array of lu_dirpages,
1174 * so that each page can be used as a single logical lu_dirpage.
1176 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1177 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1178 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1179 * value is used as a cookie to request the next lu_dirpage in a
1180 * directory listing that spans multiple pages (two in this example):
1183 * .|--------v------- -----.
1184 * |s|e|f|p|ent|ent| ... |ent|
1185 * '--|-------------- -----' Each PAGE contains a single
1186 * '------. lu_dirpage.
1187 * .---------v------- -----.
1188 * |s|e|f|p|ent| 0 | ... | 0 |
1189 * '----------------- -----'
1191 * However, on hosts where the native VM page size (PAGE_SIZE) is
1192 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1193 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1194 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1195 * after it in the same PAGE (arrows simplified for brevity, but
1196 * in general e0==s1, e1==s2, etc.):
1198 * .-------------------- -----.
1199 * |s0|e0|f0|p|ent|ent| ... |ent|
1200 * |---v---------------- -----|
1201 * |s1|e1|f1|p|ent|ent| ... |ent|
1202 * |---v---------------- -----| Here, each PAGE contains
1203 * ... multiple lu_dirpages.
1204 * |---v---------------- -----|
1205 * |s'|e'|f'|p|ent|ent| ... |ent|
1206 * '---|---------------- -----'
1208 * .----------------------------.
1211 * This structure is transformed into a single logical lu_dirpage as follows:
1213 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1214 * labeled 'next PAGE'.
1216 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1217 * a hash collision with the next page exists.
1219 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1220 * to the first entry of the next lu_dirpage.
1222 #if PAGE_SIZE > LU_PAGE_SIZE
1223 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1227 for (i = 0; i < cfs_pgs; i++) {
1228 struct lu_dirpage *dp = kmap(pages[i]);
1229 struct lu_dirpage *first = dp;
1230 struct lu_dirent *end_dirent = NULL;
1231 struct lu_dirent *ent;
1232 __u64 hash_end = dp->ldp_hash_end;
1233 __u32 flags = dp->ldp_flags;
1235 while (--lu_pgs > 0) {
1236 ent = lu_dirent_start(dp);
1237 for (end_dirent = ent; ent != NULL;
1238 end_dirent = ent, ent = lu_dirent_next(ent));
1240 /* Advance dp to next lu_dirpage. */
1241 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1243 /* Check if we've reached the end of the PAGE. */
1244 if (!((unsigned long)dp & ~PAGE_MASK))
1247 /* Save the hash and flags of this lu_dirpage. */
1248 hash_end = dp->ldp_hash_end;
1249 flags = dp->ldp_flags;
1251 /* Check if lu_dirpage contains no entries. */
1252 if (end_dirent == NULL)
1255 /* Enlarge the end entry lde_reclen from 0 to
1256 * first entry of next lu_dirpage. */
1257 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1258 end_dirent->lde_reclen =
1259 cpu_to_le16((char *)(dp->ldp_entries) -
1260 (char *)end_dirent);
1263 first->ldp_hash_end = hash_end;
1264 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1265 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1269 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1272 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1273 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1275 /* parameters for readdir page */
1276 struct readpage_param {
1277 struct md_op_data *rp_mod;
1280 struct obd_export *rp_exp;
1281 struct md_callback *rp_cb;
1284 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1285 static inline void delete_from_page_cache(struct page *page)
1287 remove_from_page_cache(page);
1293 * Read pages from server.
1295 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1296 * a header lu_dirpage which describes the start/end hash, and whether this
1297 * page is empty (contains no dir entry) or hash collide with next page.
1298 * After client receives reply, several pages will be integrated into dir page
1299 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1300 * lu_dirpage for this integrated page will be adjusted.
1302 static int mdc_read_page_remote(void *data, struct page *page0)
1304 struct readpage_param *rp = data;
1305 struct page **page_pool;
1307 struct lu_dirpage *dp;
1308 struct md_op_data *op_data = rp->rp_mod;
1309 struct ptlrpc_request *req;
1311 struct inode *inode;
1313 int rd_pgs = 0; /* number of pages actually read */
1319 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1320 inode = op_data->op_data;
1321 fid = &op_data->op_fid1;
1322 LASSERT(inode != NULL);
1324 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1325 if (page_pool != NULL) {
1326 page_pool[0] = page0;
1332 for (npages = 1; npages < max_pages; npages++) {
1333 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1337 page_pool[npages] = page;
1340 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1342 /* page0 is special, which was added into page cache early */
1343 delete_from_page_cache(page0);
1347 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1349 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1350 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1352 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1354 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1356 SetPageUptodate(page0);
1360 ptlrpc_req_finished(req);
1361 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1362 for (i = 1; i < npages; i++) {
1363 unsigned long offset;
1367 page = page_pool[i];
1369 if (rc < 0 || i >= rd_pgs) {
1374 SetPageUptodate(page);
1377 hash = le64_to_cpu(dp->ldp_hash_start);
1380 offset = hash_x_index(hash, rp->rp_hash64);
1382 prefetchw(&page->flags);
1383 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1388 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1389 " rc = %d\n", offset, ret);
1393 if (page_pool != &page0)
1394 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1400 * Read dir page from cache first, if it can not find it, read it from
1401 * server and add into the cache.
1403 * \param[in] exp MDC export
1404 * \param[in] op_data client MD stack parameters, transfering parameters
1405 * between different layers on client MD stack.
1406 * \param[in] cb_op callback required for ldlm lock enqueue during
1408 * \param[in] hash_offset the hash offset of the page to be read
1409 * \param[in] ppage the page to be read
1411 * retval = 0 get the page successfully
1412 * errno(<0) get the page failed
1414 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1415 struct md_callback *cb_op, __u64 hash_offset,
1416 struct page **ppage)
1418 struct lookup_intent it = { .it_op = IT_READDIR };
1420 struct inode *dir = op_data->op_data;
1421 struct address_space *mapping;
1422 struct lu_dirpage *dp;
1425 struct lustre_handle lockh;
1426 struct ptlrpc_request *enq_req = NULL;
1427 struct readpage_param rp_param;
1434 LASSERT(dir != NULL);
1435 mapping = dir->i_mapping;
1437 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1438 cb_op->md_blocking_ast, 0);
1439 if (enq_req != NULL)
1440 ptlrpc_req_finished(enq_req);
1443 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1444 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1449 lockh.cookie = it.it_lock_handle;
1450 mdc_set_lock_data(exp, &lockh, dir, NULL);
1452 rp_param.rp_off = hash_offset;
1453 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1454 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1455 rp_param.rp_hash64);
1457 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1458 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1459 rp_param.rp_off, PTR_ERR(page));
1460 GOTO(out_unlock, rc = PTR_ERR(page));
1461 } else if (page != NULL) {
1463 * XXX nikita: not entirely correct handling of a corner case:
1464 * suppose hash chain of entries with hash value HASH crosses
1465 * border between pages P0 and P1. First both P0 and P1 are
1466 * cached, seekdir() is called for some entry from the P0 part
1467 * of the chain. Later P0 goes out of cache. telldir(HASH)
1468 * happens and finds P1, as it starts with matching hash
1469 * value. Remaining entries from P0 part of the chain are
1470 * skipped. (Is that really a bug?)
1472 * Possible solutions: 0. don't cache P1 is such case, handle
1473 * it as an "overflow" page. 1. invalidate all pages at
1474 * once. 2. use HASH|1 as an index for P1.
1476 GOTO(hash_collision, page);
1479 rp_param.rp_exp = exp;
1480 rp_param.rp_mod = op_data;
1481 page = read_cache_page(mapping,
1482 hash_x_index(rp_param.rp_off,
1483 rp_param.rp_hash64),
1484 mdc_read_page_remote, &rp_param);
1486 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1487 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1488 rp_param.rp_off, PTR_ERR(page));
1489 GOTO(out_unlock, rc = PTR_ERR(page));
1492 wait_on_page_locked(page);
1494 if (!PageUptodate(page)) {
1495 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1496 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1497 rp_param.rp_off, -5);
1500 if (!PageChecked(page))
1501 SetPageChecked(page);
1502 if (PageError(page)) {
1503 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1504 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1505 rp_param.rp_off, -5);
1510 dp = page_address(page);
1511 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1512 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1513 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1514 rp_param.rp_off = hash_offset >> 32;
1516 start = le64_to_cpu(dp->ldp_hash_start);
1517 end = le64_to_cpu(dp->ldp_hash_end);
1518 rp_param.rp_off = hash_offset;
1521 LASSERT(start == rp_param.rp_off);
1522 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1523 #if BITS_PER_LONG == 32
1524 CWARN("Real page-wide hash collision at [%llu %llu] with "
1525 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1526 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1530 * Fetch whole overflow chain...
1538 ldlm_lock_decref(&lockh, it.it_lock_mode);
1542 mdc_release_page(page, 1);
1547 static int mdc_statfs(const struct lu_env *env,
1548 struct obd_export *exp, struct obd_statfs *osfs,
1549 time64_t max_age, __u32 flags)
1551 struct obd_device *obd = class_exp2obd(exp);
1552 struct req_format *fmt;
1553 struct ptlrpc_request *req;
1554 struct obd_statfs *msfs;
1555 struct obd_import *imp = NULL;
1560 * Since the request might also come from lprocfs, so we need
1561 * sync this with client_disconnect_export Bug15684
1563 down_read(&obd->u.cli.cl_sem);
1564 if (obd->u.cli.cl_import)
1565 imp = class_import_get(obd->u.cli.cl_import);
1566 up_read(&obd->u.cli.cl_sem);
1570 fmt = &RQF_MDS_STATFS;
1571 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1572 (flags & OBD_STATFS_SUM))
1573 fmt = &RQF_MDS_STATFS_NEW;
1574 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1577 GOTO(output, rc = -ENOMEM);
1579 if ((flags & OBD_STATFS_SUM) &&
1580 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1581 /* request aggregated states */
1582 struct mdt_body *body;
1584 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1586 GOTO(out, rc = -EPROTO);
1587 body->mbo_valid = OBD_MD_FLAGSTATFS;
1590 ptlrpc_request_set_replen(req);
1592 if (flags & OBD_STATFS_NODELAY) {
1593 /* procfs requests not want stay in wait for avoid deadlock */
1594 req->rq_no_resend = 1;
1595 req->rq_no_delay = 1;
1598 rc = ptlrpc_queue_wait(req);
1600 /* check connection error first */
1601 if (imp->imp_connect_error)
1602 rc = imp->imp_connect_error;
1606 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1608 GOTO(out, rc = -EPROTO);
1613 ptlrpc_req_finished(req);
1615 class_import_put(imp);
1619 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1621 __u32 keylen, vallen;
1625 if (gf->gf_pathlen > PATH_MAX)
1626 RETURN(-ENAMETOOLONG);
1627 if (gf->gf_pathlen < 2)
1630 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1631 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1632 sizeof(struct lu_fid));
1633 OBD_ALLOC(key, keylen);
1636 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1637 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1638 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1639 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1640 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1641 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1643 if (!fid_is_sane(&gf->gf_fid))
1644 GOTO(out, rc = -EINVAL);
1646 /* Val is struct getinfo_fid2path result plus path */
1647 vallen = sizeof(*gf) + gf->gf_pathlen;
1649 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1650 if (rc != 0 && rc != -EREMOTE)
1653 if (vallen <= sizeof(*gf))
1654 GOTO(out, rc = -EPROTO);
1655 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1656 GOTO(out, rc = -EOVERFLOW);
1658 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1659 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1660 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1661 /* only log the last 512 characters of the path */
1662 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1665 OBD_FREE(key, keylen);
1669 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1670 struct hsm_progress_kernel *hpk)
1672 struct obd_import *imp = class_exp2cliimp(exp);
1673 struct hsm_progress_kernel *req_hpk;
1674 struct ptlrpc_request *req;
1678 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1679 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1681 GOTO(out, rc = -ENOMEM);
1683 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1685 /* Copy hsm_progress struct */
1686 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1687 if (req_hpk == NULL)
1688 GOTO(out, rc = -EPROTO);
1691 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1693 ptlrpc_request_set_replen(req);
1695 mdc_get_mod_rpc_slot(req, NULL);
1696 rc = ptlrpc_queue_wait(req);
1697 mdc_put_mod_rpc_slot(req, NULL);
1701 ptlrpc_req_finished(req);
1705 * Send hsm_ct_register to MDS
1707 * \param[in] imp import
1708 * \param[in] archive_count if in bitmap format, it is the bitmap,
1709 * else it is the count of archive_ids
1710 * \param[in] archives if in bitmap format, it is NULL,
1711 * else it is archive_id lists
1713 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1716 struct ptlrpc_request *req;
1717 __u32 *archive_array;
1718 size_t archives_size;
1722 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1726 if (archives != NULL)
1727 archives_size = sizeof(*archive_array) * archive_count;
1729 archives_size = sizeof(archive_count);
1731 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1732 RCL_CLIENT, archives_size);
1734 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1736 ptlrpc_request_free(req);
1740 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1742 archive_array = req_capsule_client_get(&req->rq_pill,
1743 &RMF_MDS_HSM_ARCHIVE);
1744 if (archive_array == NULL)
1745 GOTO(out, rc = -EPROTO);
1747 if (archives != NULL)
1748 memcpy(archive_array, archives, archives_size);
1750 *archive_array = archive_count;
1752 ptlrpc_request_set_replen(req);
1754 rc = mdc_queue_wait(req);
1757 ptlrpc_req_finished(req);
1761 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1762 struct md_op_data *op_data)
1764 struct hsm_current_action *hca = op_data->op_data;
1765 struct hsm_current_action *req_hca;
1766 struct ptlrpc_request *req;
1770 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1771 &RQF_MDS_HSM_ACTION);
1775 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1777 ptlrpc_request_free(req);
1781 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1782 op_data->op_suppgids[0], 0);
1784 ptlrpc_request_set_replen(req);
1786 rc = mdc_queue_wait(req);
1790 req_hca = req_capsule_server_get(&req->rq_pill,
1791 &RMF_MDS_HSM_CURRENT_ACTION);
1792 if (req_hca == NULL)
1793 GOTO(out, rc = -EPROTO);
1799 ptlrpc_req_finished(req);
1803 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1805 struct ptlrpc_request *req;
1809 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1811 MDS_HSM_CT_UNREGISTER);
1813 GOTO(out, rc = -ENOMEM);
1815 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1817 ptlrpc_request_set_replen(req);
1819 rc = mdc_queue_wait(req);
1822 ptlrpc_req_finished(req);
1826 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1827 struct md_op_data *op_data)
1829 struct hsm_user_state *hus = op_data->op_data;
1830 struct hsm_user_state *req_hus;
1831 struct ptlrpc_request *req;
1835 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1836 &RQF_MDS_HSM_STATE_GET);
1840 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1842 ptlrpc_request_free(req);
1846 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1847 op_data->op_suppgids[0], 0);
1849 ptlrpc_request_set_replen(req);
1851 rc = mdc_queue_wait(req);
1855 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1856 if (req_hus == NULL)
1857 GOTO(out, rc = -EPROTO);
1863 ptlrpc_req_finished(req);
1867 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1868 struct md_op_data *op_data)
1870 struct hsm_state_set *hss = op_data->op_data;
1871 struct hsm_state_set *req_hss;
1872 struct ptlrpc_request *req;
1876 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1877 &RQF_MDS_HSM_STATE_SET);
1881 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1883 ptlrpc_request_free(req);
1887 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1888 op_data->op_suppgids[0], 0);
1891 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1892 if (req_hss == NULL)
1893 GOTO(out, rc = -EPROTO);
1896 ptlrpc_request_set_replen(req);
1898 mdc_get_mod_rpc_slot(req, NULL);
1899 rc = ptlrpc_queue_wait(req);
1900 mdc_put_mod_rpc_slot(req, NULL);
1904 ptlrpc_req_finished(req);
1908 static int mdc_ioc_hsm_request(struct obd_export *exp,
1909 struct hsm_user_request *hur)
1911 struct obd_import *imp = class_exp2cliimp(exp);
1912 struct ptlrpc_request *req;
1913 struct hsm_request *req_hr;
1914 struct hsm_user_item *req_hui;
1919 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1921 GOTO(out, rc = -ENOMEM);
1923 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1924 hur->hur_request.hr_itemcount
1925 * sizeof(struct hsm_user_item));
1926 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1927 hur->hur_request.hr_data_len);
1929 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1931 ptlrpc_request_free(req);
1935 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1937 /* Copy hsm_request struct */
1938 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1940 GOTO(out, rc = -EPROTO);
1941 *req_hr = hur->hur_request;
1943 /* Copy hsm_user_item structs */
1944 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1945 if (req_hui == NULL)
1946 GOTO(out, rc = -EPROTO);
1947 memcpy(req_hui, hur->hur_user_item,
1948 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1950 /* Copy opaque field */
1951 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1952 if (req_opaque == NULL)
1953 GOTO(out, rc = -EPROTO);
1954 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1956 ptlrpc_request_set_replen(req);
1958 mdc_get_mod_rpc_slot(req, NULL);
1959 rc = ptlrpc_queue_wait(req);
1960 mdc_put_mod_rpc_slot(req, NULL);
1965 ptlrpc_req_finished(req);
1969 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1970 struct lustre_kernelcomm *lk);
1972 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1973 struct obd_quotactl *oqctl)
1975 struct ptlrpc_request *req;
1976 struct obd_quotactl *oqc;
1980 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
1981 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
1986 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1989 ptlrpc_request_set_replen(req);
1990 ptlrpc_at_set_req_timeout(req);
1992 rc = ptlrpc_queue_wait(req);
1994 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
1996 if (req->rq_repmsg &&
1997 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2000 CERROR ("Can't unpack obd_quotactl\n");
2003 ptlrpc_req_finished(req);
2008 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2009 struct md_op_data *op_data)
2011 struct list_head cancels = LIST_HEAD_INIT(cancels);
2012 struct ptlrpc_request *req;
2014 struct mdc_swap_layouts *msl, *payload;
2017 msl = op_data->op_data;
2019 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2020 * first thing it will do is to cancel the 2 layout
2021 * locks held by this client.
2022 * So the client must cancel its layout locks on the 2 fids
2023 * with the request RPC to avoid extra RPC round trips.
2025 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2026 LCK_EX, MDS_INODELOCK_LAYOUT |
2027 MDS_INODELOCK_XATTR);
2028 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2029 LCK_EX, MDS_INODELOCK_LAYOUT |
2030 MDS_INODELOCK_XATTR);
2032 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2033 &RQF_MDS_SWAP_LAYOUTS);
2035 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2039 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2041 ptlrpc_request_free(req);
2045 mdc_swap_layouts_pack(req, op_data);
2047 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2052 ptlrpc_request_set_replen(req);
2054 rc = ptlrpc_queue_wait(req);
2060 ptlrpc_req_finished(req);
2064 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2065 void *karg, void __user *uarg)
2067 struct obd_device *obd = exp->exp_obd;
2068 struct obd_ioctl_data *data = karg;
2069 struct obd_import *imp = obd->u.cli.cl_import;
2073 if (!try_module_get(THIS_MODULE)) {
2074 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2075 module_name(THIS_MODULE));
2079 case OBD_IOC_FID2PATH:
2080 rc = mdc_ioc_fid2path(exp, karg);
2082 case LL_IOC_HSM_CT_START:
2083 rc = mdc_ioc_hsm_ct_start(exp, karg);
2084 /* ignore if it was already registered on this MDS. */
2088 case LL_IOC_HSM_PROGRESS:
2089 rc = mdc_ioc_hsm_progress(exp, karg);
2091 case LL_IOC_HSM_STATE_GET:
2092 rc = mdc_ioc_hsm_state_get(exp, karg);
2094 case LL_IOC_HSM_STATE_SET:
2095 rc = mdc_ioc_hsm_state_set(exp, karg);
2097 case LL_IOC_HSM_ACTION:
2098 rc = mdc_ioc_hsm_current_action(exp, karg);
2100 case LL_IOC_HSM_REQUEST:
2101 rc = mdc_ioc_hsm_request(exp, karg);
2103 case OBD_IOC_CLIENT_RECOVER:
2104 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2108 case IOC_OSC_SET_ACTIVE:
2109 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2111 case OBD_IOC_PING_TARGET:
2112 rc = ptlrpc_obd_ping(obd);
2115 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2116 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2117 * there'd be no LMV layer thus we might be called here. Eventually
2118 * this code should be removed.
2121 case IOC_OBD_STATFS: {
2122 struct obd_statfs stat_buf = {0};
2124 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2125 GOTO(out, rc = -ENODEV);
2128 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2129 min((int)data->ioc_plen2,
2130 (int)sizeof(struct obd_uuid))))
2131 GOTO(out, rc = -EFAULT);
2133 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2134 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2139 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2140 min((int) data->ioc_plen1,
2141 (int) sizeof(stat_buf))))
2142 GOTO(out, rc = -EFAULT);
2146 case OBD_IOC_QUOTACTL: {
2147 struct if_quotactl *qctl = karg;
2148 struct obd_quotactl *oqctl;
2150 OBD_ALLOC_PTR(oqctl);
2152 GOTO(out, rc = -ENOMEM);
2154 QCTL_COPY(oqctl, qctl);
2155 rc = obd_quotactl(exp, oqctl);
2157 QCTL_COPY(qctl, oqctl);
2158 qctl->qc_valid = QC_MDTIDX;
2159 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2162 OBD_FREE_PTR(oqctl);
2165 case LL_IOC_GET_CONNECT_FLAGS:
2166 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2167 sizeof(*exp_connect_flags_ptr(exp))))
2168 GOTO(out, rc = -EFAULT);
2171 case LL_IOC_LOV_SWAP_LAYOUTS:
2172 rc = mdc_ioc_swap_layouts(exp, karg);
2175 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2176 GOTO(out, rc = -ENOTTY);
2179 module_put(THIS_MODULE);
2184 static int mdc_get_info_rpc(struct obd_export *exp,
2185 u32 keylen, void *key,
2186 u32 vallen, void *val)
2188 struct obd_import *imp = class_exp2cliimp(exp);
2189 struct ptlrpc_request *req;
2194 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2198 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2199 RCL_CLIENT, keylen);
2200 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2201 RCL_CLIENT, sizeof(vallen));
2203 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2205 ptlrpc_request_free(req);
2209 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2210 memcpy(tmp, key, keylen);
2211 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2212 memcpy(tmp, &vallen, sizeof(vallen));
2214 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2215 RCL_SERVER, vallen);
2216 ptlrpc_request_set_replen(req);
2218 /* if server failed to resolve FID, and OI scrub not able to fix it, it
2219 * will return -EINPROGRESS, ptlrpc_queue_wait() will keep retrying,
2220 * set request interruptible to avoid deadlock.
2222 if (KEY_IS(KEY_FID2PATH))
2223 req->rq_allow_intr = 1;
2225 rc = ptlrpc_queue_wait(req);
2226 /* -EREMOTE means the get_info result is partial, and it needs to
2227 * continue on another MDT, see fid2path part in lmv_iocontrol */
2228 if (rc == 0 || rc == -EREMOTE) {
2229 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2230 memcpy(val, tmp, vallen);
2231 if (ptlrpc_rep_need_swab(req)) {
2232 if (KEY_IS(KEY_FID2PATH))
2233 lustre_swab_fid2path(val);
2236 ptlrpc_req_finished(req);
2241 static void lustre_swab_hai(struct hsm_action_item *h)
2243 __swab32s(&h->hai_len);
2244 __swab32s(&h->hai_action);
2245 lustre_swab_lu_fid(&h->hai_fid);
2246 lustre_swab_lu_fid(&h->hai_dfid);
2247 __swab64s(&h->hai_cookie);
2248 __swab64s(&h->hai_extent.offset);
2249 __swab64s(&h->hai_extent.length);
2250 __swab64s(&h->hai_gid);
2253 static void lustre_swab_hal(struct hsm_action_list *h)
2255 struct hsm_action_item *hai;
2258 __swab32s(&h->hal_version);
2259 __swab32s(&h->hal_count);
2260 __swab32s(&h->hal_archive_id);
2261 __swab64s(&h->hal_flags);
2263 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2264 lustre_swab_hai(hai);
2267 static void lustre_swab_kuch(struct kuc_hdr *l)
2269 __swab16s(&l->kuc_magic);
2270 /* __u8 l->kuc_transport */
2271 __swab16s(&l->kuc_msgtype);
2272 __swab16s(&l->kuc_msglen);
2275 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2276 struct lustre_kernelcomm *lk)
2278 struct obd_import *imp = class_exp2cliimp(exp);
2281 if (lk->lk_group != KUC_GRP_HSM) {
2282 CERROR("Bad copytool group %d\n", lk->lk_group);
2286 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2287 lk->lk_uid, lk->lk_group, lk->lk_flags);
2289 if (lk->lk_flags & LK_FLG_STOP) {
2290 /* Unregister with the coordinator */
2291 rc = mdc_ioc_hsm_ct_unregister(imp);
2293 __u32 *archives = NULL;
2295 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2296 archives = lk->lk_data;
2298 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2305 * Send a message to any listening copytools
2306 * @param val KUC message (kuc_hdr + hsm_action_list)
2307 * @param len total length of message
2309 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2310 size_t len, void *val)
2312 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2313 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2317 if (len < sizeof(*lh) + sizeof(*hal)) {
2318 CERROR("Short HSM message %zu < %zu\n", len,
2319 sizeof(*lh) + sizeof(*hal));
2322 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2323 lustre_swab_kuch(lh);
2324 lustre_swab_hal(hal);
2325 } else if (lh->kuc_magic != KUC_MAGIC) {
2326 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2330 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2332 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2333 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2335 /* Broadcast to HSM listeners */
2336 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2342 * callback function passed to kuc for re-registering each HSM copytool
2343 * running on MDC, after MDT shutdown/recovery.
2344 * @param data copytool registration data
2345 * @param cb_arg callback argument (obd_import)
2347 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2349 struct obd_import *imp = (struct obd_import *)cb_arg;
2350 struct kkuc_ct_data *kcd = data;
2351 __u32 *archives = NULL;
2355 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2356 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2359 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2360 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2361 "(archive=%#x)\n", imp->imp_obd->obd_name,
2362 kcd->kcd_nr_archives);
2364 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2365 "(archive nr = %u)\n",
2366 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2367 if (kcd->kcd_nr_archives != 0)
2368 archives = kcd->kcd_archives;
2371 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2372 /* ignore error if the copytool is already registered */
2373 return (rc == -EEXIST) ? 0 : rc;
2377 * Re-establish all kuc contexts with MDT
2378 * after MDT shutdown/recovery.
2380 static int mdc_kuc_reregister(struct obd_import *imp)
2382 /* re-register HSM agents */
2383 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2384 mdc_hsm_ct_reregister, imp);
2387 static int mdc_set_info_async(const struct lu_env *env,
2388 struct obd_export *exp,
2389 u32 keylen, void *key,
2390 u32 vallen, void *val,
2391 struct ptlrpc_request_set *set)
2393 struct obd_import *imp = class_exp2cliimp(exp);
2397 if (KEY_IS(KEY_READ_ONLY)) {
2398 if (vallen != sizeof(int))
2401 spin_lock(&imp->imp_lock);
2402 if (*((int *)val)) {
2403 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2404 imp->imp_connect_data.ocd_connect_flags |=
2407 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2408 imp->imp_connect_data.ocd_connect_flags &=
2409 ~OBD_CONNECT_RDONLY;
2411 spin_unlock(&imp->imp_lock);
2413 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2414 keylen, key, vallen, val, set);
2417 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2418 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2419 keylen, key, vallen, val, set);
2422 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2423 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2428 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2429 __u32 *default_easize = val;
2431 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2435 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2439 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2440 __u32 keylen, void *key, __u32 *vallen, void *val)
2444 if (KEY_IS(KEY_MAX_EASIZE)) {
2445 __u32 mdsize, *max_easize;
2447 if (*vallen != sizeof(int))
2449 mdsize = *(__u32 *)val;
2450 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2451 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2453 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2455 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2456 __u32 *default_easize;
2458 if (*vallen != sizeof(int))
2460 default_easize = val;
2461 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2463 } else if (KEY_IS(KEY_CONN_DATA)) {
2464 struct obd_import *imp = class_exp2cliimp(exp);
2465 struct obd_connect_data *data = val;
2467 if (*vallen != sizeof(*data))
2470 *data = imp->imp_connect_data;
2472 } else if (KEY_IS(KEY_TGT_COUNT)) {
2473 *((__u32 *)val) = 1;
2477 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2482 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2483 struct ptlrpc_request **request)
2485 struct ptlrpc_request *req;
2490 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2494 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2496 ptlrpc_request_free(req);
2500 mdc_pack_body(req, fid, 0, 0, -1, 0);
2502 ptlrpc_request_set_replen(req);
2504 rc = ptlrpc_queue_wait(req);
2506 ptlrpc_req_finished(req);
2512 struct mdc_rmfid_args {
2517 int mdc_rmfid_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2520 struct mdc_rmfid_args *aa;
2525 aa = ptlrpc_req_async_args(req);
2527 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2529 LASSERT(size == sizeof(int) * aa->mra_nr);
2530 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2532 LASSERT(aa->mra_rcs);
2533 LASSERT(aa->mra_nr);
2534 memcpy(aa->mra_rcs, rcs, size);
2540 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2541 int *rcs, struct ptlrpc_request_set *set)
2543 struct ptlrpc_request *req;
2544 struct mdc_rmfid_args *aa;
2550 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2554 flen = fa->fa_nr * sizeof(struct lu_fid);
2555 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2557 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2559 req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2560 RCL_SERVER, fa->fa_nr * sizeof(__u32));
2561 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2563 ptlrpc_request_free(req);
2566 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2567 memcpy(tmp, fa->fa_fids, flen);
2569 mdc_pack_body(req, NULL, 0, 0, -1, 0);
2570 b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2571 b->mbo_ctime = ktime_get_real_seconds();
2573 ptlrpc_request_set_replen(req);
2576 aa = ptlrpc_req_async_args(req);
2578 aa->mra_nr = fa->fa_nr;
2579 req->rq_interpret_reply = mdc_rmfid_interpret;
2581 ptlrpc_set_add_req(set, req);
2582 ptlrpc_check_set(NULL, set);
2587 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2588 enum obd_import_event event)
2590 struct client_obd *cli = &obd->u.cli;
2593 LASSERT(imp->imp_obd == obd);
2596 case IMP_EVENT_DISCON:
2597 spin_lock(&cli->cl_loi_list_lock);
2598 cli->cl_avail_grant = 0;
2599 cli->cl_lost_grant = 0;
2600 spin_unlock(&cli->cl_loi_list_lock);
2602 case IMP_EVENT_INACTIVE:
2604 * Flush current sequence to make client obtain new one
2605 * from server in case of disconnect/reconnect.
2607 down_read(&cli->cl_seq_rwsem);
2609 seq_client_flush(cli->cl_seq);
2610 up_read(&cli->cl_seq_rwsem);
2612 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2614 case IMP_EVENT_INVALIDATE: {
2615 struct ldlm_namespace *ns = obd->obd_namespace;
2619 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2621 env = cl_env_get(&refcheck);
2623 /* Reset grants. All pages go to failing rpcs due to
2624 * the invalid import.
2626 osc_io_unplug(env, cli, NULL);
2628 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2629 osc_ldlm_resource_invalidate,
2631 cl_env_put(env, &refcheck);
2632 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2638 case IMP_EVENT_ACTIVE:
2639 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2640 /* redo the kuc registration after reconnecting */
2642 rc = mdc_kuc_reregister(imp);
2644 case IMP_EVENT_OCD: {
2645 struct obd_connect_data *ocd = &imp->imp_connect_data;
2647 if (OCD_HAS_FLAG(ocd, GRANT))
2648 osc_init_grant(cli, ocd);
2650 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2653 case IMP_EVENT_DEACTIVATE:
2654 case IMP_EVENT_ACTIVATE:
2657 CERROR("Unknown import event %x\n", event);
2663 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2664 struct lu_fid *fid, struct md_op_data *op_data)
2666 struct client_obd *cli = &exp->exp_obd->u.cli;
2671 down_read(&cli->cl_seq_rwsem);
2673 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2674 up_read(&cli->cl_seq_rwsem);
2679 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2681 struct client_obd *cli = &exp->exp_obd->u.cli;
2682 return &cli->cl_target_uuid;
2686 * Determine whether the lock can be canceled before replaying it during
2687 * recovery, non zero value will be return if the lock can be canceled,
2688 * or zero returned for not
2690 static int mdc_cancel_weight(struct ldlm_lock *lock)
2692 if (lock->l_resource->lr_type != LDLM_IBITS)
2695 /* FIXME: if we ever get into a situation where there are too many
2696 * opened files with open locks on a single node, then we really
2697 * should replay these open locks to reget it */
2698 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2701 /* Special case for DoM locks, cancel only unused and granted locks */
2702 if (ldlm_has_dom(lock) &&
2703 (lock->l_granted_mode != lock->l_req_mode ||
2704 osc_ldlm_weigh_ast(lock) != 0))
2710 static int mdc_resource_inode_free(struct ldlm_resource *res)
2712 if (res->lr_lvb_inode)
2713 res->lr_lvb_inode = NULL;
2718 static struct ldlm_valblock_ops inode_lvbo = {
2719 .lvbo_free = mdc_resource_inode_free
2722 static int mdc_llog_init(struct obd_device *obd)
2724 struct obd_llog_group *olg = &obd->obd_olg;
2725 struct llog_ctxt *ctxt;
2730 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2735 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2736 llog_initiator_connect(ctxt);
2737 llog_ctxt_put(ctxt);
2742 static void mdc_llog_finish(struct obd_device *obd)
2744 struct llog_ctxt *ctxt;
2748 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2750 llog_cleanup(NULL, ctxt);
2755 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2761 rc = osc_setup_common(obd, cfg);
2765 rc = mdc_tunables_init(obd);
2767 GOTO(err_osc_cleanup, rc);
2769 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2771 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2773 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2775 rc = mdc_llog_init(obd);
2777 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2779 GOTO(err_llog_cleanup, rc);
2782 rc = mdc_changelog_cdev_init(obd);
2784 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2786 GOTO(err_changelog_cleanup, rc);
2791 err_changelog_cleanup:
2792 mdc_llog_finish(obd);
2794 lprocfs_free_md_stats(obd);
2795 ptlrpc_lprocfs_unregister_obd(obd);
2797 osc_cleanup_common(obd);
2801 /* Initialize the default and maximum LOV EA sizes. This allows
2802 * us to make MDS RPCs with large enough reply buffers to hold a default
2803 * sized EA without having to calculate this (via a call into the
2804 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2805 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2806 * a large number of stripes is possible. If a larger reply buffer is
2807 * required it will be reallocated in the ptlrpc layer due to overflow.
2809 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2812 struct obd_device *obd = exp->exp_obd;
2813 struct client_obd *cli = &obd->u.cli;
2816 if (cli->cl_max_mds_easize < easize)
2817 cli->cl_max_mds_easize = easize;
2819 if (cli->cl_default_mds_easize < def_easize)
2820 cli->cl_default_mds_easize = def_easize;
2825 static int mdc_precleanup(struct obd_device *obd)
2829 osc_precleanup_common(obd);
2830 mdc_changelog_cdev_finish(obd);
2832 obd_cleanup_client_import(obd);
2833 ptlrpc_lprocfs_unregister_obd(obd);
2834 lprocfs_free_md_stats(obd);
2835 mdc_llog_finish(obd);
2839 static int mdc_cleanup(struct obd_device *obd)
2841 return osc_cleanup_common(obd);
2844 int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2846 struct lustre_cfg *lcfg = buf;
2847 size_t count = class_modify_config(lcfg, PARAM_MDC,
2848 &obd->obd_kset.kobj);
2850 return count > 0 ? 0 : count;
2853 static struct obd_ops mdc_obd_ops = {
2854 .o_owner = THIS_MODULE,
2855 .o_setup = mdc_setup,
2856 .o_precleanup = mdc_precleanup,
2857 .o_cleanup = mdc_cleanup,
2858 .o_add_conn = client_import_add_conn,
2859 .o_del_conn = client_import_del_conn,
2860 .o_connect = client_connect_import,
2861 .o_reconnect = osc_reconnect,
2862 .o_disconnect = osc_disconnect,
2863 .o_iocontrol = mdc_iocontrol,
2864 .o_set_info_async = mdc_set_info_async,
2865 .o_statfs = mdc_statfs,
2866 .o_fid_init = client_fid_init,
2867 .o_fid_fini = client_fid_fini,
2868 .o_fid_alloc = mdc_fid_alloc,
2869 .o_import_event = mdc_import_event,
2870 .o_get_info = mdc_get_info,
2871 .o_process_config = mdc_process_config,
2872 .o_get_uuid = mdc_get_uuid,
2873 .o_quotactl = mdc_quotactl,
2876 static struct md_ops mdc_md_ops = {
2877 .m_get_root = mdc_get_root,
2878 .m_null_inode = mdc_null_inode,
2879 .m_close = mdc_close,
2880 .m_create = mdc_create,
2881 .m_enqueue = mdc_enqueue,
2882 .m_getattr = mdc_getattr,
2883 .m_getattr_name = mdc_getattr_name,
2884 .m_intent_lock = mdc_intent_lock,
2886 .m_rename = mdc_rename,
2887 .m_setattr = mdc_setattr,
2888 .m_setxattr = mdc_setxattr,
2889 .m_getxattr = mdc_getxattr,
2890 .m_fsync = mdc_fsync,
2891 .m_file_resync = mdc_file_resync,
2892 .m_read_page = mdc_read_page,
2893 .m_unlink = mdc_unlink,
2894 .m_cancel_unused = mdc_cancel_unused,
2895 .m_init_ea_size = mdc_init_ea_size,
2896 .m_set_lock_data = mdc_set_lock_data,
2897 .m_lock_match = mdc_lock_match,
2898 .m_get_lustre_md = mdc_get_lustre_md,
2899 .m_free_lustre_md = mdc_free_lustre_md,
2900 .m_set_open_replay_data = mdc_set_open_replay_data,
2901 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2902 .m_intent_getattr_async = mdc_intent_getattr_async,
2903 .m_revalidate_lock = mdc_revalidate_lock,
2904 .m_rmfid = mdc_rmfid,
2907 dev_t mdc_changelog_dev;
2908 struct class *mdc_changelog_class;
2909 static int __init mdc_init(void)
2912 rc = alloc_chrdev_region(&mdc_changelog_dev, 0,
2913 MDC_CHANGELOG_DEV_COUNT,
2914 MDC_CHANGELOG_DEV_NAME);
2918 mdc_changelog_class = class_create(THIS_MODULE, MDC_CHANGELOG_DEV_NAME);
2919 if (IS_ERR(mdc_changelog_class)) {
2920 rc = PTR_ERR(mdc_changelog_class);
2924 rc = class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2925 LUSTRE_MDC_NAME, &mdc_device_type);
2932 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
2936 static void __exit mdc_exit(void)
2938 class_destroy(mdc_changelog_class);
2939 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
2940 class_unregister_type(LUSTRE_MDC_NAME);
2943 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2944 MODULE_DESCRIPTION("Lustre Metadata Client");
2945 MODULE_VERSION(LUSTRE_VERSION_STRING);
2946 MODULE_LICENSE("GPL");
2948 module_init(mdc_init);
2949 module_exit(mdc_exit);