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 <lustre_swab.h>
57 #include <obd_class.h>
58 #include <lustre_osc.h>
60 #include "mdc_internal.h"
62 #define REQUEST_MINOR 244
64 static int mdc_cleanup(struct obd_device *obd);
66 static inline int mdc_queue_wait(struct ptlrpc_request *req)
68 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
71 /* obd_get_request_slot() ensures that this client has no more
72 * than cl_max_rpcs_in_flight RPCs simultaneously inf light
74 rc = obd_get_request_slot(cli);
78 rc = ptlrpc_queue_wait(req);
79 obd_put_request_slot(cli);
85 * Send MDS_GET_ROOT RPC to fetch root FID.
87 * If \a fileset is not NULL it should contain a subdirectory off
88 * the ROOT/ directory to be mounted on the client. Return the FID
89 * of the subdirectory to the client to mount onto its mountpoint.
91 * \param[in] imp MDC import
92 * \param[in] fileset fileset name, which could be NULL
93 * \param[out] rootfid root FID of this mountpoint
94 * \param[out] pc root capa will be unpacked and saved in this pointer
96 * \retval 0 on success, negative errno on failure
98 static int mdc_get_root(struct obd_export *exp, const char *fileset,
99 struct lu_fid *rootfid)
101 struct ptlrpc_request *req;
102 struct mdt_body *body;
107 if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
110 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
116 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
117 strlen(fileset) + 1);
118 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
120 ptlrpc_request_free(req);
123 mdc_pack_body(req, NULL, 0, 0, -1, 0);
124 if (fileset != NULL) {
125 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
127 memcpy(name, fileset, strlen(fileset));
129 lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
130 req->rq_send_state = LUSTRE_IMP_FULL;
132 ptlrpc_request_set_replen(req);
134 rc = ptlrpc_queue_wait(req);
138 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
140 GOTO(out, rc = -EPROTO);
142 *rootfid = body->mbo_fid1;
143 CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
144 PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
147 ptlrpc_req_finished(req);
153 * This function now is known to always saying that it will receive 4 buffers
154 * from server. Even for cases when acl_size and md_size is zero, RPC header
155 * will contain 4 fields and RPC itself will contain zero size fields. This is
156 * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
157 * and thus zero, it shrinks it, making zero size. The same story about
158 * md_size. And this is course of problem when client waits for smaller number
159 * of fields. This issue will be fixed later when client gets aware of RPC
162 static int mdc_getattr_common(struct obd_export *exp,
163 struct ptlrpc_request *req)
165 struct req_capsule *pill = &req->rq_pill;
166 struct mdt_body *body;
171 /* Request message already built. */
172 rc = ptlrpc_queue_wait(req);
176 /* sanity check for the reply */
177 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
181 CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
183 mdc_update_max_ea_from_body(exp, body);
184 if (body->mbo_eadatasize != 0) {
185 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
186 body->mbo_eadatasize);
194 static void mdc_reset_acl_req(struct ptlrpc_request *req)
196 spin_lock(&req->rq_early_free_lock);
197 sptlrpc_cli_free_repbuf(req);
198 req->rq_repbuf = NULL;
199 req->rq_repbuf_len = 0;
200 req->rq_repdata = NULL;
201 req->rq_reqdata_len = 0;
202 spin_unlock(&req->rq_early_free_lock);
205 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
206 struct ptlrpc_request **request)
208 struct ptlrpc_request *req;
209 struct obd_import *imp = class_exp2cliimp(exp);
210 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
214 /* Single MDS without an LMV case */
215 if (op_data->op_flags & MF_GET_MDT_IDX) {
221 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
225 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
227 ptlrpc_request_free(req);
232 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
233 op_data->op_mode, -1, 0);
234 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
235 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
237 ptlrpc_request_set_replen(req);
239 rc = mdc_getattr_common(exp, req);
242 acl_bufsize = MIN(imp->imp_connect_data.ocd_max_easize,
244 mdc_reset_acl_req(req);
248 ptlrpc_req_finished(req);
256 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
257 struct ptlrpc_request **request)
259 struct ptlrpc_request *req;
260 struct obd_import *imp = class_exp2cliimp(exp);
261 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
266 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
270 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
271 op_data->op_namelen + 1);
273 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
275 ptlrpc_request_free(req);
279 if (op_data->op_name) {
280 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
281 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
282 op_data->op_namelen);
283 memcpy(name, op_data->op_name, op_data->op_namelen);
287 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
288 op_data->op_mode, op_data->op_suppgids[0], 0);
289 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
291 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
292 ptlrpc_request_set_replen(req);
294 rc = mdc_getattr_common(exp, req);
297 acl_bufsize = MIN(imp->imp_connect_data.ocd_max_easize,
299 mdc_reset_acl_req(req);
303 ptlrpc_req_finished(req);
311 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
312 const struct lu_fid *fid, int opcode, u64 valid,
313 const char *xattr_name, const char *input,
314 int input_size, int output_size, int flags,
315 __u32 suppgid, struct ptlrpc_request **request)
317 struct ptlrpc_request *req;
318 int xattr_namelen = 0;
324 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
329 xattr_namelen = strlen(xattr_name) + 1;
330 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
335 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
338 /* get SELinux policy info if any */
339 rc = sptlrpc_get_sepol(req);
341 ptlrpc_request_free(req);
344 req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
345 strlen(req->rq_sepol) ?
346 strlen(req->rq_sepol) + 1 : 0);
348 /* Flush local XATTR locks to get rid of a possible cancel RPC */
349 if (opcode == MDS_REINT && fid_is_sane(fid) &&
350 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
351 struct list_head cancels = LIST_HEAD_INIT(cancels);
354 /* Without that packing would fail */
356 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
359 count = mdc_resource_get_unused(exp, fid,
361 MDS_INODELOCK_XATTR);
363 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
365 ptlrpc_request_free(req);
369 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
371 ptlrpc_request_free(req);
376 if (opcode == MDS_REINT) {
377 struct mdt_rec_setxattr *rec;
379 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
380 sizeof(struct mdt_rec_reint));
381 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
382 rec->sx_opcode = REINT_SETXATTR;
383 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
384 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
385 rec->sx_cap = cfs_curproc_cap_pack();
386 rec->sx_suppgid1 = suppgid;
387 rec->sx_suppgid2 = -1;
389 rec->sx_valid = valid | OBD_MD_FLCTIME;
390 rec->sx_time = ktime_get_real_seconds();
391 rec->sx_size = output_size;
392 rec->sx_flags = flags;
394 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
398 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
399 memcpy(tmp, xattr_name, xattr_namelen);
402 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
403 memcpy(tmp, input, input_size);
406 mdc_file_sepol_pack(req);
408 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
409 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
410 RCL_SERVER, output_size);
411 ptlrpc_request_set_replen(req);
414 if (opcode == MDS_REINT)
415 mdc_get_mod_rpc_slot(req, NULL);
417 rc = ptlrpc_queue_wait(req);
419 if (opcode == MDS_REINT)
420 mdc_put_mod_rpc_slot(req, NULL);
423 ptlrpc_req_finished(req);
429 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
430 u64 obd_md_valid, const char *name,
431 const void *value, size_t value_size,
432 unsigned int xattr_flags, u32 suppgid,
433 struct ptlrpc_request **req)
435 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
436 obd_md_valid == OBD_MD_FLXATTRRM);
438 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
439 fid, MDS_REINT, obd_md_valid, name,
440 value, value_size, 0, xattr_flags, suppgid,
444 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
445 u64 obd_md_valid, const char *name, size_t buf_size,
446 struct ptlrpc_request **req)
448 struct mdt_body *body;
451 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
452 obd_md_valid == OBD_MD_FLXATTRLS);
454 /* The below message is checked in sanity-selinux.sh test_20d */
455 CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
456 exp->exp_obd->obd_name, name, PFID(fid));
457 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
458 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
463 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
465 GOTO(out, rc = -EPROTO);
467 /* only detect the xattr size */
469 /* LU-11109: Older MDTs do not distinguish
470 * between nonexistent xattrs and zero length
471 * values in this case. Newer MDTs will return
472 * -ENODATA or set OBD_MD_FLXATTR. */
473 GOTO(out, rc = body->mbo_eadatasize);
476 if (body->mbo_eadatasize == 0) {
477 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
478 * success so that we can distinguish between
479 * zero length value and nonexistent xattr.
481 * If OBD_MD_FLXATTR is not set then we keep
482 * the old behavior and return -ENODATA for
483 * getxattr() when mbo_eadatasize is 0. But
484 * -ENODATA only makes sense for getxattr()
485 * and not for listxattr(). */
486 if (body->mbo_valid & OBD_MD_FLXATTR)
488 else if (obd_md_valid == OBD_MD_FLXATTR)
489 GOTO(out, rc = -ENODATA);
494 GOTO(out, rc = body->mbo_eadatasize);
497 ptlrpc_req_finished(*req);
504 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
505 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
507 struct req_capsule *pill = &req->rq_pill;
508 struct mdt_body *body = md->body;
509 struct posix_acl *acl;
514 if (!body->mbo_aclsize)
517 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
522 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
527 CERROR("convert xattr to acl: %d\n", rc);
531 rc = posix_acl_valid(&init_user_ns, acl);
533 CERROR("validate acl: %d\n", rc);
534 posix_acl_release(acl);
542 #define mdc_unpack_acl(req, md) 0
545 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
546 struct obd_export *dt_exp, struct obd_export *md_exp,
547 struct lustre_md *md)
549 struct req_capsule *pill = &req->rq_pill;
554 memset(md, 0, sizeof(*md));
556 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
557 LASSERT(md->body != NULL);
559 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
560 if (!S_ISREG(md->body->mbo_mode)) {
561 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
562 "regular file, but is not\n");
563 GOTO(out, rc = -EPROTO);
566 if (md->body->mbo_eadatasize == 0) {
567 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
568 "but eadatasize 0\n");
569 GOTO(out, rc = -EPROTO);
572 md->layout.lb_len = md->body->mbo_eadatasize;
573 md->layout.lb_buf = req_capsule_server_sized_get(pill,
576 if (md->layout.lb_buf == NULL)
577 GOTO(out, rc = -EPROTO);
578 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
579 const union lmv_mds_md *lmv;
582 if (!S_ISDIR(md->body->mbo_mode)) {
583 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
584 "directory, but is not\n");
585 GOTO(out, rc = -EPROTO);
588 if (md->body->mbo_valid & OBD_MD_MEA) {
589 lmv_size = md->body->mbo_eadatasize;
591 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
592 "but eadatasize 0\n");
596 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
599 GOTO(out, rc = -EPROTO);
601 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
605 if (rc < (int)sizeof(*md->lmv)) {
606 struct lmv_foreign_md *lfm = md->lfm;
608 /* short (< sizeof(struct lmv_stripe_md))
611 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
613 "lmv size too small: %d < %d\n",
614 rc, (int)sizeof(*md->lmv));
615 GOTO(out, rc = -EPROTO);
620 /* since 2.12.58 intent_getattr fetches default LMV */
621 if (md->body->mbo_valid & OBD_MD_DEFAULT_MEA) {
622 lmv_size = sizeof(struct lmv_user_md);
623 lmv = req_capsule_server_sized_get(pill,
627 GOTO(out, rc = -EPROTO);
629 rc = md_unpackmd(md_exp, &md->default_lmv, lmv,
634 if (rc < (int)sizeof(*md->default_lmv)) {
636 "default lmv size too small: %d < %d\n",
637 rc, (int)sizeof(*md->default_lmv));
638 GOTO(out, rc = -EPROTO);
644 if (md->body->mbo_valid & OBD_MD_FLACL) {
645 /* for ACL, it's possible that FLACL is set but aclsize is zero.
646 * only when aclsize != 0 there's an actual segment for ACL
649 if (md->body->mbo_aclsize) {
650 rc = mdc_unpack_acl(req, md);
653 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
655 md->posix_acl = NULL;
663 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
664 posix_acl_release(md->posix_acl);
670 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
676 void mdc_replay_open(struct ptlrpc_request *req)
678 struct md_open_data *mod = req->rq_cb_data;
679 struct ptlrpc_request *close_req;
680 struct obd_client_handle *och;
681 struct lustre_handle old_open_handle = { };
682 struct mdt_body *body;
686 DEBUG_REQ(D_ERROR, req,
687 "cannot properly replay without open data");
692 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
693 LASSERT(body != NULL);
695 spin_lock(&req->rq_lock);
697 if (och && och->och_open_handle.cookie)
698 req->rq_early_free_repbuf = 1;
700 req->rq_early_free_repbuf = 0;
701 spin_unlock(&req->rq_lock);
703 if (req->rq_early_free_repbuf) {
704 struct lustre_handle *file_open_handle;
706 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
708 file_open_handle = &och->och_open_handle;
709 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
710 file_open_handle->cookie, body->mbo_open_handle.cookie);
711 old_open_handle = *file_open_handle;
712 *file_open_handle = body->mbo_open_handle;
715 close_req = mod->mod_close_req;
717 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
718 struct mdt_ioepoch *epoch;
720 LASSERT(opc == MDS_CLOSE);
721 epoch = req_capsule_client_get(&close_req->rq_pill,
725 if (req->rq_early_free_repbuf)
726 LASSERT(old_open_handle.cookie ==
727 epoch->mio_open_handle.cookie);
729 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
730 epoch->mio_open_handle = body->mbo_open_handle;
735 void mdc_commit_open(struct ptlrpc_request *req)
737 struct md_open_data *mod = req->rq_cb_data;
742 * No need to touch md_open_data::mod_och, it holds a reference on
743 * \var mod and will zero references to each other, \var mod will be
744 * freed after that when md_open_data::mod_och will put the reference.
748 * Do not let open request to disappear as it still may be needed
749 * for close rpc to happen (it may happen on evict only, otherwise
750 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
751 * called), just mark this rpc as committed to distinguish these 2
752 * cases, see mdc_close() for details. The open request reference will
753 * be put along with freeing \var mod.
755 ptlrpc_request_addref(req);
756 spin_lock(&req->rq_lock);
757 req->rq_committed = 1;
758 spin_unlock(&req->rq_lock);
759 req->rq_cb_data = NULL;
763 int mdc_set_open_replay_data(struct obd_export *exp,
764 struct obd_client_handle *och,
765 struct lookup_intent *it)
767 struct md_open_data *mod;
768 struct mdt_rec_create *rec;
769 struct mdt_body *body;
770 struct ptlrpc_request *open_req = it->it_request;
771 struct obd_import *imp = open_req->rq_import;
774 if (!open_req->rq_replay)
777 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
778 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
779 LASSERT(rec != NULL);
780 /* Incoming message in my byte order (it's been swabbed). */
781 /* Outgoing messages always in my byte order. */
782 LASSERT(body != NULL);
784 /* Only if the import is replayable, we set replay_open data */
785 if (och && imp->imp_replayable) {
786 mod = obd_mod_alloc();
788 DEBUG_REQ(D_ERROR, open_req,
789 "cannot allocate md_open_data");
794 * Take a reference on \var mod, to be freed on mdc_close().
795 * It protects \var mod from being freed on eviction (commit
796 * callback is called despite rq_replay flag).
797 * Another reference for \var och.
802 spin_lock(&open_req->rq_lock);
805 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
806 it_disposition(it, DISP_OPEN_STRIPE);
807 mod->mod_open_req = open_req;
808 open_req->rq_cb_data = mod;
809 open_req->rq_commit_cb = mdc_commit_open;
810 open_req->rq_early_free_repbuf = 1;
811 spin_unlock(&open_req->rq_lock);
814 rec->cr_fid2 = body->mbo_fid1;
815 rec->cr_open_handle_old = body->mbo_open_handle;
816 open_req->rq_replay_cb = mdc_replay_open;
817 if (!fid_is_sane(&body->mbo_fid1)) {
818 DEBUG_REQ(D_ERROR, open_req,
819 "saving replay request with insane FID " DFID,
820 PFID(&body->mbo_fid1));
824 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
828 static void mdc_free_open(struct md_open_data *mod)
832 if (mod->mod_is_create == 0 &&
833 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
837 * No reason to asssert here if the open request has
838 * rq_replay == 1. It means that mdc_close failed, and
839 * close request wasn`t sent. It is not fatal to client.
840 * The worst thing is eviction if the client gets open lock
843 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req,
844 "free open request, rq_replay=%d",
845 mod->mod_open_req->rq_replay);
847 ptlrpc_request_committed(mod->mod_open_req, committed);
848 if (mod->mod_close_req)
849 ptlrpc_request_committed(mod->mod_close_req, committed);
852 int mdc_clear_open_replay_data(struct obd_export *exp,
853 struct obd_client_handle *och)
855 struct md_open_data *mod = och->och_mod;
859 * It is possible to not have \var mod in a case of eviction between
860 * lookup and ll_file_open().
865 LASSERT(mod != LP_POISON);
866 LASSERT(mod->mod_open_req != NULL);
868 spin_lock(&mod->mod_open_req->rq_lock);
870 mod->mod_och->och_open_handle.cookie = 0;
871 mod->mod_open_req->rq_early_free_repbuf = 0;
872 spin_unlock(&mod->mod_open_req->rq_lock);
882 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
883 struct md_open_data *mod, struct ptlrpc_request **request)
885 struct obd_device *obd = class_exp2obd(exp);
886 struct ptlrpc_request *req;
887 struct req_format *req_fmt;
888 size_t u32_count = 0;
893 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
894 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
897 if (op_data->op_bias & MDS_CLOSE_INTENT) {
898 req_fmt = &RQF_MDS_CLOSE_INTENT;
899 if (op_data->op_bias & MDS_HSM_RELEASE) {
900 /* allocate a FID for volatile file */
901 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
904 CERROR("%s: "DFID" allocating FID: rc = %d\n",
905 obd->obd_name, PFID(&op_data->op_fid1),
907 /* save the errcode and proceed to close */
911 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
912 size_t count = op_data->op_data_size / sizeof(__u32);
914 if (count > INLINE_RESYNC_ARRAY_SIZE)
918 req_fmt = &RQF_MDS_CLOSE;
922 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
925 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
927 /* Ensure that this close's handle is fixed up during replay. */
928 if (likely(mod != NULL)) {
929 LASSERTF(mod->mod_open_req != NULL &&
930 mod->mod_open_req->rq_type != LI_POISON,
931 "POISONED open %p!\n", mod->mod_open_req);
933 mod->mod_close_req = req;
935 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "matched open");
936 /* We no longer want to preserve this open for replay even
937 * though the open was committed. b=3632, b=3633 */
938 spin_lock(&mod->mod_open_req->rq_lock);
939 mod->mod_open_req->rq_replay = 0;
940 spin_unlock(&mod->mod_open_req->rq_lock);
942 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
946 * TODO: repeat close after errors
948 CWARN("%s: close of FID "DFID" failed, file reference will be "
949 "dropped when this client unmounts or is evicted\n",
950 obd->obd_name, PFID(&op_data->op_fid1));
951 GOTO(out, rc = -ENOMEM);
955 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
956 u32_count * sizeof(__u32));
958 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
960 ptlrpc_request_free(req);
965 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
966 * portal whose threads are not taking any DLM locks and are therefore
967 * always progressing */
968 req->rq_request_portal = MDS_READPAGE_PORTAL;
969 ptlrpc_at_set_req_timeout(req);
971 if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
972 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
973 OP_XVALID_LAZYBLOCKS);
975 mdc_close_pack(req, op_data);
977 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
978 obd->u.cli.cl_default_mds_easize);
980 ptlrpc_request_set_replen(req);
982 mdc_get_mod_rpc_slot(req, NULL);
983 rc = ptlrpc_queue_wait(req);
984 mdc_put_mod_rpc_slot(req, NULL);
986 if (req->rq_repmsg == NULL) {
987 CDEBUG(D_RPCTRACE, "request %p failed to send: rc = %d\n", req,
990 rc = req->rq_status ?: -EIO;
991 } else if (rc == 0 || rc == -EAGAIN) {
992 struct mdt_body *body;
994 rc = lustre_msg_get_status(req->rq_repmsg);
995 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
996 DEBUG_REQ(D_ERROR, req,
997 "type = PTL_RPC_MSG_ERR: rc = %d", rc);
1001 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1004 } else if (rc == -ESTALE) {
1006 * it can be allowed error after 3633 if open was committed and
1007 * server failed before close was sent. Let's check if mod
1008 * exists and return no error in that case
1011 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
1012 LASSERT(mod->mod_open_req != NULL);
1013 if (mod->mod_open_req->rq_committed)
1021 mod->mod_close_req = NULL;
1022 /* Since now, mod is accessed through open_req only,
1023 * thus close req does not keep a reference on mod anymore. */
1028 RETURN(rc < 0 ? rc : saved_rc);
1031 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1032 u64 offset, struct page **pages, int npages,
1033 struct ptlrpc_request **request)
1035 struct ptlrpc_request *req;
1036 struct ptlrpc_bulk_desc *desc;
1038 wait_queue_head_t waitq;
1040 struct l_wait_info lwi;
1045 init_waitqueue_head(&waitq);
1048 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1052 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1054 ptlrpc_request_free(req);
1058 req->rq_request_portal = MDS_READPAGE_PORTAL;
1059 ptlrpc_at_set_req_timeout(req);
1061 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1062 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1064 &ptlrpc_bulk_kiov_pin_ops);
1066 ptlrpc_req_finished(req);
1070 /* NB req now owns desc and will free it when it gets freed */
1071 for (i = 0; i < npages; i++)
1072 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1075 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1077 ptlrpc_request_set_replen(req);
1078 rc = ptlrpc_queue_wait(req);
1080 ptlrpc_req_finished(req);
1081 if (rc != -ETIMEDOUT)
1085 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1086 CERROR("%s: too many resend retries: rc = %d\n",
1087 exp->exp_obd->obd_name, -EIO);
1090 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1092 l_wait_event(waitq, 0, &lwi);
1097 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1098 req->rq_bulk->bd_nob_transferred);
1100 ptlrpc_req_finished(req);
1104 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1105 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1106 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1107 PAGE_SIZE * npages);
1108 ptlrpc_req_finished(req);
1116 static void mdc_release_page(struct page *page, int remove)
1120 if (likely(page->mapping != NULL))
1121 delete_from_page_cache(page);
1127 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1128 __u64 *start, __u64 *end, int hash64)
1131 * Complement of hash is used as an index so that
1132 * radix_tree_gang_lookup() can be used to find a page with starting
1133 * hash _smaller_ than one we are looking for.
1135 unsigned long offset = hash_x_index(*hash, hash64);
1139 xa_lock_irq(&mapping->i_pages);
1140 found = radix_tree_gang_lookup(&mapping->page_tree,
1141 (void **)&page, offset, 1);
1142 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1143 struct lu_dirpage *dp;
1146 xa_unlock_irq(&mapping->i_pages);
1148 * In contrast to find_lock_page() we are sure that directory
1149 * page cannot be truncated (while DLM lock is held) and,
1150 * hence, can avoid restart.
1152 * In fact, page cannot be locked here at all, because
1153 * mdc_read_page_remote does synchronous io.
1155 wait_on_page_locked(page);
1156 if (PageUptodate(page)) {
1158 if (BITS_PER_LONG == 32 && hash64) {
1159 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1160 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1161 *hash = *hash >> 32;
1163 *start = le64_to_cpu(dp->ldp_hash_start);
1164 *end = le64_to_cpu(dp->ldp_hash_end);
1166 if (unlikely(*start == 1 && *hash == 0))
1169 LASSERTF(*start <= *hash, "start = %#llx"
1170 ",end = %#llx,hash = %#llx\n",
1171 *start, *end, *hash);
1172 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1173 " hash %#llx\n", offset, *start, *end, *hash);
1176 mdc_release_page(page, 0);
1178 } else if (*end != *start && *hash == *end) {
1180 * upon hash collision, remove this page,
1181 * otherwise put page reference, and
1182 * mdc_read_page_remote() will issue RPC to
1183 * fetch the page we want.
1186 mdc_release_page(page,
1187 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1192 page = ERR_PTR(-EIO);
1195 xa_unlock_irq(&mapping->i_pages);
1202 * Adjust a set of pages, each page containing an array of lu_dirpages,
1203 * so that each page can be used as a single logical lu_dirpage.
1205 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1206 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1207 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1208 * value is used as a cookie to request the next lu_dirpage in a
1209 * directory listing that spans multiple pages (two in this example):
1212 * .|--------v------- -----.
1213 * |s|e|f|p|ent|ent| ... |ent|
1214 * '--|-------------- -----' Each PAGE contains a single
1215 * '------. lu_dirpage.
1216 * .---------v------- -----.
1217 * |s|e|f|p|ent| 0 | ... | 0 |
1218 * '----------------- -----'
1220 * However, on hosts where the native VM page size (PAGE_SIZE) is
1221 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1222 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1223 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1224 * after it in the same PAGE (arrows simplified for brevity, but
1225 * in general e0==s1, e1==s2, etc.):
1227 * .-------------------- -----.
1228 * |s0|e0|f0|p|ent|ent| ... |ent|
1229 * |---v---------------- -----|
1230 * |s1|e1|f1|p|ent|ent| ... |ent|
1231 * |---v---------------- -----| Here, each PAGE contains
1232 * ... multiple lu_dirpages.
1233 * |---v---------------- -----|
1234 * |s'|e'|f'|p|ent|ent| ... |ent|
1235 * '---|---------------- -----'
1237 * .----------------------------.
1240 * This structure is transformed into a single logical lu_dirpage as follows:
1242 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1243 * labeled 'next PAGE'.
1245 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1246 * a hash collision with the next page exists.
1248 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1249 * to the first entry of the next lu_dirpage.
1251 #if PAGE_SIZE > LU_PAGE_SIZE
1252 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1256 for (i = 0; i < cfs_pgs; i++) {
1257 struct lu_dirpage *dp = kmap(pages[i]);
1258 struct lu_dirpage *first = dp;
1259 struct lu_dirent *end_dirent = NULL;
1260 struct lu_dirent *ent;
1261 __u64 hash_end = dp->ldp_hash_end;
1262 __u32 flags = dp->ldp_flags;
1264 while (--lu_pgs > 0) {
1265 ent = lu_dirent_start(dp);
1266 for (end_dirent = ent; ent != NULL;
1267 end_dirent = ent, ent = lu_dirent_next(ent));
1269 /* Advance dp to next lu_dirpage. */
1270 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1272 /* Check if we've reached the end of the PAGE. */
1273 if (!((unsigned long)dp & ~PAGE_MASK))
1276 /* Save the hash and flags of this lu_dirpage. */
1277 hash_end = dp->ldp_hash_end;
1278 flags = dp->ldp_flags;
1280 /* Check if lu_dirpage contains no entries. */
1281 if (end_dirent == NULL)
1284 /* Enlarge the end entry lde_reclen from 0 to
1285 * first entry of next lu_dirpage. */
1286 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1287 end_dirent->lde_reclen =
1288 cpu_to_le16((char *)(dp->ldp_entries) -
1289 (char *)end_dirent);
1292 first->ldp_hash_end = hash_end;
1293 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1294 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1298 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1301 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1302 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1304 /* parameters for readdir page */
1305 struct readpage_param {
1306 struct md_op_data *rp_mod;
1309 struct obd_export *rp_exp;
1310 struct md_callback *rp_cb;
1314 * Read pages from server.
1316 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1317 * a header lu_dirpage which describes the start/end hash, and whether this
1318 * page is empty (contains no dir entry) or hash collide with next page.
1319 * After client receives reply, several pages will be integrated into dir page
1320 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1321 * lu_dirpage for this integrated page will be adjusted.
1323 static int mdc_read_page_remote(void *data, struct page *page0)
1325 struct readpage_param *rp = data;
1326 struct page **page_pool;
1328 struct lu_dirpage *dp;
1329 struct md_op_data *op_data = rp->rp_mod;
1330 struct ptlrpc_request *req;
1332 struct inode *inode;
1334 int rd_pgs = 0; /* number of pages actually read */
1340 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1341 inode = op_data->op_data;
1342 fid = &op_data->op_fid1;
1343 LASSERT(inode != NULL);
1345 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1346 if (page_pool != NULL) {
1347 page_pool[0] = page0;
1353 for (npages = 1; npages < max_pages; npages++) {
1354 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1358 page_pool[npages] = page;
1361 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1363 /* page0 is special, which was added into page cache early */
1364 delete_from_page_cache(page0);
1368 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1370 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1371 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1373 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1375 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1377 SetPageUptodate(page0);
1381 ptlrpc_req_finished(req);
1382 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1383 for (i = 1; i < npages; i++) {
1384 unsigned long offset;
1388 page = page_pool[i];
1390 if (rc < 0 || i >= rd_pgs) {
1395 SetPageUptodate(page);
1398 hash = le64_to_cpu(dp->ldp_hash_start);
1401 offset = hash_x_index(hash, rp->rp_hash64);
1403 prefetchw(&page->flags);
1404 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1409 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1410 " rc = %d\n", offset, ret);
1414 if (page_pool != &page0)
1415 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1421 * Read dir page from cache first, if it can not find it, read it from
1422 * server and add into the cache.
1424 * \param[in] exp MDC export
1425 * \param[in] op_data client MD stack parameters, transfering parameters
1426 * between different layers on client MD stack.
1427 * \param[in] cb_op callback required for ldlm lock enqueue during
1429 * \param[in] hash_offset the hash offset of the page to be read
1430 * \param[in] ppage the page to be read
1432 * retval = 0 get the page successfully
1433 * errno(<0) get the page failed
1435 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1436 struct md_callback *cb_op, __u64 hash_offset,
1437 struct page **ppage)
1439 struct lookup_intent it = { .it_op = IT_READDIR };
1441 struct inode *dir = op_data->op_data;
1442 struct address_space *mapping;
1443 struct lu_dirpage *dp;
1446 struct lustre_handle lockh;
1447 struct ptlrpc_request *enq_req = NULL;
1448 struct readpage_param rp_param;
1455 LASSERT(dir != NULL);
1456 mapping = dir->i_mapping;
1458 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1459 cb_op->md_blocking_ast, 0);
1460 if (enq_req != NULL)
1461 ptlrpc_req_finished(enq_req);
1464 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1465 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1470 lockh.cookie = it.it_lock_handle;
1471 mdc_set_lock_data(exp, &lockh, dir, NULL);
1473 rp_param.rp_off = hash_offset;
1474 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1475 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1476 rp_param.rp_hash64);
1478 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1479 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1480 rp_param.rp_off, PTR_ERR(page));
1481 GOTO(out_unlock, rc = PTR_ERR(page));
1482 } else if (page != NULL) {
1484 * XXX nikita: not entirely correct handling of a corner case:
1485 * suppose hash chain of entries with hash value HASH crosses
1486 * border between pages P0 and P1. First both P0 and P1 are
1487 * cached, seekdir() is called for some entry from the P0 part
1488 * of the chain. Later P0 goes out of cache. telldir(HASH)
1489 * happens and finds P1, as it starts with matching hash
1490 * value. Remaining entries from P0 part of the chain are
1491 * skipped. (Is that really a bug?)
1493 * Possible solutions: 0. don't cache P1 is such case, handle
1494 * it as an "overflow" page. 1. invalidate all pages at
1495 * once. 2. use HASH|1 as an index for P1.
1497 GOTO(hash_collision, page);
1500 rp_param.rp_exp = exp;
1501 rp_param.rp_mod = op_data;
1502 page = read_cache_page(mapping,
1503 hash_x_index(rp_param.rp_off,
1504 rp_param.rp_hash64),
1505 mdc_read_page_remote, &rp_param);
1507 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1508 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1509 rp_param.rp_off, PTR_ERR(page));
1510 GOTO(out_unlock, rc = PTR_ERR(page));
1513 wait_on_page_locked(page);
1515 if (!PageUptodate(page)) {
1516 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1517 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1518 rp_param.rp_off, -5);
1521 if (!PageChecked(page))
1522 SetPageChecked(page);
1523 if (PageError(page)) {
1524 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1525 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1526 rp_param.rp_off, -5);
1531 dp = page_address(page);
1532 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1533 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1534 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1535 rp_param.rp_off = hash_offset >> 32;
1537 start = le64_to_cpu(dp->ldp_hash_start);
1538 end = le64_to_cpu(dp->ldp_hash_end);
1539 rp_param.rp_off = hash_offset;
1542 LASSERT(start == rp_param.rp_off);
1543 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1544 #if BITS_PER_LONG == 32
1545 CWARN("Real page-wide hash collision at [%llu %llu] with "
1546 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1547 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1551 * Fetch whole overflow chain...
1559 ldlm_lock_decref(&lockh, it.it_lock_mode);
1563 mdc_release_page(page, 1);
1568 static int mdc_statfs_interpret(const struct lu_env *env,
1569 struct ptlrpc_request *req, void *args, int rc)
1571 struct obd_info *oinfo = args;
1572 struct obd_statfs *osfs;
1575 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1579 oinfo->oi_osfs = osfs;
1581 CDEBUG(D_CACHE, "blocks=%llu free=%llu avail=%llu "
1582 "objects=%llu free=%llu state=%x\n",
1583 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
1584 osfs->os_files, osfs->os_ffree, osfs->os_state);
1587 oinfo->oi_cb_up(oinfo, rc);
1592 static int mdc_statfs_async(struct obd_export *exp,
1593 struct obd_info *oinfo, time64_t max_age,
1594 struct ptlrpc_request_set *unused)
1596 struct ptlrpc_request *req;
1597 struct obd_info *aa;
1599 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_STATFS,
1600 LUSTRE_MDS_VERSION, MDS_STATFS);
1604 ptlrpc_request_set_replen(req);
1605 req->rq_interpret_reply = mdc_statfs_interpret;
1607 aa = ptlrpc_req_async_args(aa, req);
1610 ptlrpcd_add_req(req);
1615 static int mdc_statfs(const struct lu_env *env,
1616 struct obd_export *exp, struct obd_statfs *osfs,
1617 time64_t max_age, __u32 flags)
1619 struct obd_device *obd = class_exp2obd(exp);
1620 struct req_format *fmt;
1621 struct ptlrpc_request *req;
1622 struct obd_statfs *msfs;
1623 struct obd_import *imp = NULL;
1628 * Since the request might also come from lprocfs, so we need
1629 * sync this with client_disconnect_export Bug15684
1631 down_read(&obd->u.cli.cl_sem);
1632 if (obd->u.cli.cl_import)
1633 imp = class_import_get(obd->u.cli.cl_import);
1634 up_read(&obd->u.cli.cl_sem);
1638 fmt = &RQF_MDS_STATFS;
1639 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1640 (flags & OBD_STATFS_SUM))
1641 fmt = &RQF_MDS_STATFS_NEW;
1642 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1645 GOTO(output, rc = -ENOMEM);
1647 if ((flags & OBD_STATFS_SUM) &&
1648 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1649 /* request aggregated states */
1650 struct mdt_body *body;
1652 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1654 GOTO(out, rc = -EPROTO);
1655 body->mbo_valid = OBD_MD_FLAGSTATFS;
1658 ptlrpc_request_set_replen(req);
1660 if (flags & OBD_STATFS_NODELAY) {
1661 /* procfs requests not want stay in wait for avoid deadlock */
1662 req->rq_no_resend = 1;
1663 req->rq_no_delay = 1;
1666 rc = ptlrpc_queue_wait(req);
1668 /* check connection error first */
1669 if (imp->imp_connect_error)
1670 rc = imp->imp_connect_error;
1674 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1676 GOTO(out, rc = -EPROTO);
1681 ptlrpc_req_finished(req);
1683 class_import_put(imp);
1687 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1689 __u32 keylen, vallen;
1693 if (gf->gf_pathlen > PATH_MAX)
1694 RETURN(-ENAMETOOLONG);
1695 if (gf->gf_pathlen < 2)
1698 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1699 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1700 sizeof(struct lu_fid));
1701 OBD_ALLOC(key, keylen);
1704 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1705 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1706 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1707 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1708 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1709 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1711 if (!fid_is_sane(&gf->gf_fid))
1712 GOTO(out, rc = -EINVAL);
1714 /* Val is struct getinfo_fid2path result plus path */
1715 vallen = sizeof(*gf) + gf->gf_pathlen;
1717 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1718 if (rc != 0 && rc != -EREMOTE)
1721 if (vallen <= sizeof(*gf))
1722 GOTO(out, rc = -EPROTO);
1723 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1724 GOTO(out, rc = -EOVERFLOW);
1726 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1727 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1728 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1729 /* only log the last 512 characters of the path */
1730 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1733 OBD_FREE(key, keylen);
1737 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1738 struct hsm_progress_kernel *hpk)
1740 struct obd_import *imp = class_exp2cliimp(exp);
1741 struct hsm_progress_kernel *req_hpk;
1742 struct ptlrpc_request *req;
1746 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1747 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1749 GOTO(out, rc = -ENOMEM);
1751 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1753 /* Copy hsm_progress struct */
1754 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1755 if (req_hpk == NULL)
1756 GOTO(out, rc = -EPROTO);
1759 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1761 ptlrpc_request_set_replen(req);
1763 mdc_get_mod_rpc_slot(req, NULL);
1764 rc = ptlrpc_queue_wait(req);
1765 mdc_put_mod_rpc_slot(req, NULL);
1769 ptlrpc_req_finished(req);
1773 * Send hsm_ct_register to MDS
1775 * \param[in] imp import
1776 * \param[in] archive_count if in bitmap format, it is the bitmap,
1777 * else it is the count of archive_ids
1778 * \param[in] archives if in bitmap format, it is NULL,
1779 * else it is archive_id lists
1781 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1784 struct ptlrpc_request *req;
1785 __u32 *archive_array;
1786 size_t archives_size;
1790 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1794 if (archives != NULL)
1795 archives_size = sizeof(*archive_array) * archive_count;
1797 archives_size = sizeof(archive_count);
1799 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1800 RCL_CLIENT, archives_size);
1802 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1804 ptlrpc_request_free(req);
1808 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1810 archive_array = req_capsule_client_get(&req->rq_pill,
1811 &RMF_MDS_HSM_ARCHIVE);
1812 if (archive_array == NULL)
1813 GOTO(out, rc = -EPROTO);
1815 if (archives != NULL)
1816 memcpy(archive_array, archives, archives_size);
1818 *archive_array = archive_count;
1820 ptlrpc_request_set_replen(req);
1822 rc = mdc_queue_wait(req);
1825 ptlrpc_req_finished(req);
1829 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1830 struct md_op_data *op_data)
1832 struct hsm_current_action *hca = op_data->op_data;
1833 struct hsm_current_action *req_hca;
1834 struct ptlrpc_request *req;
1838 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1839 &RQF_MDS_HSM_ACTION);
1843 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1845 ptlrpc_request_free(req);
1849 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1850 op_data->op_suppgids[0], 0);
1852 ptlrpc_request_set_replen(req);
1854 rc = mdc_queue_wait(req);
1858 req_hca = req_capsule_server_get(&req->rq_pill,
1859 &RMF_MDS_HSM_CURRENT_ACTION);
1860 if (req_hca == NULL)
1861 GOTO(out, rc = -EPROTO);
1867 ptlrpc_req_finished(req);
1871 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1873 struct ptlrpc_request *req;
1877 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1879 MDS_HSM_CT_UNREGISTER);
1881 GOTO(out, rc = -ENOMEM);
1883 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1885 ptlrpc_request_set_replen(req);
1887 rc = mdc_queue_wait(req);
1890 ptlrpc_req_finished(req);
1894 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1895 struct md_op_data *op_data)
1897 struct hsm_user_state *hus = op_data->op_data;
1898 struct hsm_user_state *req_hus;
1899 struct ptlrpc_request *req;
1903 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1904 &RQF_MDS_HSM_STATE_GET);
1908 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1910 ptlrpc_request_free(req);
1914 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1915 op_data->op_suppgids[0], 0);
1917 ptlrpc_request_set_replen(req);
1919 rc = mdc_queue_wait(req);
1923 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1924 if (req_hus == NULL)
1925 GOTO(out, rc = -EPROTO);
1931 ptlrpc_req_finished(req);
1935 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1936 struct md_op_data *op_data)
1938 struct hsm_state_set *hss = op_data->op_data;
1939 struct hsm_state_set *req_hss;
1940 struct ptlrpc_request *req;
1944 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1945 &RQF_MDS_HSM_STATE_SET);
1949 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1951 ptlrpc_request_free(req);
1955 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1956 op_data->op_suppgids[0], 0);
1959 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1960 if (req_hss == NULL)
1961 GOTO(out, rc = -EPROTO);
1964 ptlrpc_request_set_replen(req);
1966 mdc_get_mod_rpc_slot(req, NULL);
1967 rc = ptlrpc_queue_wait(req);
1968 mdc_put_mod_rpc_slot(req, NULL);
1972 ptlrpc_req_finished(req);
1976 static int mdc_ioc_hsm_request(struct obd_export *exp,
1977 struct hsm_user_request *hur)
1979 struct obd_import *imp = class_exp2cliimp(exp);
1980 struct ptlrpc_request *req;
1981 struct hsm_request *req_hr;
1982 struct hsm_user_item *req_hui;
1987 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1989 GOTO(out, rc = -ENOMEM);
1991 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1992 hur->hur_request.hr_itemcount
1993 * sizeof(struct hsm_user_item));
1994 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1995 hur->hur_request.hr_data_len);
1997 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1999 ptlrpc_request_free(req);
2003 mdc_pack_body(req, NULL, 0, 0, -1, 0);
2005 /* Copy hsm_request struct */
2006 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
2008 GOTO(out, rc = -EPROTO);
2009 *req_hr = hur->hur_request;
2011 /* Copy hsm_user_item structs */
2012 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
2013 if (req_hui == NULL)
2014 GOTO(out, rc = -EPROTO);
2015 memcpy(req_hui, hur->hur_user_item,
2016 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
2018 /* Copy opaque field */
2019 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
2020 if (req_opaque == NULL)
2021 GOTO(out, rc = -EPROTO);
2022 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
2024 ptlrpc_request_set_replen(req);
2026 mdc_get_mod_rpc_slot(req, NULL);
2027 rc = ptlrpc_queue_wait(req);
2028 mdc_put_mod_rpc_slot(req, NULL);
2033 ptlrpc_req_finished(req);
2037 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2038 struct lustre_kernelcomm *lk);
2040 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2041 struct obd_quotactl *oqctl)
2043 struct ptlrpc_request *req;
2044 struct obd_quotactl *oqc;
2048 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
2049 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
2054 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2057 ptlrpc_request_set_replen(req);
2058 ptlrpc_at_set_req_timeout(req);
2060 rc = ptlrpc_queue_wait(req);
2062 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
2064 if (req->rq_repmsg &&
2065 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2068 CERROR ("Can't unpack obd_quotactl\n");
2071 ptlrpc_req_finished(req);
2076 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2077 struct md_op_data *op_data)
2079 struct list_head cancels = LIST_HEAD_INIT(cancels);
2080 struct ptlrpc_request *req;
2082 struct mdc_swap_layouts *msl, *payload;
2085 msl = op_data->op_data;
2087 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2088 * first thing it will do is to cancel the 2 layout
2089 * locks held by this client.
2090 * So the client must cancel its layout locks on the 2 fids
2091 * with the request RPC to avoid extra RPC round trips.
2093 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2094 LCK_EX, MDS_INODELOCK_LAYOUT |
2095 MDS_INODELOCK_XATTR);
2096 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2097 LCK_EX, MDS_INODELOCK_LAYOUT |
2098 MDS_INODELOCK_XATTR);
2100 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2101 &RQF_MDS_SWAP_LAYOUTS);
2103 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2107 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2109 ptlrpc_request_free(req);
2113 mdc_swap_layouts_pack(req, op_data);
2115 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2120 ptlrpc_request_set_replen(req);
2122 rc = ptlrpc_queue_wait(req);
2128 ptlrpc_req_finished(req);
2132 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2133 void *karg, void __user *uarg)
2135 struct obd_device *obd = exp->exp_obd;
2136 struct obd_ioctl_data *data = karg;
2137 struct obd_import *imp = obd->u.cli.cl_import;
2141 if (!try_module_get(THIS_MODULE)) {
2142 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2143 module_name(THIS_MODULE));
2147 case OBD_IOC_FID2PATH:
2148 rc = mdc_ioc_fid2path(exp, karg);
2150 case LL_IOC_HSM_CT_START:
2151 rc = mdc_ioc_hsm_ct_start(exp, karg);
2152 /* ignore if it was already registered on this MDS. */
2156 case LL_IOC_HSM_PROGRESS:
2157 rc = mdc_ioc_hsm_progress(exp, karg);
2159 case LL_IOC_HSM_STATE_GET:
2160 rc = mdc_ioc_hsm_state_get(exp, karg);
2162 case LL_IOC_HSM_STATE_SET:
2163 rc = mdc_ioc_hsm_state_set(exp, karg);
2165 case LL_IOC_HSM_ACTION:
2166 rc = mdc_ioc_hsm_current_action(exp, karg);
2168 case LL_IOC_HSM_REQUEST:
2169 rc = mdc_ioc_hsm_request(exp, karg);
2171 case OBD_IOC_CLIENT_RECOVER:
2172 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2176 case IOC_OSC_SET_ACTIVE:
2177 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2180 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2181 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2182 * there'd be no LMV layer thus we might be called here. Eventually
2183 * this code should be removed.
2186 case IOC_OBD_STATFS: {
2187 struct obd_statfs stat_buf = {0};
2189 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2190 GOTO(out, rc = -ENODEV);
2193 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2194 min((int)data->ioc_plen2,
2195 (int)sizeof(struct obd_uuid))))
2196 GOTO(out, rc = -EFAULT);
2198 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2199 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2204 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2205 min((int) data->ioc_plen1,
2206 (int) sizeof(stat_buf))))
2207 GOTO(out, rc = -EFAULT);
2211 case OBD_IOC_QUOTACTL: {
2212 struct if_quotactl *qctl = karg;
2213 struct obd_quotactl *oqctl;
2215 OBD_ALLOC_PTR(oqctl);
2217 GOTO(out, rc = -ENOMEM);
2219 QCTL_COPY(oqctl, qctl);
2220 rc = obd_quotactl(exp, oqctl);
2222 QCTL_COPY(qctl, oqctl);
2223 qctl->qc_valid = QC_MDTIDX;
2224 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2227 OBD_FREE_PTR(oqctl);
2230 case LL_IOC_GET_CONNECT_FLAGS:
2231 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2232 sizeof(*exp_connect_flags_ptr(exp))))
2233 GOTO(out, rc = -EFAULT);
2236 case LL_IOC_LOV_SWAP_LAYOUTS:
2237 rc = mdc_ioc_swap_layouts(exp, karg);
2240 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2241 GOTO(out, rc = -ENOTTY);
2244 module_put(THIS_MODULE);
2249 static int mdc_get_info_rpc(struct obd_export *exp,
2250 u32 keylen, void *key,
2251 u32 vallen, void *val)
2253 struct obd_import *imp = class_exp2cliimp(exp);
2254 struct ptlrpc_request *req;
2259 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2263 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2264 RCL_CLIENT, keylen);
2265 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2266 RCL_CLIENT, sizeof(vallen));
2268 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2270 ptlrpc_request_free(req);
2274 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2275 memcpy(tmp, key, keylen);
2276 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2277 memcpy(tmp, &vallen, sizeof(vallen));
2279 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2280 RCL_SERVER, vallen);
2281 ptlrpc_request_set_replen(req);
2283 rc = ptlrpc_queue_wait(req);
2284 /* -EREMOTE means the get_info result is partial, and it needs to
2285 * continue on another MDT, see fid2path part in lmv_iocontrol */
2286 if (rc == 0 || rc == -EREMOTE) {
2287 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2288 memcpy(val, tmp, vallen);
2289 if (ptlrpc_rep_need_swab(req)) {
2290 if (KEY_IS(KEY_FID2PATH))
2291 lustre_swab_fid2path(val);
2294 ptlrpc_req_finished(req);
2299 static void lustre_swab_hai(struct hsm_action_item *h)
2301 __swab32s(&h->hai_len);
2302 __swab32s(&h->hai_action);
2303 lustre_swab_lu_fid(&h->hai_fid);
2304 lustre_swab_lu_fid(&h->hai_dfid);
2305 __swab64s(&h->hai_cookie);
2306 __swab64s(&h->hai_extent.offset);
2307 __swab64s(&h->hai_extent.length);
2308 __swab64s(&h->hai_gid);
2311 static void lustre_swab_hal(struct hsm_action_list *h)
2313 struct hsm_action_item *hai;
2316 __swab32s(&h->hal_version);
2317 __swab32s(&h->hal_count);
2318 __swab32s(&h->hal_archive_id);
2319 __swab64s(&h->hal_flags);
2321 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2322 lustre_swab_hai(hai);
2325 static void lustre_swab_kuch(struct kuc_hdr *l)
2327 __swab16s(&l->kuc_magic);
2328 /* __u8 l->kuc_transport */
2329 __swab16s(&l->kuc_msgtype);
2330 __swab16s(&l->kuc_msglen);
2333 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2334 struct lustre_kernelcomm *lk)
2336 struct obd_import *imp = class_exp2cliimp(exp);
2339 if (lk->lk_group != KUC_GRP_HSM) {
2340 CERROR("Bad copytool group %d\n", lk->lk_group);
2344 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2345 lk->lk_uid, lk->lk_group, lk->lk_flags);
2347 if (lk->lk_flags & LK_FLG_STOP) {
2348 /* Unregister with the coordinator */
2349 rc = mdc_ioc_hsm_ct_unregister(imp);
2351 __u32 *archives = NULL;
2353 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2354 archives = lk->lk_data;
2356 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2363 * Send a message to any listening copytools
2364 * @param val KUC message (kuc_hdr + hsm_action_list)
2365 * @param len total length of message
2367 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2368 size_t len, void *val)
2370 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2371 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2375 if (len < sizeof(*lh) + sizeof(*hal)) {
2376 CERROR("Short HSM message %zu < %zu\n", len,
2377 sizeof(*lh) + sizeof(*hal));
2380 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2381 lustre_swab_kuch(lh);
2382 lustre_swab_hal(hal);
2383 } else if (lh->kuc_magic != KUC_MAGIC) {
2384 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2388 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2390 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2391 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2393 /* Broadcast to HSM listeners */
2394 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2400 * callback function passed to kuc for re-registering each HSM copytool
2401 * running on MDC, after MDT shutdown/recovery.
2402 * @param data copytool registration data
2403 * @param cb_arg callback argument (obd_import)
2405 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2407 struct obd_import *imp = (struct obd_import *)cb_arg;
2408 struct kkuc_ct_data *kcd = data;
2409 __u32 *archives = NULL;
2413 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2414 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2417 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2418 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2419 "(archive=%#x)\n", imp->imp_obd->obd_name,
2420 kcd->kcd_nr_archives);
2422 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2423 "(archive nr = %u)\n",
2424 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2425 if (kcd->kcd_nr_archives != 0)
2426 archives = kcd->kcd_archives;
2429 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2430 /* ignore error if the copytool is already registered */
2431 return (rc == -EEXIST) ? 0 : rc;
2435 * Re-establish all kuc contexts with MDT
2436 * after MDT shutdown/recovery.
2438 static int mdc_kuc_reregister(struct obd_import *imp)
2440 /* re-register HSM agents */
2441 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2442 mdc_hsm_ct_reregister, imp);
2445 static int mdc_set_info_async(const struct lu_env *env,
2446 struct obd_export *exp,
2447 u32 keylen, void *key,
2448 u32 vallen, void *val,
2449 struct ptlrpc_request_set *set)
2451 struct obd_import *imp = class_exp2cliimp(exp);
2455 if (KEY_IS(KEY_READ_ONLY)) {
2456 if (vallen != sizeof(int))
2459 spin_lock(&imp->imp_lock);
2460 if (*((int *)val)) {
2461 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2462 imp->imp_connect_data.ocd_connect_flags |=
2465 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2466 imp->imp_connect_data.ocd_connect_flags &=
2467 ~OBD_CONNECT_RDONLY;
2469 spin_unlock(&imp->imp_lock);
2471 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2472 keylen, key, vallen, val, set);
2475 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2476 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2477 keylen, key, vallen, val, set);
2480 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2481 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2486 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2487 __u32 *default_easize = val;
2489 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2493 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2497 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2498 __u32 keylen, void *key, __u32 *vallen, void *val)
2502 if (KEY_IS(KEY_MAX_EASIZE)) {
2503 __u32 mdsize, *max_easize;
2505 if (*vallen != sizeof(int))
2507 mdsize = *(__u32 *)val;
2508 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2509 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2511 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2513 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2514 __u32 *default_easize;
2516 if (*vallen != sizeof(int))
2518 default_easize = val;
2519 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2521 } else if (KEY_IS(KEY_CONN_DATA)) {
2522 struct obd_import *imp = class_exp2cliimp(exp);
2523 struct obd_connect_data *data = val;
2525 if (*vallen != sizeof(*data))
2528 *data = imp->imp_connect_data;
2530 } else if (KEY_IS(KEY_TGT_COUNT)) {
2531 *((__u32 *)val) = 1;
2535 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2540 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2541 struct ptlrpc_request **request)
2543 struct ptlrpc_request *req;
2548 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2552 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2554 ptlrpc_request_free(req);
2558 mdc_pack_body(req, fid, 0, 0, -1, 0);
2560 ptlrpc_request_set_replen(req);
2562 rc = ptlrpc_queue_wait(req);
2564 ptlrpc_req_finished(req);
2570 struct mdc_rmfid_args {
2575 int mdc_rmfid_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2578 struct mdc_rmfid_args *aa;
2583 aa = ptlrpc_req_async_args(aa, req);
2585 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2587 LASSERT(size == sizeof(int) * aa->mra_nr);
2588 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2590 LASSERT(aa->mra_rcs);
2591 LASSERT(aa->mra_nr);
2592 memcpy(aa->mra_rcs, rcs, size);
2598 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2599 int *rcs, struct ptlrpc_request_set *set)
2601 struct ptlrpc_request *req;
2602 struct mdc_rmfid_args *aa;
2608 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2612 flen = fa->fa_nr * sizeof(struct lu_fid);
2613 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2615 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2617 req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2618 RCL_SERVER, fa->fa_nr * sizeof(__u32));
2619 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2621 ptlrpc_request_free(req);
2624 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2625 memcpy(tmp, fa->fa_fids, flen);
2627 mdc_pack_body(req, NULL, 0, 0, -1, 0);
2628 b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2629 b->mbo_ctime = ktime_get_real_seconds();
2631 ptlrpc_request_set_replen(req);
2634 aa = ptlrpc_req_async_args(aa, req);
2636 aa->mra_nr = fa->fa_nr;
2637 req->rq_interpret_reply = mdc_rmfid_interpret;
2639 ptlrpc_set_add_req(set, req);
2640 ptlrpc_check_set(NULL, set);
2645 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2646 enum obd_import_event event)
2648 struct client_obd *cli = &obd->u.cli;
2651 LASSERT(imp->imp_obd == obd);
2654 case IMP_EVENT_DISCON:
2655 spin_lock(&cli->cl_loi_list_lock);
2656 cli->cl_avail_grant = 0;
2657 cli->cl_lost_grant = 0;
2658 spin_unlock(&cli->cl_loi_list_lock);
2660 case IMP_EVENT_INACTIVE:
2662 * Flush current sequence to make client obtain new one
2663 * from server in case of disconnect/reconnect.
2665 down_read(&cli->cl_seq_rwsem);
2667 seq_client_flush(cli->cl_seq);
2668 up_read(&cli->cl_seq_rwsem);
2670 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2672 case IMP_EVENT_INVALIDATE: {
2673 struct ldlm_namespace *ns = obd->obd_namespace;
2677 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2679 env = cl_env_get(&refcheck);
2681 /* Reset grants. All pages go to failing rpcs due to
2682 * the invalid import.
2684 osc_io_unplug(env, cli, NULL);
2686 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2687 osc_ldlm_resource_invalidate,
2689 cl_env_put(env, &refcheck);
2690 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2696 case IMP_EVENT_ACTIVE:
2697 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2698 /* redo the kuc registration after reconnecting */
2700 rc = mdc_kuc_reregister(imp);
2702 case IMP_EVENT_OCD: {
2703 struct obd_connect_data *ocd = &imp->imp_connect_data;
2705 if (OCD_HAS_FLAG(ocd, GRANT))
2706 osc_init_grant(cli, ocd);
2708 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2711 case IMP_EVENT_DEACTIVATE:
2712 case IMP_EVENT_ACTIVATE:
2715 CERROR("Unknown import event %x\n", event);
2721 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2722 struct lu_fid *fid, struct md_op_data *op_data)
2724 struct client_obd *cli = &exp->exp_obd->u.cli;
2729 down_read(&cli->cl_seq_rwsem);
2731 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2732 up_read(&cli->cl_seq_rwsem);
2737 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2739 struct client_obd *cli = &exp->exp_obd->u.cli;
2740 return &cli->cl_target_uuid;
2744 * Determine whether the lock can be canceled before replaying it during
2745 * recovery, non zero value will be return if the lock can be canceled,
2746 * or zero returned for not
2748 static int mdc_cancel_weight(struct ldlm_lock *lock)
2750 if (lock->l_resource->lr_type != LDLM_IBITS)
2753 /* FIXME: if we ever get into a situation where there are too many
2754 * opened files with open locks on a single node, then we really
2755 * should replay these open locks to reget it */
2756 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2759 /* Special case for DoM locks, cancel only unused and granted locks */
2760 if (ldlm_has_dom(lock) &&
2761 (lock->l_granted_mode != lock->l_req_mode ||
2762 osc_ldlm_weigh_ast(lock) != 0))
2768 static int mdc_resource_inode_free(struct ldlm_resource *res)
2770 if (res->lr_lvb_inode)
2771 res->lr_lvb_inode = NULL;
2776 static struct ldlm_valblock_ops inode_lvbo = {
2777 .lvbo_free = mdc_resource_inode_free
2780 static int mdc_llog_init(struct obd_device *obd)
2782 struct obd_llog_group *olg = &obd->obd_olg;
2783 struct llog_ctxt *ctxt;
2788 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2793 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2794 llog_initiator_connect(ctxt);
2795 llog_ctxt_put(ctxt);
2800 static void mdc_llog_finish(struct obd_device *obd)
2802 struct llog_ctxt *ctxt;
2806 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2808 llog_cleanup(NULL, ctxt);
2813 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2819 rc = osc_setup_common(obd, cfg);
2823 rc = mdc_tunables_init(obd);
2825 GOTO(err_osc_cleanup, rc);
2827 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2829 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2831 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2833 rc = mdc_llog_init(obd);
2835 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2837 GOTO(err_llog_cleanup, rc);
2840 rc = mdc_changelog_cdev_init(obd);
2842 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2844 GOTO(err_changelog_cleanup, rc);
2849 err_changelog_cleanup:
2850 mdc_llog_finish(obd);
2852 lprocfs_free_md_stats(obd);
2853 ptlrpc_lprocfs_unregister_obd(obd);
2855 osc_cleanup_common(obd);
2859 /* Initialize the default and maximum LOV EA sizes. This allows
2860 * us to make MDS RPCs with large enough reply buffers to hold a default
2861 * sized EA without having to calculate this (via a call into the
2862 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2863 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2864 * a large number of stripes is possible. If a larger reply buffer is
2865 * required it will be reallocated in the ptlrpc layer due to overflow.
2867 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2870 struct obd_device *obd = exp->exp_obd;
2871 struct client_obd *cli = &obd->u.cli;
2874 if (cli->cl_max_mds_easize < easize)
2875 cli->cl_max_mds_easize = easize;
2877 if (cli->cl_default_mds_easize < def_easize)
2878 cli->cl_default_mds_easize = def_easize;
2883 static int mdc_precleanup(struct obd_device *obd)
2887 osc_precleanup_common(obd);
2888 mdc_changelog_cdev_finish(obd);
2890 obd_cleanup_client_import(obd);
2891 ptlrpc_lprocfs_unregister_obd(obd);
2892 lprocfs_free_md_stats(obd);
2893 mdc_llog_finish(obd);
2897 static int mdc_cleanup(struct obd_device *obd)
2899 return osc_cleanup_common(obd);
2902 static struct obd_ops mdc_obd_ops = {
2903 .o_owner = THIS_MODULE,
2904 .o_setup = mdc_setup,
2905 .o_precleanup = mdc_precleanup,
2906 .o_cleanup = mdc_cleanup,
2907 .o_add_conn = client_import_add_conn,
2908 .o_del_conn = client_import_del_conn,
2909 .o_connect = client_connect_import,
2910 .o_reconnect = osc_reconnect,
2911 .o_disconnect = osc_disconnect,
2912 .o_iocontrol = mdc_iocontrol,
2913 .o_set_info_async = mdc_set_info_async,
2914 .o_statfs = mdc_statfs,
2915 .o_statfs_async = mdc_statfs_async,
2916 .o_fid_init = client_fid_init,
2917 .o_fid_fini = client_fid_fini,
2918 .o_fid_alloc = mdc_fid_alloc,
2919 .o_import_event = mdc_import_event,
2920 .o_get_info = mdc_get_info,
2921 .o_get_uuid = mdc_get_uuid,
2922 .o_quotactl = mdc_quotactl,
2925 static struct md_ops mdc_md_ops = {
2926 .m_get_root = mdc_get_root,
2927 .m_null_inode = mdc_null_inode,
2928 .m_close = mdc_close,
2929 .m_create = mdc_create,
2930 .m_enqueue = mdc_enqueue,
2931 .m_getattr = mdc_getattr,
2932 .m_getattr_name = mdc_getattr_name,
2933 .m_intent_lock = mdc_intent_lock,
2935 .m_rename = mdc_rename,
2936 .m_setattr = mdc_setattr,
2937 .m_setxattr = mdc_setxattr,
2938 .m_getxattr = mdc_getxattr,
2939 .m_fsync = mdc_fsync,
2940 .m_file_resync = mdc_file_resync,
2941 .m_read_page = mdc_read_page,
2942 .m_unlink = mdc_unlink,
2943 .m_cancel_unused = mdc_cancel_unused,
2944 .m_init_ea_size = mdc_init_ea_size,
2945 .m_set_lock_data = mdc_set_lock_data,
2946 .m_lock_match = mdc_lock_match,
2947 .m_get_lustre_md = mdc_get_lustre_md,
2948 .m_free_lustre_md = mdc_free_lustre_md,
2949 .m_set_open_replay_data = mdc_set_open_replay_data,
2950 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2951 .m_intent_getattr_async = mdc_intent_getattr_async,
2952 .m_revalidate_lock = mdc_revalidate_lock,
2953 .m_rmfid = mdc_rmfid,
2956 static int __init mdc_init(void)
2958 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2959 LUSTRE_MDC_NAME, &mdc_device_type);
2962 static void __exit mdc_exit(void)
2964 class_unregister_type(LUSTRE_MDC_NAME);
2967 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2968 MODULE_DESCRIPTION("Lustre Metadata Client");
2969 MODULE_VERSION(LUSTRE_VERSION_STRING);
2970 MODULE_LICENSE("GPL");
2972 module_init(mdc_init);
2973 module_exit(mdc_exit);