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 CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
455 exp->exp_obd->obd_name, name, PFID(fid));
456 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
457 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
462 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
464 GOTO(out, rc = -EPROTO);
466 /* only detect the xattr size */
468 /* LU-11109: Older MDTs do not distinguish
469 * between nonexistent xattrs and zero length
470 * values in this case. Newer MDTs will return
471 * -ENODATA or set OBD_MD_FLXATTR. */
472 GOTO(out, rc = body->mbo_eadatasize);
475 if (body->mbo_eadatasize == 0) {
476 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
477 * success so that we can distinguish between
478 * zero length value and nonexistent xattr.
480 * If OBD_MD_FLXATTR is not set then we keep
481 * the old behavior and return -ENODATA for
482 * getxattr() when mbo_eadatasize is 0. But
483 * -ENODATA only makes sense for getxattr()
484 * and not for listxattr(). */
485 if (body->mbo_valid & OBD_MD_FLXATTR)
487 else if (obd_md_valid == OBD_MD_FLXATTR)
488 GOTO(out, rc = -ENODATA);
493 GOTO(out, rc = body->mbo_eadatasize);
496 ptlrpc_req_finished(*req);
503 #ifdef CONFIG_FS_POSIX_ACL
504 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
506 struct req_capsule *pill = &req->rq_pill;
507 struct mdt_body *body = md->body;
508 struct posix_acl *acl;
513 if (!body->mbo_aclsize)
516 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
521 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
526 CERROR("convert xattr to acl: %d\n", rc);
530 rc = posix_acl_valid(&init_user_ns, acl);
532 CERROR("validate acl: %d\n", rc);
533 posix_acl_release(acl);
541 #define mdc_unpack_acl(req, md) 0
544 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
545 struct obd_export *dt_exp, struct obd_export *md_exp,
546 struct lustre_md *md)
548 struct req_capsule *pill = &req->rq_pill;
553 memset(md, 0, sizeof(*md));
555 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
556 LASSERT(md->body != NULL);
558 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
559 if (!S_ISREG(md->body->mbo_mode)) {
560 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
561 "regular file, but is not\n");
562 GOTO(out, rc = -EPROTO);
565 if (md->body->mbo_eadatasize == 0) {
566 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
567 "but eadatasize 0\n");
568 GOTO(out, rc = -EPROTO);
571 md->layout.lb_len = md->body->mbo_eadatasize;
572 md->layout.lb_buf = req_capsule_server_sized_get(pill,
575 if (md->layout.lb_buf == NULL)
576 GOTO(out, rc = -EPROTO);
577 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
578 const union lmv_mds_md *lmv;
581 if (!S_ISDIR(md->body->mbo_mode)) {
582 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
583 "directory, but is not\n");
584 GOTO(out, rc = -EPROTO);
587 if (md->body->mbo_valid & OBD_MD_MEA) {
588 lmv_size = md->body->mbo_eadatasize;
590 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
591 "but eadatasize 0\n");
595 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
598 GOTO(out, rc = -EPROTO);
600 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
604 if (rc < (int)sizeof(*md->lmv)) {
605 struct lmv_foreign_md *lfm = md->lfm;
607 /* short (< sizeof(struct lmv_stripe_md))
610 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
612 "lmv size too small: %d < %d\n",
613 rc, (int)sizeof(*md->lmv));
614 GOTO(out, rc = -EPROTO);
619 /* since 2.12.58 intent_getattr fetches default LMV */
620 if (md->body->mbo_valid & OBD_MD_DEFAULT_MEA) {
621 lmv_size = sizeof(struct lmv_user_md);
622 lmv = req_capsule_server_sized_get(pill,
626 GOTO(out, rc = -EPROTO);
628 rc = md_unpackmd(md_exp, &md->default_lmv, lmv,
633 if (rc < (int)sizeof(*md->default_lmv)) {
635 "default lmv size too small: %d < %d\n",
636 rc, (int)sizeof(*md->default_lmv));
637 GOTO(out, rc = -EPROTO);
643 if (md->body->mbo_valid & OBD_MD_FLACL) {
644 /* for ACL, it's possible that FLACL is set but aclsize is zero.
645 * only when aclsize != 0 there's an actual segment for ACL
648 if (md->body->mbo_aclsize) {
649 rc = mdc_unpack_acl(req, md);
652 #ifdef CONFIG_FS_POSIX_ACL
654 md->posix_acl = NULL;
662 #ifdef CONFIG_FS_POSIX_ACL
663 posix_acl_release(md->posix_acl);
669 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
675 void mdc_replay_open(struct ptlrpc_request *req)
677 struct md_open_data *mod = req->rq_cb_data;
678 struct ptlrpc_request *close_req;
679 struct obd_client_handle *och;
680 struct lustre_handle old_open_handle = { };
681 struct mdt_body *body;
685 DEBUG_REQ(D_ERROR, req,
686 "Can't properly replay without open data.");
691 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
692 LASSERT(body != NULL);
694 spin_lock(&req->rq_lock);
696 if (och && och->och_open_handle.cookie)
697 req->rq_early_free_repbuf = 1;
699 req->rq_early_free_repbuf = 0;
700 spin_unlock(&req->rq_lock);
702 if (req->rq_early_free_repbuf) {
703 struct lustre_handle *file_open_handle;
705 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
707 file_open_handle = &och->och_open_handle;
708 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
709 file_open_handle->cookie, body->mbo_open_handle.cookie);
710 old_open_handle = *file_open_handle;
711 *file_open_handle = body->mbo_open_handle;
714 close_req = mod->mod_close_req;
716 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
717 struct mdt_ioepoch *epoch;
719 LASSERT(opc == MDS_CLOSE);
720 epoch = req_capsule_client_get(&close_req->rq_pill,
724 if (req->rq_early_free_repbuf)
725 LASSERT(old_open_handle.cookie ==
726 epoch->mio_open_handle.cookie);
728 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
729 epoch->mio_open_handle = body->mbo_open_handle;
734 void mdc_commit_open(struct ptlrpc_request *req)
736 struct md_open_data *mod = req->rq_cb_data;
741 * No need to touch md_open_data::mod_och, it holds a reference on
742 * \var mod and will zero references to each other, \var mod will be
743 * freed after that when md_open_data::mod_och will put the reference.
747 * Do not let open request to disappear as it still may be needed
748 * for close rpc to happen (it may happen on evict only, otherwise
749 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
750 * called), just mark this rpc as committed to distinguish these 2
751 * cases, see mdc_close() for details. The open request reference will
752 * be put along with freeing \var mod.
754 ptlrpc_request_addref(req);
755 spin_lock(&req->rq_lock);
756 req->rq_committed = 1;
757 spin_unlock(&req->rq_lock);
758 req->rq_cb_data = NULL;
762 int mdc_set_open_replay_data(struct obd_export *exp,
763 struct obd_client_handle *och,
764 struct lookup_intent *it)
766 struct md_open_data *mod;
767 struct mdt_rec_create *rec;
768 struct mdt_body *body;
769 struct ptlrpc_request *open_req = it->it_request;
770 struct obd_import *imp = open_req->rq_import;
773 if (!open_req->rq_replay)
776 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
777 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
778 LASSERT(rec != NULL);
779 /* Incoming message in my byte order (it's been swabbed). */
780 /* Outgoing messages always in my byte order. */
781 LASSERT(body != NULL);
783 /* Only if the import is replayable, we set replay_open data */
784 if (och && imp->imp_replayable) {
785 mod = obd_mod_alloc();
787 DEBUG_REQ(D_ERROR, open_req,
788 "Can't allocate md_open_data");
793 * Take a reference on \var mod, to be freed on mdc_close().
794 * It protects \var mod from being freed on eviction (commit
795 * callback is called despite rq_replay flag).
796 * Another reference for \var och.
801 spin_lock(&open_req->rq_lock);
804 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
805 it_disposition(it, DISP_OPEN_STRIPE);
806 mod->mod_open_req = open_req;
807 open_req->rq_cb_data = mod;
808 open_req->rq_commit_cb = mdc_commit_open;
809 open_req->rq_early_free_repbuf = 1;
810 spin_unlock(&open_req->rq_lock);
813 rec->cr_fid2 = body->mbo_fid1;
814 rec->cr_open_handle_old = body->mbo_open_handle;
815 open_req->rq_replay_cb = mdc_replay_open;
816 if (!fid_is_sane(&body->mbo_fid1)) {
817 DEBUG_REQ(D_ERROR, open_req,
818 "saving replay request with insane FID " DFID,
819 PFID(&body->mbo_fid1));
823 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
827 static void mdc_free_open(struct md_open_data *mod)
831 if (mod->mod_is_create == 0 &&
832 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
836 * No reason to asssert here if the open request has
837 * rq_replay == 1. It means that mdc_close failed, and
838 * close request wasn`t sent. It is not fatal to client.
839 * The worst thing is eviction if the client gets open lock
842 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request rq_replay"
843 "= %d\n", mod->mod_open_req->rq_replay);
845 ptlrpc_request_committed(mod->mod_open_req, committed);
846 if (mod->mod_close_req)
847 ptlrpc_request_committed(mod->mod_close_req, committed);
850 int mdc_clear_open_replay_data(struct obd_export *exp,
851 struct obd_client_handle *och)
853 struct md_open_data *mod = och->och_mod;
857 * It is possible to not have \var mod in a case of eviction between
858 * lookup and ll_file_open().
863 LASSERT(mod != LP_POISON);
864 LASSERT(mod->mod_open_req != NULL);
866 spin_lock(&mod->mod_open_req->rq_lock);
868 mod->mod_och->och_open_handle.cookie = 0;
869 mod->mod_open_req->rq_early_free_repbuf = 0;
870 spin_unlock(&mod->mod_open_req->rq_lock);
880 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
881 struct md_open_data *mod, struct ptlrpc_request **request)
883 struct obd_device *obd = class_exp2obd(exp);
884 struct ptlrpc_request *req;
885 struct req_format *req_fmt;
886 size_t u32_count = 0;
891 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
892 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
895 if (op_data->op_bias & MDS_CLOSE_INTENT) {
896 req_fmt = &RQF_MDS_CLOSE_INTENT;
897 if (op_data->op_bias & MDS_HSM_RELEASE) {
898 /* allocate a FID for volatile file */
899 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
902 CERROR("%s: "DFID" allocating FID: rc = %d\n",
903 obd->obd_name, PFID(&op_data->op_fid1),
905 /* save the errcode and proceed to close */
909 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
910 size_t count = op_data->op_data_size / sizeof(__u32);
912 if (count > INLINE_RESYNC_ARRAY_SIZE)
916 req_fmt = &RQF_MDS_CLOSE;
920 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
923 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
925 /* Ensure that this close's handle is fixed up during replay. */
926 if (likely(mod != NULL)) {
927 LASSERTF(mod->mod_open_req != NULL &&
928 mod->mod_open_req->rq_type != LI_POISON,
929 "POISONED open %p!\n", mod->mod_open_req);
931 mod->mod_close_req = req;
933 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
934 /* We no longer want to preserve this open for replay even
935 * though the open was committed. b=3632, b=3633 */
936 spin_lock(&mod->mod_open_req->rq_lock);
937 mod->mod_open_req->rq_replay = 0;
938 spin_unlock(&mod->mod_open_req->rq_lock);
940 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
944 * TODO: repeat close after errors
946 CWARN("%s: close of FID "DFID" failed, file reference will be "
947 "dropped when this client unmounts or is evicted\n",
948 obd->obd_name, PFID(&op_data->op_fid1));
949 GOTO(out, rc = -ENOMEM);
953 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
954 u32_count * sizeof(__u32));
956 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
958 ptlrpc_request_free(req);
963 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
964 * portal whose threads are not taking any DLM locks and are therefore
965 * always progressing */
966 req->rq_request_portal = MDS_READPAGE_PORTAL;
967 ptlrpc_at_set_req_timeout(req);
969 if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
970 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
971 OP_XVALID_LAZYBLOCKS);
973 mdc_close_pack(req, op_data);
975 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
976 obd->u.cli.cl_default_mds_easize);
978 ptlrpc_request_set_replen(req);
980 mdc_get_mod_rpc_slot(req, NULL);
981 rc = ptlrpc_queue_wait(req);
982 mdc_put_mod_rpc_slot(req, NULL);
984 if (req->rq_repmsg == NULL) {
985 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
988 rc = req->rq_status ?: -EIO;
989 } else if (rc == 0 || rc == -EAGAIN) {
990 struct mdt_body *body;
992 rc = lustre_msg_get_status(req->rq_repmsg);
993 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
994 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
999 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1002 } else if (rc == -ESTALE) {
1004 * it can be allowed error after 3633 if open was committed and
1005 * server failed before close was sent. Let's check if mod
1006 * exists and return no error in that case
1009 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
1010 LASSERT(mod->mod_open_req != NULL);
1011 if (mod->mod_open_req->rq_committed)
1019 mod->mod_close_req = NULL;
1020 /* Since now, mod is accessed through open_req only,
1021 * thus close req does not keep a reference on mod anymore. */
1026 RETURN(rc < 0 ? rc : saved_rc);
1029 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1030 u64 offset, struct page **pages, int npages,
1031 struct ptlrpc_request **request)
1033 struct ptlrpc_request *req;
1034 struct ptlrpc_bulk_desc *desc;
1036 wait_queue_head_t waitq;
1038 struct l_wait_info lwi;
1043 init_waitqueue_head(&waitq);
1046 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1050 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1052 ptlrpc_request_free(req);
1056 req->rq_request_portal = MDS_READPAGE_PORTAL;
1057 ptlrpc_at_set_req_timeout(req);
1059 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1060 PTLRPC_BULK_PUT_SINK | PTLRPC_BULK_BUF_KIOV,
1062 &ptlrpc_bulk_kiov_pin_ops);
1064 ptlrpc_req_finished(req);
1068 /* NB req now owns desc and will free it when it gets freed */
1069 for (i = 0; i < npages; i++)
1070 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1073 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1075 ptlrpc_request_set_replen(req);
1076 rc = ptlrpc_queue_wait(req);
1078 ptlrpc_req_finished(req);
1079 if (rc != -ETIMEDOUT)
1083 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1084 CERROR("%s: too many resend retries: rc = %d\n",
1085 exp->exp_obd->obd_name, -EIO);
1088 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
1090 l_wait_event(waitq, 0, &lwi);
1095 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1096 req->rq_bulk->bd_nob_transferred);
1098 ptlrpc_req_finished(req);
1102 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1103 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1104 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1105 PAGE_SIZE * npages);
1106 ptlrpc_req_finished(req);
1114 static void mdc_release_page(struct page *page, int remove)
1118 if (likely(page->mapping != NULL))
1119 truncate_complete_page(page->mapping, page);
1125 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1126 __u64 *start, __u64 *end, int hash64)
1129 * Complement of hash is used as an index so that
1130 * radix_tree_gang_lookup() can be used to find a page with starting
1131 * hash _smaller_ than one we are looking for.
1133 unsigned long offset = hash_x_index(*hash, hash64);
1137 xa_lock_irq(&mapping->i_pages);
1138 found = radix_tree_gang_lookup(&mapping->page_tree,
1139 (void **)&page, offset, 1);
1140 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1141 struct lu_dirpage *dp;
1144 xa_unlock_irq(&mapping->i_pages);
1146 * In contrast to find_lock_page() we are sure that directory
1147 * page cannot be truncated (while DLM lock is held) and,
1148 * hence, can avoid restart.
1150 * In fact, page cannot be locked here at all, because
1151 * mdc_read_page_remote does synchronous io.
1153 wait_on_page_locked(page);
1154 if (PageUptodate(page)) {
1156 if (BITS_PER_LONG == 32 && hash64) {
1157 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1158 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1159 *hash = *hash >> 32;
1161 *start = le64_to_cpu(dp->ldp_hash_start);
1162 *end = le64_to_cpu(dp->ldp_hash_end);
1164 if (unlikely(*start == 1 && *hash == 0))
1167 LASSERTF(*start <= *hash, "start = %#llx"
1168 ",end = %#llx,hash = %#llx\n",
1169 *start, *end, *hash);
1170 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1171 " hash %#llx\n", offset, *start, *end, *hash);
1174 mdc_release_page(page, 0);
1176 } else if (*end != *start && *hash == *end) {
1178 * upon hash collision, remove this page,
1179 * otherwise put page reference, and
1180 * mdc_read_page_remote() will issue RPC to
1181 * fetch the page we want.
1184 mdc_release_page(page,
1185 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1190 page = ERR_PTR(-EIO);
1193 xa_unlock_irq(&mapping->i_pages);
1200 * Adjust a set of pages, each page containing an array of lu_dirpages,
1201 * so that each page can be used as a single logical lu_dirpage.
1203 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1204 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1205 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1206 * value is used as a cookie to request the next lu_dirpage in a
1207 * directory listing that spans multiple pages (two in this example):
1210 * .|--------v------- -----.
1211 * |s|e|f|p|ent|ent| ... |ent|
1212 * '--|-------------- -----' Each PAGE contains a single
1213 * '------. lu_dirpage.
1214 * .---------v------- -----.
1215 * |s|e|f|p|ent| 0 | ... | 0 |
1216 * '----------------- -----'
1218 * However, on hosts where the native VM page size (PAGE_SIZE) is
1219 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1220 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1221 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1222 * after it in the same PAGE (arrows simplified for brevity, but
1223 * in general e0==s1, e1==s2, etc.):
1225 * .-------------------- -----.
1226 * |s0|e0|f0|p|ent|ent| ... |ent|
1227 * |---v---------------- -----|
1228 * |s1|e1|f1|p|ent|ent| ... |ent|
1229 * |---v---------------- -----| Here, each PAGE contains
1230 * ... multiple lu_dirpages.
1231 * |---v---------------- -----|
1232 * |s'|e'|f'|p|ent|ent| ... |ent|
1233 * '---|---------------- -----'
1235 * .----------------------------.
1238 * This structure is transformed into a single logical lu_dirpage as follows:
1240 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1241 * labeled 'next PAGE'.
1243 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1244 * a hash collision with the next page exists.
1246 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1247 * to the first entry of the next lu_dirpage.
1249 #if PAGE_SIZE > LU_PAGE_SIZE
1250 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1254 for (i = 0; i < cfs_pgs; i++) {
1255 struct lu_dirpage *dp = kmap(pages[i]);
1256 struct lu_dirpage *first = dp;
1257 struct lu_dirent *end_dirent = NULL;
1258 struct lu_dirent *ent;
1259 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1260 __u32 flags = le32_to_cpu(dp->ldp_flags);
1262 while (--lu_pgs > 0) {
1263 ent = lu_dirent_start(dp);
1264 for (end_dirent = ent; ent != NULL;
1265 end_dirent = ent, ent = lu_dirent_next(ent));
1267 /* Advance dp to next lu_dirpage. */
1268 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1270 /* Check if we've reached the end of the PAGE. */
1271 if (!((unsigned long)dp & ~PAGE_MASK))
1274 /* Save the hash and flags of this lu_dirpage. */
1275 hash_end = le64_to_cpu(dp->ldp_hash_end);
1276 flags = le32_to_cpu(dp->ldp_flags);
1278 /* Check if lu_dirpage contains no entries. */
1279 if (end_dirent == NULL)
1282 /* Enlarge the end entry lde_reclen from 0 to
1283 * first entry of next lu_dirpage. */
1284 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1285 end_dirent->lde_reclen =
1286 cpu_to_le16((char *)(dp->ldp_entries) -
1287 (char *)end_dirent);
1290 first->ldp_hash_end = hash_end;
1291 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1292 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1296 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1299 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1300 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1302 /* parameters for readdir page */
1303 struct readpage_param {
1304 struct md_op_data *rp_mod;
1307 struct obd_export *rp_exp;
1308 struct md_callback *rp_cb;
1311 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1312 static inline void delete_from_page_cache(struct page *page)
1314 remove_from_page_cache(page);
1320 * Read pages from server.
1322 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1323 * a header lu_dirpage which describes the start/end hash, and whether this
1324 * page is empty (contains no dir entry) or hash collide with next page.
1325 * After client receives reply, several pages will be integrated into dir page
1326 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1327 * lu_dirpage for this integrated page will be adjusted.
1329 static int mdc_read_page_remote(void *data, struct page *page0)
1331 struct readpage_param *rp = data;
1332 struct page **page_pool;
1334 struct lu_dirpage *dp;
1335 struct md_op_data *op_data = rp->rp_mod;
1336 struct ptlrpc_request *req;
1338 struct inode *inode;
1340 int rd_pgs = 0; /* number of pages actually read */
1346 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1347 inode = op_data->op_data;
1348 fid = &op_data->op_fid1;
1349 LASSERT(inode != NULL);
1351 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1352 if (page_pool != NULL) {
1353 page_pool[0] = page0;
1359 for (npages = 1; npages < max_pages; npages++) {
1360 page = __page_cache_alloc(mapping_gfp_mask(inode->i_mapping)
1364 page_pool[npages] = page;
1367 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1369 /* page0 is special, which was added into page cache early */
1370 delete_from_page_cache(page0);
1374 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1376 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1377 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1379 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1381 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1383 SetPageUptodate(page0);
1387 ptlrpc_req_finished(req);
1388 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1389 for (i = 1; i < npages; i++) {
1390 unsigned long offset;
1394 page = page_pool[i];
1396 if (rc < 0 || i >= rd_pgs) {
1401 SetPageUptodate(page);
1404 hash = le64_to_cpu(dp->ldp_hash_start);
1407 offset = hash_x_index(hash, rp->rp_hash64);
1409 prefetchw(&page->flags);
1410 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1415 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1416 " rc = %d\n", offset, ret);
1420 if (page_pool != &page0)
1421 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1427 * Read dir page from cache first, if it can not find it, read it from
1428 * server and add into the cache.
1430 * \param[in] exp MDC export
1431 * \param[in] op_data client MD stack parameters, transfering parameters
1432 * between different layers on client MD stack.
1433 * \param[in] cb_op callback required for ldlm lock enqueue during
1435 * \param[in] hash_offset the hash offset of the page to be read
1436 * \param[in] ppage the page to be read
1438 * retval = 0 get the page successfully
1439 * errno(<0) get the page failed
1441 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1442 struct md_callback *cb_op, __u64 hash_offset,
1443 struct page **ppage)
1445 struct lookup_intent it = { .it_op = IT_READDIR };
1447 struct inode *dir = op_data->op_data;
1448 struct address_space *mapping;
1449 struct lu_dirpage *dp;
1452 struct lustre_handle lockh;
1453 struct ptlrpc_request *enq_req = NULL;
1454 struct readpage_param rp_param;
1461 LASSERT(dir != NULL);
1462 mapping = dir->i_mapping;
1464 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1465 cb_op->md_blocking_ast, 0);
1466 if (enq_req != NULL)
1467 ptlrpc_req_finished(enq_req);
1470 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1471 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1476 lockh.cookie = it.it_lock_handle;
1477 mdc_set_lock_data(exp, &lockh, dir, NULL);
1479 rp_param.rp_off = hash_offset;
1480 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1481 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1482 rp_param.rp_hash64);
1484 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1485 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1486 rp_param.rp_off, PTR_ERR(page));
1487 GOTO(out_unlock, rc = PTR_ERR(page));
1488 } else if (page != NULL) {
1490 * XXX nikita: not entirely correct handling of a corner case:
1491 * suppose hash chain of entries with hash value HASH crosses
1492 * border between pages P0 and P1. First both P0 and P1 are
1493 * cached, seekdir() is called for some entry from the P0 part
1494 * of the chain. Later P0 goes out of cache. telldir(HASH)
1495 * happens and finds P1, as it starts with matching hash
1496 * value. Remaining entries from P0 part of the chain are
1497 * skipped. (Is that really a bug?)
1499 * Possible solutions: 0. don't cache P1 is such case, handle
1500 * it as an "overflow" page. 1. invalidate all pages at
1501 * once. 2. use HASH|1 as an index for P1.
1503 GOTO(hash_collision, page);
1506 rp_param.rp_exp = exp;
1507 rp_param.rp_mod = op_data;
1508 page = read_cache_page(mapping,
1509 hash_x_index(rp_param.rp_off,
1510 rp_param.rp_hash64),
1511 mdc_read_page_remote, &rp_param);
1513 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1514 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1515 rp_param.rp_off, PTR_ERR(page));
1516 GOTO(out_unlock, rc = PTR_ERR(page));
1519 wait_on_page_locked(page);
1521 if (!PageUptodate(page)) {
1522 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1523 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1524 rp_param.rp_off, -5);
1527 if (!PageChecked(page))
1528 SetPageChecked(page);
1529 if (PageError(page)) {
1530 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1531 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1532 rp_param.rp_off, -5);
1537 dp = page_address(page);
1538 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1539 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1540 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1541 rp_param.rp_off = hash_offset >> 32;
1543 start = le64_to_cpu(dp->ldp_hash_start);
1544 end = le64_to_cpu(dp->ldp_hash_end);
1545 rp_param.rp_off = hash_offset;
1548 LASSERT(start == rp_param.rp_off);
1549 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1550 #if BITS_PER_LONG == 32
1551 CWARN("Real page-wide hash collision at [%llu %llu] with "
1552 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1553 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1557 * Fetch whole overflow chain...
1565 ldlm_lock_decref(&lockh, it.it_lock_mode);
1569 mdc_release_page(page, 1);
1574 static int mdc_statfs(const struct lu_env *env,
1575 struct obd_export *exp, struct obd_statfs *osfs,
1576 time64_t max_age, __u32 flags)
1578 struct obd_device *obd = class_exp2obd(exp);
1579 struct req_format *fmt;
1580 struct ptlrpc_request *req;
1581 struct obd_statfs *msfs;
1582 struct obd_import *imp = NULL;
1587 * Since the request might also come from lprocfs, so we need
1588 * sync this with client_disconnect_export Bug15684
1590 down_read(&obd->u.cli.cl_sem);
1591 if (obd->u.cli.cl_import)
1592 imp = class_import_get(obd->u.cli.cl_import);
1593 up_read(&obd->u.cli.cl_sem);
1597 fmt = &RQF_MDS_STATFS;
1598 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1599 (flags & OBD_STATFS_SUM))
1600 fmt = &RQF_MDS_STATFS_NEW;
1601 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1604 GOTO(output, rc = -ENOMEM);
1606 if ((flags & OBD_STATFS_SUM) &&
1607 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1608 /* request aggregated states */
1609 struct mdt_body *body;
1611 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1613 GOTO(out, rc = -EPROTO);
1614 body->mbo_valid = OBD_MD_FLAGSTATFS;
1617 ptlrpc_request_set_replen(req);
1619 if (flags & OBD_STATFS_NODELAY) {
1620 /* procfs requests not want stay in wait for avoid deadlock */
1621 req->rq_no_resend = 1;
1622 req->rq_no_delay = 1;
1625 rc = ptlrpc_queue_wait(req);
1627 /* check connection error first */
1628 if (imp->imp_connect_error)
1629 rc = imp->imp_connect_error;
1633 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1635 GOTO(out, rc = -EPROTO);
1640 ptlrpc_req_finished(req);
1642 class_import_put(imp);
1646 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1648 __u32 keylen, vallen;
1652 if (gf->gf_pathlen > PATH_MAX)
1653 RETURN(-ENAMETOOLONG);
1654 if (gf->gf_pathlen < 2)
1657 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1658 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1659 sizeof(struct lu_fid));
1660 OBD_ALLOC(key, keylen);
1663 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1664 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1665 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1666 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1667 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1668 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1670 if (!fid_is_sane(&gf->gf_fid))
1671 GOTO(out, rc = -EINVAL);
1673 /* Val is struct getinfo_fid2path result plus path */
1674 vallen = sizeof(*gf) + gf->gf_pathlen;
1676 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1677 if (rc != 0 && rc != -EREMOTE)
1680 if (vallen <= sizeof(*gf))
1681 GOTO(out, rc = -EPROTO);
1682 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1683 GOTO(out, rc = -EOVERFLOW);
1685 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1686 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1687 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1688 /* only log the last 512 characters of the path */
1689 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1692 OBD_FREE(key, keylen);
1696 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1697 struct hsm_progress_kernel *hpk)
1699 struct obd_import *imp = class_exp2cliimp(exp);
1700 struct hsm_progress_kernel *req_hpk;
1701 struct ptlrpc_request *req;
1705 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1706 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1708 GOTO(out, rc = -ENOMEM);
1710 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1712 /* Copy hsm_progress struct */
1713 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1714 if (req_hpk == NULL)
1715 GOTO(out, rc = -EPROTO);
1718 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1720 ptlrpc_request_set_replen(req);
1722 mdc_get_mod_rpc_slot(req, NULL);
1723 rc = ptlrpc_queue_wait(req);
1724 mdc_put_mod_rpc_slot(req, NULL);
1728 ptlrpc_req_finished(req);
1732 * Send hsm_ct_register to MDS
1734 * \param[in] imp import
1735 * \param[in] archive_count if in bitmap format, it is the bitmap,
1736 * else it is the count of archive_ids
1737 * \param[in] archives if in bitmap format, it is NULL,
1738 * else it is archive_id lists
1740 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1743 struct ptlrpc_request *req;
1744 __u32 *archive_array;
1745 size_t archives_size;
1749 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1753 if (archives != NULL)
1754 archives_size = sizeof(*archive_array) * archive_count;
1756 archives_size = sizeof(archive_count);
1758 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1759 RCL_CLIENT, archives_size);
1761 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1763 ptlrpc_request_free(req);
1767 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1769 archive_array = req_capsule_client_get(&req->rq_pill,
1770 &RMF_MDS_HSM_ARCHIVE);
1771 if (archive_array == NULL)
1772 GOTO(out, rc = -EPROTO);
1774 if (archives != NULL)
1775 memcpy(archive_array, archives, archives_size);
1777 *archive_array = archive_count;
1779 ptlrpc_request_set_replen(req);
1781 rc = mdc_queue_wait(req);
1784 ptlrpc_req_finished(req);
1788 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1789 struct md_op_data *op_data)
1791 struct hsm_current_action *hca = op_data->op_data;
1792 struct hsm_current_action *req_hca;
1793 struct ptlrpc_request *req;
1797 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1798 &RQF_MDS_HSM_ACTION);
1802 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1804 ptlrpc_request_free(req);
1808 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1809 op_data->op_suppgids[0], 0);
1811 ptlrpc_request_set_replen(req);
1813 rc = mdc_queue_wait(req);
1817 req_hca = req_capsule_server_get(&req->rq_pill,
1818 &RMF_MDS_HSM_CURRENT_ACTION);
1819 if (req_hca == NULL)
1820 GOTO(out, rc = -EPROTO);
1826 ptlrpc_req_finished(req);
1830 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1832 struct ptlrpc_request *req;
1836 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1838 MDS_HSM_CT_UNREGISTER);
1840 GOTO(out, rc = -ENOMEM);
1842 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1844 ptlrpc_request_set_replen(req);
1846 rc = mdc_queue_wait(req);
1849 ptlrpc_req_finished(req);
1853 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1854 struct md_op_data *op_data)
1856 struct hsm_user_state *hus = op_data->op_data;
1857 struct hsm_user_state *req_hus;
1858 struct ptlrpc_request *req;
1862 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1863 &RQF_MDS_HSM_STATE_GET);
1867 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1869 ptlrpc_request_free(req);
1873 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1874 op_data->op_suppgids[0], 0);
1876 ptlrpc_request_set_replen(req);
1878 rc = mdc_queue_wait(req);
1882 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1883 if (req_hus == NULL)
1884 GOTO(out, rc = -EPROTO);
1890 ptlrpc_req_finished(req);
1894 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1895 struct md_op_data *op_data)
1897 struct hsm_state_set *hss = op_data->op_data;
1898 struct hsm_state_set *req_hss;
1899 struct ptlrpc_request *req;
1903 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1904 &RQF_MDS_HSM_STATE_SET);
1908 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1910 ptlrpc_request_free(req);
1914 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1915 op_data->op_suppgids[0], 0);
1918 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1919 if (req_hss == NULL)
1920 GOTO(out, rc = -EPROTO);
1923 ptlrpc_request_set_replen(req);
1925 mdc_get_mod_rpc_slot(req, NULL);
1926 rc = ptlrpc_queue_wait(req);
1927 mdc_put_mod_rpc_slot(req, NULL);
1931 ptlrpc_req_finished(req);
1935 static int mdc_ioc_hsm_request(struct obd_export *exp,
1936 struct hsm_user_request *hur)
1938 struct obd_import *imp = class_exp2cliimp(exp);
1939 struct ptlrpc_request *req;
1940 struct hsm_request *req_hr;
1941 struct hsm_user_item *req_hui;
1946 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1948 GOTO(out, rc = -ENOMEM);
1950 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1951 hur->hur_request.hr_itemcount
1952 * sizeof(struct hsm_user_item));
1953 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1954 hur->hur_request.hr_data_len);
1956 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1958 ptlrpc_request_free(req);
1962 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1964 /* Copy hsm_request struct */
1965 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1967 GOTO(out, rc = -EPROTO);
1968 *req_hr = hur->hur_request;
1970 /* Copy hsm_user_item structs */
1971 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1972 if (req_hui == NULL)
1973 GOTO(out, rc = -EPROTO);
1974 memcpy(req_hui, hur->hur_user_item,
1975 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1977 /* Copy opaque field */
1978 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1979 if (req_opaque == NULL)
1980 GOTO(out, rc = -EPROTO);
1981 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1983 ptlrpc_request_set_replen(req);
1985 mdc_get_mod_rpc_slot(req, NULL);
1986 rc = ptlrpc_queue_wait(req);
1987 mdc_put_mod_rpc_slot(req, NULL);
1992 ptlrpc_req_finished(req);
1996 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1997 struct lustre_kernelcomm *lk);
1999 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2000 struct obd_quotactl *oqctl)
2002 struct ptlrpc_request *req;
2003 struct obd_quotactl *oqc;
2007 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
2008 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
2013 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2016 ptlrpc_request_set_replen(req);
2017 ptlrpc_at_set_req_timeout(req);
2019 rc = ptlrpc_queue_wait(req);
2021 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
2023 if (req->rq_repmsg &&
2024 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2027 CERROR ("Can't unpack obd_quotactl\n");
2030 ptlrpc_req_finished(req);
2035 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2036 struct md_op_data *op_data)
2038 struct list_head cancels = LIST_HEAD_INIT(cancels);
2039 struct ptlrpc_request *req;
2041 struct mdc_swap_layouts *msl, *payload;
2044 msl = op_data->op_data;
2046 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2047 * first thing it will do is to cancel the 2 layout
2048 * locks held by this client.
2049 * So the client must cancel its layout locks on the 2 fids
2050 * with the request RPC to avoid extra RPC round trips.
2052 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2053 LCK_EX, MDS_INODELOCK_LAYOUT |
2054 MDS_INODELOCK_XATTR);
2055 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2056 LCK_EX, MDS_INODELOCK_LAYOUT |
2057 MDS_INODELOCK_XATTR);
2059 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2060 &RQF_MDS_SWAP_LAYOUTS);
2062 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2066 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2068 ptlrpc_request_free(req);
2072 mdc_swap_layouts_pack(req, op_data);
2074 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2079 ptlrpc_request_set_replen(req);
2081 rc = ptlrpc_queue_wait(req);
2087 ptlrpc_req_finished(req);
2091 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2092 void *karg, void __user *uarg)
2094 struct obd_device *obd = exp->exp_obd;
2095 struct obd_ioctl_data *data = karg;
2096 struct obd_import *imp = obd->u.cli.cl_import;
2100 if (!try_module_get(THIS_MODULE)) {
2101 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2102 module_name(THIS_MODULE));
2106 case OBD_IOC_FID2PATH:
2107 rc = mdc_ioc_fid2path(exp, karg);
2109 case LL_IOC_HSM_CT_START:
2110 rc = mdc_ioc_hsm_ct_start(exp, karg);
2111 /* ignore if it was already registered on this MDS. */
2115 case LL_IOC_HSM_PROGRESS:
2116 rc = mdc_ioc_hsm_progress(exp, karg);
2118 case LL_IOC_HSM_STATE_GET:
2119 rc = mdc_ioc_hsm_state_get(exp, karg);
2121 case LL_IOC_HSM_STATE_SET:
2122 rc = mdc_ioc_hsm_state_set(exp, karg);
2124 case LL_IOC_HSM_ACTION:
2125 rc = mdc_ioc_hsm_current_action(exp, karg);
2127 case LL_IOC_HSM_REQUEST:
2128 rc = mdc_ioc_hsm_request(exp, karg);
2130 case OBD_IOC_CLIENT_RECOVER:
2131 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2135 case IOC_OSC_SET_ACTIVE:
2136 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2139 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2140 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2141 * there'd be no LMV layer thus we might be called here. Eventually
2142 * this code should be removed.
2145 case IOC_OBD_STATFS: {
2146 struct obd_statfs stat_buf = {0};
2148 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2149 GOTO(out, rc = -ENODEV);
2152 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2153 min((int)data->ioc_plen2,
2154 (int)sizeof(struct obd_uuid))))
2155 GOTO(out, rc = -EFAULT);
2157 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2158 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2163 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2164 min((int) data->ioc_plen1,
2165 (int) sizeof(stat_buf))))
2166 GOTO(out, rc = -EFAULT);
2170 case OBD_IOC_QUOTACTL: {
2171 struct if_quotactl *qctl = karg;
2172 struct obd_quotactl *oqctl;
2174 OBD_ALLOC_PTR(oqctl);
2176 GOTO(out, rc = -ENOMEM);
2178 QCTL_COPY(oqctl, qctl);
2179 rc = obd_quotactl(exp, oqctl);
2181 QCTL_COPY(qctl, oqctl);
2182 qctl->qc_valid = QC_MDTIDX;
2183 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2186 OBD_FREE_PTR(oqctl);
2189 case LL_IOC_GET_CONNECT_FLAGS:
2190 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2191 sizeof(*exp_connect_flags_ptr(exp))))
2192 GOTO(out, rc = -EFAULT);
2195 case LL_IOC_LOV_SWAP_LAYOUTS:
2196 rc = mdc_ioc_swap_layouts(exp, karg);
2199 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2200 GOTO(out, rc = -ENOTTY);
2203 module_put(THIS_MODULE);
2208 static int mdc_get_info_rpc(struct obd_export *exp,
2209 u32 keylen, void *key,
2210 u32 vallen, void *val)
2212 struct obd_import *imp = class_exp2cliimp(exp);
2213 struct ptlrpc_request *req;
2218 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2222 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2223 RCL_CLIENT, keylen);
2224 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2225 RCL_CLIENT, sizeof(vallen));
2227 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2229 ptlrpc_request_free(req);
2233 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2234 memcpy(tmp, key, keylen);
2235 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2236 memcpy(tmp, &vallen, sizeof(vallen));
2238 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2239 RCL_SERVER, vallen);
2240 ptlrpc_request_set_replen(req);
2242 rc = ptlrpc_queue_wait(req);
2243 /* -EREMOTE means the get_info result is partial, and it needs to
2244 * continue on another MDT, see fid2path part in lmv_iocontrol */
2245 if (rc == 0 || rc == -EREMOTE) {
2246 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2247 memcpy(val, tmp, vallen);
2248 if (ptlrpc_rep_need_swab(req)) {
2249 if (KEY_IS(KEY_FID2PATH))
2250 lustre_swab_fid2path(val);
2253 ptlrpc_req_finished(req);
2258 static void lustre_swab_hai(struct hsm_action_item *h)
2260 __swab32s(&h->hai_len);
2261 __swab32s(&h->hai_action);
2262 lustre_swab_lu_fid(&h->hai_fid);
2263 lustre_swab_lu_fid(&h->hai_dfid);
2264 __swab64s(&h->hai_cookie);
2265 __swab64s(&h->hai_extent.offset);
2266 __swab64s(&h->hai_extent.length);
2267 __swab64s(&h->hai_gid);
2270 static void lustre_swab_hal(struct hsm_action_list *h)
2272 struct hsm_action_item *hai;
2275 __swab32s(&h->hal_version);
2276 __swab32s(&h->hal_count);
2277 __swab32s(&h->hal_archive_id);
2278 __swab64s(&h->hal_flags);
2280 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2281 lustre_swab_hai(hai);
2284 static void lustre_swab_kuch(struct kuc_hdr *l)
2286 __swab16s(&l->kuc_magic);
2287 /* __u8 l->kuc_transport */
2288 __swab16s(&l->kuc_msgtype);
2289 __swab16s(&l->kuc_msglen);
2292 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2293 struct lustre_kernelcomm *lk)
2295 struct obd_import *imp = class_exp2cliimp(exp);
2298 if (lk->lk_group != KUC_GRP_HSM) {
2299 CERROR("Bad copytool group %d\n", lk->lk_group);
2303 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2304 lk->lk_uid, lk->lk_group, lk->lk_flags);
2306 if (lk->lk_flags & LK_FLG_STOP) {
2307 /* Unregister with the coordinator */
2308 rc = mdc_ioc_hsm_ct_unregister(imp);
2310 __u32 *archives = NULL;
2312 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2313 archives = lk->lk_data;
2315 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2322 * Send a message to any listening copytools
2323 * @param val KUC message (kuc_hdr + hsm_action_list)
2324 * @param len total length of message
2326 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2327 size_t len, void *val)
2329 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2330 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2334 if (len < sizeof(*lh) + sizeof(*hal)) {
2335 CERROR("Short HSM message %zu < %zu\n", len,
2336 sizeof(*lh) + sizeof(*hal));
2339 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2340 lustre_swab_kuch(lh);
2341 lustre_swab_hal(hal);
2342 } else if (lh->kuc_magic != KUC_MAGIC) {
2343 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2347 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2349 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2350 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2352 /* Broadcast to HSM listeners */
2353 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2359 * callback function passed to kuc for re-registering each HSM copytool
2360 * running on MDC, after MDT shutdown/recovery.
2361 * @param data copytool registration data
2362 * @param cb_arg callback argument (obd_import)
2364 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2366 struct obd_import *imp = (struct obd_import *)cb_arg;
2367 struct kkuc_ct_data *kcd = data;
2368 __u32 *archives = NULL;
2372 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2373 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2376 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2377 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2378 "(archive=%#x)\n", imp->imp_obd->obd_name,
2379 kcd->kcd_nr_archives);
2381 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2382 "(archive nr = %u)\n",
2383 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2384 if (kcd->kcd_nr_archives != 0)
2385 archives = kcd->kcd_archives;
2388 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2389 /* ignore error if the copytool is already registered */
2390 return (rc == -EEXIST) ? 0 : rc;
2394 * Re-establish all kuc contexts with MDT
2395 * after MDT shutdown/recovery.
2397 static int mdc_kuc_reregister(struct obd_import *imp)
2399 /* re-register HSM agents */
2400 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2401 mdc_hsm_ct_reregister, imp);
2404 static int mdc_set_info_async(const struct lu_env *env,
2405 struct obd_export *exp,
2406 u32 keylen, void *key,
2407 u32 vallen, void *val,
2408 struct ptlrpc_request_set *set)
2410 struct obd_import *imp = class_exp2cliimp(exp);
2414 if (KEY_IS(KEY_READ_ONLY)) {
2415 if (vallen != sizeof(int))
2418 spin_lock(&imp->imp_lock);
2419 if (*((int *)val)) {
2420 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2421 imp->imp_connect_data.ocd_connect_flags |=
2424 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2425 imp->imp_connect_data.ocd_connect_flags &=
2426 ~OBD_CONNECT_RDONLY;
2428 spin_unlock(&imp->imp_lock);
2430 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2431 keylen, key, vallen, val, set);
2434 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2435 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2436 keylen, key, vallen, val, set);
2439 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2440 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2445 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2446 __u32 *default_easize = val;
2448 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2452 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2456 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2457 __u32 keylen, void *key, __u32 *vallen, void *val)
2461 if (KEY_IS(KEY_MAX_EASIZE)) {
2462 __u32 mdsize, *max_easize;
2464 if (*vallen != sizeof(int))
2466 mdsize = *(__u32 *)val;
2467 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2468 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2470 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2472 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2473 __u32 *default_easize;
2475 if (*vallen != sizeof(int))
2477 default_easize = val;
2478 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2480 } else if (KEY_IS(KEY_CONN_DATA)) {
2481 struct obd_import *imp = class_exp2cliimp(exp);
2482 struct obd_connect_data *data = val;
2484 if (*vallen != sizeof(*data))
2487 *data = imp->imp_connect_data;
2489 } else if (KEY_IS(KEY_TGT_COUNT)) {
2490 *((__u32 *)val) = 1;
2494 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2499 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2500 struct ptlrpc_request **request)
2502 struct ptlrpc_request *req;
2507 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2511 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2513 ptlrpc_request_free(req);
2517 mdc_pack_body(req, fid, 0, 0, -1, 0);
2519 ptlrpc_request_set_replen(req);
2521 rc = ptlrpc_queue_wait(req);
2523 ptlrpc_req_finished(req);
2529 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2530 enum obd_import_event event)
2532 struct client_obd *cli = &obd->u.cli;
2535 LASSERT(imp->imp_obd == obd);
2538 case IMP_EVENT_DISCON:
2539 spin_lock(&cli->cl_loi_list_lock);
2540 cli->cl_avail_grant = 0;
2541 cli->cl_lost_grant = 0;
2542 spin_unlock(&cli->cl_loi_list_lock);
2544 case IMP_EVENT_INACTIVE:
2546 * Flush current sequence to make client obtain new one
2547 * from server in case of disconnect/reconnect.
2549 down_read(&cli->cl_seq_rwsem);
2551 seq_client_flush(cli->cl_seq);
2552 up_read(&cli->cl_seq_rwsem);
2554 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2556 case IMP_EVENT_INVALIDATE: {
2557 struct ldlm_namespace *ns = obd->obd_namespace;
2561 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2563 env = cl_env_get(&refcheck);
2565 /* Reset grants. All pages go to failing rpcs due to
2566 * the invalid import.
2568 osc_io_unplug(env, cli, NULL);
2570 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2571 osc_ldlm_resource_invalidate,
2573 cl_env_put(env, &refcheck);
2574 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2580 case IMP_EVENT_ACTIVE:
2581 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2582 /* redo the kuc registration after reconnecting */
2584 rc = mdc_kuc_reregister(imp);
2586 case IMP_EVENT_OCD: {
2587 struct obd_connect_data *ocd = &imp->imp_connect_data;
2589 if (OCD_HAS_FLAG(ocd, GRANT))
2590 osc_init_grant(cli, ocd);
2592 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2595 case IMP_EVENT_DEACTIVATE:
2596 case IMP_EVENT_ACTIVATE:
2599 CERROR("Unknown import event %x\n", event);
2605 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2606 struct lu_fid *fid, struct md_op_data *op_data)
2608 struct client_obd *cli = &exp->exp_obd->u.cli;
2613 down_read(&cli->cl_seq_rwsem);
2615 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2616 up_read(&cli->cl_seq_rwsem);
2621 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2623 struct client_obd *cli = &exp->exp_obd->u.cli;
2624 return &cli->cl_target_uuid;
2628 * Determine whether the lock can be canceled before replaying it during
2629 * recovery, non zero value will be return if the lock can be canceled,
2630 * or zero returned for not
2632 static int mdc_cancel_weight(struct ldlm_lock *lock)
2634 if (lock->l_resource->lr_type != LDLM_IBITS)
2637 /* FIXME: if we ever get into a situation where there are too many
2638 * opened files with open locks on a single node, then we really
2639 * should replay these open locks to reget it */
2640 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2643 /* Special case for DoM locks, cancel only unused and granted locks */
2644 if (ldlm_has_dom(lock) &&
2645 (lock->l_granted_mode != lock->l_req_mode ||
2646 osc_ldlm_weigh_ast(lock) != 0))
2652 static int mdc_resource_inode_free(struct ldlm_resource *res)
2654 if (res->lr_lvb_inode)
2655 res->lr_lvb_inode = NULL;
2660 static struct ldlm_valblock_ops inode_lvbo = {
2661 .lvbo_free = mdc_resource_inode_free
2664 static int mdc_llog_init(struct obd_device *obd)
2666 struct obd_llog_group *olg = &obd->obd_olg;
2667 struct llog_ctxt *ctxt;
2672 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2677 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2678 llog_initiator_connect(ctxt);
2679 llog_ctxt_put(ctxt);
2684 static void mdc_llog_finish(struct obd_device *obd)
2686 struct llog_ctxt *ctxt;
2690 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2692 llog_cleanup(NULL, ctxt);
2697 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2703 rc = osc_setup_common(obd, cfg);
2707 rc = mdc_tunables_init(obd);
2709 GOTO(err_osc_cleanup, rc);
2711 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2713 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2715 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2717 rc = mdc_llog_init(obd);
2719 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2721 GOTO(err_llog_cleanup, rc);
2724 rc = mdc_changelog_cdev_init(obd);
2726 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2728 GOTO(err_changelog_cleanup, rc);
2733 err_changelog_cleanup:
2734 mdc_llog_finish(obd);
2736 lprocfs_free_md_stats(obd);
2737 ptlrpc_lprocfs_unregister_obd(obd);
2739 osc_cleanup_common(obd);
2743 /* Initialize the default and maximum LOV EA sizes. This allows
2744 * us to make MDS RPCs with large enough reply buffers to hold a default
2745 * sized EA without having to calculate this (via a call into the
2746 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2747 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2748 * a large number of stripes is possible. If a larger reply buffer is
2749 * required it will be reallocated in the ptlrpc layer due to overflow.
2751 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2754 struct obd_device *obd = exp->exp_obd;
2755 struct client_obd *cli = &obd->u.cli;
2758 if (cli->cl_max_mds_easize < easize)
2759 cli->cl_max_mds_easize = easize;
2761 if (cli->cl_default_mds_easize < def_easize)
2762 cli->cl_default_mds_easize = def_easize;
2767 static int mdc_precleanup(struct obd_device *obd)
2771 osc_precleanup_common(obd);
2772 mdc_changelog_cdev_finish(obd);
2774 obd_cleanup_client_import(obd);
2775 ptlrpc_lprocfs_unregister_obd(obd);
2776 lprocfs_free_md_stats(obd);
2777 mdc_llog_finish(obd);
2781 static int mdc_cleanup(struct obd_device *obd)
2783 return osc_cleanup_common(obd);
2786 static struct obd_ops mdc_obd_ops = {
2787 .o_owner = THIS_MODULE,
2788 .o_setup = mdc_setup,
2789 .o_precleanup = mdc_precleanup,
2790 .o_cleanup = mdc_cleanup,
2791 .o_add_conn = client_import_add_conn,
2792 .o_del_conn = client_import_del_conn,
2793 .o_connect = client_connect_import,
2794 .o_reconnect = osc_reconnect,
2795 .o_disconnect = osc_disconnect,
2796 .o_iocontrol = mdc_iocontrol,
2797 .o_set_info_async = mdc_set_info_async,
2798 .o_statfs = mdc_statfs,
2799 .o_fid_init = client_fid_init,
2800 .o_fid_fini = client_fid_fini,
2801 .o_fid_alloc = mdc_fid_alloc,
2802 .o_import_event = mdc_import_event,
2803 .o_get_info = mdc_get_info,
2804 .o_get_uuid = mdc_get_uuid,
2805 .o_quotactl = mdc_quotactl,
2808 static struct md_ops mdc_md_ops = {
2809 .m_get_root = mdc_get_root,
2810 .m_null_inode = mdc_null_inode,
2811 .m_close = mdc_close,
2812 .m_create = mdc_create,
2813 .m_enqueue = mdc_enqueue,
2814 .m_getattr = mdc_getattr,
2815 .m_getattr_name = mdc_getattr_name,
2816 .m_intent_lock = mdc_intent_lock,
2818 .m_rename = mdc_rename,
2819 .m_setattr = mdc_setattr,
2820 .m_setxattr = mdc_setxattr,
2821 .m_getxattr = mdc_getxattr,
2822 .m_fsync = mdc_fsync,
2823 .m_file_resync = mdc_file_resync,
2824 .m_read_page = mdc_read_page,
2825 .m_unlink = mdc_unlink,
2826 .m_cancel_unused = mdc_cancel_unused,
2827 .m_init_ea_size = mdc_init_ea_size,
2828 .m_set_lock_data = mdc_set_lock_data,
2829 .m_lock_match = mdc_lock_match,
2830 .m_get_lustre_md = mdc_get_lustre_md,
2831 .m_free_lustre_md = mdc_free_lustre_md,
2832 .m_set_open_replay_data = mdc_set_open_replay_data,
2833 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2834 .m_intent_getattr_async = mdc_intent_getattr_async,
2835 .m_revalidate_lock = mdc_revalidate_lock
2838 static int __init mdc_init(void)
2840 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2841 LUSTRE_MDC_NAME, &mdc_device_type);
2844 static void __exit mdc_exit(void)
2846 class_unregister_type(LUSTRE_MDC_NAME);
2849 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2850 MODULE_DESCRIPTION("Lustre Metadata Client");
2851 MODULE_VERSION(LUSTRE_VERSION_STRING);
2852 MODULE_LICENSE("GPL");
2854 module_init(mdc_init);
2855 module_exit(mdc_exit);