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 #include <linux/delay.h>
42 #include <linux/uidgid.h>
43 #include <linux/device.h>
44 #include <linux/xarray.h>
46 #include <lustre_errno.h>
48 #include <cl_object.h>
49 #include <llog_swab.h>
50 #include <lprocfs_status.h>
51 #include <lustre_acl.h>
52 #include <lustre_compat.h>
53 #include <lustre_fid.h>
54 #include <uapi/linux/lustre/lustre_ioctl.h>
55 #include <lustre_kernelcomm.h>
56 #include <lustre_lmv.h>
57 #include <lustre_log.h>
58 #include <lustre_swab.h>
59 #include <obd_class.h>
60 #include <lustre_osc.h>
62 #include "mdc_internal.h"
64 #define REQUEST_MINOR 244
66 static int mdc_cleanup(struct obd_device *obd);
68 static inline int mdc_queue_wait(struct ptlrpc_request *req)
70 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
73 /* obd_get_request_slot() ensures that this client has no more
74 * than cl_max_rpcs_in_flight RPCs simultaneously inf light
76 rc = obd_get_request_slot(cli);
80 rc = ptlrpc_queue_wait(req);
81 obd_put_request_slot(cli);
87 * Send MDS_GET_ROOT RPC to fetch root FID.
89 * If \a fileset is not NULL it should contain a subdirectory off
90 * the ROOT/ directory to be mounted on the client. Return the FID
91 * of the subdirectory to the client to mount onto its mountpoint.
93 * \param[in] imp MDC import
94 * \param[in] fileset fileset name, which could be NULL
95 * \param[out] rootfid root FID of this mountpoint
96 * \param[out] pc root capa will be unpacked and saved in this pointer
98 * \retval 0 on success, negative errno on failure
100 static int mdc_get_root(struct obd_export *exp, const char *fileset,
101 struct lu_fid *rootfid)
103 struct ptlrpc_request *req;
104 struct mdt_body *body;
109 if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
112 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
118 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
119 strlen(fileset) + 1);
120 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
122 ptlrpc_request_free(req);
125 mdc_pack_body(req, NULL, 0, 0, -1, 0);
126 if (fileset != NULL) {
127 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
129 memcpy(name, fileset, strlen(fileset));
131 lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
132 req->rq_send_state = LUSTRE_IMP_FULL;
134 ptlrpc_request_set_replen(req);
136 rc = ptlrpc_queue_wait(req);
140 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
142 GOTO(out, rc = -EPROTO);
144 *rootfid = body->mbo_fid1;
145 CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
146 PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
149 ptlrpc_req_finished(req);
155 * This function now is known to always saying that it will receive 4 buffers
156 * from server. Even for cases when acl_size and md_size is zero, RPC header
157 * will contain 4 fields and RPC itself will contain zero size fields. This is
158 * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
159 * and thus zero, it shrinks it, making zero size. The same story about
160 * md_size. And this is course of problem when client waits for smaller number
161 * of fields. This issue will be fixed later when client gets aware of RPC
164 static int mdc_getattr_common(struct obd_export *exp,
165 struct ptlrpc_request *req)
167 struct req_capsule *pill = &req->rq_pill;
168 struct mdt_body *body;
173 /* Request message already built. */
174 rc = ptlrpc_queue_wait(req);
178 /* sanity check for the reply */
179 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
183 CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
185 mdc_update_max_ea_from_body(exp, body);
186 if (body->mbo_eadatasize != 0) {
187 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
188 body->mbo_eadatasize);
196 static void mdc_reset_acl_req(struct ptlrpc_request *req)
198 spin_lock(&req->rq_early_free_lock);
199 sptlrpc_cli_free_repbuf(req);
200 req->rq_repbuf = NULL;
201 req->rq_repbuf_len = 0;
202 req->rq_repdata = NULL;
203 req->rq_reqdata_len = 0;
204 spin_unlock(&req->rq_early_free_lock);
207 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
208 struct ptlrpc_request **request)
210 struct ptlrpc_request *req;
211 struct obd_import *imp = class_exp2cliimp(exp);
212 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
216 /* Single MDS without an LMV case */
217 if (op_data->op_flags & MF_GET_MDT_IDX) {
223 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
227 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
229 ptlrpc_request_free(req);
234 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
235 op_data->op_mode, -1, 0);
236 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
237 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
239 ptlrpc_request_set_replen(req);
241 rc = mdc_getattr_common(exp, req);
244 acl_bufsize = min_t(__u32,
245 imp->imp_connect_data.ocd_max_easize,
247 mdc_reset_acl_req(req);
251 ptlrpc_req_finished(req);
259 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
260 struct ptlrpc_request **request)
262 struct ptlrpc_request *req;
263 struct obd_import *imp = class_exp2cliimp(exp);
264 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
269 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
273 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
274 op_data->op_namelen + 1);
276 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
278 ptlrpc_request_free(req);
282 if (op_data->op_name) {
283 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
284 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
285 op_data->op_namelen);
286 memcpy(name, op_data->op_name, op_data->op_namelen);
290 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
291 op_data->op_mode, op_data->op_suppgids[0], 0);
292 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
294 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
295 ptlrpc_request_set_replen(req);
297 rc = mdc_getattr_common(exp, req);
300 acl_bufsize = min_t(__u32,
301 imp->imp_connect_data.ocd_max_easize,
303 mdc_reset_acl_req(req);
307 ptlrpc_req_finished(req);
315 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
316 const struct lu_fid *fid, int opcode, u64 valid,
317 const char *xattr_name, const char *input,
318 int input_size, int output_size, int flags,
319 __u32 suppgid, struct ptlrpc_request **request)
321 struct ptlrpc_request *req;
322 int xattr_namelen = 0;
328 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
333 xattr_namelen = strlen(xattr_name) + 1;
334 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
339 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
342 /* get SELinux policy info if any */
343 rc = sptlrpc_get_sepol(req);
345 ptlrpc_request_free(req);
348 req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
349 strlen(req->rq_sepol) ?
350 strlen(req->rq_sepol) + 1 : 0);
352 /* Flush local XATTR locks to get rid of a possible cancel RPC */
353 if (opcode == MDS_REINT && fid_is_sane(fid) &&
354 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
358 /* Without that packing would fail */
360 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
363 count = mdc_resource_get_unused(exp, fid,
365 MDS_INODELOCK_XATTR);
367 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
369 ptlrpc_request_free(req);
373 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
375 ptlrpc_request_free(req);
380 if (opcode == MDS_REINT) {
381 struct mdt_rec_setxattr *rec;
383 BUILD_BUG_ON(sizeof(struct mdt_rec_setxattr) !=
384 sizeof(struct mdt_rec_reint));
385 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
386 rec->sx_opcode = REINT_SETXATTR;
387 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
388 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
389 rec->sx_cap = cfs_curproc_cap_pack();
390 rec->sx_suppgid1 = suppgid;
391 rec->sx_suppgid2 = -1;
393 rec->sx_valid = valid | OBD_MD_FLCTIME;
394 rec->sx_time = ktime_get_real_seconds();
395 rec->sx_size = output_size;
396 rec->sx_flags = flags;
398 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
402 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
403 memcpy(tmp, xattr_name, xattr_namelen);
406 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
407 memcpy(tmp, input, input_size);
410 mdc_file_sepol_pack(req);
412 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
413 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
414 RCL_SERVER, output_size);
415 ptlrpc_request_set_replen(req);
418 if (opcode == MDS_REINT)
419 ptlrpc_get_mod_rpc_slot(req);
421 rc = ptlrpc_queue_wait(req);
423 if (opcode == MDS_REINT)
424 ptlrpc_put_mod_rpc_slot(req);
427 ptlrpc_req_finished(req);
433 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
434 u64 obd_md_valid, const char *name,
435 const void *value, size_t value_size,
436 unsigned int xattr_flags, u32 suppgid,
437 struct ptlrpc_request **req)
439 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
440 obd_md_valid == OBD_MD_FLXATTRRM);
442 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
443 fid, MDS_REINT, obd_md_valid, name,
444 value, value_size, 0, xattr_flags, suppgid,
448 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
449 u64 obd_md_valid, const char *name, size_t buf_size,
450 struct ptlrpc_request **req)
452 struct mdt_body *body;
455 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
456 obd_md_valid == OBD_MD_FLXATTRLS);
458 /* The below message is checked in sanity-selinux.sh test_20d */
459 CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
460 exp->exp_obd->obd_name, name, PFID(fid));
461 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
462 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
467 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
469 GOTO(out, rc = -EPROTO);
471 /* only detect the xattr size */
473 /* LU-11109: Older MDTs do not distinguish
474 * between nonexistent xattrs and zero length
475 * values in this case. Newer MDTs will return
476 * -ENODATA or set OBD_MD_FLXATTR. */
477 GOTO(out, rc = body->mbo_eadatasize);
480 if (body->mbo_eadatasize == 0) {
481 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
482 * success so that we can distinguish between
483 * zero length value and nonexistent xattr.
485 * If OBD_MD_FLXATTR is not set then we keep
486 * the old behavior and return -ENODATA for
487 * getxattr() when mbo_eadatasize is 0. But
488 * -ENODATA only makes sense for getxattr()
489 * and not for listxattr(). */
490 if (body->mbo_valid & OBD_MD_FLXATTR)
492 else if (obd_md_valid == OBD_MD_FLXATTR)
493 GOTO(out, rc = -ENODATA);
498 GOTO(out, rc = body->mbo_eadatasize);
501 ptlrpc_req_finished(*req);
508 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
509 struct obd_export *dt_exp, struct obd_export *md_exp,
510 struct lustre_md *md)
512 struct req_capsule *pill = &req->rq_pill;
517 memset(md, 0, sizeof(*md));
519 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
520 LASSERT(md->body != NULL);
522 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
523 if (!S_ISREG(md->body->mbo_mode)) {
524 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
525 "regular file, but is not\n");
526 GOTO(out, rc = -EPROTO);
529 if (md->body->mbo_eadatasize == 0) {
530 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
531 "but eadatasize 0\n");
532 GOTO(out, rc = -EPROTO);
535 md->layout.lb_len = md->body->mbo_eadatasize;
536 md->layout.lb_buf = req_capsule_server_sized_get(pill,
539 if (md->layout.lb_buf == NULL)
540 GOTO(out, rc = -EPROTO);
541 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
542 const union lmv_mds_md *lmv;
545 if (!S_ISDIR(md->body->mbo_mode)) {
546 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
547 "directory, but is not\n");
548 GOTO(out, rc = -EPROTO);
551 if (md->body->mbo_valid & OBD_MD_MEA) {
552 lmv_size = md->body->mbo_eadatasize;
554 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
555 "but eadatasize 0\n");
559 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
562 GOTO(out, rc = -EPROTO);
564 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
568 if (rc < (int)sizeof(*md->lmv)) {
569 struct lmv_foreign_md *lfm = md->lfm;
571 /* short (< sizeof(struct lmv_stripe_md))
574 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
576 "lmv size too small: %d < %d\n",
577 rc, (int)sizeof(*md->lmv));
578 GOTO(out, rc = -EPROTO);
583 /* since 2.12.58 intent_getattr fetches default LMV */
584 if (md->body->mbo_valid & OBD_MD_DEFAULT_MEA) {
585 lmv_size = sizeof(struct lmv_user_md);
586 lmv = req_capsule_server_sized_get(pill,
590 GOTO(out, rc = -EPROTO);
592 rc = md_unpackmd(md_exp, &md->default_lmv, lmv,
597 if (rc < (int)sizeof(*md->default_lmv)) {
599 "default lmv size too small: %d < %d\n",
600 rc, (int)sizeof(*md->default_lmv));
601 GOTO(out, rc = -EPROTO);
607 if (md->body->mbo_valid & OBD_MD_FLACL) {
608 /* for ACL, it's possible that FLACL is set but aclsize is zero.
609 * only when aclsize != 0 there's an actual segment for ACL
612 rc = mdc_unpack_acl(req, md);
625 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
631 void mdc_replay_open(struct ptlrpc_request *req)
633 struct md_open_data *mod = req->rq_cb_data;
634 struct ptlrpc_request *close_req;
635 struct obd_client_handle *och;
636 struct lustre_handle old_open_handle = { };
637 struct mdt_body *body;
641 DEBUG_REQ(D_ERROR, req,
642 "cannot properly replay without open data");
647 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
648 LASSERT(body != NULL);
650 spin_lock(&req->rq_lock);
652 if (och && och->och_open_handle.cookie)
653 req->rq_early_free_repbuf = 1;
655 req->rq_early_free_repbuf = 0;
656 spin_unlock(&req->rq_lock);
658 if (req->rq_early_free_repbuf) {
659 struct lustre_handle *file_open_handle;
661 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
663 file_open_handle = &och->och_open_handle;
664 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
665 file_open_handle->cookie, body->mbo_open_handle.cookie);
666 old_open_handle = *file_open_handle;
667 *file_open_handle = body->mbo_open_handle;
670 close_req = mod->mod_close_req;
672 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
673 struct mdt_ioepoch *epoch;
675 LASSERT(opc == MDS_CLOSE);
676 epoch = req_capsule_client_get(&close_req->rq_pill,
680 if (req->rq_early_free_repbuf)
681 LASSERT(old_open_handle.cookie ==
682 epoch->mio_open_handle.cookie);
684 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
685 epoch->mio_open_handle = body->mbo_open_handle;
690 void mdc_commit_open(struct ptlrpc_request *req)
692 struct md_open_data *mod = req->rq_cb_data;
697 * No need to touch md_open_data::mod_och, it holds a reference on
698 * \var mod and will zero references to each other, \var mod will be
699 * freed after that when md_open_data::mod_och will put the reference.
703 * Do not let open request to disappear as it still may be needed
704 * for close rpc to happen (it may happen on evict only, otherwise
705 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
706 * called), just mark this rpc as committed to distinguish these 2
707 * cases, see mdc_close() for details. The open request reference will
708 * be put along with freeing \var mod.
710 ptlrpc_request_addref(req);
711 spin_lock(&req->rq_lock);
712 req->rq_committed = 1;
713 spin_unlock(&req->rq_lock);
714 req->rq_cb_data = NULL;
718 int mdc_set_open_replay_data(struct obd_export *exp,
719 struct obd_client_handle *och,
720 struct lookup_intent *it)
722 struct md_open_data *mod;
723 struct mdt_rec_create *rec;
724 struct mdt_body *body;
725 struct ptlrpc_request *open_req = it->it_request;
726 struct obd_import *imp = open_req->rq_import;
729 if (!open_req->rq_replay)
732 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
733 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
734 LASSERT(rec != NULL);
735 /* Incoming message in my byte order (it's been swabbed). */
736 /* Outgoing messages always in my byte order. */
737 LASSERT(body != NULL);
739 /* Only if the import is replayable, we set replay_open data */
740 if (och && imp->imp_replayable) {
741 mod = obd_mod_alloc();
743 DEBUG_REQ(D_ERROR, open_req,
744 "cannot allocate md_open_data");
749 * Take a reference on \var mod, to be freed on mdc_close().
750 * It protects \var mod from being freed on eviction (commit
751 * callback is called despite rq_replay flag).
752 * Another reference for \var och.
757 spin_lock(&open_req->rq_lock);
760 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
761 it_disposition(it, DISP_OPEN_STRIPE);
762 mod->mod_open_req = open_req;
763 open_req->rq_cb_data = mod;
764 open_req->rq_commit_cb = mdc_commit_open;
765 open_req->rq_early_free_repbuf = 1;
766 spin_unlock(&open_req->rq_lock);
769 rec->cr_fid2 = body->mbo_fid1;
770 rec->cr_open_handle_old = body->mbo_open_handle;
771 open_req->rq_replay_cb = mdc_replay_open;
772 if (!fid_is_sane(&body->mbo_fid1)) {
773 DEBUG_REQ(D_ERROR, open_req,
774 "saving replay request with insane FID " DFID,
775 PFID(&body->mbo_fid1));
779 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
783 static void mdc_free_open(struct md_open_data *mod)
787 if (mod->mod_is_create == 0 &&
788 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
792 * No reason to asssert here if the open request has
793 * rq_replay == 1. It means that mdc_close failed, and
794 * close request wasn`t sent. It is not fatal to client.
795 * The worst thing is eviction if the client gets open lock
798 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req,
799 "free open request, rq_replay=%d",
800 mod->mod_open_req->rq_replay);
802 ptlrpc_request_committed(mod->mod_open_req, committed);
803 if (mod->mod_close_req)
804 ptlrpc_request_committed(mod->mod_close_req, committed);
807 int mdc_clear_open_replay_data(struct obd_export *exp,
808 struct obd_client_handle *och)
810 struct md_open_data *mod = och->och_mod;
814 * It is possible to not have \var mod in a case of eviction between
815 * lookup and ll_file_open().
820 LASSERT(mod != LP_POISON);
821 LASSERT(mod->mod_open_req != NULL);
823 spin_lock(&mod->mod_open_req->rq_lock);
825 mod->mod_och->och_open_handle.cookie = 0;
826 mod->mod_open_req->rq_early_free_repbuf = 0;
827 spin_unlock(&mod->mod_open_req->rq_lock);
837 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
838 struct md_open_data *mod, struct ptlrpc_request **request)
840 struct obd_device *obd = class_exp2obd(exp);
841 struct ptlrpc_request *req;
842 struct req_format *req_fmt;
843 size_t u32_count = 0;
848 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
849 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
852 if (op_data->op_bias & MDS_CLOSE_INTENT) {
853 req_fmt = &RQF_MDS_CLOSE_INTENT;
854 if (op_data->op_bias & MDS_HSM_RELEASE) {
855 /* allocate a FID for volatile file */
856 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
859 CERROR("%s: "DFID" allocating FID: rc = %d\n",
860 obd->obd_name, PFID(&op_data->op_fid1),
862 /* save the errcode and proceed to close */
866 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
867 size_t count = op_data->op_data_size / sizeof(__u32);
869 if (count > INLINE_RESYNC_ARRAY_SIZE)
873 req_fmt = &RQF_MDS_CLOSE;
877 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
880 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
882 /* Ensure that this close's handle is fixed up during replay. */
883 if (likely(mod != NULL)) {
884 LASSERTF(mod->mod_open_req != NULL &&
885 mod->mod_open_req->rq_type != LI_POISON,
886 "POISONED open %p!\n", mod->mod_open_req);
888 mod->mod_close_req = req;
890 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "matched open");
891 /* We no longer want to preserve this open for replay even
892 * though the open was committed. b=3632, b=3633 */
893 spin_lock(&mod->mod_open_req->rq_lock);
894 mod->mod_open_req->rq_replay = 0;
895 spin_unlock(&mod->mod_open_req->rq_lock);
897 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
901 * TODO: repeat close after errors
903 CWARN("%s: close of FID "DFID" failed, file reference will be "
904 "dropped when this client unmounts or is evicted\n",
905 obd->obd_name, PFID(&op_data->op_fid1));
906 GOTO(out, rc = -ENOMEM);
910 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
911 u32_count * sizeof(__u32));
913 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
915 ptlrpc_request_free(req);
920 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
921 * portal whose threads are not taking any DLM locks and are therefore
922 * always progressing */
923 req->rq_request_portal = MDS_READPAGE_PORTAL;
924 ptlrpc_at_set_req_timeout(req);
926 if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
927 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
928 OP_XVALID_LAZYBLOCKS);
930 mdc_close_pack(req, op_data);
932 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
933 obd->u.cli.cl_default_mds_easize);
935 ptlrpc_request_set_replen(req);
937 ptlrpc_get_mod_rpc_slot(req);
938 rc = ptlrpc_queue_wait(req);
939 ptlrpc_put_mod_rpc_slot(req);
941 if (req->rq_repmsg == NULL) {
942 CDEBUG(D_RPCTRACE, "request %p failed to send: rc = %d\n", req,
945 rc = req->rq_status ?: -EIO;
946 } else if (rc == 0 || rc == -EAGAIN) {
947 struct mdt_body *body;
949 rc = lustre_msg_get_status(req->rq_repmsg);
950 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
951 DEBUG_REQ(D_ERROR, req,
952 "type = PTL_RPC_MSG_ERR: rc = %d", rc);
956 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
959 } else if (rc == -ESTALE) {
961 * it can be allowed error after 3633 if open was committed and
962 * server failed before close was sent. Let's check if mod
963 * exists and return no error in that case
966 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
967 LASSERT(mod->mod_open_req != NULL);
968 if (mod->mod_open_req->rq_committed)
976 mod->mod_close_req = NULL;
977 /* Since now, mod is accessed through open_req only,
978 * thus close req does not keep a reference on mod anymore. */
983 RETURN(rc < 0 ? rc : saved_rc);
986 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
987 u64 offset, struct page **pages, int npages,
988 struct ptlrpc_request **request)
990 struct ptlrpc_request *req;
991 struct ptlrpc_bulk_desc *desc;
1000 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1004 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1006 ptlrpc_request_free(req);
1010 req->rq_request_portal = MDS_READPAGE_PORTAL;
1011 ptlrpc_at_set_req_timeout(req);
1013 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1014 PTLRPC_BULK_PUT_SINK,
1016 &ptlrpc_bulk_kiov_pin_ops);
1018 ptlrpc_req_finished(req);
1022 /* NB req now owns desc and will free it when it gets freed */
1023 for (i = 0; i < npages; i++)
1024 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1027 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1029 ptlrpc_request_set_replen(req);
1030 rc = ptlrpc_queue_wait(req);
1032 ptlrpc_req_finished(req);
1033 if (rc != -ETIMEDOUT)
1037 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1038 CERROR("%s: too many resend retries: rc = %d\n",
1039 exp->exp_obd->obd_name, -EIO);
1043 /* If a signal interrupts then the timeout returned will
1044 * not be zero. In that case return -EINTR
1046 if (msleep_interruptible(resends * 1000))
1052 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1053 req->rq_bulk->bd_nob_transferred);
1055 ptlrpc_req_finished(req);
1059 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1060 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1061 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1062 PAGE_SIZE * npages);
1063 ptlrpc_req_finished(req);
1071 static void mdc_release_page(struct page *page, int remove)
1075 if (likely(page->mapping != NULL))
1076 delete_from_page_cache(page);
1082 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1083 __u64 *start, __u64 *end, int hash64)
1086 * Complement of hash is used as an index so that
1087 * radix_tree_gang_lookup() can be used to find a page with starting
1088 * hash _smaller_ than one we are looking for.
1090 unsigned long offset = hash_x_index(*hash, hash64);
1092 unsigned long flags;
1095 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1096 found = radix_tree_gang_lookup(&mapping->page_tree,
1097 (void **)&page, offset, 1);
1098 if (found > 0 && !ll_xa_is_value(page)) {
1099 struct lu_dirpage *dp;
1102 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1104 * In contrast to find_lock_page() we are sure that directory
1105 * page cannot be truncated (while DLM lock is held) and,
1106 * hence, can avoid restart.
1108 * In fact, page cannot be locked here at all, because
1109 * mdc_read_page_remote does synchronous io.
1111 wait_on_page_locked(page);
1112 if (PageUptodate(page)) {
1114 if (BITS_PER_LONG == 32 && hash64) {
1115 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1116 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1117 *hash = *hash >> 32;
1119 *start = le64_to_cpu(dp->ldp_hash_start);
1120 *end = le64_to_cpu(dp->ldp_hash_end);
1122 if (unlikely(*start == 1 && *hash == 0))
1125 LASSERTF(*start <= *hash, "start = %#llx"
1126 ",end = %#llx,hash = %#llx\n",
1127 *start, *end, *hash);
1128 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1129 " hash %#llx\n", offset, *start, *end, *hash);
1132 mdc_release_page(page, 0);
1134 } else if (*end != *start && *hash == *end) {
1136 * upon hash collision, remove this page,
1137 * otherwise put page reference, and
1138 * mdc_read_page_remote() will issue RPC to
1139 * fetch the page we want.
1142 mdc_release_page(page,
1143 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1148 page = ERR_PTR(-EIO);
1151 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1158 * Adjust a set of pages, each page containing an array of lu_dirpages,
1159 * so that each page can be used as a single logical lu_dirpage.
1161 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1162 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1163 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1164 * value is used as a cookie to request the next lu_dirpage in a
1165 * directory listing that spans multiple pages (two in this example):
1168 * .|--------v------- -----.
1169 * |s|e|f|p|ent|ent| ... |ent|
1170 * '--|-------------- -----' Each PAGE contains a single
1171 * '------. lu_dirpage.
1172 * .---------v------- -----.
1173 * |s|e|f|p|ent| 0 | ... | 0 |
1174 * '----------------- -----'
1176 * However, on hosts where the native VM page size (PAGE_SIZE) is
1177 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1178 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1179 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1180 * after it in the same PAGE (arrows simplified for brevity, but
1181 * in general e0==s1, e1==s2, etc.):
1183 * .-------------------- -----.
1184 * |s0|e0|f0|p|ent|ent| ... |ent|
1185 * |---v---------------- -----|
1186 * |s1|e1|f1|p|ent|ent| ... |ent|
1187 * |---v---------------- -----| Here, each PAGE contains
1188 * ... multiple lu_dirpages.
1189 * |---v---------------- -----|
1190 * |s'|e'|f'|p|ent|ent| ... |ent|
1191 * '---|---------------- -----'
1193 * .----------------------------.
1196 * This structure is transformed into a single logical lu_dirpage as follows:
1198 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1199 * labeled 'next PAGE'.
1201 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1202 * a hash collision with the next page exists.
1204 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1205 * to the first entry of the next lu_dirpage.
1207 #if PAGE_SIZE > LU_PAGE_SIZE
1208 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1212 for (i = 0; i < cfs_pgs; i++) {
1213 struct lu_dirpage *dp = kmap(pages[i]);
1214 struct lu_dirpage *first = dp;
1215 struct lu_dirent *end_dirent = NULL;
1216 struct lu_dirent *ent;
1217 __u64 hash_end = dp->ldp_hash_end;
1218 __u32 flags = dp->ldp_flags;
1220 while (--lu_pgs > 0) {
1221 ent = lu_dirent_start(dp);
1222 for (end_dirent = ent; ent != NULL;
1223 end_dirent = ent, ent = lu_dirent_next(ent));
1225 /* Advance dp to next lu_dirpage. */
1226 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1228 /* Check if we've reached the end of the PAGE. */
1229 if (!((unsigned long)dp & ~PAGE_MASK))
1232 /* Save the hash and flags of this lu_dirpage. */
1233 hash_end = dp->ldp_hash_end;
1234 flags = dp->ldp_flags;
1236 /* Check if lu_dirpage contains no entries. */
1237 if (end_dirent == NULL)
1240 /* Enlarge the end entry lde_reclen from 0 to
1241 * first entry of next lu_dirpage. */
1242 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1243 end_dirent->lde_reclen =
1244 cpu_to_le16((char *)(dp->ldp_entries) -
1245 (char *)end_dirent);
1248 first->ldp_hash_end = hash_end;
1249 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1250 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1254 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1257 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1258 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1260 /* parameters for readdir page */
1261 struct readpage_param {
1262 struct md_op_data *rp_mod;
1265 struct obd_export *rp_exp;
1266 struct md_callback *rp_cb;
1270 * Read pages from server.
1272 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1273 * a header lu_dirpage which describes the start/end hash, and whether this
1274 * page is empty (contains no dir entry) or hash collide with next page.
1275 * After client receives reply, several pages will be integrated into dir page
1276 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1277 * lu_dirpage for this integrated page will be adjusted.
1279 static int mdc_read_page_remote(void *data, struct page *page0)
1281 struct readpage_param *rp = data;
1282 struct page **page_pool;
1284 struct lu_dirpage *dp;
1285 struct md_op_data *op_data = rp->rp_mod;
1286 struct ptlrpc_request *req;
1288 struct inode *inode;
1290 int rd_pgs = 0; /* number of pages actually read */
1296 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1297 inode = op_data->op_data;
1298 fid = &op_data->op_fid1;
1299 LASSERT(inode != NULL);
1301 OBD_ALLOC_PTR_ARRAY(page_pool, max_pages);
1302 if (page_pool != NULL) {
1303 page_pool[0] = page0;
1309 for (npages = 1; npages < max_pages; npages++) {
1310 page = page_cache_alloc(inode->i_mapping);
1313 page_pool[npages] = page;
1316 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1318 /* page0 is special, which was added into page cache early */
1319 delete_from_page_cache(page0);
1323 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1325 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1326 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1328 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1330 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1332 SetPageUptodate(page0);
1336 ptlrpc_req_finished(req);
1337 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1338 for (i = 1; i < npages; i++) {
1339 unsigned long offset;
1343 page = page_pool[i];
1345 if (rc < 0 || i >= rd_pgs) {
1350 SetPageUptodate(page);
1353 hash = le64_to_cpu(dp->ldp_hash_start);
1356 offset = hash_x_index(hash, rp->rp_hash64);
1358 prefetchw(&page->flags);
1359 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1364 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1365 " rc = %d\n", offset, ret);
1369 if (page_pool != &page0)
1370 OBD_FREE_PTR_ARRAY(page_pool, max_pages);
1376 * Read dir page from cache first, if it can not find it, read it from
1377 * server and add into the cache.
1379 * \param[in] exp MDC export
1380 * \param[in] op_data client MD stack parameters, transfering parameters
1381 * between different layers on client MD stack.
1382 * \param[in] cb_op callback required for ldlm lock enqueue during
1384 * \param[in] hash_offset the hash offset of the page to be read
1385 * \param[in] ppage the page to be read
1387 * retval = 0 get the page successfully
1388 * errno(<0) get the page failed
1390 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1391 struct md_callback *cb_op, __u64 hash_offset,
1392 struct page **ppage)
1394 struct lookup_intent it = { .it_op = IT_READDIR };
1396 struct inode *dir = op_data->op_data;
1397 struct address_space *mapping;
1398 struct lu_dirpage *dp;
1401 struct lustre_handle lockh;
1402 struct ptlrpc_request *enq_req = NULL;
1403 struct readpage_param rp_param;
1410 LASSERT(dir != NULL);
1411 mapping = dir->i_mapping;
1413 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1414 cb_op->md_blocking_ast, 0);
1415 if (enq_req != NULL)
1416 ptlrpc_req_finished(enq_req);
1419 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1420 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1425 lockh.cookie = it.it_lock_handle;
1426 mdc_set_lock_data(exp, &lockh, dir, NULL);
1428 rp_param.rp_off = hash_offset;
1429 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1430 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1431 rp_param.rp_hash64);
1433 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1434 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1435 rp_param.rp_off, PTR_ERR(page));
1436 GOTO(out_unlock, rc = PTR_ERR(page));
1437 } else if (page != NULL) {
1439 * XXX nikita: not entirely correct handling of a corner case:
1440 * suppose hash chain of entries with hash value HASH crosses
1441 * border between pages P0 and P1. First both P0 and P1 are
1442 * cached, seekdir() is called for some entry from the P0 part
1443 * of the chain. Later P0 goes out of cache. telldir(HASH)
1444 * happens and finds P1, as it starts with matching hash
1445 * value. Remaining entries from P0 part of the chain are
1446 * skipped. (Is that really a bug?)
1448 * Possible solutions: 0. don't cache P1 is such case, handle
1449 * it as an "overflow" page. 1. invalidate all pages at
1450 * once. 2. use HASH|1 as an index for P1.
1452 GOTO(hash_collision, page);
1455 rp_param.rp_exp = exp;
1456 rp_param.rp_mod = op_data;
1457 page = read_cache_page(mapping,
1458 hash_x_index(rp_param.rp_off,
1459 rp_param.rp_hash64),
1460 mdc_read_page_remote, &rp_param);
1462 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1463 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1464 rp_param.rp_off, PTR_ERR(page));
1465 GOTO(out_unlock, rc = PTR_ERR(page));
1468 wait_on_page_locked(page);
1470 if (!PageUptodate(page)) {
1471 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1472 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1473 rp_param.rp_off, -5);
1476 if (!PageChecked(page))
1477 SetPageChecked(page);
1478 if (PageError(page)) {
1479 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1480 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1481 rp_param.rp_off, -5);
1486 dp = page_address(page);
1487 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1488 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1489 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1490 rp_param.rp_off = hash_offset >> 32;
1492 start = le64_to_cpu(dp->ldp_hash_start);
1493 end = le64_to_cpu(dp->ldp_hash_end);
1494 rp_param.rp_off = hash_offset;
1497 LASSERT(start == rp_param.rp_off);
1498 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1499 #if BITS_PER_LONG == 32
1500 CWARN("Real page-wide hash collision at [%llu %llu] with "
1501 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1502 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1506 * Fetch whole overflow chain...
1514 ldlm_lock_decref(&lockh, it.it_lock_mode);
1518 mdc_release_page(page, 1);
1523 static int mdc_statfs_interpret(const struct lu_env *env,
1524 struct ptlrpc_request *req, void *args, int rc)
1526 struct obd_info *oinfo = args;
1527 struct obd_statfs *osfs;
1530 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1534 oinfo->oi_osfs = osfs;
1536 CDEBUG(D_CACHE, "blocks=%llu free=%llu avail=%llu "
1537 "objects=%llu free=%llu state=%x\n",
1538 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
1539 osfs->os_files, osfs->os_ffree, osfs->os_state);
1542 oinfo->oi_cb_up(oinfo, rc);
1547 static int mdc_statfs_async(struct obd_export *exp,
1548 struct obd_info *oinfo, time64_t max_age,
1549 struct ptlrpc_request_set *unused)
1551 struct ptlrpc_request *req;
1552 struct obd_info *aa;
1554 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_STATFS,
1555 LUSTRE_MDS_VERSION, MDS_STATFS);
1559 ptlrpc_request_set_replen(req);
1560 req->rq_interpret_reply = mdc_statfs_interpret;
1562 aa = ptlrpc_req_async_args(aa, req);
1565 ptlrpcd_add_req(req);
1570 static int mdc_statfs(const struct lu_env *env,
1571 struct obd_export *exp, struct obd_statfs *osfs,
1572 time64_t max_age, __u32 flags)
1574 struct obd_device *obd = class_exp2obd(exp);
1575 struct req_format *fmt;
1576 struct ptlrpc_request *req;
1577 struct obd_statfs *msfs;
1578 struct obd_import *imp = NULL;
1583 * Since the request might also come from lprocfs, so we need
1584 * sync this with client_disconnect_export Bug15684
1586 down_read(&obd->u.cli.cl_sem);
1587 if (obd->u.cli.cl_import)
1588 imp = class_import_get(obd->u.cli.cl_import);
1589 up_read(&obd->u.cli.cl_sem);
1593 fmt = &RQF_MDS_STATFS;
1594 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1595 (flags & OBD_STATFS_SUM))
1596 fmt = &RQF_MDS_STATFS_NEW;
1597 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1600 GOTO(output, rc = -ENOMEM);
1602 if ((flags & OBD_STATFS_SUM) &&
1603 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1604 /* request aggregated states */
1605 struct mdt_body *body;
1607 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1609 GOTO(out, rc = -EPROTO);
1610 body->mbo_valid = OBD_MD_FLAGSTATFS;
1613 ptlrpc_request_set_replen(req);
1615 if (flags & OBD_STATFS_NODELAY) {
1616 /* procfs requests not want stay in wait for avoid deadlock */
1617 req->rq_no_resend = 1;
1618 req->rq_no_delay = 1;
1621 rc = ptlrpc_queue_wait(req);
1623 /* check connection error first */
1624 if (imp->imp_connect_error)
1625 rc = imp->imp_connect_error;
1629 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1631 GOTO(out, rc = -EPROTO);
1636 ptlrpc_req_finished(req);
1638 class_import_put(imp);
1642 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1644 __u32 keylen, vallen;
1648 if (gf->gf_pathlen > PATH_MAX)
1649 RETURN(-ENAMETOOLONG);
1650 if (gf->gf_pathlen < 2)
1653 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1654 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1655 sizeof(struct lu_fid));
1656 OBD_ALLOC(key, keylen);
1659 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1660 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1661 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1662 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1663 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1664 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1666 if (!fid_is_sane(&gf->gf_fid))
1667 GOTO(out, rc = -EINVAL);
1669 /* Val is struct getinfo_fid2path result plus path */
1670 vallen = sizeof(*gf) + gf->gf_pathlen;
1672 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1673 if (rc != 0 && rc != -EREMOTE)
1676 if (vallen <= sizeof(*gf))
1677 GOTO(out, rc = -EPROTO);
1678 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1679 GOTO(out, rc = -EOVERFLOW);
1681 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1682 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1683 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1684 /* only log the last 512 characters of the path */
1685 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1688 OBD_FREE(key, keylen);
1692 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1693 struct hsm_progress_kernel *hpk)
1695 struct obd_import *imp = class_exp2cliimp(exp);
1696 struct hsm_progress_kernel *req_hpk;
1697 struct ptlrpc_request *req;
1701 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1702 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1704 GOTO(out, rc = -ENOMEM);
1706 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1708 /* Copy hsm_progress struct */
1709 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1710 if (req_hpk == NULL)
1711 GOTO(out, rc = -EPROTO);
1714 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1716 ptlrpc_request_set_replen(req);
1718 ptlrpc_get_mod_rpc_slot(req);
1719 rc = ptlrpc_queue_wait(req);
1720 ptlrpc_put_mod_rpc_slot(req);
1724 ptlrpc_req_finished(req);
1728 * Send hsm_ct_register to MDS
1730 * \param[in] imp import
1731 * \param[in] archive_count if in bitmap format, it is the bitmap,
1732 * else it is the count of archive_ids
1733 * \param[in] archives if in bitmap format, it is NULL,
1734 * else it is archive_id lists
1736 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1739 struct ptlrpc_request *req;
1740 __u32 *archive_array;
1741 size_t archives_size;
1745 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1749 if (archives != NULL)
1750 archives_size = sizeof(*archive_array) * archive_count;
1752 archives_size = sizeof(archive_count);
1754 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1755 RCL_CLIENT, archives_size);
1757 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1759 ptlrpc_request_free(req);
1763 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1765 archive_array = req_capsule_client_get(&req->rq_pill,
1766 &RMF_MDS_HSM_ARCHIVE);
1767 if (archive_array == NULL)
1768 GOTO(out, rc = -EPROTO);
1770 if (archives != NULL)
1771 memcpy(archive_array, archives, archives_size);
1773 *archive_array = archive_count;
1775 ptlrpc_request_set_replen(req);
1776 req->rq_no_resend = 1;
1778 rc = mdc_queue_wait(req);
1781 ptlrpc_req_finished(req);
1785 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1786 struct md_op_data *op_data)
1788 struct hsm_current_action *hca = op_data->op_data;
1789 struct hsm_current_action *req_hca;
1790 struct ptlrpc_request *req;
1794 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1795 &RQF_MDS_HSM_ACTION);
1799 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1801 ptlrpc_request_free(req);
1805 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1806 op_data->op_suppgids[0], 0);
1808 ptlrpc_request_set_replen(req);
1810 rc = mdc_queue_wait(req);
1814 req_hca = req_capsule_server_get(&req->rq_pill,
1815 &RMF_MDS_HSM_CURRENT_ACTION);
1816 if (req_hca == NULL)
1817 GOTO(out, rc = -EPROTO);
1823 ptlrpc_req_finished(req);
1827 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1829 struct ptlrpc_request *req;
1833 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1835 MDS_HSM_CT_UNREGISTER);
1837 GOTO(out, rc = -ENOMEM);
1839 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1841 ptlrpc_request_set_replen(req);
1843 rc = mdc_queue_wait(req);
1846 ptlrpc_req_finished(req);
1850 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1851 struct md_op_data *op_data)
1853 struct hsm_user_state *hus = op_data->op_data;
1854 struct hsm_user_state *req_hus;
1855 struct ptlrpc_request *req;
1859 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1860 &RQF_MDS_HSM_STATE_GET);
1864 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1866 ptlrpc_request_free(req);
1870 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1871 op_data->op_suppgids[0], 0);
1873 ptlrpc_request_set_replen(req);
1875 rc = mdc_queue_wait(req);
1879 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1880 if (req_hus == NULL)
1881 GOTO(out, rc = -EPROTO);
1887 ptlrpc_req_finished(req);
1891 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1892 struct md_op_data *op_data)
1894 struct hsm_state_set *hss = op_data->op_data;
1895 struct hsm_state_set *req_hss;
1896 struct ptlrpc_request *req;
1900 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1901 &RQF_MDS_HSM_STATE_SET);
1905 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1907 ptlrpc_request_free(req);
1911 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1912 op_data->op_suppgids[0], 0);
1915 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1916 if (req_hss == NULL)
1917 GOTO(out, rc = -EPROTO);
1920 ptlrpc_request_set_replen(req);
1922 ptlrpc_get_mod_rpc_slot(req);
1923 rc = ptlrpc_queue_wait(req);
1924 ptlrpc_put_mod_rpc_slot(req);
1928 ptlrpc_req_finished(req);
1932 static int mdc_ioc_hsm_request(struct obd_export *exp,
1933 struct hsm_user_request *hur)
1935 struct obd_import *imp = class_exp2cliimp(exp);
1936 struct ptlrpc_request *req;
1937 struct hsm_request *req_hr;
1938 struct hsm_user_item *req_hui;
1943 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1945 GOTO(out, rc = -ENOMEM);
1947 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1948 hur->hur_request.hr_itemcount
1949 * sizeof(struct hsm_user_item));
1950 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1951 hur->hur_request.hr_data_len);
1953 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1955 ptlrpc_request_free(req);
1959 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1961 /* Copy hsm_request struct */
1962 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1964 GOTO(out, rc = -EPROTO);
1965 *req_hr = hur->hur_request;
1967 /* Copy hsm_user_item structs */
1968 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1969 if (req_hui == NULL)
1970 GOTO(out, rc = -EPROTO);
1971 memcpy(req_hui, hur->hur_user_item,
1972 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1974 /* Copy opaque field */
1975 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1976 if (req_opaque == NULL)
1977 GOTO(out, rc = -EPROTO);
1978 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1980 ptlrpc_request_set_replen(req);
1982 ptlrpc_get_mod_rpc_slot(req);
1983 rc = ptlrpc_queue_wait(req);
1984 ptlrpc_put_mod_rpc_slot(req);
1989 ptlrpc_req_finished(req);
1993 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
1994 struct lustre_kernelcomm *lk);
1996 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
1997 struct obd_quotactl *oqctl)
1999 struct ptlrpc_request *req;
2000 struct obd_quotactl *oqc;
2004 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_QUOTACTL);
2009 if (LUSTRE_Q_CMD_IS_POOL(oqctl->qc_cmd))
2010 req_capsule_set_size(&req->rq_pill,
2013 sizeof(*oqc) + LOV_MAXPOOLNAME + 1);
2015 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION,
2018 ptlrpc_request_free(req);
2022 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2023 QCTL_COPY(oqc, oqctl);
2025 ptlrpc_request_set_replen(req);
2026 ptlrpc_at_set_req_timeout(req);
2028 rc = ptlrpc_queue_wait(req);
2030 CERROR("%s: ptlrpc_queue_wait failed: rc = %d\n",
2031 exp->exp_obd->obd_name, rc);
2035 if (req->rq_repmsg &&
2036 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2037 QCTL_COPY(oqctl, oqc);
2040 CERROR("%s: cannot unpack obd_quotactl: rc = %d\n",
2041 exp->exp_obd->obd_name, rc);
2044 ptlrpc_req_finished(req);
2049 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2050 struct md_op_data *op_data)
2053 struct ptlrpc_request *req;
2055 struct mdc_swap_layouts *msl, *payload;
2058 msl = op_data->op_data;
2060 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2061 * first thing it will do is to cancel the 2 layout
2062 * locks held by this client.
2063 * So the client must cancel its layout locks on the 2 fids
2064 * with the request RPC to avoid extra RPC round trips.
2066 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2067 LCK_EX, MDS_INODELOCK_LAYOUT |
2068 MDS_INODELOCK_XATTR);
2069 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2070 LCK_EX, MDS_INODELOCK_LAYOUT |
2071 MDS_INODELOCK_XATTR);
2073 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2074 &RQF_MDS_SWAP_LAYOUTS);
2076 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2080 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2082 ptlrpc_request_free(req);
2086 mdc_swap_layouts_pack(req, op_data);
2088 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2093 ptlrpc_request_set_replen(req);
2095 rc = ptlrpc_queue_wait(req);
2101 ptlrpc_req_finished(req);
2105 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2106 void *karg, void __user *uarg)
2108 struct obd_device *obd = exp->exp_obd;
2109 struct obd_ioctl_data *data = karg;
2110 struct obd_import *imp = obd->u.cli.cl_import;
2114 if (!try_module_get(THIS_MODULE)) {
2115 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2116 module_name(THIS_MODULE));
2120 case OBD_IOC_FID2PATH:
2121 rc = mdc_ioc_fid2path(exp, karg);
2123 case LL_IOC_HSM_CT_START:
2124 rc = mdc_ioc_hsm_ct_start(exp, karg);
2125 /* ignore if it was already registered on this MDS. */
2129 case LL_IOC_HSM_PROGRESS:
2130 rc = mdc_ioc_hsm_progress(exp, karg);
2132 case LL_IOC_HSM_STATE_GET:
2133 rc = mdc_ioc_hsm_state_get(exp, karg);
2135 case LL_IOC_HSM_STATE_SET:
2136 rc = mdc_ioc_hsm_state_set(exp, karg);
2138 case LL_IOC_HSM_ACTION:
2139 rc = mdc_ioc_hsm_current_action(exp, karg);
2141 case LL_IOC_HSM_REQUEST:
2142 rc = mdc_ioc_hsm_request(exp, karg);
2144 case OBD_IOC_CLIENT_RECOVER:
2145 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2149 case IOC_OSC_SET_ACTIVE:
2150 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2153 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2154 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2155 * there'd be no LMV layer thus we might be called here. Eventually
2156 * this code should be removed.
2159 case IOC_OBD_STATFS: {
2160 struct obd_statfs stat_buf = {0};
2162 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2163 GOTO(out, rc = -ENODEV);
2166 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2167 min((int)data->ioc_plen2,
2168 (int)sizeof(struct obd_uuid))))
2169 GOTO(out, rc = -EFAULT);
2171 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2172 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2177 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2178 min((int) data->ioc_plen1,
2179 (int) sizeof(stat_buf))))
2180 GOTO(out, rc = -EFAULT);
2184 case OBD_IOC_QUOTACTL: {
2185 struct if_quotactl *qctl = karg;
2186 struct obd_quotactl *oqctl;
2188 OBD_ALLOC_PTR(oqctl);
2190 GOTO(out, rc = -ENOMEM);
2192 QCTL_COPY(oqctl, qctl);
2193 rc = obd_quotactl(exp, oqctl);
2195 QCTL_COPY(qctl, oqctl);
2196 qctl->qc_valid = QC_MDTIDX;
2197 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2200 OBD_FREE_PTR(oqctl);
2203 case LL_IOC_GET_CONNECT_FLAGS:
2204 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2205 sizeof(*exp_connect_flags_ptr(exp))))
2206 GOTO(out, rc = -EFAULT);
2209 case LL_IOC_LOV_SWAP_LAYOUTS:
2210 rc = mdc_ioc_swap_layouts(exp, karg);
2213 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2214 GOTO(out, rc = -ENOTTY);
2217 module_put(THIS_MODULE);
2222 static int mdc_get_info_rpc(struct obd_export *exp,
2223 u32 keylen, void *key,
2224 u32 vallen, void *val)
2226 struct obd_import *imp = class_exp2cliimp(exp);
2227 struct ptlrpc_request *req;
2232 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2236 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2237 RCL_CLIENT, keylen);
2238 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2239 RCL_CLIENT, sizeof(vallen));
2241 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2243 ptlrpc_request_free(req);
2247 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2248 memcpy(tmp, key, keylen);
2249 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2250 memcpy(tmp, &vallen, sizeof(vallen));
2252 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2253 RCL_SERVER, vallen);
2254 ptlrpc_request_set_replen(req);
2256 rc = ptlrpc_queue_wait(req);
2257 /* -EREMOTE means the get_info result is partial, and it needs to
2258 * continue on another MDT, see fid2path part in lmv_iocontrol */
2259 if (rc == 0 || rc == -EREMOTE) {
2260 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2261 memcpy(val, tmp, vallen);
2262 if (ptlrpc_rep_need_swab(req)) {
2263 if (KEY_IS(KEY_FID2PATH))
2264 lustre_swab_fid2path(val);
2267 ptlrpc_req_finished(req);
2272 static void lustre_swab_hai(struct hsm_action_item *h)
2274 __swab32s(&h->hai_len);
2275 __swab32s(&h->hai_action);
2276 lustre_swab_lu_fid(&h->hai_fid);
2277 lustre_swab_lu_fid(&h->hai_dfid);
2278 __swab64s(&h->hai_cookie);
2279 __swab64s(&h->hai_extent.offset);
2280 __swab64s(&h->hai_extent.length);
2281 __swab64s(&h->hai_gid);
2284 static void lustre_swab_hal(struct hsm_action_list *h)
2286 struct hsm_action_item *hai;
2289 __swab32s(&h->hal_version);
2290 __swab32s(&h->hal_count);
2291 __swab32s(&h->hal_archive_id);
2292 __swab64s(&h->hal_flags);
2294 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2295 lustre_swab_hai(hai);
2298 static void lustre_swab_kuch(struct kuc_hdr *l)
2300 __swab16s(&l->kuc_magic);
2301 /* __u8 l->kuc_transport */
2302 __swab16s(&l->kuc_msgtype);
2303 __swab16s(&l->kuc_msglen);
2306 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2307 struct lustre_kernelcomm *lk)
2309 struct obd_import *imp = class_exp2cliimp(exp);
2312 if (lk->lk_group != KUC_GRP_HSM) {
2313 CERROR("Bad copytool group %d\n", lk->lk_group);
2317 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2318 lk->lk_uid, lk->lk_group, lk->lk_flags);
2320 if (lk->lk_flags & LK_FLG_STOP) {
2321 /* Unregister with the coordinator */
2322 rc = mdc_ioc_hsm_ct_unregister(imp);
2324 __u32 *archives = NULL;
2326 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2327 archives = lk->lk_data;
2329 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2336 * Send a message to any listening copytools
2337 * @param val KUC message (kuc_hdr + hsm_action_list)
2338 * @param len total length of message
2340 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2341 size_t len, void *val)
2343 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2344 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2348 if (len < sizeof(*lh) + sizeof(*hal)) {
2349 CERROR("Short HSM message %zu < %zu\n", len,
2350 sizeof(*lh) + sizeof(*hal));
2353 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2354 lustre_swab_kuch(lh);
2355 lustre_swab_hal(hal);
2356 } else if (lh->kuc_magic != KUC_MAGIC) {
2357 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2361 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2363 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2364 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2366 /* Broadcast to HSM listeners */
2367 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2373 * callback function passed to kuc for re-registering each HSM copytool
2374 * running on MDC, after MDT shutdown/recovery.
2375 * @param data copytool registration data
2376 * @param cb_arg callback argument (obd_import)
2378 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2380 struct obd_import *imp = (struct obd_import *)cb_arg;
2381 struct kkuc_ct_data *kcd = data;
2382 __u32 *archives = NULL;
2386 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2387 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2390 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2391 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2392 "(archive=%#x)\n", imp->imp_obd->obd_name,
2393 kcd->kcd_nr_archives);
2395 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2396 "(archive nr = %u)\n",
2397 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2398 if (kcd->kcd_nr_archives != 0)
2399 archives = kcd->kcd_archives;
2402 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2403 /* ignore error if the copytool is already registered */
2404 return (rc == -EEXIST) ? 0 : rc;
2408 * Re-establish all kuc contexts with MDT
2409 * after MDT shutdown/recovery.
2411 static int mdc_kuc_reregister(struct obd_import *imp)
2413 /* re-register HSM agents */
2414 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2415 mdc_hsm_ct_reregister, imp);
2418 static int mdc_set_info_async(const struct lu_env *env,
2419 struct obd_export *exp,
2420 u32 keylen, void *key,
2421 u32 vallen, void *val,
2422 struct ptlrpc_request_set *set)
2424 struct obd_import *imp = class_exp2cliimp(exp);
2428 if (KEY_IS(KEY_READ_ONLY)) {
2429 if (vallen != sizeof(int))
2432 spin_lock(&imp->imp_lock);
2433 if (*((int *)val)) {
2434 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2435 imp->imp_connect_data.ocd_connect_flags |=
2438 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2439 imp->imp_connect_data.ocd_connect_flags &=
2440 ~OBD_CONNECT_RDONLY;
2442 spin_unlock(&imp->imp_lock);
2444 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2445 keylen, key, vallen, val, set);
2448 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2449 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2450 keylen, key, vallen, val, set);
2453 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2454 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2459 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2460 __u32 *default_easize = val;
2462 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2466 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2470 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2471 __u32 keylen, void *key, __u32 *vallen, void *val)
2475 if (KEY_IS(KEY_MAX_EASIZE)) {
2476 __u32 mdsize, *max_easize;
2478 if (*vallen != sizeof(int))
2480 mdsize = *(__u32 *)val;
2481 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2482 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2484 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2486 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2487 __u32 *default_easize;
2489 if (*vallen != sizeof(int))
2491 default_easize = val;
2492 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2494 } else if (KEY_IS(KEY_CONN_DATA)) {
2495 struct obd_import *imp = class_exp2cliimp(exp);
2496 struct obd_connect_data *data = val;
2498 if (*vallen != sizeof(*data))
2501 *data = imp->imp_connect_data;
2503 } else if (KEY_IS(KEY_TGT_COUNT)) {
2504 *((__u32 *)val) = 1;
2508 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2513 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2514 struct ptlrpc_request **request)
2516 struct ptlrpc_request *req;
2521 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2525 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2527 ptlrpc_request_free(req);
2531 mdc_pack_body(req, fid, 0, 0, -1, 0);
2533 ptlrpc_request_set_replen(req);
2535 rc = ptlrpc_queue_wait(req);
2537 ptlrpc_req_finished(req);
2543 struct mdc_rmfid_args {
2548 int mdc_rmfid_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2551 struct mdc_rmfid_args *aa;
2556 aa = ptlrpc_req_async_args(aa, req);
2558 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2560 LASSERT(size == sizeof(int) * aa->mra_nr);
2561 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2563 LASSERT(aa->mra_rcs);
2564 LASSERT(aa->mra_nr);
2565 memcpy(aa->mra_rcs, rcs, size);
2571 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2572 int *rcs, struct ptlrpc_request_set *set)
2574 struct ptlrpc_request *req;
2575 struct mdc_rmfid_args *aa;
2581 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2585 flen = fa->fa_nr * sizeof(struct lu_fid);
2586 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2588 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2590 req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2591 RCL_SERVER, fa->fa_nr * sizeof(__u32));
2592 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2594 ptlrpc_request_free(req);
2597 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2598 memcpy(tmp, fa->fa_fids, flen);
2600 mdc_pack_body(req, NULL, 0, 0, -1, 0);
2601 b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2602 b->mbo_ctime = ktime_get_real_seconds();
2604 ptlrpc_request_set_replen(req);
2607 aa = ptlrpc_req_async_args(aa, req);
2609 aa->mra_nr = fa->fa_nr;
2610 req->rq_interpret_reply = mdc_rmfid_interpret;
2612 ptlrpc_set_add_req(set, req);
2613 ptlrpc_check_set(NULL, set);
2618 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2619 enum obd_import_event event)
2621 struct client_obd *cli = &obd->u.cli;
2624 LASSERT(imp->imp_obd == obd);
2627 case IMP_EVENT_DISCON:
2628 spin_lock(&cli->cl_loi_list_lock);
2629 cli->cl_avail_grant = 0;
2630 cli->cl_lost_grant = 0;
2631 spin_unlock(&cli->cl_loi_list_lock);
2633 case IMP_EVENT_INACTIVE:
2635 * Flush current sequence to make client obtain new one
2636 * from server in case of disconnect/reconnect.
2638 down_read(&cli->cl_seq_rwsem);
2640 seq_client_flush(cli->cl_seq);
2641 up_read(&cli->cl_seq_rwsem);
2643 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2645 case IMP_EVENT_INVALIDATE: {
2646 struct ldlm_namespace *ns = obd->obd_namespace;
2650 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2652 env = cl_env_get(&refcheck);
2654 /* Reset grants. All pages go to failing rpcs due to
2655 * the invalid import.
2657 osc_io_unplug(env, cli, NULL);
2659 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2660 osc_ldlm_resource_invalidate,
2662 cl_env_put(env, &refcheck);
2663 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2669 case IMP_EVENT_ACTIVE:
2670 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2671 /* redo the kuc registration after reconnecting */
2673 rc = mdc_kuc_reregister(imp);
2675 case IMP_EVENT_OCD: {
2676 struct obd_connect_data *ocd = &imp->imp_connect_data;
2678 if (OCD_HAS_FLAG(ocd, GRANT))
2679 osc_init_grant(cli, ocd);
2681 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2684 case IMP_EVENT_DEACTIVATE:
2685 case IMP_EVENT_ACTIVATE:
2688 CERROR("Unknown import event %x\n", event);
2694 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2695 struct lu_fid *fid, struct md_op_data *op_data)
2697 struct client_obd *cli = &exp->exp_obd->u.cli;
2702 down_read(&cli->cl_seq_rwsem);
2704 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2705 up_read(&cli->cl_seq_rwsem);
2710 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2712 struct client_obd *cli = &exp->exp_obd->u.cli;
2713 return &cli->cl_target_uuid;
2717 * Determine whether the lock can be canceled before replaying it during
2718 * recovery, non zero value will be return if the lock can be canceled,
2719 * or zero returned for not
2721 static int mdc_cancel_weight(struct ldlm_lock *lock)
2723 if (lock->l_resource->lr_type != LDLM_IBITS)
2726 /* FIXME: if we ever get into a situation where there are too many
2727 * opened files with open locks on a single node, then we really
2728 * should replay these open locks to reget it */
2729 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2732 /* Special case for DoM locks, cancel only unused and granted locks */
2733 if (ldlm_has_dom(lock) &&
2734 (lock->l_granted_mode != lock->l_req_mode ||
2735 osc_ldlm_weigh_ast(lock) != 0))
2741 static int mdc_resource_inode_free(struct ldlm_resource *res)
2743 if (res->lr_lvb_inode)
2744 res->lr_lvb_inode = NULL;
2749 static struct ldlm_valblock_ops inode_lvbo = {
2750 .lvbo_free = mdc_resource_inode_free
2753 static int mdc_llog_init(struct obd_device *obd)
2755 struct obd_llog_group *olg = &obd->obd_olg;
2756 struct llog_ctxt *ctxt;
2761 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2766 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2767 llog_initiator_connect(ctxt);
2768 llog_ctxt_put(ctxt);
2773 static void mdc_llog_finish(struct obd_device *obd)
2775 struct llog_ctxt *ctxt;
2779 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2781 llog_cleanup(NULL, ctxt);
2786 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2792 rc = osc_setup_common(obd, cfg);
2796 rc = mdc_tunables_init(obd);
2798 GOTO(err_osc_cleanup, rc);
2800 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2802 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2804 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2806 rc = mdc_llog_init(obd);
2808 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2810 GOTO(err_llog_cleanup, rc);
2813 rc = mdc_changelog_cdev_init(obd);
2815 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2817 GOTO(err_changelog_cleanup, rc);
2822 err_changelog_cleanup:
2823 mdc_llog_finish(obd);
2825 lprocfs_free_md_stats(obd);
2826 ptlrpc_lprocfs_unregister_obd(obd);
2828 osc_cleanup_common(obd);
2832 /* Initialize the default and maximum LOV EA sizes. This allows
2833 * us to make MDS RPCs with large enough reply buffers to hold a default
2834 * sized EA without having to calculate this (via a call into the
2835 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2836 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2837 * a large number of stripes is possible. If a larger reply buffer is
2838 * required it will be reallocated in the ptlrpc layer due to overflow.
2840 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2843 struct obd_device *obd = exp->exp_obd;
2844 struct client_obd *cli = &obd->u.cli;
2847 if (cli->cl_max_mds_easize < easize)
2848 cli->cl_max_mds_easize = easize;
2850 if (cli->cl_default_mds_easize < def_easize)
2851 cli->cl_default_mds_easize = def_easize;
2856 static int mdc_precleanup(struct obd_device *obd)
2860 osc_precleanup_common(obd);
2861 mdc_changelog_cdev_finish(obd);
2863 obd_cleanup_client_import(obd);
2864 ptlrpc_lprocfs_unregister_obd(obd);
2865 lprocfs_free_md_stats(obd);
2866 mdc_llog_finish(obd);
2870 static int mdc_cleanup(struct obd_device *obd)
2872 return osc_cleanup_common(obd);
2875 static const struct obd_ops mdc_obd_ops = {
2876 .o_owner = THIS_MODULE,
2877 .o_setup = mdc_setup,
2878 .o_precleanup = mdc_precleanup,
2879 .o_cleanup = mdc_cleanup,
2880 .o_add_conn = client_import_add_conn,
2881 .o_del_conn = client_import_del_conn,
2882 .o_connect = client_connect_import,
2883 .o_reconnect = osc_reconnect,
2884 .o_disconnect = osc_disconnect,
2885 .o_iocontrol = mdc_iocontrol,
2886 .o_set_info_async = mdc_set_info_async,
2887 .o_statfs = mdc_statfs,
2888 .o_statfs_async = mdc_statfs_async,
2889 .o_fid_init = client_fid_init,
2890 .o_fid_fini = client_fid_fini,
2891 .o_fid_alloc = mdc_fid_alloc,
2892 .o_import_event = mdc_import_event,
2893 .o_get_info = mdc_get_info,
2894 .o_get_uuid = mdc_get_uuid,
2895 .o_quotactl = mdc_quotactl,
2898 static const struct md_ops mdc_md_ops = {
2899 .m_get_root = mdc_get_root,
2900 .m_null_inode = mdc_null_inode,
2901 .m_close = mdc_close,
2902 .m_create = mdc_create,
2903 .m_enqueue = mdc_enqueue,
2904 .m_getattr = mdc_getattr,
2905 .m_getattr_name = mdc_getattr_name,
2906 .m_intent_lock = mdc_intent_lock,
2908 .m_rename = mdc_rename,
2909 .m_setattr = mdc_setattr,
2910 .m_setxattr = mdc_setxattr,
2911 .m_getxattr = mdc_getxattr,
2912 .m_fsync = mdc_fsync,
2913 .m_file_resync = mdc_file_resync,
2914 .m_read_page = mdc_read_page,
2915 .m_unlink = mdc_unlink,
2916 .m_cancel_unused = mdc_cancel_unused,
2917 .m_init_ea_size = mdc_init_ea_size,
2918 .m_set_lock_data = mdc_set_lock_data,
2919 .m_lock_match = mdc_lock_match,
2920 .m_get_lustre_md = mdc_get_lustre_md,
2921 .m_free_lustre_md = mdc_free_lustre_md,
2922 .m_set_open_replay_data = mdc_set_open_replay_data,
2923 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2924 .m_intent_getattr_async = mdc_intent_getattr_async,
2925 .m_revalidate_lock = mdc_revalidate_lock,
2926 .m_rmfid = mdc_rmfid,
2929 dev_t mdc_changelog_dev;
2930 struct class *mdc_changelog_class;
2931 static int __init mdc_init(void)
2934 rc = alloc_chrdev_region(&mdc_changelog_dev, 0,
2935 MDC_CHANGELOG_DEV_COUNT,
2936 MDC_CHANGELOG_DEV_NAME);
2940 mdc_changelog_class = class_create(THIS_MODULE, MDC_CHANGELOG_DEV_NAME);
2941 if (IS_ERR(mdc_changelog_class)) {
2942 rc = PTR_ERR(mdc_changelog_class);
2946 rc = class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2947 LUSTRE_MDC_NAME, &mdc_device_type);
2954 class_destroy(mdc_changelog_class);
2956 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
2960 static void __exit mdc_exit(void)
2962 class_unregister_type(LUSTRE_MDC_NAME);
2963 class_destroy(mdc_changelog_class);
2964 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
2965 idr_destroy(&mdc_changelog_minor_idr);
2968 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2969 MODULE_DESCRIPTION("Lustre Metadata Client");
2970 MODULE_VERSION(LUSTRE_VERSION_STRING);
2971 MODULE_LICENSE("GPL");
2973 module_init(mdc_init);
2974 module_exit(mdc_exit);