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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_MDC
39 #include <linux/module.h>
40 #include <linux/pagemap.h>
41 #include <linux/miscdevice.h>
42 #include <linux/init.h>
43 #include <linux/utsname.h>
44 #include <linux/kthread.h>
45 #include <linux/user_namespace.h>
46 #ifdef HAVE_UIDGID_HEADER
47 # include <linux/uidgid.h>
50 #include <lustre_acl.h>
51 #include <lustre_ioctl.h>
52 #include <obd_class.h>
53 #include <lustre_lmv.h>
54 #include <lustre_fid.h>
55 #include <lprocfs_status.h>
56 #include <lustre_param.h>
57 #include <lustre_log.h>
58 #include <lustre_kernelcomm.h>
59 #include <cl_object.h>
61 #include "mdc_internal.h"
63 #define REQUEST_MINOR 244
65 struct mdc_renew_capa_args {
66 struct obd_capa *ra_oc;
67 renew_capa_cb_t ra_cb;
70 static int mdc_cleanup(struct obd_device *obd);
72 static int mdc_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
73 const struct req_msg_field *field,
76 struct lustre_capa *capa;
80 /* swabbed already in mdc_enqueue */
81 capa = req_capsule_server_get(&req->rq_pill, field);
85 c = alloc_capa(CAPA_SITE_CLIENT);
87 CDEBUG(D_INFO, "alloc capa failed!\n");
96 static inline int mdc_queue_wait(struct ptlrpc_request *req)
98 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
101 /* obd_get_request_slot() ensures that this client has no more
102 * than cl_max_rpcs_in_flight RPCs simultaneously inf light
104 rc = obd_get_request_slot(cli);
108 rc = ptlrpc_queue_wait(req);
109 obd_put_request_slot(cli);
114 /* Helper that implements most of mdc_getstatus and signal_completed_replay. */
115 /* XXX this should become mdc_get_info("key"), sending MDS_GET_INFO RPC */
116 static int send_getstatus(struct obd_import *imp, struct lu_fid *rootfid,
117 struct obd_capa **pc, int level, int msg_flags)
119 struct ptlrpc_request *req;
120 struct mdt_body *body;
124 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_GETSTATUS,
125 LUSTRE_MDS_VERSION, MDS_GETSTATUS);
129 mdc_pack_body(req, NULL, NULL, 0, 0, -1, 0);
130 lustre_msg_add_flags(req->rq_reqmsg, msg_flags);
131 req->rq_send_state = level;
133 ptlrpc_request_set_replen(req);
135 rc = ptlrpc_queue_wait(req);
139 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
141 GOTO(out, rc = -EPROTO);
143 if (body->mbo_valid & OBD_MD_FLMDSCAPA) {
144 rc = mdc_unpack_capa(NULL, req, &RMF_CAPA1, pc);
149 *rootfid = body->mbo_fid1;
150 CDEBUG(D_NET, "root fid="DFID", last_committed="LPU64"\n",
151 PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
154 ptlrpc_req_finished(req);
158 /* This should be mdc_get_info("rootfid") */
159 static int mdc_getstatus(struct obd_export *exp, struct lu_fid *rootfid,
160 struct obd_capa **pc)
162 return send_getstatus(class_exp2cliimp(exp), rootfid, pc,
167 * This function now is known to always saying that it will receive 4 buffers
168 * from server. Even for cases when acl_size and md_size is zero, RPC header
169 * will contain 4 fields and RPC itself will contain zero size fields. This is
170 * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
171 * and thus zero, it shrinks it, making zero size. The same story about
172 * md_size. And this is course of problem when client waits for smaller number
173 * of fields. This issue will be fixed later when client gets aware of RPC
176 static int mdc_getattr_common(struct obd_export *exp,
177 struct ptlrpc_request *req)
179 struct req_capsule *pill = &req->rq_pill;
180 struct mdt_body *body;
185 /* Request message already built. */
186 rc = ptlrpc_queue_wait(req);
190 /* sanity check for the reply */
191 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
195 CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
197 mdc_update_max_ea_from_body(exp, body);
198 if (body->mbo_eadatasize != 0) {
199 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
200 body->mbo_eadatasize);
205 if (body->mbo_valid & OBD_MD_FLRMTPERM) {
206 struct mdt_remote_perm *perm;
208 LASSERT(client_is_remote(exp));
209 perm = req_capsule_server_swab_get(pill, &RMF_ACL,
210 lustre_swab_mdt_remote_perm);
215 if (body->mbo_valid & OBD_MD_FLMDSCAPA) {
216 struct lustre_capa *capa;
217 capa = req_capsule_server_get(pill, &RMF_CAPA1);
225 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
226 struct ptlrpc_request **request)
228 struct ptlrpc_request *req;
232 /* Single MDS without an LMV case */
233 if (op_data->op_flags & MF_GET_MDT_IDX) {
238 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_GETATTR);
242 mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
244 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
246 ptlrpc_request_free(req);
250 mdc_pack_body(req, &op_data->op_fid1, op_data->op_capa1,
251 op_data->op_valid, op_data->op_mode, -1, 0);
253 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
255 if (op_data->op_valid & OBD_MD_FLRMTPERM) {
256 LASSERT(client_is_remote(exp));
257 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
258 sizeof(struct mdt_remote_perm));
260 ptlrpc_request_set_replen(req);
262 rc = mdc_getattr_common(exp, req);
264 ptlrpc_req_finished(req);
270 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
271 struct ptlrpc_request **request)
273 struct ptlrpc_request *req;
278 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
279 &RQF_MDS_GETATTR_NAME);
283 mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
284 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
285 op_data->op_namelen + 1);
287 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
289 ptlrpc_request_free(req);
293 mdc_pack_body(req, &op_data->op_fid1, op_data->op_capa1,
294 op_data->op_valid, op_data->op_mode,
295 op_data->op_suppgids[0], 0);
297 if (op_data->op_name) {
298 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
299 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
300 op_data->op_namelen);
301 memcpy(name, op_data->op_name, op_data->op_namelen);
304 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
306 ptlrpc_request_set_replen(req);
308 rc = mdc_getattr_common(exp, req);
310 ptlrpc_req_finished(req);
316 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
317 const struct lu_fid *fid,
318 struct obd_capa *oc, int opcode, u64 valid,
319 const char *xattr_name, const char *input,
320 int input_size, int output_size, int flags,
321 __u32 suppgid, struct ptlrpc_request **request)
323 struct ptlrpc_request *req;
324 int xattr_namelen = 0;
330 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
334 mdc_set_capa_size(req, &RMF_CAPA1, oc);
336 xattr_namelen = strlen(xattr_name) + 1;
337 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
342 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
346 /* Flush local XATTR locks to get rid of a possible cancel RPC */
347 if (opcode == MDS_REINT && fid_is_sane(fid) &&
348 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
349 struct list_head cancels = LIST_HEAD_INIT(cancels);
352 /* Without that packing would fail */
354 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
357 count = mdc_resource_get_unused(exp, fid,
359 MDS_INODELOCK_XATTR);
361 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
363 ptlrpc_request_free(req);
367 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
369 ptlrpc_request_free(req);
374 if (opcode == MDS_REINT) {
375 struct mdt_rec_setxattr *rec;
377 CLASSERT(sizeof(struct mdt_rec_setxattr) ==
378 sizeof(struct mdt_rec_reint));
379 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
380 rec->sx_opcode = REINT_SETXATTR;
381 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
382 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
383 rec->sx_cap = cfs_curproc_cap_pack();
384 rec->sx_suppgid1 = suppgid;
385 rec->sx_suppgid2 = -1;
387 rec->sx_valid = valid | OBD_MD_FLCTIME;
388 rec->sx_time = cfs_time_current_sec();
389 rec->sx_size = output_size;
390 rec->sx_flags = flags;
392 mdc_pack_capa(req, &RMF_CAPA1, oc);
394 mdc_pack_body(req, fid, oc, 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 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
407 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
408 RCL_SERVER, output_size);
409 ptlrpc_request_set_replen(req);
412 if (opcode == MDS_REINT)
413 mdc_get_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
415 rc = ptlrpc_queue_wait(req);
417 if (opcode == MDS_REINT)
418 mdc_put_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
421 ptlrpc_req_finished(req);
427 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
428 struct obd_capa *oc, u64 valid,
429 const char *xattr_name,
430 const char *input, int input_size, int output_size,
431 int flags, __u32 suppgid,
432 struct ptlrpc_request **request)
434 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
435 fid, oc, MDS_REINT, valid, xattr_name,
436 input, input_size, output_size, flags,
440 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
441 struct obd_capa *oc, u64 valid,
442 const char *xattr_name,
443 const char *input, int input_size, int output_size,
444 int flags, struct ptlrpc_request **request)
446 return mdc_xattr_common(exp, &RQF_MDS_GETXATTR,
447 fid, oc, MDS_GETXATTR, valid, xattr_name,
448 input, input_size, output_size, flags,
452 #ifdef CONFIG_FS_POSIX_ACL
453 static int mdc_unpack_acl(struct ptlrpc_request *req, struct lustre_md *md)
455 struct req_capsule *pill = &req->rq_pill;
456 struct mdt_body *body = md->body;
457 struct posix_acl *acl;
462 if (!body->mbo_aclsize)
465 buf = req_capsule_server_sized_get(pill, &RMF_ACL, body->mbo_aclsize);
470 acl = posix_acl_from_xattr(&init_user_ns, buf, body->mbo_aclsize);
475 CERROR("convert xattr to acl: %d\n", rc);
479 rc = posix_acl_valid(acl);
481 CERROR("validate acl: %d\n", rc);
482 posix_acl_release(acl);
490 #define mdc_unpack_acl(req, md) 0
493 int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
494 struct obd_export *dt_exp, struct obd_export *md_exp,
495 struct lustre_md *md)
497 struct req_capsule *pill = &req->rq_pill;
502 memset(md, 0, sizeof(*md));
504 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
505 LASSERT(md->body != NULL);
507 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
509 struct lov_mds_md *lmm;
511 if (!S_ISREG(md->body->mbo_mode)) {
512 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
513 "regular file, but is not\n");
514 GOTO(out, rc = -EPROTO);
517 if (md->body->mbo_eadatasize == 0) {
518 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
519 "but eadatasize 0\n");
520 GOTO(out, rc = -EPROTO);
523 lmmsize = md->body->mbo_eadatasize;
524 lmm = req_capsule_server_sized_get(pill, &RMF_MDT_MD, lmmsize);
526 GOTO(out, rc = -EPROTO);
528 rc = obd_unpackmd(dt_exp, &md->lsm, lmm, lmmsize);
531 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
533 struct lov_mds_md *lmv;
535 if (!S_ISDIR(md->body->mbo_mode)) {
536 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
537 "directory, but is not\n");
538 GOTO(out, rc = -EPROTO);
541 if (md->body->mbo_eadatasize == 0) {
542 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
543 "but eadatasize 0\n");
547 if (md->body->mbo_valid & OBD_MD_MEA) {
548 lmvsize = md->body->mbo_eadatasize;
549 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
552 GOTO(out, rc = -EPROTO);
554 rc = obd_unpackmd(md_exp, (void *)&md->lmv, lmv,
559 if (rc < (typeof(rc))sizeof(*md->lmv)) {
560 CDEBUG(D_INFO, "size too small: "
561 "rc < sizeof(*md->lmv) (%d < %d)\n",
562 rc, (int)sizeof(*md->lmv));
563 GOTO(out, rc = -EPROTO);
569 if (md->body->mbo_valid & OBD_MD_FLRMTPERM) {
570 /* remote permission */
571 LASSERT(client_is_remote(exp));
572 md->remote_perm = req_capsule_server_swab_get(pill, &RMF_ACL,
573 lustre_swab_mdt_remote_perm);
574 if (!md->remote_perm)
575 GOTO(out, rc = -EPROTO);
576 } else if (md->body->mbo_valid & OBD_MD_FLACL) {
577 /* for ACL, it's possible that FLACL is set but aclsize is zero.
578 * only when aclsize != 0 there's an actual segment for ACL
581 if (md->body->mbo_aclsize) {
582 rc = mdc_unpack_acl(req, md);
585 #ifdef CONFIG_FS_POSIX_ACL
587 md->posix_acl = NULL;
591 if (md->body->mbo_valid & OBD_MD_FLMDSCAPA) {
592 struct obd_capa *oc = NULL;
594 rc = mdc_unpack_capa(NULL, req, &RMF_CAPA1, &oc);
600 if (md->body->mbo_valid & OBD_MD_FLOSSCAPA) {
601 struct obd_capa *oc = NULL;
603 rc = mdc_unpack_capa(NULL, req, &RMF_CAPA2, &oc);
613 capa_put(md->oss_capa);
617 capa_put(md->mds_capa);
620 #ifdef CONFIG_FS_POSIX_ACL
621 posix_acl_release(md->posix_acl);
624 obd_free_memmd(dt_exp, &md->lsm);
629 int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
635 void mdc_replay_open(struct ptlrpc_request *req)
637 struct md_open_data *mod = req->rq_cb_data;
638 struct ptlrpc_request *close_req;
639 struct obd_client_handle *och;
640 struct lustre_handle old;
641 struct mdt_body *body;
645 DEBUG_REQ(D_ERROR, req,
646 "Can't properly replay without open data.");
651 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
652 LASSERT(body != NULL);
656 struct lustre_handle *file_fh;
658 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
660 file_fh = &och->och_fh;
661 CDEBUG(D_HA, "updating handle from "LPX64" to "LPX64"\n",
662 file_fh->cookie, body->mbo_handle.cookie);
664 *file_fh = body->mbo_handle;
666 close_req = mod->mod_close_req;
667 if (close_req != NULL) {
668 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
669 struct mdt_ioepoch *epoch;
671 LASSERT(opc == MDS_CLOSE);
672 epoch = req_capsule_client_get(&close_req->rq_pill,
677 LASSERT(!memcmp(&old, &epoch->mio_handle, sizeof(old)));
679 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
680 epoch->mio_handle = body->mbo_handle;
685 void mdc_commit_open(struct ptlrpc_request *req)
687 struct md_open_data *mod = req->rq_cb_data;
692 * No need to touch md_open_data::mod_och, it holds a reference on
693 * \var mod and will zero references to each other, \var mod will be
694 * freed after that when md_open_data::mod_och will put the reference.
698 * Do not let open request to disappear as it still may be needed
699 * for close rpc to happen (it may happen on evict only, otherwise
700 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
701 * called), just mark this rpc as committed to distinguish these 2
702 * cases, see mdc_close() for details. The open request reference will
703 * be put along with freeing \var mod.
705 ptlrpc_request_addref(req);
706 spin_lock(&req->rq_lock);
707 req->rq_committed = 1;
708 spin_unlock(&req->rq_lock);
709 req->rq_cb_data = NULL;
713 int mdc_set_open_replay_data(struct obd_export *exp,
714 struct obd_client_handle *och,
715 struct lookup_intent *it)
717 struct md_open_data *mod;
718 struct mdt_rec_create *rec;
719 struct mdt_body *body;
720 struct ptlrpc_request *open_req = it->d.lustre.it_data;
721 struct obd_import *imp = open_req->rq_import;
724 if (!open_req->rq_replay)
727 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
728 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
729 LASSERT(rec != NULL);
730 /* Incoming message in my byte order (it's been swabbed). */
731 /* Outgoing messages always in my byte order. */
732 LASSERT(body != NULL);
734 /* Only if the import is replayable, we set replay_open data */
735 if (och && imp->imp_replayable) {
736 mod = obd_mod_alloc();
738 DEBUG_REQ(D_ERROR, open_req,
739 "Can't allocate md_open_data");
744 * Take a reference on \var mod, to be freed on mdc_close().
745 * It protects \var mod from being freed on eviction (commit
746 * callback is called despite rq_replay flag).
747 * Another reference for \var och.
752 spin_lock(&open_req->rq_lock);
755 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
756 it_disposition(it, DISP_OPEN_STRIPE);
757 mod->mod_open_req = open_req;
758 open_req->rq_cb_data = mod;
759 open_req->rq_commit_cb = mdc_commit_open;
760 spin_unlock(&open_req->rq_lock);
763 rec->cr_fid2 = body->mbo_fid1;
764 rec->cr_ioepoch = body->mbo_ioepoch;
765 rec->cr_old_handle.cookie = body->mbo_handle.cookie;
766 open_req->rq_replay_cb = mdc_replay_open;
767 if (!fid_is_sane(&body->mbo_fid1)) {
768 DEBUG_REQ(D_ERROR, open_req, "Saving replay request with "
773 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
777 static void mdc_free_open(struct md_open_data *mod)
781 if (mod->mod_is_create == 0 &&
782 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
785 LASSERT(mod->mod_open_req->rq_replay == 0);
787 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request\n");
789 ptlrpc_request_committed(mod->mod_open_req, committed);
790 if (mod->mod_close_req)
791 ptlrpc_request_committed(mod->mod_close_req, committed);
794 int mdc_clear_open_replay_data(struct obd_export *exp,
795 struct obd_client_handle *och)
797 struct md_open_data *mod = och->och_mod;
801 * It is possible to not have \var mod in a case of eviction between
802 * lookup and ll_file_open().
807 LASSERT(mod != LP_POISON);
808 LASSERT(mod->mod_open_req != NULL);
818 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
819 struct md_open_data *mod, struct ptlrpc_request **request)
821 struct obd_device *obd = class_exp2obd(exp);
822 struct ptlrpc_request *req;
823 struct req_format *req_fmt;
828 if (op_data->op_bias & MDS_HSM_RELEASE) {
829 req_fmt = &RQF_MDS_INTENT_CLOSE;
831 /* allocate a FID for volatile file */
832 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
834 CERROR("%s: "DFID" failed to allocate FID: %d\n",
835 obd->obd_name, PFID(&op_data->op_fid1), rc);
836 /* save the errcode and proceed to close */
839 } else if (op_data->op_bias & MDS_CLOSE_LAYOUT_SWAP) {
840 req_fmt = &RQF_MDS_INTENT_CLOSE;
842 req_fmt = &RQF_MDS_CLOSE;
846 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
850 mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
852 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
854 ptlrpc_request_free(req);
858 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
859 * portal whose threads are not taking any DLM locks and are therefore
860 * always progressing */
861 req->rq_request_portal = MDS_READPAGE_PORTAL;
862 ptlrpc_at_set_req_timeout(req);
864 /* Ensure that this close's handle is fixed up during replay. */
865 if (likely(mod != NULL)) {
866 LASSERTF(mod->mod_open_req != NULL &&
867 mod->mod_open_req->rq_type != LI_POISON,
868 "POISONED open %p!\n", mod->mod_open_req);
870 mod->mod_close_req = req;
872 DEBUG_REQ(D_HA, mod->mod_open_req, "matched open");
873 /* We no longer want to preserve this open for replay even
874 * though the open was committed. b=3632, b=3633 */
875 spin_lock(&mod->mod_open_req->rq_lock);
876 mod->mod_open_req->rq_replay = 0;
877 spin_unlock(&mod->mod_open_req->rq_lock);
879 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
882 mdc_close_pack(req, op_data);
884 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
885 obd->u.cli.cl_default_mds_easize);
887 ptlrpc_request_set_replen(req);
889 mdc_get_rpc_lock(obd->u.cli.cl_close_lock, NULL);
890 rc = ptlrpc_queue_wait(req);
891 mdc_put_rpc_lock(obd->u.cli.cl_close_lock, NULL);
893 if (req->rq_repmsg == NULL) {
894 CDEBUG(D_RPCTRACE, "request failed to send: %p, %d\n", req,
897 rc = req->rq_status ?: -EIO;
898 } else if (rc == 0 || rc == -EAGAIN) {
899 struct mdt_body *body;
901 rc = lustre_msg_get_status(req->rq_repmsg);
902 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
903 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err "
908 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
911 } else if (rc == -ESTALE) {
913 * it can be allowed error after 3633 if open was committed and
914 * server failed before close was sent. Let's check if mod
915 * exists and return no error in that case
918 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
919 LASSERT(mod->mod_open_req != NULL);
920 if (mod->mod_open_req->rq_committed)
927 mod->mod_close_req = NULL;
928 /* Since now, mod is accessed through open_req only,
929 * thus close req does not keep a reference on mod anymore. */
934 RETURN(rc < 0 ? rc : saved_rc);
937 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
938 __u64 offset, struct obd_capa *oc,
939 struct page **pages, int npages,
940 struct ptlrpc_request **request)
942 struct ptlrpc_request *req;
943 struct ptlrpc_bulk_desc *desc;
945 wait_queue_head_t waitq;
947 struct l_wait_info lwi;
952 init_waitqueue_head(&waitq);
955 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
959 mdc_set_capa_size(req, &RMF_CAPA1, oc);
961 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
963 ptlrpc_request_free(req);
967 req->rq_request_portal = MDS_READPAGE_PORTAL;
968 ptlrpc_at_set_req_timeout(req);
970 desc = ptlrpc_prep_bulk_imp(req, npages, 1, BULK_PUT_SINK,
973 ptlrpc_request_free(req);
977 /* NB req now owns desc and will free it when it gets freed */
978 for (i = 0; i < npages; i++)
979 ptlrpc_prep_bulk_page_pin(desc, pages[i], 0, PAGE_CACHE_SIZE);
981 mdc_readdir_pack(req, offset, PAGE_CACHE_SIZE * npages, fid, oc);
983 ptlrpc_request_set_replen(req);
984 rc = ptlrpc_queue_wait(req);
986 ptlrpc_req_finished(req);
987 if (rc != -ETIMEDOUT)
991 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
992 CERROR("%s: too many resend retries: rc = %d\n",
993 exp->exp_obd->obd_name, -EIO);
996 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(resends), NULL, NULL,
998 l_wait_event(waitq, 0, &lwi);
1003 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1004 req->rq_bulk->bd_nob_transferred);
1006 ptlrpc_req_finished(req);
1010 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1011 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1012 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1013 PAGE_CACHE_SIZE * npages);
1014 ptlrpc_req_finished(req);
1022 static void mdc_release_page(struct page *page, int remove)
1026 if (likely(page->mapping != NULL))
1027 truncate_complete_page(page->mapping, page);
1030 page_cache_release(page);
1033 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1034 __u64 *start, __u64 *end, int hash64)
1037 * Complement of hash is used as an index so that
1038 * radix_tree_gang_lookup() can be used to find a page with starting
1039 * hash _smaller_ than one we are looking for.
1041 unsigned long offset = hash_x_index(*hash, hash64);
1045 spin_lock_irq(&mapping->tree_lock);
1046 found = radix_tree_gang_lookup(&mapping->page_tree,
1047 (void **)&page, offset, 1);
1048 if (found > 0 && !radix_tree_exceptional_entry(page)) {
1049 struct lu_dirpage *dp;
1051 page_cache_get(page);
1052 spin_unlock_irq(&mapping->tree_lock);
1054 * In contrast to find_lock_page() we are sure that directory
1055 * page cannot be truncated (while DLM lock is held) and,
1056 * hence, can avoid restart.
1058 * In fact, page cannot be locked here at all, because
1059 * mdc_read_page_remote does synchronous io.
1061 wait_on_page_locked(page);
1062 if (PageUptodate(page)) {
1064 if (BITS_PER_LONG == 32 && hash64) {
1065 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1066 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1067 *hash = *hash >> 32;
1069 *start = le64_to_cpu(dp->ldp_hash_start);
1070 *end = le64_to_cpu(dp->ldp_hash_end);
1072 if (unlikely(*start == 1 && *hash == 0))
1075 LASSERTF(*start <= *hash, "start = "LPX64
1076 ",end = "LPX64",hash = "LPX64"\n",
1077 *start, *end, *hash);
1078 CDEBUG(D_VFSTRACE, "offset %lx ["LPX64" "LPX64"],"
1079 " hash "LPX64"\n", offset, *start, *end, *hash);
1082 mdc_release_page(page, 0);
1084 } else if (*end != *start && *hash == *end) {
1086 * upon hash collision, remove this page,
1087 * otherwise put page reference, and
1088 * mdc_read_page_remote() will issue RPC to
1089 * fetch the page we want.
1092 mdc_release_page(page,
1093 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1097 page_cache_release(page);
1098 page = ERR_PTR(-EIO);
1101 spin_unlock_irq(&mapping->tree_lock);
1108 * Adjust a set of pages, each page containing an array of lu_dirpages,
1109 * so that each page can be used as a single logical lu_dirpage.
1111 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1112 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1113 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1114 * value is used as a cookie to request the next lu_dirpage in a
1115 * directory listing that spans multiple pages (two in this example):
1118 * .|--------v------- -----.
1119 * |s|e|f|p|ent|ent| ... |ent|
1120 * '--|-------------- -----' Each PAGE contains a single
1121 * '------. lu_dirpage.
1122 * .---------v------- -----.
1123 * |s|e|f|p|ent| 0 | ... | 0 |
1124 * '----------------- -----'
1126 * However, on hosts where the native VM page size (PAGE_CACHE_SIZE) is
1127 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1128 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1129 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1130 * after it in the same PAGE (arrows simplified for brevity, but
1131 * in general e0==s1, e1==s2, etc.):
1133 * .-------------------- -----.
1134 * |s0|e0|f0|p|ent|ent| ... |ent|
1135 * |---v---------------- -----|
1136 * |s1|e1|f1|p|ent|ent| ... |ent|
1137 * |---v---------------- -----| Here, each PAGE contains
1138 * ... multiple lu_dirpages.
1139 * |---v---------------- -----|
1140 * |s'|e'|f'|p|ent|ent| ... |ent|
1141 * '---|---------------- -----'
1143 * .----------------------------.
1146 * This structure is transformed into a single logical lu_dirpage as follows:
1148 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1149 * labeled 'next PAGE'.
1151 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1152 * a hash collision with the next page exists.
1154 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1155 * to the first entry of the next lu_dirpage.
1157 #if PAGE_CACHE_SIZE > LU_PAGE_SIZE
1158 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1162 for (i = 0; i < cfs_pgs; i++) {
1163 struct lu_dirpage *dp = kmap(pages[i]);
1164 struct lu_dirpage *first = dp;
1165 struct lu_dirent *end_dirent = NULL;
1166 struct lu_dirent *ent;
1167 __u64 hash_end = le64_to_cpu(dp->ldp_hash_end);
1168 __u32 flags = le32_to_cpu(dp->ldp_flags);
1170 while (--lu_pgs > 0) {
1171 ent = lu_dirent_start(dp);
1172 for (end_dirent = ent; ent != NULL;
1173 end_dirent = ent, ent = lu_dirent_next(ent));
1175 /* Advance dp to next lu_dirpage. */
1176 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1178 /* Check if we've reached the end of the PAGE. */
1179 if (!((unsigned long)dp & ~PAGE_MASK))
1182 /* Save the hash and flags of this lu_dirpage. */
1183 hash_end = le64_to_cpu(dp->ldp_hash_end);
1184 flags = le32_to_cpu(dp->ldp_flags);
1186 /* Check if lu_dirpage contains no entries. */
1187 if (end_dirent == NULL)
1190 /* Enlarge the end entry lde_reclen from 0 to
1191 * first entry of next lu_dirpage. */
1192 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1193 end_dirent->lde_reclen =
1194 cpu_to_le16((char *)(dp->ldp_entries) -
1195 (char *)end_dirent);
1198 first->ldp_hash_end = hash_end;
1199 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1200 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1204 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1207 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1208 #endif /* PAGE_CACHE_SIZE > LU_PAGE_SIZE */
1210 /* parameters for readdir page */
1211 struct readpage_param {
1212 struct md_op_data *rp_mod;
1215 struct obd_export *rp_exp;
1216 struct md_callback *rp_cb;
1219 #ifndef HAVE_DELETE_FROM_PAGE_CACHE
1220 static inline void delete_from_page_cache(struct page *page)
1222 remove_from_page_cache(page);
1223 page_cache_release(page);
1228 * Read pages from server.
1230 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1231 * a header lu_dirpage which describes the start/end hash, and whether this
1232 * page is empty (contains no dir entry) or hash collide with next page.
1233 * After client receives reply, several pages will be integrated into dir page
1234 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1235 * lu_dirpage for this integrated page will be adjusted.
1237 static int mdc_read_page_remote(void *data, struct page *page0)
1239 struct readpage_param *rp = data;
1240 struct page **page_pool;
1242 struct lu_dirpage *dp;
1243 int rd_pgs = 0; /* number of pages read actually */
1245 struct md_op_data *op_data = rp->rp_mod;
1246 struct ptlrpc_request *req;
1247 int max_pages = op_data->op_max_pages;
1248 struct inode *inode;
1254 LASSERT(max_pages > 0 && max_pages <= PTLRPC_MAX_BRW_PAGES);
1255 inode = op_data->op_data;
1256 fid = &op_data->op_fid1;
1257 LASSERT(inode != NULL);
1259 OBD_ALLOC(page_pool, sizeof(page_pool[0]) * max_pages);
1260 if (page_pool != NULL) {
1261 page_pool[0] = page0;
1267 for (npages = 1; npages < max_pages; npages++) {
1268 page = page_cache_alloc_cold(inode->i_mapping);
1271 page_pool[npages] = page;
1274 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, op_data->op_capa1,
1275 page_pool, npages, &req);
1277 /* page0 is special, which was added into page cache early */
1278 delete_from_page_cache(page0);
1282 rd_pgs = (req->rq_bulk->bd_nob_transferred +
1283 PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1284 lu_pgs = req->rq_bulk->bd_nob_transferred >>
1286 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1288 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1290 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1292 SetPageUptodate(page0);
1296 ptlrpc_req_finished(req);
1297 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1298 for (i = 1; i < npages; i++) {
1299 unsigned long offset;
1303 page = page_pool[i];
1305 if (rc < 0 || i >= rd_pgs) {
1306 page_cache_release(page);
1310 SetPageUptodate(page);
1313 hash = le64_to_cpu(dp->ldp_hash_start);
1316 offset = hash_x_index(hash, rp->rp_hash64);
1318 prefetchw(&page->flags);
1319 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1324 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1325 " rc = %d\n", offset, ret);
1326 page_cache_release(page);
1329 if (page_pool != &page0)
1330 OBD_FREE(page_pool, sizeof(page_pool[0]) * max_pages);
1336 * Read dir page from cache first, if it can not find it, read it from
1337 * server and add into the cache.
1339 * \param[in] exp MDC export
1340 * \param[in] op_data client MD stack parameters, transfering parameters
1341 * between different layers on client MD stack.
1342 * \param[in] cb_op callback required for ldlm lock enqueue during
1344 * \param[in] hash_offset the hash offset of the page to be read
1345 * \param[in] ppage the page to be read
1347 * retval = 0 get the page successfully
1348 * errno(<0) get the page failed
1350 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1351 struct md_callback *cb_op, __u64 hash_offset,
1352 struct page **ppage)
1354 struct lookup_intent it = { .it_op = IT_READDIR };
1356 struct inode *dir = op_data->op_data;
1357 struct address_space *mapping;
1358 struct lu_dirpage *dp;
1361 struct lustre_handle lockh;
1362 struct ptlrpc_request *enq_req = NULL;
1363 struct readpage_param rp_param;
1370 LASSERT(dir != NULL);
1371 mapping = dir->i_mapping;
1373 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1374 cb_op->md_blocking_ast, 0);
1375 if (enq_req != NULL)
1376 ptlrpc_req_finished(enq_req);
1379 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1380 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1385 mdc_set_lock_data(exp, &it.d.lustre.it_lock_handle, dir, NULL);
1387 rp_param.rp_off = hash_offset;
1388 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1389 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1390 rp_param.rp_hash64);
1392 CERROR("%s: dir page locate: "DFID" at "LPU64": rc %ld\n",
1393 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1394 rp_param.rp_off, PTR_ERR(page));
1395 GOTO(out_unlock, rc = PTR_ERR(page));
1396 } else if (page != NULL) {
1398 * XXX nikita: not entirely correct handling of a corner case:
1399 * suppose hash chain of entries with hash value HASH crosses
1400 * border between pages P0 and P1. First both P0 and P1 are
1401 * cached, seekdir() is called for some entry from the P0 part
1402 * of the chain. Later P0 goes out of cache. telldir(HASH)
1403 * happens and finds P1, as it starts with matching hash
1404 * value. Remaining entries from P0 part of the chain are
1405 * skipped. (Is that really a bug?)
1407 * Possible solutions: 0. don't cache P1 is such case, handle
1408 * it as an "overflow" page. 1. invalidate all pages at
1409 * once. 2. use HASH|1 as an index for P1.
1411 GOTO(hash_collision, page);
1414 rp_param.rp_exp = exp;
1415 rp_param.rp_mod = op_data;
1416 page = read_cache_page(mapping,
1417 hash_x_index(rp_param.rp_off,
1418 rp_param.rp_hash64),
1419 mdc_read_page_remote, &rp_param);
1421 CDEBUG(D_INFO, "%s: read cache page: "DFID" at "LPU64": %ld\n",
1422 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1423 rp_param.rp_off, PTR_ERR(page));
1424 GOTO(out_unlock, rc = PTR_ERR(page));
1427 wait_on_page_locked(page);
1429 if (!PageUptodate(page)) {
1430 CERROR("%s: page not updated: "DFID" at "LPU64": rc %d\n",
1431 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1432 rp_param.rp_off, -5);
1435 if (!PageChecked(page))
1436 SetPageChecked(page);
1437 if (PageError(page)) {
1438 CERROR("%s: page error: "DFID" at "LPU64": rc %d\n",
1439 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1440 rp_param.rp_off, -5);
1445 dp = page_address(page);
1446 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1447 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1448 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1449 rp_param.rp_off = hash_offset >> 32;
1451 start = le64_to_cpu(dp->ldp_hash_start);
1452 end = le64_to_cpu(dp->ldp_hash_end);
1453 rp_param.rp_off = hash_offset;
1456 LASSERT(start == rp_param.rp_off);
1457 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1458 #if BITS_PER_LONG == 32
1459 CWARN("Real page-wide hash collision at ["LPU64" "LPU64"] with "
1460 "hash "LPU64"\n", le64_to_cpu(dp->ldp_hash_start),
1461 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1465 * Fetch whole overflow chain...
1473 lockh.cookie = it.d.lustre.it_lock_handle;
1474 ldlm_lock_decref(&lockh, it.d.lustre.it_lock_mode);
1475 it.d.lustre.it_lock_handle = 0;
1479 mdc_release_page(page, 1);
1485 static int mdc_statfs(const struct lu_env *env,
1486 struct obd_export *exp, struct obd_statfs *osfs,
1487 __u64 max_age, __u32 flags)
1489 struct obd_device *obd = class_exp2obd(exp);
1490 struct ptlrpc_request *req;
1491 struct obd_statfs *msfs;
1492 struct obd_import *imp = NULL;
1497 * Since the request might also come from lprocfs, so we need
1498 * sync this with client_disconnect_export Bug15684
1500 down_read(&obd->u.cli.cl_sem);
1501 if (obd->u.cli.cl_import)
1502 imp = class_import_get(obd->u.cli.cl_import);
1503 up_read(&obd->u.cli.cl_sem);
1507 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_STATFS,
1508 LUSTRE_MDS_VERSION, MDS_STATFS);
1510 GOTO(output, rc = -ENOMEM);
1512 ptlrpc_request_set_replen(req);
1514 if (flags & OBD_STATFS_NODELAY) {
1515 /* procfs requests not want stay in wait for avoid deadlock */
1516 req->rq_no_resend = 1;
1517 req->rq_no_delay = 1;
1520 rc = ptlrpc_queue_wait(req);
1522 /* check connection error first */
1523 if (imp->imp_connect_error)
1524 rc = imp->imp_connect_error;
1528 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1530 GOTO(out, rc = -EPROTO);
1535 ptlrpc_req_finished(req);
1537 class_import_put(imp);
1541 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1543 __u32 keylen, vallen;
1547 if (gf->gf_pathlen > PATH_MAX)
1548 RETURN(-ENAMETOOLONG);
1549 if (gf->gf_pathlen < 2)
1552 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1553 keylen = cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf);
1554 OBD_ALLOC(key, keylen);
1557 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1558 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1560 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n",
1561 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1563 if (!fid_is_sane(&gf->gf_fid))
1564 GOTO(out, rc = -EINVAL);
1566 /* Val is struct getinfo_fid2path result plus path */
1567 vallen = sizeof(*gf) + gf->gf_pathlen;
1569 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1570 if (rc != 0 && rc != -EREMOTE)
1573 if (vallen <= sizeof(*gf))
1574 GOTO(out, rc = -EPROTO);
1575 else if (vallen > sizeof(*gf) + gf->gf_pathlen)
1576 GOTO(out, rc = -EOVERFLOW);
1578 CDEBUG(D_IOCTL, "path get "DFID" from "LPU64" #%d\n%s\n",
1579 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno, gf->gf_path);
1582 OBD_FREE(key, keylen);
1586 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1587 struct hsm_progress_kernel *hpk)
1589 struct obd_import *imp = class_exp2cliimp(exp);
1590 struct hsm_progress_kernel *req_hpk;
1591 struct ptlrpc_request *req;
1595 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1596 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1598 GOTO(out, rc = -ENOMEM);
1600 mdc_pack_body(req, NULL, NULL, OBD_MD_FLRMTPERM, 0, -1, 0);
1602 /* Copy hsm_progress struct */
1603 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1604 if (req_hpk == NULL)
1605 GOTO(out, rc = -EPROTO);
1608 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1610 ptlrpc_request_set_replen(req);
1612 mdc_get_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
1613 rc = ptlrpc_queue_wait(req);
1614 mdc_put_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
1618 ptlrpc_req_finished(req);
1622 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archives)
1624 __u32 *archive_mask;
1625 struct ptlrpc_request *req;
1629 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_REGISTER,
1631 MDS_HSM_CT_REGISTER);
1633 GOTO(out, rc = -ENOMEM);
1635 mdc_pack_body(req, NULL, NULL, OBD_MD_FLRMTPERM, 0, -1, 0);
1637 /* Copy hsm_progress struct */
1638 archive_mask = req_capsule_client_get(&req->rq_pill,
1639 &RMF_MDS_HSM_ARCHIVE);
1640 if (archive_mask == NULL)
1641 GOTO(out, rc = -EPROTO);
1643 *archive_mask = archives;
1645 ptlrpc_request_set_replen(req);
1647 rc = mdc_queue_wait(req);
1650 ptlrpc_req_finished(req);
1654 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1655 struct md_op_data *op_data)
1657 struct hsm_current_action *hca = op_data->op_data;
1658 struct hsm_current_action *req_hca;
1659 struct ptlrpc_request *req;
1663 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1664 &RQF_MDS_HSM_ACTION);
1668 mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
1670 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1672 ptlrpc_request_free(req);
1676 mdc_pack_body(req, &op_data->op_fid1, op_data->op_capa1,
1677 OBD_MD_FLRMTPERM, 0, op_data->op_suppgids[0], 0);
1679 ptlrpc_request_set_replen(req);
1681 rc = mdc_queue_wait(req);
1685 req_hca = req_capsule_server_get(&req->rq_pill,
1686 &RMF_MDS_HSM_CURRENT_ACTION);
1687 if (req_hca == NULL)
1688 GOTO(out, rc = -EPROTO);
1694 ptlrpc_req_finished(req);
1698 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1700 struct ptlrpc_request *req;
1704 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1706 MDS_HSM_CT_UNREGISTER);
1708 GOTO(out, rc = -ENOMEM);
1710 mdc_pack_body(req, NULL, NULL, OBD_MD_FLRMTPERM, 0, -1, 0);
1712 ptlrpc_request_set_replen(req);
1714 rc = mdc_queue_wait(req);
1717 ptlrpc_req_finished(req);
1721 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1722 struct md_op_data *op_data)
1724 struct hsm_user_state *hus = op_data->op_data;
1725 struct hsm_user_state *req_hus;
1726 struct ptlrpc_request *req;
1730 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1731 &RQF_MDS_HSM_STATE_GET);
1735 mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
1737 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1739 ptlrpc_request_free(req);
1743 mdc_pack_body(req, &op_data->op_fid1, op_data->op_capa1,
1744 OBD_MD_FLRMTPERM, 0, op_data->op_suppgids[0], 0);
1746 ptlrpc_request_set_replen(req);
1748 rc = mdc_queue_wait(req);
1752 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1753 if (req_hus == NULL)
1754 GOTO(out, rc = -EPROTO);
1760 ptlrpc_req_finished(req);
1764 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1765 struct md_op_data *op_data)
1767 struct hsm_state_set *hss = op_data->op_data;
1768 struct hsm_state_set *req_hss;
1769 struct ptlrpc_request *req;
1773 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1774 &RQF_MDS_HSM_STATE_SET);
1778 mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
1780 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1782 ptlrpc_request_free(req);
1786 mdc_pack_body(req, &op_data->op_fid1, op_data->op_capa1,
1787 OBD_MD_FLRMTPERM, 0, op_data->op_suppgids[0], 0);
1790 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1791 if (req_hss == NULL)
1792 GOTO(out, rc = -EPROTO);
1795 ptlrpc_request_set_replen(req);
1797 mdc_get_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
1798 rc = ptlrpc_queue_wait(req);
1799 mdc_put_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
1803 ptlrpc_req_finished(req);
1807 static int mdc_ioc_hsm_request(struct obd_export *exp,
1808 struct hsm_user_request *hur)
1810 struct obd_import *imp = class_exp2cliimp(exp);
1811 struct ptlrpc_request *req;
1812 struct hsm_request *req_hr;
1813 struct hsm_user_item *req_hui;
1818 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1820 GOTO(out, rc = -ENOMEM);
1822 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1823 hur->hur_request.hr_itemcount
1824 * sizeof(struct hsm_user_item));
1825 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1826 hur->hur_request.hr_data_len);
1828 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1830 ptlrpc_request_free(req);
1834 mdc_pack_body(req, NULL, NULL, OBD_MD_FLRMTPERM, 0, -1, 0);
1836 /* Copy hsm_request struct */
1837 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1839 GOTO(out, rc = -EPROTO);
1840 *req_hr = hur->hur_request;
1842 /* Copy hsm_user_item structs */
1843 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1844 if (req_hui == NULL)
1845 GOTO(out, rc = -EPROTO);
1846 memcpy(req_hui, hur->hur_user_item,
1847 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1849 /* Copy opaque field */
1850 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1851 if (req_opaque == NULL)
1852 GOTO(out, rc = -EPROTO);
1853 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1855 ptlrpc_request_set_replen(req);
1857 mdc_get_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
1858 rc = ptlrpc_queue_wait(req);
1859 mdc_put_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
1864 ptlrpc_req_finished(req);
1868 static struct kuc_hdr *changelog_kuc_hdr(char *buf, size_t len, __u32 flags)
1870 struct kuc_hdr *lh = (struct kuc_hdr *)buf;
1872 LASSERT(len <= KUC_CHANGELOG_MSG_MAXSIZE);
1874 lh->kuc_magic = KUC_MAGIC;
1875 lh->kuc_transport = KUC_TRANSPORT_CHANGELOG;
1876 lh->kuc_flags = flags;
1877 lh->kuc_msgtype = CL_RECORD;
1878 lh->kuc_msglen = len;
1882 struct changelog_show {
1884 enum changelog_send_flag cs_flags;
1887 struct obd_device *cs_obd;
1890 static inline char *cs_obd_name(struct changelog_show *cs)
1892 return cs->cs_obd->obd_name;
1895 static int changelog_kkuc_cb(const struct lu_env *env, struct llog_handle *llh,
1896 struct llog_rec_hdr *hdr, void *data)
1898 struct changelog_show *cs = data;
1899 struct llog_changelog_rec *rec = (struct llog_changelog_rec *)hdr;
1905 if (rec->cr_hdr.lrh_type != CHANGELOG_REC) {
1907 CERROR("%s: not a changelog rec %x/%d: rc = %d\n",
1908 cs_obd_name(cs), rec->cr_hdr.lrh_type,
1909 rec->cr.cr_type, rc);
1913 if (rec->cr.cr_index < cs->cs_startrec) {
1914 /* Skip entries earlier than what we are interested in */
1915 CDEBUG(D_HSM, "rec="LPU64" start="LPU64"\n",
1916 rec->cr.cr_index, cs->cs_startrec);
1920 CDEBUG(D_HSM, LPU64" %02d%-5s "LPU64" 0x%x t="DFID" p="DFID" %.*s\n",
1921 rec->cr.cr_index, rec->cr.cr_type,
1922 changelog_type2str(rec->cr.cr_type), rec->cr.cr_time,
1923 rec->cr.cr_flags & CLF_FLAGMASK,
1924 PFID(&rec->cr.cr_tfid), PFID(&rec->cr.cr_pfid),
1925 rec->cr.cr_namelen, changelog_rec_name(&rec->cr));
1927 len = sizeof(*lh) + changelog_rec_size(&rec->cr) + rec->cr.cr_namelen;
1929 /* Set up the message */
1930 lh = changelog_kuc_hdr(cs->cs_buf, len, cs->cs_flags);
1931 memcpy(lh + 1, &rec->cr, len - sizeof(*lh));
1933 rc = libcfs_kkuc_msg_put(cs->cs_fp, lh);
1934 CDEBUG(D_HSM, "kucmsg fp %p len %zu rc %d\n", cs->cs_fp, len, rc);
1939 static int mdc_changelog_send_thread(void *csdata)
1941 struct changelog_show *cs = csdata;
1942 struct llog_ctxt *ctxt = NULL;
1943 struct llog_handle *llh = NULL;
1944 struct kuc_hdr *kuch;
1945 enum llog_flag flags = LLOG_F_IS_CAT;
1948 CDEBUG(D_HSM, "changelog to fp=%p start "LPU64"\n",
1949 cs->cs_fp, cs->cs_startrec);
1951 OBD_ALLOC(cs->cs_buf, KUC_CHANGELOG_MSG_MAXSIZE);
1952 if (cs->cs_buf == NULL)
1953 GOTO(out, rc = -ENOMEM);
1955 /* Set up the remote catalog handle */
1956 ctxt = llog_get_context(cs->cs_obd, LLOG_CHANGELOG_REPL_CTXT);
1958 GOTO(out, rc = -ENOENT);
1959 rc = llog_open(NULL, ctxt, &llh, NULL, CHANGELOG_CATALOG,
1962 CERROR("%s: fail to open changelog catalog: rc = %d\n",
1963 cs_obd_name(cs), rc);
1967 if (cs->cs_flags & CHANGELOG_FLAG_JOBID)
1968 flags |= LLOG_F_EXT_JOBID;
1970 rc = llog_init_handle(NULL, llh, flags, NULL);
1972 CERROR("llog_init_handle failed %d\n", rc);
1976 rc = llog_cat_process(NULL, llh, changelog_kkuc_cb, cs, 0, 0);
1978 /* Send EOF no matter what our result */
1979 kuch = changelog_kuc_hdr(cs->cs_buf, sizeof(*kuch), cs->cs_flags);
1980 kuch->kuc_msgtype = CL_EOF;
1981 libcfs_kkuc_msg_put(cs->cs_fp, kuch);
1986 llog_cat_close(NULL, llh);
1988 llog_ctxt_put(ctxt);
1990 OBD_FREE(cs->cs_buf, KUC_CHANGELOG_MSG_MAXSIZE);
1995 static int mdc_ioc_changelog_send(struct obd_device *obd,
1996 struct ioc_changelog *icc)
1998 struct changelog_show *cs;
1999 struct task_struct *task;
2002 /* Freed in mdc_changelog_send_thread */
2008 cs->cs_startrec = icc->icc_recno;
2009 /* matching fput in mdc_changelog_send_thread */
2010 cs->cs_fp = fget(icc->icc_id);
2011 cs->cs_flags = icc->icc_flags;
2014 * New thread because we should return to user app before
2015 * writing into our pipe
2017 task = kthread_run(mdc_changelog_send_thread, cs,
2018 "mdc_clg_send_thread");
2021 CERROR("%s: cannot start changelog thread: rc = %d\n",
2022 cs_obd_name(cs), rc);
2026 CDEBUG(D_HSM, "%s: started changelog thread\n",
2033 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2034 struct lustre_kernelcomm *lk);
2036 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2037 struct obd_quotactl *oqctl)
2039 struct ptlrpc_request *req;
2040 struct obd_quotactl *oqc;
2044 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
2045 &RQF_MDS_QUOTACTL, LUSTRE_MDS_VERSION,
2050 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2053 ptlrpc_request_set_replen(req);
2054 ptlrpc_at_set_req_timeout(req);
2055 req->rq_no_resend = 1;
2057 rc = ptlrpc_queue_wait(req);
2059 CERROR("ptlrpc_queue_wait failed, rc: %d\n", rc);
2061 if (req->rq_repmsg &&
2062 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2065 CERROR ("Can't unpack obd_quotactl\n");
2068 ptlrpc_req_finished(req);
2073 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2074 struct md_op_data *op_data)
2076 struct list_head cancels = LIST_HEAD_INIT(cancels);
2077 struct ptlrpc_request *req;
2079 struct mdc_swap_layouts *msl, *payload;
2082 msl = op_data->op_data;
2084 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2085 * first thing it will do is to cancel the 2 layout
2086 * locks held by this client.
2087 * So the client must cancel its layout locks on the 2 fids
2088 * with the request RPC to avoid extra RPC round trips.
2090 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2091 LCK_EX, MDS_INODELOCK_LAYOUT |
2092 MDS_INODELOCK_XATTR);
2093 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2094 LCK_EX, MDS_INODELOCK_LAYOUT |
2095 MDS_INODELOCK_XATTR);
2097 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2098 &RQF_MDS_SWAP_LAYOUTS);
2100 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2104 mdc_set_capa_size(req, &RMF_CAPA1, op_data->op_capa1);
2105 mdc_set_capa_size(req, &RMF_CAPA2, op_data->op_capa2);
2107 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2109 ptlrpc_request_free(req);
2113 mdc_swap_layouts_pack(req, op_data);
2115 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2120 ptlrpc_request_set_replen(req);
2122 rc = ptlrpc_queue_wait(req);
2128 ptlrpc_req_finished(req);
2132 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2133 void *karg, void __user *uarg)
2135 struct obd_device *obd = exp->exp_obd;
2136 struct obd_ioctl_data *data = karg;
2137 struct obd_import *imp = obd->u.cli.cl_import;
2141 if (!try_module_get(THIS_MODULE)) {
2142 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2143 module_name(THIS_MODULE));
2147 case OBD_IOC_CHANGELOG_SEND:
2148 rc = mdc_ioc_changelog_send(obd, karg);
2150 case OBD_IOC_CHANGELOG_CLEAR: {
2151 struct ioc_changelog *icc = karg;
2152 struct changelog_setinfo cs =
2153 {.cs_recno = icc->icc_recno, .cs_id = icc->icc_id};
2154 rc = obd_set_info_async(NULL, exp, strlen(KEY_CHANGELOG_CLEAR),
2155 KEY_CHANGELOG_CLEAR, sizeof(cs), &cs,
2159 case OBD_IOC_FID2PATH:
2160 rc = mdc_ioc_fid2path(exp, karg);
2162 case LL_IOC_HSM_CT_START:
2163 rc = mdc_ioc_hsm_ct_start(exp, karg);
2164 /* ignore if it was already registered on this MDS. */
2168 case LL_IOC_HSM_PROGRESS:
2169 rc = mdc_ioc_hsm_progress(exp, karg);
2171 case LL_IOC_HSM_STATE_GET:
2172 rc = mdc_ioc_hsm_state_get(exp, karg);
2174 case LL_IOC_HSM_STATE_SET:
2175 rc = mdc_ioc_hsm_state_set(exp, karg);
2177 case LL_IOC_HSM_ACTION:
2178 rc = mdc_ioc_hsm_current_action(exp, karg);
2180 case LL_IOC_HSM_REQUEST:
2181 rc = mdc_ioc_hsm_request(exp, karg);
2183 case OBD_IOC_CLIENT_RECOVER:
2184 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2188 case IOC_OSC_SET_ACTIVE:
2189 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2191 case OBD_IOC_PING_TARGET:
2192 rc = ptlrpc_obd_ping(obd);
2195 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2196 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2197 * there'd be no LMV layer thus we might be called here. Eventually
2198 * this code should be removed.
2201 case IOC_OBD_STATFS: {
2202 struct obd_statfs stat_buf = {0};
2204 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2205 GOTO(out, rc = -ENODEV);
2208 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2209 min((int)data->ioc_plen2,
2210 (int)sizeof(struct obd_uuid))))
2211 GOTO(out, rc = -EFAULT);
2213 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2214 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
2219 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2220 min((int) data->ioc_plen1,
2221 (int) sizeof(stat_buf))))
2222 GOTO(out, rc = -EFAULT);
2226 case OBD_IOC_QUOTACTL: {
2227 struct if_quotactl *qctl = karg;
2228 struct obd_quotactl *oqctl;
2230 OBD_ALLOC_PTR(oqctl);
2232 GOTO(out, rc = -ENOMEM);
2234 QCTL_COPY(oqctl, qctl);
2235 rc = obd_quotactl(exp, oqctl);
2237 QCTL_COPY(qctl, oqctl);
2238 qctl->qc_valid = QC_MDTIDX;
2239 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2242 OBD_FREE_PTR(oqctl);
2245 case LL_IOC_GET_CONNECT_FLAGS:
2246 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2247 sizeof(*exp_connect_flags_ptr(exp))))
2248 GOTO(out, rc = -EFAULT);
2251 case LL_IOC_LOV_SWAP_LAYOUTS:
2252 rc = mdc_ioc_swap_layouts(exp, karg);
2255 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2256 GOTO(out, rc = -ENOTTY);
2259 module_put(THIS_MODULE);
2264 static int mdc_get_info_rpc(struct obd_export *exp,
2265 u32 keylen, void *key,
2266 u32 vallen, void *val)
2268 struct obd_import *imp = class_exp2cliimp(exp);
2269 struct ptlrpc_request *req;
2274 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2278 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2279 RCL_CLIENT, keylen);
2280 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2281 RCL_CLIENT, sizeof(vallen));
2283 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2285 ptlrpc_request_free(req);
2289 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2290 memcpy(tmp, key, keylen);
2291 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2292 memcpy(tmp, &vallen, sizeof(vallen));
2294 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2295 RCL_SERVER, vallen);
2296 ptlrpc_request_set_replen(req);
2298 rc = ptlrpc_queue_wait(req);
2299 /* -EREMOTE means the get_info result is partial, and it needs to
2300 * continue on another MDT, see fid2path part in lmv_iocontrol */
2301 if (rc == 0 || rc == -EREMOTE) {
2302 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2303 memcpy(val, tmp, vallen);
2304 if (ptlrpc_rep_need_swab(req)) {
2305 if (KEY_IS(KEY_FID2PATH))
2306 lustre_swab_fid2path(val);
2309 ptlrpc_req_finished(req);
2314 static void lustre_swab_hai(struct hsm_action_item *h)
2316 __swab32s(&h->hai_len);
2317 __swab32s(&h->hai_action);
2318 lustre_swab_lu_fid(&h->hai_fid);
2319 lustre_swab_lu_fid(&h->hai_dfid);
2320 __swab64s(&h->hai_cookie);
2321 __swab64s(&h->hai_extent.offset);
2322 __swab64s(&h->hai_extent.length);
2323 __swab64s(&h->hai_gid);
2326 static void lustre_swab_hal(struct hsm_action_list *h)
2328 struct hsm_action_item *hai;
2331 __swab32s(&h->hal_version);
2332 __swab32s(&h->hal_count);
2333 __swab32s(&h->hal_archive_id);
2334 __swab64s(&h->hal_flags);
2336 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2337 lustre_swab_hai(hai);
2340 static void lustre_swab_kuch(struct kuc_hdr *l)
2342 __swab16s(&l->kuc_magic);
2343 /* __u8 l->kuc_transport */
2344 __swab16s(&l->kuc_msgtype);
2345 __swab16s(&l->kuc_msglen);
2348 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2349 struct lustre_kernelcomm *lk)
2351 struct obd_import *imp = class_exp2cliimp(exp);
2352 __u32 archive = lk->lk_data;
2355 if (lk->lk_group != KUC_GRP_HSM) {
2356 CERROR("Bad copytool group %d\n", lk->lk_group);
2360 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2361 lk->lk_uid, lk->lk_group, lk->lk_flags);
2363 if (lk->lk_flags & LK_FLG_STOP) {
2364 /* Unregister with the coordinator */
2365 rc = mdc_ioc_hsm_ct_unregister(imp);
2367 rc = mdc_ioc_hsm_ct_register(imp, archive);
2374 * Send a message to any listening copytools
2375 * @param val KUC message (kuc_hdr + hsm_action_list)
2376 * @param len total length of message
2378 static int mdc_hsm_copytool_send(size_t len, void *val)
2380 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2381 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2385 if (len < sizeof(*lh) + sizeof(*hal)) {
2386 CERROR("Short HSM message %zu < %zu\n", len,
2387 sizeof(*lh) + sizeof(*hal));
2390 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2391 lustre_swab_kuch(lh);
2392 lustre_swab_hal(hal);
2393 } else if (lh->kuc_magic != KUC_MAGIC) {
2394 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2398 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2400 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2401 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2403 /* Broadcast to HSM listeners */
2404 rc = libcfs_kkuc_group_put(KUC_GRP_HSM, lh);
2410 * callback function passed to kuc for re-registering each HSM copytool
2411 * running on MDC, after MDT shutdown/recovery.
2412 * @param data copytool registration data
2413 * @param cb_arg callback argument (obd_import)
2415 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2417 struct kkuc_ct_data *kcd = data;
2418 struct obd_import *imp = (struct obd_import *)cb_arg;
2421 if (kcd == NULL || kcd->kcd_magic != KKUC_CT_DATA_MAGIC)
2424 if (!obd_uuid_equals(&kcd->kcd_uuid, &imp->imp_obd->obd_uuid))
2427 CDEBUG(D_HA, "%s: recover copytool registration to MDT (archive=%#x)\n",
2428 imp->imp_obd->obd_name, kcd->kcd_archive);
2429 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_archive);
2431 /* ignore error if the copytool is already registered */
2432 return (rc == -EEXIST) ? 0 : rc;
2436 * Re-establish all kuc contexts with MDT
2437 * after MDT shutdown/recovery.
2439 static int mdc_kuc_reregister(struct obd_import *imp)
2441 /* re-register HSM agents */
2442 return libcfs_kkuc_group_foreach(KUC_GRP_HSM, mdc_hsm_ct_reregister,
2446 static int mdc_set_info_async(const struct lu_env *env,
2447 struct obd_export *exp,
2448 u32 keylen, void *key,
2449 u32 vallen, void *val,
2450 struct ptlrpc_request_set *set)
2452 struct obd_import *imp = class_exp2cliimp(exp);
2456 if (KEY_IS(KEY_READ_ONLY)) {
2457 if (vallen != sizeof(int))
2460 spin_lock(&imp->imp_lock);
2461 if (*((int *)val)) {
2462 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2463 imp->imp_connect_data.ocd_connect_flags |=
2466 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2467 imp->imp_connect_data.ocd_connect_flags &=
2468 ~OBD_CONNECT_RDONLY;
2470 spin_unlock(&imp->imp_lock);
2472 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2473 keylen, key, vallen, val, set);
2476 if (KEY_IS(KEY_SPTLRPC_CONF)) {
2477 sptlrpc_conf_client_adapt(exp->exp_obd);
2480 if (KEY_IS(KEY_FLUSH_CTX)) {
2481 sptlrpc_import_flush_my_ctx(imp);
2484 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2485 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2486 keylen, key, vallen, val, set);
2489 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2490 rc = mdc_hsm_copytool_send(vallen, val);
2494 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2495 __u32 *default_easize = val;
2497 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2501 CERROR("Unknown key %s\n", (char *)key);
2505 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2506 __u32 keylen, void *key, __u32 *vallen, void *val)
2510 if (KEY_IS(KEY_MAX_EASIZE)) {
2511 __u32 mdsize, *max_easize;
2513 if (*vallen != sizeof(int))
2515 mdsize = *(__u32 *)val;
2516 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2517 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2519 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2521 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2522 __u32 *default_easize;
2524 if (*vallen != sizeof(int))
2526 default_easize = val;
2527 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2529 } else if (KEY_IS(KEY_CONN_DATA)) {
2530 struct obd_import *imp = class_exp2cliimp(exp);
2531 struct obd_connect_data *data = val;
2533 if (*vallen != sizeof(*data))
2536 *data = imp->imp_connect_data;
2538 } else if (KEY_IS(KEY_TGT_COUNT)) {
2539 *((__u32 *)val) = 1;
2543 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2548 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2549 struct obd_capa *oc, struct ptlrpc_request **request)
2551 struct ptlrpc_request *req;
2556 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2560 mdc_set_capa_size(req, &RMF_CAPA1, oc);
2562 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2564 ptlrpc_request_free(req);
2568 mdc_pack_body(req, fid, oc, 0, 0, -1, 0);
2570 ptlrpc_request_set_replen(req);
2572 rc = ptlrpc_queue_wait(req);
2574 ptlrpc_req_finished(req);
2580 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2581 enum obd_import_event event)
2585 LASSERT(imp->imp_obd == obd);
2588 case IMP_EVENT_DISCON: {
2590 /* XXX Pass event up to OBDs stack. used only for FLD now */
2591 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_DISCON, NULL);
2595 case IMP_EVENT_INACTIVE: {
2596 struct client_obd *cli = &obd->u.cli;
2598 * Flush current sequence to make client obtain new one
2599 * from server in case of disconnect/reconnect.
2601 if (cli->cl_seq != NULL)
2602 seq_client_flush(cli->cl_seq);
2604 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE, NULL);
2607 case IMP_EVENT_INVALIDATE: {
2608 struct ldlm_namespace *ns = obd->obd_namespace;
2610 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2614 case IMP_EVENT_ACTIVE:
2615 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE, NULL);
2616 /* redo the kuc registration after reconnecting */
2618 rc = mdc_kuc_reregister(imp);
2621 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD, NULL);
2623 case IMP_EVENT_DEACTIVATE:
2624 case IMP_EVENT_ACTIVATE:
2627 CERROR("Unknown import event %x\n", event);
2633 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2634 struct lu_fid *fid, struct md_op_data *op_data)
2636 struct client_obd *cli = &exp->exp_obd->u.cli;
2637 struct lu_client_seq *seq = cli->cl_seq;
2639 RETURN(seq_client_alloc_fid(env, seq, fid));
2642 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2644 struct client_obd *cli = &exp->exp_obd->u.cli;
2645 return &cli->cl_target_uuid;
2649 * Determine whether the lock can be canceled before replaying it during
2650 * recovery, non zero value will be return if the lock can be canceled,
2651 * or zero returned for not
2653 static int mdc_cancel_weight(struct ldlm_lock *lock)
2655 if (lock->l_resource->lr_type != LDLM_IBITS)
2658 /* FIXME: if we ever get into a situation where there are too many
2659 * opened files with open locks on a single node, then we really
2660 * should replay these open locks to reget it */
2661 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2667 static int mdc_resource_inode_free(struct ldlm_resource *res)
2669 if (res->lr_lvb_inode)
2670 res->lr_lvb_inode = NULL;
2675 static struct ldlm_valblock_ops inode_lvbo = {
2676 .lvbo_free = mdc_resource_inode_free
2679 static int mdc_llog_init(struct obd_device *obd)
2681 struct obd_llog_group *olg = &obd->obd_olg;
2682 struct llog_ctxt *ctxt;
2687 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2692 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2693 llog_initiator_connect(ctxt);
2694 llog_ctxt_put(ctxt);
2699 static void mdc_llog_finish(struct obd_device *obd)
2701 struct llog_ctxt *ctxt;
2705 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2707 llog_cleanup(NULL, ctxt);
2712 static int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2714 struct client_obd *cli = &obd->u.cli;
2718 OBD_ALLOC(cli->cl_rpc_lock, sizeof (*cli->cl_rpc_lock));
2719 if (!cli->cl_rpc_lock)
2721 mdc_init_rpc_lock(cli->cl_rpc_lock);
2723 rc = ptlrpcd_addref();
2725 GOTO(err_rpc_lock, rc);
2727 OBD_ALLOC(cli->cl_close_lock, sizeof (*cli->cl_close_lock));
2728 if (!cli->cl_close_lock)
2729 GOTO(err_ptlrpcd_decref, rc = -ENOMEM);
2730 mdc_init_rpc_lock(cli->cl_close_lock);
2732 rc = client_obd_setup(obd, cfg);
2734 GOTO(err_close_lock, rc);
2735 #ifdef CONFIG_PROC_FS
2736 obd->obd_vars = lprocfs_mdc_obd_vars;
2737 lprocfs_obd_setup(obd);
2738 lprocfs_alloc_md_stats(obd, 0);
2740 sptlrpc_lprocfs_cliobd_attach(obd);
2741 ptlrpc_lprocfs_register_obd(obd);
2743 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2745 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2747 rc = mdc_llog_init(obd);
2750 CERROR("failed to setup llogging subsystems\n");
2754 spin_lock_init(&cli->cl_mod_rpcs_lock);
2755 cli->cl_max_mod_rpcs_in_flight = OBD_MAX_RIF_DEFAULT - 1;
2760 OBD_FREE(cli->cl_close_lock, sizeof (*cli->cl_close_lock));
2764 OBD_FREE(cli->cl_rpc_lock, sizeof (*cli->cl_rpc_lock));
2768 /* Initialize the default and maximum LOV EA sizes. This allows
2769 * us to make MDS RPCs with large enough reply buffers to hold a default
2770 * sized EA without having to calculate this (via a call into the
2771 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2772 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2773 * a large number of stripes is possible. If a larger reply buffer is
2774 * required it will be reallocated in the ptlrpc layer due to overflow.
2776 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2779 struct obd_device *obd = exp->exp_obd;
2780 struct client_obd *cli = &obd->u.cli;
2783 if (cli->cl_max_mds_easize < easize)
2784 cli->cl_max_mds_easize = easize;
2786 if (cli->cl_default_mds_easize < def_easize)
2787 cli->cl_default_mds_easize = def_easize;
2792 static int mdc_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2798 case OBD_CLEANUP_EARLY:
2800 case OBD_CLEANUP_EXPORTS:
2801 /* Failsafe, ok if racy */
2802 if (obd->obd_type->typ_refcnt <= 1)
2803 libcfs_kkuc_group_rem(0, KUC_GRP_HSM, NULL);
2805 obd_cleanup_client_import(obd);
2806 ptlrpc_lprocfs_unregister_obd(obd);
2807 lprocfs_obd_cleanup(obd);
2808 lprocfs_free_md_stats(obd);
2809 mdc_llog_finish(obd);
2815 static int mdc_cleanup(struct obd_device *obd)
2817 struct client_obd *cli = &obd->u.cli;
2819 OBD_FREE(cli->cl_rpc_lock, sizeof (*cli->cl_rpc_lock));
2820 OBD_FREE(cli->cl_close_lock, sizeof (*cli->cl_close_lock));
2824 return client_obd_cleanup(obd);
2827 static int mdc_process_config(struct obd_device *obd, size_t len, void *buf)
2829 struct lustre_cfg *lcfg = buf;
2830 int rc = class_process_proc_param(PARAM_MDC, obd->obd_vars, lcfg, obd);
2831 return (rc > 0 ? 0: rc);
2835 /* get remote permission for current user on fid */
2836 static int mdc_get_remote_perm(struct obd_export *exp, const struct lu_fid *fid,
2837 struct obd_capa *oc, __u32 suppgid,
2838 struct ptlrpc_request **request)
2840 struct ptlrpc_request *req;
2844 LASSERT(client_is_remote(exp));
2847 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_GETATTR);
2851 mdc_set_capa_size(req, &RMF_CAPA1, oc);
2853 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
2855 ptlrpc_request_free(req);
2859 mdc_pack_body(req, fid, oc, OBD_MD_FLRMTPERM, 0, suppgid, 0);
2861 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER,
2862 sizeof(struct mdt_remote_perm));
2864 ptlrpc_request_set_replen(req);
2866 rc = ptlrpc_queue_wait(req);
2868 ptlrpc_req_finished(req);
2874 static int mdc_interpret_renew_capa(const struct lu_env *env,
2875 struct ptlrpc_request *req, void *args,
2878 struct mdc_renew_capa_args *ra = args;
2879 struct mdt_body *body = NULL;
2880 struct lustre_capa *capa;
2884 GOTO(out, capa = ERR_PTR(status));
2886 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
2888 GOTO(out, capa = ERR_PTR(-EFAULT));
2890 if ((body->mbo_valid & OBD_MD_FLOSSCAPA) == 0)
2891 GOTO(out, capa = ERR_PTR(-ENOENT));
2893 capa = req_capsule_server_get(&req->rq_pill, &RMF_CAPA2);
2895 GOTO(out, capa = ERR_PTR(-EFAULT));
2898 ra->ra_cb(ra->ra_oc, capa);
2902 static int mdc_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2905 struct ptlrpc_request *req;
2906 struct mdc_renew_capa_args *ra;
2909 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_GETATTR,
2910 LUSTRE_MDS_VERSION, MDS_GETATTR);
2914 /* NB, OBD_MD_FLOSSCAPA is set here, but it doesn't necessarily mean the
2915 * capa to renew is oss capa.
2917 mdc_pack_body(req, &oc->c_capa.lc_fid, oc, OBD_MD_FLOSSCAPA, 0, -1, 0);
2918 ptlrpc_request_set_replen(req);
2920 CLASSERT(sizeof(*ra) <= sizeof(req->rq_async_args));
2921 ra = ptlrpc_req_async_args(req);
2924 req->rq_interpret_reply = mdc_interpret_renew_capa;
2925 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2929 static struct obd_ops mdc_obd_ops = {
2930 .o_owner = THIS_MODULE,
2931 .o_setup = mdc_setup,
2932 .o_precleanup = mdc_precleanup,
2933 .o_cleanup = mdc_cleanup,
2934 .o_add_conn = client_import_add_conn,
2935 .o_del_conn = client_import_del_conn,
2936 .o_connect = client_connect_import,
2937 .o_disconnect = client_disconnect_export,
2938 .o_iocontrol = mdc_iocontrol,
2939 .o_set_info_async = mdc_set_info_async,
2940 .o_statfs = mdc_statfs,
2941 .o_fid_init = client_fid_init,
2942 .o_fid_fini = client_fid_fini,
2943 .o_fid_alloc = mdc_fid_alloc,
2944 .o_import_event = mdc_import_event,
2945 .o_get_info = mdc_get_info,
2946 .o_process_config = mdc_process_config,
2947 .o_get_uuid = mdc_get_uuid,
2948 .o_quotactl = mdc_quotactl,
2951 static struct md_ops mdc_md_ops = {
2952 .m_getstatus = mdc_getstatus,
2953 .m_null_inode = mdc_null_inode,
2954 .m_find_cbdata = mdc_find_cbdata,
2955 .m_close = mdc_close,
2956 .m_create = mdc_create,
2957 .m_enqueue = mdc_enqueue,
2958 .m_getattr = mdc_getattr,
2959 .m_getattr_name = mdc_getattr_name,
2960 .m_intent_lock = mdc_intent_lock,
2962 .m_rename = mdc_rename,
2963 .m_setattr = mdc_setattr,
2964 .m_setxattr = mdc_setxattr,
2965 .m_getxattr = mdc_getxattr,
2966 .m_fsync = mdc_fsync,
2967 .m_read_page = mdc_read_page,
2968 .m_unlink = mdc_unlink,
2969 .m_cancel_unused = mdc_cancel_unused,
2970 .m_init_ea_size = mdc_init_ea_size,
2971 .m_set_lock_data = mdc_set_lock_data,
2972 .m_lock_match = mdc_lock_match,
2973 .m_get_lustre_md = mdc_get_lustre_md,
2974 .m_free_lustre_md = mdc_free_lustre_md,
2975 .m_set_open_replay_data = mdc_set_open_replay_data,
2976 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2977 .m_renew_capa = mdc_renew_capa,
2978 .m_unpack_capa = mdc_unpack_capa,
2979 .m_get_remote_perm = mdc_get_remote_perm,
2980 .m_intent_getattr_async = mdc_intent_getattr_async,
2981 .m_revalidate_lock = mdc_revalidate_lock
2984 static int __init mdc_init(void)
2986 return class_register_type(&mdc_obd_ops, &mdc_md_ops, true, NULL,
2987 LUSTRE_MDC_NAME, NULL);
2990 static void /*__exit*/ mdc_exit(void)
2992 class_unregister_type(LUSTRE_MDC_NAME);
2995 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2996 MODULE_DESCRIPTION("Lustre Metadata Client");
2997 MODULE_LICENSE("GPL");
2999 module_init(mdc_init);
3000 module_exit(mdc_exit);