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/
32 #define DEBUG_SUBSYSTEM S_MDC
34 #include <linux/init.h>
35 #include <linux/kthread.h>
36 #include <linux/module.h>
37 #include <linux/pagemap.h>
38 #include <linux/user_namespace.h>
39 #include <linux/utsname.h>
40 #include <linux/delay.h>
41 #include <linux/uidgid.h>
42 #include <linux/device.h>
43 #include <linux/xarray.h>
45 #include <lustre_errno.h>
47 #include <cl_object.h>
48 #include <llog_swab.h>
49 #include <lprocfs_status.h>
50 #include <lustre_acl.h>
51 #include <lustre_compat.h>
52 #include <lustre_fid.h>
53 #include <uapi/linux/lustre/lustre_ioctl.h>
54 #include <lustre_kernelcomm.h>
55 #include <lustre_lmv.h>
56 #include <lustre_log.h>
57 #include <lustre_swab.h>
58 #include <obd_class.h>
59 #include <lustre_osc.h>
61 #include "mdc_internal.h"
63 #define REQUEST_MINOR 244
65 static int mdc_cleanup(struct obd_device *obd);
67 static inline int mdc_queue_wait(struct ptlrpc_request *req)
69 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
72 /* obd_get_request_slot() ensures that this client has no more
73 * than cl_max_rpcs_in_flight RPCs simultaneously inf light
75 rc = obd_get_request_slot(cli);
79 rc = ptlrpc_queue_wait(req);
80 obd_put_request_slot(cli);
86 * Send MDS_GET_ROOT RPC to fetch root FID.
88 * If \a fileset is not NULL it should contain a subdirectory off
89 * the ROOT/ directory to be mounted on the client. Return the FID
90 * of the subdirectory to the client to mount onto its mountpoint.
92 * \param[in] imp MDC import
93 * \param[in] fileset fileset name, which could be NULL
94 * \param[out] rootfid root FID of this mountpoint
95 * \param[out] pc root capa will be unpacked and saved in this pointer
97 * \retval 0 on success, negative errno on failure
99 static int mdc_get_root(struct obd_export *exp, const char *fileset,
100 struct lu_fid *rootfid)
102 struct ptlrpc_request *req;
103 struct mdt_body *body;
108 if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
111 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
117 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
118 strlen(fileset) + 1);
119 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
121 ptlrpc_request_free(req);
124 mdc_pack_body(req, NULL, 0, 0, -1, 0);
125 if (fileset != NULL) {
126 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
128 memcpy(name, fileset, strlen(fileset));
130 lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
131 req->rq_send_state = LUSTRE_IMP_FULL;
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 *rootfid = body->mbo_fid1;
144 CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
145 PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
148 ptlrpc_req_finished(req);
154 * This function now is known to always saying that it will receive 4 buffers
155 * from server. Even for cases when acl_size and md_size is zero, RPC header
156 * will contain 4 fields and RPC itself will contain zero size fields. This is
157 * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
158 * and thus zero, it shrinks it, making zero size. The same story about
159 * md_size. And this is course of problem when client waits for smaller number
160 * of fields. This issue will be fixed later when client gets aware of RPC
163 static int mdc_getattr_common(struct obd_export *exp,
164 struct ptlrpc_request *req)
166 struct req_capsule *pill = &req->rq_pill;
167 struct mdt_body *body;
172 /* Request message already built. */
173 rc = ptlrpc_queue_wait(req);
177 /* sanity check for the reply */
178 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
182 CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
184 mdc_update_max_ea_from_body(exp, body);
185 if (body->mbo_eadatasize != 0) {
186 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
187 body->mbo_eadatasize);
195 static void mdc_reset_acl_req(struct ptlrpc_request *req)
197 spin_lock(&req->rq_early_free_lock);
198 sptlrpc_cli_free_repbuf(req);
199 req->rq_repbuf = NULL;
200 req->rq_repbuf_len = 0;
201 req->rq_repdata = NULL;
202 req->rq_reqdata_len = 0;
203 spin_unlock(&req->rq_early_free_lock);
206 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
207 struct ptlrpc_request **request)
209 struct ptlrpc_request *req;
210 struct obd_import *imp = class_exp2cliimp(exp);
211 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
215 /* Single MDS without an LMV case */
216 if (op_data->op_flags & MF_GET_MDT_IDX) {
222 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
226 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
228 ptlrpc_request_free(req);
233 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
234 op_data->op_mode, -1, 0);
235 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
236 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
238 ptlrpc_request_set_replen(req);
240 rc = mdc_getattr_common(exp, req);
243 acl_bufsize = min_t(__u32,
244 imp->imp_connect_data.ocd_max_easize,
246 mdc_reset_acl_req(req);
250 ptlrpc_req_finished(req);
258 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
259 struct ptlrpc_request **request)
261 struct ptlrpc_request *req;
262 struct obd_import *imp = class_exp2cliimp(exp);
263 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
268 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
272 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
273 op_data->op_namelen + 1);
275 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
277 ptlrpc_request_free(req);
281 if (op_data->op_name) {
282 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
283 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
284 op_data->op_namelen);
285 memcpy(name, op_data->op_name, op_data->op_namelen);
289 mdc_pack_body(req, &op_data->op_fid1, op_data->op_valid,
290 op_data->op_mode, op_data->op_suppgids[0], 0);
291 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
293 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
294 ptlrpc_request_set_replen(req);
296 rc = mdc_getattr_common(exp, req);
299 acl_bufsize = min_t(__u32,
300 imp->imp_connect_data.ocd_max_easize,
302 mdc_reset_acl_req(req);
306 ptlrpc_req_finished(req);
314 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
315 const struct lu_fid *fid, int opcode, u64 valid,
316 const char *xattr_name, const char *input,
317 int input_size, int output_size, int flags,
318 __u32 suppgid, struct ptlrpc_request **request)
320 struct ptlrpc_request *req;
321 int xattr_namelen = 0;
327 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
332 xattr_namelen = strlen(xattr_name) + 1;
333 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
338 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
341 /* get SELinux policy info if any */
342 rc = sptlrpc_get_sepol(req);
344 ptlrpc_request_free(req);
347 req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
348 strlen(req->rq_sepol) ?
349 strlen(req->rq_sepol) + 1 : 0);
351 /* Flush local XATTR locks to get rid of a possible cancel RPC */
352 if (opcode == MDS_REINT && fid_is_sane(fid) &&
353 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
357 /* Without that packing would fail */
359 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
362 count = mdc_resource_get_unused(exp, fid,
364 MDS_INODELOCK_XATTR);
366 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
368 ptlrpc_request_free(req);
372 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
374 ptlrpc_request_free(req);
379 if (opcode == MDS_REINT) {
380 struct mdt_rec_setxattr *rec;
382 BUILD_BUG_ON(sizeof(struct mdt_rec_setxattr) !=
383 sizeof(struct mdt_rec_reint));
384 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
385 rec->sx_opcode = REINT_SETXATTR;
386 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
387 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
388 rec->sx_cap = cfs_curproc_cap_pack();
389 rec->sx_suppgid1 = suppgid;
390 rec->sx_suppgid2 = -1;
392 rec->sx_valid = valid | OBD_MD_FLCTIME;
393 rec->sx_time = ktime_get_real_seconds();
394 rec->sx_size = output_size;
395 rec->sx_flags = flags;
397 mdc_pack_body(req, fid, valid, output_size, suppgid, flags);
401 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
402 memcpy(tmp, xattr_name, xattr_namelen);
405 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
406 memcpy(tmp, input, input_size);
409 mdc_file_sepol_pack(req);
411 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
412 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
413 RCL_SERVER, output_size);
414 ptlrpc_request_set_replen(req);
417 if (opcode == MDS_REINT)
418 ptlrpc_get_mod_rpc_slot(req);
420 rc = ptlrpc_queue_wait(req);
422 if (opcode == MDS_REINT)
423 ptlrpc_put_mod_rpc_slot(req);
426 ptlrpc_req_finished(req);
432 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
433 u64 obd_md_valid, const char *name,
434 const void *value, size_t value_size,
435 unsigned int xattr_flags, u32 suppgid,
436 struct ptlrpc_request **req)
438 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
439 obd_md_valid == OBD_MD_FLXATTRRM);
441 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
442 fid, MDS_REINT, obd_md_valid, name,
443 value, value_size, 0, xattr_flags, suppgid,
447 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
448 u64 obd_md_valid, const char *name, size_t buf_size,
449 struct ptlrpc_request **req)
451 struct mdt_body *body;
454 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
455 obd_md_valid == OBD_MD_FLXATTRLS);
457 /* Message below is checked in sanity-selinux test_20d
458 * and sanity-sec test_49
460 CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
461 exp->exp_obd->obd_name, name, PFID(fid));
462 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
463 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
468 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
470 GOTO(out, rc = -EPROTO);
472 /* only detect the xattr size */
474 /* LU-11109: Older MDTs do not distinguish
475 * between nonexistent xattrs and zero length
476 * values in this case. Newer MDTs will return
477 * -ENODATA or set OBD_MD_FLXATTR. */
478 GOTO(out, rc = body->mbo_eadatasize);
481 if (body->mbo_eadatasize == 0) {
482 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
483 * success so that we can distinguish between
484 * zero length value and nonexistent xattr.
486 * If OBD_MD_FLXATTR is not set then we keep
487 * the old behavior and return -ENODATA for
488 * getxattr() when mbo_eadatasize is 0. But
489 * -ENODATA only makes sense for getxattr()
490 * and not for listxattr(). */
491 if (body->mbo_valid & OBD_MD_FLXATTR)
493 else if (obd_md_valid == OBD_MD_FLXATTR)
494 GOTO(out, rc = -ENODATA);
499 GOTO(out, rc = body->mbo_eadatasize);
502 ptlrpc_req_finished(*req);
509 static int mdc_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
510 struct obd_export *dt_exp,
511 struct obd_export *md_exp,
512 struct lustre_md *md)
514 struct req_capsule *pill = &req->rq_pill;
519 memset(md, 0, sizeof(*md));
521 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
522 LASSERT(md->body != NULL);
524 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
525 if (!S_ISREG(md->body->mbo_mode)) {
526 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
527 "regular file, but is not\n");
528 GOTO(out, rc = -EPROTO);
531 if (md->body->mbo_eadatasize == 0) {
532 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
533 "but eadatasize 0\n");
534 GOTO(out, rc = -EPROTO);
537 md->layout.lb_len = md->body->mbo_eadatasize;
538 md->layout.lb_buf = req_capsule_server_sized_get(pill,
541 if (md->layout.lb_buf == NULL)
542 GOTO(out, rc = -EPROTO);
543 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
544 const union lmv_mds_md *lmv;
547 if (!S_ISDIR(md->body->mbo_mode)) {
548 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
549 "directory, but is not\n");
550 GOTO(out, rc = -EPROTO);
553 if (md->body->mbo_valid & OBD_MD_MEA) {
554 lmv_size = md->body->mbo_eadatasize;
556 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
557 "but eadatasize 0\n");
561 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
564 GOTO(out, rc = -EPROTO);
566 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
570 if (rc < (int)sizeof(*md->lmv)) {
571 struct lmv_foreign_md *lfm = md->lfm;
573 /* short (< sizeof(struct lmv_stripe_md))
576 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
578 "lmv size too small: %d < %d\n",
579 rc, (int)sizeof(*md->lmv));
580 GOTO(out, rc = -EPROTO);
585 /* since 2.12.58 intent_getattr fetches default LMV */
586 if (md->body->mbo_valid & OBD_MD_DEFAULT_MEA) {
587 lmv_size = sizeof(struct lmv_user_md);
588 lmv = req_capsule_server_sized_get(pill,
592 GOTO(out, rc = -EPROTO);
594 rc = md_unpackmd(md_exp, &md->default_lmv, lmv,
599 if (rc < (int)sizeof(*md->default_lmv)) {
601 "default lmv size too small: %d < %d\n",
602 rc, (int)sizeof(*md->default_lmv));
603 GOTO(out, rc = -EPROTO);
609 if (md->body->mbo_valid & OBD_MD_FLACL) {
610 /* for ACL, it's possible that FLACL is set but aclsize is zero.
611 * only when aclsize != 0 there's an actual segment for ACL
614 rc = mdc_unpack_acl(req, md);
627 static int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
633 void mdc_replay_open(struct ptlrpc_request *req)
635 struct md_open_data *mod = req->rq_cb_data;
636 struct ptlrpc_request *close_req;
637 struct obd_client_handle *och;
638 struct lustre_handle old_open_handle = { };
639 struct mdt_body *body;
640 struct ldlm_reply *rep;
644 DEBUG_REQ(D_ERROR, req,
645 "cannot properly replay without open data");
650 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
651 LASSERT(body != NULL);
653 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
654 if (rep != NULL && rep->lock_policy_res2 != 0)
655 DEBUG_REQ(D_ERROR, req, "Open request replay failed with %ld ",
656 (long int)rep->lock_policy_res2);
658 spin_lock(&req->rq_lock);
660 if (och && och->och_open_handle.cookie)
661 req->rq_early_free_repbuf = 1;
663 req->rq_early_free_repbuf = 0;
664 spin_unlock(&req->rq_lock);
666 if (req->rq_early_free_repbuf) {
667 struct lustre_handle *file_open_handle;
669 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
671 file_open_handle = &och->och_open_handle;
672 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
673 file_open_handle->cookie, body->mbo_open_handle.cookie);
674 old_open_handle = *file_open_handle;
675 *file_open_handle = body->mbo_open_handle;
678 close_req = mod->mod_close_req;
680 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
681 struct mdt_ioepoch *epoch;
683 LASSERT(opc == MDS_CLOSE);
684 epoch = req_capsule_client_get(&close_req->rq_pill,
688 if (req->rq_early_free_repbuf)
689 LASSERT(old_open_handle.cookie ==
690 epoch->mio_open_handle.cookie);
692 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
693 epoch->mio_open_handle = body->mbo_open_handle;
698 void mdc_commit_open(struct ptlrpc_request *req)
700 struct md_open_data *mod = req->rq_cb_data;
705 * No need to touch md_open_data::mod_och, it holds a reference on
706 * \var mod and will zero references to each other, \var mod will be
707 * freed after that when md_open_data::mod_och will put the reference.
711 * Do not let open request to disappear as it still may be needed
712 * for close rpc to happen (it may happen on evict only, otherwise
713 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
714 * called), just mark this rpc as committed to distinguish these 2
715 * cases, see mdc_close() for details. The open request reference will
716 * be put along with freeing \var mod.
718 ptlrpc_request_addref(req);
719 spin_lock(&req->rq_lock);
720 req->rq_committed = 1;
721 spin_unlock(&req->rq_lock);
722 req->rq_cb_data = NULL;
726 int mdc_set_open_replay_data(struct obd_export *exp,
727 struct obd_client_handle *och,
728 struct lookup_intent *it)
730 struct md_open_data *mod;
731 struct mdt_rec_create *rec;
732 struct mdt_body *body;
733 struct ptlrpc_request *open_req = it->it_request;
734 struct obd_import *imp = open_req->rq_import;
737 if (!open_req->rq_replay)
740 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
741 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
742 LASSERT(rec != NULL);
743 /* Incoming message in my byte order (it's been swabbed). */
744 /* Outgoing messages always in my byte order. */
745 LASSERT(body != NULL);
747 /* Only if the import is replayable, we set replay_open data */
748 if (och && imp->imp_replayable) {
749 mod = obd_mod_alloc();
751 DEBUG_REQ(D_ERROR, open_req,
752 "cannot allocate md_open_data");
757 * Take a reference on \var mod, to be freed on mdc_close().
758 * It protects \var mod from being freed on eviction (commit
759 * callback is called despite rq_replay flag).
760 * Another reference for \var och.
765 spin_lock(&open_req->rq_lock);
768 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
769 it_disposition(it, DISP_OPEN_STRIPE);
770 mod->mod_open_req = open_req;
771 open_req->rq_cb_data = mod;
772 open_req->rq_commit_cb = mdc_commit_open;
773 open_req->rq_early_free_repbuf = 1;
774 spin_unlock(&open_req->rq_lock);
777 rec->cr_fid2 = body->mbo_fid1;
778 rec->cr_open_handle_old = body->mbo_open_handle;
779 open_req->rq_replay_cb = mdc_replay_open;
780 if (!fid_is_sane(&body->mbo_fid1)) {
781 DEBUG_REQ(D_ERROR, open_req,
782 "saving replay request with insane FID " DFID,
783 PFID(&body->mbo_fid1));
787 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
791 static void mdc_free_open(struct md_open_data *mod)
795 if (mod->mod_is_create == 0 &&
796 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
800 * No reason to asssert here if the open request has
801 * rq_replay == 1. It means that mdc_close failed, and
802 * close request wasn`t sent. It is not fatal to client.
803 * The worst thing is eviction if the client gets open lock
806 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req,
807 "free open request, rq_replay=%d",
808 mod->mod_open_req->rq_replay);
810 ptlrpc_request_committed(mod->mod_open_req, committed);
811 if (mod->mod_close_req)
812 ptlrpc_request_committed(mod->mod_close_req, committed);
815 static int mdc_clear_open_replay_data(struct obd_export *exp,
816 struct obd_client_handle *och)
818 struct md_open_data *mod = och->och_mod;
822 * It is possible to not have \var mod in a case of eviction between
823 * lookup and ll_file_open().
828 LASSERT(mod != LP_POISON);
829 LASSERT(mod->mod_open_req != NULL);
831 spin_lock(&mod->mod_open_req->rq_lock);
833 mod->mod_och->och_open_handle.cookie = 0;
834 mod->mod_open_req->rq_early_free_repbuf = 0;
835 spin_unlock(&mod->mod_open_req->rq_lock);
845 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
846 struct md_open_data *mod, struct ptlrpc_request **request)
848 struct obd_device *obd = class_exp2obd(exp);
849 struct ptlrpc_request *req;
850 struct req_format *req_fmt;
851 size_t u32_count = 0;
856 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
857 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
860 if (op_data->op_bias & MDS_CLOSE_INTENT) {
861 req_fmt = &RQF_MDS_CLOSE_INTENT;
862 if (op_data->op_bias & MDS_HSM_RELEASE) {
863 /* allocate a FID for volatile file */
864 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
867 CERROR("%s: "DFID" allocating FID: rc = %d\n",
868 obd->obd_name, PFID(&op_data->op_fid1),
870 /* save the errcode and proceed to close */
874 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
875 size_t count = op_data->op_data_size / sizeof(__u32);
877 if (count > INLINE_RESYNC_ARRAY_SIZE)
881 req_fmt = &RQF_MDS_CLOSE;
885 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
888 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
890 /* Ensure that this close's handle is fixed up during replay. */
891 if (likely(mod != NULL)) {
892 LASSERTF(mod->mod_open_req != NULL &&
893 mod->mod_open_req->rq_type != LI_POISON,
894 "POISONED open %p!\n", mod->mod_open_req);
896 mod->mod_close_req = req;
898 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "matched open");
899 /* We no longer want to preserve this open for replay even
900 * though the open was committed. b=3632, b=3633 */
901 spin_lock(&mod->mod_open_req->rq_lock);
902 mod->mod_open_req->rq_replay = 0;
903 spin_unlock(&mod->mod_open_req->rq_lock);
905 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
909 * TODO: repeat close after errors
911 CWARN("%s: close of FID "DFID" failed, file reference will be "
912 "dropped when this client unmounts or is evicted\n",
913 obd->obd_name, PFID(&op_data->op_fid1));
914 GOTO(out, rc = -ENOMEM);
918 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
919 u32_count * sizeof(__u32));
921 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
923 ptlrpc_request_free(req);
928 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
929 * portal whose threads are not taking any DLM locks and are therefore
930 * always progressing */
931 req->rq_request_portal = MDS_READPAGE_PORTAL;
932 ptlrpc_at_set_req_timeout(req);
934 if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
935 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
936 OP_XVALID_LAZYBLOCKS);
938 mdc_close_pack(req, op_data);
940 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
941 obd->u.cli.cl_default_mds_easize);
943 ptlrpc_request_set_replen(req);
945 ptlrpc_get_mod_rpc_slot(req);
946 rc = ptlrpc_queue_wait(req);
947 ptlrpc_put_mod_rpc_slot(req);
949 if (req->rq_repmsg == NULL) {
950 CDEBUG(D_RPCTRACE, "request %p failed to send: rc = %d\n", req,
953 rc = req->rq_status ?: -EIO;
954 } else if (rc == 0 || rc == -EAGAIN) {
955 struct mdt_body *body;
957 rc = lustre_msg_get_status(req->rq_repmsg);
958 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
959 DEBUG_REQ(D_ERROR, req,
960 "type = PTL_RPC_MSG_ERR: rc = %d", rc);
964 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
967 } else if (rc == -ESTALE) {
969 * it can be allowed error after 3633 if open was committed and
970 * server failed before close was sent. Let's check if mod
971 * exists and return no error in that case
974 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
975 LASSERT(mod->mod_open_req != NULL);
976 if (mod->mod_open_req->rq_committed)
984 mod->mod_close_req = NULL;
985 /* Since now, mod is accessed through open_req only,
986 * thus close req does not keep a reference on mod anymore. */
991 RETURN(rc < 0 ? rc : saved_rc);
994 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
995 u64 offset, struct page **pages, int npages,
996 struct ptlrpc_request **request)
998 struct ptlrpc_request *req;
999 struct ptlrpc_bulk_desc *desc;
1008 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1012 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1014 ptlrpc_request_free(req);
1018 req->rq_request_portal = MDS_READPAGE_PORTAL;
1019 ptlrpc_at_set_req_timeout(req);
1021 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1022 PTLRPC_BULK_PUT_SINK,
1024 &ptlrpc_bulk_kiov_pin_ops);
1026 ptlrpc_req_finished(req);
1030 /* NB req now owns desc and will free it when it gets freed */
1031 for (i = 0; i < npages; i++)
1032 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1035 mdc_readdir_pack(req, offset, PAGE_SIZE * npages, fid);
1037 ptlrpc_request_set_replen(req);
1038 rc = ptlrpc_queue_wait(req);
1040 ptlrpc_req_finished(req);
1041 if (rc != -ETIMEDOUT)
1045 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1046 CERROR("%s: too many resend retries: rc = %d\n",
1047 exp->exp_obd->obd_name, -EIO);
1051 /* If a signal interrupts then the timeout returned will
1052 * not be zero. In that case return -EINTR
1054 if (msleep_interruptible(resends * 1000))
1060 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1061 req->rq_bulk->bd_nob_transferred);
1063 ptlrpc_req_finished(req);
1067 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1068 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1069 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1070 PAGE_SIZE * npages);
1071 ptlrpc_req_finished(req);
1079 static void mdc_release_page(struct page *page, int remove)
1083 if (likely(page->mapping != NULL))
1084 delete_from_page_cache(page);
1090 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1091 __u64 *start, __u64 *end, int hash64)
1094 * Complement of hash is used as an index so that
1095 * radix_tree_gang_lookup() can be used to find a page with starting
1096 * hash _smaller_ than one we are looking for.
1098 unsigned long offset = hash_x_index(*hash, hash64);
1100 unsigned long flags;
1103 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1104 found = radix_tree_gang_lookup(&mapping->page_tree,
1105 (void **)&page, offset, 1);
1106 if (found > 0 && !ll_xa_is_value(page)) {
1107 struct lu_dirpage *dp;
1110 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1112 * In contrast to find_lock_page() we are sure that directory
1113 * page cannot be truncated (while DLM lock is held) and,
1114 * hence, can avoid restart.
1116 * In fact, page cannot be locked here at all, because
1117 * mdc_read_page_remote does synchronous io.
1119 wait_on_page_locked(page);
1120 if (PageUptodate(page)) {
1122 if (BITS_PER_LONG == 32 && hash64) {
1123 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1124 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1125 *hash = *hash >> 32;
1127 *start = le64_to_cpu(dp->ldp_hash_start);
1128 *end = le64_to_cpu(dp->ldp_hash_end);
1130 if (unlikely(*start == 1 && *hash == 0))
1133 LASSERTF(*start <= *hash, "start = %#llx"
1134 ",end = %#llx,hash = %#llx\n",
1135 *start, *end, *hash);
1136 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1137 " hash %#llx\n", offset, *start, *end, *hash);
1140 mdc_release_page(page, 0);
1142 } else if (*end != *start && *hash == *end) {
1144 * upon hash collision, remove this page,
1145 * otherwise put page reference, and
1146 * mdc_read_page_remote() will issue RPC to
1147 * fetch the page we want.
1150 mdc_release_page(page,
1151 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1156 page = ERR_PTR(-EIO);
1159 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1166 * Adjust a set of pages, each page containing an array of lu_dirpages,
1167 * so that each page can be used as a single logical lu_dirpage.
1169 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1170 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1171 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1172 * value is used as a cookie to request the next lu_dirpage in a
1173 * directory listing that spans multiple pages (two in this example):
1176 * .|--------v------- -----.
1177 * |s|e|f|p|ent|ent| ... |ent|
1178 * '--|-------------- -----' Each PAGE contains a single
1179 * '------. lu_dirpage.
1180 * .---------v------- -----.
1181 * |s|e|f|p|ent| 0 | ... | 0 |
1182 * '----------------- -----'
1184 * However, on hosts where the native VM page size (PAGE_SIZE) is
1185 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1186 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1187 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1188 * after it in the same PAGE (arrows simplified for brevity, but
1189 * in general e0==s1, e1==s2, etc.):
1191 * .-------------------- -----.
1192 * |s0|e0|f0|p|ent|ent| ... |ent|
1193 * |---v---------------- -----|
1194 * |s1|e1|f1|p|ent|ent| ... |ent|
1195 * |---v---------------- -----| Here, each PAGE contains
1196 * ... multiple lu_dirpages.
1197 * |---v---------------- -----|
1198 * |s'|e'|f'|p|ent|ent| ... |ent|
1199 * '---|---------------- -----'
1201 * .----------------------------.
1204 * This structure is transformed into a single logical lu_dirpage as follows:
1206 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1207 * labeled 'next PAGE'.
1209 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1210 * a hash collision with the next page exists.
1212 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1213 * to the first entry of the next lu_dirpage.
1215 #if PAGE_SIZE > LU_PAGE_SIZE
1216 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1220 for (i = 0; i < cfs_pgs; i++) {
1221 struct lu_dirpage *dp = kmap(pages[i]);
1222 struct lu_dirpage *first = dp;
1223 struct lu_dirent *end_dirent = NULL;
1224 struct lu_dirent *ent;
1225 __u64 hash_end = dp->ldp_hash_end;
1226 __u32 flags = dp->ldp_flags;
1228 while (--lu_pgs > 0) {
1229 ent = lu_dirent_start(dp);
1230 for (end_dirent = ent; ent != NULL;
1231 end_dirent = ent, ent = lu_dirent_next(ent));
1233 /* Advance dp to next lu_dirpage. */
1234 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1236 /* Check if we've reached the end of the PAGE. */
1237 if (!((unsigned long)dp & ~PAGE_MASK))
1240 /* Save the hash and flags of this lu_dirpage. */
1241 hash_end = dp->ldp_hash_end;
1242 flags = dp->ldp_flags;
1244 /* Check if lu_dirpage contains no entries. */
1245 if (end_dirent == NULL)
1248 /* Enlarge the end entry lde_reclen from 0 to
1249 * first entry of next lu_dirpage. */
1250 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1251 end_dirent->lde_reclen =
1252 cpu_to_le16((char *)(dp->ldp_entries) -
1253 (char *)end_dirent);
1256 first->ldp_hash_end = hash_end;
1257 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1258 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1262 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1265 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1266 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1268 /* parameters for readdir page */
1269 struct readpage_param {
1270 struct md_op_data *rp_mod;
1273 struct obd_export *rp_exp;
1274 struct md_callback *rp_cb;
1278 * Read pages from server.
1280 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1281 * a header lu_dirpage which describes the start/end hash, and whether this
1282 * page is empty (contains no dir entry) or hash collide with next page.
1283 * After client receives reply, several pages will be integrated into dir page
1284 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1285 * lu_dirpage for this integrated page will be adjusted.
1287 static int mdc_read_page_remote(void *data, struct page *page0)
1289 struct readpage_param *rp = data;
1290 struct page **page_pool;
1292 struct lu_dirpage *dp;
1293 struct md_op_data *op_data = rp->rp_mod;
1294 struct ptlrpc_request *req;
1296 struct inode *inode;
1298 int rd_pgs = 0; /* number of pages actually read */
1304 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1305 inode = op_data->op_data;
1306 fid = &op_data->op_fid1;
1307 LASSERT(inode != NULL);
1309 OBD_ALLOC_PTR_ARRAY_LARGE(page_pool, max_pages);
1310 if (page_pool != NULL) {
1311 page_pool[0] = page0;
1317 for (npages = 1; npages < max_pages; npages++) {
1318 page = page_cache_alloc(inode->i_mapping);
1321 page_pool[npages] = page;
1324 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1326 /* page0 is special, which was added into page cache early */
1327 delete_from_page_cache(page0);
1331 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1333 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1334 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1336 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1338 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1340 SetPageUptodate(page0);
1344 ptlrpc_req_finished(req);
1345 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1346 for (i = 1; i < npages; i++) {
1347 unsigned long offset;
1351 page = page_pool[i];
1353 if (rc < 0 || i >= rd_pgs) {
1358 SetPageUptodate(page);
1361 hash = le64_to_cpu(dp->ldp_hash_start);
1364 offset = hash_x_index(hash, rp->rp_hash64);
1366 prefetchw(&page->flags);
1367 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1372 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1373 " rc = %d\n", offset, ret);
1377 if (page_pool != &page0)
1378 OBD_FREE_PTR_ARRAY_LARGE(page_pool, max_pages);
1384 * Read dir page from cache first, if it can not find it, read it from
1385 * server and add into the cache.
1387 * \param[in] exp MDC export
1388 * \param[in] op_data client MD stack parameters, transfering parameters
1389 * between different layers on client MD stack.
1390 * \param[in] cb_op callback required for ldlm lock enqueue during
1392 * \param[in] hash_offset the hash offset of the page to be read
1393 * \param[in] ppage the page to be read
1395 * retval = 0 get the page successfully
1396 * errno(<0) get the page failed
1398 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1399 struct md_callback *cb_op, __u64 hash_offset,
1400 struct page **ppage)
1402 struct lookup_intent it = { .it_op = IT_READDIR };
1404 struct inode *dir = op_data->op_data;
1405 struct address_space *mapping;
1406 struct lu_dirpage *dp;
1409 struct lustre_handle lockh;
1410 struct ptlrpc_request *enq_req = NULL;
1411 struct readpage_param rp_param;
1418 LASSERT(dir != NULL);
1419 mapping = dir->i_mapping;
1421 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1422 cb_op->md_blocking_ast, 0);
1423 if (enq_req != NULL)
1424 ptlrpc_req_finished(enq_req);
1427 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1428 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1433 lockh.cookie = it.it_lock_handle;
1434 mdc_set_lock_data(exp, &lockh, dir, NULL);
1436 rp_param.rp_off = hash_offset;
1437 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1438 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1439 rp_param.rp_hash64);
1441 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1442 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1443 rp_param.rp_off, PTR_ERR(page));
1444 GOTO(out_unlock, rc = PTR_ERR(page));
1445 } else if (page != NULL) {
1447 * XXX nikita: not entirely correct handling of a corner case:
1448 * suppose hash chain of entries with hash value HASH crosses
1449 * border between pages P0 and P1. First both P0 and P1 are
1450 * cached, seekdir() is called for some entry from the P0 part
1451 * of the chain. Later P0 goes out of cache. telldir(HASH)
1452 * happens and finds P1, as it starts with matching hash
1453 * value. Remaining entries from P0 part of the chain are
1454 * skipped. (Is that really a bug?)
1456 * Possible solutions: 0. don't cache P1 is such case, handle
1457 * it as an "overflow" page. 1. invalidate all pages at
1458 * once. 2. use HASH|1 as an index for P1.
1460 GOTO(hash_collision, page);
1463 rp_param.rp_exp = exp;
1464 rp_param.rp_mod = op_data;
1465 page = read_cache_page(mapping,
1466 hash_x_index(rp_param.rp_off,
1467 rp_param.rp_hash64),
1468 mdc_read_page_remote, &rp_param);
1470 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1471 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1472 rp_param.rp_off, PTR_ERR(page));
1473 GOTO(out_unlock, rc = PTR_ERR(page));
1476 wait_on_page_locked(page);
1478 if (!PageUptodate(page)) {
1479 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1480 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1481 rp_param.rp_off, -5);
1484 if (!PageChecked(page))
1485 SetPageChecked(page);
1486 if (PageError(page)) {
1487 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1488 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1489 rp_param.rp_off, -5);
1494 dp = page_address(page);
1495 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1496 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1497 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1498 rp_param.rp_off = hash_offset >> 32;
1500 start = le64_to_cpu(dp->ldp_hash_start);
1501 end = le64_to_cpu(dp->ldp_hash_end);
1502 rp_param.rp_off = hash_offset;
1505 LASSERT(start == rp_param.rp_off);
1506 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1507 #if BITS_PER_LONG == 32
1508 CWARN("Real page-wide hash collision at [%llu %llu] with "
1509 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1510 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1514 * Fetch whole overflow chain...
1522 ldlm_lock_decref(&lockh, it.it_lock_mode);
1526 mdc_release_page(page, 1);
1531 static int mdc_statfs_interpret(const struct lu_env *env,
1532 struct ptlrpc_request *req, void *args, int rc)
1534 struct obd_info *oinfo = args;
1535 struct obd_statfs *osfs;
1538 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1542 oinfo->oi_osfs = osfs;
1544 CDEBUG(D_CACHE, "blocks=%llu free=%llu avail=%llu "
1545 "objects=%llu free=%llu state=%x\n",
1546 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
1547 osfs->os_files, osfs->os_ffree, osfs->os_state);
1550 oinfo->oi_cb_up(oinfo, rc);
1555 static int mdc_statfs_async(struct obd_export *exp,
1556 struct obd_info *oinfo, time64_t max_age,
1557 struct ptlrpc_request_set *unused)
1559 struct ptlrpc_request *req;
1560 struct obd_info *aa;
1562 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_STATFS,
1563 LUSTRE_MDS_VERSION, MDS_STATFS);
1567 ptlrpc_request_set_replen(req);
1568 req->rq_interpret_reply = mdc_statfs_interpret;
1570 aa = ptlrpc_req_async_args(aa, req);
1573 ptlrpcd_add_req(req);
1578 static int mdc_statfs(const struct lu_env *env,
1579 struct obd_export *exp, struct obd_statfs *osfs,
1580 time64_t max_age, __u32 flags)
1582 struct obd_device *obd = class_exp2obd(exp);
1583 struct req_format *fmt;
1584 struct ptlrpc_request *req;
1585 struct obd_statfs *msfs;
1586 struct obd_import *imp, *imp0;
1591 * Since the request might also come from lprocfs, so we need
1592 * sync this with client_disconnect_export Bug15684
1594 with_imp_locked(obd, imp0, rc)
1595 imp = class_import_get(imp0);
1599 fmt = &RQF_MDS_STATFS;
1600 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1601 (flags & OBD_STATFS_SUM))
1602 fmt = &RQF_MDS_STATFS_NEW;
1603 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1606 GOTO(output, rc = -ENOMEM);
1607 req->rq_allow_intr = 1;
1609 if ((flags & OBD_STATFS_SUM) &&
1610 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1611 /* request aggregated states */
1612 struct mdt_body *body;
1614 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1616 GOTO(out, rc = -EPROTO);
1617 body->mbo_valid = OBD_MD_FLAGSTATFS;
1620 ptlrpc_request_set_replen(req);
1622 if (flags & OBD_STATFS_NODELAY) {
1623 /* procfs requests not want stay in wait for avoid deadlock */
1624 req->rq_no_resend = 1;
1625 req->rq_no_delay = 1;
1628 rc = ptlrpc_queue_wait(req);
1630 /* check connection error first */
1631 if (imp->imp_connect_error)
1632 rc = imp->imp_connect_error;
1636 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1638 GOTO(out, rc = -EPROTO);
1643 ptlrpc_req_finished(req);
1645 class_import_put(imp);
1649 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1651 __u32 keylen, vallen;
1655 if (gf->gf_pathlen > PATH_MAX)
1656 RETURN(-ENAMETOOLONG);
1657 if (gf->gf_pathlen < 2)
1660 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1661 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1662 sizeof(struct lu_fid));
1663 OBD_ALLOC(key, keylen);
1666 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1667 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1668 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1669 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1670 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1671 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1673 if (!fid_is_sane(&gf->gf_fid))
1674 GOTO(out, rc = -EINVAL);
1676 /* Val is struct getinfo_fid2path result plus path */
1677 vallen = sizeof(*gf) + gf->gf_pathlen;
1679 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1680 if (rc != 0 && rc != -EREMOTE)
1683 if (vallen <= sizeof(*gf))
1684 GOTO(out, rc = -EPROTO);
1685 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1686 GOTO(out, rc = -EOVERFLOW);
1688 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1689 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1690 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1691 /* only log the last 512 characters of the path */
1692 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1695 OBD_FREE(key, keylen);
1699 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1700 struct hsm_progress_kernel *hpk)
1702 struct obd_import *imp = class_exp2cliimp(exp);
1703 struct hsm_progress_kernel *req_hpk;
1704 struct ptlrpc_request *req;
1708 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1709 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1711 GOTO(out, rc = -ENOMEM);
1713 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1715 /* Copy hsm_progress struct */
1716 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1717 if (req_hpk == NULL)
1718 GOTO(out, rc = -EPROTO);
1721 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1723 ptlrpc_request_set_replen(req);
1725 ptlrpc_get_mod_rpc_slot(req);
1726 rc = ptlrpc_queue_wait(req);
1727 ptlrpc_put_mod_rpc_slot(req);
1731 ptlrpc_req_finished(req);
1735 * Send hsm_ct_register to MDS
1737 * \param[in] imp import
1738 * \param[in] archive_count if in bitmap format, it is the bitmap,
1739 * else it is the count of archive_ids
1740 * \param[in] archives if in bitmap format, it is NULL,
1741 * else it is archive_id lists
1743 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1746 struct ptlrpc_request *req;
1747 __u32 *archive_array;
1748 size_t archives_size;
1752 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1756 if (archives != NULL)
1757 archives_size = sizeof(*archive_array) * archive_count;
1759 archives_size = sizeof(archive_count);
1761 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1762 RCL_CLIENT, archives_size);
1764 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1766 ptlrpc_request_free(req);
1770 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1772 archive_array = req_capsule_client_get(&req->rq_pill,
1773 &RMF_MDS_HSM_ARCHIVE);
1774 if (archive_array == NULL)
1775 GOTO(out, rc = -EPROTO);
1777 if (archives != NULL)
1778 memcpy(archive_array, archives, archives_size);
1780 *archive_array = archive_count;
1782 ptlrpc_request_set_replen(req);
1783 req->rq_no_resend = 1;
1785 rc = mdc_queue_wait(req);
1788 ptlrpc_req_finished(req);
1792 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1793 struct md_op_data *op_data)
1795 struct hsm_current_action *hca = op_data->op_data;
1796 struct hsm_current_action *req_hca;
1797 struct ptlrpc_request *req;
1801 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1802 &RQF_MDS_HSM_ACTION);
1806 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1808 ptlrpc_request_free(req);
1812 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1813 op_data->op_suppgids[0], 0);
1815 ptlrpc_request_set_replen(req);
1817 rc = mdc_queue_wait(req);
1821 req_hca = req_capsule_server_get(&req->rq_pill,
1822 &RMF_MDS_HSM_CURRENT_ACTION);
1823 if (req_hca == NULL)
1824 GOTO(out, rc = -EPROTO);
1830 ptlrpc_req_finished(req);
1834 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1836 struct ptlrpc_request *req;
1840 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1842 MDS_HSM_CT_UNREGISTER);
1844 GOTO(out, rc = -ENOMEM);
1846 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1848 ptlrpc_request_set_replen(req);
1850 rc = mdc_queue_wait(req);
1853 ptlrpc_req_finished(req);
1857 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1858 struct md_op_data *op_data)
1860 struct hsm_user_state *hus = op_data->op_data;
1861 struct hsm_user_state *req_hus;
1862 struct ptlrpc_request *req;
1866 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1867 &RQF_MDS_HSM_STATE_GET);
1871 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1873 ptlrpc_request_free(req);
1877 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1878 op_data->op_suppgids[0], 0);
1880 ptlrpc_request_set_replen(req);
1882 rc = mdc_queue_wait(req);
1886 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1887 if (req_hus == NULL)
1888 GOTO(out, rc = -EPROTO);
1894 ptlrpc_req_finished(req);
1898 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1899 struct md_op_data *op_data)
1901 struct hsm_state_set *hss = op_data->op_data;
1902 struct hsm_state_set *req_hss;
1903 struct ptlrpc_request *req;
1907 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1908 &RQF_MDS_HSM_STATE_SET);
1912 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1914 ptlrpc_request_free(req);
1918 mdc_pack_body(req, &op_data->op_fid1, 0, 0,
1919 op_data->op_suppgids[0], 0);
1922 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1923 if (req_hss == NULL)
1924 GOTO(out, rc = -EPROTO);
1927 ptlrpc_request_set_replen(req);
1929 ptlrpc_get_mod_rpc_slot(req);
1930 rc = ptlrpc_queue_wait(req);
1931 ptlrpc_put_mod_rpc_slot(req);
1935 ptlrpc_req_finished(req);
1939 static int mdc_ioc_hsm_request(struct obd_export *exp,
1940 struct hsm_user_request *hur)
1942 struct obd_import *imp = class_exp2cliimp(exp);
1943 struct ptlrpc_request *req;
1944 struct hsm_request *req_hr;
1945 struct hsm_user_item *req_hui;
1950 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1952 GOTO(out, rc = -ENOMEM);
1954 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1955 hur->hur_request.hr_itemcount
1956 * sizeof(struct hsm_user_item));
1957 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1958 hur->hur_request.hr_data_len);
1960 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1962 ptlrpc_request_free(req);
1966 mdc_pack_body(req, NULL, 0, 0, -1, 0);
1968 /* Copy hsm_request struct */
1969 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1971 GOTO(out, rc = -EPROTO);
1972 *req_hr = hur->hur_request;
1974 /* Copy hsm_user_item structs */
1975 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1976 if (req_hui == NULL)
1977 GOTO(out, rc = -EPROTO);
1978 memcpy(req_hui, hur->hur_user_item,
1979 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1981 /* Copy opaque field */
1982 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1983 if (req_opaque == NULL)
1984 GOTO(out, rc = -EPROTO);
1985 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1987 ptlrpc_request_set_replen(req);
1989 ptlrpc_get_mod_rpc_slot(req);
1990 rc = ptlrpc_queue_wait(req);
1991 ptlrpc_put_mod_rpc_slot(req);
1996 ptlrpc_req_finished(req);
2000 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2001 struct lustre_kernelcomm *lk);
2003 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2004 struct obd_quotactl *oqctl)
2006 struct ptlrpc_request *req;
2007 struct obd_quotactl *oqc;
2011 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_QUOTACTL);
2016 if (LUSTRE_Q_CMD_IS_POOL(oqctl->qc_cmd))
2017 req_capsule_set_size(&req->rq_pill,
2020 sizeof(*oqc) + LOV_MAXPOOLNAME + 1);
2022 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION,
2025 ptlrpc_request_free(req);
2029 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2030 QCTL_COPY(oqc, oqctl);
2032 ptlrpc_request_set_replen(req);
2033 ptlrpc_at_set_req_timeout(req);
2035 rc = ptlrpc_queue_wait(req);
2037 CERROR("%s: ptlrpc_queue_wait failed: rc = %d\n",
2038 exp->exp_obd->obd_name, rc);
2042 if (req->rq_repmsg &&
2043 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2044 QCTL_COPY(oqctl, oqc);
2047 CERROR("%s: cannot unpack obd_quotactl: rc = %d\n",
2048 exp->exp_obd->obd_name, rc);
2051 ptlrpc_req_finished(req);
2056 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2057 struct md_op_data *op_data)
2060 struct ptlrpc_request *req;
2062 struct mdc_swap_layouts *msl, *payload;
2065 msl = op_data->op_data;
2067 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2068 * first thing it will do is to cancel the 2 layout
2069 * locks held by this client.
2070 * So the client must cancel its layout locks on the 2 fids
2071 * with the request RPC to avoid extra RPC round trips.
2073 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2074 LCK_EX, MDS_INODELOCK_LAYOUT |
2075 MDS_INODELOCK_XATTR);
2076 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2077 LCK_EX, MDS_INODELOCK_LAYOUT |
2078 MDS_INODELOCK_XATTR);
2080 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2081 &RQF_MDS_SWAP_LAYOUTS);
2083 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2087 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2089 ptlrpc_request_free(req);
2093 mdc_swap_layouts_pack(req, op_data);
2095 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2100 ptlrpc_request_set_replen(req);
2102 rc = ptlrpc_queue_wait(req);
2108 ptlrpc_req_finished(req);
2112 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2113 void *karg, void __user *uarg)
2115 struct obd_device *obd = exp->exp_obd;
2116 struct obd_ioctl_data *data = karg;
2117 struct obd_import *imp = obd->u.cli.cl_import;
2121 if (!try_module_get(THIS_MODULE)) {
2122 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2123 module_name(THIS_MODULE));
2127 case OBD_IOC_FID2PATH:
2128 rc = mdc_ioc_fid2path(exp, karg);
2130 case LL_IOC_HSM_CT_START:
2131 rc = mdc_ioc_hsm_ct_start(exp, karg);
2132 /* ignore if it was already registered on this MDS. */
2136 case LL_IOC_HSM_PROGRESS:
2137 rc = mdc_ioc_hsm_progress(exp, karg);
2139 case LL_IOC_HSM_STATE_GET:
2140 rc = mdc_ioc_hsm_state_get(exp, karg);
2142 case LL_IOC_HSM_STATE_SET:
2143 rc = mdc_ioc_hsm_state_set(exp, karg);
2145 case LL_IOC_HSM_ACTION:
2146 rc = mdc_ioc_hsm_current_action(exp, karg);
2148 case LL_IOC_HSM_REQUEST:
2149 rc = mdc_ioc_hsm_request(exp, karg);
2151 case OBD_IOC_CLIENT_RECOVER:
2152 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2156 case IOC_OSC_SET_ACTIVE:
2157 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2160 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2161 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2162 * there'd be no LMV layer thus we might be called here. Eventually
2163 * this code should be removed.
2166 case IOC_OBD_STATFS: {
2167 struct obd_statfs stat_buf = {0};
2169 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2170 GOTO(out, rc = -ENODEV);
2173 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2174 min((int)data->ioc_plen2,
2175 (int)sizeof(struct obd_uuid))))
2176 GOTO(out, rc = -EFAULT);
2178 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2179 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2184 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2185 min((int) data->ioc_plen1,
2186 (int) sizeof(stat_buf))))
2187 GOTO(out, rc = -EFAULT);
2191 case OBD_IOC_QUOTACTL: {
2192 struct if_quotactl *qctl = karg;
2193 struct obd_quotactl *oqctl;
2195 OBD_ALLOC_PTR(oqctl);
2197 GOTO(out, rc = -ENOMEM);
2199 QCTL_COPY(oqctl, qctl);
2200 rc = obd_quotactl(exp, oqctl);
2202 QCTL_COPY(qctl, oqctl);
2203 qctl->qc_valid = QC_MDTIDX;
2204 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2207 OBD_FREE_PTR(oqctl);
2210 case LL_IOC_GET_CONNECT_FLAGS:
2211 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2212 sizeof(*exp_connect_flags_ptr(exp))))
2213 GOTO(out, rc = -EFAULT);
2216 case LL_IOC_LOV_SWAP_LAYOUTS:
2217 rc = mdc_ioc_swap_layouts(exp, karg);
2220 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2221 GOTO(out, rc = -ENOTTY);
2224 module_put(THIS_MODULE);
2229 static int mdc_get_info_rpc(struct obd_export *exp,
2230 u32 keylen, void *key,
2231 u32 vallen, void *val)
2233 struct obd_import *imp = class_exp2cliimp(exp);
2234 struct ptlrpc_request *req;
2239 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2243 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2244 RCL_CLIENT, keylen);
2245 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2246 RCL_CLIENT, sizeof(vallen));
2248 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2250 ptlrpc_request_free(req);
2254 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2255 memcpy(tmp, key, keylen);
2256 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2257 memcpy(tmp, &vallen, sizeof(vallen));
2259 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2260 RCL_SERVER, vallen);
2261 ptlrpc_request_set_replen(req);
2263 /* if server failed to resolve FID, and OI scrub not able to fix it, it
2264 * will return -EINPROGRESS, ptlrpc_queue_wait() will keep retrying,
2265 * set request interruptible to avoid deadlock.
2267 if (KEY_IS(KEY_FID2PATH))
2268 req->rq_allow_intr = 1;
2270 rc = ptlrpc_queue_wait(req);
2271 /* -EREMOTE means the get_info result is partial, and it needs to
2272 * continue on another MDT, see fid2path part in lmv_iocontrol */
2273 if (rc == 0 || rc == -EREMOTE) {
2274 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2275 memcpy(val, tmp, vallen);
2276 if (ptlrpc_rep_need_swab(req)) {
2277 if (KEY_IS(KEY_FID2PATH))
2278 lustre_swab_fid2path(val);
2281 ptlrpc_req_finished(req);
2286 static void lustre_swab_hai(struct hsm_action_item *h)
2288 __swab32s(&h->hai_len);
2289 __swab32s(&h->hai_action);
2290 lustre_swab_lu_fid(&h->hai_fid);
2291 lustre_swab_lu_fid(&h->hai_dfid);
2292 __swab64s(&h->hai_cookie);
2293 __swab64s(&h->hai_extent.offset);
2294 __swab64s(&h->hai_extent.length);
2295 __swab64s(&h->hai_gid);
2298 static void lustre_swab_hal(struct hsm_action_list *h)
2300 struct hsm_action_item *hai;
2303 __swab32s(&h->hal_version);
2304 __swab32s(&h->hal_count);
2305 __swab32s(&h->hal_archive_id);
2306 __swab64s(&h->hal_flags);
2308 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2309 lustre_swab_hai(hai);
2312 static void lustre_swab_kuch(struct kuc_hdr *l)
2314 __swab16s(&l->kuc_magic);
2315 /* __u8 l->kuc_transport */
2316 __swab16s(&l->kuc_msgtype);
2317 __swab16s(&l->kuc_msglen);
2320 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2321 struct lustre_kernelcomm *lk)
2323 struct obd_import *imp = class_exp2cliimp(exp);
2326 if (lk->lk_group != KUC_GRP_HSM) {
2327 CERROR("Bad copytool group %d\n", lk->lk_group);
2331 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2332 lk->lk_uid, lk->lk_group, lk->lk_flags);
2334 if (lk->lk_flags & LK_FLG_STOP) {
2335 /* Unregister with the coordinator */
2336 rc = mdc_ioc_hsm_ct_unregister(imp);
2338 __u32 *archives = NULL;
2340 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2341 archives = lk->lk_data;
2343 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2350 * Send a message to any listening copytools
2351 * @param val KUC message (kuc_hdr + hsm_action_list)
2352 * @param len total length of message
2354 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2355 size_t len, void *val)
2357 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2358 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2362 if (len < sizeof(*lh) + sizeof(*hal)) {
2363 CERROR("Short HSM message %zu < %zu\n", len,
2364 sizeof(*lh) + sizeof(*hal));
2367 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2368 lustre_swab_kuch(lh);
2369 lustre_swab_hal(hal);
2370 } else if (lh->kuc_magic != KUC_MAGIC) {
2371 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2375 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2377 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2378 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2380 /* Broadcast to HSM listeners */
2381 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2387 * callback function passed to kuc for re-registering each HSM copytool
2388 * running on MDC, after MDT shutdown/recovery.
2389 * @param data copytool registration data
2390 * @param cb_arg callback argument (obd_import)
2392 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2394 struct obd_import *imp = (struct obd_import *)cb_arg;
2395 struct kkuc_ct_data *kcd = data;
2396 __u32 *archives = NULL;
2400 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2401 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2404 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2405 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2406 "(archive=%#x)\n", imp->imp_obd->obd_name,
2407 kcd->kcd_nr_archives);
2409 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2410 "(archive nr = %u)\n",
2411 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2412 if (kcd->kcd_nr_archives != 0)
2413 archives = kcd->kcd_archives;
2416 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2417 /* ignore error if the copytool is already registered */
2418 return (rc == -EEXIST) ? 0 : rc;
2422 * Re-establish all kuc contexts with MDT
2423 * after MDT shutdown/recovery.
2425 static int mdc_kuc_reregister(struct obd_import *imp)
2427 /* re-register HSM agents */
2428 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2429 mdc_hsm_ct_reregister, imp);
2432 static int mdc_set_info_async(const struct lu_env *env,
2433 struct obd_export *exp,
2434 u32 keylen, void *key,
2435 u32 vallen, void *val,
2436 struct ptlrpc_request_set *set)
2438 struct obd_import *imp = class_exp2cliimp(exp);
2442 if (KEY_IS(KEY_READ_ONLY)) {
2443 if (vallen != sizeof(int))
2446 spin_lock(&imp->imp_lock);
2447 if (*((int *)val)) {
2448 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2449 imp->imp_connect_data.ocd_connect_flags |=
2452 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2453 imp->imp_connect_data.ocd_connect_flags &=
2454 ~OBD_CONNECT_RDONLY;
2456 spin_unlock(&imp->imp_lock);
2458 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2459 keylen, key, vallen, val, set);
2462 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2463 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2464 keylen, key, vallen, val, set);
2467 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2468 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2473 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2474 __u32 *default_easize = val;
2476 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2480 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2484 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2485 __u32 keylen, void *key, __u32 *vallen, void *val)
2489 if (KEY_IS(KEY_MAX_EASIZE)) {
2490 __u32 mdsize, *max_easize;
2492 if (*vallen != sizeof(int))
2494 mdsize = *(__u32 *)val;
2495 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2496 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2498 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2500 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2501 __u32 *default_easize;
2503 if (*vallen != sizeof(int))
2505 default_easize = val;
2506 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2508 } else if (KEY_IS(KEY_CONN_DATA)) {
2509 struct obd_import *imp = class_exp2cliimp(exp);
2510 struct obd_connect_data *data = val;
2512 if (*vallen != sizeof(*data))
2515 *data = imp->imp_connect_data;
2517 } else if (KEY_IS(KEY_TGT_COUNT)) {
2518 *((__u32 *)val) = 1;
2522 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2527 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2528 struct ptlrpc_request **request)
2530 struct ptlrpc_request *req;
2535 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2539 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2541 ptlrpc_request_free(req);
2545 mdc_pack_body(req, fid, 0, 0, -1, 0);
2547 ptlrpc_request_set_replen(req);
2549 rc = ptlrpc_queue_wait(req);
2551 ptlrpc_req_finished(req);
2557 struct mdc_rmfid_args {
2562 int mdc_rmfid_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2565 struct mdc_rmfid_args *aa;
2570 aa = ptlrpc_req_async_args(aa, req);
2572 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2574 LASSERT(size == sizeof(int) * aa->mra_nr);
2575 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2577 LASSERT(aa->mra_rcs);
2578 LASSERT(aa->mra_nr);
2579 memcpy(aa->mra_rcs, rcs, size);
2585 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2586 int *rcs, struct ptlrpc_request_set *set)
2588 struct ptlrpc_request *req;
2589 struct mdc_rmfid_args *aa;
2595 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2599 flen = fa->fa_nr * sizeof(struct lu_fid);
2600 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2602 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2604 req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2605 RCL_SERVER, fa->fa_nr * sizeof(__u32));
2606 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2608 ptlrpc_request_free(req);
2611 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2612 memcpy(tmp, fa->fa_fids, flen);
2614 mdc_pack_body(req, NULL, 0, 0, -1, 0);
2615 b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2616 b->mbo_ctime = ktime_get_real_seconds();
2618 ptlrpc_request_set_replen(req);
2621 aa = ptlrpc_req_async_args(aa, req);
2623 aa->mra_nr = fa->fa_nr;
2624 req->rq_interpret_reply = mdc_rmfid_interpret;
2626 ptlrpc_set_add_req(set, req);
2627 ptlrpc_check_set(NULL, set);
2632 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2633 enum obd_import_event event)
2635 struct client_obd *cli = &obd->u.cli;
2638 LASSERT(imp->imp_obd == obd);
2641 case IMP_EVENT_DISCON:
2642 spin_lock(&cli->cl_loi_list_lock);
2643 cli->cl_avail_grant = 0;
2644 cli->cl_lost_grant = 0;
2645 spin_unlock(&cli->cl_loi_list_lock);
2647 case IMP_EVENT_INACTIVE:
2649 * Flush current sequence to make client obtain new one
2650 * from server in case of disconnect/reconnect.
2652 down_read(&cli->cl_seq_rwsem);
2654 seq_client_flush(cli->cl_seq);
2655 up_read(&cli->cl_seq_rwsem);
2657 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2659 case IMP_EVENT_INVALIDATE: {
2660 struct ldlm_namespace *ns = obd->obd_namespace;
2664 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2666 env = cl_env_get(&refcheck);
2668 /* Reset grants. All pages go to failing rpcs due to
2669 * the invalid import.
2671 osc_io_unplug(env, cli, NULL);
2673 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2674 osc_ldlm_resource_invalidate,
2676 cl_env_put(env, &refcheck);
2677 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2683 case IMP_EVENT_ACTIVE:
2684 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2685 /* redo the kuc registration after reconnecting */
2687 rc = mdc_kuc_reregister(imp);
2689 case IMP_EVENT_OCD: {
2690 struct obd_connect_data *ocd = &imp->imp_connect_data;
2692 if (OCD_HAS_FLAG(ocd, GRANT))
2693 osc_init_grant(cli, ocd);
2695 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2698 case IMP_EVENT_DEACTIVATE:
2699 case IMP_EVENT_ACTIVATE:
2702 CERROR("Unknown import event %x\n", event);
2708 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2709 struct lu_fid *fid, struct md_op_data *op_data)
2711 struct client_obd *cli = &exp->exp_obd->u.cli;
2716 down_read(&cli->cl_seq_rwsem);
2718 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2719 up_read(&cli->cl_seq_rwsem);
2724 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2726 struct client_obd *cli = &exp->exp_obd->u.cli;
2727 return &cli->cl_target_uuid;
2731 * Determine whether the lock can be canceled before replaying it during
2732 * recovery, non zero value will be return if the lock can be canceled,
2733 * or zero returned for not
2735 static int mdc_cancel_weight(struct ldlm_lock *lock)
2737 if (lock->l_resource->lr_type != LDLM_IBITS)
2740 /* FIXME: if we ever get into a situation where there are too many
2741 * opened files with open locks on a single node, then we really
2742 * should replay these open locks to reget it */
2743 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2746 /* Special case for DoM locks, cancel only unused and granted locks */
2747 if (ldlm_has_dom(lock) &&
2748 (lock->l_granted_mode != lock->l_req_mode ||
2749 osc_ldlm_weigh_ast(lock) != 0))
2755 static int mdc_resource_inode_free(struct ldlm_resource *res)
2757 if (res->lr_lvb_inode)
2758 res->lr_lvb_inode = NULL;
2763 static struct ldlm_valblock_ops inode_lvbo = {
2764 .lvbo_free = mdc_resource_inode_free
2767 static int mdc_llog_init(struct obd_device *obd)
2769 struct obd_llog_group *olg = &obd->obd_olg;
2770 struct llog_ctxt *ctxt;
2775 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2780 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2781 llog_initiator_connect(ctxt);
2782 llog_ctxt_put(ctxt);
2787 static void mdc_llog_finish(struct obd_device *obd)
2789 struct llog_ctxt *ctxt;
2793 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2795 llog_cleanup(NULL, ctxt);
2800 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2806 rc = osc_setup_common(obd, cfg);
2810 rc = mdc_tunables_init(obd);
2812 GOTO(err_osc_cleanup, rc);
2814 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2816 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2818 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2820 rc = mdc_llog_init(obd);
2822 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2824 GOTO(err_llog_cleanup, rc);
2827 rc = mdc_changelog_cdev_init(obd);
2829 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2831 GOTO(err_changelog_cleanup, rc);
2836 err_changelog_cleanup:
2837 mdc_llog_finish(obd);
2839 lprocfs_free_md_stats(obd);
2840 ptlrpc_lprocfs_unregister_obd(obd);
2842 osc_cleanup_common(obd);
2846 /* Initialize the default and maximum LOV EA sizes. This allows
2847 * us to make MDS RPCs with large enough reply buffers to hold a default
2848 * sized EA without having to calculate this (via a call into the
2849 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2850 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2851 * a large number of stripes is possible. If a larger reply buffer is
2852 * required it will be reallocated in the ptlrpc layer due to overflow.
2854 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2857 struct obd_device *obd = exp->exp_obd;
2858 struct client_obd *cli = &obd->u.cli;
2861 if (cli->cl_max_mds_easize < easize)
2862 cli->cl_max_mds_easize = easize;
2864 if (cli->cl_default_mds_easize < def_easize)
2865 cli->cl_default_mds_easize = def_easize;
2870 static int mdc_precleanup(struct obd_device *obd)
2874 osc_precleanup_common(obd);
2875 mdc_changelog_cdev_finish(obd);
2877 obd_cleanup_client_import(obd);
2878 ptlrpc_lprocfs_unregister_obd(obd);
2879 lprocfs_free_md_stats(obd);
2880 mdc_llog_finish(obd);
2884 static int mdc_cleanup(struct obd_device *obd)
2886 return osc_cleanup_common(obd);
2889 static const struct obd_ops mdc_obd_ops = {
2890 .o_owner = THIS_MODULE,
2891 .o_setup = mdc_setup,
2892 .o_precleanup = mdc_precleanup,
2893 .o_cleanup = mdc_cleanup,
2894 .o_add_conn = client_import_add_conn,
2895 .o_del_conn = client_import_del_conn,
2896 .o_connect = client_connect_import,
2897 .o_reconnect = osc_reconnect,
2898 .o_disconnect = osc_disconnect,
2899 .o_iocontrol = mdc_iocontrol,
2900 .o_set_info_async = mdc_set_info_async,
2901 .o_statfs = mdc_statfs,
2902 .o_statfs_async = mdc_statfs_async,
2903 .o_fid_init = client_fid_init,
2904 .o_fid_fini = client_fid_fini,
2905 .o_fid_alloc = mdc_fid_alloc,
2906 .o_import_event = mdc_import_event,
2907 .o_get_info = mdc_get_info,
2908 .o_get_uuid = mdc_get_uuid,
2909 .o_quotactl = mdc_quotactl,
2912 static const struct md_ops mdc_md_ops = {
2913 .m_get_root = mdc_get_root,
2914 .m_null_inode = mdc_null_inode,
2915 .m_close = mdc_close,
2916 .m_create = mdc_create,
2917 .m_enqueue = mdc_enqueue,
2918 .m_getattr = mdc_getattr,
2919 .m_getattr_name = mdc_getattr_name,
2920 .m_intent_lock = mdc_intent_lock,
2922 .m_rename = mdc_rename,
2923 .m_setattr = mdc_setattr,
2924 .m_setxattr = mdc_setxattr,
2925 .m_getxattr = mdc_getxattr,
2926 .m_fsync = mdc_fsync,
2927 .m_file_resync = mdc_file_resync,
2928 .m_read_page = mdc_read_page,
2929 .m_unlink = mdc_unlink,
2930 .m_cancel_unused = mdc_cancel_unused,
2931 .m_init_ea_size = mdc_init_ea_size,
2932 .m_set_lock_data = mdc_set_lock_data,
2933 .m_lock_match = mdc_lock_match,
2934 .m_get_lustre_md = mdc_get_lustre_md,
2935 .m_free_lustre_md = mdc_free_lustre_md,
2936 .m_set_open_replay_data = mdc_set_open_replay_data,
2937 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2938 .m_intent_getattr_async = mdc_intent_getattr_async,
2939 .m_revalidate_lock = mdc_revalidate_lock,
2940 .m_rmfid = mdc_rmfid,
2943 dev_t mdc_changelog_dev;
2944 struct class *mdc_changelog_class;
2945 static int __init mdc_init(void)
2948 rc = alloc_chrdev_region(&mdc_changelog_dev, 0,
2949 MDC_CHANGELOG_DEV_COUNT,
2950 MDC_CHANGELOG_DEV_NAME);
2954 mdc_changelog_class = class_create(THIS_MODULE, MDC_CHANGELOG_DEV_NAME);
2955 if (IS_ERR(mdc_changelog_class)) {
2956 rc = PTR_ERR(mdc_changelog_class);
2960 rc = class_register_type(&mdc_obd_ops, &mdc_md_ops, true,
2961 LUSTRE_MDC_NAME, &mdc_device_type);
2968 class_destroy(mdc_changelog_class);
2970 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
2974 static void __exit mdc_exit(void)
2976 class_unregister_type(LUSTRE_MDC_NAME);
2977 class_destroy(mdc_changelog_class);
2978 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
2979 idr_destroy(&mdc_changelog_minor_idr);
2982 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2983 MODULE_DESCRIPTION("Lustre Metadata Client");
2984 MODULE_VERSION(LUSTRE_VERSION_STRING);
2985 MODULE_LICENSE("GPL");
2987 module_init(mdc_init);
2988 module_exit(mdc_exit);