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->rq_pill, 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->rq_pill, &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->rq_pill, &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 = current_cap().cap[0];
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->rq_pill, fid, valid, output_size,
402 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
403 memcpy(tmp, xattr_name, xattr_namelen);
406 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
407 memcpy(tmp, input, input_size);
410 mdc_file_sepol_pack(&req->rq_pill);
412 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
413 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
414 RCL_SERVER, output_size);
415 ptlrpc_request_set_replen(req);
418 if (opcode == MDS_REINT)
419 ptlrpc_get_mod_rpc_slot(req);
421 rc = ptlrpc_queue_wait(req);
423 if (opcode == MDS_REINT)
424 ptlrpc_put_mod_rpc_slot(req);
427 ptlrpc_req_finished(req);
433 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
434 u64 obd_md_valid, const char *name,
435 const void *value, size_t value_size,
436 unsigned int xattr_flags, u32 suppgid,
437 struct ptlrpc_request **req)
439 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
440 obd_md_valid == OBD_MD_FLXATTRRM);
442 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
443 fid, MDS_REINT, obd_md_valid, name,
444 value, value_size, 0, xattr_flags, suppgid,
448 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
449 u64 obd_md_valid, const char *name, size_t buf_size,
450 struct ptlrpc_request **req)
452 struct mdt_body *body;
455 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
456 obd_md_valid == OBD_MD_FLXATTRLS);
458 /* Message below is checked in sanity-selinux test_20d
459 * and sanity-sec test_49
461 CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
462 exp->exp_obd->obd_name, name, PFID(fid));
463 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
464 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
469 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
471 GOTO(out, rc = -EPROTO);
473 /* only detect the xattr size */
475 /* LU-11109: Older MDTs do not distinguish
476 * between nonexistent xattrs and zero length
477 * values in this case. Newer MDTs will return
478 * -ENODATA or set OBD_MD_FLXATTR. */
479 GOTO(out, rc = body->mbo_eadatasize);
482 if (body->mbo_eadatasize == 0) {
483 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
484 * success so that we can distinguish between
485 * zero length value and nonexistent xattr.
487 * If OBD_MD_FLXATTR is not set then we keep
488 * the old behavior and return -ENODATA for
489 * getxattr() when mbo_eadatasize is 0. But
490 * -ENODATA only makes sense for getxattr()
491 * and not for listxattr(). */
492 if (body->mbo_valid & OBD_MD_FLXATTR)
494 else if (obd_md_valid == OBD_MD_FLXATTR)
495 GOTO(out, rc = -ENODATA);
500 GOTO(out, rc = body->mbo_eadatasize);
503 ptlrpc_req_finished(*req);
510 static int mdc_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
511 struct obd_export *dt_exp,
512 struct obd_export *md_exp,
513 struct lustre_md *md)
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_exp->exp_obd->obd_type->typ_lu == &mdc_device_type) {
554 CERROR("%s: no LMV, upgrading from old version?\n",
555 md_exp->exp_obd->obd_name);
557 GOTO(out_acl, rc = 0);
560 if (md->body->mbo_valid & OBD_MD_MEA) {
561 lmv_size = md->body->mbo_eadatasize;
563 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
564 "but eadatasize 0\n");
568 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
571 GOTO(out, rc = -EPROTO);
573 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
577 if (rc < (int)sizeof(*md->lmv)) {
578 struct lmv_foreign_md *lfm = md->lfm;
580 /* short (< sizeof(struct lmv_stripe_md))
583 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
585 "lmv size too small: %d < %d\n",
586 rc, (int)sizeof(*md->lmv));
587 GOTO(out, rc = -EPROTO);
592 /* since 2.12.58 intent_getattr fetches default LMV */
593 if (md->body->mbo_valid & OBD_MD_DEFAULT_MEA) {
594 lmv_size = sizeof(struct lmv_user_md);
595 lmv = req_capsule_server_sized_get(pill,
599 GOTO(out, rc = -EPROTO);
601 rc = md_unpackmd(md_exp, &md->default_lmv, lmv,
606 if (rc < (int)sizeof(*md->default_lmv)) {
608 "default lmv size too small: %d < %d\n",
609 rc, (int)sizeof(*md->default_lmv));
610 GOTO(out, rc = -EPROTO);
617 if (md->body->mbo_valid & OBD_MD_FLACL) {
618 /* for ACL, it's possible that FLACL is set but aclsize is zero.
619 * only when aclsize != 0 there's an actual segment for ACL
622 rc = mdc_unpack_acl(pill, md);
635 static int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
641 void mdc_replay_open(struct ptlrpc_request *req)
643 struct md_open_data *mod = req->rq_cb_data;
644 struct ptlrpc_request *close_req;
645 struct obd_client_handle *och;
646 struct lustre_handle old_open_handle = { };
647 struct mdt_body *body;
648 struct ldlm_reply *rep;
652 DEBUG_REQ(D_ERROR, req,
653 "cannot properly replay without open data");
658 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
659 LASSERT(body != NULL);
661 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
662 if (rep != NULL && rep->lock_policy_res2 != 0)
663 DEBUG_REQ(D_ERROR, req, "Open request replay failed with %ld ",
664 (long int)rep->lock_policy_res2);
666 spin_lock(&req->rq_lock);
668 if (och && och->och_open_handle.cookie)
669 req->rq_early_free_repbuf = 1;
671 req->rq_early_free_repbuf = 0;
672 spin_unlock(&req->rq_lock);
674 if (req->rq_early_free_repbuf) {
675 struct lustre_handle *file_open_handle;
677 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
679 file_open_handle = &och->och_open_handle;
680 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
681 file_open_handle->cookie, body->mbo_open_handle.cookie);
682 old_open_handle = *file_open_handle;
683 *file_open_handle = body->mbo_open_handle;
686 close_req = mod->mod_close_req;
688 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
689 struct mdt_ioepoch *epoch;
691 LASSERT(opc == MDS_CLOSE);
692 epoch = req_capsule_client_get(&close_req->rq_pill,
696 if (req->rq_early_free_repbuf)
697 LASSERT(old_open_handle.cookie ==
698 epoch->mio_open_handle.cookie);
700 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
701 epoch->mio_open_handle = body->mbo_open_handle;
706 void mdc_commit_open(struct ptlrpc_request *req)
708 struct md_open_data *mod = req->rq_cb_data;
713 * No need to touch md_open_data::mod_och, it holds a reference on
714 * \var mod and will zero references to each other, \var mod will be
715 * freed after that when md_open_data::mod_och will put the reference.
719 * Do not let open request to disappear as it still may be needed
720 * for close rpc to happen (it may happen on evict only, otherwise
721 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
722 * called), just mark this rpc as committed to distinguish these 2
723 * cases, see mdc_close() for details. The open request reference will
724 * be put along with freeing \var mod.
726 ptlrpc_request_addref(req);
727 spin_lock(&req->rq_lock);
728 req->rq_committed = 1;
729 spin_unlock(&req->rq_lock);
730 req->rq_cb_data = NULL;
734 int mdc_set_open_replay_data(struct obd_export *exp,
735 struct obd_client_handle *och,
736 struct lookup_intent *it)
738 struct md_open_data *mod;
739 struct mdt_rec_create *rec;
740 struct mdt_body *body;
741 struct ptlrpc_request *open_req = it->it_request;
742 struct obd_import *imp = open_req->rq_import;
745 if (!open_req->rq_replay)
748 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
749 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
750 LASSERT(rec != NULL);
751 /* Incoming message in my byte order (it's been swabbed). */
752 /* Outgoing messages always in my byte order. */
753 LASSERT(body != NULL);
755 /* Only if the import is replayable, we set replay_open data */
756 if (och && imp->imp_replayable) {
757 mod = obd_mod_alloc();
759 DEBUG_REQ(D_ERROR, open_req,
760 "cannot allocate md_open_data");
765 * Take a reference on \var mod, to be freed on mdc_close().
766 * It protects \var mod from being freed on eviction (commit
767 * callback is called despite rq_replay flag).
768 * Another reference for \var och.
773 spin_lock(&open_req->rq_lock);
776 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
777 it_disposition(it, DISP_OPEN_STRIPE);
778 mod->mod_open_req = open_req;
779 open_req->rq_cb_data = mod;
780 open_req->rq_commit_cb = mdc_commit_open;
781 open_req->rq_early_free_repbuf = 1;
782 spin_unlock(&open_req->rq_lock);
785 rec->cr_fid2 = body->mbo_fid1;
786 rec->cr_open_handle_old = body->mbo_open_handle;
787 open_req->rq_replay_cb = mdc_replay_open;
788 if (!fid_is_sane(&body->mbo_fid1)) {
789 DEBUG_REQ(D_ERROR, open_req,
790 "saving replay request with insane FID " DFID,
791 PFID(&body->mbo_fid1));
795 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
799 static void mdc_free_open(struct md_open_data *mod)
803 if (mod->mod_is_create == 0 &&
804 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
808 * No reason to asssert here if the open request has
809 * rq_replay == 1. It means that mdc_close failed, and
810 * close request wasn`t sent. It is not fatal to client.
811 * The worst thing is eviction if the client gets open lock
814 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req,
815 "free open request, rq_replay=%d",
816 mod->mod_open_req->rq_replay);
818 ptlrpc_request_committed(mod->mod_open_req, committed);
819 if (mod->mod_close_req)
820 ptlrpc_request_committed(mod->mod_close_req, committed);
823 static int mdc_clear_open_replay_data(struct obd_export *exp,
824 struct obd_client_handle *och)
826 struct md_open_data *mod = och->och_mod;
830 * It is possible to not have \var mod in a case of eviction between
831 * lookup and ll_file_open().
836 LASSERT(mod != LP_POISON);
837 LASSERT(mod->mod_open_req != NULL);
839 spin_lock(&mod->mod_open_req->rq_lock);
841 mod->mod_och->och_open_handle.cookie = 0;
842 mod->mod_open_req->rq_early_free_repbuf = 0;
843 spin_unlock(&mod->mod_open_req->rq_lock);
853 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
854 struct md_open_data *mod, struct ptlrpc_request **request)
856 struct obd_device *obd = class_exp2obd(exp);
857 struct ptlrpc_request *req;
858 struct req_format *req_fmt;
859 size_t u32_count = 0;
864 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
865 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
868 if (op_data->op_bias & MDS_CLOSE_INTENT) {
869 req_fmt = &RQF_MDS_CLOSE_INTENT;
870 if (op_data->op_bias & MDS_HSM_RELEASE) {
871 /* allocate a FID for volatile file */
872 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
875 CERROR("%s: "DFID" allocating FID: rc = %d\n",
876 obd->obd_name, PFID(&op_data->op_fid1),
878 /* save the errcode and proceed to close */
882 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
883 size_t count = op_data->op_data_size / sizeof(__u32);
885 if (count > INLINE_RESYNC_ARRAY_SIZE)
889 req_fmt = &RQF_MDS_CLOSE;
893 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
896 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
898 /* Ensure that this close's handle is fixed up during replay. */
899 if (likely(mod != NULL)) {
900 LASSERTF(mod->mod_open_req != NULL &&
901 mod->mod_open_req->rq_type != LI_POISON,
902 "POISONED open %p!\n", mod->mod_open_req);
904 mod->mod_close_req = req;
906 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "matched open");
907 /* We no longer want to preserve this open for replay even
908 * though the open was committed. b=3632, b=3633 */
909 spin_lock(&mod->mod_open_req->rq_lock);
910 mod->mod_open_req->rq_replay = 0;
911 spin_unlock(&mod->mod_open_req->rq_lock);
913 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
917 * TODO: repeat close after errors
919 CWARN("%s: close of FID "DFID" failed, file reference will be "
920 "dropped when this client unmounts or is evicted\n",
921 obd->obd_name, PFID(&op_data->op_fid1));
922 GOTO(out, rc = -ENOMEM);
926 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
927 u32_count * sizeof(__u32));
929 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
931 ptlrpc_request_free(req);
936 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
937 * portal whose threads are not taking any DLM locks and are therefore
938 * always progressing */
939 req->rq_request_portal = MDS_READPAGE_PORTAL;
940 ptlrpc_at_set_req_timeout(req);
942 if (!(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
943 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
944 OP_XVALID_LAZYBLOCKS);
946 mdc_close_pack(&req->rq_pill, op_data);
948 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
949 obd->u.cli.cl_default_mds_easize);
951 ptlrpc_request_set_replen(req);
953 ptlrpc_get_mod_rpc_slot(req);
954 rc = ptlrpc_queue_wait(req);
955 ptlrpc_put_mod_rpc_slot(req);
957 if (req->rq_repmsg == NULL) {
958 CDEBUG(D_RPCTRACE, "request %p failed to send: rc = %d\n", req,
961 rc = req->rq_status ?: -EIO;
962 } else if (rc == 0 || rc == -EAGAIN) {
963 struct mdt_body *body;
965 rc = lustre_msg_get_status(req->rq_repmsg);
966 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
967 DEBUG_REQ(D_ERROR, req,
968 "type = PTL_RPC_MSG_ERR: rc = %d", rc);
972 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
975 } else if (rc == -ESTALE) {
977 * it can be allowed error after 3633 if open was committed and
978 * server failed before close was sent. Let's check if mod
979 * exists and return no error in that case
982 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
983 LASSERT(mod->mod_open_req != NULL);
984 if (mod->mod_open_req->rq_committed)
992 mod->mod_close_req = NULL;
993 /* Since now, mod is accessed through open_req only,
994 * thus close req does not keep a reference on mod anymore. */
999 RETURN(rc < 0 ? rc : saved_rc);
1002 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1003 u64 offset, struct page **pages, int npages,
1004 struct ptlrpc_request **request)
1006 struct ptlrpc_request *req;
1007 struct ptlrpc_bulk_desc *desc;
1016 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1020 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1022 ptlrpc_request_free(req);
1026 req->rq_request_portal = MDS_READPAGE_PORTAL;
1027 ptlrpc_at_set_req_timeout(req);
1029 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1030 PTLRPC_BULK_PUT_SINK,
1032 &ptlrpc_bulk_kiov_pin_ops);
1034 ptlrpc_req_finished(req);
1038 /* NB req now owns desc and will free it when it gets freed */
1039 for (i = 0; i < npages; i++)
1040 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1043 mdc_readdir_pack(&req->rq_pill, offset, PAGE_SIZE * npages, fid);
1045 ptlrpc_request_set_replen(req);
1046 rc = ptlrpc_queue_wait(req);
1048 ptlrpc_req_finished(req);
1049 if (rc != -ETIMEDOUT)
1053 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1054 CERROR("%s: too many resend retries: rc = %d\n",
1055 exp->exp_obd->obd_name, -EIO);
1059 /* If a signal interrupts then the timeout returned will
1060 * not be zero. In that case return -EINTR
1062 if (msleep_interruptible(resends * 1000))
1068 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1069 req->rq_bulk->bd_nob_transferred);
1071 ptlrpc_req_finished(req);
1075 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1076 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1077 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1078 PAGE_SIZE * npages);
1079 ptlrpc_req_finished(req);
1087 static void mdc_release_page(struct page *page, int remove)
1091 if (likely(page->mapping != NULL))
1092 delete_from_page_cache(page);
1098 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1099 __u64 *start, __u64 *end, int hash64)
1102 * Complement of hash is used as an index so that
1103 * radix_tree_gang_lookup() can be used to find a page with starting
1104 * hash _smaller_ than one we are looking for.
1106 unsigned long offset = hash_x_index(*hash, hash64);
1108 unsigned long flags;
1111 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1112 found = radix_tree_gang_lookup(&mapping->page_tree,
1113 (void **)&page, offset, 1);
1114 if (found > 0 && !ll_xa_is_value(page)) {
1115 struct lu_dirpage *dp;
1118 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1120 * In contrast to find_lock_page() we are sure that directory
1121 * page cannot be truncated (while DLM lock is held) and,
1122 * hence, can avoid restart.
1124 * In fact, page cannot be locked here at all, because
1125 * mdc_read_page_remote does synchronous io.
1127 wait_on_page_locked(page);
1128 if (PageUptodate(page)) {
1130 if (BITS_PER_LONG == 32 && hash64) {
1131 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1132 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1133 *hash = *hash >> 32;
1135 *start = le64_to_cpu(dp->ldp_hash_start);
1136 *end = le64_to_cpu(dp->ldp_hash_end);
1138 if (unlikely(*start == 1 && *hash == 0))
1141 LASSERTF(*start <= *hash, "start = %#llx"
1142 ",end = %#llx,hash = %#llx\n",
1143 *start, *end, *hash);
1144 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1145 " hash %#llx\n", offset, *start, *end, *hash);
1148 mdc_release_page(page, 0);
1150 } else if (*end != *start && *hash == *end) {
1152 * upon hash collision, remove this page,
1153 * otherwise put page reference, and
1154 * mdc_read_page_remote() will issue RPC to
1155 * fetch the page we want.
1158 mdc_release_page(page,
1159 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1164 page = ERR_PTR(-EIO);
1167 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1174 * Adjust a set of pages, each page containing an array of lu_dirpages,
1175 * so that each page can be used as a single logical lu_dirpage.
1177 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1178 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1179 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1180 * value is used as a cookie to request the next lu_dirpage in a
1181 * directory listing that spans multiple pages (two in this example):
1184 * .|--------v------- -----.
1185 * |s|e|f|p|ent|ent| ... |ent|
1186 * '--|-------------- -----' Each PAGE contains a single
1187 * '------. lu_dirpage.
1188 * .---------v------- -----.
1189 * |s|e|f|p|ent| 0 | ... | 0 |
1190 * '----------------- -----'
1192 * However, on hosts where the native VM page size (PAGE_SIZE) is
1193 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1194 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1195 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1196 * after it in the same PAGE (arrows simplified for brevity, but
1197 * in general e0==s1, e1==s2, etc.):
1199 * .-------------------- -----.
1200 * |s0|e0|f0|p|ent|ent| ... |ent|
1201 * |---v---------------- -----|
1202 * |s1|e1|f1|p|ent|ent| ... |ent|
1203 * |---v---------------- -----| Here, each PAGE contains
1204 * ... multiple lu_dirpages.
1205 * |---v---------------- -----|
1206 * |s'|e'|f'|p|ent|ent| ... |ent|
1207 * '---|---------------- -----'
1209 * .----------------------------.
1212 * This structure is transformed into a single logical lu_dirpage as follows:
1214 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1215 * labeled 'next PAGE'.
1217 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1218 * a hash collision with the next page exists.
1220 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1221 * to the first entry of the next lu_dirpage.
1223 #if PAGE_SIZE > LU_PAGE_SIZE
1224 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1228 for (i = 0; i < cfs_pgs; i++) {
1229 struct lu_dirpage *dp = kmap(pages[i]);
1230 struct lu_dirpage *first = dp;
1231 struct lu_dirent *end_dirent = NULL;
1232 struct lu_dirent *ent;
1233 __u64 hash_end = dp->ldp_hash_end;
1234 __u32 flags = dp->ldp_flags;
1236 while (--lu_pgs > 0) {
1237 ent = lu_dirent_start(dp);
1238 for (end_dirent = ent; ent != NULL;
1239 end_dirent = ent, ent = lu_dirent_next(ent));
1241 /* Advance dp to next lu_dirpage. */
1242 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1244 /* Check if we've reached the end of the PAGE. */
1245 if (!((unsigned long)dp & ~PAGE_MASK))
1248 /* Save the hash and flags of this lu_dirpage. */
1249 hash_end = dp->ldp_hash_end;
1250 flags = dp->ldp_flags;
1252 /* Check if lu_dirpage contains no entries. */
1253 if (end_dirent == NULL)
1256 /* Enlarge the end entry lde_reclen from 0 to
1257 * first entry of next lu_dirpage. */
1258 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1259 end_dirent->lde_reclen =
1260 cpu_to_le16((char *)(dp->ldp_entries) -
1261 (char *)end_dirent);
1264 first->ldp_hash_end = hash_end;
1265 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1266 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1270 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1273 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1274 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1276 /* parameters for readdir page */
1277 struct readpage_param {
1278 struct md_op_data *rp_mod;
1281 struct obd_export *rp_exp;
1282 struct md_callback *rp_cb;
1286 * Read pages from server.
1288 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1289 * a header lu_dirpage which describes the start/end hash, and whether this
1290 * page is empty (contains no dir entry) or hash collide with next page.
1291 * After client receives reply, several pages will be integrated into dir page
1292 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1293 * lu_dirpage for this integrated page will be adjusted.
1295 static int mdc_read_page_remote(void *data, struct page *page0)
1297 struct readpage_param *rp = data;
1298 struct page **page_pool;
1300 struct lu_dirpage *dp;
1301 struct md_op_data *op_data = rp->rp_mod;
1302 struct ptlrpc_request *req;
1304 struct inode *inode;
1306 int rd_pgs = 0; /* number of pages actually read */
1312 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1313 inode = op_data->op_data;
1314 fid = &op_data->op_fid1;
1315 LASSERT(inode != NULL);
1317 OBD_ALLOC_PTR_ARRAY_LARGE(page_pool, max_pages);
1318 if (page_pool != NULL) {
1319 page_pool[0] = page0;
1325 for (npages = 1; npages < max_pages; npages++) {
1326 page = page_cache_alloc(inode->i_mapping);
1329 page_pool[npages] = page;
1332 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1334 /* page0 is special, which was added into page cache early */
1335 delete_from_page_cache(page0);
1339 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1341 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1342 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1344 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1346 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1348 SetPageUptodate(page0);
1352 ptlrpc_req_finished(req);
1353 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1354 for (i = 1; i < npages; i++) {
1355 unsigned long offset;
1359 page = page_pool[i];
1361 if (rc < 0 || i >= rd_pgs) {
1366 SetPageUptodate(page);
1369 hash = le64_to_cpu(dp->ldp_hash_start);
1372 offset = hash_x_index(hash, rp->rp_hash64);
1374 prefetchw(&page->flags);
1375 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1380 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1381 " rc = %d\n", offset, ret);
1385 if (page_pool != &page0)
1386 OBD_FREE_PTR_ARRAY_LARGE(page_pool, max_pages);
1392 * Read dir page from cache first, if it can not find it, read it from
1393 * server and add into the cache.
1395 * \param[in] exp MDC export
1396 * \param[in] op_data client MD stack parameters, transfering parameters
1397 * between different layers on client MD stack.
1398 * \param[in] cb_op callback required for ldlm lock enqueue during
1400 * \param[in] hash_offset the hash offset of the page to be read
1401 * \param[in] ppage the page to be read
1403 * retval = 0 get the page successfully
1404 * errno(<0) get the page failed
1406 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1407 struct md_callback *cb_op, __u64 hash_offset,
1408 struct page **ppage)
1410 struct lookup_intent it = { .it_op = IT_READDIR };
1412 struct inode *dir = op_data->op_data;
1413 struct address_space *mapping;
1414 struct lu_dirpage *dp;
1417 struct lustre_handle lockh;
1418 struct ptlrpc_request *enq_req = NULL;
1419 struct readpage_param rp_param;
1426 LASSERT(dir != NULL);
1427 mapping = dir->i_mapping;
1429 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1430 cb_op->md_blocking_ast, 0);
1431 if (enq_req != NULL)
1432 ptlrpc_req_finished(enq_req);
1435 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1436 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1441 lockh.cookie = it.it_lock_handle;
1442 mdc_set_lock_data(exp, &lockh, dir, NULL);
1444 rp_param.rp_off = hash_offset;
1445 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1446 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1447 rp_param.rp_hash64);
1449 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1450 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1451 rp_param.rp_off, PTR_ERR(page));
1452 GOTO(out_unlock, rc = PTR_ERR(page));
1453 } else if (page != NULL) {
1455 * XXX nikita: not entirely correct handling of a corner case:
1456 * suppose hash chain of entries with hash value HASH crosses
1457 * border between pages P0 and P1. First both P0 and P1 are
1458 * cached, seekdir() is called for some entry from the P0 part
1459 * of the chain. Later P0 goes out of cache. telldir(HASH)
1460 * happens and finds P1, as it starts with matching hash
1461 * value. Remaining entries from P0 part of the chain are
1462 * skipped. (Is that really a bug?)
1464 * Possible solutions: 0. don't cache P1 is such case, handle
1465 * it as an "overflow" page. 1. invalidate all pages at
1466 * once. 2. use HASH|1 as an index for P1.
1468 GOTO(hash_collision, page);
1471 rp_param.rp_exp = exp;
1472 rp_param.rp_mod = op_data;
1473 page = read_cache_page(mapping,
1474 hash_x_index(rp_param.rp_off,
1475 rp_param.rp_hash64),
1476 mdc_read_page_remote, &rp_param);
1478 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1479 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1480 rp_param.rp_off, PTR_ERR(page));
1481 GOTO(out_unlock, rc = PTR_ERR(page));
1484 wait_on_page_locked(page);
1486 if (!PageUptodate(page)) {
1487 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1488 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1489 rp_param.rp_off, -5);
1492 if (!PageChecked(page))
1493 SetPageChecked(page);
1494 if (PageError(page)) {
1495 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1496 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1497 rp_param.rp_off, -5);
1502 dp = page_address(page);
1503 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1504 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1505 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1506 rp_param.rp_off = hash_offset >> 32;
1508 start = le64_to_cpu(dp->ldp_hash_start);
1509 end = le64_to_cpu(dp->ldp_hash_end);
1510 rp_param.rp_off = hash_offset;
1513 LASSERT(start == rp_param.rp_off);
1514 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1515 #if BITS_PER_LONG == 32
1516 CWARN("Real page-wide hash collision at [%llu %llu] with "
1517 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1518 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1522 * Fetch whole overflow chain...
1530 ldlm_lock_decref(&lockh, it.it_lock_mode);
1534 mdc_release_page(page, 1);
1539 static int mdc_statfs_interpret(const struct lu_env *env,
1540 struct ptlrpc_request *req, void *args, int rc)
1542 struct obd_info *oinfo = args;
1543 struct obd_statfs *osfs;
1546 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1550 oinfo->oi_osfs = osfs;
1552 CDEBUG(D_CACHE, "blocks=%llu free=%llu avail=%llu "
1553 "objects=%llu free=%llu state=%x\n",
1554 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
1555 osfs->os_files, osfs->os_ffree, osfs->os_state);
1558 oinfo->oi_cb_up(oinfo, rc);
1563 static int mdc_statfs_async(struct obd_export *exp,
1564 struct obd_info *oinfo, time64_t max_age,
1565 struct ptlrpc_request_set *unused)
1567 struct ptlrpc_request *req;
1568 struct obd_info *aa;
1570 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_STATFS,
1571 LUSTRE_MDS_VERSION, MDS_STATFS);
1575 ptlrpc_request_set_replen(req);
1576 req->rq_interpret_reply = mdc_statfs_interpret;
1578 aa = ptlrpc_req_async_args(aa, req);
1581 ptlrpcd_add_req(req);
1586 static int mdc_statfs(const struct lu_env *env,
1587 struct obd_export *exp, struct obd_statfs *osfs,
1588 time64_t max_age, __u32 flags)
1590 struct obd_device *obd = class_exp2obd(exp);
1591 struct req_format *fmt;
1592 struct ptlrpc_request *req;
1593 struct obd_statfs *msfs;
1594 struct obd_import *imp, *imp0;
1599 * Since the request might also come from lprocfs, so we need
1600 * sync this with client_disconnect_export Bug15684
1602 with_imp_locked(obd, imp0, rc)
1603 imp = class_import_get(imp0);
1607 fmt = &RQF_MDS_STATFS;
1608 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1609 (flags & OBD_STATFS_SUM))
1610 fmt = &RQF_MDS_STATFS_NEW;
1611 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1614 GOTO(output, rc = -ENOMEM);
1615 req->rq_allow_intr = 1;
1617 if ((flags & OBD_STATFS_SUM) &&
1618 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1619 /* request aggregated states */
1620 struct mdt_body *body;
1622 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1624 GOTO(out, rc = -EPROTO);
1625 body->mbo_valid = OBD_MD_FLAGSTATFS;
1628 ptlrpc_request_set_replen(req);
1630 if (flags & OBD_STATFS_NODELAY) {
1631 /* procfs requests not want stay in wait for avoid deadlock */
1632 req->rq_no_resend = 1;
1633 req->rq_no_delay = 1;
1636 rc = ptlrpc_queue_wait(req);
1638 /* check connection error first */
1639 if (imp->imp_connect_error)
1640 rc = imp->imp_connect_error;
1644 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1646 GOTO(out, rc = -EPROTO);
1651 ptlrpc_req_finished(req);
1653 class_import_put(imp);
1657 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1659 __u32 keylen, vallen;
1663 if (gf->gf_pathlen > PATH_MAX)
1664 RETURN(-ENAMETOOLONG);
1665 if (gf->gf_pathlen < 2)
1668 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1669 keylen = cfs_size_round(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1670 sizeof(struct lu_fid));
1671 OBD_ALLOC(key, keylen);
1674 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1675 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)), gf, sizeof(*gf));
1676 memcpy(key + cfs_size_round(sizeof(KEY_FID2PATH)) + sizeof(*gf),
1677 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1678 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1679 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1681 if (!fid_is_sane(&gf->gf_fid))
1682 GOTO(out, rc = -EINVAL);
1684 /* Val is struct getinfo_fid2path result plus path */
1685 vallen = sizeof(*gf) + gf->gf_pathlen;
1687 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1688 if (rc != 0 && rc != -EREMOTE)
1691 if (vallen <= sizeof(*gf))
1692 GOTO(out, rc = -EPROTO);
1693 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1694 GOTO(out, rc = -EOVERFLOW);
1696 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %s\n",
1697 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1698 gf->gf_pathlen < 512 ? gf->gf_u.gf_path :
1699 /* only log the last 512 characters of the path */
1700 gf->gf_u.gf_path + gf->gf_pathlen - 512);
1703 OBD_FREE(key, keylen);
1707 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1708 struct hsm_progress_kernel *hpk)
1710 struct obd_import *imp = class_exp2cliimp(exp);
1711 struct hsm_progress_kernel *req_hpk;
1712 struct ptlrpc_request *req;
1716 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1717 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1719 GOTO(out, rc = -ENOMEM);
1721 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1723 /* Copy hsm_progress struct */
1724 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1725 if (req_hpk == NULL)
1726 GOTO(out, rc = -EPROTO);
1729 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1731 ptlrpc_request_set_replen(req);
1733 ptlrpc_get_mod_rpc_slot(req);
1734 rc = ptlrpc_queue_wait(req);
1735 ptlrpc_put_mod_rpc_slot(req);
1739 ptlrpc_req_finished(req);
1743 * Send hsm_ct_register to MDS
1745 * \param[in] imp import
1746 * \param[in] archive_count if in bitmap format, it is the bitmap,
1747 * else it is the count of archive_ids
1748 * \param[in] archives if in bitmap format, it is NULL,
1749 * else it is archive_id lists
1751 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1754 struct ptlrpc_request *req;
1755 __u32 *archive_array;
1756 size_t archives_size;
1760 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1764 if (archives != NULL)
1765 archives_size = sizeof(*archive_array) * archive_count;
1767 archives_size = sizeof(archive_count);
1769 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1770 RCL_CLIENT, archives_size);
1772 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1774 ptlrpc_request_free(req);
1778 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1780 archive_array = req_capsule_client_get(&req->rq_pill,
1781 &RMF_MDS_HSM_ARCHIVE);
1782 if (archive_array == NULL)
1783 GOTO(out, rc = -EPROTO);
1785 if (archives != NULL)
1786 memcpy(archive_array, archives, archives_size);
1788 *archive_array = archive_count;
1790 ptlrpc_request_set_replen(req);
1791 req->rq_no_resend = 1;
1793 rc = mdc_queue_wait(req);
1796 ptlrpc_req_finished(req);
1800 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1801 struct md_op_data *op_data)
1803 struct hsm_current_action *hca = op_data->op_data;
1804 struct hsm_current_action *req_hca;
1805 struct ptlrpc_request *req;
1809 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1810 &RQF_MDS_HSM_ACTION);
1814 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1816 ptlrpc_request_free(req);
1820 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1821 op_data->op_suppgids[0], 0);
1823 ptlrpc_request_set_replen(req);
1825 rc = mdc_queue_wait(req);
1829 req_hca = req_capsule_server_get(&req->rq_pill,
1830 &RMF_MDS_HSM_CURRENT_ACTION);
1831 if (req_hca == NULL)
1832 GOTO(out, rc = -EPROTO);
1838 ptlrpc_req_finished(req);
1842 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1844 struct ptlrpc_request *req;
1848 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1850 MDS_HSM_CT_UNREGISTER);
1852 GOTO(out, rc = -ENOMEM);
1854 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1856 ptlrpc_request_set_replen(req);
1858 rc = mdc_queue_wait(req);
1861 ptlrpc_req_finished(req);
1865 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1866 struct md_op_data *op_data)
1868 struct hsm_user_state *hus = op_data->op_data;
1869 struct hsm_user_state *req_hus;
1870 struct ptlrpc_request *req;
1874 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1875 &RQF_MDS_HSM_STATE_GET);
1879 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1881 ptlrpc_request_free(req);
1885 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1886 op_data->op_suppgids[0], 0);
1888 ptlrpc_request_set_replen(req);
1890 rc = mdc_queue_wait(req);
1894 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1895 if (req_hus == NULL)
1896 GOTO(out, rc = -EPROTO);
1902 ptlrpc_req_finished(req);
1906 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1907 struct md_op_data *op_data)
1909 struct hsm_state_set *hss = op_data->op_data;
1910 struct hsm_state_set *req_hss;
1911 struct ptlrpc_request *req;
1915 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1916 &RQF_MDS_HSM_STATE_SET);
1920 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1922 ptlrpc_request_free(req);
1926 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1927 op_data->op_suppgids[0], 0);
1930 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1931 if (req_hss == NULL)
1932 GOTO(out, rc = -EPROTO);
1935 ptlrpc_request_set_replen(req);
1937 ptlrpc_get_mod_rpc_slot(req);
1938 rc = ptlrpc_queue_wait(req);
1939 ptlrpc_put_mod_rpc_slot(req);
1943 ptlrpc_req_finished(req);
1947 static int mdc_ioc_hsm_request(struct obd_export *exp,
1948 struct hsm_user_request *hur)
1950 struct obd_import *imp = class_exp2cliimp(exp);
1951 struct ptlrpc_request *req;
1952 struct hsm_request *req_hr;
1953 struct hsm_user_item *req_hui;
1958 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
1960 GOTO(out, rc = -ENOMEM);
1962 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
1963 hur->hur_request.hr_itemcount
1964 * sizeof(struct hsm_user_item));
1965 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
1966 hur->hur_request.hr_data_len);
1968 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
1970 ptlrpc_request_free(req);
1974 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1976 /* Copy hsm_request struct */
1977 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
1979 GOTO(out, rc = -EPROTO);
1980 *req_hr = hur->hur_request;
1982 /* Copy hsm_user_item structs */
1983 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
1984 if (req_hui == NULL)
1985 GOTO(out, rc = -EPROTO);
1986 memcpy(req_hui, hur->hur_user_item,
1987 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
1989 /* Copy opaque field */
1990 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
1991 if (req_opaque == NULL)
1992 GOTO(out, rc = -EPROTO);
1993 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
1995 ptlrpc_request_set_replen(req);
1997 ptlrpc_get_mod_rpc_slot(req);
1998 rc = ptlrpc_queue_wait(req);
1999 ptlrpc_put_mod_rpc_slot(req);
2004 ptlrpc_req_finished(req);
2008 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2009 struct lustre_kernelcomm *lk);
2011 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2012 struct obd_quotactl *oqctl)
2014 struct ptlrpc_request *req;
2015 struct obd_quotactl *oqc;
2019 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_QUOTACTL);
2024 if (LUSTRE_Q_CMD_IS_POOL(oqctl->qc_cmd))
2025 req_capsule_set_size(&req->rq_pill,
2028 sizeof(*oqc) + LOV_MAXPOOLNAME + 1);
2030 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION,
2033 ptlrpc_request_free(req);
2037 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2038 QCTL_COPY(oqc, oqctl);
2040 ptlrpc_request_set_replen(req);
2041 ptlrpc_at_set_req_timeout(req);
2043 rc = ptlrpc_queue_wait(req);
2045 CERROR("%s: ptlrpc_queue_wait failed: rc = %d\n",
2046 exp->exp_obd->obd_name, rc);
2050 if (req->rq_repmsg &&
2051 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2052 QCTL_COPY(oqctl, oqc);
2055 CERROR("%s: cannot unpack obd_quotactl: rc = %d\n",
2056 exp->exp_obd->obd_name, rc);
2059 ptlrpc_req_finished(req);
2064 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2065 struct md_op_data *op_data)
2068 struct ptlrpc_request *req;
2070 struct mdc_swap_layouts *msl, *payload;
2073 msl = op_data->op_data;
2075 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2076 * first thing it will do is to cancel the 2 layout
2077 * locks held by this client.
2078 * So the client must cancel its layout locks on the 2 fids
2079 * with the request RPC to avoid extra RPC round trips.
2081 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2082 LCK_EX, MDS_INODELOCK_LAYOUT |
2083 MDS_INODELOCK_XATTR);
2084 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2085 LCK_EX, MDS_INODELOCK_LAYOUT |
2086 MDS_INODELOCK_XATTR);
2088 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2089 &RQF_MDS_SWAP_LAYOUTS);
2091 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2095 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2097 ptlrpc_request_free(req);
2101 mdc_swap_layouts_pack(&req->rq_pill, op_data);
2103 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2108 ptlrpc_request_set_replen(req);
2110 rc = ptlrpc_queue_wait(req);
2116 ptlrpc_req_finished(req);
2120 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2121 void *karg, void __user *uarg)
2123 struct obd_device *obd = exp->exp_obd;
2124 struct obd_ioctl_data *data = karg;
2125 struct obd_import *imp = obd->u.cli.cl_import;
2129 if (!try_module_get(THIS_MODULE)) {
2130 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2131 module_name(THIS_MODULE));
2135 case OBD_IOC_FID2PATH:
2136 rc = mdc_ioc_fid2path(exp, karg);
2138 case LL_IOC_HSM_CT_START:
2139 rc = mdc_ioc_hsm_ct_start(exp, karg);
2140 /* ignore if it was already registered on this MDS. */
2144 case LL_IOC_HSM_PROGRESS:
2145 rc = mdc_ioc_hsm_progress(exp, karg);
2147 case LL_IOC_HSM_STATE_GET:
2148 rc = mdc_ioc_hsm_state_get(exp, karg);
2150 case LL_IOC_HSM_STATE_SET:
2151 rc = mdc_ioc_hsm_state_set(exp, karg);
2153 case LL_IOC_HSM_ACTION:
2154 rc = mdc_ioc_hsm_current_action(exp, karg);
2156 case LL_IOC_HSM_REQUEST:
2157 rc = mdc_ioc_hsm_request(exp, karg);
2159 case OBD_IOC_CLIENT_RECOVER:
2160 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2164 case IOC_OSC_SET_ACTIVE:
2165 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2168 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2169 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2170 * there'd be no LMV layer thus we might be called here. Eventually
2171 * this code should be removed.
2174 case IOC_OBD_STATFS: {
2175 struct obd_statfs stat_buf = {0};
2177 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2178 GOTO(out, rc = -ENODEV);
2181 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2182 min((int)data->ioc_plen2,
2183 (int)sizeof(struct obd_uuid))))
2184 GOTO(out, rc = -EFAULT);
2186 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2187 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2192 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2193 min((int) data->ioc_plen1,
2194 (int) sizeof(stat_buf))))
2195 GOTO(out, rc = -EFAULT);
2199 case OBD_IOC_QUOTACTL: {
2200 struct if_quotactl *qctl = karg;
2201 struct obd_quotactl *oqctl;
2203 OBD_ALLOC_PTR(oqctl);
2205 GOTO(out, rc = -ENOMEM);
2207 QCTL_COPY(oqctl, qctl);
2208 rc = obd_quotactl(exp, oqctl);
2210 QCTL_COPY(qctl, oqctl);
2211 qctl->qc_valid = QC_MDTIDX;
2212 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2215 OBD_FREE_PTR(oqctl);
2218 case LL_IOC_GET_CONNECT_FLAGS:
2219 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2220 sizeof(*exp_connect_flags_ptr(exp))))
2221 GOTO(out, rc = -EFAULT);
2224 case LL_IOC_LOV_SWAP_LAYOUTS:
2225 rc = mdc_ioc_swap_layouts(exp, karg);
2228 CERROR("unrecognised ioctl: cmd = %#x\n", cmd);
2229 GOTO(out, rc = -ENOTTY);
2232 module_put(THIS_MODULE);
2237 static int mdc_get_info_rpc(struct obd_export *exp,
2238 u32 keylen, void *key,
2239 u32 vallen, void *val)
2241 struct obd_import *imp = class_exp2cliimp(exp);
2242 struct ptlrpc_request *req;
2247 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2251 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2252 RCL_CLIENT, keylen);
2253 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2254 RCL_CLIENT, sizeof(vallen));
2256 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2258 ptlrpc_request_free(req);
2262 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2263 memcpy(tmp, key, keylen);
2264 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2265 memcpy(tmp, &vallen, sizeof(vallen));
2267 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2268 RCL_SERVER, vallen);
2269 ptlrpc_request_set_replen(req);
2271 /* if server failed to resolve FID, and OI scrub not able to fix it, it
2272 * will return -EINPROGRESS, ptlrpc_queue_wait() will keep retrying,
2273 * set request interruptible to avoid deadlock.
2275 if (KEY_IS(KEY_FID2PATH))
2276 req->rq_allow_intr = 1;
2278 rc = ptlrpc_queue_wait(req);
2279 /* -EREMOTE means the get_info result is partial, and it needs to
2280 * continue on another MDT, see fid2path part in lmv_iocontrol */
2281 if (rc == 0 || rc == -EREMOTE) {
2282 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2283 memcpy(val, tmp, vallen);
2284 if (req_capsule_rep_need_swab(&req->rq_pill)) {
2285 if (KEY_IS(KEY_FID2PATH))
2286 lustre_swab_fid2path(val);
2289 ptlrpc_req_finished(req);
2294 static void lustre_swab_hai(struct hsm_action_item *h)
2296 __swab32s(&h->hai_len);
2297 __swab32s(&h->hai_action);
2298 lustre_swab_lu_fid(&h->hai_fid);
2299 lustre_swab_lu_fid(&h->hai_dfid);
2300 __swab64s(&h->hai_cookie);
2301 __swab64s(&h->hai_extent.offset);
2302 __swab64s(&h->hai_extent.length);
2303 __swab64s(&h->hai_gid);
2306 static void lustre_swab_hal(struct hsm_action_list *h)
2308 struct hsm_action_item *hai;
2311 __swab32s(&h->hal_version);
2312 __swab32s(&h->hal_count);
2313 __swab32s(&h->hal_archive_id);
2314 __swab64s(&h->hal_flags);
2316 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2317 lustre_swab_hai(hai);
2320 static void lustre_swab_kuch(struct kuc_hdr *l)
2322 __swab16s(&l->kuc_magic);
2323 /* __u8 l->kuc_transport */
2324 __swab16s(&l->kuc_msgtype);
2325 __swab16s(&l->kuc_msglen);
2328 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2329 struct lustre_kernelcomm *lk)
2331 struct obd_import *imp = class_exp2cliimp(exp);
2334 if (lk->lk_group != KUC_GRP_HSM) {
2335 CERROR("Bad copytool group %d\n", lk->lk_group);
2339 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2340 lk->lk_uid, lk->lk_group, lk->lk_flags);
2342 if (lk->lk_flags & LK_FLG_STOP) {
2343 /* Unregister with the coordinator */
2344 rc = mdc_ioc_hsm_ct_unregister(imp);
2346 __u32 *archives = NULL;
2348 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2349 archives = lk->lk_data;
2351 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2358 * Send a message to any listening copytools
2359 * @param val KUC message (kuc_hdr + hsm_action_list)
2360 * @param len total length of message
2362 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2363 size_t len, void *val)
2365 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2366 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2370 if (len < sizeof(*lh) + sizeof(*hal)) {
2371 CERROR("Short HSM message %zu < %zu\n", len,
2372 sizeof(*lh) + sizeof(*hal));
2375 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2376 lustre_swab_kuch(lh);
2377 lustre_swab_hal(hal);
2378 } else if (lh->kuc_magic != KUC_MAGIC) {
2379 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2383 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2385 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2386 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2388 /* Broadcast to HSM listeners */
2389 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2395 * callback function passed to kuc for re-registering each HSM copytool
2396 * running on MDC, after MDT shutdown/recovery.
2397 * @param data copytool registration data
2398 * @param cb_arg callback argument (obd_import)
2400 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2402 struct obd_import *imp = (struct obd_import *)cb_arg;
2403 struct kkuc_ct_data *kcd = data;
2404 __u32 *archives = NULL;
2408 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2409 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2412 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2413 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2414 "(archive=%#x)\n", imp->imp_obd->obd_name,
2415 kcd->kcd_nr_archives);
2417 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2418 "(archive nr = %u)\n",
2419 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2420 if (kcd->kcd_nr_archives != 0)
2421 archives = kcd->kcd_archives;
2424 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2425 /* ignore error if the copytool is already registered */
2426 return (rc == -EEXIST) ? 0 : rc;
2430 * Re-establish all kuc contexts with MDT
2431 * after MDT shutdown/recovery.
2433 static int mdc_kuc_reregister(struct obd_import *imp)
2435 /* re-register HSM agents */
2436 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2437 mdc_hsm_ct_reregister, imp);
2440 static int mdc_set_info_async(const struct lu_env *env,
2441 struct obd_export *exp,
2442 u32 keylen, void *key,
2443 u32 vallen, void *val,
2444 struct ptlrpc_request_set *set)
2446 struct obd_import *imp = class_exp2cliimp(exp);
2450 if (KEY_IS(KEY_READ_ONLY)) {
2451 if (vallen != sizeof(int))
2454 spin_lock(&imp->imp_lock);
2455 if (*((int *)val)) {
2456 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2457 imp->imp_connect_data.ocd_connect_flags |=
2460 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2461 imp->imp_connect_data.ocd_connect_flags &=
2462 ~OBD_CONNECT_RDONLY;
2464 spin_unlock(&imp->imp_lock);
2466 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2467 keylen, key, vallen, val, set);
2470 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2471 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2472 keylen, key, vallen, val, set);
2475 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2476 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2481 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2482 __u32 *default_easize = val;
2484 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2488 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2492 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2493 __u32 keylen, void *key, __u32 *vallen, void *val)
2497 if (KEY_IS(KEY_MAX_EASIZE)) {
2498 __u32 mdsize, *max_easize;
2500 if (*vallen != sizeof(int))
2502 mdsize = *(__u32 *)val;
2503 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2504 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2506 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2508 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2509 __u32 *default_easize;
2511 if (*vallen != sizeof(int))
2513 default_easize = val;
2514 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2516 } else if (KEY_IS(KEY_CONN_DATA)) {
2517 struct obd_import *imp = class_exp2cliimp(exp);
2518 struct obd_connect_data *data = val;
2520 if (*vallen != sizeof(*data))
2523 *data = imp->imp_connect_data;
2525 } else if (KEY_IS(KEY_TGT_COUNT)) {
2526 *((__u32 *)val) = 1;
2530 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2535 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2536 struct ptlrpc_request **request)
2538 struct ptlrpc_request *req;
2543 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2547 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2549 ptlrpc_request_free(req);
2553 mdc_pack_body(&req->rq_pill, fid, 0, 0, -1, 0);
2555 ptlrpc_request_set_replen(req);
2557 rc = ptlrpc_queue_wait(req);
2559 ptlrpc_req_finished(req);
2565 struct mdc_rmfid_args {
2570 int mdc_rmfid_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2573 struct mdc_rmfid_args *aa;
2578 aa = ptlrpc_req_async_args(aa, req);
2580 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2582 LASSERT(size == sizeof(int) * aa->mra_nr);
2583 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2585 LASSERT(aa->mra_rcs);
2586 LASSERT(aa->mra_nr);
2587 memcpy(aa->mra_rcs, rcs, size);
2593 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2594 int *rcs, struct ptlrpc_request_set *set)
2596 struct ptlrpc_request *req;
2597 struct mdc_rmfid_args *aa;
2603 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2607 flen = fa->fa_nr * sizeof(struct lu_fid);
2608 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2610 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2612 req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2613 RCL_SERVER, fa->fa_nr * sizeof(__u32));
2614 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2616 ptlrpc_request_free(req);
2619 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2620 memcpy(tmp, fa->fa_fids, flen);
2622 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
2623 b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2624 b->mbo_ctime = ktime_get_real_seconds();
2626 ptlrpc_request_set_replen(req);
2629 aa = ptlrpc_req_async_args(aa, req);
2631 aa->mra_nr = fa->fa_nr;
2632 req->rq_interpret_reply = mdc_rmfid_interpret;
2634 ptlrpc_set_add_req(set, req);
2635 ptlrpc_check_set(NULL, set);
2640 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2641 enum obd_import_event event)
2643 struct client_obd *cli = &obd->u.cli;
2646 LASSERT(imp->imp_obd == obd);
2649 case IMP_EVENT_DISCON:
2650 spin_lock(&cli->cl_loi_list_lock);
2651 cli->cl_avail_grant = 0;
2652 cli->cl_lost_grant = 0;
2653 spin_unlock(&cli->cl_loi_list_lock);
2655 case IMP_EVENT_INACTIVE:
2657 * Flush current sequence to make client obtain new one
2658 * from server in case of disconnect/reconnect.
2660 down_read(&cli->cl_seq_rwsem);
2662 seq_client_flush(cli->cl_seq);
2663 up_read(&cli->cl_seq_rwsem);
2665 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2667 case IMP_EVENT_INVALIDATE: {
2668 struct ldlm_namespace *ns = obd->obd_namespace;
2672 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2674 env = cl_env_get(&refcheck);
2676 /* Reset grants. All pages go to failing rpcs due to
2677 * the invalid import.
2679 osc_io_unplug(env, cli, NULL);
2681 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2682 osc_ldlm_resource_invalidate,
2684 cl_env_put(env, &refcheck);
2685 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2691 case IMP_EVENT_ACTIVE:
2692 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2693 /* redo the kuc registration after reconnecting */
2695 rc = mdc_kuc_reregister(imp);
2697 case IMP_EVENT_OCD: {
2698 struct obd_connect_data *ocd = &imp->imp_connect_data;
2700 if (OCD_HAS_FLAG(ocd, GRANT))
2701 osc_init_grant(cli, ocd);
2703 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2706 case IMP_EVENT_DEACTIVATE:
2707 case IMP_EVENT_ACTIVATE:
2710 CERROR("Unknown import event %x\n", event);
2716 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2717 struct lu_fid *fid, struct md_op_data *op_data)
2719 struct client_obd *cli = &exp->exp_obd->u.cli;
2724 down_read(&cli->cl_seq_rwsem);
2726 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2727 up_read(&cli->cl_seq_rwsem);
2732 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2734 struct client_obd *cli = &exp->exp_obd->u.cli;
2735 return &cli->cl_target_uuid;
2739 * Determine whether the lock can be canceled before replaying it during
2740 * recovery, non zero value will be return if the lock can be canceled,
2741 * or zero returned for not
2743 static int mdc_cancel_weight(struct ldlm_lock *lock)
2745 if (lock->l_resource->lr_type != LDLM_IBITS)
2748 /* FIXME: if we ever get into a situation where there are too many
2749 * opened files with open locks on a single node, then we really
2750 * should replay these open locks to reget it */
2751 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2754 /* Special case for DoM locks, cancel only unused and granted locks */
2755 if (ldlm_has_dom(lock) &&
2756 (lock->l_granted_mode != lock->l_req_mode ||
2757 osc_ldlm_weigh_ast(lock) != 0))
2763 static int mdc_resource_inode_free(struct ldlm_resource *res)
2765 if (res->lr_lvb_inode)
2766 res->lr_lvb_inode = NULL;
2771 static struct ldlm_valblock_ops inode_lvbo = {
2772 .lvbo_free = mdc_resource_inode_free
2775 static int mdc_llog_init(struct obd_device *obd)
2777 struct obd_llog_group *olg = &obd->obd_olg;
2778 struct llog_ctxt *ctxt;
2783 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2788 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2789 llog_initiator_connect(ctxt);
2790 llog_ctxt_put(ctxt);
2795 static void mdc_llog_finish(struct obd_device *obd)
2797 struct llog_ctxt *ctxt;
2801 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2803 llog_cleanup(NULL, ctxt);
2808 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2814 rc = osc_setup_common(obd, cfg);
2818 rc = mdc_tunables_init(obd);
2820 GOTO(err_osc_cleanup, rc);
2822 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2824 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2826 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2828 rc = mdc_llog_init(obd);
2830 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2832 GOTO(err_llog_cleanup, rc);
2835 rc = mdc_changelog_cdev_init(obd);
2837 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2839 GOTO(err_changelog_cleanup, rc);
2844 err_changelog_cleanup:
2845 mdc_llog_finish(obd);
2847 lprocfs_free_md_stats(obd);
2848 ptlrpc_lprocfs_unregister_obd(obd);
2850 osc_cleanup_common(obd);
2854 /* Initialize the default and maximum LOV EA sizes. This allows
2855 * us to make MDS RPCs with large enough reply buffers to hold a default
2856 * sized EA without having to calculate this (via a call into the
2857 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2858 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2859 * a large number of stripes is possible. If a larger reply buffer is
2860 * required it will be reallocated in the ptlrpc layer due to overflow.
2862 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2865 struct obd_device *obd = exp->exp_obd;
2866 struct client_obd *cli = &obd->u.cli;
2869 if (cli->cl_max_mds_easize < easize)
2870 cli->cl_max_mds_easize = easize;
2872 if (cli->cl_default_mds_easize < def_easize)
2873 cli->cl_default_mds_easize = def_easize;
2878 static int mdc_precleanup(struct obd_device *obd)
2882 osc_precleanup_common(obd);
2883 mdc_changelog_cdev_finish(obd);
2885 obd_cleanup_client_import(obd);
2886 ptlrpc_lprocfs_unregister_obd(obd);
2887 lprocfs_free_md_stats(obd);
2888 mdc_llog_finish(obd);
2892 static int mdc_cleanup(struct obd_device *obd)
2894 return osc_cleanup_common(obd);
2897 static const struct obd_ops mdc_obd_ops = {
2898 .o_owner = THIS_MODULE,
2899 .o_setup = mdc_setup,
2900 .o_precleanup = mdc_precleanup,
2901 .o_cleanup = mdc_cleanup,
2902 .o_add_conn = client_import_add_conn,
2903 .o_del_conn = client_import_del_conn,
2904 .o_connect = client_connect_import,
2905 .o_reconnect = osc_reconnect,
2906 .o_disconnect = osc_disconnect,
2907 .o_iocontrol = mdc_iocontrol,
2908 .o_set_info_async = mdc_set_info_async,
2909 .o_statfs = mdc_statfs,
2910 .o_statfs_async = mdc_statfs_async,
2911 .o_fid_init = client_fid_init,
2912 .o_fid_fini = client_fid_fini,
2913 .o_fid_alloc = mdc_fid_alloc,
2914 .o_import_event = mdc_import_event,
2915 .o_get_info = mdc_get_info,
2916 .o_get_uuid = mdc_get_uuid,
2917 .o_quotactl = mdc_quotactl,
2920 static const struct md_ops mdc_md_ops = {
2921 .m_get_root = mdc_get_root,
2922 .m_null_inode = mdc_null_inode,
2923 .m_close = mdc_close,
2924 .m_create = mdc_create,
2925 .m_enqueue = mdc_enqueue,
2926 .m_getattr = mdc_getattr,
2927 .m_getattr_name = mdc_getattr_name,
2928 .m_intent_lock = mdc_intent_lock,
2930 .m_rename = mdc_rename,
2931 .m_setattr = mdc_setattr,
2932 .m_setxattr = mdc_setxattr,
2933 .m_getxattr = mdc_getxattr,
2934 .m_fsync = mdc_fsync,
2935 .m_file_resync = mdc_file_resync,
2936 .m_read_page = mdc_read_page,
2937 .m_unlink = mdc_unlink,
2938 .m_cancel_unused = mdc_cancel_unused,
2939 .m_init_ea_size = mdc_init_ea_size,
2940 .m_set_lock_data = mdc_set_lock_data,
2941 .m_lock_match = mdc_lock_match,
2942 .m_get_lustre_md = mdc_get_lustre_md,
2943 .m_free_lustre_md = mdc_free_lustre_md,
2944 .m_set_open_replay_data = mdc_set_open_replay_data,
2945 .m_clear_open_replay_data = mdc_clear_open_replay_data,
2946 .m_intent_getattr_async = mdc_intent_getattr_async,
2947 .m_revalidate_lock = mdc_revalidate_lock,
2948 .m_rmfid = mdc_rmfid,
2951 dev_t mdc_changelog_dev;
2952 struct class *mdc_changelog_class;
2953 static int __init mdc_init(void)
2956 rc = alloc_chrdev_region(&mdc_changelog_dev, 0,
2957 MDC_CHANGELOG_DEV_COUNT,
2958 MDC_CHANGELOG_DEV_NAME);
2962 mdc_changelog_class = class_create(THIS_MODULE, MDC_CHANGELOG_DEV_NAME);
2963 if (IS_ERR(mdc_changelog_class)) {
2964 rc = PTR_ERR(mdc_changelog_class);
2968 rc = class_register_type(&mdc_obd_ops, &mdc_md_ops, true,
2969 LUSTRE_MDC_NAME, &mdc_device_type);
2976 class_destroy(mdc_changelog_class);
2978 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
2982 static void __exit mdc_exit(void)
2984 class_unregister_type(LUSTRE_MDC_NAME);
2985 class_destroy(mdc_changelog_class);
2986 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
2987 idr_destroy(&mdc_changelog_minor_idr);
2990 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2991 MODULE_DESCRIPTION("Lustre Metadata Client");
2992 MODULE_VERSION(LUSTRE_VERSION_STRING);
2993 MODULE_LICENSE("GPL");
2995 module_init(mdc_init);
2996 module_exit(mdc_exit);