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_ioctl_old.h>
55 #include <lustre_kernelcomm.h>
56 #include <lustre_lmv.h>
57 #include <lustre_log.h>
58 #include <lustre_osc.h>
59 #include <lustre_swab.h>
60 #include <obd_class.h>
62 #include "mdc_internal.h"
64 #define REQUEST_MINOR 244
66 static int mdc_cleanup(struct obd_device *obd);
68 static inline int mdc_queue_wait(struct ptlrpc_request *req)
70 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
73 /* obd_get_request_slot() ensures that this client has no more
74 * than cl_max_rpcs_in_flight RPCs simultaneously inf light
76 rc = obd_get_request_slot(cli);
80 rc = ptlrpc_queue_wait(req);
81 obd_put_request_slot(cli);
87 * Send MDS_GET_ROOT RPC to fetch root FID.
89 * If \a fileset is not NULL it should contain a subdirectory off
90 * the ROOT/ directory to be mounted on the client. Return the FID
91 * of the subdirectory to the client to mount onto its mountpoint.
93 * \param[in] imp MDC import
94 * \param[in] fileset fileset name, which could be NULL
95 * \param[out] rootfid root FID of this mountpoint
96 * \param[out] pc root capa will be unpacked and saved in this pointer
98 * \retval 0 on success, negative errno on failure
100 static int mdc_get_root(struct obd_export *exp, const char *fileset,
101 struct lu_fid *rootfid)
103 struct ptlrpc_request *req;
104 struct mdt_body *body;
109 if (fileset && !(exp_connect_flags(exp) & OBD_CONNECT_SUBTREE))
112 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
118 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
119 strlen(fileset) + 1);
120 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_ROOT);
122 ptlrpc_request_free(req);
125 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
126 if (fileset != NULL) {
127 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
129 memcpy(name, fileset, strlen(fileset));
131 lustre_msg_add_flags(req->rq_reqmsg, LUSTRE_IMP_FULL);
132 req->rq_send_state = LUSTRE_IMP_FULL;
134 ptlrpc_request_set_replen(req);
136 rc = ptlrpc_queue_wait(req);
140 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
142 GOTO(out, rc = -EPROTO);
144 *rootfid = body->mbo_fid1;
145 CDEBUG(D_NET, "root fid="DFID", last_committed=%llu\n",
146 PFID(rootfid), lustre_msg_get_last_committed(req->rq_repmsg));
149 ptlrpc_req_finished(req);
155 * This function now is known to always saying that it will receive 4 buffers
156 * from server. Even for cases when acl_size and md_size is zero, RPC header
157 * will contain 4 fields and RPC itself will contain zero size fields. This is
158 * because mdt_getattr*() _always_ returns 4 fields, but if acl is not needed
159 * and thus zero, it shrinks it, making zero size. The same story about
160 * md_size. And this is course of problem when client waits for smaller number
161 * of fields. This issue will be fixed later when client gets aware of RPC
164 static int mdc_getattr_common(struct obd_export *exp,
165 struct ptlrpc_request *req,
166 struct md_op_data *op_data)
168 struct req_capsule *pill = &req->rq_pill;
169 struct mdt_body *body;
174 /* Request message already built. */
175 rc = ptlrpc_queue_wait(req);
179 /* sanity check for the reply */
180 body = req_capsule_server_get(pill, &RMF_MDT_BODY);
184 CDEBUG(D_NET, "mode: %o\n", body->mbo_mode);
186 mdc_update_max_ea_from_body(exp, body);
187 if (body->mbo_eadatasize != 0) {
188 eadata = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
189 body->mbo_eadatasize);
194 /* If encryption context was returned by MDT, put it in op_data
195 * so that caller can set it on inode and save an extra getxattr.
197 if (op_data && op_data->op_valid & OBD_MD_ENCCTX &&
198 body->mbo_valid & OBD_MD_ENCCTX) {
199 op_data->op_file_encctx =
200 req_capsule_server_get(pill, &RMF_FILE_ENCCTX);
201 op_data->op_file_encctx_size =
202 req_capsule_get_size(pill, &RMF_FILE_ENCCTX,
209 static void mdc_reset_acl_req(struct ptlrpc_request *req)
211 spin_lock(&req->rq_early_free_lock);
212 sptlrpc_cli_free_repbuf(req);
213 req->rq_repbuf = NULL;
214 req->rq_repbuf_len = 0;
215 req->rq_repdata = NULL;
216 req->rq_reqdata_len = 0;
217 spin_unlock(&req->rq_early_free_lock);
220 static int mdc_getattr(struct obd_export *exp, struct md_op_data *op_data,
221 struct ptlrpc_request **request)
223 struct ptlrpc_request *req;
224 struct obd_device *obd = class_exp2obd(exp);
225 struct obd_import *imp = class_exp2cliimp(exp);
226 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
230 /* Single MDS without an LMV case */
231 if (op_data->op_flags & MF_GET_MDT_IDX) {
237 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR);
241 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR);
243 ptlrpc_request_free(req);
247 /* LU-15245: avoid deadlock with modifying RPCs on MDS_REQUEST_PORTAL */
248 req->rq_request_portal = MDS_READPAGE_PORTAL;
251 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, op_data->op_valid,
252 op_data->op_mode, -1, 0);
253 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
254 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
256 if (exp_connect_encrypt(exp) && op_data->op_valid & OBD_MD_ENCCTX)
257 req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX,
259 obd->u.cli.cl_max_mds_easize);
261 req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX,
263 ptlrpc_request_set_replen(req);
265 rc = mdc_getattr_common(exp, req, op_data);
268 acl_bufsize = min_t(__u32,
269 imp->imp_connect_data.ocd_max_easize,
271 mdc_reset_acl_req(req);
275 ptlrpc_req_finished(req);
283 static int mdc_getattr_name(struct obd_export *exp, struct md_op_data *op_data,
284 struct ptlrpc_request **request)
286 struct ptlrpc_request *req;
287 struct obd_import *imp = class_exp2cliimp(exp);
288 __u32 acl_bufsize = LUSTRE_POSIX_ACL_MAX_SIZE_OLD;
293 req = ptlrpc_request_alloc(imp, &RQF_MDS_GETATTR_NAME);
297 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
298 op_data->op_namelen + 1);
300 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GETATTR_NAME);
302 ptlrpc_request_free(req);
306 if (op_data->op_name) {
307 char *name = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
308 LASSERT(strnlen(op_data->op_name, op_data->op_namelen) ==
309 op_data->op_namelen);
310 memcpy(name, op_data->op_name, op_data->op_namelen);
314 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, op_data->op_valid,
315 op_data->op_mode, op_data->op_suppgids[0], 0);
316 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
318 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, acl_bufsize);
319 req_capsule_set_size(&req->rq_pill, &RMF_FILE_ENCCTX, RCL_SERVER, 0);
320 ptlrpc_request_set_replen(req);
321 if (op_data->op_bias & MDS_FID_OP) {
322 struct mdt_body *b = req_capsule_client_get(&req->rq_pill,
326 b->mbo_valid |= OBD_MD_NAMEHASH;
327 b->mbo_fid2 = op_data->op_fid2;
331 rc = mdc_getattr_common(exp, req, NULL);
334 acl_bufsize = min_t(__u32,
335 imp->imp_connect_data.ocd_max_easize,
337 mdc_reset_acl_req(req);
341 ptlrpc_req_finished(req);
349 static int mdc_xattr_common(struct obd_export *exp,const struct req_format *fmt,
350 const struct lu_fid *fid, int opcode, u64 valid,
351 const char *xattr_name, const char *input,
352 int input_size, int output_size, int flags,
353 __u32 suppgid, struct ptlrpc_request **request)
355 struct ptlrpc_request *req;
356 int xattr_namelen = 0;
362 req = ptlrpc_request_alloc(class_exp2cliimp(exp), fmt);
367 xattr_namelen = strlen(xattr_name) + 1;
368 req_capsule_set_size(&req->rq_pill, &RMF_NAME, RCL_CLIENT,
373 req_capsule_set_size(&req->rq_pill, &RMF_EADATA, RCL_CLIENT,
376 /* get SELinux policy info if any */
377 rc = sptlrpc_get_sepol(req);
379 ptlrpc_request_free(req);
382 req_capsule_set_size(&req->rq_pill, &RMF_SELINUX_POL, RCL_CLIENT,
383 strlen(req->rq_sepol) ?
384 strlen(req->rq_sepol) + 1 : 0);
386 /* Flush local XATTR locks to get rid of a possible cancel RPC */
387 if (opcode == MDS_REINT && fid_is_sane(fid) &&
388 exp->exp_connect_data.ocd_ibits_known & MDS_INODELOCK_XATTR) {
392 /* Without that packing would fail */
394 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
397 count = mdc_resource_get_unused(exp, fid,
399 MDS_INODELOCK_XATTR);
401 rc = mdc_prep_elc_req(exp, req, MDS_REINT, &cancels, count);
403 ptlrpc_request_free(req);
407 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, opcode);
409 ptlrpc_request_free(req);
414 if (opcode == MDS_REINT) {
415 struct mdt_rec_setxattr *rec;
417 BUILD_BUG_ON(sizeof(struct mdt_rec_setxattr) !=
418 sizeof(struct mdt_rec_reint));
419 rec = req_capsule_client_get(&req->rq_pill, &RMF_REC_REINT);
420 rec->sx_opcode = REINT_SETXATTR;
421 rec->sx_fsuid = from_kuid(&init_user_ns, current_fsuid());
422 rec->sx_fsgid = from_kgid(&init_user_ns, current_fsgid());
423 rec->sx_cap = current_cap().cap[0];
424 rec->sx_suppgid1 = suppgid;
425 rec->sx_suppgid2 = -1;
427 rec->sx_valid = valid | OBD_MD_FLCTIME;
428 rec->sx_time = ktime_get_real_seconds();
429 rec->sx_size = output_size;
430 rec->sx_flags = flags;
432 mdc_pack_body(&req->rq_pill, fid, valid, output_size,
434 /* Avoid deadlock with modifying RPCs on MDS_REQUEST_PORTAL.
437 req->rq_request_portal = MDS_READPAGE_PORTAL;
441 tmp = req_capsule_client_get(&req->rq_pill, &RMF_NAME);
442 memcpy(tmp, xattr_name, xattr_namelen);
445 tmp = req_capsule_client_get(&req->rq_pill, &RMF_EADATA);
446 memcpy(tmp, input, input_size);
449 mdc_file_sepol_pack(&req->rq_pill);
451 if (req_capsule_has_field(&req->rq_pill, &RMF_EADATA, RCL_SERVER))
452 req_capsule_set_size(&req->rq_pill, &RMF_EADATA,
453 RCL_SERVER, output_size);
454 ptlrpc_request_set_replen(req);
457 if (opcode == MDS_REINT)
458 ptlrpc_get_mod_rpc_slot(req);
460 rc = ptlrpc_queue_wait(req);
462 if (opcode == MDS_REINT)
463 ptlrpc_put_mod_rpc_slot(req);
466 ptlrpc_req_finished(req);
472 static int mdc_setxattr(struct obd_export *exp, const struct lu_fid *fid,
473 u64 obd_md_valid, const char *name,
474 const void *value, size_t value_size,
475 unsigned int xattr_flags, u32 suppgid,
476 struct ptlrpc_request **req)
478 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
479 obd_md_valid == OBD_MD_FLXATTRRM);
481 return mdc_xattr_common(exp, &RQF_MDS_REINT_SETXATTR,
482 fid, MDS_REINT, obd_md_valid, name,
483 value, value_size, 0, xattr_flags, suppgid,
487 static int mdc_getxattr(struct obd_export *exp, const struct lu_fid *fid,
488 u64 obd_md_valid, const char *name, size_t buf_size,
489 struct ptlrpc_request **req)
491 struct mdt_body *body;
494 LASSERT(obd_md_valid == OBD_MD_FLXATTR ||
495 obd_md_valid == OBD_MD_FLXATTRLS);
497 /* Message below is checked in sanity-selinux test_20d
498 * and sanity-sec test_49
500 CDEBUG(D_INFO, "%s: get xattr '%s' for "DFID"\n",
501 exp->exp_obd->obd_name, name, PFID(fid));
502 rc = mdc_xattr_common(exp, &RQF_MDS_GETXATTR, fid, MDS_GETXATTR,
503 obd_md_valid, name, NULL, 0, buf_size, 0, -1,
508 body = req_capsule_server_get(&(*req)->rq_pill, &RMF_MDT_BODY);
510 GOTO(out, rc = -EPROTO);
512 /* only detect the xattr size */
514 /* LU-11109: Older MDTs do not distinguish
515 * between nonexistent xattrs and zero length
516 * values in this case. Newer MDTs will return
517 * -ENODATA or set OBD_MD_FLXATTR. */
518 GOTO(out, rc = body->mbo_eadatasize);
521 if (body->mbo_eadatasize == 0) {
522 /* LU-11109: Newer MDTs set OBD_MD_FLXATTR on
523 * success so that we can distinguish between
524 * zero length value and nonexistent xattr.
526 * If OBD_MD_FLXATTR is not set then we keep
527 * the old behavior and return -ENODATA for
528 * getxattr() when mbo_eadatasize is 0. But
529 * -ENODATA only makes sense for getxattr()
530 * and not for listxattr(). */
531 if (body->mbo_valid & OBD_MD_FLXATTR)
533 else if (obd_md_valid == OBD_MD_FLXATTR)
534 GOTO(out, rc = -ENODATA);
539 GOTO(out, rc = body->mbo_eadatasize);
542 ptlrpc_req_finished(*req);
549 static int mdc_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
550 struct obd_export *dt_exp,
551 struct obd_export *md_exp,
552 struct lustre_md *md)
558 memset(md, 0, sizeof(*md));
560 md->body = req_capsule_server_get(pill, &RMF_MDT_BODY);
561 LASSERT(md->body != NULL);
563 if (md->body->mbo_valid & OBD_MD_FLEASIZE) {
564 if (!S_ISREG(md->body->mbo_mode)) {
565 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, should be a "
566 "regular file, but is not\n");
567 GOTO(out, rc = -EPROTO);
570 if (md->body->mbo_eadatasize == 0) {
571 CDEBUG(D_INFO, "OBD_MD_FLEASIZE set, "
572 "but eadatasize 0\n");
573 GOTO(out, rc = -EPROTO);
576 md->layout.lb_len = md->body->mbo_eadatasize;
577 md->layout.lb_buf = req_capsule_server_sized_get(pill,
580 if (md->layout.lb_buf == NULL)
581 GOTO(out, rc = -EPROTO);
582 } else if (md->body->mbo_valid & OBD_MD_FLDIREA) {
583 const union lmv_mds_md *lmv;
586 if (!S_ISDIR(md->body->mbo_mode)) {
587 CDEBUG(D_INFO, "OBD_MD_FLDIREA set, should be a "
588 "directory, but is not\n");
589 GOTO(out, rc = -EPROTO);
592 if (md_exp->exp_obd->obd_type->typ_lu == &mdc_device_type) {
593 CERROR("%s: no LMV, upgrading from old version?\n",
594 md_exp->exp_obd->obd_name);
596 GOTO(out_acl, rc = 0);
599 if (md->body->mbo_valid & OBD_MD_MEA) {
600 lmv_size = md->body->mbo_eadatasize;
602 CDEBUG(D_INFO, "OBD_MD_FLDIREA is set, "
603 "but eadatasize 0\n");
607 lmv = req_capsule_server_sized_get(pill, &RMF_MDT_MD,
610 GOTO(out, rc = -EPROTO);
612 rc = md_unpackmd(md_exp, &md->lmv, lmv, lmv_size);
616 if (rc < (int)sizeof(*md->lmv)) {
617 struct lmv_foreign_md *lfm = md->lfm;
619 /* short (< sizeof(struct lmv_stripe_md))
622 if (lfm->lfm_magic != LMV_MAGIC_FOREIGN) {
624 "lmv size too small: %d < %d\n",
625 rc, (int)sizeof(*md->lmv));
626 GOTO(out, rc = -EPROTO);
631 /* since 2.12.58 intent_getattr fetches default LMV */
632 if (md->body->mbo_valid & OBD_MD_DEFAULT_MEA) {
633 lmv_size = sizeof(struct lmv_user_md);
634 lmv = req_capsule_server_sized_get(pill,
638 GOTO(out, rc = -EPROTO);
640 rc = md_unpackmd(md_exp, &md->default_lmv, lmv,
645 if (rc < (int)sizeof(*md->default_lmv)) {
647 "default lmv size too small: %d < %d\n",
648 rc, (int)sizeof(*md->default_lmv));
649 GOTO(out, rc = -EPROTO);
656 if (md->body->mbo_valid & OBD_MD_FLACL) {
657 /* for ACL, it's possible that FLACL is set but aclsize is zero.
658 * only when aclsize != 0 there's an actual segment for ACL
661 rc = mdc_unpack_acl(pill, md);
674 static int mdc_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
680 void mdc_replay_open(struct ptlrpc_request *req)
682 struct md_open_data *mod = req->rq_cb_data;
683 struct ptlrpc_request *close_req;
684 struct obd_client_handle *och;
685 struct lustre_handle old_open_handle = { };
686 struct mdt_body *body;
687 struct ldlm_reply *rep;
691 DEBUG_REQ(D_ERROR, req,
692 "cannot properly replay without open data");
697 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
698 LASSERT(body != NULL);
700 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
701 if (rep != NULL && rep->lock_policy_res2 != 0)
702 DEBUG_REQ(D_ERROR, req, "Open request replay failed with %ld ",
703 (long int)rep->lock_policy_res2);
705 spin_lock(&req->rq_lock);
707 if (och && och->och_open_handle.cookie)
708 req->rq_early_free_repbuf = 1;
710 req->rq_early_free_repbuf = 0;
711 spin_unlock(&req->rq_lock);
713 if (req->rq_early_free_repbuf) {
714 struct lustre_handle *file_open_handle;
716 LASSERT(och->och_magic == OBD_CLIENT_HANDLE_MAGIC);
718 file_open_handle = &och->och_open_handle;
719 CDEBUG(D_HA, "updating handle from %#llx to %#llx\n",
720 file_open_handle->cookie, body->mbo_open_handle.cookie);
721 old_open_handle = *file_open_handle;
722 *file_open_handle = body->mbo_open_handle;
725 close_req = mod->mod_close_req;
727 __u32 opc = lustre_msg_get_opc(close_req->rq_reqmsg);
728 struct mdt_ioepoch *epoch;
730 LASSERT(opc == MDS_CLOSE);
731 epoch = req_capsule_client_get(&close_req->rq_pill,
735 if (req->rq_early_free_repbuf)
736 LASSERT(old_open_handle.cookie ==
737 epoch->mio_open_handle.cookie);
739 DEBUG_REQ(D_HA, close_req, "updating close body with new fh");
740 epoch->mio_open_handle = body->mbo_open_handle;
745 void mdc_commit_open(struct ptlrpc_request *req)
747 struct md_open_data *mod = req->rq_cb_data;
752 * No need to touch md_open_data::mod_och, it holds a reference on
753 * \var mod and will zero references to each other, \var mod will be
754 * freed after that when md_open_data::mod_och will put the reference.
758 * Do not let open request to disappear as it still may be needed
759 * for close rpc to happen (it may happen on evict only, otherwise
760 * ptlrpc_request::rq_replay does not let mdc_commit_open() to be
761 * called), just mark this rpc as committed to distinguish these 2
762 * cases, see mdc_close() for details. The open request reference will
763 * be put along with freeing \var mod.
765 ptlrpc_request_addref(req);
766 spin_lock(&req->rq_lock);
767 req->rq_committed = 1;
768 spin_unlock(&req->rq_lock);
769 req->rq_cb_data = NULL;
773 int mdc_set_open_replay_data(struct obd_export *exp,
774 struct obd_client_handle *och,
775 struct lookup_intent *it)
777 struct md_open_data *mod;
778 struct mdt_rec_create *rec;
779 struct mdt_body *body;
780 struct ptlrpc_request *open_req = it->it_request;
781 struct obd_import *imp = open_req->rq_import;
784 if (!open_req->rq_replay)
787 rec = req_capsule_client_get(&open_req->rq_pill, &RMF_REC_REINT);
788 body = req_capsule_server_get(&open_req->rq_pill, &RMF_MDT_BODY);
789 LASSERT(rec != NULL);
790 /* Incoming message in my byte order (it's been swabbed). */
791 /* Outgoing messages always in my byte order. */
792 LASSERT(body != NULL);
794 /* Only if the import is replayable, we set replay_open data */
795 if (och && imp->imp_replayable) {
796 mod = obd_mod_alloc();
798 DEBUG_REQ(D_ERROR, open_req,
799 "cannot allocate md_open_data");
804 * Take a reference on \var mod, to be freed on mdc_close().
805 * It protects \var mod from being freed on eviction (commit
806 * callback is called despite rq_replay flag).
807 * Another reference for \var och.
812 spin_lock(&open_req->rq_lock);
815 mod->mod_is_create = it_disposition(it, DISP_OPEN_CREATE) ||
816 it_disposition(it, DISP_OPEN_STRIPE);
817 mod->mod_open_req = open_req;
818 open_req->rq_cb_data = mod;
819 open_req->rq_commit_cb = mdc_commit_open;
820 open_req->rq_early_free_repbuf = 1;
821 spin_unlock(&open_req->rq_lock);
824 rec->cr_fid2 = body->mbo_fid1;
825 rec->cr_open_handle_old = body->mbo_open_handle;
826 open_req->rq_replay_cb = mdc_replay_open;
827 if (!fid_is_sane(&body->mbo_fid1)) {
828 DEBUG_REQ(D_ERROR, open_req,
829 "saving replay request with insane FID " DFID,
830 PFID(&body->mbo_fid1));
834 DEBUG_REQ(D_RPCTRACE, open_req, "Set up open replay data");
838 static void mdc_free_open(struct md_open_data *mod)
842 if (mod->mod_is_create == 0 &&
843 imp_connect_disp_stripe(mod->mod_open_req->rq_import))
847 * No reason to asssert here if the open request has
848 * rq_replay == 1. It means that mdc_close failed, and
849 * close request wasn`t sent. It is not fatal to client.
850 * The worst thing is eviction if the client gets open lock
853 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req,
854 "free open request, rq_replay=%d",
855 mod->mod_open_req->rq_replay);
857 ptlrpc_request_committed(mod->mod_open_req, committed);
858 if (mod->mod_close_req)
859 ptlrpc_request_committed(mod->mod_close_req, committed);
862 static int mdc_clear_open_replay_data(struct obd_export *exp,
863 struct obd_client_handle *och)
865 struct md_open_data *mod = och->och_mod;
869 * It is possible to not have \var mod in a case of eviction between
870 * lookup and ll_file_open().
875 LASSERT(mod != LP_POISON);
876 LASSERT(mod->mod_open_req != NULL);
878 spin_lock(&mod->mod_open_req->rq_lock);
880 mod->mod_och->och_open_handle.cookie = 0;
881 mod->mod_open_req->rq_early_free_repbuf = 0;
882 spin_unlock(&mod->mod_open_req->rq_lock);
892 static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
893 struct md_open_data *mod, struct ptlrpc_request **request)
895 struct obd_device *obd = class_exp2obd(exp);
896 struct ptlrpc_request *req;
897 struct req_format *req_fmt;
898 size_t u32_count = 0;
903 CDEBUG(D_INODE, "%s: "DFID" file closed with intent: %x\n",
904 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
907 if (op_data->op_bias & MDS_CLOSE_INTENT) {
908 req_fmt = &RQF_MDS_CLOSE_INTENT;
909 if (op_data->op_bias & MDS_HSM_RELEASE) {
910 /* allocate a FID for volatile file */
911 rc = mdc_fid_alloc(NULL, exp, &op_data->op_fid2,
914 CERROR("%s: "DFID" allocating FID: rc = %d\n",
915 obd->obd_name, PFID(&op_data->op_fid1),
917 /* save the errcode and proceed to close */
921 if (op_data->op_bias & MDS_CLOSE_RESYNC_DONE) {
922 size_t count = op_data->op_data_size / sizeof(__u32);
924 if (count > INLINE_RESYNC_ARRAY_SIZE)
928 req_fmt = &RQF_MDS_CLOSE;
932 if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
935 req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
937 /* Ensure that this close's handle is fixed up during replay. */
938 if (likely(mod != NULL)) {
939 LASSERTF(mod->mod_open_req != NULL &&
940 mod->mod_open_req->rq_type != LI_POISON,
941 "POISONED open %p!\n", mod->mod_open_req);
943 mod->mod_close_req = req;
945 DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "matched open");
946 /* We no longer want to preserve this open for replay even
947 * though the open was committed. b=3632, b=3633 */
948 spin_lock(&mod->mod_open_req->rq_lock);
949 mod->mod_open_req->rq_replay = 0;
950 spin_unlock(&mod->mod_open_req->rq_lock);
952 CDEBUG(D_HA, "couldn't find open req; expecting close error\n");
956 * TODO: repeat close after errors
958 CWARN("%s: close of FID "DFID" failed, file reference will be "
959 "dropped when this client unmounts or is evicted\n",
960 obd->obd_name, PFID(&op_data->op_fid1));
961 GOTO(out, rc = -ENOMEM);
965 req_capsule_set_size(&req->rq_pill, &RMF_U32, RCL_CLIENT,
966 u32_count * sizeof(__u32));
968 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
970 ptlrpc_request_free(req);
975 /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
976 * portal whose threads are not taking any DLM locks and are therefore
977 * always progressing */
978 req->rq_request_portal = MDS_READPAGE_PORTAL;
979 ptlrpc_at_set_req_timeout(req);
981 if (!obd->u.cli.cl_lsom_update ||
982 !(exp_connect_flags2(exp) & OBD_CONNECT2_LSOM))
983 op_data->op_xvalid &= ~(OP_XVALID_LAZYSIZE |
984 OP_XVALID_LAZYBLOCKS);
986 mdc_close_pack(&req->rq_pill, op_data);
988 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER,
989 obd->u.cli.cl_default_mds_easize);
991 ptlrpc_request_set_replen(req);
993 ptlrpc_get_mod_rpc_slot(req);
994 rc = ptlrpc_queue_wait(req);
995 ptlrpc_put_mod_rpc_slot(req);
997 if (req->rq_repmsg == NULL) {
998 CDEBUG(D_RPCTRACE, "request %p failed to send: rc = %d\n", req,
1001 rc = req->rq_status ?: -EIO;
1002 } else if (rc == 0 || rc == -EAGAIN) {
1003 struct mdt_body *body;
1005 rc = lustre_msg_get_status(req->rq_repmsg);
1006 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1007 DEBUG_REQ(D_ERROR, req,
1008 "type = PTL_RPC_MSG_ERR: rc = %d", rc);
1012 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1015 } else if (rc == -ESTALE) {
1017 * it can be allowed error after 3633 if open was committed and
1018 * server failed before close was sent. Let's check if mod
1019 * exists and return no error in that case
1022 DEBUG_REQ(D_HA, req, "Reset ESTALE = %d", rc);
1023 LASSERT(mod->mod_open_req != NULL);
1024 if (mod->mod_open_req->rq_committed)
1032 mod->mod_close_req = NULL;
1033 /* Since now, mod is accessed through open_req only,
1034 * thus close req does not keep a reference on mod anymore. */
1039 RETURN(rc < 0 ? rc : saved_rc);
1042 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1043 u64 offset, struct page **pages, int npages,
1044 struct ptlrpc_request **request)
1046 struct ptlrpc_request *req;
1047 struct ptlrpc_bulk_desc *desc;
1056 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1060 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1062 ptlrpc_request_free(req);
1066 req->rq_request_portal = MDS_READPAGE_PORTAL;
1067 ptlrpc_at_set_req_timeout(req);
1069 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1070 PTLRPC_BULK_PUT_SINK,
1072 &ptlrpc_bulk_kiov_pin_ops);
1074 ptlrpc_req_finished(req);
1078 /* NB req now owns desc and will free it when it gets freed */
1079 for (i = 0; i < npages; i++)
1080 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1083 mdc_readdir_pack(&req->rq_pill, offset, PAGE_SIZE * npages, fid);
1085 ptlrpc_request_set_replen(req);
1086 rc = ptlrpc_queue_wait(req);
1088 ptlrpc_req_finished(req);
1089 if (rc != -ETIMEDOUT)
1093 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1094 CERROR("%s: too many resend retries: rc = %d\n",
1095 exp->exp_obd->obd_name, -EIO);
1099 /* If a signal interrupts then the timeout returned will
1100 * not be zero. In that case return -EINTR
1102 if (msleep_interruptible(resends * 1000))
1108 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1109 req->rq_bulk->bd_nob_transferred);
1111 ptlrpc_req_finished(req);
1115 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1116 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1117 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1118 PAGE_SIZE * npages);
1119 ptlrpc_req_finished(req);
1127 static void mdc_release_page(struct page *page, int remove)
1131 if (likely(page->mapping != NULL))
1132 cfs_delete_from_page_cache(page);
1138 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1139 __u64 *start, __u64 *end, int hash64)
1142 * Complement of hash is used as an index so that
1143 * radix_tree_gang_lookup() can be used to find a page with starting
1144 * hash _smaller_ than one we are looking for.
1146 unsigned long offset = hash_x_index(*hash, hash64);
1148 unsigned long flags;
1151 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1152 found = radix_tree_gang_lookup(&mapping->page_tree,
1153 (void **)&page, offset, 1);
1154 if (found > 0 && !ll_xa_is_value(page)) {
1155 struct lu_dirpage *dp;
1158 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1160 * In contrast to find_lock_page() we are sure that directory
1161 * page cannot be truncated (while DLM lock is held) and,
1162 * hence, can avoid restart.
1164 * In fact, page cannot be locked here at all, because
1165 * mdc_read_page_remote does synchronous io.
1167 wait_on_page_locked(page);
1168 if (PageUptodate(page)) {
1170 if (BITS_PER_LONG == 32 && hash64) {
1171 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1172 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1173 *hash = *hash >> 32;
1175 *start = le64_to_cpu(dp->ldp_hash_start);
1176 *end = le64_to_cpu(dp->ldp_hash_end);
1178 if (unlikely(*start == 1 && *hash == 0))
1181 LASSERTF(*start <= *hash, "start = %#llx"
1182 ",end = %#llx,hash = %#llx\n",
1183 *start, *end, *hash);
1184 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1185 " hash %#llx\n", offset, *start, *end, *hash);
1188 mdc_release_page(page, 0);
1190 } else if (*end != *start && *hash == *end) {
1192 * upon hash collision, remove this page,
1193 * otherwise put page reference, and
1194 * mdc_read_page_remote() will issue RPC to
1195 * fetch the page we want.
1198 mdc_release_page(page,
1199 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1204 page = ERR_PTR(-EIO);
1207 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1214 * Adjust a set of pages, each page containing an array of lu_dirpages,
1215 * so that each page can be used as a single logical lu_dirpage.
1217 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1218 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1219 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1220 * value is used as a cookie to request the next lu_dirpage in a
1221 * directory listing that spans multiple pages (two in this example):
1224 * .|--------v------- -----.
1225 * |s|e|f|p|ent|ent| ... |ent|
1226 * '--|-------------- -----' Each PAGE contains a single
1227 * '------. lu_dirpage.
1228 * .---------v------- -----.
1229 * |s|e|f|p|ent| 0 | ... | 0 |
1230 * '----------------- -----'
1232 * However, on hosts where the native VM page size (PAGE_SIZE) is
1233 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1234 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1235 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1236 * after it in the same PAGE (arrows simplified for brevity, but
1237 * in general e0==s1, e1==s2, etc.):
1239 * .-------------------- -----.
1240 * |s0|e0|f0|p|ent|ent| ... |ent|
1241 * |---v---------------- -----|
1242 * |s1|e1|f1|p|ent|ent| ... |ent|
1243 * |---v---------------- -----| Here, each PAGE contains
1244 * ... multiple lu_dirpages.
1245 * |---v---------------- -----|
1246 * |s'|e'|f'|p|ent|ent| ... |ent|
1247 * '---|---------------- -----'
1249 * .----------------------------.
1252 * This structure is transformed into a single logical lu_dirpage as follows:
1254 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1255 * labeled 'next PAGE'.
1257 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1258 * a hash collision with the next page exists.
1260 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1261 * to the first entry of the next lu_dirpage.
1263 #if PAGE_SIZE > LU_PAGE_SIZE
1264 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1268 for (i = 0; i < cfs_pgs; i++) {
1269 struct lu_dirpage *dp = kmap(pages[i]);
1270 struct lu_dirpage *first = dp;
1271 struct lu_dirent *end_dirent = NULL;
1272 struct lu_dirent *ent;
1273 __u64 hash_end = dp->ldp_hash_end;
1274 __u32 flags = dp->ldp_flags;
1276 while (--lu_pgs > 0) {
1277 ent = lu_dirent_start(dp);
1278 for (end_dirent = ent; ent != NULL;
1279 end_dirent = ent, ent = lu_dirent_next(ent));
1281 /* Advance dp to next lu_dirpage. */
1282 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1284 /* Check if we've reached the end of the PAGE. */
1285 if (!((unsigned long)dp & ~PAGE_MASK))
1288 /* Save the hash and flags of this lu_dirpage. */
1289 hash_end = dp->ldp_hash_end;
1290 flags = dp->ldp_flags;
1292 /* Check if lu_dirpage contains no entries. */
1293 if (end_dirent == NULL)
1296 /* Enlarge the end entry lde_reclen from 0 to
1297 * first entry of next lu_dirpage. */
1298 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1299 end_dirent->lde_reclen =
1300 cpu_to_le16((char *)(dp->ldp_entries) -
1301 (char *)end_dirent);
1304 first->ldp_hash_end = hash_end;
1305 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1306 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1310 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1313 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1314 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1316 /* parameters for readdir page */
1317 struct readpage_param {
1318 struct md_op_data *rp_mod;
1321 struct obd_export *rp_exp;
1325 * Read pages from server.
1327 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1328 * a header lu_dirpage which describes the start/end hash, and whether this
1329 * page is empty (contains no dir entry) or hash collide with next page.
1330 * After client receives reply, several pages will be integrated into dir page
1331 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1332 * lu_dirpage for this integrated page will be adjusted.
1334 static int ll_mdc_read_page_remote(void *data, struct page *page0)
1336 struct readpage_param *rp = data;
1337 struct page **page_pool;
1339 struct lu_dirpage *dp;
1340 struct md_op_data *op_data = rp->rp_mod;
1341 struct ptlrpc_request *req;
1343 struct inode *inode;
1345 int rd_pgs = 0; /* number of pages actually read */
1351 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1352 inode = op_data->op_data;
1353 fid = &op_data->op_fid1;
1354 LASSERT(inode != NULL);
1356 OBD_ALLOC_PTR_ARRAY_LARGE(page_pool, max_pages);
1357 if (page_pool != NULL) {
1358 page_pool[0] = page0;
1364 for (npages = 1; npages < max_pages; npages++) {
1365 page = page_cache_alloc(inode->i_mapping);
1368 page_pool[npages] = page;
1371 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1373 /* page0 is special, which was added into page cache early */
1374 cfs_delete_from_page_cache(page0);
1378 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1380 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1381 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1383 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1385 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1387 SetPageUptodate(page0);
1391 ptlrpc_req_finished(req);
1392 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1393 for (i = 1; i < npages; i++) {
1394 unsigned long offset;
1398 page = page_pool[i];
1400 if (rc < 0 || i >= rd_pgs) {
1405 SetPageUptodate(page);
1408 hash = le64_to_cpu(dp->ldp_hash_start);
1411 offset = hash_x_index(hash, rp->rp_hash64);
1413 prefetchw(&page->flags);
1414 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1419 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1420 " rc = %d\n", offset, ret);
1424 if (page_pool != &page0)
1425 OBD_FREE_PTR_ARRAY_LARGE(page_pool, max_pages);
1430 #ifdef HAVE_READ_CACHE_PAGE_WANTS_FILE
1431 static inline int mdc_read_folio_remote(struct file *file, struct folio *folio)
1433 return ll_mdc_read_page_remote(file->private_data,
1434 folio_page(folio, 0));
1437 #define mdc_read_folio_remote ll_mdc_read_page_remote
1441 * Read dir page from cache first, if it can not find it, read it from
1442 * server and add into the cache.
1444 * \param[in] exp MDC export
1445 * \param[in] op_data client MD stack parameters, transfering parameters
1446 * between different layers on client MD stack.
1447 * \param[in] mrinfo callback required for ldlm lock enqueue during
1449 * \param[in] hash_offset the hash offset of the page to be read
1450 * \param[in] ppage the page to be read
1452 * retval = 0 get the page successfully
1453 * errno(<0) get the page failed
1455 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1456 struct md_readdir_info *mrinfo, __u64 hash_offset,
1457 struct page **ppage)
1459 struct lookup_intent it = { .it_op = IT_READDIR };
1461 struct inode *dir = op_data->op_data;
1462 struct address_space *mapping;
1463 struct lu_dirpage *dp;
1466 struct lustre_handle lockh;
1467 struct ptlrpc_request *enq_req = NULL;
1468 struct readpage_param rp_param;
1475 LASSERT(dir != NULL);
1476 mapping = dir->i_mapping;
1478 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1479 mrinfo->mr_blocking_ast, 0);
1480 if (enq_req != NULL)
1481 ptlrpc_req_finished(enq_req);
1484 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1485 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1490 lockh.cookie = it.it_lock_handle;
1491 mdc_set_lock_data(exp, &lockh, dir, NULL);
1493 rp_param.rp_off = hash_offset;
1494 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1495 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1496 rp_param.rp_hash64);
1498 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1499 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1500 rp_param.rp_off, PTR_ERR(page));
1501 GOTO(out_unlock, rc = PTR_ERR(page));
1502 } else if (page != NULL) {
1504 * XXX nikita: not entirely correct handling of a corner case:
1505 * suppose hash chain of entries with hash value HASH crosses
1506 * border between pages P0 and P1. First both P0 and P1 are
1507 * cached, seekdir() is called for some entry from the P0 part
1508 * of the chain. Later P0 goes out of cache. telldir(HASH)
1509 * happens and finds P1, as it starts with matching hash
1510 * value. Remaining entries from P0 part of the chain are
1511 * skipped. (Is that really a bug?)
1513 * Possible solutions: 0. don't cache P1 is such case, handle
1514 * it as an "overflow" page. 1. invalidate all pages at
1515 * once. 2. use HASH|1 as an index for P1.
1517 GOTO(hash_collision, page);
1520 rp_param.rp_exp = exp;
1521 rp_param.rp_mod = op_data;
1522 page = ll_read_cache_page(mapping,
1523 hash_x_index(rp_param.rp_off,
1524 rp_param.rp_hash64),
1525 mdc_read_folio_remote, &rp_param);
1527 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1528 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1529 rp_param.rp_off, PTR_ERR(page));
1530 GOTO(out_unlock, rc = PTR_ERR(page));
1533 wait_on_page_locked(page);
1535 if (!PageUptodate(page)) {
1536 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1537 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1538 rp_param.rp_off, -5);
1541 if (!PageChecked(page))
1542 SetPageChecked(page);
1543 if (PageError(page)) {
1544 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1545 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1546 rp_param.rp_off, -5);
1551 dp = page_address(page);
1552 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1553 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1554 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1555 rp_param.rp_off = hash_offset >> 32;
1557 start = le64_to_cpu(dp->ldp_hash_start);
1558 end = le64_to_cpu(dp->ldp_hash_end);
1559 rp_param.rp_off = hash_offset;
1562 LASSERT(start == rp_param.rp_off);
1563 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1564 #if BITS_PER_LONG == 32
1565 CWARN("Real page-wide hash collision at [%llu %llu] with "
1566 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1567 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1571 * Fetch whole overflow chain...
1579 ldlm_lock_decref(&lockh, it.it_lock_mode);
1583 mdc_release_page(page, 1);
1588 static int mdc_statfs_interpret(const struct lu_env *env,
1589 struct ptlrpc_request *req, void *args, int rc)
1591 struct obd_info *oinfo = args;
1592 struct obd_statfs *osfs;
1595 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1599 oinfo->oi_osfs = osfs;
1601 CDEBUG(D_CACHE, "blocks=%llu free=%llu avail=%llu "
1602 "objects=%llu free=%llu state=%x\n",
1603 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
1604 osfs->os_files, osfs->os_ffree, osfs->os_state);
1607 oinfo->oi_cb_up(oinfo, rc);
1612 static int mdc_statfs_async(struct obd_export *exp,
1613 struct obd_info *oinfo, time64_t max_age,
1614 struct ptlrpc_request_set *unused)
1616 struct ptlrpc_request *req;
1617 struct obd_info *aa;
1619 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_STATFS,
1620 LUSTRE_MDS_VERSION, MDS_STATFS);
1624 ptlrpc_request_set_replen(req);
1625 req->rq_interpret_reply = mdc_statfs_interpret;
1627 aa = ptlrpc_req_async_args(aa, req);
1630 ptlrpcd_add_req(req);
1635 static int mdc_statfs(const struct lu_env *env,
1636 struct obd_export *exp, struct obd_statfs *osfs,
1637 time64_t max_age, __u32 flags)
1639 struct obd_device *obd = class_exp2obd(exp);
1640 struct req_format *fmt;
1641 struct ptlrpc_request *req;
1642 struct obd_statfs *msfs;
1643 struct obd_import *imp, *imp0;
1648 * Since the request might also come from lprocfs, so we need
1649 * sync this with client_disconnect_export Bug15684
1651 with_imp_locked(obd, imp0, rc)
1652 imp = class_import_get(imp0);
1656 fmt = &RQF_MDS_STATFS;
1657 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1658 (flags & OBD_STATFS_SUM))
1659 fmt = &RQF_MDS_STATFS_NEW;
1660 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1663 GOTO(output, rc = -ENOMEM);
1664 req->rq_allow_intr = 1;
1666 if ((flags & OBD_STATFS_SUM) &&
1667 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1668 /* request aggregated states */
1669 struct mdt_body *body;
1671 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1673 GOTO(out, rc = -EPROTO);
1674 body->mbo_valid = OBD_MD_FLAGSTATFS;
1677 ptlrpc_request_set_replen(req);
1679 if (flags & OBD_STATFS_NODELAY) {
1680 /* procfs requests not want stay in wait for avoid deadlock */
1681 req->rq_no_resend = 1;
1682 req->rq_no_delay = 1;
1685 rc = ptlrpc_queue_wait(req);
1687 /* check connection error first */
1688 if (imp->imp_connect_error)
1689 rc = imp->imp_connect_error;
1693 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1695 GOTO(out, rc = -EPROTO);
1700 ptlrpc_req_finished(req);
1702 class_import_put(imp);
1706 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1708 __u32 keylen, vallen;
1712 if (gf->gf_pathlen < 2)
1715 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1716 keylen = round_up(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1717 sizeof(struct lu_fid), 8);
1718 OBD_ALLOC(key, keylen);
1721 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1722 memcpy(key + round_up(sizeof(KEY_FID2PATH), 8), gf, sizeof(*gf));
1723 memcpy(key + round_up(sizeof(KEY_FID2PATH), 8) + sizeof(*gf),
1724 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1725 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1726 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1728 if (!fid_is_sane(&gf->gf_fid))
1729 GOTO(out, rc = -EINVAL);
1731 /* Val is struct getinfo_fid2path result plus path */
1732 vallen = sizeof(*gf) + gf->gf_pathlen;
1734 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1735 if (rc != 0 && rc != -EREMOTE)
1738 if (vallen <= sizeof(*gf))
1739 GOTO(out, rc = -EPROTO);
1740 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1741 GOTO(out, rc = -EOVERFLOW);
1743 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %.*s\n",
1744 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1745 /* only log the first 512 characters of the path */
1746 512, gf->gf_u.gf_path);
1749 OBD_FREE(key, keylen);
1753 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1754 struct hsm_progress_kernel *hpk)
1756 struct obd_import *imp = class_exp2cliimp(exp);
1757 struct hsm_progress_kernel *req_hpk;
1758 struct ptlrpc_request *req;
1762 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1763 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1765 GOTO(out, rc = -ENOMEM);
1767 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1769 /* Copy hsm_progress struct */
1770 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1771 if (req_hpk == NULL)
1772 GOTO(out, rc = -EPROTO);
1775 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1777 ptlrpc_request_set_replen(req);
1779 ptlrpc_get_mod_rpc_slot(req);
1780 rc = ptlrpc_queue_wait(req);
1781 ptlrpc_put_mod_rpc_slot(req);
1785 ptlrpc_req_finished(req);
1789 * Send hsm_ct_register to MDS
1791 * \param[in] imp import
1792 * \param[in] archive_count if in bitmap format, it is the bitmap,
1793 * else it is the count of archive_ids
1794 * \param[in] archives if in bitmap format, it is NULL,
1795 * else it is archive_id lists
1797 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1800 struct ptlrpc_request *req;
1801 __u32 *archive_array;
1802 size_t archives_size;
1806 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1810 if (archives != NULL)
1811 archives_size = sizeof(*archive_array) * archive_count;
1813 archives_size = sizeof(archive_count);
1815 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1816 RCL_CLIENT, archives_size);
1818 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1820 ptlrpc_request_free(req);
1824 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1826 archive_array = req_capsule_client_get(&req->rq_pill,
1827 &RMF_MDS_HSM_ARCHIVE);
1828 if (archive_array == NULL)
1829 GOTO(out, rc = -EPROTO);
1831 if (archives != NULL)
1832 memcpy(archive_array, archives, archives_size);
1834 *archive_array = archive_count;
1836 ptlrpc_request_set_replen(req);
1837 req->rq_no_resend = 1;
1839 rc = mdc_queue_wait(req);
1842 ptlrpc_req_finished(req);
1846 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1847 struct md_op_data *op_data)
1849 struct hsm_current_action *hca = op_data->op_data;
1850 struct hsm_current_action *req_hca;
1851 struct ptlrpc_request *req;
1855 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1856 &RQF_MDS_HSM_ACTION);
1860 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1862 ptlrpc_request_free(req);
1866 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1867 op_data->op_suppgids[0], 0);
1869 ptlrpc_request_set_replen(req);
1871 rc = mdc_queue_wait(req);
1875 req_hca = req_capsule_server_get(&req->rq_pill,
1876 &RMF_MDS_HSM_CURRENT_ACTION);
1877 if (req_hca == NULL)
1878 GOTO(out, rc = -EPROTO);
1884 ptlrpc_req_finished(req);
1888 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1890 struct ptlrpc_request *req;
1894 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1896 MDS_HSM_CT_UNREGISTER);
1898 GOTO(out, rc = -ENOMEM);
1900 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1902 ptlrpc_request_set_replen(req);
1904 rc = mdc_queue_wait(req);
1907 ptlrpc_req_finished(req);
1911 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1912 struct md_op_data *op_data)
1914 struct hsm_user_state *hus = op_data->op_data;
1915 struct hsm_user_state *req_hus;
1916 struct ptlrpc_request *req;
1920 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1921 &RQF_MDS_HSM_STATE_GET);
1925 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1927 ptlrpc_request_free(req);
1931 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1932 op_data->op_suppgids[0], 0);
1934 ptlrpc_request_set_replen(req);
1936 rc = mdc_queue_wait(req);
1940 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1941 if (req_hus == NULL)
1942 GOTO(out, rc = -EPROTO);
1948 ptlrpc_req_finished(req);
1952 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1953 struct md_op_data *op_data)
1955 struct hsm_state_set *hss = op_data->op_data;
1956 struct hsm_state_set *req_hss;
1957 struct ptlrpc_request *req;
1961 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1962 &RQF_MDS_HSM_STATE_SET);
1966 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1968 ptlrpc_request_free(req);
1972 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1973 op_data->op_suppgids[0], 0);
1976 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1977 if (req_hss == NULL)
1978 GOTO(out, rc = -EPROTO);
1981 ptlrpc_request_set_replen(req);
1983 ptlrpc_get_mod_rpc_slot(req);
1984 rc = ptlrpc_queue_wait(req);
1985 ptlrpc_put_mod_rpc_slot(req);
1989 ptlrpc_req_finished(req);
1993 /* For RESTORE and RELEASE the mdt will take EX lock on the file layout.
1994 * So we can use early cancel on client side locks for that resource.
1996 static inline int mdc_hsm_request_lock_to_cancel(struct obd_export *exp,
1997 struct hsm_user_request *hur,
1998 struct list_head *cancels)
2000 struct hsm_user_item *hui = &hur->hur_user_item[0];
2001 struct hsm_request *req_hr = &hur->hur_request;
2005 if (req_hr->hr_action != HUA_RESTORE &&
2006 req_hr->hr_action != HUA_RELEASE)
2009 for (i = 0; i < req_hr->hr_itemcount; i++, hui++) {
2010 if (!fid_is_sane(&hui->hui_fid))
2012 count += mdc_resource_get_unused(exp, &hui->hui_fid, cancels,
2013 LCK_EX, MDS_INODELOCK_LAYOUT);
2019 static int mdc_ioc_hsm_request(struct obd_export *exp,
2020 struct hsm_user_request *hur)
2022 struct obd_import *imp = class_exp2cliimp(exp);
2023 struct ptlrpc_request *req;
2024 struct hsm_request *req_hr;
2025 struct hsm_user_item *req_hui;
2032 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
2036 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
2037 hur->hur_request.hr_itemcount
2038 * sizeof(struct hsm_user_item));
2039 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
2040 hur->hur_request.hr_data_len);
2042 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
2044 ptlrpc_request_free(req);
2048 /* Cancel existing locks */
2049 count = mdc_hsm_request_lock_to_cancel(exp, hur, &cancels);
2050 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
2051 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
2053 /* Copy hsm_request struct */
2054 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
2056 GOTO(out, rc = -EPROTO);
2057 *req_hr = hur->hur_request;
2059 /* Copy hsm_user_item structs */
2060 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
2061 if (req_hui == NULL)
2062 GOTO(out, rc = -EPROTO);
2063 memcpy(req_hui, hur->hur_user_item,
2064 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
2066 /* Copy opaque field */
2067 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
2068 if (req_opaque == NULL)
2069 GOTO(out, rc = -EPROTO);
2070 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
2072 ptlrpc_request_set_replen(req);
2074 ptlrpc_get_mod_rpc_slot(req);
2075 rc = ptlrpc_queue_wait(req);
2076 ptlrpc_put_mod_rpc_slot(req);
2081 ptlrpc_req_finished(req);
2085 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2086 struct lustre_kernelcomm *lk);
2088 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2089 struct obd_quotactl *oqctl)
2091 struct ptlrpc_request *req;
2092 struct obd_quotactl *oqc;
2096 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_QUOTACTL);
2101 if (LUSTRE_Q_CMD_IS_POOL(oqctl->qc_cmd))
2102 req_capsule_set_size(&req->rq_pill,
2105 sizeof(*oqc) + LOV_MAXPOOLNAME + 1);
2107 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION,
2110 ptlrpc_request_free(req);
2114 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2115 QCTL_COPY(oqc, oqctl);
2117 ptlrpc_request_set_replen(req);
2118 ptlrpc_at_set_req_timeout(req);
2120 rc = ptlrpc_queue_wait(req);
2122 CERROR("%s: ptlrpc_queue_wait failed: rc = %d\n",
2123 exp->exp_obd->obd_name, rc);
2127 if (req->rq_repmsg &&
2128 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2129 QCTL_COPY(oqctl, oqc);
2132 CERROR("%s: cannot unpack obd_quotactl: rc = %d\n",
2133 exp->exp_obd->obd_name, rc);
2136 ptlrpc_req_finished(req);
2141 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2142 struct md_op_data *op_data)
2145 struct ptlrpc_request *req;
2147 struct mdc_swap_layouts *msl, *payload;
2150 msl = op_data->op_data;
2152 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2153 * first thing it will do is to cancel the 2 layout
2154 * locks held by this client.
2155 * So the client must cancel its layout locks on the 2 fids
2156 * with the request RPC to avoid extra RPC round trips.
2158 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2159 LCK_EX, MDS_INODELOCK_LAYOUT |
2160 MDS_INODELOCK_XATTR);
2161 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2162 LCK_EX, MDS_INODELOCK_LAYOUT |
2163 MDS_INODELOCK_XATTR);
2165 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2166 &RQF_MDS_SWAP_LAYOUTS);
2168 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2172 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2174 ptlrpc_request_free(req);
2178 mdc_swap_layouts_pack(&req->rq_pill, op_data);
2180 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2185 ptlrpc_request_set_replen(req);
2187 rc = ptlrpc_queue_wait(req);
2193 ptlrpc_req_finished(req);
2197 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2198 void *karg, void __user *uarg)
2200 struct obd_device *obd = exp->exp_obd;
2201 struct obd_ioctl_data *data;
2202 struct obd_import *imp = obd->u.cli.cl_import;
2206 CDEBUG(D_IOCTL, "%s: cmd=%x len=%u karg=%pK uarg=%pK\n",
2207 obd->obd_name, cmd, len, karg, uarg);
2209 /* handle commands that do not need @karg first */
2211 case LL_IOC_GET_CONNECT_FLAGS:
2212 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2213 sizeof(*exp_connect_flags_ptr(exp))))
2218 if (unlikely(karg == NULL))
2219 RETURN(OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL", -EINVAL));
2222 if (!try_module_get(THIS_MODULE)) {
2223 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2224 module_name(THIS_MODULE));
2228 case OBD_IOC_FID2PATH:
2229 rc = mdc_ioc_fid2path(exp, karg);
2231 case LL_IOC_HSM_CT_START:
2232 rc = mdc_ioc_hsm_ct_start(exp, karg);
2233 /* ignore if it was already registered on this MDS. */
2237 case LL_IOC_HSM_PROGRESS:
2238 rc = mdc_ioc_hsm_progress(exp, karg);
2240 case LL_IOC_HSM_STATE_GET:
2241 rc = mdc_ioc_hsm_state_get(exp, karg);
2243 case LL_IOC_HSM_STATE_SET:
2244 rc = mdc_ioc_hsm_state_set(exp, karg);
2246 case LL_IOC_HSM_ACTION:
2247 rc = mdc_ioc_hsm_current_action(exp, karg);
2249 case LL_IOC_HSM_REQUEST:
2250 rc = mdc_ioc_hsm_request(exp, karg);
2252 case OBD_IOC_CLIENT_RECOVER:
2253 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2257 #ifdef IOC_OSC_SET_ACTIVE
2258 case_OBD_IOC_DEPRECATED_FT(IOC_OSC_SET_ACTIVE, obd->obd_name, 2, 17);
2260 case OBD_IOC_SET_ACTIVE:
2261 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2264 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2265 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2266 * there'd be no LMV layer thus we might be called here. Eventually
2267 * this code should be removed.
2270 case IOC_OBD_STATFS: {
2271 struct obd_statfs stat_buf = {0};
2273 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2274 GOTO(out, rc = -ENODEV);
2277 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2278 min((int)data->ioc_plen2,
2279 (int)sizeof(struct obd_uuid))))
2280 GOTO(out, rc = -EFAULT);
2282 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2283 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2288 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2289 min((int) data->ioc_plen1,
2290 (int) sizeof(stat_buf))))
2291 GOTO(out, rc = -EFAULT);
2295 case OBD_IOC_QUOTACTL: {
2296 struct if_quotactl *qctl = karg;
2297 struct obd_quotactl *oqctl;
2299 OBD_ALLOC_PTR(oqctl);
2301 GOTO(out, rc = -ENOMEM);
2303 QCTL_COPY(oqctl, qctl);
2304 rc = obd_quotactl(exp, oqctl);
2306 QCTL_COPY(qctl, oqctl);
2307 qctl->qc_valid = QC_MDTIDX;
2308 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2311 OBD_FREE_PTR(oqctl);
2314 case LL_IOC_LOV_SWAP_LAYOUTS:
2315 rc = mdc_ioc_swap_layouts(exp, karg);
2318 rc = OBD_IOC_ERROR(obd->obd_name, cmd, "unrecognized", -ENOTTY);
2322 module_put(THIS_MODULE);
2327 static int mdc_get_info_rpc(struct obd_export *exp,
2328 u32 keylen, void *key,
2329 u32 vallen, void *val)
2331 struct obd_import *imp = class_exp2cliimp(exp);
2332 struct ptlrpc_request *req;
2337 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2341 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2342 RCL_CLIENT, keylen);
2343 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2344 RCL_CLIENT, sizeof(vallen));
2346 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2348 ptlrpc_request_free(req);
2352 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2353 memcpy(tmp, key, keylen);
2354 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2355 memcpy(tmp, &vallen, sizeof(vallen));
2357 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2358 RCL_SERVER, vallen);
2359 ptlrpc_request_set_replen(req);
2361 /* if server failed to resolve FID, and OI scrub not able to fix it, it
2362 * will return -EINPROGRESS, ptlrpc_queue_wait() will keep retrying,
2363 * set request interruptible to avoid deadlock.
2365 if (KEY_IS(KEY_FID2PATH))
2366 req->rq_allow_intr = 1;
2368 rc = ptlrpc_queue_wait(req);
2369 /* -EREMOTE means the get_info result is partial, and it needs to
2370 * continue on another MDT, see fid2path part in lmv_iocontrol */
2371 if (rc == 0 || rc == -EREMOTE) {
2372 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2373 memcpy(val, tmp, vallen);
2374 if (req_capsule_rep_need_swab(&req->rq_pill)) {
2375 if (KEY_IS(KEY_FID2PATH))
2376 lustre_swab_fid2path(val);
2379 ptlrpc_req_finished(req);
2384 static void lustre_swab_hai(struct hsm_action_item *h)
2386 __swab32s(&h->hai_len);
2387 __swab32s(&h->hai_action);
2388 lustre_swab_lu_fid(&h->hai_fid);
2389 lustre_swab_lu_fid(&h->hai_dfid);
2390 __swab64s(&h->hai_cookie);
2391 __swab64s(&h->hai_extent.offset);
2392 __swab64s(&h->hai_extent.length);
2393 __swab64s(&h->hai_gid);
2396 static void lustre_swab_hal(struct hsm_action_list *h)
2398 struct hsm_action_item *hai;
2401 __swab32s(&h->hal_version);
2402 __swab32s(&h->hal_count);
2403 __swab32s(&h->hal_archive_id);
2404 __swab64s(&h->hal_flags);
2406 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2407 lustre_swab_hai(hai);
2410 static void lustre_swab_kuch(struct kuc_hdr *l)
2412 __swab16s(&l->kuc_magic);
2413 /* __u8 l->kuc_transport */
2414 __swab16s(&l->kuc_msgtype);
2415 __swab16s(&l->kuc_msglen);
2418 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2419 struct lustre_kernelcomm *lk)
2421 struct obd_import *imp = class_exp2cliimp(exp);
2424 if (lk->lk_group != KUC_GRP_HSM) {
2425 CERROR("Bad copytool group %d\n", lk->lk_group);
2429 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2430 lk->lk_uid, lk->lk_group, lk->lk_flags);
2432 if (lk->lk_flags & LK_FLG_STOP) {
2433 /* Unregister with the coordinator */
2434 rc = mdc_ioc_hsm_ct_unregister(imp);
2436 __u32 *archives = NULL;
2438 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2439 archives = lk->lk_data;
2441 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2448 * Send a message to any listening copytools
2449 * @param val KUC message (kuc_hdr + hsm_action_list)
2450 * @param len total length of message
2452 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2453 size_t len, void *val)
2455 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2456 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2460 if (len < sizeof(*lh) + sizeof(*hal)) {
2461 CERROR("Short HSM message %zu < %zu\n", len,
2462 sizeof(*lh) + sizeof(*hal));
2465 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2466 lustre_swab_kuch(lh);
2467 lustre_swab_hal(hal);
2468 } else if (lh->kuc_magic != KUC_MAGIC) {
2469 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2473 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2475 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2476 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2478 /* Broadcast to HSM listeners */
2479 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2485 * callback function passed to kuc for re-registering each HSM copytool
2486 * running on MDC, after MDT shutdown/recovery.
2487 * @param data copytool registration data
2488 * @param cb_arg callback argument (obd_import)
2490 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2492 struct obd_import *imp = (struct obd_import *)cb_arg;
2493 struct kkuc_ct_data *kcd = data;
2494 __u32 *archives = NULL;
2498 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2499 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2502 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2503 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2504 "(archive=%#x)\n", imp->imp_obd->obd_name,
2505 kcd->kcd_nr_archives);
2507 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2508 "(archive nr = %u)\n",
2509 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2510 if (kcd->kcd_nr_archives != 0)
2511 archives = kcd->kcd_archives;
2514 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2515 /* ignore error if the copytool is already registered */
2516 return (rc == -EEXIST) ? 0 : rc;
2520 * Re-establish all kuc contexts with MDT
2521 * after MDT shutdown/recovery.
2523 static int mdc_kuc_reregister(struct obd_import *imp)
2525 /* re-register HSM agents */
2526 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2527 mdc_hsm_ct_reregister, imp);
2530 static int mdc_set_info_async(const struct lu_env *env,
2531 struct obd_export *exp,
2532 u32 keylen, void *key,
2533 u32 vallen, void *val,
2534 struct ptlrpc_request_set *set)
2536 struct obd_import *imp = class_exp2cliimp(exp);
2540 if (KEY_IS(KEY_READ_ONLY)) {
2541 if (vallen != sizeof(int))
2544 spin_lock(&imp->imp_lock);
2545 if (*((int *)val)) {
2546 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2547 imp->imp_connect_data.ocd_connect_flags |=
2550 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2551 imp->imp_connect_data.ocd_connect_flags &=
2552 ~OBD_CONNECT_RDONLY;
2554 spin_unlock(&imp->imp_lock);
2556 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2557 keylen, key, vallen, val, set);
2560 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2561 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2562 keylen, key, vallen, val, set);
2565 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2566 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2571 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2572 __u32 *default_easize = val;
2574 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2578 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2582 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2583 __u32 keylen, void *key, __u32 *vallen, void *val)
2587 if (KEY_IS(KEY_MAX_EASIZE)) {
2588 __u32 mdsize, *max_easize;
2590 if (*vallen != sizeof(int))
2592 mdsize = *(__u32 *)val;
2593 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2594 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2596 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2598 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2599 __u32 *default_easize;
2601 if (*vallen != sizeof(int))
2603 default_easize = val;
2604 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2606 } else if (KEY_IS(KEY_CONN_DATA)) {
2607 struct obd_import *imp = class_exp2cliimp(exp);
2608 struct obd_connect_data *data = val;
2610 if (*vallen != sizeof(*data))
2613 *data = imp->imp_connect_data;
2615 } else if (KEY_IS(KEY_TGT_COUNT)) {
2616 *((__u32 *)val) = 1;
2620 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2625 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2626 struct ptlrpc_request **request)
2628 struct ptlrpc_request *req;
2633 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2637 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2639 ptlrpc_request_free(req);
2643 mdc_pack_body(&req->rq_pill, fid, 0, 0, -1, 0);
2645 ptlrpc_request_set_replen(req);
2647 rc = ptlrpc_queue_wait(req);
2649 ptlrpc_req_finished(req);
2655 struct mdc_rmfid_args {
2660 int mdc_rmfid_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2663 struct mdc_rmfid_args *aa;
2668 aa = ptlrpc_req_async_args(aa, req);
2670 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2672 LASSERT(size == sizeof(int) * aa->mra_nr);
2673 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2675 LASSERT(aa->mra_rcs);
2676 LASSERT(aa->mra_nr);
2677 memcpy(aa->mra_rcs, rcs, size);
2683 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2684 int *rcs, struct ptlrpc_request_set *set)
2686 struct ptlrpc_request *req;
2687 struct mdc_rmfid_args *aa;
2693 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2697 flen = fa->fa_nr * sizeof(struct lu_fid);
2698 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2700 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2702 req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2703 RCL_SERVER, fa->fa_nr * sizeof(__u32));
2704 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2706 ptlrpc_request_free(req);
2709 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2710 memcpy(tmp, fa->fa_fids, flen);
2712 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
2713 b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2714 b->mbo_ctime = ktime_get_real_seconds();
2716 ptlrpc_request_set_replen(req);
2719 aa = ptlrpc_req_async_args(aa, req);
2721 aa->mra_nr = fa->fa_nr;
2722 req->rq_interpret_reply = mdc_rmfid_interpret;
2724 ptlrpc_set_add_req(set, req);
2725 ptlrpc_check_set(NULL, set);
2730 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2731 enum obd_import_event event)
2733 struct client_obd *cli = &obd->u.cli;
2736 LASSERT(imp->imp_obd == obd);
2739 case IMP_EVENT_DISCON:
2740 spin_lock(&cli->cl_loi_list_lock);
2741 cli->cl_avail_grant = 0;
2742 cli->cl_lost_grant = 0;
2743 spin_unlock(&cli->cl_loi_list_lock);
2745 case IMP_EVENT_INACTIVE:
2747 * Flush current sequence to make client obtain new one
2748 * from server in case of disconnect/reconnect.
2750 down_read(&cli->cl_seq_rwsem);
2752 seq_client_flush(cli->cl_seq);
2753 up_read(&cli->cl_seq_rwsem);
2755 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2757 case IMP_EVENT_INVALIDATE: {
2758 struct ldlm_namespace *ns = obd->obd_namespace;
2762 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2764 env = cl_env_get(&refcheck);
2766 /* Reset grants. All pages go to failing rpcs due to
2767 * the invalid import.
2769 osc_io_unplug(env, cli, NULL);
2771 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2772 osc_ldlm_resource_invalidate,
2774 cl_env_put(env, &refcheck);
2775 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2781 case IMP_EVENT_ACTIVE:
2782 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2783 /* redo the kuc registration after reconnecting */
2785 rc = mdc_kuc_reregister(imp);
2787 case IMP_EVENT_OCD: {
2788 struct obd_connect_data *ocd = &imp->imp_connect_data;
2790 if (OCD_HAS_FLAG(ocd, GRANT))
2791 osc_init_grant(cli, ocd);
2793 md_init_ea_size(obd->obd_self_export, ocd->ocd_max_easize, 0);
2794 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2797 case IMP_EVENT_DEACTIVATE:
2798 case IMP_EVENT_ACTIVATE:
2801 CERROR("Unknown import event %x\n", event);
2807 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2808 struct lu_fid *fid, struct md_op_data *op_data)
2810 struct client_obd *cli = &exp->exp_obd->u.cli;
2815 down_read(&cli->cl_seq_rwsem);
2817 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2818 up_read(&cli->cl_seq_rwsem);
2823 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2825 struct client_obd *cli = &exp->exp_obd->u.cli;
2826 return &cli->cl_target_uuid;
2830 * Determine whether the lock can be canceled before replaying it during
2831 * recovery, non zero value will be return if the lock can be canceled,
2832 * or zero returned for not
2834 static int mdc_cancel_weight(struct ldlm_lock *lock)
2836 if (lock->l_resource->lr_type != LDLM_IBITS)
2839 /* FIXME: if we ever get into a situation where there are too many
2840 * opened files with open locks on a single node, then we really
2841 * should replay these open locks to reget it */
2842 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2845 /* Special case for DoM locks, cancel only unused and granted locks */
2846 if (ldlm_has_dom(lock) &&
2847 (lock->l_granted_mode != lock->l_req_mode ||
2848 osc_ldlm_weigh_ast(lock) != 0))
2854 static int mdc_resource_inode_free(struct ldlm_resource *res)
2856 if (res->lr_lvb_inode)
2857 res->lr_lvb_inode = NULL;
2862 static struct ldlm_valblock_ops inode_lvbo = {
2863 .lvbo_free = mdc_resource_inode_free
2866 static int mdc_llog_init(struct obd_device *obd)
2868 struct obd_llog_group *olg = &obd->obd_olg;
2869 struct llog_ctxt *ctxt;
2874 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2879 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2880 llog_initiator_connect(ctxt);
2881 llog_ctxt_put(ctxt);
2886 static void mdc_llog_finish(struct obd_device *obd)
2888 struct llog_ctxt *ctxt;
2892 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2894 llog_cleanup(NULL, ctxt);
2899 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2905 rc = osc_setup_common(obd, cfg);
2909 rc = mdc_tunables_init(obd);
2911 GOTO(err_osc_cleanup, rc);
2913 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2914 obd->u.cli.cl_lsom_update = true;
2916 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2918 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2920 rc = mdc_llog_init(obd);
2922 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2924 GOTO(err_llog_cleanup, rc);
2927 rc = mdc_changelog_cdev_init(obd);
2929 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2931 GOTO(err_changelog_cleanup, rc);
2936 err_changelog_cleanup:
2937 mdc_llog_finish(obd);
2939 lprocfs_free_md_stats(obd);
2940 ptlrpc_lprocfs_unregister_obd(obd);
2942 osc_cleanup_common(obd);
2946 /* Initialize the default and maximum LOV EA sizes. This allows
2947 * us to make MDS RPCs with large enough reply buffers to hold a default
2948 * sized EA without having to calculate this (via a call into the
2949 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2950 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2951 * a large number of stripes is possible. If a larger reply buffer is
2952 * required it will be reallocated in the ptlrpc layer due to overflow.
2954 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2957 struct obd_device *obd = exp->exp_obd;
2958 struct client_obd *cli = &obd->u.cli;
2961 if (cli->cl_max_mds_easize < easize)
2962 cli->cl_max_mds_easize = easize;
2964 if (cli->cl_default_mds_easize < def_easize)
2965 cli->cl_default_mds_easize = def_easize;
2970 static int mdc_precleanup(struct obd_device *obd)
2974 osc_precleanup_common(obd);
2976 mdc_changelog_cdev_finish(obd);
2977 mdc_llog_finish(obd);
2978 lprocfs_free_md_stats(obd);
2979 ptlrpc_lprocfs_unregister_obd(obd);
2984 static int mdc_cleanup(struct obd_device *obd)
2986 struct client_obd *cli = &obd->u.cli;
2987 LASSERT(cli->cl_mod_rpcs_in_flight == 0);
2988 return osc_cleanup_common(obd);
2991 static const struct obd_ops mdc_obd_ops = {
2992 .o_owner = THIS_MODULE,
2993 .o_setup = mdc_setup,
2994 .o_precleanup = mdc_precleanup,
2995 .o_cleanup = mdc_cleanup,
2996 .o_add_conn = client_import_add_conn,
2997 .o_del_conn = client_import_del_conn,
2998 .o_connect = client_connect_import,
2999 .o_reconnect = osc_reconnect,
3000 .o_disconnect = osc_disconnect,
3001 .o_iocontrol = mdc_iocontrol,
3002 .o_set_info_async = mdc_set_info_async,
3003 .o_statfs = mdc_statfs,
3004 .o_statfs_async = mdc_statfs_async,
3005 .o_fid_init = client_fid_init,
3006 .o_fid_fini = client_fid_fini,
3007 .o_fid_alloc = mdc_fid_alloc,
3008 .o_import_event = mdc_import_event,
3009 .o_get_info = mdc_get_info,
3010 .o_get_uuid = mdc_get_uuid,
3011 .o_quotactl = mdc_quotactl,
3014 static const struct md_ops mdc_md_ops = {
3015 .m_get_root = mdc_get_root,
3016 .m_null_inode = mdc_null_inode,
3017 .m_close = mdc_close,
3018 .m_create = mdc_create,
3019 .m_enqueue = mdc_enqueue,
3020 .m_getattr = mdc_getattr,
3021 .m_getattr_name = mdc_getattr_name,
3022 .m_intent_lock = mdc_intent_lock,
3024 .m_rename = mdc_rename,
3025 .m_setattr = mdc_setattr,
3026 .m_setxattr = mdc_setxattr,
3027 .m_getxattr = mdc_getxattr,
3028 .m_fsync = mdc_fsync,
3029 .m_file_resync = mdc_file_resync,
3030 .m_read_page = mdc_read_page,
3031 .m_unlink = mdc_unlink,
3032 .m_cancel_unused = mdc_cancel_unused,
3033 .m_init_ea_size = mdc_init_ea_size,
3034 .m_set_lock_data = mdc_set_lock_data,
3035 .m_lock_match = mdc_lock_match,
3036 .m_get_lustre_md = mdc_get_lustre_md,
3037 .m_free_lustre_md = mdc_free_lustre_md,
3038 .m_set_open_replay_data = mdc_set_open_replay_data,
3039 .m_clear_open_replay_data = mdc_clear_open_replay_data,
3040 .m_intent_getattr_async = mdc_intent_getattr_async,
3041 .m_revalidate_lock = mdc_revalidate_lock,
3042 .m_rmfid = mdc_rmfid,
3043 .m_batch_create = cli_batch_create,
3044 .m_batch_stop = cli_batch_stop,
3045 .m_batch_flush = cli_batch_flush,
3046 .m_batch_add = mdc_batch_add,
3049 dev_t mdc_changelog_dev;
3050 struct class *mdc_changelog_class;
3051 static int __init mdc_init(void)
3054 rc = alloc_chrdev_region(&mdc_changelog_dev, 0,
3055 MDC_CHANGELOG_DEV_COUNT,
3056 MDC_CHANGELOG_DEV_NAME);
3060 mdc_changelog_class = class_create(THIS_MODULE, MDC_CHANGELOG_DEV_NAME);
3061 if (IS_ERR(mdc_changelog_class)) {
3062 rc = PTR_ERR(mdc_changelog_class);
3066 rc = class_register_type(&mdc_obd_ops, &mdc_md_ops, true,
3067 LUSTRE_MDC_NAME, &mdc_device_type);
3074 class_destroy(mdc_changelog_class);
3076 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
3080 static void __exit mdc_exit(void)
3082 class_unregister_type(LUSTRE_MDC_NAME);
3083 class_destroy(mdc_changelog_class);
3084 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
3085 idr_destroy(&mdc_changelog_minor_idr);
3088 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3089 MODULE_DESCRIPTION("Lustre Metadata Client");
3090 MODULE_VERSION(LUSTRE_VERSION_STRING);
3091 MODULE_LICENSE("GPL");
3093 module_init(mdc_init);
3094 module_exit(mdc_exit);