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 = ll_capability_u32(current_cap());
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 (CFS_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 if (mod->mod_close_req)
1034 ptlrpc_request_addref(mod->mod_close_req);
1035 /* Since now, mod is accessed through open_req only,
1036 * thus close req does not keep a reference on mod anymore. */
1041 RETURN(rc < 0 ? rc : saved_rc);
1044 static int mdc_getpage(struct obd_export *exp, const struct lu_fid *fid,
1045 u64 offset, struct page **pages, int npages,
1046 struct ptlrpc_request **request)
1048 struct ptlrpc_request *req;
1049 struct ptlrpc_bulk_desc *desc;
1058 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_READPAGE);
1062 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_READPAGE);
1064 ptlrpc_request_free(req);
1068 req->rq_request_portal = MDS_READPAGE_PORTAL;
1069 ptlrpc_at_set_req_timeout(req);
1071 desc = ptlrpc_prep_bulk_imp(req, npages, 1,
1072 PTLRPC_BULK_PUT_SINK,
1074 &ptlrpc_bulk_kiov_pin_ops);
1076 ptlrpc_req_finished(req);
1080 /* NB req now owns desc and will free it when it gets freed */
1081 for (i = 0; i < npages; i++)
1082 desc->bd_frag_ops->add_kiov_frag(desc, pages[i], 0,
1085 mdc_readdir_pack(&req->rq_pill, offset, PAGE_SIZE * npages, fid);
1087 ptlrpc_request_set_replen(req);
1088 rc = ptlrpc_queue_wait(req);
1090 ptlrpc_req_finished(req);
1091 if (rc != -ETIMEDOUT)
1095 if (!client_should_resend(resends, &exp->exp_obd->u.cli)) {
1096 CERROR("%s: too many resend retries: rc = %d\n",
1097 exp->exp_obd->obd_name, -EIO);
1101 /* If a signal interrupts then the timeout returned will
1102 * not be zero. In that case return -EINTR
1104 if (msleep_interruptible(resends * 1000))
1110 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk,
1111 req->rq_bulk->bd_nob_transferred);
1113 ptlrpc_req_finished(req);
1117 if (req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK) {
1118 CERROR("%s: unexpected bytes transferred: %d (%ld expected)\n",
1119 exp->exp_obd->obd_name, req->rq_bulk->bd_nob_transferred,
1120 PAGE_SIZE * npages);
1121 ptlrpc_req_finished(req);
1129 static void mdc_release_page(struct page *page, int remove)
1133 if (likely(page->mapping != NULL))
1134 cfs_delete_from_page_cache(page);
1140 static struct page *mdc_page_locate(struct address_space *mapping, __u64 *hash,
1141 __u64 *start, __u64 *end, int hash64)
1144 * Complement of hash is used as an index so that
1145 * radix_tree_gang_lookup() can be used to find a page with starting
1146 * hash _smaller_ than one we are looking for.
1148 unsigned long offset = hash_x_index(*hash, hash64);
1150 unsigned long flags;
1153 ll_xa_lock_irqsave(&mapping->i_pages, flags);
1154 found = radix_tree_gang_lookup(&mapping->page_tree,
1155 (void **)&page, offset, 1);
1156 if (found > 0 && !ll_xa_is_value(page)) {
1157 struct lu_dirpage *dp;
1160 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1162 * In contrast to find_lock_page() we are sure that directory
1163 * page cannot be truncated (while DLM lock is held) and,
1164 * hence, can avoid restart.
1166 * In fact, page cannot be locked here at all, because
1167 * mdc_read_page_remote does synchronous io.
1169 wait_on_page_locked(page);
1170 if (PageUptodate(page)) {
1172 if (BITS_PER_LONG == 32 && hash64) {
1173 *start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1174 *end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1175 *hash = *hash >> 32;
1177 *start = le64_to_cpu(dp->ldp_hash_start);
1178 *end = le64_to_cpu(dp->ldp_hash_end);
1180 if (unlikely(*start == 1 && *hash == 0))
1183 LASSERTF(*start <= *hash, "start = %#llx"
1184 ",end = %#llx,hash = %#llx\n",
1185 *start, *end, *hash);
1186 CDEBUG(D_VFSTRACE, "offset %lx [%#llx %#llx],"
1187 " hash %#llx\n", offset, *start, *end, *hash);
1190 mdc_release_page(page, 0);
1192 } else if (*end != *start && *hash == *end) {
1194 * upon hash collision, remove this page,
1195 * otherwise put page reference, and
1196 * mdc_read_page_remote() will issue RPC to
1197 * fetch the page we want.
1200 mdc_release_page(page,
1201 le32_to_cpu(dp->ldp_flags) & LDF_COLLIDE);
1206 page = ERR_PTR(-EIO);
1209 ll_xa_unlock_irqrestore(&mapping->i_pages, flags);
1216 * Adjust a set of pages, each page containing an array of lu_dirpages,
1217 * so that each page can be used as a single logical lu_dirpage.
1219 * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
1220 * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
1221 * struct lu_dirent. It has size up to LU_PAGE_SIZE. The ldp_hash_end
1222 * value is used as a cookie to request the next lu_dirpage in a
1223 * directory listing that spans multiple pages (two in this example):
1226 * .|--------v------- -----.
1227 * |s|e|f|p|ent|ent| ... |ent|
1228 * '--|-------------- -----' Each PAGE contains a single
1229 * '------. lu_dirpage.
1230 * .---------v------- -----.
1231 * |s|e|f|p|ent| 0 | ... | 0 |
1232 * '----------------- -----'
1234 * However, on hosts where the native VM page size (PAGE_SIZE) is
1235 * larger than LU_PAGE_SIZE, a single host page may contain multiple
1236 * lu_dirpages. After reading the lu_dirpages from the MDS, the
1237 * ldp_hash_end of the first lu_dirpage refers to the one immediately
1238 * after it in the same PAGE (arrows simplified for brevity, but
1239 * in general e0==s1, e1==s2, etc.):
1241 * .-------------------- -----.
1242 * |s0|e0|f0|p|ent|ent| ... |ent|
1243 * |---v---------------- -----|
1244 * |s1|e1|f1|p|ent|ent| ... |ent|
1245 * |---v---------------- -----| Here, each PAGE contains
1246 * ... multiple lu_dirpages.
1247 * |---v---------------- -----|
1248 * |s'|e'|f'|p|ent|ent| ... |ent|
1249 * '---|---------------- -----'
1251 * .----------------------------.
1254 * This structure is transformed into a single logical lu_dirpage as follows:
1256 * - Replace e0 with e' so the request for the next lu_dirpage gets the page
1257 * labeled 'next PAGE'.
1259 * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
1260 * a hash collision with the next page exists.
1262 * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
1263 * to the first entry of the next lu_dirpage.
1265 #if PAGE_SIZE > LU_PAGE_SIZE
1266 static void mdc_adjust_dirpages(struct page **pages, int cfs_pgs, int lu_pgs)
1270 for (i = 0; i < cfs_pgs; i++) {
1271 struct lu_dirpage *dp = kmap(pages[i]);
1272 struct lu_dirpage *first = dp;
1273 struct lu_dirent *end_dirent = NULL;
1274 struct lu_dirent *ent;
1275 __u64 hash_end = dp->ldp_hash_end;
1276 __u32 flags = dp->ldp_flags;
1278 while (--lu_pgs > 0) {
1279 ent = lu_dirent_start(dp);
1280 for (end_dirent = ent; ent != NULL;
1281 end_dirent = ent, ent = lu_dirent_next(ent));
1283 /* Advance dp to next lu_dirpage. */
1284 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
1286 /* Check if we've reached the end of the PAGE. */
1287 if (!((unsigned long)dp & ~PAGE_MASK))
1290 /* Save the hash and flags of this lu_dirpage. */
1291 hash_end = dp->ldp_hash_end;
1292 flags = dp->ldp_flags;
1294 /* Check if lu_dirpage contains no entries. */
1295 if (end_dirent == NULL)
1298 /* Enlarge the end entry lde_reclen from 0 to
1299 * first entry of next lu_dirpage. */
1300 LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
1301 end_dirent->lde_reclen =
1302 cpu_to_le16((char *)(dp->ldp_entries) -
1303 (char *)end_dirent);
1306 first->ldp_hash_end = hash_end;
1307 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
1308 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
1312 LASSERTF(lu_pgs == 0, "left = %d\n", lu_pgs);
1315 #define mdc_adjust_dirpages(pages, cfs_pgs, lu_pgs) do {} while (0)
1316 #endif /* PAGE_SIZE > LU_PAGE_SIZE */
1318 /* parameters for readdir page */
1319 struct readpage_param {
1320 struct md_op_data *rp_mod;
1323 struct obd_export *rp_exp;
1327 * Read pages from server.
1329 * Page in MDS_READPAGE RPC is packed in LU_PAGE_SIZE, and each page contains
1330 * a header lu_dirpage which describes the start/end hash, and whether this
1331 * page is empty (contains no dir entry) or hash collide with next page.
1332 * After client receives reply, several pages will be integrated into dir page
1333 * in PAGE_SIZE (if PAGE_SIZE greater than LU_PAGE_SIZE), and the
1334 * lu_dirpage for this integrated page will be adjusted.
1336 static int ll_mdc_read_page_remote(void *data, struct page *page0)
1338 struct readpage_param *rp = data;
1339 struct page **page_pool;
1341 struct lu_dirpage *dp;
1342 struct md_op_data *op_data = rp->rp_mod;
1343 struct ptlrpc_request *req;
1345 struct inode *inode;
1347 int rd_pgs = 0; /* number of pages actually read */
1353 max_pages = rp->rp_exp->exp_obd->u.cli.cl_max_pages_per_rpc;
1354 inode = op_data->op_data;
1355 fid = &op_data->op_fid1;
1356 LASSERT(inode != NULL);
1358 OBD_ALLOC_PTR_ARRAY_LARGE(page_pool, max_pages);
1359 if (page_pool != NULL) {
1360 page_pool[0] = page0;
1366 for (npages = 1; npages < max_pages; npages++) {
1367 page = page_cache_alloc(inode->i_mapping);
1370 page_pool[npages] = page;
1373 rc = mdc_getpage(rp->rp_exp, fid, rp->rp_off, page_pool, npages, &req);
1375 /* page0 is special, which was added into page cache early */
1376 cfs_delete_from_page_cache(page0);
1380 rd_pgs = (req->rq_bulk->bd_nob_transferred + PAGE_SIZE - 1) >>
1382 lu_pgs = req->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
1383 LASSERT(!(req->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
1385 CDEBUG(D_INODE, "read %d(%d) pages\n", rd_pgs, lu_pgs);
1387 mdc_adjust_dirpages(page_pool, rd_pgs, lu_pgs);
1389 SetPageUptodate(page0);
1393 ptlrpc_req_finished(req);
1394 CDEBUG(D_CACHE, "read %d/%d pages\n", rd_pgs, npages);
1395 for (i = 1; i < npages; i++) {
1396 unsigned long offset;
1400 page = page_pool[i];
1402 if (rc < 0 || i >= rd_pgs) {
1407 SetPageUptodate(page);
1410 hash = le64_to_cpu(dp->ldp_hash_start);
1413 offset = hash_x_index(hash, rp->rp_hash64);
1415 prefetchw(&page->flags);
1416 ret = add_to_page_cache_lru(page, inode->i_mapping, offset,
1421 CDEBUG(D_VFSTRACE, "page %lu add to page cache failed:"
1422 " rc = %d\n", offset, ret);
1426 if (page_pool != &page0)
1427 OBD_FREE_PTR_ARRAY_LARGE(page_pool, max_pages);
1432 #ifdef HAVE_READ_CACHE_PAGE_WANTS_FILE
1433 static inline int mdc_read_folio_remote(struct file *file, struct folio *folio)
1435 return ll_mdc_read_page_remote(file->private_data,
1436 folio_page(folio, 0));
1439 #define mdc_read_folio_remote ll_mdc_read_page_remote
1443 * Read dir page from cache first, if it can not find it, read it from
1444 * server and add into the cache.
1446 * \param[in] exp MDC export
1447 * \param[in] op_data client MD stack parameters, transfering parameters
1448 * between different layers on client MD stack.
1449 * \param[in] mrinfo callback required for ldlm lock enqueue during
1451 * \param[in] hash_offset the hash offset of the page to be read
1452 * \param[in] ppage the page to be read
1454 * retval = 0 get the page successfully
1455 * errno(<0) get the page failed
1457 static int mdc_read_page(struct obd_export *exp, struct md_op_data *op_data,
1458 struct md_readdir_info *mrinfo, __u64 hash_offset,
1459 struct page **ppage)
1461 struct lookup_intent it = { .it_op = IT_READDIR };
1463 struct inode *dir = op_data->op_data;
1464 struct address_space *mapping;
1465 struct lu_dirpage *dp;
1468 struct lustre_handle lockh;
1469 struct ptlrpc_request *enq_req = NULL;
1470 struct readpage_param rp_param;
1477 LASSERT(dir != NULL);
1478 mapping = dir->i_mapping;
1480 rc = mdc_intent_lock(exp, op_data, &it, &enq_req,
1481 mrinfo->mr_blocking_ast, 0);
1482 if (enq_req != NULL)
1483 ptlrpc_req_finished(enq_req);
1486 CERROR("%s: "DFID" lock enqueue fails: rc = %d\n",
1487 exp->exp_obd->obd_name, PFID(&op_data->op_fid1), rc);
1492 lockh.cookie = it.it_lock_handle;
1493 mdc_set_lock_data(exp, &lockh, dir, NULL);
1495 rp_param.rp_off = hash_offset;
1496 rp_param.rp_hash64 = op_data->op_cli_flags & CLI_HASH64;
1497 page = mdc_page_locate(mapping, &rp_param.rp_off, &start, &end,
1498 rp_param.rp_hash64);
1500 CERROR("%s: dir page locate: "DFID" at %llu: rc %ld\n",
1501 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1502 rp_param.rp_off, PTR_ERR(page));
1503 GOTO(out_unlock, rc = PTR_ERR(page));
1504 } else if (page != NULL) {
1506 * XXX nikita: not entirely correct handling of a corner case:
1507 * suppose hash chain of entries with hash value HASH crosses
1508 * border between pages P0 and P1. First both P0 and P1 are
1509 * cached, seekdir() is called for some entry from the P0 part
1510 * of the chain. Later P0 goes out of cache. telldir(HASH)
1511 * happens and finds P1, as it starts with matching hash
1512 * value. Remaining entries from P0 part of the chain are
1513 * skipped. (Is that really a bug?)
1515 * Possible solutions: 0. don't cache P1 is such case, handle
1516 * it as an "overflow" page. 1. invalidate all pages at
1517 * once. 2. use HASH|1 as an index for P1.
1519 GOTO(hash_collision, page);
1522 rp_param.rp_exp = exp;
1523 rp_param.rp_mod = op_data;
1524 page = ll_read_cache_page(mapping,
1525 hash_x_index(rp_param.rp_off,
1526 rp_param.rp_hash64),
1527 mdc_read_folio_remote, &rp_param);
1529 CDEBUG(D_INFO, "%s: read cache page: "DFID" at %llu: %ld\n",
1530 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1531 rp_param.rp_off, PTR_ERR(page));
1532 GOTO(out_unlock, rc = PTR_ERR(page));
1535 wait_on_page_locked(page);
1537 if (!PageUptodate(page)) {
1538 CERROR("%s: page not updated: "DFID" at %llu: rc %d\n",
1539 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1540 rp_param.rp_off, -5);
1543 if (!PageChecked(page))
1544 SetPageChecked(page);
1545 if (PageError(page)) {
1546 CERROR("%s: page error: "DFID" at %llu: rc %d\n",
1547 exp->exp_obd->obd_name, PFID(&op_data->op_fid1),
1548 rp_param.rp_off, -5);
1553 dp = page_address(page);
1554 if (BITS_PER_LONG == 32 && rp_param.rp_hash64) {
1555 start = le64_to_cpu(dp->ldp_hash_start) >> 32;
1556 end = le64_to_cpu(dp->ldp_hash_end) >> 32;
1557 rp_param.rp_off = hash_offset >> 32;
1559 start = le64_to_cpu(dp->ldp_hash_start);
1560 end = le64_to_cpu(dp->ldp_hash_end);
1561 rp_param.rp_off = hash_offset;
1564 LASSERT(start == rp_param.rp_off);
1565 CWARN("Page-wide hash collision: %#lx\n", (unsigned long)end);
1566 #if BITS_PER_LONG == 32
1567 CWARN("Real page-wide hash collision at [%llu %llu] with "
1568 "hash %llu\n", le64_to_cpu(dp->ldp_hash_start),
1569 le64_to_cpu(dp->ldp_hash_end), hash_offset);
1573 * Fetch whole overflow chain...
1581 ldlm_lock_decref(&lockh, it.it_lock_mode);
1585 mdc_release_page(page, 1);
1590 static int mdc_statfs_interpret(const struct lu_env *env,
1591 struct ptlrpc_request *req, void *args, int rc)
1593 struct obd_info *oinfo = args;
1594 struct obd_statfs *osfs;
1597 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1601 oinfo->oi_osfs = osfs;
1603 CDEBUG(D_CACHE, "blocks=%llu free=%llu avail=%llu "
1604 "objects=%llu free=%llu state=%x\n",
1605 osfs->os_blocks, osfs->os_bfree, osfs->os_bavail,
1606 osfs->os_files, osfs->os_ffree, osfs->os_state);
1609 oinfo->oi_cb_up(oinfo, rc);
1614 static int mdc_statfs_async(struct obd_export *exp,
1615 struct obd_info *oinfo, time64_t max_age,
1616 struct ptlrpc_request_set *unused)
1618 struct ptlrpc_request *req;
1619 struct obd_info *aa;
1621 req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp), &RQF_MDS_STATFS,
1622 LUSTRE_MDS_VERSION, MDS_STATFS);
1626 ptlrpc_request_set_replen(req);
1627 req->rq_interpret_reply = mdc_statfs_interpret;
1629 aa = ptlrpc_req_async_args(aa, req);
1632 ptlrpcd_add_req(req);
1637 static int mdc_statfs(const struct lu_env *env,
1638 struct obd_export *exp, struct obd_statfs *osfs,
1639 time64_t max_age, __u32 flags)
1641 struct obd_device *obd = class_exp2obd(exp);
1642 struct req_format *fmt;
1643 struct ptlrpc_request *req;
1644 struct obd_statfs *msfs;
1645 struct obd_import *imp, *imp0;
1650 * Since the request might also come from lprocfs, so we need
1651 * sync this with client_disconnect_export Bug15684
1653 with_imp_locked(obd, imp0, rc)
1654 imp = class_import_get(imp0);
1658 fmt = &RQF_MDS_STATFS;
1659 if ((exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS) &&
1660 (flags & OBD_STATFS_SUM))
1661 fmt = &RQF_MDS_STATFS_NEW;
1662 req = ptlrpc_request_alloc_pack(imp, fmt, LUSTRE_MDS_VERSION,
1665 GOTO(output, rc = -ENOMEM);
1666 req->rq_allow_intr = 1;
1668 if ((flags & OBD_STATFS_SUM) &&
1669 (exp_connect_flags2(exp) & OBD_CONNECT2_SUM_STATFS)) {
1670 /* request aggregated states */
1671 struct mdt_body *body;
1673 body = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
1675 GOTO(out, rc = -EPROTO);
1676 body->mbo_valid = OBD_MD_FLAGSTATFS;
1679 ptlrpc_request_set_replen(req);
1681 if (flags & OBD_STATFS_NODELAY) {
1682 /* procfs requests not want stay in wait for avoid deadlock */
1683 req->rq_no_resend = 1;
1684 req->rq_no_delay = 1;
1687 rc = ptlrpc_queue_wait(req);
1689 /* check connection error first */
1690 if (imp->imp_connect_error)
1691 rc = imp->imp_connect_error;
1695 msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
1697 GOTO(out, rc = -EPROTO);
1702 ptlrpc_req_finished(req);
1704 class_import_put(imp);
1708 static int mdc_ioc_fid2path(struct obd_export *exp, struct getinfo_fid2path *gf)
1710 __u32 keylen, vallen;
1714 if (gf->gf_pathlen < 2)
1717 /* Key is KEY_FID2PATH + getinfo_fid2path description */
1718 keylen = round_up(sizeof(KEY_FID2PATH) + sizeof(*gf) +
1719 sizeof(struct lu_fid), 8);
1720 OBD_ALLOC(key, keylen);
1723 memcpy(key, KEY_FID2PATH, sizeof(KEY_FID2PATH));
1724 memcpy(key + round_up(sizeof(KEY_FID2PATH), 8), gf, sizeof(*gf));
1725 memcpy(key + round_up(sizeof(KEY_FID2PATH), 8) + sizeof(*gf),
1726 gf->gf_u.gf_root_fid, sizeof(struct lu_fid));
1727 CDEBUG(D_IOCTL, "path get "DFID" from %llu #%d\n",
1728 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno);
1730 if (!fid_is_sane(&gf->gf_fid))
1731 GOTO(out, rc = -EINVAL);
1733 /* Val is struct getinfo_fid2path result plus path */
1734 vallen = sizeof(*gf) + gf->gf_pathlen;
1736 rc = obd_get_info(NULL, exp, keylen, key, &vallen, gf);
1737 if (rc != 0 && rc != -EREMOTE)
1740 if (vallen <= sizeof(*gf))
1741 GOTO(out, rc = -EPROTO);
1742 if (vallen > sizeof(*gf) + gf->gf_pathlen)
1743 GOTO(out, rc = -EOVERFLOW);
1745 CDEBUG(D_IOCTL, "path got "DFID" from %llu #%d: %.*s\n",
1746 PFID(&gf->gf_fid), gf->gf_recno, gf->gf_linkno,
1747 /* only log the first 512 characters of the path */
1748 512, gf->gf_u.gf_path);
1751 OBD_FREE(key, keylen);
1755 static int mdc_ioc_hsm_progress(struct obd_export *exp,
1756 struct hsm_progress_kernel *hpk)
1758 struct obd_import *imp = class_exp2cliimp(exp);
1759 struct hsm_progress_kernel *req_hpk;
1760 struct ptlrpc_request *req;
1764 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_PROGRESS,
1765 LUSTRE_MDS_VERSION, MDS_HSM_PROGRESS);
1767 GOTO(out, rc = -ENOMEM);
1769 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1771 /* Copy hsm_progress struct */
1772 req_hpk = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_PROGRESS);
1773 if (req_hpk == NULL)
1774 GOTO(out, rc = -EPROTO);
1777 req_hpk->hpk_errval = lustre_errno_hton(hpk->hpk_errval);
1779 ptlrpc_request_set_replen(req);
1781 ptlrpc_get_mod_rpc_slot(req);
1782 rc = ptlrpc_queue_wait(req);
1783 ptlrpc_put_mod_rpc_slot(req);
1787 ptlrpc_req_finished(req);
1791 * Send hsm_ct_register to MDS
1793 * \param[in] imp import
1794 * \param[in] archive_count if in bitmap format, it is the bitmap,
1795 * else it is the count of archive_ids
1796 * \param[in] archives if in bitmap format, it is NULL,
1797 * else it is archive_id lists
1799 static int mdc_ioc_hsm_ct_register(struct obd_import *imp, __u32 archive_count,
1802 struct ptlrpc_request *req;
1803 __u32 *archive_array;
1804 size_t archives_size;
1808 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_CT_REGISTER);
1812 if (archives != NULL)
1813 archives_size = sizeof(*archive_array) * archive_count;
1815 archives_size = sizeof(archive_count);
1817 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_ARCHIVE,
1818 RCL_CLIENT, archives_size);
1820 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_CT_REGISTER);
1822 ptlrpc_request_free(req);
1826 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1828 archive_array = req_capsule_client_get(&req->rq_pill,
1829 &RMF_MDS_HSM_ARCHIVE);
1830 if (archive_array == NULL)
1831 GOTO(out, rc = -EPROTO);
1833 if (archives != NULL)
1834 memcpy(archive_array, archives, archives_size);
1836 *archive_array = archive_count;
1838 ptlrpc_request_set_replen(req);
1839 req->rq_no_resend = 1;
1841 rc = mdc_queue_wait(req);
1844 ptlrpc_req_finished(req);
1848 static int mdc_ioc_hsm_current_action(struct obd_export *exp,
1849 struct md_op_data *op_data)
1851 struct hsm_current_action *hca = op_data->op_data;
1852 struct hsm_current_action *req_hca;
1853 struct ptlrpc_request *req;
1857 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1858 &RQF_MDS_HSM_ACTION);
1862 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_ACTION);
1864 ptlrpc_request_free(req);
1868 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1869 op_data->op_suppgids[0], 0);
1871 ptlrpc_request_set_replen(req);
1873 rc = mdc_queue_wait(req);
1877 req_hca = req_capsule_server_get(&req->rq_pill,
1878 &RMF_MDS_HSM_CURRENT_ACTION);
1879 if (req_hca == NULL)
1880 GOTO(out, rc = -EPROTO);
1886 ptlrpc_req_finished(req);
1890 static int mdc_ioc_hsm_ct_unregister(struct obd_import *imp)
1892 struct ptlrpc_request *req;
1896 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_HSM_CT_UNREGISTER,
1898 MDS_HSM_CT_UNREGISTER);
1900 GOTO(out, rc = -ENOMEM);
1902 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
1904 ptlrpc_request_set_replen(req);
1906 rc = mdc_queue_wait(req);
1909 ptlrpc_req_finished(req);
1913 static int mdc_ioc_hsm_state_get(struct obd_export *exp,
1914 struct md_op_data *op_data)
1916 struct hsm_user_state *hus = op_data->op_data;
1917 struct hsm_user_state *req_hus;
1918 struct ptlrpc_request *req;
1922 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1923 &RQF_MDS_HSM_STATE_GET);
1927 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_GET);
1929 ptlrpc_request_free(req);
1933 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1934 op_data->op_suppgids[0], 0);
1936 ptlrpc_request_set_replen(req);
1938 rc = mdc_queue_wait(req);
1942 req_hus = req_capsule_server_get(&req->rq_pill, &RMF_HSM_USER_STATE);
1943 if (req_hus == NULL)
1944 GOTO(out, rc = -EPROTO);
1950 ptlrpc_req_finished(req);
1954 static int mdc_ioc_hsm_state_set(struct obd_export *exp,
1955 struct md_op_data *op_data)
1957 struct hsm_state_set *hss = op_data->op_data;
1958 struct hsm_state_set *req_hss;
1959 struct ptlrpc_request *req;
1963 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
1964 &RQF_MDS_HSM_STATE_SET);
1968 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_STATE_SET);
1970 ptlrpc_request_free(req);
1974 mdc_pack_body(&req->rq_pill, &op_data->op_fid1, 0, 0,
1975 op_data->op_suppgids[0], 0);
1978 req_hss = req_capsule_client_get(&req->rq_pill, &RMF_HSM_STATE_SET);
1979 if (req_hss == NULL)
1980 GOTO(out, rc = -EPROTO);
1983 ptlrpc_request_set_replen(req);
1985 ptlrpc_get_mod_rpc_slot(req);
1986 rc = ptlrpc_queue_wait(req);
1987 ptlrpc_put_mod_rpc_slot(req);
1991 ptlrpc_req_finished(req);
1995 /* For RESTORE and RELEASE the mdt will take EX lock on the file layout.
1996 * So we can use early cancel on client side locks for that resource.
1998 static inline int mdc_hsm_request_lock_to_cancel(struct obd_export *exp,
1999 struct hsm_user_request *hur,
2000 struct list_head *cancels)
2002 struct hsm_user_item *hui = &hur->hur_user_item[0];
2003 struct hsm_request *req_hr = &hur->hur_request;
2007 if (req_hr->hr_action != HUA_RESTORE &&
2008 req_hr->hr_action != HUA_RELEASE)
2011 for (i = 0; i < req_hr->hr_itemcount; i++, hui++) {
2012 if (!fid_is_sane(&hui->hui_fid))
2014 count += mdc_resource_get_unused(exp, &hui->hui_fid, cancels,
2015 LCK_EX, MDS_INODELOCK_LAYOUT);
2021 static int mdc_ioc_hsm_request(struct obd_export *exp,
2022 struct hsm_user_request *hur)
2024 struct obd_import *imp = class_exp2cliimp(exp);
2025 struct ptlrpc_request *req;
2026 struct hsm_request *req_hr;
2027 struct hsm_user_item *req_hui;
2034 req = ptlrpc_request_alloc(imp, &RQF_MDS_HSM_REQUEST);
2038 req_capsule_set_size(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM, RCL_CLIENT,
2039 hur->hur_request.hr_itemcount
2040 * sizeof(struct hsm_user_item));
2041 req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA, RCL_CLIENT,
2042 hur->hur_request.hr_data_len);
2044 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_HSM_REQUEST);
2046 ptlrpc_request_free(req);
2050 /* Cancel existing locks */
2051 count = mdc_hsm_request_lock_to_cancel(exp, hur, &cancels);
2052 ldlm_cli_cancel_list(&cancels, count, NULL, 0);
2053 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
2055 /* Copy hsm_request struct */
2056 req_hr = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_REQUEST);
2058 GOTO(out, rc = -EPROTO);
2059 *req_hr = hur->hur_request;
2061 /* Copy hsm_user_item structs */
2062 req_hui = req_capsule_client_get(&req->rq_pill, &RMF_MDS_HSM_USER_ITEM);
2063 if (req_hui == NULL)
2064 GOTO(out, rc = -EPROTO);
2065 memcpy(req_hui, hur->hur_user_item,
2066 hur->hur_request.hr_itemcount * sizeof(struct hsm_user_item));
2068 /* Copy opaque field */
2069 req_opaque = req_capsule_client_get(&req->rq_pill, &RMF_GENERIC_DATA);
2070 if (req_opaque == NULL)
2071 GOTO(out, rc = -EPROTO);
2072 memcpy(req_opaque, hur_data(hur), hur->hur_request.hr_data_len);
2074 ptlrpc_request_set_replen(req);
2076 ptlrpc_get_mod_rpc_slot(req);
2077 rc = ptlrpc_queue_wait(req);
2078 ptlrpc_put_mod_rpc_slot(req);
2083 ptlrpc_req_finished(req);
2087 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2088 struct lustre_kernelcomm *lk);
2090 static int mdc_quotactl(struct obd_device *unused, struct obd_export *exp,
2091 struct obd_quotactl *oqctl)
2093 struct ptlrpc_request *req;
2094 struct obd_quotactl *oqc;
2098 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_QUOTACTL);
2103 if (LUSTRE_Q_CMD_IS_POOL(oqctl->qc_cmd))
2104 req_capsule_set_size(&req->rq_pill,
2107 sizeof(*oqc) + LOV_MAXPOOLNAME + 1);
2109 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION,
2112 ptlrpc_request_free(req);
2116 oqc = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2117 QCTL_COPY(oqc, oqctl);
2119 ptlrpc_request_set_replen(req);
2120 ptlrpc_at_set_req_timeout(req);
2122 rc = ptlrpc_queue_wait(req);
2124 CERROR("%s: ptlrpc_queue_wait failed: rc = %d\n",
2125 exp->exp_obd->obd_name, rc);
2129 if (req->rq_repmsg &&
2130 (oqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL))) {
2131 QCTL_COPY(oqctl, oqc);
2134 CERROR("%s: cannot unpack obd_quotactl: rc = %d\n",
2135 exp->exp_obd->obd_name, rc);
2138 ptlrpc_req_finished(req);
2143 static int mdc_ioc_swap_layouts(struct obd_export *exp,
2144 struct md_op_data *op_data)
2147 struct ptlrpc_request *req;
2149 struct mdc_swap_layouts *msl, *payload;
2152 msl = op_data->op_data;
2154 /* When the MDT will get the MDS_SWAP_LAYOUTS RPC the
2155 * first thing it will do is to cancel the 2 layout
2156 * locks held by this client.
2157 * So the client must cancel its layout locks on the 2 fids
2158 * with the request RPC to avoid extra RPC round trips.
2160 count = mdc_resource_get_unused(exp, &op_data->op_fid1, &cancels,
2161 LCK_EX, MDS_INODELOCK_LAYOUT |
2162 MDS_INODELOCK_XATTR);
2163 count += mdc_resource_get_unused(exp, &op_data->op_fid2, &cancels,
2164 LCK_EX, MDS_INODELOCK_LAYOUT |
2165 MDS_INODELOCK_XATTR);
2167 req = ptlrpc_request_alloc(class_exp2cliimp(exp),
2168 &RQF_MDS_SWAP_LAYOUTS);
2170 ldlm_lock_list_put(&cancels, l_bl_ast, count);
2174 rc = mdc_prep_elc_req(exp, req, MDS_SWAP_LAYOUTS, &cancels, count);
2176 ptlrpc_request_free(req);
2180 mdc_swap_layouts_pack(&req->rq_pill, op_data);
2182 payload = req_capsule_client_get(&req->rq_pill, &RMF_SWAP_LAYOUTS);
2187 ptlrpc_request_set_replen(req);
2189 rc = ptlrpc_queue_wait(req);
2195 ptlrpc_req_finished(req);
2199 static int mdc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2200 void *karg, void __user *uarg)
2202 struct obd_device *obd = exp->exp_obd;
2203 struct obd_ioctl_data *data;
2204 struct obd_import *imp = obd->u.cli.cl_import;
2208 CDEBUG(D_IOCTL, "%s: cmd=%x len=%u karg=%pK uarg=%pK\n",
2209 obd->obd_name, cmd, len, karg, uarg);
2211 /* handle commands that do not need @karg first */
2213 case LL_IOC_GET_CONNECT_FLAGS:
2214 if (copy_to_user(uarg, exp_connect_flags_ptr(exp),
2215 sizeof(*exp_connect_flags_ptr(exp))))
2220 if (unlikely(karg == NULL))
2221 RETURN(OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL", -EINVAL));
2224 if (!try_module_get(THIS_MODULE)) {
2225 CERROR("%s: cannot get module '%s'\n", obd->obd_name,
2226 module_name(THIS_MODULE));
2230 case OBD_IOC_FID2PATH:
2231 rc = mdc_ioc_fid2path(exp, karg);
2233 case LL_IOC_HSM_CT_START:
2234 rc = mdc_ioc_hsm_ct_start(exp, karg);
2235 /* ignore if it was already registered on this MDS. */
2239 case LL_IOC_HSM_PROGRESS:
2240 rc = mdc_ioc_hsm_progress(exp, karg);
2242 case LL_IOC_HSM_STATE_GET:
2243 rc = mdc_ioc_hsm_state_get(exp, karg);
2245 case LL_IOC_HSM_STATE_SET:
2246 rc = mdc_ioc_hsm_state_set(exp, karg);
2248 case LL_IOC_HSM_ACTION:
2249 rc = mdc_ioc_hsm_current_action(exp, karg);
2251 case LL_IOC_HSM_REQUEST:
2252 rc = mdc_ioc_hsm_request(exp, karg);
2254 case OBD_IOC_CLIENT_RECOVER:
2255 rc = ptlrpc_recover_import(imp, data->ioc_inlbuf1, 0);
2259 #ifdef IOC_OSC_SET_ACTIVE
2260 case_OBD_IOC_DEPRECATED_FT(IOC_OSC_SET_ACTIVE, obd->obd_name, 2, 17);
2262 case OBD_IOC_SET_ACTIVE:
2263 rc = ptlrpc_set_import_active(imp, data->ioc_offset);
2266 * Normally IOC_OBD_STATFS, OBD_IOC_QUOTACTL iocontrol are handled by
2267 * LMV instead of MDC. But when the cluster is upgraded from 1.8,
2268 * there'd be no LMV layer thus we might be called here. Eventually
2269 * this code should be removed.
2272 case IOC_OBD_STATFS: {
2273 struct obd_statfs stat_buf = {0};
2275 if (*((__u32 *) data->ioc_inlbuf2) != 0)
2276 GOTO(out, rc = -ENODEV);
2279 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(obd),
2280 min((int)data->ioc_plen2,
2281 (int)sizeof(struct obd_uuid))))
2282 GOTO(out, rc = -EFAULT);
2284 rc = mdc_statfs(NULL, obd->obd_self_export, &stat_buf,
2285 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
2290 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
2291 min((int) data->ioc_plen1,
2292 (int) sizeof(stat_buf))))
2293 GOTO(out, rc = -EFAULT);
2297 case OBD_IOC_QUOTACTL: {
2298 struct if_quotactl *qctl = karg;
2299 struct obd_quotactl *oqctl;
2301 OBD_ALLOC_PTR(oqctl);
2303 GOTO(out, rc = -ENOMEM);
2305 QCTL_COPY(oqctl, qctl);
2306 rc = obd_quotactl(exp, oqctl);
2308 QCTL_COPY(qctl, oqctl);
2309 qctl->qc_valid = QC_MDTIDX;
2310 qctl->obd_uuid = obd->u.cli.cl_target_uuid;
2313 OBD_FREE_PTR(oqctl);
2316 case LL_IOC_LOV_SWAP_LAYOUTS:
2317 rc = mdc_ioc_swap_layouts(exp, karg);
2320 rc = OBD_IOC_ERROR(obd->obd_name, cmd, "unrecognized", -ENOTTY);
2324 module_put(THIS_MODULE);
2329 static int mdc_get_info_rpc(struct obd_export *exp,
2330 u32 keylen, void *key,
2331 u32 vallen, void *val)
2333 struct obd_import *imp = class_exp2cliimp(exp);
2334 struct ptlrpc_request *req;
2339 req = ptlrpc_request_alloc(imp, &RQF_MDS_GET_INFO);
2343 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_KEY,
2344 RCL_CLIENT, keylen);
2345 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VALLEN,
2346 RCL_CLIENT, sizeof(vallen));
2348 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_GET_INFO);
2350 ptlrpc_request_free(req);
2354 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_KEY);
2355 memcpy(tmp, key, keylen);
2356 tmp = req_capsule_client_get(&req->rq_pill, &RMF_GETINFO_VALLEN);
2357 memcpy(tmp, &vallen, sizeof(vallen));
2359 req_capsule_set_size(&req->rq_pill, &RMF_GETINFO_VAL,
2360 RCL_SERVER, vallen);
2361 ptlrpc_request_set_replen(req);
2363 /* if server failed to resolve FID, and OI scrub not able to fix it, it
2364 * will return -EINPROGRESS, ptlrpc_queue_wait() will keep retrying,
2365 * set request interruptible to avoid deadlock.
2367 if (KEY_IS(KEY_FID2PATH))
2368 req->rq_allow_intr = 1;
2370 rc = ptlrpc_queue_wait(req);
2371 /* -EREMOTE means the get_info result is partial, and it needs to
2372 * continue on another MDT, see fid2path part in lmv_iocontrol */
2373 if (rc == 0 || rc == -EREMOTE) {
2374 tmp = req_capsule_server_get(&req->rq_pill, &RMF_GETINFO_VAL);
2375 memcpy(val, tmp, vallen);
2376 if (req_capsule_rep_need_swab(&req->rq_pill)) {
2377 if (KEY_IS(KEY_FID2PATH))
2378 lustre_swab_fid2path(val);
2381 ptlrpc_req_finished(req);
2386 static void lustre_swab_hai(struct hsm_action_item *h)
2388 __swab32s(&h->hai_len);
2389 __swab32s(&h->hai_action);
2390 lustre_swab_lu_fid(&h->hai_fid);
2391 lustre_swab_lu_fid(&h->hai_dfid);
2392 __swab64s(&h->hai_cookie);
2393 __swab64s(&h->hai_extent.offset);
2394 __swab64s(&h->hai_extent.length);
2395 __swab64s(&h->hai_gid);
2398 static void lustre_swab_hal(struct hsm_action_list *h)
2400 struct hsm_action_item *hai;
2403 __swab32s(&h->hal_version);
2404 __swab32s(&h->hal_count);
2405 __swab32s(&h->hal_archive_id);
2406 __swab64s(&h->hal_flags);
2408 for (i = 0; i < h->hal_count; i++, hai = hai_next(hai))
2409 lustre_swab_hai(hai);
2412 static void lustre_swab_kuch(struct kuc_hdr *l)
2414 __swab16s(&l->kuc_magic);
2415 /* __u8 l->kuc_transport */
2416 __swab16s(&l->kuc_msgtype);
2417 __swab16s(&l->kuc_msglen);
2420 static int mdc_ioc_hsm_ct_start(struct obd_export *exp,
2421 struct lustre_kernelcomm *lk)
2423 struct obd_import *imp = class_exp2cliimp(exp);
2426 if (lk->lk_group != KUC_GRP_HSM) {
2427 CERROR("Bad copytool group %d\n", lk->lk_group);
2431 CDEBUG(D_HSM, "CT start r%d w%d u%d g%d f%#x\n", lk->lk_rfd, lk->lk_wfd,
2432 lk->lk_uid, lk->lk_group, lk->lk_flags);
2434 if (lk->lk_flags & LK_FLG_STOP) {
2435 /* Unregister with the coordinator */
2436 rc = mdc_ioc_hsm_ct_unregister(imp);
2438 __u32 *archives = NULL;
2440 if ((lk->lk_flags & LK_FLG_DATANR) && lk->lk_data_count > 0)
2441 archives = lk->lk_data;
2443 rc = mdc_ioc_hsm_ct_register(imp, lk->lk_data_count, archives);
2450 * Send a message to any listening copytools
2451 * @param val KUC message (kuc_hdr + hsm_action_list)
2452 * @param len total length of message
2454 static int mdc_hsm_copytool_send(const struct obd_uuid *uuid,
2455 size_t len, void *val)
2457 struct kuc_hdr *lh = (struct kuc_hdr *)val;
2458 struct hsm_action_list *hal = (struct hsm_action_list *)(lh + 1);
2462 if (len < sizeof(*lh) + sizeof(*hal)) {
2463 CERROR("Short HSM message %zu < %zu\n", len,
2464 sizeof(*lh) + sizeof(*hal));
2467 if (lh->kuc_magic == __swab16(KUC_MAGIC)) {
2468 lustre_swab_kuch(lh);
2469 lustre_swab_hal(hal);
2470 } else if (lh->kuc_magic != KUC_MAGIC) {
2471 CERROR("Bad magic %x!=%x\n", lh->kuc_magic, KUC_MAGIC);
2475 CDEBUG(D_HSM, " Received message mg=%x t=%d m=%d l=%d actions=%d "
2477 lh->kuc_magic, lh->kuc_transport, lh->kuc_msgtype,
2478 lh->kuc_msglen, hal->hal_count, hal->hal_fsname);
2480 /* Broadcast to HSM listeners */
2481 rc = libcfs_kkuc_group_put(uuid, KUC_GRP_HSM, lh);
2487 * callback function passed to kuc for re-registering each HSM copytool
2488 * running on MDC, after MDT shutdown/recovery.
2489 * @param data copytool registration data
2490 * @param cb_arg callback argument (obd_import)
2492 static int mdc_hsm_ct_reregister(void *data, void *cb_arg)
2494 struct obd_import *imp = (struct obd_import *)cb_arg;
2495 struct kkuc_ct_data *kcd = data;
2496 __u32 *archives = NULL;
2500 (kcd->kcd_magic != KKUC_CT_DATA_ARRAY_MAGIC &&
2501 kcd->kcd_magic != KKUC_CT_DATA_BITMAP_MAGIC))
2504 if (kcd->kcd_magic == KKUC_CT_DATA_BITMAP_MAGIC) {
2505 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2506 "(archive=%#x)\n", imp->imp_obd->obd_name,
2507 kcd->kcd_nr_archives);
2509 CDEBUG(D_HA, "%s: recover copytool registration to MDT "
2510 "(archive nr = %u)\n",
2511 imp->imp_obd->obd_name, kcd->kcd_nr_archives);
2512 if (kcd->kcd_nr_archives != 0)
2513 archives = kcd->kcd_archives;
2516 rc = mdc_ioc_hsm_ct_register(imp, kcd->kcd_nr_archives, archives);
2517 /* ignore error if the copytool is already registered */
2518 return (rc == -EEXIST) ? 0 : rc;
2522 * Re-establish all kuc contexts with MDT
2523 * after MDT shutdown/recovery.
2525 static int mdc_kuc_reregister(struct obd_import *imp)
2527 /* re-register HSM agents */
2528 return libcfs_kkuc_group_foreach(&imp->imp_obd->obd_uuid, KUC_GRP_HSM,
2529 mdc_hsm_ct_reregister, imp);
2532 static int mdc_set_info_async(const struct lu_env *env,
2533 struct obd_export *exp,
2534 u32 keylen, void *key,
2535 u32 vallen, void *val,
2536 struct ptlrpc_request_set *set)
2538 struct obd_import *imp = class_exp2cliimp(exp);
2542 if (KEY_IS(KEY_READ_ONLY)) {
2543 if (vallen != sizeof(int))
2546 spin_lock(&imp->imp_lock);
2547 if (*((int *)val)) {
2548 imp->imp_connect_flags_orig |= OBD_CONNECT_RDONLY;
2549 imp->imp_connect_data.ocd_connect_flags |=
2552 imp->imp_connect_flags_orig &= ~OBD_CONNECT_RDONLY;
2553 imp->imp_connect_data.ocd_connect_flags &=
2554 ~OBD_CONNECT_RDONLY;
2556 spin_unlock(&imp->imp_lock);
2558 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2559 keylen, key, vallen, val, set);
2562 if (KEY_IS(KEY_CHANGELOG_CLEAR)) {
2563 rc = do_set_info_async(imp, MDS_SET_INFO, LUSTRE_MDS_VERSION,
2564 keylen, key, vallen, val, set);
2567 if (KEY_IS(KEY_HSM_COPYTOOL_SEND)) {
2568 rc = mdc_hsm_copytool_send(&imp->imp_obd->obd_uuid, vallen,
2573 if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2574 __u32 *default_easize = val;
2576 exp->exp_obd->u.cli.cl_default_mds_easize = *default_easize;
2580 rc = osc_set_info_async(env, exp, keylen, key, vallen, val, set);
2584 static int mdc_get_info(const struct lu_env *env, struct obd_export *exp,
2585 __u32 keylen, void *key, __u32 *vallen, void *val)
2589 if (KEY_IS(KEY_MAX_EASIZE)) {
2590 __u32 mdsize, *max_easize;
2592 if (*vallen != sizeof(int))
2594 mdsize = *(__u32 *)val;
2595 if (mdsize > exp->exp_obd->u.cli.cl_max_mds_easize)
2596 exp->exp_obd->u.cli.cl_max_mds_easize = mdsize;
2598 *max_easize = exp->exp_obd->u.cli.cl_max_mds_easize;
2600 } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
2601 __u32 *default_easize;
2603 if (*vallen != sizeof(int))
2605 default_easize = val;
2606 *default_easize = exp->exp_obd->u.cli.cl_default_mds_easize;
2608 } else if (KEY_IS(KEY_CONN_DATA)) {
2609 struct obd_import *imp = class_exp2cliimp(exp);
2610 struct obd_connect_data *data = val;
2612 if (*vallen != sizeof(*data))
2615 *data = imp->imp_connect_data;
2617 } else if (KEY_IS(KEY_TGT_COUNT)) {
2618 *((__u32 *)val) = 1;
2622 rc = mdc_get_info_rpc(exp, keylen, key, *vallen, val);
2627 static int mdc_fsync(struct obd_export *exp, const struct lu_fid *fid,
2628 struct ptlrpc_request **request)
2630 struct ptlrpc_request *req;
2635 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_SYNC);
2639 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_SYNC);
2641 ptlrpc_request_free(req);
2645 mdc_pack_body(&req->rq_pill, fid, 0, 0, -1, 0);
2647 ptlrpc_request_set_replen(req);
2649 rc = ptlrpc_queue_wait(req);
2651 ptlrpc_req_finished(req);
2657 struct mdc_rmfid_args {
2662 static int mdc_rmfid_interpret(const struct lu_env *env,
2663 struct ptlrpc_request *req,
2666 struct mdc_rmfid_args *aa;
2671 aa = ptlrpc_req_async_args(aa, req);
2673 size = req_capsule_get_size(&req->rq_pill, &RMF_RCS,
2675 LASSERT(size == sizeof(int) * aa->mra_nr);
2676 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
2678 LASSERT(aa->mra_rcs);
2679 LASSERT(aa->mra_nr);
2680 memcpy(aa->mra_rcs, rcs, size);
2686 static int mdc_rmfid(struct obd_export *exp, struct fid_array *fa,
2687 int *rcs, struct ptlrpc_request_set *set)
2689 struct ptlrpc_request *req;
2690 struct mdc_rmfid_args *aa;
2696 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_MDS_RMFID);
2700 flen = fa->fa_nr * sizeof(struct lu_fid);
2701 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2703 req_capsule_set_size(&req->rq_pill, &RMF_FID_ARRAY,
2705 req_capsule_set_size(&req->rq_pill, &RMF_RCS,
2706 RCL_SERVER, fa->fa_nr * sizeof(__u32));
2707 rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_RMFID);
2709 ptlrpc_request_free(req);
2712 tmp = req_capsule_client_get(&req->rq_pill, &RMF_FID_ARRAY);
2713 memcpy(tmp, fa->fa_fids, flen);
2715 mdc_pack_body(&req->rq_pill, NULL, 0, 0, -1, 0);
2716 b = req_capsule_client_get(&req->rq_pill, &RMF_MDT_BODY);
2717 b->mbo_ctime = ktime_get_real_seconds();
2719 ptlrpc_request_set_replen(req);
2722 aa = ptlrpc_req_async_args(aa, req);
2724 aa->mra_nr = fa->fa_nr;
2725 req->rq_interpret_reply = mdc_rmfid_interpret;
2727 ptlrpc_set_add_req(set, req);
2728 ptlrpc_check_set(NULL, set);
2733 static int mdc_import_event(struct obd_device *obd, struct obd_import *imp,
2734 enum obd_import_event event)
2736 struct client_obd *cli = &obd->u.cli;
2739 LASSERT(imp->imp_obd == obd);
2742 case IMP_EVENT_DISCON:
2743 spin_lock(&cli->cl_loi_list_lock);
2744 cli->cl_avail_grant = 0;
2745 cli->cl_lost_grant = 0;
2746 spin_unlock(&cli->cl_loi_list_lock);
2748 case IMP_EVENT_INACTIVE:
2750 * Flush current sequence to make client obtain new one
2751 * from server in case of disconnect/reconnect.
2753 down_read(&cli->cl_seq_rwsem);
2755 seq_client_flush(cli->cl_seq);
2756 up_read(&cli->cl_seq_rwsem);
2758 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_INACTIVE);
2760 case IMP_EVENT_INVALIDATE: {
2761 struct ldlm_namespace *ns = obd->obd_namespace;
2765 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2767 env = cl_env_get(&refcheck);
2769 /* Reset grants. All pages go to failing rpcs due to
2770 * the invalid import.
2772 osc_io_unplug(env, cli, NULL);
2774 cfs_hash_for_each_nolock(ns->ns_rs_hash,
2775 osc_ldlm_resource_invalidate,
2777 cl_env_put(env, &refcheck);
2778 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2784 case IMP_EVENT_ACTIVE:
2785 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_ACTIVE);
2786 /* redo the kuc registration after reconnecting */
2788 rc = mdc_kuc_reregister(imp);
2790 case IMP_EVENT_OCD: {
2791 struct obd_connect_data *ocd = &imp->imp_connect_data;
2793 if (OCD_HAS_FLAG(ocd, GRANT))
2794 osc_init_grant(cli, ocd);
2796 md_init_ea_size(obd->obd_self_export, ocd->ocd_max_easize, 0);
2797 rc = obd_notify_observer(obd, obd, OBD_NOTIFY_OCD);
2800 case IMP_EVENT_DEACTIVATE:
2801 case IMP_EVENT_ACTIVATE:
2804 CERROR("Unknown import event %x\n", event);
2810 int mdc_fid_alloc(const struct lu_env *env, struct obd_export *exp,
2811 struct lu_fid *fid, struct md_op_data *op_data)
2813 struct client_obd *cli = &exp->exp_obd->u.cli;
2818 down_read(&cli->cl_seq_rwsem);
2820 rc = seq_client_alloc_fid(env, cli->cl_seq, fid);
2821 up_read(&cli->cl_seq_rwsem);
2826 static struct obd_uuid *mdc_get_uuid(struct obd_export *exp)
2828 struct client_obd *cli = &exp->exp_obd->u.cli;
2829 return &cli->cl_target_uuid;
2833 * Determine whether the lock can be canceled before replaying it during
2834 * recovery, non zero value will be return if the lock can be canceled,
2835 * or zero returned for not
2837 static int mdc_cancel_weight(struct ldlm_lock *lock)
2839 if (lock->l_resource->lr_type != LDLM_IBITS)
2842 /* FIXME: if we ever get into a situation where there are too many
2843 * opened files with open locks on a single node, then we really
2844 * should replay these open locks to reget it */
2845 if (lock->l_policy_data.l_inodebits.bits & MDS_INODELOCK_OPEN)
2848 /* Special case for DoM locks, cancel only unused and granted locks */
2849 if (ldlm_has_dom(lock) &&
2850 (lock->l_granted_mode != lock->l_req_mode ||
2851 osc_ldlm_weigh_ast(lock) != 0))
2857 static int mdc_resource_inode_free(struct ldlm_resource *res)
2859 if (res->lr_lvb_inode)
2860 res->lr_lvb_inode = NULL;
2865 static struct ldlm_valblock_ops inode_lvbo = {
2866 .lvbo_free = mdc_resource_inode_free
2869 static int mdc_llog_init(struct obd_device *obd)
2871 struct obd_llog_group *olg = &obd->obd_olg;
2872 struct llog_ctxt *ctxt;
2877 rc = llog_setup(NULL, obd, olg, LLOG_CHANGELOG_REPL_CTXT, obd,
2882 ctxt = llog_group_get_ctxt(olg, LLOG_CHANGELOG_REPL_CTXT);
2883 llog_initiator_connect(ctxt);
2884 llog_ctxt_put(ctxt);
2889 static void mdc_llog_finish(struct obd_device *obd)
2891 struct llog_ctxt *ctxt;
2895 ctxt = llog_get_context(obd, LLOG_CHANGELOG_REPL_CTXT);
2897 llog_cleanup(NULL, ctxt);
2902 int mdc_setup(struct obd_device *obd, struct lustre_cfg *cfg)
2908 rc = osc_setup_common(obd, cfg);
2912 rc = mdc_tunables_init(obd);
2914 GOTO(err_osc_cleanup, rc);
2916 obd->u.cli.cl_dom_min_inline_repsize = MDC_DOM_DEF_INLINE_REPSIZE;
2917 obd->u.cli.cl_lsom_update = true;
2919 ns_register_cancel(obd->obd_namespace, mdc_cancel_weight);
2921 obd->obd_namespace->ns_lvbo = &inode_lvbo;
2923 rc = mdc_llog_init(obd);
2925 CERROR("%s: failed to setup llogging subsystems: rc = %d\n",
2927 GOTO(err_llog_cleanup, rc);
2930 rc = mdc_changelog_cdev_init(obd);
2932 CERROR("%s: failed to setup changelog char device: rc = %d\n",
2934 GOTO(err_changelog_cleanup, rc);
2939 err_changelog_cleanup:
2940 mdc_llog_finish(obd);
2942 lprocfs_free_md_stats(obd);
2943 ptlrpc_lprocfs_unregister_obd(obd);
2945 osc_cleanup_common(obd);
2949 /* Initialize the default and maximum LOV EA sizes. This allows
2950 * us to make MDS RPCs with large enough reply buffers to hold a default
2951 * sized EA without having to calculate this (via a call into the
2952 * LOV + OSCs) each time we make an RPC. The maximum size is also tracked
2953 * but not used to avoid wastefully vmalloc()'ing large reply buffers when
2954 * a large number of stripes is possible. If a larger reply buffer is
2955 * required it will be reallocated in the ptlrpc layer due to overflow.
2957 static int mdc_init_ea_size(struct obd_export *exp, __u32 easize,
2960 struct obd_device *obd = exp->exp_obd;
2961 struct client_obd *cli = &obd->u.cli;
2964 if (cli->cl_max_mds_easize < easize)
2965 cli->cl_max_mds_easize = easize;
2967 if (cli->cl_default_mds_easize < def_easize)
2968 cli->cl_default_mds_easize = def_easize;
2973 static int mdc_precleanup(struct obd_device *obd)
2977 osc_precleanup_common(obd);
2979 mdc_changelog_cdev_finish(obd);
2980 mdc_llog_finish(obd);
2981 lprocfs_free_md_stats(obd);
2982 ptlrpc_lprocfs_unregister_obd(obd);
2987 static int mdc_cleanup(struct obd_device *obd)
2989 struct client_obd *cli = &obd->u.cli;
2990 LASSERT(cli->cl_mod_rpcs_in_flight == 0);
2991 return osc_cleanup_common(obd);
2994 static const struct obd_ops mdc_obd_ops = {
2995 .o_owner = THIS_MODULE,
2996 .o_setup = mdc_setup,
2997 .o_precleanup = mdc_precleanup,
2998 .o_cleanup = mdc_cleanup,
2999 .o_add_conn = client_import_add_conn,
3000 .o_del_conn = client_import_del_conn,
3001 .o_connect = client_connect_import,
3002 .o_reconnect = osc_reconnect,
3003 .o_disconnect = osc_disconnect,
3004 .o_iocontrol = mdc_iocontrol,
3005 .o_set_info_async = mdc_set_info_async,
3006 .o_statfs = mdc_statfs,
3007 .o_statfs_async = mdc_statfs_async,
3008 .o_fid_init = client_fid_init,
3009 .o_fid_fini = client_fid_fini,
3010 .o_fid_alloc = mdc_fid_alloc,
3011 .o_import_event = mdc_import_event,
3012 .o_get_info = mdc_get_info,
3013 .o_get_uuid = mdc_get_uuid,
3014 .o_quotactl = mdc_quotactl,
3017 static const struct md_ops mdc_md_ops = {
3018 .m_get_root = mdc_get_root,
3019 .m_null_inode = mdc_null_inode,
3020 .m_close = mdc_close,
3021 .m_create = mdc_create,
3022 .m_enqueue = mdc_enqueue,
3023 .m_getattr = mdc_getattr,
3024 .m_getattr_name = mdc_getattr_name,
3025 .m_intent_lock = mdc_intent_lock,
3027 .m_rename = mdc_rename,
3028 .m_setattr = mdc_setattr,
3029 .m_setxattr = mdc_setxattr,
3030 .m_getxattr = mdc_getxattr,
3031 .m_fsync = mdc_fsync,
3032 .m_file_resync = mdc_file_resync,
3033 .m_read_page = mdc_read_page,
3034 .m_unlink = mdc_unlink,
3035 .m_cancel_unused = mdc_cancel_unused,
3036 .m_init_ea_size = mdc_init_ea_size,
3037 .m_set_lock_data = mdc_set_lock_data,
3038 .m_lock_match = mdc_lock_match,
3039 .m_get_lustre_md = mdc_get_lustre_md,
3040 .m_free_lustre_md = mdc_free_lustre_md,
3041 .m_set_open_replay_data = mdc_set_open_replay_data,
3042 .m_clear_open_replay_data = mdc_clear_open_replay_data,
3043 .m_intent_getattr_async = mdc_intent_getattr_async,
3044 .m_revalidate_lock = mdc_revalidate_lock,
3045 .m_rmfid = mdc_rmfid,
3046 .m_batch_create = cli_batch_create,
3047 .m_batch_stop = cli_batch_stop,
3048 .m_batch_flush = cli_batch_flush,
3049 .m_batch_add = mdc_batch_add,
3052 dev_t mdc_changelog_dev;
3053 struct class *mdc_changelog_class;
3054 static int __init mdc_init(void)
3057 rc = alloc_chrdev_region(&mdc_changelog_dev, 0,
3058 MDC_CHANGELOG_DEV_COUNT,
3059 MDC_CHANGELOG_DEV_NAME);
3063 mdc_changelog_class = class_create(THIS_MODULE, MDC_CHANGELOG_DEV_NAME);
3064 if (IS_ERR(mdc_changelog_class)) {
3065 rc = PTR_ERR(mdc_changelog_class);
3069 rc = class_register_type(&mdc_obd_ops, &mdc_md_ops, true,
3070 LUSTRE_MDC_NAME, &mdc_device_type);
3077 class_destroy(mdc_changelog_class);
3079 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
3083 static void __exit mdc_exit(void)
3085 class_unregister_type(LUSTRE_MDC_NAME);
3086 class_destroy(mdc_changelog_class);
3087 unregister_chrdev_region(mdc_changelog_dev, MDC_CHANGELOG_DEV_COUNT);
3088 idr_destroy(&mdc_changelog_minor_idr);
3091 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3092 MODULE_DESCRIPTION("Lustre Metadata Client");
3093 MODULE_VERSION(LUSTRE_VERSION_STRING);
3094 MODULE_LICENSE("GPL");
3096 module_init(mdc_init);
3097 module_exit(mdc_exit);