4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_LMV
38 #include <linux/slab.h>
39 #include <linux/module.h>
40 #include <linux/init.h>
41 #include <linux/slab.h>
42 #include <linux/pagemap.h>
43 #include <linux/math64.h>
44 #include <linux/seq_file.h>
45 #include <linux/namei.h>
46 #include <linux/lustre_intent.h>
48 #include <obd_support.h>
49 #include <lustre/lustre_idl.h>
50 #include <lustre_lib.h>
51 #include <lustre_net.h>
52 #include <lustre_dlm.h>
53 #include <lustre_mdc.h>
54 #include <obd_class.h>
55 #include <lprocfs_status.h>
56 #include "lmv_internal.h"
58 static int lmv_intent_remote(struct obd_export *exp, struct lookup_intent *it,
59 const struct lu_fid *parent_fid,
60 struct ptlrpc_request **reqp,
61 ldlm_blocking_callback cb_blocking,
62 __u64 extra_lock_flags)
64 struct obd_device *obd = exp->exp_obd;
65 struct lmv_obd *lmv = &obd->u.lmv;
66 struct ptlrpc_request *req = NULL;
67 struct lustre_handle plock;
68 struct md_op_data *op_data;
69 struct lmv_tgt_desc *tgt;
70 struct mdt_body *body;
75 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
79 LASSERT((body->mbo_valid & OBD_MD_MDS));
82 * Unfortunately, we have to lie to MDC/MDS to retrieve
83 * attributes llite needs and provideproper locking.
85 if (it->it_op & IT_LOOKUP)
86 it->it_op = IT_GETATTR;
89 * We got LOOKUP lock, but we really need attrs.
91 pmode = it->d.lustre.it_lock_mode;
93 plock.cookie = it->d.lustre.it_lock_handle;
94 it->d.lustre.it_lock_mode = 0;
95 it->d.lustre.it_data = NULL;
98 LASSERT(fid_is_sane(&body->mbo_fid1));
100 tgt = lmv_find_target(lmv, &body->mbo_fid1);
102 GOTO(out, rc = PTR_ERR(tgt));
104 OBD_ALLOC_PTR(op_data);
106 GOTO(out, rc = -ENOMEM);
108 op_data->op_fid1 = body->mbo_fid1;
109 /* Sent the parent FID to the remote MDT */
110 if (parent_fid != NULL) {
111 /* The parent fid is only for remote open to
112 * check whether the open is from OBF,
113 * see mdt_cross_open */
114 LASSERT(it->it_op & IT_OPEN);
115 op_data->op_fid2 = *parent_fid;
118 op_data->op_bias = MDS_CROSS_REF;
119 CDEBUG(D_INODE, "REMOTE_INTENT with fid="DFID" -> mds #%d\n",
120 PFID(&body->mbo_fid1), tgt->ltd_idx);
122 rc = md_intent_lock(tgt->ltd_exp, op_data, it, &req, cb_blocking,
125 GOTO(out_free_op_data, rc);
128 * LLite needs LOOKUP lock to track dentry revocation in order to
129 * maintain dcache consistency. Thus drop UPDATE|PERM lock here
130 * and put LOOKUP in request.
132 if (it->d.lustre.it_lock_mode != 0) {
133 it->d.lustre.it_remote_lock_handle =
134 it->d.lustre.it_lock_handle;
135 it->d.lustre.it_remote_lock_mode = it->d.lustre.it_lock_mode;
139 it->d.lustre.it_lock_handle = plock.cookie;
140 it->d.lustre.it_lock_mode = pmode;
145 OBD_FREE_PTR(op_data);
148 ldlm_lock_decref(&plock, pmode);
150 ptlrpc_req_finished(*reqp);
155 int lmv_revalidate_slaves(struct obd_export *exp, struct mdt_body *mbody,
156 struct lmv_stripe_md *lsm,
157 ldlm_blocking_callback cb_blocking,
158 int extra_lock_flags)
160 struct obd_device *obd = exp->exp_obd;
161 struct lmv_obd *lmv = &obd->u.lmv;
162 struct ptlrpc_request *req = NULL;
163 struct mdt_body *body;
164 struct md_op_data *op_data;
165 unsigned long size = 0;
166 unsigned long nlink = 0;
176 * revalidate slaves has some problems, temporarily return,
177 * we may not need that
179 OBD_ALLOC_PTR(op_data);
184 * Loop over the stripe information, check validity and update them
185 * from MDS if needed.
187 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
189 struct lookup_intent it = { .it_op = IT_GETATTR };
190 struct lustre_handle *lockh = NULL;
191 struct lmv_tgt_desc *tgt = NULL;
194 fid = lsm->lsm_md_oinfo[i].lmo_fid;
195 inode = lsm->lsm_md_oinfo[i].lmo_root;
198 * Prepare op_data for revalidating. Note that @fid2 shluld be
199 * defined otherwise it will go to server and take new lock
200 * which is not needed here.
202 memset(op_data, 0, sizeof(*op_data));
203 op_data->op_fid1 = fid;
204 op_data->op_fid2 = fid;
206 tgt = lmv_locate_mds(lmv, op_data, &fid);
208 GOTO(cleanup, rc = PTR_ERR(tgt));
210 CDEBUG(D_INODE, "Revalidate slave "DFID" -> mds #%d\n",
211 PFID(&fid), tgt->ltd_idx);
214 ptlrpc_req_finished(req);
218 rc = md_intent_lock(tgt->ltd_exp, op_data, &it, &req,
219 cb_blocking, extra_lock_flags);
223 lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
224 if (rc > 0 && req == NULL) {
225 /* slave inode is still valid */
226 CDEBUG(D_INODE, "slave "DFID" is still valid.\n",
230 /* refresh slave from server */
231 body = req_capsule_server_get(&req->rq_pill,
233 LASSERT(body != NULL);
234 if (unlikely(body->mbo_nlink < 2)) {
235 CERROR("%s: nlink %d < 2 corrupt stripe %d "DFID
237 obd->obd_name, body->mbo_nlink, i,
238 PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
239 PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
241 if (it.d.lustre.it_lock_mode && lockh) {
242 ldlm_lock_decref(lockh,
243 it.d.lustre.it_lock_mode);
244 it.d.lustre.it_lock_mode = 0;
247 GOTO(cleanup, rc = -EIO);
251 i_size_write(inode, body->mbo_size);
252 set_nlink(inode, body->mbo_nlink);
253 LTIME_S(inode->i_atime) = body->mbo_atime;
254 LTIME_S(inode->i_ctime) = body->mbo_ctime;
255 LTIME_S(inode->i_mtime) = body->mbo_mtime;
258 md_set_lock_data(tgt->ltd_exp, &lockh->cookie, inode, NULL);
260 size += i_size_read(inode);
263 nlink += inode->i_nlink - 2;
265 nlink += inode->i_nlink;
267 atime = LTIME_S(inode->i_atime) > atime ?
268 LTIME_S(inode->i_atime) : atime;
269 ctime = LTIME_S(inode->i_ctime) > ctime ?
270 LTIME_S(inode->i_ctime) : ctime;
271 mtime = LTIME_S(inode->i_mtime) > mtime ?
272 LTIME_S(inode->i_mtime) : mtime;
274 if (it.d.lustre.it_lock_mode != 0 && lockh != NULL) {
275 ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
276 it.d.lustre.it_lock_mode = 0;
279 CDEBUG(D_INODE, "i %d "DFID" size %llu, nlink %u, atime "
280 "%lu, mtime %lu, ctime %lu.\n", i, PFID(&fid),
281 i_size_read(inode), inode->i_nlink,
282 LTIME_S(inode->i_atime), LTIME_S(inode->i_mtime),
283 LTIME_S(inode->i_ctime));
287 * update attr of master request.
289 CDEBUG(D_INODE, "Return refreshed attrs: size = %lu nlink %lu atime "
290 LPU64 "ctime "LPU64" mtime "LPU64" for " DFID"\n", size, nlink,
291 atime, ctime, mtime, PFID(&lsm->lsm_md_oinfo[0].lmo_fid));
294 mbody->mbo_atime = atime;
295 mbody->mbo_ctime = ctime;
296 mbody->mbo_mtime = mtime;
300 ptlrpc_req_finished(req);
302 OBD_FREE_PTR(op_data);
308 * IT_OPEN is intended to open (and create, possible) an object. Parent (pid)
311 int lmv_intent_open(struct obd_export *exp, struct md_op_data *op_data,
312 struct lookup_intent *it, struct ptlrpc_request **reqp,
313 ldlm_blocking_callback cb_blocking, __u64 extra_lock_flags)
315 struct obd_device *obd = exp->exp_obd;
316 struct lmv_obd *lmv = &obd->u.lmv;
317 struct lmv_tgt_desc *tgt;
318 struct mdt_body *body;
322 if (it->it_flags & MDS_OPEN_BY_FID) {
323 LASSERT(fid_is_sane(&op_data->op_fid2));
325 /* for striped directory, we can't know parent stripe fid
326 * without name, but we can set it to child fid, and MDT
327 * will obtain it from linkea in open in such case. */
328 if (op_data->op_mea1 != NULL)
329 op_data->op_fid1 = op_data->op_fid2;
331 tgt = lmv_find_target(lmv, &op_data->op_fid2);
333 RETURN(PTR_ERR(tgt));
335 op_data->op_mds = tgt->ltd_idx;
337 LASSERT(fid_is_sane(&op_data->op_fid1));
338 LASSERT(fid_is_zero(&op_data->op_fid2));
339 LASSERT(op_data->op_name != NULL);
341 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
343 RETURN(PTR_ERR(tgt));
346 /* If it is ready to open the file by FID, do not need
347 * allocate FID at all, otherwise it will confuse MDT */
348 if ((it->it_op & IT_CREAT) && !(it->it_flags & MDS_OPEN_BY_FID)) {
350 * For lookup(IT_CREATE) cases allocate new fid and setup FLD
353 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
358 CDEBUG(D_INODE, "OPEN_INTENT with fid1="DFID", fid2="DFID","
359 " name='%s' -> mds #%d\n", PFID(&op_data->op_fid1),
360 PFID(&op_data->op_fid2), op_data->op_name, tgt->ltd_idx);
362 rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
367 * Nothing is found, do not access body->fid1 as it is zero and thus
370 if ((it->d.lustre.it_disposition & DISP_LOOKUP_NEG) &&
371 !(it->d.lustre.it_disposition & DISP_OPEN_CREATE) &&
372 !(it->d.lustre.it_disposition & DISP_OPEN_OPEN))
375 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
379 /* Not cross-ref case, just get out of here. */
380 if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
381 rc = lmv_intent_remote(exp, it, &op_data->op_fid1, reqp,
382 cb_blocking, extra_lock_flags);
386 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
395 * Handler for: getattr, lookup and revalidate cases.
398 lmv_intent_lookup(struct obd_export *exp, struct md_op_data *op_data,
399 struct lookup_intent *it, struct ptlrpc_request **reqp,
400 ldlm_blocking_callback cb_blocking,
401 __u64 extra_lock_flags)
403 struct obd_device *obd = exp->exp_obd;
404 struct lmv_obd *lmv = &obd->u.lmv;
405 struct lmv_tgt_desc *tgt = NULL;
406 struct mdt_body *body;
407 struct lmv_stripe_md *lsm = op_data->op_mea1;
411 /* If it returns ERR_PTR(-EBADFD) then it is an unknown hash type
412 * it will try all stripes to locate the object */
413 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
414 if (IS_ERR(tgt) && (PTR_ERR(tgt) != -EBADFD))
415 RETURN(PTR_ERR(tgt));
417 /* Both migrating dir and unknown hash dir need to try
418 * all of sub-stripes */
419 if (lsm != NULL && !lmv_is_known_hash_type(lsm)) {
420 struct lmv_oinfo *oinfo;
422 oinfo = &lsm->lsm_md_oinfo[0];
424 op_data->op_fid1 = oinfo->lmo_fid;
425 op_data->op_mds = oinfo->lmo_mds;
426 tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
428 RETURN(PTR_ERR(tgt));
431 if (!fid_is_sane(&op_data->op_fid2))
432 fid_zero(&op_data->op_fid2);
434 CDEBUG(D_INODE, "LOOKUP_INTENT with fid1="DFID", fid2="DFID
435 ", name='%s' -> mds #%d lsm=%p lsm_magic=%x\n",
436 PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
437 op_data->op_name ? op_data->op_name : "<NULL>",
438 tgt->ltd_idx, lsm, lsm == NULL ? -1 : lsm->lsm_md_magic);
440 op_data->op_bias &= ~MDS_CROSS_REF;
442 rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp, cb_blocking,
448 /* If RPC happens, lsm information will be revalidated
449 * during update_inode process (see ll_update_lsm_md) */
450 if (op_data->op_mea2 != NULL) {
451 rc = lmv_revalidate_slaves(exp, NULL, op_data->op_mea2,
458 } else if (it_disposition(it, DISP_LOOKUP_NEG) && lsm != NULL &&
459 lmv_need_try_all_stripes(lsm)) {
460 /* For migrating and unknown hash type directory, it will
461 * try to target the entry on other stripes */
464 for (stripe_index = 1;
465 stripe_index < lsm->lsm_md_stripe_count &&
466 it_disposition(it, DISP_LOOKUP_NEG); stripe_index++) {
467 struct lmv_oinfo *oinfo;
469 /* release the previous request */
470 ptlrpc_req_finished(*reqp);
471 it->d.lustre.it_data = NULL;
474 oinfo = &lsm->lsm_md_oinfo[stripe_index];
475 tgt = lmv_find_target(lmv, &oinfo->lmo_fid);
477 RETURN(PTR_ERR(tgt));
479 CDEBUG(D_INODE, "Try other stripes " DFID"\n",
480 PFID(&oinfo->lmo_fid));
482 op_data->op_fid1 = oinfo->lmo_fid;
483 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
484 rc = md_intent_lock(tgt->ltd_exp, op_data, it, reqp,
485 cb_blocking, extra_lock_flags);
492 * MDS has returned success. Probably name has been resolved in
493 * remote inode. Let's check this.
495 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
499 /* Not cross-ref case, just get out of here. */
500 if (unlikely((body->mbo_valid & OBD_MD_MDS))) {
501 rc = lmv_intent_remote(exp, it, NULL, reqp, cb_blocking,
505 body = req_capsule_server_get(&(*reqp)->rq_pill, &RMF_MDT_BODY);
513 int lmv_intent_lock(struct obd_export *exp, struct md_op_data *op_data,
514 struct lookup_intent *it, struct ptlrpc_request **reqp,
515 ldlm_blocking_callback cb_blocking,
516 __u64 extra_lock_flags)
518 struct obd_device *obd = exp->exp_obd;
523 LASSERT(fid_is_sane(&op_data->op_fid1));
525 CDEBUG(D_INODE, "INTENT LOCK '%s' for "DFID" '%.*s' on "DFID"\n",
526 LL_IT2STR(it), PFID(&op_data->op_fid2), op_data->op_namelen,
527 op_data->op_name, PFID(&op_data->op_fid1));
529 rc = lmv_check_connect(obd);
533 if (it->it_op & (IT_LOOKUP | IT_GETATTR | IT_LAYOUT))
534 rc = lmv_intent_lookup(exp, op_data, it, reqp, cb_blocking,
536 else if (it->it_op & IT_OPEN)
537 rc = lmv_intent_open(exp, op_data, it, reqp, cb_blocking,
543 struct lustre_handle lock_handle;
545 if (it->d.lustre.it_lock_mode != 0) {
546 lock_handle.cookie = it->d.lustre.it_lock_handle;
547 ldlm_lock_decref(&lock_handle,
548 it->d.lustre.it_lock_mode);
551 it->d.lustre.it_lock_handle = 0;
552 it->d.lustre.it_lock_mode = 0;
554 if (it->d.lustre.it_remote_lock_mode != 0) {
555 lock_handle.cookie = it->d.lustre.it_remote_lock_handle;
556 ldlm_lock_decref(&lock_handle,
557 it->d.lustre.it_remote_lock_mode);
560 it->d.lustre.it_remote_lock_handle = 0;
561 it->d.lustre.it_remote_lock_mode = 0;