1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 # define EXPORT_SYMTAB
25 #define DEBUG_SUBSYSTEM S_LMV
27 #include <linux/slab.h>
28 #include <linux/module.h>
29 #include <linux/init.h>
30 #include <linux/slab.h>
31 #include <linux/pagemap.h>
32 #include <asm/div64.h>
33 #include <linux/seq_file.h>
35 #include <liblustre.h>
38 #include <linux/obd_support.h>
39 #include <linux/lustre_lib.h>
40 #include <linux/lustre_net.h>
41 #include <linux/lustre_idl.h>
42 #include <linux/lustre_dlm.h>
43 #include <linux/lustre_mds.h>
44 #include <linux/obd_class.h>
45 #include <linux/obd_ost.h>
46 #include <linux/lprocfs_status.h>
47 #include <linux/lustre_fsfilt.h>
48 #include <linux/obd_lmv.h>
49 #include "lmv_internal.h"
52 static inline void lmv_drop_intent_lock(struct lookup_intent *it)
54 if (it->d.lustre.it_lock_mode != 0)
55 ldlm_lock_decref((void *)&it->d.lustre.it_lock_handle,
56 it->d.lustre.it_lock_mode);
59 int lmv_handle_remote_inode(struct obd_export *exp,
60 void *lmm, int lmmsize,
61 struct lookup_intent *it, int flags,
62 struct ptlrpc_request **reqp,
63 ldlm_blocking_callback cb_blocking)
65 struct obd_device *obd = exp->exp_obd;
66 struct lmv_obd *lmv = &obd->u.lmv;
67 struct mds_body *body = NULL;
71 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
72 LASSERT(body != NULL);
74 if (body->valid & OBD_MD_MDS) {
75 /* oh, MDS reports that this is remote inode case
76 * i.e. we have to ask for real attrs on another MDS */
77 struct ptlrpc_request *req = NULL;
79 struct lustre_handle plock;
82 if (it->it_op == IT_LOOKUP) {
83 /* unfortunately, we have to lie to MDC/MDS to
84 * retrieve attributes llite needs */
85 it->it_op = IT_GETATTR;
88 /* we got LOOKUP lock, but we really need attrs */
89 pmode = it->d.lustre.it_lock_mode;
91 memcpy(&plock, &it->d.lustre.it_lock_handle,
93 it->d.lustre.it_lock_mode = 0;
97 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
98 rc = md_intent_lock(lmv->tgts[nfid.mds].ltd_exp, &nfid,
99 NULL, 0, lmm, lmmsize, NULL, it, flags,
102 /* llite needs LOOKUP lock to track dentry revocation in
103 * order to maintain dcache consistency. thus drop UPDATE
104 * lock here and put LOOKUP in request */
106 lmv_drop_intent_lock(it);
107 memcpy(&it->d.lustre.it_lock_handle, &plock,
109 it->d.lustre.it_lock_mode = pmode;
112 ldlm_lock_decref(&plock, pmode);
114 ptlrpc_req_finished(*reqp);
120 int lmv_intent_open(struct obd_export *exp,
121 struct ll_fid *pfid, const char *name, int len,
122 void *lmm, int lmmsize, struct ll_fid *cfid,
123 struct lookup_intent *it, int flags,
124 struct ptlrpc_request **reqp,
125 ldlm_blocking_callback cb_blocking)
127 struct obd_device *obd = exp->exp_obd;
128 struct lmv_obd *lmv = &obd->u.lmv;
129 struct mds_body *body = NULL;
130 struct ll_fid rpfid = *pfid;
133 int rc, mds, loop = 0;
136 /* IT_OPEN is intended to open (and create, possible) an object. Parent
137 * (pfid) may be splitted dir */
140 LASSERT(++loop <= 2);
142 obj = lmv_grab_obj(obd, &rpfid);
144 /* directory is already splitted, so we have to forward
145 * request to the right MDS */
146 mds = raw_name2idx(obj->hashtype, obj->objcount, (char *)name, len);
147 CDEBUG(D_OTHER, "forward to MDS #%u (%lu/%lu/%lu)\n", mds,
148 (unsigned long) rpfid.mds, (unsigned long) rpfid.id,
149 (unsigned long) rpfid.generation);
150 rpfid = obj->objs[mds].fid;
154 rc = md_intent_lock(lmv->tgts[rpfid.mds].ltd_exp, &rpfid, name,
155 len, lmm, lmmsize, cfid, it, flags, reqp,
157 if (rc == -ERESTART) {
158 /* directory got splitted. time to update local object
159 * and repeat the request with proper MDS */
160 LASSERT(fid_equal(pfid, &rpfid));
161 rc = lmv_get_mea_and_update_object(exp, &rpfid);
163 ptlrpc_req_finished(*reqp);
170 /* okay, MDS has returned success. Probably name has been resolved in
172 rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it,
173 flags, reqp, cb_blocking);
176 CERROR("can't handle remote %s: dir %lu/%lu/%lu(%lu/%lu/%lu):"
177 "%*s: %d\n", LL_IT2STR(it),
178 (unsigned long) pfid->mds,
179 (unsigned long) pfid->id,
180 (unsigned long) pfid->generation,
181 (unsigned long) rpfid.mds,
182 (unsigned long) rpfid.id,
183 (unsigned long) rpfid.generation,
188 /* caller may use attrs MDS returns on IT_OPEN lock request so, we have
189 * to update them for splitted dir */
190 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
191 LASSERT(body != NULL);
194 obj = lmv_grab_obj(obd, cfid);
195 if (!obj && (mea = body_of_splitted_dir(*reqp, 1))) {
196 /* wow! this is splitted dir, we'd like to handle it */
197 obj = lmv_create_obj(exp, &body->fid1, mea);
199 RETURN(PTR_ERR(obj));
203 /* this is splitted dir and we'd want to get attrs */
204 CDEBUG(D_OTHER, "attrs from slaves for %lu/%lu/%lu\n",
205 (unsigned long)cfid->mds, (unsigned long)cfid->id,
206 (unsigned long)cfid->generation);
207 rc = lmv_revalidate_slaves(exp, reqp, cfid,
209 } else if (S_ISDIR(body->mode)) {
210 /*CWARN("hmmm, %lu/%lu/%lu has not lmv obj?!\n",
211 (unsigned long) cfid->mds,
212 (unsigned long) cfid->id,
213 (unsigned long) cfid->generation);*/
222 int lmv_intent_getattr(struct obd_export *exp,
223 struct ll_fid *pfid, const char *name, int len,
224 void *lmm, int lmmsize, struct ll_fid *cfid,
225 struct lookup_intent *it, int flags,
226 struct ptlrpc_request **reqp,
227 ldlm_blocking_callback cb_blocking)
229 struct obd_device *obd = exp->exp_obd;
230 struct lmv_obd *lmv = &obd->u.lmv;
231 struct mds_body *body = NULL;
232 struct ll_fid rpfid = *pfid;
233 struct lmv_obj *obj, *obj2;
239 /* caller wants to revalidate attrs of obj we have to revalidate
240 * slaves if requested object is splitted directory */
241 CDEBUG(D_OTHER, "revalidate attrs for %lu/%lu/%lu\n",
242 (unsigned long)cfid->mds, (unsigned long)cfid->id,
243 (unsigned long)cfid->generation);
245 obj = lmv_grab_obj(obd, cfid);
247 /* in fact, we need not this with current intent_lock(),
248 * but it may change some day */
249 if (!fid_equal(pfid, cfid)){
250 rpfid = obj->objs[mds].fid;
256 CDEBUG(D_OTHER, "INTENT getattr for %*s on %lu/%lu/%lu\n",
257 len, name, (unsigned long)pfid->mds, (unsigned long)pfid->id,
258 (unsigned long)pfid->generation);
260 obj = lmv_grab_obj(obd, pfid);
262 /* directory is already splitted. calculate mds */
263 mds = raw_name2idx(obj->hashtype, obj->objcount, (char *) name, len);
264 rpfid = obj->objs[mds].fid;
268 CDEBUG(D_OTHER, "forward to MDS #%u (slave %lu/%lu/%lu)\n",
269 mds, (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
270 (unsigned long)rpfid.generation);
273 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, &rpfid, name,
274 len, lmm, lmmsize, cfid, it, flags, reqp,
280 /* this is splitted dir. In order to optimize things a
281 * bit, we consider obj valid updating missing parts.
283 * FIXME: do we need to return any lock here? It would
284 * be fine if we don't. this means that nobody should
285 * use UPDATE lock to notify about object * removal */
287 "revalidate slaves for %lu/%lu/%lu, rc %d\n",
288 (unsigned long)cfid->mds, (unsigned long)cfid->id,
289 (unsigned long)cfid->generation, rc);
292 rc = lmv_revalidate_slaves(exp, reqp, cfid, it, rc,
300 /* okay, MDS has returned success. probably name has been
301 * resolved in remote inode */
302 rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it,
303 flags, reqp, cb_blocking);
307 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
308 LASSERT(body != NULL);
311 obj2 = lmv_grab_obj(obd, cfid);
313 if (!obj2 && (mea = body_of_splitted_dir(*reqp, 1))) {
314 /* wow! this is splitted dir, we'd like to handle it. */
315 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
316 LASSERT(body != NULL);
318 obj2 = lmv_create_obj(exp, &body->fid1, mea);
320 RETURN(PTR_ERR(obj2));
324 /* this is splitted dir and we'd want to get attrs */
325 CDEBUG(D_OTHER, "attrs from slaves for %lu/%lu/%lu, rc %d\n",
326 (unsigned long)cfid->mds, (unsigned long)cfid->id,
327 (unsigned long)cfid->generation, rc);
329 rc = lmv_revalidate_slaves(exp, reqp, cfid, it, 1, cb_blocking);
335 void lmv_update_body_from_obj(struct mds_body *body, struct lmv_inode *obj)
338 body->size += obj->size;
339 /* body->atime = obj->atime;
340 body->ctime = obj->ctime;
341 body->mtime = obj->mtime;
342 body->nlink = obj->nlink;*/
345 int lmv_lookup_slaves(struct obd_export *exp, struct ptlrpc_request **reqp)
347 struct obd_device *obd = exp->exp_obd;
348 struct lmv_obd *lmv = &obd->u.lmv;
349 struct mds_body *body = NULL;
350 struct lustre_handle *lockh;
351 struct ldlm_lock *lock;
352 struct mds_body *body2;
360 /* master is locked. we'd like to take locks on slaves and update
361 * attributes to be returned from the slaves it's important that lookup
362 * is called in two cases:
364 * - for first time (dcache has no such a resolving yet).
365 * - ->d_revalidate() returned false.
367 * last case possible only if all the objs (master and all slaves aren't
370 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
371 LASSERT(body != NULL);
373 obj = lmv_grab_obj(obd, &body->fid1);
374 LASSERT(obj != NULL);
376 CDEBUG(D_OTHER, "lookup slaves for %lu/%lu/%lu\n",
377 (unsigned long)body->fid1.mds,
378 (unsigned long)body->fid1.id,
379 (unsigned long)body->fid1.generation);
383 for (i = 0; i < obj->objcount; i++) {
384 struct ll_fid fid = obj->objs[i].fid;
385 struct ptlrpc_request *req = NULL;
386 struct lookup_intent it;
388 if (fid_equal(&fid, &obj->fid))
389 /* skip master obj */
392 CDEBUG(D_OTHER, "lookup slave %lu/%lu/%lu\n",
393 (unsigned long)fid.mds, (unsigned long)fid.id,
394 (unsigned long)fid.generation);
397 memset(&it, 0, sizeof(it));
398 it.it_op = IT_GETATTR;
399 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &fid,
400 NULL, 0, NULL, 0, &fid, &it, 0, &req,
401 lmv_dirobj_blocking_ast);
403 lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
404 if (rc > 0 && req == NULL) {
405 /* nice, this slave is valid */
406 LASSERT(req == NULL);
407 CDEBUG(D_OTHER, "cached\n");
412 /* error during lookup */
415 lock = ldlm_handle2lock(lockh);
418 lock->l_ast_data = lmv_get_obj(obj);
420 body2 = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body2));
423 obj->objs[i].size = body2->size;
425 CDEBUG(D_OTHER, "fresh: %lu\n",
426 (unsigned long)obj->objs[i].size);
431 ptlrpc_req_finished(req);
433 lmv_update_body_from_obj(body, obj->objs + i);
435 if (it.d.lustre.it_lock_mode)
436 ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
444 int lmv_intent_lookup(struct obd_export *exp,
445 struct ll_fid *pfid, const char *name, int len,
446 void *lmm, int lmmsize, struct ll_fid *cfid,
447 struct lookup_intent *it, int flags,
448 struct ptlrpc_request **reqp,
449 ldlm_blocking_callback cb_blocking)
451 struct obd_device *obd = exp->exp_obd;
452 struct lmv_obd *lmv = &obd->u.lmv;
453 struct mds_body *body = NULL;
454 struct ll_fid rpfid = *pfid;
457 int rc, mds, loop = 0;
460 /* IT_LOOKUP is intended to produce name -> fid resolving (let's call
461 * this lookup below) or to confirm requested resolving is still valid
462 * (let's call this revalidation) cfid != NULL specifies revalidation */
465 /* this is revalidation: we have to check is LOOKUP lock still
466 * valid for given fid. very important part is that we have to
467 * choose right mds because namespace is per mds */
469 obj = lmv_grab_obj(obd, pfid);
471 mds = raw_name2idx(obj->hashtype, obj->objcount,
473 rpfid = obj->objs[mds].fid;
478 CDEBUG(D_OTHER, "revalidate lookup for %lu/%lu/%lu to %d MDS\n",
479 (unsigned long)cfid->mds, (unsigned long)cfid->id,
480 (unsigned long)cfid->generation, mds);
485 LASSERT(++loop <= 2);
486 /* this is lookup. during lookup we have to update all the
487 * attributes, because returned values will be put in struct
490 obj = lmv_grab_obj(obd, pfid);
493 /* directory is already splitted. calculate mds */
494 mds = raw_name2idx(obj->hashtype, obj->objcount,
496 rpfid = obj->objs[mds].fid;
502 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, pfid, name,
503 len, lmm, lmmsize, cfid, it, flags, reqp,
510 /* very interesting. it seems object is still valid but for some
511 * reason llite calls lookup, not revalidate */
512 CDEBUG(D_OTHER, "lookup for %lu/%lu/%lu and data should be uptodate\n",
513 (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
514 (unsigned long)rpfid.generation);
515 LASSERT(*reqp == NULL);
519 if (rc == 0 && *reqp == NULL) {
520 /* once again, we're asked for lookup, not revalidate */
521 CDEBUG(D_OTHER, "lookup for %lu/%lu/%lu and data should be uptodate\n",
522 (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
523 (unsigned long)rpfid.generation);
527 if (rc == -ERESTART) {
528 /* directory got splitted since last update. this shouldn't be
529 * becasue splitting causes lock revocation, so revalidate had
530 * to fail and lookup on dir had to return mea */
531 CWARN("we haven't knew about directory splitting!\n");
532 LASSERT(obj == NULL);
534 obj = lmv_create_obj(exp, &rpfid, NULL);
536 RETURN(PTR_ERR(obj));
544 /* okay, MDS has returned success. probably name has been resolved in
546 rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it, flags,
549 if (rc == 0 && (mea = body_of_splitted_dir(*reqp, 1))) {
550 /* wow! this is splitted dir, we'd like to handle it */
551 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
552 LASSERT(body != NULL);
554 obj = lmv_grab_obj(obd, &body->fid1);
556 obj = lmv_create_obj(exp, &body->fid1, mea);
558 RETURN(PTR_ERR(obj));
566 int lmv_intent_lock(struct obd_export *exp,
567 struct ll_fid *pfid, const char *name, int len,
568 void *lmm, int lmmsize, struct ll_fid *cfid,
569 struct lookup_intent *it, int flags,
570 struct ptlrpc_request **reqp,
571 ldlm_blocking_callback cb_blocking)
573 struct obd_device *obd = exp->exp_obd;
580 CDEBUG(D_OTHER, "INTENT LOCK '%s' for '%*s' on %lu/%lu -> %u\n",
581 LL_IT2STR(it), len, name, (unsigned long) pfid->id,
582 (unsigned long) pfid->generation, pfid->mds);
584 rc = lmv_check_connect(obd);
588 if (it->it_op == IT_LOOKUP)
589 rc = lmv_intent_lookup(exp, pfid, name, len, lmm,
590 lmmsize, cfid, it, flags, reqp,
592 else if (it->it_op & IT_OPEN)
593 rc = lmv_intent_open(exp, pfid, name, len, lmm,
594 lmmsize, cfid, it, flags, reqp,
596 else if (it->it_op == IT_GETATTR || it->it_op == IT_CHDIR)
597 rc = lmv_intent_getattr(exp, pfid, name, len, lmm,
598 lmmsize, cfid, it, flags, reqp,
605 int lmv_revalidate_slaves(struct obd_export *exp, struct ptlrpc_request **reqp,
606 struct ll_fid *mfid, struct lookup_intent *oit,
607 int master_valid, ldlm_blocking_callback cb_blocking)
609 struct obd_device *obd = exp->exp_obd;
610 struct ptlrpc_request *mreq = *reqp;
611 struct lmv_obd *lmv = &obd->u.lmv;
612 struct lustre_handle master_lockh;
613 struct ldlm_lock *lock;
614 unsigned long size = 0;
615 struct mds_body *body;
617 int master_lock_mode;
621 /* we have to loop over the subobjects, check validity and update them
622 * from MDSs if needed. it's very useful that we need not to update all
623 * the fields. say, common fields (that are equal on all the subojects
624 * need not to be update, another fields (i_size, for example) are
625 * cached all the time */
626 obj = lmv_grab_obj(obd, mfid);
627 LASSERT(obj != NULL);
629 master_lock_mode = 0;
633 for (i = 0; i < obj->objcount; i++) {
634 struct ll_fid fid = obj->objs[i].fid;
635 struct lustre_handle *lockh = NULL;
636 struct ptlrpc_request *req = NULL;
637 ldlm_blocking_callback cb;
638 struct lookup_intent it;
641 CDEBUG(D_OTHER, "revalidate subobj %lu/%lu/%lu\n",
642 (unsigned long)fid.mds, (unsigned long)fid.id,
643 (unsigned long) fid.generation);
645 memset(&it, 0, sizeof(it));
646 it.it_op = IT_GETATTR;
647 cb = lmv_dirobj_blocking_ast;
649 if (fid_equal(&fid, &obj->fid)) {
651 /* lmv_intent_getattr() already checked
652 * validness and took the lock */
654 /* it even got the reply refresh attrs
656 body = lustre_msg_buf(mreq->rq_repmsg,
658 LASSERT(body != NULL);
661 /* take already cached attrs into account */
663 "master is locked and cached\n");
671 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &fid,
672 NULL, 0, NULL, 0, &fid, &it, 0, &req, cb);
673 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
674 if (rc > 0 && req == NULL) {
675 /* nice, this slave is valid */
676 LASSERT(req == NULL);
677 CDEBUG(D_OTHER, "cached\n");
682 /* error during revalidation */
686 LASSERT(master_valid == 0);
687 /* save lock on master to be returned to the caller */
688 CDEBUG(D_OTHER, "no lock on master yet\n");
689 memcpy(&master_lockh, lockh, sizeof(master_lockh));
690 master_lock_mode = it.d.lustre.it_lock_mode;
691 it.d.lustre.it_lock_mode = 0;
693 /* this is slave. we want to control it */
694 lock = ldlm_handle2lock(lockh);
696 lock->l_ast_data = lmv_get_obj(obj);
701 /* this is first reply, we'll use it to return
702 * updated data back to the caller */
704 ptlrpc_request_addref(req);
709 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
713 obj->objs[i].size = body->size;
715 CDEBUG(D_OTHER, "fresh: %lu\n",
716 (unsigned long)obj->objs[i].size);
719 ptlrpc_req_finished(req);
721 size += obj->objs[i].size;
723 if (it.d.lustre.it_lock_mode)
724 ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
728 /* some attrs got refreshed, we have reply and it's time to put
729 * fresh attrs to it */
730 CDEBUG(D_OTHER, "return refreshed attrs: size = %lu\n",
731 (unsigned long)size);
733 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
736 /* FIXME: what about other attributes? */
740 /* very important to maintain lli->mds the same because
741 * of revalidation. mreq == NULL means that caller has
742 * no reply and the only attr we can return is size */
743 body->valid = OBD_MD_FLSIZE;
744 body->mds = obj->fid.mds;
746 if (master_valid == 0) {
747 memcpy(&oit->d.lustre.it_lock_handle,
748 &master_lockh, sizeof(master_lockh));
749 oit->d.lustre.it_lock_mode = master_lock_mode;
753 /* it seems all the attrs are fresh and we did no request */
754 CDEBUG(D_OTHER, "all the attrs were fresh\n");
755 if (master_valid == 0)
756 oit->d.lustre.it_lock_mode = master_lock_mode;