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, struct ll_uctxt *uctxt,
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;
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, uctxt, &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, struct ll_uctxt *uctxt,
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;
136 /* IT_OPEN is intended to open (and create, possible) an object. Parent
137 * (pfid) may be splitted dir */
141 obj = lmv_grab_obj(obd, &rpfid);
143 /* directory is already splitted, so we have to forward
144 * request to the right MDS */
145 mds = raw_name2idx(obj->objcount, (char *)name, len);
146 CDEBUG(D_OTHER, "forward to MDS #%u\n", mds);
148 rpfid = obj->objs[mds].fid;
152 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, uctxt, &rpfid, name,
153 len, lmm, lmmsize, cfid, it, flags, reqp,
155 if (rc == -ERESTART) {
156 /* directory got splitted. time to update local object
157 * and repeat the request with proper MDS */
158 LASSERT(fid_equal(pfid, &rpfid));
159 rc = lmv_get_mea_and_update_object(exp, &rpfid);
161 ptlrpc_req_finished(*reqp);
168 /* okay, MDS has returned success. Probably name has been resolved in
170 rc = lmv_handle_remote_inode(exp, uctxt, lmm, lmmsize, it,
171 flags, reqp, cb_blocking);
177 /* caller may use attrs MDS returns on IT_OPEN lock request so, we have
178 * to update them for splitted dir */
179 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
180 LASSERT(body != NULL);
183 obj = lmv_grab_obj(obd, cfid);
184 if (!obj && (mea = is_body_of_splitted_dir(*reqp, 1))) {
185 /* wow! this is splitted dir, we'd like to handle it */
186 rc = lmv_create_obj(exp, &body->fid1, mea);
190 obj = lmv_grab_obj(obd, cfid);
194 /* this is splitted dir and we'd want to get attrs */
195 CDEBUG(D_OTHER, "attrs from slaves for %lu/%lu/%lu\n",
196 (unsigned long)cfid->mds, (unsigned long)cfid->id,
197 (unsigned long)cfid->generation);
198 rc = lmv_revalidate_slaves(exp, reqp, cfid,
200 } else if (S_ISDIR(body->mode)) {
201 /*CWARN("hmmm, %lu/%lu/%lu has not lmv obj?!\n",
202 (unsigned long) cfid->mds,
203 (unsigned long) cfid->id,
204 (unsigned long) cfid->generation);*/
213 int lmv_intent_getattr(struct obd_export *exp, struct ll_uctxt *uctxt,
214 struct ll_fid *pfid, const char *name, int len,
215 void *lmm, int lmmsize, struct ll_fid *cfid,
216 struct lookup_intent *it, int flags,
217 struct ptlrpc_request **reqp,
218 ldlm_blocking_callback cb_blocking)
220 struct obd_device *obd = exp->exp_obd;
221 struct lmv_obd *lmv = &obd->u.lmv;
222 struct mds_body *body = NULL;
223 struct ll_fid rpfid = *pfid;
224 struct lmv_obj *obj, *obj2;
230 /* caller wants to revalidate attrs of obj we have to revalidate
231 * slaves if requested object is splitted directory */
232 CDEBUG(D_OTHER, "revalidate attrs for %lu/%lu/%lu\n",
233 (unsigned long)cfid->mds, (unsigned long)cfid->id,
234 (unsigned long)cfid->generation);
236 obj = lmv_grab_obj(obd, cfid);
238 /* in fact, we need not this with current intent_lock(),
239 * but it may change some day */
240 rpfid = obj->objs[mds].fid;
243 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, uctxt, &rpfid, name,
244 len, lmm, lmmsize, cfid, it, flags, reqp,
246 if (obj && rc >= 0) {
247 /* this is splitted dir. In order to optimize things a
248 * bit, we consider obj valid updating missing parts.
250 * FIXME: do we need to return any lock here? It would
251 * be fine if we don't. this means that nobody should
252 * use UPDATE lock to notify about object * removal */
254 "revalidate slaves for %lu/%lu/%lu, rc %d\n",
255 (unsigned long)cfid->mds, (unsigned long)cfid->id,
256 (unsigned long)cfid->generation, rc);
258 rc = lmv_revalidate_slaves(exp, reqp, cfid, it, rc,
265 CDEBUG(D_OTHER, "INTENT getattr for %*s on %lu/%lu/%lu\n",
266 len, name, (unsigned long)pfid->mds, (unsigned long)pfid->id,
267 (unsigned long)pfid->generation);
270 obj = lmv_grab_obj(obd, pfid);
272 /* directory is already splitted. calculate mds */
273 mds = raw_name2idx(obj->objcount, (char *) name, len);
274 rpfid = obj->objs[mds].fid;
277 CDEBUG(D_OTHER, "forward to MDS #%u (slave %lu/%lu/%lu)\n",
278 mds, (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
279 (unsigned long)rpfid.generation);
282 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, uctxt, &rpfid, name,
283 len, lmm, lmmsize, NULL, it, flags, reqp,
291 /* okay, MDS has returned success. probably name has been
292 * resolved in remote inode */
293 rc = lmv_handle_remote_inode(exp, uctxt, lmm, lmmsize, it,
294 flags, reqp, cb_blocking);
298 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
299 LASSERT(body != NULL);
302 obj2 = lmv_grab_obj(obd, cfid);
304 if (!obj2 && (mea = is_body_of_splitted_dir(*reqp, 1))) {
305 /* wow! this is splitted dir, we'd like to handle it. */
306 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
307 LASSERT(body != NULL);
309 rc = lmv_create_obj(exp, &body->fid1, mea);
313 obj2 = lmv_grab_obj(obd, cfid);
317 /* this is splitted dir and we'd want to get attrs */
318 CDEBUG(D_OTHER, "attrs from slaves for %lu/%lu/%lu, rc %d\n",
319 (unsigned long)cfid->mds, (unsigned long)cfid->id,
320 (unsigned long)cfid->generation, rc);
322 rc = lmv_revalidate_slaves(exp, reqp, cfid, it, 1, cb_blocking);
328 void lmv_update_body_from_obj(struct mds_body *body, struct lmv_inode *obj)
331 body->size += obj->size;
339 int lmv_lookup_slaves(struct obd_export *exp, struct ptlrpc_request **reqp)
341 struct obd_device *obd = exp->exp_obd;
342 struct lmv_obd *lmv = &obd->u.lmv;
343 struct mds_body *body = NULL;
344 struct lustre_handle *lockh;
345 struct ldlm_lock *lock;
346 struct mds_body *body2;
347 struct ll_uctxt uctxt;
355 /* master is locked. we'd like to take locks on slaves and update
356 * attributes to be returned from the slaves it's important that lookup
357 * is called in two cases:
359 * - for first time (dcache has no such a resolving yet.
360 * - ->d_revalidate() returned false.
362 * last case possible only if all the objs (master and all slaves aren't
365 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
366 LASSERT(body != NULL);
368 obj = lmv_grab_obj(obd, &body->fid1);
371 CDEBUG(D_OTHER, "lookup slaves for %lu/%lu/%lu\n",
372 (unsigned long)body->fid1.mds,
373 (unsigned long)body->fid1.id,
374 (unsigned long)body->fid1.generation);
379 for (i = 0; i < obj->objcount; i++) {
380 struct ll_fid fid = obj->objs[i].fid;
381 struct ptlrpc_request *req = NULL;
382 struct lookup_intent it;
384 if (fid_equal(&fid, &obj->fid))
385 /* skip master obj */
388 CDEBUG(D_OTHER, "lookup slave %lu/%lu/%lu\n",
389 (unsigned long)fid.mds, (unsigned long)fid.id,
390 (unsigned long)fid.generation);
393 memset(&it, 0, sizeof(it));
394 it.it_op = IT_GETATTR;
395 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &uctxt, &fid,
396 NULL, 0, NULL, 0, &fid, &it, 0, &req,
397 lmv_dirobj_blocking_ast);
398 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
400 /* nice, this slave is valid */
401 LASSERT(req == NULL);
402 CDEBUG(D_OTHER, "cached\n");
407 /* error during revalidation */
410 /* rc == 0, this means we have no such a lock and can't think
411 * obj is still valid. lookup it again */
412 LASSERT(req == NULL);
414 memset(&it, 0, sizeof(it));
415 it.it_op = IT_GETATTR;
416 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &uctxt, &fid,
417 NULL, 0, NULL, 0, NULL, &it, 0, &req,
418 lmv_dirobj_blocking_ast);
420 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
424 /* error during lookup */
427 lock = ldlm_handle2lock(lockh);
430 lock->l_ast_data = lmv_get_obj(obj);
432 body2 = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body2));
435 obj->objs[i].size = body2->size;
436 CDEBUG(D_OTHER, "fresh: %lu\n",
437 (unsigned long) obj->objs[i].size);
442 ptlrpc_req_finished(req);
444 lmv_update_body_from_obj(body, obj->objs + i);
445 if (it.d.lustre.it_lock_mode)
446 ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
454 int lmv_intent_lookup(struct obd_export *exp, struct ll_uctxt *uctxt,
455 struct ll_fid *pfid, const char *name, int len,
456 void *lmm, int lmmsize, struct ll_fid *cfid,
457 struct lookup_intent *it, int flags,
458 struct ptlrpc_request **reqp,
459 ldlm_blocking_callback cb_blocking)
461 struct obd_device *obd = exp->exp_obd;
462 struct lmv_obd *lmv = &obd->u.lmv;
463 struct mds_body *body = NULL;
464 struct ll_fid rpfid = *pfid;
470 /* IT_LOOKUP is intended to produce name -> fid resolving (let's call
471 * this lookup below) or to confirm requested resolving is still valid
472 * (let's call this revalidation) cfid != NULL specifies revalidation */
475 /* this is revalidation: we have to check is LOOKUP lock still
476 * valid for given fid. very important part is that we have to
477 * choose right mds because namespace is per mds */
479 obj = lmv_grab_obj(obd, pfid);
481 mds = raw_name2idx(obj->objcount, (char *) name, len);
482 rpfid = obj->objs[mds].fid;
487 CDEBUG(D_OTHER, "revalidate lookup for %lu/%lu/%lu to %d MDS\n",
488 (unsigned long)cfid->mds, (unsigned long)cfid->id,
489 (unsigned long)cfid->generation, mds);
491 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, uctxt, pfid, name,
492 len, lmm, lmmsize, cfid, it, flags,
499 /* this is lookup. during lookup we have to update all the attributes,
500 * because returned values will be put in struct inode */
502 obj = lmv_grab_obj(obd, pfid);
504 /* directory is already splitted. calculate mds */
505 mds = raw_name2idx(obj->objcount, (char *)name, len);
506 rpfid = obj->objs[mds].fid;
510 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, uctxt, &rpfid, name,
511 len, lmm, lmmsize, NULL, it, flags, reqp,
514 /* very interesting. it seems object is still valid but for some
515 * reason llite calls lookup, not revalidate */
516 CWARN("lookup for %lu/%lu/%lu and data should be uptodate\n",
517 (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
518 (unsigned long)rpfid.generation);
520 LASSERT(*reqp == NULL);
524 if (rc == 0 && *reqp == NULL) {
525 /* once again, we're asked for lookup, not revalidate */
526 CWARN("lookup for %lu/%lu/%lu and data should be uptodate\n",
527 (unsigned long)rpfid.mds, (unsigned long)rpfid.id,
528 (unsigned long)rpfid.generation);
532 if (rc == -ERESTART) {
533 /* directory got splitted since last update. this shouldn't
534 * be becasue splitting causes lock revocation, so revalidate
535 * had to fail and lookup on dir had to return mea */
536 CWARN("we haven't knew about directory splitting!\n");
537 LASSERT(obj == NULL);
538 rc = lmv_create_obj(exp, &rpfid, NULL);
547 /* okay, MDS has returned success. probably name has been
548 * resolved in remote inode */
549 rc = lmv_handle_remote_inode(exp, uctxt, lmm, lmmsize, it, flags,
552 if (rc == 0 && (mea = is_body_of_splitted_dir(*reqp, 1))) {
553 /* wow! this is splitted dir, we'd like to handle it */
554 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
555 LASSERT(body != NULL);
557 if ((obj = lmv_grab_obj(obd, &body->fid1)))
560 rc = lmv_create_obj(exp, &body->fid1, mea);
566 int lmv_intent_lock(struct obd_export *exp, struct ll_uctxt *uctxt,
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, uctxt, pfid, name, len, lmm,
590 lmmsize, cfid, it, flags, reqp,
592 else if (it->it_op & IT_OPEN)
593 rc = lmv_intent_open(exp, uctxt, 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, uctxt, 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 unsigned long size = 0;
614 struct ldlm_lock *lock;
615 struct mds_body *body;
616 struct ll_uctxt uctxt;
618 int master_lock_mode;
622 /* we have to loop over the subobjects, check validity and update them
623 * from MDSs if needed. it's very useful that we need not to update all
624 * the fields. say, common fields (that are equal on all the subojects
625 * need not to be update, another fields (i_size, for example) are
626 * cached all the time */
627 obj = lmv_grab_obj(obd, mfid);
632 master_lock_mode = 0;
634 for (i = 0; i < obj->objcount; i++) {
635 struct ll_fid fid = obj->objs[i].fid;
636 struct lustre_handle *lockh = NULL;
637 struct ptlrpc_request *req = NULL;
638 ldlm_blocking_callback cb;
639 struct lookup_intent it;
642 CDEBUG(D_OTHER, "revalidate subobj %lu/%lu/%lu\n",
643 (unsigned long)fid.mds, (unsigned long)fid.id,
644 (unsigned long) fid.generation);
646 memset(&it, 0, sizeof(it));
647 it.it_op = IT_GETATTR;
648 cb = lmv_dirobj_blocking_ast;
650 if (fid_equal(&fid, &obj->fid)) {
652 /* lmv_intent_getattr() already checked
653 * validness and took the lock */
655 /* it even got the reply refresh attrs
657 body = lustre_msg_buf(mreq->rq_repmsg,
659 LASSERT(body != NULL);
662 /* take already cached attrs into account */
664 "master is locked and cached\n");
672 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &uctxt, &fid,
673 NULL, 0, NULL, 0, &fid, &it, 0, &req, cb);
674 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
676 /* nice, this slave is valid */
677 LASSERT(req == NULL);
678 CDEBUG(D_OTHER, "cached\n");
683 /* error during revalidation */
686 /* rc == 0, this means we have no such a lock and can't think
687 * obj is still valid. lookup it again */
688 LASSERT(req == NULL);
690 memset(&it, 0, sizeof(it));
691 it.it_op = IT_GETATTR;
692 rc = md_intent_lock(lmv->tgts[fid.mds].ltd_exp, &uctxt, &fid,
693 NULL, 0, NULL, 0, NULL, &it, 0, &req, cb);
694 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
697 /* error during lookup */
701 LASSERT(master_valid == 0);
702 /* save lock on master to be returned to the caller */
703 CDEBUG(D_OTHER, "no lock on master yet\n");
704 memcpy(&master_lockh, lockh, sizeof(master_lockh));
705 master_lock_mode = it.d.lustre.it_lock_mode;
706 it.d.lustre.it_lock_mode = 0;
708 /* this is slave. we want to control it */
709 lock = ldlm_handle2lock(lockh);
711 lock->l_ast_data = lmv_get_obj(obj);
716 /* this is first reply, we'll use it to return
717 * updated data back to the caller */
719 ptlrpc_request_addref(req);
724 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
728 obj->objs[i].size = body->size;
729 CDEBUG(D_OTHER, "fresh: %lu\n",
730 (unsigned long) obj->objs[i].size);
733 ptlrpc_req_finished(req);
735 size += obj->objs[i].size;
736 if (it.d.lustre.it_lock_mode)
737 ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
741 /* some attrs got refreshed, we have reply and it's time
742 * to put fresh attrs to it */
743 CDEBUG(D_OTHER, "return refreshed attrs: size = %lu\n",
744 (unsigned long) size);
745 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
748 /* FIXME: what about another attributes? */
751 /* very important to maintain lli->mds the same because
752 * of revalidation. mreq == NULL means that caller has
753 * no reply and the only attr we can return is size */
754 body->valid = OBD_MD_FLSIZE;
755 body->mds = obj->fid.mds;
757 if (master_valid == 0) {
758 memcpy(&oit->d.lustre.it_lock_handle,
759 &master_lockh, sizeof(master_lockh));
760 oit->d.lustre.it_lock_mode = master_lock_mode;
764 /* it seems all the attrs are fresh and we did no request */
765 CDEBUG(D_OTHER, "all the attrs were fresh\n");
766 if (master_valid == 0)
767 oit->d.lustre.it_lock_mode = master_lock_mode;