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, void *lmm,
60 int lmmsize, struct lookup_intent *it,
61 int flags, struct ptlrpc_request **reqp,
62 ldlm_blocking_callback cb_blocking)
64 struct obd_device *obd = exp->exp_obd;
65 struct lmv_obd *lmv = &obd->u.lmv;
66 struct mds_body *body = NULL;
70 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
71 LASSERT(body != NULL);
73 if (body->valid & OBD_MD_MDS) {
75 * oh, MDS reports that this is remote inode case i.e. we have
76 * to ask for real attrs on another MDS.
78 struct ptlrpc_request *req = NULL;
79 struct lustre_handle plock;
83 if (it->it_op == IT_LOOKUP || it->it_op == IT_CHDIR) {
85 * unfortunately, we have to lie to MDC/MDS to retrieve
86 * attributes llite needs.
88 it->it_op = IT_GETATTR;
91 /* we got LOOKUP lock, but we really need attrs */
92 pmode = it->d.lustre.it_lock_mode;
94 memcpy(&plock, &it->d.lustre.it_lock_handle,
96 it->d.lustre.it_lock_mode = 0;
100 it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
101 rc = md_intent_lock(lmv->tgts[id_group(&nid)].ltd_exp, &nid,
102 NULL, 0, lmm, lmmsize, NULL, it, flags,
106 * llite needs LOOKUP lock to track dentry revocation in order
107 * to maintain dcache consistency. Thus drop UPDATE lock here
108 * and put LOOKUP in request.
111 lmv_drop_intent_lock(it);
112 memcpy(&it->d.lustre.it_lock_handle, &plock,
114 it->d.lustre.it_lock_mode = pmode;
117 ldlm_lock_decref(&plock, pmode);
119 ptlrpc_req_finished(*reqp);
125 int lmv_intent_open(struct obd_export *exp, struct lustre_id *pid,
126 const char *name, int len, void *lmm, int lmmsize,
127 struct lustre_id *cid, struct lookup_intent *it,
128 int flags, struct ptlrpc_request **reqp,
129 ldlm_blocking_callback cb_blocking)
131 struct obd_device *obd = exp->exp_obd;
132 struct lmv_obd *lmv = &obd->u.lmv;
133 struct mds_body *body = NULL;
134 struct lustre_id rpid = *pid;
135 int rc, mds, loop = 0;
140 /* IT_OPEN is intended to open (and create, possible) an object. Parent
141 * (pid) may be splitted dir */
144 LASSERT(++loop <= 2);
145 mds = id_group(&rpid);
146 obj = lmv_grab_obj(obd, &rpid);
148 /* directory is already splitted, so we have to forward
149 * request to the right MDS */
150 mds = raw_name2idx(obj->hashtype, obj->objcount,
153 CDEBUG(D_OTHER, "forward to MDS #%u ("DLID4")\n",
155 rpid = obj->objs[mds].id;
159 rc = md_intent_lock(lmv->tgts[id_group(&rpid)].ltd_exp, &rpid, name,
160 len, lmm, lmmsize, cid, it, flags, reqp, cb_blocking);
161 if (rc == -ERESTART) {
162 /* directory got splitted. time to update local object and
163 * repeat the request with proper MDS */
164 LASSERT(id_equal_fid(pid, &rpid));
165 rc = lmv_get_mea_and_update_object(exp, &rpid);
167 ptlrpc_req_finished(*reqp);
174 /* okay, MDS has returned success. Probably name has been resolved in
176 rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it,
177 flags, reqp, cb_blocking);
182 * this is possible, that some userspace application will try to
183 * open file as directory and we will have error -20 here. As
184 * this is "usual" situation, we should not print error here,
187 CDEBUG(D_OTHER, "can't handle remote %s: dir "DLID4"("DLID4"):"
188 "%*s: %d\n", LL_IT2STR(it), OLID4(pid), OLID4(&rpid),
193 /* caller may use attrs MDS returns on IT_OPEN lock request so, we have
194 * to update them for splitted dir */
195 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
196 LASSERT(body != NULL);
199 obj = lmv_grab_obj(obd, cid);
200 if (!obj && (mea = lmv_splitted_dir_body(*reqp, 1))) {
201 /* wow! this is splitted dir, we'd like to handle it */
202 obj = lmv_create_obj(exp, &body->id1, mea);
204 RETURN(PTR_ERR(obj));
208 /* this is splitted dir and we'd want to get attrs */
209 CDEBUG(D_OTHER, "attrs from slaves for "DLID4"\n",
212 rc = lmv_revalidate_slaves(exp, reqp, cid, it, 1,
214 } else if (S_ISDIR(body->mode)) {
215 CDEBUG(D_OTHER, "object "DLID4" has not lmv obj?\n",
225 int lmv_intent_getattr(struct obd_export *exp, struct lustre_id *pid,
226 const char *name, int len, void *lmm, int lmmsize,
227 struct lustre_id *cid, struct lookup_intent *it,
228 int flags, struct ptlrpc_request **reqp,
229 ldlm_blocking_callback cb_blocking)
231 struct obd_device *obd = exp->exp_obd;
232 struct lmv_obd *lmv = &obd->u.lmv;
233 struct mds_body *body = NULL;
234 struct lustre_id rpid = *pid;
235 struct lmv_obj *obj = NULL, *obj2 = NULL;
241 /* caller wants to revalidate attrs of obj we have to revalidate
242 * slaves if requested object is splitted directory */
243 CDEBUG(D_OTHER, "revalidate attrs for "DLID4"\n", OLID4(cid));
246 obj = lmv_grab_obj(obd, cid);
248 /* in fact, we need not this with current intent_lock(),
249 * but it may change some day */
250 if (!id_equal_fid(pid, cid)){
251 rpid = obj->objs[mds].id;
252 mds = id_group(&rpid);
258 CDEBUG(D_OTHER, "INTENT getattr for %*s on "DLID4"\n",
259 len, name, OLID4(pid));
261 obj = lmv_grab_obj(obd, pid);
263 /* directory is already splitted. calculate mds */
264 mds = raw_name2idx(obj->hashtype, obj->objcount,
266 rpid = obj->objs[mds].id;
267 mds = id_group(&rpid);
270 CDEBUG(D_OTHER, "forward to MDS #%u (slave "DLID4")\n",
275 /* the same about fid returning. */
276 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, &rpid, name,
277 len, lmm, lmmsize, cid, it, flags,
283 /* this is splitted dir. In order to optimize things a
284 * bit, we consider obj valid updating missing parts.
286 * FIXME: do we need to return any lock here? It would
287 * be fine if we don't. this means that nobody should
288 * use UPDATE lock to notify about object * removal */
290 "revalidate slaves for "DLID4", rc %d\n",
294 rc = lmv_revalidate_slaves(exp, reqp, cid, it, rc,
302 /* okay, MDS has returned success. probably name has been
303 * resolved in remote inode */
304 rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it,
305 flags, reqp, cb_blocking);
309 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
310 LASSERT(body != NULL);
313 obj2 = lmv_grab_obj(obd, cid);
315 if (!obj2 && (mea = lmv_splitted_dir_body(*reqp, 1))) {
316 /* wow! this is splitted dir, we'd like to handle it. */
317 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
318 LASSERT(body != NULL);
320 obj2 = lmv_create_obj(exp, &body->id1, mea);
322 RETURN(PTR_ERR(obj2));
326 /* this is splitted dir and we'd want to get attrs */
327 CDEBUG(D_OTHER, "attrs from slaves for "DLID4", rc %d\n",
330 rc = lmv_revalidate_slaves(exp, reqp, cid, it, 1,
337 void lmv_update_body_from_obj(struct mds_body *body, struct lmv_inode *obj)
340 body->size += obj->size;
343 int lmv_lookup_slaves(struct obd_export *exp, struct ptlrpc_request **reqp)
345 struct obd_device *obd = exp->exp_obd;
346 struct lmv_obd *lmv = &obd->u.lmv;
347 struct mds_body *body = NULL;
348 struct lustre_handle *lockh;
349 struct ldlm_lock *lock;
350 struct mds_body *body2;
358 /* master is locked. we'd like to take locks on slaves and update
359 * attributes to be returned from the slaves it's important that lookup
360 * is called in two cases:
362 * - for first time (dcache has no such a resolving yet).
363 * - ->d_revalidate() returned false.
365 * last case possible only if all the objs (master and all slaves aren't
368 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
369 LASSERT(body != NULL);
371 obj = lmv_grab_obj(obd, &body->id1);
372 LASSERT(obj != NULL);
374 CDEBUG(D_OTHER, "lookup slaves for "DLID4"\n",
379 for (i = 0; i < obj->objcount; i++) {
380 struct lustre_id id = obj->objs[i].id;
381 struct ptlrpc_request *req = NULL;
382 struct lookup_intent it;
384 if (id_equal_fid(&id, &obj->id))
385 /* skip master obj */
388 CDEBUG(D_OTHER, "lookup slave "DLID4"\n", OLID4(&id));
391 memset(&it, 0, sizeof(it));
392 it.it_op = IT_GETATTR;
393 rc = md_intent_lock(lmv->tgts[id_group(&id)].ltd_exp, &id,
394 NULL, 0, NULL, 0, &id, &it, 0, &req,
395 lmv_dirobj_blocking_ast);
397 lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
398 if (rc > 0 && req == NULL) {
399 /* nice, this slave is valid */
400 LASSERT(req == NULL);
401 CDEBUG(D_OTHER, "cached\n");
406 /* error during lookup */
409 lock = ldlm_handle2lock(lockh);
412 lock->l_ast_data = lmv_get_obj(obj);
414 body2 = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body2));
417 obj->objs[i].size = body2->size;
419 CDEBUG(D_OTHER, "fresh: %lu\n",
420 (unsigned long)obj->objs[i].size);
425 ptlrpc_req_finished(req);
427 lmv_update_body_from_obj(body, obj->objs + i);
429 if (it.d.lustre.it_lock_mode)
430 ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
440 int lmv_intent_lookup(struct obd_export *exp, struct lustre_id *pid,
441 const char *name, int len, void *lmm, int lmmsize,
442 struct lustre_id *cid, struct lookup_intent *it,
443 int flags, struct ptlrpc_request **reqp,
444 ldlm_blocking_callback cb_blocking)
446 struct obd_device *obd = exp->exp_obd;
447 struct lmv_obd *lmv = &obd->u.lmv;
448 struct mds_body *body = NULL;
449 struct lustre_id rpid = *pid;
452 int rc, mds, loop = 0;
456 * IT_LOOKUP is intended to produce name -> id resolving (let's call
457 * this lookup below) or to confirm requested resolving is still valid
458 * (let's call this revalidation) cid != NULL specifies revalidation.
462 * this is revalidation: we have to check is LOOKUP lock still
463 * valid for given id. Very important part is that we have to
464 * choose right mds because namespace is per mds.
467 obj = lmv_grab_obj(obd, pid);
469 mds = raw_name2idx(obj->hashtype, obj->objcount,
471 rpid = obj->objs[mds].id;
474 mds = id_group(&rpid);
476 CDEBUG(D_OTHER, "revalidate lookup for "DLID4" to %d MDS\n",
482 LASSERT(++loop <= 2);
484 /* this is lookup. during lookup we have to update all the
485 * attributes, because returned values will be put in struct
488 obj = lmv_grab_obj(obd, pid);
491 /* directory is already splitted. calculate mds */
492 mds = raw_name2idx(obj->hashtype, obj->objcount,
494 rpid = obj->objs[mds].id;
495 mds = id_group(&rpid);
500 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, &rpid, name,
501 len, lmm, lmmsize, cid, it, flags,
508 /* very interesting. it seems object is still valid but for some
509 * reason llite calls lookup, not revalidate */
510 CDEBUG(D_OTHER, "lookup for "DLID4" and data should be uptodate\n",
512 LASSERT(*reqp == NULL);
516 if (rc == 0 && *reqp == NULL) {
517 /* once again, we're asked for lookup, not revalidate */
518 CDEBUG(D_OTHER, "lookup for "DLID4" and data should be uptodate\n",
523 if (rc == -ERESTART) {
524 /* directory got splitted since last update. this shouldn't be
525 * becasue splitting causes lock revocation, so revalidate had
526 * to fail and lookup on dir had to return mea */
527 CWARN("we haven't knew about directory splitting!\n");
528 LASSERT(obj == NULL);
530 obj = lmv_create_obj(exp, &rpid, NULL);
532 RETURN(PTR_ERR(obj));
540 /* okay, MDS has returned success. Probably name has been resolved in
542 rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it,
543 flags, reqp, cb_blocking);
545 if (rc == 0 && (mea = lmv_splitted_dir_body(*reqp, 1))) {
546 /* wow! this is splitted dir, we'd like to handle it */
547 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
548 LASSERT(body != NULL);
550 obj = lmv_grab_obj(obd, &body->id1);
552 obj = lmv_create_obj(exp, &body->id1, mea);
554 RETURN(PTR_ERR(obj));
562 int lmv_intent_lock(struct obd_export *exp, struct lustre_id *pid,
563 const char *name, int len, void *lmm, int lmmsize,
564 struct lustre_id *cid, struct lookup_intent *it,
565 int flags, struct ptlrpc_request **reqp,
566 ldlm_blocking_callback cb_blocking)
568 struct obd_device *obd = exp->exp_obd;
575 CDEBUG(D_OTHER, "INTENT LOCK '%s' for '%*s' on %lu/%lu -> %lu\n",
576 LL_IT2STR(it), len, name, (unsigned long)id_ino(pid),
577 (unsigned long)id_gen(pid), (unsigned long)id_group(pid));
579 rc = lmv_check_connect(obd);
583 if (it->it_op == IT_LOOKUP)
584 rc = lmv_intent_lookup(exp, pid, name, len, lmm,
585 lmmsize, cid, it, flags, reqp,
587 else if (it->it_op & IT_OPEN)
588 rc = lmv_intent_open(exp, pid, name, len, lmm,
589 lmmsize, cid, it, flags, reqp,
591 else if (it->it_op == IT_GETATTR || it->it_op == IT_CHDIR)
592 rc = lmv_intent_getattr(exp, pid, name, len, lmm,
593 lmmsize, cid, it, flags, reqp,
600 int lmv_revalidate_slaves(struct obd_export *exp, struct ptlrpc_request **reqp,
601 struct lustre_id *mid, struct lookup_intent *oit,
602 int master_valid, ldlm_blocking_callback cb_blocking)
604 struct obd_device *obd = exp->exp_obd;
605 struct ptlrpc_request *mreq = *reqp;
606 struct lmv_obd *lmv = &obd->u.lmv;
607 struct lustre_handle master_lockh;
608 struct ldlm_lock *lock;
609 unsigned long size = 0;
610 struct mds_body *body;
612 int master_lock_mode;
616 /* we have to loop over the subobjects, check validity and update them
617 * from MDSs if needed. it's very useful that we need not to update all
618 * the fields. say, common fields (that are equal on all the subojects
619 * need not to be update, another fields (i_size, for example) are
620 * cached all the time */
621 obj = lmv_grab_obj(obd, mid);
622 LASSERT(obj != NULL);
624 master_lock_mode = 0;
628 for (i = 0; i < obj->objcount; i++) {
629 struct lustre_id id = obj->objs[i].id;
630 struct lustre_handle *lockh = NULL;
631 struct ptlrpc_request *req = NULL;
632 ldlm_blocking_callback cb;
633 struct lookup_intent it;
636 CDEBUG(D_OTHER, "revalidate subobj "DLID4"\n",
639 memset(&it, 0, sizeof(it));
640 it.it_op = IT_GETATTR;
641 cb = lmv_dirobj_blocking_ast;
643 if (id_equal_fid(&id, &obj->id)) {
645 /* lmv_intent_getattr() already checked
646 * validness and took the lock */
648 /* it even got the reply refresh attrs
650 body = lustre_msg_buf(mreq->rq_repmsg,
652 LASSERT(body != NULL);
655 /* take already cached attrs into account */
657 "master is locked and cached\n");
665 rc = md_intent_lock(lmv->tgts[id_group(&id)].ltd_exp,
666 &id, NULL, 0, NULL, 0, &id, &it, 0,
668 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
669 if (rc > 0 && req == NULL) {
670 /* nice, this slave is valid */
671 LASSERT(req == NULL);
672 CDEBUG(D_OTHER, "cached\n");
677 /* error during revalidation */
681 LASSERT(master_valid == 0);
682 /* save lock on master to be returned to the caller */
683 CDEBUG(D_OTHER, "no lock on master yet\n");
684 memcpy(&master_lockh, lockh, sizeof(master_lockh));
685 master_lock_mode = it.d.lustre.it_lock_mode;
686 it.d.lustre.it_lock_mode = 0;
688 /* this is slave. we want to control it */
689 lock = ldlm_handle2lock(lockh);
691 lock->l_ast_data = lmv_get_obj(obj);
696 /* this is first reply, we'll use it to return updated
697 * data back to the caller */
699 ptlrpc_request_addref(req);
704 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
708 obj->objs[i].size = body->size;
710 CDEBUG(D_OTHER, "fresh: %lu\n",
711 (unsigned long)obj->objs[i].size);
714 ptlrpc_req_finished(req);
716 size += obj->objs[i].size;
718 if (it.d.lustre.it_lock_mode)
719 ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
723 /* some attrs got refreshed, we have reply and it's time to put
724 * fresh attrs to it */
725 CDEBUG(D_OTHER, "return refreshed attrs: size = %lu\n",
726 (unsigned long)size);
728 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
731 /* FIXME: what about other attributes? */
735 /* very important to maintain id_group(lli->lli_id) the
736 * same because of revalidation. mreq == NULL means that
737 * caller has no reply and the only attr we can return
739 body->valid = OBD_MD_FLSIZE;
740 // body->mds = id_group(&obj->id);
742 if (master_valid == 0) {
743 memcpy(&oit->d.lustre.it_lock_handle,
744 &master_lockh, sizeof(master_lockh));
745 oit->d.lustre.it_lock_mode = master_lock_mode;
749 /* it seems all the attrs are fresh and we did no request */
750 CDEBUG(D_OTHER, "all the attrs were fresh\n");
751 if (master_valid == 0)
752 oit->d.lustre.it_lock_mode = master_lock_mode;