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,
162 if (rc == -ERESTART) {
163 /* directory got splitted. time to update local object and
164 * repeat the request with proper MDS */
165 LASSERT(id_equal_fid(pid, &rpid));
166 rc = lmv_get_mea_and_update_object(exp, &rpid);
168 ptlrpc_req_finished(*reqp);
175 /* okay, MDS has returned success. Probably name has been resolved in
177 rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it,
178 flags, reqp, cb_blocking);
183 * this is possible, that some userspace application will try to
184 * open file as directory and we will have error -20 here. As
185 * this is "usual" situation, we should not print error here,
188 CDEBUG(D_OTHER, "can't handle remote %s: dir "DLID4"("DLID4"):"
189 "%*s: %d\n", LL_IT2STR(it), OLID4(pid), OLID4(&rpid),
194 /* caller may use attrs MDS returns on IT_OPEN lock request so, we have
195 * to update them for splitted dir */
196 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
197 LASSERT(body != NULL);
200 obj = lmv_grab_obj(obd, cid);
201 if (!obj && (mea = lmv_splitted_dir_body(*reqp, 1))) {
202 /* wow! this is splitted dir, we'd like to handle it */
203 obj = lmv_create_obj(exp, &body->id1, mea);
205 RETURN(PTR_ERR(obj));
209 /* this is splitted dir and we'd want to get attrs */
210 CDEBUG(D_OTHER, "attrs from slaves for "DLID4"\n",
213 rc = lmv_revalidate_slaves(exp, reqp, cid, it, 1,
215 } else if (S_ISDIR(body->mode)) {
216 CDEBUG(D_OTHER, "object "DLID4" has not lmv obj?\n",
226 int lmv_intent_getattr(struct obd_export *exp, struct lustre_id *pid,
227 const char *name, int len, void *lmm, int lmmsize,
228 struct lustre_id *cid, struct lookup_intent *it,
229 int flags, struct ptlrpc_request **reqp,
230 ldlm_blocking_callback cb_blocking)
232 struct obd_device *obd = exp->exp_obd;
233 struct lmv_obd *lmv = &obd->u.lmv;
234 struct mds_body *body = NULL;
235 struct lustre_id rpid = *pid;
236 struct lmv_obj *obj, *obj2;
242 /* caller wants to revalidate attrs of obj we have to revalidate
243 * slaves if requested object is splitted directory */
244 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);
257 CDEBUG(D_OTHER, "INTENT getattr for %*s on "DLID4"\n",
258 len, name, OLID4(pid));
260 obj = lmv_grab_obj(obd, pid);
262 /* directory is already splitted. calculate mds */
263 mds = raw_name2idx(obj->hashtype, obj->objcount,
265 rpid = obj->objs[mds].id;
266 mds = id_group(&rpid);
269 CDEBUG(D_OTHER, "forward to MDS #%u (slave "DLID4")\n",
273 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, &rpid, name,
274 len, lmm, lmmsize, cid, 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 "DLID4", rc %d\n",
291 rc = lmv_revalidate_slaves(exp, reqp, cid, it, rc,
299 /* okay, MDS has returned success. probably name has been
300 * resolved in remote inode */
301 rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it,
302 flags, reqp, cb_blocking);
306 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
307 LASSERT(body != NULL);
310 obj2 = lmv_grab_obj(obd, cid);
312 if (!obj2 && (mea = lmv_splitted_dir_body(*reqp, 1))) {
313 /* wow! this is splitted dir, we'd like to handle it. */
314 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
315 LASSERT(body != NULL);
317 obj2 = lmv_create_obj(exp, &body->id1, mea);
319 RETURN(PTR_ERR(obj2));
323 /* this is splitted dir and we'd want to get attrs */
324 CDEBUG(D_OTHER, "attrs from slaves for "DLID4", rc %d\n",
327 rc = lmv_revalidate_slaves(exp, reqp, cid, it, 1, cb_blocking);
333 void lmv_update_body_from_obj(struct mds_body *body, struct lmv_inode *obj)
336 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;
354 /* master is locked. we'd like to take locks on slaves and update
355 * attributes to be returned from the slaves it's important that lookup
356 * is called in two cases:
358 * - for first time (dcache has no such a resolving yet).
359 * - ->d_revalidate() returned false.
361 * last case possible only if all the objs (master and all slaves aren't
364 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
365 LASSERT(body != NULL);
367 obj = lmv_grab_obj(obd, &body->id1);
368 LASSERT(obj != NULL);
370 CDEBUG(D_OTHER, "lookup slaves for "DLID4"\n",
375 for (i = 0; i < obj->objcount; i++) {
376 struct lustre_id id = obj->objs[i].id;
377 struct ptlrpc_request *req = NULL;
378 struct lookup_intent it;
380 if (id_equal_fid(&id, &obj->id))
381 /* skip master obj */
384 CDEBUG(D_OTHER, "lookup slave "DLID4"\n", OLID4(&id));
387 memset(&it, 0, sizeof(it));
388 it.it_op = IT_GETATTR;
389 rc = md_intent_lock(lmv->tgts[id_group(&id)].ltd_exp, &id,
390 NULL, 0, NULL, 0, &id, &it, 0, &req,
391 lmv_dirobj_blocking_ast);
393 lockh = (struct lustre_handle *)&it.d.lustre.it_lock_handle;
394 if (rc > 0 && req == NULL) {
395 /* nice, this slave is valid */
396 LASSERT(req == NULL);
397 CDEBUG(D_OTHER, "cached\n");
402 /* error during lookup */
405 lock = ldlm_handle2lock(lockh);
408 lock->l_ast_data = lmv_get_obj(obj);
410 body2 = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body2));
413 obj->objs[i].size = body2->size;
415 CDEBUG(D_OTHER, "fresh: %lu\n",
416 (unsigned long)obj->objs[i].size);
421 ptlrpc_req_finished(req);
423 lmv_update_body_from_obj(body, obj->objs + i);
425 if (it.d.lustre.it_lock_mode)
426 ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
434 int lmv_intent_lookup(struct obd_export *exp, struct lustre_id *pid,
435 const char *name, int len, void *lmm, int lmmsize,
436 struct lustre_id *cid, struct lookup_intent *it,
437 int flags, struct ptlrpc_request **reqp,
438 ldlm_blocking_callback cb_blocking)
440 struct obd_device *obd = exp->exp_obd;
441 struct lmv_obd *lmv = &obd->u.lmv;
442 struct mds_body *body = NULL;
443 struct lustre_id rpid = *pid;
446 int rc, mds, loop = 0;
450 * IT_LOOKUP is intended to produce name -> id resolving (let's call
451 * this lookup below) or to confirm requested resolving is still valid
452 * (let's call this revalidation) cid != NULL specifies revalidation.
456 * this is revalidation: we have to check is LOOKUP lock still
457 * valid for given id. Very important part is that we have to
458 * choose right mds because namespace is per mds.
461 obj = lmv_grab_obj(obd, pid);
463 mds = raw_name2idx(obj->hashtype, obj->objcount,
465 rpid = obj->objs[mds].id;
468 mds = id_group(&rpid);
470 CDEBUG(D_OTHER, "revalidate lookup for "DLID4" to %d MDS\n",
476 LASSERT(++loop <= 2);
478 /* this is lookup. during lookup we have to update all the
479 * attributes, because returned values will be put in struct
482 obj = lmv_grab_obj(obd, pid);
485 /* directory is already splitted. calculate mds */
486 mds = raw_name2idx(obj->hashtype, obj->objcount,
488 rpid = obj->objs[mds].id;
489 mds = id_group(&rpid);
494 rc = md_intent_lock(lmv->tgts[mds].ltd_exp, &rpid, name,
495 len, lmm, lmmsize, cid, it, flags, reqp,
502 /* very interesting. it seems object is still valid but for some
503 * reason llite calls lookup, not revalidate */
504 CDEBUG(D_OTHER, "lookup for "DLID4" and data should be uptodate\n",
506 LASSERT(*reqp == NULL);
510 if (rc == 0 && *reqp == NULL) {
511 /* once again, we're asked for lookup, not revalidate */
512 CDEBUG(D_OTHER, "lookup for "DLID4" and data should be uptodate\n",
517 if (rc == -ERESTART) {
518 /* directory got splitted since last update. this shouldn't be
519 * becasue splitting causes lock revocation, so revalidate had
520 * to fail and lookup on dir had to return mea */
521 CWARN("we haven't knew about directory splitting!\n");
522 LASSERT(obj == NULL);
524 obj = lmv_create_obj(exp, &rpid, NULL);
526 RETURN(PTR_ERR(obj));
534 /* okay, MDS has returned success. Probably name has been resolved in
536 rc = lmv_handle_remote_inode(exp, lmm, lmmsize, it, flags,
539 if (rc == 0 && (mea = lmv_splitted_dir_body(*reqp, 1))) {
540 /* wow! this is splitted dir, we'd like to handle it */
541 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
542 LASSERT(body != NULL);
544 obj = lmv_grab_obj(obd, &body->id1);
546 obj = lmv_create_obj(exp, &body->id1, mea);
548 RETURN(PTR_ERR(obj));
556 int lmv_intent_lock(struct obd_export *exp, struct lustre_id *pid,
557 const char *name, int len, void *lmm, int lmmsize,
558 struct lustre_id *cid, struct lookup_intent *it,
559 int flags, struct ptlrpc_request **reqp,
560 ldlm_blocking_callback cb_blocking)
562 struct obd_device *obd = exp->exp_obd;
569 CDEBUG(D_OTHER, "INTENT LOCK '%s' for '%*s' on %lu/%lu -> %lu\n",
570 LL_IT2STR(it), len, name, (unsigned long)id_ino(pid),
571 (unsigned long)id_gen(pid), (unsigned long)id_group(pid));
573 rc = lmv_check_connect(obd);
577 if (it->it_op == IT_LOOKUP)
578 rc = lmv_intent_lookup(exp, pid, name, len, lmm,
579 lmmsize, cid, it, flags, reqp,
581 else if (it->it_op & IT_OPEN)
582 rc = lmv_intent_open(exp, pid, name, len, lmm,
583 lmmsize, cid, it, flags, reqp,
585 else if (it->it_op == IT_GETATTR || it->it_op == IT_CHDIR)
586 rc = lmv_intent_getattr(exp, pid, name, len, lmm,
587 lmmsize, cid, it, flags, reqp,
594 int lmv_revalidate_slaves(struct obd_export *exp, struct ptlrpc_request **reqp,
595 struct lustre_id *mid, struct lookup_intent *oit,
596 int master_valid, ldlm_blocking_callback cb_blocking)
598 struct obd_device *obd = exp->exp_obd;
599 struct ptlrpc_request *mreq = *reqp;
600 struct lmv_obd *lmv = &obd->u.lmv;
601 struct lustre_handle master_lockh;
602 struct ldlm_lock *lock;
603 unsigned long size = 0;
604 struct mds_body *body;
606 int master_lock_mode;
610 /* we have to loop over the subobjects, check validity and update them
611 * from MDSs if needed. it's very useful that we need not to update all
612 * the fields. say, common fields (that are equal on all the subojects
613 * need not to be update, another fields (i_size, for example) are
614 * cached all the time */
615 obj = lmv_grab_obj(obd, mid);
616 LASSERT(obj != NULL);
618 master_lock_mode = 0;
622 for (i = 0; i < obj->objcount; i++) {
623 struct lustre_id id = obj->objs[i].id;
624 struct lustre_handle *lockh = NULL;
625 struct ptlrpc_request *req = NULL;
626 ldlm_blocking_callback cb;
627 struct lookup_intent it;
630 CDEBUG(D_OTHER, "revalidate subobj "DLID4"\n",
633 memset(&it, 0, sizeof(it));
634 it.it_op = IT_GETATTR;
635 cb = lmv_dirobj_blocking_ast;
637 if (id_equal_fid(&id, &obj->id)) {
639 /* lmv_intent_getattr() already checked
640 * validness and took the lock */
642 /* it even got the reply refresh attrs
644 body = lustre_msg_buf(mreq->rq_repmsg,
646 LASSERT(body != NULL);
649 /* take already cached attrs into account */
651 "master is locked and cached\n");
659 rc = md_intent_lock(lmv->tgts[id_group(&id)].ltd_exp,
660 &id, NULL, 0, NULL, 0, &id, &it, 0,
662 lockh = (struct lustre_handle *) &it.d.lustre.it_lock_handle;
663 if (rc > 0 && req == NULL) {
664 /* nice, this slave is valid */
665 LASSERT(req == NULL);
666 CDEBUG(D_OTHER, "cached\n");
671 /* error during revalidation */
675 LASSERT(master_valid == 0);
676 /* save lock on master to be returned to the caller */
677 CDEBUG(D_OTHER, "no lock on master yet\n");
678 memcpy(&master_lockh, lockh, sizeof(master_lockh));
679 master_lock_mode = it.d.lustre.it_lock_mode;
680 it.d.lustre.it_lock_mode = 0;
682 /* this is slave. we want to control it */
683 lock = ldlm_handle2lock(lockh);
685 lock->l_ast_data = lmv_get_obj(obj);
690 /* this is first reply, we'll use it to return updated
691 * data back to the caller */
693 ptlrpc_request_addref(req);
698 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof(*body));
702 obj->objs[i].size = body->size;
704 CDEBUG(D_OTHER, "fresh: %lu\n",
705 (unsigned long)obj->objs[i].size);
708 ptlrpc_req_finished(req);
710 size += obj->objs[i].size;
712 if (it.d.lustre.it_lock_mode)
713 ldlm_lock_decref(lockh, it.d.lustre.it_lock_mode);
717 /* some attrs got refreshed, we have reply and it's time to put
718 * fresh attrs to it */
719 CDEBUG(D_OTHER, "return refreshed attrs: size = %lu\n",
720 (unsigned long)size);
722 body = lustre_msg_buf((*reqp)->rq_repmsg, 1, sizeof(*body));
725 /* FIXME: what about other attributes? */
729 /* very important to maintain id_group(lli->lli_id) the
730 * same because of revalidation. mreq == NULL means that
731 * caller has no reply and the only attr we can return
733 body->valid = OBD_MD_FLSIZE;
734 // body->mds = id_group(&obj->id);
736 if (master_valid == 0) {
737 memcpy(&oit->d.lustre.it_lock_handle,
738 &master_lockh, sizeof(master_lockh));
739 oit->d.lustre.it_lock_mode = master_lock_mode;
743 /* it seems all the attrs are fresh and we did no request */
744 CDEBUG(D_OTHER, "all the attrs were fresh\n");
745 if (master_valid == 0)
746 oit->d.lustre.it_lock_mode = master_lock_mode;