1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2001-2003 Cluster File Systems, Inc.
6 * This file is part of Lustre, http://www.sf.net/projects/lustre/
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_MDC
28 # include <linux/module.h>
29 # include <linux/pagemap.h>
30 # include <linux/miscdevice.h>
31 # include <linux/init.h>
33 # include <liblustre.h>
36 #include <linux/obd_class.h>
37 #include <linux/lustre_mds.h>
38 #include <linux/lustre_dlm.h>
39 #include <linux/lustre_sec.h>
40 #include <linux/lprocfs_status.h>
41 #include <linux/lustre_acl.h>
42 #include <linux/lustre_lite.h>
43 #include "mdc_internal.h"
45 int it_disposition(struct lookup_intent *it, int flag)
47 return LUSTRE_IT(it)->it_disposition & flag;
49 EXPORT_SYMBOL(it_disposition);
51 void it_set_disposition(struct lookup_intent *it, int flag)
53 LUSTRE_IT(it)->it_disposition |= flag;
55 EXPORT_SYMBOL(it_set_disposition);
57 static void mdc_id2mdc_data(struct mdc_op_data *data,
61 int namelen, int mode)
72 data->namelen = namelen;
73 data->create_mode = mode;
74 data->mod_time = LTIME_S(CURRENT_TIME);
77 static int it_to_lock_mode(struct lookup_intent *it)
79 /* CREAT needs to be tested before open (both could be set) */
80 if (it->it_op & IT_CREAT)
82 else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP |
90 int it_open_error(int phase, struct lookup_intent *it)
92 if (it_disposition(it, DISP_OPEN_OPEN)) {
93 if (phase == DISP_OPEN_OPEN)
94 return LUSTRE_IT(it)->it_status;
99 if (it_disposition(it, DISP_OPEN_CREATE)) {
100 if (phase == DISP_OPEN_CREATE)
101 return LUSTRE_IT(it)->it_status;
106 if (it_disposition(it, DISP_LOOKUP_EXECD)) {
107 if (phase == DISP_LOOKUP_EXECD)
108 return LUSTRE_IT(it)->it_status;
113 if (it_disposition(it, DISP_IT_EXECD)) {
114 if (phase == DISP_IT_EXECD)
115 return LUSTRE_IT(it)->it_status;
119 CERROR("it disp: %X, status: %d\n", LUSTRE_IT(it)->it_disposition,
120 LUSTRE_IT(it)->it_status);
124 EXPORT_SYMBOL(it_open_error);
126 /* this must be called on a lockh that is known to have a referenced lock */
127 int mdc_set_lock_data(struct obd_export *exp, __u64 *l, void *data)
129 struct ldlm_lock *lock;
130 struct lustre_handle *lockh = (struct lustre_handle *)l;
138 lock = ldlm_handle2lock(lockh);
140 LASSERT(lock != NULL);
141 lock_res_and_lock(lock);
143 if (lock->l_ast_data && lock->l_ast_data != data) {
144 struct inode *new_inode = data;
145 struct inode *old_inode = lock->l_ast_data;
146 if (!(old_inode->i_state & I_FREEING)) {
147 CERROR("Found existing inode %p/%lu/%u state %lu "
148 "in lock: setting data to %p/%lu/%u\n",
149 old_inode, old_inode->i_ino,
150 old_inode->i_generation, old_inode->i_state,
151 new_inode, new_inode->i_ino,
152 new_inode->i_generation);
153 unlock_res_and_lock(lock);
158 lock->l_ast_data = data;
159 unlock_res_and_lock(lock);
165 EXPORT_SYMBOL(mdc_set_lock_data);
167 int mdc_change_cbdata(struct obd_export *exp, struct lustre_id *id,
168 ldlm_iterator_t it, void *data)
170 struct ldlm_res_id res_id = { .name = {0} };
173 res_id.name[0] = id_fid(id);
174 res_id.name[1] = id_group(id);
176 ldlm_change_cbdata(class_exp2obd(exp)->obd_namespace,
184 mdc_clear_replay_flag(struct ptlrpc_request *req, int rc)
186 /* Don't hold error requests for replay. */
187 if (req->rq_replay) {
188 unsigned long irqflags;
189 spin_lock_irqsave(&req->rq_lock, irqflags);
191 spin_unlock_irqrestore(&req->rq_lock, irqflags);
193 if (rc && req->rq_transno != 0) {
194 DEBUG_REQ(D_ERROR, req, "transno returned on error rc %d", rc);
199 /* We always reserve enough space in the reply packet for a stripe MD, because
200 * we don't know in advance the file type. */
201 int mdc_enqueue(struct obd_export *exp,
203 struct lookup_intent *it,
205 struct mdc_op_data *data,
206 struct lustre_handle *lockh,
209 ldlm_completion_callback cb_completion,
210 ldlm_blocking_callback cb_blocking,
213 struct ptlrpc_request *req;
214 struct ldlm_res_id res_id = {
215 .name = {id_fid(&data->id1), id_group(&data->id1)}
217 struct obd_device *obddev = class_exp2obd(exp);
218 ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
219 struct ldlm_intent *lit;
220 struct ldlm_request *lockreq;
221 int reqsize[6] = {[MDS_REQ_SECDESC_OFF] = 0,
222 [MDS_REQ_INTENT_LOCKREQ_OFF] = sizeof(*lockreq),
223 [MDS_REQ_INTENT_IT_OFF] = sizeof(*lit)};
224 int repsize[5] = {sizeof(struct ldlm_reply),
225 sizeof(struct mds_body),
226 obddev->u.cli.cl_max_mds_easize};
227 int req_buffers = 3, reply_buffers = 0;
228 int rc, flags = LDLM_FL_HAS_INTENT;
229 struct ldlm_reply *dlm_rep = NULL;
231 unsigned long irqflags;
234 // LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
235 // ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
237 reqsize[0] = lustre_secdesc_size();
239 if (it->it_op & IT_OPEN) {
240 it->it_create_mode |= S_IFREG;
241 it->it_create_mode &= ~current->fs->umask;
243 reqsize[req_buffers++] = sizeof(struct mds_rec_create);
244 reqsize[req_buffers++] = data->namelen + 1;
245 reqsize[req_buffers++] = obddev->u.cli.cl_max_mds_easize;
247 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
248 LDLM_ENQUEUE, req_buffers, reqsize, NULL);
252 spin_lock_irqsave (&req->rq_lock, irqflags);
254 spin_unlock_irqrestore (&req->rq_lock, irqflags);
256 /* pack the intent */
257 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
259 lit->opc = (__u64)it->it_op;
261 /* pack the intended request */
262 mdc_open_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF, data,
263 it->it_create_mode, 0, it->it_flags, lmm, lmmsize);
264 /* get ready for the reply */
266 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
268 req->rq_replen = lustre_msg_size(5, repsize);
269 } else if (it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR)) {
270 __u64 valid = data->valid | OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE |
273 /* we don't expect xattr retrieve could reach here */
274 LASSERT(!(valid & (OBD_MD_FLXATTR | OBD_MD_FLXATTRLIST)));
276 reqsize[req_buffers++] = sizeof(struct mds_body);
277 reqsize[req_buffers++] = data->namelen + 1;
279 if (it->it_op & IT_GETATTR)
280 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
282 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
283 LDLM_ENQUEUE, req_buffers, reqsize, NULL);
288 /* pack the intent */
289 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
291 lit->opc = (__u64)it->it_op;
293 /* pack the intended request */
294 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
295 valid, it->it_flags, data);
297 /* get ready for the reply */
299 repsize[4] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
301 req->rq_replen = lustre_msg_size(5, repsize);
302 } else if (it->it_op == IT_READDIR) {
303 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
304 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
305 LDLM_ENQUEUE, 2, reqsize, NULL);
309 /* get ready for the reply */
311 req->rq_replen = lustre_msg_size(1, repsize);
312 } else if (it->it_op == IT_UNLINK) {
313 reqsize[req_buffers++] = sizeof(struct mds_body);
314 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
315 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
316 LDLM_ENQUEUE, req_buffers, reqsize, NULL);
320 /* pack the intended request */
321 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
324 /* pack the intent */
325 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
327 lit->opc = (__u64)it->it_op;
329 /* get ready for the reply */
331 req->rq_replen = lustre_msg_size(3, repsize);
337 lustre_pack_secdesc(req, reqsize[0]);
339 mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
340 rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
341 lock_type, &policy, lock_mode, &flags,cb_blocking,
342 cb_completion, NULL, cb_data, NULL, 0, NULL,
344 mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
346 /* Similarly, if we're going to replay this request, we don't want to
347 * actually get a lock, just perform the intent. */
348 if (req->rq_transno || req->rq_replay) {
349 lockreq = lustre_msg_buf(req->rq_reqmsg,
350 MDS_REQ_INTENT_LOCKREQ_OFF,
352 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
355 /* This can go when we're sure that this can never happen */
356 LASSERT(rc != -ENOENT);
357 /* We need dlm_rep to be assigned this early, to check lock mode of
358 returned lock from request to avoid possible race with lock
360 if (rc == ELDLM_LOCK_ABORTED || !rc) {
361 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
362 LASSERT(dlm_rep != NULL); /* checked by ldlm_cli_enqueue() */
364 if (rc == ELDLM_LOCK_ABORTED) {
366 memset(lockh, 0, sizeof(*lockh));
368 } else if (rc != 0) {
369 CERROR("ldlm_cli_enqueue: %d\n", rc);
370 LASSERTF(rc < 0, "rc = %d\n", rc);
371 mdc_clear_replay_flag(req, rc);
372 ptlrpc_req_finished(req);
374 } else { /* rc = 0 */
375 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
378 /* If the server gave us back a different lock mode, we should
379 * fix up our variables. */
380 if (dlm_rep->lock_desc.l_req_mode != lock_mode) {
381 ldlm_lock_addref(lockh, dlm_rep->lock_desc.l_req_mode);
382 ldlm_lock_decref(lockh, lock_mode);
383 lock_mode = dlm_rep->lock_desc.l_req_mode;
386 ldlm_lock_allow_match(lock);
390 LASSERT_REPSWABBED(req, 0); /* swabbed by ldlm_cli_enqueue() */
392 LUSTRE_IT(it)->it_disposition = (int) dlm_rep->lock_policy_res1;
393 LUSTRE_IT(it)->it_status = (int) dlm_rep->lock_policy_res2;
394 LUSTRE_IT(it)->it_lock_mode = lock_mode;
395 LUSTRE_IT(it)->it_data = req;
397 if (LUSTRE_IT(it)->it_status < 0 && req->rq_replay)
398 mdc_clear_replay_flag(req, LUSTRE_IT(it)->it_status);
400 DEBUG_REQ(D_DLMTRACE, req, "disposition: %x, status: %d",
401 LUSTRE_IT(it)->it_disposition, LUSTRE_IT(it)->it_status);
403 /* We know what to expect, so we do any byte flipping required here */
404 LASSERT(reply_buffers == 5 || reply_buffers == 4 ||
405 reply_buffers == 3 || reply_buffers == 1);
406 if (reply_buffers >= 3) {
407 struct mds_body *body;
409 body = lustre_swab_repbuf(req, 1, sizeof (*body),
410 lustre_swab_mds_body);
412 CERROR ("Can't swab mds_body\n");
416 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
417 /* The eadata is opaque; just check that it is there.
418 * Eventually, obd_unpackmd() will check the contents */
419 eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
421 if (eadata == NULL) {
422 CERROR ("Missing/short eadata\n");
425 if (it->it_op & IT_OPEN) {
428 replayea = lustre_msg_buf(req->rq_reqmsg,
429 MDS_REQ_INTENT_REC_OFF + 2,
432 memcpy(replayea, eadata, body->eadatasize);
434 LASSERT(req->rq_reqmsg->bufcount == 6);
435 req->rq_reqmsg->buflens[5] = body->eadatasize;
436 /* If this isn't the last buffer, we might
437 * have to shift other data around. */
444 EXPORT_SYMBOL(mdc_enqueue);
447 * This long block is all about fixing up the lock and request state
448 * so that it is correct as of the moment _before_ the operation was
449 * applied; that way, the VFS will think that everything is normal and
450 * call Lustre's regular VFS methods.
452 * If we're performing a creation, that means that unless the creation
453 * failed with EEXIST, we should fake up a negative dentry.
455 * For everything else, we want to lookup to succeed.
457 * One additional note: if CREATE or OPEN succeeded, we add an extra
458 * reference to the request because we need to keep it around until
459 * ll_create/ll_open gets called.
461 * The server will return to us, in it_disposition, an indication of
462 * exactly what d.lustre->it_status refers to.
464 * If DISP_OPEN_OPEN is set, then d.lustre->it_status refers to the open() call,
465 * otherwise if DISP_OPEN_CREATE is set, then it status is the
466 * creation failure mode. In either case, one of DISP_LOOKUP_NEG or
467 * DISP_LOOKUP_POS will be set, indicating whether the child lookup
470 * Else, if DISP_LOOKUP_EXECD then d.lustre->it_status is the rc of the
473 int mdc_intent_lock(struct obd_export *exp, struct lustre_id *pid,
474 const char *name, int len, void *lmm, int lmmsize,
475 struct lustre_id *cid, struct lookup_intent *it,
476 int lookup_flags, struct ptlrpc_request **reqp,
477 ldlm_blocking_callback cb_blocking)
479 struct lustre_handle lockh;
480 struct ptlrpc_request *request;
481 struct mds_body *mds_body;
482 struct lustre_handle old_lock;
483 struct ldlm_lock *lock;
488 CDEBUG(D_DLMTRACE, "name: %*s in obj "DLID4", intent: %s flags %#o\n",
489 len, name, OLID4(pid), ldlm_it2str(it->it_op), it->it_flags);
491 if (cid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
492 it->it_op == IT_CHDIR)) {
493 /* We could just return 1 immediately, but since we should only
494 * be called in revalidate_it if we already have a lock, let's
496 struct ldlm_res_id res_id = {.name = {id_fid(cid),
498 struct lustre_handle lockh;
499 ldlm_policy_data_t policy;
502 /* For the GETATTR case, ll_revalidate_it issues two separate
503 queries - for LOOKUP and for UPDATE lock because it cannot
504 check them together - we might have those two bits to be
505 present in two separate granted locks */
506 policy.l_inodebits.bits = (it->it_op == IT_GETATTR) ?
507 MDS_INODELOCK_UPDATE : MDS_INODELOCK_LOOKUP;
510 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
511 LDLM_FL_BLOCK_GRANTED, &res_id,
512 LDLM_IBITS, &policy, mode,
517 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
518 LDLM_FL_BLOCK_GRANTED, &res_id,
519 LDLM_IBITS, &policy, mode,
524 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
525 LDLM_FL_BLOCK_GRANTED, &res_id,
526 LDLM_IBITS, &policy, mode,
531 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
532 LDLM_FL_BLOCK_GRANTED, &res_id,
533 LDLM_IBITS, &policy, mode,
537 if (ptlrpcs_check_cred(exp->exp_obd->u.cli.cl_import)) {
538 /* return immediately if no credential held */
539 ldlm_lock_decref(&lockh, mode);
542 memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh,
544 LUSTRE_IT(it)->it_lock_mode = mode;
547 /* Only return failure if it was not GETATTR by cid (from
553 /* lookup_it may be called only after revalidate_it has run, because
554 * revalidate_it cannot return errors, only zero. Returning zero causes
555 * this call to lookup, which *can* return an error.
557 * We only want to execute the request associated with the intent one
558 * time, however, so don't send the request again. Instead, skip past
559 * this and use the request from revalidate. In this case, revalidate
560 * never dropped its reference, so the refcounts are all OK */
561 if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
562 struct mdc_op_data *op_data;
564 OBD_ALLOC(op_data, sizeof(*op_data));
568 mdc_id2mdc_data(op_data, pid, cid, name, len, 0);
571 op_data->valid |= OBD_MD_FID;
573 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
574 op_data, &lockh, lmm, lmmsize,
575 ldlm_completion_ast, cb_blocking, NULL);
576 OBD_FREE(op_data, sizeof(*op_data));
580 memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh, sizeof(lockh));
582 request = *reqp = LUSTRE_IT(it)->it_data;
583 LASSERT(request != NULL);
585 /* If we're doing an IT_OPEN which did not result in an actual
586 * successful open, then we need to remove the bit which saves this
587 * request for unconditional replay.
589 * It's important that we do this first! Otherwise we might exit the
590 * function without doing so, and try to replay a failed create (bug
592 if (it->it_op & IT_OPEN && request->rq_replay &&
593 (!it_disposition(it, DISP_OPEN_OPEN) || LUSTRE_IT(it)->it_status != 0))
594 mdc_clear_replay_flag(request, LUSTRE_IT(it)->it_status);
596 if (!it_disposition(it, DISP_IT_EXECD)) {
597 /* The server failed before it even started executing the
598 * intent, i.e. because it couldn't unpack the request. */
599 LASSERT(LUSTRE_IT(it)->it_status != 0);
600 RETURN(LUSTRE_IT(it)->it_status);
602 rc = it_open_error(DISP_IT_EXECD, it);
606 mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
607 LASSERT(mds_body != NULL); /* mdc_enqueue checked */
608 LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
610 /* If we were revalidating a fid/name pair, mark the intent in case we
611 * fail and get called again from lookup */
613 it_set_disposition(it, DISP_ENQ_COMPLETE);
614 /* Also: did we find the same inode? */
616 /* we have to compare all the fields but type, because MDS can
617 * return fid/mds/ino/gen if inode lives on another MDS -bzzz */
618 if (!(lookup_flags & LOOKUP_COBD) && !id_equal(cid, &mds_body->id1))
622 rc = it_open_error(DISP_LOOKUP_EXECD, it);
627 * keep requests around for the multiple phases of the call this shows
628 * the DISP_XX must guarantee we make it into the call.
630 if (it_disposition(it, DISP_OPEN_CREATE) &&
631 !it_open_error(DISP_OPEN_CREATE, it))
632 ptlrpc_request_addref(request); /* balanced in ll_create_node */
633 if (it_disposition(it, DISP_OPEN_OPEN) &&
634 !it_open_error(DISP_OPEN_OPEN, it))
635 ptlrpc_request_addref(request); /* balanced in ll_file_open */
637 if (it->it_op & IT_CREAT) {
638 /* XXX this belongs in ll_create_it */
639 } else if (it->it_op == IT_OPEN) {
640 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
642 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR));
646 * if we already have a matching lock, then cancel the new one. We have
647 * to set the data here instead of in mdc_enqueue, because we need to
648 * use the child's inode as the l_ast_data to match, and that's not
649 * available until intent_finish has performed the iget().)
651 lock = ldlm_handle2lock(&lockh);
653 ldlm_policy_data_t policy = lock->l_policy_data;
654 LDLM_DEBUG(lock, "matching against this");
657 LASSERTF(id_fid(&mds_body->id1) == lock->l_resource->lr_name.name[0] &&
658 id_group(&mds_body->id1) == lock->l_resource->lr_name.name[1],
659 "Invalid lock is returned to client. Lock res_is: %lu/%lu, "
660 "response res_id: %lu/%lu.\n",
661 (unsigned long)lock->l_resource->lr_name.name[0],
662 (unsigned long)lock->l_resource->lr_name.name[1],
663 (unsigned long)id_fid(&mds_body->id1),
664 (unsigned long)id_group(&mds_body->id1));
666 memcpy(&old_lock, &lockh, sizeof(lockh));
667 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
668 LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
669 ldlm_lock_decref_and_cancel(&lockh,
670 LUSTRE_IT(it)->it_lock_mode);
671 memcpy(&lockh, &old_lock, sizeof(old_lock));
672 memcpy(&LUSTRE_IT(it)->it_lock_handle, &lockh,
676 CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
677 len, name, ldlm_it2str(it->it_op), LUSTRE_IT(it)->it_status,
678 LUSTRE_IT(it)->it_disposition, rc);
682 EXPORT_SYMBOL(mdc_intent_lock);