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_smfs.h>
40 //#include <linux/lustre_snap.h>
41 #include <linux/lprocfs_status.h>
42 #include "mdc_internal.h"
44 int it_disposition(struct lookup_intent *it, int flag)
46 return it->d.lustre.it_disposition & flag;
48 EXPORT_SYMBOL(it_disposition);
50 void it_set_disposition(struct lookup_intent *it, int flag)
52 it->d.lustre.it_disposition |= flag;
54 EXPORT_SYMBOL(it_set_disposition);
56 static void mdc_fid2mdc_op_data(struct mdc_op_data *data,
57 struct ll_fid *f1, struct ll_fid *f2,
58 const char *name, int namelen, int mode)
67 memset(&data->fid2, 0, sizeof(data->fid2));
69 data->namelen = namelen;
70 data->create_mode = mode;
71 data->mod_time = LTIME_S(CURRENT_TIME);
74 static int it_to_lock_mode(struct lookup_intent *it)
76 /* CREAT needs to be tested before open (both could be set) */
77 if (it->it_op & IT_CREAT)
79 else if (it->it_op & (IT_READDIR | IT_GETATTR | IT_OPEN | IT_LOOKUP |
87 int it_open_error(int phase, struct lookup_intent *it)
89 if (it_disposition(it, DISP_OPEN_OPEN)) {
90 if (phase == DISP_OPEN_OPEN)
91 return it->d.lustre.it_status;
96 if (it_disposition(it, DISP_OPEN_CREATE)) {
97 if (phase == DISP_OPEN_CREATE)
98 return it->d.lustre.it_status;
103 if (it_disposition(it, DISP_LOOKUP_EXECD)) {
104 if (phase == DISP_LOOKUP_EXECD)
105 return it->d.lustre.it_status;
110 if (it_disposition(it, DISP_IT_EXECD)) {
111 if (phase == DISP_IT_EXECD)
112 return it->d.lustre.it_status;
116 CERROR("it disp: %X, status: %d\n", it->d.lustre.it_disposition,
117 it->d.lustre.it_status);
121 EXPORT_SYMBOL(it_open_error);
123 /* this must be called on a lockh that is known to have a referenced lock */
124 int mdc_set_lock_data(struct obd_export *exp, __u64 *l, void *data)
126 struct ldlm_lock *lock;
127 struct lustre_handle *lockh = (struct lustre_handle *)l;
135 lock = ldlm_handle2lock(lockh);
137 LASSERT(lock != NULL);
138 l_lock(&lock->l_resource->lr_namespace->ns_lock);
140 if (lock->l_ast_data && lock->l_ast_data != data) {
141 struct inode *new_inode = data;
142 struct inode *old_inode = lock->l_ast_data;
143 LASSERTF(old_inode->i_state & I_FREEING,
144 "Found existing inode %p/%lu/%u state %lu in lock: "
145 "setting data to %p/%lu/%u\n", old_inode,
146 old_inode->i_ino, old_inode->i_generation,
148 new_inode, new_inode->i_ino, new_inode->i_generation);
151 lock->l_ast_data = data;
152 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
158 EXPORT_SYMBOL(mdc_set_lock_data);
160 int mdc_change_cbdata(struct obd_export *exp, struct ll_fid *fid,
161 ldlm_iterator_t it, void *data)
163 struct ldlm_res_id res_id = { .name = {0} };
166 res_id.name[0] = fid->id;
167 res_id.name[1] = fid->generation;
169 ldlm_change_cbdata(class_exp2obd(exp)->obd_namespace, &res_id, it,
176 /* We always reserve enough space in the reply packet for a stripe MD, because
177 * we don't know in advance the file type. */
178 int mdc_enqueue(struct obd_export *exp,
180 struct lookup_intent *it,
182 struct mdc_op_data *data,
183 struct lustre_handle *lockh,
186 ldlm_completion_callback cb_completion,
187 ldlm_blocking_callback cb_blocking,
190 struct ptlrpc_request *req;
191 struct obd_device *obddev = class_exp2obd(exp);
192 struct ldlm_res_id res_id =
193 { .name = {data->fid1.id, data->fid1.generation} };
194 ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
195 struct ldlm_intent *lit;
196 struct ldlm_request *lockreq;
197 struct ldlm_reply *dlm_rep;
198 int reqsize[6] = {[MDS_REQ_SECDESC_OFF] = 0,
199 [MDS_REQ_INTENT_LOCKREQ_OFF] = sizeof(*lockreq),
200 [MDS_REQ_INTENT_IT_OFF] = sizeof(*lit)};
201 int repsize[4] = {sizeof(struct ldlm_reply),
202 sizeof(struct mds_body),
203 obddev->u.cli.cl_max_mds_easize,
204 obddev->u.cli.cl_max_mds_cookiesize};
205 int req_buffers = 3, reply_buffers = 0;
206 int rc, flags = LDLM_FL_HAS_INTENT;
208 unsigned long irqflags;
211 // LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
212 // ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
214 reqsize[0] = mdc_get_secdesc_size();
216 if (it->it_op & IT_OPEN) {
217 it->it_create_mode |= S_IFREG;
218 it->it_create_mode &= ~current->fs->umask;
220 reqsize[req_buffers++] = sizeof(struct mds_rec_create);
221 reqsize[req_buffers++] = data->namelen + 1;
222 reqsize[req_buffers++] = obddev->u.cli.cl_max_mds_easize;
223 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
224 LDLM_ENQUEUE, req_buffers, reqsize, NULL);
228 spin_lock_irqsave (&req->rq_lock, irqflags);
230 spin_unlock_irqrestore (&req->rq_lock, irqflags);
232 /* pack the intent */
233 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
235 lit->opc = (__u64)it->it_op;
237 /* pack the intended request */
238 mdc_open_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF, data,
239 it->it_create_mode, 0, it->it_flags,
241 /* get ready for the reply */
243 req->rq_replen = lustre_msg_size(3, repsize);
244 } else if (it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR)) {
245 int valid = OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE;
247 reqsize[req_buffers++] = sizeof(struct mds_body);
248 reqsize[req_buffers++] = data->namelen + 1;
250 if (it->it_op & IT_GETATTR)
251 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
252 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
253 LDLM_ENQUEUE, req_buffers, reqsize, NULL);
258 /* pack the intent */
259 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
261 lit->opc = (__u64)it->it_op;
263 /* pack the intended request */
264 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
265 valid, it->it_flags, data);
266 /* get ready for the reply */
268 req->rq_replen = lustre_msg_size(3, repsize);
269 } else if (it->it_op == IT_READDIR) {
270 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
271 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
272 LDLM_ENQUEUE, 2, reqsize, NULL);
276 /* get ready for the reply */
278 req->rq_replen = lustre_msg_size(1, repsize);
279 } else if (it->it_op == IT_UNLINK) {
280 reqsize[req_buffers++] = sizeof(struct mds_body);
281 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);
287 /* pack the intended request */
288 mdc_getattr_pack(req->rq_reqmsg, MDS_REQ_INTENT_REC_OFF,
291 /* pack the intent */
292 lit = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_INTENT_IT_OFF,
294 lit->opc = (__u64)it->it_op;
296 /* get ready for the reply */
298 req->rq_replen = lustre_msg_size(3, repsize);
304 mdc_pack_secdesc(req, reqsize[0]);
306 mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
307 rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
308 lock_type, &policy, lock_mode, &flags,cb_blocking,
309 cb_completion, NULL, cb_data, NULL, 0, NULL,
311 mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
313 /* Similarly, if we're going to replay this request, we don't want to
314 * actually get a lock, just perform the intent. */
315 if (req->rq_transno || req->rq_replay) {
316 lockreq = lustre_msg_buf(req->rq_reqmsg,
317 MDS_REQ_INTENT_LOCKREQ_OFF,
319 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
322 /* This can go when we're sure that this can never happen */
323 LASSERT(rc != -ENOENT);
324 if (rc == ELDLM_LOCK_ABORTED) {
326 memset(lockh, 0, sizeof(*lockh));
328 } else if (rc != 0) {
329 CERROR("ldlm_cli_enqueue: %d\n", rc);
331 ptlrpc_req_finished(req);
333 } else { /* rc = 0 */
334 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
337 /* If the server gave us back a different lock mode, we should
338 * fix up our variables. */
339 if (lock->l_req_mode != lock_mode) {
340 ldlm_lock_addref(lockh, lock->l_req_mode);
341 ldlm_lock_decref(lockh, lock_mode);
342 lock_mode = lock->l_req_mode;
345 ldlm_lock_allow_match(lock);
349 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
350 LASSERT(dlm_rep != NULL); /* checked by ldlm_cli_enqueue() */
351 LASSERT_REPSWABBED(req, 0); /* swabbed by ldlm_cli_enqueue() */
353 it->d.lustre.it_disposition = (int) dlm_rep->lock_policy_res1;
354 it->d.lustre.it_status = (int) dlm_rep->lock_policy_res2;
355 it->d.lustre.it_lock_mode = lock_mode;
356 it->d.lustre.it_data = req;
358 if (it->d.lustre.it_status < 0 && req->rq_replay) {
359 LASSERT(req->rq_transno == 0);
360 /* Don't hold error requests for replay. */
361 spin_lock(&req->rq_lock);
363 spin_unlock(&req->rq_lock);
366 DEBUG_REQ(D_RPCTRACE, req, "disposition: %x, status: %d",
367 it->d.lustre.it_disposition, it->d.lustre.it_status);
369 /* We know what to expect, so we do any byte flipping required here */
370 LASSERT(reply_buffers == 4 || reply_buffers == 3 || reply_buffers == 1);
371 if (reply_buffers >= 3) {
372 struct mds_body *body;
374 body = lustre_swab_repbuf(req, 1, sizeof (*body),
375 lustre_swab_mds_body);
377 CERROR ("Can't swab mds_body\n");
381 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
382 /* The eadata is opaque; just check that it is there.
383 * Eventually, obd_unpackmd() will check the contents */
384 eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
386 if (eadata == NULL) {
387 CERROR ("Missing/short eadata\n");
390 if (it->it_op & IT_OPEN) {
393 replayea = lustre_msg_buf(req->rq_reqmsg,
394 MDS_REQ_INTENT_REC_OFF + 2,
397 memcpy(replayea, eadata, body->eadatasize);
399 LASSERT(req->rq_reqmsg->bufcount == 6);
400 req->rq_reqmsg->buflens[5] = body->eadatasize;
401 /* If this isn't the last buffer, we might
402 * have to shift other data around. */
409 EXPORT_SYMBOL(mdc_enqueue);
412 * This long block is all about fixing up the lock and request state
413 * so that it is correct as of the moment _before_ the operation was
414 * applied; that way, the VFS will think that everything is normal and
415 * call Lustre's regular VFS methods.
417 * If we're performing a creation, that means that unless the creation
418 * failed with EEXIST, we should fake up a negative dentry.
420 * For everything else, we want to lookup to succeed.
422 * One additional note: if CREATE or OPEN succeeded, we add an extra
423 * reference to the request because we need to keep it around until
424 * ll_create/ll_open gets called.
426 * The server will return to us, in it_disposition, an indication of
427 * exactly what d.lustre.it_status refers to.
429 * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
430 * otherwise if DISP_OPEN_CREATE is set, then it status is the
431 * creation failure mode. In either case, one of DISP_LOOKUP_NEG or
432 * DISP_LOOKUP_POS will be set, indicating whether the child lookup
435 * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
438 int mdc_intent_lock(struct obd_export *exp,
439 struct ll_fid *pfid, const char *name, int len,
440 void *lmm, int lmmsize, struct ll_fid *cfid,
441 struct lookup_intent *it, int lookup_flags,
442 struct ptlrpc_request **reqp,
443 ldlm_blocking_callback cb_blocking)
445 struct lustre_handle lockh;
446 struct ptlrpc_request *request;
448 struct mds_body *mds_body;
449 struct lustre_handle old_lock;
450 struct ldlm_lock *lock;
454 CDEBUG(D_DLMTRACE, "name: %*s in inode "LPU64", intent: %s flags %#o\n",
455 len, name, pfid->id, ldlm_it2str(it->it_op), it->it_flags);
457 if (cfid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
458 it->it_op == IT_CHDIR)) {
459 /* We could just return 1 immediately, but since we should only
460 * be called in revalidate_it if we already have a lock, let's
462 struct ldlm_res_id res_id ={.name = {cfid->id,
464 struct lustre_handle lockh;
465 ldlm_policy_data_t policy;
468 /* For the GETATTR case, ll_revalidate_it issues two separate
469 queries - for LOOKUP and for UPDATE lock because if cannot
470 check them together - we might have those two bits to be
471 present in two separate granted locks */
472 policy.l_inodebits.bits =
473 (it->it_op == IT_GETATTR)?MDS_INODELOCK_UPDATE:
474 MDS_INODELOCK_LOOKUP;
476 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
477 LDLM_FL_BLOCK_GRANTED, &res_id,
478 LDLM_IBITS, &policy, LCK_PR, &lockh);
481 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
482 LDLM_FL_BLOCK_GRANTED, &res_id,
483 LDLM_IBITS, &policy, LCK_PW,
487 memcpy(&it->d.lustre.it_lock_handle, &lockh,
489 it->d.lustre.it_lock_mode = mode;
492 /* Only return failure if it was not GETATTR by cfid
493 (from inode_revalidate) */
498 /* lookup_it may be called only after revalidate_it has run, because
499 * revalidate_it cannot return errors, only zero. Returning zero causes
500 * this call to lookup, which *can* return an error.
502 * We only want to execute the request associated with the intent one
503 * time, however, so don't send the request again. Instead, skip past
504 * this and use the request from revalidate. In this case, revalidate
505 * never dropped its reference, so the refcounts are all OK */
506 if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
507 struct mdc_op_data op_data;
508 mdc_fid2mdc_op_data(&op_data, pfid, cfid, name, len, 0);
510 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
511 &op_data, &lockh, lmm, lmmsize,
512 ldlm_completion_ast, cb_blocking, NULL);
515 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
517 request = *reqp = it->d.lustre.it_data;
518 LASSERT(request != NULL);
520 /* If we're doing an IT_OPEN which did not result in an actual
521 * successful open, then we need to remove the bit which saves
522 * this request for unconditional replay.
524 * It's important that we do this first! Otherwise we might exit the
525 * function without doing so, and try to replay a failed create
527 if (it->it_op & IT_OPEN) {
528 if (!it_disposition(it, DISP_OPEN_OPEN) ||
529 it->d.lustre.it_status != 0) {
530 unsigned long irqflags;
532 spin_lock_irqsave(&request->rq_lock, irqflags);
533 request->rq_replay = 0;
534 spin_unlock_irqrestore(&request->rq_lock, irqflags);
537 if (!it_disposition(it, DISP_IT_EXECD)) {
538 /* The server failed before it even started executing the
539 * intent, i.e. because it couldn't unpack the request. */
540 LASSERT(it->d.lustre.it_status != 0);
541 RETURN(it->d.lustre.it_status);
543 rc = it_open_error(DISP_IT_EXECD, it);
547 mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
548 LASSERT(mds_body != NULL); /* mdc_enqueue checked */
549 LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
551 /* If we were revalidating a fid/name pair, mark the intent in
552 * case we fail and get called again from lookup */
554 it_set_disposition(it, DISP_ENQ_COMPLETE);
555 /* Also: did we find the same inode? */
556 /* we have to compare all the fields but type, because
557 * MDS can return mds/ino/generation triple if inode
558 * lives on another MDS -bzzz */
559 if (cfid->generation != mds_body->fid1.generation ||
560 cfid->id != mds_body->fid1.id ||
561 cfid->mds != mds_body->fid1.mds)
565 rc = it_open_error(DISP_LOOKUP_EXECD, it);
569 /* keep requests around for the multiple phases of the call
570 * this shows the DISP_XX must guarantee we make it into the call
572 if (it_disposition(it, DISP_OPEN_CREATE) &&
573 !it_open_error(DISP_OPEN_CREATE, it))
574 ptlrpc_request_addref(request); /* balanced in ll_create_node */
575 if (it_disposition(it, DISP_OPEN_OPEN) &&
576 !it_open_error(DISP_OPEN_OPEN, it))
577 ptlrpc_request_addref(request); /* balanced in ll_file_open */
579 if (it->it_op & IT_CREAT) {
580 /* XXX this belongs in ll_create_it */
581 } else if (it->it_op == IT_OPEN) {
582 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
584 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR));
587 /* If we already have a matching lock, then cancel the new
588 * one. We have to set the data here instead of in
589 * mdc_enqueue, because we need to use the child's inode as
590 * the l_ast_data to match, and that's not available until
591 * intent_finish has performed the iget().) */
592 lock = ldlm_handle2lock(&lockh);
594 ldlm_policy_data_t policy = lock->l_policy_data;
595 LDLM_DEBUG(lock, "matching against this");
597 memcpy(&old_lock, &lockh, sizeof(lockh));
598 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
599 LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
600 ldlm_lock_decref_and_cancel(&lockh,
601 it->d.lustre.it_lock_mode);
602 memcpy(&lockh, &old_lock, sizeof(old_lock));
603 memcpy(&it->d.lustre.it_lock_handle, &lockh,
607 CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
608 len, name, ldlm_it2str(it->it_op), it->d.lustre.it_status,
609 it->d.lustre.it_disposition, rc);
613 EXPORT_SYMBOL(mdc_intent_lock);