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/lprocfs_status.h>
40 #include "mdc_internal.h"
42 int it_disposition(struct lookup_intent *it, int flag)
44 return it->d.lustre.it_disposition & flag;
46 EXPORT_SYMBOL(it_disposition);
48 void it_set_disposition(struct lookup_intent *it, int flag)
50 it->d.lustre.it_disposition |= flag;
52 EXPORT_SYMBOL(it_set_disposition);
54 static void mdc_fid2mdc_op_data(struct mdc_op_data *data, struct ll_uctxt *ctxt,
55 struct ll_fid *f1, struct ll_fid *f2,
56 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))
86 int it_open_error(int phase, struct lookup_intent *it)
88 if (it_disposition(it, DISP_OPEN_OPEN)) {
89 if (phase == DISP_OPEN_OPEN)
90 return it->d.lustre.it_status;
95 if (it_disposition(it, DISP_OPEN_CREATE)) {
96 if (phase == DISP_OPEN_CREATE)
97 return it->d.lustre.it_status;
102 if (it_disposition(it, DISP_LOOKUP_EXECD)) {
103 if (phase == DISP_LOOKUP_EXECD)
104 return it->d.lustre.it_status;
109 if (it_disposition(it, DISP_IT_EXECD)) {
110 if (phase == DISP_IT_EXECD)
111 return it->d.lustre.it_status;
115 CERROR("it disp: %X, status: %d\n", it->d.lustre.it_disposition,
116 it->d.lustre.it_status);
120 EXPORT_SYMBOL(it_open_error);
122 /* this must be called on a lockh that is known to have a referenced lock */
123 void mdc_set_lock_data(__u64 *l, void *data)
125 struct ldlm_lock *lock;
126 struct lustre_handle *lockh = (struct lustre_handle *)l;
134 lock = ldlm_handle2lock(lockh);
136 LASSERT(lock != NULL);
137 l_lock(&lock->l_resource->lr_namespace->ns_lock);
139 if (lock->l_ast_data && lock->l_ast_data != data) {
140 struct inode *new_inode = data;
141 struct inode *old_inode = lock->l_ast_data;
142 LASSERTF(old_inode->i_state & I_FREEING,
143 "Found existing inode %p/%lu/%u state %lu in lock: "
144 "setting data to %p/%lu/%u\n", old_inode,
145 old_inode->i_ino, old_inode->i_generation,
147 new_inode, new_inode->i_ino, new_inode->i_generation);
150 lock->l_ast_data = data;
151 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
156 EXPORT_SYMBOL(mdc_set_lock_data);
158 int mdc_change_cbdata(struct obd_export *exp, struct ll_fid *fid,
159 ldlm_iterator_t it, void *data)
161 struct ldlm_res_id res_id = { .name = {0} };
164 res_id.name[0] = fid->id;
165 res_id.name[1] = fid->generation;
167 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 int size[6] = {sizeof(struct ldlm_request), sizeof(struct ldlm_intent)};
196 int rc, flags = LDLM_FL_HAS_INTENT;
197 int repsize[4] = {sizeof(struct ldlm_reply),
198 sizeof(struct mds_body),
199 obddev->u.cli.cl_max_mds_easize,
200 obddev->u.cli.cl_max_mds_cookiesize};
201 struct ldlm_reply *dlm_rep;
202 struct ldlm_intent *lit;
203 struct ldlm_request *lockreq;
205 unsigned long irqflags;
206 int reply_buffers = 0;
209 // LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
210 // ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
212 if (it->it_op & IT_OPEN) {
213 it->it_create_mode |= S_IFREG;
214 it->it_create_mode &= ~current->fs->umask;
216 size[2] = sizeof(struct mds_rec_create);
217 size[3] = data->namelen + 1;
218 size[4] = obddev->u.cli.cl_max_mds_easize;
219 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE,
224 spin_lock_irqsave (&req->rq_lock, irqflags);
226 spin_unlock_irqrestore (&req->rq_lock, irqflags);
228 /* pack the intent */
229 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
230 lit->opc = (__u64)it->it_op;
232 /* pack the intended request */
233 mdc_open_pack(req, 2, data, it->it_create_mode, 0,
234 it->it_flags, lmm, lmmsize);
235 /* get ready for the reply */
237 req->rq_replen = lustre_msg_size(3, repsize);
238 } else if (it->it_op & IT_UNLINK) {
239 size[2] = sizeof(struct mds_rec_unlink);
240 size[3] = data->namelen + 1;
241 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 4,
246 /* pack the intent */
247 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
248 lit->opc = (__u64)it->it_op;
250 /* pack the intended request */
251 mdc_unlink_pack(req, 2, data);
252 /* get ready for the reply */
254 req->rq_replen = lustre_msg_size(4, repsize);
255 } else if (it->it_op & (IT_GETATTR | IT_LOOKUP)) {
256 int valid = OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE;
257 size[2] = sizeof(struct mds_body);
258 size[3] = data->namelen + 1;
260 if (it->it_op & IT_GETATTR)
261 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
263 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 4,
268 /* pack the intent */
269 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
270 lit->opc = (__u64)it->it_op;
272 /* pack the intended request */
273 mdc_getattr_pack(req, valid, 2, it->it_flags, data);
274 /* get ready for the reply */
276 req->rq_replen = lustre_msg_size(3, repsize);
277 } else if (it->it_op == IT_READDIR) {
278 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
279 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 1,
284 /* get ready for the reply */
286 req->rq_replen = lustre_msg_size(1, repsize);
291 mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
292 rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
293 lock_type, &policy, lock_mode, &flags,cb_blocking,
294 cb_completion, NULL, cb_data, NULL, 0, NULL,
296 mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
298 /* Similarly, if we're going to replay this request, we don't want to
299 * actually get a lock, just perform the intent. */
300 if (req->rq_transno || req->rq_replay) {
301 lockreq = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*lockreq));
302 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
305 /* This can go when we're sure that this can never happen */
306 LASSERT(rc != -ENOENT);
307 if (rc == ELDLM_LOCK_ABORTED) {
309 memset(lockh, 0, sizeof(*lockh));
311 } else if (rc != 0) {
312 CERROR("ldlm_cli_enqueue: %d\n", rc);
314 ptlrpc_req_finished(req);
316 } else { /* rc = 0 */
317 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
320 /* If the server gave us back a different lock mode, we should
321 * fix up our variables. */
322 if (lock->l_req_mode != lock_mode) {
323 ldlm_lock_addref(lockh, lock->l_req_mode);
324 ldlm_lock_decref(lockh, lock_mode);
325 lock_mode = lock->l_req_mode;
328 ldlm_lock_allow_match(lock);
332 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
333 LASSERT(dlm_rep != NULL); /* checked by ldlm_cli_enqueue() */
334 LASSERT_REPSWABBED(req, 0); /* swabbed by ldlm_cli_enqueue() */
336 it->d.lustre.it_disposition = (int) dlm_rep->lock_policy_res1;
337 it->d.lustre.it_status = (int) dlm_rep->lock_policy_res2;
338 it->d.lustre.it_lock_mode = lock_mode;
339 it->d.lustre.it_data = req;
341 if (it->d.lustre.it_status < 0 && req->rq_replay) {
342 LASSERT(req->rq_transno == 0);
343 /* Don't hold error requests for replay. */
344 spin_lock(&req->rq_lock);
346 spin_unlock(&req->rq_lock);
349 /* We know what to expect, so we do any byte flipping required here */
350 LASSERT(reply_buffers == 4 || reply_buffers == 3 || reply_buffers == 1);
351 if (reply_buffers >= 3) {
352 struct mds_body *body;
354 body = lustre_swab_repbuf(req, 1, sizeof (*body),
355 lustre_swab_mds_body);
357 CERROR ("Can't swab mds_body\n");
361 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
363 /* The eadata is opaque; just check that it is
364 * there. Eventually, obd_unpackmd() will check
366 eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
368 if (eadata == NULL) {
369 CERROR ("Missing/short eadata\n");
372 if (it->it_op & IT_OPEN) {
373 replayea = lustre_msg_buf(req->rq_reqmsg, 4,
374 obddev->u.cli.cl_max_mds_easize);
376 memcpy(replayea, eadata, body->eadatasize);
383 EXPORT_SYMBOL(mdc_enqueue);
386 * This long block is all about fixing up the lock and request state
387 * so that it is correct as of the moment _before_ the operation was
388 * applied; that way, the VFS will think that everything is normal and
389 * call Lustre's regular VFS methods.
391 * If we're performing a creation, that means that unless the creation
392 * failed with EEXIST, we should fake up a negative dentry.
394 * For everything else, we want to lookup to succeed.
396 * One additional note: if CREATE or OPEN succeeded, we add an extra
397 * reference to the request because we need to keep it around until
398 * ll_create/ll_open gets called.
400 * The server will return to us, in it_disposition, an indication of
401 * exactly what d.lustre.it_status refers to.
403 * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
404 * otherwise if DISP_OPEN_CREATE is set, then it status is the
405 * creation failure mode. In either case, one of DISP_LOOKUP_NEG or
406 * DISP_LOOKUP_POS will be set, indicating whether the child lookup
409 * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
412 int mdc_intent_lock(struct obd_export *exp, struct ll_uctxt *uctxt,
413 struct ll_fid *pfid, const char *name, int len,
414 void *lmm, int lmmsize,
415 struct ll_fid *cfid, struct lookup_intent *it,
416 int lookup_flags, struct ptlrpc_request **reqp,
417 ldlm_blocking_callback cb_blocking)
419 struct lustre_handle lockh;
420 struct ptlrpc_request *request;
422 struct mds_body *mds_body;
423 struct lustre_handle old_lock;
424 struct ldlm_lock *lock;
428 CDEBUG(D_DLMTRACE, "name: %*s in %ld, intent: %s\n", len, name,
429 (unsigned long)pfid->id, ldlm_it2str(it->it_op));
431 if (cfid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR)) {
432 /* We could just return 1 immediately, but since we should only
433 * be called in revalidate_it if we already have a lock, let's
435 struct ldlm_res_id res_id ={.name = {cfid->id,
437 struct lustre_handle lockh;
438 ldlm_policy_data_t policy;
441 /* For the GETATTR case, ll_revalidate_it issues two separate
442 queries - for LOOKUP and for UPDATE lock because if cannot
443 check them together - we might have those two bits to be
444 present in two separate granted locks */
445 policy.l_inodebits.bits =
446 (it->it_op == IT_GETATTR)?MDS_INODELOCK_UPDATE:
447 MDS_INODELOCK_LOOKUP;
449 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
450 LDLM_FL_BLOCK_GRANTED, &res_id,
451 LDLM_IBITS, &policy, LCK_PR, &lockh);
454 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
455 LDLM_FL_BLOCK_GRANTED, &res_id,
456 LDLM_IBITS, &policy, LCK_PW,
460 memcpy(&it->d.lustre.it_lock_handle, &lockh,
462 it->d.lustre.it_lock_mode = mode;
467 /* lookup_it may be called only after revalidate_it has run, because
468 * revalidate_it cannot return errors, only zero. Returning zero causes
469 * this call to lookup, which *can* return an error.
471 * We only want to execute the request associated with the intent one
472 * time, however, so don't send the request again. Instead, skip past
473 * this and use the request from revalidate. In this case, revalidate
474 * never dropped its reference, so the refcounts are all OK */
475 if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
476 struct mdc_op_data op_data;
477 mdc_fid2mdc_op_data(&op_data, uctxt, pfid, cfid, name, len, 0);
479 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
480 &op_data, &lockh, lmm, lmmsize,
481 ldlm_completion_ast, cb_blocking, NULL);
484 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
486 request = *reqp = it->d.lustre.it_data;
487 LASSERT(request != NULL);
489 if (!it_disposition(it, DISP_IT_EXECD)) {
490 /* The server failed before it even started executing the
491 * intent, i.e. because it couldn't unpack the request. */
492 LASSERT(it->d.lustre.it_status != 0);
493 RETURN(it->d.lustre.it_status);
495 rc = it_open_error(DISP_IT_EXECD, it);
499 mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
500 LASSERT(mds_body != NULL); /* mdc_enqueue checked */
501 LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
503 /* If we were revalidating a fid/name pair, mark the intent in
504 * case we fail and get called again from lookup */
506 it_set_disposition(it, DISP_ENQ_COMPLETE);
507 /* Also: did we find the same inode? */
508 if (memcmp(cfid, &mds_body->fid1, sizeof(*cfid)))
512 /* If we're doing an IT_OPEN which did not result in an actual
513 * successful open, then we need to remove the bit which saves
514 * this request for unconditional replay. */
515 if (it->it_op & IT_OPEN) {
516 if (!it_disposition(it, DISP_OPEN_OPEN) ||
517 it->d.lustre.it_status != 0) {
518 unsigned long irqflags;
520 spin_lock_irqsave(&request->rq_lock, irqflags);
521 request->rq_replay = 0;
522 spin_unlock_irqrestore(&request->rq_lock, irqflags);
526 rc = it_open_error(DISP_LOOKUP_EXECD, it);
530 /* keep requests around for the multiple phases of the call
531 * this shows the DISP_XX must guarantee we make it into the call
533 if (it_disposition(it, DISP_OPEN_CREATE) &&
534 !it_open_error(DISP_OPEN_CREATE, it))
535 ptlrpc_request_addref(request);
536 if (it_disposition(it, DISP_OPEN_OPEN) &&
537 !it_open_error(DISP_OPEN_OPEN, it))
538 ptlrpc_request_addref(request);
540 if (it->it_op & IT_CREAT) {
541 /* XXX this belongs in ll_create_iit */
542 } else if (it->it_op == IT_OPEN) {
543 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
545 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP));
548 /* If we already have a matching lock, then cancel the new
549 * one. We have to set the data here instead of in
550 * mdc_enqueue, because we need to use the child's inode as
551 * the l_ast_data to match, and that's not available until
552 * intent_finish has performed the iget().) */
553 lock = ldlm_handle2lock(&lockh);
555 ldlm_policy_data_t policy = lock->l_policy_data;
556 LDLM_DEBUG(lock, "matching against this");
558 memcpy(&old_lock, &lockh, sizeof(lockh));
559 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
560 LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
561 ldlm_lock_decref_and_cancel(&lockh,
562 it->d.lustre.it_lock_mode);
563 memcpy(&lockh, &old_lock, sizeof(old_lock));
564 memcpy(&it->d.lustre.it_lock_handle, &lockh,
568 CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
569 len, name, ldlm_it2str(it->it_op), it->d.lustre.it_status,
570 it->d.lustre.it_disposition, rc);
574 EXPORT_SYMBOL(mdc_intent_lock);