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 |
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,
178 /* We always reserve enough space in the reply packet for a stripe MD, because
179 * we don't know in advance the file type. */
180 int mdc_enqueue(struct obd_export *exp,
182 struct lookup_intent *it,
184 struct mdc_op_data *data,
185 struct lustre_handle *lockh,
188 ldlm_completion_callback cb_completion,
189 ldlm_blocking_callback cb_blocking,
192 struct ptlrpc_request *req;
193 struct obd_device *obddev = class_exp2obd(exp);
194 struct ldlm_res_id res_id =
195 { .name = {data->fid1.id, data->fid1.generation} };
196 ldlm_policy_data_t policy = { .l_inodebits = { MDS_INODELOCK_LOOKUP } };
197 int size[6] = {sizeof(struct ldlm_request), sizeof(struct ldlm_intent)};
198 int rc, flags = LDLM_FL_HAS_INTENT;
199 int repsize[4] = {sizeof(struct ldlm_reply),
200 sizeof(struct mds_body),
201 obddev->u.cli.cl_max_mds_easize,
202 obddev->u.cli.cl_max_mds_cookiesize};
203 struct ldlm_reply *dlm_rep;
204 struct ldlm_intent *lit;
205 struct ldlm_request *lockreq;
207 unsigned long irqflags;
208 int reply_buffers = 0;
211 // LDLM_DEBUG_NOLOCK("mdsintent=%s,name=%s,dir=%lu",
212 // ldlm_it2str(it->it_op), it_name, it_inode->i_ino);
214 if (it->it_op & IT_OPEN) {
215 it->it_create_mode |= S_IFREG;
216 it->it_create_mode &= ~current->fs->umask;
218 size[2] = sizeof(struct mds_rec_create);
219 size[3] = data->namelen + 1;
220 size[4] = obddev->u.cli.cl_max_mds_easize;
221 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE,
226 spin_lock_irqsave (&req->rq_lock, irqflags);
228 spin_unlock_irqrestore (&req->rq_lock, irqflags);
230 /* pack the intent */
231 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
232 lit->opc = (__u64)it->it_op;
234 /* pack the intended request */
235 mdc_open_pack(req->rq_reqmsg, 2, data, it->it_create_mode, 0,
236 it->it_flags, lmm, lmmsize);
237 /* get ready for the reply */
239 req->rq_replen = lustre_msg_size(3, repsize);
240 } else if (it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR)) {
241 int valid = OBD_MD_FLNOTOBD | OBD_MD_FLEASIZE;
242 size[2] = sizeof(struct mds_body);
243 size[3] = data->namelen + 1;
245 if (it->it_op & IT_GETATTR)
246 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
248 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 4,
253 /* pack the intent */
254 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
255 lit->opc = (__u64)it->it_op;
257 /* pack the intended request */
258 mdc_getattr_pack(req->rq_reqmsg, valid, 2, it->it_flags, data);
259 /* get ready for the reply */
261 req->rq_replen = lustre_msg_size(3, repsize);
262 } else if (it->it_op == IT_READDIR) {
263 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
264 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 1,
269 /* get ready for the reply */
271 req->rq_replen = lustre_msg_size(1, repsize);
272 } else if (it->it_op == IT_UNLINK) {
273 size[2] = sizeof(struct mds_body);
274 policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
275 req = ptlrpc_prep_req(class_exp2cliimp(exp), LDLM_ENQUEUE, 3,
280 /* pack the intended request */
281 mdc_getattr_pack(req->rq_reqmsg, 0, 2, 0, data);
283 /* pack the intent */
284 lit = lustre_msg_buf(req->rq_reqmsg, 1, sizeof (*lit));
285 lit->opc = (__u64)it->it_op;
287 /* get ready for the reply */
289 req->rq_replen = lustre_msg_size(3, repsize);
294 mdc_get_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
295 rc = ldlm_cli_enqueue(exp, req, obddev->obd_namespace, res_id,
296 lock_type, &policy, lock_mode, &flags,cb_blocking,
297 cb_completion, NULL, cb_data, NULL, 0, NULL,
299 mdc_put_rpc_lock(obddev->u.cli.cl_rpc_lock, it);
301 /* Similarly, if we're going to replay this request, we don't want to
302 * actually get a lock, just perform the intent. */
303 if (req->rq_transno || req->rq_replay) {
304 lockreq = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*lockreq));
305 lockreq->lock_flags |= LDLM_FL_INTENT_ONLY;
308 /* This can go when we're sure that this can never happen */
309 LASSERT(rc != -ENOENT);
310 if (rc == ELDLM_LOCK_ABORTED) {
312 memset(lockh, 0, sizeof(*lockh));
314 } else if (rc != 0) {
315 CERROR("ldlm_cli_enqueue: %d\n", rc);
317 ptlrpc_req_finished(req);
319 } else { /* rc = 0 */
320 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
323 /* If the server gave us back a different lock mode, we should
324 * fix up our variables. */
325 if (lock->l_req_mode != lock_mode) {
326 ldlm_lock_addref(lockh, lock->l_req_mode);
327 ldlm_lock_decref(lockh, lock_mode);
328 lock_mode = lock->l_req_mode;
331 ldlm_lock_allow_match(lock);
335 dlm_rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*dlm_rep));
336 LASSERT(dlm_rep != NULL); /* checked by ldlm_cli_enqueue() */
337 LASSERT_REPSWABBED(req, 0); /* swabbed by ldlm_cli_enqueue() */
339 it->d.lustre.it_disposition = (int) dlm_rep->lock_policy_res1;
340 it->d.lustre.it_status = (int) dlm_rep->lock_policy_res2;
341 it->d.lustre.it_lock_mode = lock_mode;
342 it->d.lustre.it_data = req;
344 if (it->d.lustre.it_status < 0 && req->rq_replay) {
345 LASSERT(req->rq_transno == 0);
346 /* Don't hold error requests for replay. */
347 spin_lock(&req->rq_lock);
349 spin_unlock(&req->rq_lock);
352 DEBUG_REQ(D_RPCTRACE, req, "disposition: %x, status: %d",
353 it->d.lustre.it_disposition, it->d.lustre.it_status);
355 /* We know what to expect, so we do any byte flipping required here */
356 LASSERT(reply_buffers == 4 || reply_buffers == 3 || reply_buffers == 1);
357 if (reply_buffers >= 3) {
358 struct mds_body *body;
360 body = lustre_swab_repbuf(req, 1, sizeof (*body),
361 lustre_swab_mds_body);
363 CERROR ("Can't swab mds_body\n");
367 if ((body->valid & OBD_MD_FLEASIZE) != 0) {
369 /* The eadata is opaque; just check that it is
370 * there. Eventually, obd_unpackmd() will check
372 eadata = lustre_swab_repbuf(req, 2, body->eadatasize,
374 if (eadata == NULL) {
375 CERROR ("Missing/short eadata\n");
378 if (it->it_op & IT_OPEN) {
379 replayea = lustre_msg_buf(req->rq_reqmsg, 4,
380 obddev->u.cli.cl_max_mds_easize);
382 memcpy(replayea, eadata, body->eadatasize);
389 EXPORT_SYMBOL(mdc_enqueue);
392 * This long block is all about fixing up the lock and request state
393 * so that it is correct as of the moment _before_ the operation was
394 * applied; that way, the VFS will think that everything is normal and
395 * call Lustre's regular VFS methods.
397 * If we're performing a creation, that means that unless the creation
398 * failed with EEXIST, we should fake up a negative dentry.
400 * For everything else, we want to lookup to succeed.
402 * One additional note: if CREATE or OPEN succeeded, we add an extra
403 * reference to the request because we need to keep it around until
404 * ll_create/ll_open gets called.
406 * The server will return to us, in it_disposition, an indication of
407 * exactly what d.lustre.it_status refers to.
409 * If DISP_OPEN_OPEN is set, then d.lustre.it_status refers to the open() call,
410 * otherwise if DISP_OPEN_CREATE is set, then it status is the
411 * creation failure mode. In either case, one of DISP_LOOKUP_NEG or
412 * DISP_LOOKUP_POS will be set, indicating whether the child lookup
415 * Else, if DISP_LOOKUP_EXECD then d.lustre.it_status is the rc of the
418 int mdc_intent_lock(struct obd_export *exp, struct ll_uctxt *uctxt,
419 struct ll_fid *pfid, const char *name, int len,
420 void *lmm, int lmmsize, struct ll_fid *cfid,
421 struct lookup_intent *it, int lookup_flags,
422 struct ptlrpc_request **reqp,
423 ldlm_blocking_callback cb_blocking)
425 struct lustre_handle lockh;
426 struct ptlrpc_request *request;
428 struct mds_body *mds_body;
429 struct lustre_handle old_lock;
430 struct ldlm_lock *lock;
434 CDEBUG(D_DLMTRACE, "name: %*s in %ld, intent: %s\n", len, name,
435 pfid ? (unsigned long) pfid->id : 0 , ldlm_it2str(it->it_op));
437 if (cfid && (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR ||
438 it->it_op == IT_CHDIR)) {
439 /* We could just return 1 immediately, but since we should only
440 * be called in revalidate_it if we already have a lock, let's
442 struct ldlm_res_id res_id ={.name = {cfid->id,
444 struct lustre_handle lockh;
445 ldlm_policy_data_t policy;
448 /* For the GETATTR case, ll_revalidate_it issues two separate
449 queries - for LOOKUP and for UPDATE lock because if cannot
450 check them together - we might have those two bits to be
451 present in two separate granted locks */
452 policy.l_inodebits.bits =
453 (it->it_op == IT_GETATTR)?MDS_INODELOCK_UPDATE:
454 MDS_INODELOCK_LOOKUP;
456 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
457 LDLM_FL_BLOCK_GRANTED, &res_id,
458 LDLM_IBITS, &policy, LCK_PR, &lockh);
461 rc = ldlm_lock_match(exp->exp_obd->obd_namespace,
462 LDLM_FL_BLOCK_GRANTED, &res_id,
463 LDLM_IBITS, &policy, LCK_PW,
467 memcpy(&it->d.lustre.it_lock_handle, &lockh,
469 it->d.lustre.it_lock_mode = mode;
474 /* lookup_it may be called only after revalidate_it has run, because
475 * revalidate_it cannot return errors, only zero. Returning zero causes
476 * this call to lookup, which *can* return an error.
478 * We only want to execute the request associated with the intent one
479 * time, however, so don't send the request again. Instead, skip past
480 * this and use the request from revalidate. In this case, revalidate
481 * never dropped its reference, so the refcounts are all OK */
482 if (!it_disposition(it, DISP_ENQ_COMPLETE)) {
483 struct mdc_op_data op_data;
484 mdc_fid2mdc_op_data(&op_data, uctxt, pfid, cfid, name, len, 0);
486 rc = mdc_enqueue(exp, LDLM_IBITS, it, it_to_lock_mode(it),
487 &op_data, &lockh, lmm, lmmsize,
488 ldlm_completion_ast, cb_blocking, NULL);
491 memcpy(&it->d.lustre.it_lock_handle, &lockh, sizeof(lockh));
493 request = *reqp = it->d.lustre.it_data;
494 LASSERT(request != NULL);
496 if (!it_disposition(it, DISP_IT_EXECD)) {
497 /* The server failed before it even started executing the
498 * intent, i.e. because it couldn't unpack the request. */
499 LASSERT(it->d.lustre.it_status != 0);
500 RETURN(it->d.lustre.it_status);
502 rc = it_open_error(DISP_IT_EXECD, it);
506 mds_body = lustre_msg_buf(request->rq_repmsg, 1, sizeof(*mds_body));
507 LASSERT(mds_body != NULL); /* mdc_enqueue checked */
508 LASSERT_REPSWABBED(request, 1); /* mdc_enqueue swabbed */
510 /* If we were revalidating a fid/name pair, mark the intent in
511 * case we fail and get called again from lookup */
513 it_set_disposition(it, DISP_ENQ_COMPLETE);
514 /* Also: did we find the same inode? */
515 /* we have to compare all the fields but type, because
516 * MDS can return mds/ino/generation triple if inode
517 * lives on another MDS -bzzz */
518 if (cfid->generation != mds_body->fid1.generation ||
519 cfid->id != mds_body->fid1.id ||
520 cfid->mds != mds_body->fid1.mds)
524 /* If we're doing an IT_OPEN which did not result in an actual
525 * successful open, then we need to remove the bit which saves
526 * this request for unconditional replay. */
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);
538 rc = it_open_error(DISP_LOOKUP_EXECD, it);
542 /* keep requests around for the multiple phases of the call
543 * this shows the DISP_XX must guarantee we make it into the call
545 if (it_disposition(it, DISP_OPEN_CREATE) &&
546 !it_open_error(DISP_OPEN_CREATE, it))
547 ptlrpc_request_addref(request);
548 if (it_disposition(it, DISP_OPEN_OPEN) &&
549 !it_open_error(DISP_OPEN_OPEN, it))
550 ptlrpc_request_addref(request);
552 if (it->it_op & IT_CREAT) {
553 /* XXX this belongs in ll_create_iit */
554 } else if (it->it_op == IT_OPEN) {
555 LASSERT(!it_disposition(it, DISP_OPEN_CREATE));
557 LASSERT(it->it_op & (IT_GETATTR | IT_LOOKUP | IT_CHDIR));
560 /* If we already have a matching lock, then cancel the new
561 * one. We have to set the data here instead of in
562 * mdc_enqueue, because we need to use the child's inode as
563 * the l_ast_data to match, and that's not available until
564 * intent_finish has performed the iget().) */
565 lock = ldlm_handle2lock(&lockh);
567 ldlm_policy_data_t policy = lock->l_policy_data;
568 LDLM_DEBUG(lock, "matching against this");
570 memcpy(&old_lock, &lockh, sizeof(lockh));
571 if (ldlm_lock_match(NULL, LDLM_FL_BLOCK_GRANTED, NULL,
572 LDLM_IBITS, &policy, LCK_NL, &old_lock)) {
573 ldlm_lock_decref_and_cancel(&lockh,
574 it->d.lustre.it_lock_mode);
575 memcpy(&lockh, &old_lock, sizeof(old_lock));
576 memcpy(&it->d.lustre.it_lock_handle, &lockh,
580 CDEBUG(D_DENTRY, "D_IT dentry %*s intent: %s status %d disp %x rc %d\n",
581 len, name, ldlm_it2str(it->it_op), it->d.lustre.it_status,
582 it->d.lustre.it_disposition, rc);
586 EXPORT_SYMBOL(mdc_intent_lock);