1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
5 * Lustre Metadata Server (mds) request handler
7 * Copyright (c) 2001-2005 Cluster File Systems, Inc.
8 * Author: Peter Braam <braam@clusterfs.com>
9 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
11 * Author: Mike Shaver <shaver@clusterfs.com>
13 * This file is part of the Lustre file system, http://www.lustre.org
14 * Lustre is a trademark of Cluster File Systems, Inc.
16 * You may have signed or agreed to another license before downloading
17 * this software. If so, you are bound by the terms and conditions
18 * of that agreement, and the following does not apply to you. See the
19 * LICENSE file included with this distribution for more information.
21 * If you did not agree to a different license, then this copy of Lustre
22 * is open source software; you can redistribute it and/or modify it
23 * under the terms of version 2 of the GNU General Public License as
24 * published by the Free Software Foundation.
26 * In either case, Lustre is distributed in the hope that it will be
27 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
28 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
29 * license text for more details.
33 # define EXPORT_SYMTAB
35 #define DEBUG_SUBSYSTEM S_MDS
37 #include <lustre_mds.h>
38 #include <linux/module.h>
39 #include <linux/init.h>
40 #include <linux/random.h>
42 #include <linux/jbd.h>
43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
44 # include <linux/smp_lock.h>
45 # include <linux/buffer_head.h>
46 # include <linux/workqueue.h>
47 # include <linux/mount.h>
49 # include <linux/locks.h>
52 #include <obd_class.h>
53 #include <lustre_dlm.h>
55 #include <lustre_fsfilt.h>
56 #include <lprocfs_status.h>
57 #include <lustre_commit_confd.h>
58 #include <lustre_quota.h>
59 #include <lustre_disk.h>
60 #include <lustre_param.h>
62 #include "mds_internal.h"
65 CFS_MODULE_PARM(mds_num_threads, "i", int, 0444,
66 "number of MDS service threads to start");
68 static int mds_intent_policy(struct ldlm_namespace *ns,
69 struct ldlm_lock **lockp, void *req_cookie,
70 ldlm_mode_t mode, int flags, void *data);
71 static int mds_postsetup(struct obd_device *obd);
72 static int mds_cleanup(struct obd_device *obd);
74 /* Assumes caller has already pushed into the kernel filesystem context */
75 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
76 loff_t offset, int count)
78 struct ptlrpc_bulk_desc *desc;
79 struct l_wait_info lwi;
82 int rc = 0, npages, i, tmpcount, tmpsize = 0;
85 LASSERT((offset & ~CFS_PAGE_MASK) == 0); /* I'm dubious about this */
87 npages = (count + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
88 OBD_ALLOC(pages, sizeof(*pages) * npages);
90 GOTO(out, rc = -ENOMEM);
92 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
95 GOTO(out_free, rc = -ENOMEM);
97 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
98 tmpsize = tmpcount > CFS_PAGE_SIZE ? CFS_PAGE_SIZE : tmpcount;
100 OBD_PAGE_ALLOC(pages[i], CFS_ALLOC_STD);
101 if (pages[i] == NULL)
102 GOTO(cleanup_buf, rc = -ENOMEM);
104 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
107 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
108 tmpsize = tmpcount > CFS_PAGE_SIZE ? CFS_PAGE_SIZE : tmpcount;
109 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
110 tmpsize, offset, file->f_dentry->d_inode->i_ino,
111 i_size_read(file->f_dentry->d_inode));
113 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
114 kmap(pages[i]), tmpsize, &offset);
118 GOTO(cleanup_buf, rc = -EIO);
121 LASSERT(desc->bd_nob == count);
123 rc = ptlrpc_start_bulk_transfer(desc);
125 GOTO(cleanup_buf, rc);
127 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
128 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
129 OBD_FAIL_MDS_SENDPAGE, rc);
130 GOTO(abort_bulk, rc);
133 timeout = (int)req->rq_deadline - (int)cfs_time_current_sec();
135 CERROR("Req deadline already passed %lu (now: %lu)\n",
136 req->rq_deadline, cfs_time_current_sec());
138 lwi = LWI_TIMEOUT(max(timeout, 1) * HZ, NULL, NULL);
139 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
140 LASSERT (rc == 0 || rc == -ETIMEDOUT);
143 if (desc->bd_success &&
144 desc->bd_nob_transferred == count)
145 GOTO(cleanup_buf, rc);
147 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
150 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
151 (rc == -ETIMEDOUT) ? "timeout" : "network error",
152 desc->bd_nob_transferred, count,
153 req->rq_export->exp_client_uuid.uuid,
154 req->rq_export->exp_connection->c_remote_uuid.uuid);
156 class_fail_export(req->rq_export);
160 ptlrpc_abort_bulk (desc);
162 for (i = 0; i < npages; i++)
164 OBD_PAGE_FREE(pages[i]);
166 ptlrpc_free_bulk(desc);
168 OBD_FREE(pages, sizeof(*pages) * npages);
173 /* only valid locked dentries or errors should be returned */
174 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
175 struct vfsmount **mnt, int lock_mode,
176 struct lustre_handle *lockh,
177 char *name, int namelen, __u64 lockpart)
179 struct mds_obd *mds = &obd->u.mds;
180 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
181 struct ldlm_res_id res_id = { .name = {0} };
182 int flags = LDLM_FL_ATOMIC_CB, rc;
183 ldlm_policy_data_t policy = { .l_inodebits = { lockpart} };
189 res_id.name[0] = de->d_inode->i_ino;
190 res_id.name[1] = de->d_inode->i_generation;
191 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id,
192 LDLM_IBITS, &policy, lock_mode, &flags,
193 ldlm_blocking_ast, ldlm_completion_ast,
194 NULL, NULL, 0, NULL, lockh);
195 if (rc != ELDLM_OK) {
197 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
203 /* Look up an entry by inode number. */
204 /* this function ONLY returns valid dget'd dentries with an initialized inode
206 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
207 struct vfsmount **mnt)
210 unsigned long ino = fid->id;
211 __u32 generation = fid->generation;
213 struct dentry *result;
216 RETURN(ERR_PTR(-ESTALE));
218 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
220 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
221 ino, generation, mds->mds_obt.obt_sb);
223 /* under ext3 this is neither supposed to return bad inodes
225 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
229 inode = result->d_inode;
231 RETURN(ERR_PTR(-ENOENT));
233 if (inode->i_nlink == 0) {
234 if (inode->i_mode == 0 &&
235 LTIME_S(inode->i_ctime) == 0 ) {
236 struct obd_device *obd = container_of(mds, struct
238 LCONSOLE_WARN("Found inode with zero nlink, mode and "
239 "ctime -- this may indicate disk"
240 "corruption (device %s, inode %lu, link:"
241 " %lu, count: %d)\n", obd->obd_name, inode->i_ino,
242 (unsigned long)inode->i_nlink,
243 atomic_read(&inode->i_count));
246 RETURN(ERR_PTR(-ENOENT));
249 if (generation && inode->i_generation != generation) {
250 /* we didn't find the right inode.. */
251 CDEBUG(D_INODE, "found wrong generation: inode %lu, link: %lu, "
252 "count: %d, generation %u/%u\n", inode->i_ino,
253 (unsigned long)inode->i_nlink,
254 atomic_read(&inode->i_count), inode->i_generation,
257 RETURN(ERR_PTR(-ENOENT));
261 *mnt = mds->mds_vfsmnt;
268 static int mds_connect_internal(struct obd_export *exp,
269 struct obd_connect_data *data)
271 struct obd_device *obd = exp->exp_obd;
273 data->ocd_connect_flags &= MDS_CONNECT_SUPPORTED;
274 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
276 /* If no known bits (which should not happen, probably,
277 as everybody should support LOOKUP and UPDATE bits at least)
278 revert to compat mode with plain locks. */
279 if (!data->ocd_ibits_known &&
280 data->ocd_connect_flags & OBD_CONNECT_IBITS)
281 data->ocd_connect_flags &= ~OBD_CONNECT_IBITS;
283 if (!obd->u.mds.mds_fl_acl)
284 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
286 if (!obd->u.mds.mds_fl_user_xattr)
287 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
289 exp->exp_connect_flags = data->ocd_connect_flags;
290 data->ocd_version = LUSTRE_VERSION_CODE;
291 exp->exp_mds_data.med_ibits_known = data->ocd_ibits_known;
294 if (obd->u.mds.mds_fl_acl &&
295 ((exp->exp_connect_flags & OBD_CONNECT_ACL) == 0)) {
296 CWARN("%s: MDS requires ACL support but client does not\n",
303 static int mds_reconnect(struct obd_export *exp, struct obd_device *obd,
304 struct obd_uuid *cluuid,
305 struct obd_connect_data *data)
310 if (exp == NULL || obd == NULL || cluuid == NULL)
313 rc = mds_connect_internal(exp, data);
318 /* Establish a connection to the MDS.
320 * This will set up an export structure for the client to hold state data
321 * about that client, like open files, the last operation number it did
322 * on the server, etc.
324 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
325 struct obd_uuid *cluuid, struct obd_connect_data *data,
328 struct obd_export *exp;
329 struct mds_export_data *med;
330 struct mds_client_data *mcd = NULL;
331 lnet_nid_t *client_nid = (lnet_nid_t *)localdata;
332 int rc, abort_recovery;
335 if (!conn || !obd || !cluuid)
338 /* Check for aborted recovery. */
339 spin_lock_bh(&obd->obd_processing_task_lock);
340 abort_recovery = obd->obd_abort_recovery;
341 spin_unlock_bh(&obd->obd_processing_task_lock);
343 target_abort_recovery(obd);
345 /* XXX There is a small race between checking the list and adding a
346 * new connection for the same UUID, but the real threat (list
347 * corruption when multiple different clients connect) is solved.
349 * There is a second race between adding the export to the list,
350 * and filling in the client data below. Hence skipping the case
351 * of NULL mcd above. We should already be controlling multiple
352 * connects at the client, and we can't hold the spinlock over
353 * memory allocations without risk of deadlocking.
355 rc = class_connect(conn, obd, cluuid);
358 exp = class_conn2export(conn);
360 med = &exp->exp_mds_data;
362 rc = mds_connect_internal(exp, data);
366 OBD_ALLOC(mcd, sizeof(*mcd));
368 GOTO(out, rc = -ENOMEM);
370 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
373 rc = mds_client_add(obd, exp, -1, *client_nid);
379 OBD_FREE(mcd, sizeof(*mcd));
382 class_disconnect(exp);
384 class_export_put(exp);
390 int mds_init_export(struct obd_export *exp)
392 struct mds_export_data *med = &exp->exp_mds_data;
394 INIT_LIST_HEAD(&med->med_open_head);
395 spin_lock_init(&med->med_open_lock);
397 spin_lock(&exp->exp_lock);
398 exp->exp_connecting = 1;
399 spin_unlock(&exp->exp_lock);
404 static int mds_destroy_export(struct obd_export *export)
406 struct mds_export_data *med;
407 struct obd_device *obd = export->exp_obd;
408 struct mds_obd *mds = &obd->u.mds;
409 struct lvfs_run_ctxt saved;
410 struct lov_mds_md *lmm;
411 struct llog_cookie *logcookies;
415 med = &export->exp_mds_data;
416 target_destroy_export(export);
418 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
421 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
422 /* Close any open files (which may also cause orphan unlinking). */
424 OBD_ALLOC(lmm, mds->mds_max_mdsize);
426 CWARN("%s: allocation failure during cleanup; can not force "
427 "close file handles on this service.\n", obd->obd_name);
428 GOTO(out, rc = -ENOMEM);
431 OBD_ALLOC(logcookies, mds->mds_max_cookiesize);
432 if (logcookies == NULL) {
433 CWARN("%s: allocation failure during cleanup; can not force "
434 "close file handles on this service.\n", obd->obd_name);
435 OBD_FREE(lmm, mds->mds_max_mdsize);
436 GOTO(out, rc = -ENOMEM);
439 spin_lock(&med->med_open_lock);
440 while (!list_empty(&med->med_open_head)) {
441 struct list_head *tmp = med->med_open_head.next;
442 struct mds_file_data *mfd =
443 list_entry(tmp, struct mds_file_data, mfd_list);
444 int lmm_size = mds->mds_max_mdsize;
445 umode_t mode = mfd->mfd_dentry->d_inode->i_mode;
448 /* Remove mfd handle so it can't be found again.
449 * We are consuming the mfd_list reference here. */
450 mds_mfd_unlink(mfd, 0);
451 spin_unlock(&med->med_open_lock);
453 /* If you change this message, be sure to update
454 * replay_single:test_46 */
455 CDEBUG(D_INODE|D_IOCTL, "%s: force closing file handle for "
456 "%.*s (ino %lu)\n", obd->obd_name,
457 mfd->mfd_dentry->d_name.len,mfd->mfd_dentry->d_name.name,
458 mfd->mfd_dentry->d_inode->i_ino);
460 rc = mds_get_md(obd, mfd->mfd_dentry->d_inode, lmm,&lmm_size,1);
462 CWARN("mds_get_md failure, rc=%d\n", rc);
464 valid |= OBD_MD_FLEASIZE;
466 /* child orphan sem protects orphan_dec_test and
467 * is_orphan race, mds_mfd_close drops it */
468 MDS_DOWN_WRITE_ORPHAN_SEM(mfd->mfd_dentry->d_inode);
470 rc = mds_mfd_close(NULL, REQ_REC_OFF, obd, mfd,
471 !(export->exp_flags & OBD_OPT_FAILOVER),
472 lmm, lmm_size, logcookies,
473 mds->mds_max_cookiesize,
477 CDEBUG(D_INODE|D_IOCTL, "Error closing file: %d\n", rc);
479 if (valid & OBD_MD_FLCOOKIE) {
480 rc = mds_osc_destroy_orphan(obd, mode, lmm,
481 lmm_size, logcookies, 1);
483 CDEBUG(D_INODE, "%s: destroy of orphan failed,"
484 " rc = %d\n", obd->obd_name, rc);
487 valid &= ~OBD_MD_FLCOOKIE;
490 spin_lock(&med->med_open_lock);
493 OBD_FREE(logcookies, mds->mds_max_cookiesize);
494 OBD_FREE(lmm, mds->mds_max_mdsize);
496 spin_unlock(&med->med_open_lock);
498 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
499 mds_client_free(export);
505 static int mds_disconnect(struct obd_export *exp)
511 class_export_get(exp);
513 /* Disconnect early so that clients can't keep using export */
514 rc = class_disconnect(exp);
515 if (exp->exp_obd->obd_namespace != NULL)
516 ldlm_cancel_locks_for_export(exp);
518 /* complete all outstanding replies */
519 spin_lock(&exp->exp_lock);
520 while (!list_empty(&exp->exp_outstanding_replies)) {
521 struct ptlrpc_reply_state *rs =
522 list_entry(exp->exp_outstanding_replies.next,
523 struct ptlrpc_reply_state, rs_exp_list);
524 struct ptlrpc_service *svc = rs->rs_service;
526 spin_lock(&svc->srv_lock);
527 list_del_init(&rs->rs_exp_list);
528 ptlrpc_schedule_difficult_reply(rs);
529 spin_unlock(&svc->srv_lock);
531 spin_unlock(&exp->exp_lock);
533 class_export_put(exp);
537 static int mds_getstatus(struct ptlrpc_request *req)
539 struct mds_obd *mds = mds_req2mds(req);
540 struct mds_body *body;
541 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
544 rc = lustre_pack_reply(req, 2, size, NULL);
545 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
546 CERROR("mds: out of memory for message\n");
547 req->rq_status = -ENOMEM; /* superfluous? */
551 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
552 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
554 /* the last_committed and last_xid fields are filled in for all
555 * replies already - no need to do so here also.
560 /* get the LOV EA from @inode and store it into @md. It can be at most
561 * @size bytes, and @size is updated with the actual EA size.
562 * The EA size is also returned on success, and -ve errno on failure.
563 * If there is no EA then 0 is returned. */
564 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
571 LOCK_INODE_MUTEX(inode);
572 rc = fsfilt_get_md(obd, inode, md, *size, "lov");
575 CERROR("Error %d reading eadata for ino %lu\n",
579 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
591 UNLOCK_INODE_MUTEX(inode);
597 /* Call with lock=1 if you want mds_pack_md to take the i_mutex.
598 * Call with lock=0 if the caller has already taken the i_mutex. */
599 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
600 struct mds_body *body, struct inode *inode, int lock)
602 struct mds_obd *mds = &obd->u.mds;
608 lmm = lustre_msg_buf(msg, offset, 0);
610 /* Some problem with getting eadata when I sized the reply
612 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
616 lmm_size = lustre_msg_buflen(msg, offset);
618 /* I don't really like this, but it is a sanity check on the client
619 * MD request. However, if the client doesn't know how much space
620 * to reserve for the MD, it shouldn't be bad to have too much space.
622 if (lmm_size > mds->mds_max_mdsize) {
623 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
624 inode->i_ino, lmm_size, mds->mds_max_mdsize);
628 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
630 if (S_ISDIR(inode->i_mode))
631 body->valid |= OBD_MD_FLDIREA;
633 body->valid |= OBD_MD_FLEASIZE;
634 body->eadatasize = lmm_size;
641 #ifdef CONFIG_FS_POSIX_ACL
643 int mds_pack_posix_acl(struct inode *inode, struct lustre_msg *repmsg,
644 struct mds_body *repbody, int repoff)
646 struct dentry de = { .d_inode = inode };
650 LASSERT(repbody->aclsize == 0);
651 LASSERT(lustre_msg_bufcount(repmsg) > repoff);
653 buflen = lustre_msg_buflen(repmsg, repoff);
657 if (!inode->i_op || !inode->i_op->getxattr)
661 rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
662 lustre_msg_buf(repmsg, repoff, buflen),
667 repbody->aclsize = rc;
668 else if (rc != -ENODATA) {
669 CERROR("buflen %d, get acl: %d\n", buflen, rc);
674 repbody->valid |= OBD_MD_FLACL;
678 #define mds_pack_posix_acl(inode, repmsg, repbody, repoff) 0
681 int mds_pack_acl(struct mds_export_data *med, struct inode *inode,
682 struct lustre_msg *repmsg, struct mds_body *repbody,
685 return mds_pack_posix_acl(inode, repmsg, repbody, repoff);
688 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
689 struct ptlrpc_request *req,
690 struct mds_body *reqbody, int reply_off)
692 struct mds_body *body;
693 struct inode *inode = dentry->d_inode;
700 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
701 LASSERT(body != NULL); /* caller prepped reply */
703 mds_pack_inode2fid(&body->fid1, inode);
704 body->flags = reqbody->flags; /* copy MDS_BFLAG_EXT_FLAGS if present */
705 mds_pack_inode2body(body, inode);
708 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
709 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
710 rc = mds_pack_md(obd, req->rq_repmsg, reply_off, body,
713 /* If we have LOV EA data, the OST holds size, atime, mtime */
714 if (!(body->valid & OBD_MD_FLEASIZE) &&
715 !(body->valid & OBD_MD_FLDIREA))
716 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
717 OBD_MD_FLATIME | OBD_MD_FLMTIME);
719 lustre_shrink_reply(req, reply_off, body->eadatasize, 0);
720 if (body->eadatasize)
722 } else if (S_ISLNK(inode->i_mode) &&
723 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
724 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off, 0);
727 LASSERT (symname != NULL); /* caller prepped reply */
728 len = lustre_msg_buflen(req->rq_repmsg, reply_off);
730 rc = inode->i_op->readlink(dentry, symname, len);
732 CERROR("readlink failed: %d\n", rc);
733 } else if (rc != len - 1) {
734 CERROR ("Unexpected readlink rc %d: expecting %d\n",
738 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
739 body->valid |= OBD_MD_LINKNAME;
740 body->eadatasize = rc + 1;
741 symname[rc] = 0; /* NULL terminate */
745 } else if (reqbody->valid == OBD_MD_FLFLAGS &&
746 reqbody->flags & MDS_BFLAG_EXT_FLAGS) {
749 /* We only return the full set of flags on ioctl, otherwise we
750 * get enough flags from the inode in mds_pack_inode2body(). */
751 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_GETFLAGS,
754 body->flags = flags | MDS_BFLAG_EXT_FLAGS;
757 if (reqbody->valid & OBD_MD_FLMODEASIZE) {
758 struct mds_obd *mds = mds_req2mds(req);
759 body->max_cookiesize = mds->mds_max_cookiesize;
760 body->max_mdsize = mds->mds_max_mdsize;
761 body->valid |= OBD_MD_FLMODEASIZE;
767 #ifdef CONFIG_FS_POSIX_ACL
768 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
769 (reqbody->valid & OBD_MD_FLACL)) {
770 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
771 inode, req->rq_repmsg,
774 lustre_shrink_reply(req, reply_off, body->aclsize, 0);
783 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
786 struct mds_obd *mds = mds_req2mds(req);
787 struct mds_body *body;
788 int rc, bufcount = 2;
789 int size[4] = { sizeof(struct ptlrpc_body), sizeof(*body) };
792 LASSERT(offset == REQ_REC_OFF); /* non-intent */
794 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
795 LASSERT(body != NULL); /* checked by caller */
796 LASSERT(lustre_req_swabbed(req, offset)); /* swabbed by caller */
798 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
799 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
800 LOCK_INODE_MUTEX(inode);
801 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
803 UNLOCK_INODE_MUTEX(inode);
804 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
807 if (rc != -ENODATA) {
808 CERROR("error getting inode %lu MD: rc = %d\n",
813 } else if (rc > mds->mds_max_mdsize) {
815 CERROR("MD size %d larger than maximum possible %u\n",
816 rc, mds->mds_max_mdsize);
821 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
822 if (i_size_read(inode) + 1 != body->eadatasize)
823 CERROR("symlink size: %Lu, reply space: %d\n",
824 i_size_read(inode) + 1, body->eadatasize);
825 size[bufcount] = min_t(int, i_size_read(inode) + 1,
828 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
829 i_size_read(inode) + 1, body->eadatasize);
832 #ifdef CONFIG_FS_POSIX_ACL
833 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
834 (body->valid & OBD_MD_FLACL)) {
835 struct dentry de = { .d_inode = inode };
838 if (inode->i_op && inode->i_op->getxattr) {
840 rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
845 if (rc != -ENODATA) {
846 CERROR("got acl size: %d\n", rc);
856 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
857 CERROR("failed MDS_GETATTR_PACK test\n");
858 req->rq_status = -ENOMEM;
862 rc = lustre_pack_reply(req, bufcount, size, NULL);
864 CERROR("lustre_pack_reply failed: rc %d\n", rc);
872 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
873 int child_part, struct lustre_handle *child_lockh)
875 struct obd_device *obd = req->rq_export->exp_obd;
876 struct mds_obd *mds = &obd->u.mds;
877 struct ldlm_reply *rep = NULL;
878 struct lvfs_run_ctxt saved;
879 struct mds_body *body;
880 struct dentry *dparent = NULL, *dchild = NULL;
881 struct lvfs_ucred uc = {NULL,};
882 struct lustre_handle parent_lockh;
884 int rc = 0, cleanup_phase = 0, resent_req = 0;
888 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
890 /* Swab now, before anyone looks inside the request */
891 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
892 lustre_swab_mds_body);
894 CERROR("Can't swab mds_body\n");
898 lustre_set_req_swabbed(req, offset + 1);
899 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
901 CERROR("Can't unpack name\n");
904 namesize = lustre_msg_buflen(req->rq_reqmsg, offset + 1);
905 /* namesize less than 2 means we have empty name, probably came from
906 revalidate by cfid, so no point in having name to be set */
910 rc = mds_init_ucred(&uc, req, offset);
914 LASSERT(offset == REQ_REC_OFF || offset == DLM_INTENT_REC_OFF);
915 /* if requests were at offset 2, the getattr reply goes back at 1 */
916 if (offset == DLM_INTENT_REC_OFF) {
917 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
919 offset = DLM_REPLY_REC_OFF;
922 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
923 cleanup_phase = 1; /* kernel context */
924 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
926 /* FIXME: handle raw lookup */
928 if (body->valid == OBD_MD_FLID) {
929 struct mds_body *mds_reply;
930 int size = sizeof(*mds_reply);
932 // The user requested ONLY the inode number, so do a raw lookup
933 rc = lustre_pack_reply(req, 1, &size, NULL);
935 CERROR("out of memory\n");
939 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
941 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
943 mds_reply->fid1.id = inum;
944 mds_reply->valid = OBD_MD_FLID;
949 if (lustre_handle_is_used(child_lockh)) {
950 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
954 if (resent_req == 0) {
956 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout*2);
957 rc = mds_get_parent_child_locked(obd, &obd->u.mds,
961 MDS_INODELOCK_UPDATE,
963 child_lockh, &dchild,
966 /* For revalidate by fid we always take UPDATE lock */
967 dchild = mds_fid2locked_dentry(obd, &body->fid2, NULL,
969 NULL, 0, child_part);
972 rc = PTR_ERR(dchild);
977 struct ldlm_lock *granted_lock;
978 struct ll_fid child_fid;
979 struct ldlm_resource *res;
980 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
981 granted_lock = ldlm_handle2lock(child_lockh);
982 LASSERTF(granted_lock != NULL, LPU64"/%u lockh "LPX64"\n",
983 body->fid1.id, body->fid1.generation,
984 child_lockh->cookie);
987 res = granted_lock->l_resource;
988 child_fid.id = res->lr_name.name[0];
989 child_fid.generation = res->lr_name.name[1];
990 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
991 LASSERT(!IS_ERR(dchild));
992 LDLM_LOCK_PUT(granted_lock);
995 cleanup_phase = 2; /* dchild, dparent, locks */
997 if (dchild->d_inode == NULL) {
998 intent_set_disposition(rep, DISP_LOOKUP_NEG);
999 /* in the intent case, the policy clears this error:
1000 the disposition is enough */
1001 GOTO(cleanup, rc = -ENOENT);
1003 intent_set_disposition(rep, DISP_LOOKUP_POS);
1006 if (req->rq_repmsg == NULL) {
1007 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
1009 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1014 rc = mds_getattr_internal(obd, dchild, req, body, offset);
1015 GOTO(cleanup, rc); /* returns the lock to the client */
1018 switch (cleanup_phase) {
1020 if (resent_req == 0) {
1021 if (rc && dchild->d_inode)
1022 ldlm_lock_decref(child_lockh, LCK_CR);
1024 ldlm_lock_decref(&parent_lockh, LCK_CR);
1030 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1032 mds_exit_ucred(&uc, mds);
1033 if (!req->rq_packed_final) {
1034 req->rq_status = rc;
1035 lustre_pack_reply(req, 1, NULL, NULL);
1041 static int mds_getattr(struct ptlrpc_request *req, int offset)
1043 struct mds_obd *mds = mds_req2mds(req);
1044 struct obd_device *obd = req->rq_export->exp_obd;
1045 struct lvfs_run_ctxt saved;
1047 struct mds_body *body;
1048 struct lvfs_ucred uc = { NULL, };
1052 OBD_COUNTER_INCREMENT(obd, getattr);
1054 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1055 lustre_swab_mds_body);
1059 rc = mds_init_ucred(&uc, req, offset);
1061 GOTO(out_ucred, rc);
1063 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1064 de = mds_fid2dentry(mds, &body->fid1, NULL);
1066 rc = req->rq_status = PTR_ERR(de);
1070 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1072 CERROR("mds_getattr_pack_msg: %d\n", rc);
1076 req->rq_status = mds_getattr_internal(obd, de, req, body,REPLY_REC_OFF);
1081 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1083 if (!req->rq_packed_final) {
1084 req->rq_status = rc;
1085 lustre_pack_reply(req, 1, NULL, NULL);
1087 mds_exit_ucred(&uc, mds);
1091 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1096 spin_lock(&obd->obd_osfs_lock);
1097 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
1099 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1100 spin_unlock(&obd->obd_osfs_lock);
1105 static int mds_statfs(struct ptlrpc_request *req)
1107 struct obd_device *obd = req->rq_export->exp_obd;
1108 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1109 int rc, size[2] = { sizeof(struct ptlrpc_body),
1110 sizeof(struct obd_statfs) };
1113 /* This will trigger a watchdog timeout */
1114 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1115 (MDS_SERVICE_WATCHDOG_FACTOR *
1116 at_get(&svc->srv_at_estimate) / 1000) + 1);
1117 OBD_COUNTER_INCREMENT(obd, statfs);
1119 rc = lustre_pack_reply(req, 2, size, NULL);
1120 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1121 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1125 /* We call this so that we can cache a bit - 1 jiffie worth */
1126 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1127 size[REPLY_REC_OFF]),
1128 cfs_time_current_64() - HZ);
1130 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1136 req->rq_status = rc;
1140 static int mds_sync(struct ptlrpc_request *req, int offset)
1142 struct obd_device *obd = req->rq_export->exp_obd;
1143 struct mds_obd *mds = &obd->u.mds;
1144 struct mds_body *body;
1145 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
1148 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1149 lustre_swab_mds_body);
1151 GOTO(out, rc = -EFAULT);
1153 rc = lustre_pack_reply(req, 2, size, NULL);
1154 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1155 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1159 if (body->fid1.id == 0) {
1160 /* a fid of zero is taken to mean "sync whole filesystem" */
1161 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
1166 de = mds_fid2dentry(mds, &body->fid1, NULL);
1168 GOTO(out, rc = PTR_ERR(de));
1170 /* The file parameter isn't used for anything */
1171 if (de->d_inode->i_fop && de->d_inode->i_fop->fsync)
1172 rc = de->d_inode->i_fop->fsync(NULL, de, 1);
1174 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1176 mds_pack_inode2fid(&body->fid1, de->d_inode);
1177 mds_pack_inode2body(body, de->d_inode);
1184 req->rq_status = rc;
1188 /* mds_readpage does not take a DLM lock on the inode, because the client must
1189 * already have a PR lock.
1191 * If we were to take another one here, a deadlock will result, if another
1192 * thread is already waiting for a PW lock. */
1193 static int mds_readpage(struct ptlrpc_request *req, int offset)
1195 struct obd_device *obd = req->rq_export->exp_obd;
1196 struct mds_obd *mds = &obd->u.mds;
1197 struct vfsmount *mnt;
1200 struct mds_body *body, *repbody;
1201 struct lvfs_run_ctxt saved;
1202 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
1203 struct lvfs_ucred uc = {NULL,};
1206 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1209 rc = lustre_pack_reply(req, 2, size, NULL);
1211 CERROR("error packing readpage reply: rc %d\n", rc);
1215 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1216 lustre_swab_mds_body);
1218 GOTO (out, rc = -EFAULT);
1220 rc = mds_init_ucred(&uc, req, offset);
1224 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1225 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1227 GOTO(out_pop, rc = PTR_ERR(de));
1229 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1231 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1232 /* note: in case of an error, dentry_open puts dentry */
1234 GOTO(out_pop, rc = PTR_ERR(file));
1236 /* body->size is actually the offset -eeb */
1237 if ((body->size & (de->d_inode->i_sb->s_blocksize - 1)) != 0) {
1238 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1239 body->size, de->d_inode->i_sb->s_blocksize);
1240 GOTO(out_file, rc = -EFAULT);
1243 /* body->nlink is actually the #bytes to read -eeb */
1244 if (body->nlink & (de->d_inode->i_sb->s_blocksize - 1)) {
1245 CERROR("size %u is not multiple of blocksize %lu\n",
1246 body->nlink, de->d_inode->i_sb->s_blocksize);
1247 GOTO(out_file, rc = -EFAULT);
1250 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1252 repbody->size = i_size_read(file->f_dentry->d_inode);
1253 repbody->valid = OBD_MD_FLSIZE;
1255 /* to make this asynchronous make sure that the handling function
1256 doesn't send a reply when this function completes. Instead a
1257 callback function would send the reply */
1258 /* body->size is actually the offset -eeb */
1259 rc = mds_sendpage(req, file, body->size, body->nlink);
1262 filp_close(file, 0);
1264 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1266 mds_exit_ucred(&uc, mds);
1267 req->rq_status = rc;
1271 int mds_reint(struct ptlrpc_request *req, int offset,
1272 struct lustre_handle *lockh)
1274 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
1277 OBD_ALLOC(rec, sizeof(*rec));
1281 rc = mds_update_unpack(req, offset, rec);
1282 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1283 CERROR("invalid record\n");
1284 GOTO(out, req->rq_status = -EINVAL);
1287 /* rc will be used to interrupt a for loop over multiple records */
1288 rc = mds_reint_rec(rec, offset, req, lockh);
1290 OBD_FREE(rec, sizeof(*rec));
1294 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1295 struct obd_device *obd, int *process)
1297 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1298 case MDS_CONNECT: /* This will never get here, but for completeness. */
1299 case OST_CONNECT: /* This will never get here, but for completeness. */
1300 case MDS_DISCONNECT:
1301 case OST_DISCONNECT:
1306 case MDS_SYNC: /* used in unmounting */
1310 *process = target_queue_recovery_request(req, obd);
1314 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1316 /* XXX what should we set rq_status to here? */
1317 req->rq_status = -EAGAIN;
1318 RETURN(ptlrpc_error(req));
1322 static char *reint_names[] = {
1323 [REINT_SETATTR] "setattr",
1324 [REINT_CREATE] "create",
1325 [REINT_LINK] "link",
1326 [REINT_UNLINK] "unlink",
1327 [REINT_RENAME] "rename",
1328 [REINT_OPEN] "open",
1331 static int mds_set_info_rpc(struct obd_export *exp, struct ptlrpc_request *req)
1334 int keylen, vallen, rc = 0;
1337 key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
1339 DEBUG_REQ(D_HA, req, "no set_info key");
1342 keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
1344 val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, 0);
1345 vallen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1);
1347 rc = lustre_pack_reply(req, 1, NULL, NULL);
1350 lustre_msg_set_status(req->rq_repmsg, 0);
1352 if (KEY_IS("read-only")) {
1353 if (val == NULL || vallen < sizeof(__u32)) {
1354 DEBUG_REQ(D_HA, req, "no set_info val");
1359 exp->exp_connect_flags |= OBD_CONNECT_RDONLY;
1361 exp->exp_connect_flags &= ~OBD_CONNECT_RDONLY;
1369 static int mds_handle_quotacheck(struct ptlrpc_request *req)
1371 struct obd_quotactl *oqctl;
1375 oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1376 lustre_swab_obd_quotactl);
1380 rc = lustre_pack_reply(req, 1, NULL, NULL);
1382 CERROR("mds: out of memory while packing quotacheck reply\n");
1386 req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1390 static int mds_handle_quotactl(struct ptlrpc_request *req)
1392 struct obd_quotactl *oqctl, *repoqc;
1393 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) };
1396 oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1397 lustre_swab_obd_quotactl);
1401 rc = lustre_pack_reply(req, 2, size, NULL);
1405 repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc));
1407 req->rq_status = obd_quotactl(req->rq_export, oqctl);
1412 static int mds_msg_check_version(struct lustre_msg *msg)
1416 switch (lustre_msg_get_opc(msg)) {
1418 case MDS_DISCONNECT:
1420 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1422 CERROR("bad opc %u version %08x, expecting %08x\n",
1423 lustre_msg_get_opc(msg),
1424 lustre_msg_get_version(msg),
1425 LUSTRE_OBD_VERSION);
1429 case MDS_GETATTR_NAME:
1434 case MDS_DONE_WRITING:
1440 case MDS_QUOTACHECK:
1444 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
1446 CERROR("bad opc %u version %08x, expecting %08x\n",
1447 lustre_msg_get_opc(msg),
1448 lustre_msg_get_version(msg),
1449 LUSTRE_MDS_VERSION);
1453 case LDLM_BL_CALLBACK:
1454 case LDLM_CP_CALLBACK:
1455 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1457 CERROR("bad opc %u version %08x, expecting %08x\n",
1458 lustre_msg_get_opc(msg),
1459 lustre_msg_get_version(msg),
1460 LUSTRE_DLM_VERSION);
1462 case OBD_LOG_CANCEL:
1463 case LLOG_ORIGIN_HANDLE_CREATE:
1464 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1465 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1466 case LLOG_ORIGIN_HANDLE_CLOSE:
1467 case LLOG_ORIGIN_HANDLE_DESTROY:
1468 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1470 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1472 CERROR("bad opc %u version %08x, expecting %08x\n",
1473 lustre_msg_get_opc(msg),
1474 lustre_msg_get_version(msg),
1475 LUSTRE_LOG_VERSION);
1478 CERROR("MDS unknown opcode %d\n", lustre_msg_get_opc(msg));
1484 int mds_handle(struct ptlrpc_request *req)
1486 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1488 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1489 struct obd_device *obd = NULL;
1492 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1494 LASSERT(current->journal_info == NULL);
1496 rc = mds_msg_check_version(req->rq_reqmsg);
1498 CERROR("MDS drop mal-formed request\n");
1502 /* XXX identical to OST */
1503 if (lustre_msg_get_opc(req->rq_reqmsg) != MDS_CONNECT) {
1504 struct mds_export_data *med;
1505 int recovering, abort_recovery;
1507 if (req->rq_export == NULL) {
1508 CERROR("operation %d on unconnected MDS from %s\n",
1509 lustre_msg_get_opc(req->rq_reqmsg),
1510 libcfs_id2str(req->rq_peer));
1511 req->rq_status = -ENOTCONN;
1512 GOTO(out, rc = -ENOTCONN);
1515 med = &req->rq_export->exp_mds_data;
1516 obd = req->rq_export->exp_obd;
1519 /* sanity check: if the xid matches, the request must
1520 * be marked as a resent or replayed */
1521 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
1522 req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid))
1523 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
1524 (MSG_RESENT | MSG_REPLAY))) {
1525 CERROR("rq_xid "LPU64" matches last_xid, "
1526 "expected RESENT flag\n",
1528 req->rq_status = -ENOTCONN;
1529 GOTO(out, rc = -EFAULT);
1531 /* else: note the opposite is not always true; a
1532 * RESENT req after a failover will usually not match
1533 * the last_xid, since it was likely never
1534 * committed. A REPLAYed request will almost never
1535 * match the last xid, however it could for a
1536 * committed, but still retained, open. */
1538 /* Check for aborted recovery. */
1539 spin_lock_bh(&obd->obd_processing_task_lock);
1540 abort_recovery = obd->obd_abort_recovery;
1541 recovering = obd->obd_recovering;
1542 spin_unlock_bh(&obd->obd_processing_task_lock);
1543 if (abort_recovery) {
1544 target_abort_recovery(obd);
1545 } else if (recovering) {
1546 rc = mds_filter_recovery_request(req, obd,
1548 if (rc || !should_process)
1553 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1555 DEBUG_REQ(D_INODE, req, "connect");
1556 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1557 rc = target_handle_connect(req, mds_handle);
1559 /* Now that we have an export, set mds. */
1560 obd = req->rq_export->exp_obd;
1561 mds = mds_req2mds(req);
1565 case MDS_DISCONNECT:
1566 DEBUG_REQ(D_INODE, req, "disconnect");
1567 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1568 rc = target_handle_disconnect(req);
1569 req->rq_status = rc; /* superfluous? */
1573 DEBUG_REQ(D_INODE, req, "getstatus");
1574 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1575 rc = mds_getstatus(req);
1579 DEBUG_REQ(D_INODE, req, "getattr");
1580 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1581 rc = mds_getattr(req, REQ_REC_OFF);
1585 DEBUG_REQ(D_INODE, req, "setxattr");
1586 OBD_FAIL_RETURN(OBD_FAIL_MDS_SETXATTR_NET, 0);
1587 rc = mds_setxattr(req);
1591 DEBUG_REQ(D_INODE, req, "getxattr");
1592 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETXATTR_NET, 0);
1593 rc = mds_getxattr(req);
1596 case MDS_GETATTR_NAME: {
1597 struct lustre_handle lockh = { 0 };
1598 DEBUG_REQ(D_INODE, req, "getattr_name");
1599 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1601 /* If this request gets a reconstructed reply, we won't be
1602 * acquiring any new locks in mds_getattr_lock, so we don't
1605 rc = mds_getattr_lock(req, REQ_REC_OFF, MDS_INODELOCK_UPDATE,
1607 /* this non-intent call (from an ioctl) is special */
1608 req->rq_status = rc;
1609 if (rc == 0 && lustre_handle_is_used(&lockh))
1610 ldlm_lock_decref(&lockh, LCK_CR);
1614 DEBUG_REQ(D_INODE, req, "statfs");
1615 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1616 rc = mds_statfs(req);
1620 DEBUG_REQ(D_INODE, req, "readpage");
1621 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1622 rc = mds_readpage(req, REQ_REC_OFF);
1624 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1631 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF,
1634 int size[4] = { sizeof(struct ptlrpc_body),
1635 sizeof(struct mds_body),
1636 mds->mds_max_mdsize,
1637 mds->mds_max_cookiesize };
1640 /* NB only peek inside req now; mds_reint() will swab it */
1642 CERROR ("Can't inspect opcode\n");
1647 if (lustre_msg_swabbed(req->rq_reqmsg))
1650 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1651 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1652 reint_names[opc] == NULL) ? reint_names[opc] :
1655 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1657 if (opc == REINT_UNLINK || opc == REINT_RENAME)
1659 else if (opc == REINT_OPEN)
1664 rc = lustre_pack_reply(req, bufcount, size, NULL);
1668 rc = mds_reint(req, REQ_REC_OFF, NULL);
1669 fail = OBD_FAIL_MDS_REINT_NET_REP;
1674 DEBUG_REQ(D_INODE, req, "close");
1675 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1676 rc = mds_close(req, REQ_REC_OFF);
1679 case MDS_DONE_WRITING:
1680 DEBUG_REQ(D_INODE, req, "done_writing");
1681 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1682 rc = mds_done_writing(req, REQ_REC_OFF);
1686 DEBUG_REQ(D_INODE, req, "pin");
1687 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1688 rc = mds_pin(req, REQ_REC_OFF);
1692 DEBUG_REQ(D_INODE, req, "sync");
1693 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1694 rc = mds_sync(req, REQ_REC_OFF);
1698 DEBUG_REQ(D_INODE, req, "set_info");
1699 rc = mds_set_info_rpc(req->rq_export, req);
1702 case MDS_QUOTACHECK:
1703 DEBUG_REQ(D_INODE, req, "quotacheck");
1704 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACHECK_NET, 0);
1705 rc = mds_handle_quotacheck(req);
1709 DEBUG_REQ(D_INODE, req, "quotactl");
1710 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACTL_NET, 0);
1711 rc = mds_handle_quotactl(req);
1715 DEBUG_REQ(D_INODE, req, "ping");
1716 rc = target_handle_ping(req);
1719 case OBD_LOG_CANCEL:
1720 CDEBUG(D_INODE, "log cancel\n");
1721 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1722 rc = -ENOTSUPP; /* la la la */
1726 DEBUG_REQ(D_INODE, req, "enqueue");
1727 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1728 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1729 ldlm_server_blocking_ast, NULL);
1730 fail = OBD_FAIL_LDLM_REPLY;
1733 DEBUG_REQ(D_INODE, req, "convert");
1734 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1735 rc = ldlm_handle_convert(req);
1737 case LDLM_BL_CALLBACK:
1738 case LDLM_CP_CALLBACK:
1739 DEBUG_REQ(D_INODE, req, "callback");
1740 CERROR("callbacks should not happen on MDS\n");
1742 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1744 case LLOG_ORIGIN_HANDLE_CREATE:
1745 DEBUG_REQ(D_INODE, req, "llog_init");
1746 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1747 rc = llog_origin_handle_create(req);
1749 case LLOG_ORIGIN_HANDLE_DESTROY:
1750 DEBUG_REQ(D_INODE, req, "llog_init");
1751 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1752 rc = llog_origin_handle_destroy(req);
1754 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1755 DEBUG_REQ(D_INODE, req, "llog next block");
1756 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1757 rc = llog_origin_handle_next_block(req);
1759 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1760 DEBUG_REQ(D_INODE, req, "llog prev block");
1761 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1762 rc = llog_origin_handle_prev_block(req);
1764 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1765 DEBUG_REQ(D_INODE, req, "llog read header");
1766 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1767 rc = llog_origin_handle_read_header(req);
1769 case LLOG_ORIGIN_HANDLE_CLOSE:
1770 DEBUG_REQ(D_INODE, req, "llog close");
1771 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1772 rc = llog_origin_handle_close(req);
1775 DEBUG_REQ(D_INODE, req, "llog catinfo");
1776 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1777 rc = llog_catinfo(req);
1780 req->rq_status = -ENOTSUPP;
1781 rc = ptlrpc_error(req);
1785 LASSERT(current->journal_info == NULL);
1787 /* If we're DISCONNECTing, the mds_export_data is already freed */
1788 if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != MDS_DISCONNECT) {
1789 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1791 /* I don't think last_xid is used for anyway, so I'm not sure
1792 if we need to care about last_close_xid here.*/
1793 lustre_msg_set_last_xid(req->rq_repmsg,
1794 le64_to_cpu(med->med_mcd->mcd_last_xid));
1796 target_committed_to_req(req);
1802 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1803 if (obd && obd->obd_recovering) {
1804 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1805 return target_queue_last_replay_reply(req, rc);
1807 /* Lost a race with recovery; let the error path DTRT. */
1808 rc = req->rq_status = -ENOTCONN;
1811 target_send_reply(req, rc, fail);
1815 /* Update the server data on disk. This stores the new mount_count and
1816 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1817 * then the server last_rcvd value may be less than that of the clients.
1818 * This will alert us that we may need to do client recovery.
1820 * Also assumes for mds_last_transno that we are not modifying it (no locking).
1822 int mds_update_server_data(struct obd_device *obd, int force_sync)
1824 struct mds_obd *mds = &obd->u.mds;
1825 struct lr_server_data *lsd = mds->mds_server_data;
1826 struct file *filp = mds->mds_rcvd_filp;
1827 struct lvfs_run_ctxt saved;
1832 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1833 mds->mds_mount_count, mds->mds_last_transno);
1835 spin_lock(&mds->mds_transno_lock);
1836 lsd->lsd_last_transno = cpu_to_le64(mds->mds_last_transno);
1837 spin_unlock(&mds->mds_transno_lock);
1839 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1840 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off,force_sync);
1841 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1843 CERROR("error writing MDS server data: rc = %d\n", rc);
1848 static void fsoptions_to_mds_flags(struct mds_obd *mds, char *options)
1858 while (*p && *p != ',')
1862 if (len == sizeof("user_xattr") - 1 &&
1863 memcmp(options, "user_xattr", len) == 0) {
1864 mds->mds_fl_user_xattr = 1;
1865 LCONSOLE_INFO("Enabling user_xattr\n");
1866 } else if (len == sizeof("nouser_xattr") - 1 &&
1867 memcmp(options, "nouser_xattr", len) == 0) {
1868 mds->mds_fl_user_xattr = 0;
1869 LCONSOLE_INFO("Disabling user_xattr\n");
1870 } else if (len == sizeof("acl") - 1 &&
1871 memcmp(options, "acl", len) == 0) {
1872 #ifdef CONFIG_FS_POSIX_ACL
1873 mds->mds_fl_acl = 1;
1874 LCONSOLE_INFO("Enabling ACL\n");
1876 CWARN("ignoring unsupported acl mount option\n");
1878 } else if (len == sizeof("noacl") - 1 &&
1879 memcmp(options, "noacl", len) == 0) {
1880 #ifdef CONFIG_FS_POSIX_ACL
1881 mds->mds_fl_acl = 0;
1882 LCONSOLE_INFO("Disabling ACL\n");
1890 static int mds_nid_stats_clear_read(char *page, char **start, off_t off,
1891 int count, int *eof, void *data)
1894 return snprintf(page, count, "%s\n",
1895 "Write into this file to clear all nid stats and "
1896 "stale nid entries");
1899 static int mds_nid_stats_clear_write(struct file *file, const char *buffer,
1900 unsigned long count, void *data)
1902 struct obd_device *obd = (struct obd_device *)data;
1903 struct list_head *nids= &obd->obd_proc_nid_list;
1904 nid_stat_t *client_stat = NULL, *nxt;
1906 spin_lock(&obd->nid_lock);
1908 list_for_each_entry_safe (client_stat, nxt, nids, nid_chain) {
1909 if (!client_stat->nid_exp_ref_count)
1910 lprocfs_free_client_stats(client_stat);
1911 else if (client_stat->nid_stats) {
1912 lprocfs_clear_stats(client_stat->nid_stats);
1916 spin_unlock(&obd->nid_lock);
1922 /* mount the file system (secretly). lustre_cfg parameters are:
1928 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1930 struct lprocfs_static_vars lvars;
1931 struct lustre_cfg* lcfg = buf;
1932 struct mds_obd *mds = &obd->u.mds;
1933 struct lustre_sb_info *lsi;
1934 struct lustre_mount_info *lmi;
1935 struct vfsmount *mnt;
1936 struct obd_uuid uuid;
1943 /* setup 1:/dev/loop/0 2:ext3 3:mdsA 4:errors=remount-ro,iopen_nopriv */
1945 CLASSERT(offsetof(struct obd_device, u.obt) ==
1946 offsetof(struct obd_device, u.mds.mds_obt));
1948 if (lcfg->lcfg_bufcount < 3)
1951 if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
1954 lmi = server_get_mount(obd->obd_name);
1956 CERROR("Not mounted in lustre_fill_super?\n");
1960 /* We mounted in lustre_fill_super.
1961 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1962 lsi = s2lsi(lmi->lmi_sb);
1963 fsoptions_to_mds_flags(mds, lsi->lsi_ldd->ldd_mount_opts);
1964 fsoptions_to_mds_flags(mds, lsi->lsi_lmd->lmd_opts);
1966 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1967 if (IS_ERR(obd->obd_fsops))
1968 GOTO(err_put, rc = PTR_ERR(obd->obd_fsops));
1970 CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
1972 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1974 sema_init(&mds->mds_epoch_sem, 1);
1975 spin_lock_init(&mds->mds_transno_lock);
1976 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1977 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1978 mds->mds_atime_diff = MAX_ATIME_DIFF;
1979 mds->mds_evict_ost_nids = 1;
1981 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
1982 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1983 LDLM_NAMESPACE_GREEDY);
1984 if (obd->obd_namespace == NULL) {
1986 GOTO(err_ops, rc = -ENOMEM);
1988 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1990 lprocfs_init_vars(mds, &lvars);
1991 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
1992 lprocfs_alloc_obd_stats(obd, LPROC_MDS_LAST) == 0) {
1993 /* Init private stats here */
1994 mds_stats_counter_init(obd->obd_stats);
1995 obd->obd_proc_exports_entry = proc_mkdir("exports",
1996 obd->obd_proc_entry);
1999 rc = mds_fs_setup(obd, mnt);
2001 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
2006 if (obd->obd_proc_exports_entry)
2007 lprocfs_add_simple(obd->obd_proc_exports_entry,
2008 "clear", mds_nid_stats_clear_read,
2009 mds_nid_stats_clear_write, obd);
2011 if (lcfg->lcfg_bufcount >= 4 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2014 ll_generate_random_uuid(uuid);
2015 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
2017 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
2018 if (mds->mds_profile == NULL)
2019 GOTO(err_fs, rc = -ENOMEM);
2021 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
2022 LUSTRE_CFG_BUFLEN(lcfg, 3));
2025 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2026 "mds_ldlm_client", &obd->obd_ldlm_client);
2027 obd->obd_replayable = 1;
2029 rc = lquota_setup(mds_quota_interface_ref, obd);
2033 mds->mds_group_hash = upcall_cache_init(obd->obd_name);
2034 if (IS_ERR(mds->mds_group_hash)) {
2035 rc = PTR_ERR(mds->mds_group_hash);
2036 mds->mds_group_hash = NULL;
2037 GOTO(err_qctxt, rc);
2040 /* Don't wait for mds_postrecov trying to clear orphans */
2041 obd->obd_async_recov = 1;
2042 rc = mds_postsetup(obd);
2043 /* Bug 11557 - allow async abort_recov start
2044 FIXME can remove most of this obd_async_recov plumbing
2045 obd->obd_async_recov = 0;
2048 GOTO(err_qctxt, rc);
2050 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2051 if (uuid_ptr != NULL) {
2052 class_uuid_unparse(uuid_ptr, &uuid);
2058 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2059 if (obd->obd_recovering) {
2060 LCONSOLE_WARN("MDT %s now serving %s (%s%s%s), but will be in "
2061 "recovery until %d %s reconnect, or if no clients"
2062 " reconnect for %d:%.02d; during that time new "
2063 "clients will not be allowed to connect. "
2064 "Recovery progress can be monitored by watching "
2065 "/proc/fs/lustre/mds/%s/recovery_status.\n",
2066 obd->obd_name, lustre_cfg_string(lcfg, 1),
2067 label ?: "", label ? "/" : "", str,
2068 obd->obd_recoverable_clients,
2069 (obd->obd_recoverable_clients == 1) ?
2070 "client" : "clients",
2071 obd->obd_recovery_timeout / 60,
2072 obd->obd_recovery_timeout % 60,
2075 LCONSOLE_INFO("MDT %s now serving %s (%s%s%s) with recovery "
2076 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2077 label ?: "", label ? "/" : "", str,
2078 obd->obd_replayable ? "enabled" : "disabled");
2081 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
2087 lquota_cleanup(mds_quota_interface_ref, obd);
2089 /* No extra cleanup needed for llog_init_commit_thread() */
2090 mds_fs_cleanup(obd);
2091 upcall_cache_cleanup(mds->mds_group_hash);
2092 mds->mds_group_hash = NULL;
2094 lprocfs_obd_cleanup(obd);
2095 lprocfs_free_obd_stats(obd);
2096 ldlm_namespace_free(obd->obd_namespace, 0);
2097 obd->obd_namespace = NULL;
2099 fsfilt_put_ops(obd->obd_fsops);
2101 server_put_mount(obd->obd_name, mnt);
2102 obd->u.obt.obt_sb = NULL;
2106 static int mds_lov_clean(struct obd_device *obd)
2108 struct mds_obd *mds = &obd->u.mds;
2109 struct obd_device *osc = mds->mds_osc_obd;
2112 if (mds->mds_profile) {
2113 class_del_profile(mds->mds_profile);
2114 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2115 mds->mds_profile = NULL;
2118 /* There better be a lov */
2122 RETURN(PTR_ERR(osc));
2124 obd_register_observer(osc, NULL);
2126 /* Give lov our same shutdown flags */
2127 osc->obd_force = obd->obd_force;
2128 osc->obd_fail = obd->obd_fail;
2130 /* Cleanup the lov */
2131 obd_disconnect(mds->mds_osc_exp);
2132 class_manual_cleanup(osc);
2133 mds->mds_osc_exp = NULL;
2138 static int mds_postsetup(struct obd_device *obd)
2140 struct mds_obd *mds = &obd->u.mds;
2144 rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
2149 rc = llog_setup(obd, LLOG_LOVEA_ORIG_CTXT, obd, 0, NULL,
2154 if (mds->mds_profile) {
2155 struct lustre_profile *lprof;
2156 /* The profile defines which osc and mdc to connect to, for a
2157 client. We reuse that here to figure out the name of the
2158 lov to use (and ignore lprof->lp_mdc).
2159 The profile was set in the config log with
2160 LCFG_MOUNTOPT profilenm oscnm mdcnm */
2161 lprof = class_get_profile(mds->mds_profile);
2162 if (lprof == NULL) {
2163 CERROR("No profile found: %s\n", mds->mds_profile);
2164 GOTO(err_cleanup, rc = -ENOENT);
2166 rc = mds_lov_connect(obd, lprof->lp_osc);
2168 GOTO(err_cleanup, rc);
2175 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2176 llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2180 int mds_postrecov(struct obd_device *obd)
2182 struct llog_ctxt *ctxt;
2189 LASSERT(!obd->obd_recovering);
2190 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
2191 LASSERT(ctxt != NULL);
2192 llog_ctxt_put(ctxt);
2194 /* set nextid first, so we are sure it happens */
2195 mutex_down(&obd->obd_dev_sem);
2196 rc = mds_lov_set_nextid(obd);
2197 mutex_up(&obd->obd_dev_sem);
2199 CERROR("%s: mds_lov_set_nextid failed %d\n",
2204 /* clean PENDING dir */
2205 rc = mds_cleanup_pending(obd);
2209 /* FIXME Does target_finish_recovery really need this to block? */
2210 /* Notify the LOV, which will in turn call mds_notify for each tgt */
2211 /* This means that we have to hack obd_notify to think we're obd_set_up
2212 during mds_lov_connect. */
2213 obd_notify(obd->u.mds.mds_osc_obd, NULL,
2214 obd->obd_async_recov ? OBD_NOTIFY_SYNC_NONBLOCK :
2215 OBD_NOTIFY_SYNC, NULL);
2217 /* quota recovery */
2218 lquota_recovery(mds_quota_interface_ref, obd);
2224 /* We need to be able to stop an mds_lov_synchronize */
2225 static int mds_lov_early_clean(struct obd_device *obd)
2227 struct mds_obd *mds = &obd->u.mds;
2228 struct obd_device *osc = mds->mds_osc_obd;
2230 if (!osc || (!obd->obd_force && !obd->obd_fail))
2233 CDEBUG(D_HA, "abort inflight\n");
2234 return (obd_precleanup(osc, OBD_CLEANUP_EARLY));
2237 static int mds_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2243 case OBD_CLEANUP_EARLY:
2245 case OBD_CLEANUP_EXPORTS:
2246 target_cleanup_recovery(obd);
2247 mds_lov_early_clean(obd);
2249 case OBD_CLEANUP_SELF_EXP:
2250 mds_lov_disconnect(obd);
2252 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2253 llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2254 rc = obd_llog_finish(obd, 0);
2256 case OBD_CLEANUP_OBD:
2262 static int mds_cleanup(struct obd_device *obd)
2264 struct mds_obd *mds = &obd->u.mds;
2265 lvfs_sbdev_type save_dev;
2268 if (obd->u.obt.obt_sb == NULL)
2270 save_dev = lvfs_sbdev(obd->u.obt.obt_sb);
2272 if (mds->mds_osc_exp)
2273 /* lov export was disconnected by mds_lov_clean;
2274 we just need to drop our ref */
2275 class_export_put(mds->mds_osc_exp);
2277 lprocfs_free_per_client_stats(obd);
2278 remove_proc_entry("clear", obd->obd_proc_exports_entry);
2279 lprocfs_obd_cleanup(obd);
2280 lprocfs_free_obd_stats(obd);
2282 lquota_cleanup(mds_quota_interface_ref, obd);
2284 mds_update_server_data(obd, 1);
2285 if (mds->mds_lov_objids != NULL)
2286 OBD_FREE(mds->mds_lov_objids, mds->mds_lov_objids_size);
2287 mds_fs_cleanup(obd);
2289 upcall_cache_cleanup(mds->mds_group_hash);
2290 mds->mds_group_hash = NULL;
2292 server_put_mount(obd->obd_name, mds->mds_vfsmnt);
2293 obd->u.obt.obt_sb = NULL;
2295 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2297 spin_lock_bh(&obd->obd_processing_task_lock);
2298 if (obd->obd_recovering) {
2299 target_cancel_recovery_timer(obd);
2300 obd->obd_recovering = 0;
2302 spin_unlock_bh(&obd->obd_processing_task_lock);
2304 fsfilt_put_ops(obd->obd_fsops);
2306 LCONSOLE_INFO("MDT %s has stopped.\n", obd->obd_name);
2311 static void fixup_handle_for_resent_req(struct ptlrpc_request *req, int offset,
2312 struct ldlm_lock *new_lock,
2313 struct ldlm_lock **old_lock,
2314 struct lustre_handle *lockh)
2316 struct obd_export *exp = req->rq_export;
2317 struct ldlm_request *dlmreq =
2318 lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*dlmreq));
2319 struct lustre_handle remote_hdl = dlmreq->lock_handle[0];
2320 struct list_head *iter;
2322 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2325 spin_lock(&exp->exp_ldlm_data.led_lock);
2326 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2327 struct ldlm_lock *lock;
2328 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2329 if (lock == new_lock)
2331 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2332 lockh->cookie = lock->l_handle.h_cookie;
2333 LDLM_DEBUG(lock, "restoring lock cookie");
2334 DEBUG_REQ(D_DLMTRACE, req,"restoring lock cookie "LPX64,
2337 *old_lock = LDLM_LOCK_GET(lock);
2338 spin_unlock(&exp->exp_ldlm_data.led_lock);
2342 spin_unlock(&exp->exp_ldlm_data.led_lock);
2344 /* If the xid matches, then we know this is a resent request,
2345 * and allow it. (It's probably an OPEN, for which we don't
2348 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
2352 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
2355 /* This remote handle isn't enqueued, so we never received or
2356 * processed this request. Clear MSG_RESENT, because it can
2357 * be handled like any normal request now. */
2359 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2361 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
2365 int intent_disposition(struct ldlm_reply *rep, int flag)
2369 return (rep->lock_policy_res1 & flag);
2372 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2376 rep->lock_policy_res1 |= flag;
2379 #define IS_CLIENT_DISCONNECT_ERROR(error) \
2380 (error == -ENOTCONN || error == -ENODEV)
2382 static int mds_intent_policy(struct ldlm_namespace *ns,
2383 struct ldlm_lock **lockp, void *req_cookie,
2384 ldlm_mode_t mode, int flags, void *data)
2386 struct ptlrpc_request *req = req_cookie;
2387 struct ldlm_lock *lock = *lockp;
2388 struct ldlm_intent *it;
2389 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2390 struct ldlm_reply *rep;
2391 struct lustre_handle lockh = { 0 };
2392 struct ldlm_lock *new_lock = NULL;
2393 int getattr_part = MDS_INODELOCK_UPDATE;
2394 int repsize[5] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
2395 [DLM_LOCKREPLY_OFF] = sizeof(struct ldlm_reply),
2396 [DLM_REPLY_REC_OFF] = sizeof(struct mds_body),
2397 [DLM_REPLY_REC_OFF+1] = mds->mds_max_mdsize };
2398 int repbufcnt = 4, rc;
2401 LASSERT(req != NULL);
2403 if (lustre_msg_bufcount(req->rq_reqmsg) <= DLM_INTENT_IT_OFF) {
2404 /* No intent was provided */
2405 rc = lustre_pack_reply(req, 2, repsize, NULL);
2410 it = lustre_swab_reqbuf(req, DLM_INTENT_IT_OFF, sizeof(*it),
2411 lustre_swab_ldlm_intent);
2413 CERROR("Intent missing\n");
2414 RETURN(req->rq_status = -EFAULT);
2417 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2419 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
2420 (it->opc & (IT_OPEN | IT_GETATTR | IT_LOOKUP)))
2421 /* we should never allow OBD_CONNECT_ACL if not configured */
2422 repsize[repbufcnt++] = LUSTRE_POSIX_ACL_MAX_SIZE;
2423 else if (it->opc & IT_UNLINK)
2424 repsize[repbufcnt++] = mds->mds_max_cookiesize;
2426 rc = lustre_pack_reply(req, repbufcnt, repsize, NULL);
2428 RETURN(req->rq_status = rc);
2430 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
2431 intent_set_disposition(rep, DISP_IT_EXECD);
2434 /* execute policy */
2435 switch ((long)it->opc) {
2437 case IT_CREAT|IT_OPEN:
2438 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
2439 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock, NULL,
2441 /* XXX swab here to assert that an mds_open reint
2442 * packet is following */
2443 rep->lock_policy_res2 = mds_reint(req, DLM_INTENT_REC_OFF,
2446 /* We abort the lock if the lookup was negative and
2447 * we did not make it to the OPEN portion */
2448 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2449 RETURN(ELDLM_LOCK_ABORTED);
2450 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2451 !intent_disposition(rep, DISP_OPEN_OPEN))
2454 /* If there was an error of some sort or if we are not
2455 * returning any locks */
2456 if (rep->lock_policy_res2 ||
2457 !intent_disposition(rep, DISP_OPEN_LOCK)) {
2458 /* If it is the disconnect error (ENODEV & ENOCONN)
2459 * ptlrpc layer should know this imediately, it should
2460 * be replied by rq_stats, otherwise, return it by
2463 if (IS_CLIENT_DISCONNECT_ERROR(rep->lock_policy_res2))
2464 RETURN(rep->lock_policy_res2);
2466 RETURN(ELDLM_LOCK_ABORTED);
2470 getattr_part = MDS_INODELOCK_LOOKUP;
2472 getattr_part |= MDS_INODELOCK_LOOKUP;
2473 OBD_COUNTER_INCREMENT(req->rq_export->exp_obd, getattr);
2475 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock,
2478 /* INODEBITS_INTEROP: if this lock was converted from a
2479 * plain lock (client does not support inodebits), then
2480 * child lock must be taken with both lookup and update
2481 * bits set for all operations.
2483 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS))
2484 getattr_part = MDS_INODELOCK_LOOKUP |
2485 MDS_INODELOCK_UPDATE;
2487 rep->lock_policy_res2 = mds_getattr_lock(req,DLM_INTENT_REC_OFF,
2488 getattr_part, &lockh);
2489 /* FIXME: LDLM can set req->rq_status. MDS sets
2490 policy_res{1,2} with disposition and status.
2491 - replay: returns 0 & req->status is old status
2492 - otherwise: returns req->status */
2493 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2494 rep->lock_policy_res2 = 0;
2495 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2496 rep->lock_policy_res2)
2497 RETURN(ELDLM_LOCK_ABORTED);
2498 if (req->rq_status != 0) {
2500 rep->lock_policy_res2 = req->rq_status;
2501 RETURN(ELDLM_LOCK_ABORTED);
2505 CERROR("Unhandled intent "LPD64"\n", it->opc);
2509 /* By this point, whatever function we called above must have either
2510 * filled in 'lockh', been an intent replay, or returned an error. We
2511 * want to allow replayed RPCs to not get a lock, since we would just
2512 * drop it below anyways because lock replay is done separately by the
2513 * client afterwards. For regular RPCs we want to give the new lock to
2514 * the client instead of whatever lock it was about to get. */
2515 if (new_lock == NULL)
2516 new_lock = ldlm_handle2lock(&lockh);
2517 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2520 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
2521 it->opc, lockh.cookie);
2523 /* If we've already given this lock to a client once, then we should
2524 * have no readers or writers. Otherwise, we should have one reader
2525 * _or_ writer ref (which will be zeroed below) before returning the
2526 * lock to a client. */
2527 if (new_lock->l_export == req->rq_export) {
2528 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2530 LASSERT(new_lock->l_export == NULL);
2531 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2536 if (new_lock->l_export == req->rq_export) {
2537 /* Already gave this to the client, which means that we
2538 * reconstructed a reply. */
2539 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2541 RETURN(ELDLM_LOCK_REPLACED);
2544 /* Fixup the lock to be given to the client */
2545 lock_res_and_lock(new_lock);
2546 new_lock->l_readers = 0;
2547 new_lock->l_writers = 0;
2549 new_lock->l_export = class_export_get(req->rq_export);
2550 spin_lock(&req->rq_export->exp_ldlm_data.led_lock);
2551 list_add(&new_lock->l_export_chain,
2552 &new_lock->l_export->exp_ldlm_data.led_held_locks);
2553 spin_unlock(&req->rq_export->exp_ldlm_data.led_lock);
2555 new_lock->l_blocking_ast = lock->l_blocking_ast;
2556 new_lock->l_completion_ast = lock->l_completion_ast;
2558 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2559 sizeof(lock->l_remote_handle));
2561 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2563 unlock_res_and_lock(new_lock);
2564 LDLM_LOCK_PUT(new_lock);
2566 RETURN(ELDLM_LOCK_REPLACED);
2569 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
2571 struct mds_obd *mds = &obd->u.mds;
2572 struct lprocfs_static_vars lvars;
2573 int mds_min_threads;
2574 int mds_max_threads;
2578 lprocfs_init_vars(mdt, &lvars);
2579 lprocfs_obd_setup(obd, lvars.obd_vars);
2581 sema_init(&mds->mds_health_sem, 1);
2583 if (mds_num_threads) {
2584 /* If mds_num_threads is set, it is the min and the max. */
2585 if (mds_num_threads > MDS_THREADS_MAX)
2586 mds_num_threads = MDS_THREADS_MAX;
2587 if (mds_num_threads < MDS_THREADS_MIN)
2588 mds_num_threads = MDS_THREADS_MIN;
2589 mds_max_threads = mds_min_threads = mds_num_threads;
2591 /* Base min threads on memory and cpus */
2592 mds_min_threads = num_possible_cpus() * num_physpages >>
2593 (27 - CFS_PAGE_SHIFT);
2594 if (mds_min_threads < MDS_THREADS_MIN)
2595 mds_min_threads = MDS_THREADS_MIN;
2596 /* Largest auto threads start value */
2597 if (mds_min_threads > 32)
2598 mds_min_threads = 32;
2599 mds_max_threads = min(MDS_THREADS_MAX, mds_min_threads * 4);
2603 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2604 MDS_MAXREPSIZE, MDS_REQUEST_PORTAL,
2605 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_FACTOR,
2606 mds_handle, LUSTRE_MDS_NAME,
2607 obd->obd_proc_entry, target_print_req,
2608 mds_min_threads, mds_max_threads, "ll_mdt");
2610 if (!mds->mds_service) {
2611 CERROR("failed to start service\n");
2612 GOTO(err_lprocfs, rc = -ENOMEM);
2615 rc = ptlrpc_start_threads(obd, mds->mds_service);
2617 GOTO(err_thread, rc);
2619 mds->mds_setattr_service =
2620 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2621 MDS_MAXREPSIZE, MDS_SETATTR_PORTAL,
2622 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_FACTOR,
2623 mds_handle, "mds_setattr",
2624 obd->obd_proc_entry, target_print_req,
2625 mds_min_threads, mds_max_threads,
2627 if (!mds->mds_setattr_service) {
2628 CERROR("failed to start getattr service\n");
2629 GOTO(err_thread, rc = -ENOMEM);
2632 rc = ptlrpc_start_threads(obd, mds->mds_setattr_service);
2634 GOTO(err_thread2, rc);
2636 mds->mds_readpage_service =
2637 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2638 MDS_MAXREPSIZE, MDS_READPAGE_PORTAL,
2639 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_FACTOR,
2640 mds_handle, "mds_readpage",
2641 obd->obd_proc_entry, target_print_req,
2642 MDS_THREADS_MIN_READPAGE, mds_max_threads,
2644 if (!mds->mds_readpage_service) {
2645 CERROR("failed to start readpage service\n");
2646 GOTO(err_thread2, rc = -ENOMEM);
2649 rc = ptlrpc_start_threads(obd, mds->mds_readpage_service);
2652 GOTO(err_thread3, rc);
2654 ping_evictor_start();
2659 ptlrpc_unregister_service(mds->mds_readpage_service);
2660 mds->mds_readpage_service = NULL;
2662 ptlrpc_unregister_service(mds->mds_setattr_service);
2663 mds->mds_setattr_service = NULL;
2665 ptlrpc_unregister_service(mds->mds_service);
2666 mds->mds_service = NULL;
2668 lprocfs_obd_cleanup(obd);
2672 static int mdt_cleanup(struct obd_device *obd)
2674 struct mds_obd *mds = &obd->u.mds;
2677 ping_evictor_stop();
2679 down(&mds->mds_health_sem);
2680 ptlrpc_unregister_service(mds->mds_readpage_service);
2681 ptlrpc_unregister_service(mds->mds_setattr_service);
2682 ptlrpc_unregister_service(mds->mds_service);
2683 mds->mds_readpage_service = NULL;
2684 mds->mds_setattr_service = NULL;
2685 mds->mds_service = NULL;
2686 up(&mds->mds_health_sem);
2688 lprocfs_obd_cleanup(obd);
2693 static int mdt_health_check(struct obd_device *obd)
2695 struct mds_obd *mds = &obd->u.mds;
2698 down(&mds->mds_health_sem);
2699 rc |= ptlrpc_service_health_check(mds->mds_readpage_service);
2700 rc |= ptlrpc_service_health_check(mds->mds_setattr_service);
2701 rc |= ptlrpc_service_health_check(mds->mds_service);
2702 up(&mds->mds_health_sem);
2705 * health_check to return 0 on healthy
2706 * and 1 on unhealthy.
2714 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2717 struct obd_device *obd = data;
2720 fid.generation = gen;
2721 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2724 static int mds_health_check(struct obd_device *obd)
2726 struct obd_device_target *odt = &obd->u.obt;
2727 #ifdef USE_HEALTH_CHECK_WRITE
2728 struct mds_obd *mds = &obd->u.mds;
2732 if (odt->obt_sb->s_flags & MS_RDONLY)
2735 #ifdef USE_HEALTH_CHECK_WRITE
2736 LASSERT(mds->mds_health_check_filp != NULL);
2737 rc |= !!lvfs_check_io_health(obd, mds->mds_health_check_filp);
2743 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
2745 struct lustre_cfg *lcfg = buf;
2746 struct lprocfs_static_vars lvars;
2749 lprocfs_init_vars(mds, &lvars);
2751 rc = class_process_proc_param(PARAM_MDT, lvars.obd_vars, lcfg, obd);
2756 struct lvfs_callback_ops mds_lvfs_ops = {
2757 l_fid2dentry: mds_lvfs_fid2dentry,
2760 /* use obd ops to offer management infrastructure */
2761 static struct obd_ops mds_obd_ops = {
2762 .o_owner = THIS_MODULE,
2763 .o_connect = mds_connect,
2764 .o_reconnect = mds_reconnect,
2765 .o_init_export = mds_init_export,
2766 .o_destroy_export = mds_destroy_export,
2767 .o_disconnect = mds_disconnect,
2768 .o_setup = mds_setup,
2769 .o_precleanup = mds_precleanup,
2770 .o_cleanup = mds_cleanup,
2771 .o_postrecov = mds_postrecov,
2772 .o_statfs = mds_obd_statfs,
2773 .o_iocontrol = mds_iocontrol,
2774 .o_create = mds_obd_create,
2775 .o_destroy = mds_obd_destroy,
2776 .o_llog_init = mds_llog_init,
2777 .o_llog_finish = mds_llog_finish,
2778 .o_notify = mds_notify,
2779 .o_health_check = mds_health_check,
2780 .o_process_config = mds_process_config,
2783 static struct obd_ops mdt_obd_ops = {
2784 .o_owner = THIS_MODULE,
2785 .o_setup = mdt_setup,
2786 .o_cleanup = mdt_cleanup,
2787 .o_health_check = mdt_health_check,
2790 quota_interface_t *mds_quota_interface_ref;
2791 extern quota_interface_t mds_quota_interface;
2793 static int __init mds_init(void)
2796 struct lprocfs_static_vars lvars;
2798 request_module("lquota");
2799 mds_quota_interface_ref = PORTAL_SYMBOL_GET(mds_quota_interface);
2800 rc = lquota_init(mds_quota_interface_ref);
2802 if (mds_quota_interface_ref)
2803 PORTAL_SYMBOL_PUT(mds_quota_interface);
2806 init_obd_quota_ops(mds_quota_interface_ref, &mds_obd_ops);
2808 lprocfs_init_vars(mds, &lvars);
2809 class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
2810 lprocfs_init_vars(mdt, &lvars);
2811 class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
2816 static void /*__exit*/ mds_exit(void)
2818 lquota_exit(mds_quota_interface_ref);
2819 if (mds_quota_interface_ref)
2820 PORTAL_SYMBOL_PUT(mds_quota_interface);
2822 class_unregister_type(LUSTRE_MDS_NAME);
2823 class_unregister_type(LUSTRE_MDT_NAME);
2826 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2827 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2828 MODULE_LICENSE("GPL");
2830 module_init(mds_init);
2831 module_exit(mds_exit);