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,
463 CWARN("mds_get_md failure, rc=%d\n", rc);
465 valid |= OBD_MD_FLEASIZE;
467 /* child orphan sem protects orphan_dec_test and
468 * is_orphan race, mds_mfd_close drops it */
469 MDS_DOWN_WRITE_ORPHAN_SEM(mfd->mfd_dentry->d_inode);
471 rc = mds_mfd_close(NULL, REQ_REC_OFF, obd, mfd,
472 !(export->exp_flags & OBD_OPT_FAILOVER),
473 lmm, lmm_size, logcookies,
474 mds->mds_max_cookiesize,
478 CDEBUG(D_INODE|D_IOCTL, "Error closing file: %d\n", rc);
480 if (valid & OBD_MD_FLCOOKIE) {
481 rc = mds_osc_destroy_orphan(obd, mode, lmm,
482 lmm_size, logcookies, 1);
484 CDEBUG(D_INODE, "%s: destroy of orphan failed,"
485 " rc = %d\n", obd->obd_name, rc);
488 valid &= ~OBD_MD_FLCOOKIE;
491 spin_lock(&med->med_open_lock);
494 OBD_FREE(logcookies, mds->mds_max_cookiesize);
495 OBD_FREE(lmm, mds->mds_max_mdsize);
497 spin_unlock(&med->med_open_lock);
499 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
500 mds_client_free(export);
506 static int mds_disconnect(struct obd_export *exp)
512 class_export_get(exp);
514 /* Disconnect early so that clients can't keep using export */
515 rc = class_disconnect(exp);
516 if (exp->exp_obd->obd_namespace != NULL)
517 ldlm_cancel_locks_for_export(exp);
519 /* complete all outstanding replies */
520 spin_lock(&exp->exp_lock);
521 while (!list_empty(&exp->exp_outstanding_replies)) {
522 struct ptlrpc_reply_state *rs =
523 list_entry(exp->exp_outstanding_replies.next,
524 struct ptlrpc_reply_state, rs_exp_list);
525 struct ptlrpc_service *svc = rs->rs_service;
527 spin_lock(&svc->srv_lock);
528 list_del_init(&rs->rs_exp_list);
529 ptlrpc_schedule_difficult_reply(rs);
530 spin_unlock(&svc->srv_lock);
532 spin_unlock(&exp->exp_lock);
534 class_export_put(exp);
538 static int mds_getstatus(struct ptlrpc_request *req)
540 struct mds_obd *mds = mds_req2mds(req);
541 struct mds_body *body;
542 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
545 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_PACK, req->rq_status = -ENOMEM);
546 rc = lustre_pack_reply(req, 2, size, NULL);
548 RETURN(req->rq_status = rc);
550 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
551 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
553 /* the last_committed and last_xid fields are filled in for all
554 * replies already - no need to do so here also.
559 /* get the LOV EA from @inode and store it into @md. It can be at most
560 * @size bytes, and @size is updated with the actual EA size.
561 * The EA size is also returned on success, and -ve errno on failure.
562 * If there is no EA then 0 is returned. */
563 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
564 int *size, int lock, int flags)
570 LOCK_INODE_MUTEX(inode);
571 rc = fsfilt_get_md(obd, inode, md, *size, "lov");
573 if (rc == 0 && flags == MDS_GETATTR)
574 rc = mds_get_default_md(obd, md, &lmm_size);
577 CERROR("Error %d reading eadata for ino %lu\n",
581 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
593 UNLOCK_INODE_MUTEX(inode);
599 /* Call with lock=1 if you want mds_pack_md to take the i_mutex.
600 * Call with lock=0 if the caller has already taken the i_mutex. */
601 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
602 struct mds_body *body, struct inode *inode, int lock, int flags)
604 struct mds_obd *mds = &obd->u.mds;
610 lmm = lustre_msg_buf(msg, offset, 0);
612 /* Some problem with getting eadata when I sized the reply
614 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
618 lmm_size = lustre_msg_buflen(msg, offset);
620 /* I don't really like this, but it is a sanity check on the client
621 * MD request. However, if the client doesn't know how much space
622 * to reserve for the MD, it shouldn't be bad to have too much space.
624 if (lmm_size > mds->mds_max_mdsize) {
625 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
626 inode->i_ino, lmm_size, mds->mds_max_mdsize);
630 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, flags);
632 if (S_ISDIR(inode->i_mode))
633 body->valid |= OBD_MD_FLDIREA;
635 body->valid |= OBD_MD_FLEASIZE;
636 body->eadatasize = lmm_size;
643 #ifdef CONFIG_FS_POSIX_ACL
645 int mds_pack_posix_acl(struct inode *inode, struct lustre_msg *repmsg,
646 struct mds_body *repbody, int repoff)
648 struct dentry de = { .d_inode = inode };
652 LASSERT(repbody->aclsize == 0);
653 LASSERT(lustre_msg_bufcount(repmsg) > repoff);
655 buflen = lustre_msg_buflen(repmsg, repoff);
659 if (!inode->i_op || !inode->i_op->getxattr)
663 rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
664 lustre_msg_buf(repmsg, repoff, buflen),
669 repbody->aclsize = rc;
670 else if (rc != -ENODATA) {
671 CERROR("buflen %d, get acl: %d\n", buflen, rc);
676 repbody->valid |= OBD_MD_FLACL;
680 #define mds_pack_posix_acl(inode, repmsg, repbody, repoff) 0
683 int mds_pack_acl(struct mds_export_data *med, struct inode *inode,
684 struct lustre_msg *repmsg, struct mds_body *repbody,
687 return mds_pack_posix_acl(inode, repmsg, repbody, repoff);
690 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
691 struct ptlrpc_request *req,
692 struct mds_body *reqbody, int reply_off)
694 struct mds_body *body;
695 struct inode *inode = dentry->d_inode;
703 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
704 LASSERT(body != NULL); /* caller prepped reply */
706 mds_pack_inode2fid(&body->fid1, inode);
707 body->flags = reqbody->flags; /* copy MDS_BFLAG_EXT_FLAGS if present */
708 mds_pack_inode2body(body, inode);
711 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
712 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
713 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR &&
714 ((S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))))
717 rc = mds_pack_md(obd, req->rq_repmsg, reply_off, body,
720 /* If we have LOV EA data, the OST holds size, atime, mtime */
721 if (!(body->valid & OBD_MD_FLEASIZE) &&
722 !(body->valid & OBD_MD_FLDIREA))
723 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
724 OBD_MD_FLATIME | OBD_MD_FLMTIME);
726 lustre_shrink_reply(req, reply_off, body->eadatasize, 0);
727 if (body->eadatasize)
729 } else if (S_ISLNK(inode->i_mode) &&
730 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
731 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off, 0);
734 LASSERT (symname != NULL); /* caller prepped reply */
735 len = lustre_msg_buflen(req->rq_repmsg, reply_off);
737 rc = inode->i_op->readlink(dentry, symname, len);
739 CERROR("readlink failed: %d\n", rc);
740 } else if (rc != len - 1) {
741 CERROR ("Unexpected readlink rc %d: expecting %d\n",
745 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
746 body->valid |= OBD_MD_LINKNAME;
747 body->eadatasize = rc + 1;
748 symname[rc] = 0; /* NULL terminate */
752 } else if (reqbody->valid == OBD_MD_FLFLAGS &&
753 reqbody->flags & MDS_BFLAG_EXT_FLAGS) {
756 /* We only return the full set of flags on ioctl, otherwise we
757 * get enough flags from the inode in mds_pack_inode2body(). */
758 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_GETFLAGS,
761 body->flags = flags | MDS_BFLAG_EXT_FLAGS;
764 if (reqbody->valid & OBD_MD_FLMODEASIZE) {
765 struct mds_obd *mds = mds_req2mds(req);
766 body->max_cookiesize = mds->mds_max_cookiesize;
767 body->max_mdsize = mds->mds_max_mdsize;
768 body->valid |= OBD_MD_FLMODEASIZE;
774 #ifdef CONFIG_FS_POSIX_ACL
775 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
776 (reqbody->valid & OBD_MD_FLACL)) {
777 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
778 inode, req->rq_repmsg,
781 lustre_shrink_reply(req, reply_off, body->aclsize, 0);
790 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
793 struct mds_obd *mds = mds_req2mds(req);
794 struct mds_body *body;
795 int rc, bufcount = 2;
796 int size[4] = { sizeof(struct ptlrpc_body), sizeof(*body) };
799 LASSERT(offset == REQ_REC_OFF); /* non-intent */
801 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
802 LASSERT(body != NULL); /* checked by caller */
803 LASSERT(lustre_req_swabbed(req, offset)); /* swabbed by caller */
805 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
806 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
807 LOCK_INODE_MUTEX(inode);
808 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
810 UNLOCK_INODE_MUTEX(inode);
811 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
813 if ((rc == 0) && (lustre_msg_get_opc(req->rq_reqmsg) == MDS_GETATTR) &&
814 ((S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))))
815 rc = sizeof(struct lov_mds_md);
817 if (rc != -ENODATA) {
818 CERROR("error getting inode %lu MD: rc = %d\n",
823 } else if (rc > mds->mds_max_mdsize) {
825 CERROR("MD size %d larger than maximum possible %u\n",
826 rc, mds->mds_max_mdsize);
831 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
832 if (i_size_read(inode) + 1 != body->eadatasize)
833 CERROR("symlink size: %Lu, reply space: %d\n",
834 i_size_read(inode) + 1, body->eadatasize);
835 size[bufcount] = min_t(int, i_size_read(inode) + 1,
838 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
839 i_size_read(inode) + 1, body->eadatasize);
842 #ifdef CONFIG_FS_POSIX_ACL
843 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
844 (body->valid & OBD_MD_FLACL)) {
845 struct dentry de = { .d_inode = inode };
848 if (inode->i_op && inode->i_op->getxattr) {
850 rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
855 if (rc != -ENODATA) {
856 CERROR("got acl size: %d\n", rc);
866 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
867 CERROR("failed MDS_GETATTR_PACK test\n");
868 req->rq_status = -ENOMEM;
872 rc = lustre_pack_reply(req, bufcount, size, NULL);
881 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
882 int child_part, struct lustre_handle *child_lockh)
884 struct obd_device *obd = req->rq_export->exp_obd;
885 struct mds_obd *mds = &obd->u.mds;
886 struct ldlm_reply *rep = NULL;
887 struct lvfs_run_ctxt saved;
888 struct mds_body *body;
889 struct dentry *dparent = NULL, *dchild = NULL;
890 struct lvfs_ucred uc = {NULL,};
891 struct lustre_handle parent_lockh;
893 int rc = 0, cleanup_phase = 0, resent_req = 0;
897 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
899 /* Swab now, before anyone looks inside the request */
900 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
901 lustre_swab_mds_body);
903 CERROR("Can't swab mds_body\n");
907 lustre_set_req_swabbed(req, offset + 1);
908 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
910 CERROR("Can't unpack name\n");
913 namesize = lustre_msg_buflen(req->rq_reqmsg, offset + 1);
914 /* namesize less than 2 means we have empty name, probably came from
915 revalidate by cfid, so no point in having name to be set */
919 rc = mds_init_ucred(&uc, req, offset);
923 LASSERT(offset == REQ_REC_OFF || offset == DLM_INTENT_REC_OFF);
924 /* if requests were at offset 2, the getattr reply goes back at 1 */
925 if (offset == DLM_INTENT_REC_OFF) {
926 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
928 offset = DLM_REPLY_REC_OFF;
931 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
932 cleanup_phase = 1; /* kernel context */
933 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
935 /* FIXME: handle raw lookup */
937 if (body->valid == OBD_MD_FLID) {
938 struct mds_body *mds_reply;
939 int size = sizeof(*mds_reply);
941 // The user requested ONLY the inode number, so do a raw lookup
942 rc = lustre_pack_reply(req, 1, &size, NULL);
944 CERROR("out of memory\n");
948 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
950 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
952 mds_reply->fid1.id = inum;
953 mds_reply->valid = OBD_MD_FLID;
958 if (lustre_handle_is_used(child_lockh)) {
959 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
963 if (resent_req == 0) {
965 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout*2);
966 rc = mds_get_parent_child_locked(obd, &obd->u.mds,
970 MDS_INODELOCK_UPDATE,
972 child_lockh, &dchild,
975 /* For revalidate by fid we always take UPDATE lock */
976 dchild = mds_fid2locked_dentry(obd, &body->fid2, NULL,
978 NULL, 0, child_part);
981 rc = PTR_ERR(dchild);
986 struct ldlm_lock *granted_lock;
987 struct ll_fid child_fid;
988 struct ldlm_resource *res;
989 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
990 granted_lock = ldlm_handle2lock(child_lockh);
991 LASSERTF(granted_lock != NULL, LPU64"/%u lockh "LPX64"\n",
992 body->fid1.id, body->fid1.generation,
993 child_lockh->cookie);
996 res = granted_lock->l_resource;
997 child_fid.id = res->lr_name.name[0];
998 child_fid.generation = res->lr_name.name[1];
999 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
1000 LASSERT(!IS_ERR(dchild));
1001 LDLM_LOCK_PUT(granted_lock);
1004 cleanup_phase = 2; /* dchild, dparent, locks */
1006 if (dchild->d_inode == NULL) {
1007 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1008 /* in the intent case, the policy clears this error:
1009 the disposition is enough */
1010 GOTO(cleanup, rc = -ENOENT);
1012 intent_set_disposition(rep, DISP_LOOKUP_POS);
1015 if (req->rq_repmsg == NULL) {
1016 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
1018 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1023 rc = mds_getattr_internal(obd, dchild, req, body, offset);
1024 GOTO(cleanup, rc); /* returns the lock to the client */
1027 switch (cleanup_phase) {
1029 if (resent_req == 0) {
1030 if (rc && dchild->d_inode)
1031 ldlm_lock_decref(child_lockh, LCK_CR);
1033 ldlm_lock_decref(&parent_lockh, LCK_CR);
1039 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1041 mds_exit_ucred(&uc, mds);
1042 if (!req->rq_packed_final) {
1043 int rc2 = lustre_pack_reply(req, 1, NULL, NULL);
1046 req->rq_status = rc;
1052 static int mds_getattr(struct ptlrpc_request *req, int offset)
1054 struct mds_obd *mds = mds_req2mds(req);
1055 struct obd_device *obd = req->rq_export->exp_obd;
1056 struct lvfs_run_ctxt saved;
1058 struct mds_body *body;
1059 struct lvfs_ucred uc = { NULL, };
1063 OBD_COUNTER_INCREMENT(obd, getattr);
1065 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1066 lustre_swab_mds_body);
1070 rc = mds_init_ucred(&uc, req, offset);
1072 GOTO(out_ucred, rc);
1074 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1075 de = mds_fid2dentry(mds, &body->fid1, NULL);
1077 rc = req->rq_status = PTR_ERR(de);
1081 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1083 CERROR("mds_getattr_pack_msg: %d\n", rc);
1087 req->rq_status = mds_getattr_internal(obd, de, req, body,REPLY_REC_OFF);
1092 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1094 if (!req->rq_packed_final) {
1095 int rc2 = lustre_pack_reply(req, 1, NULL, NULL);
1098 req->rq_status = rc;
1100 mds_exit_ucred(&uc, mds);
1104 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1109 spin_lock(&obd->obd_osfs_lock);
1110 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
1112 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1113 spin_unlock(&obd->obd_osfs_lock);
1118 static int mds_statfs(struct ptlrpc_request *req)
1120 struct obd_device *obd = req->rq_export->exp_obd;
1121 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1122 int rc, size[2] = { sizeof(struct ptlrpc_body),
1123 sizeof(struct obd_statfs) };
1126 /* This will trigger a watchdog timeout */
1127 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1128 (MDS_SERVICE_WATCHDOG_FACTOR *
1129 at_get(&svc->srv_at_estimate) / 1000) + 1);
1130 OBD_COUNTER_INCREMENT(obd, statfs);
1132 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK))
1133 GOTO(out, rc = -ENOMEM);
1134 rc = lustre_pack_reply(req, 2, size, NULL);
1138 /* We call this so that we can cache a bit - 1 jiffie worth */
1139 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1140 size[REPLY_REC_OFF]),
1141 cfs_time_current_64() - HZ);
1143 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1149 req->rq_status = rc;
1153 static int mds_sync(struct ptlrpc_request *req, int offset)
1155 struct obd_device *obd = req->rq_export->exp_obd;
1156 struct mds_obd *mds = &obd->u.mds;
1157 struct mds_body *body;
1158 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
1161 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1162 lustre_swab_mds_body);
1164 GOTO(out, rc = -EFAULT);
1166 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK))
1167 GOTO(out, rc = -ENOMEM);
1168 rc = lustre_pack_reply(req, 2, size, NULL);
1172 if (body->fid1.id == 0) {
1173 /* a fid of zero is taken to mean "sync whole filesystem" */
1174 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
1179 de = mds_fid2dentry(mds, &body->fid1, NULL);
1181 GOTO(out, rc = PTR_ERR(de));
1183 /* The file parameter isn't used for anything */
1184 if (de->d_inode->i_fop && de->d_inode->i_fop->fsync)
1185 rc = de->d_inode->i_fop->fsync(NULL, de, 1);
1187 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1189 mds_pack_inode2fid(&body->fid1, de->d_inode);
1190 mds_pack_inode2body(body, de->d_inode);
1197 req->rq_status = rc;
1201 /* mds_readpage does not take a DLM lock on the inode, because the client must
1202 * already have a PR lock.
1204 * If we were to take another one here, a deadlock will result, if another
1205 * thread is already waiting for a PW lock. */
1206 static int mds_readpage(struct ptlrpc_request *req, int offset)
1208 struct obd_device *obd = req->rq_export->exp_obd;
1209 struct mds_obd *mds = &obd->u.mds;
1210 struct vfsmount *mnt;
1213 struct mds_body *body, *repbody;
1214 struct lvfs_run_ctxt saved;
1215 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
1216 struct lvfs_ucred uc = {NULL,};
1219 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_PACK, -ENOMEM);
1220 rc = lustre_pack_reply(req, 2, size, NULL);
1224 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1225 lustre_swab_mds_body);
1227 GOTO (out, rc = -EFAULT);
1229 rc = mds_init_ucred(&uc, req, offset);
1233 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1234 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1236 GOTO(out_pop, rc = PTR_ERR(de));
1238 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1240 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1241 /* note: in case of an error, dentry_open puts dentry */
1243 GOTO(out_pop, rc = PTR_ERR(file));
1245 /* body->size is actually the offset -eeb */
1246 if ((body->size & (de->d_inode->i_sb->s_blocksize - 1)) != 0) {
1247 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1248 body->size, de->d_inode->i_sb->s_blocksize);
1249 GOTO(out_file, rc = -EFAULT);
1252 /* body->nlink is actually the #bytes to read -eeb */
1253 if (body->nlink & (de->d_inode->i_sb->s_blocksize - 1)) {
1254 CERROR("size %u is not multiple of blocksize %lu\n",
1255 body->nlink, de->d_inode->i_sb->s_blocksize);
1256 GOTO(out_file, rc = -EFAULT);
1259 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1261 repbody->size = i_size_read(file->f_dentry->d_inode);
1262 repbody->valid = OBD_MD_FLSIZE;
1264 /* to make this asynchronous make sure that the handling function
1265 doesn't send a reply when this function completes. Instead a
1266 callback function would send the reply */
1267 /* body->size is actually the offset -eeb */
1268 rc = mds_sendpage(req, file, body->size, body->nlink);
1271 filp_close(file, 0);
1273 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1275 mds_exit_ucred(&uc, mds);
1276 req->rq_status = rc;
1280 int mds_reint(struct ptlrpc_request *req, int offset,
1281 struct lustre_handle *lockh)
1283 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
1286 OBD_ALLOC(rec, sizeof(*rec));
1290 rc = mds_update_unpack(req, offset, rec);
1291 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1292 CERROR("invalid record\n");
1293 GOTO(out, req->rq_status = -EINVAL);
1296 mds_root_squash(&req->rq_export->exp_obd->u.mds, &req->rq_peer.nid,
1297 &rec->ur_uc.luc_fsuid, &rec->ur_uc.luc_fsgid,
1298 &rec->ur_uc.luc_cap, &rec->ur_uc.luc_suppgid1,
1299 &rec->ur_uc.luc_suppgid2);
1301 /* rc will be used to interrupt a for loop over multiple records */
1302 rc = mds_reint_rec(rec, offset, req, lockh);
1304 OBD_FREE(rec, sizeof(*rec));
1308 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1309 struct obd_device *obd, int *process)
1311 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1312 case MDS_CONNECT: /* This will never get here, but for completeness. */
1313 case OST_CONNECT: /* This will never get here, but for completeness. */
1314 case MDS_DISCONNECT:
1315 case OST_DISCONNECT:
1320 case MDS_SYNC: /* used in unmounting */
1324 *process = target_queue_recovery_request(req, obd);
1328 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1330 /* XXX what should we set rq_status to here? */
1331 req->rq_status = -EAGAIN;
1332 RETURN(ptlrpc_error(req));
1336 static char *reint_names[] = {
1337 [REINT_SETATTR] "setattr",
1338 [REINT_CREATE] "create",
1339 [REINT_LINK] "link",
1340 [REINT_UNLINK] "unlink",
1341 [REINT_RENAME] "rename",
1342 [REINT_OPEN] "open",
1345 static int mds_set_info_rpc(struct obd_export *exp, struct ptlrpc_request *req)
1348 int keylen, vallen, rc = 0;
1351 key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
1353 DEBUG_REQ(D_HA, req, "no set_info key");
1356 keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
1358 val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, 0);
1359 vallen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1);
1361 rc = lustre_pack_reply(req, 1, NULL, NULL);
1365 lustre_msg_set_status(req->rq_repmsg, 0);
1367 if (KEY_IS("read-only")) {
1368 if (val == NULL || vallen < sizeof(__u32)) {
1369 DEBUG_REQ(D_HA, req, "no set_info val");
1374 exp->exp_connect_flags |= OBD_CONNECT_RDONLY;
1376 exp->exp_connect_flags &= ~OBD_CONNECT_RDONLY;
1384 static int mds_handle_quotacheck(struct ptlrpc_request *req)
1386 struct obd_quotactl *oqctl;
1390 oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1391 lustre_swab_obd_quotactl);
1395 rc = lustre_pack_reply(req, 1, NULL, NULL);
1399 req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1403 static int mds_handle_quotactl(struct ptlrpc_request *req)
1405 struct obd_quotactl *oqctl, *repoqc;
1406 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) };
1409 oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1410 lustre_swab_obd_quotactl);
1414 rc = lustre_pack_reply(req, 2, size, NULL);
1418 repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc));
1420 req->rq_status = obd_quotactl(req->rq_export, oqctl);
1425 static int mds_msg_check_version(struct lustre_msg *msg)
1429 switch (lustre_msg_get_opc(msg)) {
1431 case MDS_DISCONNECT:
1433 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1435 CERROR("bad opc %u version %08x, expecting %08x\n",
1436 lustre_msg_get_opc(msg),
1437 lustre_msg_get_version(msg),
1438 LUSTRE_OBD_VERSION);
1442 case MDS_GETATTR_NAME:
1447 case MDS_DONE_WRITING:
1453 case MDS_QUOTACHECK:
1457 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
1459 CERROR("bad opc %u version %08x, expecting %08x\n",
1460 lustre_msg_get_opc(msg),
1461 lustre_msg_get_version(msg),
1462 LUSTRE_MDS_VERSION);
1466 case LDLM_BL_CALLBACK:
1467 case LDLM_CP_CALLBACK:
1468 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1470 CERROR("bad opc %u version %08x, expecting %08x\n",
1471 lustre_msg_get_opc(msg),
1472 lustre_msg_get_version(msg),
1473 LUSTRE_DLM_VERSION);
1475 case OBD_LOG_CANCEL:
1476 case LLOG_ORIGIN_HANDLE_CREATE:
1477 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1478 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1479 case LLOG_ORIGIN_HANDLE_CLOSE:
1480 case LLOG_ORIGIN_HANDLE_DESTROY:
1481 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1483 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1485 CERROR("bad opc %u version %08x, expecting %08x\n",
1486 lustre_msg_get_opc(msg),
1487 lustre_msg_get_version(msg),
1488 LUSTRE_LOG_VERSION);
1491 CERROR("MDS unknown opcode %d\n", lustre_msg_get_opc(msg));
1497 int mds_handle(struct ptlrpc_request *req)
1499 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1501 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1502 struct obd_device *obd = NULL;
1505 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1507 LASSERT(current->journal_info == NULL);
1509 rc = mds_msg_check_version(req->rq_reqmsg);
1511 CERROR("MDS drop mal-formed request\n");
1515 /* XXX identical to OST */
1516 if (lustre_msg_get_opc(req->rq_reqmsg) != MDS_CONNECT) {
1517 struct mds_export_data *med;
1518 int recovering, abort_recovery;
1520 if (req->rq_export == NULL) {
1521 CERROR("operation %d on unconnected MDS from %s\n",
1522 lustre_msg_get_opc(req->rq_reqmsg),
1523 libcfs_id2str(req->rq_peer));
1524 req->rq_status = -ENOTCONN;
1525 GOTO(out, rc = -ENOTCONN);
1528 med = &req->rq_export->exp_mds_data;
1529 obd = req->rq_export->exp_obd;
1532 /* sanity check: if the xid matches, the request must
1533 * be marked as a resent or replayed */
1534 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
1535 req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid))
1536 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
1537 (MSG_RESENT | MSG_REPLAY))) {
1538 CERROR("rq_xid "LPU64" matches last_xid, "
1539 "expected RESENT flag\n",
1541 req->rq_status = -ENOTCONN;
1542 GOTO(out, rc = -EFAULT);
1544 /* else: note the opposite is not always true; a
1545 * RESENT req after a failover will usually not match
1546 * the last_xid, since it was likely never
1547 * committed. A REPLAYed request will almost never
1548 * match the last xid, however it could for a
1549 * committed, but still retained, open. */
1551 /* Check for aborted recovery. */
1552 spin_lock_bh(&obd->obd_processing_task_lock);
1553 abort_recovery = obd->obd_abort_recovery;
1554 recovering = obd->obd_recovering;
1555 spin_unlock_bh(&obd->obd_processing_task_lock);
1556 if (abort_recovery) {
1557 target_abort_recovery(obd);
1558 } else if (recovering) {
1559 rc = mds_filter_recovery_request(req, obd,
1561 if (rc || !should_process)
1566 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1568 DEBUG_REQ(D_INODE, req, "connect");
1569 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1570 rc = target_handle_connect(req, mds_handle);
1572 /* Now that we have an export, set mds. */
1573 obd = req->rq_export->exp_obd;
1574 mds = mds_req2mds(req);
1578 case MDS_DISCONNECT:
1579 DEBUG_REQ(D_INODE, req, "disconnect");
1580 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1581 rc = target_handle_disconnect(req);
1582 req->rq_status = rc; /* superfluous? */
1586 DEBUG_REQ(D_INODE, req, "getstatus");
1587 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1588 rc = mds_getstatus(req);
1592 DEBUG_REQ(D_INODE, req, "getattr");
1593 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1594 rc = mds_getattr(req, REQ_REC_OFF);
1598 DEBUG_REQ(D_INODE, req, "setxattr");
1599 OBD_FAIL_RETURN(OBD_FAIL_MDS_SETXATTR_NET, 0);
1600 rc = mds_setxattr(req);
1604 DEBUG_REQ(D_INODE, req, "getxattr");
1605 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETXATTR_NET, 0);
1606 rc = mds_getxattr(req);
1609 case MDS_GETATTR_NAME: {
1610 struct lustre_handle lockh = { 0 };
1611 DEBUG_REQ(D_INODE, req, "getattr_name");
1612 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1614 /* If this request gets a reconstructed reply, we won't be
1615 * acquiring any new locks in mds_getattr_lock, so we don't
1618 rc = mds_getattr_lock(req, REQ_REC_OFF, MDS_INODELOCK_UPDATE,
1620 /* this non-intent call (from an ioctl) is special */
1621 req->rq_status = rc;
1622 if (rc == 0 && lustre_handle_is_used(&lockh))
1623 ldlm_lock_decref(&lockh, LCK_CR);
1627 DEBUG_REQ(D_INODE, req, "statfs");
1628 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1629 rc = mds_statfs(req);
1633 DEBUG_REQ(D_INODE, req, "readpage");
1634 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1635 rc = mds_readpage(req, REQ_REC_OFF);
1637 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1644 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF,
1648 int size[4] = { sizeof(struct ptlrpc_body),
1649 sizeof(struct mds_body),
1650 mds->mds_max_mdsize,
1651 mds->mds_max_cookiesize };
1654 /* NB only peek inside req now; mds_reint() will swab it */
1656 CERROR ("Can't inspect opcode\n");
1661 if (lustre_msg_swabbed(req->rq_reqmsg))
1664 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1665 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1666 reint_names[opc] == NULL) ? reint_names[opc] :
1671 op = PTLRPC_LAST_CNTR + MDS_REINT_CREATE;
1674 op = PTLRPC_LAST_CNTR + MDS_REINT_LINK;
1677 op = PTLRPC_LAST_CNTR + MDS_REINT_OPEN;
1680 op = PTLRPC_LAST_CNTR + MDS_REINT_SETATTR;
1683 op = PTLRPC_LAST_CNTR + MDS_REINT_RENAME;
1686 op = PTLRPC_LAST_CNTR + MDS_REINT_UNLINK;
1693 if (op && req->rq_rqbd->rqbd_service->srv_stats)
1694 lprocfs_counter_incr(
1695 req->rq_rqbd->rqbd_service->srv_stats, op);
1697 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1699 if (opc == REINT_UNLINK || opc == REINT_RENAME)
1701 else if (opc == REINT_OPEN)
1706 rc = lustre_pack_reply(req, bufcount, size, NULL);
1710 rc = mds_reint(req, REQ_REC_OFF, NULL);
1711 fail = OBD_FAIL_MDS_REINT_NET_REP;
1716 DEBUG_REQ(D_INODE, req, "close");
1717 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1718 rc = mds_close(req, REQ_REC_OFF);
1719 fail = OBD_FAIL_MDS_CLOSE_NET_REP;
1722 case MDS_DONE_WRITING:
1723 DEBUG_REQ(D_INODE, req, "done_writing");
1724 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1725 rc = mds_done_writing(req, REQ_REC_OFF);
1729 DEBUG_REQ(D_INODE, req, "pin");
1730 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1731 rc = mds_pin(req, REQ_REC_OFF);
1735 DEBUG_REQ(D_INODE, req, "sync");
1736 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1737 rc = mds_sync(req, REQ_REC_OFF);
1741 DEBUG_REQ(D_INODE, req, "set_info");
1742 rc = mds_set_info_rpc(req->rq_export, req);
1745 case MDS_QUOTACHECK:
1746 DEBUG_REQ(D_INODE, req, "quotacheck");
1747 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACHECK_NET, 0);
1748 rc = mds_handle_quotacheck(req);
1752 DEBUG_REQ(D_INODE, req, "quotactl");
1753 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACTL_NET, 0);
1754 rc = mds_handle_quotactl(req);
1758 DEBUG_REQ(D_INODE, req, "ping");
1759 rc = target_handle_ping(req);
1762 case OBD_LOG_CANCEL:
1763 CDEBUG(D_INODE, "log cancel\n");
1764 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1765 rc = -ENOTSUPP; /* la la la */
1769 DEBUG_REQ(D_INODE, req, "enqueue");
1770 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1771 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1772 ldlm_server_blocking_ast, NULL);
1773 fail = OBD_FAIL_LDLM_REPLY;
1776 DEBUG_REQ(D_INODE, req, "convert");
1777 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1778 rc = ldlm_handle_convert(req);
1780 case LDLM_BL_CALLBACK:
1781 case LDLM_CP_CALLBACK:
1782 DEBUG_REQ(D_INODE, req, "callback");
1783 CERROR("callbacks should not happen on MDS\n");
1785 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1787 case LLOG_ORIGIN_HANDLE_CREATE:
1788 DEBUG_REQ(D_INODE, req, "llog_init");
1789 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1790 rc = llog_origin_handle_create(req);
1792 case LLOG_ORIGIN_HANDLE_DESTROY:
1793 DEBUG_REQ(D_INODE, req, "llog_init");
1794 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1795 rc = llog_origin_handle_destroy(req);
1797 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1798 DEBUG_REQ(D_INODE, req, "llog next block");
1799 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1800 rc = llog_origin_handle_next_block(req);
1802 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1803 DEBUG_REQ(D_INODE, req, "llog prev block");
1804 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1805 rc = llog_origin_handle_prev_block(req);
1807 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1808 DEBUG_REQ(D_INODE, req, "llog read header");
1809 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1810 rc = llog_origin_handle_read_header(req);
1812 case LLOG_ORIGIN_HANDLE_CLOSE:
1813 DEBUG_REQ(D_INODE, req, "llog close");
1814 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1815 rc = llog_origin_handle_close(req);
1818 DEBUG_REQ(D_INODE, req, "llog catinfo");
1819 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1820 rc = llog_catinfo(req);
1823 req->rq_status = -ENOTSUPP;
1824 rc = ptlrpc_error(req);
1828 LASSERT(current->journal_info == NULL);
1830 /* If we're DISCONNECTing, the mds_export_data is already freed */
1831 if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != MDS_DISCONNECT) {
1832 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1834 /* I don't think last_xid is used for anyway, so I'm not sure
1835 if we need to care about last_close_xid here.*/
1836 lustre_msg_set_last_xid(req->rq_repmsg,
1837 le64_to_cpu(med->med_mcd->mcd_last_xid));
1839 target_committed_to_req(req);
1845 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1846 if (obd && obd->obd_recovering) {
1847 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1848 return target_queue_last_replay_reply(req, rc);
1850 /* Lost a race with recovery; let the error path DTRT. */
1851 rc = req->rq_status = -ENOTCONN;
1854 target_send_reply(req, rc, fail);
1858 /* Update the server data on disk. This stores the new mount_count and
1859 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1860 * then the server last_rcvd value may be less than that of the clients.
1861 * This will alert us that we may need to do client recovery.
1863 * Also assumes for mds_last_transno that we are not modifying it (no locking).
1865 int mds_update_server_data(struct obd_device *obd, int force_sync)
1867 struct mds_obd *mds = &obd->u.mds;
1868 struct lr_server_data *lsd = mds->mds_server_data;
1869 struct file *filp = mds->mds_rcvd_filp;
1870 struct lvfs_run_ctxt saved;
1875 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1876 mds->mds_mount_count, mds->mds_last_transno);
1878 spin_lock(&mds->mds_transno_lock);
1879 lsd->lsd_last_transno = cpu_to_le64(mds->mds_last_transno);
1880 spin_unlock(&mds->mds_transno_lock);
1882 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1883 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off,force_sync);
1884 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1886 CERROR("error writing MDS server data: rc = %d\n", rc);
1891 static void fsoptions_to_mds_flags(struct mds_obd *mds, char *options)
1901 while (*p && *p != ',')
1905 if (len == sizeof("user_xattr") - 1 &&
1906 memcmp(options, "user_xattr", len) == 0) {
1907 mds->mds_fl_user_xattr = 1;
1908 LCONSOLE_INFO("Enabling user_xattr\n");
1909 } else if (len == sizeof("nouser_xattr") - 1 &&
1910 memcmp(options, "nouser_xattr", len) == 0) {
1911 mds->mds_fl_user_xattr = 0;
1912 LCONSOLE_INFO("Disabling user_xattr\n");
1913 } else if (len == sizeof("acl") - 1 &&
1914 memcmp(options, "acl", len) == 0) {
1915 #ifdef CONFIG_FS_POSIX_ACL
1916 mds->mds_fl_acl = 1;
1917 LCONSOLE_INFO("Enabling ACL\n");
1919 CWARN("ignoring unsupported acl mount option\n");
1921 } else if (len == sizeof("noacl") - 1 &&
1922 memcmp(options, "noacl", len) == 0) {
1923 #ifdef CONFIG_FS_POSIX_ACL
1924 mds->mds_fl_acl = 0;
1925 LCONSOLE_INFO("Disabling ACL\n");
1933 /* mount the file system (secretly). lustre_cfg parameters are:
1939 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
1941 struct lprocfs_static_vars lvars;
1942 struct lustre_cfg* lcfg = buf;
1943 struct mds_obd *mds = &obd->u.mds;
1944 struct lustre_sb_info *lsi;
1945 struct lustre_mount_info *lmi;
1946 struct vfsmount *mnt;
1947 struct obd_uuid uuid;
1954 /* setup 1:/dev/loop/0 2:ext3 3:mdsA 4:errors=remount-ro,iopen_nopriv */
1956 CLASSERT(offsetof(struct obd_device, u.obt) ==
1957 offsetof(struct obd_device, u.mds.mds_obt));
1959 if (lcfg->lcfg_bufcount < 3)
1962 if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
1965 lmi = server_get_mount(obd->obd_name);
1967 CERROR("Not mounted in lustre_fill_super?\n");
1971 /* We mounted in lustre_fill_super.
1972 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1973 lsi = s2lsi(lmi->lmi_sb);
1974 fsoptions_to_mds_flags(mds, lsi->lsi_ldd->ldd_mount_opts);
1975 fsoptions_to_mds_flags(mds, lsi->lsi_lmd->lmd_opts);
1977 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1978 if (IS_ERR(obd->obd_fsops))
1979 GOTO(err_put, rc = PTR_ERR(obd->obd_fsops));
1981 CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
1983 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1985 sema_init(&mds->mds_epoch_sem, 1);
1986 spin_lock_init(&mds->mds_transno_lock);
1987 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1988 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1989 mds->mds_atime_diff = MAX_ATIME_DIFF;
1990 mds->mds_evict_ost_nids = 1;
1992 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
1993 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1994 LDLM_NAMESPACE_GREEDY);
1995 if (obd->obd_namespace == NULL) {
1997 GOTO(err_ops, rc = -ENOMEM);
1999 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
2001 lprocfs_mds_init_vars(&lvars);
2002 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
2003 lprocfs_alloc_obd_stats(obd, LPROC_MDS_LAST) == 0) {
2004 /* Init private stats here */
2005 mds_stats_counter_init(obd->obd_stats);
2006 obd->obd_proc_exports_entry = proc_mkdir("exports",
2007 obd->obd_proc_entry);
2010 rc = mds_fs_setup(obd, mnt);
2012 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
2017 if (obd->obd_proc_exports_entry)
2018 lprocfs_add_simple(obd->obd_proc_exports_entry,
2019 "clear", lprocfs_nid_stats_clear_read,
2020 lprocfs_nid_stats_clear_write, obd);
2022 if (lcfg->lcfg_bufcount >= 4 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
2025 ll_generate_random_uuid(uuid);
2026 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
2028 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
2029 if (mds->mds_profile == NULL)
2030 GOTO(err_fs, rc = -ENOMEM);
2032 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
2033 LUSTRE_CFG_BUFLEN(lcfg, 3));
2036 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2037 "mds_ldlm_client", &obd->obd_ldlm_client);
2038 obd->obd_replayable = 1;
2040 rc = lquota_setup(mds_quota_interface_ref, obd);
2044 mds->mds_group_hash = upcall_cache_init(obd->obd_name);
2045 if (IS_ERR(mds->mds_group_hash)) {
2046 rc = PTR_ERR(mds->mds_group_hash);
2047 mds->mds_group_hash = NULL;
2048 GOTO(err_qctxt, rc);
2051 /* Don't wait for mds_postrecov trying to clear orphans */
2052 obd->obd_async_recov = 1;
2053 rc = mds_postsetup(obd);
2054 /* Bug 11557 - allow async abort_recov start
2055 FIXME can remove most of this obd_async_recov plumbing
2056 obd->obd_async_recov = 0;
2059 GOTO(err_qctxt, rc);
2061 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2062 if (uuid_ptr != NULL) {
2063 class_uuid_unparse(uuid_ptr, &uuid);
2069 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2070 if (obd->obd_recovering) {
2071 LCONSOLE_WARN("MDT %s now serving %s (%s%s%s), but will be in "
2072 "recovery for at least %d:%.02d, or until %d "
2073 "client%s reconnect. During this time new clients"
2074 " will not be allowed to connect. "
2075 "Recovery progress can be monitored by watching "
2076 "/proc/fs/lustre/mds/%s/recovery_status.\n",
2077 obd->obd_name, lustre_cfg_string(lcfg, 1),
2078 label ?: "", label ? "/" : "", str,
2079 obd->obd_recovery_timeout / 60,
2080 obd->obd_recovery_timeout % 60,
2081 obd->obd_max_recoverable_clients,
2082 (obd->obd_max_recoverable_clients == 1) ? "":"s",
2085 LCONSOLE_INFO("MDT %s now serving %s (%s%s%s) with recovery "
2086 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2087 label ?: "", label ? "/" : "", str,
2088 obd->obd_replayable ? "enabled" : "disabled");
2091 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
2097 lquota_cleanup(mds_quota_interface_ref, obd);
2099 /* No extra cleanup needed for llog_init_commit_thread() */
2100 mds_fs_cleanup(obd);
2101 upcall_cache_cleanup(mds->mds_group_hash);
2102 mds->mds_group_hash = NULL;
2104 lprocfs_free_obd_stats(obd);
2105 lprocfs_obd_cleanup(obd);
2106 ldlm_namespace_free(obd->obd_namespace, 0);
2107 obd->obd_namespace = NULL;
2109 fsfilt_put_ops(obd->obd_fsops);
2111 server_put_mount(obd->obd_name, mnt);
2112 obd->u.obt.obt_sb = NULL;
2116 static int mds_lov_clean(struct obd_device *obd)
2118 struct mds_obd *mds = &obd->u.mds;
2119 struct obd_device *osc = mds->mds_osc_obd;
2122 if (mds->mds_profile) {
2123 class_del_profile(mds->mds_profile);
2124 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2125 mds->mds_profile = NULL;
2128 /* There better be a lov */
2132 RETURN(PTR_ERR(osc));
2134 obd_register_observer(osc, NULL);
2136 /* Give lov our same shutdown flags */
2137 osc->obd_force = obd->obd_force;
2138 osc->obd_fail = obd->obd_fail;
2140 /* Cleanup the lov */
2141 obd_disconnect(mds->mds_osc_exp);
2142 class_manual_cleanup(osc);
2143 mds->mds_osc_exp = NULL;
2148 static int mds_postsetup(struct obd_device *obd)
2150 struct mds_obd *mds = &obd->u.mds;
2154 rc = llog_setup(obd, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
2159 rc = llog_setup(obd, LLOG_LOVEA_ORIG_CTXT, obd, 0, NULL,
2164 if (mds->mds_profile) {
2165 struct lustre_profile *lprof;
2166 /* The profile defines which osc and mdc to connect to, for a
2167 client. We reuse that here to figure out the name of the
2168 lov to use (and ignore lprof->lp_mdc).
2169 The profile was set in the config log with
2170 LCFG_MOUNTOPT profilenm oscnm mdcnm */
2171 lprof = class_get_profile(mds->mds_profile);
2172 if (lprof == NULL) {
2173 CERROR("No profile found: %s\n", mds->mds_profile);
2174 GOTO(err_cleanup, rc = -ENOENT);
2176 rc = mds_lov_connect(obd, lprof->lp_osc);
2178 GOTO(err_cleanup, rc);
2185 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2186 llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2190 int mds_postrecov(struct obd_device *obd)
2192 struct llog_ctxt *ctxt;
2199 LASSERT(!obd->obd_recovering);
2200 ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
2201 LASSERT(ctxt != NULL);
2202 llog_ctxt_put(ctxt);
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 remove_proc_entry("clear", obd->obd_proc_exports_entry);
2278 lprocfs_free_per_client_stats(obd);
2279 lprocfs_free_obd_stats(obd);
2280 lprocfs_obd_cleanup(obd);
2282 lquota_cleanup(mds_quota_interface_ref, obd);
2284 mds_update_server_data(obd, 1);
2285 mds_fs_cleanup(obd);
2287 upcall_cache_cleanup(mds->mds_group_hash);
2288 mds->mds_group_hash = NULL;
2290 server_put_mount(obd->obd_name, mds->mds_vfsmnt);
2291 obd->u.obt.obt_sb = NULL;
2293 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2295 spin_lock_bh(&obd->obd_processing_task_lock);
2296 if (obd->obd_recovering) {
2297 target_cancel_recovery_timer(obd);
2298 obd->obd_recovering = 0;
2300 spin_unlock_bh(&obd->obd_processing_task_lock);
2302 fsfilt_put_ops(obd->obd_fsops);
2304 LCONSOLE_INFO("MDT %s has stopped.\n", obd->obd_name);
2309 static void fixup_handle_for_resent_req(struct ptlrpc_request *req, int offset,
2310 struct ldlm_lock *new_lock,
2311 struct ldlm_lock **old_lock,
2312 struct lustre_handle *lockh)
2314 struct obd_export *exp = req->rq_export;
2315 struct ldlm_request *dlmreq =
2316 lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*dlmreq));
2317 struct lustre_handle remote_hdl = dlmreq->lock_handle[0];
2318 struct list_head *iter;
2320 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2323 spin_lock(&exp->exp_ldlm_data.led_lock);
2324 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2325 struct ldlm_lock *lock;
2326 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2327 if (lock == new_lock)
2329 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2330 lockh->cookie = lock->l_handle.h_cookie;
2331 LDLM_DEBUG(lock, "restoring lock cookie");
2332 DEBUG_REQ(D_DLMTRACE, req,"restoring lock cookie "LPX64,
2335 *old_lock = LDLM_LOCK_GET(lock);
2336 spin_unlock(&exp->exp_ldlm_data.led_lock);
2340 spin_unlock(&exp->exp_ldlm_data.led_lock);
2342 /* If the xid matches, then we know this is a resent request,
2343 * and allow it. (It's probably an OPEN, for which we don't
2346 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
2350 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
2353 /* This remote handle isn't enqueued, so we never received or
2354 * processed this request. Clear MSG_RESENT, because it can
2355 * be handled like any normal request now. */
2357 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2359 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
2363 int intent_disposition(struct ldlm_reply *rep, int flag)
2367 return (rep->lock_policy_res1 & flag);
2370 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2374 rep->lock_policy_res1 |= flag;
2377 #define IS_CLIENT_DISCONNECT_ERROR(error) \
2378 (error == -ENOTCONN || error == -ENODEV)
2380 static int mds_intent_policy(struct ldlm_namespace *ns,
2381 struct ldlm_lock **lockp, void *req_cookie,
2382 ldlm_mode_t mode, int flags, void *data)
2384 struct ptlrpc_request *req = req_cookie;
2385 struct ldlm_lock *lock = *lockp;
2386 struct ldlm_intent *it;
2387 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2388 struct ldlm_reply *rep;
2389 struct lustre_handle lockh = { 0 };
2390 struct ldlm_lock *new_lock = NULL;
2391 int getattr_part = MDS_INODELOCK_UPDATE;
2392 int repsize[5] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
2393 [DLM_LOCKREPLY_OFF] = sizeof(struct ldlm_reply),
2394 [DLM_REPLY_REC_OFF] = sizeof(struct mds_body),
2395 [DLM_REPLY_REC_OFF+1] = mds->mds_max_mdsize };
2396 int repbufcnt = 4, rc;
2399 LASSERT(req != NULL);
2401 if (lustre_msg_bufcount(req->rq_reqmsg) <= DLM_INTENT_IT_OFF) {
2402 /* No intent was provided */
2403 rc = lustre_pack_reply(req, 2, repsize, NULL);
2409 it = lustre_swab_reqbuf(req, DLM_INTENT_IT_OFF, sizeof(*it),
2410 lustre_swab_ldlm_intent);
2412 CERROR("Intent missing\n");
2413 RETURN(req->rq_status = -EFAULT);
2416 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2418 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
2419 (it->opc & (IT_OPEN | IT_GETATTR | IT_LOOKUP)))
2420 /* we should never allow OBD_CONNECT_ACL if not configured */
2421 repsize[repbufcnt++] = LUSTRE_POSIX_ACL_MAX_SIZE;
2422 else if (it->opc & IT_UNLINK)
2423 repsize[repbufcnt++] = mds->mds_max_cookiesize;
2425 rc = lustre_pack_reply(req, repbufcnt, repsize, NULL);
2427 RETURN(req->rq_status = rc);
2429 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
2430 intent_set_disposition(rep, DISP_IT_EXECD);
2433 /* execute policy */
2434 switch ((long)it->opc) {
2436 case IT_CREAT|IT_OPEN:
2437 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
2438 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock, NULL,
2440 /* XXX swab here to assert that an mds_open reint
2441 * packet is following */
2442 rep->lock_policy_res2 = mds_reint(req, DLM_INTENT_REC_OFF,
2445 /* We abort the lock if the lookup was negative and
2446 * we did not make it to the OPEN portion */
2447 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2448 RETURN(ELDLM_LOCK_ABORTED);
2449 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2450 !intent_disposition(rep, DISP_OPEN_OPEN))
2453 /* If there was an error of some sort or if we are not
2454 * returning any locks */
2455 if (rep->lock_policy_res2 ||
2456 !intent_disposition(rep, DISP_OPEN_LOCK)) {
2457 /* If it is the disconnect error (ENODEV & ENOCONN)
2458 * ptlrpc layer should know this imediately, it should
2459 * be replied by rq_stats, otherwise, return it by
2462 if (IS_CLIENT_DISCONNECT_ERROR(rep->lock_policy_res2))
2463 RETURN(rep->lock_policy_res2);
2465 RETURN(ELDLM_LOCK_ABORTED);
2469 getattr_part = MDS_INODELOCK_LOOKUP;
2471 getattr_part |= MDS_INODELOCK_LOOKUP;
2472 OBD_COUNTER_INCREMENT(req->rq_export->exp_obd, getattr);
2474 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock,
2477 /* INODEBITS_INTEROP: if this lock was converted from a
2478 * plain lock (client does not support inodebits), then
2479 * child lock must be taken with both lookup and update
2480 * bits set for all operations.
2482 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS))
2483 getattr_part = MDS_INODELOCK_LOOKUP |
2484 MDS_INODELOCK_UPDATE;
2486 rep->lock_policy_res2 = mds_getattr_lock(req,DLM_INTENT_REC_OFF,
2487 getattr_part, &lockh);
2488 /* FIXME: LDLM can set req->rq_status. MDS sets
2489 policy_res{1,2} with disposition and status.
2490 - replay: returns 0 & req->status is old status
2491 - otherwise: returns req->status */
2492 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2493 rep->lock_policy_res2 = 0;
2494 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2495 rep->lock_policy_res2)
2496 RETURN(ELDLM_LOCK_ABORTED);
2497 if (req->rq_status != 0) {
2499 rep->lock_policy_res2 = req->rq_status;
2500 RETURN(ELDLM_LOCK_ABORTED);
2504 CERROR("Unhandled intent "LPD64"\n", it->opc);
2508 /* By this point, whatever function we called above must have either
2509 * filled in 'lockh', been an intent replay, or returned an error. We
2510 * want to allow replayed RPCs to not get a lock, since we would just
2511 * drop it below anyways because lock replay is done separately by the
2512 * client afterwards. For regular RPCs we want to give the new lock to
2513 * the client instead of whatever lock it was about to get. */
2514 if (new_lock == NULL)
2515 new_lock = ldlm_handle2lock(&lockh);
2516 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2519 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
2520 it->opc, lockh.cookie);
2522 /* If we've already given this lock to a client once, then we should
2523 * have no readers or writers. Otherwise, we should have one reader
2524 * _or_ writer ref (which will be zeroed below) before returning the
2525 * lock to a client. */
2526 if (new_lock->l_export == req->rq_export) {
2527 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2529 LASSERT(new_lock->l_export == NULL);
2530 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2535 if (new_lock->l_export == req->rq_export) {
2536 /* Already gave this to the client, which means that we
2537 * reconstructed a reply. */
2538 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2540 RETURN(ELDLM_LOCK_REPLACED);
2543 /* Fixup the lock to be given to the client */
2544 lock_res_and_lock(new_lock);
2545 new_lock->l_readers = 0;
2546 new_lock->l_writers = 0;
2548 new_lock->l_export = class_export_get(req->rq_export);
2549 spin_lock(&req->rq_export->exp_ldlm_data.led_lock);
2550 list_add(&new_lock->l_export_chain,
2551 &new_lock->l_export->exp_ldlm_data.led_held_locks);
2552 spin_unlock(&req->rq_export->exp_ldlm_data.led_lock);
2554 new_lock->l_blocking_ast = lock->l_blocking_ast;
2555 new_lock->l_completion_ast = lock->l_completion_ast;
2557 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2558 sizeof(lock->l_remote_handle));
2560 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2562 unlock_res_and_lock(new_lock);
2563 LDLM_LOCK_PUT(new_lock);
2565 RETURN(ELDLM_LOCK_REPLACED);
2568 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
2570 struct mds_obd *mds = &obd->u.mds;
2571 struct lprocfs_static_vars lvars;
2572 int mds_min_threads;
2573 int mds_max_threads;
2577 lprocfs_mdt_init_vars(&lvars);
2578 lprocfs_obd_setup(obd, lvars.obd_vars);
2580 sema_init(&mds->mds_health_sem, 1);
2582 if (mds_num_threads) {
2583 /* If mds_num_threads is set, it is the min and the max. */
2584 if (mds_num_threads > MDS_THREADS_MAX)
2585 mds_num_threads = MDS_THREADS_MAX;
2586 if (mds_num_threads < MDS_THREADS_MIN)
2587 mds_num_threads = MDS_THREADS_MIN;
2588 mds_max_threads = mds_min_threads = mds_num_threads;
2590 /* Base min threads on memory and cpus */
2591 mds_min_threads = num_possible_cpus() * num_physpages >>
2592 (27 - CFS_PAGE_SHIFT);
2593 if (mds_min_threads < MDS_THREADS_MIN)
2594 mds_min_threads = MDS_THREADS_MIN;
2595 /* Largest auto threads start value */
2596 if (mds_min_threads > 32)
2597 mds_min_threads = 32;
2598 mds_max_threads = min(MDS_THREADS_MAX, mds_min_threads * 4);
2602 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2603 MDS_MAXREPSIZE, MDS_REQUEST_PORTAL,
2604 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_FACTOR,
2605 mds_handle, LUSTRE_MDS_NAME,
2606 obd->obd_proc_entry, target_print_req,
2607 mds_min_threads, mds_max_threads, "ll_mdt");
2609 if (!mds->mds_service) {
2610 CERROR("failed to start service\n");
2611 GOTO(err_lprocfs, rc = -ENOMEM);
2614 rc = ptlrpc_start_threads(obd, mds->mds_service);
2616 GOTO(err_thread, rc);
2618 mds->mds_setattr_service =
2619 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2620 MDS_MAXREPSIZE, MDS_SETATTR_PORTAL,
2621 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_FACTOR,
2622 mds_handle, "mds_setattr",
2623 obd->obd_proc_entry, target_print_req,
2624 mds_min_threads, mds_max_threads,
2626 if (!mds->mds_setattr_service) {
2627 CERROR("failed to start getattr service\n");
2628 GOTO(err_thread, rc = -ENOMEM);
2631 rc = ptlrpc_start_threads(obd, mds->mds_setattr_service);
2633 GOTO(err_thread2, rc);
2635 mds->mds_readpage_service =
2636 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2637 MDS_MAXREPSIZE, MDS_READPAGE_PORTAL,
2638 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_FACTOR,
2639 mds_handle, "mds_readpage",
2640 obd->obd_proc_entry, target_print_req,
2641 MDS_THREADS_MIN_READPAGE, mds_max_threads,
2643 if (!mds->mds_readpage_service) {
2644 CERROR("failed to start readpage service\n");
2645 GOTO(err_thread2, rc = -ENOMEM);
2648 rc = ptlrpc_start_threads(obd, mds->mds_readpage_service);
2651 GOTO(err_thread3, rc);
2653 ping_evictor_start();
2658 ptlrpc_unregister_service(mds->mds_readpage_service);
2659 mds->mds_readpage_service = NULL;
2661 ptlrpc_unregister_service(mds->mds_setattr_service);
2662 mds->mds_setattr_service = NULL;
2664 ptlrpc_unregister_service(mds->mds_service);
2665 mds->mds_service = NULL;
2667 lprocfs_obd_cleanup(obd);
2671 static int mdt_cleanup(struct obd_device *obd)
2673 struct mds_obd *mds = &obd->u.mds;
2676 ping_evictor_stop();
2678 down(&mds->mds_health_sem);
2679 ptlrpc_unregister_service(mds->mds_readpage_service);
2680 ptlrpc_unregister_service(mds->mds_setattr_service);
2681 ptlrpc_unregister_service(mds->mds_service);
2682 mds->mds_readpage_service = NULL;
2683 mds->mds_setattr_service = NULL;
2684 mds->mds_service = NULL;
2685 up(&mds->mds_health_sem);
2687 lprocfs_obd_cleanup(obd);
2692 static int mdt_health_check(struct obd_device *obd)
2694 struct mds_obd *mds = &obd->u.mds;
2697 down(&mds->mds_health_sem);
2698 rc |= ptlrpc_service_health_check(mds->mds_readpage_service);
2699 rc |= ptlrpc_service_health_check(mds->mds_setattr_service);
2700 rc |= ptlrpc_service_health_check(mds->mds_service);
2701 up(&mds->mds_health_sem);
2704 * health_check to return 0 on healthy
2705 * and 1 on unhealthy.
2713 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2716 struct obd_device *obd = data;
2719 fid.generation = gen;
2720 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2723 static int mds_health_check(struct obd_device *obd)
2725 struct obd_device_target *odt = &obd->u.obt;
2726 #ifdef USE_HEALTH_CHECK_WRITE
2727 struct mds_obd *mds = &obd->u.mds;
2731 if (odt->obt_sb->s_flags & MS_RDONLY)
2734 #ifdef USE_HEALTH_CHECK_WRITE
2735 LASSERT(mds->mds_health_check_filp != NULL);
2736 rc |= !!lvfs_check_io_health(obd, mds->mds_health_check_filp);
2742 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
2744 struct lustre_cfg *lcfg = buf;
2745 struct lprocfs_static_vars lvars;
2748 lprocfs_mds_init_vars(&lvars);
2750 rc = class_process_proc_param(PARAM_MDT, lvars.obd_vars, lcfg, obd);
2755 struct lvfs_callback_ops mds_lvfs_ops = {
2756 l_fid2dentry: mds_lvfs_fid2dentry,
2759 /* use obd ops to offer management infrastructure */
2760 static struct obd_ops mds_obd_ops = {
2761 .o_owner = THIS_MODULE,
2762 .o_connect = mds_connect,
2763 .o_reconnect = mds_reconnect,
2764 .o_init_export = mds_init_export,
2765 .o_destroy_export = mds_destroy_export,
2766 .o_disconnect = mds_disconnect,
2767 .o_setup = mds_setup,
2768 .o_precleanup = mds_precleanup,
2769 .o_cleanup = mds_cleanup,
2770 .o_postrecov = mds_postrecov,
2771 .o_statfs = mds_obd_statfs,
2772 .o_iocontrol = mds_iocontrol,
2773 .o_create = mds_obd_create,
2774 .o_destroy = mds_obd_destroy,
2775 .o_llog_init = mds_llog_init,
2776 .o_llog_finish = mds_llog_finish,
2777 .o_notify = mds_notify,
2778 .o_health_check = mds_health_check,
2779 .o_process_config = mds_process_config,
2782 static struct obd_ops mdt_obd_ops = {
2783 .o_owner = THIS_MODULE,
2784 .o_setup = mdt_setup,
2785 .o_cleanup = mdt_cleanup,
2786 .o_health_check = mdt_health_check,
2789 quota_interface_t *mds_quota_interface_ref;
2790 extern quota_interface_t mds_quota_interface;
2792 static int __init mds_init(void)
2795 struct lprocfs_static_vars lvars;
2797 request_module("lquota");
2798 mds_quota_interface_ref = PORTAL_SYMBOL_GET(mds_quota_interface);
2799 rc = lquota_init(mds_quota_interface_ref);
2801 if (mds_quota_interface_ref)
2802 PORTAL_SYMBOL_PUT(mds_quota_interface);
2805 init_obd_quota_ops(mds_quota_interface_ref, &mds_obd_ops);
2807 lprocfs_mds_init_vars(&lvars);
2808 class_register_type(&mds_obd_ops, lvars.module_vars, LUSTRE_MDS_NAME);
2809 lprocfs_mdt_init_vars(&lvars);
2810 class_register_type(&mdt_obd_ops, lvars.module_vars, LUSTRE_MDT_NAME);
2815 static void /*__exit*/ mds_exit(void)
2817 lquota_exit(mds_quota_interface_ref);
2818 if (mds_quota_interface_ref)
2819 PORTAL_SYMBOL_PUT(mds_quota_interface);
2821 class_unregister_type(LUSTRE_MDS_NAME);
2822 class_unregister_type(LUSTRE_MDT_NAME);
2825 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2826 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2827 MODULE_LICENSE("GPL");
2829 module_init(mds_init);
2830 module_exit(mds_exit);