* lustre/mds/handler.c
* Lustre Metadata Server (mds) request handler
*
- * Copyright (c) 2001, 2002 Cluster File Systems, Inc.
+ * Copyright (c) 2001-2003 Cluster File Systems, Inc.
* Author: Peter Braam <braam@clusterfs.com>
* Author: Andreas Dilger <adilger@clusterfs.com>
* Author: Phil Schwan <phil@clusterfs.com>
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
*/
-#define EXPORT_SYMTAB
+#ifndef EXPORT_SYMTAB
+# define EXPORT_SYMTAB
+#endif
#define DEBUG_SUBSYSTEM S_MDS
#include <linux/module.h>
#include <linux/init.h>
#include <linux/obd_class.h>
#include <linux/random.h>
-#include <linux/locks.h>
+#include <linux/fs.h>
+#include <linux/jbd.h>
+#include <linux/ext3_fs.h>
#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
-#include <linux/buffer_head.h>
+# include <linux/smp_lock.h>
+# include <linux/buffer_head.h>
+# include <linux/workqueue.h>
+# include <linux/mount.h>
+#else
+# include <linux/locks.h>
#endif
#include <linux/obd_lov.h>
+#include <linux/lustre_mds.h>
+#include <linux/lustre_fsfilt.h>
#include <linux/lprocfs_status.h>
+#include <linux/lustre_commit_confd.h>
-static kmem_cache_t *mds_file_cache;
-
-extern int mds_get_lovtgts(struct mds_obd *obd, int tgt_count,
- obd_uuid_t *uuidarray);
-extern int mds_get_lovdesc(struct mds_obd *obd, struct lov_desc *desc);
-extern void mds_start_transno(struct mds_obd *mds);
-extern int mds_finish_transno(struct mds_obd *mds, void *handle,
- struct ptlrpc_request *req, int rc);
-static int mds_cleanup(struct obd_device * obddev);
-
-extern struct lprocfs_vars status_var_nm_1[];
-extern struct lprocfs_vars status_class_var[];
-
-inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
-{
- return &req->rq_export->exp_obd->u.mds;
-}
-
-static int mds_bulk_timeout(void *data)
-{
- struct ptlrpc_bulk_desc *desc = data;
+#include "mds_internal.h"
- ENTRY;
- recovd_conn_fail(desc->bd_connection);
- RETURN(1);
-}
+static int mds_intent_policy(struct ldlm_namespace *ns,
+ struct ldlm_lock **lockp, void *req_cookie,
+ ldlm_mode_t mode, int flags, void *data);
+static int mds_postsetup(struct obd_device *obd);
+static int mds_cleanup(struct obd_device *obd, int flags);
/* Assumes caller has already pushed into the kernel filesystem context */
static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
- __u64 offset)
+ loff_t offset, int count)
{
- int rc = 0;
- struct mds_obd *mds = mds_req2mds(req);
struct ptlrpc_bulk_desc *desc;
- struct ptlrpc_bulk_page *bulk;
struct l_wait_info lwi;
- char *buf;
+ struct page **pages;
+ int rc = 0, npages, i, tmpcount, tmpsize = 0;
ENTRY;
- desc = ptlrpc_prep_bulk(req->rq_connection);
- if (desc == NULL)
+ LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
+
+ npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
+ OBD_ALLOC(pages, sizeof(*pages) * npages);
+ if (!pages)
GOTO(out, rc = -ENOMEM);
- bulk = ptlrpc_prep_bulk_page(desc);
- if (bulk == NULL)
- GOTO(cleanup_bulk, rc = -ENOMEM);
+ desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
+ MDS_BULK_PORTAL);
+ if (desc == NULL)
+ GOTO(out_free, rc = -ENOMEM);
+
+ for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
+ tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
+
+ pages[i] = alloc_pages(GFP_KERNEL, 0);
+ if (pages[i] == NULL)
+ GOTO(cleanup_buf, rc = -ENOMEM);
+
+ ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
+ }
- OBD_ALLOC(buf, PAGE_SIZE);
- if (buf == NULL)
- GOTO(cleanup_bulk, rc = -ENOMEM);
+ for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
+ tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
+ CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
+ tmpsize, offset, file->f_dentry->d_inode->i_ino,
+ file->f_dentry->d_inode->i_size);
- rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
+ rc = fsfilt_readpage(req->rq_export->exp_obd, file,
+ kmap(pages[i]), tmpsize, &offset);
+ kunmap(pages[i]);
- if (rc != PAGE_SIZE)
- GOTO(cleanup_buf, rc = -EIO);
+ if (rc != tmpsize)
+ GOTO(cleanup_buf, rc = -EIO);
+ }
- bulk->bp_xid = req->rq_xid;
- bulk->bp_buf = buf;
- bulk->bp_buflen = PAGE_SIZE;
- desc->bd_portal = MDS_BULK_PORTAL;
+ LASSERT(desc->bd_nob == count);
- rc = ptlrpc_send_bulk(desc);
+ rc = ptlrpc_start_bulk_transfer(desc);
if (rc)
GOTO(cleanup_buf, rc);
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
OBD_FAIL_MDS_SENDPAGE, rc);
- ptlrpc_abort_bulk(desc);
- GOTO(cleanup_buf, rc);
+ GOTO(abort_bulk, rc);
}
- lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
- rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_SENT,
- &lwi);
- if (rc) {
- if (rc != -ETIMEDOUT)
- LBUG();
- GOTO(cleanup_buf, rc);
+ lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
+ rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
+ LASSERT (rc == 0 || rc == -ETIMEDOUT);
+
+ if (rc == 0) {
+ if (desc->bd_success &&
+ desc->bd_nob_transferred == count)
+ GOTO(cleanup_buf, rc);
+
+ rc = -ETIMEDOUT; /* XXX should this be a different errno? */
}
+ DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
+ (rc == -ETIMEDOUT) ? "timeout" : "network error",
+ desc->bd_nob_transferred, count,
+ req->rq_export->exp_client_uuid.uuid,
+ req->rq_export->exp_connection->c_remote_uuid.uuid);
+
+ ptlrpc_fail_export(req->rq_export);
+
EXIT;
+ abort_bulk:
+ ptlrpc_abort_bulk (desc);
cleanup_buf:
- OBD_FREE(buf, PAGE_SIZE);
- cleanup_bulk:
+ for (i = 0; i < npages; i++)
+ if (pages[i])
+ __free_pages(pages[i], 0);
+
ptlrpc_free_bulk(desc);
+ out_free:
+ OBD_FREE(pages, sizeof(*pages) * npages);
out:
return rc;
}
-/*
- * Look up a named entry in a directory, and get an LDLM lock on it.
- * 'dir' is a inode for which an LDLM lock has already been taken.
- *
- * If we do not need an exclusive or write lock on this entry (e.g.
- * a read lock for attribute lookup only) then we do not hold the
- * directory on return. It is up to the caller to know what type
- * of lock it is getting, and clean up appropriately.
- */
-struct dentry *mds_name2locked_dentry(struct obd_device *obd,
- struct dentry *dir, struct vfsmount **mnt,
- char *name, int namelen, int lock_mode,
- struct lustre_handle *lockh,
- int dir_lock_mode)
+int mds_lock_mode_for_dir(struct obd_device *obd,
+ struct dentry *dentry, int mode)
{
- struct dentry *dchild;
- int flags = 0, rc;
- __u64 res_id[3] = {0};
- ENTRY;
-
- down(&dir->d_inode->i_sem);
- dchild = lookup_one_len(name, dir, namelen);
- if (IS_ERR(dchild)) {
- CERROR("child lookup error %ld\n", PTR_ERR(dchild));
- up(&dir->d_inode->i_sem);
- LBUG();
- RETURN(dchild);
- }
- if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
- up(&dir->d_inode->i_sem);
- ldlm_lock_decref(lockh, dir_lock_mode);
- }
-
- if (lock_mode == 0 || !dchild->d_inode)
- RETURN(dchild);
-
- res_id[0] = dchild->d_inode->i_ino;
- res_id[1] = dchild->d_inode->i_generation;
- rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
- res_id, LDLM_PLAIN, NULL, 0, lock_mode,
- &flags, ldlm_completion_ast,
- mds_blocking_ast, NULL, 0, lockh);
- if (rc != ELDLM_OK) {
- l_dput(dchild);
- up(&dir->d_inode->i_sem);
- RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
+ int ret_mode;
+
+ /* any dir access needs couple locks:
+ * 1) on part of dir we gonna lookup/modify in
+ * 2) on a whole dir to protect it from concurrent splitting
+ * and to flush client's cache for readdir()
+ * so, for a given mode and dentry this routine decides what
+ * lock mode to use for lock #2:
+ * 1) if caller's gonna lookup in dir then we need to protect
+ * dir from being splitted only - LCK_CR
+ * 2) if caller's gonna modify dir then we need to protect
+ * dir from being splitted and to flush cache - LCK_CW
+ * 3) if caller's gonna modify dir and that dir seems ready
+ * for splitting then we need to protect it from any
+ * type of access (lookup/modify/split) - LCK_EX -bzzz */
+
+ if (mode == LCK_PR) {
+ ret_mode = LCK_CR;
+ } else if (mode == LCK_PW) {
+ /* caller gonna modify directory.we use concurrent
+ write lock here to retract client's cache for readdir */
+ ret_mode = LCK_CW;
+ if (mds_splitting_expected(obd, dentry)) {
+ /* splitting possible. serialize any access */
+ CERROR("%s: gonna split %lu/%lu\n",
+ obd->obd_name,
+ (unsigned long) dentry->d_inode->i_ino,
+ (unsigned long) dentry->d_inode->i_generation);
+ ret_mode = LCK_EX;
+ }
+ } else {
+ CWARN("unexpected lock mode %d\n", mode);
+ ret_mode = LCK_EX;
}
-
- RETURN(dchild);
+ return ret_mode;
}
+/* only valid locked dentries or errors should be returned */
struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
struct vfsmount **mnt, int lock_mode,
- struct lustre_handle *lockh)
+ struct lustre_handle *lockh, int *mode,
+ char *name, int namelen, __u64 lockpart)
{
struct mds_obd *mds = &obd->u.mds;
struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
+ struct ldlm_res_id res_id = { .name = {0} };
int flags = 0, rc;
- __u64 res_id[3] = {0};
+ ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
+
ENTRY;
if (IS_ERR(de))
RETURN(de);
- res_id[0] = de->d_inode->i_ino;
- res_id[1] = de->d_inode->i_generation;
- rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
- res_id, LDLM_PLAIN, NULL, 0, lock_mode,
- &flags, ldlm_completion_ast,
- mds_blocking_ast, NULL, 0, lockh);
+ res_id.name[0] = de->d_inode->i_ino;
+ res_id.name[1] = de->d_inode->i_generation;
+ lockh[1].cookie = 0;
+#ifdef S_PDIROPS
+ if (name && IS_PDIROPS(de->d_inode)) {
+ ldlm_policy_data_t cpolicy =
+ { .l_inodebits = { MDS_INODELOCK_UPDATE } };
+ LASSERT(mode != NULL);
+ *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
+ rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
+ res_id, LDLM_IBITS,
+ &cpolicy, *mode, &flags,
+ mds_blocking_ast,
+ ldlm_completion_ast, NULL, NULL,
+ NULL, 0, NULL, lockh + 1);
+ if (rc != ELDLM_OK) {
+ l_dput(de);
+ RETURN(ERR_PTR(-ENOLCK));
+ }
+ flags = 0;
+
+ res_id.name[2] = full_name_hash(name, namelen);
+
+ CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
+ de->d_inode->i_ino, de->d_inode->i_generation,
+ res_id.name[2]);
+ }
+#endif
+ rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
+ LDLM_IBITS, &policy, lock_mode, &flags,
+ mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
+ NULL, 0, NULL, lockh);
if (rc != ELDLM_OK) {
l_dput(de);
- retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
+ retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
+#ifdef S_PDIROPS
+ if (lockh[1].cookie)
+ ldlm_lock_decref(lockh + 1, LCK_CW);
+#endif
}
RETURN(retval);
#define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
#endif
+
/* Look up an entry by inode number. */
+/* this function ONLY returns valid dget'd dentries with an initialized inode
+ or errors */
struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
struct vfsmount **mnt)
{
- /* stolen from NFS */
- struct super_block *sb = mds->mds_sb;
+ char fid_name[32];
unsigned long ino = fid->id;
__u32 generation = fid->generation;
struct inode *inode;
- struct list_head *lp;
struct dentry *result;
if (ino == 0)
RETURN(ERR_PTR(-ESTALE));
- inode = iget(sb, ino);
- if (inode == NULL)
- RETURN(ERR_PTR(-ENOMEM));
+ snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
+
+ CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
+ ino, generation, mds->mds_sb);
- CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
+ /* under ext3 this is neither supposed to return bad inodes
+ nor NULL inodes. */
+ result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
+ if (IS_ERR(result))
+ RETURN(result);
+
+ inode = result->d_inode;
+ if (!inode)
+ RETURN(ERR_PTR(-ENOENT));
- if (is_bad_inode(inode) ||
- (generation && inode->i_generation != generation)) {
+#if 0
+ /* here we disabled generation check, as root inode i_generation
+ * of cache mds and real mds are different. */
+ if (generation && inode->i_generation != generation) {
/* we didn't find the right inode.. */
- CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
- inode->i_ino, inode->i_nlink,
+ CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
+ inode->i_ino, (unsigned long)inode->i_nlink,
atomic_read(&inode->i_count), inode->i_generation,
generation);
- iput(inode);
+ dput(result);
RETURN(ERR_PTR(-ENOENT));
}
+#endif
- /* now to find a dentry. If possible, get a well-connected one */
- if (mnt)
+ if (mnt) {
*mnt = mds->mds_vfsmnt;
- spin_lock(&dcache_lock);
- list_for_each(lp, &inode->i_dentry) {
- result = list_entry(lp, struct dentry, d_alias);
- if (!(result->d_flags & DCACHE_DISCONNECTED)) {
- dget_locked(result);
- result->d_vfs_flags |= DCACHE_REFERENCED;
- spin_unlock(&dcache_lock);
- iput(inode);
- if (mnt)
- mntget(*mnt);
- return result;
- }
- }
- spin_unlock(&dcache_lock);
- result = d_alloc_root(inode);
- if (result == NULL) {
- iput(inode);
- return ERR_PTR(-ENOMEM);
- }
- if (mnt)
mntget(*mnt);
- result->d_flags |= DCACHE_DISCONNECTED;
- return result;
+ }
+
+ RETURN(result);
}
+
/* Establish a connection to the MDS.
*
* This will set up an export structure for the client to hold state data
* on the server, etc.
*/
static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
- obd_uuid_t cluuid, struct recovd_obd *recovd,
- ptlrpc_recovery_cb_t recover)
+ struct obd_uuid *cluuid)
{
struct obd_export *exp;
- struct mds_export_data *med;
+ struct mds_export_data *med; /* */
struct mds_client_data *mcd;
- struct list_head *p;
- int rc;
+ int rc, abort_recovery;
ENTRY;
if (!conn || !obd || !cluuid)
RETURN(-EINVAL);
- MOD_INC_USE_COUNT;
-
- spin_lock(&obd->obd_dev_lock);
- list_for_each(p, &obd->obd_exports) {
- exp = list_entry(p, struct obd_export, exp_obd_chain);
- mcd = exp->exp_mds_data.med_mcd;
- if (!mcd) {
- CERROR("FYI: NULL mcd - simultaneous connects\n");
- continue;
- }
- if (!memcmp(cluuid, mcd->mcd_uuid, sizeof mcd->mcd_uuid)) {
- /* XXX make handle-found-export a subroutine */
- LASSERT(exp->exp_obd == obd);
-
- spin_unlock(&obd->obd_dev_lock);
- if (exp->exp_connection) {
- struct lustre_handle *hdl;
- hdl = &exp->exp_ldlm_data.led_import.imp_handle;
- /* Might be a re-connect after a partition. */
- if (!memcmp(conn, hdl, sizeof *conn)) {
- CERROR("%s reconnecting\n", cluuid);
- conn->addr = (__u64) (unsigned long)exp;
- conn->cookie = exp->exp_cookie;
- rc = EALREADY;
- } else {
- CERROR("%s reconnecting from %s, "
- "handle mismatch (ours %Lx/%Lx, "
- "theirs %Lx/%Lx)\n", cluuid,
- exp->exp_connection->
- c_remote_uuid, hdl->addr,
- hdl->cookie, conn->addr,
- conn->cookie);
- /* XXX disconnect them here? */
- memset(conn, 0, sizeof *conn);
- rc = -EALREADY;
- }
- MOD_DEC_USE_COUNT;
- RETURN(rc);
- }
- conn->addr = (__u64) (unsigned long)exp;
- conn->cookie = exp->exp_cookie;
- CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
- cluuid, exp);
- CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
- (long long)conn->addr, (long long)conn->cookie);
- RETURN(0);
- }
- }
- spin_unlock(&obd->obd_dev_lock);
-
- if (obd->u.mds.mds_recoverable_clients != 0) {
- CERROR("denying connection for new client %s: in recovery\n",
- cluuid);
- MOD_DEC_USE_COUNT;
- RETURN(-EBUSY);
- }
+ /* Check for aborted recovery. */
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ abort_recovery = obd->obd_abort_recovery;
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ if (abort_recovery)
+ target_abort_recovery(obd);
/* XXX There is a small race between checking the list and adding a
* new connection for the same UUID, but the real threat (list
*/
rc = class_connect(conn, obd, cluuid);
if (rc)
- GOTO(out_dec, rc);
+ RETURN(rc);
exp = class_conn2export(conn);
LASSERT(exp);
med = &exp->exp_mds_data;
OBD_ALLOC(mcd, sizeof(*mcd));
if (!mcd) {
CERROR("mds: out of memory for client data\n");
- GOTO(out_export, rc = -ENOMEM);
+ GOTO(out, rc = -ENOMEM);
}
memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
med->med_mcd = mcd;
- INIT_LIST_HEAD(&med->med_open_head);
- spin_lock_init(&med->med_open_lock);
-
- rc = mds_client_add(&obd->u.mds, med, -1);
- if (rc)
- GOTO(out_mcd, rc);
-
- RETURN(0);
-
-out_mcd:
- OBD_FREE(mcd, sizeof(*mcd));
-out_export:
- class_disconnect(conn);
-out_dec:
- MOD_DEC_USE_COUNT;
+ rc = mds_client_add(obd, &obd->u.mds, med, -1);
+ if (rc == 0)
+ EXIT;
+out:
+ if (rc) {
+ OBD_FREE(mcd, sizeof(*mcd));
+ class_disconnect(exp, 0);
+ }
+ class_export_put(exp);
return rc;
}
-/* Call with med->med_open_lock held, please. */
-inline int mds_close_mfd(struct mds_file_data *mfd, struct mds_export_data *med)
+static int mds_init_export(struct obd_export *exp)
{
- struct file *file = mfd->mfd_file;
- LASSERT(file->private_data == mfd);
-
- list_del(&mfd->mfd_list);
- mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
- kmem_cache_free(mds_file_cache, mfd);
+ struct mds_export_data *med = &exp->exp_mds_data;
- return filp_close(file, 0);
+ INIT_LIST_HEAD(&med->med_open_head);
+ spin_lock_init(&med->med_open_lock);
+ RETURN(0);
}
-static int mds_disconnect(struct lustre_handle *conn)
+static int mds_destroy_export(struct obd_export *export)
{
- struct obd_export *export = class_conn2export(conn);
- struct list_head *tmp, *n;
- struct mds_export_data *med = &export->exp_mds_data;
- int rc;
+ struct mds_export_data *med;
+ struct obd_device *obd = export->exp_obd;
+ struct lvfs_run_ctxt saved;
+ int rc = 0;
ENTRY;
- /*
- * Close any open files.
- */
+ med = &export->exp_mds_data;
+ target_destroy_export(export);
+
+ if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
+ GOTO(out, 0);
+
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
+
+ /* Close any open files (which may also cause orphan unlinking). */
spin_lock(&med->med_open_lock);
- list_for_each_safe(tmp, n, &med->med_open_head) {
+ while (!list_empty(&med->med_open_head)) {
+ struct list_head *tmp = med->med_open_head.next;
struct mds_file_data *mfd =
list_entry(tmp, struct mds_file_data, mfd_list);
- rc = mds_close_mfd(mfd, med);
- if (rc) {
- /* XXX better diagnostics, with file path and stuff */
- CDEBUG(D_INODE, "Error %d closing mfd %p\n", rc, mfd);
- }
+ BDEVNAME_DECLARE_STORAGE(btmp);
+
+ /* bug 1579: fix force-closing for 2.5 */
+ struct dentry *dentry = mfd->mfd_dentry;
+
+ list_del(&mfd->mfd_list);
+ spin_unlock(&med->med_open_lock);
+
+ /* If you change this message, be sure to update
+ * replay_single:test_46 */
+ CERROR("force closing client file handle for %*s (%s:%lu)\n",
+ dentry->d_name.len, dentry->d_name.name,
+ ll_bdevname(dentry->d_inode->i_sb, btmp),
+ dentry->d_inode->i_ino);
+ rc = mds_mfd_close(NULL, obd, mfd,
+ !(export->exp_flags & OBD_OPT_FAILOVER));
+
+ if (rc)
+ CDEBUG(D_INODE, "Error closing file: %d\n", rc);
+ spin_lock(&med->med_open_lock);
}
spin_unlock(&med->med_open_lock);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
- ldlm_cancel_locks_for_export(export);
- mds_client_free(export);
-
- rc = class_disconnect(conn);
- if (!rc)
- MOD_DEC_USE_COUNT;
+out:
+ mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
RETURN(rc);
}
-/*
- * XXX This is NOT guaranteed to flush all transactions to disk (even though
- * it is equivalent to calling sync()) because it only _starts_ the flush
- * and does not wait for completion. It's better than nothing though.
- * What we really want is a mild form of fsync_dev_lockfs(), but it is
- * non-standard, or enabling do_sync_supers in ext3, just for this call.
- */
-static void mds_fsync_super(struct super_block *sb)
+static int mds_disconnect(struct obd_export *exp, int flags)
{
- lock_kernel();
- lock_super(sb);
- if (sb->s_dirt && sb->s_op && sb->s_op->write_super)
- sb->s_op->write_super(sb);
- unlock_super(sb);
- unlock_kernel();
+ struct obd_device *obd;
+ struct mds_obd *mds;
+ unsigned long irqflags;
+ int rc;
+ ENTRY;
+
+ LASSERT(exp);
+ class_export_get(exp);
+
+ obd = class_exp2obd(exp);
+ if (obd == NULL) {
+ CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
+ exp->exp_handle.h_cookie);
+ RETURN(-EINVAL);
+ }
+ mds = &obd->u.mds;
+
+ if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
+ && !atomic_read(&mds->mds_real_clients)) {
+ /* there was no client at all */
+ mds_lmv_disconnect(obd, flags);
+ }
+
+ if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
+ && atomic_dec_and_test(&mds->mds_real_clients)) {
+ /* time to drop LMV connections */
+ CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
+ "Disconnnect from LMV now\n",
+ obd->obd_name, exp->exp_client_uuid.uuid);
+ mds_lmv_disconnect(obd, flags);
+ }
+
+ spin_lock_irqsave(&exp->exp_lock, irqflags);
+ exp->exp_flags = flags;
+ spin_unlock_irqrestore(&exp->exp_lock, irqflags);
+
+ /* Disconnect early so that clients can't keep using export */
+ rc = class_disconnect(exp, flags);
+ ldlm_cancel_locks_for_export(exp);
+
+ /* complete all outstanding replies */
+ spin_lock_irqsave(&exp->exp_lock, irqflags);
+ while (!list_empty(&exp->exp_outstanding_replies)) {
+ struct ptlrpc_reply_state *rs =
+ list_entry(exp->exp_outstanding_replies.next,
+ struct ptlrpc_reply_state, rs_exp_list);
+ struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
+
+ spin_lock(&svc->srv_lock);
+ list_del_init(&rs->rs_exp_list);
+ ptlrpc_schedule_difficult_reply(rs);
+ spin_unlock(&svc->srv_lock);
+ }
+ spin_unlock_irqrestore(&exp->exp_lock, irqflags);
+
+ class_export_put(exp);
+ RETURN(rc);
}
static int mds_getstatus(struct ptlrpc_request *req)
int rc, size = sizeof(*body);
ENTRY;
- rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
+ rc = lustre_pack_reply(req, 1, &size, NULL);
if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
CERROR("mds: out of memory for message: size=%d\n", size);
- req->rq_status = -ENOMEM;
- RETURN(0);
+ req->rq_status = -ENOMEM; /* superfluous? */
+ RETURN(-ENOMEM);
}
- /* Flush any outstanding transactions to disk so the client will
- * get the latest last_committed value and can drop their local
- * requests if they have any. This would be fsync_super() if it
- * was exported.
- */
- mds_fsync_super(mds->mds_sb);
-
- body = lustre_msg_buf(req->rq_repmsg, 0);
+ body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
/* the last_committed and last_xid fields are filled in for all
RETURN(0);
}
-static int mds_getlovinfo(struct ptlrpc_request *req)
-{
- struct mds_obd *mds = mds_req2mds(req);
- struct mds_status_req *streq;
- struct lov_desc *desc;
- int tgt_count;
- int rc, size[2] = {sizeof(*desc)};
- ENTRY;
-
- streq = lustre_msg_buf(req->rq_reqmsg, 0);
- streq->flags = NTOH__u32(streq->flags);
- streq->repbuf = NTOH__u32(streq->repbuf);
- size[1] = streq->repbuf;
-
- rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
- if (rc) {
- CERROR("mds: out of memory for message: size=%d\n", size[1]);
- req->rq_status = -ENOMEM;
- RETURN(0);
- }
-
- desc = lustre_msg_buf(req->rq_repmsg, 0);
- rc = mds_get_lovdesc(mds, desc);
- if (rc) {
- CERROR("mds_get_lovdesc error %d", rc);
- req->rq_status = rc;
- RETURN(0);
- }
-
- tgt_count = le32_to_cpu(desc->ld_tgt_count);
- if (tgt_count * sizeof(obd_uuid_t) > streq->repbuf) {
- CERROR("too many targets, enlarge client buffers\n");
- req->rq_status = -ENOSPC;
- RETURN(0);
- }
-
- /* XXX the MDS should not really know about this */
- mds->mds_max_mdsize = lov_mds_md_size(tgt_count);
- rc = mds_get_lovtgts(mds, tgt_count,
- lustre_msg_buf(req->rq_repmsg, 1));
- if (rc) {
- CERROR("get_lovtgts error %d\n", rc);
- req->rq_status = rc;
- RETURN(0);
- }
- RETURN(0);
-}
-
int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
- void *data, __u32 data_len, int flag)
+ void *data, int flag)
{
int do_ast;
ENTRY;
/* XXX layering violation! -phil */
l_lock(&lock->l_resource->lr_namespace->ns_lock);
+ /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
+ * such that mds_blocking_ast is called just before l_i_p takes the
+ * ns_lock, then by the time we get the lock, we might not be the
+ * correct blocking function anymore. So check, and return early, if
+ * so. */
+ if (lock->l_blocking_ast != mds_blocking_ast) {
+ l_unlock(&lock->l_resource->lr_namespace->ns_lock);
+ RETURN(0);
+ }
+
lock->l_flags |= LDLM_FL_CBPENDING;
do_ast = (!lock->l_readers && !lock->l_writers);
l_unlock(&lock->l_resource->lr_namespace->ns_lock);
rc = ldlm_cli_cancel(&lockh);
if (rc < 0)
CERROR("ldlm_cli_cancel: %d\n", rc);
- } else
- LDLM_DEBUG(lock, "Lock still has references, will be"
+ } else {
+ LDLM_DEBUG(lock, "Lock still has references, will be "
"cancelled later");
+ }
RETURN(0);
}
-int mds_pack_md(struct mds_obd *mds, struct ptlrpc_request *req,
- int offset, struct mds_body *body, struct inode *inode)
+int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
+ int *size, int lock)
+{
+ int rc = 0;
+ int lmm_size;
+
+ if (lock)
+ down(&inode->i_sem);
+ rc = fsfilt_get_md(obd, inode, md, *size);
+ if (lock)
+ up(&inode->i_sem);
+
+ if (rc < 0) {
+ CERROR("Error %d reading eadata for ino %lu\n",
+ rc, inode->i_ino);
+ } else if (rc > 0) {
+ lmm_size = rc;
+
+ if (S_ISREG(inode->i_mode))
+ rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
+
+ if (rc == 0) {
+ *size = lmm_size;
+ rc = lmm_size;
+ } else if (rc > 0) {
+ *size = rc;
+ }
+ }
+
+ RETURN (rc);
+}
+
+
+/* Call with lock=1 if you want mds_pack_md to take the i_sem.
+ * Call with lock=0 if the caller has already taken the i_sem. */
+int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
+ struct mds_body *body, struct inode *inode, int lock)
{
- struct lov_mds_md *lmm;
- int lmm_size = req->rq_repmsg->buflens[offset];
+ struct mds_obd *mds = &obd->u.mds;
+ void *lmm;
+ int lmm_size;
int rc;
+ ENTRY;
- if (lmm_size == 0) {
- CDEBUG(D_INFO, "no space reserved for inode %u MD\n", inode->i_ino);
+ lmm = lustre_msg_buf(msg, offset, 0);
+ if (lmm == NULL) {
+ /* Some problem with getting eadata when I sized the reply
+ * buffer... */
+ CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
+ inode->i_ino);
RETURN(0);
}
-
- lmm = lustre_msg_buf(req->rq_repmsg, offset);
+ lmm_size = msg->buflens[offset];
/* I don't really like this, but it is a sanity check on the client
* MD request. However, if the client doesn't know how much space
- * to reserve for the MD, this shouldn't be fatal either...
+ * to reserve for the MD, it shouldn't be bad to have too much space.
*/
if (lmm_size > mds->mds_max_mdsize) {
- CERROR("Reading MD for inode %u of %d bytes > max %d\n",
+ CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
inode->i_ino, lmm_size, mds->mds_max_mdsize);
// RETURN(-EINVAL);
}
- /* We don't need to store the reply size, because this buffer is
- * discarded right after unpacking, and the LOV can figure out the
- * size itself from the ost count.
- */
- if ((rc = mds_fs_get_md(mds, inode, lmm, lmm_size)) < 0) {
- CDEBUG(D_INFO, "No md for ino %u: rc = %d\n", inode->i_ino, rc);
- } else if (rc > 0) {
- body->valid |= OBD_MD_FLEASIZE;
+ rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
+ if (rc > 0) {
+ if (S_ISDIR(inode->i_mode))
+ body->valid |= OBD_MD_FLDIREA;
+ else
+ body->valid |= OBD_MD_FLEASIZE;
+ body->eadatasize = lmm_size;
rc = 0;
}
- return rc;
+ RETURN(rc);
}
-static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
+static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
struct ptlrpc_request *req,
struct mds_body *reqbody, int reply_off)
{
int rc = 0;
ENTRY;
- if (inode == NULL)
+ if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
RETURN(-ENOENT);
- body = lustre_msg_buf(req->rq_repmsg, reply_off);
+ body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
+ LASSERT(body != NULL); /* caller prepped reply */
+
+ if (dentry->d_flags & DCACHE_CROSS_REF) {
+ CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
+ (unsigned long) dentry->d_mdsnum,
+ (unsigned long) dentry->d_inum,
+ (unsigned long) dentry->d_generation);
+ body->valid |= OBD_MD_FLID | OBD_MD_MDS;
+ body->fid1.id = dentry->d_inum;
+ body->fid1.mds = dentry->d_mdsnum;
+ body->fid1.generation = dentry->d_generation;
+ RETURN(0);
+ }
+ mds_pack_inode2fid(obd, &body->fid1, inode);
+ mds_pack_inode2body(obd, body, inode);
+
+ if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
+ (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
+ rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
+ inode, 1);
- mds_pack_inode2fid(&body->fid1, inode);
- mds_pack_inode2body(body, inode);
+ /* If we have LOV EA data, the OST holds size, atime, mtime */
+ if (!(body->valid & OBD_MD_FLEASIZE) &&
+ !(body->valid & OBD_MD_FLDIREA))
+ body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
+ OBD_MD_FLATIME | OBD_MD_FLMTIME);
+ } else if (S_ISLNK(inode->i_mode) &&
+ (reqbody->valid & OBD_MD_LINKNAME) != 0) {
+ char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
+ int len;
- if (S_ISREG(inode->i_mode)) {
- rc = mds_pack_md(mds, req, reply_off + 1, body, inode);
- } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
- char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
- int len = req->rq_repmsg->buflens[reply_off + 1];
+ LASSERT (symname != NULL); /* caller prepped reply */
+ len = req->rq_repmsg->buflens[reply_off + 1];
rc = inode->i_op->readlink(dentry, symname, len);
if (rc < 0) {
CERROR("readlink failed: %d\n", rc);
+ } else if (rc != len - 1) {
+ CERROR ("Unexpected readlink rc %d: expecting %d\n",
+ rc, len - 1);
+ rc = -EINVAL;
} else {
CDEBUG(D_INODE, "read symlink dest %s\n", symname);
body->valid |= OBD_MD_LINKNAME;
+ body->eadatasize = rc + 1;
+ symname[rc] = 0; /* NULL terminate */
+ rc = 0;
}
}
+
RETURN(rc);
}
-static int mds_getattr_name(int offset, struct ptlrpc_request *req)
+static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
+ struct dentry *dentry,
+ int offset)
{
- struct mds_obd *mds = mds_req2mds(req);
- struct obd_device *obd = req->rq_export->exp_obd;
- struct obd_run_ctxt saved;
- struct mds_body *body;
- struct dentry *de = NULL, *dchild = NULL;
- struct inode *dir;
- struct lustre_handle lockh;
- char *name;
- int namelen, flags = 0, lock_mode, rc = 0;
- struct obd_ucred uc;
- __u64 res_id[3] = {0, 0, 0};
+ int rc = 0, size[1] = {sizeof(struct mds_body)};
ENTRY;
- LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
-
- if (req->rq_reqmsg->bufcount <= offset + 1) {
- LBUG();
- GOTO(out_pre_de, rc = -EINVAL);
- }
-
- body = lustre_msg_buf(req->rq_reqmsg, offset);
- name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
- namelen = req->rq_reqmsg->buflens[offset + 1];
- /* requests were at offset 2, replies go back at 1 */
- if (offset)
- offset = 1;
-
- uc.ouc_fsuid = body->fsuid;
- uc.ouc_fsgid = body->fsgid;
- uc.ouc_cap = body->capability;
- push_ctxt(&saved, &mds->mds_ctxt, &uc);
- de = mds_fid2dentry(mds, &body->fid1, NULL);
- if (IS_ERR(de)) {
- GOTO(out_pre_de, rc = -ENOENT);
- }
-
- dir = de->d_inode;
- CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
-
- lock_mode = LCK_PR;
- res_id[0] = dir->i_ino;
- res_id[1] = dir->i_generation;
-
- rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
- NULL, 0, lock_mode, &lockh);
- if (rc == 0) {
- LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
- rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
- res_id, LDLM_PLAIN, NULL, 0, lock_mode,
- &flags, ldlm_completion_ast,
- mds_blocking_ast, NULL, 0, &lockh);
- if (rc != ELDLM_OK) {
- CERROR("lock enqueue: err: %d\n", rc);
- GOTO(out_create_de, rc = -EIO);
- }
+ if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
+ CERROR("failed MDS_GETATTR_PACK test\n");
+ req->rq_status = -ENOMEM;
+ GOTO(out, rc = -ENOMEM);
}
- ldlm_lock_dump((void *)(unsigned long)lockh.addr);
- down(&dir->i_sem);
- dchild = lookup_one_len(name, de, namelen - 1);
- if (IS_ERR(dchild)) {
- CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
- up(&dir->i_sem);
- GOTO(out_create_dchild, rc = PTR_ERR(dchild));
+ rc = lustre_pack_reply(req, 1, size, NULL);
+ if (rc) {
+ CERROR("out of memory\n");
+ GOTO(out, req->rq_status = rc);
}
- rc = mds_getattr_internal(mds, dchild, req, body, offset);
-
EXIT;
-out_create_dchild:
- l_dput(dchild);
- up(&dir->i_sem);
- ldlm_lock_decref(&lockh, lock_mode);
-out_create_de:
- l_dput(de);
-out_pre_de:
- req->rq_status = rc;
- pop_ctxt(&saved);
- return 0;
+ out:
+ return(rc);
}
-static int mds_getattr(int offset, struct ptlrpc_request *req)
+static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
+ int offset)
{
struct mds_obd *mds = mds_req2mds(req);
- struct obd_run_ctxt saved;
- struct dentry *de;
- struct inode *inode;
struct mds_body *body;
- struct obd_ucred uc;
int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
ENTRY;
- body = lustre_msg_buf(req->rq_reqmsg, offset);
- uc.ouc_fsuid = body->fsuid;
- uc.ouc_fsgid = body->fsgid;
- uc.ouc_cap = body->capability;
- push_ctxt(&saved, &mds->mds_ctxt, &uc);
- de = mds_fid2dentry(mds, &body->fid1, NULL);
- if (IS_ERR(de)) {
- rc = req->rq_status = -ENOENT;
- GOTO(out_pop, PTR_ERR(de));
- }
+ body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
+ LASSERT(body != NULL); /* checked by caller */
+ LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
- inode = de->d_inode;
- if (S_ISREG(body->fid1.f_type)) {
- int rc = mds_fs_get_md(mds, inode, NULL, 0);
- CDEBUG(D_INODE, "got %d bytes MD data for inode %u\n",
+ if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
+ (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
+ int rc;
+ down(&inode->i_sem);
+ rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
+ up(&inode->i_sem);
+ CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
rc, inode->i_ino);
if (rc < 0) {
if (rc != -ENODATA)
- CERROR("error getting inode %u MD: rc = %d\n",
+ CERROR("error getting inode %lu MD: rc = %d\n",
inode->i_ino, rc);
size[bufcount] = 0;
} else if (rc > mds->mds_max_mdsize) {
size[bufcount] = 0;
CERROR("MD size %d larger than maximum possible %u\n",
rc, mds->mds_max_mdsize);
- } else
+ } else {
size[bufcount] = rc;
+ }
bufcount++;
- } else if (body->valid & OBD_MD_LINKNAME) {
- size[bufcount] = MIN(inode->i_size + 1, body->size);
+ } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
+ if (inode->i_size + 1 != body->eadatasize)
+ CERROR("symlink size: %Lu, reply space: %d\n",
+ inode->i_size + 1, body->eadatasize);
+ size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
bufcount++;
- CDEBUG(D_INODE, "symlink size: %d, reply space: %d\n",
- inode->i_size + 1, body->size);
+ CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
+ inode->i_size + 1, body->eadatasize);
}
if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
GOTO(out, rc = -ENOMEM);
}
- rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
- &req->rq_repmsg);
+ rc = lustre_pack_reply(req, bufcount, size, NULL);
if (rc) {
- CERROR("out of memoryK\n");
- req->rq_status = rc;
- GOTO(out, rc);
+ CERROR("out of memory\n");
+ GOTO(out, req->rq_status = rc);
}
- req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
+ EXIT;
+ out:
+ return(rc);
+}
-out:
- l_dput(de);
-out_pop:
- pop_ctxt(&saved);
+int mds_check_mds_num(struct obd_device *obd, struct inode* inode,
+ char *name, int namelen)
+{
+ struct mea *mea = NULL;
+ int mea_size, rc = 0;
+ ENTRY;
+
+ rc = mds_get_lmv_attr(obd, inode, &mea, &mea_size);
+ if (rc)
+ RETURN(rc);
+ if (mea != NULL) {
+ /* dir is already splitted, check is requested filename
+ * should live at this MDS or at another one */
+ int i;
+ i = mea_name2idx(mea, name, namelen - 1);
+ if (mea->mea_master != i) {
+ CERROR("inapropriate MDS(%d) for %s. should be %d\n",
+ mea->mea_master, name, i);
+ rc = -ERESTART;
+ }
+ }
+
+ if (mea)
+ OBD_FREE(mea, mea_size);
RETURN(rc);
}
-static int mds_statfs(struct ptlrpc_request *req)
+static int mds_getattr_name(int offset, struct ptlrpc_request *req,
+ struct lustre_handle *child_lockh, int child_part)
{
- struct mds_obd *mds = mds_req2mds(req);
- struct obd_statfs *osfs;
- struct statfs sfs;
- int rc, size = sizeof(*osfs);
+ struct obd_device *obd = req->rq_export->exp_obd;
+ struct ldlm_reply *rep = NULL;
+ struct lvfs_run_ctxt saved;
+ struct mds_body *body;
+ struct dentry *dparent = NULL, *dchild = NULL;
+ struct lvfs_ucred uc;
+ struct lustre_handle parent_lockh[2];
+ int namesize, update_mode;
+ int rc = 0, cleanup_phase = 0, resent_req = 0;
+ char *name;
ENTRY;
- rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
- if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
- CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
- GOTO(out, rc);
+ LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
+
+ /* Swab now, before anyone looks inside the request */
+
+ body = lustre_swab_reqbuf(req, offset, sizeof(*body),
+ lustre_swab_mds_body);
+ if (body == NULL) {
+ CERROR("Can't swab mds_body\n");
+ GOTO(cleanup, rc = -EFAULT);
}
- rc = mds_fs_statfs(mds, &sfs);
- if (rc) {
- CERROR("mds: statfs failed: rc %d\n", rc);
- GOTO(out, rc);
+ LASSERT_REQSWAB(req, offset + 1);
+ name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
+ if (name == NULL) {
+ CERROR("Can't unpack name\n");
+ GOTO(cleanup, rc = -EFAULT);
}
- osfs = lustre_msg_buf(req->rq_repmsg, 0);
- memset(osfs, 0, size);
- statfs_pack(osfs, &sfs);
- obd_statfs_pack(osfs, osfs);
-
-out:
- req->rq_status = rc;
- RETURN(0);
-}
+ namesize = req->rq_reqmsg->buflens[offset + 1];
-static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
-{
- struct mds_file_data *mfd = NULL;
+ LASSERT (offset == 0 || offset == 2);
+ /* if requests were at offset 2, the getattr reply goes back at 1 */
+ if (offset) {
+ rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
+ offset = 1;
+ }
- if (!handle || !handle->addr)
- RETURN(NULL);
+ uc.luc_fsuid = body->fsuid;
+ uc.luc_fsgid = body->fsgid;
+ uc.luc_cap = body->capability;
+ uc.luc_suppgid1 = body->suppgid;
+ uc.luc_suppgid2 = -1;
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
+ cleanup_phase = 1; /* kernel context */
+ intent_set_disposition(rep, DISP_LOOKUP_EXECD);
+
+ LASSERT(namesize > 0);
+ if (namesize == 1) {
+ /* we have no dentry here, drop LOOKUP bit */
+ child_part &= ~MDS_INODELOCK_LOOKUP;
+ CDEBUG(D_OTHER, "%s: request to retrieve attrs for %lu/%lu\n",
+ obd->obd_name, (unsigned long) body->fid1.id,
+ (unsigned long) body->fid1.generation);
+ dchild = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR,
+ parent_lockh, &update_mode,
+ NULL, 0, child_part);
+ if (IS_ERR(dchild)) {
+ CERROR("can't find inode: %d\n", (int) PTR_ERR(dchild));
+ GOTO(cleanup, rc = PTR_ERR(dchild));
+ }
+ memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
+#ifdef S_PDIROPS
+ if (parent_lockh[1].cookie)
+ ldlm_lock_decref(parent_lockh + 1, update_mode);
+#endif
+ cleanup_phase = 2;
+ goto fill_inode;
+ }
+
+ /* FIXME: handle raw lookup */
+#if 0
+ if (body->valid == OBD_MD_FLID) {
+ struct mds_body *mds_reply;
+ int size = sizeof(*mds_reply);
+ ino_t inum;
+ // The user requested ONLY the inode number, so do a raw lookup
+ rc = lustre_pack_reply(req, 1, &size, NULL);
+ if (rc) {
+ CERROR("out of memory\n");
+ GOTO(cleanup, rc);
+ }
- mfd = (struct mds_file_data *)(unsigned long)(handle->addr);
- if (!kmem_cache_validate(mds_file_cache, mfd))
- RETURN(NULL);
+ rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
- if (mfd->mfd_servercookie != handle->cookie)
- RETURN(NULL);
+ mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
+ sizeof(*mds_reply));
+ mds_reply->fid1.id = inum;
+ mds_reply->valid = OBD_MD_FLID;
+ GOTO(cleanup, rc);
+ }
+#endif
- return mfd;
-}
+ if (child_lockh->cookie != 0) {
+ LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
+ resent_req = 1;
+ }
-static int mds_store_md(struct mds_obd *mds, struct ptlrpc_request *req,
- int offset, struct mds_body *body, struct inode *inode)
-{
- struct lov_mds_md *lmm = lustre_msg_buf(req->rq_reqmsg, offset);
- int lmm_size = req->rq_reqmsg->buflens[offset];
- struct obd_run_ctxt saved;
- struct obd_ucred uc;
- void *handle;
- int rc, rc2;
- ENTRY;
+ if (resent_req == 0) {
+ rc = mds_get_parent_child_locked(obd, &obd->u.mds, &body->fid1,
+ parent_lockh, &dparent,
+ LCK_PR, MDS_INODELOCK_LOOKUP,
+ &update_mode, name, namesize,
+ child_lockh, &dchild, LCK_PR,
+ child_part);
+ if (rc)
+ GOTO(cleanup, rc);
+ } else {
+ struct ldlm_lock *granted_lock;
+ struct ll_fid child_fid;
+ struct ldlm_resource *res;
+ DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
+ granted_lock = ldlm_handle2lock(child_lockh);
+ LASSERT(granted_lock);
- /* I don't really like this, but it is a sanity check on the client
- * MD request.
- */
- if (lmm_size > mds->mds_max_mdsize) {
- CERROR("Saving MD for inode %u of %d bytes > max %d\n",
- inode->i_ino, lmm_size, mds->mds_max_mdsize);
- //RETURN(-EINVAL);
+ res = granted_lock->l_resource;
+ child_fid.id = res->lr_name.name[0];
+ child_fid.generation = res->lr_name.name[1];
+ dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
+ LASSERT(dchild);
+ LDLM_LOCK_PUT(granted_lock);
}
- CDEBUG(D_INODE, "storing %d bytes MD for inode %u\n",
- lmm_size, inode->i_ino);
- uc.ouc_fsuid = body->fsuid;
- uc.ouc_fsgid = body->fsgid;
- uc.ouc_cap = body->capability;
- push_ctxt(&saved, &mds->mds_ctxt, &uc);
- mds_start_transno(mds);
- handle = mds_fs_start(mds, inode, MDS_FSOP_SETATTR);
- if (IS_ERR(handle)) {
- rc = PTR_ERR(handle);
- mds_finish_transno(mds, handle, req, rc);
- GOTO(out_ea, rc);
- }
+ cleanup_phase = 2; /* dchild, dparent, locks */
- rc = mds_fs_set_md(mds, inode, handle, lmm, lmm_size);
- rc = mds_finish_transno(mds, handle, req, rc);
+ /* let's make sure this name should leave on this mds node */
+ rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
+ if (rc)
+ GOTO(cleanup, rc);
- rc2 = mds_fs_commit(mds, inode, handle);
- if (rc2 && !rc)
- rc = rc2;
-out_ea:
- pop_ctxt(&saved);
+fill_inode:
- RETURN(rc);
+ if (!DENTRY_VALID(dchild)) {
+ intent_set_disposition(rep, DISP_LOOKUP_NEG);
+ /* in the intent case, the policy clears this error:
+ the disposition is enough */
+ rc = -ENOENT;
+ GOTO(cleanup, rc);
+ } else {
+ intent_set_disposition(rep, DISP_LOOKUP_POS);
+ }
+
+ if (req->rq_repmsg == NULL) {
+ if (dchild->d_flags & DCACHE_CROSS_REF)
+ rc = mds_getattr_pack_msg_cf(req, dchild, offset);
+ else
+ rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
+ if (rc != 0) {
+ CERROR ("mds_getattr_pack_msg: %d\n", rc);
+ GOTO (cleanup, rc);
+ }
+ }
+
+ rc = mds_getattr_internal(obd, dchild, req, body, offset);
+ GOTO(cleanup, rc); /* returns the lock to the client */
+
+ cleanup:
+ switch (cleanup_phase) {
+ case 2:
+ if (resent_req == 0) {
+ if (rc && DENTRY_VALID(dchild))
+ ldlm_lock_decref(child_lockh, LCK_PR);
+ if (dparent) {
+ ldlm_lock_decref(parent_lockh, LCK_PR);
+#ifdef S_PDIROPS
+ if (parent_lockh[1].cookie != 0)
+ ldlm_lock_decref(parent_lockh + 1,
+ update_mode);
+#endif
+ }
+ if (dparent)
+ l_dput(dparent);
+ }
+ l_dput(dchild);
+ case 1:
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
+ default: ;
+ }
+ return rc;
}
-static int mds_open(struct ptlrpc_request *req)
+static int mds_getattr(int offset, struct ptlrpc_request *req)
{
struct mds_obd *mds = mds_req2mds(req);
- struct mds_body *body;
- struct mds_export_data *med;
- struct mds_file_data *mfd;
+ struct obd_device *obd = req->rq_export->exp_obd;
+ struct lvfs_run_ctxt saved;
struct dentry *de;
- struct file *file;
- struct vfsmount *mnt;
- __u32 flags;
- struct list_head *tmp;
- int rc, size = sizeof(*body);
+ struct mds_body *body;
+ struct lvfs_ucred uc;
+ int rc = 0;
ENTRY;
- if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
- CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
- req->rq_status = -ENOMEM;
- RETURN(-ENOMEM);
+ body = lustre_swab_reqbuf (req, offset, sizeof (*body),
+ lustre_swab_mds_body);
+ if (body == NULL) {
+ CERROR ("Can't unpack body\n");
+ RETURN (-EFAULT);
}
- rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
- if (rc) {
- CERROR("mds: pack error: rc = %d\n", rc);
- req->rq_status = rc;
- RETURN(rc);
+ uc.luc_fsuid = body->fsuid;
+ uc.luc_fsgid = body->fsgid;
+ uc.luc_cap = body->capability;
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
+ de = mds_fid2dentry(mds, &body->fid1, NULL);
+ if (IS_ERR(de)) {
+ rc = req->rq_status = -ENOENT;
+ GOTO(out_pop, PTR_ERR(de));
}
- body = lustre_msg_buf(req->rq_reqmsg, 0);
-
- /* was this animal open already and the client lost the reply? */
- /* XXX need some way to detect a reopen, to avoid locked list walks */
- med = &req->rq_export->exp_mds_data;
- spin_lock(&med->med_open_lock);
- list_for_each(tmp, &med->med_open_head) {
- mfd = list_entry(tmp, typeof(*mfd), mfd_list);
- if (!memcmp(&mfd->mfd_clienthandle, &body->handle,
- sizeof(mfd->mfd_clienthandle)) &&
- body->fid1.id == mfd->mfd_file->f_dentry->d_inode->i_ino) {
- de = mfd->mfd_file->f_dentry;
- spin_unlock(&med->med_open_lock);
- CERROR("Re opening "LPD64"\n", body->fid1.id);
- GOTO(out_pack, rc = 0);
- }
+ rc = mds_getattr_pack_msg(req, de->d_inode, offset);
+ if (rc != 0) {
+ CERROR ("mds_getattr_pack_msg: %d\n", rc);
+ GOTO (out_pop, rc);
}
- spin_unlock(&med->med_open_lock);
- mfd = kmem_cache_alloc(mds_file_cache, GFP_KERNEL);
- if (!mfd) {
- CERROR("mds: out of memory\n");
- req->rq_status = -ENOMEM;
- RETURN(0);
- }
+ req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
- de = mds_fid2dentry(mds, &body->fid1, &mnt);
- if (IS_ERR(de))
- GOTO(out_free, rc = PTR_ERR(de));
+ l_dput(de);
+ GOTO(out_pop, rc);
+out_pop:
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
+ return rc;
+}
- /* check if this inode has seen a delayed object creation */
- if (lustre_msg_get_op_flags(req->rq_reqmsg) & MDS_OPEN_HAS_EA) {
- rc = mds_store_md(mds, req, 1, body, de->d_inode);
- if (rc) {
- l_dput(de);
- mntput(mnt);
- GOTO(out_free, rc);
- }
- }
- flags = body->flags;
- /* dentry_open does a dput(de) and mntput(mnt) on error */
- file = dentry_open(de, mnt, flags & ~O_DIRECT);
- if (IS_ERR(file)) {
- rc = PTR_ERR(file);
- GOTO(out_free, 0);
- }
+static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
+ unsigned long max_age)
+{
+ int rc;
- file->private_data = mfd;
- mfd->mfd_file = file;
- memcpy(&mfd->mfd_clienthandle, &body->handle, sizeof(body->handle));
- get_random_bytes(&mfd->mfd_servercookie, sizeof(mfd->mfd_servercookie));
- spin_lock(&med->med_open_lock);
- list_add(&mfd->mfd_list, &med->med_open_head);
- spin_unlock(&med->med_open_lock);
+ spin_lock(&obd->obd_osfs_lock);
+ rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
+ if (rc == 0)
+ memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
+ spin_unlock(&obd->obd_osfs_lock);
-out_pack:
- body = lustre_msg_buf(req->rq_repmsg, 0);
- mds_pack_inode2fid(&body->fid1, de->d_inode);
- mds_pack_inode2body(body, de->d_inode);
- body->handle.addr = (__u64)(unsigned long)mfd;
- body->handle.cookie = mfd->mfd_servercookie;
- CDEBUG(D_INODE, "llite file "LPX64": addr %p, cookie "LPX64"\n",
- mfd->mfd_clienthandle.addr, mfd, mfd->mfd_servercookie);
- RETURN(0);
+ return rc;
+}
+
+static int mds_statfs(struct ptlrpc_request *req)
+{
+ struct obd_device *obd = req->rq_export->exp_obd;
+ int rc, size = sizeof(struct obd_statfs);
+ ENTRY;
+
+ rc = lustre_pack_reply(req, 1, &size, NULL);
+ if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
+ CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
+ GOTO(out, rc);
+ }
+
+ /* We call this so that we can cache a bit - 1 jiffie worth */
+ rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
+ jiffies - HZ);
+ if (rc) {
+ CERROR("mds_obd_statfs failed: rc %d\n", rc);
+ GOTO(out, rc);
+ }
-out_free:
- mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
- kmem_cache_free(mds_file_cache, mfd);
+ EXIT;
+out:
req->rq_status = rc;
- RETURN(0);
+ return 0;
}
-static int mds_close(struct ptlrpc_request *req)
+static int mds_sync(struct ptlrpc_request *req)
{
- struct mds_export_data *med = &req->rq_export->exp_mds_data;
+ struct obd_device *obd = req->rq_export->exp_obd;
+ struct mds_obd *mds = &obd->u.mds;
struct mds_body *body;
- struct mds_file_data *mfd;
- int rc;
+ int rc, size = sizeof(*body);
ENTRY;
- body = lustre_msg_buf(req->rq_reqmsg, 0);
+ body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
+ if (body == NULL)
+ GOTO(out, rc = -EPROTO);
- mfd = mds_handle2mfd(&body->handle);
- if (!mfd) {
- CERROR("no handle for file close "LPD64
- ": addr "LPX64", cookie "LPX64"\n",
- body->fid1.id, body->handle.addr, body->handle.cookie);
- RETURN(-ESTALE);
+ rc = lustre_pack_reply(req, 1, &size, NULL);
+ if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
+ CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
+ GOTO(out, rc);
}
- spin_lock(&med->med_open_lock);
- req->rq_status = mds_close_mfd(mfd, med);
- spin_unlock(&med->med_open_lock);
+ if (body->fid1.id == 0) {
+ /* a fid of zero is taken to mean "sync whole filesystem" */
+ rc = fsfilt_sync(obd, mds->mds_sb);
+ if (rc)
+ GOTO(out, rc);
+ } else {
+ /* just any file to grab fsync method - "file" arg unused */
+ struct file *file = mds->mds_rcvd_filp;
+ struct dentry *de;
- if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
- CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
- req->rq_status = -ENOMEM;
- RETURN(-ENOMEM);
- }
+ de = mds_fid2dentry(mds, &body->fid1, NULL);
+ if (IS_ERR(de))
+ GOTO(out, rc = PTR_ERR(de));
- rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
- if (rc) {
- CERROR("mds: lustre_pack_msg: rc = %d\n", rc);
- req->rq_status = rc;
- }
+ rc = file->f_op->fsync(NULL, de, 1);
+ l_dput(de);
+ if (rc)
+ GOTO(out, rc);
- RETURN(0);
+ body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
+ mds_pack_inode2fid(obd, &body->fid1, de->d_inode);
+ mds_pack_inode2body(obd, body, de->d_inode);
+ }
+out:
+ req->rq_status = rc;
+ return 0;
}
+/* mds_readpage does not take a DLM lock on the inode, because the client must
+ * already have a PR lock.
+ *
+ * If we were to take another one here, a deadlock will result, if another
+ * thread is already waiting for a PW lock. */
static int mds_readpage(struct ptlrpc_request *req)
{
- struct mds_obd *mds = mds_req2mds(req);
+ struct obd_device *obd = req->rq_export->exp_obd;
struct vfsmount *mnt;
struct dentry *de;
struct file *file;
struct mds_body *body, *repbody;
- struct obd_run_ctxt saved;
- int rc, size = sizeof(*body);
- struct obd_ucred uc;
+ struct lvfs_run_ctxt saved;
+ int rc, size = sizeof(*repbody);
+ struct lvfs_ucred uc;
ENTRY;
- rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
+ rc = lustre_pack_reply(req, 1, &size, NULL);
if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
CERROR("mds: out of memory\n");
GOTO(out, rc = -ENOMEM);
}
- body = lustre_msg_buf(req->rq_reqmsg, 0);
- uc.ouc_fsuid = body->fsuid;
- uc.ouc_fsgid = body->fsgid;
- uc.ouc_cap = body->capability;
- push_ctxt(&saved, &mds->mds_ctxt, &uc);
- de = mds_fid2dentry(mds, &body->fid1, &mnt);
+ body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
+ if (body == NULL)
+ GOTO (out, rc = -EFAULT);
+
+ uc.luc_fsuid = body->fsuid;
+ uc.luc_fsgid = body->fsgid;
+ uc.luc_cap = body->capability;
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
+ de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
if (IS_ERR(de))
GOTO(out_pop, rc = PTR_ERR(de));
- CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
+ CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
/* note: in case of an error, dentry_open puts dentry */
if (IS_ERR(file))
GOTO(out_pop, rc = PTR_ERR(file));
- repbody = lustre_msg_buf(req->rq_repmsg, 0);
+ /* body->size is actually the offset -eeb */
+ if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
+ CERROR("offset "LPU64" not on a block boundary of %lu\n",
+ body->size, de->d_inode->i_blksize);
+ GOTO(out_file, rc = -EFAULT);
+ }
+
+ /* body->nlink is actually the #bytes to read -eeb */
+ if (body->nlink & (de->d_inode->i_blksize - 1)) {
+ CERROR("size %u is not multiple of blocksize %lu\n",
+ body->nlink, de->d_inode->i_blksize);
+ GOTO(out_file, rc = -EFAULT);
+ }
+
+ repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
repbody->size = file->f_dentry->d_inode->i_size;
repbody->valid = OBD_MD_FLSIZE;
/* to make this asynchronous make sure that the handling function
doesn't send a reply when this function completes. Instead a
callback function would send the reply */
- /* note: in case of an error, dentry_open puts dentry */
- rc = mds_sendpage(req, file, body->size);
+ /* body->size is actually the offset -eeb */
+ rc = mds_sendpage(req, file, body->size, body->nlink);
+out_file:
filp_close(file, 0);
out_pop:
- pop_ctxt(&saved);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
out:
req->rq_status = rc;
RETURN(0);
}
-int mds_reint(struct ptlrpc_request *req, int offset)
+int mds_reint(struct ptlrpc_request *req, int offset,
+ struct lustre_handle *lockh)
{
+ struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
int rc;
- struct mds_update_record rec;
+ ENTRY;
+
+ OBD_ALLOC(rec, sizeof(*rec));
+ if (rec == NULL)
+ RETURN(-ENOMEM);
- rc = mds_update_unpack(req, offset, &rec);
+ rc = mds_update_unpack(req, offset, rec);
if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
CERROR("invalid record\n");
- req->rq_status = -EINVAL;
- RETURN(0);
+ GOTO(out, req->rq_status = -EINVAL);
}
/* rc will be used to interrupt a for loop over multiple records */
- rc = mds_reint_rec(&rec, offset, req);
- return rc;
+ rc = mds_reint_rec(rec, offset, req, lockh);
+ out:
+ OBD_FREE(rec, sizeof(*rec));
+ RETURN(rc);
}
-/* forward declaration */
-int mds_handle(struct ptlrpc_request *req);
-
-static int check_for_next_transno(struct mds_obd *mds)
+static int mds_filter_recovery_request(struct ptlrpc_request *req,
+ struct obd_device *obd, int *process)
{
- struct ptlrpc_request *req;
- req = list_entry(mds->mds_recovery_queue.next,
- struct ptlrpc_request, rq_list);
- return req->rq_reqmsg->transno == mds->mds_next_recovery_transno;
+ switch (req->rq_reqmsg->opc) {
+ case MDS_CONNECT: /* This will never get here, but for completeness. */
+ case OST_CONNECT: /* This will never get here, but for completeness. */
+ case MDS_DISCONNECT:
+ case OST_DISCONNECT:
+ *process = 1;
+ RETURN(0);
+
+ case MDS_CLOSE:
+ case MDS_SYNC: /* used in unmounting */
+ case OBD_PING:
+ case MDS_REINT:
+ case LDLM_ENQUEUE:
+ case OST_CREATE:
+ *process = target_queue_recovery_request(req, obd);
+ RETURN(0);
+
+ default:
+ DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
+ *process = 0;
+ /* XXX what should we set rq_status to here? */
+ req->rq_status = -EAGAIN;
+ RETURN(ptlrpc_error(req));
+ }
}
-static void process_recovery_queue(struct mds_obd *mds)
+static char *reint_names[] = {
+ [REINT_SETATTR] "setattr",
+ [REINT_CREATE] "create",
+ [REINT_LINK] "link",
+ [REINT_UNLINK] "unlink",
+ [REINT_RENAME] "rename",
+ [REINT_OPEN] "open",
+};
+
+#define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER |\
+ OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ|\
+ OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME|\
+ OBD_MD_FLID)
+
+static void reconstruct_create(struct ptlrpc_request *req)
{
- struct ptlrpc_request *req;
-
- for (;;) {
- spin_lock(&mds->mds_processing_task_lock);
- req = list_entry(mds->mds_recovery_queue.next,
- struct ptlrpc_request, rq_list);
-
- if (req->rq_reqmsg->transno != mds->mds_next_recovery_transno) {
- spin_unlock(&mds->mds_processing_task_lock);
- wait_event(mds->mds_next_transno_waitq,
- check_for_next_transno(mds));
- continue;
- }
- list_del(&req->rq_list);
- spin_unlock(&mds->mds_processing_task_lock);
+ struct mds_export_data *med = &req->rq_export->exp_mds_data;
+ struct mds_client_data *mcd = med->med_mcd;
+ struct ost_body *body;
- DEBUG_REQ(D_HA, req, "");
- mds_handle(req);
-
- if (list_empty(&mds->mds_recovery_queue))
- break;
- }
+ body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
+
+ /* copy rc, transno and disp; steal locks */
+ mds_req_from_mcd(req, mcd);
+ CERROR("reconstruct reply for x"LPU64"\n", req->rq_xid);
}
-static int queue_recovery_request(struct ptlrpc_request *req,
- struct mds_obd *mds)
+static int mdt_obj_create(struct ptlrpc_request *req)
{
- struct list_head *tmp;
- int inserted = 0, transno = req->rq_reqmsg->transno;
+ struct obd_device *obd = req->rq_export->exp_obd;
+ struct ldlm_res_id res_id = { .name = {0} };
+ struct mds_obd *mds = &obd->u.mds;
+ struct ost_body *body, *repbody;
+ int rc, size = sizeof(*repbody);
+ char fidname[LL_FID_NAMELEN];
+ struct inode *parent_inode;
+ struct lustre_handle lockh;
+ struct lvfs_run_ctxt saved;
+ ldlm_policy_data_t policy;
+ struct dentry *new = NULL;
+ struct dentry_params dp;
+ int mealen, flags = 0;
+ unsigned int tmpname;
+ struct lvfs_ucred uc;
+ struct mea *mea;
+ void *handle;
+ ENTRY;
+
+ DEBUG_REQ(D_HA, req, "create remote object");
- if (!transno) {
- DEBUG_REQ(D_HA, req, "not queueing");
- return 1;
- }
+ parent_inode = mds->mds_objects_dir->d_inode;
- spin_lock(&mds->mds_processing_task_lock);
+ body = lustre_swab_reqbuf(req, 0, sizeof(*body),
+ lustre_swab_ost_body);
+ if (body == NULL)
+ RETURN(-EFAULT);
- if (mds->mds_processing_task == current->pid) {
- /* Processing the queue right now, don't re-add. */
- spin_unlock(&mds->mds_processing_task_lock);
- return 1;
- }
+ MDS_CHECK_RESENT(req, reconstruct_create(req));
- /* XXX O(n^2) */
- list_for_each(tmp, &mds->mds_recovery_queue) {
- struct ptlrpc_request *reqiter =
- list_entry(tmp, struct ptlrpc_request, rq_list);
- if (reqiter->rq_reqmsg->transno > transno) {
- list_add_tail(&req->rq_list, &reqiter->rq_list);
- inserted = 1;
- break;
+ uc.luc_fsuid = body->oa.o_uid;
+ uc.luc_fsgid = body->oa.o_gid;
+
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
+
+ rc = lustre_pack_reply(req, 1, &size, NULL);
+ if (rc)
+ RETURN(rc);
+
+ repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
+
+ if (body->oa.o_flags & OBD_FL_RECREATE_OBJS) {
+ /* this is re-create request from MDS holding directory name.
+ * we have to lookup given ino/generation first. if it exists
+ * (good case) then there is nothing to do. if it does not
+ * then we have to recreate it */
+ struct ll_fid fid;
+ fid.id = body->oa.o_id;
+ fid.generation = body->oa.o_generation;
+ new = mds_fid2dentry(mds, &fid, NULL);
+ if (!IS_ERR(new) && new->d_inode) {
+ CWARN("mkdir() repairing is on its way: %lu/%lu\n",
+ (unsigned long) fid.id,
+ (unsigned long) fid.generation);
+ obdo_from_inode(&repbody->oa, new->d_inode,
+ FILTER_VALID_FLAGS);
+ repbody->oa.o_id = new->d_inode->i_ino;
+ repbody->oa.o_generation = new->d_inode->i_generation;
+ repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
+ GOTO(cleanup2, rc = 0);
}
+ CWARN("hmm. for some reason dir %lu/%lu (or reply) got lost\n",
+ (unsigned long) fid.id, (unsigned long) fid.generation);
+ LASSERT(new->d_inode == NULL ||
+ new->d_inode->i_generation != fid.generation);
+ l_dput(new);
}
+
+ down(&parent_inode->i_sem);
+ handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
+ LASSERT(!IS_ERR(handle));
+
+repeat:
+ tmpname = ll_insecure_random_int();
+ rc = sprintf(fidname, "%u", tmpname);
+ new = lookup_one_len(fidname, mds->mds_objects_dir, rc);
+ if (IS_ERR(new)) {
+ CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
+ obd->obd_name, fidname, (int) PTR_ERR(new));
+ fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
+ up(&parent_inode->i_sem);
+ RETURN(PTR_ERR(new));
+ } else if (new->d_inode) {
+ CERROR("%s: name exists. repeat\n", obd->obd_name);
+ goto repeat;
+ }
+
+ new->d_fsdata = (void *) &dp;
+ dp.p_inum = 0;
+ dp.p_ptr = req;
+
+ if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
+ DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
+ (unsigned long) body->oa.o_id,
+ (unsigned long) body->oa.o_generation);
+ dp.p_inum = body->oa.o_id;
+ dp.p_generation = body->oa.o_generation;
+ }
+ rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
+ if (rc == 0) {
+ obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
+ repbody->oa.o_id = new->d_inode->i_ino;
+ repbody->oa.o_generation = new->d_inode->i_generation;
+ repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
- if (!inserted)
- list_add_tail(&req->rq_list, &mds->mds_recovery_queue);
+ rc = fsfilt_del_dir_entry(obd, new);
+ up(&parent_inode->i_sem);
- if (mds->mds_processing_task != 0) {
- /* Someone else is processing this queue, we'll leave it to
- * them.
- */
- spin_unlock(&mds->mds_processing_task_lock);
- if (transno == mds->mds_next_recovery_transno)
- wake_up(&mds->mds_next_transno_waitq);
- return 0;
+ if (rc) {
+ CERROR("can't remove name for object: %d\n", rc);
+ GOTO(cleanup, rc);
+ }
+
+ /* this lock should be taken to serialize MDS modifications
+ * in failure case */
+ res_id.name[0] = new->d_inode->i_ino;
+ res_id.name[1] = new->d_inode->i_generation;
+ policy.l_inodebits.bits = MDS_INODELOCK_UPDATE;
+ rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
+ res_id, LDLM_IBITS, &policy,
+ LCK_EX, &flags, mds_blocking_ast,
+ ldlm_completion_ast, NULL, NULL,
+ NULL, 0, NULL, &lockh);
+ if (rc != ELDLM_OK)
+ GOTO(cleanup, rc);
+
+ CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
+ (unsigned long) new->d_inode->i_ino,
+ (unsigned long) new->d_inode->i_generation,
+ (unsigned) new->d_inode->i_mode);
+ } else {
+ up(&parent_inode->i_sem);
+ CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
+ }
+
+ if (rc == 0 && body->oa.o_valid & OBD_MD_FLID) {
+ /* this is new object for splitted dir. we have to
+ * prevent recursive splitting on it -bzzz */
+ mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
+ OBD_ALLOC(mea, mealen);
+ if (mea == NULL)
+ GOTO(cleanup, rc = -ENOMEM);
+ mea->mea_count = 0;
+ down(&new->d_inode->i_sem);
+ rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
+ up(&new->d_inode->i_sem);
+ OBD_FREE(mea, mealen);
+ }
+
+cleanup:
+ rc = mds_finish_transno(mds, parent_inode, handle, req, rc, 0);
+ if (rc == 0)
+ ptlrpc_save_lock(req, &lockh, LCK_EX);
+ else
+ ldlm_lock_decref(&lockh, LCK_EX);
+cleanup2:
+ l_dput(new);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
+ RETURN(rc);
+}
+
+static int mdt_get_info(struct ptlrpc_request *req)
+{
+ char *key;
+ struct obd_export *exp = req->rq_export;
+ int keylen, rc = 0, size = sizeof(obd_id);
+ obd_id *reply;
+ ENTRY;
+
+ key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
+ if (key == NULL) {
+ DEBUG_REQ(D_HA, req, "no get_info key");
+ RETURN(-EFAULT);
}
+ keylen = req->rq_reqmsg->buflens[0];
- /* Nobody is processing, and we know there's (at least) one to process
- * now, so we'll do the honours.
- */
- mds->mds_processing_task = current->pid;
- spin_unlock(&mds->mds_processing_task_lock);
+ if (keylen < strlen("mdsize") || memcmp(key, "mdsize", 6) != 0)
+ RETURN(-EPROTO);
- process_recovery_queue(mds);
- return 0;
+ rc = lustre_pack_reply(req, 1, &size, NULL);
+ if (rc)
+ RETURN(rc);
+
+ reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
+ rc = obd_get_info(exp, keylen, key, &size, reply);
+ req->rq_repmsg->status = 0;
+ RETURN(rc);
}
-static int filter_recovery_request(struct ptlrpc_request *req,
- struct mds_obd *mds, int *process)
+static int mds_set_info(struct obd_export *exp, __u32 keylen,
+ void *key, __u32 vallen, void *val)
{
- switch (req->rq_reqmsg->opc) {
- case MDS_CONNECT:
- case MDS_DISCONNECT:
- case MDS_OPEN:
- *process = 1;
- RETURN(0);
-
- case MDS_GETSTATUS: /* used in unmounting */
- case MDS_REINT:
- case LDLM_ENQUEUE:
- *process = queue_recovery_request(req, mds);
+ struct obd_device *obd;
+ struct mds_obd *mds;
+ int rc = 0;
+ ENTRY;
+
+ obd = class_exp2obd(exp);
+ if (obd == NULL) {
+ CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
+ exp->exp_handle.h_cookie);
+ RETURN(-EINVAL);
+ }
+
+ mds = &obd->u.mds;
+ if (keylen == strlen("mds_num") &&
+ memcmp(key, "mds_num", keylen) == 0) {
+ int valsize;
+ __u32 group;
+ CDEBUG(D_IOCTL, "set mds num %d\n", *(int*)val);
+ mds->mds_num = *(int*)val;
+ group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
+ valsize = sizeof(group);
+ /*mds number has been changed, so the corresponding obdfilter exp
+ *need to be changed too*/
+ rc = obd_set_info(mds->mds_osc_exp, strlen("mds_conn"), "mds_conn",
+ valsize, &group);
+ RETURN(rc);
+ } else if (keylen == strlen("client") &&
+ memcmp(key, "client", keylen) == 0) {
+ if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
+ atomic_inc(&mds->mds_real_clients);
+ CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
+ obd->obd_name, exp->exp_client_uuid.uuid,
+ atomic_read(&mds->mds_real_clients));
+ exp->exp_flags |= OBD_OPT_REAL_CLIENT;
+ }
+ if (mds->mds_lmv_name) {
+ rc = mds_lmv_connect(obd, mds->mds_lmv_name);
+ LASSERT(rc == 0);
+ }
RETURN(0);
-
- default:
- DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
- *process = 0;
- RETURN(ptlrpc_error(req->rq_svc, req));
}
+ CDEBUG(D_IOCTL, "invalid key\n");
+ RETURN(-EINVAL);
}
-static int mds_queue_final_reply(struct ptlrpc_request *req, int rc)
+static int mdt_set_info(struct ptlrpc_request *req)
{
- struct mds_obd *mds = mds_req2mds(req);
+ char *key, *val;
+ struct obd_export *exp = req->rq_export;
+ int keylen, rc = 0, vallen;
+ ENTRY;
- if (rc) {
- /* Just like ptlrpc_error, but without the sending. */
- lustre_pack_msg(0, NULL, NULL, &req->rq_replen,
- &req->rq_repmsg);
- req->rq_type = PTL_RPC_MSG_ERR;
- }
-
- list_add(&req->rq_list, &mds->mds_delayed_reply_queue);
- if (--mds->mds_recoverable_clients == 0) {
- struct list_head *tmp, *n;
-
- CDEBUG(D_HA,
- "all clients recovered, sending delayed replies\n");
- list_for_each_safe(tmp, n, &mds->mds_delayed_reply_queue) {
- req = list_entry(tmp, struct ptlrpc_request, rq_list);
- DEBUG_REQ(D_HA, req, "delayed:");
- ptlrpc_reply(req->rq_svc, req);
- }
- } else {
- CDEBUG(D_HA, "%d recoverable clients remain\n",
- mds->mds_recoverable_clients);
+ key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
+ if (key == NULL) {
+ DEBUG_REQ(D_HA, req, "no set_info key");
+ RETURN(-EFAULT);
}
+ keylen = req->rq_reqmsg->buflens[0];
- return 1;
-}
+ if (keylen == strlen("mds_num") &&
+ memcmp(key, "mds_num", keylen) == 0) {
+ rc = lustre_pack_reply(req, 0, NULL, NULL);
+ if (rc)
+ RETURN(rc);
+ val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
-static char *reint_names[] = {
- [REINT_SETATTR] "setattr",
- [REINT_CREATE] "create",
- [REINT_LINK] "link",
- [REINT_UNLINK] "unlink",
- [REINT_RENAME] "rename"
-};
+ vallen = req->rq_reqmsg->buflens[1];
+ rc = obd_set_info(exp, keylen, key, vallen, val);
+ req->rq_repmsg->status = 0;
+ RETURN(rc);
+ } else if (keylen == strlen("client") &&
+ memcmp(key, "client", keylen) == 0) {
+ rc = lustre_pack_reply(req, 0, NULL, NULL);
+ if (rc)
+ RETURN(rc);
+ rc = obd_set_info(exp, keylen, key, sizeof(obd_id), NULL);
+ req->rq_repmsg->status = 0;
+ RETURN(rc);
+ }
+ CDEBUG(D_IOCTL, "invalid key\n");
+ RETURN(-EINVAL);
+}
+
+extern int ost_brw_write(struct ptlrpc_request *, struct obd_trans_info *);
int mds_handle(struct ptlrpc_request *req)
{
- int rc;
- int should_process;
+ int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
+ int rc = 0;
struct mds_obd *mds = NULL; /* quell gcc overwarning */
+ struct obd_device *obd = NULL;
ENTRY;
- rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
- if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
- CERROR("lustre_mds: Invalid request\n");
- GOTO(out, rc);
- }
-
- LASSERT(!strcmp(req->rq_obd->obd_type->typ_name, LUSTRE_MDT_NAME));
+ OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
+ LASSERT(current->journal_info == NULL);
+ /* XXX identical to OST */
if (req->rq_reqmsg->opc != MDS_CONNECT) {
- if (req->rq_export == NULL)
+ struct mds_export_data *med;
+ int recovering, abort_recovery;
+
+ if (req->rq_export == NULL) {
+ CERROR("lustre_mds: operation %d on unconnected MDS\n",
+ req->rq_reqmsg->opc);
+ req->rq_status = -ENOTCONN;
GOTO(out, rc = -ENOTCONN);
+ }
- mds = mds_req2mds(req);
- if (mds->mds_recoverable_clients != 0) {
- rc = filter_recovery_request(req, mds, &should_process);
+ med = &req->rq_export->exp_mds_data;
+ obd = req->rq_export->exp_obd;
+ mds = &obd->u.mds;
+
+ /* sanity check: if the xid matches, the request must
+ * be marked as a resent or replayed */
+ if (req->rq_xid == med->med_mcd->mcd_last_xid)
+ LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
+ (MSG_RESENT | MSG_REPLAY),
+ "rq_xid "LPU64" matches last_xid, "
+ "expected RESENT flag\n",
+ req->rq_xid);
+ /* else: note the opposite is not always true; a
+ * RESENT req after a failover will usually not match
+ * the last_xid, since it was likely never
+ * committed. A REPLAYed request will almost never
+ * match the last xid, however it could for a
+ * committed, but still retained, open. */
+
+ /* Check for aborted recovery. */
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ abort_recovery = obd->obd_abort_recovery;
+ recovering = obd->obd_recovering;
+ spin_unlock_bh(&obd->obd_processing_task_lock);
+ if (abort_recovery) {
+ target_abort_recovery(obd);
+ } else if (recovering) {
+ rc = mds_filter_recovery_request(req, obd,
+ &should_process);
if (rc || !should_process)
RETURN(rc);
}
case MDS_CONNECT:
DEBUG_REQ(D_INODE, req, "connect");
OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
- rc = target_handle_connect(req);
- /* Make sure that last_rcvd is correct. */
- if (!rc) {
+ rc = target_handle_connect(req, mds_handle);
+ if (!rc)
/* Now that we have an export, set mds. */
mds = mds_req2mds(req);
- mds_fsync_super(mds->mds_sb);
- }
break;
case MDS_DISCONNECT:
DEBUG_REQ(D_INODE, req, "disconnect");
OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
rc = target_handle_disconnect(req);
- /* Make sure that last_rcvd is correct. */
- if (!rc)
- mds_fsync_super(mds->mds_sb);
- goto out;
+ req->rq_status = rc; /* superfluous? */
+ break;
case MDS_GETSTATUS:
DEBUG_REQ(D_INODE, req, "getstatus");
rc = mds_getstatus(req);
break;
- case MDS_GETLOVINFO:
- DEBUG_REQ(D_INODE, req, "getlovinfo");
- rc = mds_getlovinfo(req);
- break;
-
case MDS_GETATTR:
DEBUG_REQ(D_INODE, req, "getattr");
OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
rc = mds_getattr(0, req);
break;
+ case MDS_GETATTR_NAME: {
+ struct lustre_handle lockh;
+ DEBUG_REQ(D_INODE, req, "getattr_name");
+ OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
+
+ /* If this request gets a reconstructed reply, we won't be
+ * acquiring any new locks in mds_getattr_name, so we don't
+ * want to cancel.
+ */
+ lockh.cookie = 0;
+ rc = mds_getattr_name(0, req, &lockh, MDS_INODELOCK_UPDATE);
+ /* this non-intent call (from an ioctl) is special */
+ req->rq_status = rc;
+ if (rc == 0 && lockh.cookie)
+ ldlm_lock_decref(&lockh, LCK_PR);
+ break;
+ }
case MDS_STATFS:
DEBUG_REQ(D_INODE, req, "statfs");
OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
break;
case MDS_READPAGE:
- DEBUG_REQ(D_INODE, req, "readpage\n");
+ DEBUG_REQ(D_INODE, req, "readpage");
OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
rc = mds_readpage(req);
- if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
- return 0;
+ if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
+ if (req->rq_reply_state) {
+ lustre_free_reply_state (req->rq_reply_state);
+ req->rq_reply_state = NULL;
+ }
+ RETURN(0);
+ }
+
break;
case MDS_REINT: {
- int size = sizeof(struct mds_body);
- int opc = *(u32 *)lustre_msg_buf(req->rq_reqmsg, 0),
- realopc = opc & REINT_OPCODE_MASK;
-
- DEBUG_REQ(D_INODE, req, "reint (%s%s)",
- reint_names[realopc],
- opc & REINT_REPLAYING ? "|REPLAYING" : "");
-
+ __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*opcp));
+ __u32 opc;
+ int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
+ mds->mds_max_cookiesize};
+ int bufcount;
+
+ /* NB only peek inside req now; mds_reint() will swab it */
+ if (opcp == NULL) {
+ CERROR ("Can't inspect opcode\n");
+ rc = -EINVAL;
+ break;
+ }
+ opc = *opcp;
+ if (lustre_msg_swabbed (req->rq_reqmsg))
+ __swab32s(&opc);
+
+ DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
+ (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
+ reint_names[opc] == NULL) ? reint_names[opc] :
+ "unknown opcode");
+
OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
- rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
- &req->rq_repmsg);
- if (rc) {
- req->rq_status = rc;
+ if (opc == REINT_UNLINK)
+ bufcount = 3;
+ else if (opc == REINT_OPEN || opc == REINT_RENAME)
+ bufcount = 2;
+ else
+ bufcount = 1;
+
+ rc = lustre_pack_reply(req, bufcount, size, NULL);
+ if (rc)
break;
- }
- rc = mds_reint(req, 0);
- OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
- break;
- }
- case MDS_OPEN:
- DEBUG_REQ(D_INODE, req, "open");
- OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
- rc = mds_open(req);
+ rc = mds_reint(req, 0, NULL);
+ fail = OBD_FAIL_MDS_REINT_NET_REP;
break;
+ }
case MDS_CLOSE:
DEBUG_REQ(D_INODE, req, "close");
rc = mds_close(req);
break;
+ case MDS_DONE_WRITING:
+ DEBUG_REQ(D_INODE, req, "done_writing");
+ OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
+ rc = mds_done_writing(req);
+ break;
+
+ case MDS_PIN:
+ DEBUG_REQ(D_INODE, req, "pin");
+ OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
+ rc = mds_pin(req);
+ break;
+
+ case MDS_SYNC:
+ DEBUG_REQ(D_INODE, req, "sync");
+ OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
+ rc = mds_sync(req);
+ break;
+
+ case OBD_PING:
+ DEBUG_REQ(D_INODE, req, "ping");
+ rc = target_handle_ping(req);
+ break;
+
+ case OBD_LOG_CANCEL:
+ CDEBUG(D_INODE, "log cancel\n");
+ OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
+ rc = -ENOTSUPP; /* la la la */
+ break;
+
case LDLM_ENQUEUE:
DEBUG_REQ(D_INODE, req, "enqueue");
OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
- rc = ldlm_handle_enqueue(req);
+ rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
+ ldlm_server_blocking_ast, NULL);
break;
case LDLM_CONVERT:
DEBUG_REQ(D_INODE, req, "convert");
LBUG();
OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
break;
+ case LLOG_ORIGIN_HANDLE_CREATE:
+ DEBUG_REQ(D_INODE, req, "llog_init");
+ OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
+ rc = llog_origin_handle_create(req);
+ break;
+ case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
+ DEBUG_REQ(D_INODE, req, "llog next block");
+ OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
+ rc = llog_origin_handle_next_block(req);
+ break;
+ case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
+ DEBUG_REQ(D_INODE, req, "llog prev block");
+ OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
+ rc = llog_origin_handle_prev_block(req);
+ break;
+ case LLOG_ORIGIN_HANDLE_READ_HEADER:
+ DEBUG_REQ(D_INODE, req, "llog read header");
+ OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
+ rc = llog_origin_handle_read_header(req);
+ break;
+ case LLOG_ORIGIN_HANDLE_CLOSE:
+ DEBUG_REQ(D_INODE, req, "llog close");
+ OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
+ rc = llog_origin_handle_close(req);
+ break;
+ case OST_CREATE:
+ DEBUG_REQ(D_INODE, req, "ost_create");
+ rc = mdt_obj_create(req);
+ break;
+ case OST_GET_INFO:
+ DEBUG_REQ(D_INODE, req, "get_info");
+ rc = mdt_get_info(req);
+ break;
+ case OST_SET_INFO:
+ DEBUG_REQ(D_INODE, req, "set_info");
+ rc = mdt_set_info(req);
+ break;
+ case OST_WRITE:
+ CDEBUG(D_INODE, "write\n");
+ OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
+ rc = ost_brw_write(req, NULL);
+ LASSERT(current->journal_info == NULL);
+ /* mdt_brw sends its own replies */
+ RETURN(rc);
+ break;
+ case LLOG_CATINFO:
+ DEBUG_REQ(D_INODE, req, "llog catinfo");
+ OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
+ rc = llog_catinfo(req);
+ break;
default:
- rc = ptlrpc_error(req->rq_svc, req);
+ req->rq_status = -ENOTSUPP;
+ rc = ptlrpc_error(req);
RETURN(rc);
}
+ LASSERT(current->journal_info == NULL);
+
EXIT;
- if (!rc) {
+ /* If we're DISCONNECTing, the mds_export_data is already freed */
+ if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
struct mds_export_data *med = &req->rq_export->exp_mds_data;
-
+ struct obd_device *obd = list_entry(mds, struct obd_device,
+ u.mds);
req->rq_repmsg->last_xid =
- HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
- req->rq_repmsg->last_committed =
- HTON__u64(mds->mds_last_committed);
- CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
- (unsigned long long)mds->mds_last_rcvd,
- (unsigned long long)mds->mds_last_committed,
- cpu_to_le32(req->rq_xid));
+ le64_to_cpu(med->med_mcd->mcd_last_xid);
+
+ if (!obd->obd_no_transno) {
+ req->rq_repmsg->last_committed =
+ obd->obd_last_committed;
+ } else {
+ DEBUG_REQ(D_IOCTL, req,
+ "not sending last_committed update");
+ }
+ CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
+ ", xid "LPU64"\n",
+ mds->mds_last_transno, obd->obd_last_committed,
+ req->rq_xid);
}
out:
if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
- struct mds_obd *mds = mds_req2mds(req);
- LASSERT(mds->mds_recoverable_clients);
- DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
- return mds_queue_final_reply(req, rc);
- }
-
- /* MDS_CONNECT / EALREADY (note: not -EALREADY!) isn't an error */
- if (rc && (req->rq_reqmsg->opc != MDS_CONNECT ||
- rc != EALREADY)) {
- CERROR("mds: processing error (opcode %d): %d\n",
- req->rq_reqmsg->opc, rc);
- ptlrpc_error(req->rq_svc, req);
- } else {
- CDEBUG(D_NET, "sending reply\n");
- ptlrpc_reply(req->rq_svc, req);
+ if (obd && obd->obd_recovering) {
+ DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
+ return target_queue_final_reply(req, rc);
+ }
+ /* Lost a race with recovery; let the error path DTRT. */
+ rc = req->rq_status = -ENOTCONN;
}
+
+ target_send_reply(req, rc, fail);
return 0;
}
* then the server last_rcvd value may be less than that of the clients.
* This will alert us that we may need to do client recovery.
*
- * Assumes we are already in the server filesystem context.
- *
- * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
+ * Also assumes for mds_last_transno that we are not modifying it (no locking).
*/
-int mds_update_server_data(struct mds_obd *mds)
+int mds_update_server_data(struct obd_device *obd, int force_sync)
{
+ struct mds_obd *mds = &obd->u.mds;
struct mds_server_data *msd = mds->mds_server_data;
struct file *filp = mds->mds_rcvd_filp;
+ struct lvfs_run_ctxt saved;
loff_t off = 0;
int rc;
+ ENTRY;
- msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
- msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
+ msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
- CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
- (unsigned long long)mds->mds_mount_count,
- (unsigned long long)mds->mds_last_rcvd);
- rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
- if (rc != sizeof(*msd)) {
- CERROR("error writing MDS server data: rc = %d\n", rc);
- if (rc > 0)
- RETURN(-EIO);
- RETURN(rc);
- }
-#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
- rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
-#else
- rc = file_fsync(filp, filp->f_dentry, 1);
-#endif
+ CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
+ mds->mds_mount_count, mds->mds_last_transno);
+ rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
if (rc)
- CERROR("error flushing MDS server data: rc = %d\n", rc);
-
- return 0;
-}
-
-/* Do recovery actions for the MDS */
-static int mds_recovery_complete(struct obd_device *obddev)
-{
- struct mds_obd *mds = &obddev->u.mds;
- struct obd_run_ctxt saved;
- int rc;
-
- LASSERT(mds->mds_recoverable_clients == 0);
-
- /* This happens at the end when recovery is complete */
- ++mds->mds_mount_count;
- push_ctxt(&saved, &mds->mds_ctxt, NULL);
- rc = mds_update_server_data(mds);
- pop_ctxt(&saved);
+ CERROR("error writing MDS server data: rc = %d\n", rc);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
- return rc;
+ RETURN(rc);
}
/* mount the file system (secretly) */
-static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
+static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
{
- struct obd_ioctl_data* data = buf;
- struct mds_obd *mds = &obddev->u.mds;
+ struct lustre_cfg* lcfg = buf;
+ struct mds_obd *mds = &obd->u.mds;
+ char *options = NULL;
struct vfsmount *mnt;
+ unsigned long page;
int rc = 0;
ENTRY;
- MOD_INC_USE_COUNT;
-#ifdef CONFIG_DEV_RDONLY
dev_clear_rdonly(2);
-#endif
- if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
- GOTO(err_dec, rc = -EINVAL);
- mds->mds_fstype = strdup(data->ioc_inlbuf2);
+ if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
+ RETURN(rc = -EINVAL);
+
+ obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
+ if (IS_ERR(obd->obd_fsops))
+ RETURN(rc = PTR_ERR(obd->obd_fsops));
+
+ mds->mds_max_mdsize = sizeof(struct lov_mds_md);
+
+ page = __get_free_page(GFP_KERNEL);
+ if (!page)
+ RETURN(-ENOMEM);
+
+ options = (char *)page;
+ memset(options, 0, PAGE_SIZE);
+
+ /* here we use "iopen_nopriv" hardcoded, because it affects MDS utility
+ * and the rest of options are passed by mount options. Probably this
+ * should be moved to somewhere else like startup scripts or lconf. */
+ sprintf(options, "iopen_nopriv");
+
+ if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
+ sprintf(options + strlen(options), ",%s",
+ lcfg->lcfg_inlbuf4);
+
+ /* we have to know mdsnum before touching underlying fs -bzzz */
+ if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
+ strcmp(lcfg->lcfg_inlbuf5, "dumb")) {
+ class_uuid_t uuid;
+
+ CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
+ obd->obd_name, lcfg->lcfg_inlbuf5);
+
+ generate_random_uuid(uuid);
+ class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
+
+ OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen5);
+ if (mds->mds_lmv_name == NULL)
+ RETURN(rc = -ENOMEM);
+
+ memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf5,
+ lcfg->lcfg_inllen5);
+
+ rc = mds_lmv_connect(obd, mds->mds_lmv_name);
+ if (rc) {
+ OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen5);
+ GOTO(err_ops, rc);
+ }
+ }
+
+ /* FIXME-WANGDI: this should be reworked when we will use lmv along
+ * with cobd, because correct mdsnum is set in mds_lmv_connect(). */
+ if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 && !mds->mds_lmv_obd &&
+ strcmp(lcfg->lcfg_inlbuf6, "dumb")) {
+ if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master")) &&
+ mds->mds_num == 0) {
+ mds->mds_num = REAL_MDS_NUMBER;
+ } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache")) &&
+ mds->mds_num == 0) {
+ mds->mds_num = CACHE_MDS_NUMBER;
+ }
+ }
+
+ mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
+ lcfg->lcfg_inlbuf1, options);
+
+ free_page(page);
- mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
if (IS_ERR(mnt)) {
rc = PTR_ERR(mnt);
CERROR("do_kern_mount failed: rc = %d\n", rc);
- GOTO(err_kfree, rc);
+ GOTO(err_ops, rc);
}
- CERROR("%s: mnt is %p\n", data->ioc_inlbuf1, mnt);
- mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
- if (!mds->mds_sb)
- GOTO(err_put, rc = -ENODEV);
+ CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
- init_MUTEX(&mds->mds_transno_sem);
- mds->mds_max_mdsize = sizeof(struct lov_mds_md);
- rc = mds_fs_setup(obddev, mnt);
+ sema_init(&mds->mds_orphan_recovery_sem, 1);
+ sema_init(&mds->mds_epoch_sem, 1);
+ spin_lock_init(&mds->mds_transno_lock);
+ mds->mds_max_cookiesize = sizeof(struct llog_cookie);
+ atomic_set(&mds->mds_open_count, 0);
+ atomic_set(&mds->mds_real_clients, 0);
+
+ obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
+ LDLM_NAMESPACE_SERVER);
+ if (obd->obd_namespace == NULL) {
+ mds_cleanup(obd, 0);
+ GOTO(err_put, rc = -ENOMEM);
+ }
+ ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
+
+ rc = mds_fs_setup(obd, mnt);
if (rc) {
CERROR("MDS filesystem method init failed: rc = %d\n", rc);
- GOTO(err_put, rc);
+ GOTO(err_ns, rc);
}
- obddev->obd_namespace =
- ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
- if (obddev->obd_namespace == NULL) {
- mds_cleanup(obddev);
- GOTO(err_fs, rc = -ENOMEM);
+ rc = llog_start_commit_thread();
+ if (rc < 0)
+ GOTO(err_fs, rc);
+
+ if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3) {
+ class_uuid_t uuid;
+
+ generate_random_uuid(uuid);
+ class_uuid_unparse(uuid, &mds->mds_lov_uuid);
+
+ OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
+ if (mds->mds_profile == NULL)
+ GOTO(err_fs, rc = -ENOMEM);
+
+ memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
+ lcfg->lcfg_inllen3);
}
ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
- "mds_ldlm_client", &obddev->obd_ldlm_client);
+ "mds_ldlm_client", &obd->obd_ldlm_client);
+ obd->obd_replayable = 1;
+
+ rc = mds_postsetup(obd);
+ if (rc)
+ GOTO(err_fs, rc);
- spin_lock_init(&mds->mds_processing_task_lock);
- mds->mds_processing_task = 0;
- INIT_LIST_HEAD(&mds->mds_recovery_queue);
- INIT_LIST_HEAD(&mds->mds_delayed_reply_queue);
-
RETURN(0);
err_fs:
- mds_fs_cleanup(obddev);
+ /* No extra cleanup needed for llog_init_commit_thread() */
+ mds_fs_cleanup(obd, 0);
+err_ns:
+ ldlm_namespace_free(obd->obd_namespace, 0);
+ obd->obd_namespace = NULL;
err_put:
unlock_kernel();
mntput(mds->mds_vfsmnt);
mds->mds_sb = 0;
lock_kernel();
-err_kfree:
- kfree(mds->mds_fstype);
-err_dec:
- MOD_DEC_USE_COUNT;
- RETURN(rc);
+err_ops:
+ fsfilt_put_ops(obd->obd_fsops);
+ return rc;
}
-static int mds_cleanup(struct obd_device *obddev)
+static int mds_postsetup(struct obd_device *obd)
{
- struct super_block *sb;
- struct mds_obd *mds = &obddev->u.mds;
- struct obd_run_ctxt saved;
+ struct mds_obd *mds = &obd->u.mds;
+ int rc = 0;
ENTRY;
- sb = mds->mds_sb;
- if (!mds->mds_sb)
- RETURN(0);
+ rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
+ obd, 0, NULL, &llog_lvfs_ops);
+ if (rc)
+ RETURN(rc);
- push_ctxt(&saved, &mds->mds_ctxt, NULL);
- mds_update_server_data(mds);
+ /* This check for @dumb string is needed to handle mounting MDS
+ with smfs. Read lconf:MDSDEV.write_conf() for more detail
+ explanation. */
+ if (mds->mds_profile && strcmp(mds->mds_profile, "dumb")) {
+ struct lvfs_run_ctxt saved;
+ struct lustre_profile *lprof;
+ struct config_llog_instance cfg;
+
+ cfg.cfg_instance = NULL;
+ cfg.cfg_uuid = mds->mds_lov_uuid;
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
+ rc = class_config_parse_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT),
+ mds->mds_profile, &cfg);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
+ if (rc)
+ GOTO(err_llog, rc);
- if (mds->mds_rcvd_filp) {
- int rc = filp_close(mds->mds_rcvd_filp, 0);
- mds->mds_rcvd_filp = NULL;
+ lprof = class_get_profile(mds->mds_profile);
+ if (lprof == NULL) {
+ CERROR("No profile found: %s\n", mds->mds_profile);
+ GOTO(err_cleanup, rc = -ENOENT);
+ }
+ rc = mds_lov_connect(obd, lprof->lp_osc);
+ if (rc)
+ GOTO(err_cleanup, rc);
+ rc = mds_lmv_postsetup(obd);
if (rc)
- CERROR("last_rcvd file won't close, rc=%d\n", rc);
+ GOTO(err_cleanup, rc);
+ }
+
+ RETURN(rc);
+
+err_cleanup:
+ mds_lov_clean(obd);
+err_llog:
+ obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
+ LLOG_CONFIG_ORIG_CTXT));
+ RETURN(rc);
+}
+
+int mds_postrecov(struct obd_device *obd)
+{
+ struct mds_obd *mds = &obd->u.mds;
+ struct llog_ctxt *ctxt;
+ int rc, item = 0;
+ ENTRY;
+
+ LASSERT(!obd->obd_recovering);
+ ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
+ LASSERT(ctxt != NULL);
+
+ /* set nextid first, so we are sure it happens */
+ rc = mds_lov_set_nextid(obd);
+ if (rc) {
+ CERROR("%s: mds_lov_set_nextid failed\n", obd->obd_name);
+ GOTO(out, rc);
+ }
+
+ /* clean PENDING dir */
+ rc = mds_cleanup_orphans(obd);
+ if (rc < 0)
+ GOTO(out, rc);
+ item = rc;
+
+ rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
+ NULL, NULL, NULL);
+ if (rc) {
+ CERROR("%s: failed at llog_origin_connect: %d\n",
+ obd->obd_name, rc);
+ GOTO(out, rc);
}
- pop_ctxt(&saved);
+
+ /* remove the orphaned precreated objects */
+ rc = mds_lov_clearorphans(mds, NULL /* all OSTs */);
+ if (rc)
+ GOTO(err_llog, rc);
+
+out:
+ RETURN(rc < 0 ? rc : item);
+
+err_llog:
+ /* cleanup all llogging subsystems */
+ rc = obd_llog_finish(obd, &obd->obd_llogs,
+ mds->mds_lov_desc.ld_tgt_count);
+ if (rc)
+ CERROR("%s: failed to cleanup llogging subsystems\n",
+ obd->obd_name);
+ goto out;
+}
+
+int mds_lov_clean(struct obd_device *obd)
+{
+ struct mds_obd *mds = &obd->u.mds;
+
+ if (mds->mds_profile) {
+ char * cln_prof;
+ struct config_llog_instance cfg;
+ struct lvfs_run_ctxt saved;
+ int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
+
+ OBD_ALLOC(cln_prof, len);
+ sprintf(cln_prof, "%s-clean", mds->mds_profile);
+
+ cfg.cfg_instance = NULL;
+ cfg.cfg_uuid = mds->mds_lov_uuid;
+
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
+ class_config_parse_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT),
+ cln_prof, &cfg);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
+
+ OBD_FREE(cln_prof, len);
+ OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
+ mds->mds_profile = NULL;
+ }
+ RETURN(0);
+}
+
+int mds_lmv_clean(struct obd_device *obd)
+{
+ struct mds_obd *mds = &obd->u.mds;
+
+ if (mds->mds_lmv_name) {
+ OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
+ mds->mds_lmv_name = NULL;
+ }
+ RETURN(0);
+}
+
+static int mds_precleanup(struct obd_device *obd, int flags)
+{
+ int rc = 0;
+ ENTRY;
+
+ mds_lmv_clean(obd);
+ mds_lov_disconnect(obd, flags);
+ mds_lov_clean(obd);
+ obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
+ RETURN(rc);
+}
+
+static int mds_cleanup(struct obd_device *obd, int flags)
+{
+ struct mds_obd *mds = &obd->u.mds;
+ ENTRY;
+
+ if (mds->mds_sb == NULL)
+ RETURN(0);
+
+ mds_update_server_data(obd, 1);
+ if (mds->mds_lov_objids != NULL) {
+ OBD_FREE(mds->mds_lov_objids,
+ mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
+ }
+ mds_fs_cleanup(obd, flags);
unlock_kernel();
+
+ /* 2 seems normal on mds, (may_umount() also expects 2
+ fwiw), but we only see 1 at this point in obdfilter. */
+ if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
+ CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
+ atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
+
mntput(mds->mds_vfsmnt);
+
mds->mds_sb = 0;
- kfree(mds->mds_fstype);
- ldlm_namespace_free(obddev->obd_namespace);
+ ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
+
+ spin_lock_bh(&obd->obd_processing_task_lock);
+ if (obd->obd_recovering) {
+ target_cancel_recovery_timer(obd);
+ obd->obd_recovering = 0;
+ }
+ spin_unlock_bh(&obd->obd_processing_task_lock);
lock_kernel();
-#ifdef CONFIG_DEV_RDONLY
dev_clear_rdonly(2);
-#endif
- mds_fs_cleanup(obddev);
+ fsfilt_put_ops(obd->obd_fsops);
- MOD_DEC_USE_COUNT;
RETURN(0);
}
-static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
- ldlm_mode_t mode, int flags, void *data)
+static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
+ struct ldlm_lock *new_lock,
+ struct lustre_handle *lockh)
{
- struct ptlrpc_request *req = req_cookie;
- int rc = 0;
- ENTRY;
+ struct obd_export *exp = req->rq_export;
+ struct obd_device *obd = exp->exp_obd;
+ struct ldlm_request *dlmreq =
+ lustre_msg_buf(req->rq_reqmsg, 0, sizeof (*dlmreq));
+ struct lustre_handle remote_hdl = dlmreq->lock_handle1;
+ struct list_head *iter;
+
+ if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
+ return;
+
+ l_lock(&obd->obd_namespace->ns_lock);
+ list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
+ struct ldlm_lock *lock;
+ lock = list_entry(iter, struct ldlm_lock, l_export_chain);
+ if (lock == new_lock)
+ continue;
+ if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
+ lockh->cookie = lock->l_handle.h_cookie;
+ DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
+ lockh->cookie);
+ l_unlock(&obd->obd_namespace->ns_lock);
+ return;
+ }
+ }
+ l_unlock(&obd->obd_namespace->ns_lock);
- if (!req_cookie)
- RETURN(0);
+ /* If the xid matches, then we know this is a resent request,
+ * and allow it. (It's probably an OPEN, for which we don't
+ * send a lock */
+ if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
+ return;
- if (req->rq_reqmsg->bufcount > 1) {
- /* an intent needs to be considered */
- struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
- struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
- struct mds_body *mds_rep;
- struct ldlm_reply *rep;
- __u64 new_resid[3] = {0, 0, 0}, old_res;
- int rc, size[3] = {sizeof(struct ldlm_reply),
- sizeof(struct mds_body),
- mds->mds_max_mdsize};
+ /* This remote handle isn't enqueued, so we never received or
+ * processed this request. Clear MSG_RESENT, because it can
+ * be handled like any normal request now. */
- it->opc = NTOH__u64(it->opc);
+ lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
- LDLM_DEBUG(lock, "intent policy, opc: %s",
- ldlm_it2str(it->opc));
+ DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
+ remote_hdl.cookie);
+}
- rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
- &req->rq_repmsg);
- if (rc) {
- rc = req->rq_status = -ENOMEM;
- RETURN(rc);
- }
+int intent_disposition(struct ldlm_reply *rep, int flag)
+{
+ if (!rep)
+ return 0;
+ return (rep->lock_policy_res1 & flag);
+}
- rep = lustre_msg_buf(req->rq_repmsg, 0);
- rep->lock_policy_res1 = 1;
+void intent_set_disposition(struct ldlm_reply *rep, int flag)
+{
+ if (!rep)
+ return;
+ rep->lock_policy_res1 |= flag;
+}
- /* execute policy */
- switch ((long)it->opc) {
- case IT_CREAT|IT_OPEN:
- rc = mds_reint(req, 2);
- if (rc || (req->rq_status != 0 &&
- req->rq_status != -EEXIST)) {
- rep->lock_policy_res2 = req->rq_status;
- RETURN(ELDLM_LOCK_ABORTED);
- }
- break;
- case IT_CREAT:
- case IT_MKDIR:
- case IT_MKNOD:
- case IT_RENAME2:
- case IT_LINK2:
- case IT_RMDIR:
- case IT_SYMLINK:
- case IT_UNLINK:
- rc = mds_reint(req, 2);
- if (rc || (req->rq_status != 0 &&
- req->rq_status != -EISDIR &&
- req->rq_status != -ENOTDIR)) {
- rep->lock_policy_res2 = req->rq_status;
- RETURN(ELDLM_LOCK_ABORTED);
- }
- break;
- case IT_GETATTR:
- case IT_LOOKUP:
- case IT_OPEN:
- case IT_READDIR:
- case IT_READLINK:
- case IT_RENAME:
- case IT_LINK:
- case IT_SETATTR:
- rc = mds_getattr_name(2, req);
- /* FIXME: we need to sit down and decide on who should
- * set req->rq_status, who should return negative and
- * positive return values, and what they all mean. */
- if (rc || req->rq_status != 0) {
- rep->lock_policy_res2 = req->rq_status;
+static int mds_intent_policy(struct ldlm_namespace *ns,
+ struct ldlm_lock **lockp, void *req_cookie,
+ ldlm_mode_t mode, int flags, void *data)
+{
+ struct ptlrpc_request *req = req_cookie;
+ struct ldlm_lock *lock = *lockp;
+ struct ldlm_intent *it;
+ struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
+ struct ldlm_reply *rep;
+ struct lustre_handle lockh = { 0 };
+ struct ldlm_lock *new_lock;
+ int getattr_part = MDS_INODELOCK_UPDATE;
+ int rc, offset = 2, repsize[4] = {sizeof(struct ldlm_reply),
+ sizeof(struct mds_body),
+ mds->mds_max_mdsize,
+ mds->mds_max_cookiesize};
+ ENTRY;
+
+ LASSERT(req != NULL);
+
+ if (req->rq_reqmsg->bufcount <= 1) {
+ /* No intent was provided */
+ int size = sizeof(struct ldlm_reply);
+ rc = lustre_pack_reply(req, 1, &size, NULL);
+ LASSERT(rc == 0);
+ RETURN(0);
+ }
+
+ it = lustre_swab_reqbuf(req, 1, sizeof(*it), lustre_swab_ldlm_intent);
+ if (it == NULL) {
+ CERROR("Intent missing\n");
+ RETURN(req->rq_status = -EFAULT);
+ }
+
+ LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
+
+ rc = lustre_pack_reply(req, it->opc == IT_UNLINK ? 4 : 3, repsize,
+ NULL);
+ if (rc)
+ RETURN(req->rq_status = rc);
+
+ rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
+ intent_set_disposition(rep, DISP_IT_EXECD);
+
+ fixup_handle_for_resent_req(req, lock, &lockh);
+
+ /* execute policy */
+ switch ((long)it->opc) {
+ case IT_OPEN:
+ case IT_CREAT|IT_OPEN:
+ /* XXX swab here to assert that an mds_open reint
+ * packet is following */
+ rep->lock_policy_res2 = mds_reint(req, offset, &lockh);
+#if 0
+ /* We abort the lock if the lookup was negative and
+ * we did not make it to the OPEN portion */
+ if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
+ RETURN(ELDLM_LOCK_ABORTED);
+ if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
+ !intent_disposition(rep, DISP_OPEN_OPEN))
+#endif
+ if (rep->lock_policy_res2) {
+ /* mds_open returns ENOLCK where it
+ * should return zero, but it has no
+ * lock to return */
+ if (rep->lock_policy_res2 == ENOLCK)
+ rep->lock_policy_res2 = 0;
RETURN(ELDLM_LOCK_ABORTED);
}
- break;
- case IT_READDIR|IT_OPEN:
- LBUG();
- break;
- default:
- CERROR("Unhandled intent "LPD64"\n", it->opc);
+ break;
+ case IT_LOOKUP:
+ getattr_part = MDS_INODELOCK_LOOKUP;
+ case IT_CHDIR:
+ case IT_GETATTR:
+ getattr_part |= MDS_INODELOCK_LOOKUP;
+ case IT_READDIR:
+ rep->lock_policy_res2 = mds_getattr_name(offset, req, &lockh,
+ getattr_part);
+ /* FIXME: LDLM can set req->rq_status. MDS sets
+ policy_res{1,2} with disposition and status.
+ - replay: returns 0 & req->status is old status
+ - otherwise: returns req->status */
+ if (intent_disposition(rep, DISP_LOOKUP_NEG))
+ rep->lock_policy_res2 = 0;
+ if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
+ rep->lock_policy_res2)
+ RETURN(ELDLM_LOCK_ABORTED);
+ if (req->rq_status != 0) {
LBUG();
+ rep->lock_policy_res2 = req->rq_status;
+ RETURN(ELDLM_LOCK_ABORTED);
}
+ break;
+ default:
+ CERROR("Unhandled intent "LPD64"\n", it->opc);
+ LBUG();
+ }
- /* We don't bother returning a lock to the client for a file
- * or directory we are removing.
- *
- * As for link and rename, there is no reason for the client
- * to get a lock on the target at this point. If they are
- * going to modify the file/directory later they will get a
- * lock at that time.
- */
- if (it->opc & (IT_UNLINK | IT_RMDIR | IT_LINK | IT_LINK2 |
- IT_RENAME | IT_RENAME2))
- RETURN(ELDLM_LOCK_ABORTED);
+ /* By this point, whatever function we called above must have either
+ * filled in 'lockh', been an intent replay, or returned an error. We
+ * want to allow replayed RPCs to not get a lock, since we would just
+ * drop it below anyways because lock replay is done separately by the
+ * client afterwards. For regular RPCs we want to give the new lock to
+ * the client instead of whatever lock it was about to get. */
+ new_lock = ldlm_handle2lock(&lockh);
+ if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
+ RETURN(0);
- rep->lock_policy_res2 = req->rq_status;
- mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
+ LASSERT(new_lock != NULL);
- /* If the client is about to open a file that doesn't have an MD
- * stripe record, it's going to need a write lock. */
- if (it->opc & IT_OPEN &&
- !(lustre_msg_get_op_flags(req->rq_reqmsg)&MDS_OPEN_HAS_EA)){
- LDLM_DEBUG(lock, "open with no EA; returning PW lock");
- lock->l_req_mode = LCK_PW;
- }
+ /* If we've already given this lock to a client once, then we should
+ * have no readers or writers. Otherwise, we should have one reader
+ * _or_ writer ref (which will be zeroed below) before returning the
+ * lock to a client. */
+ if (new_lock->l_export == req->rq_export) {
+ LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
+ } else {
+ LASSERT(new_lock->l_export == NULL);
+ LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
+ }
- if (flags & LDLM_FL_INTENT_ONLY) {
- LDLM_DEBUG(lock, "INTENT_ONLY, aborting lock");
- RETURN(ELDLM_LOCK_ABORTED);
- }
- /* Give the client a lock on the child object, instead of the
- * parent that it requested. */
- new_resid[0] = NTOH__u32(mds_rep->ino);
- new_resid[1] = NTOH__u32(mds_rep->generation);
- if (new_resid[0] == 0)
- LBUG();
- old_res = lock->l_resource->lr_name[0];
+ *lockp = new_lock;
- ldlm_lock_change_resource(lock, new_resid);
- if (lock->l_resource == NULL) {
- LBUG();
- RETURN(-ENOMEM);
- }
- LDLM_DEBUG(lock, "intent policy, old res %ld",
- (long)old_res);
- RETURN(ELDLM_LOCK_CHANGED);
- } else {
- int size = sizeof(struct ldlm_reply);
- rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
- &req->rq_repmsg);
- if (rc) {
- LBUG();
- RETURN(-ENOMEM);
- }
+ if (new_lock->l_export == req->rq_export) {
+ /* Already gave this to the client, which means that we
+ * reconstructed a reply. */
+ LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
+ MSG_RESENT);
+ RETURN(ELDLM_LOCK_REPLACED);
}
- RETURN(rc);
+
+ /* Fixup the lock to be given to the client */
+ l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
+ new_lock->l_readers = 0;
+ new_lock->l_writers = 0;
+
+ new_lock->l_export = class_export_get(req->rq_export);
+ list_add(&new_lock->l_export_chain,
+ &new_lock->l_export->exp_ldlm_data.led_held_locks);
+
+ new_lock->l_blocking_ast = lock->l_blocking_ast;
+ new_lock->l_completion_ast = lock->l_completion_ast;
+
+ memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
+ sizeof(lock->l_remote_handle));
+
+ new_lock->l_flags &= ~LDLM_FL_LOCAL;
+
+ LDLM_LOCK_PUT(new_lock);
+ l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
+
+ RETURN(ELDLM_LOCK_REPLACED);
}
int mds_attach(struct obd_device *dev, obd_count len, void *data)
{
- return lprocfs_reg_obd(dev, status_var_nm_1, dev);
+ struct lprocfs_static_vars lvars;
+
+ lprocfs_init_multi_vars(0, &lvars);
+ return lprocfs_obd_attach(dev, lvars.obd_vars);
}
int mds_detach(struct obd_device *dev)
{
- return lprocfs_dereg_obd(dev);
+ return lprocfs_obd_detach(dev);
+}
+
+int mdt_attach(struct obd_device *dev, obd_count len, void *data)
+{
+ struct lprocfs_static_vars lvars;
+
+ lprocfs_init_multi_vars(1, &lvars);
+ return lprocfs_obd_attach(dev, lvars.obd_vars);
}
-static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
+int mdt_detach(struct obd_device *dev)
{
- int i;
- // struct obd_ioctl_data* data = buf;
- struct mds_obd *mds = &obddev->u.mds;
+ return lprocfs_obd_detach(dev);
+}
+
+static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
+{
+ struct mds_obd *mds = &obd->u.mds;
int rc = 0;
ENTRY;
- MOD_INC_USE_COUNT;
+ mds->mds_service =
+ ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
+ MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
+ mds_handle, "mds", obd->obd_proc_entry);
- mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
- MDS_BUFSIZE, MDS_MAXREQSIZE,
- MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
- "self", mds_handle, "mds");
if (!mds->mds_service) {
CERROR("failed to start service\n");
- GOTO(err_dec, rc = -EINVAL);
+ RETURN(-ENOMEM);
}
- for (i = 0; i < MDT_NUM_THREADS; i++) {
- char name[32];
- sprintf(name, "ll_mdt_%02d", i);
- rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
- if (rc) {
- CERROR("cannot start MDT thread #%d: rc %d\n", i, rc);
- GOTO(err_thread, rc);
- }
+ rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
+ "ll_mdt");
+ if (rc)
+ GOTO(err_thread, rc);
+
+ mds->mds_setattr_service =
+ ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
+ MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
+ mds_handle, "mds_setattr",
+ obd->obd_proc_entry);
+ if (!mds->mds_setattr_service) {
+ CERROR("failed to start getattr service\n");
+ GOTO(err_thread, rc = -ENOMEM);
}
+ rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
+ MDT_NUM_THREADS, "ll_mdt_attr");
+ if (rc)
+ GOTO(err_thread2, rc);
+
+ mds->mds_readpage_service =
+ ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
+ MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
+ mds_handle, "mds_readpage",
+ obd->obd_proc_entry);
+ if (!mds->mds_readpage_service) {
+ CERROR("failed to start readpage service\n");
+ GOTO(err_thread2, rc = -ENOMEM);
+ }
+
+ rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
+ MDT_NUM_THREADS, "ll_mdt_rdpg");
+
+ if (rc)
+ GOTO(err_thread3, rc);
+
RETURN(0);
+err_thread3:
+ ptlrpc_unregister_service(mds->mds_readpage_service);
+err_thread2:
+ ptlrpc_unregister_service(mds->mds_setattr_service);
err_thread:
- ptlrpc_stop_all_threads(mds->mds_service);
ptlrpc_unregister_service(mds->mds_service);
-err_dec:
- MOD_DEC_USE_COUNT;
- RETURN(rc);
+ return rc;
}
-
-static int mdt_cleanup(struct obd_device *obddev)
+static int mdt_cleanup(struct obd_device *obd, int flags)
{
- struct mds_obd *mds = &obddev->u.mds;
+ struct mds_obd *mds = &obd->u.mds;
ENTRY;
+ ptlrpc_stop_all_threads(mds->mds_readpage_service);
+ ptlrpc_unregister_service(mds->mds_readpage_service);
+
+ ptlrpc_stop_all_threads(mds->mds_setattr_service);
+ ptlrpc_unregister_service(mds->mds_setattr_service);
+
ptlrpc_stop_all_threads(mds->mds_service);
ptlrpc_unregister_service(mds->mds_service);
- MOD_DEC_USE_COUNT;
RETURN(0);
}
-extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
- int len, void *karg, void *uarg);
+static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
+ void *data)
+{
+ struct obd_device *obd = data;
+ struct ll_fid fid;
+ fid.id = id;
+ fid.generation = gen;
+ return mds_fid2dentry(&obd->u.mds, &fid, NULL);
+}
+
+static int mds_get_info(struct obd_export *exp, __u32 keylen,
+ void *key, __u32 *vallen, void *val)
+{
+ struct obd_device *obd;
+ struct mds_obd *mds;
+ ENTRY;
+
+ obd = class_exp2obd(exp);
+ if (obd == NULL) {
+ CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
+ exp->exp_handle.h_cookie);
+ RETURN(-EINVAL);
+ }
+
+ if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
+ /*Get log_context handle*/
+ unsigned long *llh_handle = val;
+ *vallen = sizeof(unsigned long);
+ *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
+ RETURN(0);
+ }
+ if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
+ /*Get log_context handle*/
+ unsigned long *sb = val;
+ *vallen = sizeof(unsigned long);
+ *sb = (unsigned long)obd->u.mds.mds_sb;
+ RETURN(0);
+ }
+
+ mds = &obd->u.mds;
+ keylen == strlen("mdsize");
+ if (keylen && memcmp(key, "mdsize", keylen) == 0) {
+ __u32 *mdsize = val;
+ *vallen = sizeof(*mdsize);
+ *mdsize = mds->mds_max_mdsize;
+ RETURN(0);
+ }
+
+ CDEBUG(D_IOCTL, "invalid key\n");
+ RETURN(-EINVAL);
+
+}
+struct lvfs_callback_ops mds_lvfs_ops = {
+ l_fid2dentry: mds_lvfs_fid2dentry,
+};
+
+int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
+ int objcount, struct obd_ioobj *obj,
+ int niocount, struct niobuf_remote *nb,
+ struct niobuf_local *res,
+ struct obd_trans_info *oti);
+int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
+ int objcount, struct obd_ioobj *obj, int niocount,
+ struct niobuf_local *res, struct obd_trans_info *oti,
+ int rc);
/* use obd ops to offer management infrastructure */
static struct obd_ops mds_obd_ops = {
- o_attach: mds_attach,
- o_detach: mds_detach,
- o_connect: mds_connect,
- o_disconnect: mds_disconnect,
- o_setup: mds_setup,
- o_cleanup: mds_cleanup,
- o_iocontrol: mds_iocontrol
+ .o_owner = THIS_MODULE,
+ .o_attach = mds_attach,
+ .o_detach = mds_detach,
+ .o_connect = mds_connect,
+ .o_init_export = mds_init_export,
+ .o_destroy_export = mds_destroy_export,
+ .o_disconnect = mds_disconnect,
+ .o_setup = mds_setup,
+ .o_precleanup = mds_precleanup,
+ .o_cleanup = mds_cleanup,
+ .o_postrecov = mds_postrecov,
+ .o_statfs = mds_obd_statfs,
+ .o_iocontrol = mds_iocontrol,
+ .o_create = mds_obd_create,
+ .o_destroy = mds_obd_destroy,
+ .o_llog_init = mds_llog_init,
+ .o_llog_finish = mds_llog_finish,
+ .o_notify = mds_notify,
+ .o_get_info = mds_get_info,
+ .o_set_info = mds_set_info,
+ .o_preprw = mds_preprw,
+ .o_commitrw = mds_commitrw,
};
static struct obd_ops mdt_obd_ops = {
- o_setup: mdt_setup,
- o_cleanup: mdt_cleanup,
+ .o_owner = THIS_MODULE,
+ .o_attach = mdt_attach,
+ .o_detach = mdt_detach,
+ .o_setup = mdt_setup,
+ .o_cleanup = mdt_cleanup,
+ .o_attach = mdt_attach,
+ .o_detach = mdt_detach,
};
-
static int __init mds_init(void)
{
+ struct lprocfs_static_vars lvars;
- mds_file_cache = kmem_cache_create("ll_mds_file_data",
- sizeof(struct mds_file_data),
- 0, 0, NULL, NULL);
- if (mds_file_cache == NULL)
- return -ENOMEM;
+ lprocfs_init_multi_vars(0, &lvars);
+ class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
+ LUSTRE_MDS_NAME);
+ lprocfs_init_multi_vars(1, &lvars);
+ class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
+ LUSTRE_MDT_NAME);
- class_register_type(&mds_obd_ops, status_class_var, LUSTRE_MDS_NAME);
- class_register_type(&mdt_obd_ops, 0, LUSTRE_MDT_NAME);
- ldlm_register_intent(ldlm_intent_policy);
return 0;
-
}
-static void __exit mds_exit(void)
+static void /*__exit*/ mds_exit(void)
{
-
-
- ldlm_unregister_intent();
class_unregister_type(LUSTRE_MDS_NAME);
class_unregister_type(LUSTRE_MDT_NAME);
- if (kmem_cache_destroy(mds_file_cache))
- CERROR("couldn't free MDS file cache\n");
-
}
-MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
-MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
+MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
+MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
MODULE_LICENSE("GPL");
module_init(mds_init);
git://git.whamcloud.com / fs / lustre-release.git / blobdiff