// XXX - add transaction sequence numbers
#define EXPORT_SYMTAB
-
-#include <linux/version.h>
-#include <linux/module.h>
-#include <linux/fs.h>
-#include <linux/stat.h>
-#include <linux/locks.h>
-#include <linux/quotaops.h>
-#include <asm/unistd.h>
-#include <asm/uaccess.h>
-
#define DEBUG_SUBSYSTEM S_MDS
#include <linux/obd_support.h>
#include <linux/lustre_lib.h>
#include <linux/lustre_idl.h>
#include <linux/lustre_mds.h>
+#include <linux/lustre_dlm.h>
#include <linux/obd_class.h>
+extern inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req);
+
struct mds_client_info *mds_uuid_to_mci(struct mds_obd *mds, __u8 *uuid)
{
struct list_head *p;
return NULL;
}
+/* Assumes caller has already pushed us into the kernel context. */
int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
struct ptlrpc_request *req)
{
/* get from req->rq_connection-> or req->rq_client */
- struct obd_run_ctxt saved;
struct mds_client_info *mci;
loff_t off;
int rc;
req->rq_repmsg->transno = HTON__u64(mds->mds_last_rcvd);
mci->mci_mcd->mcd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
mci->mci_mcd->mcd_mount_count = cpu_to_le64(mds->mds_mount_count);
- mci->mci_mcd->mcd_last_xid = cpu_to_le32(req->rq_reqmsg->xid);
+ mci->mci_mcd->mcd_last_xid = cpu_to_le64(req->rq_xid);
mds_fs_set_last_rcvd(mds, handle);
- push_ctxt(&saved, &mds->mds_ctxt);
rc = lustre_fwrite(mds->mds_rcvd_filp, (char *)mci->mci_mcd,
sizeof(*mci->mci_mcd), &off);
- pop_ctxt(&saved);
CDEBUG(D_INODE, "wrote trans #%Ld for client '%s' at #%d: rc = %d\n",
mds->mds_last_rcvd, mci->mci_mcd->mcd_uuid, mci->mci_off, rc);
// store new value and last committed value in req struct
return rc;
}
-static int mds_reint_setattr(struct mds_update_record *rec,
+static int mds_reint_setattr(struct mds_update_record *rec, int offset,
struct ptlrpc_request *req)
{
- struct mds_obd *mds = &req->rq_obd->u.mds;
+ struct mds_obd *mds = mds_req2mds(req);
struct dentry *de;
void *handle;
int rc = 0;
return(0);
}
-static int mds_reint_recreate(struct mds_update_record *rec,
+static int mds_reint_recreate(struct mds_update_record *rec, int offset,
struct ptlrpc_request *req)
{
struct dentry *de = NULL;
- struct mds_obd *mds = &req->rq_obd->u.mds;
+ struct mds_obd *mds = mds_req2mds(req);
struct dentry *dchild = NULL;
struct inode *dir;
int rc = 0;
body = lustre_msg_buf(req->rq_repmsg, 0);
body->ino = dchild->d_inode->i_ino;
body->generation = dchild->d_inode->i_generation;
- } else {
+ } else {
CERROR("child doesn't exist (dir %ld, name %s)\n",
dir->i_ino, rec->ur_name);
rc = -ENOENT;
return 0;
}
-static int mds_reint_create(struct mds_update_record *rec,
+static int mds_reint_create(struct mds_update_record *rec, int offset,
struct ptlrpc_request *req)
{
struct dentry *de = NULL;
- struct mds_obd *mds = &req->rq_obd->u.mds;
+ struct mds_obd *mds = mds_req2mds(req);
struct dentry *dchild = NULL;
struct inode *dir;
void *handle;
- int rc = 0, type = rec->ur_mode & S_IFMT;
- int err;
+ struct lustre_handle lockh;
+ int rc = 0, err, flags, lock_mode, type = rec->ur_mode & S_IFMT;
+ __u64 res_id[3] = {0,0,0};
ENTRY;
+ /* requests were at offset 2, replies go back at 1 */
+ if (offset)
+ offset = 1;
+
+ if (strcmp(req->rq_export->export_obd->obd_type->typ_name, "mds") != 0)
+ LBUG();
+
de = mds_fid2dentry(mds, rec->ur_fid1, NULL);
if (IS_ERR(de) || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE)) {
LBUG();
GOTO(out_create_de, rc = -ESTALE);
}
dir = de->d_inode;
- CDEBUG(D_INODE, "parent ino %ld name %s mode %o\n",
+ CDEBUG(D_INODE, "parent ino %ld name %s mode %o\n",
dir->i_ino, rec->ur_name, rec->ur_mode);
+ lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
+ res_id[0] = dir->i_ino;
+
+ rc = ldlm_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
+ NULL, 0, lock_mode, &lockh);
+ if (rc == 0) {
+ LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
+ rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
+ (struct lustre_handle *)&mds->mds_connh,
+ NULL, mds->mds_local_namespace, NULL,
+ res_id, LDLM_PLAIN, NULL, 0, lock_mode,
+ &flags, (void *)mds_lock_callback, NULL,
+ 0, &lockh);
+ if (rc != ELDLM_OK) {
+ CERROR("lock enqueue: err: %d\n", rc);
+ GOTO(out_create_de, rc = -EIO);
+ }
+ }
+ ldlm_lock_dump((void *)(unsigned long)lockh.addr);
+
down(&dir->i_sem);
dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
if (IS_ERR(dchild)) {
}
if (dchild->d_inode) {
- CERROR("child exists (dir %ld, name %s, ino %ld)\n",
+ struct mds_body *body;
+ struct obdo *obdo;
+ struct inode *inode = dchild->d_inode;
+ CDEBUG(D_INODE, "child exists (dir %ld, name %s, ino %ld)\n",
dir->i_ino, rec->ur_name, dchild->d_inode->i_ino);
- LBUG();
+
+ body = lustre_msg_buf(req->rq_repmsg, offset);
+ mds_pack_inode2fid(&body->fid1, inode);
+ mds_pack_inode2body(body, inode);
+#warning FIXME: This ext3/N-specific code does not belong here
+ /* If i_file_acl is set, this inode has an EA */
+ if (S_ISREG(inode->i_mode) && inode->u.ext3_i.i_file_acl) {
+ obdo = lustre_msg_buf(req->rq_repmsg, offset + 1);
+ mds_fs_get_obdo(mds, inode, obdo);
+ }
+ /* now a normal case for intent locking */
GOTO(out_create_dchild, rc = -EEXIST);
}
OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_CREATE_WRITE, dir->i_sb->s_dev);
+ if (dir->i_mode & S_ISGID) {
+ rec->ur_gid = dir->i_gid;
+ if (S_ISDIR(rec->ur_mode))
+ rec->ur_mode |= S_ISGID;
+ }
+
switch (type) {
case S_IFREG: {
handle = mds_fs_start(mds, dir, MDS_FSOP_CREATE);
case S_IFBLK:
case S_IFIFO:
case S_IFSOCK: {
- int rdev = rec->ur_id;
+ int rdev = rec->ur_rdev;
handle = mds_fs_start(mds, dir, MDS_FSOP_MKNOD);
if (!handle)
GOTO(out_create_dchild, PTR_ERR(handle));
if (rc) {
CERROR("error during create: %d\n", rc);
- if (rc != -ENOSPC) LBUG();
+ if (rc != -ENOSPC)
+ LBUG();
GOTO(out_create_commit, rc);
} else {
struct iattr iattr;
struct mds_body *body;
CDEBUG(D_INODE, "created ino %ld\n", dchild->d_inode->i_ino);
- if (type == S_IFREG) {
- rc = mds_fs_set_objid(mds, inode, handle, rec->ur_id);
- if (rc)
- CERROR("error %d setting objid for %ld\n",
+ if (!offset && type == S_IFREG) {
+ struct obdo *obdo;
+ obdo = lustre_msg_buf(req->rq_reqmsg, 2);
+ rc = mds_fs_set_obdo(mds, inode, handle, obdo);
+ if (rc) {
+ CERROR("error %d setting obdo for %ld\n",
rc, inode->i_ino);
+ GOTO(out_create_unlink, rc);
+ }
}
iattr.ia_atime = rec->ur_time;
/* XXX should we abort here in case of error? */
}
- body = lustre_msg_buf(req->rq_repmsg, 0);
+ body = lustre_msg_buf(req->rq_repmsg, offset);
body->ino = inode->i_ino;
body->generation = inode->i_generation;
}
-
+ EXIT;
out_create_commit:
err = mds_fs_commit(mds, dir, handle);
if (err) {
out_create_dchild:
l_dput(dchild);
up(&dir->i_sem);
+ ldlm_lock_decref(&lockh, lock_mode);
out_create_de:
l_dput(de);
req->rq_status = rc;
return 0;
+
+out_create_unlink:
+ /* Destroy the file we just created. This should not need extra
+ * journal credits, as we have already modified all of the blocks
+ * needed in order to create the file in the first place.
+ */
+ switch(type) {
+ case S_IFDIR:
+ err = vfs_rmdir(dir, dchild);
+ if (err)
+ CERROR("failed rmdir in error path: rc = %d\n", err);
+ break;
+ default:
+ err = vfs_unlink(dir, dchild);
+ if (err)
+ CERROR("failed unlink in error path: rc = %d\n", err);
+ }
+
+ goto out_create_commit;
}
-static int mds_reint_unlink(struct mds_update_record *rec,
+static int mds_reint_unlink(struct mds_update_record *rec, int offset,
struct ptlrpc_request *req)
{
struct dentry *de = NULL;
struct dentry *dchild = NULL;
- struct mds_obd *mds = &req->rq_obd->u.mds;
+ struct mds_obd *mds = mds_req2mds(req);
+ struct obdo *obdo;
struct inode *dir, *inode;
+ int lock_mode, flags;
+ __u64 res_id[3] = {0};
+ struct lustre_handle lockh;
void *handle;
int rc = 0;
int err;
LBUG();
GOTO(out_unlink, rc = -ESTALE);
}
+
dir = de->d_inode;
CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
+ lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
+ res_id[0] = dir->i_ino;
+
+ rc = ldlm_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
+ NULL, 0, lock_mode, &lockh);
+ if (rc == 0) {
+ LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
+ rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
+ (struct lustre_handle *)&mds->mds_connh,
+ NULL, mds->mds_local_namespace, NULL,
+ res_id, LDLM_PLAIN, NULL, 0, lock_mode,
+ &flags, (void *)mds_lock_callback, NULL,
+ 0, &lockh);
+ if (rc != ELDLM_OK) {
+ CERROR("lock enqueue: err: %d\n", rc);
+ GOTO(out_unlink_de, rc = -EIO);
+ }
+ } else
+
+ ldlm_lock_dump((void *)(unsigned long)lockh.addr);
down(&dir->i_sem);
dchild = lookup_one_len(rec->ur_name, de, rec->ur_namelen - 1);
if (!inode) {
CERROR("child doesn't exist (dir %ld, name %s\n",
dir->i_ino, rec->ur_name);
- LBUG();
- GOTO(out_unlink_dchild, rc = -ESTALE);
+ GOTO(out_unlink_dchild, rc = -ENOENT);
}
+#if 0 /* in intent case the client doesn't have the inode */
if (inode->i_ino != rec->ur_fid2->id) {
CERROR("inode and FID ID do not match (%ld != %Ld)\n",
inode->i_ino, rec->ur_fid2->id);
LBUG();
GOTO(out_unlink_dchild, rc = -ESTALE);
}
+#endif
OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_UNLINK_WRITE, dir->i_sb->s_dev);
rc = vfs_rmdir(dir, dchild);
break;
default:
+ if (offset) {
+ obdo = lustre_msg_buf(req->rq_repmsg, 1);
+ rc = mds_fs_get_obdo(mds, inode, obdo);
+ if (rc < 0)
+ CDEBUG(D_INFO, "No obdo for ino %ld err %d\n",
+ inode->i_ino, rc);
+ }
+
handle = mds_fs_start(mds, dir, MDS_FSOP_UNLINK);
if (!handle)
GOTO(out_unlink_dchild, rc = PTR_ERR(handle));
EXIT;
out_unlink_dchild:
+ if (!rc)
+ res_id[0] = inode->i_ino;
l_dput(dchild);
out_unlink_de:
up(&dir->i_sem);
+ ldlm_lock_decref(&lockh
+, lock_mode);
+ if (!rc) {
+ /* Take an exclusive lock on the resource that we're
+ * about to free, to force everyone to drop their
+ * locks. */
+ LDLM_DEBUG_NOLOCK("getting EX lock res %Lu", res_id[0]);
+ rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
+ (struct lustre_handle *)&mds->mds_connh,
+ NULL, mds->mds_local_namespace, NULL,
+ res_id,
+ LDLM_PLAIN, NULL, 0, LCK_EX, &flags,
+ (void *)mds_lock_callback, NULL, 0,
+ &lockh);
+ if (rc)
+ CERROR("failed to get child inode lock (child ino %Ld, "
+ "dir ino %ld)\n",
+ res_id[0], de->d_inode->i_ino);
+ }
+
l_dput(de);
+
+ if (!rc) {
+ ldlm_lock_decref(&lockh, LCK_EX);
+ rc = ldlm_cli_cancel(&lockh, NULL);
+ if (rc < 0)
+ CERROR("failed to cancel child inode lock ino "
+ "%Ld: %d\n", res_id[0], rc);
+ }
+
out_unlink:
req->rq_status = rc;
return 0;
}
-static int mds_reint_link(struct mds_update_record *rec,
+static int mds_reint_link(struct mds_update_record *rec, int offset,
struct ptlrpc_request *req)
{
struct dentry *de_src = NULL;
struct dentry *de_tgt_dir = NULL;
struct dentry *dchild = NULL;
- struct mds_obd *mds = &req->rq_obd->u.mds;
+ struct mds_obd *mds = mds_req2mds(req);
void *handle;
int rc = 0;
int err;
}
OBD_FAIL_WRITE(OBD_FAIL_MDS_REINT_LINK_WRITE,
- dchild->d_inode->i_sb->s_dev);
+ de_src->d_inode->i_sb->s_dev);
handle = mds_fs_start(mds, de_tgt_dir->d_inode, MDS_FSOP_LINK);
if (!handle)
return 0;
}
-static int mds_reint_rename(struct mds_update_record *rec,
+static int mds_reint_rename(struct mds_update_record *rec, int offset,
struct ptlrpc_request *req)
{
struct dentry *de_srcdir = NULL;
struct dentry *de_tgtdir = NULL;
struct dentry *de_old = NULL;
struct dentry *de_new = NULL;
- struct mds_obd *mds = &req->rq_obd->u.mds;
+ struct mds_obd *mds = mds_req2mds(req);
void *handle;
int rc = 0;
int err;
GOTO(out_rename_srcdir, rc = -ESTALE);
}
+#warning FIXME: This needs locking attention
+
de_old = lookup_one_len(rec->ur_name, de_srcdir, rec->ur_namelen - 1);
if (IS_ERR(de_old)) {
CERROR("old child lookup error %ld\n", PTR_ERR(de_old));
handle = mds_fs_start(mds, de_tgtdir->d_inode, MDS_FSOP_RENAME);
if (!handle)
GOTO(out_rename_denew, rc = PTR_ERR(handle));
- rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new);
+ rc = vfs_rename(de_srcdir->d_inode, de_old, de_tgtdir->d_inode, de_new,
+ NULL);
if (!rc)
rc = mds_update_last_rcvd(mds, handle, req);
return 0;
}
-typedef int (*mds_reinter)(struct mds_update_record *, struct ptlrpc_request*);
+typedef int (*mds_reinter)(struct mds_update_record *, int offset,
+ struct ptlrpc_request *);
static mds_reinter reinters[REINT_MAX+1] = {
[REINT_SETATTR] mds_reint_setattr,
[REINT_RECREATE] mds_reint_recreate,
};
-int mds_reint_rec(struct mds_update_record *rec, struct ptlrpc_request *req)
+int mds_reint_rec(struct mds_update_record *rec, int offset,
+ struct ptlrpc_request *req)
{
- int rc, size = sizeof(struct mds_body);
+ struct mds_obd *mds = mds_req2mds(req);
+ struct obd_run_ctxt saved;
+
+ int rc;
if (rec->ur_opcode < 1 || rec->ur_opcode > REINT_MAX) {
CERROR("opcode %d not valid\n", rec->ur_opcode);
RETURN(rc);
}
- rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
- if (rc) {
- CERROR("mds: out of memory\n");
- rc = req->rq_status = -ENOMEM;
- RETURN(rc);
- }
-
- rc = reinters[rec->ur_opcode](rec, req);
+ push_ctxt(&saved, &mds->mds_ctxt);
+ rc = reinters[rec->ur_opcode](rec, offset, req);
+ pop_ctxt(&saved);
return rc;
}