#include <linux/lustre_fsfilt.h>
#include <portals/list.h>
+#include <linux/lustre_smfs.h>
#include "mds_internal.h"
/* This limit is arbitrary, but for now we fit it in 1 page (32k clients) */
cl_idx, med->med_mcd->mcd_uuid);
med->med_idx = cl_idx;
- med->med_off = MDS_LR_CLIENT_START + (cl_idx * MDS_LR_CLIENT_SIZE);
+ med->med_off = le32_to_cpu(mds->mds_server_data->msd_client_start) +
+ (cl_idx * le16_to_cpu(mds->mds_server_data->msd_client_size));
if (new_client) {
- struct obd_run_ctxt saved;
+ struct lvfs_run_ctxt saved;
loff_t off = med->med_off;
struct file *file = mds->mds_rcvd_filp;
int rc;
- push_ctxt(&saved, &obd->obd_ctxt, NULL);
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
rc = fsfilt_write_record(obd, file, med->med_mcd,
sizeof(*med->med_mcd), &off, 1);
- pop_ctxt(&saved, &obd->obd_ctxt, NULL);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
if (rc)
return rc;
struct mds_obd *mds = &exp->exp_obd->u.mds;
struct obd_device *obd = exp->exp_obd;
struct mds_client_data zero_mcd;
- struct obd_run_ctxt saved;
+ struct lvfs_run_ctxt saved;
int rc;
unsigned long *bitmap = mds->mds_client_bitmap;
- LASSERT(bitmap);
if (!med->med_mcd)
RETURN(0);
CDEBUG(D_INFO, "freeing client at idx %u (%lld)with UUID '%s'\n",
med->med_idx, med->med_off, med->med_mcd->mcd_uuid);
+ LASSERT(bitmap);
+
/* Clear the bit _after_ zeroing out the client so we don't
race with mds_client_add and zero out new clients.*/
if (!test_bit(med->med_idx, bitmap)) {
if (clear_client) {
memset(&zero_mcd, 0, sizeof zero_mcd);
- push_ctxt(&saved, &obd->obd_ctxt, NULL);
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
sizeof(zero_mcd), &med->med_off, 1);
- pop_ctxt(&saved, &obd->obd_ctxt, NULL);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
CDEBUG(rc == 0 ? D_INFO : D_ERROR,
"zeroing out client %s idx %u in %s rc %d\n",
ENTRY;
/* ensure padding in the struct is the correct size */
- LASSERT (offsetof(struct mds_server_data, msd_padding) +
- sizeof(msd->msd_padding) == MDS_LR_SERVER_SIZE);
- LASSERT (offsetof(struct mds_client_data, mcd_padding) +
- sizeof(mcd->mcd_padding) == MDS_LR_CLIENT_SIZE);
+ LASSERT(offsetof(struct mds_server_data, msd_padding) +
+ sizeof(msd->msd_padding) == MDS_LR_SERVER_SIZE);
+ LASSERT(offsetof(struct mds_client_data, mcd_padding) +
+ sizeof(mcd->mcd_padding) == MDS_LR_CLIENT_SIZE);
OBD_ALLOC_WAIT(msd, sizeof(*msd));
if (!msd)
memcpy(msd->msd_uuid, obd->obd_uuid.uuid,sizeof(msd->msd_uuid));
msd->msd_last_transno = 0;
- mount_count = msd->msd_mount_count = 0;
+ mount_count = msd->msd_mount_count = 0;
msd->msd_server_size = cpu_to_le32(MDS_LR_SERVER_SIZE);
msd->msd_client_start = cpu_to_le32(MDS_LR_CLIENT_START);
msd->msd_client_size = cpu_to_le16(MDS_LR_CLIENT_SIZE);
CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
obd->obd_name, last_rcvd_size);
CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
- last_rcvd_size <= MDS_LR_CLIENT_START ? 0 :
- (last_rcvd_size - MDS_LR_CLIENT_START) / MDS_LR_CLIENT_SIZE);
+ last_rcvd_size <= le32_to_cpu(msd->msd_client_start) ? 0 :
+ (last_rcvd_size - le32_to_cpu(msd->msd_client_start)) /
+ le16_to_cpu(msd->msd_client_size));
/* When we do a clean MDS shutdown, we save the last_transno into
* the header. If we find clients with higher last_transno values
spin_lock_init(&med->med_open_lock);
mcd = NULL;
+ exp->exp_replay_needed = 1;
obd->obd_recoverable_clients++;
obd->obd_max_recoverable_clients++;
class_export_put(exp);
RETURN(rc);
}
+static int mds_fs_post_setup(struct obd_device *obd)
+{
+ struct mds_obd *mds = &obd->u.mds;
+ struct dentry *de = mds_fid2dentry(mds, &mds->mds_rootfid, NULL);
+ int rc = 0;
+
+ rc = fsfilt_post_setup(obd);
+ if (rc)
+ GOTO(out, rc);
+
+ fsfilt_set_fs_flags(obd, de->d_inode,
+ SM_DO_REC | SM_DO_COW);
+ fsfilt_set_fs_flags(obd, mds->mds_pending_dir->d_inode,
+ SM_DO_REC | SM_DO_COW);
+ fsfilt_set_mds_flags(obd, mds->mds_sb);
+out:
+ l_dput(de);
+ return rc;
+}
+
int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
{
struct mds_obd *mds = &obd->u.mds;
- struct obd_run_ctxt saved;
+ struct lvfs_run_ctxt saved;
struct dentry *dentry;
struct file *file;
int rc;
ENTRY;
-
- /* Get rid of unneeded supplementary groups */
- current->ngroups = 0;
- memset(current->groups, 0, sizeof(current->groups));
+ rc = cleanup_group_info();
+ if (rc)
+ RETURN(rc);
mds->mds_vfsmnt = mnt;
mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
fsfilt_setup(obd, mds->mds_sb);
- OBD_SET_CTXT_MAGIC(&obd->obd_ctxt);
- obd->obd_ctxt.pwdmnt = mnt;
- obd->obd_ctxt.pwd = mnt->mnt_root;
- obd->obd_ctxt.fs = get_ds();
- obd->obd_ctxt.cb_ops = mds_lvfs_ops;
+ OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
+ obd->obd_lvfs_ctxt.pwdmnt = mnt;
+ obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
+ obd->obd_lvfs_ctxt.fs = get_ds();
+ obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
/* setup the directory tree */
- push_ctxt(&saved, &obd->obd_ctxt, NULL);
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
if (IS_ERR(dentry)) {
rc = PTR_ERR(dentry);
dentry = lookup_one_len("__iopen__", current->fs->pwd,
strlen("__iopen__"));
- if (IS_ERR(dentry) || !dentry->d_inode) {
- rc = (IS_ERR(dentry)) ? PTR_ERR(dentry): -ENOENT;
- CERROR("cannot open iopen FH directory: rc = %d\n", rc);
+ if (IS_ERR(dentry)) {
+ rc = PTR_ERR(dentry);
+ CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
GOTO(err_pop, rc);
}
+ if (!dentry->d_inode) {
+ rc = -ENOENT;
+ CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
+ GOTO(err_fid, rc);
+ }
mds->mds_fid_de = dentry;
dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
GOTO(err_fid, rc);
}
mds->mds_pending_dir = dentry;
-
+
dentry = simple_mkdir(current->fs->pwd, "LOGS", 0777, 1);
if (IS_ERR(dentry)) {
rc = PTR_ERR(dentry);
}
mds->mds_objects_dir = dentry;
+ dentry = simple_mkdir(current->fs->pwd, "FIDS", 0777, 1);
+ if (IS_ERR(dentry)) {
+ rc = PTR_ERR(dentry);
+ CERROR("cannot create FIDS directory: rc = %d\n", rc);
+ GOTO(err_fids, rc);
+ }
+ mds->mds_fids_dir = dentry;
+
/* open and test the last rcvd file */
file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
if (IS_ERR(file)) {
GOTO(err_lov_objid, rc = -ENOENT);
}
err_pop:
- pop_ctxt(&saved, &obd->obd_ctxt, NULL);
-
+ if (!rc) {
+ rc = mds_fs_post_setup(obd);
+ if (rc)
+ CERROR("can not post setup fsfilt\n");
+ }
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
return rc;
err_lov_objid:
err_last_rcvd:
if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
CERROR("can't close %s after error\n", LAST_RCVD);
+err_fids:
+ dput(mds->mds_fids_dir);
err_objects:
dput(mds->mds_objects_dir);
err_logs:
goto err_pop;
}
+static int mds_fs_post_cleanup(struct obd_device *obd)
+{
+ int rc = 0;
+ rc = fsfilt_post_cleanup(obd);
+ return rc;
+}
int mds_fs_cleanup(struct obd_device *obd, int flags)
{
struct mds_obd *mds = &obd->u.mds;
- struct obd_run_ctxt saved;
+ struct lvfs_run_ctxt saved;
int rc = 0;
if (flags & OBD_OPT_FAILOVER)
class_disconnect_exports(obd, flags); /* cleans up client info too */
mds_server_free_data(mds);
- push_ctxt(&saved, &obd->obd_ctxt, NULL);
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
if (mds->mds_rcvd_filp) {
rc = filp_close(mds->mds_rcvd_filp, 0);
mds->mds_rcvd_filp = NULL;
if (rc)
CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
}
+ if (mds->mds_fids_dir != NULL) {
+ l_dput(mds->mds_fids_dir);
+ mds->mds_fids_dir = NULL;
+ }
if (mds->mds_objects_dir != NULL) {
l_dput(mds->mds_objects_dir);
mds->mds_objects_dir = NULL;
l_dput(mds->mds_pending_dir);
mds->mds_pending_dir = NULL;
}
- pop_ctxt(&saved, &obd->obd_ctxt, NULL);
+ rc = mds_fs_post_cleanup(obd);
+
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
shrink_dcache_parent(mds->mds_fid_de);
dput(mds->mds_fid_de);
* performance sensitive, it is accomplished by creating a file, checking the
* fid, and renaming it. */
int mds_obd_create(struct obd_export *exp, struct obdo *oa,
- struct lov_stripe_md **ea, struct obd_trans_info *oti)
+ struct lov_stripe_md **ea, struct obd_trans_info *oti)
{
struct mds_obd *mds = &exp->exp_obd->u.mds;
struct inode *parent_inode = mds->mds_objects_dir->d_inode;
unsigned int tmpname = ll_insecure_random_int();
struct file *filp;
struct dentry *new_child;
- struct obd_run_ctxt saved;
+ struct lvfs_run_ctxt saved;
char fidname[LL_FID_NAMELEN];
void *handle;
int rc = 0, err, namelen;
ENTRY;
- push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
+ push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
sprintf(fidname, "OBJECTS/%u", tmpname);
filp = filp_open(fidname, O_CREAT | O_EXCL, 0644);
handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
FSFILT_OP_RENAME, NULL);
- if (IS_ERR(handle))
+ if (IS_ERR(handle))
GOTO(out_dput, rc = PTR_ERR(handle));
-
+
lock_kernel();
rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
mds->mds_objects_dir->d_inode, new_child);
CERROR("error renaming new object "LPU64":%u: rc %d\n",
oa->o_id, oa->o_generation, rc);
- err = fsfilt_commit(exp->exp_obd, mds->mds_objects_dir->d_inode,
- handle, 0);
- if (!err)
- oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
+ err = fsfilt_commit(exp->exp_obd, mds->mds_sb,
+ mds->mds_objects_dir->d_inode, handle, 0);
+ if (!err) {
+ oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
+ oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLGROUP;
+ }
else if (!rc)
rc = err;
out_dput:
rc = err;
}
out_pop:
- pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
+ pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
RETURN(rc);
}
struct mds_obd *mds = &exp->exp_obd->u.mds;
struct inode *parent_inode = mds->mds_objects_dir->d_inode;
struct obd_device *obd = exp->exp_obd;
- struct obd_run_ctxt saved;
+ struct lvfs_run_ctxt saved;
char fidname[LL_FID_NAMELEN];
struct dentry *de;
void *handle;
int err, namelen, rc = 0;
ENTRY;
- push_ctxt(&saved, &obd->obd_ctxt, NULL);
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
down(&parent_inode->i_sem);
de = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
if (de == NULL || de->d_inode == NULL) {
- CERROR("destroying non-existent object "LPU64"\n", oa->o_id);
+ CERROR("destroying non-existent object "LPU64" %s\n",
+ oa->o_id, fidname);
GOTO(out_dput, rc = IS_ERR(de) ? PTR_ERR(de) : -ENOENT);
}
- handle = fsfilt_start(obd, mds->mds_objects_dir->d_inode,
- FSFILT_OP_UNLINK_LOG, oti);
- if (IS_ERR(handle)) {
+ /* Stripe count is 1 here since this is some MDS specific stuff
+ that is unlinked, not spanned across multiple OSTs */
+ handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
+ FSFILT_OP_UNLINK, oti, 1);
+
+ if (IS_ERR(handle))
GOTO(out_dput, rc = PTR_ERR(handle));
- }
rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
if (rc)
CERROR("error destroying object "LPU64":%u: rc %d\n",
oa->o_id, oa->o_generation, rc);
- err = fsfilt_commit(obd, mds->mds_objects_dir->d_inode, handle, 0);
+ err = fsfilt_commit(obd, mds->mds_sb, mds->mds_objects_dir->d_inode,
+ handle, 0);
if (err && !rc)
rc = err;
out_dput:
if (de != NULL)
l_dput(de);
up(&parent_inode->i_sem);
- pop_ctxt(&saved, &obd->obd_ctxt, NULL);
+ pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
RETURN(rc);
}