* 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/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) */
#define MDS_MAX_CLIENT_WORDS (MDS_MAX_CLIENTS / sizeof(unsigned long))
#define LAST_RCVD "last_rcvd"
+#define LOV_OBJID "lov_objid"
/* Add client data to the MDS. We use a bitmap to locate a free space
* in the last_rcvd file if cl_off is -1 (i.e. a new client).
{
unsigned long *bitmap = mds->mds_client_bitmap;
int new_client = (cl_idx == -1);
+ ENTRY;
LASSERT(bitmap != NULL);
/* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
- if (!strcmp(med->med_mcd->mcd_uuid, "OBD_CLASS_UUID"))
+ if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
RETURN(0);
/* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
return -ENOMEM;
}
if (test_and_set_bit(cl_idx, bitmap)) {
- CERROR("MDS client %d: found bit is set in bitmap\n",
- cl_idx);
cl_idx = find_next_zero_bit(bitmap, MDS_MAX_CLIENTS,
cl_idx);
goto repeat;
}
}
- CDEBUG(D_INFO, "client at index %d with UUID '%s' added\n",
+ CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
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;
- ssize_t written;
- void *handle;
-
- push_ctxt(&saved, &mds->mds_ctxt, NULL);
- /* We need to start a transaction here first, to avoid a
- * possible ordering deadlock on last_rcvd->i_sem and the
- * journal lock. In most places we start the journal handle
- * first (because we do compound transactions), and then
- * later do the write into last_rcvd, which gets i_sem.
- *
- * Without this transaction, clients connecting at the same
- * time other MDS operations are ongoing get last_rcvd->i_sem
- * first (in generic_file_write()) and start the journal
- * transaction afterwards, and can deadlock with other ops.
- *
- * We use FSFILT_OP_SETATTR because it is smallest, but all
- * ops include enough space for the last_rcvd update so we
- * could use any of them, or maybe an FSFILT_OP_NONE is best?
- */
- handle = fsfilt_start(obd,mds->mds_rcvd_filp->f_dentry->d_inode,
- FSFILT_OP_SETATTR, NULL);
- if (IS_ERR(handle)) {
- written = PTR_ERR(handle);
- CERROR("unable to start transaction: rc %d\n",
- (int)written);
- } else {
- written = fsfilt_write_record(obd, mds->mds_rcvd_filp,
- med->med_mcd,
- sizeof(*med->med_mcd),
- &off);
- fsfilt_commit(obd,mds->mds_rcvd_filp->f_dentry->d_inode,
- handle, 0);
- }
- pop_ctxt(&saved, &mds->mds_ctxt, NULL);
+ struct file *file = mds->mds_rcvd_filp;
+ int rc;
- if (written != sizeof(*med->med_mcd)) {
- if (written < 0)
- RETURN(written);
- RETURN(-EIO);
- }
+ 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_lvfs_ctxt, NULL);
+
+ if (rc)
+ return rc;
CDEBUG(D_INFO, "wrote client mcd at idx %u off %llu (len %u)\n",
med->med_idx, med->med_off,
(unsigned int)sizeof(*med->med_mcd));
return 0;
}
-int mds_client_free(struct obd_export *exp)
+int mds_client_free(struct obd_export *exp, int clear_client)
{
struct mds_export_data *med = &exp->exp_mds_data;
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;
- int written;
+ struct lvfs_run_ctxt saved;
+ int rc;
unsigned long *bitmap = mds->mds_client_bitmap;
- LASSERT(bitmap);
if (!med->med_mcd)
RETURN(0);
/* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
- if (!strcmp(med->med_mcd->mcd_uuid, "OBD_CLASS_UUID"))
+ if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
GOTO(free_and_out, 0);
- CDEBUG(D_INFO, "freeing client at index %u (%lld)with UUID '%s'\n",
+ CDEBUG(D_INFO, "freeing client at idx %u (%lld)with UUID '%s'\n",
med->med_idx, med->med_off, med->med_mcd->mcd_uuid);
- if (!test_and_clear_bit(med->med_idx, bitmap)) {
+ 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)) {
CERROR("MDS client %u: bit already clear in bitmap!!\n",
med->med_idx);
LBUG();
}
- memset(&zero_mcd, 0, sizeof zero_mcd);
- push_ctxt(&saved, &mds->mds_ctxt, NULL);
- written = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
- sizeof(zero_mcd), &med->med_off);
- pop_ctxt(&saved, &mds->mds_ctxt, NULL);
+ if (clear_client) {
+ memset(&zero_mcd, 0, sizeof zero_mcd);
+ 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_lvfs_ctxt, NULL);
- if (written != sizeof(zero_mcd)) {
- CERROR("error zeroing out client %s index %d in %s: %d\n",
- med->med_mcd->mcd_uuid, med->med_idx, LAST_RCVD,
- written);
- } else {
- CDEBUG(D_INFO, "zeroed out disconnecting client %s at off %d\n",
- med->med_mcd->mcd_uuid, med->med_idx);
+ CDEBUG(rc == 0 ? D_INFO : D_ERROR,
+ "zeroing out client %s idx %u in %s rc %d\n",
+ med->med_mcd->mcd_uuid, med->med_idx, LAST_RCVD, rc);
}
+ if (!test_and_clear_bit(med->med_idx, bitmap)) {
+ CERROR("MDS client %u: bit already clear in bitmap!!\n",
+ med->med_idx);
+ LBUG();
+ }
+
+
+ /* Make sure the server's last_transno is up to date. Do this
+ * after the client is freed so we know all the client's
+ * transactions have been committed. */
+ mds_update_server_data(exp->exp_obd, 1);
+
free_and_out:
OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
struct mds_server_data *msd;
struct mds_client_data *mcd = NULL;
loff_t off = 0;
- int cl_idx;
unsigned long last_rcvd_size = file->f_dentry->d_inode->i_size;
- __u64 last_transno = 0;
__u64 mount_count;
- int rc = 0;
+ int cl_idx, rc = 0;
+ ENTRY;
- LASSERT(sizeof(struct mds_client_data) == MDS_LR_CLIENT_SIZE);
- LASSERT(sizeof(struct mds_server_data) <= MDS_LR_SERVER_SIZE);
+ /* 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);
- OBD_ALLOC(msd, sizeof(*msd));
+ OBD_ALLOC_WAIT(msd, sizeof(*msd));
if (!msd)
RETURN(-ENOMEM);
- OBD_ALLOC(mds->mds_client_bitmap,
+ OBD_ALLOC_WAIT(mds->mds_client_bitmap,
MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
if (!mds->mds_client_bitmap) {
OBD_FREE(msd, sizeof(*msd));
mds->mds_server_data = msd;
if (last_rcvd_size == 0) {
- int written;
CWARN("%s: initializing new %s\n", obd->obd_name, LAST_RCVD);
+
memcpy(msd->msd_uuid, obd->obd_uuid.uuid,sizeof(msd->msd_uuid));
+ msd->msd_last_transno = 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);
- written = fsfilt_write_record(obd, file, msd, sizeof(*msd),
- &off);
-
- if (written == sizeof(*msd))
- RETURN(0);
- CERROR("%s: error writing new MSD: %d\n", obd->obd_name,
- written);
- GOTO(err_msd, rc = (written < 0 ? written : -EIO));
- }
-
- rc = fsfilt_read_record(obd, file, msd, sizeof(*msd), &off);
-
- if (rc != sizeof(*msd)) {
- CERROR("error reading MDS %s: rc = %d\n", LAST_RCVD,rc);
- if (rc > 0)
- rc = -EIO;
- GOTO(err_msd, rc);
+ msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
+ } else {
+ rc = fsfilt_read_record(obd, file, msd, sizeof(*msd), &off);
+ if (rc) {
+ CERROR("error reading MDS %s: rc = %d\n", LAST_RCVD, rc);
+ GOTO(err_msd, rc);
+ }
+ if (strcmp(msd->msd_uuid, obd->obd_uuid.uuid) != 0) {
+ CERROR("OBD UUID %s does not match last_rcvd UUID %s\n",
+ obd->obd_uuid.uuid, msd->msd_uuid);
+ GOTO(err_msd, rc = -EINVAL);
+ }
+ mount_count = le64_to_cpu(msd->msd_mount_count);
}
- if (!msd->msd_server_size)
- msd->msd_server_size = cpu_to_le32(MDS_LR_SERVER_SIZE);
- if (!msd->msd_client_start)
- msd->msd_client_start = cpu_to_le32(MDS_LR_CLIENT_START);
- if (!msd->msd_client_size)
- msd->msd_client_size = cpu_to_le16(MDS_LR_CLIENT_SIZE);
-
- if (msd->msd_feature_incompat) {
+ if (msd->msd_feature_incompat & ~cpu_to_le32(MDS_INCOMPAT_SUPP)) {
CERROR("unsupported incompat feature %x\n",
- le32_to_cpu(msd->msd_feature_incompat));
+ le32_to_cpu(msd->msd_feature_incompat) &
+ ~MDS_INCOMPAT_SUPP);
GOTO(err_msd, rc = -EINVAL);
}
- if (msd->msd_feature_rocompat) {
+ /* XXX updating existing b_devel fs only, can be removed in future */
+ msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
+ if (msd->msd_feature_rocompat & ~cpu_to_le32(MDS_ROCOMPAT_SUPP)) {
CERROR("unsupported read-only feature %x\n",
- le32_to_cpu(msd->msd_feature_rocompat));
+ le32_to_cpu(msd->msd_feature_rocompat) &
+ ~MDS_ROCOMPAT_SUPP);
/* Do something like remount filesystem read-only */
GOTO(err_msd, rc = -EINVAL);
}
- last_transno = le64_to_cpu(msd->msd_last_transno);
- mds->mds_last_transno = last_transno;
-
- mount_count = le64_to_cpu(msd->msd_mount_count);
- mds->mds_mount_count = mount_count;
+ mds->mds_last_transno = le64_to_cpu(msd->msd_last_transno);
CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
- obd->obd_name, last_transno);
+ obd->obd_name, mds->mds_last_transno);
CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
- obd->obd_name, mount_count);
+ obd->obd_name, mount_count + 1);
CDEBUG(D_INODE, "%s: server data size: %u\n",
obd->obd_name, le32_to_cpu(msd->msd_server_size));
CDEBUG(D_INODE, "%s: per-client data start: %u\n",
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) / 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 FILTER shutdown, we save the last_transno into
+ /* When we do a clean MDS shutdown, we save the last_transno into
* the header. If we find clients with higher last_transno values
* then those clients may need recovery done. */
- for (cl_idx = 0; off < last_rcvd_size; cl_idx++) {
+ for (cl_idx = 0, off = le32_to_cpu(msd->msd_client_start);
+ off < last_rcvd_size; cl_idx++) {
__u64 last_transno;
- int mount_age;
+ struct obd_export *exp;
+ struct mds_export_data *med;
if (!mcd) {
- OBD_ALLOC(mcd, sizeof(*mcd));
+ OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
if (!mcd)
- GOTO(err_msd, rc = -ENOMEM);
+ GOTO(err_client, rc = -ENOMEM);
}
- /* Don't assume off is incremented properly, in case
- * sizeof(fsd) isn't the same as fsd->fsd_client_size.
- */
+ /* Don't assume off is incremented properly by
+ * fsfilt_read_record(), in case sizeof(*mcd)
+ * isn't the same as msd->msd_client_size. */
off = le32_to_cpu(msd->msd_client_start) +
cl_idx * le16_to_cpu(msd->msd_client_size);
rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
- if (rc != sizeof(*mcd)) {
- CERROR("error reading MDS %s offset %d: rc = %d\n",
- LAST_RCVD, cl_idx, rc);
- if (rc > 0) /* XXX fatal error or just abort reading? */
- rc = -EIO;
- break;
+ if (rc) {
+ CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
+ LAST_RCVD, cl_idx, off, rc);
+ break; /* read error shouldn't cause startup to fail */
}
if (mcd->mcd_uuid[0] == '\0') {
/* These exports are cleaned up by mds_disconnect(), so they
* need to be set up like real exports as mds_connect() does.
*/
- mount_age = mount_count - le64_to_cpu(mcd->mcd_mount_count);
- if (mount_age < MDS_MOUNT_RECOV) {
- struct obd_export *exp = class_new_export(obd);
- struct mds_export_data *med;
- CERROR("RCVRNG CLIENT uuid: %s off: %d lr: "LPU64
- "srv lr: "LPU64" mnt: "LPU64" last mount: "LPU64
- "\n", mcd->mcd_uuid, cl_idx,
- last_transno, le64_to_cpu(msd->msd_last_transno),
- le64_to_cpu(mcd->mcd_mount_count), mount_count);
-
- if (!exp) {
- rc = -ENOMEM;
- break;
- }
-
- memcpy(&exp->exp_client_uuid.uuid, mcd->mcd_uuid,
- sizeof exp->exp_client_uuid.uuid);
- med = &exp->exp_mds_data;
- med->med_mcd = mcd;
- mds_client_add(obd, mds, med, cl_idx);
- /* create helper if export init gets more complex */
- INIT_LIST_HEAD(&med->med_open_head);
- spin_lock_init(&med->med_open_lock);
-
- mcd = NULL;
- obd->obd_recoverable_clients++;
- class_export_put(exp);
- } else {
- CDEBUG(D_INFO, "discarded client %d, UUID '%s', count "
- LPU64"\n", cl_idx, mcd->mcd_uuid,
- le64_to_cpu(mcd->mcd_mount_count));
- }
-
- CDEBUG(D_OTHER, "client at offset %d has last_transno = "
- LPU64"\n", cl_idx, last_transno);
+ CDEBUG(D_HA|D_WARNING,"RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
+ " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
+ last_transno, le64_to_cpu(msd->msd_last_transno),
+ mcd->mcd_last_xid);
+
+ exp = class_new_export(obd);
+ if (exp == NULL)
+ GOTO(err_client, rc = -ENOMEM);
+
+ memcpy(&exp->exp_client_uuid.uuid, mcd->mcd_uuid,
+ sizeof exp->exp_client_uuid.uuid);
+ med = &exp->exp_mds_data;
+ med->med_mcd = mcd;
+ mds_client_add(obd, mds, med, cl_idx);
+ /* create helper if export init gets more complex */
+ INIT_LIST_HEAD(&med->med_open_head);
+ 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);
+
+ CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
+ cl_idx, last_transno);
if (last_transno > mds->mds_last_transno)
- mds->mds_last_transno = last_transno;
+ mds->mds_last_transno = last_transno;
}
-
+ if (mcd)
+ OBD_FREE(mcd, sizeof(*mcd));
obd->obd_last_committed = mds->mds_last_transno;
if (obd->obd_recoverable_clients) {
- CERROR("RECOVERY: %d recoverable clients, last_transno "
- LPU64"\n",
- obd->obd_recoverable_clients, mds->mds_last_transno);
- obd->obd_next_recovery_transno = obd->obd_last_committed
- + 1;
- obd->obd_recovering = 1;
+ CWARN("RECOVERY: service %s, %d recoverable clients, "
+ "last_transno "LPU64"\n", obd->obd_name,
+ obd->obd_recoverable_clients, mds->mds_last_transno);
+ obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
+ target_start_recovery_thread(obd, mds_handle);
+ obd->obd_recovery_start = LTIME_S(CURRENT_TIME);
}
+
+ mds->mds_mount_count = mount_count + 1;
+ msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
- if (mcd)
- OBD_FREE(mcd, sizeof(*mcd));
+ /* save it, so mount count and last_transno is current */
+ rc = mds_update_server_data(obd, 1);
+ if (rc)
+ GOTO(err_client, rc);
- return 0;
+ RETURN(0);
+err_client:
+ class_disconnect_exports(obd, 0);
err_msd:
mds_server_free_data(mds);
- return rc;
+ RETURN(rc);
}
-static int mds_fs_prep(struct obd_device *obd)
+static int mds_fs_post_setup(struct obd_device *obd)
{
struct mds_obd *mds = &obd->u.mds;
- struct obd_run_ctxt saved;
+ struct dentry *de = mds_fid2dentry(mds, &mds->mds_rootfid, NULL);
+ int rc = 0;
+
+ rc = fsfilt_post_setup(obd, de);
+ 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 lvfs_run_ctxt saved;
struct dentry *dentry;
struct file *file;
int rc;
+ ENTRY;
+
+ rc = cleanup_group_info();
+ if (rc)
+ RETURN(rc);
- push_ctxt(&saved, &mds->mds_ctxt, NULL);
- dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755);
+ 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_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_lvfs_ctxt, NULL);
+ dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
if (IS_ERR(dentry)) {
rc = PTR_ERR(dentry);
CERROR("cannot create ROOT directory: rc = %d\n", rc);
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);
+ dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
if (IS_ERR(dentry)) {
rc = PTR_ERR(dentry);
CERROR("cannot create PENDING directory: rc = %d\n", rc);
GOTO(err_fid, rc);
}
mds->mds_pending_dir = dentry;
-
- dentry = simple_mkdir(current->fs->pwd, "LOGS", 0700);
+
+ dentry = simple_mkdir(current->fs->pwd, "LOGS", 0777, 1);
if (IS_ERR(dentry)) {
rc = PTR_ERR(dentry);
CERROR("cannot create LOGS directory: rc = %d\n", rc);
}
mds->mds_logs_dir = dentry;
+ dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
+ if (IS_ERR(dentry)) {
+ rc = PTR_ERR(dentry);
+ CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
+ GOTO(err_logs, rc);
+ }
+ 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;
+
+ dentry = simple_mkdir(current->fs->pwd, "UNNAMED", 0777, 1);
+ if (IS_ERR(dentry)) {
+ rc = PTR_ERR(dentry);
+ CERROR("cannot create UNNAMED directory: rc = %d\n", rc);
+ GOTO(err_unnamed, rc);
+ }
+ mds->mds_unnamed_dir = dentry;
+
+ /* open and test the last rcvd file */
file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
if (IS_ERR(file)) {
rc = PTR_ERR(file);
CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
-
- GOTO(err_logs, rc = PTR_ERR(file));
+ GOTO(err_objects, rc = PTR_ERR(file));
}
+ mds->mds_rcvd_filp = file;
if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
file->f_dentry->d_inode->i_mode);
- GOTO(err_filp, rc = -ENOENT);
- }
-
- rc = fsfilt_journal_data(obd, file);
- if (rc) {
- CERROR("cannot journal data on %s: rc = %d\n", LAST_RCVD, rc);
- GOTO(err_filp, rc);
+ GOTO(err_last_rcvd, rc = -ENOENT);
}
rc = mds_read_last_rcvd(obd, file);
if (rc) {
CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
- GOTO(err_client, rc);
+ GOTO(err_last_rcvd, rc);
}
- mds->mds_rcvd_filp = file;
-#ifdef I_SKIP_PDFLUSH
- /*
- * we need this to protect from deadlock
- * pdflush vs. lustre_fwrite()
- */
- file->f_dentry->d_inode->i_flags |= I_SKIP_PDFLUSH;
-#endif
-err_pop:
- pop_ctxt(&saved, &mds->mds_ctxt, NULL);
+ /* open and test the lov objd file */
+ file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
+ if (IS_ERR(file)) {
+ rc = PTR_ERR(file);
+ CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
+ GOTO(err_client, rc = PTR_ERR(file));
+ }
+ mds->mds_lov_objid_filp = file;
+ if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
+ CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
+ file->f_dentry->d_inode->i_mode);
+ GOTO(err_lov_objid, rc = -ENOENT);
+ }
+err_pop:
+ 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:
+ if (mds->mds_lov_objid_filp && filp_close(mds->mds_lov_objid_filp, 0))
+ CERROR("can't close %s after error\n", LOV_OBJID);
err_client:
class_disconnect_exports(obd, 0);
-err_filp:
- if (filp_close(file, 0))
+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_unnamed:
+ dput(mds->mds_unnamed_dir);
+err_fids:
+ dput(mds->mds_fids_dir);
+err_objects:
+ dput(mds->mds_objects_dir);
err_logs:
dput(mds->mds_logs_dir);
err_pending:
goto err_pop;
}
-int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
+static int mds_fs_post_cleanup(struct obd_device *obd)
{
- struct mds_obd *mds = &obd->u.mds;
- ENTRY;
-
- mds->mds_vfsmnt = mnt;
-
- OBD_SET_CTXT_MAGIC(&mds->mds_ctxt);
- mds->mds_ctxt.pwdmnt = mnt;
- mds->mds_ctxt.pwd = mnt->mnt_root;
- mds->mds_ctxt.fs = get_ds();
- RETURN(mds_fs_prep(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)
" be preserved.\n", obd->obd_name);
class_disconnect_exports(obd, flags); /* cleans up client info too */
+ target_cleanup_recovery(obd);
mds_server_free_data(mds);
- push_ctxt(&saved, &mds->mds_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", LAST_RCVD, rc);
}
+ if (mds->mds_lov_objid_filp) {
+ rc = filp_close(mds->mds_lov_objid_filp, 0);
+ mds->mds_lov_objid_filp = NULL;
+ if (rc)
+ CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
+ }
+ if (mds->mds_unnamed_dir != NULL) {
+ l_dput(mds->mds_unnamed_dir);
+ mds->mds_unnamed_dir = NULL;
+ }
+ 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;
+ }
if (mds->mds_logs_dir) {
l_dput(mds->mds_logs_dir);
mds->mds_logs_dir = NULL;
l_dput(mds->mds_pending_dir);
mds->mds_pending_dir = NULL;
}
- pop_ctxt(&saved, &mds->mds_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);
return rc;
}
-/* This is a callback from the llog_* functions.
- * Assumes caller has already pushed us into the kernel context. */
-int mds_log_close(struct llog_handle *cathandle, struct llog_handle *loghandle)
+/* Creates an object with the same name as its fid. Because this is not at all
+ * 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 llog_object_hdr *llh = loghandle->lgh_hdr;
- struct mds_obd *mds = &cathandle->lgh_obd->u.mds;
- struct dentry *dchild = NULL;
- int rc;
- ENTRY;
-
- /* If we are going to delete this log, grab a ref before we close
- * it so we don't have to immediately do another lookup.
- */
- if (llh->llh_hdr.lth_type != LLOG_CATALOG_MAGIC && llh->llh_count == 0){
- CDEBUG(D_INODE, "deleting log file "LPX64":%x\n",
- loghandle->lgh_cookie.lgc_lgl.lgl_oid,
- loghandle->lgh_cookie.lgc_lgl.lgl_ogen);
- down(&mds->mds_logs_dir->d_inode->i_sem);
- dchild = dget(loghandle->lgh_file->f_dentry);
- llog_delete_log(cathandle, loghandle);
- } else {
- CDEBUG(D_INODE, "closing log file "LPX64":%x\n",
- loghandle->lgh_cookie.lgc_lgl.lgl_oid,
- loghandle->lgh_cookie.lgc_lgl.lgl_ogen);
- }
-
- rc = filp_close(loghandle->lgh_file, 0);
-
- llog_free_handle(loghandle); /* also removes loghandle from list */
-
- if (dchild) {
- int err = vfs_unlink(mds->mds_logs_dir->d_inode, dchild);
- if (err) {
- CERROR("error unlinking empty log %*s: rc %d\n",
- dchild->d_name.len, dchild->d_name.name, err);
- if (!rc)
- rc = err;
- }
- l_dput(dchild);
- up(&mds->mds_logs_dir->d_inode->i_sem);
- }
- RETURN(rc);
-}
-
-/* This is a callback from the llog_* functions.
- * Assumes caller has already pushed us into the kernel context. */
-struct llog_handle *mds_log_open(struct obd_device *obd,
- struct llog_cookie *logcookie)
-{
- struct ll_fid fid = { .id = logcookie->lgc_lgl.lgl_oid,
- .generation = logcookie->lgc_lgl.lgl_ogen,
- .f_type = S_IFREG };
- struct llog_handle *loghandle;
+ struct mds_obd *mds = &exp->exp_obd->u.mds;
+ struct inode *parent_inode = mds->mds_objects_dir->d_inode;
+ struct file *filp;
struct dentry *dchild;
- int rc;
+ struct lvfs_run_ctxt saved;
+ char fidname[LL_FID_NAMELEN];
+ void *handle;
+ int rc = 0, err, namelen;
ENTRY;
- loghandle = llog_alloc_handle();
- if (loghandle == NULL)
- RETURN(ERR_PTR(-ENOMEM));
-
- down(&obd->u.mds.mds_logs_dir->d_inode->i_sem);
- dchild = mds_fid2dentry(&obd->u.mds, &fid, NULL);
- up(&obd->u.mds.mds_logs_dir->d_inode->i_sem);
- if (IS_ERR(dchild)) {
- rc = PTR_ERR(dchild);
- CERROR("error looking up log file "LPX64":%x: rc %d\n",
- fid.id, fid.generation, rc);
- GOTO(out, rc);
- }
-
- if (dchild->d_inode == NULL) {
- rc = -ENOENT;
- CERROR("nonexistent log file "LPX64":%x: rc %d\n",
- fid.id, fid.generation, rc);
- GOTO(out_put, rc);
- }
-
- /* dentry_open does a dput(de) and mntput(mds->mds_vfsmnt) on error */
- mntget(obd->u.mds.mds_vfsmnt);
- loghandle->lgh_file = dentry_open(dchild, obd->u.mds.mds_vfsmnt,
- O_RDWR | O_LARGEFILE);
- if (IS_ERR(loghandle->lgh_file)) {
- rc = PTR_ERR(loghandle->lgh_file);
- CERROR("error opening logfile "LPX64":%x: rc %d\n",
- fid.id, fid.generation, rc);
- GOTO(out, rc);
+ push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
+ down(&parent_inode->i_sem);
+ if (oa->o_id) {
+ namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
+
+ dchild = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
+ if (IS_ERR(dchild))
+ GOTO(out_pop, rc = PTR_ERR(dchild));
+
+ if (dchild->d_inode == NULL) {
+ struct dentry_params dp;
+ struct inode *inode;
+
+ CWARN("creating log with ID "LPU64"\n", oa->o_id);
+
+ dchild->d_fsdata = (void *) &dp;
+ dp.p_ptr = NULL;
+ dp.p_inum = oa->o_id;
+ rc = ll_vfs_create(parent_inode, dchild, S_IFREG, NULL);
+ if (dchild->d_fsdata == (void *)(unsigned long)oa->o_id)
+ dchild->d_fsdata = NULL;
+ if (rc) {
+ CDEBUG(D_INODE, "err during create: %d\n", rc);
+ dput(dchild);
+ GOTO(out_pop, rc);
+ }
+ inode = dchild->d_inode;
+ LASSERT(inode->i_ino == oa->o_id);
+ inode->i_generation = oa->o_generation;
+ CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
+ inode->i_ino, inode->i_generation);
+ mark_inode_dirty(inode);
+ } else {
+ CWARN("it should be here!\n");
+ }
+ GOTO(out_pop, rc);
}
- memcpy(&loghandle->lgh_cookie, logcookie, sizeof(*logcookie));
- loghandle->lgh_log_create = mds_log_create;
- loghandle->lgh_log_open = mds_log_open;
- loghandle->lgh_log_close = mds_log_close;
- loghandle->lgh_obd = obd;
-
- RETURN(loghandle);
-out_put:
- l_dput(dchild);
-out:
- llog_free_handle(loghandle);
- return ERR_PTR(rc);
-}
-
-/* This is a callback from the llog_* functions.
- * Assumes caller has already pushed us into the kernel context. */
-struct llog_handle *mds_log_create(struct obd_device *obd)
-{
- char logbuf[24], *logname; /* logSSSSSSSSSS.count */
- struct llog_handle *loghandle;
- int rc, open_flags = O_RDWR | O_CREAT | O_LARGEFILE;
- ENTRY;
-
- loghandle = llog_alloc_handle();
- if (!loghandle)
- RETURN(ERR_PTR(-ENOMEM));
-
-retry:
- if (!obd->u.mds.mds_catalog) {
- logname = "LOGS/catalog";
- } else {
- sprintf(logbuf, "LOGS/log%lu.%u\n",
- CURRENT_SECONDS, obd->u.mds.mds_catalog->lgh_index++);
- open_flags |= O_EXCL;
- logname = logbuf;
- }
- loghandle->lgh_file = filp_open(logname, open_flags, 0644);
- if (IS_ERR(loghandle->lgh_file)) {
- rc = PTR_ERR(loghandle->lgh_file);
+ sprintf(fidname, "OBJECTS/%u.%u",ll_insecure_random_int(),current->pid);
+ filp = filp_open(fidname, O_CREAT | O_EXCL, 0644);
+ if (IS_ERR(filp)) {
+ rc = PTR_ERR(filp);
if (rc == -EEXIST) {
- CDEBUG(D_HA, "collision in logfile %s creation\n",
- logname);
- obd->u.mds.mds_catalog->lgh_index++;
- goto retry;
+ CERROR("impossible object name collision %s\n",
+ fidname);
+ LBUG();
}
- CERROR("error opening/creating %s: rc %d\n", logname, rc);
- GOTO(out_handle, rc);
+ CERROR("error creating tmp object %s: rc %d\n", fidname, rc);
+ GOTO(out_pop, rc);
}
- loghandle->lgh_cookie.lgc_lgl.lgl_oid =
- loghandle->lgh_file->f_dentry->d_inode->i_ino;
- loghandle->lgh_cookie.lgc_lgl.lgl_ogen =
- loghandle->lgh_file->f_dentry->d_inode->i_generation;
- loghandle->lgh_log_create = mds_log_create;
- loghandle->lgh_log_open = mds_log_open;
- loghandle->lgh_log_close = mds_log_close;
- loghandle->lgh_obd = obd;
+ LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
- RETURN(loghandle);
+ oa->o_id = filp->f_dentry->d_inode->i_ino;
+ oa->o_generation = filp->f_dentry->d_inode->i_generation;
+ namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
+ CWARN("created log anonymous "LPU64"/%u\n", oa->o_id, oa->o_generation);
-out_handle:
- llog_free_handle(loghandle);
- return ERR_PTR(rc);
-}
+ dchild = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
-struct llog_handle *mds_get_catalog(struct obd_device *obd)
-{
- struct mds_server_data *msd = obd->u.mds.mds_server_data;
- struct obd_run_ctxt saved;
- struct llog_handle *cathandle = NULL;
- int rc = 0;
- ENTRY;
-
- push_ctxt(&saved, &obd->u.mds.mds_ctxt, NULL);
-
- if (msd->msd_catalog_oid) {
- struct llog_cookie catcookie;
-
- catcookie.lgc_lgl.lgl_oid = le64_to_cpu(msd->msd_catalog_oid);
- catcookie.lgc_lgl.lgl_ogen = le32_to_cpu(msd->msd_catalog_ogen);
- cathandle = mds_log_open(obd, &catcookie);
- if (IS_ERR(cathandle)) {
- CERROR("error opening catalog "LPX64":%x: rc %d\n",
- catcookie.lgc_lgl.lgl_oid,
- catcookie.lgc_lgl.lgl_ogen,
- (int)PTR_ERR(cathandle));
- msd->msd_catalog_oid = 0;
- msd->msd_catalog_ogen = 0;
- }
- /* ORPHANS FIXME: compare catalog UUID to msd_peeruuid */
+ if (IS_ERR(dchild)) {
+ CERROR("getting neg dentry for obj rename: %d\n", rc);
+ GOTO(out_close, rc = PTR_ERR(dchild));
}
-
- if (!msd->msd_catalog_oid) {
- struct llog_logid *lgl;
-
- cathandle = mds_log_create(obd);
- if (IS_ERR(cathandle)) {
- CERROR("error creating new catalog: rc %d\n",
- (int)PTR_ERR(cathandle));
- GOTO(out, cathandle);
- }
- lgl = &cathandle->lgh_cookie.lgc_lgl;
- msd->msd_catalog_oid = cpu_to_le64(lgl->lgl_oid);
- msd->msd_catalog_ogen = cpu_to_le32(lgl->lgl_ogen);
- rc = mds_update_server_data(obd);
- if (rc) {
- CERROR("error writing new catalog to disk: rc %d\n",rc);
- GOTO(out_handle, rc);
- }
+ if (dchild->d_inode != NULL) {
+ CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
+ oa->o_id, oa->o_generation);
+ LBUG();
}
- rc = llog_init_catalog(cathandle, &obd->u.mds.mds_osc_uuid);
-
-out:
- pop_ctxt(&saved, &obd->u.mds.mds_ctxt, NULL);
- RETURN(cathandle);
-
-out_handle:
- mds_log_close(cathandle, cathandle);
- cathandle = ERR_PTR(rc);
- goto out;
+ handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
+ FSFILT_OP_RENAME, NULL);
+ 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, dchild);
+ unlock_kernel();
+ if (rc)
+ 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_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:
+ dput(dchild);
+out_close:
+ err = filp_close(filp, 0);
+ if (err) {
+ CERROR("closing tmpfile %s: rc %d\n", fidname, rc);
+ if (!rc)
+ rc = err;
+ }
+out_pop:
+ up(&parent_inode->i_sem);
+ pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
+ RETURN(rc);
}
-void mds_put_catalog(struct llog_handle *cathandle)
+int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
+ struct lov_stripe_md *ea, struct obd_trans_info *oti)
{
- struct llog_handle *loghandle, *n;
- int 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 lvfs_run_ctxt saved;
+ char fidname[LL_FID_NAMELEN];
+ struct dentry *de;
+ void *handle;
+ int err, namelen, rc = 0;
ENTRY;
- list_for_each_entry_safe(loghandle, n, &cathandle->lgh_list, lgh_list)
- mds_log_close(cathandle, loghandle);
+ push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
- rc = filp_close(cathandle->lgh_file, 0);
- if (rc)
- CERROR("error closing catalog: rc %d\n", rc);
+ namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
- llog_free_handle(cathandle);
- EXIT;
+ down(&parent_inode->i_sem);
+ de = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
+ if (IS_ERR(de) || de->d_inode == NULL) {
+ rc = IS_ERR(de) ? PTR_ERR(de) : -ENOENT;
+ CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
+ oa->o_id, fidname, rc);
+ GOTO(out_dput, rc);
+ }
+ /* 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_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_lvfs_ctxt, NULL);
+ RETURN(rc);
}