lustre/mds/mds_fs.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/mds/mds_fs.c
   5  *
   6  *  Lustre Metadata Server (MDS) filesystem interface code
   7  *
   8  *  Copyright (C) 2002 Cluster File Systems, Inc.
   9  *
  10  *  This code is issued under the GNU General Public License.
  11  *  See the file COPYING in this distribution
  12  *
  13  *  by Andreas Dilger <adilger@clusterfs.com>
  14  *
  15  */
  16
  17 #define EXPORT_SYMTAB
  18 #define DEBUG_SUBSYSTEM S_MDS
  19
  20 #include <linux/module.h>
  21 #include <linux/kmod.h>
  22 #include <linux/lustre_mds.h>
  23 #include <linux/obd_class.h>
  24 #include <linux/obd_support.h>
  25 #include <linux/lustre_lib.h>
  26 #include <linux/lustre_fsfilt.h>
  27
  28 /* This limit is arbitrary, but for now we fit it in 1 page (32k clients) */
  29 #define MDS_MAX_CLIENTS (PAGE_SIZE * 8)
  30 #define MDS_MAX_CLIENT_WORDS (MDS_MAX_CLIENTS / sizeof(unsigned long))
  31
  32 static unsigned long last_rcvd_slots[MDS_MAX_CLIENT_WORDS];
  33
  34 #define LAST_RCVD "last_rcvd"
  35
  36 /* Add client data to the MDS.  We use a bitmap to locate a free space
  37  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
  38  * Otherwise, we have just read the data from the last_rcvd file and
  39  * we know its offset.
  40  */
  41 int mds_client_add(struct mds_obd *mds, struct mds_export_data *med, int cl_off)
  42 {
  43         int new_client = (cl_off == -1);
  44
  45         /* the bitmap operations can handle cl_off > sizeof(long) * 8, so
  46          * there's no need for extra complication here
  47          */
  48         if (new_client) {
  49                 cl_off = find_first_zero_bit(last_rcvd_slots, MDS_MAX_CLIENTS);
  50         repeat:
  51                 if (cl_off >= MDS_MAX_CLIENTS) {
  52                         CERROR("no room for clients - fix MDS_MAX_CLIENTS\n");
  53                         return -ENOMEM;
  54                 }
  55                 if (test_and_set_bit(cl_off, last_rcvd_slots)) {
  56                         CERROR("MDS client %d: found bit is set in bitmap\n",
  57                                cl_off);
  58                         cl_off = find_next_zero_bit(last_rcvd_slots,
  59                                                     MDS_MAX_CLIENTS, cl_off);
  60                         goto repeat;
  61                 }
  62         } else {
  63                 if (test_and_set_bit(cl_off, last_rcvd_slots)) {
  64                         CERROR("MDS client %d: bit already set in bitmap!!\n",
  65                                cl_off);
  66                         LBUG();
  67                 }
  68         }
  69
  70         CDEBUG(D_INFO, "client at offset %d with UUID '%s' added\n",
  71                cl_off, med->med_mcd->mcd_uuid);
  72
  73         med->med_off = cl_off;
  74
  75         if (new_client) {
  76                 struct obd_run_ctxt saved;
  77                 loff_t off = MDS_LR_CLIENT + (cl_off * MDS_LR_SIZE);
  78                 ssize_t written;
  79
  80                 push_ctxt(&saved, &mds->mds_ctxt, NULL);
  81                 written = lustre_fwrite(mds->mds_rcvd_filp,
  82                                                 (char *)med->med_mcd,
  83                                                 sizeof(*med->med_mcd), &off);
  84                 pop_ctxt(&saved, &mds->mds_ctxt, NULL);
  85
  86                 if (written != sizeof(*med->med_mcd)) {
  87                         if (written < 0)
  88                                 RETURN(written);
  89                         RETURN(-EIO);
  90                 }
  91                 CDEBUG(D_INFO, "wrote client mcd at off %u (len %u)\n",
  92                        MDS_LR_CLIENT + (cl_off * MDS_LR_SIZE),
  93                        (unsigned int)sizeof(*med->med_mcd));
  94         }
  95         return 0;
  96 }
  97
  98 int mds_client_free(struct obd_export *exp)
  99 {
 100         struct mds_export_data *med = &exp->exp_mds_data;
 101         struct mds_obd *mds = &exp->exp_obd->u.mds;
 102         struct mds_client_data zero_mcd;
 103         struct obd_run_ctxt saved;
 104         int written;
 105         loff_t off;
 106
 107         if (!med->med_mcd)
 108                 RETURN(0);
 109
 110         off = MDS_LR_CLIENT + (med->med_off * MDS_LR_SIZE);
 111
 112         CDEBUG(D_INFO, "freeing client at offset %u (%lld)with UUID '%s'\n",
 113                med->med_off, off, med->med_mcd->mcd_uuid);
 114
 115         if (!test_and_clear_bit(med->med_off, last_rcvd_slots)) {
 116                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 117                        med->med_off);
 118                 LBUG();
 119         }
 120
 121         memset(&zero_mcd, 0, sizeof zero_mcd);
 122         push_ctxt(&saved, &mds->mds_ctxt, NULL);
 123         written = lustre_fwrite(mds->mds_rcvd_filp, (const char *)&zero_mcd,
 124                                 sizeof(zero_mcd), &off);
 125         pop_ctxt(&saved, &mds->mds_ctxt, NULL);
 126
 127         if (written != sizeof(zero_mcd)) {
 128                 CERROR("error zeroing out client %s off %d in %s: %d\n",
 129                        med->med_mcd->mcd_uuid, med->med_off, LAST_RCVD,
 130                        written);
 131         } else {
 132                 CDEBUG(D_INFO, "zeroed out disconnecting client %s at off %d\n",
 133                        med->med_mcd->mcd_uuid, med->med_off);
 134         }
 135
 136         OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
 137
 138         return 0;
 139 }
 140
 141 static int mds_server_free_data(struct mds_obd *mds)
 142 {
 143         OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
 144         mds->mds_server_data = NULL;
 145
 146         return 0;
 147 }
 148
 149 static int mds_read_last_rcvd(struct obd_device *obddev, struct file *f)
 150 {
 151         struct mds_obd *mds = &obddev->u.mds;
 152         struct mds_server_data *msd;
 153         struct mds_client_data *mcd = NULL;
 154         loff_t off = 0;
 155         int cl_off;
 156         unsigned long last_rcvd_size = f->f_dentry->d_inode->i_size;
 157         __u64 last_rcvd = 0;
 158         __u64 last_mount;
 159         int rc = 0;
 160
 161         OBD_ALLOC(msd, sizeof(*msd));
 162         if (!msd)
 163                 RETURN(-ENOMEM);
 164         rc = lustre_fread(f, (char *)msd, sizeof(*msd), &off);
 165
 166         mds->mds_server_data = msd;
 167         if (rc == 0) {
 168                 CERROR("empty MDS %s, new MDS?\n", LAST_RCVD);
 169                 RETURN(0);
 170         }
 171
 172         if (rc != sizeof(*msd)) {
 173                 CERROR("error reading MDS %s: rc = %d\n", LAST_RCVD, rc);
 174                 if (rc > 0)
 175                         rc = -EIO;
 176                 GOTO(err_msd, rc);
 177         }
 178
 179         CDEBUG(D_INODE, "last_rcvd has size %lu (msd + %lu clients)\n",
 180                last_rcvd_size, (last_rcvd_size - sizeof *msd) / sizeof *mcd);
 181
 182         /*
 183          * When we do a clean MDS shutdown, we save the last_rcvd into
 184          * the header.  If we find clients with higher last_rcvd values
 185          * then those clients may need recovery done.
 186          */
 187         last_rcvd = le64_to_cpu(msd->msd_last_rcvd);
 188         mds->mds_last_rcvd = last_rcvd;
 189         CDEBUG(D_INODE, "got "LPU64" for server last_rcvd value\n", last_rcvd);
 190
 191         last_mount = le64_to_cpu(msd->msd_mount_count);
 192         mds->mds_mount_count = last_mount;
 193         CDEBUG(D_INODE, "got "LPU64" for server last_mount value\n",last_mount);
 194
 195         /* off is adjusted by lustre_fread, so we don't adjust it in the loop */
 196         for (off = MDS_LR_CLIENT, cl_off = 0; off < last_rcvd_size; cl_off++) {
 197                 int mount_age;
 198
 199                 if (!mcd) {
 200                         OBD_ALLOC(mcd, sizeof(*mcd));
 201                         if (!mcd)
 202                                 GOTO(err_msd, rc = -ENOMEM);
 203                 }
 204
 205                 rc = lustre_fread(f, (char *)mcd, sizeof(*mcd), &off);
 206                 if (rc != sizeof(*mcd)) {
 207                         CERROR("error reading MDS %s offset %d: rc = %d\n",
 208                                LAST_RCVD, cl_off, rc);
 209                         if (rc > 0) /* XXX fatal error or just abort reading? */
 210                                 rc = -EIO;
 211                         break;
 212                 }
 213
 214                 if (mcd->mcd_uuid[0] == '\0') {
 215                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 216                                cl_off);
 217                         continue;
 218                 }
 219
 220                 last_rcvd = le64_to_cpu(mcd->mcd_last_rcvd);
 221
 222                 /* These exports are cleaned up by mds_disconnect(), so they
 223                  * need to be set up like real exports as mds_connect() does.
 224                  */
 225                 mount_age = last_mount - le64_to_cpu(mcd->mcd_mount_count);
 226                 if (mount_age < MDS_MOUNT_RECOV) {
 227                         struct obd_export *exp = class_new_export(obddev);
 228                         struct mds_export_data *med;
 229
 230                         if (!exp) {
 231                                 rc = -ENOMEM;
 232                                 break;
 233                         }
 234
 235                         med = &exp->exp_mds_data;
 236                         med->med_mcd = mcd;
 237                         mds_client_add(mds, med, cl_off);
 238                         /* create helper if export init gets more complex */
 239                         INIT_LIST_HEAD(&med->med_open_head);
 240                         spin_lock_init(&med->med_open_lock);
 241
 242                         mcd = NULL;
 243                         mds->mds_recoverable_clients++;
 244                 } else {
 245                         CDEBUG(D_INFO,
 246                                "discarded client %d, UUID '%s', count %Ld\n",
 247                                cl_off, mcd->mcd_uuid,
 248                                (long long)le64_to_cpu(mcd->mcd_mount_count));
 249                 }
 250
 251                 CDEBUG(D_OTHER, "client at offset %d has last_rcvd = %Lu\n",
 252                        cl_off, (unsigned long long)last_rcvd);
 253
 254                 if (last_rcvd > mds->mds_last_rcvd)
 255                         mds->mds_last_rcvd = last_rcvd;
 256         }
 257
 258         mds->mds_last_committed = mds->mds_last_rcvd;
 259         if (mds->mds_recoverable_clients) {
 260                 CERROR("RECOVERY: %d recoverable clients, last_rcvd "LPU64"\n",
 261                        mds->mds_recoverable_clients, mds->mds_last_rcvd);
 262                 mds->mds_next_recovery_transno = mds->mds_last_committed + 1;
 263         }
 264
 265         if (mcd)
 266                 OBD_FREE(mcd, sizeof(*mcd));
 267
 268         return 0;
 269
 270 err_msd:
 271         mds_server_free_data(mds);
 272         return rc;
 273 }
 274
 275 static int mds_fs_prep(struct obd_device *obddev)
 276 {
 277         struct mds_obd *mds = &obddev->u.mds;
 278         struct obd_run_ctxt saved;
 279         struct dentry *dentry;
 280         struct file *f;
 281         int rc;
 282
 283         push_ctxt(&saved, &mds->mds_ctxt, NULL);
 284         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755);
 285         if (IS_ERR(dentry)) {
 286                 rc = PTR_ERR(dentry);
 287                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 288                 GOTO(err_pop, rc);
 289         }
 290
 291         mds->mds_rootfid.id = dentry->d_inode->i_ino;
 292         mds->mds_rootfid.generation = dentry->d_inode->i_generation;
 293         mds->mds_rootfid.f_type = S_IFDIR;
 294
 295         dput(dentry);
 296
 297         dentry = simple_mkdir(current->fs->pwd, "FH", 0700);
 298         if (IS_ERR(dentry)) {
 299                 rc = PTR_ERR(dentry);
 300                 CERROR("cannot create FH directory: rc = %d\n", rc);
 301                 GOTO(err_pop, rc);
 302         }
 303         /* XXX probably want to hold on to this later... */
 304         dput(dentry);
 305
 306         f = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 307         if (IS_ERR(f)) {
 308                 rc = PTR_ERR(f);
 309                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 310                 GOTO(err_pop, rc = PTR_ERR(f));
 311         }
 312         if (!S_ISREG(f->f_dentry->d_inode->i_mode)) {
 313                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 314                        f->f_dentry->d_inode->i_mode);
 315                 GOTO(err_filp, rc = -ENOENT);
 316         }
 317
 318         rc = fsfilt_journal_data(obddev, f);
 319         if (rc) {
 320                 CERROR("cannot journal data on %s: rc = %d\n", LAST_RCVD, rc);
 321                 GOTO(err_filp, rc);
 322         }
 323
 324         rc = mds_read_last_rcvd(obddev, f);
 325         if (rc) {
 326                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 327                 GOTO(err_client, rc);
 328         }
 329         mds->mds_rcvd_filp = f;
 330 err_pop:
 331         pop_ctxt(&saved, &mds->mds_ctxt, NULL);
 332
 333         return rc;
 334
 335 err_client:
 336         class_disconnect_all(obddev);
 337 err_filp:
 338         if (filp_close(f, 0))
 339                 CERROR("can't close %s after error\n", LAST_RCVD);
 340         goto err_pop;
 341 }
 342
 343 int mds_fs_setup(struct obd_device *obddev, struct vfsmount *mnt)
 344 {
 345         struct mds_obd *mds = &obddev->u.mds;
 346         ENTRY;
 347
 348         mds->mds_vfsmnt = mnt;
 349
 350         OBD_SET_CTXT_MAGIC(&mds->mds_ctxt);
 351         mds->mds_ctxt.pwdmnt = mnt;
 352         mds->mds_ctxt.pwd = mnt->mnt_root;
 353         mds->mds_ctxt.fs = get_ds();
 354
 355         RETURN(mds_fs_prep(obddev));
 356 }
 357
 358 int mds_fs_cleanup(struct obd_device *obddev)
 359 {
 360         struct mds_obd *mds = &obddev->u.mds;
 361         struct obd_run_ctxt saved;
 362         int rc = 0;
 363
 364         class_disconnect_all(obddev); /* this cleans up client info too */
 365         mds_server_free_data(mds);
 366
 367         push_ctxt(&saved, &mds->mds_ctxt, NULL);
 368         if (mds->mds_rcvd_filp) {
 369                 rc = filp_close(mds->mds_rcvd_filp, 0);
 370                 mds->mds_rcvd_filp = NULL;
 371
 372                 if (rc)
 373                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
 374         }
 375         pop_ctxt(&saved, &mds->mds_ctxt, NULL);
 376
 377         return rc;
 378 }