fs/ocfs2/stackglue.c

   1 /* -*- mode: c; c-basic-offset: 8; -*-
   2  * vim: noexpandtab sw=8 ts=8 sts=0:
   3  *
   4  * stackglue.c
   5  *
   6  * Code which implements an OCFS2 specific interface to underlying
   7  * cluster stacks.
   8  *
   9  * Copyright (C) 2007 Oracle.  All rights reserved.
  10  *
  11  * This program is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU General Public
  13  * License as published by the Free Software Foundation, version 2.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * General Public License for more details.
  19  */
  20
  21 #include <linux/slab.h>
  22 #include <linux/crc32.h>
  23 #include <linux/kmod.h>
  24
  25 /* Needed for AOP_TRUNCATED_PAGE in mlog_errno() */
  26 #include <linux/fs.h>
  27
  28 #include "cluster/masklog.h"
  29 #include "cluster/nodemanager.h"
  30 #include "cluster/heartbeat.h"
  31
  32 #include "stackglue.h"
  33
  34 static struct ocfs2_locking_protocol *lproto;
  35
  36 struct o2dlm_private {
  37         struct dlm_eviction_cb op_eviction_cb;
  38 };
  39
  40 /* These should be identical */
  41 #if (DLM_LOCK_IV != LKM_IVMODE)
  42 # error Lock modes do not match
  43 #endif
  44 #if (DLM_LOCK_NL != LKM_NLMODE)
  45 # error Lock modes do not match
  46 #endif
  47 #if (DLM_LOCK_CR != LKM_CRMODE)
  48 # error Lock modes do not match
  49 #endif
  50 #if (DLM_LOCK_CW != LKM_CWMODE)
  51 # error Lock modes do not match
  52 #endif
  53 #if (DLM_LOCK_PR != LKM_PRMODE)
  54 # error Lock modes do not match
  55 #endif
  56 #if (DLM_LOCK_PW != LKM_PWMODE)
  57 # error Lock modes do not match
  58 #endif
  59 #if (DLM_LOCK_EX != LKM_EXMODE)
  60 # error Lock modes do not match
  61 #endif
  62 static inline int mode_to_o2dlm(int mode)
  63 {
  64         BUG_ON(mode > LKM_MAXMODE);
  65
  66         return mode;
  67 }
  68
  69 #define map_flag(_generic, _o2dlm)              \
  70         if (flags & (_generic)) {               \
  71                 flags &= ~(_generic);           \
  72                 o2dlm_flags |= (_o2dlm);        \
  73         }
  74 static int flags_to_o2dlm(u32 flags)
  75 {
  76         int o2dlm_flags = 0;
  77
  78         map_flag(DLM_LKF_NOQUEUE, LKM_NOQUEUE);
  79         map_flag(DLM_LKF_CANCEL, LKM_CANCEL);
  80         map_flag(DLM_LKF_CONVERT, LKM_CONVERT);
  81         map_flag(DLM_LKF_VALBLK, LKM_VALBLK);
  82         map_flag(DLM_LKF_IVVALBLK, LKM_INVVALBLK);
  83         map_flag(DLM_LKF_ORPHAN, LKM_ORPHAN);
  84         map_flag(DLM_LKF_FORCEUNLOCK, LKM_FORCE);
  85         map_flag(DLM_LKF_TIMEOUT, LKM_TIMEOUT);
  86         map_flag(DLM_LKF_LOCAL, LKM_LOCAL);
  87
  88         /* map_flag() should have cleared every flag passed in */
  89         BUG_ON(flags != 0);
  90
  91         return o2dlm_flags;
  92 }
  93 #undef map_flag
  94
  95 /*
  96  * Map an o2dlm status to standard errno values.
  97  *
  98  * o2dlm only uses a handful of these, and returns even fewer to the
  99  * caller. Still, we try to assign sane values to each error.
 100  *
 101  * The following value pairs have special meanings to dlmglue, thus
 102  * the right hand side needs to stay unique - never duplicate the
 103  * mapping elsewhere in the table!
 104  *
 105  * DLM_NORMAL:          0
 106  * DLM_NOTQUEUED:       -EAGAIN
 107  * DLM_CANCELGRANT:     -EBUSY
 108  * DLM_CANCEL:          -DLM_ECANCEL
 109  */
 110 /* Keep in sync with dlmapi.h */
 111 static int status_map[] = {
 112         [DLM_NORMAL]                    = 0,            /* Success */
 113         [DLM_GRANTED]                   = -EINVAL,
 114         [DLM_DENIED]                    = -EACCES,
 115         [DLM_DENIED_NOLOCKS]            = -EACCES,
 116         [DLM_WORKING]                   = -EACCES,
 117         [DLM_BLOCKED]                   = -EINVAL,
 118         [DLM_BLOCKED_ORPHAN]            = -EINVAL,
 119         [DLM_DENIED_GRACE_PERIOD]       = -EACCES,
 120         [DLM_SYSERR]                    = -ENOMEM,      /* It is what it is */
 121         [DLM_NOSUPPORT]                 = -EPROTO,
 122         [DLM_CANCELGRANT]               = -EBUSY,       /* Cancel after grant */
 123         [DLM_IVLOCKID]                  = -EINVAL,
 124         [DLM_SYNC]                      = -EINVAL,
 125         [DLM_BADTYPE]                   = -EINVAL,
 126         [DLM_BADRESOURCE]               = -EINVAL,
 127         [DLM_MAXHANDLES]                = -ENOMEM,
 128         [DLM_NOCLINFO]                  = -EINVAL,
 129         [DLM_NOLOCKMGR]                 = -EINVAL,
 130         [DLM_NOPURGED]                  = -EINVAL,
 131         [DLM_BADARGS]                   = -EINVAL,
 132         [DLM_VOID]                      = -EINVAL,
 133         [DLM_NOTQUEUED]                 = -EAGAIN,      /* Trylock failed */
 134         [DLM_IVBUFLEN]                  = -EINVAL,
 135         [DLM_CVTUNGRANT]                = -EPERM,
 136         [DLM_BADPARAM]                  = -EINVAL,
 137         [DLM_VALNOTVALID]               = -EINVAL,
 138         [DLM_REJECTED]                  = -EPERM,
 139         [DLM_ABORT]                     = -EINVAL,
 140         [DLM_CANCEL]                    = -DLM_ECANCEL, /* Successful cancel */
 141         [DLM_IVRESHANDLE]               = -EINVAL,
 142         [DLM_DEADLOCK]                  = -EDEADLK,
 143         [DLM_DENIED_NOASTS]             = -EINVAL,
 144         [DLM_FORWARD]                   = -EINVAL,
 145         [DLM_TIMEOUT]                   = -ETIMEDOUT,
 146         [DLM_IVGROUPID]                 = -EINVAL,
 147         [DLM_VERS_CONFLICT]             = -EOPNOTSUPP,
 148         [DLM_BAD_DEVICE_PATH]           = -ENOENT,
 149         [DLM_NO_DEVICE_PERMISSION]      = -EPERM,
 150         [DLM_NO_CONTROL_DEVICE]         = -ENOENT,
 151         [DLM_RECOVERING]                = -ENOTCONN,
 152         [DLM_MIGRATING]                 = -ERESTART,
 153         [DLM_MAXSTATS]                  = -EINVAL,
 154 };
 155
 156 static int dlm_status_to_errno(enum dlm_status status)
 157 {
 158         BUG_ON(status > (sizeof(status_map) / sizeof(status_map[0])));
 159
 160         return status_map[status];
 161 }
 162
 163 static void o2dlm_lock_ast_wrapper(void *astarg)
 164 {
 165         BUG_ON(lproto == NULL);
 166
 167         lproto->lp_lock_ast(astarg);
 168 }
 169
 170 static void o2dlm_blocking_ast_wrapper(void *astarg, int level)
 171 {
 172         BUG_ON(lproto == NULL);
 173
 174         lproto->lp_blocking_ast(astarg, level);
 175 }
 176
 177 static void o2dlm_unlock_ast_wrapper(void *astarg, enum dlm_status status)
 178 {
 179         int error = dlm_status_to_errno(status);
 180
 181         BUG_ON(lproto == NULL);
 182
 183         /*
 184          * In o2dlm, you can get both the lock_ast() for the lock being
 185          * granted and the unlock_ast() for the CANCEL failing.  A
 186          * successful cancel sends DLM_NORMAL here.  If the
 187          * lock grant happened before the cancel arrived, you get
 188          * DLM_CANCELGRANT.
 189          *
 190          * There's no need for the double-ast.  If we see DLM_CANCELGRANT,
 191          * we just ignore it.  We expect the lock_ast() to handle the
 192          * granted lock.
 193          */
 194         if (status == DLM_CANCELGRANT)
 195                 return;
 196
 197         lproto->lp_unlock_ast(astarg, error);
 198 }
 199
 200 int ocfs2_dlm_lock(struct ocfs2_cluster_connection *conn,
 201                    int mode,
 202                    union ocfs2_dlm_lksb *lksb,
 203                    u32 flags,
 204                    void *name,
 205                    unsigned int namelen,
 206                    void *astarg)
 207 {
 208         enum dlm_status status;
 209         int o2dlm_mode = mode_to_o2dlm(mode);
 210         int o2dlm_flags = flags_to_o2dlm(flags);
 211         int ret;
 212
 213         BUG_ON(lproto == NULL);
 214
 215         status = dlmlock(conn->cc_lockspace, o2dlm_mode, &lksb->lksb_o2dlm,
 216                          o2dlm_flags, name, namelen,
 217                          o2dlm_lock_ast_wrapper, astarg,
 218                          o2dlm_blocking_ast_wrapper);
 219         ret = dlm_status_to_errno(status);
 220         return ret;
 221 }
 222
 223 int ocfs2_dlm_unlock(struct ocfs2_cluster_connection *conn,
 224                      union ocfs2_dlm_lksb *lksb,
 225                      u32 flags,
 226                      void *astarg)
 227 {
 228         enum dlm_status status;
 229         int o2dlm_flags = flags_to_o2dlm(flags);
 230         int ret;
 231
 232         BUG_ON(lproto == NULL);
 233
 234         status = dlmunlock(conn->cc_lockspace, &lksb->lksb_o2dlm,
 235                            o2dlm_flags, o2dlm_unlock_ast_wrapper, astarg);
 236         ret = dlm_status_to_errno(status);
 237         return ret;
 238 }
 239
 240 int ocfs2_dlm_lock_status(union ocfs2_dlm_lksb *lksb)
 241 {
 242         return dlm_status_to_errno(lksb->lksb_o2dlm.status);
 243 }
 244
 245 /*
 246  * Why don't we cast to ocfs2_meta_lvb?  The "clean" answer is that we
 247  * don't cast at the glue level.  The real answer is that the header
 248  * ordering is nigh impossible.
 249  */
 250 void *ocfs2_dlm_lvb(union ocfs2_dlm_lksb *lksb)
 251 {
 252         return (void *)(lksb->lksb_o2dlm.lvb);
 253 }
 254
 255 /*
 256  * Called from the dlm when it's about to evict a node. This is how the
 257  * classic stack signals node death.
 258  */
 259 static void o2dlm_eviction_cb(int node_num, void *data)
 260 {
 261         struct ocfs2_cluster_connection *conn = data;
 262
 263         mlog(ML_NOTICE, "o2dlm has evicted node %d from group %.*s\n",
 264              node_num, conn->cc_namelen, conn->cc_name);
 265
 266         conn->cc_recovery_handler(node_num, conn->cc_recovery_data);
 267 }
 268
 269 int ocfs2_cluster_connect(const char *group,
 270                           int grouplen,
 271                           void (*recovery_handler)(int node_num,
 272                                                    void *recovery_data),
 273                           void *recovery_data,
 274                           struct ocfs2_cluster_connection **conn)
 275 {
 276         int rc = 0;
 277         struct ocfs2_cluster_connection *new_conn;
 278         u32 dlm_key;
 279         struct dlm_ctxt *dlm;
 280         struct o2dlm_private *priv;
 281         struct dlm_protocol_version dlm_version;
 282
 283         BUG_ON(group == NULL);
 284         BUG_ON(conn == NULL);
 285         BUG_ON(recovery_handler == NULL);
 286
 287         if (grouplen > GROUP_NAME_MAX) {
 288                 rc = -EINVAL;
 289                 goto out;
 290         }
 291
 292         /* for now we only have one cluster/node, make sure we see it
 293          * in the heartbeat universe */
 294         if (!o2hb_check_local_node_heartbeating()) {
 295                 rc = -EINVAL;
 296                 goto out;
 297         }
 298
 299         new_conn = kzalloc(sizeof(struct ocfs2_cluster_connection),
 300                            GFP_KERNEL);
 301         if (!new_conn) {
 302                 rc = -ENOMEM;
 303                 goto out;
 304         }
 305
 306         memcpy(new_conn->cc_name, group, grouplen);
 307         new_conn->cc_namelen = grouplen;
 308         new_conn->cc_recovery_handler = recovery_handler;
 309         new_conn->cc_recovery_data = recovery_data;
 310
 311         /* Start the new connection at our maximum compatibility level */
 312         new_conn->cc_version = lproto->lp_max_version;
 313
 314         priv = kzalloc(sizeof(struct o2dlm_private), GFP_KERNEL);
 315         if (!priv) {
 316                 rc = -ENOMEM;
 317                 goto out_free;
 318         }
 319
 320         /* This just fills the structure in.  It is safe to use new_conn. */
 321         dlm_setup_eviction_cb(&priv->op_eviction_cb, o2dlm_eviction_cb,
 322                               new_conn);
 323
 324         new_conn->cc_private = priv;
 325
 326         /* used by the dlm code to make message headers unique, each
 327          * node in this domain must agree on this. */
 328         dlm_key = crc32_le(0, group, grouplen);
 329         dlm_version.pv_major = new_conn->cc_version.pv_major;
 330         dlm_version.pv_minor = new_conn->cc_version.pv_minor;
 331
 332         dlm = dlm_register_domain(group, dlm_key, &dlm_version);
 333         if (IS_ERR(dlm)) {
 334                 rc = PTR_ERR(dlm);
 335                 mlog_errno(rc);
 336                 goto out_free;
 337         }
 338
 339         new_conn->cc_version.pv_major = dlm_version.pv_major;
 340         new_conn->cc_version.pv_minor = dlm_version.pv_minor;
 341         new_conn->cc_lockspace = dlm;
 342
 343         dlm_register_eviction_cb(dlm, &priv->op_eviction_cb);
 344
 345         *conn = new_conn;
 346
 347 out_free:
 348         if (rc) {
 349                 kfree(new_conn->cc_private);
 350                 kfree(new_conn);
 351         }
 352
 353 out:
 354         return rc;
 355 }
 356
 357
 358 int ocfs2_cluster_disconnect(struct ocfs2_cluster_connection *conn)
 359 {
 360         struct dlm_ctxt *dlm = conn->cc_lockspace;
 361         struct o2dlm_private *priv = conn->cc_private;
 362
 363         dlm_unregister_eviction_cb(&priv->op_eviction_cb);
 364         dlm_unregister_domain(dlm);
 365
 366         kfree(priv);
 367         kfree(conn);
 368
 369         return 0;
 370 }
 371
 372 static void o2hb_stop(const char *group)
 373 {
 374         int ret;
 375         char *argv[5], *envp[3];
 376
 377         argv[0] = (char *)o2nm_get_hb_ctl_path();
 378         argv[1] = "-K";
 379         argv[2] = "-u";
 380         argv[3] = (char *)group;
 381         argv[4] = NULL;
 382
 383         mlog(0, "Run: %s %s %s %s\n", argv[0], argv[1], argv[2], argv[3]);
 384
 385         /* minimal command environment taken from cpu_run_sbin_hotplug */
 386         envp[0] = "HOME=/";
 387         envp[1] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
 388         envp[2] = NULL;
 389
 390         ret = call_usermodehelper(argv[0], argv, envp, UMH_WAIT_PROC);
 391         if (ret < 0)
 392                 mlog_errno(ret);
 393 }
 394
 395 /*
 396  * Hangup is a hack for tools compatibility.  Older ocfs2-tools software
 397  * expects the filesystem to call "ocfs2_hb_ctl" during unmount.  This
 398  * happens regardless of whether the DLM got started, so we can't do it
 399  * in ocfs2_cluster_disconnect().  We bring the o2hb_stop() function into
 400  * the glue and provide a "hangup" API for super.c to call.
 401  *
 402  * Other stacks will eventually provide a NULL ->hangup() pointer.
 403  */
 404 void ocfs2_cluster_hangup(const char *group, int grouplen)
 405 {
 406         BUG_ON(group == NULL);
 407         BUG_ON(group[grouplen] != '\0');
 408
 409         o2hb_stop(group);
 410 }
 411
 412 int ocfs2_cluster_this_node(unsigned int *node)
 413 {
 414         int node_num;
 415
 416         node_num = o2nm_this_node();
 417         if (node_num == O2NM_INVALID_NODE_NUM)
 418                 return -ENOENT;
 419
 420         if (node_num >= O2NM_MAX_NODES)
 421                 return -EOVERFLOW;
 422
 423         *node = node_num;
 424         return 0;
 425 }
 426
 427 void o2cb_get_stack(struct ocfs2_locking_protocol *proto)
 428 {
 429         BUG_ON(proto == NULL);
 430
 431         lproto = proto;
 432 }
 433
 434 void o2cb_put_stack(void)
 435 {
 436         lproto = NULL;
 437 }