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
  24 /* Needed for AOP_TRUNCATED_PAGE in mlog_errno() */
  25 #include <linux/fs.h>
  26
  27 #include "cluster/masklog.h"
  28 #include "cluster/nodemanager.h"
  29
  30 #include "stackglue.h"
  31
  32 static struct ocfs2_locking_protocol *lproto;
  33
  34 struct o2dlm_private {
  35         struct dlm_eviction_cb op_eviction_cb;
  36 };
  37
  38 /* These should be identical */
  39 #if (DLM_LOCK_IV != LKM_IVMODE)
  40 # error Lock modes do not match
  41 #endif
  42 #if (DLM_LOCK_NL != LKM_NLMODE)
  43 # error Lock modes do not match
  44 #endif
  45 #if (DLM_LOCK_CR != LKM_CRMODE)
  46 # error Lock modes do not match
  47 #endif
  48 #if (DLM_LOCK_CW != LKM_CWMODE)
  49 # error Lock modes do not match
  50 #endif
  51 #if (DLM_LOCK_PR != LKM_PRMODE)
  52 # error Lock modes do not match
  53 #endif
  54 #if (DLM_LOCK_PW != LKM_PWMODE)
  55 # error Lock modes do not match
  56 #endif
  57 #if (DLM_LOCK_EX != LKM_EXMODE)
  58 # error Lock modes do not match
  59 #endif
  60 static inline int mode_to_o2dlm(int mode)
  61 {
  62         BUG_ON(mode > LKM_MAXMODE);
  63
  64         return mode;
  65 }
  66
  67 #define map_flag(_generic, _o2dlm)              \
  68         if (flags & (_generic)) {               \
  69                 flags &= ~(_generic);           \
  70                 o2dlm_flags |= (_o2dlm);        \
  71         }
  72 static int flags_to_o2dlm(u32 flags)
  73 {
  74         int o2dlm_flags = 0;
  75
  76         map_flag(DLM_LKF_NOQUEUE, LKM_NOQUEUE);
  77         map_flag(DLM_LKF_CANCEL, LKM_CANCEL);
  78         map_flag(DLM_LKF_CONVERT, LKM_CONVERT);
  79         map_flag(DLM_LKF_VALBLK, LKM_VALBLK);
  80         map_flag(DLM_LKF_IVVALBLK, LKM_INVVALBLK);
  81         map_flag(DLM_LKF_ORPHAN, LKM_ORPHAN);
  82         map_flag(DLM_LKF_FORCEUNLOCK, LKM_FORCE);
  83         map_flag(DLM_LKF_TIMEOUT, LKM_TIMEOUT);
  84         map_flag(DLM_LKF_LOCAL, LKM_LOCAL);
  85
  86         /* map_flag() should have cleared every flag passed in */
  87         BUG_ON(flags != 0);
  88
  89         return o2dlm_flags;
  90 }
  91 #undef map_flag
  92
  93 /*
  94  * Map an o2dlm status to standard errno values.
  95  *
  96  * o2dlm only uses a handful of these, and returns even fewer to the
  97  * caller. Still, we try to assign sane values to each error.
  98  *
  99  * The following value pairs have special meanings to dlmglue, thus
 100  * the right hand side needs to stay unique - never duplicate the
 101  * mapping elsewhere in the table!
 102  *
 103  * DLM_NORMAL:          0
 104  * DLM_NOTQUEUED:       -EAGAIN
 105  * DLM_CANCELGRANT:     -DLM_ECANCEL
 106  * DLM_CANCEL:          -DLM_EUNLOCK
 107  */
 108 /* Keep in sync with dlmapi.h */
 109 static int status_map[] = {
 110         [DLM_NORMAL]                    = 0,            /* Success */
 111         [DLM_GRANTED]                   = -EINVAL,
 112         [DLM_DENIED]                    = -EACCES,
 113         [DLM_DENIED_NOLOCKS]            = -EACCES,
 114         [DLM_WORKING]                   = -EBUSY,
 115         [DLM_BLOCKED]                   = -EINVAL,
 116         [DLM_BLOCKED_ORPHAN]            = -EINVAL,
 117         [DLM_DENIED_GRACE_PERIOD]       = -EACCES,
 118         [DLM_SYSERR]                    = -ENOMEM,      /* It is what it is */
 119         [DLM_NOSUPPORT]                 = -EPROTO,
 120         [DLM_CANCELGRANT]               = -DLM_ECANCEL, /* Cancel after grant */
 121         [DLM_IVLOCKID]                  = -EINVAL,
 122         [DLM_SYNC]                      = -EINVAL,
 123         [DLM_BADTYPE]                   = -EINVAL,
 124         [DLM_BADRESOURCE]               = -EINVAL,
 125         [DLM_MAXHANDLES]                = -ENOMEM,
 126         [DLM_NOCLINFO]                  = -EINVAL,
 127         [DLM_NOLOCKMGR]                 = -EINVAL,
 128         [DLM_NOPURGED]                  = -EINVAL,
 129         [DLM_BADARGS]                   = -EINVAL,
 130         [DLM_VOID]                      = -EINVAL,
 131         [DLM_NOTQUEUED]                 = -EAGAIN,      /* Trylock failed */
 132         [DLM_IVBUFLEN]                  = -EINVAL,
 133         [DLM_CVTUNGRANT]                = -EPERM,
 134         [DLM_BADPARAM]                  = -EINVAL,
 135         [DLM_VALNOTVALID]               = -EINVAL,
 136         [DLM_REJECTED]                  = -EPERM,
 137         [DLM_ABORT]                     = -EINVAL,
 138         [DLM_CANCEL]                    = -DLM_EUNLOCK, /* Successful cancel */
 139         [DLM_IVRESHANDLE]               = -EINVAL,
 140         [DLM_DEADLOCK]                  = -EDEADLK,
 141         [DLM_DENIED_NOASTS]             = -EINVAL,
 142         [DLM_FORWARD]                   = -EINVAL,
 143         [DLM_TIMEOUT]                   = -ETIMEDOUT,
 144         [DLM_IVGROUPID]                 = -EINVAL,
 145         [DLM_VERS_CONFLICT]             = -EOPNOTSUPP,
 146         [DLM_BAD_DEVICE_PATH]           = -ENOENT,
 147         [DLM_NO_DEVICE_PERMISSION]      = -EPERM,
 148         [DLM_NO_CONTROL_DEVICE]         = -ENOENT,
 149         [DLM_RECOVERING]                = -ENOTCONN,
 150         [DLM_MIGRATING]                 = -ERESTART,
 151         [DLM_MAXSTATS]                  = -EINVAL,
 152 };
 153 static int dlm_status_to_errno(enum dlm_status status)
 154 {
 155         BUG_ON(status > (sizeof(status_map) / sizeof(status_map[0])));
 156
 157         return status_map[status];
 158 }
 159
 160 static void o2dlm_lock_ast_wrapper(void *astarg)
 161 {
 162         BUG_ON(lproto == NULL);
 163
 164         lproto->lp_lock_ast(astarg);
 165 }
 166
 167 static void o2dlm_blocking_ast_wrapper(void *astarg, int level)
 168 {
 169         BUG_ON(lproto == NULL);
 170
 171         lproto->lp_blocking_ast(astarg, level);
 172 }
 173
 174 static void o2dlm_unlock_ast_wrapper(void *astarg, enum dlm_status status)
 175 {
 176         int error;
 177
 178         BUG_ON(lproto == NULL);
 179
 180         /*
 181          * XXX: CANCEL values are sketchy.
 182          *
 183          * Currently we have preserved the o2dlm paradigm.  You can get
 184          * unlock_ast() whether the cancel succeded or not.
 185          *
 186          * First, we're going to pass DLM_EUNLOCK just like fs/dlm does for
 187          * successful unlocks.  That is a clean behavior.
 188          *
 189          * In o2dlm, you can get both the lock_ast() for the lock being
 190          * granted and the unlock_ast() for the CANCEL failing.  A
 191          * successful cancel sends DLM_NORMAL here.  If the
 192          * lock grant happened before the cancel arrived, you get
 193          * DLM_CANCELGRANT.  For now, we'll use DLM_ECANCEL to signify
 194          * CANCELGRANT - the CANCEL was supposed to happen but didn't.  We
 195          * can then use DLM_EUNLOCK to signify a successful CANCEL -
 196          * effectively, the CANCEL caused the lock to roll back.
 197          *
 198          * In the future, we will likely move the o2dlm to send only one
 199          * ast - either unlock_ast() for a successful CANCEL or lock_ast()
 200          * when the grant succeeds.  At that point, we'll send DLM_ECANCEL
 201          * for all cancel results (CANCELGRANT will no longer exist).
 202          */
 203         error = dlm_status_to_errno(status);
 204
 205         /* Successful unlock is DLM_EUNLOCK */
 206         if (!error)
 207                 error = -DLM_EUNLOCK;
 208
 209         lproto->lp_unlock_ast(astarg, error);
 210 }
 211
 212 int ocfs2_dlm_lock(struct ocfs2_cluster_connection *conn,
 213                    int mode,
 214                    union ocfs2_dlm_lksb *lksb,
 215                    u32 flags,
 216                    void *name,
 217                    unsigned int namelen,
 218                    void *astarg)
 219 {
 220         enum dlm_status status;
 221         int o2dlm_mode = mode_to_o2dlm(mode);
 222         int o2dlm_flags = flags_to_o2dlm(flags);
 223         int ret;
 224
 225         BUG_ON(lproto == NULL);
 226
 227         status = dlmlock(conn->cc_lockspace, o2dlm_mode, &lksb->lksb_o2dlm,
 228                          o2dlm_flags, name, namelen,
 229                          o2dlm_lock_ast_wrapper, astarg,
 230                          o2dlm_blocking_ast_wrapper);
 231         ret = dlm_status_to_errno(status);
 232         return ret;
 233 }
 234
 235 int ocfs2_dlm_unlock(struct ocfs2_cluster_connection *conn,
 236                      union ocfs2_dlm_lksb *lksb,
 237                      u32 flags,
 238                      void *astarg)
 239 {
 240         enum dlm_status status;
 241         int o2dlm_flags = flags_to_o2dlm(flags);
 242         int ret;
 243
 244         BUG_ON(lproto == NULL);
 245
 246         status = dlmunlock(conn->cc_lockspace, &lksb->lksb_o2dlm,
 247                            o2dlm_flags, o2dlm_unlock_ast_wrapper, astarg);
 248         ret = dlm_status_to_errno(status);
 249         return ret;
 250 }
 251
 252 int ocfs2_dlm_lock_status(union ocfs2_dlm_lksb *lksb)
 253 {
 254         return dlm_status_to_errno(lksb->lksb_o2dlm.status);
 255 }
 256
 257 /*
 258  * Why don't we cast to ocfs2_meta_lvb?  The "clean" answer is that we
 259  * don't cast at the glue level.  The real answer is that the header
 260  * ordering is nigh impossible.
 261  */
 262 void *ocfs2_dlm_lvb(union ocfs2_dlm_lksb *lksb)
 263 {
 264         return (void *)(lksb->lksb_o2dlm.lvb);
 265 }
 266
 267 /*
 268  * Called from the dlm when it's about to evict a node. This is how the
 269  * classic stack signals node death.
 270  */
 271 static void o2dlm_eviction_cb(int node_num, void *data)
 272 {
 273         struct ocfs2_cluster_connection *conn = data;
 274
 275         mlog(ML_NOTICE, "o2dlm has evicted node %d from group %.*s\n",
 276              node_num, conn->cc_namelen, conn->cc_name);
 277
 278         conn->cc_recovery_handler(node_num, conn->cc_recovery_data);
 279 }
 280
 281 int ocfs2_cluster_connect(const char *group,
 282                           int grouplen,
 283                           void (*recovery_handler)(int node_num,
 284                                                    void *recovery_data),
 285                           void *recovery_data,
 286                           struct ocfs2_cluster_connection **conn)
 287 {
 288         int rc = 0;
 289         struct ocfs2_cluster_connection *new_conn;
 290         u32 dlm_key;
 291         struct dlm_ctxt *dlm;
 292         struct o2dlm_private *priv;
 293         struct dlm_protocol_version dlm_version;
 294
 295         BUG_ON(group == NULL);
 296         BUG_ON(conn == NULL);
 297         BUG_ON(recovery_handler == NULL);
 298
 299         if (grouplen > GROUP_NAME_MAX) {
 300                 rc = -EINVAL;
 301                 goto out;
 302         }
 303
 304         new_conn = kzalloc(sizeof(struct ocfs2_cluster_connection),
 305                            GFP_KERNEL);
 306         if (!new_conn) {
 307                 rc = -ENOMEM;
 308                 goto out;
 309         }
 310
 311         memcpy(new_conn->cc_name, group, grouplen);
 312         new_conn->cc_namelen = grouplen;
 313         new_conn->cc_recovery_handler = recovery_handler;
 314         new_conn->cc_recovery_data = recovery_data;
 315
 316         /* Start the new connection at our maximum compatibility level */
 317         new_conn->cc_version = lproto->lp_max_version;
 318
 319         priv = kzalloc(sizeof(struct o2dlm_private), GFP_KERNEL);
 320         if (!priv) {
 321                 rc = -ENOMEM;
 322                 goto out_free;
 323         }
 324
 325         /* This just fills the structure in.  It is safe to use new_conn. */
 326         dlm_setup_eviction_cb(&priv->op_eviction_cb, o2dlm_eviction_cb,
 327                               new_conn);
 328
 329         new_conn->cc_private = priv;
 330
 331         /* used by the dlm code to make message headers unique, each
 332          * node in this domain must agree on this. */
 333         dlm_key = crc32_le(0, group, grouplen);
 334         dlm_version.pv_major = new_conn->cc_version.pv_major;
 335         dlm_version.pv_minor = new_conn->cc_version.pv_minor;
 336
 337         dlm = dlm_register_domain(group, dlm_key, &dlm_version);
 338         if (IS_ERR(dlm)) {
 339                 rc = PTR_ERR(dlm);
 340                 mlog_errno(rc);
 341                 goto out_free;
 342         }
 343
 344         new_conn->cc_version.pv_major = dlm_version.pv_major;
 345         new_conn->cc_version.pv_minor = dlm_version.pv_minor;
 346         new_conn->cc_lockspace = dlm;
 347
 348         dlm_register_eviction_cb(dlm, &priv->op_eviction_cb);
 349
 350         *conn = new_conn;
 351
 352 out_free:
 353         if (rc) {
 354                 kfree(new_conn->cc_private);
 355                 kfree(new_conn);
 356         }
 357
 358 out:
 359         return rc;
 360 }
 361
 362 int ocfs2_cluster_disconnect(struct ocfs2_cluster_connection *conn)
 363 {
 364         struct dlm_ctxt *dlm = conn->cc_lockspace;
 365         struct o2dlm_private *priv = conn->cc_private;
 366
 367         dlm_unregister_eviction_cb(&priv->op_eviction_cb);
 368         dlm_unregister_domain(dlm);
 369
 370         kfree(priv);
 371         kfree(conn);
 372
 373         return 0;
 374 }
 375
 376 int ocfs2_cluster_this_node(unsigned int *node)
 377 {
 378         int node_num;
 379
 380         node_num = o2nm_this_node();
 381         if (node_num == O2NM_INVALID_NODE_NUM)
 382                 return -ENOENT;
 383
 384         if (node_num >= O2NM_MAX_NODES)
 385                 return -EOVERFLOW;
 386
 387         *node = node_num;
 388         return 0;
 389 }
 390
 391 void o2cb_get_stack(struct ocfs2_locking_protocol *proto)
 392 {
 393         BUG_ON(proto == NULL);
 394
 395         lproto = proto;
 396 }
 397
 398 void o2cb_put_stack(void)
 399 {
 400         lproto = NULL;
 401 }