err.no Git - linux-2.6/blob - 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 "stackglue.h"
  29
  30 static struct ocfs2_locking_protocol *lproto;
  31
  32 struct o2dlm_private {
  33         struct dlm_eviction_cb op_eviction_cb;
  34 };
  35
  36 /* These should be identical */
  37 #if (DLM_LOCK_IV != LKM_IVMODE)
  38 # error Lock modes do not match
  39 #endif
  40 #if (DLM_LOCK_NL != LKM_NLMODE)
  41 # error Lock modes do not match
  42 #endif
  43 #if (DLM_LOCK_CR != LKM_CRMODE)
  44 # error Lock modes do not match
  45 #endif
  46 #if (DLM_LOCK_CW != LKM_CWMODE)
  47 # error Lock modes do not match
  48 #endif
  49 #if (DLM_LOCK_PR != LKM_PRMODE)
  50 # error Lock modes do not match
  51 #endif
  52 #if (DLM_LOCK_PW != LKM_PWMODE)
  53 # error Lock modes do not match
  54 #endif
  55 #if (DLM_LOCK_EX != LKM_EXMODE)
  56 # error Lock modes do not match
  57 #endif
  58 static inline int mode_to_o2dlm(int mode)
  59 {
  60         BUG_ON(mode > LKM_MAXMODE);
  61
  62         return mode;
  63 }
  64
  65 #define map_flag(_generic, _o2dlm)              \
  66         if (flags & (_generic)) {               \
  67                 flags &= ~(_generic);           \
  68                 o2dlm_flags |= (_o2dlm);        \
  69         }
  70 static int flags_to_o2dlm(u32 flags)
  71 {
  72         int o2dlm_flags = 0;
  73
  74         map_flag(DLM_LKF_NOQUEUE, LKM_NOQUEUE);
  75         map_flag(DLM_LKF_CANCEL, LKM_CANCEL);
  76         map_flag(DLM_LKF_CONVERT, LKM_CONVERT);
  77         map_flag(DLM_LKF_VALBLK, LKM_VALBLK);
  78         map_flag(DLM_LKF_IVVALBLK, LKM_INVVALBLK);
  79         map_flag(DLM_LKF_ORPHAN, LKM_ORPHAN);
  80         map_flag(DLM_LKF_FORCEUNLOCK, LKM_FORCE);
  81         map_flag(DLM_LKF_TIMEOUT, LKM_TIMEOUT);
  82         map_flag(DLM_LKF_LOCAL, LKM_LOCAL);
  83
  84         /* map_flag() should have cleared every flag passed in */
  85         BUG_ON(flags != 0);
  86
  87         return o2dlm_flags;
  88 }
  89 #undef map_flag
  90
  91 /*
  92  * Map an o2dlm status to standard errno values.
  93  *
  94  * o2dlm only uses a handful of these, and returns even fewer to the
  95  * caller. Still, we try to assign sane values to each error.
  96  *
  97  * The following value pairs have special meanings to dlmglue, thus
  98  * the right hand side needs to stay unique - never duplicate the
  99  * mapping elsewhere in the table!
 100  *
 101  * DLM_NORMAL:          0
 102  * DLM_NOTQUEUED:       -EAGAIN
 103  * DLM_CANCELGRANT:     -DLM_ECANCEL
 104  * DLM_CANCEL:          -DLM_EUNLOCK
 105  */
 106 /* Keep in sync with dlmapi.h */
 107 static int status_map[] = {
 108         [DLM_NORMAL]                    = 0,            /* Success */
 109         [DLM_GRANTED]                   = -EINVAL,
 110         [DLM_DENIED]                    = -EACCES,
 111         [DLM_DENIED_NOLOCKS]            = -EACCES,
 112         [DLM_WORKING]                   = -EBUSY,
 113         [DLM_BLOCKED]                   = -EINVAL,
 114         [DLM_BLOCKED_ORPHAN]            = -EINVAL,
 115         [DLM_DENIED_GRACE_PERIOD]       = -EACCES,
 116         [DLM_SYSERR]                    = -ENOMEM,      /* It is what it is */
 117         [DLM_NOSUPPORT]                 = -EPROTO,
 118         [DLM_CANCELGRANT]               = -DLM_ECANCEL, /* Cancel after grant */
 119         [DLM_IVLOCKID]                  = -EINVAL,
 120         [DLM_SYNC]                      = -EINVAL,
 121         [DLM_BADTYPE]                   = -EINVAL,
 122         [DLM_BADRESOURCE]               = -EINVAL,
 123         [DLM_MAXHANDLES]                = -ENOMEM,
 124         [DLM_NOCLINFO]                  = -EINVAL,
 125         [DLM_NOLOCKMGR]                 = -EINVAL,
 126         [DLM_NOPURGED]                  = -EINVAL,
 127         [DLM_BADARGS]                   = -EINVAL,
 128         [DLM_VOID]                      = -EINVAL,
 129         [DLM_NOTQUEUED]                 = -EAGAIN,      /* Trylock failed */
 130         [DLM_IVBUFLEN]                  = -EINVAL,
 131         [DLM_CVTUNGRANT]                = -EPERM,
 132         [DLM_BADPARAM]                  = -EINVAL,
 133         [DLM_VALNOTVALID]               = -EINVAL,
 134         [DLM_REJECTED]                  = -EPERM,
 135         [DLM_ABORT]                     = -EINVAL,
 136         [DLM_CANCEL]                    = -DLM_EUNLOCK, /* Successful cancel */
 137         [DLM_IVRESHANDLE]               = -EINVAL,
 138         [DLM_DEADLOCK]                  = -EDEADLK,
 139         [DLM_DENIED_NOASTS]             = -EINVAL,
 140         [DLM_FORWARD]                   = -EINVAL,
 141         [DLM_TIMEOUT]                   = -ETIMEDOUT,
 142         [DLM_IVGROUPID]                 = -EINVAL,
 143         [DLM_VERS_CONFLICT]             = -EOPNOTSUPP,
 144         [DLM_BAD_DEVICE_PATH]           = -ENOENT,
 145         [DLM_NO_DEVICE_PERMISSION]      = -EPERM,
 146         [DLM_NO_CONTROL_DEVICE]         = -ENOENT,
 147         [DLM_RECOVERING]                = -ENOTCONN,
 148         [DLM_MIGRATING]                 = -ERESTART,
 149         [DLM_MAXSTATS]                  = -EINVAL,
 150 };
 151 static int dlm_status_to_errno(enum dlm_status status)
 152 {
 153         BUG_ON(status > (sizeof(status_map) / sizeof(status_map[0])));
 154
 155         return status_map[status];
 156 }
 157
 158 static void o2dlm_lock_ast_wrapper(void *astarg)
 159 {
 160         BUG_ON(lproto == NULL);
 161
 162         lproto->lp_lock_ast(astarg);
 163 }
 164
 165 static void o2dlm_blocking_ast_wrapper(void *astarg, int level)
 166 {
 167         BUG_ON(lproto == NULL);
 168
 169         lproto->lp_blocking_ast(astarg, level);
 170 }
 171
 172 static void o2dlm_unlock_ast_wrapper(void *astarg, enum dlm_status status)
 173 {
 174         int error;
 175
 176         BUG_ON(lproto == NULL);
 177
 178         /*
 179          * XXX: CANCEL values are sketchy.
 180          *
 181          * Currently we have preserved the o2dlm paradigm.  You can get
 182          * unlock_ast() whether the cancel succeded or not.
 183          *
 184          * First, we're going to pass DLM_EUNLOCK just like fs/dlm does for
 185          * successful unlocks.  That is a clean behavior.
 186          *
 187          * In o2dlm, you can get both the lock_ast() for the lock being
 188          * granted and the unlock_ast() for the CANCEL failing.  A
 189          * successful cancel sends DLM_NORMAL here.  If the
 190          * lock grant happened before the cancel arrived, you get
 191          * DLM_CANCELGRANT.  For now, we'll use DLM_ECANCEL to signify
 192          * CANCELGRANT - the CANCEL was supposed to happen but didn't.  We
 193          * can then use DLM_EUNLOCK to signify a successful CANCEL -
 194          * effectively, the CANCEL caused the lock to roll back.
 195          *
 196          * In the future, we will likely move the o2dlm to send only one
 197          * ast - either unlock_ast() for a successful CANCEL or lock_ast()
 198          * when the grant succeeds.  At that point, we'll send DLM_ECANCEL
 199          * for all cancel results (CANCELGRANT will no longer exist).
 200          */
 201         error = dlm_status_to_errno(status);
 202
 203         /* Successful unlock is DLM_EUNLOCK */
 204         if (!error)
 205                 error = -DLM_EUNLOCK;
 206
 207         lproto->lp_unlock_ast(astarg, error);
 208 }
 209
 210 int ocfs2_dlm_lock(struct ocfs2_cluster_connection *conn,
 211                    int mode,
 212                    union ocfs2_dlm_lksb *lksb,
 213                    u32 flags,
 214                    void *name,
 215                    unsigned int namelen,
 216                    void *astarg)
 217 {
 218         enum dlm_status status;
 219         int o2dlm_mode = mode_to_o2dlm(mode);
 220         int o2dlm_flags = flags_to_o2dlm(flags);
 221         int ret;
 222
 223         BUG_ON(lproto == NULL);
 224
 225         status = dlmlock(conn->cc_lockspace, o2dlm_mode, &lksb->lksb_o2dlm,
 226                          o2dlm_flags, name, namelen,
 227                          o2dlm_lock_ast_wrapper, astarg,
 228                          o2dlm_blocking_ast_wrapper);
 229         ret = dlm_status_to_errno(status);
 230         return ret;
 231 }
 232
 233 int ocfs2_dlm_unlock(struct ocfs2_cluster_connection *conn,
 234                      union ocfs2_dlm_lksb *lksb,
 235                      u32 flags,
 236                      void *astarg)
 237 {
 238         enum dlm_status status;
 239         int o2dlm_flags = flags_to_o2dlm(flags);
 240         int ret;
 241
 242         BUG_ON(lproto == NULL);
 243
 244         status = dlmunlock(conn->cc_lockspace, &lksb->lksb_o2dlm,
 245                            o2dlm_flags, o2dlm_unlock_ast_wrapper, astarg);
 246         ret = dlm_status_to_errno(status);
 247         return ret;
 248 }
 249
 250 int ocfs2_dlm_lock_status(union ocfs2_dlm_lksb *lksb)
 251 {
 252         return dlm_status_to_errno(lksb->lksb_o2dlm.status);
 253 }
 254
 255 /*
 256  * Why don't we cast to ocfs2_meta_lvb?  The "clean" answer is that we
 257  * don't cast at the glue level.  The real answer is that the header
 258  * ordering is nigh impossible.
 259  */
 260 void *ocfs2_dlm_lvb(union ocfs2_dlm_lksb *lksb)
 261 {
 262         return (void *)(lksb->lksb_o2dlm.lvb);
 263 }
 264
 265 /*
 266  * Called from the dlm when it's about to evict a node. This is how the
 267  * classic stack signals node death.
 268  */
 269 static void o2dlm_eviction_cb(int node_num, void *data)
 270 {
 271         struct ocfs2_cluster_connection *conn = data;
 272
 273         mlog(ML_NOTICE, "o2dlm has evicted node %d from group %.*s\n",
 274              node_num, conn->cc_namelen, conn->cc_name);
 275
 276         conn->cc_recovery_handler(node_num, conn->cc_recovery_data);
 277 }
 278
 279 int ocfs2_cluster_connect(const char *group,
 280                           int grouplen,
 281                           void (*recovery_handler)(int node_num,
 282                                                    void *recovery_data),
 283                           void *recovery_data,
 284                           struct ocfs2_cluster_connection **conn)
 285 {
 286         int rc = 0;
 287         struct ocfs2_cluster_connection *new_conn;
 288         u32 dlm_key;
 289         struct dlm_ctxt *dlm;
 290         struct o2dlm_private *priv;
 291         struct dlm_protocol_version dlm_version;
 292
 293         BUG_ON(group == NULL);
 294         BUG_ON(conn == NULL);
 295         BUG_ON(recovery_handler == NULL);
 296
 297         if (grouplen > GROUP_NAME_MAX) {
 298                 rc = -EINVAL;
 299                 goto out;
 300         }
 301
 302         new_conn = kzalloc(sizeof(struct ocfs2_cluster_connection),
 303                            GFP_KERNEL);
 304         if (!new_conn) {
 305                 rc = -ENOMEM;
 306                 goto out;
 307         }
 308
 309         memcpy(new_conn->cc_name, group, grouplen);
 310         new_conn->cc_namelen = grouplen;
 311         new_conn->cc_recovery_handler = recovery_handler;
 312         new_conn->cc_recovery_data = recovery_data;
 313
 314         /* Start the new connection at our maximum compatibility level */
 315         new_conn->cc_version = lproto->lp_max_version;
 316
 317         priv = kzalloc(sizeof(struct o2dlm_private), GFP_KERNEL);
 318         if (!priv) {
 319                 rc = -ENOMEM;
 320                 goto out_free;
 321         }
 322
 323         /* This just fills the structure in.  It is safe to use new_conn. */
 324         dlm_setup_eviction_cb(&priv->op_eviction_cb, o2dlm_eviction_cb,
 325                               new_conn);
 326
 327         new_conn->cc_private = priv;
 328
 329         /* used by the dlm code to make message headers unique, each
 330          * node in this domain must agree on this. */
 331         dlm_key = crc32_le(0, group, grouplen);
 332         dlm_version.pv_major = new_conn->cc_version.pv_major;
 333         dlm_version.pv_minor = new_conn->cc_version.pv_minor;
 334
 335         dlm = dlm_register_domain(group, dlm_key, &dlm_version);
 336         if (IS_ERR(dlm)) {
 337                 rc = PTR_ERR(dlm);
 338                 mlog_errno(rc);
 339                 goto out_free;
 340         }
 341
 342         new_conn->cc_version.pv_major = dlm_version.pv_major;
 343         new_conn->cc_version.pv_minor = dlm_version.pv_minor;
 344         new_conn->cc_lockspace = dlm;
 345
 346         dlm_register_eviction_cb(dlm, &priv->op_eviction_cb);
 347
 348         *conn = new_conn;
 349
 350 out_free:
 351         if (rc) {
 352                 kfree(new_conn->cc_private);
 353                 kfree(new_conn);
 354         }
 355
 356 out:
 357         return rc;
 358 }
 359
 360 int ocfs2_cluster_disconnect(struct ocfs2_cluster_connection *conn)
 361 {
 362         struct dlm_ctxt *dlm = conn->cc_lockspace;
 363         struct o2dlm_private *priv = conn->cc_private;
 364
 365         dlm_unregister_eviction_cb(&priv->op_eviction_cb);
 366         dlm_unregister_domain(dlm);
 367
 368         kfree(priv);
 369         kfree(conn);
 370
 371         return 0;
 372 }
 373
 374 void o2cb_get_stack(struct ocfs2_locking_protocol *proto)
 375 {
 376         BUG_ON(proto == NULL);
 377
 378         lproto = proto;
 379 }
 380
 381 void o2cb_put_stack(void)
 382 {
 383         lproto = NULL;
 384 }