lustre/mdt/mdt_capa.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mdt/mdt_capa.c
  37  *
  38  * Lustre Metadata Target (mdt) capability key read/write/update.
  39  *
  40  * Author: Lai Siyao <lsy@clusterfs.com>
  41  */
  42
  43 #ifndef EXPORT_SYMTAB
  44 # define EXPORT_SYMTAB
  45 #endif
  46 #define DEBUG_SUBSYSTEM S_MDS
  47
  48 #include "mdt_internal.h"
  49
  50 static inline void set_capa_key_expiry(struct mdt_device *mdt)
  51 {
  52         mdt->mdt_ck_expiry = jiffies + mdt->mdt_ck_timeout * CFS_HZ;
  53 }
  54
  55 static void make_capa_key(struct lustre_capa_key *key,
  56                           mdsno_t mdsnum, int keyid)
  57 {
  58         key->lk_mdsid = mdsnum;
  59         key->lk_keyid = keyid + 1;
  60         ll_get_random_bytes(key->lk_key, sizeof(key->lk_key));
  61 }
  62
  63 static inline void lck_cpu_to_le(struct lustre_capa_key *tgt,
  64                                  struct lustre_capa_key *src)
  65 {
  66         tgt->lk_mdsid   = cpu_to_le64(src->lk_mdsid);
  67         tgt->lk_keyid   = cpu_to_le32(src->lk_keyid);
  68         tgt->lk_padding = cpu_to_le32(src->lk_padding);
  69         memcpy(tgt->lk_key, src->lk_key, sizeof(src->lk_key));
  70 }
  71
  72 static inline void lck_le_to_cpu(struct lustre_capa_key *tgt,
  73                                  struct lustre_capa_key *src)
  74 {
  75         tgt->lk_mdsid   = le64_to_cpu(src->lk_mdsid);
  76         tgt->lk_keyid   = le32_to_cpu(src->lk_keyid);
  77         tgt->lk_padding = le32_to_cpu(src->lk_padding);
  78         memcpy(tgt->lk_key, src->lk_key, sizeof(src->lk_key));
  79 }
  80
  81 static int write_capa_keys(const struct lu_env *env,
  82                            struct mdt_device *mdt,
  83                            struct lustre_capa_key *keys)
  84 {
  85         struct mdt_thread_info *mti;
  86         struct lustre_capa_key *tmp;
  87         struct thandle *th;
  88         loff_t off = 0;
  89         int i, rc;
  90
  91         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
  92         mdt_trans_credit_init(env, mdt, MDT_TXN_CAPA_KEYS_WRITE_OP);
  93         th = mdt_trans_start(env, mdt);
  94         if (IS_ERR(th))
  95                 RETURN(PTR_ERR(th));
  96
  97         tmp = &mti->mti_capa_key;
  98
  99         for (i = 0; i < 2; i++) {
 100                 lck_cpu_to_le(tmp, &keys[i]);
 101
 102                 rc = dt_record_write(env, mdt->mdt_ck_obj,
 103                                      mdt_buf_const(env, tmp, sizeof(*tmp)),
 104                                      &off, th);
 105                 if (rc)
 106                         break;
 107         }
 108
 109         mdt_trans_stop(env, mdt, th);
 110
 111         CDEBUG(D_INFO, "write capability keys rc = %d:\n", rc);
 112         return rc;
 113 }
 114
 115 static int read_capa_keys(const struct lu_env *env,
 116                           struct mdt_device *mdt,
 117                           struct lustre_capa_key *keys)
 118 {
 119         struct mdt_thread_info *mti;
 120         struct lustre_capa_key *tmp;
 121         loff_t off = 0;
 122         int i, rc;
 123
 124         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 125         tmp = &mti->mti_capa_key;
 126
 127         for (i = 0; i < 2; i++) {
 128                 rc = dt_record_read(env, mdt->mdt_ck_obj,
 129                                     mdt_buf(env, tmp, sizeof(*tmp)), &off);
 130                 if (rc)
 131                         return rc;
 132
 133                 lck_le_to_cpu(&keys[i], tmp);
 134                 DEBUG_CAPA_KEY(D_SEC, &keys[i], "read");
 135         }
 136
 137         return 0;
 138 }
 139
 140 int mdt_capa_keys_init(const struct lu_env *env, struct mdt_device *mdt)
 141 {
 142         struct lustre_capa_key  *keys = mdt->mdt_capa_keys;
 143         struct mdt_thread_info  *mti;
 144         struct dt_object        *obj;
 145         struct lu_attr          *la;
 146         mdsno_t                  mdsnum;
 147         unsigned long            size;
 148         int                      rc;
 149         ENTRY;
 150
 151         mdsnum = mdt_md_site(mdt)->ms_node_id;
 152
 153         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 154         LASSERT(mti != NULL);
 155         la = &mti->mti_attr.ma_attr;
 156
 157         obj = mdt->mdt_ck_obj;
 158         rc = obj->do_ops->do_attr_get(env, mdt->mdt_ck_obj, la, BYPASS_CAPA);
 159         if (rc)
 160                 RETURN(rc);
 161
 162         size = (unsigned long)la->la_size;
 163         if (size == 0) {
 164                 int i;
 165
 166                 for (i = 0; i < 2; i++) {
 167                         make_capa_key(&keys[i], mdsnum, i);
 168                         DEBUG_CAPA_KEY(D_SEC, &keys[i], "initializing");
 169                 }
 170
 171                 rc = write_capa_keys(env, mdt, keys);
 172                 if (rc) {
 173                         CERROR("error writing MDS %s: rc %d\n", CAPA_KEYS, rc);
 174                         RETURN(rc);
 175                 }
 176         } else {
 177                 rc = read_capa_keys(env, mdt, keys);
 178                 if (rc) {
 179                         CERROR("error reading MDS %s: rc %d\n", CAPA_KEYS, rc);
 180                         RETURN(rc);
 181                 }
 182         }
 183         set_capa_key_expiry(mdt);
 184         cfs_timer_arm(&mdt->mdt_ck_timer, mdt->mdt_ck_expiry);
 185         CDEBUG(D_SEC, "mds_ck_timer %lu\n", mdt->mdt_ck_expiry);
 186         RETURN(0);
 187 }
 188
 189 void mdt_ck_timer_callback(unsigned long castmeharder)
 190 {
 191         struct mdt_device *mdt = (struct mdt_device *)castmeharder;
 192         struct ptlrpc_thread *thread = &mdt->mdt_ck_thread;
 193
 194         ENTRY;
 195         thread->t_flags |= SVC_EVENT;
 196         cfs_waitq_signal(&thread->t_ctl_waitq);
 197         EXIT;
 198 }
 199
 200 static int mdt_ck_thread_main(void *args)
 201 {
 202         struct mdt_device      *mdt = args;
 203         struct ptlrpc_thread   *thread = &mdt->mdt_ck_thread;
 204         struct lustre_capa_key *bkey = &mdt->mdt_capa_keys[0],
 205                                *rkey = &mdt->mdt_capa_keys[1];
 206         struct lustre_capa_key *tmp;
 207         struct lu_env           env;
 208         struct mdt_thread_info *info;
 209         struct md_device       *next;
 210         struct l_wait_info      lwi = { 0 };
 211         mdsno_t                 mdsnum;
 212         int                     rc;
 213         ENTRY;
 214
 215         cfs_daemonize_ctxt("mdt_ck");
 216         cfs_block_allsigs();
 217
 218         thread->t_flags = SVC_RUNNING;
 219         cfs_waitq_signal(&thread->t_ctl_waitq);
 220
 221         rc = lu_env_init(&env, LCT_MD_THREAD|LCT_REMEMBER|LCT_NOREF);
 222         if (rc)
 223                 RETURN(rc);
 224
 225         thread->t_env = &env;
 226         env.le_ctx.lc_thread = thread;
 227         env.le_ctx.lc_cookie = 0x1;
 228
 229         info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
 230         LASSERT(info != NULL);
 231
 232         tmp = &info->mti_capa_key;
 233         mdsnum = mdt_md_site(mdt)->ms_node_id;
 234         while (1) {
 235                 l_wait_event(thread->t_ctl_waitq,
 236                              thread->t_flags & (SVC_STOPPING | SVC_EVENT),
 237                              &lwi);
 238
 239                 if (thread->t_flags & SVC_STOPPING)
 240                         break;
 241                 thread->t_flags &= ~SVC_EVENT;
 242
 243                 if (cfs_time_before(cfs_time_current(), mdt->mdt_ck_expiry))
 244                         break;
 245
 246                 *tmp = *rkey;
 247                 make_capa_key(tmp, mdsnum, rkey->lk_keyid);
 248
 249                 next = mdt->mdt_child;
 250                 rc = next->md_ops->mdo_update_capa_key(&env, next, tmp);
 251                 if (!rc) {
 252                         cfs_spin_lock(&capa_lock);
 253                         *bkey = *rkey;
 254                         *rkey = *tmp;
 255                         cfs_spin_unlock(&capa_lock);
 256
 257                         rc = write_capa_keys(&env, mdt, mdt->mdt_capa_keys);
 258                         if (rc) {
 259                                 cfs_spin_lock(&capa_lock);
 260                                 *rkey = *bkey;
 261                                 memset(bkey, 0, sizeof(*bkey));
 262                                 cfs_spin_unlock(&capa_lock);
 263                         } else {
 264                                 set_capa_key_expiry(mdt);
 265                                 DEBUG_CAPA_KEY(D_SEC, rkey, "new");
 266                         }
 267                 }
 268                 if (rc) {
 269                         DEBUG_CAPA_KEY(D_ERROR, rkey, "update failed for");
 270                         /* next retry is in 300 sec */
 271                         mdt->mdt_ck_expiry = jiffies + 300 * CFS_HZ;
 272                 }
 273
 274                 cfs_timer_arm(&mdt->mdt_ck_timer, mdt->mdt_ck_expiry);
 275                 CDEBUG(D_SEC, "mdt_ck_timer %lu\n", mdt->mdt_ck_expiry);
 276         }
 277         lu_env_fini(&env);
 278
 279         thread->t_flags = SVC_STOPPED;
 280         cfs_waitq_signal(&thread->t_ctl_waitq);
 281         RETURN(0);
 282 }
 283
 284 int mdt_ck_thread_start(struct mdt_device *mdt)
 285 {
 286         struct ptlrpc_thread *thread = &mdt->mdt_ck_thread;
 287         int rc;
 288
 289         cfs_waitq_init(&thread->t_ctl_waitq);
 290         rc = cfs_kernel_thread(mdt_ck_thread_main, mdt,
 291                                (CLONE_VM | CLONE_FILES));
 292         if (rc < 0) {
 293                 CERROR("cannot start mdt_ck thread, rc = %d\n", rc);
 294                 return rc;
 295         }
 296
 297         l_cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING);
 298         return 0;
 299 }
 300
 301 void mdt_ck_thread_stop(struct mdt_device *mdt)
 302 {
 303         struct ptlrpc_thread *thread = &mdt->mdt_ck_thread;
 304
 305         if (!(thread->t_flags & SVC_RUNNING))
 306                 return;
 307
 308         thread->t_flags = SVC_STOPPING;
 309         cfs_waitq_signal(&thread->t_ctl_waitq);
 310         l_cfs_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
 311 }