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 (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/mdt/mdt_capa.c
  40  *
  41  * Lustre Metadata Target (mdt) capability key read/write/update.
  42  *
  43  * Author: Lai Siyao <lsy@clusterfs.com>
  44  */
  45
  46 #ifndef EXPORT_SYMTAB
  47 # define EXPORT_SYMTAB
  48 #endif
  49 #define DEBUG_SUBSYSTEM S_MDS
  50
  51 #include "mdt_internal.h"
  52
  53 static inline void set_capa_key_expiry(struct mdt_device *mdt)
  54 {
  55         mdt->mdt_ck_expiry = jiffies + mdt->mdt_ck_timeout * CFS_HZ;
  56 }
  57
  58 static void make_capa_key(struct lustre_capa_key *key,
  59                           mdsno_t mdsnum, int keyid)
  60 {
  61         key->lk_seq = mdsnum;
  62         key->lk_keyid = keyid + 1;
  63         cfs_get_random_bytes(key->lk_key, sizeof(key->lk_key));
  64 }
  65
  66 static inline void lck_cpu_to_le(struct lustre_capa_key *tgt,
  67                                  struct lustre_capa_key *src)
  68 {
  69         tgt->lk_seq   = cpu_to_le64(src->lk_seq);
  70         tgt->lk_keyid   = cpu_to_le32(src->lk_keyid);
  71         tgt->lk_padding = cpu_to_le32(src->lk_padding);
  72         memcpy(tgt->lk_key, src->lk_key, sizeof(src->lk_key));
  73 }
  74
  75 static inline void lck_le_to_cpu(struct lustre_capa_key *tgt,
  76                                  struct lustre_capa_key *src)
  77 {
  78         tgt->lk_seq   = le64_to_cpu(src->lk_seq);
  79         tgt->lk_keyid   = le32_to_cpu(src->lk_keyid);
  80         tgt->lk_padding = le32_to_cpu(src->lk_padding);
  81         memcpy(tgt->lk_key, src->lk_key, sizeof(src->lk_key));
  82 }
  83
  84 static int write_capa_keys(const struct lu_env *env,
  85                            struct mdt_device *mdt,
  86                            struct lustre_capa_key *keys)
  87 {
  88         struct mdt_thread_info *mti;
  89         struct lustre_capa_key *tmp;
  90         struct thandle *th;
  91         loff_t off = 0;
  92         int i, rc;
  93
  94         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
  95         mdt_trans_credit_init(env, mdt, MDT_TXN_CAPA_KEYS_WRITE_OP);
  96         th = mdt_trans_start(env, mdt);
  97         if (IS_ERR(th))
  98                 RETURN(PTR_ERR(th));
  99
 100         tmp = &mti->mti_capa_key;
 101
 102         for (i = 0; i < 2; i++) {
 103                 lck_cpu_to_le(tmp, &keys[i]);
 104
 105                 rc = dt_record_write(env, mdt->mdt_ck_obj,
 106                                      mdt_buf_const(env, tmp, sizeof(*tmp)),
 107                                      &off, th);
 108                 if (rc)
 109                         break;
 110         }
 111
 112         mdt_trans_stop(env, mdt, th);
 113
 114         CDEBUG(D_INFO, "write capability keys rc = %d:\n", rc);
 115         return rc;
 116 }
 117
 118 static int read_capa_keys(const struct lu_env *env,
 119                           struct mdt_device *mdt,
 120                           struct lustre_capa_key *keys)
 121 {
 122         struct mdt_thread_info *mti;
 123         struct lustre_capa_key *tmp;
 124         loff_t off = 0;
 125         int i, rc;
 126
 127         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 128         tmp = &mti->mti_capa_key;
 129
 130         for (i = 0; i < 2; i++) {
 131                 rc = dt_record_read(env, mdt->mdt_ck_obj,
 132                                     mdt_buf(env, tmp, sizeof(*tmp)), &off);
 133                 if (rc)
 134                         return rc;
 135
 136                 lck_le_to_cpu(&keys[i], tmp);
 137                 DEBUG_CAPA_KEY(D_SEC, &keys[i], "read");
 138         }
 139
 140         return 0;
 141 }
 142
 143 int mdt_capa_keys_init(const struct lu_env *env, struct mdt_device *mdt)
 144 {
 145         struct lustre_capa_key  *keys = mdt->mdt_capa_keys;
 146         struct mdt_thread_info  *mti;
 147         struct dt_object        *obj;
 148         struct lu_attr          *la;
 149         mdsno_t                  mdsnum;
 150         unsigned long            size;
 151         int                      rc;
 152         ENTRY;
 153
 154         mdsnum = mdt_md_site(mdt)->ms_node_id;
 155
 156         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 157         LASSERT(mti != NULL);
 158         la = &mti->mti_attr.ma_attr;
 159
 160         obj = mdt->mdt_ck_obj;
 161         rc = obj->do_ops->do_attr_get(env, mdt->mdt_ck_obj, la, BYPASS_CAPA);
 162         if (rc)
 163                 RETURN(rc);
 164
 165         size = (unsigned long)la->la_size;
 166         if (size == 0) {
 167                 int i;
 168
 169                 for (i = 0; i < 2; i++) {
 170                         make_capa_key(&keys[i], mdsnum, i);
 171                         DEBUG_CAPA_KEY(D_SEC, &keys[i], "initializing");
 172                 }
 173
 174                 rc = write_capa_keys(env, mdt, keys);
 175                 if (rc) {
 176                         CERROR("error writing MDS %s: rc %d\n", CAPA_KEYS, rc);
 177                         RETURN(rc);
 178                 }
 179         } else {
 180                 rc = read_capa_keys(env, mdt, keys);
 181                 if (rc) {
 182                         CERROR("error reading MDS %s: rc %d\n", CAPA_KEYS, rc);
 183                         RETURN(rc);
 184                 }
 185         }
 186         set_capa_key_expiry(mdt);
 187         cfs_timer_arm(&mdt->mdt_ck_timer, mdt->mdt_ck_expiry);
 188         CDEBUG(D_SEC, "mds_ck_timer %lu\n", mdt->mdt_ck_expiry);
 189         RETURN(0);
 190 }
 191
 192 void mdt_ck_timer_callback(unsigned long castmeharder)
 193 {
 194         struct mdt_device *mdt = (struct mdt_device *)castmeharder;
 195         struct ptlrpc_thread *thread = &mdt->mdt_ck_thread;
 196
 197         ENTRY;
 198         thread_add_flags(thread, SVC_EVENT);
 199         cfs_waitq_signal(&thread->t_ctl_waitq);
 200         EXIT;
 201 }
 202
 203 static int mdt_ck_thread_main(void *args)
 204 {
 205         struct mdt_device      *mdt = args;
 206         struct ptlrpc_thread   *thread = &mdt->mdt_ck_thread;
 207         struct lustre_capa_key *bkey = &mdt->mdt_capa_keys[0],
 208                                *rkey = &mdt->mdt_capa_keys[1];
 209         struct lustre_capa_key *tmp;
 210         struct lu_env           env;
 211         struct mdt_thread_info *info;
 212         struct md_device       *next;
 213         struct l_wait_info      lwi = { 0 };
 214         mdsno_t                 mdsnum;
 215         int                     rc;
 216         ENTRY;
 217
 218         cfs_daemonize_ctxt("mdt_ck");
 219         cfs_block_allsigs();
 220
 221         thread_set_flags(thread, SVC_RUNNING);
 222         cfs_waitq_signal(&thread->t_ctl_waitq);
 223
 224         rc = lu_env_init(&env, LCT_MD_THREAD|LCT_REMEMBER|LCT_NOREF);
 225         if (rc)
 226                 RETURN(rc);
 227
 228         thread->t_env = &env;
 229         env.le_ctx.lc_thread = thread;
 230         env.le_ctx.lc_cookie = 0x1;
 231
 232         info = lu_context_key_get(&env.le_ctx, &mdt_thread_key);
 233         LASSERT(info != NULL);
 234
 235         tmp = &info->mti_capa_key;
 236         mdsnum = mdt_md_site(mdt)->ms_node_id;
 237         while (1) {
 238                 l_wait_event(thread->t_ctl_waitq,
 239                              thread_is_stopping(thread) ||
 240                              thread_is_event(thread),
 241                              &lwi);
 242
 243                 if (thread_is_stopping(thread))
 244                         break;
 245                 thread_clear_flags(thread, SVC_EVENT);
 246
 247                 if (cfs_time_before(cfs_time_current(), mdt->mdt_ck_expiry))
 248                         break;
 249
 250                 *tmp = *rkey;
 251                 make_capa_key(tmp, mdsnum, rkey->lk_keyid);
 252
 253                 next = mdt->mdt_child;
 254                 rc = next->md_ops->mdo_update_capa_key(&env, next, tmp);
 255                 if (!rc) {
 256                         cfs_spin_lock(&capa_lock);
 257                         *bkey = *rkey;
 258                         *rkey = *tmp;
 259                         cfs_spin_unlock(&capa_lock);
 260
 261                         rc = write_capa_keys(&env, mdt, mdt->mdt_capa_keys);
 262                         if (rc) {
 263                                 cfs_spin_lock(&capa_lock);
 264                                 *rkey = *bkey;
 265                                 memset(bkey, 0, sizeof(*bkey));
 266                                 cfs_spin_unlock(&capa_lock);
 267                         } else {
 268                                 set_capa_key_expiry(mdt);
 269                                 DEBUG_CAPA_KEY(D_SEC, rkey, "new");
 270                         }
 271                 }
 272                 if (rc) {
 273                         DEBUG_CAPA_KEY(D_ERROR, rkey, "update failed for");
 274                         /* next retry is in 300 sec */
 275                         mdt->mdt_ck_expiry = jiffies + 300 * CFS_HZ;
 276                 }
 277
 278                 cfs_timer_arm(&mdt->mdt_ck_timer, mdt->mdt_ck_expiry);
 279                 CDEBUG(D_SEC, "mdt_ck_timer %lu\n", mdt->mdt_ck_expiry);
 280         }
 281         lu_env_fini(&env);
 282
 283         thread_set_flags(thread, SVC_STOPPED);
 284         cfs_waitq_signal(&thread->t_ctl_waitq);
 285         RETURN(0);
 286 }
 287
 288 int mdt_ck_thread_start(struct mdt_device *mdt)
 289 {
 290         struct ptlrpc_thread *thread = &mdt->mdt_ck_thread;
 291         int rc;
 292
 293         cfs_waitq_init(&thread->t_ctl_waitq);
 294         rc = cfs_create_thread(mdt_ck_thread_main, mdt, CFS_DAEMON_FLAGS);
 295         if (rc < 0) {
 296                 CERROR("cannot start mdt_ck thread, rc = %d\n", rc);
 297                 return rc;
 298         }
 299
 300         l_wait_condition(thread->t_ctl_waitq, thread_is_running(thread));
 301         return 0;
 302 }
 303
 304 void mdt_ck_thread_stop(struct mdt_device *mdt)
 305 {
 306         struct ptlrpc_thread *thread = &mdt->mdt_ck_thread;
 307
 308         if (!thread_is_running(thread))
 309                 return;
 310
 311         thread_set_flags(thread, SVC_STOPPING);
 312         cfs_waitq_signal(&thread->t_ctl_waitq);
 313         l_wait_condition(thread->t_ctl_waitq, thread_is_stopped(thread));
 314 }