lustre/cmobd/cm_oss_reint.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002 Cluster File Systems, Inc. <info@clusterfs.com>
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #define DEBUG_SUBSYSTEM S_CMOBD
  23
  24 #include <linux/version.h>
  25 #include <linux/init.h>
  26 #include <linux/obd_support.h>
  27 #include <linux/lustre_lib.h>
  28 #include <linux/lustre_net.h>
  29 #include <linux/lustre_idl.h>
  30 #include <linux/lustre_dlm.h>
  31 #include <linux/obd_class.h>
  32 #include <linux/lustre_log.h>
  33 #include <linux/lustre_cmobd.h>
  34 #include <linux/lustre_fsfilt.h>
  35 #include <linux/lustre_smfs.h>
  36
  37 #include "cm_internal.h"
  38
  39 void lov_free_memmd(struct lov_stripe_md **lsmp);
  40
  41 int lov_alloc_memmd(struct lov_stripe_md **lsmp, int stripe_count,
  42                     int pattern);
  43
  44 int smfs_rec_unpack(struct smfs_proc_args *args, char *record,
  45                     char **pbuf, int *opcode);
  46
  47 /* helper functions for cmobd to construct pseudo lsm */
  48 int cmobd_dummy_lsm(struct lov_stripe_md **lsmp, int stripe_cnt,
  49                     struct obdo *oa, __u32 stripe_size)
  50 {
  51         int i, rc;
  52         ENTRY;
  53
  54         rc = lov_alloc_memmd(lsmp, stripe_cnt, LOV_PATTERN_CMOBD);
  55         if (rc < 0)
  56                 RETURN(rc);
  57
  58         for (i = 0; i < stripe_cnt; i++) {
  59                 (*lsmp)->lsm_oinfo[i].loi_id = oa->o_id;
  60                 (*lsmp)->lsm_object_id = oa->o_id;
  61                 if (oa->o_valid & OBD_MD_FLGROUP) {
  62                         (*lsmp)->lsm_oinfo[i].loi_gr = oa->o_gr;
  63                         (*lsmp)->lsm_object_gr = oa->o_gr;
  64                 }
  65                 (*lsmp)->lsm_oinfo[i].loi_ost_idx = i;
  66                 (*lsmp)->lsm_stripe_size = stripe_size;
  67         }
  68         RETURN(0);
  69 }
  70
  71 void cmobd_free_lsm(struct lov_stripe_md **lsmp)
  72 {
  73         ENTRY;
  74         lov_free_memmd(lsmp);
  75         EXIT;
  76 }
  77
  78 /* reintegration functions */
  79 static int cmobd_setattr_reint(struct obd_device *obd, void *rec)
  80 {
  81         int rc = 0;
  82         struct lov_stripe_md *lsm;
  83         struct cm_obd *cmobd = &obd->u.cm;
  84         struct obd_export *exp = cmobd->master_exp;
  85         struct obdo *oa = (struct obdo *)rec;
  86         ENTRY;
  87
  88         rc = cmobd_dummy_lsm(&lsm, cmobd->master_desc.ld_tgt_count, oa,
  89                              (__u32)cmobd->master_desc.ld_default_stripe_size);
  90         if (rc)
  91                 GOTO(out, rc);
  92
  93         rc = obd_setattr(exp, oa, lsm, NULL);
  94
  95         cmobd_free_lsm(&lsm);
  96 out:
  97         RETURN(rc);
  98 }
  99
 100 static int cmobd_create_reint(struct obd_device *obd, void *rec)
 101 {
 102         struct cm_obd *cmobd = &obd->u.cm;
 103         struct obd_export *exp = cmobd->master_exp;
 104         struct obdo *oa = (struct obdo *)rec;
 105         struct obd_trans_info oti = { 0 };
 106         struct lov_stripe_md *lsm;
 107         int rc;
 108         ENTRY;
 109
 110         rc = cmobd_dummy_lsm(&lsm, cmobd->master_desc.ld_tgt_count, oa,
 111                              (__u32)cmobd->master_desc.ld_default_stripe_size);
 112         if (rc)
 113                 GOTO(out, rc);
 114         if (cmobd->master_group != oa->o_gr) {
 115                 int group = oa->o_gr;
 116                 int valsize = sizeof(group);
 117
 118                 rc = obd_set_info(exp, strlen("mds_conn"),
 119                                   "mds_conn", valsize, &group);
 120                 if (rc)
 121                         GOTO(out, rc);
 122                 cmobd->master_group = oa->o_gr;
 123         }
 124
 125         oti.oti_flags |= OBD_MODE_CROW;
 126         rc = obd_create(exp, oa, NULL, 0, &lsm, &oti);
 127         cmobd_free_lsm(&lsm);
 128         EXIT;
 129 out:
 130         return rc;
 131 }
 132
 133 /* direct cut-n-paste of filter_blocking_ast() */
 134 static int cache_blocking_ast(struct ldlm_lock *lock,
 135                               struct ldlm_lock_desc *desc,
 136                               void *data, int flag)
 137 {
 138         int rc, do_ast;
 139         ENTRY;
 140
 141         if (flag == LDLM_CB_CANCELING) {
 142                 /* Don't need to do anything here. */
 143                 RETURN(0);
 144         }
 145
 146         /* XXX layering violation!  -phil */
 147         lock_res_and_lock(lock);
 148
 149         /* get this: if filter_blocking_ast() is racing with ldlm_intent_policy,
 150          * such that filter_blocking_ast is called just before l_i_p takes the
 151          * ns_lock, then by the time we get the lock, we might not be the
 152          * correct blocking function anymore.  So check, and return early, if
 153          * so. */
 154         if (lock->l_blocking_ast != cache_blocking_ast) {
 155                 unlock_res_and_lock(lock);
 156                 RETURN(0);
 157         }
 158
 159         lock->l_flags |= LDLM_FL_CBPENDING;
 160         do_ast = (!lock->l_readers && !lock->l_writers);
 161         unlock_res_and_lock(lock);
 162
 163         if (do_ast) {
 164                 struct lustre_handle lockh;
 165                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 166                 ldlm_lock2handle(lock, &lockh);
 167                 rc = ldlm_cli_cancel(&lockh);
 168                 if (rc < 0)
 169                         CERROR("ldlm_cli_cancel: %d\n", rc);
 170         } else {
 171                 LDLM_DEBUG(lock, "Lock still has references, will be "
 172                            "cancelled later");
 173         }
 174         RETURN(0);
 175 }
 176
 177 static int master_blocking_ast(struct ldlm_lock *lock,
 178                                struct ldlm_lock_desc *desc,
 179                                void *data, int flag)
 180 {
 181         int rc;
 182         struct lustre_handle lockh;
 183         ENTRY;
 184
 185         switch (flag) {
 186         case LDLM_CB_BLOCKING:
 187                 ldlm_lock2handle(lock, &lockh);
 188                 rc = ldlm_cli_cancel(&lockh);
 189                 if (rc < 0) {
 190                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
 191                         RETURN(rc);
 192                 }
 193                 break;
 194         case LDLM_CB_CANCELING:
 195                 /* do nothing here by now */
 196                 break;
 197         default:
 198                 LBUG();
 199         }
 200         RETURN(0);
 201 }
 202
 203 static int cmobd_write_extents(struct obd_device *obd, struct obdo *oa,
 204                                struct ldlm_extent *extent)
 205 {
 206         struct cm_obd *cmobd = &obd->u.cm;
 207         struct obd_device *cache = cmobd->cache_exp->exp_obd;
 208         struct lustre_handle lockh_src = { 0 };
 209         struct lustre_handle lockh_dst = { 0 };
 210         struct ldlm_res_id res_id;
 211         ldlm_policy_data_t policy;
 212         struct lov_stripe_md *lsm;
 213         int flags = 0, err, rc = 0;
 214         ENTRY;
 215
 216         /* XXX for debug write replay without smfs and kml */
 217         res_id.name[0]= oa->o_id;
 218         res_id.name[1]= oa->o_gr;
 219         policy.l_extent.start = extent->start;
 220         policy.l_extent.end = extent->end;
 221
 222         /* get extent read lock on the source replay file */
 223         rc = ldlm_cli_enqueue(NULL, NULL, cache->obd_namespace, res_id,
 224                               LDLM_EXTENT, &policy, LCK_PR,
 225                               &flags, cache_blocking_ast, ldlm_completion_ast,
 226                               NULL, NULL, NULL, 0, NULL, &lockh_src);
 227         if (rc != ELDLM_OK)
 228                 RETURN(rc);
 229
 230         /* construct the pseudo lsm */
 231         rc = cmobd_dummy_lsm(&lsm, cmobd->master_desc.ld_tgt_count, oa,
 232                              (__u32)cmobd->master_desc.ld_default_stripe_size);
 233         if (rc)
 234                 GOTO(out_lock, rc);
 235
 236         rc = obd_enqueue(cmobd->master_exp, lsm, LDLM_EXTENT, &policy,
 237                          LCK_PW, &flags, master_blocking_ast,
 238                          ldlm_completion_ast, NULL,
 239                          NULL, 0, NULL, &lockh_dst);
 240         if (rc != ELDLM_OK)
 241                 GOTO(out_lsm, rc);
 242
 243         err = cmobd_replay_write(obd, oa, &policy.l_extent);
 244
 245         rc = obd_cancel(cmobd->master_exp, lsm, LCK_PW, &lockh_dst);
 246         if (rc)
 247                 GOTO(out_lsm, rc);
 248
 249         /* XXX in fact, I just want to cancel the only lockh_dst instantly. */
 250         rc = obd_cancel_unused(cmobd->master_exp, lsm, 0, NULL);
 251         if (err)
 252                 rc = err;
 253 out_lsm:
 254         cmobd_free_lsm(&lsm);
 255 out_lock:
 256         ldlm_lock_decref(&lockh_src, LCK_PR);
 257         RETURN(rc);
 258 }
 259
 260 static int cmobd_write_reint(struct obd_device *obd, void *rec)
 261 {
 262         struct obdo *oa = (struct obdo *)rec;
 263         struct cm_obd *cmobd = &obd->u.cm;
 264         struct ldlm_extent *extent = NULL;
 265         char *extents_buf = NULL;
 266         struct obd_device *cache;
 267         int rc = 0, ext_num = 0;
 268         unsigned long csb, ino;
 269         __u32 size = 0;
 270         ENTRY;
 271
 272         size = sizeof(csb);
 273         obd_get_info(cmobd->cache_exp, strlen("cache_sb") + 1,
 274                      "cache_sb", &size, &csb);
 275
 276         ino = *(int*)(&oa->o_inline[0]);
 277
 278         cache = cmobd->cache_exp->exp_obd;
 279         rc = fsfilt_get_ino_write_extents(cache, (struct super_block *)csb,
 280                                           ino, &extents_buf, &ext_num);
 281         if (rc)
 282                 GOTO(out, rc);
 283         extent = (struct ldlm_extent *)extents_buf;
 284         size = ext_num;
 285         while (extent && size --) {
 286                 rc = cmobd_write_extents(obd, oa, extent);
 287                 if (rc)
 288                         GOTO(out, rc);
 289                 extent ++;
 290         }
 291 out:
 292         if (extents_buf)
 293                 fsfilt_free_write_extents(cache, (struct super_block *)csb,
 294                                           ino, extents_buf, ext_num);
 295         RETURN(rc);
 296 }
 297
 298 int cmobd_reint_oss(struct obd_device *obd, void *record, int opcode)
 299 {
 300         switch (opcode) {
 301         case OST_CREATE:
 302                 return cmobd_create_reint(obd, record);
 303         case OST_SETATTR:
 304                 return cmobd_setattr_reint(obd, record);
 305         case OST_WRITE:
 306                 return cmobd_write_reint(obd, record);
 307         default:
 308                 CERROR("unrecognized oss reint opcode %d\n",
 309                        opcode);
 310                 return -EINVAL;
 311         }
 312 }