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 obdo *oa = (struct obdo *)rec;
 103         struct cm_obd *cmobd = &obd->u.cm;
 104         struct obd_export *exp = cmobd->master_exp;
 105         struct lov_stripe_md *lsm;
 106         struct obd_trans_info oti = { 0 };
 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                 rc = obd_set_info(exp, strlen("mds_conn"), "mds_conn",
 118                                   valsize, &group);
 119                 if (rc)
 120                         GOTO(out, rc = -EINVAL);
 121                 cmobd->master_group = oa->o_gr;
 122         }
 123         rc = obd_create(exp, oa, NULL, 0, &lsm, &oti);
 124
 125         cmobd_free_lsm(&lsm);
 126 out:
 127         RETURN(rc);
 128 }
 129
 130 /* direct cut-n-paste of filter_blocking_ast() */
 131 static int cache_blocking_ast(struct ldlm_lock *lock,
 132                               struct ldlm_lock_desc *desc,
 133                               void *data, int flag)
 134 {
 135         int do_ast;
 136         ENTRY;
 137
 138         if (flag == LDLM_CB_CANCELING) {
 139                 /* Don't need to do anything here. */
 140                 RETURN(0);
 141         }
 142
 143         /* XXX layering violation!  -phil */
 144         lock_res_and_lock(lock);
 145
 146         /* Get this: if filter_blocking_ast is racing with ldlm_intent_policy,
 147          * such that filter_blocking_ast is called just before l_i_p takes the
 148          * ns_lock, then by the time we get the lock, we might not be the
 149          * correct blocking function anymore.  So check, and return early, if
 150          * so. */
 151         if (lock->l_blocking_ast != cache_blocking_ast) {
 152                 unlock_res_and_lock(lock);
 153                 RETURN(0);
 154         }
 155
 156         lock->l_flags |= LDLM_FL_CBPENDING;
 157         do_ast = (!lock->l_readers && !lock->l_writers);
 158         unlock_res_and_lock(lock);
 159
 160         if (do_ast) {
 161                 struct lustre_handle lockh;
 162                 int rc;
 163
 164                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 165                 ldlm_lock2handle(lock, &lockh);
 166                 rc = ldlm_cli_cancel(&lockh);
 167                 if (rc < 0)
 168                         CERROR("ldlm_cli_cancel: %d\n", rc);
 169         } else {
 170                 LDLM_DEBUG(lock, "Lock still has references, will be "
 171                            "cancelled later");
 172         }
 173         RETURN(0);
 174 }
 175
 176 static int master_blocking_ast(struct ldlm_lock *lock,
 177                                struct ldlm_lock_desc *desc,
 178                                void *data, int flag)
 179 {
 180         int rc;
 181         struct lustre_handle lockh;
 182         ENTRY;
 183
 184         switch (flag) {
 185         case LDLM_CB_BLOCKING:
 186                 ldlm_lock2handle(lock, &lockh);
 187                 rc = ldlm_cli_cancel(&lockh);
 188                 if (rc < 0) {
 189                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
 190                         RETURN(rc);
 191                 }
 192                 break;
 193         case LDLM_CB_CANCELING:
 194                 /* do nothing here by now */
 195                 break;
 196         default:
 197                 LBUG();
 198         }
 199         RETURN(0);
 200 }
 201
 202 static int cmobd_write_extents(struct obd_device *obd, struct obdo *oa,
 203                                struct ldlm_extent *extent)
 204 {
 205         struct cm_obd *cmobd = &obd->u.cm;
 206         struct obd_device *cache = cmobd->cache_exp->exp_obd;
 207         struct lustre_handle lockh_src = { 0 };
 208         struct lustre_handle lockh_dst = { 0 };
 209         struct ldlm_res_id res_id;
 210         ldlm_policy_data_t policy;
 211         struct lov_stripe_md *lsm;
 212         int flags = 0, err, rc = 0;
 213         ENTRY;
 214
 215         /* XXX for debug write replay without smfs and kml */
 216         res_id.name[0]= oa->o_id;
 217         res_id.name[1]= oa->o_gr;
 218         policy.l_extent.start = extent->start;
 219         policy.l_extent.end = extent->end;
 220
 221         /* get extent read lock on the source replay file */
 222         rc = ldlm_cli_enqueue(NULL, NULL, cache->obd_namespace, res_id,
 223                               LDLM_EXTENT, &policy, LCK_PR,
 224                               &flags, cache_blocking_ast, ldlm_completion_ast,
 225                               NULL, NULL, NULL, 0, NULL, &lockh_src);
 226         if (rc != ELDLM_OK)
 227                 RETURN(rc);
 228
 229         /* construct the pseudo lsm */
 230
 231         /*
 232          * it is not good to access lov fields like @desc directly. This is
 233          * layering violation. It should be accessed via some interface method,
 234          * like llite does. --umka
 235          */
 236         rc = cmobd_dummy_lsm(&lsm, cmobd->master_desc.ld_tgt_count, oa,
 237                              (__u32)cmobd->master_desc.ld_default_stripe_size);
 238         if (rc)
 239                 GOTO(out_lock, rc);
 240
 241         rc = obd_enqueue(cmobd->master_exp, lsm, LDLM_EXTENT, &policy,
 242                          LCK_PW, &flags, master_blocking_ast,
 243                          ldlm_completion_ast, NULL,
 244                          NULL, 0, NULL, &lockh_dst);
 245         if (rc != ELDLM_OK)
 246                 GOTO(out_lsm, rc);
 247
 248         err = cmobd_replay_write(obd, oa, &policy.l_extent);
 249
 250         rc = obd_cancel(cmobd->master_exp, lsm, LCK_PW, &lockh_dst);
 251         if (rc)
 252                 GOTO(out_lsm, rc);
 253         /* XXX in fact, I just want to cancel the only lockh_dst
 254          *     instantly. */
 255         rc = obd_cancel_unused(cmobd->master_exp, lsm, 0, NULL);
 256         if (err)
 257                 rc = err;
 258 out_lsm:
 259         cmobd_free_lsm(&lsm);
 260 out_lock:
 261         ldlm_lock_decref(&lockh_src, LCK_PR);
 262         RETURN(rc);
 263 }
 264
 265 static int cmobd_write_reint(struct obd_device *obd, void *rec)
 266 {
 267         struct cm_obd *cmobd = &obd->u.cm;
 268         struct obdo *oa = (struct obdo *)rec;
 269         struct ldlm_extent *extent = NULL;
 270         char *extents_buf = NULL;
 271         struct obd_device *cache;
 272         int rc = 0, ext_num = 0;
 273         unsigned long csb, ino;
 274         __u32 size = 0;
 275         ENTRY;
 276
 277         size = sizeof(csb);
 278         obd_get_info(cmobd->cache_exp, strlen("cache_sb") + 1,
 279                      "cache_sb", &size, &csb);
 280
 281         ino = *(int*)(&oa->o_inline[0]);
 282
 283         cache = cmobd->cache_exp->exp_obd;
 284         rc = fsfilt_get_ino_write_extents(cache, (struct super_block *)csb,
 285                                           ino, &extents_buf, &ext_num);
 286         if (rc)
 287                 GOTO(out, rc);
 288         extent = (struct ldlm_extent *)extents_buf;
 289         size = ext_num;
 290         while (extent && size --) {
 291                 rc = cmobd_write_extents(obd, oa, extent);
 292                 if (rc)
 293                         GOTO(out, rc);
 294                 extent ++;
 295         }
 296 out:
 297         if (extents_buf)
 298                 fsfilt_free_write_extents(cache, (struct super_block *)csb,
 299                                           ino, extents_buf, ext_num);
 300         RETURN(rc);
 301 }
 302
 303 int cmobd_reint_oss(struct obd_device *obd, void *record, int opcode)
 304 {
 305         switch (opcode) {
 306         case OST_CREATE:
 307                 return cmobd_create_reint(obd, record);
 308         case OST_SETATTR:
 309                 return cmobd_setattr_reint(obd, record);
 310         case OST_WRITE:
 311                 return cmobd_write_reint(obd, record);
 312         default:
 313                 CERROR("unrecognized oss reint opcode %d\n",
 314                        opcode);
 315                 return -EINVAL;
 316         }
 317 }