lustre/osc/osc_create.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2001-2003 Cluster File Systems, Inc.
   5  *   Author Peter Braam <braam@clusterfs.com>
   6  *
   7  *   This file is part of the Lustre file system, http://www.lustre.org
   8  *   Lustre is a trademark of Cluster File Systems, Inc.
   9  *
  10  *   You may have signed or agreed to another license before downloading
  11  *   this software.  If so, you are bound by the terms and conditions
  12  *   of that agreement, and the following does not apply to you.  See the
  13  *   LICENSE file included with this distribution for more information.
  14  *
  15  *   If you did not agree to a different license, then this copy of Lustre
  16  *   is open source software; you can redistribute it and/or modify it
  17  *   under the terms of version 2 of the GNU General Public License as
  18  *   published by the Free Software Foundation.
  19  *
  20  *   In either case, Lustre is distributed in the hope that it will be
  21  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  22  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  23  *   license text for more details.
  24  *
  25  *  For testing and management it is treated as an obd_device,
  26  *  although * it does not export a full OBD method table (the
  27  *  requests are coming * in over the wire, so object target modules
  28  *  do not have a full * method table.)
  29  *
  30  */
  31
  32 #ifndef EXPORT_SYMTAB
  33 # define EXPORT_SYMTAB
  34 #endif
  35 #define DEBUG_SUBSYSTEM S_OSC
  36
  37 #ifdef __KERNEL__
  38 # include <libcfs/libcfs.h>
  39 #else /* __KERNEL__ */
  40 # include <liblustre.h>
  41 #endif
  42
  43 #ifdef  __CYGWIN__
  44 # include <ctype.h>
  45 #endif
  46
  47 # include <lustre_dlm.h>
  48 #include <obd_class.h>
  49 #include "osc_internal.h"
  50
  51 static int osc_interpret_create(struct ptlrpc_request *req, void *data, int rc)
  52 {
  53         struct osc_creator *oscc;
  54         struct ost_body *body = NULL;
  55         ENTRY;
  56
  57         if (req->rq_repmsg) {
  58                 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
  59                                           lustre_swab_ost_body);
  60                 if (body == NULL && rc == 0)
  61                         rc = -EPROTO;
  62         }
  63
  64         oscc = req->rq_async_args.pointer_arg[0];
  65         LASSERT(oscc && (oscc->oscc_obd != LP_POISON));
  66
  67         spin_lock(&oscc->oscc_lock);
  68         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  69         switch (rc) {
  70         case 0: {
  71                 if (body) {
  72                         int diff = body->oa.o_id - oscc->oscc_last_id;
  73
  74                         if (diff < oscc->oscc_grow_count)
  75                                 oscc->oscc_grow_count =
  76                                         max(diff/3, OST_MIN_PRECREATE);
  77                         else
  78                                 oscc->oscc_flags &= ~OSCC_FLAG_LOW;
  79                         oscc->oscc_last_id = body->oa.o_id;
  80                 }
  81                 spin_unlock(&oscc->oscc_lock);
  82                 break;
  83         }
  84         case -ENOSPC:
  85         case -EROFS: {
  86                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
  87                 if (body && rc == -ENOSPC) {
  88                         oscc->oscc_grow_count = OST_MIN_PRECREATE;
  89                         oscc->oscc_last_id = body->oa.o_id;
  90                 }
  91                 spin_unlock(&oscc->oscc_lock);
  92                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
  93                 break;
  94         }
  95         case -EIO: {
  96                 /* filter always set body->oa.o_id as the last_id
  97                  * of filter (see filter_handle_precreate for detail)*/
  98                 if (body && body->oa.o_id > oscc->oscc_last_id)
  99                         oscc->oscc_last_id = body->oa.o_id;
 100                 spin_unlock(&oscc->oscc_lock);
 101                 break;
 102         }
 103         default: {
 104                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 105                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 106                 spin_unlock(&oscc->oscc_lock);
 107                 DEBUG_REQ(D_ERROR, req,
 108                           "unknown rc %d from async create: failing oscc", rc);
 109                 ptlrpc_fail_import(req->rq_import,
 110                                    lustre_msg_get_conn_cnt(req->rq_reqmsg));
 111         }
 112         }
 113
 114         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64", next to use "LPU64"\n",
 115                oscc->oscc_last_id, oscc->oscc_next_id);
 116
 117         cfs_waitq_signal(&oscc->oscc_waitq);
 118         RETURN(rc);
 119 }
 120
 121 static int oscc_internal_create(struct osc_creator *oscc)
 122 {
 123         struct ptlrpc_request *request;
 124         struct ost_body *body;
 125         int size[] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 126         ENTRY;
 127
 128         spin_lock(&oscc->oscc_lock);
 129         if (oscc->oscc_grow_count < OST_MAX_PRECREATE &&
 130             !(oscc->oscc_flags & (OSCC_FLAG_LOW | OSCC_FLAG_RECOVERING)) &&
 131             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 132                    (oscc->oscc_grow_count / 4 + 1)) {
 133                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 134                 oscc->oscc_grow_count *= 2;
 135         }
 136
 137         if (oscc->oscc_grow_count > OST_MAX_PRECREATE / 2)
 138                 oscc->oscc_grow_count = OST_MAX_PRECREATE / 2;
 139
 140         if (oscc->oscc_flags & OSCC_FLAG_CREATING ||
 141             oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 142                 spin_unlock(&oscc->oscc_lock);
 143                 RETURN(0);
 144         }
 145         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 146         spin_unlock(&oscc->oscc_lock);
 147
 148         request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
 149                                   LUSTRE_OST_VERSION, OST_CREATE, 2,
 150                                   size, NULL);
 151         if (request == NULL) {
 152                 spin_lock(&oscc->oscc_lock);
 153                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 154                 spin_unlock(&oscc->oscc_lock);
 155                 RETURN(-ENOMEM);
 156         }
 157
 158         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
 159         body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
 160
 161         spin_lock(&oscc->oscc_lock);
 162         body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
 163         body->oa.o_valid |= OBD_MD_FLID;
 164         spin_unlock(&oscc->oscc_lock);
 165         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
 166                body->oa.o_id, oscc->oscc_last_id);
 167
 168         ptlrpc_req_set_repsize(request, 2, size);
 169
 170         request->rq_async_args.pointer_arg[0] = oscc;
 171         request->rq_interpret_reply = osc_interpret_create;
 172         ptlrpcd_add_req(request);
 173
 174         RETURN(0);
 175 }
 176
 177 static int oscc_has_objects(struct osc_creator *oscc, int count)
 178 {
 179         int have_objs;
 180         spin_lock(&oscc->oscc_lock);
 181         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 182         spin_unlock(&oscc->oscc_lock);
 183
 184         if (!have_objs)
 185                 oscc_internal_create(oscc);
 186
 187         return have_objs;
 188 }
 189
 190 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 191 {
 192         int have_objs;
 193         int ost_full;
 194         int osc_invalid;
 195
 196         have_objs = oscc_has_objects(oscc, count);
 197
 198         spin_lock(&oscc->oscc_lock);
 199         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 200         spin_unlock(&oscc->oscc_lock);
 201
 202         osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 203
 204         return have_objs || ost_full || osc_invalid;
 205 }
 206
 207 static int oscc_precreate(struct osc_creator *oscc, int wait)
 208 {
 209         struct l_wait_info lwi = { 0 };
 210         int rc = 0;
 211         ENTRY;
 212
 213         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 214                 RETURN(0);
 215
 216         if (!wait)
 217                 RETURN(0);
 218
 219         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 220         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 221
 222         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 223                 rc = -ENOSPC;
 224
 225         if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
 226                 rc = -EIO;
 227
 228         RETURN(rc);
 229 }
 230
 231 int oscc_recovering(struct osc_creator *oscc)
 232 {
 233         int recov = 0;
 234
 235         spin_lock(&oscc->oscc_lock);
 236         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 237         spin_unlock(&oscc->oscc_lock);
 238
 239         return recov;
 240 }
 241
 242 /* decide if the OST has remaining object, return value :
 243         0 : the OST has remaining object, and don't need to do precreate.
 244         1 : the OST has no remaining object, and will send a RPC for precreate.
 245         2 : the OST has no remaining object, and will not get any for
 246             a potentially very long time
 247  */
 248 int osc_precreate(struct obd_export *exp, int need_create)
 249 {
 250         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 251         struct obd_import *imp = exp->exp_imp_reverse;
 252         ENTRY;
 253
 254         LASSERT(oscc != NULL);
 255         if (imp != NULL && imp->imp_deactive)
 256                 RETURN(2);
 257
 258         if (oscc->oscc_last_id < oscc->oscc_next_id) {
 259                 spin_lock(&oscc->oscc_lock);
 260                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 261                         spin_unlock(&oscc->oscc_lock);
 262                         RETURN(1);
 263                 }
 264                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC ||
 265                     oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 266                         spin_unlock(&oscc->oscc_lock);
 267                         RETURN(2);
 268                 }
 269                 spin_unlock(&oscc->oscc_lock);
 270
 271                 if (oscc->oscc_flags & OSCC_FLAG_CREATING)
 272                         RETURN(1);
 273
 274                 if (!need_create)
 275                         RETURN(1);
 276
 277                 oscc_internal_create(oscc);
 278                 RETURN(1);
 279         }
 280         RETURN(0);
 281 }
 282
 283 int osc_create(struct obd_export *exp, struct obdo *oa,
 284                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 285 {
 286         struct lov_stripe_md *lsm;
 287         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 288         int try_again = 1, rc = 0;
 289         ENTRY;
 290         LASSERT(oa);
 291         LASSERT(ea);
 292
 293         if ((oa->o_valid & OBD_MD_FLGROUP) && (oa->o_gr != 0))
 294                 RETURN(osc_real_create(exp, oa, ea, oti));
 295
 296         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 297             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 298                 RETURN(osc_real_create(exp, oa, ea, oti));
 299         }
 300
 301         /* this is the special case where create removes orphans */
 302         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 303             oa->o_flags == OBD_FL_DELORPHAN) {
 304                 spin_lock(&oscc->oscc_lock);
 305                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 306                         spin_unlock(&oscc->oscc_lock);
 307                         RETURN(-EBUSY);
 308                 }
 309                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 310                         spin_unlock(&oscc->oscc_lock);
 311                         RETURN(0);
 312                 }
 313                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 314                 spin_unlock(&oscc->oscc_lock);
 315                 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
 316                        oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
 317
 318                 /* delete from next_id on up */
 319                 oa->o_valid |= OBD_MD_FLID;
 320                 oa->o_id = oscc->oscc_next_id - 1;
 321
 322                 rc = osc_real_create(exp, oa, ea, NULL);
 323
 324                 spin_lock(&oscc->oscc_lock);
 325                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 326                 if (rc == 0 || rc == -ENOSPC) {
 327                         if (rc == -ENOSPC)
 328                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 329                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 330                         oscc->oscc_last_id = oa->o_id;
 331                         CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
 332                                LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
 333                                oscc->oscc_last_id, rc);
 334                         cfs_waitq_signal(&oscc->oscc_waitq);
 335                 } else {
 336                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 337                                oscc->oscc_obd->obd_name, rc);
 338                 }
 339                 spin_unlock(&oscc->oscc_lock);
 340
 341
 342                 RETURN(rc);
 343         }
 344
 345         lsm = *ea;
 346         if (lsm == NULL) {
 347                 rc = obd_alloc_memmd(exp, &lsm);
 348                 if (rc < 0)
 349                         RETURN(rc);
 350         }
 351
 352         while (try_again) {
 353                 /* If orphans are being recovered, then we must wait until
 354                    it is finished before we can continue with create. */
 355                 if (oscc_recovering(oscc)) {
 356                         struct l_wait_info lwi;
 357
 358                         CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
 359                                oscc->oscc_obd->obd_name);
 360
 361                         lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(obd_timeout/4)),
 362                                           NULL, NULL);
 363                         rc = l_wait_event(oscc->oscc_waitq,
 364                                           !oscc_recovering(oscc), &lwi);
 365                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 366                         if (rc == -ETIMEDOUT) {
 367                                 CDEBUG(D_HA,"%s: timeout waiting on recovery\n",
 368                                        oscc->oscc_obd->obd_name);
 369                                 RETURN(rc);
 370                         }
 371                         CDEBUG(D_HA, "%s: oscc recovery over, waking up\n",
 372                                oscc->oscc_obd->obd_name);
 373                 }
 374
 375                 spin_lock(&oscc->oscc_lock);
 376                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 377                         spin_unlock(&oscc->oscc_lock);
 378                         break;
 379                 }
 380
 381                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 382                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 383                         oa->o_id = oscc->oscc_next_id;
 384                         lsm->lsm_object_id = oscc->oscc_next_id;
 385                         *ea = lsm;
 386                         oscc->oscc_next_id++;
 387                         try_again = 0;
 388
 389                         CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
 390                                exp->exp_obd->obd_name, oscc->oscc_next_id);
 391                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 392                         rc = -ENOSPC;
 393                         spin_unlock(&oscc->oscc_lock);
 394                         break;
 395                 }
 396                 spin_unlock(&oscc->oscc_lock);
 397                 rc = oscc_precreate(oscc, try_again);
 398                 if (rc)
 399                         break;
 400         }
 401
 402         if (rc == 0)
 403                 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
 404                        obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
 405         else if (*ea == NULL)
 406                 obd_free_memmd(exp, &lsm);
 407         RETURN(rc);
 408 }
 409
 410 void oscc_init(struct obd_device *obd)
 411 {
 412         struct osc_creator *oscc;
 413
 414         if (obd == NULL)
 415                 return;
 416
 417         oscc = &obd->u.cli.cl_oscc;
 418
 419         memset(oscc, 0, sizeof(*oscc));
 420         CFS_INIT_LIST_HEAD(&oscc->oscc_list);
 421         cfs_waitq_init(&oscc->oscc_waitq);
 422         spin_lock_init(&oscc->oscc_lock);
 423         oscc->oscc_obd = obd;
 424         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 425
 426         oscc->oscc_next_id = 2;
 427         oscc->oscc_last_id = 1;
 428         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 429         /* XXX the export handle should give the oscc the last object */
 430         /* oed->oed_oscc.oscc_last_id = exph->....; */
 431 }