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         if (rc == -ENOSPC || rc == -EROFS) {
  70                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
  71                 if (body && rc == -ENOSPC) {
  72                         oscc->oscc_grow_count = OST_MIN_PRECREATE;
  73                         oscc->oscc_last_id = body->oa.o_id;
  74                 }
  75                 spin_unlock(&oscc->oscc_lock);
  76                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
  77         } else if (rc != 0 && rc != -EIO) {
  78                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
  79                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
  80                 spin_unlock(&oscc->oscc_lock);
  81                 DEBUG_REQ(D_ERROR, req,
  82                           "Unknown rc %d from async create: failing oscc", rc);
  83                 ptlrpc_fail_import(req->rq_import,
  84                                    lustre_msg_get_conn_cnt(req->rq_reqmsg));
  85         } else {
  86                 if (rc == 0) {
  87                          if (body) {
  88                                 int diff = body->oa.o_id - oscc->oscc_last_id;
  89
  90                                 if (diff < oscc->oscc_grow_count)
  91                                         oscc->oscc_grow_count =
  92                                                 max(diff/3, OST_MIN_PRECREATE);
  93                                 else
  94                                         oscc->oscc_flags &= ~OSCC_FLAG_LOW;
  95                                 oscc->oscc_last_id = body->oa.o_id;
  96                         }
  97                 } else {
  98                         /* filter always set body->oa.o_id as the last_id
  99                          * of filter (see filter_handle_precreate for detail)*/
 100                         if (body && body->oa.o_id > oscc->oscc_last_id)
 101                                 oscc->oscc_last_id = body->oa.o_id;
 102                 }
 103                 spin_unlock(&oscc->oscc_lock);
 104
 105         }
 106
 107         CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
 108                oscc->oscc_last_id, oscc->oscc_next_id);
 109
 110         cfs_waitq_signal(&oscc->oscc_waitq);
 111         RETURN(rc);
 112 }
 113
 114 static int oscc_internal_create(struct osc_creator *oscc)
 115 {
 116         struct ptlrpc_request *request;
 117         struct ost_body *body;
 118         int size[] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 119         ENTRY;
 120
 121         spin_lock(&oscc->oscc_lock);
 122         if (oscc->oscc_grow_count < OST_MAX_PRECREATE &&
 123             !(oscc->oscc_flags & (OSCC_FLAG_LOW | OSCC_FLAG_RECOVERING)) &&
 124             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 125                    (oscc->oscc_grow_count / 4 + 1)) {
 126                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 127                 oscc->oscc_grow_count *= 2;
 128         }
 129
 130         if (oscc->oscc_grow_count > OST_MAX_PRECREATE / 2)
 131                 oscc->oscc_grow_count = OST_MAX_PRECREATE / 2;
 132
 133         if (oscc->oscc_flags & OSCC_FLAG_CREATING ||
 134             oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 135                 spin_unlock(&oscc->oscc_lock);
 136                 RETURN(0);
 137         }
 138         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 139         spin_unlock(&oscc->oscc_lock);
 140
 141         request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
 142                                   LUSTRE_OST_VERSION, OST_CREATE, 2,
 143                                   size, NULL);
 144         if (request == NULL) {
 145                 spin_lock(&oscc->oscc_lock);
 146                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 147                 spin_unlock(&oscc->oscc_lock);
 148                 RETURN(-ENOMEM);
 149         }
 150
 151         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
 152         body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
 153
 154         spin_lock(&oscc->oscc_lock);
 155         body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
 156         body->oa.o_gr = oscc->oscc_oa.o_gr;
 157         LASSERT(body->oa.o_gr > 0);
 158         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 159         spin_unlock(&oscc->oscc_lock);
 160         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
 161                body->oa.o_id, oscc->oscc_last_id);
 162
 163         ptlrpc_req_set_repsize(request, 2, size);
 164
 165         request->rq_async_args.pointer_arg[0] = oscc;
 166         request->rq_interpret_reply = osc_interpret_create;
 167         ptlrpcd_add_req(request);
 168
 169         RETURN(0);
 170 }
 171
 172 static int oscc_has_objects(struct osc_creator *oscc, int count)
 173 {
 174         int have_objs;
 175         spin_lock(&oscc->oscc_lock);
 176         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 177         spin_unlock(&oscc->oscc_lock);
 178
 179         if (!have_objs)
 180                 oscc_internal_create(oscc);
 181
 182         return have_objs;
 183 }
 184
 185 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 186 {
 187         int have_objs;
 188         int ost_full;
 189         int osc_invalid;
 190
 191         have_objs = oscc_has_objects(oscc, count);
 192
 193         spin_lock(&oscc->oscc_lock);
 194         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 195         spin_unlock(&oscc->oscc_lock);
 196
 197         osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 198
 199         return have_objs || ost_full || osc_invalid;
 200 }
 201
 202 static int oscc_precreate(struct osc_creator *oscc, int wait)
 203 {
 204         struct l_wait_info lwi = { 0 };
 205         int rc = 0;
 206         ENTRY;
 207
 208         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 209                 RETURN(0);
 210
 211         if (!wait)
 212                 RETURN(0);
 213
 214         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 215         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 216
 217         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 218                 rc = -ENOSPC;
 219
 220         if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
 221                 rc = -EIO;
 222
 223         RETURN(rc);
 224 }
 225
 226 int oscc_recovering(struct osc_creator *oscc)
 227 {
 228         int recov = 0;
 229
 230         spin_lock(&oscc->oscc_lock);
 231         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 232         spin_unlock(&oscc->oscc_lock);
 233
 234         return recov;
 235 }
 236
 237 /* decide if the OST has remaining object, return value :
 238         0 : the OST has remaining object, and don't need to do precreate.
 239         1 : the OST has no remaining object, and will send a RPC for precreate.
 240         2 : the OST has no remaining object, and will not get any for
 241             a potentially very long time
 242      1000 : unusable
 243  */
 244 int osc_precreate(struct obd_export *exp, int need_create)
 245 {
 246         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 247         struct obd_import *imp = exp->exp_imp_reverse;
 248         ENTRY;
 249
 250         LASSERT(oscc != NULL);
 251         if (imp != NULL && imp->imp_deactive)
 252                 RETURN(1000);
 253
 254         if (oscc->oscc_last_id < oscc->oscc_next_id) {
 255                 spin_lock(&oscc->oscc_lock);
 256                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 257                         spin_unlock(&oscc->oscc_lock);
 258                         RETURN(1000);
 259                 }
 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_RECOVERING) {
 265                         spin_unlock(&oscc->oscc_lock);
 266                         RETURN(2);
 267                 }
 268                 spin_unlock(&oscc->oscc_lock);
 269
 270                 if (oscc->oscc_flags & OSCC_FLAG_CREATING)
 271                         RETURN(1);
 272
 273                 if (!need_create)
 274                         RETURN(1);
 275
 276                 oscc_internal_create(oscc);
 277                 RETURN(1);
 278         }
 279         RETURN(0);
 280 }
 281
 282 int osc_create(struct obd_export *exp, struct obdo *oa,
 283                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 284 {
 285         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 286         struct lov_stripe_md *lsm;
 287         int try_again = 1, rc = 0;
 288         ENTRY;
 289
 290         LASSERT(oa);
 291         LASSERT(ea);
 292         LASSERT(oa->o_gr > 0);
 293         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 294
 295         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 296             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 297                 RETURN(osc_real_create(exp, oa, ea, oti));
 298         }
 299
 300         if (oa->o_gr == FILTER_GROUP_LLOG || oa->o_gr == FILTER_GROUP_ECHO)
 301                 RETURN(osc_real_create(exp, oa, ea, oti));
 302
 303         /* this is the special case where create removes orphans */
 304         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 305             oa->o_flags == OBD_FL_DELORPHAN) {
 306                 spin_lock(&oscc->oscc_lock);
 307                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 308                         spin_unlock(&oscc->oscc_lock);
 309                         RETURN(-EBUSY);
 310                 }
 311                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 312                         spin_unlock(&oscc->oscc_lock);
 313                         RETURN(0);
 314                 }
 315                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 316                 spin_unlock(&oscc->oscc_lock);
 317                 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
 318                        oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
 319
 320                 /* delete from next_id on up */
 321                 oa->o_valid |= OBD_MD_FLID;
 322                 oa->o_id = oscc->oscc_next_id - 1;
 323
 324                 rc = osc_real_create(exp, oa, ea, NULL);
 325
 326                 spin_lock(&oscc->oscc_lock);
 327                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 328                 if (rc == 0 || rc == -ENOSPC) {
 329                         if (rc == -ENOSPC)
 330                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 331                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 332                         oscc->oscc_last_id = oa->o_id;
 333                         CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
 334                                LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
 335                                oscc->oscc_last_id, rc);
 336                         cfs_waitq_signal(&oscc->oscc_waitq);
 337                 } else {
 338                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 339                                oscc->oscc_obd->obd_name, rc);
 340                 }
 341                 spin_unlock(&oscc->oscc_lock);
 342
 343
 344                 RETURN(rc);
 345         }
 346
 347         lsm = *ea;
 348         if (lsm == NULL) {
 349                 rc = obd_alloc_memmd(exp, &lsm);
 350                 if (rc < 0)
 351                         RETURN(rc);
 352         }
 353
 354         while (try_again) {
 355                 /* If orphans are being recovered, then we must wait until
 356                    it is finished before we can continue with create. */
 357                 if (oscc_recovering(oscc)) {
 358                         struct l_wait_info lwi;
 359
 360                         CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
 361                                oscc->oscc_obd->obd_name);
 362
 363                         lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(obd_timeout/4)),
 364                                           NULL, NULL);
 365                         rc = l_wait_event(oscc->oscc_waitq,
 366                                           !oscc_recovering(oscc), &lwi);
 367                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 368                         if (rc == -ETIMEDOUT) {
 369                                 CDEBUG(D_HA,"%s: timeout waiting on recovery\n",
 370                                        oscc->oscc_obd->obd_name);
 371                                 RETURN(rc);
 372                         }
 373                         CDEBUG(D_HA, "%s: oscc recovery over, waking up\n",
 374                                oscc->oscc_obd->obd_name);
 375                 }
 376
 377                 spin_lock(&oscc->oscc_lock);
 378                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 379                         spin_unlock(&oscc->oscc_lock);
 380                         break;
 381                 }
 382
 383                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 384                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 385                         oa->o_id = oscc->oscc_next_id;
 386                         lsm->lsm_object_id = oscc->oscc_next_id;
 387                         *ea = lsm;
 388                         oscc->oscc_next_id++;
 389                         try_again = 0;
 390
 391                         CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
 392                                exp->exp_obd->obd_name, oscc->oscc_next_id);
 393                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 394                         rc = -ENOSPC;
 395                         spin_unlock(&oscc->oscc_lock);
 396                         break;
 397                 }
 398                 spin_unlock(&oscc->oscc_lock);
 399                 rc = oscc_precreate(oscc, try_again);
 400                 if (rc)
 401                         break;
 402         }
 403
 404         if (rc == 0)
 405                 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
 406                        obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
 407         else if (*ea == NULL)
 408                 obd_free_memmd(exp, &lsm);
 409         RETURN(rc);
 410 }
 411
 412 void oscc_init(struct obd_device *obd)
 413 {
 414         struct osc_creator *oscc;
 415
 416         if (obd == NULL)
 417                 return;
 418
 419         oscc = &obd->u.cli.cl_oscc;
 420
 421         memset(oscc, 0, sizeof(*oscc));
 422         CFS_INIT_LIST_HEAD(&oscc->oscc_list);
 423         cfs_waitq_init(&oscc->oscc_waitq);
 424         spin_lock_init(&oscc->oscc_lock);
 425         oscc->oscc_obd = obd;
 426         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 427
 428         oscc->oscc_next_id = 2;
 429         oscc->oscc_last_id = 1;
 430         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 431         /* XXX the export handle should give the oscc the last object */
 432         /* oed->oed_oscc.oscc_last_id = exph->....; */
 433 }