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  * 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 [sun.com URL with a
  20  * copy of GPLv2].
  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  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osc/osc_create.c
  37  * For testing and management it is treated as an obd_device,
  38  * although * it does not export a full OBD method table (the
  39  * requests are coming * in over the wire, so object target modules
  40  * do not have a full * method table.)
  41  *
  42  * Author: Peter Braam <braam@clusterfs.com>
  43  */
  44
  45 #ifndef EXPORT_SYMTAB
  46 # define EXPORT_SYMTAB
  47 #endif
  48 #define DEBUG_SUBSYSTEM S_OSC
  49
  50 #ifdef __KERNEL__
  51 # include <libcfs/libcfs.h>
  52 #else /* __KERNEL__ */
  53 # include <liblustre.h>
  54 #endif
  55
  56 #ifdef  __CYGWIN__
  57 # include <ctype.h>
  58 #endif
  59
  60 # include <lustre_dlm.h>
  61 #include <obd_class.h>
  62 #include "osc_internal.h"
  63
  64 static int osc_interpret_create(struct ptlrpc_request *req, void *data, int rc)
  65 {
  66         struct osc_creator *oscc;
  67         struct ost_body *body = NULL;
  68         ENTRY;
  69
  70         if (req->rq_repmsg) {
  71                 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
  72                                           lustre_swab_ost_body);
  73                 if (body == NULL && rc == 0)
  74                         rc = -EPROTO;
  75         }
  76
  77         oscc = req->rq_async_args.pointer_arg[0];
  78         LASSERT(oscc && (oscc->oscc_obd != LP_POISON));
  79
  80         spin_lock(&oscc->oscc_lock);
  81         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  82         switch (rc) {
  83         case 0: {
  84                 if (body) {
  85                         int diff = body->oa.o_id - oscc->oscc_last_id;
  86
  87                         if (diff < oscc->oscc_grow_count)
  88                                 oscc->oscc_grow_count =
  89                                         max(diff/3, OST_MIN_PRECREATE);
  90                         else
  91                                 oscc->oscc_flags &= ~OSCC_FLAG_LOW;
  92                         oscc->oscc_last_id = body->oa.o_id;
  93                 }
  94                 spin_unlock(&oscc->oscc_lock);
  95                 break;
  96         }
  97         case -EAGAIN:
  98                 /* valid race delorphan vs create, or somthing after resend */
  99                 spin_unlock(&oscc->oscc_lock);
 100                 DEBUG_REQ(D_INODE, req, "Got EGAIN - resend \n");
 101                 break;
 102         case -ENOSPC:
 103         case -EROFS:
 104         case -EFBIG: {
 105                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 106                 if (body && rc == -ENOSPC) {
 107                         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 108                         oscc->oscc_last_id = body->oa.o_id;
 109                 }
 110                 spin_unlock(&oscc->oscc_lock);
 111                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
 112                 break;
 113         }
 114         case -EIO: {
 115                 /* filter always set body->oa.o_id as the last_id
 116                  * of filter (see filter_handle_precreate for detail)*/
 117                 if (body && body->oa.o_id > oscc->oscc_last_id)
 118                         oscc->oscc_last_id = body->oa.o_id;
 119                 spin_unlock(&oscc->oscc_lock);
 120                 break;
 121         }
 122         default: {
 123                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 124                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 125                 spin_unlock(&oscc->oscc_lock);
 126                 DEBUG_REQ(D_ERROR, req,
 127                           "Unknown rc %d from async create: failing oscc", rc);
 128                 ptlrpc_fail_import(req->rq_import,
 129                                    lustre_msg_get_conn_cnt(req->rq_reqmsg));
 130         }
 131         }
 132
 133         CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
 134                oscc->oscc_last_id, oscc->oscc_next_id);
 135
 136         cfs_waitq_signal(&oscc->oscc_waitq);
 137         RETURN(rc);
 138 }
 139
 140 static int oscc_internal_create(struct osc_creator *oscc)
 141 {
 142         struct ptlrpc_request *request;
 143         struct ost_body *body;
 144         __u32 size[] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 145         ENTRY;
 146
 147         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 148
 149         if (oscc->oscc_flags & OSCC_FLAG_CREATING ||
 150             oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 151                 spin_unlock(&oscc->oscc_lock);
 152                 RETURN(0);
 153         }
 154
 155         if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
 156             ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
 157             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 158                    (oscc->oscc_grow_count / 4 + 1)) {
 159                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 160                 oscc->oscc_grow_count *= 2;
 161         }
 162
 163         if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
 164                 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
 165
 166         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 167         spin_unlock(&oscc->oscc_lock);
 168
 169         request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
 170                                   LUSTRE_OST_VERSION, OST_CREATE, 2,
 171                                   size, NULL);
 172         if (request == NULL) {
 173                 spin_lock(&oscc->oscc_lock);
 174                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 175                 spin_unlock(&oscc->oscc_lock);
 176                 RETURN(-ENOMEM);
 177         }
 178
 179         request->rq_request_portal = OST_CREATE_PORTAL;
 180         ptlrpc_at_set_req_timeout(request);
 181         body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
 182
 183         spin_lock(&oscc->oscc_lock);
 184         body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
 185         body->oa.o_gr = oscc->oscc_oa.o_gr;
 186         LASSERT(body->oa.o_gr > 0);
 187         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 188         spin_unlock(&oscc->oscc_lock);
 189         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
 190                body->oa.o_id, oscc->oscc_last_id);
 191
 192         ptlrpc_req_set_repsize(request, 2, size);
 193
 194         request->rq_async_args.pointer_arg[0] = oscc;
 195         request->rq_interpret_reply = osc_interpret_create;
 196         ptlrpcd_add_req(request);
 197
 198         RETURN(0);
 199 }
 200
 201 static int oscc_has_objects(struct osc_creator *oscc, int count)
 202 {
 203         int have_objs;
 204         spin_lock(&oscc->oscc_lock);
 205         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 206
 207         if (!have_objs) {
 208                 oscc_internal_create(oscc);
 209         } else {
 210                 spin_unlock(&oscc->oscc_lock);
 211         }
 212
 213         return have_objs;
 214 }
 215
 216 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 217 {
 218         int have_objs;
 219         int ost_full;
 220         int osc_invalid;
 221
 222         have_objs = oscc_has_objects(oscc, count);
 223
 224         spin_lock(&oscc->oscc_lock);
 225         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 226         spin_unlock(&oscc->oscc_lock);
 227
 228         osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 229
 230         return have_objs || ost_full || osc_invalid;
 231 }
 232
 233 static int oscc_precreate(struct osc_creator *oscc, int wait)
 234 {
 235         struct l_wait_info lwi = { 0 };
 236         int rc = 0;
 237         ENTRY;
 238
 239         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 240                 RETURN(0);
 241
 242         if (!wait)
 243                 RETURN(0);
 244
 245         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 246         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 247
 248         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 249                 rc = -ENOSPC;
 250
 251         if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
 252                 rc = -EIO;
 253
 254         RETURN(rc);
 255 }
 256
 257 int oscc_recovering(struct osc_creator *oscc)
 258 {
 259         int recov = 0;
 260
 261         spin_lock(&oscc->oscc_lock);
 262         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 263         spin_unlock(&oscc->oscc_lock);
 264
 265         return recov;
 266 }
 267
 268 /* decide if the OST has remaining object, return value :
 269         0 : the OST has remaining object, and don't need to do precreate.
 270         1 : the OST has no remaining object, and will send a RPC for precreate.
 271         2 : the OST has no remaining object, and will not get any for
 272             a potentially very long time
 273      1000 : unusable
 274  */
 275 int osc_precreate(struct obd_export *exp)
 276 {
 277         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 278         struct obd_import *imp = exp->exp_imp_reverse;
 279         ENTRY;
 280
 281         LASSERT(oscc != NULL);
 282         if (imp != NULL && imp->imp_deactive)
 283                 RETURN(1000);
 284
 285         if (oscc->oscc_last_id < oscc->oscc_next_id) {
 286                 spin_lock(&oscc->oscc_lock);
 287                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 288                         spin_unlock(&oscc->oscc_lock);
 289                         RETURN(1000);
 290                 }
 291                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 292                         spin_unlock(&oscc->oscc_lock);
 293                         RETURN(1);
 294                 }
 295                 if (oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 296                         spin_unlock(&oscc->oscc_lock);
 297                         RETURN(2);
 298                 }
 299
 300                 if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
 301                         spin_unlock(&oscc->oscc_lock);
 302                         RETURN(1);
 303                 }
 304
 305                 oscc_internal_create(oscc);
 306                 RETURN(1);
 307         }
 308         RETURN(0);
 309 }
 310
 311 int osc_create(struct obd_export *exp, struct obdo *oa,
 312                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 313 {
 314         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 315         struct lov_stripe_md *lsm;
 316         int try_again = 1, rc = 0;
 317         ENTRY;
 318
 319         LASSERT(oa);
 320         LASSERT(ea);
 321         LASSERT(oa->o_gr > 0);
 322         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 323
 324         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 325             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 326                 RETURN(osc_real_create(exp, oa, ea, oti));
 327         }
 328
 329         if (oa->o_gr == FILTER_GROUP_LLOG || oa->o_gr == FILTER_GROUP_ECHO)
 330                 RETURN(osc_real_create(exp, oa, ea, oti));
 331
 332         /* this is the special case where create removes orphans */
 333         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 334             oa->o_flags == OBD_FL_DELORPHAN) {
 335                 spin_lock(&oscc->oscc_lock);
 336                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 337                         spin_unlock(&oscc->oscc_lock);
 338                         RETURN(-EBUSY);
 339                 }
 340                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 341                         spin_unlock(&oscc->oscc_lock);
 342                         RETURN(0);
 343                 }
 344                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 345                 /* seting flag LOW we prevent extra grow precreate size
 346                  * and enforce use last assigned size */
 347                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 348                 spin_unlock(&oscc->oscc_lock);
 349                 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
 350                        oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
 351
 352                 /* delete from next_id on up */
 353                 oa->o_valid |= OBD_MD_FLID;
 354                 oa->o_id = oscc->oscc_next_id - 1;
 355
 356                 rc = osc_real_create(exp, oa, ea, NULL);
 357
 358                 spin_lock(&oscc->oscc_lock);
 359                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 360                 if (rc == 0 || rc == -ENOSPC) {
 361                         if (rc == -ENOSPC)
 362                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 363                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 364                         oscc->oscc_last_id = oa->o_id;
 365                         CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
 366                                LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
 367                                oscc->oscc_last_id, rc);
 368                         cfs_waitq_signal(&oscc->oscc_waitq);
 369                 } else {
 370                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 371                                oscc->oscc_obd->obd_name, rc);
 372                 }
 373                 spin_unlock(&oscc->oscc_lock);
 374
 375
 376                 RETURN(rc);
 377         }
 378
 379         lsm = *ea;
 380         if (lsm == NULL) {
 381                 rc = obd_alloc_memmd(exp, &lsm);
 382                 if (rc < 0)
 383                         RETURN(rc);
 384         }
 385
 386         while (try_again) {
 387                 /* If orphans are being recovered, then we must wait until
 388                    it is finished before we can continue with create. */
 389                 if (oscc_recovering(oscc)) {
 390                         struct l_wait_info lwi;
 391
 392                         CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
 393                                oscc->oscc_obd->obd_name);
 394
 395                         lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(
 396                                 obd_timeout / 4)), NULL, NULL);
 397                         rc = l_wait_event(oscc->oscc_waitq,
 398                                           !oscc_recovering(oscc), &lwi);
 399                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 400                         if (rc == -ETIMEDOUT) {
 401                                 CDEBUG(D_HA,"%s: timeout waiting on recovery\n",
 402                                        oscc->oscc_obd->obd_name);
 403                                 RETURN(rc);
 404                         }
 405                         CDEBUG(D_HA, "%s: oscc recovery over, waking up\n",
 406                                oscc->oscc_obd->obd_name);
 407                 }
 408
 409                 spin_lock(&oscc->oscc_lock);
 410                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 411                         spin_unlock(&oscc->oscc_lock);
 412                         break;
 413                 }
 414
 415                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 416                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 417                         oa->o_id = oscc->oscc_next_id;
 418                         lsm->lsm_object_id = oscc->oscc_next_id;
 419                         *ea = lsm;
 420                         oscc->oscc_next_id++;
 421                         try_again = 0;
 422
 423                         CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
 424                                exp->exp_obd->obd_name, oscc->oscc_next_id);
 425                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 426                         rc = -ENOSPC;
 427                         spin_unlock(&oscc->oscc_lock);
 428                         break;
 429                 }
 430                 spin_unlock(&oscc->oscc_lock);
 431                 rc = oscc_precreate(oscc, try_again);
 432                 if (rc)
 433                         break;
 434         }
 435
 436         if (rc == 0)
 437                 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
 438                        obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
 439         else if (*ea == NULL)
 440                 obd_free_memmd(exp, &lsm);
 441         RETURN(rc);
 442 }
 443
 444 void oscc_init(struct obd_device *obd)
 445 {
 446         struct osc_creator *oscc;
 447
 448         if (obd == NULL)
 449                 return;
 450
 451         oscc = &obd->u.cli.cl_oscc;
 452
 453         memset(oscc, 0, sizeof(*oscc));
 454         CFS_INIT_LIST_HEAD(&oscc->oscc_list);
 455         cfs_waitq_init(&oscc->oscc_waitq);
 456         spin_lock_init(&oscc->oscc_lock);
 457         oscc->oscc_obd = obd;
 458         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 459         oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
 460
 461         oscc->oscc_next_id = 2;
 462         oscc->oscc_last_id = 1;
 463         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 464         /* XXX the export handle should give the oscc the last object */
 465         /* oed->oed_oscc.oscc_last_id = exph->....; */
 466 }