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
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  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(const struct lu_env *env,
  65                                 struct ptlrpc_request *req, void *data, int rc)
  66 {
  67         struct osc_creator *oscc;
  68         struct ost_body *body = NULL;
  69         ENTRY;
  70
  71         if (req->rq_repmsg) {
  72                 body = lustre_swab_repbuf(req, REPLY_REC_OFF, sizeof(*body),
  73                                           lustre_swab_ost_body);
  74                 if (body == NULL && rc == 0)
  75                         rc = -EPROTO;
  76         }
  77
  78         oscc = req->rq_async_args.pointer_arg[0];
  79         LASSERT(oscc && (oscc->oscc_obd != LP_POISON));
  80
  81         spin_lock(&oscc->oscc_lock);
  82         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  83         switch (rc) {
  84         case 0: {
  85                 if (body) {
  86                         int diff = body->oa.o_id - oscc->oscc_last_id;
  87
  88                         /* oscc_internal_create() stores the original value of
  89                          * grow_count in rq_async_args.space[0].
  90                          * We can't compare against oscc_grow_count directly,
  91                          * because it may have been increased while the RPC
  92                          * is in flight, so we would always find ourselves
  93                          * having created fewer objects and decreasing the
  94                          * precreate request size.  b=18577 */
  95                         if (diff < (int) req->rq_async_args.space[0]) {
  96                                 /* the OST has not managed to create all the
  97                                  * objects we asked for */
  98                                 oscc->oscc_grow_count = max(diff,
  99                                                             OST_MIN_PRECREATE);
 100                                 /* don't bump grow_count next time */
 101                                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 102                         } else {
 103                                 /* the OST is able to keep up with the work,
 104                                  * we could consider increasing grow_count
 105                                  * next time if needed */
 106                                 oscc->oscc_flags &= ~OSCC_FLAG_LOW;
 107                         }
 108                         oscc->oscc_last_id = body->oa.o_id;
 109                 }
 110                 spin_unlock(&oscc->oscc_lock);
 111                 break;
 112         }
 113         case -EAGAIN:
 114                 /* valid race delorphan vs create, or somthing after resend */
 115                 spin_unlock(&oscc->oscc_lock);
 116                 DEBUG_REQ(D_INODE, req, "Got EAGAIN - resend \n");
 117                 break;
 118         case -ENOSPC:
 119         case -EROFS:
 120         case -EFBIG: {
 121                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 122                 if (body && rc == -ENOSPC) {
 123                         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 124                         oscc->oscc_last_id = body->oa.o_id;
 125                 }
 126                 spin_unlock(&oscc->oscc_lock);
 127                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
 128                 break;
 129         }
 130         case -EIO: {
 131                 /* filter always set body->oa.o_id as the last_id
 132                  * of filter (see filter_handle_precreate for detail)*/
 133                 if (body && body->oa.o_id > oscc->oscc_last_id)
 134                         oscc->oscc_last_id = body->oa.o_id;
 135                 spin_unlock(&oscc->oscc_lock);
 136                 break;
 137         }
 138         default: {
 139                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 140                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 141                 spin_unlock(&oscc->oscc_lock);
 142                 DEBUG_REQ(D_ERROR, req,
 143                           "Unknown rc %d from async create: failing oscc", rc);
 144                 ptlrpc_fail_import(req->rq_import,
 145                                    lustre_msg_get_conn_cnt(req->rq_reqmsg));
 146         }
 147         }
 148
 149         CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
 150                oscc->oscc_last_id, oscc->oscc_next_id);
 151
 152         cfs_waitq_signal(&oscc->oscc_waitq);
 153         RETURN(rc);
 154 }
 155
 156 static int oscc_internal_create(struct osc_creator *oscc)
 157 {
 158         struct ptlrpc_request *request;
 159         struct ost_body *body;
 160         __u32 size[] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 161         ENTRY;
 162
 163         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 164
 165         if (oscc->oscc_flags & OSCC_FLAG_CREATING ||
 166             oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 167                 spin_unlock(&oscc->oscc_lock);
 168                 RETURN(0);
 169         }
 170
 171         if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
 172             ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
 173             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 174                    (oscc->oscc_grow_count / 4 + 1)) {
 175                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 176                 oscc->oscc_grow_count *= 2;
 177         }
 178
 179         if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
 180                 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
 181
 182         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 183         spin_unlock(&oscc->oscc_lock);
 184
 185         request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
 186                                   LUSTRE_OST_VERSION, OST_CREATE, 2,
 187                                   size, NULL);
 188         if (request == NULL) {
 189                 spin_lock(&oscc->oscc_lock);
 190                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 191                 spin_unlock(&oscc->oscc_lock);
 192                 RETURN(-ENOMEM);
 193         }
 194
 195         request->rq_request_portal = OST_CREATE_PORTAL;
 196         ptlrpc_at_set_req_timeout(request);
 197         body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
 198
 199         spin_lock(&oscc->oscc_lock);
 200         body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
 201         body->oa.o_gr = oscc->oscc_oa.o_gr;
 202         LASSERT_MDS_GROUP(body->oa.o_gr);
 203         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 204         request->rq_async_args.space[0] = oscc->oscc_grow_count;
 205         spin_unlock(&oscc->oscc_lock);
 206         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
 207                body->oa.o_id, oscc->oscc_last_id);
 208
 209         ptlrpc_req_set_repsize(request, 2, size);
 210
 211         request->rq_async_args.pointer_arg[0] = oscc;
 212         request->rq_interpret_reply = osc_interpret_create;
 213         ptlrpcd_add_req(request, PSCOPE_OTHER);
 214
 215         RETURN(0);
 216 }
 217
 218 static int oscc_has_objects(struct osc_creator *oscc, int count)
 219 {
 220         int have_objs;
 221         spin_lock(&oscc->oscc_lock);
 222         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 223
 224         if (!have_objs) {
 225                 oscc_internal_create(oscc);
 226         } else {
 227                 spin_unlock(&oscc->oscc_lock);
 228         }
 229
 230         return have_objs;
 231 }
 232
 233 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 234 {
 235         int have_objs;
 236         int ost_full;
 237         int osc_invalid;
 238
 239         have_objs = oscc_has_objects(oscc, count);
 240
 241         spin_lock(&oscc->oscc_lock);
 242         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 243         spin_unlock(&oscc->oscc_lock);
 244
 245         osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 246
 247         return have_objs || ost_full || osc_invalid;
 248 }
 249
 250 static int oscc_precreate(struct osc_creator *oscc, int wait)
 251 {
 252         struct l_wait_info lwi = { 0 };
 253         int rc = 0;
 254         ENTRY;
 255
 256         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 257                 RETURN(0);
 258
 259         if (!wait)
 260                 RETURN(0);
 261
 262         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 263         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 264
 265         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 266                 rc = -ENOSPC;
 267
 268         if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
 269                 rc = -EIO;
 270
 271         RETURN(rc);
 272 }
 273
 274 int oscc_recovering(struct osc_creator *oscc)
 275 {
 276         int recov = 0;
 277
 278         spin_lock(&oscc->oscc_lock);
 279         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 280         spin_unlock(&oscc->oscc_lock);
 281
 282         return recov;
 283 }
 284
 285 /* decide if the OST has remaining object, return value :
 286         0 : the OST has remaining object, and don't need to do precreate.
 287         1 : the OST has no remaining object, and will send a RPC for precreate.
 288         2 : the OST has no remaining object, and will not get any for
 289             a potentially very long time
 290      1000 : unusable
 291  */
 292 int osc_precreate(struct obd_export *exp)
 293 {
 294         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 295         struct obd_import *imp = exp->exp_imp_reverse;
 296         ENTRY;
 297
 298         LASSERT(oscc != NULL);
 299         if (imp != NULL && imp->imp_deactive)
 300                 RETURN(1000);
 301
 302         if (oscc_recovering(oscc))
 303                 RETURN(2);
 304
 305         if (oscc->oscc_flags & OSCC_FLAG_NOSPC)
 306                 RETURN(1000);
 307
 308         if (oscc->oscc_last_id < oscc->oscc_next_id) {
 309                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS)
 310                         RETURN(1);
 311
 312                 spin_lock(&oscc->oscc_lock);
 313                 if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
 314                         spin_unlock(&oscc->oscc_lock);
 315                         RETURN(1);
 316                 }
 317
 318                 oscc_internal_create(oscc);
 319                 RETURN(1);
 320         }
 321         RETURN(0);
 322 }
 323
 324 int osc_create(struct obd_export *exp, struct obdo *oa,
 325                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 326 {
 327         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 328         struct obd_import  *imp  = exp->exp_obd->u.cli.cl_import;
 329         struct lov_stripe_md *lsm;
 330         int try_again = 1, rc = 0;
 331         ENTRY;
 332
 333         LASSERT(oa);
 334         LASSERT(ea);
 335         LASSERT_MDS_GROUP(oa->o_gr);
 336         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 337
 338         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 339             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 340                 RETURN(osc_real_create(exp, oa, ea, oti));
 341         }
 342
 343         if (oa->o_gr == FILTER_GROUP_LLOG || oa->o_gr == FILTER_GROUP_ECHO)
 344                 RETURN(osc_real_create(exp, oa, ea, oti));
 345
 346         /* this is the special case where create removes orphans */
 347         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 348             oa->o_flags == OBD_FL_DELORPHAN) {
 349                 spin_lock(&oscc->oscc_lock);
 350                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 351                         spin_unlock(&oscc->oscc_lock);
 352                         RETURN(-EBUSY);
 353                 }
 354                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 355                         spin_unlock(&oscc->oscc_lock);
 356                         RETURN(0);
 357                 }
 358                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 359                 /* seting flag LOW we prevent extra grow precreate size
 360                  * and enforce use last assigned size */
 361                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 362                 spin_unlock(&oscc->oscc_lock);
 363                 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
 364                        oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
 365
 366                 /* delete from next_id on up */
 367                 oa->o_valid |= OBD_MD_FLID;
 368                 oa->o_id = oscc->oscc_next_id - 1;
 369
 370                 rc = osc_real_create(exp, oa, ea, NULL);
 371
 372                 spin_lock(&oscc->oscc_lock);
 373                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 374                 if (rc == 0 || rc == -ENOSPC) {
 375                         struct obd_connect_data *ocd;
 376
 377                         if (rc == -ENOSPC)
 378                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 379                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 380
 381                         oscc->oscc_last_id = oa->o_id;
 382                         ocd = &imp->imp_connect_data;
 383                         if (ocd->ocd_connect_flags & OBD_CONNECT_SKIP_ORPHAN) {
 384                                 CDEBUG(D_HA, "%s: Skip orphan set, reset last "
 385                                        "objid\n", oscc->oscc_obd->obd_name);
 386                                 oscc->oscc_next_id = oa->o_id + 1;
 387                         }
 388
 389                         /* sanity check for next objid. see bug 17025 */
 390                         LASSERT(oscc->oscc_next_id == oa->o_id + 1);
 391
 392                         CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
 393                                LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
 394                                oscc->oscc_last_id, rc);
 395                         cfs_waitq_signal(&oscc->oscc_waitq);
 396                 } else {
 397                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 398                                oscc->oscc_obd->obd_name, rc);
 399                 }
 400                 spin_unlock(&oscc->oscc_lock);
 401
 402
 403                 RETURN(rc);
 404         }
 405
 406         lsm = *ea;
 407         if (lsm == NULL) {
 408                 rc = obd_alloc_memmd(exp, &lsm);
 409                 if (rc < 0)
 410                         RETURN(rc);
 411         }
 412
 413         while (try_again) {
 414                 /* If orphans are being recovered, then we must wait until
 415                    it is finished before we can continue with create. */
 416                 if (oscc_recovering(oscc)) {
 417                         struct l_wait_info lwi;
 418
 419                         CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
 420                                oscc->oscc_obd->obd_name);
 421
 422                         lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(
 423                                 obd_timeout / 4)), NULL, NULL);
 424                         rc = l_wait_event(oscc->oscc_waitq,
 425                                           !oscc_recovering(oscc), &lwi);
 426                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 427                         if (rc == -ETIMEDOUT) {
 428                                 CDEBUG(D_HA,"%s: timeout waiting on recovery\n",
 429                                        oscc->oscc_obd->obd_name);
 430                                 RETURN(rc);
 431                         }
 432                         CDEBUG(D_HA, "%s: oscc recovery over, waking up\n",
 433                                oscc->oscc_obd->obd_name);
 434                 }
 435
 436                 spin_lock(&oscc->oscc_lock);
 437                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 438                         spin_unlock(&oscc->oscc_lock);
 439                         break;
 440                 }
 441
 442                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 443                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 444                         oa->o_id = oscc->oscc_next_id;
 445                         lsm->lsm_object_id = oscc->oscc_next_id;
 446                         *ea = lsm;
 447                         oscc->oscc_next_id++;
 448                         try_again = 0;
 449
 450                         CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
 451                                exp->exp_obd->obd_name, oscc->oscc_next_id);
 452                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 453                         rc = -ENOSPC;
 454                         spin_unlock(&oscc->oscc_lock);
 455                         break;
 456                 }
 457                 spin_unlock(&oscc->oscc_lock);
 458                 rc = oscc_precreate(oscc, try_again);
 459                 if (rc)
 460                         break;
 461         }
 462
 463         if (rc == 0)
 464                 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
 465                        obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
 466         else if (*ea == NULL)
 467                 obd_free_memmd(exp, &lsm);
 468         RETURN(rc);
 469 }
 470
 471 void oscc_init(struct obd_device *obd)
 472 {
 473         struct osc_creator *oscc;
 474
 475         if (obd == NULL)
 476                 return;
 477
 478         oscc = &obd->u.cli.cl_oscc;
 479
 480         memset(oscc, 0, sizeof(*oscc));
 481         CFS_INIT_LIST_HEAD(&oscc->oscc_list);
 482         cfs_waitq_init(&oscc->oscc_waitq);
 483         spin_lock_init(&oscc->oscc_lock);
 484         oscc->oscc_obd = obd;
 485         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 486         oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
 487
 488         oscc->oscc_next_id = 2;
 489         oscc->oscc_last_id = 1;
 490         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 491         /* XXX the export handle should give the oscc the last object */
 492         /* oed->oed_oscc.oscc_last_id = exph->....; */
 493 }