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                         if (diff < oscc->oscc_grow_count)
  89                                 oscc->oscc_grow_count =
  90                                         max(diff/3, OST_MIN_PRECREATE);
  91                         else
  92                                 oscc->oscc_flags &= ~OSCC_FLAG_LOW;
  93                         oscc->oscc_last_id = body->oa.o_id;
  94                 }
  95                 spin_unlock(&oscc->oscc_lock);
  96                 break;
  97         }
  98         case -EAGAIN:
  99                 /* valid race delorphan vs create, or somthing after resend */
 100                 spin_unlock(&oscc->oscc_lock);
 101                 DEBUG_REQ(D_INODE, req, "Got EAGAIN - resend \n");
 102                 break;
 103         case -ENOSPC:
 104         case -EROFS:
 105         case -EFBIG: {
 106                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 107                 if (body && rc == -ENOSPC) {
 108                         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 109                         oscc->oscc_last_id = body->oa.o_id;
 110                 }
 111                 spin_unlock(&oscc->oscc_lock);
 112                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
 113                 break;
 114         }
 115         case -EIO: {
 116                 /* filter always set body->oa.o_id as the last_id
 117                  * of filter (see filter_handle_precreate for detail)*/
 118                 if (body && body->oa.o_id > oscc->oscc_last_id)
 119                         oscc->oscc_last_id = body->oa.o_id;
 120                 spin_unlock(&oscc->oscc_lock);
 121                 break;
 122         }
 123         default: {
 124                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 125                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
 126                 spin_unlock(&oscc->oscc_lock);
 127                 DEBUG_REQ(D_ERROR, req,
 128                           "Unknown rc %d from async create: failing oscc", rc);
 129                 ptlrpc_fail_import(req->rq_import,
 130                                    lustre_msg_get_conn_cnt(req->rq_reqmsg));
 131         }
 132         }
 133
 134         CDEBUG(D_HA, "preallocated through id "LPU64" (next to use "LPU64")\n",
 135                oscc->oscc_last_id, oscc->oscc_next_id);
 136
 137         cfs_waitq_signal(&oscc->oscc_waitq);
 138         RETURN(rc);
 139 }
 140
 141 static int oscc_internal_create(struct osc_creator *oscc)
 142 {
 143         struct ptlrpc_request *request;
 144         struct ost_body *body;
 145         __u32 size[] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 146         ENTRY;
 147
 148         LASSERT_SPIN_LOCKED(&oscc->oscc_lock);
 149
 150         if (oscc->oscc_flags & OSCC_FLAG_CREATING ||
 151             oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 152                 spin_unlock(&oscc->oscc_lock);
 153                 RETURN(0);
 154         }
 155
 156         if (oscc->oscc_grow_count < oscc->oscc_max_grow_count &&
 157             ((oscc->oscc_flags & OSCC_FLAG_LOW) == 0) &&
 158             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 159                    (oscc->oscc_grow_count / 4 + 1)) {
 160                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 161                 oscc->oscc_grow_count *= 2;
 162         }
 163
 164         if (oscc->oscc_grow_count > oscc->oscc_max_grow_count / 2)
 165                 oscc->oscc_grow_count = oscc->oscc_max_grow_count / 2;
 166
 167         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 168         spin_unlock(&oscc->oscc_lock);
 169
 170         request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
 171                                   LUSTRE_OST_VERSION, OST_CREATE, 2,
 172                                   size, NULL);
 173         if (request == NULL) {
 174                 spin_lock(&oscc->oscc_lock);
 175                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 176                 spin_unlock(&oscc->oscc_lock);
 177                 RETURN(-ENOMEM);
 178         }
 179
 180         request->rq_request_portal = OST_CREATE_PORTAL;
 181         ptlrpc_at_set_req_timeout(request);
 182         body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
 183
 184         spin_lock(&oscc->oscc_lock);
 185         body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
 186         body->oa.o_gr = oscc->oscc_oa.o_gr;
 187         LASSERT(body->oa.o_gr > 0);
 188         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 189         spin_unlock(&oscc->oscc_lock);
 190         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
 191                body->oa.o_id, oscc->oscc_last_id);
 192
 193         ptlrpc_req_set_repsize(request, 2, size);
 194
 195         request->rq_async_args.pointer_arg[0] = oscc;
 196         request->rq_interpret_reply = osc_interpret_create;
 197         ptlrpcd_add_req(request);
 198
 199         RETURN(0);
 200 }
 201
 202 static int oscc_has_objects(struct osc_creator *oscc, int count)
 203 {
 204         int have_objs;
 205         spin_lock(&oscc->oscc_lock);
 206         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 207
 208         if (!have_objs) {
 209                 oscc_internal_create(oscc);
 210         } else {
 211                 spin_unlock(&oscc->oscc_lock);
 212         }
 213
 214         return have_objs;
 215 }
 216
 217 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 218 {
 219         int have_objs;
 220         int ost_full;
 221         int osc_invalid;
 222
 223         have_objs = oscc_has_objects(oscc, count);
 224
 225         spin_lock(&oscc->oscc_lock);
 226         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 227         spin_unlock(&oscc->oscc_lock);
 228
 229         osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 230
 231         return have_objs || ost_full || osc_invalid;
 232 }
 233
 234 static int oscc_precreate(struct osc_creator *oscc, int wait)
 235 {
 236         struct l_wait_info lwi = { 0 };
 237         int rc = 0;
 238         ENTRY;
 239
 240         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 241                 RETURN(0);
 242
 243         if (!wait)
 244                 RETURN(0);
 245
 246         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 247         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 248
 249         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 250                 rc = -ENOSPC;
 251
 252         if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
 253                 rc = -EIO;
 254
 255         RETURN(rc);
 256 }
 257
 258 int oscc_recovering(struct osc_creator *oscc)
 259 {
 260         int recov = 0;
 261
 262         spin_lock(&oscc->oscc_lock);
 263         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 264         spin_unlock(&oscc->oscc_lock);
 265
 266         return recov;
 267 }
 268
 269 /* decide if the OST has remaining object, return value :
 270         0 : the OST has remaining object, and don't need to do precreate.
 271         1 : the OST has no remaining object, and will send a RPC for precreate.
 272         2 : the OST has no remaining object, and will not get any for
 273             a potentially very long time
 274      1000 : unusable
 275  */
 276 int osc_precreate(struct obd_export *exp)
 277 {
 278         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 279         struct obd_import *imp = exp->exp_imp_reverse;
 280         ENTRY;
 281
 282         LASSERT(oscc != NULL);
 283         if (imp != NULL && imp->imp_deactive)
 284                 RETURN(1000);
 285
 286         if (oscc->oscc_last_id < oscc->oscc_next_id) {
 287                 spin_lock(&oscc->oscc_lock);
 288                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 289                         spin_unlock(&oscc->oscc_lock);
 290                         RETURN(1000);
 291                 }
 292                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 293                         spin_unlock(&oscc->oscc_lock);
 294                         RETURN(1);
 295                 }
 296                 if (oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 297                         spin_unlock(&oscc->oscc_lock);
 298                         RETURN(2);
 299                 }
 300
 301                 if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
 302                         spin_unlock(&oscc->oscc_lock);
 303                         RETURN(1);
 304                 }
 305
 306                 oscc_internal_create(oscc);
 307                 RETURN(1);
 308         }
 309         RETURN(0);
 310 }
 311
 312 int osc_create(struct obd_export *exp, struct obdo *oa,
 313                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 314 {
 315         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 316         struct lov_stripe_md *lsm;
 317         int try_again = 1, rc = 0;
 318         ENTRY;
 319
 320         LASSERT(oa);
 321         LASSERT(ea);
 322         LASSERT(oa->o_gr > 0);
 323         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 324
 325         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 326             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 327                 RETURN(osc_real_create(exp, oa, ea, oti));
 328         }
 329
 330         if (oa->o_gr == FILTER_GROUP_LLOG || oa->o_gr == FILTER_GROUP_ECHO)
 331                 RETURN(osc_real_create(exp, oa, ea, oti));
 332
 333         /* this is the special case where create removes orphans */
 334         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 335             oa->o_flags == OBD_FL_DELORPHAN) {
 336                 spin_lock(&oscc->oscc_lock);
 337                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 338                         spin_unlock(&oscc->oscc_lock);
 339                         RETURN(-EBUSY);
 340                 }
 341                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 342                         spin_unlock(&oscc->oscc_lock);
 343                         RETURN(0);
 344                 }
 345                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 346                 /* seting flag LOW we prevent extra grow precreate size
 347                  * and enforce use last assigned size */
 348                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 349                 spin_unlock(&oscc->oscc_lock);
 350                 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
 351                        oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
 352
 353                 /* delete from next_id on up */
 354                 oa->o_valid |= OBD_MD_FLID;
 355                 oa->o_id = oscc->oscc_next_id - 1;
 356
 357                 rc = osc_real_create(exp, oa, ea, NULL);
 358
 359                 spin_lock(&oscc->oscc_lock);
 360                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 361                 if (rc == 0 || rc == -ENOSPC) {
 362                         if (rc == -ENOSPC)
 363                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 364                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 365                         oscc->oscc_last_id = oa->o_id;
 366                         CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
 367                                LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
 368                                oscc->oscc_last_id, rc);
 369                         cfs_waitq_signal(&oscc->oscc_waitq);
 370                 } else {
 371                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 372                                oscc->oscc_obd->obd_name, rc);
 373                 }
 374                 spin_unlock(&oscc->oscc_lock);
 375
 376
 377                 RETURN(rc);
 378         }
 379
 380         lsm = *ea;
 381         if (lsm == NULL) {
 382                 rc = obd_alloc_memmd(exp, &lsm);
 383                 if (rc < 0)
 384                         RETURN(rc);
 385         }
 386
 387         while (try_again) {
 388                 /* If orphans are being recovered, then we must wait until
 389                    it is finished before we can continue with create. */
 390                 if (oscc_recovering(oscc)) {
 391                         struct l_wait_info lwi;
 392
 393                         CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
 394                                oscc->oscc_obd->obd_name);
 395
 396                         lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(
 397                                 obd_timeout / 4)), NULL, NULL);
 398                         rc = l_wait_event(oscc->oscc_waitq,
 399                                           !oscc_recovering(oscc), &lwi);
 400                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 401                         if (rc == -ETIMEDOUT) {
 402                                 CDEBUG(D_HA,"%s: timeout waiting on recovery\n",
 403                                        oscc->oscc_obd->obd_name);
 404                                 RETURN(rc);
 405                         }
 406                         CDEBUG(D_HA, "%s: oscc recovery over, waking up\n",
 407                                oscc->oscc_obd->obd_name);
 408                 }
 409
 410                 spin_lock(&oscc->oscc_lock);
 411                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 412                         spin_unlock(&oscc->oscc_lock);
 413                         break;
 414                 }
 415
 416                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 417                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 418                         oa->o_id = oscc->oscc_next_id;
 419                         lsm->lsm_object_id = oscc->oscc_next_id;
 420                         *ea = lsm;
 421                         oscc->oscc_next_id++;
 422                         try_again = 0;
 423
 424                         CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
 425                                exp->exp_obd->obd_name, oscc->oscc_next_id);
 426                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 427                         rc = -ENOSPC;
 428                         spin_unlock(&oscc->oscc_lock);
 429                         break;
 430                 }
 431                 spin_unlock(&oscc->oscc_lock);
 432                 rc = oscc_precreate(oscc, try_again);
 433                 if (rc)
 434                         break;
 435         }
 436
 437         if (rc == 0)
 438                 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
 439                        obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
 440         else if (*ea == NULL)
 441                 obd_free_memmd(exp, &lsm);
 442         RETURN(rc);
 443 }
 444
 445 void oscc_init(struct obd_device *obd)
 446 {
 447         struct osc_creator *oscc;
 448
 449         if (obd == NULL)
 450                 return;
 451
 452         oscc = &obd->u.cli.cl_oscc;
 453
 454         memset(oscc, 0, sizeof(*oscc));
 455         CFS_INIT_LIST_HEAD(&oscc->oscc_list);
 456         cfs_waitq_init(&oscc->oscc_waitq);
 457         spin_lock_init(&oscc->oscc_lock);
 458         oscc->oscc_obd = obd;
 459         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 460         oscc->oscc_max_grow_count = OST_MAX_PRECREATE;
 461
 462         oscc->oscc_next_id = 2;
 463         oscc->oscc_last_id = 1;
 464         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 465         /* XXX the export handle should give the oscc the last object */
 466         /* oed->oed_oscc.oscc_last_id = exph->....; */
 467 }