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