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;
 167         ptlrpc_at_set_req_timeout(request);
 168         body = lustre_msg_buf(request->rq_reqmsg, REQ_REC_OFF, sizeof(*body));
 169
 170         spin_lock(&oscc->oscc_lock);
 171         body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
 172         body->oa.o_gr = oscc->oscc_oa.o_gr;
 173         LASSERT(body->oa.o_gr > 0);
 174         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 175         spin_unlock(&oscc->oscc_lock);
 176         CDEBUG(D_RPCTRACE, "prealloc through id "LPU64" (last seen "LPU64")\n",
 177                body->oa.o_id, oscc->oscc_last_id);
 178
 179         ptlrpc_req_set_repsize(request, 2, size);
 180
 181         request->rq_async_args.pointer_arg[0] = oscc;
 182         request->rq_interpret_reply = osc_interpret_create;
 183         ptlrpcd_add_req(request);
 184
 185         RETURN(0);
 186 }
 187
 188 static int oscc_has_objects(struct osc_creator *oscc, int count)
 189 {
 190         int have_objs;
 191         spin_lock(&oscc->oscc_lock);
 192         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 193
 194         if (!have_objs) {
 195                 oscc_internal_create(oscc);
 196         } else {
 197                 spin_unlock(&oscc->oscc_lock);
 198         }
 199
 200         return have_objs;
 201 }
 202
 203 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 204 {
 205         int have_objs;
 206         int ost_full;
 207         int osc_invalid;
 208
 209         have_objs = oscc_has_objects(oscc, count);
 210
 211         spin_lock(&oscc->oscc_lock);
 212         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 213         spin_unlock(&oscc->oscc_lock);
 214
 215         osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 216
 217         return have_objs || ost_full || osc_invalid;
 218 }
 219
 220 static int oscc_precreate(struct osc_creator *oscc, int wait)
 221 {
 222         struct l_wait_info lwi = { 0 };
 223         int rc = 0;
 224         ENTRY;
 225
 226         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 227                 RETURN(0);
 228
 229         if (!wait)
 230                 RETURN(0);
 231
 232         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 233         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 234
 235         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 236                 rc = -ENOSPC;
 237
 238         if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
 239                 rc = -EIO;
 240
 241         RETURN(rc);
 242 }
 243
 244 int oscc_recovering(struct osc_creator *oscc)
 245 {
 246         int recov = 0;
 247
 248         spin_lock(&oscc->oscc_lock);
 249         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 250         spin_unlock(&oscc->oscc_lock);
 251
 252         return recov;
 253 }
 254
 255 /* decide if the OST has remaining object, return value :
 256         0 : the OST has remaining object, and don't need to do precreate.
 257         1 : the OST has no remaining object, and will send a RPC for precreate.
 258         2 : the OST has no remaining object, and will not get any for
 259             a potentially very long time
 260      1000 : unusable
 261  */
 262 int osc_precreate(struct obd_export *exp)
 263 {
 264         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 265         struct obd_import *imp = exp->exp_imp_reverse;
 266         ENTRY;
 267
 268         LASSERT(oscc != NULL);
 269         if (imp != NULL && imp->imp_deactive)
 270                 RETURN(1000);
 271
 272         if (oscc->oscc_last_id < oscc->oscc_next_id) {
 273                 spin_lock(&oscc->oscc_lock);
 274                 if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 275                         spin_unlock(&oscc->oscc_lock);
 276                         RETURN(1000);
 277                 }
 278                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 279                         spin_unlock(&oscc->oscc_lock);
 280                         RETURN(1);
 281                 }
 282                 if (oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 283                         spin_unlock(&oscc->oscc_lock);
 284                         RETURN(2);
 285                 }
 286
 287                 if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
 288                         spin_unlock(&oscc->oscc_lock);
 289                         RETURN(1);
 290                 }
 291
 292                 oscc_internal_create(oscc);
 293                 RETURN(1);
 294         }
 295         RETURN(0);
 296 }
 297
 298 int osc_create(struct obd_export *exp, struct obdo *oa,
 299                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 300 {
 301         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 302         struct lov_stripe_md *lsm;
 303         int try_again = 1, rc = 0;
 304         ENTRY;
 305
 306         LASSERT(oa);
 307         LASSERT(ea);
 308         LASSERT(oa->o_gr > 0);
 309         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 310
 311         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 312             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 313                 RETURN(osc_real_create(exp, oa, ea, oti));
 314         }
 315
 316         if (oa->o_gr == FILTER_GROUP_LLOG || oa->o_gr == FILTER_GROUP_ECHO)
 317                 RETURN(osc_real_create(exp, oa, ea, oti));
 318
 319         /* this is the special case where create removes orphans */
 320         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 321             oa->o_flags == OBD_FL_DELORPHAN) {
 322                 spin_lock(&oscc->oscc_lock);
 323                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 324                         spin_unlock(&oscc->oscc_lock);
 325                         RETURN(-EBUSY);
 326                 }
 327                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 328                         spin_unlock(&oscc->oscc_lock);
 329                         RETURN(0);
 330                 }
 331                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 332                 /* seting flag LOW we prevent extra grow precreate size
 333                  * and enforce use last assigned size */
 334                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 335                 spin_unlock(&oscc->oscc_lock);
 336                 CDEBUG(D_HA, "%s: oscc recovery started - delete to "LPU64"\n",
 337                        oscc->oscc_obd->obd_name, oscc->oscc_next_id - 1);
 338
 339                 /* delete from next_id on up */
 340                 oa->o_valid |= OBD_MD_FLID;
 341                 oa->o_id = oscc->oscc_next_id - 1;
 342
 343                 rc = osc_real_create(exp, oa, ea, NULL);
 344
 345                 spin_lock(&oscc->oscc_lock);
 346                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 347                 if (rc == 0 || rc == -ENOSPC) {
 348                         if (rc == -ENOSPC)
 349                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 350                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 351                         oscc->oscc_last_id = oa->o_id;
 352                         CDEBUG(D_HA, "%s: oscc recovery finished, last_id: "
 353                                LPU64", rc: %d\n", oscc->oscc_obd->obd_name,
 354                                oscc->oscc_last_id, rc);
 355                         cfs_waitq_signal(&oscc->oscc_waitq);
 356                 } else {
 357                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 358                                oscc->oscc_obd->obd_name, rc);
 359                 }
 360                 spin_unlock(&oscc->oscc_lock);
 361
 362
 363                 RETURN(rc);
 364         }
 365
 366         lsm = *ea;
 367         if (lsm == NULL) {
 368                 rc = obd_alloc_memmd(exp, &lsm);
 369                 if (rc < 0)
 370                         RETURN(rc);
 371         }
 372
 373         while (try_again) {
 374                 /* If orphans are being recovered, then we must wait until
 375                    it is finished before we can continue with create. */
 376                 if (oscc_recovering(oscc)) {
 377                         struct l_wait_info lwi;
 378
 379                         CDEBUG(D_HA,"%s: oscc recovery in progress, waiting\n",
 380                                oscc->oscc_obd->obd_name);
 381
 382                         lwi = LWI_TIMEOUT(cfs_timeout_cap(cfs_time_seconds(
 383                                 obd_timeout / 4)), NULL, NULL);
 384                         rc = l_wait_event(oscc->oscc_waitq,
 385                                           !oscc_recovering(oscc), &lwi);
 386                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 387                         if (rc == -ETIMEDOUT) {
 388                                 CDEBUG(D_HA,"%s: timeout waiting on recovery\n",
 389                                        oscc->oscc_obd->obd_name);
 390                                 RETURN(rc);
 391                         }
 392                         CDEBUG(D_HA, "%s: oscc recovery over, waking up\n",
 393                                oscc->oscc_obd->obd_name);
 394                 }
 395
 396                 spin_lock(&oscc->oscc_lock);
 397                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 398                         spin_unlock(&oscc->oscc_lock);
 399                         break;
 400                 }
 401
 402                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 403                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 404                         oa->o_id = oscc->oscc_next_id;
 405                         lsm->lsm_object_id = oscc->oscc_next_id;
 406                         *ea = lsm;
 407                         oscc->oscc_next_id++;
 408                         try_again = 0;
 409
 410                         CDEBUG(D_RPCTRACE, "%s: set oscc_next_id = "LPU64"\n",
 411                                exp->exp_obd->obd_name, oscc->oscc_next_id);
 412                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 413                         rc = -ENOSPC;
 414                         spin_unlock(&oscc->oscc_lock);
 415                         break;
 416                 }
 417                 spin_unlock(&oscc->oscc_lock);
 418                 rc = oscc_precreate(oscc, try_again);
 419                 if (rc)
 420                         break;
 421         }
 422
 423         if (rc == 0)
 424                 CDEBUG(D_INFO, "%s: returning objid "LPU64"\n",
 425                        obd2cli_tgt(oscc->oscc_obd), lsm->lsm_object_id);
 426         else if (*ea == NULL)
 427                 obd_free_memmd(exp, &lsm);
 428         RETURN(rc);
 429 }
 430
 431 void oscc_init(struct obd_device *obd)
 432 {
 433         struct osc_creator *oscc;
 434
 435         if (obd == NULL)
 436                 return;
 437
 438         oscc = &obd->u.cli.cl_oscc;
 439
 440         memset(oscc, 0, sizeof(*oscc));
 441         CFS_INIT_LIST_HEAD(&oscc->oscc_list);
 442         cfs_waitq_init(&oscc->oscc_waitq);
 443         spin_lock_init(&oscc->oscc_lock);
 444         oscc->oscc_obd = obd;
 445         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 446
 447         oscc->oscc_next_id = 2;
 448         oscc->oscc_last_id = 1;
 449         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 450         /* XXX the export handle should give the oscc the last object */
 451         /* oed->oed_oscc.oscc_last_id = exph->....; */
 452 }