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