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