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