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 Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  *  For testing and management it is treated as an obd_device,
  23  *  although * it does not export a full OBD method table (the
  24  *  requests are coming * in over the wire, so object target modules
  25  *  do not have a full * method table.)
  26  *
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_OSC
  33
  34 #ifdef __KERNEL__
  35 # include <linux/version.h>
  36 # include <linux/module.h>
  37 # include <linux/mm.h>
  38 # include <linux/highmem.h>
  39 # include <linux/ctype.h>
  40 # include <linux/init.h>
  41 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  42 #  include <linux/workqueue.h>
  43 #  include <linux/smp_lock.h>
  44 # else
  45 #  include <linux/locks.h>
  46 # endif
  47 #else /* __KERNEL__ */
  48 # include <liblustre.h>
  49 #endif
  50
  51 #ifdef  __CYGWIN__
  52 # include <ctype.h>
  53 #endif
  54
  55 # include <linux/lustre_dlm.h>
  56 #include <linux/obd_class.h>
  57 #include "osc_internal.h"
  58
  59 static int osc_interpret_create(struct ptlrpc_request *req, void *data,
  60                                 int rc)
  61 {
  62         struct osc_creator *oscc;
  63         struct ost_body *body = NULL;
  64         ENTRY;
  65
  66         if (req->rq_repmsg) {
  67                 body = lustre_swab_repbuf(req, 0, sizeof(*body),
  68                                           lustre_swab_ost_body);
  69                 if (body == NULL && rc == 0)
  70                         rc = -EPROTO;
  71         }
  72
  73         oscc = req->rq_async_args.pointer_arg[0];
  74         spin_lock(&oscc->oscc_lock);
  75         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  76         if (body)
  77                 oscc->oscc_last_id = body->oa.o_id;
  78         if (rc == -ENOSPC) {
  79                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
  80                 spin_unlock(&oscc->oscc_lock);
  81                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
  82         } else if (rc != 0 && rc != -EIO) {
  83                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
  84                 spin_unlock(&oscc->oscc_lock);
  85                 DEBUG_REQ(D_ERROR, req,
  86                           "unknown rc %d from async create: failing oscc", rc);
  87                 ptlrpc_fail_import(req->rq_import, req->rq_import_generation);
  88         } else {
  89                 spin_unlock(&oscc->oscc_lock);
  90         }
  91
  92         CDEBUG(D_HA, "preallocated through id "LPU64" (last used "LPU64")\n",
  93                oscc->oscc_last_id, oscc->oscc_next_id);
  94
  95         wake_up(&oscc->oscc_waitq);
  96         RETURN(rc);
  97 }
  98
  99 static int oscc_internal_create(struct osc_creator *oscc)
 100 {
 101         struct ptlrpc_request *request;
 102         struct ost_body *body;
 103         int size = sizeof(*body);
 104         ENTRY;
 105
 106         spin_lock(&oscc->oscc_lock);
 107         if (oscc->oscc_flags & OSCC_FLAG_CREATING ||
 108             oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 109                 spin_unlock(&oscc->oscc_lock);
 110                 RETURN(0);
 111         }
 112         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 113         spin_unlock(&oscc->oscc_lock);
 114
 115         request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
 116                                   LUSTRE_OBD_VERSION, OST_CREATE,
 117                                   1, &size, NULL);
 118         if (request == NULL) {
 119                 spin_lock(&oscc->oscc_lock);
 120                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 121                 spin_unlock(&oscc->oscc_lock);
 122                 RETURN(-ENOMEM);
 123         }
 124
 125         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
 126         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
 127
 128         spin_lock(&oscc->oscc_lock);
 129         body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
 130         /* probably we should take frequence of request into account? -bzzz */
 131         if (oscc->oscc_grow_count < oscc->oscc_max_grow_count) {
 132                 oscc->oscc_grow_count *= 2;
 133                 if (oscc->oscc_grow_count > oscc->oscc_max_grow_count)
 134                         oscc->oscc_grow_count = oscc->oscc_max_grow_count;
 135         }
 136         body->oa.o_gr = oscc->oscc_gr;
 137         LASSERT(body->oa.o_gr > 0);
 138         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 139         spin_unlock(&oscc->oscc_lock);
 140         CDEBUG(D_INFO, "preallocating through id "LPU64" (last used "LPU64")\n",
 141                body->oa.o_id, oscc->oscc_next_id);
 142
 143         request->rq_replen = lustre_msg_size(1, &size);
 144
 145         request->rq_async_args.pointer_arg[0] = oscc;
 146         request->rq_interpret_reply = osc_interpret_create;
 147         ptlrpcd_add_req(request);
 148
 149         RETURN(0);
 150 }
 151
 152 static int oscc_has_objects(struct osc_creator *oscc, int count)
 153 {
 154         int have_objs;
 155         spin_lock(&oscc->oscc_lock);
 156         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 157         spin_unlock(&oscc->oscc_lock);
 158
 159         if (!have_objs)
 160                 oscc_internal_create(oscc);
 161
 162         return have_objs;
 163 }
 164
 165 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 166 {
 167         int have_objs;
 168         int ost_full;
 169         int osc_invalid;
 170
 171         have_objs = oscc_has_objects(oscc, count);
 172
 173         spin_lock(&oscc->oscc_lock);
 174         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 175         spin_unlock(&oscc->oscc_lock);
 176
 177         osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 178
 179         return have_objs || ost_full || osc_invalid;
 180 }
 181
 182 static int oscc_precreate(struct osc_creator *oscc, int wait)
 183 {
 184         struct l_wait_info lwi = { 0 };
 185         int rc = 0;
 186         ENTRY;
 187
 188         if (oscc_has_objects(oscc, oscc->oscc_kick_barrier))
 189                 RETURN(0);
 190
 191         if (!wait)
 192                 RETURN(0);
 193
 194         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 195         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 196
 197         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 198                 rc = -ENOSPC;
 199
 200         if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
 201                 rc = -EIO;
 202
 203         RETURN(rc);
 204 }
 205
 206 int oscc_recovering(struct osc_creator *oscc)
 207 {
 208         int recov = 0;
 209
 210         spin_lock(&oscc->oscc_lock);
 211         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 212         spin_unlock(&oscc->oscc_lock);
 213
 214         return recov;
 215 }
 216
 217 int osc_create(struct obd_export *exp, struct obdo *oa,
 218                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 219 {
 220         struct lov_stripe_md *lsm;
 221         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 222         int try_again = 1, rc = 0;
 223         ENTRY;
 224         LASSERT(oa);
 225         LASSERT(ea);
 226         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 227         LASSERT(oa->o_gr > 0);
 228
 229         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 230             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 231                 /* Exceptional case where we are trying to repair missing
 232                  * objects for various groups.  We have already validated that
 233                  * this is a valid group for the file.  Don't set oscc->oscc_gr.
 234                  */
 235                 RETURN(osc_real_create(exp, oa, ea, oti));
 236         }
 237
 238         LASSERT(oscc->oscc_gr == 0 || oscc->oscc_gr == oa->o_gr);
 239         oscc->oscc_gr = oa->o_gr;
 240
 241         if (oa->o_gr == FILTER_GROUP_LLOG || oa->o_gr == FILTER_GROUP_ECHO)
 242                 RETURN(osc_real_create(exp, oa, ea, oti));
 243
 244         /* this is the special case where create removes orphans */
 245         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 246             oa->o_flags == OBD_FL_DELORPHAN) {
 247                 spin_lock(&oscc->oscc_lock);
 248                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 249                         spin_unlock(&oscc->oscc_lock);
 250                         return -EBUSY;
 251                 }
 252                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 253                         spin_unlock(&oscc->oscc_lock);
 254                         return 0;
 255                 }
 256                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 257                 spin_unlock(&oscc->oscc_lock);
 258                 CDEBUG(D_HA, "%s; oscc recovery started\n",
 259                        exp->exp_obd->obd_name);
 260                 LASSERT(oscc->oscc_flags & OSCC_FLAG_RECOVERING);
 261
 262                 /* delete from next_id on up */
 263                 oa->o_valid |= OBD_MD_FLID;
 264                 oa->o_id = oscc->oscc_next_id - 1;
 265
 266                 CDEBUG(D_HA, "%s: deleting to next_id: "LPU64"\n",
 267                        exp->exp_obd->obd_name, oa->o_id);
 268
 269                 rc = osc_real_create(exp, oa, ea, NULL);
 270                 if (oscc->oscc_obd == NULL) {
 271                         CWARN("the obd for oscc %p has been freed\n", oscc);
 272                         RETURN(rc);
 273                 }
 274
 275                 spin_lock(&oscc->oscc_lock);
 276                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 277                 if (rc == 0 || rc == -ENOSPC) {
 278                         if (rc == -ENOSPC)
 279                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 280                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 281                         oscc->oscc_last_id = oa->o_id;
 282
 283                         CDEBUG(D_HA, "%s: oscc recovery finished: %d\n",
 284                                exp->exp_obd->obd_name, rc);
 285                         wake_up(&oscc->oscc_waitq);
 286
 287                 } else {
 288                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 289                                exp->exp_obd->obd_name, rc);
 290                 }
 291                 spin_unlock(&oscc->oscc_lock);
 292
 293                 RETURN(rc);
 294         }
 295
 296         lsm = *ea;
 297         if (lsm == NULL) {
 298                 rc = obd_alloc_memmd(exp, &lsm);
 299                 if (rc < 0)
 300                         RETURN(rc);
 301         }
 302
 303         while (try_again) {
 304                 /* If orphans are being recovered, then we must wait until
 305                    it is finished before we can continue with create. */
 306                 if (oscc_recovering(oscc)) {
 307                         struct l_wait_info lwi;
 308
 309                         CDEBUG(D_HA,"%p: oscc recovery in progress, waiting\n",
 310                                oscc);
 311                         lwi = LWI_TIMEOUT(MAX(obd_timeout*HZ/4, 1), NULL, NULL);
 312                         rc = l_wait_event(oscc->oscc_waitq,
 313                                           !oscc_recovering(oscc), &lwi);
 314
 315                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 316                         if (rc == -ETIMEDOUT) {
 317                                 CDEBUG(D_HA,"%p: timeout waiting on recovery\n",
 318                                        oscc);
 319                                 RETURN(rc);
 320                         }
 321                         CDEBUG(D_HA, "%s: oscc recovery over, waking up\n",
 322                                exp->exp_obd->obd_name);
 323                 }
 324
 325                 spin_lock(&oscc->oscc_lock);
 326                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 327                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 328                         oa->o_id = oscc->oscc_next_id;
 329                         oa->o_gr = oscc->oscc_gr;
 330                         lsm->lsm_object_id = oscc->oscc_next_id;
 331                         lsm->lsm_object_gr = oscc->oscc_gr;
 332                         *ea = lsm;
 333                         oscc->oscc_next_id++;
 334                         try_again = 0;
 335                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 336                         rc = -ENOSPC;
 337                         spin_unlock(&oscc->oscc_lock);
 338                         break;
 339                 }
 340                 spin_unlock(&oscc->oscc_lock);
 341                 rc = oscc_precreate(oscc, try_again);
 342                 if (rc == -EIO)
 343                         break;
 344         }
 345
 346         if (rc == 0)
 347                 CDEBUG(D_HA, "%s: returning objid "LPU64"\n",
 348                        oscc->oscc_obd->u.cli.cl_import->imp_target_uuid.uuid,
 349                        lsm->lsm_object_id);
 350         else if (*ea == NULL)
 351                 obd_free_memmd(exp, &lsm);
 352         RETURN(rc);
 353 }
 354
 355 void oscc_init(struct obd_device *obd)
 356 {
 357         struct osc_creator *oscc;
 358
 359         if (obd == NULL)
 360                 return;
 361
 362         oscc = &obd->u.cli.cl_oscc;
 363
 364         memset(oscc, 0, sizeof(*oscc));
 365         INIT_LIST_HEAD(&oscc->oscc_list);
 366         init_waitqueue_head(&oscc->oscc_waitq);
 367         spin_lock_init(&oscc->oscc_lock);
 368         oscc->oscc_obd = obd;
 369         oscc->oscc_kick_barrier = 100;
 370         oscc->oscc_grow_count = 36;
 371         oscc->oscc_max_grow_count = 2000;
 372
 373         oscc->oscc_next_id = 2;
 374         oscc->oscc_last_id = 1;
 375         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 376         /* XXX the export handle should give the oscc the last object */
 377         /* oed->oed_oscc.oscc_last_id = exph->....; */
 378 }