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         if (body)
  76                 oscc->oscc_last_id = body->oa.o_id;
  77         if (rc == -ENOSPC) {
  78                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
  79                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
  80         } else if (rc != 0 && rc != -EIO) {
  81                 DEBUG_REQ(D_ERROR, req,
  82                           "unknown rc %d from async create: failing oscc",
  83                           rc);
  84                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
  85                 ptlrpc_fail_import(req->rq_import, req->rq_import_generation);
  86         }
  87         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  88         spin_unlock(&oscc->oscc_lock);
  89
  90         CDEBUG(D_HA, "preallocated through id "LPU64" (last used "LPU64")\n",
  91                oscc->oscc_last_id, oscc->oscc_next_id);
  92
  93         wake_up(&oscc->oscc_waitq);
  94         RETURN(rc);
  95 }
  96
  97 static int oscc_internal_create(struct osc_creator *oscc)
  98 {
  99         struct ptlrpc_request *request;
 100         struct ost_body *body;
 101         int size = sizeof(*body);
 102         ENTRY;
 103
 104         spin_lock(&oscc->oscc_lock);
 105         if (oscc->oscc_flags & OSCC_FLAG_CREATING ||
 106             oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 107                 spin_unlock(&oscc->oscc_lock);
 108                 RETURN(0);
 109         }
 110         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 111         spin_unlock(&oscc->oscc_lock);
 112
 113         request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import, OST_CREATE,
 114                                   1, &size, NULL);
 115         if (request == NULL) {
 116                 spin_lock(&oscc->oscc_lock);
 117                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 118                 spin_unlock(&oscc->oscc_lock);
 119                 RETURN(-ENOMEM);
 120         }
 121
 122         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
 123         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
 124
 125         spin_lock(&oscc->oscc_lock);
 126         body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
 127         /* probably we should take frequence of request into account? -bzzz */
 128         if (oscc->oscc_grow_count < oscc->oscc_max_grow_count) {
 129                 oscc->oscc_grow_count *= 2;
 130                 if (oscc->oscc_grow_count > oscc->oscc_max_grow_count)
 131                         oscc->oscc_grow_count = oscc->oscc_max_grow_count;
 132         }
 133         body->oa.o_gr = oscc->oscc_gr;
 134         LASSERT(body->oa.o_gr > 0);
 135         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 136         CDEBUG(D_INFO, "preallocating through id "LPU64" (last used "LPU64")\n",
 137                body->oa.o_id, oscc->oscc_next_id);
 138         spin_unlock(&oscc->oscc_lock);
 139
 140         request->rq_replen = lustre_msg_size(1, &size);
 141
 142         request->rq_async_args.pointer_arg[0] = oscc;
 143         request->rq_interpret_reply = osc_interpret_create;
 144         ptlrpcd_add_req(request);
 145
 146         RETURN(0);
 147 }
 148
 149 static int oscc_has_objects(struct osc_creator *oscc, int count)
 150 {
 151         int have_objs;
 152         spin_lock(&oscc->oscc_lock);
 153         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 154         spin_unlock(&oscc->oscc_lock);
 155
 156         if (!have_objs)
 157                 oscc_internal_create(oscc);
 158
 159         return have_objs;
 160 }
 161
 162 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 163 {
 164         int have_objs;
 165         int ost_full;
 166         int osc_invalid;
 167
 168         have_objs = oscc_has_objects(oscc, count);
 169
 170         spin_lock(&oscc->oscc_lock);
 171         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 172         spin_unlock(&oscc->oscc_lock);
 173
 174         osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 175
 176         return have_objs || ost_full || osc_invalid;
 177 }
 178
 179 static int oscc_precreate(struct osc_creator *oscc, int wait)
 180 {
 181         struct l_wait_info lwi = { 0 };
 182         int rc = 0;
 183         ENTRY;
 184
 185         if (oscc_has_objects(oscc, oscc->oscc_kick_barrier))
 186                 RETURN(0);
 187
 188         if (!wait)
 189                 RETURN(0);
 190
 191         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 192         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 193
 194         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 195                 rc = -ENOSPC;
 196
 197         if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
 198                 rc = -EIO;
 199
 200         RETURN(rc);
 201 }
 202
 203 int oscc_recovering(struct osc_creator *oscc)
 204 {
 205         int recov = 0;
 206
 207         spin_lock(&oscc->oscc_lock);
 208         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 209         spin_unlock(&oscc->oscc_lock);
 210
 211         return recov;
 212 }
 213
 214 int osc_create(struct obd_export *exp, struct obdo *oa,
 215                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 216 {
 217         struct lov_stripe_md *lsm;
 218         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 219         int try_again = 1, rc = 0;
 220         ENTRY;
 221         LASSERT(oa);
 222         LASSERT(ea);
 223         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 224         LASSERT(oa->o_gr > 0);
 225
 226         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 227             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 228                 /* Exceptional case where we are trying to repair missing
 229                  * objects for various groups.  We have already validated that
 230                  * this is a valid group for the file.  Don't set oscc->oscc_gr.
 231                  */
 232                 RETURN(osc_real_create(exp, oa, ea, oti));
 233         }
 234
 235         LASSERT(oscc->oscc_gr == 0 || oscc->oscc_gr == oa->o_gr);
 236         oscc->oscc_gr = oa->o_gr;
 237
 238         if (oa->o_gr == FILTER_GROUP_LLOG || oa->o_gr == FILTER_GROUP_ECHO)
 239                 RETURN(osc_real_create(exp, oa, ea, oti));
 240
 241         lsm = *ea;
 242         if (lsm == NULL) {
 243                 rc = obd_alloc_memmd(exp, &lsm);
 244                 if (rc < 0)
 245                         RETURN(rc);
 246         }
 247
 248         /* this is the special case where create removes orphans */
 249         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 250             oa->o_flags == OBD_FL_DELORPHAN) {
 251                 CDEBUG(D_HA, "%s; oscc recovery started\n",
 252                        exp->exp_obd->obd_name);
 253                 LASSERT(oscc->oscc_flags & OSCC_FLAG_RECOVERING);
 254
 255                 /* delete from next_id on up */
 256                 oa->o_valid |= OBD_MD_FLID;
 257                 oa->o_id = oscc->oscc_next_id - 1;
 258
 259                 CDEBUG(D_HA, "%s: deleting to next_id: "LPU64"\n",
 260                        exp->exp_obd->obd_name, oa->o_id);
 261
 262                 rc = osc_real_create(exp, oa, ea, NULL);
 263                 if (oscc->oscc_obd == NULL) {
 264                         CWARN("the obd for oscc %p has been freed\n", oscc);
 265                         RETURN(rc);
 266                 }
 267
 268                 spin_lock(&oscc->oscc_lock);
 269                 if (rc == 0 || rc == -ENOSPC) {
 270                         if (rc == -ENOSPC)
 271                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 272                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 273                         oscc->oscc_last_id = oa->o_id;
 274
 275                         CDEBUG(D_HA, "%s: oscc recovery finished: %d\n",
 276                                exp->exp_obd->obd_name, rc);
 277                         wake_up(&oscc->oscc_waitq);
 278
 279                 } else {
 280                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 281                                exp->exp_obd->obd_name, rc);
 282                 }
 283                 spin_unlock(&oscc->oscc_lock);
 284
 285                 RETURN(rc);
 286         }
 287
 288         while (try_again) {
 289                 /* If orphans are being recovered, then we must wait until
 290                    it is finished before we can continue with create. */
 291                 if (oscc_recovering(oscc)) {
 292                         struct l_wait_info lwi;
 293
 294                         CDEBUG(D_HA,"%s: oscc sync in progress, waiting\n",
 295                                exp->exp_obd->obd_name);
 296
 297                         lwi = LWI_TIMEOUT(MAX(obd_timeout * HZ, 1), NULL, NULL);
 298                         rc = l_wait_event(oscc->oscc_waitq,
 299                                           !oscc_recovering(oscc), &lwi);
 300                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 301                         if (rc == -ETIMEDOUT) {
 302                                 CDEBUG(D_HA, "%s: timed out waiting for sync\n",
 303                                        exp->exp_obd->obd_name);
 304                                 RETURN(rc);
 305                         }
 306                         CDEBUG(D_HA, "%s: oscc sync over, waking up\n",
 307                                exp->exp_obd->obd_name);
 308                 }
 309
 310                 spin_lock(&oscc->oscc_lock);
 311                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 312                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 313                         oa->o_id = oscc->oscc_next_id;
 314                         oa->o_gr = oscc->oscc_gr;
 315                         lsm->lsm_object_id = oscc->oscc_next_id;
 316                         lsm->lsm_object_gr = oscc->oscc_gr;
 317                         *ea = lsm;
 318                         oscc->oscc_next_id++;
 319                         try_again = 0;
 320                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 321                         rc = -ENOSPC;
 322                         spin_unlock(&oscc->oscc_lock);
 323                         break;
 324                 }
 325                 spin_unlock(&oscc->oscc_lock);
 326                 rc = oscc_precreate(oscc, try_again);
 327                 if (rc == -EIO)
 328                         break;
 329         }
 330
 331         if (rc == 0)
 332                 CDEBUG(D_INFO, "returning objid "LPU64"/"LPU64"\n",
 333                                 lsm->lsm_object_id, lsm->lsm_object_gr);
 334         else if (*ea == NULL)
 335                 obd_free_memmd(exp, &lsm);
 336         RETURN(rc);
 337 }
 338
 339 void oscc_init(struct obd_device *obd)
 340 {
 341         struct osc_creator *oscc;
 342
 343         if (obd == NULL)
 344                 return;
 345
 346         oscc = &obd->u.cli.cl_oscc;
 347
 348         memset(oscc, 0, sizeof(*oscc));
 349         INIT_LIST_HEAD(&oscc->oscc_list);
 350         init_waitqueue_head(&oscc->oscc_waitq);
 351         spin_lock_init(&oscc->oscc_lock);
 352         oscc->oscc_obd = obd;
 353         oscc->oscc_kick_barrier = 100;
 354         oscc->oscc_grow_count = 36;
 355         oscc->oscc_max_grow_count = 2000;
 356
 357         oscc->oscc_next_id = 2;
 358         oscc->oscc_last_id = 1;
 359         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 360         /* XXX the export handle should give the oscc the last object */
 361         /* oed->oed_oscc.oscc_last_id = exph->....; */
 362 }