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, int rc)
  60 {
  61         struct osc_creator *oscc;
  62         struct ost_body *body = NULL;
  63         ENTRY;
  64
  65         if (req->rq_repmsg) {
  66                 body = lustre_swab_repbuf(req, 0, sizeof(*body),
  67                                           lustre_swab_ost_body);
  68                 if (body == NULL && rc == 0)
  69                         rc = -EPROTO;
  70         }
  71
  72         oscc = req->rq_async_args.pointer_arg[0];
  73         LASSERT(oscc && (oscc->oscc_obd != LP_POISON));
  74
  75         spin_lock(&oscc->oscc_lock);
  76         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  77         if (rc == -ENOSPC || rc == -EROFS) {
  78                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
  79                 if (body && rc == -ENOSPC) {
  80                         oscc->oscc_grow_count = OST_MIN_PRECREATE;
  81                         oscc->oscc_last_id = body->oa.o_id;
  82                 }
  83                 spin_unlock(&oscc->oscc_lock);
  84                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
  85         } else if (rc != 0 && rc != -EIO) {
  86                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
  87                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
  88                 spin_unlock(&oscc->oscc_lock);
  89                 DEBUG_REQ(D_ERROR, req,
  90                           "unknown rc %d from async create: failing oscc", rc);
  91                 ptlrpc_fail_import(req->rq_import, req->rq_import_generation);
  92         } else {
  93                 if (rc == 0) {
  94                         oscc->oscc_flags &= ~OSCC_FLAG_LOW;
  95                         if (body) {
  96                                 int diff = body->oa.o_id - oscc->oscc_last_id;
  97                                 if (diff != oscc->oscc_grow_count)
  98                                         oscc->oscc_grow_count =
  99                                                 max(diff/3, OST_MIN_PRECREATE);
 100                                 oscc->oscc_last_id = body->oa.o_id;
 101                         }
 102                 }
 103                 spin_unlock(&oscc->oscc_lock);
 104         }
 105
 106         CDEBUG(D_HA, "preallocated through id "LPU64" (last used "LPU64")\n",
 107                oscc->oscc_last_id, oscc->oscc_next_id);
 108
 109         wake_up(&oscc->oscc_waitq);
 110         RETURN(rc);
 111 }
 112
 113 static int oscc_internal_create(struct osc_creator *oscc)
 114 {
 115         struct ptlrpc_request *request;
 116         struct ost_body *body;
 117         int size = sizeof(*body);
 118         ENTRY;
 119
 120         spin_lock(&oscc->oscc_lock);
 121         if (oscc->oscc_grow_count < OST_MAX_PRECREATE &&
 122             !(oscc->oscc_flags & (OSCC_FLAG_LOW | OSCC_FLAG_RECOVERING)) &&
 123             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 124                    (oscc->oscc_grow_count / 4 + 1)) {
 125                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 126                 oscc->oscc_grow_count *= 2;
 127         }
 128
 129         if (oscc->oscc_grow_count > OST_MAX_PRECREATE / 2)
 130                 oscc->oscc_grow_count = OST_MAX_PRECREATE / 2;
 131
 132         if (oscc->oscc_flags & OSCC_FLAG_CREATING ||
 133             oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 134                 spin_unlock(&oscc->oscc_lock);
 135                 RETURN(0);
 136         }
 137         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 138         spin_unlock(&oscc->oscc_lock);
 139
 140         request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import, OST_CREATE,
 141                                   1, &size, NULL);
 142         if (request == NULL) {
 143                 spin_lock(&oscc->oscc_lock);
 144                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 145                 spin_unlock(&oscc->oscc_lock);
 146                 RETURN(-ENOMEM);
 147         }
 148
 149         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
 150         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
 151
 152         spin_lock(&oscc->oscc_lock);
 153         body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
 154         body->oa.o_valid |= OBD_MD_FLID;
 155         spin_unlock(&oscc->oscc_lock);
 156         CDEBUG(D_HA, "preallocating through id "LPU64" (last used "LPU64")\n",
 157                body->oa.o_id, oscc->oscc_next_id);
 158
 159         request->rq_replen = lustre_msg_size(1, &size);
 160
 161         request->rq_async_args.pointer_arg[0] = oscc;
 162         request->rq_interpret_reply = osc_interpret_create;
 163         ptlrpcd_add_req(request);
 164
 165         RETURN(0);
 166 }
 167
 168 static int oscc_has_objects(struct osc_creator *oscc, int count)
 169 {
 170         int have_objs;
 171         spin_lock(&oscc->oscc_lock);
 172         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 173         spin_unlock(&oscc->oscc_lock);
 174
 175         if (!have_objs)
 176                 oscc_internal_create(oscc);
 177
 178         return have_objs;
 179 }
 180
 181 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 182 {
 183         int have_objs;
 184         int ost_full;
 185         int osc_invalid;
 186
 187         have_objs = oscc_has_objects(oscc, count);
 188
 189         spin_lock(&oscc->oscc_lock);
 190         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 191         spin_unlock(&oscc->oscc_lock);
 192
 193         osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 194
 195         return have_objs || ost_full || osc_invalid;
 196 }
 197
 198 static int oscc_precreate(struct osc_creator *oscc, int wait)
 199 {
 200         struct l_wait_info lwi = { 0 };
 201         int rc = 0;
 202         ENTRY;
 203
 204         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 205                 RETURN(0);
 206
 207         if (!wait)
 208                 RETURN(0);
 209
 210         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 211         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 212
 213         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 214                 rc = -ENOSPC;
 215
 216         if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
 217                 rc = -EIO;
 218
 219         RETURN(rc);
 220 }
 221
 222 int oscc_recovering(struct osc_creator *oscc)
 223 {
 224         int recov = 0;
 225
 226         spin_lock(&oscc->oscc_lock);
 227         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 228         spin_unlock(&oscc->oscc_lock);
 229
 230         return recov;
 231 }
 232
 233 int osc_create(struct obd_export *exp, struct obdo *oa,
 234                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 235 {
 236         struct lov_stripe_md *lsm;
 237         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 238         int try_again = 1, rc = 0;
 239         ENTRY;
 240         LASSERT(oa);
 241         LASSERT(ea);
 242
 243         if ((oa->o_valid & OBD_MD_FLGROUP) && (oa->o_gr != 0))
 244                 RETURN(osc_real_create(exp, oa, ea, oti));
 245
 246         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 247             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 248                 RETURN(osc_real_create(exp, oa, ea, oti));
 249         }
 250
 251         /* this is the special case where create removes orphans */
 252         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 253             oa->o_flags == OBD_FL_DELORPHAN) {
 254                 spin_lock(&oscc->oscc_lock);
 255                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 256                         spin_unlock(&oscc->oscc_lock);
 257                         return -EBUSY;
 258                 }
 259                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 260                         spin_unlock(&oscc->oscc_lock);
 261                         return 0;
 262                 }
 263                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 264                 spin_unlock(&oscc->oscc_lock);
 265                 CDEBUG(D_HA, "%s: oscc recovery started\n",
 266                        oscc->oscc_obd->obd_name);
 267
 268                 /* delete from next_id on up */
 269                 oa->o_valid |= OBD_MD_FLID;
 270                 oa->o_id = oscc->oscc_next_id - 1;
 271
 272                 CDEBUG(D_HA, "%s: deleting to next_id: "LPU64"\n",
 273                        oscc->oscc_obd->obd_name, oa->o_id);
 274
 275                 rc = osc_real_create(exp, oa, ea, NULL);
 276
 277                 spin_lock(&oscc->oscc_lock);
 278                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 279                 if (rc == 0 || rc == -ENOSPC) {
 280                         if (rc == -ENOSPC)
 281                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 282                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 283                         oscc->oscc_last_id = oa->o_id;
 284                         CDEBUG(D_HA, "%s: oscc recovery finished: %d\n",
 285                                oscc->oscc_obd->obd_name, rc);
 286                         wake_up(&oscc->oscc_waitq);
 287                 } else {
 288                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 289                                oscc->oscc_obd->obd_name, rc);
 290                 }
 291                 spin_unlock(&oscc->oscc_lock);
 292
 293
 294                 RETURN(rc);
 295         }
 296
 297         lsm = *ea;
 298         if (lsm == NULL) {
 299                 rc = obd_alloc_memmd(exp, &lsm);
 300                 if (rc < 0)
 301                         RETURN(rc);
 302         }
 303
 304         while (try_again) {
 305                 /* If orphans are being recovered, then we must wait until
 306                    it is finished before we can continue with create. */
 307                 if (oscc_recovering(oscc)) {
 308                         struct l_wait_info lwi;
 309
 310                         CDEBUG(D_HA,"%p: oscc recovery in progress, waiting\n",
 311                                oscc);
 312
 313                         lwi = LWI_TIMEOUT(MAX(obd_timeout*HZ/4, 1), NULL, NULL);
 314                         rc = l_wait_event(oscc->oscc_waitq,
 315                                           !oscc_recovering(oscc), &lwi);
 316                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 317                         if (rc == -ETIMEDOUT) {
 318                                 CDEBUG(D_HA,"%p: timeout waiting on recovery\n",
 319                                        oscc);
 320                                 RETURN(rc);
 321                         }
 322                         CDEBUG(D_HA, "%p: oscc recovery over, waking up\n",
 323                                oscc);
 324                 }
 325
 326                 spin_lock(&oscc->oscc_lock);
 327                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 328                         spin_unlock(&oscc->oscc_lock);
 329                         break;
 330                 }
 331
 332                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 333                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 334                         oa->o_id = oscc->oscc_next_id;
 335                         lsm->lsm_object_id = oscc->oscc_next_id;
 336                         *ea = lsm;
 337                         oscc->oscc_next_id++;
 338                         try_again = 0;
 339                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 340                         rc = -ENOSPC;
 341                         spin_unlock(&oscc->oscc_lock);
 342                         break;
 343                 }
 344                 spin_unlock(&oscc->oscc_lock);
 345                 rc = oscc_precreate(oscc, try_again);
 346                 if (rc)
 347                         break;
 348         }
 349
 350         if (rc == 0)
 351                 CDEBUG(D_HA, "%s: returning objid "LPU64"\n",
 352                        oscc->oscc_obd->u.cli.cl_import->imp_target_uuid.uuid,
 353                        lsm->lsm_object_id);
 354         else if (*ea == NULL)
 355                 obd_free_memmd(exp, &lsm);
 356         RETURN(rc);
 357 }
 358
 359 void oscc_init(struct obd_device *obd)
 360 {
 361         struct osc_creator *oscc;
 362
 363         if (obd == NULL)
 364                 return;
 365
 366         oscc = &obd->u.cli.cl_oscc;
 367
 368         memset(oscc, 0, sizeof(*oscc));
 369         INIT_LIST_HEAD(&oscc->oscc_list);
 370         init_waitqueue_head(&oscc->oscc_waitq);
 371         spin_lock_init(&oscc->oscc_lock);
 372         oscc->oscc_obd = obd;
 373         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 374
 375         oscc->oscc_next_id = 2;
 376         oscc->oscc_last_id = 1;
 377         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 378         /* XXX the export handle should give the oscc the last object */
 379         /* oed->oed_oscc.oscc_last_id = exph->....; */
 380 }