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         spin_lock(&oscc->oscc_lock);
  74         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  75         if (rc == -ENOSPC || rc == -EROFS) {
  76                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
  77                 if (body && rc == -ENOSPC) {
  78                         oscc->oscc_grow_count = OST_MIN_PRECREATE;
  79                         oscc->oscc_last_id = body->oa.o_id;
  80                 }
  81                 spin_unlock(&oscc->oscc_lock);
  82                 DEBUG_REQ(D_INODE, req, "OST out of space, flagging");
  83         } else if (rc != 0 && rc != -EIO) {
  84                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
  85                 oscc->oscc_grow_count = OST_MIN_PRECREATE;
  86                 spin_unlock(&oscc->oscc_lock);
  87                 DEBUG_REQ(D_ERROR, req,
  88                           "unknown rc %d from async create: failing oscc", rc);
  89                 ptlrpc_fail_import(req->rq_import, req->rq_import_generation);
  90         } else {
  91                 if (rc == 0) {
  92                         oscc->oscc_flags &= ~OSCC_FLAG_LOW;
  93                         if (body) {
  94                                 int diff = body->oa.o_id - oscc->oscc_last_id;
  95                                 if (diff != oscc->oscc_grow_count)
  96                                         oscc->oscc_grow_count =
  97                                                 max(diff/3, OST_MIN_PRECREATE);
  98                                 oscc->oscc_last_id = body->oa.o_id;
  99                         }
 100                 }
 101                 spin_unlock(&oscc->oscc_lock);
 102         }
 103
 104         CDEBUG(D_HA, "preallocated through id "LPU64" (last used "LPU64")\n",
 105                oscc->oscc_last_id, oscc->oscc_next_id);
 106
 107         wake_up(&oscc->oscc_waitq);
 108         RETURN(rc);
 109 }
 110
 111 static int oscc_internal_create(struct osc_creator *oscc)
 112 {
 113         struct ptlrpc_request *request;
 114         struct ost_body *body;
 115         int size = sizeof(*body);
 116         ENTRY;
 117
 118         spin_lock(&oscc->oscc_lock);
 119         if (oscc->oscc_grow_count < OST_MAX_PRECREATE &&
 120             !(oscc->oscc_flags & (OSCC_FLAG_LOW | OSCC_FLAG_RECOVERING)) &&
 121             (__s64)(oscc->oscc_last_id - oscc->oscc_next_id) <=
 122                    (oscc->oscc_grow_count / 4 + 1)) {
 123                 oscc->oscc_flags |= OSCC_FLAG_LOW;
 124                 oscc->oscc_grow_count *= 2;
 125         }
 126
 127         if (oscc->oscc_grow_count > OST_MAX_PRECREATE / 2)
 128                 oscc->oscc_grow_count = OST_MAX_PRECREATE / 2;
 129
 130         if (oscc->oscc_flags & OSCC_FLAG_CREATING ||
 131             oscc->oscc_flags & OSCC_FLAG_RECOVERING) {
 132                 spin_unlock(&oscc->oscc_lock);
 133                 RETURN(0);
 134         }
 135         oscc->oscc_flags |= OSCC_FLAG_CREATING;
 136         spin_unlock(&oscc->oscc_lock);
 137
 138         request = ptlrpc_prep_req(oscc->oscc_obd->u.cli.cl_import,
 139                                   LUSTRE_OBD_VERSION, OST_CREATE,
 140                                   1, &size, NULL);
 141         if (request == NULL) {
 142                 spin_lock(&oscc->oscc_lock);
 143                 oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
 144                 spin_unlock(&oscc->oscc_lock);
 145                 RETURN(-ENOMEM);
 146         }
 147
 148         request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
 149         body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
 150
 151         spin_lock(&oscc->oscc_lock);
 152         body->oa.o_id = oscc->oscc_last_id + oscc->oscc_grow_count;
 153         /* probably we should take frequence of request into account? -bzzz */
 154         if (oscc->oscc_grow_count < oscc->oscc_max_grow_count) {
 155                 oscc->oscc_grow_count *= 2;
 156                 if (oscc->oscc_grow_count > oscc->oscc_max_grow_count)
 157                         oscc->oscc_grow_count = oscc->oscc_max_grow_count;
 158         }
 159         body->oa.o_gr = oscc->oscc_gr;
 160         LASSERT(body->oa.o_gr > 0);
 161         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGROUP;
 162         spin_unlock(&oscc->oscc_lock);
 163         CDEBUG(D_INFO, "preallocating through id "LPU64" (last used "LPU64")\n",
 164                body->oa.o_id, oscc->oscc_next_id);
 165
 166         request->rq_replen = lustre_msg_size(1, &size);
 167
 168         request->rq_async_args.pointer_arg[0] = oscc;
 169         request->rq_interpret_reply = osc_interpret_create;
 170         ptlrpcd_add_req(request);
 171
 172         RETURN(0);
 173 }
 174
 175 static int oscc_has_objects(struct osc_creator *oscc, int count)
 176 {
 177         int have_objs;
 178         spin_lock(&oscc->oscc_lock);
 179         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 180         spin_unlock(&oscc->oscc_lock);
 181
 182         if (!have_objs)
 183                 oscc_internal_create(oscc);
 184
 185         return have_objs;
 186 }
 187
 188 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 189 {
 190         int have_objs;
 191         int ost_full;
 192         int osc_invalid;
 193
 194         have_objs = oscc_has_objects(oscc, count);
 195
 196         spin_lock(&oscc->oscc_lock);
 197         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 198         spin_unlock(&oscc->oscc_lock);
 199
 200         osc_invalid = oscc->oscc_obd->u.cli.cl_import->imp_invalid;
 201
 202         return have_objs || ost_full || osc_invalid;
 203 }
 204
 205 static int oscc_precreate(struct osc_creator *oscc, int wait)
 206 {
 207         struct l_wait_info lwi = { 0 };
 208         int rc = 0;
 209         ENTRY;
 210
 211         if (oscc_has_objects(oscc, oscc->oscc_grow_count / 2))
 212                 RETURN(0);
 213
 214         if (!wait)
 215                 RETURN(0);
 216
 217         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 218         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 219
 220         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 221                 rc = -ENOSPC;
 222
 223         if (oscc->oscc_obd->u.cli.cl_import->imp_invalid)
 224                 rc = -EIO;
 225
 226         RETURN(rc);
 227 }
 228
 229 int oscc_recovering(struct osc_creator *oscc)
 230 {
 231         int recov = 0;
 232
 233         spin_lock(&oscc->oscc_lock);
 234         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 235         spin_unlock(&oscc->oscc_lock);
 236
 237         return recov;
 238 }
 239
 240 int osc_create(struct obd_export *exp, struct obdo *oa,
 241                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 242 {
 243         struct lov_stripe_md *lsm;
 244         struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
 245         int try_again = 1, rc = 0;
 246         ENTRY;
 247         LASSERT(oa);
 248         LASSERT(ea);
 249         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
 250         LASSERT(oa->o_gr > 0);
 251
 252         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 253             oa->o_flags == OBD_FL_RECREATE_OBJS) {
 254                 /* Exceptional case where we are trying to repair missing
 255                  * objects for various groups.  We have already validated that
 256                  * this is a valid group for the file.  Don't set oscc->oscc_gr.
 257                  */
 258                 RETURN(osc_real_create(exp, oa, ea, oti));
 259         }
 260
 261         LASSERT(oscc->oscc_gr == 0 || oscc->oscc_gr == oa->o_gr);
 262         oscc->oscc_gr = oa->o_gr;
 263
 264         if (oa->o_gr == FILTER_GROUP_LLOG || oa->o_gr == FILTER_GROUP_ECHO)
 265                 RETURN(osc_real_create(exp, oa, ea, oti));
 266
 267         /* this is the special case where create removes orphans */
 268         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 269             oa->o_flags == OBD_FL_DELORPHAN) {
 270                 spin_lock(&oscc->oscc_lock);
 271                 if (oscc->oscc_flags & OSCC_FLAG_SYNC_IN_PROGRESS) {
 272                         spin_unlock(&oscc->oscc_lock);
 273                         return -EBUSY;
 274                 }
 275                 if (!(oscc->oscc_flags & OSCC_FLAG_RECOVERING)) {
 276                         spin_unlock(&oscc->oscc_lock);
 277                         return 0;
 278                 }
 279                 oscc->oscc_flags |= OSCC_FLAG_SYNC_IN_PROGRESS;
 280                 spin_unlock(&oscc->oscc_lock);
 281                 CDEBUG(D_HA, "%s; oscc recovery started\n",
 282                        exp->exp_obd->obd_name);
 283                 LASSERT(oscc->oscc_flags & OSCC_FLAG_RECOVERING);
 284
 285                 /* delete from next_id on up */
 286                 oa->o_valid |= OBD_MD_FLID;
 287                 oa->o_id = oscc->oscc_next_id - 1;
 288
 289                 CDEBUG(D_HA, "%s: deleting to next_id: "LPU64"\n",
 290                        oscc->oscc_obd->obd_name, oa->o_id);
 291
 292                 rc = osc_real_create(exp, oa, ea, NULL);
 293                 if (oscc->oscc_obd == NULL) {
 294                         CWARN("the obd for oscc %p has been freed\n", oscc);
 295                         RETURN(rc);
 296                 }
 297
 298                 spin_lock(&oscc->oscc_lock);
 299                 oscc->oscc_flags &= ~OSCC_FLAG_SYNC_IN_PROGRESS;
 300                 if (rc == 0 || rc == -ENOSPC) {
 301                         if (rc == -ENOSPC)
 302                                 oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 303                         oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 304                         oscc->oscc_last_id = oa->o_id;
 305
 306                         CDEBUG(D_HA, "%s: oscc recovery finished: %d\n",
 307                                oscc->oscc_obd->obd_name, rc);
 308                         wake_up(&oscc->oscc_waitq);
 309
 310                 } else {
 311                         CDEBUG(D_ERROR, "%s: oscc recovery failed: %d\n",
 312                                oscc->oscc_obd->obd_name, rc);
 313                 }
 314                 spin_unlock(&oscc->oscc_lock);
 315
 316                 RETURN(rc);
 317         }
 318
 319         lsm = *ea;
 320         if (lsm == NULL) {
 321                 rc = obd_alloc_memmd(exp, &lsm);
 322                 if (rc < 0)
 323                         RETURN(rc);
 324         }
 325
 326         while (try_again) {
 327                 /* If orphans are being recovered, then we must wait until
 328                    it is finished before we can continue with create. */
 329                 if (oscc_recovering(oscc)) {
 330                         struct l_wait_info lwi;
 331
 332                         CDEBUG(D_HA,"%p: oscc recovery in progress, waiting\n",
 333                                oscc);
 334                         lwi = LWI_TIMEOUT(MAX(obd_timeout*HZ/4, 1), NULL, NULL);
 335                         rc = l_wait_event(oscc->oscc_waitq,
 336                                           !oscc_recovering(oscc), &lwi);
 337
 338                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
 339                         if (rc == -ETIMEDOUT) {
 340                                 CDEBUG(D_HA,"%p: timeout waiting on recovery\n",
 341                                        oscc);
 342                                 RETURN(rc);
 343                         }
 344                         CDEBUG(D_HA, "%s: oscc recovery over, waking up\n",
 345                                exp->exp_obd->obd_name);
 346                 }
 347
 348                 spin_lock(&oscc->oscc_lock);
 349                 if (oscc->oscc_flags & OSCC_FLAG_EXITING) {
 350                         spin_unlock(&oscc->oscc_lock);
 351                         break;
 352                 }
 353
 354                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 355                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 356                         oa->o_id = oscc->oscc_next_id;
 357                         oa->o_gr = oscc->oscc_gr;
 358                         lsm->lsm_object_id = oscc->oscc_next_id;
 359                         lsm->lsm_object_gr = oscc->oscc_gr;
 360                         *ea = lsm;
 361                         oscc->oscc_next_id++;
 362                         try_again = 0;
 363                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 364                         rc = -ENOSPC;
 365                         spin_unlock(&oscc->oscc_lock);
 366                         break;
 367                 }
 368                 spin_unlock(&oscc->oscc_lock);
 369                 rc = oscc_precreate(oscc, try_again);
 370                 if (rc)
 371                         break;
 372         }
 373
 374         if (rc == 0)
 375                 CDEBUG(D_HA, "%s: returning objid "LPU64"\n",
 376                        oscc->oscc_obd->u.cli.cl_import->imp_target_uuid.uuid,
 377                        lsm->lsm_object_id);
 378         else if (*ea == NULL)
 379                 obd_free_memmd(exp, &lsm);
 380         RETURN(rc);
 381 }
 382
 383 void oscc_init(struct obd_device *obd)
 384 {
 385         struct osc_creator *oscc;
 386
 387         if (obd == NULL)
 388                 return;
 389
 390         oscc = &obd->u.cli.cl_oscc;
 391
 392         memset(oscc, 0, sizeof(*oscc));
 393         INIT_LIST_HEAD(&oscc->oscc_list);
 394         init_waitqueue_head(&oscc->oscc_waitq);
 395         spin_lock_init(&oscc->oscc_lock);
 396         oscc->oscc_obd = obd;
 397         oscc->oscc_kick_barrier = 100;
 398         oscc->oscc_max_grow_count = 2000;
 399         oscc->oscc_grow_count = OST_MIN_PRECREATE;
 400
 401         oscc->oscc_next_id = 2;
 402         oscc->oscc_last_id = 1;
 403         oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
 404         /* XXX the export handle should give the oscc the last object */
 405         /* oed->oed_oscc.oscc_last_id = exph->....; */
 406 }