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