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