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/lustre_dlm.h>
  40 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  41 #  include <linux/workqueue.h>
  42 #  include <linux/smp_lock.h>
  43 # else
  44 #  include <linux/locks.h>
  45 # endif
  46 #else /* __KERNEL__ */
  47 # include <liblustre.h>
  48 #endif
  49
  50 #ifndef  __CYGWIN__
  51 # include <linux/ctype.h>
  52 # include <linux/init.h>
  53 #else
  54 # include <ctype.h>
  55 #endif
  56
  57 #include <linux/obd_class.h>
  58 #include "osc_internal.h"
  59
  60 static int osc_interpret_create(struct ptlrpc_request *req, void *data,
  61                                 int rc)
  62 {
  63         struct osc_creator *oscc;
  64         struct ost_body *body = NULL;
  65         ENTRY;
  66
  67         if (req->rq_repmsg) {
  68                 body = lustre_swab_repbuf(req, 0, sizeof(*body),
  69                                           lustre_swab_ost_body);
  70                 if (body == NULL && rc == 0)
  71                         rc = -EPROTO;
  72         }
  73
  74         oscc = req->rq_async_args.pointer_arg[0];
  75         spin_lock(&oscc->oscc_lock);
  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         } else if (rc != 0 && rc != -EIO) {
  82                 DEBUG_REQ(D_ERROR, req,
  83                           "unknown rc %d from async create: failing oscc",
  84                           rc);
  85                 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
  86                 ptlrpc_fail_import(req->rq_import, req->rq_import_generation);
  87         }
  88         oscc->oscc_flags &= ~OSCC_FLAG_CREATING;
  89         spin_unlock(&oscc->oscc_lock);
  90
  91         CDEBUG(D_INFO, "preallocated through id "LPU64" (last used "LPU64")\n",
  92                oscc->oscc_last_id, oscc->oscc_next_id);
  93
  94         wake_up(&oscc->oscc_waitq);
  95         RETURN(rc);
  96 }
  97
  98 static int oscc_internal_create(struct osc_creator *oscc)
  99 {
 100         struct ptlrpc_request *request;
 101         struct ost_body *body;
 102         int size = sizeof(*body);
 103         ENTRY;
 104
 105         spin_lock(&oscc->oscc_lock);
 106         if (oscc->oscc_flags & OSCC_FLAG_CREATING) {
 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(class_exp2cliimp(oscc->oscc_exp), 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         body->oa.o_valid |= OBD_MD_FLID;
 128         CDEBUG(D_INFO, "preallocating through id "LPU64" (last used "LPU64")\n",
 129                body->oa.o_id, oscc->oscc_next_id);
 130         spin_unlock(&oscc->oscc_lock);
 131
 132         request->rq_replen = lustre_msg_size(1, &size);
 133
 134         request->rq_async_args.pointer_arg[0] = oscc;
 135         request->rq_interpret_reply = osc_interpret_create;
 136         ptlrpcd_add_req(request);
 137
 138         RETURN(0);
 139 }
 140
 141 static int oscc_has_objects(struct osc_creator *oscc, int count)
 142 {
 143         int have_objs;
 144         spin_lock(&oscc->oscc_lock);
 145         have_objs = ((__s64)(oscc->oscc_last_id - oscc->oscc_next_id) >= count);
 146         spin_unlock(&oscc->oscc_lock);
 147
 148         if (!have_objs)
 149                 oscc_internal_create(oscc);
 150
 151         return have_objs;
 152 }
 153
 154 static int oscc_wait_for_objects(struct osc_creator *oscc, int count)
 155 {
 156         int have_objs;
 157         int ost_full;
 158         int osc_invalid;
 159
 160         have_objs = oscc_has_objects(oscc, count);
 161
 162         spin_lock(&oscc->oscc_lock);
 163         ost_full = (oscc->oscc_flags & OSCC_FLAG_NOSPC);
 164         spin_unlock(&oscc->oscc_lock);
 165
 166         osc_invalid = class_exp2cliimp(oscc->oscc_exp)->imp_invalid;
 167
 168         return have_objs || ost_full || osc_invalid;
 169 }
 170
 171 static int oscc_precreate(struct osc_creator *oscc, int wait)
 172 {
 173         struct l_wait_info lwi = { 0 };
 174         int rc = 0;
 175         ENTRY;
 176
 177         if (oscc_has_objects(oscc, oscc->oscc_kick_barrier))
 178                 RETURN(0);
 179
 180         if (!wait)
 181                 RETURN(0);
 182
 183         /* no rc check -- a no-INTR, no-TIMEOUT wait can't fail */
 184         l_wait_event(oscc->oscc_waitq, oscc_wait_for_objects(oscc, 1), &lwi);
 185
 186         if (!oscc_has_objects(oscc, 1) && (oscc->oscc_flags & OSCC_FLAG_NOSPC))
 187                 rc = -ENOSPC;
 188
 189         if (class_exp2cliimp(oscc->oscc_exp)->imp_invalid)
 190                 rc = -EIO;
 191
 192         RETURN(rc);
 193 }
 194
 195 int oscc_recovering(struct osc_creator *oscc)
 196 {
 197         int recov = 0;
 198
 199         spin_lock(&oscc->oscc_lock);
 200         recov = oscc->oscc_flags & OSCC_FLAG_RECOVERING;
 201         spin_unlock(&oscc->oscc_lock);
 202
 203         return recov;
 204 }
 205
 206 int osc_create(struct obd_export *exp, struct obdo *oa,
 207                struct lov_stripe_md **ea, struct obd_trans_info *oti)
 208 {
 209         struct lov_stripe_md *lsm;
 210         struct osc_creator *oscc = &exp->u.eu_osc_data.oed_oscc;
 211         int try_again = 1, rc = 0;
 212         ENTRY;
 213         LASSERT(oa);
 214         LASSERT(ea);
 215
 216         if ((oa->o_valid & OBD_MD_FLGROUP) && (oa->o_gr != 0))
 217                 RETURN(osc_real_create(exp, oa, ea, oti));
 218
 219         lsm = *ea;
 220         if (lsm == NULL) {
 221                 rc = obd_alloc_memmd(exp, &lsm);
 222                 if (rc < 0)
 223                         RETURN(rc);
 224         }
 225
 226         /* this is the special case where create removes orphans */
 227         if ((oa->o_valid & OBD_MD_FLFLAGS) &&
 228             oa->o_flags == OBD_FL_DELORPHAN) {
 229                 CDEBUG(D_HA, "%p: oscc recovery started\n", oscc);
 230                 /* delete from next_id on up */
 231                 oa->o_valid |= OBD_MD_FLID;
 232                 oa->o_id = oscc->oscc_next_id - 1;
 233
 234                 rc = osc_real_create(oscc->oscc_exp, oa, ea, NULL);
 235
 236                 spin_lock(&oscc->oscc_lock);
 237                 if (rc == -ENOSPC)
 238                         oscc->oscc_flags |= OSCC_FLAG_NOSPC;
 239                 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
 240                 oscc->oscc_last_id = oa->o_id;
 241                 wake_up(&oscc->oscc_waitq);
 242                 spin_unlock(&oscc->oscc_lock);
 243
 244                 CDEBUG(D_HA, "%p: oscc recovery finished\n", oscc);
 245
 246                 RETURN(rc);
 247         }
 248
 249         /* If orphans are being recovered, then we must wait until it is
 250            finished before we can continue with create. */
 251         if (oscc_recovering(oscc)) {
 252                 struct l_wait_info lwi;
 253
 254                 CDEBUG(D_HA, "%p: oscc recovery in progress, waiting\n", oscc);
 255
 256                 lwi = LWI_TIMEOUT(MAX(obd_timeout * HZ, 1), NULL, NULL);
 257                 rc = l_wait_event(oscc->oscc_waitq, !oscc_recovering(oscc),
 258                                   &lwi);
 259                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 260                 if (rc == -ETIMEDOUT)
 261                         RETURN(rc);
 262                 CDEBUG(D_HA, "%p: oscc recovery over, waking up\n", oscc);
 263         }
 264
 265
 266         while (try_again) {
 267                 spin_lock(&oscc->oscc_lock);
 268                 if (oscc->oscc_last_id >= oscc->oscc_next_id) {
 269                         memcpy(oa, &oscc->oscc_oa, sizeof(*oa));
 270                         oa->o_id = oscc->oscc_next_id;
 271                         lsm->lsm_object_id = oscc->oscc_next_id;
 272                         *ea = lsm;
 273                         oscc->oscc_next_id++;
 274                         try_again = 0;
 275                 } else if (oscc->oscc_flags & OSCC_FLAG_NOSPC) {
 276                         rc = -ENOSPC;
 277                         spin_unlock(&oscc->oscc_lock);
 278                         break;
 279                 }
 280                 spin_unlock(&oscc->oscc_lock);
 281                 rc = oscc_precreate(oscc, try_again);
 282                 if (rc == -EIO)
 283                         break;
 284         }
 285
 286         if (rc == 0)
 287                 CDEBUG(D_INFO, "returning objid "LPU64"\n", lsm->lsm_object_id);
 288         else if (*ea == NULL)
 289                 obd_free_memmd(exp, &lsm);
 290         RETURN(rc);
 291 }
 292
 293 void oscc_init(struct obd_export *exp)
 294 {
 295         struct osc_export_data *oed;
 296
 297         if (exp == NULL)
 298                 return;
 299
 300         oed = &exp->exp_osc_data;
 301         memset(oed, 0, sizeof(*oed));
 302         INIT_LIST_HEAD(&oed->oed_oscc.oscc_list);
 303         init_waitqueue_head(&oed->oed_oscc.oscc_waitq);
 304         spin_lock_init(&oed->oed_oscc.oscc_lock);
 305         oed->oed_oscc.oscc_exp = exp;
 306         oed->oed_oscc.oscc_kick_barrier = 100;
 307         oed->oed_oscc.oscc_grow_count = 2000;
 308         oed->oed_oscc.oscc_initial_create_count = 2000;
 309
 310         oed->oed_oscc.oscc_next_id = 2;
 311         oed->oed_oscc.oscc_last_id = 1;
 312         oed->oed_oscc.oscc_flags |= OSCC_FLAG_RECOVERING;
 313         /* XXX the export handle should give the oscc the last object */
 314         /* oed->oed_oscc.oscc_last_id = exph->....; */
 315 }