drivers/staging/lustre/lustre/llite/llite_close.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/llite/llite_close.c
  37  *
  38  * Lustre Lite routines to issue a secondary close after writeback
  39  */
  40
  41 #include <linux/module.h>
  42
  43 #define DEBUG_SUBSYSTEM S_LLITE
  44
  45 #include <lustre_lite.h>
  46 #include "llite_internal.h"
  47
  48 /** records that a write is in flight */
  49 void vvp_write_pending(struct ccc_object *club, struct ccc_page *page)
  50 {
  51         struct ll_inode_info *lli = ll_i2info(club->cob_inode);
  52
  53         ENTRY;
  54         spin_lock(&lli->lli_lock);
  55         lli->lli_flags |= LLIF_SOM_DIRTY;
  56         if (page != NULL && list_empty(&page->cpg_pending_linkage))
  57                 list_add(&page->cpg_pending_linkage,
  58                              &club->cob_pending_list);
  59         spin_unlock(&lli->lli_lock);
  60         EXIT;
  61 }
  62
  63 /** records that a write has completed */
  64 void vvp_write_complete(struct ccc_object *club, struct ccc_page *page)
  65 {
  66         struct ll_inode_info *lli = ll_i2info(club->cob_inode);
  67         int rc = 0;
  68
  69         ENTRY;
  70         spin_lock(&lli->lli_lock);
  71         if (page != NULL && !list_empty(&page->cpg_pending_linkage)) {
  72                 list_del_init(&page->cpg_pending_linkage);
  73                 rc = 1;
  74         }
  75         spin_unlock(&lli->lli_lock);
  76         if (rc)
  77                 ll_queue_done_writing(club->cob_inode, 0);
  78         EXIT;
  79 }
  80
  81 /** Queues DONE_WRITING if
  82  * - done writing is allowed;
  83  * - inode has no no dirty pages; */
  84 void ll_queue_done_writing(struct inode *inode, unsigned long flags)
  85 {
  86         struct ll_inode_info *lli = ll_i2info(inode);
  87         struct ccc_object *club = cl2ccc(ll_i2info(inode)->lli_clob);
  88         ENTRY;
  89
  90         spin_lock(&lli->lli_lock);
  91         lli->lli_flags |= flags;
  92
  93         if ((lli->lli_flags & LLIF_DONE_WRITING) &&
  94             list_empty(&club->cob_pending_list)) {
  95                 struct ll_close_queue *lcq = ll_i2sbi(inode)->ll_lcq;
  96
  97                 if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
  98                         CWARN("ino %lu/%u(flags %u) som valid it just after "
  99                               "recovery\n",
 100                               inode->i_ino, inode->i_generation,
 101                               lli->lli_flags);
 102                 /* DONE_WRITING is allowed and inode has no dirty page. */
 103                 spin_lock(&lcq->lcq_lock);
 104
 105                 LASSERT(list_empty(&lli->lli_close_list));
 106                 CDEBUG(D_INODE, "adding inode %lu/%u to close list\n",
 107                        inode->i_ino, inode->i_generation);
 108                 list_add_tail(&lli->lli_close_list, &lcq->lcq_head);
 109
 110                 /* Avoid a concurrent insertion into the close thread queue:
 111                  * an inode is already in the close thread, open(), write(),
 112                  * close() happen, epoch is closed as the inode is marked as
 113                  * LLIF_EPOCH_PENDING. When pages are written inode should not
 114                  * be inserted into the queue again, clear this flag to avoid
 115                  * it. */
 116                 lli->lli_flags &= ~LLIF_DONE_WRITING;
 117
 118                 wake_up(&lcq->lcq_waitq);
 119                 spin_unlock(&lcq->lcq_lock);
 120         }
 121         spin_unlock(&lli->lli_lock);
 122         EXIT;
 123 }
 124
 125 /** Pack SOM attributes info @opdata for CLOSE, DONE_WRITING rpc. */
 126 void ll_done_writing_attr(struct inode *inode, struct md_op_data *op_data)
 127 {
 128         struct ll_inode_info *lli = ll_i2info(inode);
 129         ENTRY;
 130
 131         op_data->op_flags |= MF_SOM_CHANGE;
 132         /* Check if Size-on-MDS attributes are valid. */
 133         if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
 134                 CERROR("ino %lu/%u(flags %u) som valid it just after "
 135                        "recovery\n", inode->i_ino, inode->i_generation,
 136                        lli->lli_flags);
 137
 138         if (!cl_local_size(inode)) {
 139                 /* Send Size-on-MDS Attributes if valid. */
 140                 op_data->op_attr.ia_valid |= ATTR_MTIME_SET | ATTR_CTIME_SET |
 141                                 ATTR_ATIME_SET | ATTR_SIZE | ATTR_BLOCKS;
 142         }
 143         EXIT;
 144 }
 145
 146 /** Closes ioepoch and packs Size-on-MDS attribute if needed into @op_data. */
 147 void ll_ioepoch_close(struct inode *inode, struct md_op_data *op_data,
 148                       struct obd_client_handle **och, unsigned long flags)
 149 {
 150         struct ll_inode_info *lli = ll_i2info(inode);
 151         struct ccc_object *club = cl2ccc(ll_i2info(inode)->lli_clob);
 152         ENTRY;
 153
 154         spin_lock(&lli->lli_lock);
 155         if (!(list_empty(&club->cob_pending_list))) {
 156                 if (!(lli->lli_flags & LLIF_EPOCH_PENDING)) {
 157                         LASSERT(*och != NULL);
 158                         LASSERT(lli->lli_pending_och == NULL);
 159                         /* Inode is dirty and there is no pending write done
 160                          * request yet, DONE_WRITE is to be sent later. */
 161                         lli->lli_flags |= LLIF_EPOCH_PENDING;
 162                         lli->lli_pending_och = *och;
 163                         spin_unlock(&lli->lli_lock);
 164
 165                         inode = igrab(inode);
 166                         LASSERT(inode);
 167                         GOTO(out, 0);
 168                 }
 169                 if (flags & LLIF_DONE_WRITING) {
 170                         /* Some pages are still dirty, it is early to send
 171                          * DONE_WRITE. Wait untill all pages will be flushed
 172                          * and try DONE_WRITE again later. */
 173                         LASSERT(!(lli->lli_flags & LLIF_DONE_WRITING));
 174                         lli->lli_flags |= LLIF_DONE_WRITING;
 175                         spin_unlock(&lli->lli_lock);
 176
 177                         inode = igrab(inode);
 178                         LASSERT(inode);
 179                         GOTO(out, 0);
 180                 }
 181         }
 182         CDEBUG(D_INODE, "Epoch "LPU64" closed on "DFID"\n",
 183                ll_i2info(inode)->lli_ioepoch, PFID(&lli->lli_fid));
 184         op_data->op_flags |= MF_EPOCH_CLOSE;
 185
 186         if (flags & LLIF_DONE_WRITING) {
 187                 LASSERT(lli->lli_flags & LLIF_SOM_DIRTY);
 188                 LASSERT(!(lli->lli_flags & LLIF_DONE_WRITING));
 189                 *och = lli->lli_pending_och;
 190                 lli->lli_pending_och = NULL;
 191                 lli->lli_flags &= ~LLIF_EPOCH_PENDING;
 192         } else {
 193                 /* Pack Size-on-MDS inode attributes only if they has changed */
 194                 if (!(lli->lli_flags & LLIF_SOM_DIRTY)) {
 195                         spin_unlock(&lli->lli_lock);
 196                         GOTO(out, 0);
 197                 }
 198
 199                 /* There is a pending DONE_WRITE -- close epoch with no
 200                  * attribute change. */
 201                 if (lli->lli_flags & LLIF_EPOCH_PENDING) {
 202                         spin_unlock(&lli->lli_lock);
 203                         GOTO(out, 0);
 204                 }
 205         }
 206
 207         LASSERT(list_empty(&club->cob_pending_list));
 208         lli->lli_flags &= ~LLIF_SOM_DIRTY;
 209         spin_unlock(&lli->lli_lock);
 210         ll_done_writing_attr(inode, op_data);
 211
 212         EXIT;
 213 out:
 214         return;
 215 }
 216
 217 /**
 218  * Cliens updates SOM attributes on MDS (including llog cookies):
 219  * obd_getattr with no lock and md_setattr.
 220  */
 221 int ll_som_update(struct inode *inode, struct md_op_data *op_data)
 222 {
 223         struct ll_inode_info *lli = ll_i2info(inode);
 224         struct ptlrpc_request *request = NULL;
 225         __u32 old_flags;
 226         struct obdo *oa;
 227         int rc;
 228         ENTRY;
 229
 230         LASSERT(op_data != NULL);
 231         if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
 232                 CERROR("ino %lu/%u(flags %u) som valid it just after "
 233                        "recovery\n", inode->i_ino, inode->i_generation,
 234                        lli->lli_flags);
 235
 236         OBDO_ALLOC(oa);
 237         if (!oa) {
 238                 CERROR("can't allocate memory for Size-on-MDS update.\n");
 239                 RETURN(-ENOMEM);
 240         }
 241
 242         old_flags = op_data->op_flags;
 243         op_data->op_flags = MF_SOM_CHANGE;
 244
 245         /* If inode is already in another epoch, skip getattr from OSTs. */
 246         if (lli->lli_ioepoch == op_data->op_ioepoch) {
 247                 rc = ll_inode_getattr(inode, oa, op_data->op_ioepoch,
 248                                       old_flags & MF_GETATTR_LOCK);
 249                 if (rc) {
 250                         oa->o_valid = 0;
 251                         if (rc != -ENOENT)
 252                                 CERROR("inode_getattr failed (%d): unable to "
 253                                        "send a Size-on-MDS attribute update "
 254                                        "for inode %lu/%u\n", rc, inode->i_ino,
 255                                        inode->i_generation);
 256                 } else {
 257                         CDEBUG(D_INODE, "Size-on-MDS update on "DFID"\n",
 258                                PFID(&lli->lli_fid));
 259                 }
 260                 /* Install attributes into op_data. */
 261                 md_from_obdo(op_data, oa, oa->o_valid);
 262         }
 263
 264         rc = md_setattr(ll_i2sbi(inode)->ll_md_exp, op_data,
 265                         NULL, 0, NULL, 0, &request, NULL);
 266         ptlrpc_req_finished(request);
 267
 268         OBDO_FREE(oa);
 269         RETURN(rc);
 270 }
 271
 272 /**
 273  * Closes the ioepoch and packs all the attributes into @op_data for
 274  * DONE_WRITING rpc.
 275  */
 276 static void ll_prepare_done_writing(struct inode *inode,
 277                                     struct md_op_data *op_data,
 278                                     struct obd_client_handle **och)
 279 {
 280         ll_ioepoch_close(inode, op_data, och, LLIF_DONE_WRITING);
 281         /* If there is no @och, we do not do D_W yet. */
 282         if (*och == NULL)
 283                 return;
 284
 285         ll_pack_inode2opdata(inode, op_data, &(*och)->och_fh);
 286         ll_prep_md_op_data(op_data, inode, NULL, NULL,
 287                            0, 0, LUSTRE_OPC_ANY, NULL);
 288 }
 289
 290 /** Send a DONE_WRITING rpc. */
 291 static void ll_done_writing(struct inode *inode)
 292 {
 293         struct obd_client_handle *och = NULL;
 294         struct md_op_data *op_data;
 295         int rc;
 296         ENTRY;
 297
 298         LASSERT(exp_connect_som(ll_i2mdexp(inode)));
 299
 300         OBD_ALLOC_PTR(op_data);
 301         if (op_data == NULL) {
 302                 CERROR("can't allocate op_data\n");
 303                 EXIT;
 304                 return;
 305         }
 306
 307         ll_prepare_done_writing(inode, op_data, &och);
 308         /* If there is no @och, we do not do D_W yet. */
 309         if (och == NULL)
 310                 GOTO(out, 0);
 311
 312         rc = md_done_writing(ll_i2sbi(inode)->ll_md_exp, op_data, NULL);
 313         if (rc == -EAGAIN) {
 314                 /* MDS has instructed us to obtain Size-on-MDS attribute from
 315                  * OSTs and send setattr to back to MDS. */
 316                 rc = ll_som_update(inode, op_data);
 317         } else if (rc) {
 318                 CERROR("inode %lu mdc done_writing failed: rc = %d\n",
 319                        inode->i_ino, rc);
 320         }
 321 out:
 322         ll_finish_md_op_data(op_data);
 323         if (och) {
 324                 md_clear_open_replay_data(ll_i2sbi(inode)->ll_md_exp, och);
 325                 OBD_FREE_PTR(och);
 326         }
 327         EXIT;
 328 }
 329
 330 static struct ll_inode_info *ll_close_next_lli(struct ll_close_queue *lcq)
 331 {
 332         struct ll_inode_info *lli = NULL;
 333
 334         spin_lock(&lcq->lcq_lock);
 335
 336         if (!list_empty(&lcq->lcq_head)) {
 337                 lli = list_entry(lcq->lcq_head.next, struct ll_inode_info,
 338                                      lli_close_list);
 339                 list_del_init(&lli->lli_close_list);
 340         } else if (atomic_read(&lcq->lcq_stop))
 341                 lli = ERR_PTR(-EALREADY);
 342
 343         spin_unlock(&lcq->lcq_lock);
 344         return lli;
 345 }
 346
 347 static int ll_close_thread(void *arg)
 348 {
 349         struct ll_close_queue *lcq = arg;
 350         ENTRY;
 351
 352         complete(&lcq->lcq_comp);
 353
 354         while (1) {
 355                 struct l_wait_info lwi = { 0 };
 356                 struct ll_inode_info *lli;
 357                 struct inode *inode;
 358
 359                 l_wait_event_exclusive(lcq->lcq_waitq,
 360                                        (lli = ll_close_next_lli(lcq)) != NULL,
 361                                        &lwi);
 362                 if (IS_ERR(lli))
 363                         break;
 364
 365                 inode = ll_info2i(lli);
 366                 CDEBUG(D_INFO, "done_writting for inode %lu/%u\n",
 367                        inode->i_ino, inode->i_generation);
 368                 ll_done_writing(inode);
 369                 iput(inode);
 370         }
 371
 372         CDEBUG(D_INFO, "ll_close exiting\n");
 373         complete(&lcq->lcq_comp);
 374         RETURN(0);
 375 }
 376
 377 int ll_close_thread_start(struct ll_close_queue **lcq_ret)
 378 {
 379         struct ll_close_queue *lcq;
 380         task_t *task;
 381
 382         if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CLOSE_THREAD))
 383                 return -EINTR;
 384
 385         OBD_ALLOC(lcq, sizeof(*lcq));
 386         if (lcq == NULL)
 387                 return -ENOMEM;
 388
 389         spin_lock_init(&lcq->lcq_lock);
 390         INIT_LIST_HEAD(&lcq->lcq_head);
 391         init_waitqueue_head(&lcq->lcq_waitq);
 392         init_completion(&lcq->lcq_comp);
 393
 394         task = kthread_run(ll_close_thread, lcq, "ll_close");
 395         if (IS_ERR(task)) {
 396                 OBD_FREE(lcq, sizeof(*lcq));
 397                 return PTR_ERR(task);
 398         }
 399
 400         wait_for_completion(&lcq->lcq_comp);
 401         *lcq_ret = lcq;
 402         return 0;
 403 }
 404
 405 void ll_close_thread_shutdown(struct ll_close_queue *lcq)
 406 {
 407         init_completion(&lcq->lcq_comp);
 408         atomic_inc(&lcq->lcq_stop);
 409         wake_up(&lcq->lcq_waitq);
 410         wait_for_completion(&lcq->lcq_comp);
 411         OBD_FREE(lcq, sizeof(*lcq));
 412 }