LXR linux/drivers/staging/lustre/lustre/llite/llite

   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.gnu.org/licenses/gpl-2.0.html
  19 *
  20 * GPL HEADER END
  21 */
  22/*
  23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24 * Use is subject to license terms.
  25 *
  26 * Copyright (c) 2011, 2012, Intel Corporation.
  27 */
  28/*
  29 * This file is part of Lustre, http://www.lustre.org/
  30 * Lustre is a trademark of Sun Microsystems, Inc.
  31 *
  32 * lustre/llite/llite_close.c
  33 *
  34 * Lustre Lite routines to issue a secondary close after writeback
  35 */
  36
  37#include <linux/module.h>
  38
  39#define DEBUG_SUBSYSTEM S_LLITE
  40
  41#include "llite_internal.h"
  42
  43/** records that a write is in flight */
  44void vvp_write_pending(struct vvp_object *club, struct vvp_page *page)
  45{
  46        struct ll_inode_info *lli = ll_i2info(club->vob_inode);
  47
  48        spin_lock(&lli->lli_lock);
  49        lli->lli_flags |= LLIF_SOM_DIRTY;
  50        if (page && list_empty(&page->vpg_pending_linkage))
  51                list_add(&page->vpg_pending_linkage, &club->vob_pending_list);
  52        spin_unlock(&lli->lli_lock);
  53}
  54
  55/** records that a write has completed */
  56void vvp_write_complete(struct vvp_object *club, struct vvp_page *page)
  57{
  58        struct ll_inode_info *lli = ll_i2info(club->vob_inode);
  59        int rc = 0;
  60
  61        spin_lock(&lli->lli_lock);
  62        if (page && !list_empty(&page->vpg_pending_linkage)) {
  63                list_del_init(&page->vpg_pending_linkage);
  64                rc = 1;
  65        }
  66        spin_unlock(&lli->lli_lock);
  67        if (rc)
  68                ll_queue_done_writing(club->vob_inode, 0);
  69}
  70
  71/** Queues DONE_WRITING if
  72 * - done writing is allowed;
  73 * - inode has no no dirty pages;
  74 */
  75void ll_queue_done_writing(struct inode *inode, unsigned long flags)
  76{
  77        struct ll_inode_info *lli = ll_i2info(inode);
  78        struct vvp_object *club = cl2vvp(ll_i2info(inode)->lli_clob);
  79
  80        spin_lock(&lli->lli_lock);
  81        lli->lli_flags |= flags;
  82
  83        if ((lli->lli_flags & LLIF_DONE_WRITING) &&
  84            list_empty(&club->vob_pending_list)) {
  85                struct ll_close_queue *lcq = ll_i2sbi(inode)->ll_lcq;
  86
  87                if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
  88                        CWARN("%s: file "DFID"(flags %u) Size-on-MDS valid, done writing allowed and no diry pages\n",
  89                              ll_get_fsname(inode->i_sb, NULL, 0),
  90                              PFID(ll_inode2fid(inode)), lli->lli_flags);
  91                /* DONE_WRITING is allowed and inode has no dirty page. */
  92                spin_lock(&lcq->lcq_lock);
  93
  94                LASSERT(list_empty(&lli->lli_close_list));
  95                CDEBUG(D_INODE, "adding inode "DFID" to close list\n",
  96                       PFID(ll_inode2fid(inode)));
  97                list_add_tail(&lli->lli_close_list, &lcq->lcq_head);
  98
  99                /* Avoid a concurrent insertion into the close thread queue:
 100                 * an inode is already in the close thread, open(), write(),
 101                 * close() happen, epoch is closed as the inode is marked as
 102                 * LLIF_EPOCH_PENDING. When pages are written inode should not
 103                 * be inserted into the queue again, clear this flag to avoid
 104                 * it.
 105                 */
 106                lli->lli_flags &= ~LLIF_DONE_WRITING;
 107
 108                wake_up(&lcq->lcq_waitq);
 109                spin_unlock(&lcq->lcq_lock);
 110        }
 111        spin_unlock(&lli->lli_lock);
 112}
 113
 114/** Pack SOM attributes info @opdata for CLOSE, DONE_WRITING rpc. */
 115void ll_done_writing_attr(struct inode *inode, struct md_op_data *op_data)
 116{
 117        struct ll_inode_info *lli = ll_i2info(inode);
 118
 119        op_data->op_flags |= MF_SOM_CHANGE;
 120        /* Check if Size-on-MDS attributes are valid. */
 121        if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
 122                CERROR("%s: inode "DFID"(flags %u) MDS holds lock on Size-on-MDS attributes\n",
 123                       ll_get_fsname(inode->i_sb, NULL, 0),
 124                       PFID(ll_inode2fid(inode)), lli->lli_flags);
 125
 126        if (!cl_local_size(inode)) {
 127                /* Send Size-on-MDS Attributes if valid. */
 128                op_data->op_attr.ia_valid |= ATTR_MTIME_SET | ATTR_CTIME_SET |
 129                                ATTR_ATIME_SET | ATTR_SIZE | ATTR_BLOCKS;
 130        }
 131}
 132
 133/** Closes ioepoch and packs Size-on-MDS attribute if needed into @op_data. */
 134void ll_ioepoch_close(struct inode *inode, struct md_op_data *op_data,
 135                      struct obd_client_handle **och, unsigned long flags)
 136{
 137        struct ll_inode_info *lli = ll_i2info(inode);
 138        struct vvp_object *club = cl2vvp(ll_i2info(inode)->lli_clob);
 139
 140        spin_lock(&lli->lli_lock);
 141        if (!(list_empty(&club->vob_pending_list))) {
 142                if (!(lli->lli_flags & LLIF_EPOCH_PENDING)) {
 143                        LASSERT(*och);
 144                        LASSERT(!lli->lli_pending_och);
 145                        /* Inode is dirty and there is no pending write done
 146                         * request yet, DONE_WRITE is to be sent later.
 147                         */
 148                        lli->lli_flags |= LLIF_EPOCH_PENDING;
 149                        lli->lli_pending_och = *och;
 150                        spin_unlock(&lli->lli_lock);
 151
 152                        inode = igrab(inode);
 153                        LASSERT(inode);
 154                        goto out;
 155                }
 156                if (flags & LLIF_DONE_WRITING) {
 157                        /* Some pages are still dirty, it is early to send
 158                         * DONE_WRITE. Wait until all pages will be flushed
 159                         * and try DONE_WRITE again later.
 160                         */
 161                        LASSERT(!(lli->lli_flags & LLIF_DONE_WRITING));
 162                        lli->lli_flags |= LLIF_DONE_WRITING;
 163                        spin_unlock(&lli->lli_lock);
 164
 165                        inode = igrab(inode);
 166                        LASSERT(inode);
 167                        goto out;
 168                }
 169        }
 170        CDEBUG(D_INODE, "Epoch %llu closed on "DFID"\n",
 171               ll_i2info(inode)->lli_ioepoch, PFID(&lli->lli_fid));
 172        op_data->op_flags |= MF_EPOCH_CLOSE;
 173
 174        if (flags & LLIF_DONE_WRITING) {
 175                LASSERT(lli->lli_flags & LLIF_SOM_DIRTY);
 176                LASSERT(!(lli->lli_flags & LLIF_DONE_WRITING));
 177                *och = lli->lli_pending_och;
 178                lli->lli_pending_och = NULL;
 179                lli->lli_flags &= ~LLIF_EPOCH_PENDING;
 180        } else {
 181                /* Pack Size-on-MDS inode attributes only if they has changed */
 182                if (!(lli->lli_flags & LLIF_SOM_DIRTY)) {
 183                        spin_unlock(&lli->lli_lock);
 184                        goto out;
 185                }
 186
 187                /* There is a pending DONE_WRITE -- close epoch with no
 188                 * attribute change.
 189                 */
 190                if (lli->lli_flags & LLIF_EPOCH_PENDING) {
 191                        spin_unlock(&lli->lli_lock);
 192                        goto out;
 193                }
 194        }
 195
 196        LASSERT(list_empty(&club->vob_pending_list));
 197        lli->lli_flags &= ~LLIF_SOM_DIRTY;
 198        spin_unlock(&lli->lli_lock);
 199        ll_done_writing_attr(inode, op_data);
 200
 201out:
 202        return;
 203}
 204
 205/**
 206 * Cliens updates SOM attributes on MDS (including llog cookies):
 207 * obd_getattr with no lock and md_setattr.
 208 */
 209int ll_som_update(struct inode *inode, struct md_op_data *op_data)
 210{
 211        struct ll_inode_info *lli = ll_i2info(inode);
 212        struct ptlrpc_request *request = NULL;
 213        __u32 old_flags;
 214        struct obdo *oa;
 215        int rc;
 216
 217        LASSERT(op_data);
 218        if (lli->lli_flags & LLIF_MDS_SIZE_LOCK)
 219                CERROR("%s: inode "DFID"(flags %u) MDS holds lock on Size-on-MDS attributes\n",
 220                       ll_get_fsname(inode->i_sb, NULL, 0),
 221                       PFID(ll_inode2fid(inode)), lli->lli_flags);
 222
 223        oa = kmem_cache_zalloc(obdo_cachep, GFP_NOFS);
 224        if (!oa) {
 225                CERROR("can't allocate memory for Size-on-MDS update.\n");
 226                return -ENOMEM;
 227        }
 228
 229        old_flags = op_data->op_flags;
 230        op_data->op_flags = MF_SOM_CHANGE;
 231
 232        /* If inode is already in another epoch, skip getattr from OSTs. */
 233        if (lli->lli_ioepoch == op_data->op_ioepoch) {
 234                rc = ll_inode_getattr(inode, oa, op_data->op_ioepoch,
 235                                      old_flags & MF_GETATTR_LOCK);
 236                if (rc) {
 237                        oa->o_valid = 0;
 238                        if (rc != -ENOENT)
 239                                CERROR("%s: inode_getattr failed - unable to send a Size-on-MDS attribute update for inode "DFID": rc = %d\n",
 240                                       ll_get_fsname(inode->i_sb, NULL, 0),
 241                                       PFID(ll_inode2fid(inode)), rc);
 242                } else {
 243                        CDEBUG(D_INODE, "Size-on-MDS update on "DFID"\n",
 244                               PFID(&lli->lli_fid));
 245                }
 246                /* Install attributes into op_data. */
 247                md_from_obdo(op_data, oa, oa->o_valid);
 248        }
 249
 250        rc = md_setattr(ll_i2sbi(inode)->ll_md_exp, op_data,
 251                        NULL, 0, NULL, 0, &request, NULL);
 252        ptlrpc_req_finished(request);
 253
 254        kmem_cache_free(obdo_cachep, oa);
 255        return rc;
 256}
 257
 258/**
 259 * Closes the ioepoch and packs all the attributes into @op_data for
 260 * DONE_WRITING rpc.
 261 */
 262static void ll_prepare_done_writing(struct inode *inode,
 263                                    struct md_op_data *op_data,
 264                                    struct obd_client_handle **och)
 265{
 266        ll_ioepoch_close(inode, op_data, och, LLIF_DONE_WRITING);
 267        /* If there is no @och, we do not do D_W yet. */
 268        if (!*och)
 269                return;
 270
 271        ll_pack_inode2opdata(inode, op_data, &(*och)->och_fh);
 272        ll_prep_md_op_data(op_data, inode, NULL, NULL,
 273                           0, 0, LUSTRE_OPC_ANY, NULL);
 274}
 275
 276/** Send a DONE_WRITING rpc. */
 277static void ll_done_writing(struct inode *inode)
 278{
 279        struct obd_client_handle *och = NULL;
 280        struct md_op_data *op_data;
 281        int rc;
 282
 283        LASSERT(exp_connect_som(ll_i2mdexp(inode)));
 284
 285        op_data = kzalloc(sizeof(*op_data), GFP_NOFS);
 286        if (!op_data)
 287                return;
 288
 289        ll_prepare_done_writing(inode, op_data, &och);
 290        /* If there is no @och, we do not do D_W yet. */
 291        if (!och)
 292                goto out;
 293
 294        rc = md_done_writing(ll_i2sbi(inode)->ll_md_exp, op_data, NULL);
 295        if (rc == -EAGAIN)
 296                /* MDS has instructed us to obtain Size-on-MDS attribute from
 297                 * OSTs and send setattr to back to MDS.
 298                 */
 299                rc = ll_som_update(inode, op_data);
 300        else if (rc) {
 301                CERROR("%s: inode "DFID" mdc done_writing failed: rc = %d\n",
 302                       ll_get_fsname(inode->i_sb, NULL, 0),
 303                       PFID(ll_inode2fid(inode)), rc);
 304        }
 305out:
 306        ll_finish_md_op_data(op_data);
 307        if (och) {
 308                md_clear_open_replay_data(ll_i2sbi(inode)->ll_md_exp, och);
 309                kfree(och);
 310        }
 311}
 312
 313static struct ll_inode_info *ll_close_next_lli(struct ll_close_queue *lcq)
 314{
 315        struct ll_inode_info *lli = NULL;
 316
 317        spin_lock(&lcq->lcq_lock);
 318
 319        if (!list_empty(&lcq->lcq_head)) {
 320                lli = list_entry(lcq->lcq_head.next, struct ll_inode_info,
 321                                 lli_close_list);
 322                list_del_init(&lli->lli_close_list);
 323        } else if (atomic_read(&lcq->lcq_stop)) {
 324                lli = ERR_PTR(-EALREADY);
 325        }
 326
 327        spin_unlock(&lcq->lcq_lock);
 328        return lli;
 329}
 330
 331static int ll_close_thread(void *arg)
 332{
 333        struct ll_close_queue *lcq = arg;
 334
 335        complete(&lcq->lcq_comp);
 336
 337        while (1) {
 338                struct l_wait_info lwi = { 0 };
 339                struct ll_inode_info *lli;
 340                struct inode *inode;
 341
 342                l_wait_event_exclusive(lcq->lcq_waitq,
 343                                       (lli = ll_close_next_lli(lcq)) != NULL,
 344                                       &lwi);
 345                if (IS_ERR(lli))
 346                        break;
 347
 348                inode = ll_info2i(lli);
 349                CDEBUG(D_INFO, "done_writing for inode "DFID"\n",
 350                       PFID(ll_inode2fid(inode)));
 351                ll_done_writing(inode);
 352                iput(inode);
 353        }
 354
 355        CDEBUG(D_INFO, "ll_close exiting\n");
 356        complete(&lcq->lcq_comp);
 357        return 0;
 358}
 359
 360int ll_close_thread_start(struct ll_close_queue **lcq_ret)
 361{
 362        struct ll_close_queue *lcq;
 363        struct task_struct *task;
 364
 365        if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CLOSE_THREAD))
 366                return -EINTR;
 367
 368        lcq = kzalloc(sizeof(*lcq), GFP_NOFS);
 369        if (!lcq)
 370                return -ENOMEM;
 371
 372        spin_lock_init(&lcq->lcq_lock);
 373        INIT_LIST_HEAD(&lcq->lcq_head);
 374        init_waitqueue_head(&lcq->lcq_waitq);
 375        init_completion(&lcq->lcq_comp);
 376
 377        task = kthread_run(ll_close_thread, lcq, "ll_close");
 378        if (IS_ERR(task)) {
 379                kfree(lcq);
 380                return PTR_ERR(task);
 381        }
 382
 383        wait_for_completion(&lcq->lcq_comp);
 384        *lcq_ret = lcq;
 385        return 0;
 386}
 387
 388void ll_close_thread_shutdown(struct ll_close_queue *lcq)
 389{
 390        init_completion(&lcq->lcq_comp);
 391        atomic_inc(&lcq->lcq_stop);
 392        wake_up(&lcq->lcq_waitq);
 393        wait_for_completion(&lcq->lcq_comp);
 394        kfree(lcq);
 395}
 396