lustre/llite/llite_close.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite routines to issue a secondary close after writeback
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #include <linux/module.h>
  25
  26 #define DEBUG_SUBSYSTEM S_LLITE
  27
  28 #include <lustre_lite.h>
  29 #include "llite_internal.h"
  30
  31 #ifdef HAVE_CLOSE_THREAD
  32 /* record that a write is in flight */
  33 void llap_write_pending(struct inode *inode, struct ll_async_page *llap)
  34 {
  35         struct ll_inode_info *lli = ll_i2info(inode);
  36         spin_lock(&lli->lli_lock);
  37         list_add(&llap->llap_pending_write, &lli->lli_pending_write_llaps);
  38         spin_unlock(&lli->lli_lock);
  39 }
  40
  41 /* record that a write has completed */
  42 void llap_write_complete(struct inode *inode, struct ll_async_page *llap)
  43 {
  44         struct ll_inode_info *lli = ll_i2info(inode);
  45         spin_lock(&lli->lli_lock);
  46         list_del_init(&llap->llap_pending_write);
  47         spin_unlock(&lli->lli_lock);
  48 }
  49
  50 void ll_open_complete(struct inode *inode)
  51 {
  52         struct ll_inode_info *lli = ll_i2info(inode);
  53         spin_lock(&lli->lli_lock);
  54         lli->lli_send_done_writing = 0;
  55         spin_unlock(&lli->lli_lock);
  56 }
  57
  58 /* if we close with writes in flight then we want the completion or cancelation
  59  * of those writes to send a DONE_WRITING rpc to the MDS */
  60 int ll_is_inode_dirty(struct inode *inode)
  61 {
  62         struct ll_inode_info *lli = ll_i2info(inode);
  63         int rc = 0;
  64         ENTRY;
  65
  66         spin_lock(&lli->lli_lock);
  67         if (!list_empty(&lli->lli_pending_write_llaps))
  68                 rc = 1;
  69         spin_unlock(&lli->lli_lock);
  70         RETURN(rc);
  71 }
  72
  73 void ll_try_done_writing(struct inode *inode)
  74 {
  75         struct ll_inode_info *lli = ll_i2info(inode);
  76         struct ll_close_queue *lcq = ll_i2sbi(inode)->ll_lcq;
  77
  78         spin_lock(&lli->lli_lock);
  79
  80         if (lli->lli_send_done_writing &&
  81             list_empty(&lli->lli_pending_write_llaps)) {
  82
  83                 spin_lock(&lcq->lcq_lock);
  84                 if (list_empty(&lli->lli_close_item)) {
  85                         CDEBUG(D_INODE, "adding inode %lu/%u to close list\n",
  86                                inode->i_ino, inode->i_generation);
  87                         igrab(inode);
  88                         list_add_tail(&lli->lli_close_item, &lcq->lcq_list);
  89                         wake_up(&lcq->lcq_waitq);
  90                 }
  91                 spin_unlock(&lcq->lcq_lock);
  92         }
  93
  94         spin_unlock(&lli->lli_lock);
  95 }
  96
  97 /* The MDS needs us to get the real file attributes, then send a DONE_WRITING */
  98 void ll_queue_done_writing(struct inode *inode)
  99 {
 100         struct ll_inode_info *lli = ll_i2info(inode);
 101         ENTRY;
 102
 103         spin_lock(&lli->lli_lock);
 104         lli->lli_send_done_writing = 1;
 105         spin_unlock(&lli->lli_lock);
 106
 107         ll_try_done_writing(inode);
 108         EXIT;
 109 }
 110
 111 /* If we know the file size and have the cookies:
 112  *  - send a DONE_WRITING rpc
 113  *
 114  * Otherwise:
 115  *  - get a whole-file lock
 116  *  - get the authoritative size and all cookies with GETATTRs
 117  *  - send a DONE_WRITING rpc
 118  */
 119 static void ll_close_done_writing(struct inode *inode)
 120 {
 121         struct ll_inode_info *lli = ll_i2info(inode);
 122         ldlm_policy_data_t policy = { .l_extent = {0, OBD_OBJECT_EOF } };
 123         struct lustre_handle lockh = { 0 };
 124         struct obdo obdo;
 125         obd_flag valid;
 126         int rc, ast_flags = 0;
 127         ENTRY;
 128
 129         memset(&obdo, 0, sizeof(obdo));
 130         if (test_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags))
 131                 goto rpc;
 132
 133         rc = ll_extent_lock(NULL, inode, lli->lli_smd, LCK_PW, &policy, &lockh,
 134                             ast_flags);
 135         if (rc != 0) {
 136                 CERROR("lock acquisition failed (%d): unable to send "
 137                        "DONE_WRITING for inode %lu/%u\n", rc, inode->i_ino,
 138                        inode->i_generation);
 139                 GOTO(out, rc);
 140         }
 141
 142         rc = ll_lsm_getattr(ll_i2obdexp(inode), lli->lli_smd, &obdo);
 143         if (rc) {
 144                 CERROR("inode_getattr failed (%d): unable to send DONE_WRITING "
 145                        "for inode %lu/%u\n", rc, inode->i_ino,
 146                        inode->i_generation);
 147                 ll_extent_unlock(NULL, inode, lli->lli_smd, LCK_PW, &lockh);
 148                 GOTO(out, rc);
 149         }
 150
 151         obdo_refresh_inode(inode, &obdo, valid);
 152
 153         CDEBUG(D_INODE, "objid "LPX64" size %Lu, blocks %lu, blksize %lu\n",
 154                lli->lli_smd->lsm_object_id, i_size_read(inode), inode->i_blocks,
 155                1<<inode->i_blkbits);
 156
 157         set_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags);
 158
 159         rc = ll_extent_unlock(NULL, inode, lli->lli_smd, LCK_PW, &lockh);
 160         if (rc != ELDLM_OK)
 161                 CERROR("unlock failed (%d)?  proceeding anyways...\n", rc);
 162
 163  rpc:
 164         obdo.o_id = inode->i_ino;
 165         obdo.o_size = i_size_read(inode);
 166         obdo.o_blocks = inode->i_blocks;
 167         obdo.o_valid = OBD_MD_FLID | OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 168
 169         rc = mdc_done_writing(ll_i2sbi(inode)->ll_mdc_exp, &obdo);
 170  out:
 171 }
 172
 173
 174 static struct ll_inode_info *ll_close_next_lli(struct ll_close_queue *lcq)
 175 {
 176         struct ll_inode_info *lli = NULL;
 177
 178         spin_lock(&lcq->lcq_lock);
 179
 180         if (lcq->lcq_list.next == NULL)
 181                 lli = ERR_PTR(-1);
 182         else if (!list_empty(&lcq->lcq_list)) {
 183                 lli = list_entry(lcq->lcq_list.next, struct ll_inode_info,
 184                                  lli_close_item);
 185                 list_del(&lli->lli_close_item);
 186         }
 187
 188         spin_unlock(&lcq->lcq_lock);
 189         return lli;
 190 }
 191 #else
 192 static struct ll_inode_info *ll_close_next_lli(struct ll_close_queue *lcq)
 193 {
 194         if (lcq->lcq_list.next == NULL)
 195                 return ERR_PTR(-1);
 196
 197         return NULL;
 198 }
 199 #endif
 200
 201 static int ll_close_thread(void *arg)
 202 {
 203         struct ll_close_queue *lcq = arg;
 204         ENTRY;
 205
 206         {
 207                 char name[CFS_CURPROC_COMM_MAX];
 208                 snprintf(name, sizeof(name) - 1, "ll_close");
 209                 cfs_daemonize(name);
 210         }
 211
 212         complete(&lcq->lcq_comp);
 213
 214         while (1) {
 215                 struct l_wait_info lwi = { 0 };
 216                 struct ll_inode_info *lli;
 217                 //struct inode *inode;
 218
 219                 l_wait_event_exclusive(lcq->lcq_waitq,
 220                                        (lli = ll_close_next_lli(lcq)) != NULL,
 221                                        &lwi);
 222                 if (IS_ERR(lli))
 223                         break;
 224
 225                 //inode = ll_info2i(lli);
 226                 //ll_close_done_writing(inode);
 227                 //iput(inode);
 228         }
 229
 230         complete(&lcq->lcq_comp);
 231         RETURN(0);
 232 }
 233
 234 int ll_close_thread_start(struct ll_close_queue **lcq_ret)
 235 {
 236         struct ll_close_queue *lcq;
 237         pid_t pid;
 238
 239         OBD_FAIL_RETURN(OBD_FAIL_LDLM_CLOSE_THREAD, -EINTR);
 240         OBD_ALLOC(lcq, sizeof(*lcq));
 241         if (lcq == NULL)
 242                 return -ENOMEM;
 243
 244         spin_lock_init(&lcq->lcq_lock);
 245         INIT_LIST_HEAD(&lcq->lcq_list);
 246         init_waitqueue_head(&lcq->lcq_waitq);
 247         init_completion(&lcq->lcq_comp);
 248
 249         pid = kernel_thread(ll_close_thread, lcq, 0);
 250         if (pid < 0) {
 251                 OBD_FREE(lcq, sizeof(*lcq));
 252                 return pid;
 253         }
 254
 255         wait_for_completion(&lcq->lcq_comp);
 256         *lcq_ret = lcq;
 257         return 0;
 258 }
 259
 260 void ll_close_thread_shutdown(struct ll_close_queue *lcq)
 261 {
 262         init_completion(&lcq->lcq_comp);
 263         lcq->lcq_list.next = NULL;
 264         wake_up(&lcq->lcq_waitq);
 265         wait_for_completion(&lcq->lcq_comp);
 266         OBD_FREE(lcq, sizeof(*lcq));
 267 }
 268
 269