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 <linux/lustre_mds.h>
  29 #include <linux/lustre_lite.h>
  30 #include "llite_internal.h"
  31
  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         if (!list_empty(&llap->llap_pending_write))
  47                 list_del_init(&llap->llap_pending_write);
  48         spin_unlock(&lli->lli_lock);
  49 }
  50
  51 void ll_open_complete(struct inode *inode)
  52 {
  53         struct ll_inode_info *lli = ll_i2info(inode);
  54         spin_lock(&lli->lli_lock);
  55         lli->lli_send_done_writing = 0;
  56         spin_unlock(&lli->lli_lock);
  57 }
  58
  59 /* if we close with writes in flight then we want the completion or cancelation
  60  * of those writes to send a DONE_WRITING rpc to the MDS */
  61 int ll_is_inode_dirty(struct inode *inode)
  62 {
  63         struct ll_inode_info *lli = ll_i2info(inode);
  64         int rc = 0;
  65         ENTRY;
  66
  67         spin_lock(&lli->lli_lock);
  68         if (!list_empty(&lli->lli_pending_write_llaps))
  69                 rc = 1;
  70         spin_unlock(&lli->lli_lock);
  71         RETURN(rc);
  72 }
  73
  74 void ll_try_done_writing(struct inode *inode)
  75 {
  76         struct ll_inode_info *lli = ll_i2info(inode);
  77         struct ll_close_queue *lcq = ll_i2sbi(inode)->ll_lcq;
  78
  79         spin_lock(&lli->lli_lock);
  80
  81         if (lli->lli_send_done_writing &&
  82             list_empty(&lli->lli_pending_write_llaps)) {
  83
  84                 spin_lock(&lcq->lcq_lock);
  85                 if (list_empty(&lli->lli_close_item)) {
  86                         CDEBUG(D_INODE, "adding inode %lu/%u to close list\n",
  87                                inode->i_ino, inode->i_generation);
  88                         LASSERT(igrab(inode) == inode);
  89                         list_add_tail(&lli->lli_close_item, &lcq->lcq_list);
  90                         wake_up(&lcq->lcq_waitq);
  91                 }
  92                 spin_unlock(&lcq->lcq_lock);
  93         }
  94
  95         spin_unlock(&lli->lli_lock);
  96 }
  97
  98 /* The MDS needs us to get the real file attributes, then send a DONE_WRITING */
  99 void ll_queue_done_writing(struct inode *inode)
 100 {
 101         struct ll_inode_info *lli = ll_i2info(inode);
 102         ENTRY;
 103
 104         spin_lock(&lli->lli_lock);
 105         lli->lli_send_done_writing = 1;
 106         spin_unlock(&lli->lli_lock);
 107
 108         ll_try_done_writing(inode);
 109         EXIT;
 110 }
 111
 112 /* If we know the file size and have the cookies:
 113  *  - send a DONE_WRITING rpc
 114  *
 115  * Otherwise:
 116  *  - get a whole-file lock
 117  *  - get the authoritative size and all cookies with GETATTRs
 118  *  - send a DONE_WRITING rpc
 119  */
 120 static void ll_close_done_writing(struct inode *inode)
 121 {
 122         struct ll_inode_info *lli = ll_i2info(inode);
 123         struct ldlm_extent extent = { .start = 0, .end = OBD_OBJECT_EOF };
 124         struct lustre_handle lockh = { 0 };
 125         struct obdo obdo;
 126         obd_flag valid;
 127         int rc, ast_flags = 0;
 128         ENTRY;
 129
 130         memset(&obdo, 0, sizeof(obdo));
 131         if (test_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags))
 132                 goto rpc;
 133
 134         rc = ll_extent_lock_no_validate(NULL, inode, lli->lli_smd, LCK_PW,
 135                                         &extent, &lockh, ast_flags);
 136         if (rc != ELDLM_OK) {
 137                 CERROR("lock acquisition failed (%d): unable to send "
 138                        "DONE_WRITING for inode %lu/%u\n", rc, inode->i_ino,
 139                        inode->i_generation);
 140                 GOTO(out, rc);
 141         }
 142
 143         rc = ll_lsm_getattr(ll_i2obdexp(inode), lli->lli_smd, &obdo);
 144         if (rc) {
 145                 CERROR("inode_getattr failed (%d): unable to send DONE_WRITING "
 146                        "for inode %lu/%u\n", rc, inode->i_ino,
 147                        inode->i_generation);
 148                 ll_extent_unlock(NULL, inode, lli->lli_smd, LCK_PW, &lockh);
 149                 GOTO(out, rc);
 150         }
 151
 152         obdo_refresh_inode(inode, &obdo, valid);
 153
 154         CDEBUG(D_INODE, "objid "LPX64" size %Lu, blocks %lu, blksize %lu\n",
 155                lli->lli_smd->lsm_object_id, inode->i_size, inode->i_blocks,
 156                inode->i_blksize);
 157
 158         set_bit(LLI_F_HAVE_OST_SIZE_LOCK, &lli->lli_flags);
 159
 160         rc = ll_extent_unlock(NULL, inode, lli->lli_smd, LCK_PW, &lockh);
 161         if (rc != ELDLM_OK)
 162                 CERROR("unlock failed (%d)?  proceeding anyways...\n", rc);
 163
 164  rpc:
 165         obdo.o_id = inode->i_ino;
 166         obdo.o_size = inode->i_size;
 167         obdo.o_blocks = inode->i_blocks;
 168         obdo.o_valid = OBD_MD_FLID | OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
 169
 170         rc = mdc_done_writing(ll_i2sbi(inode)->ll_mdc_exp, &obdo);
 171  out:
 172         iput(inode);
 173 }
 174
 175 static struct ll_inode_info *ll_close_next_lli(struct ll_close_queue *lcq)
 176 {
 177         struct ll_inode_info *lli = NULL;
 178
 179         spin_lock(&lcq->lcq_lock);
 180
 181         if (lcq->lcq_list.next == NULL)
 182                 lli = ERR_PTR(-1);
 183         else if (!list_empty(&lcq->lcq_list)) {
 184                 lli = list_entry(lcq->lcq_list.next, struct ll_inode_info,
 185                                  lli_close_item);
 186                 list_del(&lli->lli_close_item);
 187         }
 188
 189         spin_unlock(&lcq->lcq_lock);
 190         return lli;
 191 }
 192
 193 static int ll_close_thread(void *arg)
 194 {
 195         struct ll_close_queue *lcq = arg;
 196         ENTRY;
 197
 198         /* XXX boiler-plate */
 199         {
 200                 char name[sizeof(current->comm)];
 201                 unsigned long flags;
 202                 snprintf(name, sizeof(name) - 1, "ll_close");
 203                 kportal_daemonize(name);
 204                 SIGNAL_MASK_LOCK(current, flags);
 205                 sigfillset(&current->blocked);
 206                 RECALC_SIGPENDING;
 207                 SIGNAL_MASK_UNLOCK(current, flags);
 208         }
 209
 210         complete(&lcq->lcq_comp);
 211
 212         while (1) {
 213                 struct l_wait_info lwi = { 0 };
 214                 struct ll_inode_info *lli;
 215                 struct inode *inode;
 216
 217                 l_wait_event_exclusive(lcq->lcq_waitq,
 218                                        (lli = ll_close_next_lli(lcq)) != NULL,
 219                                        &lwi);
 220                 if (IS_ERR(lli))
 221                         break;
 222
 223                 inode = ll_info2i(lli);
 224                 ll_close_done_writing(inode);
 225         }
 226
 227         complete(&lcq->lcq_comp);
 228         RETURN(0);
 229 }
 230
 231 int ll_close_thread_start(struct ll_close_queue **lcq_ret)
 232 {
 233         struct ll_close_queue *lcq;
 234         pid_t pid;
 235
 236         OBD_ALLOC(lcq, sizeof(*lcq));
 237         if (lcq == NULL)
 238                 return -ENOMEM;
 239
 240         spin_lock_init(&lcq->lcq_lock);
 241         INIT_LIST_HEAD(&lcq->lcq_list);
 242         init_waitqueue_head(&lcq->lcq_waitq);
 243         init_completion(&lcq->lcq_comp);
 244
 245         pid = kernel_thread(ll_close_thread, lcq, 0);
 246         if (pid < 0) {
 247                 OBD_FREE(lcq, sizeof(*lcq));
 248                 return pid;
 249         }
 250
 251         wait_for_completion(&lcq->lcq_comp);
 252         *lcq_ret = lcq;
 253         return 0;
 254 }
 255
 256 void ll_close_thread_shutdown(struct ll_close_queue *lcq)
 257 {
 258         init_completion(&lcq->lcq_comp);
 259         lcq->lcq_list.next = NULL;
 260         wake_up(&lcq->lcq_waitq);
 261         wait_for_completion(&lcq->lcq_comp);
 262         OBD_FREE(lcq, sizeof(*lcq));
 263 }