lustre/lib/page.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.sf.net/projects/lustre/
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  */
  22
  23
  24
  25 #include <linux/config.h>
  26 #include <linux/kernel.h>
  27 #include <linux/mm.h>
  28 #include <linux/string.h>
  29 #include <linux/stat.h>
  30 #include <linux/errno.h>
  31 #include <linux/locks.h>
  32 #include <linux/unistd.h>
  33 #include <linux/version.h>
  34
  35 #include <asm/system.h>
  36 #include <asm/uaccess.h>
  37
  38 #include <linux/fs.h>
  39 #include <linux/stat.h>
  40 #include <asm/uaccess.h>
  41 #include <asm/segment.h>
  42 #include <linux/mm.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/smp_lock.h>
  45
  46 #define DEBUG_SUBSYSTEM S_OST
  47
  48 #include <linux/obd_class.h>
  49 #include <linux/lustre_net.h>
  50 #include <linux/lustre_lib.h>
  51 #include <linux/lustre_ha.h>
  52
  53 static int sync_io_timeout(void *data)
  54 {
  55         struct io_cb_data *cbd = data;
  56         struct ptlrpc_bulk_desc *desc = cbd->desc;
  57
  58         ENTRY;
  59         desc->b_connection->c_level = LUSTRE_CONN_RECOVD;
  60         desc->b_flags |= PTL_RPC_FL_TIMEOUT;
  61         if (desc->b_connection && desc->b_connection->c_recovd &&
  62             class_signal_connection_failure) {
  63                 /* XXXshaver Do we need a resend strategy, or do we just
  64                  * XXXshaver return -ERESTARTSYS and punt it?
  65                  */
  66                 CERROR("signalling failure of conn %p\n", desc->b_connection);
  67                 class_signal_connection_failure(desc->b_connection);
  68
  69                 /* We go back to sleep, until we're resumed or interrupted. */
  70                 RETURN(0);
  71         }
  72
  73         /* If we can't be recovered, just abort the syscall with -ETIMEDOUT. */
  74         RETURN(1);
  75 }
  76
  77 static int sync_io_intr(void *data)
  78 {
  79         struct io_cb_data *cbd = data;
  80         struct ptlrpc_bulk_desc *desc = cbd->desc;
  81
  82         ENTRY;
  83         desc->b_flags |= PTL_RPC_FL_INTR;
  84         RETURN(1); /* ignored, as of this writing */
  85 }
  86
  87 int ll_sync_io_cb(struct io_cb_data *data, int err, int phase)
  88 {
  89         int ret;
  90         ENTRY;
  91
  92         if (phase == CB_PHASE_START) {
  93                 struct l_wait_info lwi;
  94                 lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, sync_io_timeout,
  95                                        sync_io_intr, data);
  96                 ret = l_wait_event(data->waitq, data->complete, &lwi);
  97                 if (atomic_dec_and_test(&data->refcount))
  98                         OBD_FREE(data, sizeof(*data));
  99                 if (ret == -ERESTARTSYS)
 100                         return ret;
 101         } else if (phase == CB_PHASE_FINISH) {
 102                 data->err = err;
 103                 data->complete = 1;
 104                 wake_up(&data->waitq);
 105                 if (atomic_dec_and_test(&data->refcount))
 106                         OBD_FREE(data, sizeof(*data));
 107                 return err;
 108         } else
 109                 LBUG();
 110         EXIT;
 111         return 0;
 112 }
 113
 114 struct io_cb_data *ll_init_cb(void)
 115 {
 116         struct io_cb_data *d;
 117
 118
 119         OBD_ALLOC(d, sizeof(*d));
 120         if (d) {
 121                 init_waitqueue_head(&d->waitq);
 122                 atomic_set(&d->refcount, 2);
 123         }
 124         RETURN(d);
 125 }
 126
 127 /*
 128  * Remove page from dirty list
 129  */
 130 static void __set_page_clean(struct page *page)
 131 {
 132         struct address_space *mapping = page->mapping;
 133         struct inode *inode;
 134
 135         if (!mapping)
 136                 return;
 137
 138 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,4,9))
 139         spin_lock(&pagecache_lock);
 140 #endif
 141
 142         list_del(&page->list);
 143         list_add(&page->list, &mapping->clean_pages);
 144
 145         inode = mapping->host;
 146         if (list_empty(&mapping->dirty_pages)) {
 147                 CDEBUG(D_INODE, "inode clean\n");
 148                 inode->i_state &= ~I_DIRTY_PAGES;
 149         }
 150 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,4,10))
 151         spin_unlock(&pagecache_lock);
 152 #endif
 153         EXIT;
 154 }
 155
 156 inline void set_page_clean(struct page *page)
 157 {
 158         if (PageDirty(page)) {
 159                 ClearPageDirty(page);
 160                 __set_page_clean(page);
 161         }
 162 }
 163
 164 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,4,10))
 165 /*
 166  * Add a page to the dirty page list.
 167  */
 168 void __set_page_dirty(struct page *page)
 169 {
 170         struct address_space *mapping;
 171         spinlock_t *pg_lock;
 172
 173         pg_lock = PAGECACHE_LOCK(page);
 174         spin_lock(pg_lock);
 175
 176         mapping = page->mapping;
 177         spin_lock(&mapping->page_lock);
 178
 179         list_del(&page->list);
 180         list_add(&page->list, &mapping->dirty_pages);
 181
 182         spin_unlock(&mapping->page_lock);
 183         spin_unlock(pg_lock);
 184
 185         if (mapping->host)
 186                 mark_inode_dirty_pages(mapping->host);
 187 }
 188 #else
 189 /*
 190  * Add a page to the dirty page list.
 191  */
 192 void set_page_dirty(struct page *page)
 193 {
 194         if (!test_and_set_bit(PG_dirty, &page->flags)) {
 195                 struct address_space *mapping = page->mapping;
 196
 197                 if (mapping) {
 198                         spin_lock(&pagecache_lock);
 199                         list_del(&page->list);
 200                         list_add(&page->list, &mapping->dirty_pages);
 201                         spin_unlock(&pagecache_lock);
 202
 203                         if (mapping->host)
 204                                 mark_inode_dirty_pages(mapping->host);
 205                 }
 206         }
 207 }
 208 #endif
 209
 210 inline void lustre_put_page(struct page *page)
 211 {
 212         kunmap(page);
 213         page_cache_release(page);
 214 }
 215
 216 struct page *lustre_get_page_read(struct inode *inode, unsigned long index)
 217 {
 218         struct address_space *mapping = inode->i_mapping;
 219         struct page *page;
 220         int rc;
 221
 222         page = read_cache_page(mapping, index,
 223                                (filler_t*)mapping->a_ops->readpage, NULL);
 224         if (!IS_ERR(page)) {
 225                 wait_on_page(page);
 226                 kmap(page);
 227                 if (!Page_Uptodate(page)) {
 228                         CERROR("page index %lu not uptodate\n", index);
 229                         GOTO(err_page, rc = -EIO);
 230                 }
 231                 if (PageError(page)) {
 232                         CERROR("page index %lu has error\n", index);
 233                         GOTO(err_page, rc = -EIO);
 234                 }
 235         }
 236         return page;
 237
 238 err_page:
 239         lustre_put_page(page);
 240         return ERR_PTR(rc);
 241 }
 242
 243 struct page *lustre_get_page_write(struct inode *inode, unsigned long index)
 244 {
 245         struct address_space *mapping = inode->i_mapping;
 246         struct page *page;
 247         int rc;
 248
 249         page = grab_cache_page(mapping, index); /* locked page */
 250
 251         if (!IS_ERR(page)) {
 252                 kmap(page);
 253                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
 254                  * a no-op for most filesystems, because we write the whole
 255                  * page.  For partial-page I/O this will read in the page.
 256                  */
 257                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
 258                 if (rc) {
 259                         CERROR("page index %lu, rc = %d\n", index, rc);
 260                         if (rc != -ENOSPC)
 261                                 LBUG();
 262                         GOTO(err_unlock, rc);
 263                 }
 264                 /* XXX not sure if we need this if we are overwriting page */
 265                 if (PageError(page)) {
 266                         CERROR("error on page index %lu, rc = %d\n", index, rc);
 267                         LBUG();
 268                         GOTO(err_unlock, rc = -EIO);
 269                 }
 270         }
 271         return page;
 272
 273 err_unlock:
 274         unlock_page(page);
 275         lustre_put_page(page);
 276         return ERR_PTR(rc);
 277 }
 278
 279 int lustre_commit_write(struct page *page, unsigned from, unsigned to)
 280 {
 281         struct inode *inode = page->mapping->host;
 282         int err;
 283
 284         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
 285         if (!err && IS_SYNC(inode))
 286                 err = waitfor_one_page(page);
 287
 288         //SetPageUptodate(page); // the client commit_write will do this
 289
 290         SetPageReferenced(page);
 291         unlock_page(page);
 292         lustre_put_page(page);
 293         return err;
 294 }