lustre/llite/rw.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite I/O Page Cache
   5  *
   6  * Copyright (C) 2002 Cluster File Systems, Inc.
   7  */
   8
   9 #include <linux/config.h>
  10 #include <linux/kernel.h>
  11 #include <linux/mm.h>
  12 #include <linux/string.h>
  13 #include <linux/stat.h>
  14 #include <linux/iobuf.h>
  15 #include <linux/errno.h>
  16 #include <linux/smp_lock.h>
  17 #include <linux/unistd.h>
  18 #include <linux/version.h>
  19 #include <asm/system.h>
  20 #include <asm/uaccess.h>
  21
  22 #include <linux/fs.h>
  23 #include <linux/stat.h>
  24 #include <asm/uaccess.h>
  25 #include <asm/segment.h>
  26 #include <linux/mm.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/smp_lock.h>
  29
  30 #define DEBUG_SUBSYSTEM S_LLITE
  31
  32 #include <linux/lustre_mds.h>
  33 #include <linux/lustre_lite.h>
  34 #include <linux/lustre_lib.h>
  35
  36 /*
  37  * Remove page from dirty list
  38  */
  39 static void __set_page_clean(struct page *page)
  40 {
  41         struct address_space *mapping = page->mapping;
  42         struct inode *inode;
  43
  44         if (!mapping)
  45                 return;
  46
  47 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
  48         spin_lock(&pagecache_lock);
  49 #endif
  50
  51         list_del(&page->list);
  52         list_add(&page->list, &mapping->clean_pages);
  53
  54         inode = mapping->host;
  55         if (list_empty(&mapping->dirty_pages)) {
  56                 CDEBUG(D_INODE, "inode clean\n");
  57                 inode->i_state &= ~I_DIRTY_PAGES;
  58         }
  59 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
  60         spin_unlock(&pagecache_lock);
  61 #endif
  62         EXIT;
  63 }
  64
  65 inline void set_page_clean(struct page *page)
  66 {
  67         if (PageDirty(page)) {
  68                 ClearPageDirty(page);
  69                 __set_page_clean(page);
  70         }
  71 }
  72
  73 /* SYNCHRONOUS I/O to object storage for an inode */
  74 static int ll_brw(int cmd, struct inode *inode, struct page *page, int create)
  75 {
  76         struct ll_inode_info *lli = ll_i2info(inode);
  77         struct lov_stripe_md *lsm = lli->lli_smd;
  78         struct brw_cb_data *brw_cbd = ll_init_brw_cb_data();
  79         struct brw_page pg;
  80         int err;
  81         ENTRY;
  82
  83         if (!brw_cbd)
  84                 RETURN(-ENOMEM);
  85
  86         pg.pg = page;
  87         pg.off = ((obd_off)page->index) << PAGE_SHIFT;
  88
  89         if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size))
  90                 pg.count = inode->i_size % PAGE_SIZE;
  91         else
  92                 pg.count = PAGE_SIZE;
  93
  94         pg.flag = create ? OBD_BRW_CREATE : 0;
  95
  96         err = obd_brw(cmd, ll_i2obdconn(inode),lsm, 1, &pg, ll_sync_brw_cb,
  97                       brw_cbd);
  98
  99         RETURN(err);
 100 } /* ll_brw */
 101
 102 /* returns the page unlocked, but with a reference */
 103 static int ll_readpage(struct file *file, struct page *page)
 104 {
 105         struct inode *inode = page->mapping->host;
 106         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 107         int rc = 0;
 108         ENTRY;
 109
 110         if (!PageLocked(page))
 111                 LBUG();
 112
 113         if (inode->i_size <= offset) {
 114                 memset(kmap(page), 0, PAGE_SIZE);
 115                 kunmap(page);
 116                 GOTO(readpage_out, rc);
 117         }
 118
 119         if (PageUptodate(page)) {
 120                 CERROR("Explain this please?\n");
 121                 GOTO(readpage_out, rc);
 122         }
 123
 124         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
 125         EXIT;
 126
 127  readpage_out:
 128         if (!rc)
 129                 SetPageUptodate(page);
 130         unlock_page(page);
 131         return 0;
 132 } /* ll_readpage */
 133
 134 void ll_truncate(struct inode *inode)
 135 {
 136         struct obdo oa = {0};
 137         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 138         struct lustre_handle *lockhs = NULL;
 139         int err;
 140         ENTRY;
 141
 142         if (!lsm) {
 143                 /* object not yet allocated */
 144                 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 145                 return;
 146         }
 147
 148         oa.o_id = lsm->lsm_object_id;
 149         oa.o_mode = inode->i_mode;
 150         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 151
 152         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after "LPD64")\n",
 153                oa.o_id, inode->i_size);
 154
 155         err = ll_size_lock(inode, lsm, inode->i_size, LCK_PW, &lockhs);
 156         if (err) {
 157                 CERROR("ll_size_lock failed: %d\n", err);
 158                 /* FIXME: What to do here?  It's too late to back out... */
 159                 LBUG();
 160         }
 161
 162         /* truncate == punch from new size to absolute end of file */
 163         err = obd_punch(ll_i2obdconn(inode), &oa, lsm, inode->i_size,
 164                         OBD_OBJECT_EOF);
 165         if (err)
 166                 CERROR("obd_truncate fails (%d)\n", err);
 167         else
 168                 obdo_to_inode(inode, &oa, oa.o_valid);
 169
 170         err = ll_size_unlock(inode, lsm, LCK_PW, lockhs);
 171         if (err)
 172                 CERROR("ll_size_unlock failed: %d\n", err);
 173
 174         EXIT;
 175         return;
 176 } /* ll_truncate */
 177
 178 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 179
 180 static int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 181                             unsigned to)
 182 {
 183         struct inode *inode = page->mapping->host;
 184         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 185         int rc = 0;
 186         char *addr;
 187         ENTRY;
 188
 189         addr = kmap(page);
 190         if (!PageLocked(page))
 191                 LBUG();
 192
 193         if (Page_Uptodate(page))
 194                 GOTO(prepare_done, rc);
 195
 196         /* We're completely overwriting an existing page, so _don't_ set it up
 197          * to date until commit_write */
 198         if (from == 0 && to == PAGE_SIZE)
 199                 RETURN(0);
 200
 201         /* We are writing to a new page, no need to read old data */
 202         if (inode->i_size <= offset) {
 203                 memset(addr, 0, PAGE_SIZE);
 204                 GOTO(prepare_done, rc=0);
 205         }
 206
 207         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
 208
 209         EXIT;
 210  prepare_done:
 211         if (!rc)
 212                 SetPageUptodate(page);
 213
 214         return rc;
 215 }
 216
 217 /* returns the page unlocked, but with a reference */
 218 static int ll_writepage(struct page *page)
 219 {
 220         struct inode *inode = page->mapping->host;
 221         int err;
 222         ENTRY;
 223
 224         if (!PageLocked(page))
 225                 LBUG();
 226
 227         err = ll_brw(OBD_BRW_WRITE, inode, page, 1);
 228         if ( !err ) {
 229                 //SetPageUptodate(page);
 230                 set_page_clean(page);
 231         } else {
 232                 CERROR("ll_brw failure %d\n", err);
 233         }
 234         unlock_page(page);
 235         RETURN(err);
 236 }
 237
 238
 239 /* SYNCHRONOUS I/O to object storage for an inode -- object attr will be updated
 240  * too */
 241 static int ll_commit_write(struct file *file, struct page *page,
 242                            unsigned from, unsigned to)
 243 {
 244         int create = 1;
 245         struct inode *inode = page->mapping->host;
 246         struct ll_inode_info *lli = ll_i2info(inode);
 247         struct lov_stripe_md *md = lli->lli_smd;
 248         struct brw_page pg;
 249         int err;
 250         loff_t size;
 251         struct brw_cb_data *cbd = ll_init_brw_cb_data();
 252         ENTRY;
 253
 254         pg.pg = page;
 255         pg.count = to;
 256         pg.off = (((obd_off)page->index) << PAGE_SHIFT);
 257         pg.flag = create ? OBD_BRW_CREATE : 0;
 258
 259         if (!cbd)
 260                 RETURN(-ENOMEM);
 261
 262         SetPageUptodate(page);
 263
 264         if (!PageLocked(page))
 265                 LBUG();
 266
 267         CDEBUG(D_INODE, "commit_page writing (off "LPD64"), count "LPD64"\n",
 268                pg.off, pg.count);
 269
 270         err = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode), md,
 271                       1, &pg, ll_sync_brw_cb, cbd);
 272         kunmap(page);
 273
 274         size = pg.off + pg.count;
 275         /* do NOT truncate when writing in the middle of a file */
 276         if (size > inode->i_size)
 277                 inode->i_size = size;
 278
 279         RETURN(err);
 280 } /* ll_commit_write */
 281
 282
 283 static int ll_direct_IO(int rw, struct inode *inode, struct kiobuf *iobuf,
 284                         unsigned long blocknr, int blocksize)
 285 {
 286         obd_count bufs_per_obdo = iobuf->nr_pages;
 287         struct ll_inode_info *lli = ll_i2info(inode);
 288         struct lov_stripe_md *lsm = lli->lli_smd;
 289         struct brw_page *pga;
 290         int i, rc = 0;
 291         struct brw_cb_data *cbd;
 292
 293         ENTRY;
 294         if (!lsm || !lsm->lsm_object_id)
 295                 RETURN(-ENOMEM);
 296
 297         if (blocksize != PAGE_SIZE) {
 298                 CERROR("direct_IO blocksize != PAGE_SIZE\n");
 299                 RETURN(-EINVAL);
 300         }
 301
 302         cbd = ll_init_brw_cb_data();
 303         if (!cbd)
 304                 RETURN(-ENOMEM);
 305
 306         OBD_ALLOC(pga, sizeof(*pga) * bufs_per_obdo);
 307         if (!pga) {
 308                 OBD_FREE(cbd, sizeof(*cbd));
 309                 RETURN(-ENOMEM);
 310         }
 311
 312         /* NB: we can't use iobuf->maplist[i]->index for the offset
 313          * instead of "blocknr" because ->index contains garbage.
 314          */
 315         for (i = 0; i < bufs_per_obdo; i++, blocknr++) {
 316                 pga[i].pg = iobuf->maplist[i];
 317                 pga[i].count = PAGE_SIZE;
 318                 pga[i].off = (obd_off)blocknr << PAGE_SHIFT;
 319                 pga[i].flag = OBD_BRW_CREATE;
 320         }
 321
 322         rc = obd_brw(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 323                      ll_i2obdconn(inode), lsm, bufs_per_obdo, pga,
 324                      ll_sync_brw_cb, cbd);
 325         if (rc == 0)
 326                 rc = bufs_per_obdo * PAGE_SIZE;
 327
 328         OBD_FREE(pga, sizeof(*pga) * bufs_per_obdo);
 329         RETURN(rc);
 330 }
 331
 332 int ll_flush_inode_pages(struct inode * inode)
 333 {
 334         obd_count        bufs_per_obdo = 0;
 335         obd_size         *count = NULL;
 336         obd_off          *offset = NULL;
 337         obd_flag         *flags = NULL;
 338         int              err = 0;
 339
 340         ENTRY;
 341
 342         spin_lock(&pagecache_lock);
 343
 344         spin_unlock(&pagecache_lock);
 345
 346
 347         OBD_ALLOC(count, sizeof(*count) * bufs_per_obdo);
 348         OBD_ALLOC(offset, sizeof(*offset) * bufs_per_obdo);
 349         OBD_ALLOC(flags, sizeof(*flags) * bufs_per_obdo);
 350         if (!count || !offset || !flags)
 351                 GOTO(out, err=-ENOMEM);
 352
 353 #if 0
 354         for (i = 0 ; i < bufs_per_obdo ; i++) {
 355                 count[i] = PAGE_SIZE;
 356                 offset[i] = ((obd_off)(iobuf->maplist[i])->index) << PAGE_SHIFT;
 357                 flags[i] = OBD_BRW_CREATE;
 358         }
 359
 360         err = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode),
 361                       ll_i2info(inode)->lli_smd, bufs_per_obdo,
 362                       iobuf->maplist, count, offset, flags, NULL, NULL);
 363         if (err == 0)
 364                 err = bufs_per_obdo * 4096;
 365 #endif
 366  out:
 367         OBD_FREE(flags, sizeof(*flags) * bufs_per_obdo);
 368         OBD_FREE(count, sizeof(*count) * bufs_per_obdo);
 369         OBD_FREE(offset, sizeof(*offset) * bufs_per_obdo);
 370         RETURN(err);
 371 }
 372
 373 #endif
 374
 375
 376 struct address_space_operations ll_aops = {
 377         readpage: ll_readpage,
 378 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
 379         direct_IO: ll_direct_IO,
 380         writepage: ll_writepage,
 381         sync_page: block_sync_page,
 382         prepare_write: ll_prepare_write,
 383         commit_write: ll_commit_write,
 384         bmap: NULL
 385 #endif
 386 };