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/locks.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(2,4,9))
  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(2,4,10))
  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 *md = lli->lli_smd;
  78         struct io_cb_data *cbd = ll_init_cb();
  79         struct brw_page pg;
  80         int err;
  81         ENTRY;
  82
  83         CHECK_MOUNT_EPOCH(inode);
  84
  85         if (!cbd)
  86                 RETURN(-ENOMEM);
  87
  88         pg.pg = page;
  89         pg.count = PAGE_SIZE;
  90         pg.off = ((obd_off)page->index) << PAGE_SHIFT;
  91         pg.flag = create ? OBD_BRW_CREATE : 0;
  92
  93         err = obd_brw(cmd, ll_i2obdconn(inode), md, 1, &pg, ll_sync_io_cb, cbd);
  94
  95         RETURN(err);
  96 } /* ll_brw */
  97
  98 /* returns the page unlocked, but with a reference */
  99 static int ll_readpage(struct file *file, struct page *page)
 100 {
 101         struct inode *inode = page->mapping->host;
 102         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 103         int rc = 0;
 104         ENTRY;
 105
 106         if (!PageLocked(page))
 107                 LBUG();
 108
 109         if (inode->i_size <= offset) {
 110                 memset(kmap(page), 0, PAGE_SIZE);
 111                 kunmap(page);
 112                 GOTO(readpage_out, rc);
 113         }
 114
 115         if (Page_Uptodate(page)) {
 116                 CERROR("Explain this please?\n");
 117                 GOTO(readpage_out, rc);
 118         }
 119
 120         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
 121         EXIT;
 122
 123  readpage_out:
 124         if (!rc)
 125                 SetPageUptodate(page);
 126         UnlockPage(page);
 127         return 0;
 128 } /* ll_readpage */
 129
 130
 131 static int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 132                             unsigned to)
 133 {
 134         struct inode *inode = page->mapping->host;
 135         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 136         int rc = 0;
 137         char *addr;
 138         ENTRY;
 139
 140         addr = kmap(page);
 141         if (!PageLocked(page))
 142                 LBUG();
 143
 144         if (Page_Uptodate(page))
 145                 GOTO(prepare_done, rc);
 146
 147         /* We're completely overwriting an existing page, so _don't_ set it up
 148          * to date until commit_write */
 149         if (from == 0 && to == PAGE_SIZE)
 150                 RETURN(0);
 151
 152         /* We are writing to a new page, no need to read old data */
 153         if (inode->i_size <= offset) {
 154                 memset(addr, 0, PAGE_SIZE);
 155                 goto prepare_done;
 156         }
 157
 158         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
 159
 160         EXIT;
 161  prepare_done:
 162         if (!rc)
 163                 SetPageUptodate(page);
 164
 165         return rc;
 166 }
 167
 168 /* returns the page unlocked, but with a reference */
 169 static int ll_writepage(struct page *page)
 170 {
 171         struct inode *inode = page->mapping->host;
 172         int err;
 173         ENTRY;
 174
 175         if (!PageLocked(page))
 176                 LBUG();
 177
 178         err = ll_brw(OBD_BRW_WRITE, inode, page, 1);
 179         if ( !err ) {
 180                 //SetPageUptodate(page);
 181                 set_page_clean(page);
 182         } else {
 183                 CERROR("ll_brw failure %d\n", err);
 184         }
 185         unlock_page(page);
 186         RETURN(err);
 187 }
 188
 189
 190 /* SYNCHRONOUS I/O to object storage for an inode -- object attr will be updated
 191  * too */
 192 static int ll_commit_write(struct file *file, struct page *page,
 193                            unsigned from, unsigned to)
 194 {
 195         int create = 1;
 196         struct inode *inode = page->mapping->host;
 197         struct ll_inode_info *lli = ll_i2info(inode);
 198         struct lov_stripe_md *md = lli->lli_smd;
 199         struct brw_page pg;
 200         int err;
 201         loff_t size;
 202         struct io_cb_data *cbd = ll_init_cb();
 203         ENTRY;
 204
 205         CHECK_MOUNT_EPOCH(inode);
 206
 207         pg.pg = page;
 208         pg.count = to;
 209         pg.off = (((obd_off)page->index) << PAGE_SHIFT);
 210         pg.flag = create ? OBD_BRW_CREATE : 0;
 211
 212         if (!cbd)
 213                 RETURN(-ENOMEM);
 214
 215         SetPageUptodate(page);
 216
 217         if (!PageLocked(page))
 218                 LBUG();
 219
 220         CDEBUG(D_INODE, "commit_page writing (off "LPD64"), count %ld\n",
 221                pg.off, pg.count);
 222
 223         err = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode), md,
 224                       1, &pg, ll_sync_io_cb, cbd);
 225         kunmap(page);
 226
 227         size = pg.off + pg.count;
 228         /* do NOT truncate when writing in the middle of a file */
 229         if (size > inode->i_size)
 230                 inode->i_size = size;
 231
 232         RETURN(err);
 233 } /* ll_commit_write */
 234
 235 void ll_truncate(struct inode *inode)
 236 {
 237         struct obdo oa = {0};
 238         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 239         struct lustre_handle *lockhs = NULL;
 240         int err;
 241         ENTRY;
 242
 243         if (!lsm) {
 244                 /* object not yet allocated */
 245                 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 246                 return;
 247         }
 248
 249         oa.o_id = lsm->lsm_object_id;
 250         oa.o_mode = inode->i_mode;
 251         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 252
 253         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after "LPD64")\n",
 254                oa.o_id, inode->i_size);
 255
 256         err = ll_size_lock(inode, lsm, inode->i_size, LCK_PW, &lockhs);
 257         if (err) {
 258                 CERROR("ll_size_lock failed: %d\n", err);
 259                 /* FIXME: What to do here?  It's too late to back out... */
 260                 LBUG();
 261         }
 262
 263         /* truncate == punch to/from start from/to end:
 264            set end to -1 for that. */
 265         err = obd_punch(ll_i2obdconn(inode), &oa, lsm, inode->i_size,
 266                         OBD_PUNCH_EOF);
 267         if (err)
 268                 CERROR("obd_truncate fails (%d)\n", err);
 269         else
 270                 obdo_to_inode(inode, &oa, oa.o_valid);
 271
 272         err = ll_size_unlock(inode, lsm, LCK_PW, lockhs);
 273         if (err)
 274                 CERROR("ll_size_unlock failed: %d\n", err);
 275
 276         EXIT;
 277         return;
 278 } /* ll_truncate */
 279
 280 static int ll_direct_IO(int rw, struct inode *inode, struct kiobuf *iobuf,
 281                         unsigned long blocknr, int blocksize)
 282 {
 283         obd_count bufs_per_obdo = iobuf->nr_pages;
 284         struct ll_inode_info *lli = ll_i2info(inode);
 285         struct lov_stripe_md *lsm = lli->lli_smd;
 286         struct brw_page *pga;
 287         int i, rc = 0;
 288         struct io_cb_data *cbd;
 289
 290         CHECK_MOUNT_EPOCH(inode);
 291
 292         ENTRY;
 293         if (!lsm || !lsm->lsm_object_id)
 294                 RETURN(-ENOMEM);
 295
 296         if (blocksize != PAGE_SIZE) {
 297                 CERROR("direct_IO blocksize != PAGE_SIZE\n");
 298                 RETURN(-EINVAL);
 299         }
 300
 301         cbd = ll_init_cb();
 302         if (!cbd)
 303                 RETURN(-ENOMEM);
 304
 305         OBD_ALLOC(pga, sizeof(*pga) * bufs_per_obdo);
 306         if (!pga) {
 307                 OBD_FREE(cbd, sizeof(*cbd));
 308                 RETURN(-ENOMEM);
 309         }
 310
 311         /* NB: we can't use iobuf->maplist[i]->index for the offset
 312          * instead of "blocknr" because ->index contains garbage.
 313          */
 314         for (i = 0; i < bufs_per_obdo; i++, blocknr++) {
 315                 pga[i].pg = iobuf->maplist[i];
 316                 pga[i].count = PAGE_SIZE;
 317                 pga[i].off = (obd_off)blocknr << PAGE_SHIFT;
 318                 pga[i].flag = OBD_BRW_CREATE;
 319         }
 320
 321         rc = obd_brw(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 322                      ll_i2obdconn(inode), lsm, bufs_per_obdo, pga,
 323                      ll_sync_io_cb, cbd);
 324         if (rc == 0)
 325                 rc = bufs_per_obdo * PAGE_SIZE;
 326
 327         OBD_FREE(pga, sizeof(*pga) * bufs_per_obdo);
 328         RETURN(rc);
 329 }
 330
 331 int ll_flush_inode_pages(struct inode * inode)
 332 {
 333         obd_count        bufs_per_obdo = 0;
 334         obd_size         *count = NULL;
 335         obd_off          *offset = NULL;
 336         obd_flag         *flags = NULL;
 337         int              err = 0;
 338
 339         ENTRY;
 340
 341         spin_lock(&pagecache_lock);
 342
 343         spin_unlock(&pagecache_lock);
 344
 345
 346         OBD_ALLOC(count, sizeof(*count) * bufs_per_obdo);
 347         OBD_ALLOC(offset, sizeof(*offset) * bufs_per_obdo);
 348         OBD_ALLOC(flags, sizeof(*flags) * bufs_per_obdo);
 349         if (!count || !offset || !flags)
 350                 GOTO(out, err=-ENOMEM);
 351
 352 #if 0
 353         for (i = 0 ; i < bufs_per_obdo ; i++) {
 354                 count[i] = PAGE_SIZE;
 355                 offset[i] = ((obd_off)(iobuf->maplist[i])->index) << PAGE_SHIFT;
 356                 flags[i] = OBD_BRW_CREATE;
 357         }
 358
 359         err = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode),
 360                       ll_i2info(inode)->lli_smd, bufs_per_obdo,
 361                       iobuf->maplist, count, offset, flags, NULL, NULL);
 362         if (err == 0)
 363                 err = bufs_per_obdo * 4096;
 364 #endif
 365  out:
 366         OBD_FREE(flags, sizeof(*flags) * bufs_per_obdo);
 367         OBD_FREE(count, sizeof(*count) * bufs_per_obdo);
 368         OBD_FREE(offset, sizeof(*offset) * bufs_per_obdo);
 369         RETURN(err);
 370 }
 371
 372 struct address_space_operations ll_aops = {
 373         readpage: ll_readpage,
 374         writepage: ll_writepage,
 375 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,4,17))
 376         direct_IO: ll_direct_IO,
 377 #endif
 378         sync_page: block_sync_page,
 379         prepare_write: ll_prepare_write,
 380         commit_write: ll_commit_write,
 381         bmap: NULL
 382 };