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 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.off = ((obd_off)page->index) << PAGE_SHIFT;
  90
  91         if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size))
  92                 pg.count = inode->i_size % PAGE_SIZE;
  93         else
  94                 pg.count = PAGE_SIZE;
  95
  96         pg.flag = create ? OBD_BRW_CREATE : 0;
  97
  98         err = obd_brw(cmd, ll_i2obdconn(inode),lsm, 1, &pg, ll_sync_io_cb, cbd);
  99
 100         RETURN(err);
 101 } /* ll_brw */
 102
 103 /* returns the page unlocked, but with a reference */
 104 static int ll_readpage(struct file *file, struct page *page)
 105 {
 106         struct inode *inode = page->mapping->host;
 107         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 108         int rc = 0;
 109         ENTRY;
 110
 111         if (!PageLocked(page))
 112                 LBUG();
 113
 114         if (inode->i_size <= offset) {
 115                 memset(kmap(page), 0, PAGE_SIZE);
 116                 kunmap(page);
 117                 GOTO(readpage_out, rc);
 118         }
 119
 120         if (PageUptodate(page)) {
 121                 CERROR("Explain this please?\n");
 122                 GOTO(readpage_out, rc);
 123         }
 124
 125         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
 126         EXIT;
 127
 128  readpage_out:
 129         if (!rc)
 130                 SetPageUptodate(page);
 131         unlock_page(page);
 132         return 0;
 133 } /* ll_readpage */
 134
 135 void ll_truncate(struct inode *inode)
 136 {
 137         struct obdo oa = {0};
 138         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 139         struct lustre_handle *lockhs = NULL;
 140         int err;
 141         ENTRY;
 142
 143         if (!lsm) {
 144                 /* object not yet allocated */
 145                 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 146                 return;
 147         }
 148
 149         oa.o_id = lsm->lsm_object_id;
 150         oa.o_mode = inode->i_mode;
 151         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 152
 153         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after "LPD64")\n",
 154                oa.o_id, inode->i_size);
 155
 156         err = ll_size_lock(inode, lsm, inode->i_size, LCK_PW, &lockhs);
 157         if (err) {
 158                 CERROR("ll_size_lock failed: %d\n", err);
 159                 /* FIXME: What to do here?  It's too late to back out... */
 160                 LBUG();
 161         }
 162
 163         /* truncate == punch from new size to absolute end of file */
 164         err = obd_punch(ll_i2obdconn(inode), &oa, lsm, inode->i_size,
 165                         OBD_OBJECT_EOF);
 166         if (err)
 167                 CERROR("obd_truncate fails (%d)\n", err);
 168         else
 169                 obdo_to_inode(inode, &oa, oa.o_valid);
 170
 171         err = ll_size_unlock(inode, lsm, LCK_PW, lockhs);
 172         if (err)
 173                 CERROR("ll_size_unlock failed: %d\n", err);
 174
 175         EXIT;
 176         return;
 177 } /* ll_truncate */
 178
 179 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 180
 181 static int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 182                             unsigned to)
 183 {
 184         struct inode *inode = page->mapping->host;
 185         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 186         int rc = 0;
 187         char *addr;
 188         ENTRY;
 189
 190         addr = kmap(page);
 191         if (!PageLocked(page))
 192                 LBUG();
 193
 194         if (Page_Uptodate(page))
 195                 GOTO(prepare_done, rc);
 196
 197         /* We're completely overwriting an existing page, so _don't_ set it up
 198          * to date until commit_write */
 199         if (from == 0 && to == PAGE_SIZE)
 200                 RETURN(0);
 201
 202         /* We are writing to a new page, no need to read old data */
 203         if (inode->i_size <= offset) {
 204                 memset(addr, 0, PAGE_SIZE);
 205                 GOTO(prepare_done, rc=0);
 206         }
 207
 208         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
 209
 210         EXIT;
 211  prepare_done:
 212         if (!rc)
 213                 SetPageUptodate(page);
 214
 215         return rc;
 216 }
 217
 218 /* returns the page unlocked, but with a reference */
 219 static int ll_writepage(struct page *page)
 220 {
 221         struct inode *inode = page->mapping->host;
 222         int err;
 223         ENTRY;
 224
 225         if (!PageLocked(page))
 226                 LBUG();
 227
 228         err = ll_brw(OBD_BRW_WRITE, inode, page, 1);
 229         if ( !err ) {
 230                 //SetPageUptodate(page);
 231                 set_page_clean(page);
 232         } else {
 233                 CERROR("ll_brw failure %d\n", err);
 234         }
 235         unlock_page(page);
 236         RETURN(err);
 237 }
 238
 239
 240 /* SYNCHRONOUS I/O to object storage for an inode -- object attr will be updated
 241  * too */
 242 static int ll_commit_write(struct file *file, struct page *page,
 243                            unsigned from, unsigned to)
 244 {
 245         int create = 1;
 246         struct inode *inode = page->mapping->host;
 247         struct ll_inode_info *lli = ll_i2info(inode);
 248         struct lov_stripe_md *md = lli->lli_smd;
 249         struct brw_page pg;
 250         int err;
 251         loff_t size;
 252         struct io_cb_data *cbd = ll_init_cb();
 253         ENTRY;
 254
 255         CHECK_MOUNT_EPOCH(inode);
 256
 257         pg.pg = page;
 258         pg.count = to;
 259         pg.off = (((obd_off)page->index) << PAGE_SHIFT);
 260         pg.flag = create ? OBD_BRW_CREATE : 0;
 261
 262         if (!cbd)
 263                 RETURN(-ENOMEM);
 264
 265         SetPageUptodate(page);
 266
 267         if (!PageLocked(page))
 268                 LBUG();
 269
 270         CDEBUG(D_INODE, "commit_page writing (off "LPD64"), count "LPD64"\n",
 271                pg.off, pg.count);
 272
 273         err = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode), md,
 274                       1, &pg, ll_sync_io_cb, cbd);
 275         kunmap(page);
 276
 277         size = pg.off + pg.count;
 278         /* do NOT truncate when writing in the middle of a file */
 279         if (size > inode->i_size)
 280                 inode->i_size = size;
 281
 282         RETURN(err);
 283 } /* ll_commit_write */
 284
 285
 286 static int ll_direct_IO(int rw, struct inode *inode, struct kiobuf *iobuf,
 287                         unsigned long blocknr, int blocksize)
 288 {
 289         obd_count bufs_per_obdo = iobuf->nr_pages;
 290         struct ll_inode_info *lli = ll_i2info(inode);
 291         struct lov_stripe_md *lsm = lli->lli_smd;
 292         struct brw_page *pga;
 293         int i, rc = 0;
 294         struct io_cb_data *cbd;
 295
 296         CHECK_MOUNT_EPOCH(inode);
 297
 298         ENTRY;
 299         if (!lsm || !lsm->lsm_object_id)
 300                 RETURN(-ENOMEM);
 301
 302         if (blocksize != PAGE_SIZE) {
 303                 CERROR("direct_IO blocksize != PAGE_SIZE\n");
 304                 RETURN(-EINVAL);
 305         }
 306
 307         cbd = ll_init_cb();
 308         if (!cbd)
 309                 RETURN(-ENOMEM);
 310
 311         OBD_ALLOC(pga, sizeof(*pga) * bufs_per_obdo);
 312         if (!pga) {
 313                 OBD_FREE(cbd, sizeof(*cbd));
 314                 RETURN(-ENOMEM);
 315         }
 316
 317         /* NB: we can't use iobuf->maplist[i]->index for the offset
 318          * instead of "blocknr" because ->index contains garbage.
 319          */
 320         for (i = 0; i < bufs_per_obdo; i++, blocknr++) {
 321                 pga[i].pg = iobuf->maplist[i];
 322                 pga[i].count = PAGE_SIZE;
 323                 pga[i].off = (obd_off)blocknr << PAGE_SHIFT;
 324                 pga[i].flag = OBD_BRW_CREATE;
 325         }
 326
 327         rc = obd_brw(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 328                      ll_i2obdconn(inode), lsm, bufs_per_obdo, pga,
 329                      ll_sync_io_cb, cbd);
 330         if (rc == 0)
 331                 rc = bufs_per_obdo * PAGE_SIZE;
 332
 333         OBD_FREE(pga, sizeof(*pga) * bufs_per_obdo);
 334         RETURN(rc);
 335 }
 336
 337 int ll_flush_inode_pages(struct inode * inode)
 338 {
 339         obd_count        bufs_per_obdo = 0;
 340         obd_size         *count = NULL;
 341         obd_off          *offset = NULL;
 342         obd_flag         *flags = NULL;
 343         int              err = 0;
 344
 345         ENTRY;
 346
 347         spin_lock(&pagecache_lock);
 348
 349         spin_unlock(&pagecache_lock);
 350
 351
 352         OBD_ALLOC(count, sizeof(*count) * bufs_per_obdo);
 353         OBD_ALLOC(offset, sizeof(*offset) * bufs_per_obdo);
 354         OBD_ALLOC(flags, sizeof(*flags) * bufs_per_obdo);
 355         if (!count || !offset || !flags)
 356                 GOTO(out, err=-ENOMEM);
 357
 358 #if 0
 359         for (i = 0 ; i < bufs_per_obdo ; i++) {
 360                 count[i] = PAGE_SIZE;
 361                 offset[i] = ((obd_off)(iobuf->maplist[i])->index) << PAGE_SHIFT;
 362                 flags[i] = OBD_BRW_CREATE;
 363         }
 364
 365         err = obd_brw(OBD_BRW_WRITE, ll_i2obdconn(inode),
 366                       ll_i2info(inode)->lli_smd, bufs_per_obdo,
 367                       iobuf->maplist, count, offset, flags, NULL, NULL);
 368         if (err == 0)
 369                 err = bufs_per_obdo * 4096;
 370 #endif
 371  out:
 372         OBD_FREE(flags, sizeof(*flags) * bufs_per_obdo);
 373         OBD_FREE(count, sizeof(*count) * bufs_per_obdo);
 374         OBD_FREE(offset, sizeof(*offset) * bufs_per_obdo);
 375         RETURN(err);
 376 }
 377
 378 #endif
 379
 380
 381 struct address_space_operations ll_aops = {
 382         readpage: ll_readpage,
 383 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
 384         direct_IO: ll_direct_IO,
 385         writepage: ll_writepage,
 386         sync_page: block_sync_page,
 387         prepare_write: ll_prepare_write,
 388         commit_write: ll_commit_write,
 389         bmap: NULL
 390 #endif
 391 };