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