lustre/obdfs/rw.c

   1 /*
   2  * OBDFS Super operations
   3  *
   4  * Copyright (C) 1996, 1997, Olaf Kirch <okir@monad.swb.de>
   5  * Copryright (C) 1999 Stelias Computing Inc,
   6  *                (author Peter J. Braam <braam@stelias.com>)
   7  * Copryright (C) 1999 Seagate Technology Inc.
   8 */
   9
  10
  11 #include <linux/config.h>
  12 #include <linux/kernel.h>
  13 #include <linux/mm.h>
  14 #include <linux/string.h>
  15 #include <linux/stat.h>
  16 #include <linux/errno.h>
  17 #include <linux/locks.h>
  18 #include <linux/unistd.h>
  19
  20 #include <asm/system.h>
  21 #include <asm/uaccess.h>
  22
  23 #include <linux/fs.h>
  24 #include <linux/stat.h>
  25 #include <asm/uaccess.h>
  26 #include <linux/vmalloc.h>
  27 #include <asm/segment.h>
  28 #include <linux/mm.h>
  29 #include <linux/pagemap.h>
  30 #include <linux/smp_lock.h>
  31
  32 #include <linux/obd_support.h>
  33 #include <linux/obd_ext2.h>
  34 #include <linux/obdfs.h>
  35
  36
  37 /* SYNCHRONOUS I/O for an inode */
  38 static int obdfs_brw(int rw, struct inode *inode, struct page *page, int create)
  39 {
  40         obd_count        num_obdo = 1;
  41         obd_count        bufs_per_obdo = 1;
  42         struct obdo     *oa;
  43         char            *buf = (char *)page_address(page);
  44         obd_size         count = PAGE_SIZE;
  45         obd_off          offset = ((obd_off)page->index) << PAGE_SHIFT;
  46         obd_flag         flags = create ? OBD_BRW_CREATE : 0;
  47         int              err;
  48
  49         ENTRY;
  50         if (IOPS(inode, brw) == NULL) {
  51                 printk(KERN_ERR __FUNCTION__ ": no brw method!\n");
  52                 EXIT;
  53                 return -EIO;
  54         }
  55
  56         oa = obdo_fromid(IID(inode), inode->i_ino, OBD_MD_FLNOTOBD);
  57         if ( IS_ERR(oa) ) {
  58                 EXIT;
  59                 return PTR_ERR(oa);
  60         }
  61         obdfs_from_inode(oa, inode);
  62
  63         err = IOPS(inode, brw)(rw, IID(inode), num_obdo, &oa, &bufs_per_obdo,
  64                                &buf, &count, &offset, &flags);
  65
  66         if ( !err )
  67                 obdfs_to_inode(inode, oa); /* copy o_blocks to i_blocks */
  68
  69         obdo_free(oa);
  70
  71         EXIT;
  72         return err;
  73 } /* obdfs_brw */
  74
  75 /* returns the page unlocked, but with a reference */
  76 int obdfs_readpage(struct dentry *dentry, struct page *page)
  77 {
  78         struct inode *inode = dentry->d_inode;
  79         int rc;
  80
  81         ENTRY;
  82         /* PDEBUG(page, "READ"); */
  83         rc = obdfs_brw(READ, inode, page, 0);
  84         if ( !rc ) {
  85                 SetPageUptodate(page);
  86                 UnlockPage(page);
  87         }
  88         /* PDEBUG(page, "READ"); */
  89         EXIT;
  90         return rc;
  91 } /* obdfs_readpage */
  92
  93 static kmem_cache_t *obdfs_pgrq_cachep = NULL;
  94
  95 int obdfs_init_pgrqcache(void)
  96 {
  97         ENTRY;
  98         if (obdfs_pgrq_cachep == NULL) {
  99                 CDEBUG(D_CACHE, "allocating obdfs_pgrq_cache\n");
 100                 obdfs_pgrq_cachep = kmem_cache_create("obdfs_pgrq",
 101                                                       sizeof(struct obdfs_pgrq),
 102                                                       0, SLAB_HWCACHE_ALIGN,
 103                                                       NULL, NULL);
 104                 if (obdfs_pgrq_cachep == NULL) {
 105                         EXIT;
 106                         return -ENOMEM;
 107                 } else {
 108                         CDEBUG(D_CACHE, "allocated cache at %p\n",
 109                                obdfs_pgrq_cachep);
 110                 }
 111         } else {
 112                 CDEBUG(D_CACHE, "using existing cache at %p\n",
 113                        obdfs_pgrq_cachep);
 114         }
 115         EXIT;
 116         return 0;
 117 } /* obdfs_init_wreqcache */
 118
 119 inline void obdfs_pgrq_del(struct obdfs_pgrq *pgrq)
 120 {
 121         --obdfs_cache_count;
 122         CDEBUG(D_INFO, "deleting page %p from list [count %ld]\n",
 123                pgrq->rq_page, obdfs_cache_count);
 124         OBDClearCachePage(pgrq->rq_page);
 125         list_del(&pgrq->rq_plist);
 126         kmem_cache_free(obdfs_pgrq_cachep, pgrq);
 127 }
 128
 129 void obdfs_cleanup_pgrqcache(void)
 130 {
 131         ENTRY;
 132         if (obdfs_pgrq_cachep != NULL) {
 133                 CDEBUG(D_CACHE, "destroying obdfs_pgrqcache at %p, count %ld\n",
 134                        obdfs_pgrq_cachep, obdfs_cache_count);
 135                 if (kmem_cache_destroy(obdfs_pgrq_cachep))
 136                         printk(KERN_INFO __FUNCTION__
 137                                ": unable to free all of cache\n");
 138                 obdfs_pgrq_cachep = NULL;
 139         } else
 140                 printk(KERN_INFO __FUNCTION__ ": called with NULL pointer\n");
 141
 142         EXIT;
 143 } /* obdfs_cleanup_wreqcache */
 144
 145
 146 /*
 147  * See whether a specific page in the page cache.
 148  * Called with the list lock held.
 149  */
 150 #ifdef PG_obdcache
 151 #define obdfs_find_in_page_list(inode, page) OBDAddCachePage(page)
 152 #else
 153 static int obdfs_find_in_page_list(struct inode *inode, struct page *page)
 154 {
 155         struct list_head *page_list = obdfs_iplist(inode);
 156         struct list_head *tmp;
 157
 158         ENTRY;
 159
 160         CDEBUG(D_INFO, "looking for inode %ld page %p\n", inode->i_ino, page);
 161         OIDEBUG(inode);
 162
 163         if (list_empty(page_list)) {
 164                 CDEBUG(D_INFO, "empty list\n");
 165                 EXIT;
 166                 return 0;
 167         }
 168         tmp = page_list;
 169         while ( (tmp = tmp->next) != page_list ) {
 170                 struct obdfs_pgrq *pgrq;
 171
 172                 pgrq = list_entry(tmp, struct obdfs_pgrq, rq_plist);
 173                 if (pgrq->rq_page == page) {
 174                         CDEBUG(D_INFO, "found page %p in list\n", page);
 175                         EXIT;
 176                         return 1;
 177                 }
 178         }
 179
 180         EXIT;
 181         return 0;
 182 } /* obdfs_find_in_page_list */
 183 #endif
 184
 185
 186 /* called with the list lock held */
 187 static struct page *obdfs_find_page_index(struct inode *inode,
 188                                           unsigned long index)
 189 {
 190         struct list_head *page_list = obdfs_iplist(inode);
 191         struct list_head *tmp;
 192         struct page *page;
 193
 194         ENTRY;
 195
 196         CDEBUG(D_INFO, "looking for inode %ld pageindex %ld\n",
 197                inode->i_ino, index);
 198         OIDEBUG(inode);
 199
 200         if (list_empty(page_list)) {
 201                 EXIT;
 202                 return NULL;
 203         }
 204         tmp = page_list;
 205         while ( (tmp = tmp->next) != page_list ) {
 206                 struct obdfs_pgrq *pgrq;
 207
 208                 pgrq = list_entry(tmp, struct obdfs_pgrq, rq_plist);
 209                 page = pgrq->rq_page;
 210                 if (index == page->index) {
 211                         CDEBUG(D_INFO,
 212                                "INDEX SEARCH found page %p, index %ld\n",
 213                                page, index);
 214                         EXIT;
 215                         return page;
 216                 }
 217         }
 218
 219         EXIT;
 220         return NULL;
 221 } /* obdfs_find_page_index */
 222
 223
 224 /* call and free pages from Linux page cache: called with io lock on inodes */
 225 int obdfs_do_vec_wr(struct inode **inodes, obd_count num_io,
 226                     obd_count num_obdos, struct obdo **obdos,
 227                     obd_count *oa_bufs, struct page **pages, char **bufs,
 228                     obd_size *counts, obd_off *offsets, obd_flag *flags)
 229 {
 230         int err;
 231
 232         ENTRY;
 233         if (IOPS(inodes[0], brw) == NULL) {
 234                 printk(KERN_ERR __FUNCTION__ ": no brw method!\n");
 235                 EXIT;
 236                 return -EIO;
 237         }
 238
 239         CDEBUG(D_INFO, "writing %d page(s), %d obdo(s) in vector\n",
 240                num_io, num_obdos);
 241         { /* DEBUGGING */
 242                 int i;
 243                 printk("OBDOS: ");
 244                 for (i = 0; i < num_obdos; i++)
 245                         printk("%ld:0x%p ", (long)obdos[i]->o_id, obdos[i]);
 246
 247                 printk("\nPAGES: ");
 248                 for (i = 0; i < num_io; i++)
 249                         printk("0x%p ", pages[i]);
 250                 printk("\n");
 251         }
 252
 253         err = IOPS(inodes[0], brw)(WRITE, IID(inodes[0]), num_obdos, obdos,
 254                                   oa_bufs, bufs, counts, offsets, flags);
 255
 256         CDEBUG(D_CACHE, "BRW done\n");
 257         /* release the pages from the page cache */
 258         while ( num_io > 0 ) {
 259                 --num_io;
 260                 CDEBUG(D_INFO, "calling put_page for %p, index %ld\n",
 261                        pages[num_io], pages[num_io]->index);
 262                 /* PDEBUG(pages[num_io], "do_vec_wr"); */
 263                 put_page(pages[num_io]);
 264                 /* PDEBUG(pages[num_io], "do_vec_wr"); */
 265         }
 266         CDEBUG(D_CACHE, "put_page done\n");
 267
 268         while ( num_obdos > 0) {
 269                 --num_obdos;
 270                 CDEBUG(D_INFO, "free obdo %ld\n",(long)obdos[num_obdos]->o_id);
 271                 obdfs_to_inode(inodes[num_obdos], obdos[num_obdos]);
 272                 obdo_free(obdos[num_obdos]);
 273         }
 274         CDEBUG(D_CACHE, "obdo_free done\n");
 275         EXIT;
 276         return err;
 277 }
 278
 279
 280 /*
 281  * Add a page to the write request cache list for later writing
 282  * ASYNCHRONOUS write method.
 283  */
 284 static int obdfs_add_page_to_cache(struct inode *inode, struct page *page)
 285 {
 286         int res = 0;
 287
 288         ENTRY;
 289
 290         /* If this page isn't already in the inode page list, add it */
 291         obd_down(&obdfs_i2sbi(inode)->osi_list_mutex);
 292         if ( !obdfs_find_in_page_list(inode, page) ) {
 293                 struct obdfs_pgrq *pgrq;
 294                 pgrq = kmem_cache_alloc(obdfs_pgrq_cachep, SLAB_KERNEL);
 295                 if (!pgrq) {
 296                         EXIT;
 297                         obd_up(&obdfs_i2sbi(inode)->osi_list_mutex);
 298                         return -ENOMEM;
 299                 }
 300                 memset(pgrq, 0, sizeof(*pgrq));
 301
 302                 pgrq->rq_page = page;
 303                 pgrq->rq_jiffies = jiffies;
 304                 get_page(pgrq->rq_page);
 305                 list_add(&pgrq->rq_plist, obdfs_iplist(inode));
 306                 obdfs_cache_count++;
 307                 CDEBUG(D_INFO,
 308                        "added inode %ld page %p, pgrq: %p, cache count [%ld]\n",
 309                        inode->i_ino, page, pgrq, obdfs_cache_count);
 310         }
 311
 312         /* If inode isn't already on the superblock inodes list, add it,
 313          * and increase ref count on inode so it doesn't disappear on us.
 314          *
 315          * We increment the reference count on the inode to keep it from
 316          * being freed from memory.  This _should_ be an iget() with an
 317          * iput() in both flush_reqs() and put_inode(), but since put_inode()
 318          * is called from iput() we can't call iput() again there.  Instead
 319          * we just increment/decrement i_count, which is essentially what
 320          * iget/iput do for an inode already in memory.
 321          */
 322         if ( list_empty(obdfs_islist(inode)) ) {
 323                 inode->i_count++;
 324                 CDEBUG(D_INFO, "adding inode %ld to superblock list %p\n",
 325                        inode->i_ino, obdfs_slist(inode));
 326                 list_add(obdfs_islist(inode), obdfs_slist(inode));
 327         }
 328         obd_up(&obdfs_i2sbi(inode)->osi_list_mutex);
 329
 330         /* XXX For testing purposes, we write out the page here.
 331          *     In the future, a flush daemon will write out the page.
 332         res = obdfs_flush_reqs(obdfs_slist(inode), ~0UL);
 333          */
 334
 335         EXIT;
 336         return res;
 337 } /* obdfs_add_page_to_cache */
 338
 339
 340 /* select between SYNC and ASYNC I/O methods */
 341 int obdfs_do_writepage(struct inode *inode, struct page *page, int sync)
 342 {
 343         int err;
 344
 345         ENTRY;
 346         /* PDEBUG(page, "WRITEPAGE"); */
 347         if ( sync )
 348                 err = obdfs_brw(WRITE, inode, page, 1);
 349         else {
 350                 err = obdfs_add_page_to_cache(inode, page);
 351                 CDEBUG(D_INFO, "DO_WR ino: %ld, page %p, err %d, uptodate %d\n",
 352                        inode->i_ino, page, err, Page_Uptodate(page));
 353         }
 354
 355         if ( !err )
 356                 SetPageUptodate(page);
 357         /* PDEBUG(page,"WRITEPAGE"); */
 358         EXIT;
 359         return err;
 360 } /* obdfs_do_writepage */
 361
 362 /* returns the page unlocked, but with a reference */
 363 int obdfs_writepage(struct dentry *dentry, struct page *page)
 364 {
 365         return obdfs_do_writepage(dentry->d_inode, page, 0);
 366 }
 367
 368 /*
 369  * This does the "real" work of the write. The generic routine has
 370  * allocated the page, locked it, done all the page alignment stuff
 371  * calculations etc. Now we should just copy the data from user
 372  * space and write it back to the real medium..
 373  *
 374  * If the writer ends up delaying the write, the writer needs to
 375  * increment the page use counts until he is done with the page.
 376  *
 377  * Return value is the number of bytes written.
 378  */
 379 int obdfs_write_one_page(struct file *file, struct page *page,
 380                          unsigned long offset, unsigned long bytes,
 381                          const char * buf)
 382 {
 383         struct inode *inode = file->f_dentry->d_inode;
 384         int err;
 385
 386         ENTRY;
 387         /* We check for complete page writes here, as we then don't have to
 388          * get the page before writing over everything anyways.
 389          */
 390         if ( !Page_Uptodate(page) && (offset != 0 || bytes != PAGE_SIZE) ) {
 391                 err = obdfs_brw(READ, inode, page, 0);
 392                 if ( err )
 393                         return err;
 394                 SetPageUptodate(page);
 395         }
 396
 397         if (copy_from_user((u8*)page_address(page) + offset, buf, bytes))
 398                 return -EFAULT;
 399
 400         lock_kernel();
 401         err = obdfs_writepage(file->f_dentry, page);
 402         unlock_kernel();
 403
 404         return (err < 0 ? err : bytes);
 405 } /* obdfs_write_one_page */
 406
 407 /*
 408  * return an up to date page:
 409  *  - if locked is true then is returned locked
 410  *  - if create is true the corresponding disk blocks are created
 411  *  - page is held, i.e. caller must release the page
 412  *
 413  * modeled on NFS code.
 414  */
 415 struct page *obdfs_getpage(struct inode *inode, unsigned long offset,
 416                            int create, int locked)
 417 {
 418         struct page * page;
 419         int index;
 420         int err;
 421
 422         ENTRY;
 423
 424         offset = offset & PAGE_CACHE_MASK;
 425         CDEBUG(D_INFO, "ino: %ld, offset %ld, create %d, locked %d\n",
 426                inode->i_ino, offset, create, locked);
 427         index = offset >> PAGE_CACHE_SHIFT;
 428
 429         page = grab_cache_page(&inode->i_data, index);
 430
 431         /* Yuck, no page */
 432         if (! page) {
 433             printk(KERN_WARNING " grab_cache_page says no dice ...\n");
 434             EXIT;
 435             return NULL;
 436         }
 437
 438         /* PDEBUG(page, "GETPAGE: got page - before reading\n"); */
 439         /* now check if the data in the page is up to date */
 440         if ( Page_Uptodate(page)) {
 441                 if (!locked)
 442                         UnlockPage(page);
 443                 EXIT;
 444                 return page;
 445         }
 446
 447
 448 #ifdef EXT2_OBD_DEBUG
 449         if ((obd_debug_level & D_INFO) && obdfs_find_page_index(inode, index)) {
 450                 CDEBUG(D_INFO, "OVERWRITE: found dirty page %p, index %ld\n",
 451                        page, page->index);
 452         }
 453 #endif
 454
 455         err = obdfs_brw(READ, inode, page, create);
 456
 457         if ( err ) {
 458                 SetPageError(page);
 459                 UnlockPage(page);
 460                 EXIT;
 461                 return page;
 462         }
 463
 464         if ( !locked )
 465                 UnlockPage(page);
 466         SetPageUptodate(page);
 467         /* PDEBUG(page,"GETPAGE - after reading"); */
 468         EXIT;
 469         return page;
 470 } /* obdfs_getpage */
 471
 472
 473 void obdfs_truncate(struct inode *inode)
 474 {
 475         struct obdo *oa;
 476         int err;
 477         ENTRY;
 478
 479         obdfs_dequeue_pages(inode);
 480
 481         if (IOPS(inode, punch) == NULL) {
 482                 printk(KERN_ERR __FUNCTION__ ": no punch method!\n");
 483                 EXIT;
 484                 return;
 485         }
 486         oa = obdo_alloc();
 487         if ( !oa ) {
 488                 /* XXX This would give an inconsistent FS, so deal with it as
 489                  * best we can for now - an obdo on the stack is not pretty.
 490                  */
 491                 struct obdo obdo;
 492
 493                 printk(__FUNCTION__ ": obdo_alloc failed - using stack!\n");
 494
 495                 obdo.o_valid = OBD_MD_FLNOTOBD;
 496                 obdfs_from_inode(&obdo, inode);
 497
 498                 err = IOPS(inode, punch)(IID(inode), &obdo, obdo.o_size, 0);
 499         } else {
 500                 oa->o_valid = OBD_MD_FLNOTOBD;
 501                 obdfs_from_inode(oa, inode);
 502
 503                 CDEBUG(D_PUNCH, "calling punch for %ld (%Lu bytes at 0)\n",
 504                        (long)oa->o_id, oa->o_size);
 505                 err = IOPS(inode, punch)(IID(inode), oa, oa->o_size, 0);
 506
 507                 obdo_free(oa);
 508         }
 509
 510         if (err) {
 511                 printk(__FUNCTION__ ": obd_truncate fails (%d)\n", err);
 512                 EXIT;
 513                 return;
 514         }
 515         EXIT;
 516 }