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         list_del(&pgrq->rq_plist);
 125         kmem_cache_free(obdfs_pgrq_cachep, pgrq);
 126 }
 127
 128 void obdfs_cleanup_pgrqcache(void)
 129 {
 130         ENTRY;
 131         if (obdfs_pgrq_cachep != NULL) {
 132                 CDEBUG(D_CACHE, "destroying obdfs_pgrqcache at %p, count %ld\n",
 133                        obdfs_pgrq_cachep, obdfs_cache_count);
 134                 if (kmem_cache_destroy(obdfs_pgrq_cachep))
 135                         printk(KERN_INFO __FUNCTION__
 136                                ": unable to free all of cache\n");
 137                 obdfs_pgrq_cachep = NULL;
 138         } else
 139                 printk(KERN_INFO __FUNCTION__ ": called with NULL pointer\n");
 140
 141         EXIT;
 142 } /* obdfs_cleanup_wreqcache */
 143
 144
 145 /*
 146  * Find a specific page in the page cache.  If it is found, we return
 147  * the write request struct associated with it, if not found return NULL.
 148  * Called with the list lock held.
 149  */
 150 static struct obdfs_pgrq *
 151 obdfs_find_in_page_list(struct inode *inode, struct page *page)
 152 {
 153         struct list_head *page_list = obdfs_iplist(inode);
 154         struct list_head *tmp;
 155
 156         ENTRY;
 157
 158         CDEBUG(D_INFO, "looking for inode %ld page %p\n", inode->i_ino, page);
 159         OIDEBUG(inode);
 160
 161         if (list_empty(page_list)) {
 162                 CDEBUG(D_INFO, "empty list\n");
 163                 EXIT;
 164                 return NULL;
 165         }
 166         tmp = page_list;
 167         while ( (tmp = tmp->next) != page_list ) {
 168                 struct obdfs_pgrq *pgrq;
 169
 170                 pgrq = list_entry(tmp, struct obdfs_pgrq, rq_plist);
 171                 if (pgrq->rq_page == page) {
 172                         CDEBUG(D_INFO, "found page %p in list\n", page);
 173                         EXIT;
 174                         return pgrq;
 175                 }
 176         }
 177
 178         EXIT;
 179         return NULL;
 180 } /* obdfs_find_in_page_list */
 181
 182
 183 /* called with the list lock held */
 184 static struct page *obdfs_find_page_index(struct inode *inode,
 185                                           unsigned long index)
 186 {
 187         struct list_head *page_list = obdfs_iplist(inode);
 188         struct list_head *tmp;
 189         struct page *page;
 190
 191         ENTRY;
 192
 193         CDEBUG(D_INFO, "looking for inode %ld pageindex %ld\n",
 194                inode->i_ino, index);
 195         OIDEBUG(inode);
 196
 197         if (list_empty(page_list)) {
 198                 EXIT;
 199                 return NULL;
 200         }
 201         tmp = page_list;
 202         while ( (tmp = tmp->next) != page_list ) {
 203                 struct obdfs_pgrq *pgrq;
 204
 205                 pgrq = list_entry(tmp, struct obdfs_pgrq, rq_plist);
 206                 page = pgrq->rq_page;
 207                 if (index == page->index) {
 208                         CDEBUG(D_INFO,
 209                                "INDEX SEARCH found page %p, index %ld\n",
 210                                page, index);
 211                         EXIT;
 212                         return page;
 213                 }
 214         }
 215
 216         EXIT;
 217         return NULL;
 218 } /* obdfs_find_page_index */
 219
 220
 221 /* call and free pages from Linux page cache: called with io lock on inodes */
 222 int obdfs_do_vec_wr(struct inode **inodes, obd_count num_io,
 223                     obd_count num_obdos, struct obdo **obdos,
 224                     obd_count *oa_bufs, struct page **pages, char **bufs,
 225                     obd_size *counts, obd_off *offsets, obd_flag *flags)
 226 {
 227         int err;
 228
 229         ENTRY;
 230         if (IOPS(inodes[0], brw) == NULL) {
 231                 printk(KERN_ERR __FUNCTION__ ": no brw method!\n");
 232                 EXIT;
 233                 return -EIO;
 234         }
 235
 236         CDEBUG(D_INFO, "writing %d page(s), %d obdo(s) in vector\n",
 237                num_io, num_obdos);
 238         { /* DEBUGGING */
 239                 int i;
 240                 printk("OBDOS: ");
 241                 for (i = 0; i < num_obdos; i++)
 242                         printk("%ld:0x%p ", (long)obdos[i]->o_id, obdos[i]);
 243
 244                 printk("\nPAGES: ");
 245                 for (i = 0; i < num_io; i++)
 246                         printk("0x%p ", pages[i]);
 247                 printk("\n");
 248         }
 249
 250         err = IOPS(inodes[0], brw)(WRITE, IID(inodes[0]), num_obdos, obdos,
 251                                   oa_bufs, bufs, counts, offsets, flags);
 252
 253         CDEBUG(D_CACHE, "BRW done\n");
 254         /* release the pages from the page cache */
 255         while ( num_io > 0 ) {
 256                 --num_io;
 257                 CDEBUG(D_INFO, "calling put_page for %p, index %ld\n",
 258                        pages[num_io], pages[num_io]->index);
 259                 /* PDEBUG(pages[num_io], "do_vec_wr"); */
 260                 put_page(pages[num_io]);
 261                 /* PDEBUG(pages[num_io], "do_vec_wr"); */
 262         }
 263         CDEBUG(D_CACHE, "put_page done\n");
 264
 265         while ( num_obdos > 0) {
 266                 --num_obdos;
 267                 CDEBUG(D_INFO, "free obdo %ld\n",(long)obdos[num_obdos]->o_id);
 268                 obdfs_to_inode(inodes[num_obdos], obdos[num_obdos]);
 269                 obdo_free(obdos[num_obdos]);
 270         }
 271         CDEBUG(D_CACHE, "obdo_free done\n");
 272         EXIT;
 273         return err;
 274 }
 275
 276
 277 /*
 278  * Add a page to the write request cache list for later writing
 279  * ASYNCHRONOUS write method.
 280  */
 281 static int obdfs_add_page_to_cache(struct inode *inode, struct page *page)
 282 {
 283         int res = 0;
 284
 285         ENTRY;
 286
 287         /* If this page isn't already in the inode page list, add it */
 288         obd_down(&obdfs_i2sbi(inode)->osi_list_mutex);
 289         if ( !obdfs_find_in_page_list(inode, page) ) {
 290                 struct obdfs_pgrq *pgrq;
 291                 pgrq = kmem_cache_alloc(obdfs_pgrq_cachep, SLAB_KERNEL);
 292                 if (!pgrq) {
 293                         EXIT;
 294                         obd_up(&obdfs_i2sbi(inode)->osi_list_mutex);
 295                         return -ENOMEM;
 296                 }
 297                 memset(pgrq, 0, sizeof(*pgrq));
 298
 299                 pgrq->rq_page = page;
 300                 pgrq->rq_jiffies = jiffies;
 301                 get_page(pgrq->rq_page);
 302                 list_add(&pgrq->rq_plist, obdfs_iplist(inode));
 303                 obdfs_cache_count++;
 304                 CDEBUG(D_INFO,
 305                        "added inode %ld page %p, pgrq: %p, cache count [%ld]\n",
 306                        inode->i_ino, page, pgrq, obdfs_cache_count);
 307         }
 308
 309         /* If inode isn't already on the superblock inodes list, add it,
 310          * and increase ref count on inode so it doesn't disappear on us.
 311          *
 312          * We increment the reference count on the inode to keep it from
 313          * being freed from memory.  This _should_ be an iget() with an
 314          * iput() in both flush_reqs() and put_inode(), but since put_inode()
 315          * is called from iput() we can't call iput() again there.  Instead
 316          * we just increment/decrement i_count, which is essentially what
 317          * iget/iput do for an inode already in memory.
 318          */
 319         if ( list_empty(obdfs_islist(inode)) ) {
 320                 inode->i_count++;
 321                 CDEBUG(D_INFO, "adding inode %ld to superblock list %p\n",
 322                        inode->i_ino, obdfs_slist(inode));
 323                 list_add(obdfs_islist(inode), obdfs_slist(inode));
 324         }
 325         obd_up(&obdfs_i2sbi(inode)->osi_list_mutex);
 326
 327         /* XXX For testing purposes, we write out the page here.
 328          *     In the future, a flush daemon will write out the page.
 329         res = obdfs_flush_reqs(obdfs_slist(inode), ~0UL);
 330          */
 331
 332         EXIT;
 333         return res;
 334 } /* obdfs_add_page_to_cache */
 335
 336
 337 /* select between SYNC and ASYNC I/O methods */
 338 int obdfs_do_writepage(struct inode *inode, struct page *page, int sync)
 339 {
 340         int err;
 341
 342         ENTRY;
 343         /* PDEBUG(page, "WRITEPAGE"); */
 344         if ( sync )
 345                 err = obdfs_brw(WRITE, inode, page, 1);
 346         else {
 347                 err = obdfs_add_page_to_cache(inode, page);
 348                 CDEBUG(D_INFO, "DO_WR ino: %ld, page %p, err %d, uptodate %d\n",
 349                        inode->i_ino, page, err, Page_Uptodate(page));
 350         }
 351
 352         if ( !err )
 353                 SetPageUptodate(page);
 354         /* PDEBUG(page,"WRITEPAGE"); */
 355         EXIT;
 356         return err;
 357 } /* obdfs_do_writepage */
 358
 359 /* returns the page unlocked, but with a reference */
 360 int obdfs_writepage(struct dentry *dentry, struct page *page)
 361 {
 362         return obdfs_do_writepage(dentry->d_inode, page, 0);
 363 }
 364
 365 /*
 366  * This does the "real" work of the write. The generic routine has
 367  * allocated the page, locked it, done all the page alignment stuff
 368  * calculations etc. Now we should just copy the data from user
 369  * space and write it back to the real medium..
 370  *
 371  * If the writer ends up delaying the write, the writer needs to
 372  * increment the page use counts until he is done with the page.
 373  *
 374  * Return value is the number of bytes written.
 375  */
 376 int obdfs_write_one_page(struct file *file, struct page *page,
 377                          unsigned long offset, unsigned long bytes,
 378                          const char * buf)
 379 {
 380         struct inode *inode = file->f_dentry->d_inode;
 381         int err;
 382
 383         ENTRY;
 384         /* We check for complete page writes here, as we then don't have to
 385          * get the page before writing over everything anyways.
 386          */
 387         if ( !Page_Uptodate(page) && (offset != 0 || bytes != PAGE_SIZE) ) {
 388                 err = obdfs_brw(READ, inode, page, 0);
 389                 if ( err )
 390                         return err;
 391                 SetPageUptodate(page);
 392         }
 393
 394         if (copy_from_user((u8*)page_address(page) + offset, buf, bytes))
 395                 return -EFAULT;
 396
 397         lock_kernel();
 398         err = obdfs_writepage(file->f_dentry, page);
 399         unlock_kernel();
 400
 401         return (err < 0 ? err : bytes);
 402 } /* obdfs_write_one_page */
 403
 404 /*
 405  * return an up to date page:
 406  *  - if locked is true then is returned locked
 407  *  - if create is true the corresponding disk blocks are created
 408  *  - page is held, i.e. caller must release the page
 409  *
 410  * modeled on NFS code.
 411  */
 412 struct page *obdfs_getpage(struct inode *inode, unsigned long offset,
 413                            int create, int locked)
 414 {
 415         struct page * page;
 416         int index;
 417         int err;
 418
 419         ENTRY;
 420
 421         offset = offset & PAGE_CACHE_MASK;
 422         CDEBUG(D_INFO, "ino: %ld, offset %ld, create %d, locked %d\n",
 423                inode->i_ino, offset, create, locked);
 424         index = offset >> PAGE_CACHE_SHIFT;
 425
 426         page = grab_cache_page(&inode->i_data, index);
 427
 428         /* Yuck, no page */
 429         if (! page) {
 430             printk(KERN_WARNING " grab_cache_page says no dice ...\n");
 431             EXIT;
 432             return NULL;
 433         }
 434
 435         /* PDEBUG(page, "GETPAGE: got page - before reading\n"); */
 436         /* now check if the data in the page is up to date */
 437         if ( Page_Uptodate(page)) {
 438                 if (!locked)
 439                         UnlockPage(page);
 440                 EXIT;
 441                 return page;
 442         }
 443
 444
 445 #ifdef EXT2_OBD_DEBUG
 446         if ((obd_debug_level & D_INFO) && obdfs_find_page_index(inode, index)) {
 447                 CDEBUG(D_INFO, "OVERWRITE: found dirty page %p, index %ld\n",
 448                        page, page->index);
 449         }
 450 #endif
 451
 452         err = obdfs_brw(READ, inode, page, create);
 453
 454         if ( err ) {
 455                 SetPageError(page);
 456                 UnlockPage(page);
 457                 EXIT;
 458                 return page;
 459         }
 460
 461         if ( !locked )
 462                 UnlockPage(page);
 463         SetPageUptodate(page);
 464         /* PDEBUG(page,"GETPAGE - after reading"); */
 465         EXIT;
 466         return page;
 467 } /* obdfs_getpage */
 468
 469
 470 void obdfs_truncate(struct inode *inode)
 471 {
 472         struct obdo *oa;
 473         int err;
 474         ENTRY;
 475
 476         obdfs_dequeue_pages(inode);
 477
 478         if (IOPS(inode, punch) == NULL) {
 479                 printk(KERN_ERR __FUNCTION__ ": no punch method!\n");
 480                 EXIT;
 481                 return;
 482         }
 483         oa = obdo_alloc();
 484         if ( !oa ) {
 485                 /* XXX This would give an inconsistent FS, so deal with it as
 486                  * best we can for now - an obdo on the stack is not pretty.
 487                  */
 488                 struct obdo obdo;
 489
 490                 printk(__FUNCTION__ ": obdo_alloc failed - using stack!\n");
 491
 492                 obdo.o_valid = OBD_MD_FLNOTOBD;
 493                 obdfs_from_inode(&obdo, inode);
 494
 495                 err = IOPS(inode, punch)(IID(inode), &obdo, obdo.o_size, 0);
 496         } else {
 497                 oa->o_valid = OBD_MD_FLNOTOBD;
 498                 obdfs_from_inode(oa, inode);
 499
 500                 CDEBUG(D_PUNCH, "calling punch for %ld (%Lu bytes at 0)\n",
 501                        (long)oa->o_id, oa->o_size);
 502                 err = IOPS(inode, punch)(IID(inode), oa, oa->o_size, 0);
 503
 504                 obdo_free(oa);
 505         }
 506
 507         if (err) {
 508                 printk(__FUNCTION__ ": obd_truncate fails (%d)\n", err);
 509                 EXIT;
 510                 return;
 511         }
 512         EXIT;
 513 }