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         oa = obdo_fromid(IID(inode), inode->i_ino, OBD_MD_FLNOTOBD);
  51         if ( IS_ERR(oa) ) {
  52                 EXIT;
  53                 return PTR_ERR(oa);
  54         }
  55         obdfs_from_inode(oa, inode);
  56
  57         err = IOPS(inode, brw)(rw, IID(inode), num_obdo, &oa, &bufs_per_obdo,
  58                                &buf, &count, &offset, &flags);
  59
  60         if ( !err )
  61                 obdfs_to_inode(inode, oa); /* copy o_blocks to i_blocks */
  62
  63         obdo_free(oa);
  64
  65         EXIT;
  66         return err;
  67 } /* obdfs_brw */
  68
  69 /* returns the page unlocked, but with a reference */
  70 int obdfs_readpage(struct dentry *dentry, struct page *page)
  71 {
  72         struct inode *inode = dentry->d_inode;
  73         int rc;
  74
  75         ENTRY;
  76         /* PDEBUG(page, "READ"); */
  77         rc = obdfs_brw(READ, inode, page, 0);
  78         if ( !rc ) {
  79                 SetPageUptodate(page);
  80                 UnlockPage(page);
  81         }
  82         /* PDEBUG(page, "READ"); */
  83         EXIT;
  84         return rc;
  85 } /* obdfs_readpage */
  86
  87 static kmem_cache_t *obdfs_pgrq_cachep = NULL;
  88
  89 int obdfs_init_pgrqcache(void)
  90 {
  91         ENTRY;
  92         if (obdfs_pgrq_cachep == NULL) {
  93                 CDEBUG(D_CACHE, "allocating obdfs_pgrq_cache\n");
  94                 obdfs_pgrq_cachep = kmem_cache_create("obdfs_pgrq",
  95                                                       sizeof(struct obdfs_pgrq),
  96                                                       0, SLAB_HWCACHE_ALIGN,
  97                                                       NULL, NULL);
  98                 if (obdfs_pgrq_cachep == NULL) {
  99                         EXIT;
 100                         return -ENOMEM;
 101                 } else {
 102                         CDEBUG(D_CACHE, "allocated cache at %p\n",
 103                                obdfs_pgrq_cachep);
 104                 }
 105         } else {
 106                 CDEBUG(D_CACHE, "using existing cache at %p\n",
 107                        obdfs_pgrq_cachep);
 108         }
 109         EXIT;
 110         return 0;
 111 } /* obdfs_init_wreqcache */
 112
 113 inline void obdfs_pgrq_del(struct obdfs_pgrq *pgrq)
 114 {
 115         --obdfs_cache_count;
 116         CDEBUG(D_INFO, "deleting page %p from list [count %ld]\n",
 117                pgrq->rq_page, obdfs_cache_count);
 118         list_del(&pgrq->rq_plist);
 119         kmem_cache_free(obdfs_pgrq_cachep, pgrq);
 120 }
 121
 122 void obdfs_cleanup_pgrqcache(void)
 123 {
 124         ENTRY;
 125         if (obdfs_pgrq_cachep != NULL) {
 126                 CDEBUG(D_CACHE, "destroying obdfs_pgrqcache at %p, count %ld\n",
 127                        obdfs_pgrq_cachep, obdfs_cache_count);
 128                 if (kmem_cache_destroy(obdfs_pgrq_cachep))
 129                         printk(KERN_INFO __FUNCTION__
 130                                ": unable to free all of cache\n");
 131                 obdfs_pgrq_cachep = NULL;
 132         } else
 133                 printk(KERN_INFO __FUNCTION__ ": called with NULL pointer\n");
 134
 135         EXIT;
 136 } /* obdfs_cleanup_wreqcache */
 137
 138
 139 /*
 140  * Find a specific page in the page cache.  If it is found, we return
 141  * the write request struct associated with it, if not found return NULL.
 142  * Called with the list lock held.
 143  */
 144 static struct obdfs_pgrq *
 145 obdfs_find_in_page_list(struct inode *inode, struct page *page)
 146 {
 147         struct list_head *page_list = obdfs_iplist(inode);
 148         struct list_head *tmp;
 149
 150         ENTRY;
 151
 152         CDEBUG(D_INFO, "looking for inode %ld page %p\n", inode->i_ino, page);
 153         OIDEBUG(inode);
 154
 155         if (list_empty(page_list)) {
 156                 CDEBUG(D_INFO, "empty list\n");
 157                 EXIT;
 158                 return NULL;
 159         }
 160         tmp = page_list;
 161         while ( (tmp = tmp->next) != page_list ) {
 162                 struct obdfs_pgrq *pgrq;
 163
 164                 pgrq = list_entry(tmp, struct obdfs_pgrq, rq_plist);
 165                 if (pgrq->rq_page == page) {
 166                         CDEBUG(D_INFO, "found page %p in list\n", page);
 167                         EXIT;
 168                         return pgrq;
 169                 }
 170         }
 171
 172         EXIT;
 173         return NULL;
 174 } /* obdfs_find_in_page_list */
 175
 176
 177 /* called with the list lock held */
 178 static struct page *obdfs_find_page_index(struct inode *inode,
 179                                           unsigned long index)
 180 {
 181         struct list_head *page_list = obdfs_iplist(inode);
 182         struct list_head *tmp;
 183         struct page *page;
 184
 185         ENTRY;
 186
 187         CDEBUG(D_INFO, "looking for inode %ld pageindex %ld\n",
 188                inode->i_ino, index);
 189         OIDEBUG(inode);
 190
 191         if (list_empty(page_list)) {
 192                 EXIT;
 193                 return NULL;
 194         }
 195         tmp = page_list;
 196         while ( (tmp = tmp->next) != page_list ) {
 197                 struct obdfs_pgrq *pgrq;
 198
 199                 pgrq = list_entry(tmp, struct obdfs_pgrq, rq_plist);
 200                 page = pgrq->rq_page;
 201                 if (index == page->index) {
 202                         CDEBUG(D_INFO,
 203                                "INDEX SEARCH found page %p, index %ld\n",
 204                                page, index);
 205                         EXIT;
 206                         return page;
 207                 }
 208         }
 209
 210         EXIT;
 211         return NULL;
 212 } /* obdfs_find_page_index */
 213
 214
 215 /* call and free pages from Linux page cache: called with io lock on inodes */
 216 int obdfs_do_vec_wr(struct inode **inodes, obd_count num_io,
 217                     obd_count num_obdos, struct obdo **obdos,
 218                     obd_count *oa_bufs, struct page **pages, char **bufs,
 219                     obd_size *counts, obd_off *offsets, obd_flag *flags)
 220 {
 221         struct super_block *sb = inodes[0]->i_sb;
 222         struct obdfs_sb_info *sbi = (struct obdfs_sb_info *)&sb->u.generic_sbp;
 223         int err;
 224
 225         ENTRY;
 226         CDEBUG(D_INFO, "writing %d page(s), %d obdo(s) in vector\n",
 227                num_io, num_obdos);
 228         { /* DEBUGGING */
 229                 int i;
 230                 printk("OBDOS: ");
 231                 for (i = 0; i < num_obdos; i++)
 232                         printk("%ld:0x%p ", (long)obdos[i]->o_id, obdos[i]);
 233
 234                 printk("\nPAGES: ");
 235                 for (i = 0; i < num_io; i++)
 236                         printk("0x%p ", pages[i]);
 237                 printk("\n");
 238         }
 239
 240         err = OPS(sb, brw)(WRITE, &sbi->osi_conn, num_obdos, obdos, oa_bufs,
 241                                 bufs, counts, offsets, flags);
 242
 243         CDEBUG(D_CACHE, "BRW done\n");
 244         /* release the pages from the page cache */
 245         while ( num_io > 0 ) {
 246                 --num_io;
 247                 CDEBUG(D_INFO, "calling put_page for %p, index %ld\n",
 248                        pages[num_io], pages[num_io]->index);
 249                 /* PDEBUG(pages[num_io], "do_vec_wr"); */
 250                 put_page(pages[num_io]);
 251                 /* PDEBUG(pages[num_io], "do_vec_wr"); */
 252         }
 253         CDEBUG(D_CACHE, "put_page done\n");
 254
 255         while ( num_obdos > 0) {
 256                 --num_obdos;
 257                 CDEBUG(D_INFO, "free obdo %ld\n",(long)obdos[num_obdos]->o_id);
 258                 obdfs_to_inode(inodes[num_obdos], obdos[num_obdos]);
 259                 obdo_free(obdos[num_obdos]);
 260         }
 261         CDEBUG(D_CACHE, "obdo_free done\n");
 262         EXIT;
 263         return err;
 264 }
 265
 266
 267 /*
 268  * Add a page to the write request cache list for later writing
 269  * ASYNCHRONOUS write method.
 270  */
 271 static int obdfs_add_page_to_cache(struct inode *inode, struct page *page)
 272 {
 273         int res = 0;
 274
 275         ENTRY;
 276
 277         /* If this page isn't already in the inode page list, add it */
 278         obd_down(&obdfs_i2sbi(inode)->osi_list_mutex);
 279         if ( !obdfs_find_in_page_list(inode, page) ) {
 280                 struct obdfs_pgrq *pgrq;
 281                 pgrq = kmem_cache_alloc(obdfs_pgrq_cachep, SLAB_KERNEL);
 282                 if (!pgrq) {
 283                         EXIT;
 284                         obd_up(&obdfs_i2sbi(inode)->osi_list_mutex);
 285                         return -ENOMEM;
 286                 }
 287                 memset(pgrq, 0, sizeof(*pgrq));
 288
 289                 pgrq->rq_page = page;
 290                 pgrq->rq_jiffies = jiffies;
 291                 get_page(pgrq->rq_page);
 292                 list_add(&pgrq->rq_plist, obdfs_iplist(inode));
 293                 obdfs_cache_count++;
 294                 CDEBUG(D_INFO,
 295                        "added inode %ld page %p, pgrq: %p, cache count [%ld]\n",
 296                        inode->i_ino, page, pgrq, obdfs_cache_count);
 297         }
 298
 299         /* If inode isn't already on the superblock inodes list, add it,
 300          * and increase ref count on inode so it doesn't disappear on us.
 301          *
 302          * We increment the reference count on the inode to keep it from
 303          * being freed from memory.  This _should_ be an iget() with an
 304          * iput() in both flush_reqs() and put_inode(), but since put_inode()
 305          * is called from iput() we can't call iput() again there.  Instead
 306          * we just increment/decrement i_count, which is essentially what
 307          * iget/iput do for an inode already in memory.
 308          */
 309         if ( list_empty(obdfs_islist(inode)) ) {
 310                 inode->i_count++;
 311                 CDEBUG(D_INFO, "adding inode %ld to superblock list %p\n",
 312                        inode->i_ino, obdfs_slist(inode));
 313                 list_add(obdfs_islist(inode), obdfs_slist(inode));
 314         }
 315         obd_up(&obdfs_i2sbi(inode)->osi_list_mutex);
 316
 317         /* XXX For testing purposes, we write out the page here.
 318          *     In the future, a flush daemon will write out the page.
 319         res = obdfs_flush_reqs(obdfs_slist(inode), ~0UL);
 320          */
 321
 322         EXIT;
 323         return res;
 324 } /* obdfs_add_page_to_cache */
 325
 326
 327 /* select between SYNC and ASYNC I/O methods */
 328 int obdfs_do_writepage(struct inode *inode, struct page *page, int sync)
 329 {
 330         int err;
 331
 332         ENTRY;
 333         /* PDEBUG(page, "WRITEPAGE"); */
 334         if ( sync )
 335                 err = obdfs_brw(WRITE, inode, page, 1);
 336         else {
 337                 err = obdfs_add_page_to_cache(inode, page);
 338                 CDEBUG(D_INFO, "DO_WR ino: %ld, page %p, err %d, uptodate %d\n",
 339                        inode->i_ino, page, err, Page_Uptodate(page));
 340         }
 341
 342         if ( !err )
 343                 SetPageUptodate(page);
 344         /* PDEBUG(page,"WRITEPAGE"); */
 345         EXIT;
 346         return err;
 347 } /* obdfs_do_writepage */
 348
 349 /* returns the page unlocked, but with a reference */
 350 int obdfs_writepage(struct dentry *dentry, struct page *page)
 351 {
 352         return obdfs_do_writepage(dentry->d_inode, page, 0);
 353 }
 354
 355 /*
 356  * This does the "real" work of the write. The generic routine has
 357  * allocated the page, locked it, done all the page alignment stuff
 358  * calculations etc. Now we should just copy the data from user
 359  * space and write it back to the real medium..
 360  *
 361  * If the writer ends up delaying the write, the writer needs to
 362  * increment the page use counts until he is done with the page.
 363  *
 364  * Return value is the number of bytes written.
 365  */
 366 int obdfs_write_one_page(struct file *file, struct page *page,
 367                          unsigned long offset, unsigned long bytes,
 368                          const char * buf)
 369 {
 370         struct inode *inode = file->f_dentry->d_inode;
 371         int err;
 372
 373         ENTRY;
 374         /* We check for complete page writes here, as we then don't have to
 375          * get the page before writing over everything anyways.
 376          */
 377         if ( !Page_Uptodate(page) && (offset != 0 || bytes != PAGE_SIZE) ) {
 378                 err = obdfs_brw(READ, inode, page, 0);
 379                 if ( err )
 380                         return err;
 381                 SetPageUptodate(page);
 382         }
 383
 384         if (copy_from_user((u8*)page_address(page) + offset, buf, bytes))
 385                 return -EFAULT;
 386
 387         lock_kernel(); /* XXX do we really need to lock the kernel to write? */
 388         err = obdfs_writepage(file->f_dentry, page);
 389         unlock_kernel();
 390
 391         return (err < 0 ? err : bytes);
 392 } /* obdfs_write_one_page */
 393
 394 /*
 395  * return an up to date page:
 396  *  - if locked is true then is returned locked
 397  *  - if create is true the corresponding disk blocks are created
 398  *  - page is held, i.e. caller must release the page
 399  *
 400  * modeled on NFS code.
 401  */
 402 struct page *obdfs_getpage(struct inode *inode, unsigned long offset,
 403                            int create, int locked)
 404 {
 405         struct page * page;
 406         int index;
 407         int err;
 408
 409         ENTRY;
 410
 411         offset = offset & PAGE_CACHE_MASK;
 412         CDEBUG(D_INFO, "ino: %ld, offset %ld, create %d, locked %d\n",
 413                inode->i_ino, offset, create, locked);
 414         index = offset >> PAGE_CACHE_SHIFT;
 415
 416         page = grab_cache_page(&inode->i_data, index);
 417
 418         /* Yuck, no page */
 419         if (! page) {
 420             printk(KERN_WARNING " grab_cache_page says no dice ...\n");
 421             EXIT;
 422             return NULL;
 423         }
 424
 425         /* PDEBUG(page, "GETPAGE: got page - before reading\n"); */
 426         /* now check if the data in the page is up to date */
 427         if ( Page_Uptodate(page)) {
 428                 if (!locked)
 429                         UnlockPage(page);
 430                 EXIT;
 431                 return page;
 432         }
 433
 434
 435 #ifdef EXT2_OBD_DEBUG
 436         if ((obd_debug_level & D_INFO) && obdfs_find_page_index(inode, index)) {
 437                 CDEBUG(D_INFO, "OVERWRITE: found dirty page %p, index %ld\n",
 438                        page, page->index);
 439         }
 440 #endif
 441
 442         err = obdfs_brw(READ, inode, page, create);
 443
 444         if ( err ) {
 445                 SetPageError(page);
 446                 UnlockPage(page);
 447                 EXIT;
 448                 return page;
 449         }
 450
 451         if ( !locked )
 452                 UnlockPage(page);
 453         SetPageUptodate(page);
 454         /* PDEBUG(page,"GETPAGE - after reading"); */
 455         EXIT;
 456         return page;
 457 } /* obdfs_getpage */
 458
 459