Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Copyright 2000 by Hans Reiser, licensing governed by reiserfs/README |
| 3 | */ |
| 4 | |
| 5 | |
| 6 | #include <linux/time.h> |
| 7 | #include <linux/reiserfs_fs.h> |
| 8 | #include <linux/reiserfs_acl.h> |
| 9 | #include <linux/reiserfs_xattr.h> |
| 10 | #include <linux/smp_lock.h> |
| 11 | #include <asm/uaccess.h> |
| 12 | #include <linux/pagemap.h> |
| 13 | #include <linux/swap.h> |
| 14 | #include <linux/writeback.h> |
| 15 | #include <linux/blkdev.h> |
| 16 | #include <linux/buffer_head.h> |
| 17 | #include <linux/quotaops.h> |
| 18 | |
| 19 | /* |
| 20 | ** We pack the tails of files on file close, not at the time they are written. |
| 21 | ** This implies an unnecessary copy of the tail and an unnecessary indirect item |
| 22 | ** insertion/balancing, for files that are written in one write. |
| 23 | ** It avoids unnecessary tail packings (balances) for files that are written in |
| 24 | ** multiple writes and are small enough to have tails. |
| 25 | ** |
| 26 | ** file_release is called by the VFS layer when the file is closed. If |
| 27 | ** this is the last open file descriptor, and the file |
| 28 | ** small enough to have a tail, and the tail is currently in an |
| 29 | ** unformatted node, the tail is converted back into a direct item. |
| 30 | ** |
| 31 | ** We use reiserfs_truncate_file to pack the tail, since it already has |
| 32 | ** all the conditions coded. |
| 33 | */ |
| 34 | static int reiserfs_file_release (struct inode * inode, struct file * filp) |
| 35 | { |
| 36 | |
| 37 | struct reiserfs_transaction_handle th ; |
| 38 | int err; |
| 39 | int jbegin_failure = 0; |
| 40 | |
| 41 | if (!S_ISREG (inode->i_mode)) |
| 42 | BUG (); |
| 43 | |
| 44 | /* fast out for when nothing needs to be done */ |
| 45 | if ((atomic_read(&inode->i_count) > 1 || |
| 46 | !(REISERFS_I(inode)->i_flags & i_pack_on_close_mask) || |
| 47 | !tail_has_to_be_packed(inode)) && |
| 48 | REISERFS_I(inode)->i_prealloc_count <= 0) { |
| 49 | return 0; |
| 50 | } |
| 51 | |
| 52 | reiserfs_write_lock(inode->i_sb); |
| 53 | down (&inode->i_sem); |
| 54 | /* freeing preallocation only involves relogging blocks that |
| 55 | * are already in the current transaction. preallocation gets |
| 56 | * freed at the end of each transaction, so it is impossible for |
| 57 | * us to log any additional blocks (including quota blocks) |
| 58 | */ |
| 59 | err = journal_begin(&th, inode->i_sb, 1); |
| 60 | if (err) { |
| 61 | /* uh oh, we can't allow the inode to go away while there |
| 62 | * is still preallocation blocks pending. Try to join the |
| 63 | * aborted transaction |
| 64 | */ |
| 65 | jbegin_failure = err; |
| 66 | err = journal_join_abort(&th, inode->i_sb, 1); |
| 67 | |
| 68 | if (err) { |
| 69 | /* hmpf, our choices here aren't good. We can pin the inode |
| 70 | * which will disallow unmount from every happening, we can |
| 71 | * do nothing, which will corrupt random memory on unmount, |
| 72 | * or we can forcibly remove the file from the preallocation |
| 73 | * list, which will leak blocks on disk. Lets pin the inode |
| 74 | * and let the admin know what is going on. |
| 75 | */ |
| 76 | igrab(inode); |
| 77 | reiserfs_warning(inode->i_sb, "pinning inode %lu because the " |
| 78 | "preallocation can't be freed"); |
| 79 | goto out; |
| 80 | } |
| 81 | } |
| 82 | reiserfs_update_inode_transaction(inode) ; |
| 83 | |
| 84 | #ifdef REISERFS_PREALLOCATE |
| 85 | reiserfs_discard_prealloc (&th, inode); |
| 86 | #endif |
| 87 | err = journal_end(&th, inode->i_sb, 1); |
| 88 | |
| 89 | /* copy back the error code from journal_begin */ |
| 90 | if (!err) |
| 91 | err = jbegin_failure; |
| 92 | |
| 93 | if (!err && atomic_read(&inode->i_count) <= 1 && |
| 94 | (REISERFS_I(inode)->i_flags & i_pack_on_close_mask) && |
| 95 | tail_has_to_be_packed (inode)) { |
| 96 | /* if regular file is released by last holder and it has been |
| 97 | appended (we append by unformatted node only) or its direct |
| 98 | item(s) had to be converted, then it may have to be |
| 99 | indirect2direct converted */ |
| 100 | err = reiserfs_truncate_file(inode, 0) ; |
| 101 | } |
| 102 | out: |
| 103 | up (&inode->i_sem); |
| 104 | reiserfs_write_unlock(inode->i_sb); |
| 105 | return err; |
| 106 | } |
| 107 | |
| 108 | static void reiserfs_vfs_truncate_file(struct inode *inode) { |
| 109 | reiserfs_truncate_file(inode, 1) ; |
| 110 | } |
| 111 | |
| 112 | /* Sync a reiserfs file. */ |
| 113 | |
| 114 | /* |
| 115 | * FIXME: sync_mapping_buffers() never has anything to sync. Can |
| 116 | * be removed... |
| 117 | */ |
| 118 | |
| 119 | static int reiserfs_sync_file( |
| 120 | struct file * p_s_filp, |
| 121 | struct dentry * p_s_dentry, |
| 122 | int datasync |
| 123 | ) { |
| 124 | struct inode * p_s_inode = p_s_dentry->d_inode; |
| 125 | int n_err; |
| 126 | int barrier_done; |
| 127 | |
| 128 | if (!S_ISREG(p_s_inode->i_mode)) |
| 129 | BUG (); |
| 130 | n_err = sync_mapping_buffers(p_s_inode->i_mapping) ; |
| 131 | reiserfs_write_lock(p_s_inode->i_sb); |
| 132 | barrier_done = reiserfs_commit_for_inode(p_s_inode); |
| 133 | reiserfs_write_unlock(p_s_inode->i_sb); |
| 134 | if (barrier_done != 1) |
| 135 | blkdev_issue_flush(p_s_inode->i_sb->s_bdev, NULL); |
| 136 | if (barrier_done < 0) |
| 137 | return barrier_done; |
| 138 | return ( n_err < 0 ) ? -EIO : 0; |
| 139 | } |
| 140 | |
| 141 | /* I really do not want to play with memory shortage right now, so |
| 142 | to simplify the code, we are not going to write more than this much pages at |
| 143 | a time. This still should considerably improve performance compared to 4k |
| 144 | at a time case. This is 32 pages of 4k size. */ |
| 145 | #define REISERFS_WRITE_PAGES_AT_A_TIME (128 * 1024) / PAGE_CACHE_SIZE |
| 146 | |
| 147 | /* Allocates blocks for a file to fulfil write request. |
| 148 | Maps all unmapped but prepared pages from the list. |
| 149 | Updates metadata with newly allocated blocknumbers as needed */ |
| 150 | static int reiserfs_allocate_blocks_for_region( |
| 151 | struct reiserfs_transaction_handle *th, |
| 152 | struct inode *inode, /* Inode we work with */ |
| 153 | loff_t pos, /* Writing position */ |
| 154 | int num_pages, /* number of pages write going |
| 155 | to touch */ |
| 156 | int write_bytes, /* amount of bytes to write */ |
| 157 | struct page **prepared_pages, /* array of |
| 158 | prepared pages |
| 159 | */ |
| 160 | int blocks_to_allocate /* Amount of blocks we |
| 161 | need to allocate to |
| 162 | fit the data into file |
| 163 | */ |
| 164 | ) |
| 165 | { |
| 166 | struct cpu_key key; // cpu key of item that we are going to deal with |
| 167 | struct item_head *ih; // pointer to item head that we are going to deal with |
| 168 | struct buffer_head *bh; // Buffer head that contains items that we are going to deal with |
| 169 | __u32 * item; // pointer to item we are going to deal with |
| 170 | INITIALIZE_PATH(path); // path to item, that we are going to deal with. |
| 171 | b_blocknr_t *allocated_blocks; // Pointer to a place where allocated blocknumbers would be stored. |
| 172 | reiserfs_blocknr_hint_t hint; // hint structure for block allocator. |
| 173 | size_t res; // return value of various functions that we call. |
| 174 | int curr_block; // current block used to keep track of unmapped blocks. |
| 175 | int i; // loop counter |
| 176 | int itempos; // position in item |
| 177 | unsigned int from = (pos & (PAGE_CACHE_SIZE - 1)); // writing position in |
| 178 | // first page |
| 179 | unsigned int to = ((pos + write_bytes - 1) & (PAGE_CACHE_SIZE - 1)) + 1; /* last modified byte offset in last page */ |
| 180 | __u64 hole_size ; // amount of blocks for a file hole, if it needed to be created. |
| 181 | int modifying_this_item = 0; // Flag for items traversal code to keep track |
| 182 | // of the fact that we already prepared |
| 183 | // current block for journal |
| 184 | int will_prealloc = 0; |
| 185 | RFALSE(!blocks_to_allocate, "green-9004: tried to allocate zero blocks?"); |
| 186 | |
| 187 | /* only preallocate if this is a small write */ |
| 188 | if (REISERFS_I(inode)->i_prealloc_count || |
| 189 | (!(write_bytes & (inode->i_sb->s_blocksize -1)) && |
| 190 | blocks_to_allocate < |
| 191 | REISERFS_SB(inode->i_sb)->s_alloc_options.preallocsize)) |
| 192 | will_prealloc = REISERFS_SB(inode->i_sb)->s_alloc_options.preallocsize; |
| 193 | |
| 194 | allocated_blocks = kmalloc((blocks_to_allocate + will_prealloc) * |
| 195 | sizeof(b_blocknr_t), GFP_NOFS); |
| 196 | |
| 197 | /* First we compose a key to point at the writing position, we want to do |
| 198 | that outside of any locking region. */ |
| 199 | make_cpu_key (&key, inode, pos+1, TYPE_ANY, 3/*key length*/); |
| 200 | |
| 201 | /* If we came here, it means we absolutely need to open a transaction, |
| 202 | since we need to allocate some blocks */ |
| 203 | reiserfs_write_lock(inode->i_sb); // Journaling stuff and we need that. |
| 204 | res = journal_begin(th, inode->i_sb, JOURNAL_PER_BALANCE_CNT * 3 + 1 + 2 * REISERFS_QUOTA_TRANS_BLOCKS); // Wish I know if this number enough |
| 205 | if (res) |
| 206 | goto error_exit; |
| 207 | reiserfs_update_inode_transaction(inode) ; |
| 208 | |
| 209 | /* Look for the in-tree position of our write, need path for block allocator */ |
| 210 | res = search_for_position_by_key(inode->i_sb, &key, &path); |
| 211 | if ( res == IO_ERROR ) { |
| 212 | res = -EIO; |
| 213 | goto error_exit; |
| 214 | } |
| 215 | |
| 216 | /* Allocate blocks */ |
| 217 | /* First fill in "hint" structure for block allocator */ |
| 218 | hint.th = th; // transaction handle. |
| 219 | hint.path = &path; // Path, so that block allocator can determine packing locality or whatever it needs to determine. |
| 220 | hint.inode = inode; // Inode is needed by block allocator too. |
| 221 | hint.search_start = 0; // We have no hint on where to search free blocks for block allocator. |
| 222 | hint.key = key.on_disk_key; // on disk key of file. |
| 223 | hint.block = inode->i_blocks>>(inode->i_sb->s_blocksize_bits-9); // Number of disk blocks this file occupies already. |
| 224 | hint.formatted_node = 0; // We are allocating blocks for unformatted node. |
| 225 | hint.preallocate = will_prealloc; |
| 226 | |
| 227 | /* Call block allocator to allocate blocks */ |
| 228 | res = reiserfs_allocate_blocknrs(&hint, allocated_blocks, blocks_to_allocate, blocks_to_allocate); |
| 229 | if ( res != CARRY_ON ) { |
| 230 | if ( res == NO_DISK_SPACE ) { |
| 231 | /* We flush the transaction in case of no space. This way some |
| 232 | blocks might become free */ |
| 233 | SB_JOURNAL(inode->i_sb)->j_must_wait = 1; |
| 234 | res = restart_transaction(th, inode, &path); |
| 235 | if (res) |
| 236 | goto error_exit; |
| 237 | |
| 238 | /* We might have scheduled, so search again */ |
| 239 | res = search_for_position_by_key(inode->i_sb, &key, &path); |
| 240 | if ( res == IO_ERROR ) { |
| 241 | res = -EIO; |
| 242 | goto error_exit; |
| 243 | } |
| 244 | |
| 245 | /* update changed info for hint structure. */ |
| 246 | res = reiserfs_allocate_blocknrs(&hint, allocated_blocks, blocks_to_allocate, blocks_to_allocate); |
| 247 | if ( res != CARRY_ON ) { |
| 248 | res = -ENOSPC; |
| 249 | pathrelse(&path); |
| 250 | goto error_exit; |
| 251 | } |
| 252 | } else { |
| 253 | res = -ENOSPC; |
| 254 | pathrelse(&path); |
| 255 | goto error_exit; |
| 256 | } |
| 257 | } |
| 258 | |
| 259 | #ifdef __BIG_ENDIAN |
| 260 | // Too bad, I have not found any way to convert a given region from |
| 261 | // cpu format to little endian format |
| 262 | { |
| 263 | int i; |
| 264 | for ( i = 0; i < blocks_to_allocate ; i++) |
| 265 | allocated_blocks[i]=cpu_to_le32(allocated_blocks[i]); |
| 266 | } |
| 267 | #endif |
| 268 | |
| 269 | /* Blocks allocating well might have scheduled and tree might have changed, |
| 270 | let's search the tree again */ |
| 271 | /* find where in the tree our write should go */ |
| 272 | res = search_for_position_by_key(inode->i_sb, &key, &path); |
| 273 | if ( res == IO_ERROR ) { |
| 274 | res = -EIO; |
| 275 | goto error_exit_free_blocks; |
| 276 | } |
| 277 | |
| 278 | bh = get_last_bh( &path ); // Get a bufferhead for last element in path. |
| 279 | ih = get_ih( &path ); // Get a pointer to last item head in path. |
| 280 | item = get_item( &path ); // Get a pointer to last item in path |
| 281 | |
| 282 | /* Let's see what we have found */ |
| 283 | if ( res != POSITION_FOUND ) { /* position not found, this means that we |
| 284 | might need to append file with holes |
| 285 | first */ |
| 286 | // Since we are writing past the file's end, we need to find out if |
| 287 | // there is a hole that needs to be inserted before our writing |
| 288 | // position, and how many blocks it is going to cover (we need to |
| 289 | // populate pointers to file blocks representing the hole with zeros) |
| 290 | |
| 291 | { |
| 292 | int item_offset = 1; |
| 293 | /* |
| 294 | * if ih is stat data, its offset is 0 and we don't want to |
| 295 | * add 1 to pos in the hole_size calculation |
| 296 | */ |
| 297 | if (is_statdata_le_ih(ih)) |
| 298 | item_offset = 0; |
| 299 | hole_size = (pos + item_offset - |
| 300 | (le_key_k_offset( get_inode_item_key_version(inode), |
| 301 | &(ih->ih_key)) + |
| 302 | op_bytes_number(ih, inode->i_sb->s_blocksize))) >> |
| 303 | inode->i_sb->s_blocksize_bits; |
| 304 | } |
| 305 | |
| 306 | if ( hole_size > 0 ) { |
| 307 | int to_paste = min_t(__u64, hole_size, MAX_ITEM_LEN(inode->i_sb->s_blocksize)/UNFM_P_SIZE ); // How much data to insert first time. |
| 308 | /* area filled with zeroes, to supply as list of zero blocknumbers |
| 309 | We allocate it outside of loop just in case loop would spin for |
| 310 | several iterations. */ |
| 311 | char *zeros = kmalloc(to_paste*UNFM_P_SIZE, GFP_ATOMIC); // We cannot insert more than MAX_ITEM_LEN bytes anyway. |
| 312 | if ( !zeros ) { |
| 313 | res = -ENOMEM; |
| 314 | goto error_exit_free_blocks; |
| 315 | } |
| 316 | memset ( zeros, 0, to_paste*UNFM_P_SIZE); |
| 317 | do { |
| 318 | to_paste = min_t(__u64, hole_size, MAX_ITEM_LEN(inode->i_sb->s_blocksize)/UNFM_P_SIZE ); |
| 319 | if ( is_indirect_le_ih(ih) ) { |
| 320 | /* Ok, there is existing indirect item already. Need to append it */ |
| 321 | /* Calculate position past inserted item */ |
| 322 | make_cpu_key( &key, inode, le_key_k_offset( get_inode_item_key_version(inode), &(ih->ih_key)) + op_bytes_number(ih, inode->i_sb->s_blocksize), TYPE_INDIRECT, 3); |
| 323 | res = reiserfs_paste_into_item( th, &path, &key, inode, (char *)zeros, UNFM_P_SIZE*to_paste); |
| 324 | if ( res ) { |
| 325 | kfree(zeros); |
| 326 | goto error_exit_free_blocks; |
| 327 | } |
| 328 | } else if ( is_statdata_le_ih(ih) ) { |
| 329 | /* No existing item, create it */ |
| 330 | /* item head for new item */ |
| 331 | struct item_head ins_ih; |
| 332 | |
| 333 | /* create a key for our new item */ |
| 334 | make_cpu_key( &key, inode, 1, TYPE_INDIRECT, 3); |
| 335 | |
| 336 | /* Create new item head for our new item */ |
| 337 | make_le_item_head (&ins_ih, &key, key.version, 1, |
| 338 | TYPE_INDIRECT, to_paste*UNFM_P_SIZE, |
| 339 | 0 /* free space */); |
| 340 | |
| 341 | /* Find where such item should live in the tree */ |
| 342 | res = search_item (inode->i_sb, &key, &path); |
| 343 | if ( res != ITEM_NOT_FOUND ) { |
| 344 | /* item should not exist, otherwise we have error */ |
| 345 | if ( res != -ENOSPC ) { |
| 346 | reiserfs_warning (inode->i_sb, |
| 347 | "green-9008: search_by_key (%K) returned %d", |
| 348 | &key, res); |
| 349 | } |
| 350 | res = -EIO; |
| 351 | kfree(zeros); |
| 352 | goto error_exit_free_blocks; |
| 353 | } |
| 354 | res = reiserfs_insert_item( th, &path, &key, &ins_ih, inode, (char *)zeros); |
| 355 | } else { |
| 356 | reiserfs_panic(inode->i_sb, "green-9011: Unexpected key type %K\n", &key); |
| 357 | } |
| 358 | if ( res ) { |
| 359 | kfree(zeros); |
| 360 | goto error_exit_free_blocks; |
| 361 | } |
| 362 | /* Now we want to check if transaction is too full, and if it is |
| 363 | we restart it. This will also free the path. */ |
| 364 | if (journal_transaction_should_end(th, th->t_blocks_allocated)) { |
| 365 | res = restart_transaction(th, inode, &path); |
| 366 | if (res) { |
| 367 | pathrelse (&path); |
| 368 | kfree(zeros); |
| 369 | goto error_exit; |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | /* Well, need to recalculate path and stuff */ |
| 374 | set_cpu_key_k_offset( &key, cpu_key_k_offset(&key) + (to_paste << inode->i_blkbits)); |
| 375 | res = search_for_position_by_key(inode->i_sb, &key, &path); |
| 376 | if ( res == IO_ERROR ) { |
| 377 | res = -EIO; |
| 378 | kfree(zeros); |
| 379 | goto error_exit_free_blocks; |
| 380 | } |
| 381 | bh=get_last_bh(&path); |
| 382 | ih=get_ih(&path); |
| 383 | item = get_item(&path); |
| 384 | hole_size -= to_paste; |
| 385 | } while ( hole_size ); |
| 386 | kfree(zeros); |
| 387 | } |
| 388 | } |
| 389 | |
| 390 | // Go through existing indirect items first |
| 391 | // replace all zeroes with blocknumbers from list |
| 392 | // Note that if no corresponding item was found, by previous search, |
| 393 | // it means there are no existing in-tree representation for file area |
| 394 | // we are going to overwrite, so there is nothing to scan through for holes. |
| 395 | for ( curr_block = 0, itempos = path.pos_in_item ; curr_block < blocks_to_allocate && res == POSITION_FOUND ; ) { |
| 396 | retry: |
| 397 | |
| 398 | if ( itempos >= ih_item_len(ih)/UNFM_P_SIZE ) { |
| 399 | /* We run out of data in this indirect item, let's look for another |
| 400 | one. */ |
| 401 | /* First if we are already modifying current item, log it */ |
| 402 | if ( modifying_this_item ) { |
| 403 | journal_mark_dirty (th, inode->i_sb, bh); |
| 404 | modifying_this_item = 0; |
| 405 | } |
| 406 | /* Then set the key to look for a new indirect item (offset of old |
| 407 | item is added to old item length */ |
| 408 | set_cpu_key_k_offset( &key, le_key_k_offset( get_inode_item_key_version(inode), &(ih->ih_key)) + op_bytes_number(ih, inode->i_sb->s_blocksize)); |
| 409 | /* Search ofor position of new key in the tree. */ |
| 410 | res = search_for_position_by_key(inode->i_sb, &key, &path); |
| 411 | if ( res == IO_ERROR) { |
| 412 | res = -EIO; |
| 413 | goto error_exit_free_blocks; |
| 414 | } |
| 415 | bh=get_last_bh(&path); |
| 416 | ih=get_ih(&path); |
| 417 | item = get_item(&path); |
| 418 | itempos = path.pos_in_item; |
| 419 | continue; // loop to check all kinds of conditions and so on. |
| 420 | } |
| 421 | /* Ok, we have correct position in item now, so let's see if it is |
| 422 | representing file hole (blocknumber is zero) and fill it if needed */ |
| 423 | if ( !item[itempos] ) { |
| 424 | /* Ok, a hole. Now we need to check if we already prepared this |
| 425 | block to be journaled */ |
| 426 | while ( !modifying_this_item ) { // loop until succeed |
| 427 | /* Well, this item is not journaled yet, so we must prepare |
| 428 | it for journal first, before we can change it */ |
| 429 | struct item_head tmp_ih; // We copy item head of found item, |
| 430 | // here to detect if fs changed under |
| 431 | // us while we were preparing for |
| 432 | // journal. |
| 433 | int fs_gen; // We store fs generation here to find if someone |
| 434 | // changes fs under our feet |
| 435 | |
| 436 | copy_item_head (&tmp_ih, ih); // Remember itemhead |
| 437 | fs_gen = get_generation (inode->i_sb); // remember fs generation |
| 438 | reiserfs_prepare_for_journal(inode->i_sb, bh, 1); // Prepare a buffer within which indirect item is stored for changing. |
| 439 | if (fs_changed (fs_gen, inode->i_sb) && item_moved (&tmp_ih, &path)) { |
| 440 | // Sigh, fs was changed under us, we need to look for new |
| 441 | // location of item we are working with |
| 442 | |
| 443 | /* unmark prepaerd area as journaled and search for it's |
| 444 | new position */ |
| 445 | reiserfs_restore_prepared_buffer(inode->i_sb, bh); |
| 446 | res = search_for_position_by_key(inode->i_sb, &key, &path); |
| 447 | if ( res == IO_ERROR) { |
| 448 | res = -EIO; |
| 449 | goto error_exit_free_blocks; |
| 450 | } |
| 451 | bh=get_last_bh(&path); |
| 452 | ih=get_ih(&path); |
| 453 | item = get_item(&path); |
| 454 | itempos = path.pos_in_item; |
| 455 | goto retry; |
| 456 | } |
| 457 | modifying_this_item = 1; |
| 458 | } |
| 459 | item[itempos] = allocated_blocks[curr_block]; // Assign new block |
| 460 | curr_block++; |
| 461 | } |
| 462 | itempos++; |
| 463 | } |
| 464 | |
| 465 | if ( modifying_this_item ) { // We need to log last-accessed block, if it |
| 466 | // was modified, but not logged yet. |
| 467 | journal_mark_dirty (th, inode->i_sb, bh); |
| 468 | } |
| 469 | |
| 470 | if ( curr_block < blocks_to_allocate ) { |
| 471 | // Oh, well need to append to indirect item, or to create indirect item |
| 472 | // if there weren't any |
| 473 | if ( is_indirect_le_ih(ih) ) { |
| 474 | // Existing indirect item - append. First calculate key for append |
| 475 | // position. We do not need to recalculate path as it should |
| 476 | // already point to correct place. |
| 477 | make_cpu_key( &key, inode, le_key_k_offset( get_inode_item_key_version(inode), &(ih->ih_key)) + op_bytes_number(ih, inode->i_sb->s_blocksize), TYPE_INDIRECT, 3); |
| 478 | res = reiserfs_paste_into_item( th, &path, &key, inode, (char *)(allocated_blocks+curr_block), UNFM_P_SIZE*(blocks_to_allocate-curr_block)); |
| 479 | if ( res ) { |
| 480 | goto error_exit_free_blocks; |
| 481 | } |
| 482 | } else if (is_statdata_le_ih(ih) ) { |
| 483 | // Last found item was statdata. That means we need to create indirect item. |
| 484 | struct item_head ins_ih; /* itemhead for new item */ |
| 485 | |
| 486 | /* create a key for our new item */ |
| 487 | make_cpu_key( &key, inode, 1, TYPE_INDIRECT, 3); // Position one, |
| 488 | // because that's |
| 489 | // where first |
| 490 | // indirect item |
| 491 | // begins |
| 492 | /* Create new item head for our new item */ |
| 493 | make_le_item_head (&ins_ih, &key, key.version, 1, TYPE_INDIRECT, |
| 494 | (blocks_to_allocate-curr_block)*UNFM_P_SIZE, |
| 495 | 0 /* free space */); |
| 496 | /* Find where such item should live in the tree */ |
| 497 | res = search_item (inode->i_sb, &key, &path); |
| 498 | if ( res != ITEM_NOT_FOUND ) { |
| 499 | /* Well, if we have found such item already, or some error |
| 500 | occured, we need to warn user and return error */ |
| 501 | if ( res != -ENOSPC ) { |
| 502 | reiserfs_warning (inode->i_sb, |
| 503 | "green-9009: search_by_key (%K) " |
| 504 | "returned %d", &key, res); |
| 505 | } |
| 506 | res = -EIO; |
| 507 | goto error_exit_free_blocks; |
| 508 | } |
| 509 | /* Insert item into the tree with the data as its body */ |
| 510 | res = reiserfs_insert_item( th, &path, &key, &ins_ih, inode, (char *)(allocated_blocks+curr_block)); |
| 511 | } else { |
| 512 | reiserfs_panic(inode->i_sb, "green-9010: unexpected item type for key %K\n",&key); |
| 513 | } |
| 514 | } |
| 515 | |
| 516 | // the caller is responsible for closing the transaction |
| 517 | // unless we return an error, they are also responsible for logging |
| 518 | // the inode. |
| 519 | // |
| 520 | pathrelse(&path); |
| 521 | /* |
| 522 | * cleanup prellocation from previous writes |
| 523 | * if this is a partial block write |
| 524 | */ |
| 525 | if (write_bytes & (inode->i_sb->s_blocksize -1)) |
| 526 | reiserfs_discard_prealloc(th, inode); |
| 527 | reiserfs_write_unlock(inode->i_sb); |
| 528 | |
| 529 | // go through all the pages/buffers and map the buffers to newly allocated |
| 530 | // blocks (so that system knows where to write these pages later). |
| 531 | curr_block = 0; |
| 532 | for ( i = 0; i < num_pages ; i++ ) { |
| 533 | struct page *page=prepared_pages[i]; //current page |
| 534 | struct buffer_head *head = page_buffers(page);// first buffer for a page |
| 535 | int block_start, block_end; // in-page offsets for buffers. |
| 536 | |
| 537 | if (!page_buffers(page)) |
| 538 | reiserfs_panic(inode->i_sb, "green-9005: No buffers for prepared page???"); |
| 539 | |
| 540 | /* For each buffer in page */ |
| 541 | for(bh = head, block_start = 0; bh != head || !block_start; |
| 542 | block_start=block_end, bh = bh->b_this_page) { |
| 543 | if (!bh) |
| 544 | reiserfs_panic(inode->i_sb, "green-9006: Allocated but absent buffer for a page?"); |
| 545 | block_end = block_start+inode->i_sb->s_blocksize; |
| 546 | if (i == 0 && block_end <= from ) |
| 547 | /* if this buffer is before requested data to map, skip it */ |
| 548 | continue; |
| 549 | if (i == num_pages - 1 && block_start >= to) |
| 550 | /* If this buffer is after requested data to map, abort |
| 551 | processing of current page */ |
| 552 | break; |
| 553 | |
| 554 | if ( !buffer_mapped(bh) ) { // Ok, unmapped buffer, need to map it |
| 555 | map_bh( bh, inode->i_sb, le32_to_cpu(allocated_blocks[curr_block])); |
| 556 | curr_block++; |
| 557 | set_buffer_new(bh); |
| 558 | } |
| 559 | } |
| 560 | } |
| 561 | |
| 562 | RFALSE( curr_block > blocks_to_allocate, "green-9007: Used too many blocks? weird"); |
| 563 | |
| 564 | kfree(allocated_blocks); |
| 565 | return 0; |
| 566 | |
| 567 | // Need to deal with transaction here. |
| 568 | error_exit_free_blocks: |
| 569 | pathrelse(&path); |
| 570 | // free blocks |
| 571 | for( i = 0; i < blocks_to_allocate; i++ ) |
| 572 | reiserfs_free_block(th, inode, le32_to_cpu(allocated_blocks[i]), 1); |
| 573 | |
| 574 | error_exit: |
| 575 | if (th->t_trans_id) { |
| 576 | int err; |
| 577 | // update any changes we made to blk count |
| 578 | reiserfs_update_sd(th, inode); |
| 579 | err = journal_end(th, inode->i_sb, JOURNAL_PER_BALANCE_CNT * 3 + 1 + 2 * REISERFS_QUOTA_TRANS_BLOCKS); |
| 580 | if (err) |
| 581 | res = err; |
| 582 | } |
| 583 | reiserfs_write_unlock(inode->i_sb); |
| 584 | kfree(allocated_blocks); |
| 585 | |
| 586 | return res; |
| 587 | } |
| 588 | |
| 589 | /* Unlock pages prepared by reiserfs_prepare_file_region_for_write */ |
| 590 | static void reiserfs_unprepare_pages(struct page **prepared_pages, /* list of locked pages */ |
| 591 | size_t num_pages /* amount of pages */) { |
| 592 | int i; // loop counter |
| 593 | |
| 594 | for (i=0; i < num_pages ; i++) { |
| 595 | struct page *page = prepared_pages[i]; |
| 596 | |
| 597 | try_to_free_buffers(page); |
| 598 | unlock_page(page); |
| 599 | page_cache_release(page); |
| 600 | } |
| 601 | } |
| 602 | |
| 603 | /* This function will copy data from userspace to specified pages within |
| 604 | supplied byte range */ |
| 605 | static int reiserfs_copy_from_user_to_file_region( |
| 606 | loff_t pos, /* In-file position */ |
| 607 | int num_pages, /* Number of pages affected */ |
| 608 | int write_bytes, /* Amount of bytes to write */ |
| 609 | struct page **prepared_pages, /* pointer to |
| 610 | array to |
| 611 | prepared pages |
| 612 | */ |
| 613 | const char __user *buf /* Pointer to user-supplied |
| 614 | data*/ |
| 615 | ) |
| 616 | { |
| 617 | long page_fault=0; // status of copy_from_user. |
| 618 | int i; // loop counter. |
| 619 | int offset; // offset in page |
| 620 | |
| 621 | for ( i = 0, offset = (pos & (PAGE_CACHE_SIZE-1)); i < num_pages ; i++,offset=0) { |
| 622 | size_t count = min_t(size_t,PAGE_CACHE_SIZE-offset,write_bytes); // How much of bytes to write to this page |
| 623 | struct page *page=prepared_pages[i]; // Current page we process. |
| 624 | |
| 625 | fault_in_pages_readable( buf, count); |
| 626 | |
| 627 | /* Copy data from userspace to the current page */ |
| 628 | kmap(page); |
| 629 | page_fault = __copy_from_user(page_address(page)+offset, buf, count); // Copy the data. |
| 630 | /* Flush processor's dcache for this page */ |
| 631 | flush_dcache_page(page); |
| 632 | kunmap(page); |
| 633 | buf+=count; |
| 634 | write_bytes-=count; |
| 635 | |
| 636 | if (page_fault) |
| 637 | break; // Was there a fault? abort. |
| 638 | } |
| 639 | |
| 640 | return page_fault?-EFAULT:0; |
| 641 | } |
| 642 | |
| 643 | /* taken fs/buffer.c:__block_commit_write */ |
| 644 | int reiserfs_commit_page(struct inode *inode, struct page *page, |
| 645 | unsigned from, unsigned to) |
| 646 | { |
| 647 | unsigned block_start, block_end; |
| 648 | int partial = 0; |
| 649 | unsigned blocksize; |
| 650 | struct buffer_head *bh, *head; |
| 651 | unsigned long i_size_index = inode->i_size >> PAGE_CACHE_SHIFT; |
| 652 | int new; |
| 653 | int logit = reiserfs_file_data_log(inode); |
| 654 | struct super_block *s = inode->i_sb; |
| 655 | int bh_per_page = PAGE_CACHE_SIZE / s->s_blocksize; |
| 656 | struct reiserfs_transaction_handle th; |
| 657 | int ret = 0; |
| 658 | |
| 659 | th.t_trans_id = 0; |
| 660 | blocksize = 1 << inode->i_blkbits; |
| 661 | |
| 662 | if (logit) { |
| 663 | reiserfs_write_lock(s); |
| 664 | ret = journal_begin(&th, s, bh_per_page + 1); |
| 665 | if (ret) |
| 666 | goto drop_write_lock; |
| 667 | reiserfs_update_inode_transaction(inode); |
| 668 | } |
| 669 | for(bh = head = page_buffers(page), block_start = 0; |
| 670 | bh != head || !block_start; |
| 671 | block_start=block_end, bh = bh->b_this_page) |
| 672 | { |
| 673 | |
| 674 | new = buffer_new(bh); |
| 675 | clear_buffer_new(bh); |
| 676 | block_end = block_start + blocksize; |
| 677 | if (block_end <= from || block_start >= to) { |
| 678 | if (!buffer_uptodate(bh)) |
| 679 | partial = 1; |
| 680 | } else { |
| 681 | set_buffer_uptodate(bh); |
| 682 | if (logit) { |
| 683 | reiserfs_prepare_for_journal(s, bh, 1); |
| 684 | journal_mark_dirty(&th, s, bh); |
| 685 | } else if (!buffer_dirty(bh)) { |
| 686 | mark_buffer_dirty(bh); |
| 687 | /* do data=ordered on any page past the end |
| 688 | * of file and any buffer marked BH_New. |
| 689 | */ |
| 690 | if (reiserfs_data_ordered(inode->i_sb) && |
| 691 | (new || page->index >= i_size_index)) { |
| 692 | reiserfs_add_ordered_list(inode, bh); |
| 693 | } |
| 694 | } |
| 695 | } |
| 696 | } |
| 697 | if (logit) { |
| 698 | ret = journal_end(&th, s, bh_per_page + 1); |
| 699 | drop_write_lock: |
| 700 | reiserfs_write_unlock(s); |
| 701 | } |
| 702 | /* |
| 703 | * If this is a partial write which happened to make all buffers |
| 704 | * uptodate then we can optimize away a bogus readpage() for |
| 705 | * the next read(). Here we 'discover' whether the page went |
| 706 | * uptodate as a result of this (potentially partial) write. |
| 707 | */ |
| 708 | if (!partial) |
| 709 | SetPageUptodate(page); |
| 710 | return ret; |
| 711 | } |
| 712 | |
| 713 | |
| 714 | /* Submit pages for write. This was separated from actual file copying |
| 715 | because we might want to allocate block numbers in-between. |
| 716 | This function assumes that caller will adjust file size to correct value. */ |
| 717 | static int reiserfs_submit_file_region_for_write( |
| 718 | struct reiserfs_transaction_handle *th, |
| 719 | struct inode *inode, |
| 720 | loff_t pos, /* Writing position offset */ |
| 721 | size_t num_pages, /* Number of pages to write */ |
| 722 | size_t write_bytes, /* number of bytes to write */ |
| 723 | struct page **prepared_pages /* list of pages */ |
| 724 | ) |
| 725 | { |
| 726 | int status; // return status of block_commit_write. |
| 727 | int retval = 0; // Return value we are going to return. |
| 728 | int i; // loop counter |
| 729 | int offset; // Writing offset in page. |
| 730 | int orig_write_bytes = write_bytes; |
| 731 | int sd_update = 0; |
| 732 | |
| 733 | for ( i = 0, offset = (pos & (PAGE_CACHE_SIZE-1)); i < num_pages ; i++,offset=0) { |
| 734 | int count = min_t(int,PAGE_CACHE_SIZE-offset,write_bytes); // How much of bytes to write to this page |
| 735 | struct page *page=prepared_pages[i]; // Current page we process. |
| 736 | |
| 737 | status = reiserfs_commit_page(inode, page, offset, offset+count); |
| 738 | if ( status ) |
| 739 | retval = status; // To not overcomplicate matters We are going to |
| 740 | // submit all the pages even if there was error. |
| 741 | // we only remember error status to report it on |
| 742 | // exit. |
| 743 | write_bytes-=count; |
| 744 | } |
| 745 | /* now that we've gotten all the ordered buffers marked dirty, |
| 746 | * we can safely update i_size and close any running transaction |
| 747 | */ |
| 748 | if ( pos + orig_write_bytes > inode->i_size) { |
| 749 | inode->i_size = pos + orig_write_bytes; // Set new size |
| 750 | /* If the file have grown so much that tail packing is no |
| 751 | * longer possible, reset "need to pack" flag */ |
| 752 | if ( (have_large_tails (inode->i_sb) && |
| 753 | inode->i_size > i_block_size (inode)*4) || |
| 754 | (have_small_tails (inode->i_sb) && |
| 755 | inode->i_size > i_block_size(inode)) ) |
| 756 | REISERFS_I(inode)->i_flags &= ~i_pack_on_close_mask ; |
| 757 | else if ( (have_large_tails (inode->i_sb) && |
| 758 | inode->i_size < i_block_size (inode)*4) || |
| 759 | (have_small_tails (inode->i_sb) && |
| 760 | inode->i_size < i_block_size(inode)) ) |
| 761 | REISERFS_I(inode)->i_flags |= i_pack_on_close_mask ; |
| 762 | |
| 763 | if (th->t_trans_id) { |
| 764 | reiserfs_write_lock(inode->i_sb); |
| 765 | reiserfs_update_sd(th, inode); // And update on-disk metadata |
| 766 | reiserfs_write_unlock(inode->i_sb); |
| 767 | } else |
| 768 | inode->i_sb->s_op->dirty_inode(inode); |
| 769 | |
| 770 | sd_update = 1; |
| 771 | } |
| 772 | if (th->t_trans_id) { |
| 773 | reiserfs_write_lock(inode->i_sb); |
| 774 | if (!sd_update) |
| 775 | reiserfs_update_sd(th, inode); |
| 776 | status = journal_end(th, th->t_super, th->t_blocks_allocated); |
| 777 | if (status) |
| 778 | retval = status; |
| 779 | reiserfs_write_unlock(inode->i_sb); |
| 780 | } |
| 781 | th->t_trans_id = 0; |
| 782 | |
| 783 | /* |
| 784 | * we have to unlock the pages after updating i_size, otherwise |
| 785 | * we race with writepage |
| 786 | */ |
| 787 | for ( i = 0; i < num_pages ; i++) { |
| 788 | struct page *page=prepared_pages[i]; |
| 789 | unlock_page(page); |
| 790 | mark_page_accessed(page); |
| 791 | page_cache_release(page); |
| 792 | } |
| 793 | return retval; |
| 794 | } |
| 795 | |
| 796 | /* Look if passed writing region is going to touch file's tail |
| 797 | (if it is present). And if it is, convert the tail to unformatted node */ |
| 798 | static int reiserfs_check_for_tail_and_convert( struct inode *inode, /* inode to deal with */ |
| 799 | loff_t pos, /* Writing position */ |
| 800 | int write_bytes /* amount of bytes to write */ |
| 801 | ) |
| 802 | { |
| 803 | INITIALIZE_PATH(path); // needed for search_for_position |
| 804 | struct cpu_key key; // Key that would represent last touched writing byte. |
| 805 | struct item_head *ih; // item header of found block; |
| 806 | int res; // Return value of various functions we call. |
| 807 | int cont_expand_offset; // We will put offset for generic_cont_expand here |
| 808 | // This can be int just because tails are created |
| 809 | // only for small files. |
| 810 | |
| 811 | /* this embodies a dependency on a particular tail policy */ |
| 812 | if ( inode->i_size >= inode->i_sb->s_blocksize*4 ) { |
| 813 | /* such a big files do not have tails, so we won't bother ourselves |
| 814 | to look for tails, simply return */ |
| 815 | return 0; |
| 816 | } |
| 817 | |
| 818 | reiserfs_write_lock(inode->i_sb); |
| 819 | /* find the item containing the last byte to be written, or if |
| 820 | * writing past the end of the file then the last item of the |
| 821 | * file (and then we check its type). */ |
| 822 | make_cpu_key (&key, inode, pos+write_bytes+1, TYPE_ANY, 3/*key length*/); |
| 823 | res = search_for_position_by_key(inode->i_sb, &key, &path); |
| 824 | if ( res == IO_ERROR ) { |
| 825 | reiserfs_write_unlock(inode->i_sb); |
| 826 | return -EIO; |
| 827 | } |
| 828 | ih = get_ih(&path); |
| 829 | res = 0; |
| 830 | if ( is_direct_le_ih(ih) ) { |
| 831 | /* Ok, closest item is file tail (tails are stored in "direct" |
| 832 | * items), so we need to unpack it. */ |
| 833 | /* To not overcomplicate matters, we just call generic_cont_expand |
| 834 | which will in turn call other stuff and finally will boil down to |
| 835 | reiserfs_get_block() that would do necessary conversion. */ |
| 836 | cont_expand_offset = le_key_k_offset(get_inode_item_key_version(inode), &(ih->ih_key)); |
| 837 | pathrelse(&path); |
| 838 | res = generic_cont_expand( inode, cont_expand_offset); |
| 839 | } else |
| 840 | pathrelse(&path); |
| 841 | |
| 842 | reiserfs_write_unlock(inode->i_sb); |
| 843 | return res; |
| 844 | } |
| 845 | |
| 846 | /* This function locks pages starting from @pos for @inode. |
| 847 | @num_pages pages are locked and stored in |
| 848 | @prepared_pages array. Also buffers are allocated for these pages. |
| 849 | First and last page of the region is read if it is overwritten only |
| 850 | partially. If last page did not exist before write (file hole or file |
| 851 | append), it is zeroed, then. |
| 852 | Returns number of unallocated blocks that should be allocated to cover |
| 853 | new file data.*/ |
| 854 | static int reiserfs_prepare_file_region_for_write( |
| 855 | struct inode *inode /* Inode of the file */, |
| 856 | loff_t pos, /* position in the file */ |
| 857 | size_t num_pages, /* number of pages to |
| 858 | prepare */ |
| 859 | size_t write_bytes, /* Amount of bytes to be |
| 860 | overwritten from |
| 861 | @pos */ |
| 862 | struct page **prepared_pages /* pointer to array |
| 863 | where to store |
| 864 | prepared pages */ |
| 865 | ) |
| 866 | { |
| 867 | int res=0; // Return values of different functions we call. |
| 868 | unsigned long index = pos >> PAGE_CACHE_SHIFT; // Offset in file in pages. |
| 869 | int from = (pos & (PAGE_CACHE_SIZE - 1)); // Writing offset in first page |
| 870 | int to = ((pos + write_bytes - 1) & (PAGE_CACHE_SIZE - 1)) + 1; |
| 871 | /* offset of last modified byte in last |
| 872 | page */ |
| 873 | struct address_space *mapping = inode->i_mapping; // Pages are mapped here. |
| 874 | int i; // Simple counter |
| 875 | int blocks = 0; /* Return value (blocks that should be allocated) */ |
| 876 | struct buffer_head *bh, *head; // Current bufferhead and first bufferhead |
| 877 | // of a page. |
| 878 | unsigned block_start, block_end; // Starting and ending offsets of current |
| 879 | // buffer in the page. |
| 880 | struct buffer_head *wait[2], **wait_bh=wait; // Buffers for page, if |
| 881 | // Page appeared to be not up |
| 882 | // to date. Note how we have |
| 883 | // at most 2 buffers, this is |
| 884 | // because we at most may |
| 885 | // partially overwrite two |
| 886 | // buffers for one page. One at // the beginning of write area |
| 887 | // and one at the end. |
| 888 | // Everything inthe middle gets // overwritten totally. |
| 889 | |
| 890 | struct cpu_key key; // cpu key of item that we are going to deal with |
| 891 | struct item_head *ih = NULL; // pointer to item head that we are going to deal with |
| 892 | struct buffer_head *itembuf=NULL; // Buffer head that contains items that we are going to deal with |
| 893 | INITIALIZE_PATH(path); // path to item, that we are going to deal with. |
| 894 | __u32 * item=NULL; // pointer to item we are going to deal with |
| 895 | int item_pos=-1; /* Position in indirect item */ |
| 896 | |
| 897 | |
| 898 | if ( num_pages < 1 ) { |
| 899 | reiserfs_warning (inode->i_sb, |
| 900 | "green-9001: reiserfs_prepare_file_region_for_write " |
| 901 | "called with zero number of pages to process"); |
| 902 | return -EFAULT; |
| 903 | } |
| 904 | |
| 905 | /* We have 2 loops for pages. In first loop we grab and lock the pages, so |
| 906 | that nobody would touch these until we release the pages. Then |
| 907 | we'd start to deal with mapping buffers to blocks. */ |
| 908 | for ( i = 0; i < num_pages; i++) { |
| 909 | prepared_pages[i] = grab_cache_page(mapping, index + i); // locks the page |
| 910 | if ( !prepared_pages[i]) { |
| 911 | res = -ENOMEM; |
| 912 | goto failed_page_grabbing; |
| 913 | } |
| 914 | if (!page_has_buffers(prepared_pages[i])) |
| 915 | create_empty_buffers(prepared_pages[i], inode->i_sb->s_blocksize, 0); |
| 916 | } |
| 917 | |
| 918 | /* Let's count amount of blocks for a case where all the blocks |
| 919 | overwritten are new (we will substract already allocated blocks later)*/ |
| 920 | if ( num_pages > 2 ) |
| 921 | /* These are full-overwritten pages so we count all the blocks in |
| 922 | these pages are counted as needed to be allocated */ |
| 923 | blocks = (num_pages - 2) << (PAGE_CACHE_SHIFT - inode->i_blkbits); |
| 924 | |
| 925 | /* count blocks needed for first page (possibly partially written) */ |
| 926 | blocks += ((PAGE_CACHE_SIZE - from) >> inode->i_blkbits) + |
| 927 | !!(from & (inode->i_sb->s_blocksize-1)); /* roundup */ |
| 928 | |
| 929 | /* Now we account for last page. If last page == first page (we |
| 930 | overwrite only one page), we substract all the blocks past the |
| 931 | last writing position in a page out of already calculated number |
| 932 | of blocks */ |
| 933 | blocks += ((num_pages > 1) << (PAGE_CACHE_SHIFT-inode->i_blkbits)) - |
| 934 | ((PAGE_CACHE_SIZE - to) >> inode->i_blkbits); |
| 935 | /* Note how we do not roundup here since partial blocks still |
| 936 | should be allocated */ |
| 937 | |
| 938 | /* Now if all the write area lies past the file end, no point in |
| 939 | maping blocks, since there is none, so we just zero out remaining |
| 940 | parts of first and last pages in write area (if needed) */ |
| 941 | if ( (pos & ~((loff_t)PAGE_CACHE_SIZE - 1)) > inode->i_size ) { |
| 942 | if ( from != 0 ) {/* First page needs to be partially zeroed */ |
| 943 | char *kaddr = kmap_atomic(prepared_pages[0], KM_USER0); |
| 944 | memset(kaddr, 0, from); |
| 945 | kunmap_atomic( kaddr, KM_USER0); |
| 946 | } |
| 947 | if ( to != PAGE_CACHE_SIZE ) { /* Last page needs to be partially zeroed */ |
| 948 | char *kaddr = kmap_atomic(prepared_pages[num_pages-1], KM_USER0); |
| 949 | memset(kaddr+to, 0, PAGE_CACHE_SIZE - to); |
| 950 | kunmap_atomic( kaddr, KM_USER0); |
| 951 | } |
| 952 | |
| 953 | /* Since all blocks are new - use already calculated value */ |
| 954 | return blocks; |
| 955 | } |
| 956 | |
| 957 | /* Well, since we write somewhere into the middle of a file, there is |
| 958 | possibility we are writing over some already allocated blocks, so |
| 959 | let's map these blocks and substract number of such blocks out of blocks |
| 960 | we need to allocate (calculated above) */ |
| 961 | /* Mask write position to start on blocksize, we do it out of the |
| 962 | loop for performance reasons */ |
| 963 | pos &= ~((loff_t) inode->i_sb->s_blocksize - 1); |
| 964 | /* Set cpu key to the starting position in a file (on left block boundary)*/ |
| 965 | make_cpu_key (&key, inode, 1 + ((pos) & ~((loff_t) inode->i_sb->s_blocksize - 1)), TYPE_ANY, 3/*key length*/); |
| 966 | |
| 967 | reiserfs_write_lock(inode->i_sb); // We need that for at least search_by_key() |
| 968 | for ( i = 0; i < num_pages ; i++ ) { |
| 969 | |
| 970 | head = page_buffers(prepared_pages[i]); |
| 971 | /* For each buffer in the page */ |
| 972 | for(bh = head, block_start = 0; bh != head || !block_start; |
| 973 | block_start=block_end, bh = bh->b_this_page) { |
| 974 | if (!bh) |
| 975 | reiserfs_panic(inode->i_sb, "green-9002: Allocated but absent buffer for a page?"); |
| 976 | /* Find where this buffer ends */ |
| 977 | block_end = block_start+inode->i_sb->s_blocksize; |
| 978 | if (i == 0 && block_end <= from ) |
| 979 | /* if this buffer is before requested data to map, skip it*/ |
| 980 | continue; |
| 981 | |
| 982 | if (i == num_pages - 1 && block_start >= to) { |
| 983 | /* If this buffer is after requested data to map, abort |
| 984 | processing of current page */ |
| 985 | break; |
| 986 | } |
| 987 | |
| 988 | if ( buffer_mapped(bh) && bh->b_blocknr !=0 ) { |
| 989 | /* This is optimisation for a case where buffer is mapped |
| 990 | and have blocknumber assigned. In case significant amount |
| 991 | of such buffers are present, we may avoid some amount |
| 992 | of search_by_key calls. |
| 993 | Probably it would be possible to move parts of this code |
| 994 | out of BKL, but I afraid that would overcomplicate code |
| 995 | without any noticeable benefit. |
| 996 | */ |
| 997 | item_pos++; |
| 998 | /* Update the key */ |
| 999 | set_cpu_key_k_offset( &key, cpu_key_k_offset(&key) + inode->i_sb->s_blocksize); |
| 1000 | blocks--; // Decrease the amount of blocks that need to be |
| 1001 | // allocated |
| 1002 | continue; // Go to the next buffer |
| 1003 | } |
| 1004 | |
| 1005 | if ( !itembuf || /* if first iteration */ |
| 1006 | item_pos >= ih_item_len(ih)/UNFM_P_SIZE) |
| 1007 | { /* or if we progressed past the |
| 1008 | current unformatted_item */ |
| 1009 | /* Try to find next item */ |
| 1010 | res = search_for_position_by_key(inode->i_sb, &key, &path); |
| 1011 | /* Abort if no more items */ |
| 1012 | if ( res != POSITION_FOUND ) { |
| 1013 | /* make sure later loops don't use this item */ |
| 1014 | itembuf = NULL; |
| 1015 | item = NULL; |
| 1016 | break; |
| 1017 | } |
| 1018 | |
| 1019 | /* Update information about current indirect item */ |
| 1020 | itembuf = get_last_bh( &path ); |
| 1021 | ih = get_ih( &path ); |
| 1022 | item = get_item( &path ); |
| 1023 | item_pos = path.pos_in_item; |
| 1024 | |
| 1025 | RFALSE( !is_indirect_le_ih (ih), "green-9003: indirect item expected"); |
| 1026 | } |
| 1027 | |
| 1028 | /* See if there is some block associated with the file |
| 1029 | at that position, map the buffer to this block */ |
| 1030 | if ( get_block_num(item,item_pos) ) { |
| 1031 | map_bh(bh, inode->i_sb, get_block_num(item,item_pos)); |
| 1032 | blocks--; // Decrease the amount of blocks that need to be |
| 1033 | // allocated |
| 1034 | } |
| 1035 | item_pos++; |
| 1036 | /* Update the key */ |
| 1037 | set_cpu_key_k_offset( &key, cpu_key_k_offset(&key) + inode->i_sb->s_blocksize); |
| 1038 | } |
| 1039 | } |
| 1040 | pathrelse(&path); // Free the path |
| 1041 | reiserfs_write_unlock(inode->i_sb); |
| 1042 | |
| 1043 | /* Now zero out unmappend buffers for the first and last pages of |
| 1044 | write area or issue read requests if page is mapped. */ |
| 1045 | /* First page, see if it is not uptodate */ |
| 1046 | if ( !PageUptodate(prepared_pages[0]) ) { |
| 1047 | head = page_buffers(prepared_pages[0]); |
| 1048 | |
| 1049 | /* For each buffer in page */ |
| 1050 | for(bh = head, block_start = 0; bh != head || !block_start; |
| 1051 | block_start=block_end, bh = bh->b_this_page) { |
| 1052 | |
| 1053 | if (!bh) |
| 1054 | reiserfs_panic(inode->i_sb, "green-9002: Allocated but absent buffer for a page?"); |
| 1055 | /* Find where this buffer ends */ |
| 1056 | block_end = block_start+inode->i_sb->s_blocksize; |
| 1057 | if ( block_end <= from ) |
| 1058 | /* if this buffer is before requested data to map, skip it*/ |
| 1059 | continue; |
| 1060 | if ( block_start < from ) { /* Aha, our partial buffer */ |
| 1061 | if ( buffer_mapped(bh) ) { /* If it is mapped, we need to |
| 1062 | issue READ request for it to |
| 1063 | not loose data */ |
| 1064 | ll_rw_block(READ, 1, &bh); |
| 1065 | *wait_bh++=bh; |
| 1066 | } else { /* Not mapped, zero it */ |
| 1067 | char *kaddr = kmap_atomic(prepared_pages[0], KM_USER0); |
| 1068 | memset(kaddr+block_start, 0, from-block_start); |
| 1069 | kunmap_atomic( kaddr, KM_USER0); |
| 1070 | set_buffer_uptodate(bh); |
| 1071 | } |
| 1072 | } |
| 1073 | } |
| 1074 | } |
| 1075 | |
| 1076 | /* Last page, see if it is not uptodate, or if the last page is past the end of the file. */ |
| 1077 | if ( !PageUptodate(prepared_pages[num_pages-1]) || |
| 1078 | ((pos+write_bytes)>>PAGE_CACHE_SHIFT) > (inode->i_size>>PAGE_CACHE_SHIFT) ) { |
| 1079 | head = page_buffers(prepared_pages[num_pages-1]); |
| 1080 | |
| 1081 | /* for each buffer in page */ |
| 1082 | for(bh = head, block_start = 0; bh != head || !block_start; |
| 1083 | block_start=block_end, bh = bh->b_this_page) { |
| 1084 | |
| 1085 | if (!bh) |
| 1086 | reiserfs_panic(inode->i_sb, "green-9002: Allocated but absent buffer for a page?"); |
| 1087 | /* Find where this buffer ends */ |
| 1088 | block_end = block_start+inode->i_sb->s_blocksize; |
| 1089 | if ( block_start >= to ) |
| 1090 | /* if this buffer is after requested data to map, skip it*/ |
| 1091 | break; |
| 1092 | if ( block_end > to ) { /* Aha, our partial buffer */ |
| 1093 | if ( buffer_mapped(bh) ) { /* If it is mapped, we need to |
| 1094 | issue READ request for it to |
| 1095 | not loose data */ |
| 1096 | ll_rw_block(READ, 1, &bh); |
| 1097 | *wait_bh++=bh; |
| 1098 | } else { /* Not mapped, zero it */ |
| 1099 | char *kaddr = kmap_atomic(prepared_pages[num_pages-1], KM_USER0); |
| 1100 | memset(kaddr+to, 0, block_end-to); |
| 1101 | kunmap_atomic( kaddr, KM_USER0); |
| 1102 | set_buffer_uptodate(bh); |
| 1103 | } |
| 1104 | } |
| 1105 | } |
| 1106 | } |
| 1107 | |
| 1108 | /* Wait for read requests we made to happen, if necessary */ |
| 1109 | while(wait_bh > wait) { |
| 1110 | wait_on_buffer(*--wait_bh); |
| 1111 | if (!buffer_uptodate(*wait_bh)) { |
| 1112 | res = -EIO; |
| 1113 | goto failed_read; |
| 1114 | } |
| 1115 | } |
| 1116 | |
| 1117 | return blocks; |
| 1118 | failed_page_grabbing: |
| 1119 | num_pages = i; |
| 1120 | failed_read: |
| 1121 | reiserfs_unprepare_pages(prepared_pages, num_pages); |
| 1122 | return res; |
| 1123 | } |
| 1124 | |
| 1125 | /* Write @count bytes at position @ppos in a file indicated by @file |
| 1126 | from the buffer @buf. |
| 1127 | |
| 1128 | generic_file_write() is only appropriate for filesystems that are not seeking to optimize performance and want |
| 1129 | something simple that works. It is not for serious use by general purpose filesystems, excepting the one that it was |
| 1130 | written for (ext2/3). This is for several reasons: |
| 1131 | |
| 1132 | * It has no understanding of any filesystem specific optimizations. |
| 1133 | |
| 1134 | * It enters the filesystem repeatedly for each page that is written. |
| 1135 | |
| 1136 | * It depends on reiserfs_get_block() function which if implemented by reiserfs performs costly search_by_key |
| 1137 | * operation for each page it is supplied with. By contrast reiserfs_file_write() feeds as much as possible at a time |
| 1138 | * to reiserfs which allows for fewer tree traversals. |
| 1139 | |
| 1140 | * Each indirect pointer insertion takes a lot of cpu, because it involves memory moves inside of blocks. |
| 1141 | |
| 1142 | * Asking the block allocation code for blocks one at a time is slightly less efficient. |
| 1143 | |
| 1144 | All of these reasons for not using only generic file write were understood back when reiserfs was first miscoded to |
| 1145 | use it, but we were in a hurry to make code freeze, and so it couldn't be revised then. This new code should make |
| 1146 | things right finally. |
| 1147 | |
| 1148 | Future Features: providing search_by_key with hints. |
| 1149 | |
| 1150 | */ |
| 1151 | static ssize_t reiserfs_file_write( struct file *file, /* the file we are going to write into */ |
| 1152 | const char __user *buf, /* pointer to user supplied data |
| 1153 | (in userspace) */ |
| 1154 | size_t count, /* amount of bytes to write */ |
| 1155 | loff_t *ppos /* pointer to position in file that we start writing at. Should be updated to |
| 1156 | * new current position before returning. */ ) |
| 1157 | { |
| 1158 | size_t already_written = 0; // Number of bytes already written to the file. |
| 1159 | loff_t pos; // Current position in the file. |
| 1160 | ssize_t res; // return value of various functions that we call. |
| 1161 | int err = 0; |
| 1162 | struct inode *inode = file->f_dentry->d_inode; // Inode of the file that we are writing to. |
| 1163 | /* To simplify coding at this time, we store |
| 1164 | locked pages in array for now */ |
| 1165 | struct page * prepared_pages[REISERFS_WRITE_PAGES_AT_A_TIME]; |
| 1166 | struct reiserfs_transaction_handle th; |
| 1167 | th.t_trans_id = 0; |
| 1168 | |
| 1169 | if ( file->f_flags & O_DIRECT) { // Direct IO needs treatment |
| 1170 | ssize_t result, after_file_end = 0; |
| 1171 | if ( (*ppos + count >= inode->i_size) || (file->f_flags & O_APPEND) ) { |
| 1172 | /* If we are appending a file, we need to put this savelink in here. |
| 1173 | If we will crash while doing direct io, finish_unfinished will |
| 1174 | cut the garbage from the file end. */ |
| 1175 | reiserfs_write_lock(inode->i_sb); |
| 1176 | err = journal_begin(&th, inode->i_sb, JOURNAL_PER_BALANCE_CNT ); |
| 1177 | if (err) { |
| 1178 | reiserfs_write_unlock (inode->i_sb); |
| 1179 | return err; |
| 1180 | } |
| 1181 | reiserfs_update_inode_transaction(inode); |
| 1182 | add_save_link (&th, inode, 1 /* Truncate */); |
| 1183 | after_file_end = 1; |
| 1184 | err = journal_end(&th, inode->i_sb, JOURNAL_PER_BALANCE_CNT ); |
| 1185 | reiserfs_write_unlock(inode->i_sb); |
| 1186 | if (err) |
| 1187 | return err; |
| 1188 | } |
| 1189 | result = generic_file_write(file, buf, count, ppos); |
| 1190 | |
| 1191 | if ( after_file_end ) { /* Now update i_size and remove the savelink */ |
| 1192 | struct reiserfs_transaction_handle th; |
| 1193 | reiserfs_write_lock(inode->i_sb); |
| 1194 | err = journal_begin(&th, inode->i_sb, 1); |
| 1195 | if (err) { |
| 1196 | reiserfs_write_unlock (inode->i_sb); |
| 1197 | return err; |
| 1198 | } |
| 1199 | reiserfs_update_inode_transaction(inode); |
| 1200 | reiserfs_update_sd(&th, inode); |
| 1201 | err = journal_end(&th, inode->i_sb, 1); |
| 1202 | if (err) { |
| 1203 | reiserfs_write_unlock (inode->i_sb); |
| 1204 | return err; |
| 1205 | } |
| 1206 | err = remove_save_link (inode, 1/* truncate */); |
| 1207 | reiserfs_write_unlock(inode->i_sb); |
| 1208 | if (err) |
| 1209 | return err; |
| 1210 | } |
| 1211 | |
| 1212 | return result; |
| 1213 | } |
| 1214 | |
| 1215 | if ( unlikely((ssize_t) count < 0 )) |
| 1216 | return -EINVAL; |
| 1217 | |
| 1218 | if (unlikely(!access_ok(VERIFY_READ, buf, count))) |
| 1219 | return -EFAULT; |
| 1220 | |
| 1221 | down(&inode->i_sem); // locks the entire file for just us |
| 1222 | |
| 1223 | pos = *ppos; |
| 1224 | |
| 1225 | /* Check if we can write to specified region of file, file |
| 1226 | is not overly big and this kind of stuff. Adjust pos and |
| 1227 | count, if needed */ |
| 1228 | res = generic_write_checks(file, &pos, &count, 0); |
| 1229 | if (res) |
| 1230 | goto out; |
| 1231 | |
| 1232 | if ( count == 0 ) |
| 1233 | goto out; |
| 1234 | |
| 1235 | res = remove_suid(file->f_dentry); |
| 1236 | if (res) |
| 1237 | goto out; |
| 1238 | |
| 1239 | inode_update_time(inode, 1); /* Both mtime and ctime */ |
| 1240 | |
| 1241 | // Ok, we are done with all the checks. |
| 1242 | |
| 1243 | // Now we should start real work |
| 1244 | |
| 1245 | /* If we are going to write past the file's packed tail or if we are going |
| 1246 | to overwrite part of the tail, we need that tail to be converted into |
| 1247 | unformatted node */ |
| 1248 | res = reiserfs_check_for_tail_and_convert( inode, pos, count); |
| 1249 | if (res) |
| 1250 | goto out; |
| 1251 | |
| 1252 | while ( count > 0) { |
| 1253 | /* This is the main loop in which we running until some error occures |
| 1254 | or until we write all of the data. */ |
| 1255 | size_t num_pages;/* amount of pages we are going to write this iteration */ |
| 1256 | size_t write_bytes; /* amount of bytes to write during this iteration */ |
| 1257 | size_t blocks_to_allocate; /* how much blocks we need to allocate for this iteration */ |
| 1258 | |
| 1259 | /* (pos & (PAGE_CACHE_SIZE-1)) is an idiom for offset into a page of pos*/ |
| 1260 | num_pages = !!((pos+count) & (PAGE_CACHE_SIZE - 1)) + /* round up partial |
| 1261 | pages */ |
| 1262 | ((count + (pos & (PAGE_CACHE_SIZE-1))) >> PAGE_CACHE_SHIFT); |
| 1263 | /* convert size to amount of |
| 1264 | pages */ |
| 1265 | reiserfs_write_lock(inode->i_sb); |
| 1266 | if ( num_pages > REISERFS_WRITE_PAGES_AT_A_TIME |
| 1267 | || num_pages > reiserfs_can_fit_pages(inode->i_sb) ) { |
| 1268 | /* If we were asked to write more data than we want to or if there |
| 1269 | is not that much space, then we shorten amount of data to write |
| 1270 | for this iteration. */ |
| 1271 | num_pages = min_t(size_t, REISERFS_WRITE_PAGES_AT_A_TIME, reiserfs_can_fit_pages(inode->i_sb)); |
| 1272 | /* Also we should not forget to set size in bytes accordingly */ |
| 1273 | write_bytes = (num_pages << PAGE_CACHE_SHIFT) - |
| 1274 | (pos & (PAGE_CACHE_SIZE-1)); |
| 1275 | /* If position is not on the |
| 1276 | start of the page, we need |
| 1277 | to substract the offset |
| 1278 | within page */ |
| 1279 | } else |
| 1280 | write_bytes = count; |
| 1281 | |
| 1282 | /* reserve the blocks to be allocated later, so that later on |
| 1283 | we still have the space to write the blocks to */ |
| 1284 | reiserfs_claim_blocks_to_be_allocated(inode->i_sb, num_pages << (PAGE_CACHE_SHIFT - inode->i_blkbits)); |
| 1285 | reiserfs_write_unlock(inode->i_sb); |
| 1286 | |
Jan Kara | 127144d | 2005-05-01 08:59:07 -0700 | [diff] [blame^] | 1287 | if ( !num_pages ) { /* If we do not have enough space even for a single page... */ |
| 1288 | if ( pos > inode->i_size+inode->i_sb->s_blocksize-(pos & (inode->i_sb->s_blocksize-1))) { |
| 1289 | res = -ENOSPC; |
| 1290 | break; // In case we are writing past the end of the last file block, break. |
| 1291 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1292 | // Otherwise we are possibly overwriting the file, so |
| 1293 | // let's set write size to be equal or less than blocksize. |
| 1294 | // This way we get it correctly for file holes. |
| 1295 | // But overwriting files on absolutelly full volumes would not |
| 1296 | // be very efficient. Well, people are not supposed to fill |
| 1297 | // 100% of disk space anyway. |
| 1298 | write_bytes = min_t(size_t, count, inode->i_sb->s_blocksize - (pos & (inode->i_sb->s_blocksize - 1))); |
| 1299 | num_pages = 1; |
| 1300 | // No blocks were claimed before, so do it now. |
| 1301 | reiserfs_claim_blocks_to_be_allocated(inode->i_sb, 1 << (PAGE_CACHE_SHIFT - inode->i_blkbits)); |
| 1302 | } |
| 1303 | |
| 1304 | /* Prepare for writing into the region, read in all the |
| 1305 | partially overwritten pages, if needed. And lock the pages, |
| 1306 | so that nobody else can access these until we are done. |
| 1307 | We get number of actual blocks needed as a result.*/ |
| 1308 | blocks_to_allocate = reiserfs_prepare_file_region_for_write(inode, pos, num_pages, write_bytes, prepared_pages); |
| 1309 | if ( blocks_to_allocate < 0 ) { |
| 1310 | res = blocks_to_allocate; |
| 1311 | reiserfs_release_claimed_blocks(inode->i_sb, num_pages << (PAGE_CACHE_SHIFT - inode->i_blkbits)); |
| 1312 | break; |
| 1313 | } |
| 1314 | |
| 1315 | /* First we correct our estimate of how many blocks we need */ |
| 1316 | reiserfs_release_claimed_blocks(inode->i_sb, (num_pages << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits)) - blocks_to_allocate ); |
| 1317 | |
| 1318 | if ( blocks_to_allocate > 0) {/*We only allocate blocks if we need to*/ |
| 1319 | /* Fill in all the possible holes and append the file if needed */ |
| 1320 | res = reiserfs_allocate_blocks_for_region(&th, inode, pos, num_pages, write_bytes, prepared_pages, blocks_to_allocate); |
| 1321 | } |
| 1322 | |
| 1323 | /* well, we have allocated the blocks, so it is time to free |
| 1324 | the reservation we made earlier. */ |
| 1325 | reiserfs_release_claimed_blocks(inode->i_sb, blocks_to_allocate); |
| 1326 | if ( res ) { |
| 1327 | reiserfs_unprepare_pages(prepared_pages, num_pages); |
| 1328 | break; |
| 1329 | } |
| 1330 | |
| 1331 | /* NOTE that allocating blocks and filling blocks can be done in reverse order |
| 1332 | and probably we would do that just to get rid of garbage in files after a |
| 1333 | crash */ |
| 1334 | |
| 1335 | /* Copy data from user-supplied buffer to file's pages */ |
| 1336 | res = reiserfs_copy_from_user_to_file_region(pos, num_pages, write_bytes, prepared_pages, buf); |
| 1337 | if ( res ) { |
| 1338 | reiserfs_unprepare_pages(prepared_pages, num_pages); |
| 1339 | break; |
| 1340 | } |
| 1341 | |
| 1342 | /* Send the pages to disk and unlock them. */ |
| 1343 | res = reiserfs_submit_file_region_for_write(&th, inode, pos, num_pages, |
| 1344 | write_bytes,prepared_pages); |
| 1345 | if ( res ) |
| 1346 | break; |
| 1347 | |
| 1348 | already_written += write_bytes; |
| 1349 | buf += write_bytes; |
| 1350 | *ppos = pos += write_bytes; |
| 1351 | count -= write_bytes; |
| 1352 | balance_dirty_pages_ratelimited(inode->i_mapping); |
| 1353 | } |
| 1354 | |
| 1355 | /* this is only true on error */ |
| 1356 | if (th.t_trans_id) { |
| 1357 | reiserfs_write_lock(inode->i_sb); |
| 1358 | err = journal_end(&th, th.t_super, th.t_blocks_allocated); |
| 1359 | reiserfs_write_unlock(inode->i_sb); |
| 1360 | if (err) { |
| 1361 | res = err; |
| 1362 | goto out; |
| 1363 | } |
| 1364 | } |
| 1365 | |
| 1366 | if ((file->f_flags & O_SYNC) || IS_SYNC(inode)) |
| 1367 | res = generic_osync_inode(inode, file->f_mapping, OSYNC_METADATA|OSYNC_DATA); |
| 1368 | |
| 1369 | up(&inode->i_sem); |
| 1370 | reiserfs_async_progress_wait(inode->i_sb); |
| 1371 | return (already_written != 0)?already_written:res; |
| 1372 | |
| 1373 | out: |
| 1374 | up(&inode->i_sem); // unlock the file on exit. |
| 1375 | return res; |
| 1376 | } |
| 1377 | |
| 1378 | static ssize_t reiserfs_aio_write(struct kiocb *iocb, const char __user *buf, |
| 1379 | size_t count, loff_t pos) |
| 1380 | { |
| 1381 | return generic_file_aio_write(iocb, buf, count, pos); |
| 1382 | } |
| 1383 | |
| 1384 | |
| 1385 | |
| 1386 | struct file_operations reiserfs_file_operations = { |
| 1387 | .read = generic_file_read, |
| 1388 | .write = reiserfs_file_write, |
| 1389 | .ioctl = reiserfs_ioctl, |
| 1390 | .mmap = generic_file_mmap, |
| 1391 | .release = reiserfs_file_release, |
| 1392 | .fsync = reiserfs_sync_file, |
| 1393 | .sendfile = generic_file_sendfile, |
| 1394 | .aio_read = generic_file_aio_read, |
| 1395 | .aio_write = reiserfs_aio_write, |
| 1396 | }; |
| 1397 | |
| 1398 | |
| 1399 | struct inode_operations reiserfs_file_inode_operations = { |
| 1400 | .truncate = reiserfs_vfs_truncate_file, |
| 1401 | .setattr = reiserfs_setattr, |
| 1402 | .setxattr = reiserfs_setxattr, |
| 1403 | .getxattr = reiserfs_getxattr, |
| 1404 | .listxattr = reiserfs_listxattr, |
| 1405 | .removexattr = reiserfs_removexattr, |
| 1406 | .permission = reiserfs_permission, |
| 1407 | }; |
| 1408 | |
| 1409 | |