| blob | abfada2f52db7b2a54f282d968e3a634199df0ad |
1 /*
2 * Copyright 2000 by Hans Reiser, licensing governed by reiserfs/README
3 */
5 #include <linux/time.h>
6 #include <linux/reiserfs_fs.h>
7 #include <linux/reiserfs_acl.h>
8 #include <linux/reiserfs_xattr.h>
9 #include <linux/smp_lock.h>
10 #include <asm/uaccess.h>
11 #include <linux/pagemap.h>
12 #include <linux/swap.h>
13 #include <linux/writeback.h>
14 #include <linux/blkdev.h>
15 #include <linux/buffer_head.h>
16 #include <linux/quotaops.h>
18 /*
19 ** We pack the tails of files on file close, not at the time they are written.
20 ** This implies an unnecessary copy of the tail and an unnecessary indirect item
21 ** insertion/balancing, for files that are written in one write.
22 ** It avoids unnecessary tail packings (balances) for files that are written in
23 ** multiple writes and are small enough to have tails.
24 **
25 ** file_release is called by the VFS layer when the file is closed. If
26 ** this is the last open file descriptor, and the file
27 ** small enough to have a tail, and the tail is currently in an
28 ** unformatted node, the tail is converted back into a direct item.
29 **
30 ** We use reiserfs_truncate_file to pack the tail, since it already has
31 ** all the conditions coded.
32 */
34 {
42 /* fast out for when nothing needs to be done */
48 }
57 /* freeing preallocation only involves relogging blocks that
58 * are already in the current transaction. preallocation gets
59 * freed at the end of each transaction, so it is impossible for
60 * us to log any additional blocks (including quota blocks)
61 */
64 /* uh oh, we can't allow the inode to go away while there
65 * is still preallocation blocks pending. Try to join the
66 * aborted transaction
67 */
72 /* hmpf, our choices here aren't good. We can pin the inode
73 * which will disallow unmount from every happening, we can
74 * do nothing, which will corrupt random memory on unmount,
75 * or we can forcibly remove the file from the preallocation
76 * list, which will leak blocks on disk. Lets pin the inode
77 * and let the admin know what is going on.
78 */
81 "pinning inode %lu because the "
85 }
86 }
89 #ifdef REISERFS_PREALLOCATE
91 #endif
94 /* copy back the error code from journal_begin */
101 /* if regular file is released by last holder and it has been
102 appended (we append by unformatted node only) or its direct
103 item(s) had to be converted, then it may have to be
104 indirect2direct converted */
106 }
107 out:
112 }
115 {
124 }
127 {
129 }
131 /* Sync a reiserfs file. */
133 /*
134 * FIXME: sync_mapping_buffers() never has anything to sync. Can
135 * be removed...
136 */
140 {
155 }
157 /* I really do not want to play with memory shortage right now, so
158 to simplify the code, we are not going to write more than this much pages at
159 a time. This still should considerably improve performance compared to 4k
160 at a time case. This is 32 pages of 4k size. */
161 #define REISERFS_WRITE_PAGES_AT_A_TIME (128 * 1024) / PAGE_CACHE_SIZE
163 /* Allocates blocks for a file to fulfil write request.
164 Maps all unmapped but prepared pages from the list.
165 Updates metadata with newly allocated blocknumbers as needed */
166 static int reiserfs_allocate_blocks_for_region(struct reiserfs_transaction_handle *th, struct inode *inode, /* Inode we work with */
169 to touch */
172 prepared pages
173 */
175 need to allocate to
176 fit the data into file
177 */
178 )
179 {
185 b_blocknr_t *allocated_blocks; // Pointer to a place where allocated blocknumbers would be stored.
192 // first page
193 unsigned int to = ((pos + write_bytes - 1) & (PAGE_CACHE_SIZE - 1)) + 1; /* last modified byte offset in last page */
196 // of the fact that we already prepared
197 // current block for journal
202 /* only preallocate if this is a small write */
205 blocks_to_allocate <
207 will_prealloc =
215 /* First we compose a key to point at the writing position, we want to do
216 that outside of any locking region. */
219 /* If we came here, it means we absolutely need to open a transaction,
220 since we need to allocate some blocks */
222 res = journal_begin(th, inode->i_sb, JOURNAL_PER_BALANCE_CNT * 3 + 1 + 2 * REISERFS_QUOTA_TRANS_BLOCKS(inode->i_sb)); // Wish I know if this number enough
227 /* Look for the in-tree position of our write, need path for block allocator */
232 }
234 /* Allocate blocks */
235 /* First fill in "hint" structure for block allocator */
237 hint.path = &path; // Path, so that block allocator can determine packing locality or whatever it needs to determine.
241 hint.block = inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9); // Number of disk blocks this file occupies already.
245 /* Call block allocator to allocate blocks */
246 res =
251 /* We flush the transaction in case of no space. This way some
252 blocks might become free */
258 /* We might have scheduled, so search again */
259 res =
265 }
267 /* update changed info for hint structure. */
268 res =
270 blocks_to_allocate,
271 blocks_to_allocate);
276 }
281 }
282 }
283 #ifdef __BIG_ENDIAN
284 // Too bad, I have not found any way to convert a given region from
285 // cpu format to little endian format
286 {
290 }
291 #endif
293 /* Blocks allocating well might have scheduled and tree might have changed,
294 let's search the tree again */
295 /* find where in the tree our write should go */
300 }
306 /* Let's see what we have found */
308 might need to append file with holes
309 first */
310 // Since we are writing past the file's end, we need to find out if
311 // there is a hole that needs to be inserted before our writing
312 // position, and how many blocks it is going to cover (we need to
313 // populate pointers to file blocks representing the hole with zeros)
315 {
317 /*
318 * if ih is stat data, its offset is 0 and we don't want to
319 * add 1 to pos in the hole_size calculation
320 */
324 (le_key_k_offset
327 inode->
328 i_sb->
329 s_blocksize)))
331 }
334 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.
335 /* area filled with zeroes, to supply as list of zero blocknumbers
336 We allocate it outside of loop just in case loop would spin for
337 several iterations. */
338 char *zeros = kzalloc(to_paste * UNFM_P_SIZE, GFP_ATOMIC); // We cannot insert more than MAX_ITEM_LEN bytes anyway.
342 }
344 to_paste =
347 s_blocksize) /
348 UNFM_P_SIZE);
350 /* Ok, there is existing indirect item already. Need to append it */
351 /* Calculate position past inserted item */
353 le_key_k_offset
354 (get_inode_item_key_version
358 inode->
359 i_sb->
360 s_blocksize),
362 res =
365 inode,
367 zeros,
368 UNFM_P_SIZE
369 *
370 to_paste);
374 }
376 /* No existing item, create it */
377 /* item head for new item */
380 /* create a key for our new item */
384 /* Create new item head for our new item */
387 TYPE_INDIRECT,
388 to_paste *
389 UNFM_P_SIZE,
392 /* Find where such item should live in the tree */
393 res =
397 /* item should not exist, otherwise we have error */
400 i_sb,
403 res);
404 }
408 }
409 res =
412 inode,
418 }
422 }
423 /* Now we want to check if transaction is too full, and if it is
424 we restart it. This will also free the path. */
429 res =
436 }
437 }
439 /* Well, need to recalculate path and stuff */
443 i_blkbits));
444 res =
451 }
458 }
459 }
460 // Go through existing indirect items first
461 // replace all zeroes with blocknumbers from list
462 // Note that if no corresponding item was found, by previous search,
463 // it means there are no existing in-tree representation for file area
464 // we are going to overwrite, so there is nothing to scan through for holes.
467 retry:
470 /* We run out of data in this indirect item, let's look for another
471 one. */
472 /* First if we are already modifying current item, log it */
476 }
477 /* Then set the key to look for a new indirect item (offset of old
478 item is added to old item length */
480 le_key_k_offset
485 s_blocksize));
486 /* Search ofor position of new key in the tree. */
487 res =
493 }
499 }
500 /* Ok, we have correct position in item now, so let's see if it is
501 representing file hole (blocknumber is zero) and fill it if needed */
503 /* Ok, a hole. Now we need to check if we already prepared this
504 block to be journaled */
506 /* Well, this item is not journaled yet, so we must prepare
507 it for journal first, before we can change it */
509 // here to detect if fs changed under
510 // us while we were preparing for
511 // journal.
513 // changes fs under our feet
517 reiserfs_prepare_for_journal(inode->i_sb, bh, 1); // Prepare a buffer within which indirect item is stored for changing.
520 // Sigh, fs was changed under us, we need to look for new
521 // location of item we are working with
523 /* unmark prepaerd area as journaled and search for it's
524 new position */
526 i_sb,
527 bh);
528 res =
530 i_sb,
536 }
542 }
544 }
546 curr_block++;
547 }
548 itempos++;
549 }
552 // was modified, but not logged yet.
554 }
557 // Oh, well need to append to indirect item, or to create indirect item
558 // if there weren't any
560 // Existing indirect item - append. First calculate key for append
561 // position. We do not need to recalculate path as it should
562 // already point to correct place.
570 res =
573 curr_block),
574 UNFM_P_SIZE *
576 curr_block));
579 }
581 // Last found item was statdata. That means we need to create indirect item.
584 /* create a key for our new item */
586 // because that's
587 // where first
588 // indirect item
589 // begins
590 /* Create new item head for our new item */
592 TYPE_INDIRECT,
596 /* Find where such item should live in the tree */
599 /* Well, if we have found such item already, or some error
600 occured, we need to warn user and return error */
603 "green-9009: search_by_key (%K) "
605 res);
606 }
609 }
610 /* Insert item into the tree with the data as its body */
611 res =
613 inode,
615 curr_block));
620 }
621 }
622 // the caller is responsible for closing the transaction
623 // unless we return an error, they are also responsible for logging
624 // the inode.
625 //
627 /*
628 * cleanup prellocation from previous writes
629 * if this is a partial block write
630 */
635 // go through all the pages/buffers and map the buffers to newly allocated
636 // blocks (so that system knows where to write these pages later).
647 /* For each buffer in page */
655 /* if this buffer is before requested data to map, skip it */
658 /* If this buffer is after requested data to map, abort
659 processing of current page */
666 curr_block++;
668 }
669 }
670 }
678 // Need to deal with transaction here.
679 error_exit_free_blocks:
681 // free blocks
686 error_exit:
689 // update any changes we made to blk count
691 err =
697 }
702 }
704 /* Unlock pages prepared by reiserfs_prepare_file_region_for_write */
707 {
716 }
717 }
719 /* This function will copy data from userspace to specified pages within
720 supplied byte range */
725 array to
726 prepared pages
727 */
729 data */
730 )
731 {
738 size_t count = min_t(size_t, PAGE_CACHE_SIZE - offset, write_bytes); // How much of bytes to write to this page
743 /* Copy data from userspace to the current page */
746 /* Flush processor's dcache for this page */
754 }
757 }
759 /* taken fs/buffer.c:__block_commit_write */
762 {
784 }
802 /* do data=ordered on any page past the end
803 * of file and any buffer marked BH_New.
804 */
808 }
809 }
810 }
811 }
814 drop_write_lock:
816 }
817 /*
818 * If this is a partial write which happened to make all buffers
819 * uptodate then we can optimize away a bogus readpage() for
820 * the next read(). Here we 'discover' whether the page went
821 * uptodate as a result of this (potentially partial) write.
822 */
826 }
828 /* Submit pages for write. This was separated from actual file copying
829 because we might want to allocate block numbers in-between.
830 This function assumes that caller will adjust file size to correct value. */
831 static int reiserfs_submit_file_region_for_write(struct reiserfs_transaction_handle *th, struct inode *inode, loff_t pos, /* Writing position offset */
835 )
836 {
846 int count = min_t(int, PAGE_CACHE_SIZE - offset, write_bytes); // How much of bytes to write to this page
849 status =
853 // submit all the pages even if there was error.
854 // we only remember error status to report it on
855 // exit.
857 }
858 /* now that we've gotten all the ordered buffers marked dirty,
859 * we can safely update i_size and close any running transaction
860 */
863 /* If the file have grown so much that tail packing is no
864 * longer possible, reset "need to pack" flag */
878 // this sets the proper flags for O_SYNC to trigger a commit
886 }
889 }
898 }
901 /*
902 * we have to unlock the pages after updating i_size, otherwise
903 * we race with writepage
904 */
910 }
912 }
914 /* Look if passed writing region is going to touch file's tail
915 (if it is present). And if it is, convert the tail to unformatted node */
919 )
920 {
926 // This can be int just because tails are created
927 // only for small files.
929 /* this embodies a dependency on a particular tail policy */
931 /* such a big files do not have tails, so we won't bother ourselves
932 to look for tails, simply return */
934 }
937 /* find the item containing the last byte to be written, or if
938 * writing past the end of the file then the last item of the
939 * file (and then we check its type). */
946 }
950 /* Ok, closest item is file tail (tails are stored in "direct"
951 * items), so we need to unpack it. */
952 /* To not overcomplicate matters, we just call generic_cont_expand
953 which will in turn call other stuff and finally will boil down to
954 reiserfs_get_block() that would do necessary conversion. */
955 cont_expand_offset =
965 }
967 /* This function locks pages starting from @pos for @inode.
968 @num_pages pages are locked and stored in
969 @prepared_pages array. Also buffers are allocated for these pages.
970 First and last page of the region is read if it is overwritten only
971 partially. If last page did not exist before write (file hole or file
972 append), it is zeroed, then.
973 Returns number of unallocated blocks that should be allocated to cover
974 new file data.*/
979 prepare */
981 overwritten from
982 @pos */
984 where to store
985 prepared pages */
986 )
987 {
992 /* offset of last modified byte in last
993 page */
998 // of a page.
1000 // buffer in the page.
1002 // Page appeared to be not up
1003 // to date. Note how we have
1004 // at most 2 buffers, this is
1005 // because we at most may
1006 // partially overwrite two
1007 // buffers for one page. One at // the beginning of write area
1008 // and one at the end.
1009 // Everything inthe middle gets // overwritten totally.
1013 struct buffer_head *itembuf = NULL; // Buffer head that contains items that we are going to deal with
1020 "green-9001: reiserfs_prepare_file_region_for_write "
1023 }
1025 /* We have 2 loops for pages. In first loop we grab and lock the pages, so
1026 that nobody would touch these until we release the pages. Then
1027 we'd start to deal with mapping buffers to blocks. */
1033 }
1037 }
1039 /* Let's count amount of blocks for a case where all the blocks
1040 overwritten are new (we will substract already allocated blocks later) */
1042 /* These are full-overwritten pages so we count all the blocks in
1043 these pages are counted as needed to be allocated */
1044 blocks =
1047 /* count blocks needed for first page (possibly partially written) */
1048 blocks += ((PAGE_CACHE_SIZE - from) >> inode->i_blkbits) + !!(from & (inode->i_sb->s_blocksize - 1)); /* roundup */
1050 /* Now we account for last page. If last page == first page (we
1051 overwrite only one page), we substract all the blocks past the
1052 last writing position in a page out of already calculated number
1053 of blocks */
1056 /* Note how we do not roundup here since partial blocks still
1057 should be allocated */
1059 /* Now if all the write area lies past the file end, no point in
1060 maping blocks, since there is none, so we just zero out remaining
1061 parts of first and last pages in write area (if needed) */
1068 }
1072 KM_USER0);
1076 }
1078 /* Since all blocks are new - use already calculated value */
1080 }
1082 /* Well, since we write somewhere into the middle of a file, there is
1083 possibility we are writing over some already allocated blocks, so
1084 let's map these blocks and substract number of such blocks out of blocks
1085 we need to allocate (calculated above) */
1086 /* Mask write position to start on blocksize, we do it out of the
1087 loop for performance reasons */
1089 /* Set cpu key to the starting position in a file (on left block boundary) */
1098 /* For each buffer in the page */
1104 /* Find where this buffer ends */
1107 /* if this buffer is before requested data to map, skip it */
1111 /* If this buffer is after requested data to map, abort
1112 processing of current page */
1114 }
1117 /* This is optimisation for a case where buffer is mapped
1118 and have blocknumber assigned. In case significant amount
1119 of such buffers are present, we may avoid some amount
1120 of search_by_key calls.
1121 Probably it would be possible to move parts of this code
1122 out of BKL, but I afraid that would overcomplicate code
1123 without any noticeable benefit.
1124 */
1125 item_pos++;
1126 /* Update the key */
1131 // allocated
1133 }
1137 current unformatted_item */
1138 /* Try to find next item */
1139 res =
1142 /* Abort if no more items */
1144 /* make sure later loops don't use this item */
1148 }
1150 /* Update information about current indirect item */
1158 }
1160 /* See if there is some block associated with the file
1161 at that position, map the buffer to this block */
1166 // allocated
1167 }
1168 item_pos++;
1169 /* Update the key */
1173 }
1174 }
1178 /* Now zero out unmappend buffers for the first and last pages of
1179 write area or issue read requests if page is mapped. */
1180 /* First page, see if it is not uptodate */
1184 /* For each buffer in page */
1191 /* Find where this buffer ends */
1194 /* if this buffer is before requested data to map, skip it */
1198 issue READ request for it to
1199 not loose data */
1205 KM_USER0);
1211 }
1212 }
1213 }
1214 }
1216 /* Last page, see if it is not uptodate, or if the last page is past the end of the file. */
1222 /* for each buffer in page */
1229 /* Find where this buffer ends */
1232 /* if this buffer is after requested data to map, skip it */
1236 issue READ request for it to
1237 not loose data */
1244 KM_USER0);
1249 }
1250 }
1251 }
1252 }
1254 /* Wait for read requests we made to happen, if necessary */
1260 }
1261 }
1264 failed_page_grabbing:
1266 failed_read:
1269 }
1271 /* Write @count bytes at position @ppos in a file indicated by @file
1272 from the buffer @buf.
1274 generic_file_write() is only appropriate for filesystems that are not seeking to optimize performance and want
1275 something simple that works. It is not for serious use by general purpose filesystems, excepting the one that it was
1276 written for (ext2/3). This is for several reasons:
1278 * It has no understanding of any filesystem specific optimizations.
1280 * It enters the filesystem repeatedly for each page that is written.
1282 * It depends on reiserfs_get_block() function which if implemented by reiserfs performs costly search_by_key
1283 * operation for each page it is supplied with. By contrast reiserfs_file_write() feeds as much as possible at a time
1284 * to reiserfs which allows for fewer tree traversals.
1286 * Each indirect pointer insertion takes a lot of cpu, because it involves memory moves inside of blocks.
1288 * Asking the block allocation code for blocks one at a time is slightly less efficient.
1290 All of these reasons for not using only generic file write were understood back when reiserfs was first miscoded to
1291 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
1292 things right finally.
1294 Future Features: providing search_by_key with hints.
1296 */
1297 static ssize_t reiserfs_file_write(struct file *file, /* the file we are going to write into */
1299 (in userspace) */
1302 * new current position before returning. */
1303 )
1304 {
1309 struct inode *inode = file->f_path.dentry->d_inode; // Inode of the file that we are writing to.
1310 /* To simplify coding at this time, we store
1311 locked pages in array for now */
1316 /* If a filesystem is converted from 3.5 to 3.6, we'll have v3.5 items
1317 * lying around (most of the disk, in fact). Despite the filesystem
1318 * now being a v3.6 format, the old items still can't support large
1319 * file sizes. Catch this case here, as the rest of the VFS layer is
1320 * oblivious to the different limitations between old and new items.
1321 * reiserfs_setattr catches this for truncates. This chunk is lifted
1322 * from generic_write_checks. */
1328 }
1331 }
1346 /* Check if we can write to specified region of file, file
1347 is not overly big and this kind of stuff. Adjust pos and
1348 count, if needed */
1362 // Ok, we are done with all the checks.
1364 // Now we should start real work
1366 /* If we are going to write past the file's packed tail or if we are going
1367 to overwrite part of the tail, we need that tail to be converted into
1368 unformatted node */
1374 /* This is the main loop in which we running until some error occures
1375 or until we write all of the data. */
1380 /* (pos & (PAGE_CACHE_SIZE-1)) is an idiom for offset into a page of pos */
1382 pages */
1385 /* convert size to amount of
1386 pages */
1390 /* If we were asked to write more data than we want to or if there
1391 is not that much space, then we shorten amount of data to write
1392 for this iteration. */
1393 num_pages =
1396 /* Also we should not forget to set size in bytes accordingly */
1399 /* If position is not on the
1400 start of the page, we need
1401 to substract the offset
1402 within page */
1406 /* reserve the blocks to be allocated later, so that later on
1407 we still have the space to write the blocks to */
1409 num_pages <<
1420 }
1421 // Otherwise we are possibly overwriting the file, so
1422 // let's set write size to be equal or less than blocksize.
1423 // This way we get it correctly for file holes.
1424 // But overwriting files on absolutelly full volumes would not
1425 // be very efficient. Well, people are not supposed to fill
1426 // 100% of disk space anyway.
1427 write_bytes =
1432 // No blocks were claimed before, so do it now.
1435 (PAGE_CACHE_SHIFT
1436 -
1437 inode->
1438 i_blkbits));
1439 }
1441 /* Prepare for writing into the region, read in all the
1442 partially overwritten pages, if needed. And lock the pages,
1443 so that nobody else can access these until we are done.
1444 We get number of actual blocks needed as a result. */
1446 num_pages,
1447 write_bytes,
1448 prepared_pages);
1451 num_pages <<
1455 }
1459 /* First we correct our estimate of how many blocks we need */
1464 s_blocksize_bits)) -
1465 blocks_to_allocate);
1468 /* Fill in all the possible holes and append the file if needed */
1469 res =
1471 num_pages,
1472 write_bytes,
1473 prepared_pages,
1474 blocks_to_allocate);
1475 }
1477 /* well, we have allocated the blocks, so it is time to free
1478 the reservation we made earlier. */
1480 blocks_to_allocate);
1484 }
1486 /* NOTE that allocating blocks and filling blocks can be done in reverse order
1487 and probably we would do that just to get rid of garbage in files after a
1488 crash */
1490 /* Copy data from user-supplied buffer to file's pages */
1491 res =
1493 write_bytes,
1498 }
1500 /* Send the pages to disk and unlock them. */
1501 res =
1503 num_pages,
1504 write_bytes,
1505 prepared_pages);
1514 }
1516 /* this is only true on error */
1524 }
1525 }
1536 out:
1539 }
1545 #ifdef CONFIG_COMPAT
1547 #endif
1557 };
1567 };