1 /*
2 * e4defrag.c - ext4 filesystem defragmenter
3 *
4 * Copyright (C) 2009 NEC Software Tohoku, Ltd.
5 *
6 * Author: Akira Fujita <a-fujita@rs.jp.nec.com>
7 * Takashi Sato <t-sato@yk.jp.nec.com>
8 */
9
10 #ifndef _LARGEFILE_SOURCE
11 #define _LARGEFILE_SOURCE
12 #endif
13
14 #ifndef _LARGEFILE64_SOURCE
15 #define _LARGEFILE64_SOURCE
16 #endif
17
18 #ifndef _GNU_SOURCE
19 #define _GNU_SOURCE
20 #endif
21
22 #include "config.h"
23 #include <ctype.h>
24 #include <dirent.h>
25 #include <endian.h>
26 #include <errno.h>
27 #include <fcntl.h>
28 #include <ftw.h>
29 #include <limits.h>
30 #include <mntent.h>
31 #include <stdio.h>
32 #include <stdlib.h>
33 #include <string.h>
34 #include <unistd.h>
35 #include <ext2fs/ext2_types.h>
36 #include <ext2fs/ext2fs.h>
37 #include <sys/ioctl.h>
38 #include <ext2fs/fiemap.h>
39 #include <sys/mman.h>
40 #include <sys/stat.h>
41 #include <sys/statfs.h>
42 #include <sys/vfs.h>
43
44 /* A relatively new ioctl interface ... */
45 #ifndef EXT4_IOC_MOVE_EXT
46 #define EXT4_IOC_MOVE_EXT _IOWR('f', 15, struct move_extent)
47 #endif
48
49 /* Macro functions */
50 #define PRINT_ERR_MSG(msg) fprintf(stderr, "%s\n", (msg))
51 #define IN_FTW_PRINT_ERR_MSG(msg) \
52 fprintf(stderr, "\t%s\t\t[ NG ]\n", (msg))
53 #define PRINT_FILE_NAME(file) fprintf(stderr, " \"%s\"\n", (file))
54 #define PRINT_ERR_MSG_WITH_ERRNO(msg) \
55 fprintf(stderr, "\t%s:%s\t[ NG ]\n", (msg), strerror(errno))
56 #define STATISTIC_ERR_MSG(msg) \
57 fprintf(stderr, "\t%s\n", (msg))
58 #define STATISTIC_ERR_MSG_WITH_ERRNO(msg) \
59 fprintf(stderr, "\t%s:%s\n", (msg), strerror(errno))
60 #define min(x, y) (((x) > (y)) ? (y) : (x))
61 #define CALC_SCORE(ratio) \
62 ((ratio) > 10 ? (80 + 20 * (ratio) / 100) : (8 * (ratio)))
63 /* Wrap up the free function */
64 #define FREE(tmp) \
65 do { \
66 if ((tmp) != NULL) \
67 free(tmp); \
68 } while (0) \
69 /* Insert list2 after list1 */
70 #define insert(list1, list2) \
71 do { \
72 list2->next = list1->next; \
73 list1->next->prev = list2; \
74 list2->prev = list1; \
75 list1->next = list2; \
76 } while (0)
77
78 /* To delete unused warning */
79 #ifdef __GNUC__
80 #define EXT2FS_ATTR(x) __attribute__(x)
81 #else
82 #define EXT2FS_ATTR(x)
83 #endif
84
85 /* The mode of defrag */
86 #define DETAIL 0x01
87 #define STATISTIC 0x02
88
89 #define DEVNAME 0
90 #define DIRNAME 1
91 #define FILENAME 2
92
93 #define FTW_OPEN_FD 2000
94
95 #define FS_EXT4 "ext4"
96 #define ROOT_UID 0
97
98 #define BOUND_SCORE 55
99 #define SHOW_FRAG_FILES 5
100
101 /* Magic number for ext4 */
102 #define EXT4_SUPER_MAGIC 0xEF53
103
104 /* Definition of flex_bg */
105 #define EXT4_FEATURE_INCOMPAT_FLEX_BG 0x0200
106
107 /* The following macro is used for ioctl FS_IOC_FIEMAP
108 * EXTENT_MAX_COUNT: the maximum number of extents for exchanging between
109 * kernel-space and user-space per ioctl
110 */
111 #define EXTENT_MAX_COUNT 512
112
113 /* The following macros are error message */
114 #define MSG_USAGE \
115 "Usage : e4defrag [-v] file...| directory...| device...\n\
116 : e4defrag -c file...| directory...| device...\n"
117
118 #define NGMSG_EXT4 "Filesystem is not ext4 filesystem"
119 #define NGMSG_FILE_EXTENT "Failed to get file extents"
120 #define NGMSG_FILE_INFO "Failed to get file information"
121 #define NGMSG_FILE_OPEN "Failed to open"
122 #define NGMSG_FILE_UNREG "File is not regular file"
123 #define NGMSG_LOST_FOUND "Can not process \"lost+found\""
124
125 /* Data type for filesystem-wide blocks number */
126 typedef unsigned long long ext4_fsblk_t;
127
128 struct fiemap_extent_data {
129 __u64 len; /* blocks count */
130 __u64 logical; /* start logical block number */
131 ext4_fsblk_t physical; /* start physical block number */
132 };
133
134 struct fiemap_extent_list {
135 struct fiemap_extent_list *prev;
136 struct fiemap_extent_list *next;
137 struct fiemap_extent_data data; /* extent belong to file */
138 };
139
140 struct fiemap_extent_group {
141 struct fiemap_extent_group *prev;
142 struct fiemap_extent_group *next;
143 __u64 len; /* length of this continuous region */
144 struct fiemap_extent_list *start; /* start ext */
145 struct fiemap_extent_list *end; /* end ext */
146 };
147
148 struct move_extent {
149 __s32 reserved; /* original file descriptor */
150 __u32 donor_fd; /* donor file descriptor */
151 __u64 orig_start; /* logical start offset in block for orig */
152 __u64 donor_start; /* logical start offset in block for donor */
153 __u64 len; /* block length to be moved */
154 __u64 moved_len; /* moved block length */
155 };
156
157 struct frag_statistic_ino {
158 int now_count; /* the file's extents count of before defrag */
159 int best_count; /* the best file's extents count */
160 __u64 size_per_ext; /* size(KB) per extent */
161 float ratio; /* the ratio of fragmentation */
162 char msg_buffer[PATH_MAX + 1]; /* pathname of the file */
163 };
164
165 static char lost_found_dir[PATH_MAX + 1];
166 static int block_size;
167 static int extents_before_defrag;
168 static int extents_after_defrag;
169 static int mode_flag;
170 static unsigned int current_uid;
171 static unsigned int defraged_file_count;
172 static unsigned int frag_files_before_defrag;
173 static unsigned int frag_files_after_defrag;
174 static unsigned int regular_count;
175 static unsigned int succeed_cnt;
176 static unsigned int total_count;
177 static __u8 log_groups_per_flex;
178 static __u32 blocks_per_group;
179 static __u32 feature_incompat;
180 static ext4_fsblk_t files_block_count;
181 static struct frag_statistic_ino frag_rank[SHOW_FRAG_FILES];
182
183
184 /*
185 * We prefer posix_fadvise64 when available, as it allows 64bit offset on
186 * 32bit systems
187 */
188 #if defined(HAVE_POSIX_FADVISE64)
189 #define posix_fadvise posix_fadvise64
190 #elif defined(HAVE_FADVISE64)
191 #define posix_fadvise fadvise64
192 #elif !defined(HAVE_POSIX_FADVISE)
193 #error posix_fadvise not available!
194 #endif
195
196 #ifndef HAVE_FALLOCATE64
197 #error fallocate64 not available!
198 #endif /* ! HAVE_FALLOCATE64 */
199
200 /*
201 * get_mount_point() - Get device's mount point.
202 *
203 * @devname: the device's name.
204 * @mount_point: the mount point.
205 * @dir_path_len: the length of directory.
206 */
get_mount_point(const char * devname,char * mount_point,int dir_path_len)207 static int get_mount_point(const char *devname, char *mount_point,
208 int dir_path_len)
209 {
210 /* Refer to /etc/mtab */
211 const char *mtab = MOUNTED;
212 FILE *fp = NULL;
213 struct mntent *mnt = NULL;
214 struct stat64 sb;
215
216 if (stat64(devname, &sb) < 0) {
217 perror(NGMSG_FILE_INFO);
218 PRINT_FILE_NAME(devname);
219 return -1;
220 }
221
222 fp = setmntent(mtab, "r");
223 if (fp == NULL) {
224 perror("Couldn't access /etc/mtab");
225 return -1;
226 }
227
228 while ((mnt = getmntent(fp)) != NULL) {
229 struct stat64 ms;
230
231 /*
232 * To handle device symlinks, we see if the
233 * device number matches, not the name
234 */
235 if (stat64(mnt->mnt_fsname, &ms) < 0)
236 continue;
237 if (sb.st_rdev != ms.st_rdev)
238 continue;
239
240 endmntent(fp);
241 if (strcmp(mnt->mnt_type, FS_EXT4) == 0) {
242 strncpy(mount_point, mnt->mnt_dir,
243 dir_path_len);
244 return 0;
245 }
246 PRINT_ERR_MSG(NGMSG_EXT4);
247 return -1;
248 }
249 endmntent(fp);
250 PRINT_ERR_MSG("Filesystem is not mounted");
251 return -1;
252 }
253
254 /*
255 * is_ext4() - Whether on an ext4 filesystem.
256 *
257 * @file: the file's name.
258 */
is_ext4(const char * file,char * devname)259 static int is_ext4(const char *file, char *devname)
260 {
261 int maxlen = 0;
262 int len, ret;
263 FILE *fp = NULL;
264 char *mnt_type = NULL;
265 /* Refer to /etc/mtab */
266 const char *mtab = MOUNTED;
267 char file_path[PATH_MAX + 1];
268 struct mntent *mnt = NULL;
269 struct statfs64 fsbuf;
270
271 /* Get full path */
272 if (realpath(file, file_path) == NULL) {
273 perror("Couldn't get full path");
274 PRINT_FILE_NAME(file);
275 return -1;
276 }
277
278 if (statfs64(file_path, &fsbuf) < 0) {
279 perror("Failed to get filesystem information");
280 PRINT_FILE_NAME(file);
281 return -1;
282 }
283
284 if (fsbuf.f_type != EXT4_SUPER_MAGIC) {
285 PRINT_ERR_MSG(NGMSG_EXT4);
286 return -1;
287 }
288
289 fp = setmntent(mtab, "r");
290 if (fp == NULL) {
291 perror("Couldn't access /etc/mtab");
292 return -1;
293 }
294
295 while ((mnt = getmntent(fp)) != NULL) {
296 if (mnt->mnt_fsname[0] != '/')
297 continue;
298 len = strlen(mnt->mnt_dir);
299 ret = memcmp(file_path, mnt->mnt_dir, len);
300 if (ret != 0)
301 continue;
302
303 if (maxlen >= len)
304 continue;
305
306 maxlen = len;
307
308 mnt_type = realloc(mnt_type, strlen(mnt->mnt_type) + 1);
309 if (mnt_type == NULL) {
310 endmntent(fp);
311 return -1;
312 }
313 memset(mnt_type, 0, strlen(mnt->mnt_type) + 1);
314 strncpy(mnt_type, mnt->mnt_type, strlen(mnt->mnt_type));
315 strncpy(lost_found_dir, mnt->mnt_dir, PATH_MAX);
316 strncpy(devname, mnt->mnt_fsname, strlen(mnt->mnt_fsname) + 1);
317 }
318
319 endmntent(fp);
320 if (mnt_type && strcmp(mnt_type, FS_EXT4) == 0) {
321 FREE(mnt_type);
322 return 0;
323 } else {
324 FREE(mnt_type);
325 PRINT_ERR_MSG(NGMSG_EXT4);
326 return -1;
327 }
328 }
329
330 /*
331 * calc_entry_counts() - Calculate file counts.
332 *
333 * @file: file name.
334 * @buf: file info.
335 * @flag: file type.
336 * @ftwbuf: the pointer of a struct FTW.
337 */
calc_entry_counts(const char * file EXT2FS_ATTR ((unused)),const struct stat64 * buf,int flag EXT2FS_ATTR ((unused)),struct FTW * ftwbuf EXT2FS_ATTR ((unused)))338 static int calc_entry_counts(const char *file EXT2FS_ATTR((unused)),
339 const struct stat64 *buf, int flag EXT2FS_ATTR((unused)),
340 struct FTW *ftwbuf EXT2FS_ATTR((unused)))
341 {
342 if (S_ISREG(buf->st_mode))
343 regular_count++;
344
345 total_count++;
346
347 return 0;
348 }
349
350 /*
351 * page_in_core() - Get information on whether pages are in core.
352 *
353 * @fd: defrag target file's descriptor.
354 * @defrag_data: data used for defrag.
355 * @vec: page state array.
356 * @page_num: page number.
357 */
page_in_core(int fd,struct move_extent defrag_data,unsigned char ** vec,unsigned int * page_num)358 static int page_in_core(int fd, struct move_extent defrag_data,
359 unsigned char **vec, unsigned int *page_num)
360 {
361 long pagesize;
362 void *page = NULL;
363 loff_t offset, end_offset, length;
364
365 if (vec == NULL || *vec != NULL)
366 return -1;
367
368 pagesize = sysconf(_SC_PAGESIZE);
369 if (pagesize < 0)
370 return -1;
371 /* In mmap, offset should be a multiple of the page size */
372 offset = (loff_t)defrag_data.orig_start * block_size;
373 length = (loff_t)defrag_data.len * block_size;
374 end_offset = offset + length;
375 /* Round the offset down to the nearest multiple of pagesize */
376 offset = (offset / pagesize) * pagesize;
377 length = end_offset - offset;
378
379 page = mmap(NULL, length, PROT_READ, MAP_SHARED, fd, offset);
380 if (page == MAP_FAILED)
381 return -1;
382
383 *page_num = 0;
384 *page_num = (length + pagesize - 1) / pagesize;
385 *vec = (unsigned char *)calloc(*page_num, 1);
386 if (*vec == NULL) {
387 munmap(page, length);
388 return -1;
389 }
390
391 /* Get information on whether pages are in core */
392 if (mincore(page, (size_t)length, *vec) == -1 ||
393 munmap(page, length) == -1) {
394 FREE(*vec);
395 return -1;
396 }
397
398 return 0;
399 }
400
401 /*
402 * defrag_fadvise() - Predeclare an access pattern for file data.
403 *
404 * @fd: defrag target file's descriptor.
405 * @defrag_data: data used for defrag.
406 * @vec: page state array.
407 * @page_num: page number.
408 */
defrag_fadvise(int fd,struct move_extent defrag_data,unsigned char * vec,unsigned int page_num)409 static int defrag_fadvise(int fd, struct move_extent defrag_data,
410 unsigned char *vec, unsigned int page_num)
411 {
412 int flag = 1;
413 long pagesize = sysconf(_SC_PAGESIZE);
414 int fadvise_flag = POSIX_FADV_DONTNEED;
415 int sync_flag = SYNC_FILE_RANGE_WAIT_BEFORE |
416 SYNC_FILE_RANGE_WRITE |
417 SYNC_FILE_RANGE_WAIT_AFTER;
418 unsigned int i;
419 loff_t offset;
420
421 if (pagesize < 1)
422 return -1;
423
424 offset = (loff_t)defrag_data.orig_start * block_size;
425 offset = (offset / pagesize) * pagesize;
426
427 #ifdef HAVE_SYNC_FILE_RANGE
428 /* Sync file for fadvise process */
429 if (sync_file_range(fd, offset,
430 (loff_t)pagesize * page_num, sync_flag) < 0)
431 return -1;
432 #endif
433
434 /* Try to release buffer cache which this process used,
435 * then other process can use the released buffer
436 */
437 for (i = 0; i < page_num; i++) {
438 if ((vec[i] & 0x1) == 0) {
439 offset += pagesize;
440 continue;
441 }
442 if (posix_fadvise(fd, offset, pagesize, fadvise_flag) < 0) {
443 if ((mode_flag & DETAIL) && flag) {
444 perror("\tFailed to fadvise");
445 flag = 0;
446 }
447 }
448 offset += pagesize;
449 }
450
451 return 0;
452 }
453
454 /*
455 * check_free_size() - Check if there's enough disk space.
456 *
457 * @fd: defrag target file's descriptor.
458 * @file: file name.
459 * @blk_count: file blocks.
460 */
check_free_size(int fd,const char * file,ext4_fsblk_t blk_count)461 static int check_free_size(int fd, const char *file, ext4_fsblk_t blk_count)
462 {
463 ext4_fsblk_t free_blk_count;
464 struct statfs64 fsbuf;
465
466 if (fstatfs64(fd, &fsbuf) < 0) {
467 if (mode_flag & DETAIL) {
468 PRINT_FILE_NAME(file);
469 PRINT_ERR_MSG_WITH_ERRNO(
470 "Failed to get filesystem information");
471 }
472 return -1;
473 }
474
475 /* Compute free space for root and normal user separately */
476 if (current_uid == ROOT_UID)
477 free_blk_count = fsbuf.f_bfree;
478 else
479 free_blk_count = fsbuf.f_bavail;
480
481 if (free_blk_count >= blk_count)
482 return 0;
483
484 return -ENOSPC;
485 }
486
487 /*
488 * file_frag_count() - Get file fragment count.
489 *
490 * @fd: defrag target file's descriptor.
491 */
file_frag_count(int fd)492 static int file_frag_count(int fd)
493 {
494 int ret;
495 struct fiemap fiemap_buf;
496
497 /* When fm_extent_count is 0,
498 * ioctl just get file fragment count.
499 */
500 memset(&fiemap_buf, 0, sizeof(struct fiemap));
501 fiemap_buf.fm_start = 0;
502 fiemap_buf.fm_length = FIEMAP_MAX_OFFSET;
503 fiemap_buf.fm_flags |= FIEMAP_FLAG_SYNC;
504
505 ret = ioctl(fd, FS_IOC_FIEMAP, &fiemap_buf);
506 if (ret < 0)
507 return ret;
508
509 return fiemap_buf.fm_mapped_extents;
510 }
511
512 /*
513 * file_check() - Check file's attributes.
514 *
515 * @fd: defrag target file's descriptor.
516 * @buf: a pointer of the struct stat64.
517 * @file: file name.
518 * @extents: file extents.
519 * @blk_count: file blocks.
520 */
file_check(int fd,const struct stat64 * buf,const char * file,int extents,ext4_fsblk_t blk_count)521 static int file_check(int fd, const struct stat64 *buf, const char *file,
522 int extents, ext4_fsblk_t blk_count)
523 {
524 int ret;
525 struct flock lock;
526
527 /* Write-lock check is more reliable */
528 lock.l_type = F_WRLCK;
529 lock.l_start = 0;
530 lock.l_whence = SEEK_SET;
531 lock.l_len = 0;
532
533 /* Free space */
534 ret = check_free_size(fd, file, blk_count);
535 if (ret < 0) {
536 if ((mode_flag & DETAIL) && ret == -ENOSPC) {
537 printf("\033[79;0H\033[K[%u/%u] \"%s\"\t\t"
538 " extents: %d -> %d\n", defraged_file_count,
539 total_count, file, extents, extents);
540 IN_FTW_PRINT_ERR_MSG(
541 "Defrag size is larger than filesystem's free space");
542 }
543 return -1;
544 }
545
546 /* Access authority */
547 if (current_uid != ROOT_UID &&
548 buf->st_uid != current_uid) {
549 if (mode_flag & DETAIL) {
550 printf("\033[79;0H\033[K[%u/%u] \"%s\"\t\t"
551 " extents: %d -> %d\n", defraged_file_count,
552 total_count, file, extents, extents);
553 IN_FTW_PRINT_ERR_MSG(
554 "File is not current user's file"
555 " or current user is not root");
556 }
557 return -1;
558 }
559
560 /* Lock status */
561 if (fcntl(fd, F_GETLK, &lock) < 0) {
562 if (mode_flag & DETAIL) {
563 PRINT_FILE_NAME(file);
564 PRINT_ERR_MSG_WITH_ERRNO(
565 "Failed to get lock information");
566 }
567 return -1;
568 } else if (lock.l_type != F_UNLCK) {
569 if (mode_flag & DETAIL) {
570 PRINT_FILE_NAME(file);
571 IN_FTW_PRINT_ERR_MSG("File has been locked");
572 }
573 return -1;
574 }
575
576 return 0;
577 }
578
579 /*
580 * insert_extent_by_logical() - Sequentially insert extent by logical.
581 *
582 * @ext_list_head: the head of logical extent list.
583 * @ext: the extent element which will be inserted.
584 */
insert_extent_by_logical(struct fiemap_extent_list ** ext_list_head,struct fiemap_extent_list * ext)585 static int insert_extent_by_logical(struct fiemap_extent_list **ext_list_head,
586 struct fiemap_extent_list *ext)
587 {
588 struct fiemap_extent_list *ext_list_tmp = *ext_list_head;
589
590 if (ext == NULL)
591 goto out;
592
593 /* First element */
594 if (*ext_list_head == NULL) {
595 (*ext_list_head) = ext;
596 (*ext_list_head)->prev = *ext_list_head;
597 (*ext_list_head)->next = *ext_list_head;
598 return 0;
599 }
600
601 if (ext->data.logical <= ext_list_tmp->data.logical) {
602 /* Insert before head */
603 if (ext_list_tmp->data.logical <
604 ext->data.logical + ext->data.len)
605 /* Overlap */
606 goto out;
607 /* Adjust head */
608 *ext_list_head = ext;
609 } else {
610 /* Insert into the middle or last of the list */
611 do {
612 if (ext->data.logical < ext_list_tmp->data.logical)
613 break;
614 ext_list_tmp = ext_list_tmp->next;
615 } while (ext_list_tmp != (*ext_list_head));
616 if (ext->data.logical <
617 ext_list_tmp->prev->data.logical +
618 ext_list_tmp->prev->data.len)
619 /* Overlap */
620 goto out;
621
622 if (ext_list_tmp != *ext_list_head &&
623 ext_list_tmp->data.logical <
624 ext->data.logical + ext->data.len)
625 /* Overlap */
626 goto out;
627 }
628 ext_list_tmp = ext_list_tmp->prev;
629 /* Insert "ext" after "ext_list_tmp" */
630 insert(ext_list_tmp, ext);
631 return 0;
632 out:
633 errno = EINVAL;
634 return -1;
635 }
636
637 /*
638 * insert_extent_by_physical() - Sequentially insert extent by physical.
639 *
640 * @ext_list_head: the head of physical extent list.
641 * @ext: the extent element which will be inserted.
642 */
insert_extent_by_physical(struct fiemap_extent_list ** ext_list_head,struct fiemap_extent_list * ext)643 static int insert_extent_by_physical(struct fiemap_extent_list **ext_list_head,
644 struct fiemap_extent_list *ext)
645 {
646 struct fiemap_extent_list *ext_list_tmp = *ext_list_head;
647
648 if (ext == NULL)
649 goto out;
650
651 /* First element */
652 if (*ext_list_head == NULL) {
653 (*ext_list_head) = ext;
654 (*ext_list_head)->prev = *ext_list_head;
655 (*ext_list_head)->next = *ext_list_head;
656 return 0;
657 }
658
659 if (ext->data.physical <= ext_list_tmp->data.physical) {
660 /* Insert before head */
661 if (ext_list_tmp->data.physical <
662 ext->data.physical + ext->data.len)
663 /* Overlap */
664 goto out;
665 /* Adjust head */
666 *ext_list_head = ext;
667 } else {
668 /* Insert into the middle or last of the list */
669 do {
670 if (ext->data.physical < ext_list_tmp->data.physical)
671 break;
672 ext_list_tmp = ext_list_tmp->next;
673 } while (ext_list_tmp != (*ext_list_head));
674 if (ext->data.physical <
675 ext_list_tmp->prev->data.physical +
676 ext_list_tmp->prev->data.len)
677 /* Overlap */
678 goto out;
679
680 if (ext_list_tmp != *ext_list_head &&
681 ext_list_tmp->data.physical <
682 ext->data.physical + ext->data.len)
683 /* Overlap */
684 goto out;
685 }
686 ext_list_tmp = ext_list_tmp->prev;
687 /* Insert "ext" after "ext_list_tmp" */
688 insert(ext_list_tmp, ext);
689 return 0;
690 out:
691 errno = EINVAL;
692 return -1;
693 }
694
695 /*
696 * insert_exts_group() - Insert a exts_group.
697 *
698 * @ext_group_head: the head of a exts_group list.
699 * @exts_group: the exts_group element which will be inserted.
700 */
insert_exts_group(struct fiemap_extent_group ** ext_group_head,struct fiemap_extent_group * exts_group)701 static int insert_exts_group(struct fiemap_extent_group **ext_group_head,
702 struct fiemap_extent_group *exts_group)
703 {
704 struct fiemap_extent_group *ext_group_tmp = NULL;
705
706 if (exts_group == NULL) {
707 errno = EINVAL;
708 return -1;
709 }
710
711 /* Initialize list */
712 if (*ext_group_head == NULL) {
713 (*ext_group_head) = exts_group;
714 (*ext_group_head)->prev = *ext_group_head;
715 (*ext_group_head)->next = *ext_group_head;
716 return 0;
717 }
718
719 ext_group_tmp = (*ext_group_head)->prev;
720 insert(ext_group_tmp, exts_group);
721
722 return 0;
723 }
724
725 /*
726 * join_extents() - Find continuous region(exts_group).
727 *
728 * @ext_list_head: the head of the extent list.
729 * @ext_group_head: the head of the target exts_group list.
730 */
join_extents(struct fiemap_extent_list * ext_list_head,struct fiemap_extent_group ** ext_group_head)731 static int join_extents(struct fiemap_extent_list *ext_list_head,
732 struct fiemap_extent_group **ext_group_head)
733 {
734 __u64 len = ext_list_head->data.len;
735 struct fiemap_extent_list *ext_list_start = ext_list_head;
736 struct fiemap_extent_list *ext_list_tmp = ext_list_head->next;
737
738 do {
739 struct fiemap_extent_group *ext_group_tmp = NULL;
740
741 /* This extent and previous extent are not continuous,
742 * so, all previous extents are treated as an extent group.
743 */
744 if ((ext_list_tmp->prev->data.logical +
745 ext_list_tmp->prev->data.len)
746 != ext_list_tmp->data.logical) {
747 ext_group_tmp =
748 malloc(sizeof(struct fiemap_extent_group));
749 if (ext_group_tmp == NULL)
750 return -1;
751
752 memset(ext_group_tmp, 0,
753 sizeof(struct fiemap_extent_group));
754 ext_group_tmp->len = len;
755 ext_group_tmp->start = ext_list_start;
756 ext_group_tmp->end = ext_list_tmp->prev;
757
758 if (insert_exts_group(ext_group_head,
759 ext_group_tmp) < 0) {
760 FREE(ext_group_tmp);
761 return -1;
762 }
763 ext_list_start = ext_list_tmp;
764 len = ext_list_tmp->data.len;
765 ext_list_tmp = ext_list_tmp->next;
766 continue;
767 }
768
769 /* This extent and previous extent are continuous,
770 * so, they belong to the same extent group, and we check
771 * if the next extent belongs to the same extent group.
772 */
773 len += ext_list_tmp->data.len;
774 ext_list_tmp = ext_list_tmp->next;
775 } while (ext_list_tmp != ext_list_head->next);
776
777 return 0;
778 }
779
780 /*
781 * get_file_extents() - Get file's extent list.
782 *
783 * @fd: defrag target file's descriptor.
784 * @ext_list_head: the head of the extent list.
785 */
get_file_extents(int fd,struct fiemap_extent_list ** ext_list_head)786 static int get_file_extents(int fd, struct fiemap_extent_list **ext_list_head)
787 {
788 __u32 i;
789 int ret;
790 int ext_buf_size, fie_buf_size;
791 __u64 pos = 0;
792 struct fiemap *fiemap_buf = NULL;
793 struct fiemap_extent *ext_buf = NULL;
794 struct fiemap_extent_list *ext_list = NULL;
795
796 /* Convert units, in bytes.
797 * Be careful : now, physical block number in extent is 48bit,
798 * and the maximum blocksize for ext4 is 4K(12bit),
799 * so there is no overflow, but in future it may be changed.
800 */
801
802 /* Alloc space for fiemap */
803 ext_buf_size = EXTENT_MAX_COUNT * sizeof(struct fiemap_extent);
804 fie_buf_size = sizeof(struct fiemap) + ext_buf_size;
805
806 fiemap_buf = malloc(fie_buf_size);
807 if (fiemap_buf == NULL)
808 return -1;
809
810 ext_buf = fiemap_buf->fm_extents;
811 memset(fiemap_buf, 0, fie_buf_size);
812 fiemap_buf->fm_length = FIEMAP_MAX_OFFSET;
813 fiemap_buf->fm_flags |= FIEMAP_FLAG_SYNC;
814 fiemap_buf->fm_extent_count = EXTENT_MAX_COUNT;
815
816 do {
817 fiemap_buf->fm_start = pos;
818 memset(ext_buf, 0, ext_buf_size);
819 ret = ioctl(fd, FS_IOC_FIEMAP, fiemap_buf);
820 if (ret < 0 || fiemap_buf->fm_mapped_extents == 0)
821 goto out;
822 for (i = 0; i < fiemap_buf->fm_mapped_extents; i++) {
823 ext_list = NULL;
824 ext_list = malloc(sizeof(struct fiemap_extent_list));
825 if (ext_list == NULL)
826 goto out;
827
828 ext_list->data.physical = ext_buf[i].fe_physical
829 / block_size;
830 ext_list->data.logical = ext_buf[i].fe_logical
831 / block_size;
832 ext_list->data.len = ext_buf[i].fe_length
833 / block_size;
834
835 ret = insert_extent_by_physical(
836 ext_list_head, ext_list);
837 if (ret < 0) {
838 FREE(ext_list);
839 goto out;
840 }
841 }
842 /* Record file's logical offset this time */
843 pos = ext_buf[EXTENT_MAX_COUNT-1].fe_logical +
844 ext_buf[EXTENT_MAX_COUNT-1].fe_length;
845 /*
846 * If fm_extents array has been filled and
847 * there are extents left, continue to cycle.
848 */
849 } while (fiemap_buf->fm_mapped_extents
850 == EXTENT_MAX_COUNT &&
851 !(ext_buf[EXTENT_MAX_COUNT-1].fe_flags
852 & FIEMAP_EXTENT_LAST));
853
854 FREE(fiemap_buf);
855 return 0;
856 out:
857 FREE(fiemap_buf);
858 return -1;
859 }
860
861 /*
862 * get_logical_count() - Get the file logical extents count.
863 *
864 * @logical_list_head: the head of the logical extent list.
865 */
get_logical_count(struct fiemap_extent_list * logical_list_head)866 static int get_logical_count(struct fiemap_extent_list *logical_list_head)
867 {
868 int ret = 0;
869 struct fiemap_extent_list *ext_list_tmp = logical_list_head;
870
871 do {
872 ret++;
873 ext_list_tmp = ext_list_tmp->next;
874 } while (ext_list_tmp != logical_list_head);
875
876 return ret;
877 }
878
879 /*
880 * get_physical_count() - Get the file physical extents count.
881 *
882 * @physical_list_head: the head of the physical extent list.
883 */
get_physical_count(struct fiemap_extent_list * physical_list_head)884 static int get_physical_count(struct fiemap_extent_list *physical_list_head)
885 {
886 int ret = 0;
887 struct fiemap_extent_list *ext_list_tmp = physical_list_head;
888
889 do {
890 if ((ext_list_tmp->data.physical + ext_list_tmp->data.len)
891 != ext_list_tmp->next->data.physical ||
892 (ext_list_tmp->data.logical + ext_list_tmp->data.len)
893 != ext_list_tmp->next->data.logical) {
894 /* This extent and next extent are not continuous. */
895 ret++;
896 }
897
898 ext_list_tmp = ext_list_tmp->next;
899 } while (ext_list_tmp != physical_list_head);
900
901 return ret;
902 }
903
904 /*
905 * change_physical_to_logical() - Change list from physical to logical.
906 *
907 * @physical_list_head: the head of physical extent list.
908 * @logical_list_head: the head of logical extent list.
909 */
change_physical_to_logical(struct fiemap_extent_list ** physical_list_head,struct fiemap_extent_list ** logical_list_head)910 static int change_physical_to_logical(
911 struct fiemap_extent_list **physical_list_head,
912 struct fiemap_extent_list **logical_list_head)
913 {
914 int ret;
915 struct fiemap_extent_list *ext_list_tmp = *physical_list_head;
916 struct fiemap_extent_list *ext_list_next = ext_list_tmp->next;
917
918 while (1) {
919 if (ext_list_tmp == ext_list_next) {
920 ret = insert_extent_by_logical(
921 logical_list_head, ext_list_tmp);
922 if (ret < 0)
923 return -1;
924
925 *physical_list_head = NULL;
926 break;
927 }
928
929 ext_list_tmp->prev->next = ext_list_tmp->next;
930 ext_list_tmp->next->prev = ext_list_tmp->prev;
931 *physical_list_head = ext_list_next;
932
933 ret = insert_extent_by_logical(
934 logical_list_head, ext_list_tmp);
935 if (ret < 0) {
936 FREE(ext_list_tmp);
937 return -1;
938 }
939 ext_list_tmp = ext_list_next;
940 ext_list_next = ext_list_next->next;
941 }
942
943 return 0;
944 }
945
946 /* get_file_blocks() - Get total file blocks.
947 *
948 * @ext_list_head: the extent list head of the target file
949 */
get_file_blocks(struct fiemap_extent_list * ext_list_head)950 static ext4_fsblk_t get_file_blocks(struct fiemap_extent_list *ext_list_head)
951 {
952 ext4_fsblk_t blk_count = 0;
953 struct fiemap_extent_list *ext_list_tmp = ext_list_head;
954
955 do {
956 blk_count += ext_list_tmp->data.len;
957 ext_list_tmp = ext_list_tmp->next;
958 } while (ext_list_tmp != ext_list_head);
959
960 return blk_count;
961 }
962
963 /*
964 * free_ext() - Free the extent list.
965 *
966 * @ext_list_head: the extent list head of which will be free.
967 */
free_ext(struct fiemap_extent_list * ext_list_head)968 static void free_ext(struct fiemap_extent_list *ext_list_head)
969 {
970 struct fiemap_extent_list *ext_list_tmp = NULL;
971
972 if (ext_list_head == NULL)
973 return;
974
975 while (ext_list_head->next != ext_list_head) {
976 ext_list_tmp = ext_list_head;
977 ext_list_head->prev->next = ext_list_head->next;
978 ext_list_head->next->prev = ext_list_head->prev;
979 ext_list_head = ext_list_head->next;
980 free(ext_list_tmp);
981 }
982 free(ext_list_head);
983 }
984
985 /*
986 * free_exts_group() - Free the exts_group.
987 *
988 * @*ext_group_head: the exts_group list head which will be free.
989 */
free_exts_group(struct fiemap_extent_group * ext_group_head)990 static void free_exts_group(struct fiemap_extent_group *ext_group_head)
991 {
992 struct fiemap_extent_group *ext_group_tmp = NULL;
993
994 if (ext_group_head == NULL)
995 return;
996
997 while (ext_group_head->next != ext_group_head) {
998 ext_group_tmp = ext_group_head;
999 ext_group_head->prev->next = ext_group_head->next;
1000 ext_group_head->next->prev = ext_group_head->prev;
1001 ext_group_head = ext_group_head->next;
1002 free(ext_group_tmp);
1003 }
1004 free(ext_group_head);
1005 }
1006
1007 /*
1008 * get_best_count() - Get the file best extents count.
1009 *
1010 * @block_count: the file's physical block count.
1011 */
get_best_count(ext4_fsblk_t block_count)1012 static int get_best_count(ext4_fsblk_t block_count)
1013 {
1014 int ret;
1015 unsigned int flex_bg_num;
1016
1017 /* Calcuate best extents count */
1018 if (feature_incompat & EXT4_FEATURE_INCOMPAT_FLEX_BG) {
1019 flex_bg_num = 1 << log_groups_per_flex;
1020 ret = ((block_count - 1) /
1021 ((ext4_fsblk_t)blocks_per_group *
1022 flex_bg_num)) + 1;
1023 } else
1024 ret = ((block_count - 1) / blocks_per_group) + 1;
1025
1026 return ret;
1027 }
1028
1029
1030 /*
1031 * file_statistic() - Get statistic info of the file's fragments.
1032 *
1033 * @file: the file's name.
1034 * @buf: the pointer of the struct stat64.
1035 * @flag: file type.
1036 * @ftwbuf: the pointer of a struct FTW.
1037 */
file_statistic(const char * file,const struct stat64 * buf,int flag EXT2FS_ATTR ((unused)),struct FTW * ftwbuf EXT2FS_ATTR ((unused)))1038 static int file_statistic(const char *file, const struct stat64 *buf,
1039 int flag EXT2FS_ATTR((unused)),
1040 struct FTW *ftwbuf EXT2FS_ATTR((unused)))
1041 {
1042 int fd;
1043 int ret;
1044 int now_ext_count, best_ext_count = 0, physical_ext_count;
1045 int i, j;
1046 __u64 size_per_ext = 0;
1047 float ratio = 0.0;
1048 ext4_fsblk_t blk_count = 0;
1049 char msg_buffer[PATH_MAX + 24];
1050 struct fiemap_extent_list *physical_list_head = NULL;
1051 struct fiemap_extent_list *logical_list_head = NULL;
1052
1053 defraged_file_count++;
1054
1055 if (mode_flag & DETAIL) {
1056 if (total_count == 1 && regular_count == 1)
1057 printf("<File>\n");
1058 else {
1059 printf("[%u/%u]", defraged_file_count, total_count);
1060 fflush(stdout);
1061 }
1062 }
1063 if (lost_found_dir[0] != '\0' &&
1064 !memcmp(file, lost_found_dir, strnlen(lost_found_dir, PATH_MAX))) {
1065 if (mode_flag & DETAIL) {
1066 PRINT_FILE_NAME(file);
1067 STATISTIC_ERR_MSG(NGMSG_LOST_FOUND);
1068 }
1069 return 0;
1070 }
1071
1072 if (!S_ISREG(buf->st_mode)) {
1073 if (mode_flag & DETAIL) {
1074 PRINT_FILE_NAME(file);
1075 STATISTIC_ERR_MSG(NGMSG_FILE_UNREG);
1076 }
1077 return 0;
1078 }
1079
1080 /* Access authority */
1081 if (current_uid != ROOT_UID &&
1082 buf->st_uid != current_uid) {
1083 if (mode_flag & DETAIL) {
1084 PRINT_FILE_NAME(file);
1085 STATISTIC_ERR_MSG(
1086 "File is not current user's file"
1087 " or current user is not root");
1088 }
1089 return 0;
1090 }
1091
1092 /* Empty file */
1093 if (buf->st_size == 0) {
1094 if (mode_flag & DETAIL) {
1095 PRINT_FILE_NAME(file);
1096 STATISTIC_ERR_MSG("File size is 0");
1097 }
1098 return 0;
1099 }
1100
1101 /* Has no blocks */
1102 if (buf->st_blocks == 0) {
1103 if (mode_flag & DETAIL) {
1104 PRINT_FILE_NAME(file);
1105 STATISTIC_ERR_MSG("File has no blocks");
1106 }
1107 return 0;
1108 }
1109
1110 fd = open64(file, O_RDONLY);
1111 if (fd < 0) {
1112 if (mode_flag & DETAIL) {
1113 PRINT_FILE_NAME(file);
1114 STATISTIC_ERR_MSG_WITH_ERRNO(NGMSG_FILE_OPEN);
1115 }
1116 return 0;
1117 }
1118
1119 /* Get file's physical extents */
1120 ret = get_file_extents(fd, &physical_list_head);
1121 if (ret < 0) {
1122 if (mode_flag & DETAIL) {
1123 PRINT_FILE_NAME(file);
1124 STATISTIC_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1125 }
1126 goto out;
1127 }
1128
1129 /* Get the count of file's continuous physical region */
1130 physical_ext_count = get_physical_count(physical_list_head);
1131
1132 /* Change list from physical to logical */
1133 ret = change_physical_to_logical(&physical_list_head,
1134 &logical_list_head);
1135 if (ret < 0) {
1136 if (mode_flag & DETAIL) {
1137 PRINT_FILE_NAME(file);
1138 STATISTIC_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1139 }
1140 goto out;
1141 }
1142
1143 /* Count file fragments before defrag */
1144 now_ext_count = get_logical_count(logical_list_head);
1145
1146 if (current_uid == ROOT_UID) {
1147 /* Calculate the size per extent */
1148 blk_count = get_file_blocks(logical_list_head);
1149
1150 best_ext_count = get_best_count(blk_count);
1151
1152 /* e4defrag rounds size_per_ext up to a block size boundary */
1153 size_per_ext = blk_count * (buf->st_blksize / 1024) /
1154 now_ext_count;
1155
1156 ratio = (float)(physical_ext_count - best_ext_count) * 100 /
1157 blk_count;
1158
1159 extents_before_defrag += now_ext_count;
1160 extents_after_defrag += best_ext_count;
1161 files_block_count += blk_count;
1162 }
1163
1164 if (total_count == 1 && regular_count == 1) {
1165 /* File only */
1166 if (mode_flag & DETAIL) {
1167 int count = 0;
1168 struct fiemap_extent_list *ext_list_tmp =
1169 logical_list_head;
1170
1171 /* Print extents info */
1172 do {
1173 count++;
1174 printf("[ext %d]:\tstart %llu:\tlogical "
1175 "%llu:\tlen %llu\n", count,
1176 ext_list_tmp->data.physical,
1177 ext_list_tmp->data.logical,
1178 ext_list_tmp->data.len);
1179 ext_list_tmp = ext_list_tmp->next;
1180 } while (ext_list_tmp != logical_list_head);
1181
1182 } else {
1183 printf("%-40s%10s/%-10s%9s\n",
1184 "<File>", "now", "best", "size/ext");
1185 if (current_uid == ROOT_UID) {
1186 if (strlen(file) > 40)
1187 printf("%s\n%50d/%-10d%6llu KB\n",
1188 file, now_ext_count,
1189 best_ext_count, size_per_ext);
1190 else
1191 printf("%-40s%10d/%-10d%6llu KB\n",
1192 file, now_ext_count,
1193 best_ext_count, size_per_ext);
1194 } else {
1195 if (strlen(file) > 40)
1196 printf("%s\n%50d/%-10s%7s\n",
1197 file, now_ext_count,
1198 "-", "-");
1199 else
1200 printf("%-40s%10d/%-10s%7s\n",
1201 file, now_ext_count,
1202 "-", "-");
1203 }
1204 }
1205 succeed_cnt++;
1206 goto out;
1207 }
1208
1209 if (mode_flag & DETAIL) {
1210 /* Print statistic info */
1211 sprintf(msg_buffer, "[%u/%u]%s",
1212 defraged_file_count, total_count, file);
1213 if (current_uid == ROOT_UID) {
1214 if (strlen(msg_buffer) > 40)
1215 printf("\033[79;0H\033[K%s\n"
1216 "%50d/%-10d%6llu KB\n",
1217 msg_buffer, now_ext_count,
1218 best_ext_count, size_per_ext);
1219 else
1220 printf("\033[79;0H\033[K%-40s"
1221 "%10d/%-10d%6llu KB\n",
1222 msg_buffer, now_ext_count,
1223 best_ext_count, size_per_ext);
1224 } else {
1225 if (strlen(msg_buffer) > 40)
1226 printf("\033[79;0H\033[K%s\n%50d/%-10s%7s\n",
1227 msg_buffer, now_ext_count,
1228 "-", "-");
1229 else
1230 printf("\033[79;0H\033[K%-40s%10d/%-10s%7s\n",
1231 msg_buffer, now_ext_count,
1232 "-", "-");
1233 }
1234 }
1235
1236 for (i = 0; i < SHOW_FRAG_FILES; i++) {
1237 if (ratio >= frag_rank[i].ratio) {
1238 for (j = SHOW_FRAG_FILES - 1; j > i; j--) {
1239 memset(&frag_rank[j], 0,
1240 sizeof(struct frag_statistic_ino));
1241 strncpy(frag_rank[j].msg_buffer,
1242 frag_rank[j - 1].msg_buffer,
1243 strnlen(frag_rank[j - 1].msg_buffer,
1244 PATH_MAX));
1245 frag_rank[j].now_count =
1246 frag_rank[j - 1].now_count;
1247 frag_rank[j].best_count =
1248 frag_rank[j - 1].best_count;
1249 frag_rank[j].size_per_ext =
1250 frag_rank[j - 1].size_per_ext;
1251 frag_rank[j].ratio =
1252 frag_rank[j - 1].ratio;
1253 }
1254 memset(&frag_rank[i], 0,
1255 sizeof(struct frag_statistic_ino));
1256 strncpy(frag_rank[i].msg_buffer, file,
1257 strnlen(file, PATH_MAX));
1258 frag_rank[i].now_count = now_ext_count;
1259 frag_rank[i].best_count = best_ext_count;
1260 frag_rank[i].size_per_ext = size_per_ext;
1261 frag_rank[i].ratio = ratio;
1262 break;
1263 }
1264 }
1265
1266 succeed_cnt++;
1267
1268 out:
1269 close(fd);
1270 free_ext(physical_list_head);
1271 free_ext(logical_list_head);
1272 return 0;
1273 }
1274
1275 /*
1276 * print_progress - Print defrag progress
1277 *
1278 * @file: file name.
1279 * @start: logical offset for defrag target file
1280 * @file_size: defrag target filesize
1281 */
print_progress(const char * file,loff_t start,loff_t file_size)1282 static void print_progress(const char *file, loff_t start, loff_t file_size)
1283 {
1284 int percent = (start * 100) / file_size;
1285 printf("\033[79;0H\033[K[%u/%u]%s:\t%3d%%",
1286 defraged_file_count, total_count, file, min(percent, 100));
1287 fflush(stdout);
1288
1289 return;
1290 }
1291
1292 /*
1293 * call_defrag() - Execute the defrag program.
1294 *
1295 * @fd: target file descriptor.
1296 * @donor_fd: donor file descriptor.
1297 * @file: target file name.
1298 * @buf: pointer of the struct stat64.
1299 * @ext_list_head: head of the extent list.
1300 */
call_defrag(int fd,int donor_fd,const char * file,const struct stat64 * buf,struct fiemap_extent_list * ext_list_head)1301 static int call_defrag(int fd, int donor_fd, const char *file,
1302 const struct stat64 *buf, struct fiemap_extent_list *ext_list_head)
1303 {
1304 loff_t start = 0;
1305 unsigned int page_num;
1306 unsigned char *vec = NULL;
1307 int defraged_ret = 0;
1308 int ret;
1309 struct move_extent move_data;
1310 struct fiemap_extent_list *ext_list_tmp = NULL;
1311
1312 memset(&move_data, 0, sizeof(struct move_extent));
1313 move_data.donor_fd = donor_fd;
1314
1315 /* Print defrag progress */
1316 print_progress(file, start, buf->st_size);
1317
1318 ext_list_tmp = ext_list_head;
1319 do {
1320 move_data.orig_start = ext_list_tmp->data.logical;
1321 /* Logical offset of orig and donor should be same */
1322 move_data.donor_start = move_data.orig_start;
1323 move_data.len = ext_list_tmp->data.len;
1324 move_data.moved_len = 0;
1325
1326 ret = page_in_core(fd, move_data, &vec, &page_num);
1327 if (ret < 0) {
1328 if (mode_flag & DETAIL) {
1329 printf("\n");
1330 PRINT_ERR_MSG_WITH_ERRNO(
1331 "Failed to get file map");
1332 } else {
1333 printf("\t[ NG ]\n");
1334 }
1335 return -1;
1336 }
1337
1338 /* EXT4_IOC_MOVE_EXT */
1339 defraged_ret =
1340 ioctl(fd, EXT4_IOC_MOVE_EXT, &move_data);
1341
1342 /* Free pages */
1343 ret = defrag_fadvise(fd, move_data, vec, page_num);
1344 if (vec) {
1345 free(vec);
1346 vec = NULL;
1347 }
1348 if (ret < 0) {
1349 if (mode_flag & DETAIL) {
1350 printf("\n");
1351 PRINT_ERR_MSG_WITH_ERRNO(
1352 "Failed to free page");
1353 } else {
1354 printf("\t[ NG ]\n");
1355 }
1356 return -1;
1357 }
1358
1359 if (defraged_ret < 0) {
1360 if (mode_flag & DETAIL) {
1361 printf("\n");
1362 PRINT_ERR_MSG_WITH_ERRNO(
1363 "Failed to defrag with "
1364 "EXT4_IOC_MOVE_EXT ioctl");
1365 if (errno == ENOTTY)
1366 printf("\tAt least 2.6.31-rc1 of "
1367 "vanilla kernel is required\n");
1368 } else {
1369 printf("\t[ NG ]\n");
1370 }
1371 return -1;
1372 }
1373 /* Adjust logical offset for next ioctl */
1374 move_data.orig_start += move_data.moved_len;
1375 move_data.donor_start = move_data.orig_start;
1376
1377 start = move_data.orig_start * buf->st_blksize;
1378
1379 /* Print defrag progress */
1380 print_progress(file, start, buf->st_size);
1381
1382 /* End of file */
1383 if (start >= buf->st_size)
1384 break;
1385
1386 ext_list_tmp = ext_list_tmp->next;
1387 } while (ext_list_tmp != ext_list_head);
1388
1389 return 0;
1390 }
1391
1392 /*
1393 * file_defrag() - Check file attributes and call ioctl to defrag.
1394 *
1395 * @file: the file's name.
1396 * @buf: the pointer of the struct stat64.
1397 * @flag: file type.
1398 * @ftwbuf: the pointer of a struct FTW.
1399 */
file_defrag(const char * file,const struct stat64 * buf,int flag EXT2FS_ATTR ((unused)),struct FTW * ftwbuf EXT2FS_ATTR ((unused)))1400 static int file_defrag(const char *file, const struct stat64 *buf,
1401 int flag EXT2FS_ATTR((unused)),
1402 struct FTW *ftwbuf EXT2FS_ATTR((unused)))
1403 {
1404 int fd;
1405 int donor_fd = -1;
1406 int ret;
1407 int best;
1408 int file_frags_start, file_frags_end;
1409 int orig_physical_cnt, donor_physical_cnt = 0;
1410 char tmp_inode_name[PATH_MAX + 8];
1411 ext4_fsblk_t blk_count = 0;
1412 struct fiemap_extent_list *orig_list_physical = NULL;
1413 struct fiemap_extent_list *orig_list_logical = NULL;
1414 struct fiemap_extent_list *donor_list_physical = NULL;
1415 struct fiemap_extent_list *donor_list_logical = NULL;
1416 struct fiemap_extent_group *orig_group_head = NULL;
1417 struct fiemap_extent_group *orig_group_tmp = NULL;
1418
1419 defraged_file_count++;
1420
1421 if (mode_flag & DETAIL) {
1422 printf("[%u/%u]", defraged_file_count, total_count);
1423 fflush(stdout);
1424 }
1425
1426 if (lost_found_dir[0] != '\0' &&
1427 !memcmp(file, lost_found_dir, strnlen(lost_found_dir, PATH_MAX))) {
1428 if (mode_flag & DETAIL) {
1429 PRINT_FILE_NAME(file);
1430 IN_FTW_PRINT_ERR_MSG(NGMSG_LOST_FOUND);
1431 }
1432 return 0;
1433 }
1434
1435 if (!S_ISREG(buf->st_mode)) {
1436 if (mode_flag & DETAIL) {
1437 PRINT_FILE_NAME(file);
1438 IN_FTW_PRINT_ERR_MSG(NGMSG_FILE_UNREG);
1439 }
1440 return 0;
1441 }
1442
1443 /* Empty file */
1444 if (buf->st_size == 0) {
1445 if (mode_flag & DETAIL) {
1446 PRINT_FILE_NAME(file);
1447 IN_FTW_PRINT_ERR_MSG("File size is 0");
1448 }
1449 return 0;
1450 }
1451
1452 /* Has no blocks */
1453 if (buf->st_blocks == 0) {
1454 if (mode_flag & DETAIL) {
1455 PRINT_FILE_NAME(file);
1456 STATISTIC_ERR_MSG("File has no blocks");
1457 }
1458 return 0;
1459 }
1460
1461 fd = open64(file, O_RDWR);
1462 if (fd < 0) {
1463 if (mode_flag & DETAIL) {
1464 PRINT_FILE_NAME(file);
1465 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_OPEN);
1466 }
1467 return 0;
1468 }
1469
1470 /* Get file's extents */
1471 ret = get_file_extents(fd, &orig_list_physical);
1472 if (ret < 0) {
1473 if (mode_flag & DETAIL) {
1474 PRINT_FILE_NAME(file);
1475 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1476 }
1477 goto out;
1478 }
1479
1480 /* Get the count of file's continuous physical region */
1481 orig_physical_cnt = get_physical_count(orig_list_physical);
1482
1483 /* Change list from physical to logical */
1484 ret = change_physical_to_logical(&orig_list_physical,
1485 &orig_list_logical);
1486 if (ret < 0) {
1487 if (mode_flag & DETAIL) {
1488 PRINT_FILE_NAME(file);
1489 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1490 }
1491 goto out;
1492 }
1493
1494 /* Count file fragments before defrag */
1495 file_frags_start = get_logical_count(orig_list_logical);
1496
1497 blk_count = get_file_blocks(orig_list_logical);
1498 if (file_check(fd, buf, file, file_frags_start, blk_count) < 0)
1499 goto out;
1500
1501 if (fsync(fd) < 0) {
1502 if (mode_flag & DETAIL) {
1503 PRINT_FILE_NAME(file);
1504 PRINT_ERR_MSG_WITH_ERRNO("Failed to sync(fsync)");
1505 }
1506 goto out;
1507 }
1508
1509 if (current_uid == ROOT_UID)
1510 best = get_best_count(blk_count);
1511 else
1512 best = 1;
1513
1514 if (file_frags_start <= best)
1515 goto check_improvement;
1516
1517 /* Combine extents to group */
1518 ret = join_extents(orig_list_logical, &orig_group_head);
1519 if (ret < 0) {
1520 if (mode_flag & DETAIL) {
1521 PRINT_FILE_NAME(file);
1522 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1523 }
1524 goto out;
1525 }
1526
1527 /* Create donor inode */
1528 memset(tmp_inode_name, 0, PATH_MAX + 8);
1529 sprintf(tmp_inode_name, "%.*s.defrag",
1530 (int)strnlen(file, PATH_MAX), file);
1531 donor_fd = open64(tmp_inode_name, O_WRONLY | O_CREAT | O_EXCL, S_IRUSR);
1532 if (donor_fd < 0) {
1533 if (mode_flag & DETAIL) {
1534 PRINT_FILE_NAME(file);
1535 if (errno == EEXIST)
1536 PRINT_ERR_MSG_WITH_ERRNO(
1537 "File is being defraged by other program");
1538 else
1539 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_OPEN);
1540 }
1541 goto out;
1542 }
1543
1544 /* Unlink donor inode */
1545 ret = unlink(tmp_inode_name);
1546 if (ret < 0) {
1547 if (mode_flag & DETAIL) {
1548 PRINT_FILE_NAME(file);
1549 PRINT_ERR_MSG_WITH_ERRNO("Failed to unlink");
1550 }
1551 goto out;
1552 }
1553
1554 /* Allocate space for donor inode */
1555 orig_group_tmp = orig_group_head;
1556 do {
1557 ret = fallocate64(donor_fd, 0,
1558 (loff_t)orig_group_tmp->start->data.logical * block_size,
1559 (loff_t)orig_group_tmp->len * block_size);
1560 if (ret < 0) {
1561 if (mode_flag & DETAIL) {
1562 PRINT_FILE_NAME(file);
1563 PRINT_ERR_MSG_WITH_ERRNO("Failed to fallocate");
1564 }
1565 goto out;
1566 }
1567
1568 orig_group_tmp = orig_group_tmp->next;
1569 } while (orig_group_tmp != orig_group_head);
1570
1571 /* Get donor inode's extents */
1572 ret = get_file_extents(donor_fd, &donor_list_physical);
1573 if (ret < 0) {
1574 if (mode_flag & DETAIL) {
1575 PRINT_FILE_NAME(file);
1576 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1577 }
1578 goto out;
1579 }
1580
1581 /* Calcuate donor inode's continuous physical region */
1582 donor_physical_cnt = get_physical_count(donor_list_physical);
1583
1584 /* Change donor extent list from physical to logical */
1585 ret = change_physical_to_logical(&donor_list_physical,
1586 &donor_list_logical);
1587 if (ret < 0) {
1588 if (mode_flag & DETAIL) {
1589 PRINT_FILE_NAME(file);
1590 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1591 }
1592 goto out;
1593 }
1594
1595 check_improvement:
1596 if (mode_flag & DETAIL) {
1597 if (file_frags_start != 1)
1598 frag_files_before_defrag++;
1599
1600 extents_before_defrag += file_frags_start;
1601 }
1602
1603 if (file_frags_start <= best ||
1604 orig_physical_cnt <= donor_physical_cnt) {
1605 printf("\033[79;0H\033[K[%u/%u]%s:\t%3d%%",
1606 defraged_file_count, total_count, file, 100);
1607 if (mode_flag & DETAIL)
1608 printf(" extents: %d -> %d",
1609 file_frags_start, file_frags_start);
1610
1611 printf("\t[ OK ]\n");
1612 succeed_cnt++;
1613
1614 if (file_frags_start != 1)
1615 frag_files_after_defrag++;
1616
1617 extents_after_defrag += file_frags_start;
1618 goto out;
1619 }
1620
1621 /* Defrag the file */
1622 ret = call_defrag(fd, donor_fd, file, buf, donor_list_logical);
1623
1624 /* Count file fragments after defrag and print extents info */
1625 if (mode_flag & DETAIL) {
1626 file_frags_end = file_frag_count(fd);
1627 if (file_frags_end < 0) {
1628 printf("\n");
1629 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_INFO);
1630 goto out;
1631 }
1632
1633 if (file_frags_end != 1)
1634 frag_files_after_defrag++;
1635
1636 extents_after_defrag += file_frags_end;
1637
1638 if (ret < 0)
1639 goto out;
1640
1641 printf(" extents: %d -> %d",
1642 file_frags_start, file_frags_end);
1643 fflush(stdout);
1644 }
1645
1646 if (ret < 0)
1647 goto out;
1648
1649 printf("\t[ OK ]\n");
1650 fflush(stdout);
1651 succeed_cnt++;
1652
1653 out:
1654 close(fd);
1655 if (donor_fd != -1)
1656 close(donor_fd);
1657 free_ext(orig_list_physical);
1658 free_ext(orig_list_logical);
1659 free_ext(donor_list_physical);
1660 free_exts_group(orig_group_head);
1661 return 0;
1662 }
1663
1664 /*
1665 * main() - Ext4 online defrag.
1666 *
1667 * @argc: the number of parameter.
1668 * @argv[]: the pointer array of parameter.
1669 */
main(int argc,char * argv[])1670 int main(int argc, char *argv[])
1671 {
1672 int opt;
1673 int i, j, ret = 0;
1674 int flags = FTW_PHYS | FTW_MOUNT;
1675 int arg_type = -1;
1676 int success_flag = 0;
1677 char dir_name[PATH_MAX + 1];
1678 char dev_name[PATH_MAX + 1];
1679 struct stat64 buf;
1680 ext2_filsys fs = NULL;
1681
1682 /* Parse arguments */
1683 if (argc == 1)
1684 goto out;
1685
1686 while ((opt = getopt(argc, argv, "vc")) != EOF) {
1687 switch (opt) {
1688 case 'v':
1689 mode_flag |= DETAIL;
1690 break;
1691 case 'c':
1692 mode_flag |= STATISTIC;
1693 break;
1694 default:
1695 goto out;
1696 }
1697 }
1698
1699 if (argc == optind)
1700 goto out;
1701
1702 current_uid = getuid();
1703
1704 /* Main process */
1705 for (i = optind; i < argc; i++) {
1706 succeed_cnt = 0;
1707 regular_count = 0;
1708 total_count = 0;
1709 frag_files_before_defrag = 0;
1710 frag_files_after_defrag = 0;
1711 extents_before_defrag = 0;
1712 extents_after_defrag = 0;
1713 defraged_file_count = 0;
1714 files_block_count = 0;
1715 blocks_per_group = 0;
1716 feature_incompat = 0;
1717 log_groups_per_flex = 0;
1718
1719 memset(dir_name, 0, PATH_MAX + 1);
1720 memset(dev_name, 0, PATH_MAX + 1);
1721 memset(lost_found_dir, 0, PATH_MAX + 1);
1722 memset(frag_rank, 0,
1723 sizeof(struct frag_statistic_ino) * SHOW_FRAG_FILES);
1724
1725 if ((mode_flag & STATISTIC) && i > optind)
1726 printf("\n");
1727
1728 #if BYTE_ORDER != BIG_ENDIAN && BYTE_ORDER != LITTLE_ENDIAN
1729 PRINT_ERR_MSG("Endian's type is not big/little endian");
1730 PRINT_FILE_NAME(argv[i]);
1731 continue;
1732 #endif
1733
1734 if (lstat64(argv[i], &buf) < 0) {
1735 perror(NGMSG_FILE_INFO);
1736 PRINT_FILE_NAME(argv[i]);
1737 continue;
1738 }
1739
1740 /* Handle i.e. lvm device symlinks */
1741 if (S_ISLNK(buf.st_mode)) {
1742 struct stat64 buf2;
1743
1744 if (stat64(argv[i], &buf2) == 0 &&
1745 S_ISBLK(buf2.st_mode))
1746 buf = buf2;
1747 }
1748
1749 if (S_ISBLK(buf.st_mode)) {
1750 /* Block device */
1751 strncpy(dev_name, argv[i], strnlen(argv[i], PATH_MAX));
1752 if (get_mount_point(argv[i], dir_name, PATH_MAX) < 0)
1753 continue;
1754 if (lstat64(dir_name, &buf) < 0) {
1755 perror(NGMSG_FILE_INFO);
1756 PRINT_FILE_NAME(argv[i]);
1757 continue;
1758 }
1759 arg_type = DEVNAME;
1760 if (!(mode_flag & STATISTIC))
1761 printf("ext4 defragmentation for device(%s)\n",
1762 argv[i]);
1763 } else if (S_ISDIR(buf.st_mode)) {
1764 /* Directory */
1765 if (access(argv[i], R_OK) < 0) {
1766 perror(argv[i]);
1767 continue;
1768 }
1769 arg_type = DIRNAME;
1770 strncpy(dir_name, argv[i], strnlen(argv[i], PATH_MAX));
1771 } else if (S_ISREG(buf.st_mode)) {
1772 /* Regular file */
1773 arg_type = FILENAME;
1774 } else {
1775 /* Irregular file */
1776 PRINT_ERR_MSG(NGMSG_FILE_UNREG);
1777 PRINT_FILE_NAME(argv[i]);
1778 continue;
1779 }
1780
1781 /* Set blocksize */
1782 block_size = buf.st_blksize;
1783
1784 /* For device case,
1785 * filesystem type checked in get_mount_point()
1786 */
1787 if (arg_type == FILENAME || arg_type == DIRNAME) {
1788 if (is_ext4(argv[i], dev_name) < 0)
1789 continue;
1790 if (realpath(argv[i], dir_name) == NULL) {
1791 perror("Couldn't get full path");
1792 PRINT_FILE_NAME(argv[i]);
1793 continue;
1794 }
1795 }
1796
1797 if (current_uid == ROOT_UID) {
1798 /* Get super block info */
1799 ret = ext2fs_open(dev_name, EXT2_FLAG_64BITS, 0,
1800 block_size, unix_io_manager, &fs);
1801 if (ret) {
1802 if (mode_flag & DETAIL)
1803 com_err(argv[1], ret,
1804 "while trying to open file system: %s",
1805 dev_name);
1806 continue;
1807 }
1808
1809 blocks_per_group = fs->super->s_blocks_per_group;
1810 feature_incompat = fs->super->s_feature_incompat;
1811 log_groups_per_flex = fs->super->s_log_groups_per_flex;
1812
1813 ext2fs_close_free(&fs);
1814 }
1815
1816 switch (arg_type) {
1817 int mount_dir_len = 0;
1818
1819 case DIRNAME:
1820 if (!(mode_flag & STATISTIC))
1821 printf("ext4 defragmentation "
1822 "for directory(%s)\n", argv[i]);
1823
1824 mount_dir_len = strnlen(lost_found_dir, PATH_MAX);
1825
1826 strncat(lost_found_dir, "/lost+found",
1827 PATH_MAX - strnlen(lost_found_dir, PATH_MAX));
1828
1829 /* Not the case("e4defrag mount_piont_dir") */
1830 if (dir_name[mount_dir_len] != '\0') {
1831 /*
1832 * "e4defrag mount_piont_dir/lost+found"
1833 * or "e4defrag mount_piont_dir/lost+found/"
1834 */
1835 if (strncmp(lost_found_dir, dir_name,
1836 strnlen(lost_found_dir,
1837 PATH_MAX)) == 0 &&
1838 (dir_name[strnlen(lost_found_dir,
1839 PATH_MAX)] == '\0' ||
1840 dir_name[strnlen(lost_found_dir,
1841 PATH_MAX)] == '/')) {
1842 PRINT_ERR_MSG(NGMSG_LOST_FOUND);
1843 PRINT_FILE_NAME(argv[i]);
1844 continue;
1845 }
1846
1847 /* "e4defrag mount_piont_dir/else_dir" */
1848 memset(lost_found_dir, 0, PATH_MAX + 1);
1849 }
1850 case DEVNAME:
1851 if (arg_type == DEVNAME) {
1852 strncpy(lost_found_dir, dir_name,
1853 strnlen(dir_name, PATH_MAX));
1854 strncat(lost_found_dir, "/lost+found/",
1855 PATH_MAX - strnlen(lost_found_dir,
1856 PATH_MAX));
1857 }
1858
1859 nftw64(dir_name, calc_entry_counts, FTW_OPEN_FD, flags);
1860
1861 if (mode_flag & STATISTIC) {
1862 if (mode_flag & DETAIL)
1863 printf("%-40s%10s/%-10s%9s\n",
1864 "<File>", "now", "best", "size/ext");
1865
1866 if (!(mode_flag & DETAIL) &&
1867 current_uid != ROOT_UID) {
1868 printf(" Done.\n");
1869 success_flag = 1;
1870 continue;
1871 }
1872
1873 nftw64(dir_name, file_statistic,
1874 FTW_OPEN_FD, flags);
1875
1876 if (succeed_cnt != 0 &&
1877 current_uid == ROOT_UID) {
1878 if (mode_flag & DETAIL)
1879 printf("\n");
1880 printf("%-40s%10s/%-10s%9s\n",
1881 "<Fragmented files>", "now",
1882 "best", "size/ext");
1883 for (j = 0; j < SHOW_FRAG_FILES; j++) {
1884 if (strlen(frag_rank[j].
1885 msg_buffer) > 37) {
1886 printf("%d. %s\n%50d/"
1887 "%-10d%6llu KB\n",
1888 j + 1,
1889 frag_rank[j].msg_buffer,
1890 frag_rank[j].now_count,
1891 frag_rank[j].best_count,
1892 frag_rank[j].
1893 size_per_ext);
1894 } else if (strlen(frag_rank[j].
1895 msg_buffer) > 0) {
1896 printf("%d. %-37s%10d/"
1897 "%-10d%6llu KB\n",
1898 j + 1,
1899 frag_rank[j].msg_buffer,
1900 frag_rank[j].now_count,
1901 frag_rank[j].best_count,
1902 frag_rank[j].
1903 size_per_ext);
1904 } else
1905 break;
1906 }
1907 }
1908 break;
1909 }
1910 /* File tree walk */
1911 nftw64(dir_name, file_defrag, FTW_OPEN_FD, flags);
1912 printf("\n\tSuccess:\t\t\t[ %u/%u ]\n", succeed_cnt,
1913 total_count);
1914 printf("\tFailure:\t\t\t[ %u/%u ]\n",
1915 total_count - succeed_cnt, total_count);
1916 if (mode_flag & DETAIL) {
1917 printf("\tTotal extents:\t\t\t%4d->%d\n",
1918 extents_before_defrag,
1919 extents_after_defrag);
1920 printf("\tFragmented percentage:\t\t"
1921 "%3llu%%->%llu%%\n",
1922 !regular_count ? 0 :
1923 ((unsigned long long)
1924 frag_files_before_defrag * 100) /
1925 regular_count,
1926 !regular_count ? 0 :
1927 ((unsigned long long)
1928 frag_files_after_defrag * 100) /
1929 regular_count);
1930 }
1931 break;
1932 case FILENAME:
1933 total_count = 1;
1934 regular_count = 1;
1935 strncat(lost_found_dir, "/lost+found/",
1936 PATH_MAX - strnlen(lost_found_dir,
1937 PATH_MAX));
1938 if (strncmp(lost_found_dir, dir_name,
1939 strnlen(lost_found_dir,
1940 PATH_MAX)) == 0) {
1941 PRINT_ERR_MSG(NGMSG_LOST_FOUND);
1942 PRINT_FILE_NAME(argv[i]);
1943 continue;
1944 }
1945
1946 if (mode_flag & STATISTIC) {
1947 file_statistic(argv[i], &buf, FTW_F, NULL);
1948 break;
1949 } else
1950 printf("ext4 defragmentation for %s\n",
1951 argv[i]);
1952 /* Defrag single file process */
1953 file_defrag(argv[i], &buf, FTW_F, NULL);
1954 if (succeed_cnt != 0)
1955 printf(" Success:\t\t\t[1/1]\n");
1956 else
1957 printf(" Success:\t\t\t[0/1]\n");
1958
1959 break;
1960 }
1961
1962 if (succeed_cnt != 0)
1963 success_flag = 1;
1964 if (mode_flag & STATISTIC) {
1965 if (current_uid != ROOT_UID) {
1966 printf(" Done.\n");
1967 continue;
1968 }
1969
1970 if (!succeed_cnt) {
1971 if (mode_flag & DETAIL)
1972 printf("\n");
1973
1974 if (arg_type == DEVNAME)
1975 printf(" In this device(%s), "
1976 "none can be defragmented.\n", argv[i]);
1977 else if (arg_type == DIRNAME)
1978 printf(" In this directory(%s), "
1979 "none can be defragmented.\n", argv[i]);
1980 else
1981 printf(" This file(%s) "
1982 "can't be defragmented.\n", argv[i]);
1983 } else {
1984 float files_ratio = 0.0;
1985 float score = 0.0;
1986 __u64 size_per_ext = files_block_count *
1987 (buf.st_blksize / 1024) /
1988 extents_before_defrag;
1989 files_ratio = (float)(extents_before_defrag -
1990 extents_after_defrag) *
1991 100 / files_block_count;
1992 score = CALC_SCORE(files_ratio);
1993 printf("\n Total/best extents\t\t\t\t%d/%d\n"
1994 " Average size per extent"
1995 "\t\t\t%llu KB\n"
1996 " Fragmentation score\t\t\t\t%.0f\n",
1997 extents_before_defrag,
1998 extents_after_defrag,
1999 size_per_ext, score);
2000 printf(" [0-30 no problem:"
2001 " 31-55 a little bit fragmented:"
2002 " 56- needs defrag]\n");
2003
2004 if (arg_type == DEVNAME)
2005 printf(" This device (%s) ", argv[i]);
2006 else if (arg_type == DIRNAME)
2007 printf(" This directory (%s) ",
2008 argv[i]);
2009 else
2010 printf(" This file (%s) ", argv[i]);
2011
2012 if (score > BOUND_SCORE)
2013 printf("needs defragmentation.\n");
2014 else
2015 printf("does not need "
2016 "defragmentation.\n");
2017 }
2018 printf(" Done.\n");
2019 }
2020
2021 }
2022
2023 if (success_flag)
2024 return 0;
2025
2026 exit(1);
2027
2028 out:
2029 printf(MSG_USAGE);
2030 exit(1);
2031 }
2032
2033