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 #include "../version.h"
45
46 /* A relatively new ioctl interface ... */
47 #ifndef EXT4_IOC_MOVE_EXT
48 #define EXT4_IOC_MOVE_EXT _IOWR('f', 15, struct move_extent)
49 #endif
50
51 /* Macro functions */
52 #define PRINT_ERR_MSG(msg) fprintf(stderr, "%s\n", (msg))
53 #define IN_FTW_PRINT_ERR_MSG(msg) \
54 fprintf(stderr, "\t%s\t\t[ NG ]\n", (msg))
55 #define PRINT_FILE_NAME(file) fprintf(stderr, " \"%s\"\n", (file))
56 #define PRINT_ERR_MSG_WITH_ERRNO(msg) \
57 fprintf(stderr, "\t%s:%s\t[ NG ]\n", (msg), strerror(errno))
58 #define STATISTIC_ERR_MSG(msg) \
59 fprintf(stderr, "\t%s\n", (msg))
60 #define STATISTIC_ERR_MSG_WITH_ERRNO(msg) \
61 fprintf(stderr, "\t%s:%s\n", (msg), strerror(errno))
62 #define min(x, y) (((x) > (y)) ? (y) : (x))
63 #define CALC_SCORE(ratio) \
64 ((ratio) > 10 ? (80 + 20 * (ratio) / 100) : (8 * (ratio)))
65 /* Wrap up the free function */
66 #define FREE(tmp) \
67 do { \
68 if ((tmp) != NULL) \
69 free(tmp); \
70 } while (0) \
71 /* Insert list2 after list1 */
72 #define insert(list1, list2) \
73 do { \
74 list2->next = list1->next; \
75 list1->next->prev = list2; \
76 list2->prev = list1; \
77 list1->next = list2; \
78 } while (0)
79
80 /* To delete unused warning */
81 #ifdef __GNUC__
82 #define EXT2FS_ATTR(x) __attribute__(x)
83 #else
84 #define EXT2FS_ATTR(x)
85 #endif
86
87 /* The mode of defrag */
88 #define DETAIL 0x01
89 #define STATISTIC 0x02
90
91 #define DEVNAME 0
92 #define DIRNAME 1
93 #define FILENAME 2
94
95 #define FTW_OPEN_FD 2000
96
97 #define FS_EXT4 "ext4"
98 #define ROOT_UID 0
99
100 #define BOUND_SCORE 55
101 #define SHOW_FRAG_FILES 5
102
103 /* Magic number for ext4 */
104 #define EXT4_SUPER_MAGIC 0xEF53
105
106 /* Definition of flex_bg */
107 #define EXT4_FEATURE_INCOMPAT_FLEX_BG 0x0200
108
109 /* The following macro is used for ioctl FS_IOC_FIEMAP
110 * EXTENT_MAX_COUNT: the maximum number of extents for exchanging between
111 * kernel-space and user-space per ioctl
112 */
113 #define EXTENT_MAX_COUNT 512
114
115 /* The following macros are error message */
116 #define MSG_USAGE \
117 "Usage : e4defrag [-v] file...| directory...| device...\n\
118 : e4defrag -c file...| directory...| device...\n"
119
120 #define NGMSG_EXT4 "Filesystem is not ext4 filesystem"
121 #define NGMSG_FILE_EXTENT "Failed to get file extents"
122 #define NGMSG_FILE_INFO "Failed to get file information"
123 #define NGMSG_FILE_OPEN "Failed to open"
124 #define NGMSG_FILE_UNREG "File is not regular file"
125 #define NGMSG_LOST_FOUND "Can not process \"lost+found\""
126
127 /* Data type for filesystem-wide blocks number */
128 typedef unsigned long long ext4_fsblk_t;
129
130 struct fiemap_extent_data {
131 __u64 len; /* blocks count */
132 __u64 logical; /* start logical block number */
133 ext4_fsblk_t physical; /* start physical block number */
134 };
135
136 struct fiemap_extent_list {
137 struct fiemap_extent_list *prev;
138 struct fiemap_extent_list *next;
139 struct fiemap_extent_data data; /* extent belong to file */
140 };
141
142 struct fiemap_extent_group {
143 struct fiemap_extent_group *prev;
144 struct fiemap_extent_group *next;
145 __u64 len; /* length of this continuous region */
146 struct fiemap_extent_list *start; /* start ext */
147 struct fiemap_extent_list *end; /* end ext */
148 };
149
150 struct move_extent {
151 __s32 reserved; /* original file descriptor */
152 __u32 donor_fd; /* donor file descriptor */
153 __u64 orig_start; /* logical start offset in block for orig */
154 __u64 donor_start; /* logical start offset in block for donor */
155 __u64 len; /* block length to be moved */
156 __u64 moved_len; /* moved block length */
157 };
158
159 struct frag_statistic_ino {
160 int now_count; /* the file's extents count of before defrag */
161 int best_count; /* the best file's extents count */
162 __u64 size_per_ext; /* size(KB) per extent */
163 float ratio; /* the ratio of fragmentation */
164 char msg_buffer[PATH_MAX + 1]; /* pathname of the file */
165 };
166
167 static char lost_found_dir[PATH_MAX + 1];
168 static int block_size;
169 static int extents_before_defrag;
170 static int extents_after_defrag;
171 static int mode_flag;
172 static unsigned int current_uid;
173 static unsigned int defraged_file_count;
174 static unsigned int frag_files_before_defrag;
175 static unsigned int frag_files_after_defrag;
176 static unsigned int regular_count;
177 static unsigned int succeed_cnt;
178 static unsigned int total_count;
179 static __u8 log_groups_per_flex;
180 static __u32 blocks_per_group;
181 static __u32 feature_incompat;
182 static ext4_fsblk_t files_block_count;
183 static struct frag_statistic_ino frag_rank[SHOW_FRAG_FILES];
184
185
186 /*
187 * We prefer posix_fadvise64 when available, as it allows 64bit offset on
188 * 32bit systems
189 */
190 #if defined(HAVE_POSIX_FADVISE64)
191 #define posix_fadvise posix_fadvise64
192 #elif defined(HAVE_FADVISE64)
193 #define posix_fadvise fadvise64
194 #elif !defined(HAVE_POSIX_FADVISE)
195 #error posix_fadvise not available!
196 #endif
197
198 #ifndef HAVE_FALLOCATE64
199 #error fallocate64 not available!
200 #endif /* ! HAVE_FALLOCATE64 */
201
202 /*
203 * get_mount_point() - Get device's mount point.
204 *
205 * @devname: the device's name.
206 * @mount_point: the mount point.
207 * @dir_path_len: the length of directory.
208 */
get_mount_point(const char * devname,char * mount_point,int dir_path_len)209 static int get_mount_point(const char *devname, char *mount_point,
210 int dir_path_len)
211 {
212 /* Refer to /etc/mtab */
213 const char *mtab = MOUNTED;
214 FILE *fp = NULL;
215 struct mntent *mnt = NULL;
216 struct stat64 sb;
217
218 if (stat64(devname, &sb) < 0) {
219 perror(NGMSG_FILE_INFO);
220 PRINT_FILE_NAME(devname);
221 return -1;
222 }
223
224 fp = setmntent(mtab, "r");
225 if (fp == NULL) {
226 perror("Couldn't access /etc/mtab");
227 return -1;
228 }
229
230 while ((mnt = getmntent(fp)) != NULL) {
231 struct stat64 ms;
232
233 /*
234 * To handle device symlinks, we see if the
235 * device number matches, not the name
236 */
237 if (stat64(mnt->mnt_fsname, &ms) < 0)
238 continue;
239 if (sb.st_rdev != ms.st_rdev)
240 continue;
241
242 endmntent(fp);
243 if (strcmp(mnt->mnt_type, FS_EXT4) == 0) {
244 strncpy(mount_point, mnt->mnt_dir,
245 dir_path_len);
246 return 0;
247 }
248 PRINT_ERR_MSG(NGMSG_EXT4);
249 return -1;
250 }
251 endmntent(fp);
252 PRINT_ERR_MSG("Filesystem is not mounted");
253 return -1;
254 }
255
256 /*
257 * is_ext4() - Whether on an ext4 filesystem.
258 *
259 * @file: the file's name.
260 */
is_ext4(const char * file,char * devname)261 static int is_ext4(const char *file, char *devname)
262 {
263 int maxlen = 0;
264 int len, ret;
265 FILE *fp = NULL;
266 char *mnt_type = NULL;
267 /* Refer to /etc/mtab */
268 const char *mtab = MOUNTED;
269 char file_path[PATH_MAX + 1];
270 struct mntent *mnt = NULL;
271 struct statfs64 fsbuf;
272
273 /* Get full path */
274 if (realpath(file, file_path) == NULL) {
275 perror("Couldn't get full path");
276 PRINT_FILE_NAME(file);
277 return -1;
278 }
279
280 if (statfs64(file_path, &fsbuf) < 0) {
281 perror("Failed to get filesystem information");
282 PRINT_FILE_NAME(file);
283 return -1;
284 }
285
286 if (fsbuf.f_type != EXT4_SUPER_MAGIC) {
287 PRINT_ERR_MSG(NGMSG_EXT4);
288 return -1;
289 }
290
291 fp = setmntent(mtab, "r");
292 if (fp == NULL) {
293 perror("Couldn't access /etc/mtab");
294 return -1;
295 }
296
297 while ((mnt = getmntent(fp)) != NULL) {
298 if (mnt->mnt_fsname[0] != '/')
299 continue;
300 len = strlen(mnt->mnt_dir);
301 ret = memcmp(file_path, mnt->mnt_dir, len);
302 if (ret != 0)
303 continue;
304
305 if (maxlen >= len)
306 continue;
307
308 maxlen = len;
309
310 mnt_type = realloc(mnt_type, strlen(mnt->mnt_type) + 1);
311 if (mnt_type == NULL) {
312 endmntent(fp);
313 return -1;
314 }
315 memset(mnt_type, 0, strlen(mnt->mnt_type) + 1);
316 strncpy(mnt_type, mnt->mnt_type, strlen(mnt->mnt_type));
317 strncpy(lost_found_dir, mnt->mnt_dir, PATH_MAX);
318 strncpy(devname, mnt->mnt_fsname, strlen(mnt->mnt_fsname) + 1);
319 }
320
321 endmntent(fp);
322 if (mnt_type && strcmp(mnt_type, FS_EXT4) == 0) {
323 FREE(mnt_type);
324 return 0;
325 } else {
326 FREE(mnt_type);
327 PRINT_ERR_MSG(NGMSG_EXT4);
328 return -1;
329 }
330 }
331
332 /*
333 * calc_entry_counts() - Calculate file counts.
334 *
335 * @file: file name.
336 * @buf: file info.
337 * @flag: file type.
338 * @ftwbuf: the pointer of a struct FTW.
339 */
calc_entry_counts(const char * file EXT2FS_ATTR ((unused)),const struct stat64 * buf,int flag EXT2FS_ATTR ((unused)),struct FTW * ftwbuf EXT2FS_ATTR ((unused)))340 static int calc_entry_counts(const char *file EXT2FS_ATTR((unused)),
341 const struct stat64 *buf, int flag EXT2FS_ATTR((unused)),
342 struct FTW *ftwbuf EXT2FS_ATTR((unused)))
343 {
344 if (S_ISREG(buf->st_mode))
345 regular_count++;
346
347 total_count++;
348
349 return 0;
350 }
351
352 /*
353 * page_in_core() - Get information on whether pages are in core.
354 *
355 * @fd: defrag target file's descriptor.
356 * @defrag_data: data used for defrag.
357 * @vec: page state array.
358 * @page_num: page number.
359 */
page_in_core(int fd,struct move_extent defrag_data,unsigned char ** vec,unsigned int * page_num)360 static int page_in_core(int fd, struct move_extent defrag_data,
361 unsigned char **vec, unsigned int *page_num)
362 {
363 long pagesize;
364 void *page = NULL;
365 loff_t offset, end_offset, length;
366
367 if (vec == NULL || *vec != NULL)
368 return -1;
369
370 pagesize = sysconf(_SC_PAGESIZE);
371 if (pagesize < 0)
372 return -1;
373 /* In mmap, offset should be a multiple of the page size */
374 offset = (loff_t)defrag_data.orig_start * block_size;
375 length = (loff_t)defrag_data.len * block_size;
376 end_offset = offset + length;
377 /* Round the offset down to the nearest multiple of pagesize */
378 offset = (offset / pagesize) * pagesize;
379 length = end_offset - offset;
380
381 page = mmap(NULL, length, PROT_READ, MAP_SHARED, fd, offset);
382 if (page == MAP_FAILED)
383 return -1;
384
385 *page_num = 0;
386 *page_num = (length + pagesize - 1) / pagesize;
387 *vec = (unsigned char *)calloc(*page_num, 1);
388 if (*vec == NULL) {
389 munmap(page, length);
390 return -1;
391 }
392
393 /* Get information on whether pages are in core */
394 if (mincore(page, (size_t)length, *vec) == -1 ||
395 munmap(page, length) == -1) {
396 FREE(*vec);
397 return -1;
398 }
399
400 return 0;
401 }
402
403 /*
404 * defrag_fadvise() - Predeclare an access pattern for file data.
405 *
406 * @fd: defrag target file's descriptor.
407 * @defrag_data: data used for defrag.
408 * @vec: page state array.
409 * @page_num: page number.
410 */
defrag_fadvise(int fd,struct move_extent defrag_data,unsigned char * vec,unsigned int page_num)411 static int defrag_fadvise(int fd, struct move_extent defrag_data,
412 unsigned char *vec, unsigned int page_num)
413 {
414 int flag = 1;
415 long pagesize = sysconf(_SC_PAGESIZE);
416 int fadvise_flag = POSIX_FADV_DONTNEED;
417 int sync_flag = SYNC_FILE_RANGE_WAIT_BEFORE |
418 SYNC_FILE_RANGE_WRITE |
419 SYNC_FILE_RANGE_WAIT_AFTER;
420 unsigned int i;
421 loff_t offset;
422
423 if (pagesize < 1)
424 return -1;
425
426 offset = (loff_t)defrag_data.orig_start * block_size;
427 offset = (offset / pagesize) * pagesize;
428
429 #ifdef HAVE_SYNC_FILE_RANGE
430 /* Sync file for fadvise process */
431 if (sync_file_range(fd, offset,
432 (loff_t)pagesize * page_num, sync_flag) < 0)
433 return -1;
434 #endif
435
436 /* Try to release buffer cache which this process used,
437 * then other process can use the released buffer
438 */
439 for (i = 0; i < page_num; i++) {
440 if ((vec[i] & 0x1) == 0) {
441 offset += pagesize;
442 continue;
443 }
444 if (posix_fadvise(fd, offset, pagesize, fadvise_flag) < 0) {
445 if ((mode_flag & DETAIL) && flag) {
446 perror("\tFailed to fadvise");
447 flag = 0;
448 }
449 }
450 offset += pagesize;
451 }
452
453 return 0;
454 }
455
456 /*
457 * check_free_size() - Check if there's enough disk space.
458 *
459 * @fd: defrag target file's descriptor.
460 * @file: file name.
461 * @blk_count: file blocks.
462 */
check_free_size(int fd,const char * file,ext4_fsblk_t blk_count)463 static int check_free_size(int fd, const char *file, ext4_fsblk_t blk_count)
464 {
465 ext4_fsblk_t free_blk_count;
466 struct statfs64 fsbuf;
467
468 if (fstatfs64(fd, &fsbuf) < 0) {
469 if (mode_flag & DETAIL) {
470 PRINT_FILE_NAME(file);
471 PRINT_ERR_MSG_WITH_ERRNO(
472 "Failed to get filesystem information");
473 }
474 return -1;
475 }
476
477 /* Compute free space for root and normal user separately */
478 if (current_uid == ROOT_UID)
479 free_blk_count = fsbuf.f_bfree;
480 else
481 free_blk_count = fsbuf.f_bavail;
482
483 if (free_blk_count >= blk_count)
484 return 0;
485
486 return -ENOSPC;
487 }
488
489 /*
490 * file_frag_count() - Get file fragment count.
491 *
492 * @fd: defrag target file's descriptor.
493 */
file_frag_count(int fd)494 static int file_frag_count(int fd)
495 {
496 int ret;
497 struct fiemap fiemap_buf;
498
499 /* When fm_extent_count is 0,
500 * ioctl just get file fragment count.
501 */
502 memset(&fiemap_buf, 0, sizeof(struct fiemap));
503 fiemap_buf.fm_start = 0;
504 fiemap_buf.fm_length = FIEMAP_MAX_OFFSET;
505 fiemap_buf.fm_flags |= FIEMAP_FLAG_SYNC;
506
507 ret = ioctl(fd, FS_IOC_FIEMAP, &fiemap_buf);
508 if (ret < 0)
509 return ret;
510
511 return fiemap_buf.fm_mapped_extents;
512 }
513
514 /*
515 * file_check() - Check file's attributes.
516 *
517 * @fd: defrag target file's descriptor.
518 * @buf: a pointer of the struct stat64.
519 * @file: file name.
520 * @extents: file extents.
521 * @blk_count: file blocks.
522 */
file_check(int fd,const struct stat64 * buf,const char * file,int extents,ext4_fsblk_t blk_count)523 static int file_check(int fd, const struct stat64 *buf, const char *file,
524 int extents, ext4_fsblk_t blk_count)
525 {
526 int ret;
527 struct flock lock;
528
529 /* Write-lock check is more reliable */
530 lock.l_type = F_WRLCK;
531 lock.l_start = 0;
532 lock.l_whence = SEEK_SET;
533 lock.l_len = 0;
534
535 /* Free space */
536 ret = check_free_size(fd, file, blk_count);
537 if (ret < 0) {
538 if ((mode_flag & DETAIL) && ret == -ENOSPC) {
539 printf("\033[79;0H\033[K[%u/%u] \"%s\"\t\t"
540 " extents: %d -> %d\n", defraged_file_count,
541 total_count, file, extents, extents);
542 IN_FTW_PRINT_ERR_MSG(
543 "Defrag size is larger than filesystem's free space");
544 }
545 return -1;
546 }
547
548 /* Access authority */
549 if (current_uid != ROOT_UID &&
550 buf->st_uid != current_uid) {
551 if (mode_flag & DETAIL) {
552 printf("\033[79;0H\033[K[%u/%u] \"%s\"\t\t"
553 " extents: %d -> %d\n", defraged_file_count,
554 total_count, file, extents, extents);
555 IN_FTW_PRINT_ERR_MSG(
556 "File is not current user's file"
557 " or current user is not root");
558 }
559 return -1;
560 }
561
562 /* Lock status */
563 if (fcntl(fd, F_GETLK, &lock) < 0) {
564 if (mode_flag & DETAIL) {
565 PRINT_FILE_NAME(file);
566 PRINT_ERR_MSG_WITH_ERRNO(
567 "Failed to get lock information");
568 }
569 return -1;
570 } else if (lock.l_type != F_UNLCK) {
571 if (mode_flag & DETAIL) {
572 PRINT_FILE_NAME(file);
573 IN_FTW_PRINT_ERR_MSG("File has been locked");
574 }
575 return -1;
576 }
577
578 return 0;
579 }
580
581 /*
582 * insert_extent_by_logical() - Sequentially insert extent by logical.
583 *
584 * @ext_list_head: the head of logical extent list.
585 * @ext: the extent element which will be inserted.
586 */
insert_extent_by_logical(struct fiemap_extent_list ** ext_list_head,struct fiemap_extent_list * ext)587 static int insert_extent_by_logical(struct fiemap_extent_list **ext_list_head,
588 struct fiemap_extent_list *ext)
589 {
590 struct fiemap_extent_list *ext_list_tmp = *ext_list_head;
591
592 if (ext == NULL)
593 goto out;
594
595 /* First element */
596 if (*ext_list_head == NULL) {
597 (*ext_list_head) = ext;
598 (*ext_list_head)->prev = *ext_list_head;
599 (*ext_list_head)->next = *ext_list_head;
600 return 0;
601 }
602
603 if (ext->data.logical <= ext_list_tmp->data.logical) {
604 /* Insert before head */
605 if (ext_list_tmp->data.logical <
606 ext->data.logical + ext->data.len)
607 /* Overlap */
608 goto out;
609 /* Adjust head */
610 *ext_list_head = ext;
611 } else {
612 /* Insert into the middle or last of the list */
613 do {
614 if (ext->data.logical < ext_list_tmp->data.logical)
615 break;
616 ext_list_tmp = ext_list_tmp->next;
617 } while (ext_list_tmp != (*ext_list_head));
618 if (ext->data.logical <
619 ext_list_tmp->prev->data.logical +
620 ext_list_tmp->prev->data.len)
621 /* Overlap */
622 goto out;
623
624 if (ext_list_tmp != *ext_list_head &&
625 ext_list_tmp->data.logical <
626 ext->data.logical + ext->data.len)
627 /* Overlap */
628 goto out;
629 }
630 ext_list_tmp = ext_list_tmp->prev;
631 /* Insert "ext" after "ext_list_tmp" */
632 insert(ext_list_tmp, ext);
633 return 0;
634 out:
635 errno = EINVAL;
636 return -1;
637 }
638
639 /*
640 * insert_extent_by_physical() - Sequentially insert extent by physical.
641 *
642 * @ext_list_head: the head of physical extent list.
643 * @ext: the extent element which will be inserted.
644 */
insert_extent_by_physical(struct fiemap_extent_list ** ext_list_head,struct fiemap_extent_list * ext)645 static int insert_extent_by_physical(struct fiemap_extent_list **ext_list_head,
646 struct fiemap_extent_list *ext)
647 {
648 struct fiemap_extent_list *ext_list_tmp = *ext_list_head;
649
650 if (ext == NULL)
651 goto out;
652
653 /* First element */
654 if (*ext_list_head == NULL) {
655 (*ext_list_head) = ext;
656 (*ext_list_head)->prev = *ext_list_head;
657 (*ext_list_head)->next = *ext_list_head;
658 return 0;
659 }
660
661 if (ext->data.physical <= ext_list_tmp->data.physical) {
662 /* Insert before head */
663 if (ext_list_tmp->data.physical <
664 ext->data.physical + ext->data.len)
665 /* Overlap */
666 goto out;
667 /* Adjust head */
668 *ext_list_head = ext;
669 } else {
670 /* Insert into the middle or last of the list */
671 do {
672 if (ext->data.physical < ext_list_tmp->data.physical)
673 break;
674 ext_list_tmp = ext_list_tmp->next;
675 } while (ext_list_tmp != (*ext_list_head));
676 if (ext->data.physical <
677 ext_list_tmp->prev->data.physical +
678 ext_list_tmp->prev->data.len)
679 /* Overlap */
680 goto out;
681
682 if (ext_list_tmp != *ext_list_head &&
683 ext_list_tmp->data.physical <
684 ext->data.physical + ext->data.len)
685 /* Overlap */
686 goto out;
687 }
688 ext_list_tmp = ext_list_tmp->prev;
689 /* Insert "ext" after "ext_list_tmp" */
690 insert(ext_list_tmp, ext);
691 return 0;
692 out:
693 errno = EINVAL;
694 return -1;
695 }
696
697 /*
698 * insert_exts_group() - Insert a exts_group.
699 *
700 * @ext_group_head: the head of a exts_group list.
701 * @exts_group: the exts_group element which will be inserted.
702 */
insert_exts_group(struct fiemap_extent_group ** ext_group_head,struct fiemap_extent_group * exts_group)703 static int insert_exts_group(struct fiemap_extent_group **ext_group_head,
704 struct fiemap_extent_group *exts_group)
705 {
706 struct fiemap_extent_group *ext_group_tmp = NULL;
707
708 if (exts_group == NULL) {
709 errno = EINVAL;
710 return -1;
711 }
712
713 /* Initialize list */
714 if (*ext_group_head == NULL) {
715 (*ext_group_head) = exts_group;
716 (*ext_group_head)->prev = *ext_group_head;
717 (*ext_group_head)->next = *ext_group_head;
718 return 0;
719 }
720
721 ext_group_tmp = (*ext_group_head)->prev;
722 insert(ext_group_tmp, exts_group);
723
724 return 0;
725 }
726
727 /*
728 * join_extents() - Find continuous region(exts_group).
729 *
730 * @ext_list_head: the head of the extent list.
731 * @ext_group_head: the head of the target exts_group list.
732 */
join_extents(struct fiemap_extent_list * ext_list_head,struct fiemap_extent_group ** ext_group_head)733 static int join_extents(struct fiemap_extent_list *ext_list_head,
734 struct fiemap_extent_group **ext_group_head)
735 {
736 __u64 len = ext_list_head->data.len;
737 struct fiemap_extent_list *ext_list_start = ext_list_head;
738 struct fiemap_extent_list *ext_list_tmp = ext_list_head->next;
739
740 do {
741 struct fiemap_extent_group *ext_group_tmp = NULL;
742
743 /* This extent and previous extent are not continuous,
744 * so, all previous extents are treated as an extent group.
745 */
746 if ((ext_list_tmp->prev->data.logical +
747 ext_list_tmp->prev->data.len)
748 != ext_list_tmp->data.logical) {
749 ext_group_tmp =
750 malloc(sizeof(struct fiemap_extent_group));
751 if (ext_group_tmp == NULL)
752 return -1;
753
754 memset(ext_group_tmp, 0,
755 sizeof(struct fiemap_extent_group));
756 ext_group_tmp->len = len;
757 ext_group_tmp->start = ext_list_start;
758 ext_group_tmp->end = ext_list_tmp->prev;
759
760 if (insert_exts_group(ext_group_head,
761 ext_group_tmp) < 0) {
762 FREE(ext_group_tmp);
763 return -1;
764 }
765 ext_list_start = ext_list_tmp;
766 len = ext_list_tmp->data.len;
767 ext_list_tmp = ext_list_tmp->next;
768 continue;
769 }
770
771 /* This extent and previous extent are continuous,
772 * so, they belong to the same extent group, and we check
773 * if the next extent belongs to the same extent group.
774 */
775 len += ext_list_tmp->data.len;
776 ext_list_tmp = ext_list_tmp->next;
777 } while (ext_list_tmp != ext_list_head->next);
778
779 return 0;
780 }
781
782 /*
783 * get_file_extents() - Get file's extent list.
784 *
785 * @fd: defrag target file's descriptor.
786 * @ext_list_head: the head of the extent list.
787 */
get_file_extents(int fd,struct fiemap_extent_list ** ext_list_head)788 static int get_file_extents(int fd, struct fiemap_extent_list **ext_list_head)
789 {
790 __u32 i;
791 int ret;
792 int ext_buf_size, fie_buf_size;
793 __u64 pos = 0;
794 struct fiemap *fiemap_buf = NULL;
795 struct fiemap_extent *ext_buf = NULL;
796 struct fiemap_extent_list *ext_list = NULL;
797
798 /* Convert units, in bytes.
799 * Be careful : now, physical block number in extent is 48bit,
800 * and the maximum blocksize for ext4 is 4K(12bit),
801 * so there is no overflow, but in future it may be changed.
802 */
803
804 /* Alloc space for fiemap */
805 ext_buf_size = EXTENT_MAX_COUNT * sizeof(struct fiemap_extent);
806 fie_buf_size = sizeof(struct fiemap) + ext_buf_size;
807
808 fiemap_buf = malloc(fie_buf_size);
809 if (fiemap_buf == NULL)
810 return -1;
811
812 ext_buf = fiemap_buf->fm_extents;
813 memset(fiemap_buf, 0, fie_buf_size);
814 fiemap_buf->fm_length = FIEMAP_MAX_OFFSET;
815 fiemap_buf->fm_flags |= FIEMAP_FLAG_SYNC;
816 fiemap_buf->fm_extent_count = EXTENT_MAX_COUNT;
817
818 do {
819 fiemap_buf->fm_start = pos;
820 memset(ext_buf, 0, ext_buf_size);
821 ret = ioctl(fd, FS_IOC_FIEMAP, fiemap_buf);
822 if (ret < 0 || fiemap_buf->fm_mapped_extents == 0)
823 goto out;
824 for (i = 0; i < fiemap_buf->fm_mapped_extents; i++) {
825 ext_list = NULL;
826 ext_list = malloc(sizeof(struct fiemap_extent_list));
827 if (ext_list == NULL)
828 goto out;
829
830 ext_list->data.physical = ext_buf[i].fe_physical
831 / block_size;
832 ext_list->data.logical = ext_buf[i].fe_logical
833 / block_size;
834 ext_list->data.len = ext_buf[i].fe_length
835 / block_size;
836
837 ret = insert_extent_by_physical(
838 ext_list_head, ext_list);
839 if (ret < 0) {
840 FREE(ext_list);
841 goto out;
842 }
843 }
844 /* Record file's logical offset this time */
845 pos = ext_buf[EXTENT_MAX_COUNT-1].fe_logical +
846 ext_buf[EXTENT_MAX_COUNT-1].fe_length;
847 /*
848 * If fm_extents array has been filled and
849 * there are extents left, continue to cycle.
850 */
851 } while (fiemap_buf->fm_mapped_extents
852 == EXTENT_MAX_COUNT &&
853 !(ext_buf[EXTENT_MAX_COUNT-1].fe_flags
854 & FIEMAP_EXTENT_LAST));
855
856 FREE(fiemap_buf);
857 return 0;
858 out:
859 FREE(fiemap_buf);
860 return -1;
861 }
862
863 /*
864 * get_logical_count() - Get the file logical extents count.
865 *
866 * @logical_list_head: the head of the logical extent list.
867 */
get_logical_count(struct fiemap_extent_list * logical_list_head)868 static int get_logical_count(struct fiemap_extent_list *logical_list_head)
869 {
870 int ret = 0;
871 struct fiemap_extent_list *ext_list_tmp = logical_list_head;
872
873 do {
874 ret++;
875 ext_list_tmp = ext_list_tmp->next;
876 } while (ext_list_tmp != logical_list_head);
877
878 return ret;
879 }
880
881 /*
882 * get_physical_count() - Get the file physical extents count.
883 *
884 * @physical_list_head: the head of the physical extent list.
885 */
get_physical_count(struct fiemap_extent_list * physical_list_head)886 static int get_physical_count(struct fiemap_extent_list *physical_list_head)
887 {
888 int ret = 0;
889 struct fiemap_extent_list *ext_list_tmp = physical_list_head;
890
891 do {
892 if ((ext_list_tmp->data.physical + ext_list_tmp->data.len)
893 != ext_list_tmp->next->data.physical ||
894 (ext_list_tmp->data.logical + ext_list_tmp->data.len)
895 != ext_list_tmp->next->data.logical) {
896 /* This extent and next extent are not continuous. */
897 ret++;
898 }
899
900 ext_list_tmp = ext_list_tmp->next;
901 } while (ext_list_tmp != physical_list_head);
902
903 return ret;
904 }
905
906 /*
907 * change_physical_to_logical() - Change list from physical to logical.
908 *
909 * @physical_list_head: the head of physical extent list.
910 * @logical_list_head: the head of logical extent list.
911 */
change_physical_to_logical(struct fiemap_extent_list ** physical_list_head,struct fiemap_extent_list ** logical_list_head)912 static int change_physical_to_logical(
913 struct fiemap_extent_list **physical_list_head,
914 struct fiemap_extent_list **logical_list_head)
915 {
916 int ret;
917 struct fiemap_extent_list *ext_list_tmp = *physical_list_head;
918 struct fiemap_extent_list *ext_list_next = ext_list_tmp->next;
919
920 while (1) {
921 if (ext_list_tmp == ext_list_next) {
922 ret = insert_extent_by_logical(
923 logical_list_head, ext_list_tmp);
924 if (ret < 0)
925 return -1;
926
927 *physical_list_head = NULL;
928 break;
929 }
930
931 ext_list_tmp->prev->next = ext_list_tmp->next;
932 ext_list_tmp->next->prev = ext_list_tmp->prev;
933 *physical_list_head = ext_list_next;
934
935 ret = insert_extent_by_logical(
936 logical_list_head, ext_list_tmp);
937 if (ret < 0) {
938 FREE(ext_list_tmp);
939 return -1;
940 }
941 ext_list_tmp = ext_list_next;
942 ext_list_next = ext_list_next->next;
943 }
944
945 return 0;
946 }
947
948 /* get_file_blocks() - Get total file blocks.
949 *
950 * @ext_list_head: the extent list head of the target file
951 */
get_file_blocks(struct fiemap_extent_list * ext_list_head)952 static ext4_fsblk_t get_file_blocks(struct fiemap_extent_list *ext_list_head)
953 {
954 ext4_fsblk_t blk_count = 0;
955 struct fiemap_extent_list *ext_list_tmp = ext_list_head;
956
957 do {
958 blk_count += ext_list_tmp->data.len;
959 ext_list_tmp = ext_list_tmp->next;
960 } while (ext_list_tmp != ext_list_head);
961
962 return blk_count;
963 }
964
965 /*
966 * free_ext() - Free the extent list.
967 *
968 * @ext_list_head: the extent list head of which will be free.
969 */
free_ext(struct fiemap_extent_list * ext_list_head)970 static void free_ext(struct fiemap_extent_list *ext_list_head)
971 {
972 struct fiemap_extent_list *ext_list_tmp = NULL;
973
974 if (ext_list_head == NULL)
975 return;
976
977 while (ext_list_head->next != ext_list_head) {
978 ext_list_tmp = ext_list_head;
979 ext_list_head->prev->next = ext_list_head->next;
980 ext_list_head->next->prev = ext_list_head->prev;
981 ext_list_head = ext_list_head->next;
982 free(ext_list_tmp);
983 }
984 free(ext_list_head);
985 }
986
987 /*
988 * free_exts_group() - Free the exts_group.
989 *
990 * @*ext_group_head: the exts_group list head which will be free.
991 */
free_exts_group(struct fiemap_extent_group * ext_group_head)992 static void free_exts_group(struct fiemap_extent_group *ext_group_head)
993 {
994 struct fiemap_extent_group *ext_group_tmp = NULL;
995
996 if (ext_group_head == NULL)
997 return;
998
999 while (ext_group_head->next != ext_group_head) {
1000 ext_group_tmp = ext_group_head;
1001 ext_group_head->prev->next = ext_group_head->next;
1002 ext_group_head->next->prev = ext_group_head->prev;
1003 ext_group_head = ext_group_head->next;
1004 free(ext_group_tmp);
1005 }
1006 free(ext_group_head);
1007 }
1008
1009 /*
1010 * get_best_count() - Get the file best extents count.
1011 *
1012 * @block_count: the file's physical block count.
1013 */
get_best_count(ext4_fsblk_t block_count)1014 static int get_best_count(ext4_fsblk_t block_count)
1015 {
1016 int ret;
1017 unsigned int flex_bg_num;
1018
1019 /* Calculate best extents count */
1020 if (feature_incompat & EXT4_FEATURE_INCOMPAT_FLEX_BG) {
1021 flex_bg_num = 1 << log_groups_per_flex;
1022 ret = ((block_count - 1) /
1023 ((ext4_fsblk_t)blocks_per_group *
1024 flex_bg_num)) + 1;
1025 } else
1026 ret = ((block_count - 1) / blocks_per_group) + 1;
1027
1028 return ret;
1029 }
1030
1031
1032 /*
1033 * file_statistic() - Get statistic info of the file's fragments.
1034 *
1035 * @file: the file's name.
1036 * @buf: the pointer of the struct stat64.
1037 * @flag: file type.
1038 * @ftwbuf: the pointer of a struct FTW.
1039 */
file_statistic(const char * file,const struct stat64 * buf,int flag EXT2FS_ATTR ((unused)),struct FTW * ftwbuf EXT2FS_ATTR ((unused)))1040 static int file_statistic(const char *file, const struct stat64 *buf,
1041 int flag EXT2FS_ATTR((unused)),
1042 struct FTW *ftwbuf EXT2FS_ATTR((unused)))
1043 {
1044 int fd;
1045 int ret;
1046 int now_ext_count, best_ext_count = 0, physical_ext_count;
1047 int i, j;
1048 __u64 size_per_ext = 0;
1049 float ratio = 0.0;
1050 ext4_fsblk_t blk_count = 0;
1051 char msg_buffer[PATH_MAX + 24];
1052 struct fiemap_extent_list *physical_list_head = NULL;
1053 struct fiemap_extent_list *logical_list_head = NULL;
1054
1055 defraged_file_count++;
1056
1057 if (mode_flag & DETAIL) {
1058 if (total_count == 1 && regular_count == 1)
1059 printf("<File>\n");
1060 else {
1061 printf("[%u/%u]", defraged_file_count, total_count);
1062 fflush(stdout);
1063 }
1064 }
1065 if (lost_found_dir[0] != '\0' &&
1066 !memcmp(file, lost_found_dir, strnlen(lost_found_dir, PATH_MAX))) {
1067 if (mode_flag & DETAIL) {
1068 PRINT_FILE_NAME(file);
1069 STATISTIC_ERR_MSG(NGMSG_LOST_FOUND);
1070 }
1071 return 0;
1072 }
1073
1074 if (!S_ISREG(buf->st_mode)) {
1075 if (mode_flag & DETAIL) {
1076 PRINT_FILE_NAME(file);
1077 STATISTIC_ERR_MSG(NGMSG_FILE_UNREG);
1078 }
1079 return 0;
1080 }
1081
1082 /* Access authority */
1083 if (current_uid != ROOT_UID &&
1084 buf->st_uid != current_uid) {
1085 if (mode_flag & DETAIL) {
1086 PRINT_FILE_NAME(file);
1087 STATISTIC_ERR_MSG(
1088 "File is not current user's file"
1089 " or current user is not root");
1090 }
1091 return 0;
1092 }
1093
1094 /* Empty file */
1095 if (buf->st_size == 0) {
1096 if (mode_flag & DETAIL) {
1097 PRINT_FILE_NAME(file);
1098 STATISTIC_ERR_MSG("File size is 0");
1099 }
1100 return 0;
1101 }
1102
1103 /* Has no blocks */
1104 if (buf->st_blocks == 0) {
1105 if (mode_flag & DETAIL) {
1106 PRINT_FILE_NAME(file);
1107 STATISTIC_ERR_MSG("File has no blocks");
1108 }
1109 return 0;
1110 }
1111
1112 fd = open64(file, O_RDONLY);
1113 if (fd < 0) {
1114 if (mode_flag & DETAIL) {
1115 PRINT_FILE_NAME(file);
1116 STATISTIC_ERR_MSG_WITH_ERRNO(NGMSG_FILE_OPEN);
1117 }
1118 return 0;
1119 }
1120
1121 /* Get file's physical extents */
1122 ret = get_file_extents(fd, &physical_list_head);
1123 if (ret < 0) {
1124 if (mode_flag & DETAIL) {
1125 PRINT_FILE_NAME(file);
1126 STATISTIC_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1127 }
1128 goto out;
1129 }
1130
1131 /* Get the count of file's continuous physical region */
1132 physical_ext_count = get_physical_count(physical_list_head);
1133
1134 /* Change list from physical to logical */
1135 ret = change_physical_to_logical(&physical_list_head,
1136 &logical_list_head);
1137 if (ret < 0) {
1138 if (mode_flag & DETAIL) {
1139 PRINT_FILE_NAME(file);
1140 STATISTIC_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1141 }
1142 goto out;
1143 }
1144
1145 /* Count file fragments before defrag */
1146 now_ext_count = get_logical_count(logical_list_head);
1147
1148 if (current_uid == ROOT_UID) {
1149 /* Calculate the size per extent */
1150 blk_count = get_file_blocks(logical_list_head);
1151
1152 best_ext_count = get_best_count(blk_count);
1153
1154 /* e4defrag rounds size_per_ext up to a block size boundary */
1155 size_per_ext = blk_count * (buf->st_blksize / 1024) /
1156 now_ext_count;
1157
1158 ratio = (float)(physical_ext_count - best_ext_count) * 100 /
1159 blk_count;
1160
1161 extents_before_defrag += now_ext_count;
1162 extents_after_defrag += best_ext_count;
1163 files_block_count += blk_count;
1164 }
1165
1166 if (total_count == 1 && regular_count == 1) {
1167 /* File only */
1168 if (mode_flag & DETAIL) {
1169 int count = 0;
1170 struct fiemap_extent_list *ext_list_tmp =
1171 logical_list_head;
1172
1173 /* Print extents info */
1174 do {
1175 count++;
1176 printf("[ext %d]:\tstart %llu:\tlogical "
1177 "%llu:\tlen %llu\n", count,
1178 ext_list_tmp->data.physical,
1179 ext_list_tmp->data.logical,
1180 ext_list_tmp->data.len);
1181 ext_list_tmp = ext_list_tmp->next;
1182 } while (ext_list_tmp != logical_list_head);
1183
1184 } else {
1185 printf("%-40s%10s/%-10s%9s\n",
1186 "<File>", "now", "best", "size/ext");
1187 if (current_uid == ROOT_UID) {
1188 if (strlen(file) > 40)
1189 printf("%s\n%50d/%-10d%6llu KB\n",
1190 file, now_ext_count,
1191 best_ext_count, size_per_ext);
1192 else
1193 printf("%-40s%10d/%-10d%6llu KB\n",
1194 file, now_ext_count,
1195 best_ext_count, size_per_ext);
1196 } else {
1197 if (strlen(file) > 40)
1198 printf("%s\n%50d/%-10s%7s\n",
1199 file, now_ext_count,
1200 "-", "-");
1201 else
1202 printf("%-40s%10d/%-10s%7s\n",
1203 file, now_ext_count,
1204 "-", "-");
1205 }
1206 }
1207 succeed_cnt++;
1208 goto out;
1209 }
1210
1211 if (mode_flag & DETAIL) {
1212 /* Print statistic info */
1213 sprintf(msg_buffer, "[%u/%u]%s",
1214 defraged_file_count, total_count, file);
1215 if (current_uid == ROOT_UID) {
1216 if (strlen(msg_buffer) > 40)
1217 printf("\033[79;0H\033[K%s\n"
1218 "%50d/%-10d%6llu KB\n",
1219 msg_buffer, now_ext_count,
1220 best_ext_count, size_per_ext);
1221 else
1222 printf("\033[79;0H\033[K%-40s"
1223 "%10d/%-10d%6llu KB\n",
1224 msg_buffer, now_ext_count,
1225 best_ext_count, size_per_ext);
1226 } else {
1227 if (strlen(msg_buffer) > 40)
1228 printf("\033[79;0H\033[K%s\n%50d/%-10s%7s\n",
1229 msg_buffer, now_ext_count,
1230 "-", "-");
1231 else
1232 printf("\033[79;0H\033[K%-40s%10d/%-10s%7s\n",
1233 msg_buffer, now_ext_count,
1234 "-", "-");
1235 }
1236 }
1237
1238 for (i = 0; i < SHOW_FRAG_FILES; i++) {
1239 if (ratio >= frag_rank[i].ratio) {
1240 for (j = SHOW_FRAG_FILES - 1; j > i; j--) {
1241 memset(&frag_rank[j], 0,
1242 sizeof(struct frag_statistic_ino));
1243 strncpy(frag_rank[j].msg_buffer,
1244 frag_rank[j - 1].msg_buffer,
1245 strnlen(frag_rank[j - 1].msg_buffer,
1246 PATH_MAX));
1247 frag_rank[j].now_count =
1248 frag_rank[j - 1].now_count;
1249 frag_rank[j].best_count =
1250 frag_rank[j - 1].best_count;
1251 frag_rank[j].size_per_ext =
1252 frag_rank[j - 1].size_per_ext;
1253 frag_rank[j].ratio =
1254 frag_rank[j - 1].ratio;
1255 }
1256 memset(&frag_rank[i], 0,
1257 sizeof(struct frag_statistic_ino));
1258 strncpy(frag_rank[i].msg_buffer, file,
1259 strnlen(file, PATH_MAX));
1260 frag_rank[i].now_count = now_ext_count;
1261 frag_rank[i].best_count = best_ext_count;
1262 frag_rank[i].size_per_ext = size_per_ext;
1263 frag_rank[i].ratio = ratio;
1264 break;
1265 }
1266 }
1267
1268 succeed_cnt++;
1269
1270 out:
1271 close(fd);
1272 free_ext(physical_list_head);
1273 free_ext(logical_list_head);
1274 return 0;
1275 }
1276
1277 /*
1278 * print_progress - Print defrag progress
1279 *
1280 * @file: file name.
1281 * @start: logical offset for defrag target file
1282 * @file_size: defrag target filesize
1283 */
print_progress(const char * file,loff_t start,loff_t file_size)1284 static void print_progress(const char *file, loff_t start, loff_t file_size)
1285 {
1286 int percent = (start * 100) / file_size;
1287 printf("\033[79;0H\033[K[%u/%u]%s:\t%3d%%",
1288 defraged_file_count, total_count, file, min(percent, 100));
1289 fflush(stdout);
1290
1291 return;
1292 }
1293
1294 /*
1295 * call_defrag() - Execute the defrag program.
1296 *
1297 * @fd: target file descriptor.
1298 * @donor_fd: donor file descriptor.
1299 * @file: target file name.
1300 * @buf: pointer of the struct stat64.
1301 * @ext_list_head: head of the extent list.
1302 */
call_defrag(int fd,int donor_fd,const char * file,const struct stat64 * buf,struct fiemap_extent_list * ext_list_head)1303 static int call_defrag(int fd, int donor_fd, const char *file,
1304 const struct stat64 *buf, struct fiemap_extent_list *ext_list_head)
1305 {
1306 loff_t start = 0;
1307 unsigned int page_num;
1308 unsigned char *vec = NULL;
1309 int defraged_ret = 0;
1310 int ret;
1311 struct move_extent move_data;
1312 struct fiemap_extent_list *ext_list_tmp = NULL;
1313
1314 memset(&move_data, 0, sizeof(struct move_extent));
1315 move_data.donor_fd = donor_fd;
1316
1317 /* Print defrag progress */
1318 print_progress(file, start, buf->st_size);
1319
1320 ext_list_tmp = ext_list_head;
1321 do {
1322 move_data.orig_start = ext_list_tmp->data.logical;
1323 /* Logical offset of orig and donor should be same */
1324 move_data.donor_start = move_data.orig_start;
1325 move_data.len = ext_list_tmp->data.len;
1326 move_data.moved_len = 0;
1327
1328 ret = page_in_core(fd, move_data, &vec, &page_num);
1329 if (ret < 0) {
1330 if (mode_flag & DETAIL) {
1331 printf("\n");
1332 PRINT_ERR_MSG_WITH_ERRNO(
1333 "Failed to get file map");
1334 } else {
1335 printf("\t[ NG ]\n");
1336 }
1337 return -1;
1338 }
1339
1340 /* EXT4_IOC_MOVE_EXT */
1341 defraged_ret =
1342 ioctl(fd, EXT4_IOC_MOVE_EXT, &move_data);
1343
1344 /* Free pages */
1345 ret = defrag_fadvise(fd, move_data, vec, page_num);
1346 if (vec) {
1347 free(vec);
1348 vec = NULL;
1349 }
1350 if (ret < 0) {
1351 if (mode_flag & DETAIL) {
1352 printf("\n");
1353 PRINT_ERR_MSG_WITH_ERRNO(
1354 "Failed to free page");
1355 } else {
1356 printf("\t[ NG ]\n");
1357 }
1358 return -1;
1359 }
1360
1361 if (defraged_ret < 0) {
1362 if (mode_flag & DETAIL) {
1363 printf("\n");
1364 PRINT_ERR_MSG_WITH_ERRNO(
1365 "Failed to defrag with "
1366 "EXT4_IOC_MOVE_EXT ioctl");
1367 if (errno == ENOTTY)
1368 printf("\tAt least 2.6.31-rc1 of "
1369 "vanilla kernel is required\n");
1370 } else {
1371 printf("\t[ NG ]\n");
1372 }
1373 return -1;
1374 }
1375 /* Adjust logical offset for next ioctl */
1376 move_data.orig_start += move_data.moved_len;
1377 move_data.donor_start = move_data.orig_start;
1378
1379 start = move_data.orig_start * buf->st_blksize;
1380
1381 /* Print defrag progress */
1382 print_progress(file, start, buf->st_size);
1383
1384 /* End of file */
1385 if (start >= buf->st_size)
1386 break;
1387
1388 ext_list_tmp = ext_list_tmp->next;
1389 } while (ext_list_tmp != ext_list_head);
1390
1391 return 0;
1392 }
1393
1394 /*
1395 * file_defrag() - Check file attributes and call ioctl to defrag.
1396 *
1397 * @file: the file's name.
1398 * @buf: the pointer of the struct stat64.
1399 * @flag: file type.
1400 * @ftwbuf: the pointer of a struct FTW.
1401 */
file_defrag(const char * file,const struct stat64 * buf,int flag EXT2FS_ATTR ((unused)),struct FTW * ftwbuf EXT2FS_ATTR ((unused)))1402 static int file_defrag(const char *file, const struct stat64 *buf,
1403 int flag EXT2FS_ATTR((unused)),
1404 struct FTW *ftwbuf EXT2FS_ATTR((unused)))
1405 {
1406 int fd;
1407 int donor_fd = -1;
1408 int ret;
1409 int best;
1410 int file_frags_start, file_frags_end;
1411 int orig_physical_cnt, donor_physical_cnt = 0;
1412 char tmp_inode_name[PATH_MAX + 8];
1413 ext4_fsblk_t blk_count = 0;
1414 struct fiemap_extent_list *orig_list_physical = NULL;
1415 struct fiemap_extent_list *orig_list_logical = NULL;
1416 struct fiemap_extent_list *donor_list_physical = NULL;
1417 struct fiemap_extent_list *donor_list_logical = NULL;
1418 struct fiemap_extent_group *orig_group_head = NULL;
1419 struct fiemap_extent_group *orig_group_tmp = NULL;
1420
1421 defraged_file_count++;
1422
1423 if (mode_flag & DETAIL) {
1424 printf("[%u/%u]", defraged_file_count, total_count);
1425 fflush(stdout);
1426 }
1427
1428 if (lost_found_dir[0] != '\0' &&
1429 !memcmp(file, lost_found_dir, strnlen(lost_found_dir, PATH_MAX))) {
1430 if (mode_flag & DETAIL) {
1431 PRINT_FILE_NAME(file);
1432 IN_FTW_PRINT_ERR_MSG(NGMSG_LOST_FOUND);
1433 }
1434 return 0;
1435 }
1436
1437 if (!S_ISREG(buf->st_mode)) {
1438 if (mode_flag & DETAIL) {
1439 PRINT_FILE_NAME(file);
1440 IN_FTW_PRINT_ERR_MSG(NGMSG_FILE_UNREG);
1441 }
1442 return 0;
1443 }
1444
1445 /* Empty file */
1446 if (buf->st_size == 0) {
1447 if (mode_flag & DETAIL) {
1448 PRINT_FILE_NAME(file);
1449 IN_FTW_PRINT_ERR_MSG("File size is 0");
1450 }
1451 return 0;
1452 }
1453
1454 /* Has no blocks */
1455 if (buf->st_blocks == 0) {
1456 if (mode_flag & DETAIL) {
1457 PRINT_FILE_NAME(file);
1458 STATISTIC_ERR_MSG("File has no blocks");
1459 }
1460 return 0;
1461 }
1462
1463 fd = open64(file, O_RDWR);
1464 if (fd < 0) {
1465 if (mode_flag & DETAIL) {
1466 PRINT_FILE_NAME(file);
1467 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_OPEN);
1468 }
1469 return 0;
1470 }
1471
1472 /* Get file's extents */
1473 ret = get_file_extents(fd, &orig_list_physical);
1474 if (ret < 0) {
1475 if (mode_flag & DETAIL) {
1476 PRINT_FILE_NAME(file);
1477 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1478 }
1479 goto out;
1480 }
1481
1482 /* Get the count of file's continuous physical region */
1483 orig_physical_cnt = get_physical_count(orig_list_physical);
1484
1485 /* Change list from physical to logical */
1486 ret = change_physical_to_logical(&orig_list_physical,
1487 &orig_list_logical);
1488 if (ret < 0) {
1489 if (mode_flag & DETAIL) {
1490 PRINT_FILE_NAME(file);
1491 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1492 }
1493 goto out;
1494 }
1495
1496 /* Count file fragments before defrag */
1497 file_frags_start = get_logical_count(orig_list_logical);
1498
1499 blk_count = get_file_blocks(orig_list_logical);
1500 if (file_check(fd, buf, file, file_frags_start, blk_count) < 0)
1501 goto out;
1502
1503 if (fsync(fd) < 0) {
1504 if (mode_flag & DETAIL) {
1505 PRINT_FILE_NAME(file);
1506 PRINT_ERR_MSG_WITH_ERRNO("Failed to sync(fsync)");
1507 }
1508 goto out;
1509 }
1510
1511 if (current_uid == ROOT_UID)
1512 best = get_best_count(blk_count);
1513 else
1514 best = 1;
1515
1516 if (file_frags_start <= best)
1517 goto check_improvement;
1518
1519 /* Combine extents to group */
1520 ret = join_extents(orig_list_logical, &orig_group_head);
1521 if (ret < 0) {
1522 if (mode_flag & DETAIL) {
1523 PRINT_FILE_NAME(file);
1524 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1525 }
1526 goto out;
1527 }
1528
1529 /* Create donor inode */
1530 memset(tmp_inode_name, 0, PATH_MAX + 8);
1531 sprintf(tmp_inode_name, "%.*s.defrag",
1532 (int)strnlen(file, PATH_MAX), file);
1533 donor_fd = open64(tmp_inode_name, O_WRONLY | O_CREAT | O_EXCL, S_IRUSR);
1534 if (donor_fd < 0) {
1535 if (mode_flag & DETAIL) {
1536 PRINT_FILE_NAME(file);
1537 if (errno == EEXIST)
1538 PRINT_ERR_MSG_WITH_ERRNO(
1539 "File is being defraged by other program");
1540 else
1541 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_OPEN);
1542 }
1543 goto out;
1544 }
1545
1546 /* Unlink donor inode */
1547 ret = unlink(tmp_inode_name);
1548 if (ret < 0) {
1549 if (mode_flag & DETAIL) {
1550 PRINT_FILE_NAME(file);
1551 PRINT_ERR_MSG_WITH_ERRNO("Failed to unlink");
1552 }
1553 goto out;
1554 }
1555
1556 /* Allocate space for donor inode */
1557 orig_group_tmp = orig_group_head;
1558 do {
1559 ret = fallocate64(donor_fd, 0,
1560 (loff_t)orig_group_tmp->start->data.logical * block_size,
1561 (loff_t)orig_group_tmp->len * block_size);
1562 if (ret < 0) {
1563 if (mode_flag & DETAIL) {
1564 PRINT_FILE_NAME(file);
1565 PRINT_ERR_MSG_WITH_ERRNO("Failed to fallocate");
1566 }
1567 goto out;
1568 }
1569
1570 orig_group_tmp = orig_group_tmp->next;
1571 } while (orig_group_tmp != orig_group_head);
1572
1573 /* Get donor inode's extents */
1574 ret = get_file_extents(donor_fd, &donor_list_physical);
1575 if (ret < 0) {
1576 if (mode_flag & DETAIL) {
1577 PRINT_FILE_NAME(file);
1578 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1579 }
1580 goto out;
1581 }
1582
1583 /* Calculate donor inode's continuous physical region */
1584 donor_physical_cnt = get_physical_count(donor_list_physical);
1585
1586 /* Change donor extent list from physical to logical */
1587 ret = change_physical_to_logical(&donor_list_physical,
1588 &donor_list_logical);
1589 if (ret < 0) {
1590 if (mode_flag & DETAIL) {
1591 PRINT_FILE_NAME(file);
1592 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_EXTENT);
1593 }
1594 goto out;
1595 }
1596
1597 check_improvement:
1598 if (mode_flag & DETAIL) {
1599 if (file_frags_start != 1)
1600 frag_files_before_defrag++;
1601
1602 extents_before_defrag += file_frags_start;
1603 }
1604
1605 if (file_frags_start <= best ||
1606 orig_physical_cnt <= donor_physical_cnt) {
1607 printf("\033[79;0H\033[K[%u/%u]%s:\t%3d%%",
1608 defraged_file_count, total_count, file, 100);
1609 if (mode_flag & DETAIL)
1610 printf(" extents: %d -> %d",
1611 file_frags_start, file_frags_start);
1612
1613 printf("\t[ OK ]\n");
1614 succeed_cnt++;
1615
1616 if (file_frags_start != 1)
1617 frag_files_after_defrag++;
1618
1619 extents_after_defrag += file_frags_start;
1620 goto out;
1621 }
1622
1623 /* Defrag the file */
1624 ret = call_defrag(fd, donor_fd, file, buf, donor_list_logical);
1625
1626 /* Count file fragments after defrag and print extents info */
1627 if (mode_flag & DETAIL) {
1628 file_frags_end = file_frag_count(fd);
1629 if (file_frags_end < 0) {
1630 printf("\n");
1631 PRINT_ERR_MSG_WITH_ERRNO(NGMSG_FILE_INFO);
1632 goto out;
1633 }
1634
1635 if (file_frags_end != 1)
1636 frag_files_after_defrag++;
1637
1638 extents_after_defrag += file_frags_end;
1639
1640 if (ret < 0)
1641 goto out;
1642
1643 printf(" extents: %d -> %d",
1644 file_frags_start, file_frags_end);
1645 fflush(stdout);
1646 }
1647
1648 if (ret < 0)
1649 goto out;
1650
1651 printf("\t[ OK ]\n");
1652 fflush(stdout);
1653 succeed_cnt++;
1654
1655 out:
1656 close(fd);
1657 if (donor_fd != -1)
1658 close(donor_fd);
1659 free_ext(orig_list_physical);
1660 free_ext(orig_list_logical);
1661 free_ext(donor_list_physical);
1662 free_exts_group(orig_group_head);
1663 return 0;
1664 }
1665
1666 /*
1667 * main() - Ext4 online defrag.
1668 *
1669 * @argc: the number of parameter.
1670 * @argv[]: the pointer array of parameter.
1671 */
main(int argc,char * argv[])1672 int main(int argc, char *argv[])
1673 {
1674 int opt;
1675 int i, j, ret = 0;
1676 int flags = FTW_PHYS | FTW_MOUNT;
1677 int arg_type = -1;
1678 int mount_dir_len = 0;
1679 int success_flag = 0;
1680 char dir_name[PATH_MAX + 1];
1681 char dev_name[PATH_MAX + 1];
1682 struct stat64 buf;
1683 ext2_filsys fs = NULL;
1684
1685 printf("e4defrag %s (%s)\n", E2FSPROGS_VERSION, E2FSPROGS_DATE);
1686
1687 /* Parse arguments */
1688 if (argc == 1)
1689 goto out;
1690
1691 while ((opt = getopt(argc, argv, "vc")) != EOF) {
1692 switch (opt) {
1693 case 'v':
1694 mode_flag |= DETAIL;
1695 break;
1696 case 'c':
1697 mode_flag |= STATISTIC;
1698 break;
1699 default:
1700 goto out;
1701 }
1702 }
1703
1704 if (argc == optind)
1705 goto out;
1706
1707 current_uid = getuid();
1708
1709 /* Main process */
1710 for (i = optind; i < argc; i++) {
1711 succeed_cnt = 0;
1712 regular_count = 0;
1713 total_count = 0;
1714 frag_files_before_defrag = 0;
1715 frag_files_after_defrag = 0;
1716 extents_before_defrag = 0;
1717 extents_after_defrag = 0;
1718 defraged_file_count = 0;
1719 files_block_count = 0;
1720 blocks_per_group = 0;
1721 feature_incompat = 0;
1722 log_groups_per_flex = 0;
1723
1724 memset(dir_name, 0, PATH_MAX + 1);
1725 memset(dev_name, 0, PATH_MAX + 1);
1726 memset(lost_found_dir, 0, PATH_MAX + 1);
1727 memset(frag_rank, 0,
1728 sizeof(struct frag_statistic_ino) * SHOW_FRAG_FILES);
1729
1730 if ((mode_flag & STATISTIC) && i > optind)
1731 printf("\n");
1732
1733 #if BYTE_ORDER != BIG_ENDIAN && BYTE_ORDER != LITTLE_ENDIAN
1734 PRINT_ERR_MSG("Endian's type is not big/little endian");
1735 PRINT_FILE_NAME(argv[i]);
1736 continue;
1737 #endif
1738
1739 if (lstat64(argv[i], &buf) < 0) {
1740 perror(NGMSG_FILE_INFO);
1741 PRINT_FILE_NAME(argv[i]);
1742 continue;
1743 }
1744
1745 /* Handle i.e. lvm device symlinks */
1746 if (S_ISLNK(buf.st_mode)) {
1747 struct stat64 buf2;
1748
1749 if (stat64(argv[i], &buf2) == 0 &&
1750 S_ISBLK(buf2.st_mode))
1751 buf = buf2;
1752 }
1753
1754 if (S_ISBLK(buf.st_mode)) {
1755 /* Block device */
1756 strncpy(dev_name, argv[i], strnlen(argv[i], PATH_MAX));
1757 if (get_mount_point(argv[i], dir_name, PATH_MAX) < 0)
1758 continue;
1759 if (lstat64(dir_name, &buf) < 0) {
1760 perror(NGMSG_FILE_INFO);
1761 PRINT_FILE_NAME(argv[i]);
1762 continue;
1763 }
1764 arg_type = DEVNAME;
1765 if (!(mode_flag & STATISTIC))
1766 printf("ext4 defragmentation for device(%s)\n",
1767 argv[i]);
1768 } else if (S_ISDIR(buf.st_mode)) {
1769 /* Directory */
1770 if (access(argv[i], R_OK) < 0) {
1771 perror(argv[i]);
1772 continue;
1773 }
1774 arg_type = DIRNAME;
1775 strncpy(dir_name, argv[i], strnlen(argv[i], PATH_MAX));
1776 } else if (S_ISREG(buf.st_mode)) {
1777 /* Regular file */
1778 arg_type = FILENAME;
1779 } else {
1780 /* Irregular file */
1781 PRINT_ERR_MSG(NGMSG_FILE_UNREG);
1782 PRINT_FILE_NAME(argv[i]);
1783 continue;
1784 }
1785
1786 /* Set blocksize */
1787 block_size = buf.st_blksize;
1788
1789 /* For device case,
1790 * filesystem type checked in get_mount_point()
1791 */
1792 if (arg_type == FILENAME || arg_type == DIRNAME) {
1793 if (is_ext4(argv[i], dev_name) < 0)
1794 continue;
1795 if (realpath(argv[i], dir_name) == NULL) {
1796 perror("Couldn't get full path");
1797 PRINT_FILE_NAME(argv[i]);
1798 continue;
1799 }
1800 }
1801
1802 if (current_uid == ROOT_UID) {
1803 /* Get super block info */
1804 ret = ext2fs_open(dev_name, EXT2_FLAG_64BITS, 0,
1805 block_size, unix_io_manager, &fs);
1806 if (ret) {
1807 if (mode_flag & DETAIL)
1808 com_err(argv[1], ret,
1809 "while trying to open file system: %s",
1810 dev_name);
1811 continue;
1812 }
1813
1814 blocks_per_group = fs->super->s_blocks_per_group;
1815 feature_incompat = fs->super->s_feature_incompat;
1816 log_groups_per_flex = fs->super->s_log_groups_per_flex;
1817
1818 ext2fs_close_free(&fs);
1819 }
1820
1821 switch (arg_type) {
1822
1823 case DIRNAME:
1824 if (!(mode_flag & STATISTIC))
1825 printf("ext4 defragmentation "
1826 "for directory(%s)\n", argv[i]);
1827
1828 mount_dir_len = strnlen(lost_found_dir, PATH_MAX);
1829
1830 strncat(lost_found_dir, "/lost+found",
1831 PATH_MAX - strnlen(lost_found_dir, PATH_MAX));
1832
1833 /* Not the case("e4defrag mount_point_dir") */
1834 if (dir_name[mount_dir_len] != '\0') {
1835 /*
1836 * "e4defrag mount_point_dir/lost+found"
1837 * or "e4defrag mount_point_dir/lost+found/"
1838 */
1839 if (strncmp(lost_found_dir, dir_name,
1840 strnlen(lost_found_dir,
1841 PATH_MAX)) == 0 &&
1842 (dir_name[strnlen(lost_found_dir,
1843 PATH_MAX)] == '\0' ||
1844 dir_name[strnlen(lost_found_dir,
1845 PATH_MAX)] == '/')) {
1846 PRINT_ERR_MSG(NGMSG_LOST_FOUND);
1847 PRINT_FILE_NAME(argv[i]);
1848 continue;
1849 }
1850
1851 /* "e4defrag mount_point_dir/else_dir" */
1852 memset(lost_found_dir, 0, PATH_MAX + 1);
1853 }
1854 /* fall through */
1855 case DEVNAME:
1856 if (arg_type == DEVNAME) {
1857 strncpy(lost_found_dir, dir_name,
1858 strnlen(dir_name, PATH_MAX));
1859 strncat(lost_found_dir, "/lost+found/",
1860 PATH_MAX - strnlen(lost_found_dir,
1861 PATH_MAX));
1862 }
1863
1864 nftw64(dir_name, calc_entry_counts, FTW_OPEN_FD, flags);
1865
1866 if (mode_flag & STATISTIC) {
1867 if (mode_flag & DETAIL)
1868 printf("%-40s%10s/%-10s%9s\n",
1869 "<File>", "now", "best", "size/ext");
1870
1871 if (!(mode_flag & DETAIL) &&
1872 current_uid != ROOT_UID) {
1873 printf(" Done.\n");
1874 success_flag = 1;
1875 continue;
1876 }
1877
1878 nftw64(dir_name, file_statistic,
1879 FTW_OPEN_FD, flags);
1880
1881 if (succeed_cnt != 0 &&
1882 current_uid == ROOT_UID) {
1883 if (mode_flag & DETAIL)
1884 printf("\n");
1885 printf("%-40s%10s/%-10s%9s\n",
1886 "<Fragmented files>", "now",
1887 "best", "size/ext");
1888 for (j = 0; j < SHOW_FRAG_FILES; j++) {
1889 if (strlen(frag_rank[j].
1890 msg_buffer) > 37) {
1891 printf("%d. %s\n%50d/"
1892 "%-10d%6llu KB\n",
1893 j + 1,
1894 frag_rank[j].msg_buffer,
1895 frag_rank[j].now_count,
1896 frag_rank[j].best_count,
1897 frag_rank[j].
1898 size_per_ext);
1899 } else if (strlen(frag_rank[j].
1900 msg_buffer) > 0) {
1901 printf("%d. %-37s%10d/"
1902 "%-10d%6llu KB\n",
1903 j + 1,
1904 frag_rank[j].msg_buffer,
1905 frag_rank[j].now_count,
1906 frag_rank[j].best_count,
1907 frag_rank[j].
1908 size_per_ext);
1909 } else
1910 break;
1911 }
1912 }
1913 break;
1914 }
1915 /* File tree walk */
1916 nftw64(dir_name, file_defrag, FTW_OPEN_FD, flags);
1917 printf("\n\tSuccess:\t\t\t[ %u/%u ]\n", succeed_cnt,
1918 total_count);
1919 printf("\tFailure:\t\t\t[ %u/%u ]\n",
1920 total_count - succeed_cnt, total_count);
1921 if (mode_flag & DETAIL) {
1922 printf("\tTotal extents:\t\t\t%4d->%d\n",
1923 extents_before_defrag,
1924 extents_after_defrag);
1925 printf("\tFragmented percentage:\t\t"
1926 "%3llu%%->%llu%%\n",
1927 !regular_count ? 0 :
1928 ((unsigned long long)
1929 frag_files_before_defrag * 100) /
1930 regular_count,
1931 !regular_count ? 0 :
1932 ((unsigned long long)
1933 frag_files_after_defrag * 100) /
1934 regular_count);
1935 }
1936 break;
1937 case FILENAME:
1938 total_count = 1;
1939 regular_count = 1;
1940 strncat(lost_found_dir, "/lost+found/",
1941 PATH_MAX - strnlen(lost_found_dir,
1942 PATH_MAX));
1943 if (strncmp(lost_found_dir, dir_name,
1944 strnlen(lost_found_dir,
1945 PATH_MAX)) == 0) {
1946 PRINT_ERR_MSG(NGMSG_LOST_FOUND);
1947 PRINT_FILE_NAME(argv[i]);
1948 continue;
1949 }
1950
1951 if (mode_flag & STATISTIC) {
1952 file_statistic(argv[i], &buf, FTW_F, NULL);
1953 break;
1954 } else
1955 printf("ext4 defragmentation for %s\n",
1956 argv[i]);
1957 /* Defrag single file process */
1958 file_defrag(argv[i], &buf, FTW_F, NULL);
1959 if (succeed_cnt != 0)
1960 printf(" Success:\t\t\t[1/1]\n");
1961 else
1962 printf(" Success:\t\t\t[0/1]\n");
1963
1964 break;
1965 }
1966
1967 if (succeed_cnt != 0)
1968 success_flag = 1;
1969 if (mode_flag & STATISTIC) {
1970 if (current_uid != ROOT_UID) {
1971 printf(" Done.\n");
1972 continue;
1973 }
1974
1975 if (!succeed_cnt) {
1976 if (mode_flag & DETAIL)
1977 printf("\n");
1978
1979 if (arg_type == DEVNAME)
1980 printf(" In this device(%s), "
1981 "none can be defragmented.\n", argv[i]);
1982 else if (arg_type == DIRNAME)
1983 printf(" In this directory(%s), "
1984 "none can be defragmented.\n", argv[i]);
1985 else
1986 printf(" This file(%s) "
1987 "can't be defragmented.\n", argv[i]);
1988 } else {
1989 float files_ratio = 0.0;
1990 float score = 0.0;
1991 __u64 size_per_ext = files_block_count *
1992 (buf.st_blksize / 1024) /
1993 extents_before_defrag;
1994 files_ratio = (float)(extents_before_defrag -
1995 extents_after_defrag) *
1996 100 / files_block_count;
1997 score = CALC_SCORE(files_ratio);
1998 printf("\n Total/best extents\t\t\t\t%d/%d\n"
1999 " Average size per extent"
2000 "\t\t\t%llu KB\n"
2001 " Fragmentation score\t\t\t\t%.0f\n",
2002 extents_before_defrag,
2003 extents_after_defrag,
2004 size_per_ext, score);
2005 printf(" [0-30 no problem:"
2006 " 31-55 a little bit fragmented:"
2007 " 56- needs defrag]\n");
2008
2009 if (arg_type == DEVNAME)
2010 printf(" This device (%s) ", argv[i]);
2011 else if (arg_type == DIRNAME)
2012 printf(" This directory (%s) ",
2013 argv[i]);
2014 else
2015 printf(" This file (%s) ", argv[i]);
2016
2017 if (score > BOUND_SCORE)
2018 printf("needs defragmentation.\n");
2019 else
2020 printf("does not need "
2021 "defragmentation.\n");
2022 }
2023 printf(" Done.\n");
2024 }
2025
2026 }
2027
2028 if (success_flag)
2029 return 0;
2030
2031 exit(1);
2032
2033 out:
2034 printf(MSG_USAGE);
2035 exit(1);
2036 }
2037
2038