ls: --color now highlights hard linked files, too
[coreutils/bo.git] / src / shred.c
blob6b7d9016750ec62e306a195b8f6fd592383b3712
1 /* shred.c - overwrite files and devices to make it harder to recover data
3 Copyright (C) 1999-2008 Free Software Foundation, Inc.
4 Copyright (C) 1997, 1998, 1999 Colin Plumb.
6 This program is free software: you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation, either version 3 of the License, or
9 (at your option) any later version.
11 This program is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with this program. If not, see <http://www.gnu.org/licenses/>.
19 Written by Colin Plumb. */
21 /* TODO:
22 - use consistent non-capitalization in error messages
23 - add standard GNU copyleft comment
25 - Add -r/-R/--recursive
26 - Add -i/--interactive
27 - Reserve -d
28 - Add -L
29 - Add an unlink-all option to emulate rm.
33 * Do a more secure overwrite of given files or devices, to make it harder
34 * for even very expensive hardware probing to recover the data.
36 * Although this process is also known as "wiping", I prefer the longer
37 * name both because I think it is more evocative of what is happening and
38 * because a longer name conveys a more appropriate sense of deliberateness.
40 * For the theory behind this, see "Secure Deletion of Data from Magnetic
41 * and Solid-State Memory", on line at
42 * http://www.cs.auckland.ac.nz/~pgut001/pubs/secure_del.html
44 * Just for the record, reversing one or two passes of disk overwrite
45 * is not terribly difficult with hardware help. Hook up a good-quality
46 * digitizing oscilloscope to the output of the head preamplifier and copy
47 * the high-res digitized data to a computer for some off-line analysis.
48 * Read the "current" data and average all the pulses together to get an
49 * "average" pulse on the disk. Subtract this average pulse from all of
50 * the actual pulses and you can clearly see the "echo" of the previous
51 * data on the disk.
53 * Real hard drives have to balance the cost of the media, the head,
54 * and the read circuitry. They use better-quality media than absolutely
55 * necessary to limit the cost of the read circuitry. By throwing that
56 * assumption out, and the assumption that you want the data processed
57 * as fast as the hard drive can spin, you can do better.
59 * If asked to wipe a file, this also unlinks it, renaming it to in a
60 * clever way to try to leave no trace of the original filename.
62 * This was inspired by a desire to improve on some code titled:
63 * Wipe V1.0-- Overwrite and delete files. S. 2/3/96
64 * but I've rewritten everything here so completely that no trace of
65 * the original remains.
67 * Thanks to:
68 * Bob Jenkins, for his good RNG work and patience with the FSF copyright
69 * paperwork.
70 * Jim Meyering, for his work merging this into the GNU fileutils while
71 * still letting me feel a sense of ownership and pride. Getting me to
72 * tolerate the GNU brace style was quite a feat of diplomacy.
73 * Paul Eggert, for lots of useful discussion and code. I disagree with
74 * an awful lot of his suggestions, but they're disagreements worth having.
76 * Things to think about:
77 * - Security: Is there any risk to the race
78 * between overwriting and unlinking a file? Will it do anything
79 * drastically bad if told to attack a named pipe or socket?
82 /* The official name of this program (e.g., no `g' prefix). */
83 #define PROGRAM_NAME "shred"
85 #define AUTHORS proper_name ("Colin Plumb")
87 #include <config.h>
89 #include <getopt.h>
90 #include <stdio.h>
91 #include <assert.h>
92 #include <setjmp.h>
93 #include <sys/types.h>
95 #include "system.h"
96 #include "xstrtol.h"
97 #include "error.h"
98 #include "fcntl--.h"
99 #include "human.h"
100 #include "quotearg.h" /* For quotearg_colon */
101 #include "randint.h"
102 #include "randread.h"
104 /* Default number of times to overwrite. */
105 enum { DEFAULT_PASSES = 25 };
107 /* How many seconds to wait before checking whether to output another
108 verbose output line. */
109 enum { VERBOSE_UPDATE = 5 };
111 /* Sector size and corresponding mask, for recovering after write failures.
112 The size must be a power of 2. */
113 enum { SECTOR_SIZE = 512 };
114 enum { SECTOR_MASK = SECTOR_SIZE - 1 };
115 verify (0 < SECTOR_SIZE && (SECTOR_SIZE & SECTOR_MASK) == 0);
117 struct Options
119 bool force; /* -f flag: chmod files if necessary */
120 size_t n_iterations; /* -n flag: Number of iterations */
121 off_t size; /* -s flag: size of file */
122 bool remove_file; /* -u flag: remove file after shredding */
123 bool verbose; /* -v flag: Print progress */
124 bool exact; /* -x flag: Do not round up file size */
125 bool zero_fill; /* -z flag: Add a final zero pass */
128 /* For long options that have no equivalent short option, use a
129 non-character as a pseudo short option, starting with CHAR_MAX + 1. */
130 enum
132 RANDOM_SOURCE_OPTION = CHAR_MAX + 1
135 static struct option const long_opts[] =
137 {"exact", no_argument, NULL, 'x'},
138 {"force", no_argument, NULL, 'f'},
139 {"iterations", required_argument, NULL, 'n'},
140 {"size", required_argument, NULL, 's'},
141 {"random-source", required_argument, NULL, RANDOM_SOURCE_OPTION},
142 {"remove", no_argument, NULL, 'u'},
143 {"verbose", no_argument, NULL, 'v'},
144 {"zero", no_argument, NULL, 'z'},
145 {GETOPT_HELP_OPTION_DECL},
146 {GETOPT_VERSION_OPTION_DECL},
147 {NULL, 0, NULL, 0}
150 void
151 usage (int status)
153 if (status != EXIT_SUCCESS)
154 fprintf (stderr, _("Try `%s --help' for more information.\n"),
155 program_name);
156 else
158 printf (_("Usage: %s [OPTION]... FILE...\n"), program_name);
159 fputs (_("\
160 Overwrite the specified FILE(s) repeatedly, in order to make it harder\n\
161 for even very expensive hardware probing to recover the data.\n\
163 "), stdout);
164 fputs (_("\
165 Mandatory arguments to long options are mandatory for short options too.\n\
166 "), stdout);
167 printf (_("\
168 -f, --force change permissions to allow writing if necessary\n\
169 -n, --iterations=N overwrite N times instead of the default (%d)\n\
170 --random-source=FILE get random bytes from FILE (default /dev/urandom)\n\
171 -s, --size=N shred this many bytes (suffixes like K, M, G accepted)\n\
172 "), DEFAULT_PASSES);
173 fputs (_("\
174 -u, --remove truncate and remove file after overwriting\n\
175 -v, --verbose show progress\n\
176 -x, --exact do not round file sizes up to the next full block;\n\
177 this is the default for non-regular files\n\
178 -z, --zero add a final overwrite with zeros to hide shredding\n\
179 "), stdout);
180 fputs (HELP_OPTION_DESCRIPTION, stdout);
181 fputs (VERSION_OPTION_DESCRIPTION, stdout);
182 fputs (_("\
184 If FILE is -, shred standard output.\n\
186 Delete FILE(s) if --remove (-u) is specified. The default is not to remove\n\
187 the files because it is common to operate on device files like /dev/hda,\n\
188 and those files usually should not be removed. When operating on regular\n\
189 files, most people use the --remove option.\n\
191 "), stdout);
192 fputs (_("\
193 CAUTION: Note that shred relies on a very important assumption:\n\
194 that the file system overwrites data in place. This is the traditional\n\
195 way to do things, but many modern file system designs do not satisfy this\n\
196 assumption. The following are examples of file systems on which shred is\n\
197 not effective, or is not guaranteed to be effective in all file system modes:\n\
199 "), stdout);
200 fputs (_("\
201 * log-structured or journaled file systems, such as those supplied with\n\
202 AIX and Solaris (and JFS, ReiserFS, XFS, Ext3, etc.)\n\
204 * file systems that write redundant data and carry on even if some writes\n\
205 fail, such as RAID-based file systems\n\
207 * file systems that make snapshots, such as Network Appliance's NFS server\n\
209 "), stdout);
210 fputs (_("\
211 * file systems that cache in temporary locations, such as NFS\n\
212 version 3 clients\n\
214 * compressed file systems\n\
216 "), stdout);
217 fputs (_("\
218 In the case of ext3 file systems, the above disclaimer applies\n\
219 (and shred is thus of limited effectiveness) only in data=journal mode,\n\
220 which journals file data in addition to just metadata. In both the\n\
221 data=ordered (default) and data=writeback modes, shred works as usual.\n\
222 Ext3 journaling modes can be changed by adding the data=something option\n\
223 to the mount options for a particular file system in the /etc/fstab file,\n\
224 as documented in the mount man page (man mount).\n\
226 "), stdout);
227 fputs (_("\
228 In addition, file system backups and remote mirrors may contain copies\n\
229 of the file that cannot be removed, and that will allow a shredded file\n\
230 to be recovered later.\n\
231 "), stdout);
232 emit_bug_reporting_address ();
234 exit (status);
239 * Fill a buffer with a fixed pattern.
241 * The buffer must be at least 3 bytes long, even if
242 * size is less. Larger sizes are filled exactly.
244 static void
245 fillpattern (int type, unsigned char *r, size_t size)
247 size_t i;
248 unsigned int bits = type & 0xfff;
250 bits |= bits << 12;
251 r[0] = (bits >> 4) & 255;
252 r[1] = (bits >> 8) & 255;
253 r[2] = bits & 255;
254 for (i = 3; i < size / 2; i *= 2)
255 memcpy (r + i, r, i);
256 if (i < size)
257 memcpy (r + i, r, size - i);
259 /* Invert the first bit of every sector. */
260 if (type & 0x1000)
261 for (i = 0; i < size; i += SECTOR_SIZE)
262 r[i] ^= 0x80;
266 * Generate a 6-character (+ nul) pass name string
267 * FIXME: allow translation of "random".
269 #define PASS_NAME_SIZE 7
270 static void
271 passname (unsigned char const *data, char name[PASS_NAME_SIZE])
273 if (data)
274 sprintf (name, "%02x%02x%02x", data[0], data[1], data[2]);
275 else
276 memcpy (name, "random", PASS_NAME_SIZE);
279 /* Return true when it's ok to ignore an fsync or fdatasync
280 failure that set errno to ERRNO_VAL. */
281 static bool
282 ignorable_sync_errno (int errno_val)
284 return (errno_val == EINVAL
285 || errno_val == EBADF
286 /* HP-UX does this */
287 || errno_val == EISDIR);
290 /* Request that all data for FD be transferred to the corresponding
291 storage device. QNAME is the file name (quoted for colons).
292 Report any errors found. Return 0 on success, -1
293 (setting errno) on failure. It is not an error if fdatasync and/or
294 fsync is not supported for this file, or if the file is not a
295 writable file descriptor. */
296 static int
297 dosync (int fd, char const *qname)
299 int err;
301 #if HAVE_FDATASYNC
302 if (fdatasync (fd) == 0)
303 return 0;
304 err = errno;
305 if ( ! ignorable_sync_errno (err))
307 error (0, err, _("%s: fdatasync failed"), qname);
308 errno = err;
309 return -1;
311 #endif
313 if (fsync (fd) == 0)
314 return 0;
315 err = errno;
316 if ( ! ignorable_sync_errno (err))
318 error (0, err, _("%s: fsync failed"), qname);
319 errno = err;
320 return -1;
323 sync ();
324 return 0;
327 /* Turn on or off direct I/O mode for file descriptor FD, if possible.
328 Try to turn it on if ENABLE is true. Otherwise, try to turn it off. */
329 static void
330 direct_mode (int fd, bool enable)
332 if (O_DIRECT)
334 int fd_flags = fcntl (fd, F_GETFL);
335 if (0 < fd_flags)
337 int new_flags = (enable
338 ? (fd_flags | O_DIRECT)
339 : (fd_flags & ~O_DIRECT));
340 if (new_flags != fd_flags)
341 fcntl (fd, F_SETFL, new_flags);
345 #if HAVE_DIRECTIO && defined DIRECTIO_ON && defined DIRECTIO_OFF
346 /* This is Solaris-specific. See the following for details:
347 http://docs.sun.com/db/doc/816-0213/6m6ne37so?q=directio&a=view */
348 directio (fd, enable ? DIRECTIO_ON : DIRECTIO_OFF);
349 #endif
353 * Do pass number k of n, writing "size" bytes of the given pattern "type"
354 * to the file descriptor fd. Qname, k and n are passed in only for verbose
355 * progress message purposes. If n == 0, no progress messages are printed.
357 * If *sizep == -1, the size is unknown, and it will be filled in as soon
358 * as writing fails.
360 * Return 1 on write error, -1 on other error, 0 on success.
362 static int
363 dopass (int fd, char const *qname, off_t *sizep, int type,
364 struct randread_source *s, unsigned long int k, unsigned long int n)
366 off_t size = *sizep;
367 off_t offset; /* Current file posiiton */
368 time_t thresh IF_LINT (= 0); /* Time to maybe print next status update */
369 time_t now = 0; /* Current time */
370 size_t lim; /* Amount of data to try writing */
371 size_t soff; /* Offset into buffer for next write */
372 ssize_t ssize; /* Return value from write */
374 /* Fill pattern buffer. Aligning it to a 32-bit boundary speeds up randread
375 in some cases. */
376 typedef uint32_t fill_pattern_buffer[3 * 1024];
377 union
379 fill_pattern_buffer buffer;
380 char c[sizeof (fill_pattern_buffer)];
381 unsigned char u[sizeof (fill_pattern_buffer)];
382 } r;
384 off_t sizeof_r = sizeof r;
385 char pass_string[PASS_NAME_SIZE]; /* Name of current pass */
386 bool write_error = false;
387 bool first_write = true;
389 /* Printable previous offset into the file */
390 char previous_offset_buf[LONGEST_HUMAN_READABLE + 1];
391 char const *previous_human_offset IF_LINT (= 0);
393 if (lseek (fd, 0, SEEK_SET) == -1)
395 error (0, errno, _("%s: cannot rewind"), qname);
396 return -1;
399 /* Constant fill patterns need only be set up once. */
400 if (type >= 0)
402 lim = (0 <= size && size < sizeof_r ? size : sizeof r);
403 fillpattern (type, r.u, lim);
404 passname (r.u, pass_string);
406 else
408 passname (0, pass_string);
411 /* Set position if first status update */
412 if (n)
414 error (0, 0, _("%s: pass %lu/%lu (%s)..."), qname, k, n, pass_string);
415 thresh = time (NULL) + VERBOSE_UPDATE;
416 previous_human_offset = "";
419 offset = 0;
420 for (;;)
422 /* How much to write this time? */
423 lim = sizeof r;
424 if (0 <= size && size - offset < sizeof_r)
426 if (size < offset)
427 break;
428 lim = size - offset;
429 if (!lim)
430 break;
432 if (type < 0)
433 randread (s, &r, lim);
434 /* Loop to retry partial writes. */
435 for (soff = 0; soff < lim; soff += ssize, first_write = false)
437 ssize = write (fd, r.c + soff, lim - soff);
438 if (ssize <= 0)
440 if (size < 0 && (ssize == 0 || errno == ENOSPC))
442 /* Ah, we have found the end of the file */
443 *sizep = size = offset + soff;
444 break;
446 else
448 int errnum = errno;
449 char buf[INT_BUFSIZE_BOUND (uintmax_t)];
451 /* If the first write of the first pass for a given file
452 has just failed with EINVAL, turn off direct mode I/O
453 and try again. This works around a bug in linux-2.4
454 whereby opening with O_DIRECT would succeed for some
455 file system types (e.g., ext3), but any attempt to
456 access a file through the resulting descriptor would
457 fail with EINVAL. */
458 if (k == 1 && first_write && errno == EINVAL)
460 direct_mode (fd, false);
461 ssize = 0;
462 continue;
464 error (0, errnum, _("%s: error writing at offset %s"),
465 qname, umaxtostr (offset + soff, buf));
467 /* 'shred' is often used on bad media, before throwing it
468 out. Thus, it shouldn't give up on bad blocks. This
469 code works because lim is always a multiple of
470 SECTOR_SIZE, except at the end. */
471 verify (sizeof r % SECTOR_SIZE == 0);
472 if (errnum == EIO && 0 <= size && (soff | SECTOR_MASK) < lim)
474 size_t soff1 = (soff | SECTOR_MASK) + 1;
475 if (lseek (fd, offset + soff1, SEEK_SET) != -1)
477 /* Arrange to skip this block. */
478 ssize = soff1 - soff;
479 write_error = true;
480 continue;
482 error (0, errno, _("%s: lseek failed"), qname);
484 return -1;
489 /* Okay, we have written "soff" bytes. */
491 if (offset + soff < offset)
493 error (0, 0, _("%s: file too large"), qname);
494 return -1;
497 offset += soff;
499 /* Time to print progress? */
500 if (n
501 && ((offset == size && *previous_human_offset)
502 || thresh <= (now = time (NULL))))
504 char offset_buf[LONGEST_HUMAN_READABLE + 1];
505 char size_buf[LONGEST_HUMAN_READABLE + 1];
506 int human_progress_opts = (human_autoscale | human_SI
507 | human_base_1024 | human_B);
508 char const *human_offset
509 = human_readable (offset, offset_buf,
510 human_floor | human_progress_opts, 1, 1);
512 if (offset == size
513 || !STREQ (previous_human_offset, human_offset))
515 if (size < 0)
516 error (0, 0, _("%s: pass %lu/%lu (%s)...%s"),
517 qname, k, n, pass_string, human_offset);
518 else
520 uintmax_t off = offset;
521 int percent = (size == 0
522 ? 100
523 : (off <= TYPE_MAXIMUM (uintmax_t) / 100
524 ? off * 100 / size
525 : off / (size / 100)));
526 char const *human_size
527 = human_readable (size, size_buf,
528 human_ceiling | human_progress_opts,
529 1, 1);
530 if (offset == size)
531 human_offset = human_size;
532 error (0, 0, _("%s: pass %lu/%lu (%s)...%s/%s %d%%"),
533 qname, k, n, pass_string, human_offset, human_size,
534 percent);
537 strcpy (previous_offset_buf, human_offset);
538 previous_human_offset = previous_offset_buf;
539 thresh = now + VERBOSE_UPDATE;
542 * Force periodic syncs to keep displayed progress accurate
543 * FIXME: Should these be present even if -v is not enabled,
544 * to keep the buffer cache from filling with dirty pages?
545 * It's a common problem with programs that do lots of writes,
546 * like mkfs.
548 if (dosync (fd, qname) != 0)
550 if (errno != EIO)
551 return -1;
552 write_error = true;
558 /* Force what we just wrote to hit the media. */
559 if (dosync (fd, qname) != 0)
561 if (errno != EIO)
562 return -1;
563 write_error = true;
566 return write_error;
570 * The passes start and end with a random pass, and the passes in between
571 * are done in random order. The idea is to deprive someone trying to
572 * reverse the process of knowledge of the overwrite patterns, so they
573 * have the additional step of figuring out what was done to the disk
574 * before they can try to reverse or cancel it.
576 * First, all possible 1-bit patterns. There are two of them.
577 * Then, all possible 2-bit patterns. There are four, but the two
578 * which are also 1-bit patterns can be omitted.
579 * Then, all possible 3-bit patterns. Likewise, 8-2 = 6.
580 * Then, all possible 4-bit patterns. 16-4 = 12.
582 * The basic passes are:
583 * 1-bit: 0x000, 0xFFF
584 * 2-bit: 0x555, 0xAAA
585 * 3-bit: 0x249, 0x492, 0x924, 0x6DB, 0xB6D, 0xDB6 (+ 1-bit)
586 * 100100100100 110110110110
587 * 9 2 4 D B 6
588 * 4-bit: 0x111, 0x222, 0x333, 0x444, 0x666, 0x777,
589 * 0x888, 0x999, 0xBBB, 0xCCC, 0xDDD, 0xEEE (+ 1-bit, 2-bit)
590 * Adding three random passes at the beginning, middle and end
591 * produces the default 25-pass structure.
593 * The next extension would be to 5-bit and 6-bit patterns.
594 * There are 30 uncovered 5-bit patterns and 64-8-2 = 46 uncovered
595 * 6-bit patterns, so they would increase the time required
596 * significantly. 4-bit patterns are enough for most purposes.
598 * The main gotcha is that this would require a trickier encoding,
599 * since lcm(2,3,4) = 12 bits is easy to fit into an int, but
600 * lcm(2,3,4,5) = 60 bits is not.
602 * One extension that is included is to complement the first bit in each
603 * 512-byte block, to alter the phase of the encoded data in the more
604 * complex encodings. This doesn't apply to MFM, so the 1-bit patterns
605 * are considered part of the 3-bit ones and the 2-bit patterns are
606 * considered part of the 4-bit patterns.
609 * How does the generalization to variable numbers of passes work?
611 * Here's how...
612 * Have an ordered list of groups of passes. Each group is a set.
613 * Take as many groups as will fit, plus a random subset of the
614 * last partial group, and place them into the passes list.
615 * Then shuffle the passes list into random order and use that.
617 * One extra detail: if we can't include a large enough fraction of the
618 * last group to be interesting, then just substitute random passes.
620 * If you want more passes than the entire list of groups can
621 * provide, just start repeating from the beginning of the list.
623 static int const
624 patterns[] =
626 -2, /* 2 random passes */
627 2, 0x000, 0xFFF, /* 1-bit */
628 2, 0x555, 0xAAA, /* 2-bit */
629 -1, /* 1 random pass */
630 6, 0x249, 0x492, 0x6DB, 0x924, 0xB6D, 0xDB6, /* 3-bit */
631 12, 0x111, 0x222, 0x333, 0x444, 0x666, 0x777,
632 0x888, 0x999, 0xBBB, 0xCCC, 0xDDD, 0xEEE, /* 4-bit */
633 -1, /* 1 random pass */
634 /* The following patterns have the frst bit per block flipped */
635 8, 0x1000, 0x1249, 0x1492, 0x16DB, 0x1924, 0x1B6D, 0x1DB6, 0x1FFF,
636 14, 0x1111, 0x1222, 0x1333, 0x1444, 0x1555, 0x1666, 0x1777,
637 0x1888, 0x1999, 0x1AAA, 0x1BBB, 0x1CCC, 0x1DDD, 0x1EEE,
638 -1, /* 1 random pass */
639 0 /* End */
643 * Generate a random wiping pass pattern with num passes.
644 * This is a two-stage process. First, the passes to include
645 * are chosen, and then they are shuffled into the desired
646 * order.
648 static void
649 genpattern (int *dest, size_t num, struct randint_source *s)
651 size_t randpasses;
652 int const *p;
653 int *d;
654 size_t n;
655 size_t accum, top, swap;
656 int k;
658 if (!num)
659 return;
661 /* Stage 1: choose the passes to use */
662 p = patterns;
663 randpasses = 0;
664 d = dest; /* Destination for generated pass list */
665 n = num; /* Passes remaining to fill */
667 for (;;)
669 k = *p++; /* Block descriptor word */
670 if (!k)
671 { /* Loop back to the beginning */
672 p = patterns;
674 else if (k < 0)
675 { /* -k random passes */
676 k = -k;
677 if ((size_t) k >= n)
679 randpasses += n;
680 n = 0;
681 break;
683 randpasses += k;
684 n -= k;
686 else if ((size_t) k <= n)
687 { /* Full block of patterns */
688 memcpy (d, p, k * sizeof (int));
689 p += k;
690 d += k;
691 n -= k;
693 else if (n < 2 || 3 * n < (size_t) k)
694 { /* Finish with random */
695 randpasses += n;
696 break;
698 else
699 { /* Pad out with k of the n available */
702 if (n == (size_t) k || randint_choose (s, k) < n)
704 *d++ = *p;
705 n--;
707 p++;
709 while (n);
710 break;
713 top = num - randpasses; /* Top of initialized data */
714 /* assert (d == dest+top); */
717 * We now have fixed patterns in the dest buffer up to
718 * "top", and we need to scramble them, with "randpasses"
719 * random passes evenly spaced among them.
721 * We want one at the beginning, one at the end, and
722 * evenly spaced in between. To do this, we basically
723 * use Bresenham's line draw (a.k.a DDA) algorithm
724 * to draw a line with slope (randpasses-1)/(num-1).
725 * (We use a positive accumulator and count down to
726 * do this.)
728 * So for each desired output value, we do the following:
729 * - If it should be a random pass, copy the pass type
730 * to top++, out of the way of the other passes, and
731 * set the current pass to -1 (random).
732 * - If it should be a normal pattern pass, choose an
733 * entry at random between here and top-1 (inclusive)
734 * and swap the current entry with that one.
736 randpasses--; /* To speed up later math */
737 accum = randpasses; /* Bresenham DDA accumulator */
738 for (n = 0; n < num; n++)
740 if (accum <= randpasses)
742 accum += num - 1;
743 dest[top++] = dest[n];
744 dest[n] = -1;
746 else
748 swap = n + randint_choose (s, top - n);
749 k = dest[n];
750 dest[n] = dest[swap];
751 dest[swap] = k;
753 accum -= randpasses;
755 /* assert (top == num); */
759 * The core routine to actually do the work. This overwrites the first
760 * size bytes of the given fd. Return true if successful.
762 static bool
763 do_wipefd (int fd, char const *qname, struct randint_source *s,
764 struct Options const *flags)
766 size_t i;
767 struct stat st;
768 off_t size; /* Size to write, size to read */
769 unsigned long int n; /* Number of passes for printing purposes */
770 int *passarray;
771 bool ok = true;
772 struct randread_source *rs;
774 n = 0; /* dopass takes n -- 0 to mean "don't print progress" */
775 if (flags->verbose)
776 n = flags->n_iterations + flags->zero_fill;
778 if (fstat (fd, &st))
780 error (0, errno, _("%s: fstat failed"), qname);
781 return false;
784 /* If we know that we can't possibly shred the file, give up now.
785 Otherwise, we may go into a infinite loop writing data before we
786 find that we can't rewind the device. */
787 if ((S_ISCHR (st.st_mode) && isatty (fd))
788 || S_ISFIFO (st.st_mode)
789 || S_ISSOCK (st.st_mode))
791 error (0, 0, _("%s: invalid file type"), qname);
792 return false;
795 direct_mode (fd, true);
797 /* Allocate pass array */
798 passarray = xnmalloc (flags->n_iterations, sizeof *passarray);
800 size = flags->size;
801 if (size == -1)
803 /* Accept a length of zero only if it's a regular file.
804 For any other type of file, try to get the size another way. */
805 if (S_ISREG (st.st_mode))
807 size = st.st_size;
808 if (size < 0)
810 error (0, 0, _("%s: file has negative size"), qname);
811 return false;
814 else
816 size = lseek (fd, 0, SEEK_END);
817 if (size <= 0)
819 /* We are unable to determine the length, up front.
820 Let dopass do that as part of its first iteration. */
821 size = -1;
825 /* Allow `rounding up' only for regular files. */
826 if (0 <= size && !(flags->exact) && S_ISREG (st.st_mode))
828 size += ST_BLKSIZE (st) - 1 - (size - 1) % ST_BLKSIZE (st);
830 /* If in rounding up, we've just overflowed, use the maximum. */
831 if (size < 0)
832 size = TYPE_MAXIMUM (off_t);
836 /* Schedule the passes in random order. */
837 genpattern (passarray, flags->n_iterations, s);
839 rs = randint_get_source (s);
841 /* Do the work */
842 for (i = 0; i < flags->n_iterations; i++)
844 int err = dopass (fd, qname, &size, passarray[i], rs, i + 1, n);
845 if (err)
847 if (err < 0)
849 memset (passarray, 0, flags->n_iterations * sizeof (int));
850 free (passarray);
851 return false;
853 ok = false;
857 memset (passarray, 0, flags->n_iterations * sizeof (int));
858 free (passarray);
860 if (flags->zero_fill)
862 int err = dopass (fd, qname, &size, 0, rs, flags->n_iterations + 1, n);
863 if (err)
865 if (err < 0)
866 return false;
867 ok = false;
871 /* Okay, now deallocate the data. The effect of ftruncate on
872 non-regular files is unspecified, so don't worry about any
873 errors reported for them. */
874 if (flags->remove_file && ftruncate (fd, 0) != 0
875 && S_ISREG (st.st_mode))
877 error (0, errno, _("%s: error truncating"), qname);
878 return false;
881 return ok;
884 /* A wrapper with a little more checking for fds on the command line */
885 static bool
886 wipefd (int fd, char const *qname, struct randint_source *s,
887 struct Options const *flags)
889 int fd_flags = fcntl (fd, F_GETFL);
891 if (fd_flags < 0)
893 error (0, errno, _("%s: fcntl failed"), qname);
894 return false;
896 if (fd_flags & O_APPEND)
898 error (0, 0, _("%s: cannot shred append-only file descriptor"), qname);
899 return false;
901 return do_wipefd (fd, qname, s, flags);
904 /* --- Name-wiping code --- */
906 /* Characters allowed in a file name - a safe universal set. */
907 static char const nameset[] =
908 "0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ_.";
910 /* Increment NAME (with LEN bytes). NAME must be a big-endian base N
911 number with the digits taken from nameset. Return true if
912 successful if not (because NAME already has the greatest possible
913 value. */
915 static bool
916 incname (char *name, size_t len)
918 while (len--)
920 char const *p = strchr (nameset, name[len]);
922 /* If this character has a successor, use it. */
923 if (p[1])
925 name[len] = p[1];
926 return true;
929 /* Otherwise, set this digit to 0 and increment the prefix. */
930 name[len] = nameset[0];
933 return false;
937 * Repeatedly rename a file with shorter and shorter names,
938 * to obliterate all traces of the file name on any system that
939 * adds a trailing delimiter to on-disk file names and reuses
940 * the same directory slot. Finally, unlink it.
941 * The passed-in filename is modified in place to the new filename.
942 * (Which is unlinked if this function succeeds, but is still present if
943 * it fails for some reason.)
945 * The main loop is written carefully to not get stuck if all possible
946 * names of a given length are occupied. It counts down the length from
947 * the original to 0. While the length is non-zero, it tries to find an
948 * unused file name of the given length. It continues until either the
949 * name is available and the rename succeeds, or it runs out of names
950 * to try (incname wraps and returns 1). Finally, it unlinks the file.
952 * The unlink is Unix-specific, as ANSI-standard remove has more
953 * portability problems with C libraries making it "safe". rename
954 * is ANSI-standard.
956 * To force the directory data out, we try to open the directory and
957 * invoke fdatasync and/or fsync on it. This is non-standard, so don't
958 * insist that it works: just fall back to a global sync in that case.
959 * This is fairly significantly Unix-specific. Of course, on any
960 * file system with synchronous metadata updates, this is unnecessary.
962 static bool
963 wipename (char *oldname, char const *qoldname, struct Options const *flags)
965 char *newname = xstrdup (oldname);
966 char *base = last_component (newname);
967 size_t len = base_len (base);
968 char *dir = dir_name (newname);
969 char *qdir = xstrdup (quotearg_colon (dir));
970 bool first = true;
971 bool ok = true;
973 int dir_fd = open (dir, O_RDONLY | O_DIRECTORY | O_NOCTTY | O_NONBLOCK);
975 if (flags->verbose)
976 error (0, 0, _("%s: removing"), qoldname);
978 while (len)
980 memset (base, nameset[0], len);
981 base[len] = 0;
984 struct stat st;
985 if (lstat (newname, &st) < 0)
987 if (rename (oldname, newname) == 0)
989 if (0 <= dir_fd && dosync (dir_fd, qdir) != 0)
990 ok = false;
991 if (flags->verbose)
994 * People seem to understand this better than talking
995 * about renaming oldname. newname doesn't need
996 * quoting because we picked it. oldname needs to
997 * be quoted only the first time.
999 char const *old = (first ? qoldname : oldname);
1000 error (0, 0, _("%s: renamed to %s"), old, newname);
1001 first = false;
1003 memcpy (oldname + (base - newname), base, len + 1);
1004 break;
1006 else
1008 /* The rename failed: give up on this length. */
1009 break;
1012 else
1014 /* newname exists, so increment BASE so we use another */
1017 while (incname (base, len));
1018 len--;
1020 if (unlink (oldname) != 0)
1022 error (0, errno, _("%s: failed to remove"), qoldname);
1023 ok = false;
1025 else if (flags->verbose)
1026 error (0, 0, _("%s: removed"), qoldname);
1027 if (0 <= dir_fd)
1029 if (dosync (dir_fd, qdir) != 0)
1030 ok = false;
1031 if (close (dir_fd) != 0)
1033 error (0, errno, _("%s: failed to close"), qdir);
1034 ok = false;
1037 free (newname);
1038 free (dir);
1039 free (qdir);
1040 return ok;
1044 * Finally, the function that actually takes a filename and grinds
1045 * it into hamburger.
1047 * FIXME
1048 * Detail to note: since we do not restore errno to EACCES after
1049 * a failed chmod, we end up printing the error code from the chmod.
1050 * This is actually the error that stopped us from proceeding, so
1051 * it's arguably the right one, and in practice it'll be either EACCES
1052 * again or EPERM, which both give similar error messages.
1053 * Does anyone disagree?
1055 static bool
1056 wipefile (char *name, char const *qname,
1057 struct randint_source *s, struct Options const *flags)
1059 bool ok;
1060 int fd;
1062 fd = open (name, O_WRONLY | O_NOCTTY | O_BINARY);
1063 if (fd < 0
1064 && (errno == EACCES && flags->force)
1065 && chmod (name, S_IWUSR) == 0)
1066 fd = open (name, O_WRONLY | O_NOCTTY | O_BINARY);
1067 if (fd < 0)
1069 error (0, errno, _("%s: failed to open for writing"), qname);
1070 return false;
1073 ok = do_wipefd (fd, qname, s, flags);
1074 if (close (fd) != 0)
1076 error (0, errno, _("%s: failed to close"), qname);
1077 ok = false;
1079 if (ok && flags->remove_file)
1080 ok = wipename (name, qname, flags);
1081 return ok;
1085 /* Buffers for random data. */
1086 static struct randint_source *randint_source;
1088 /* Just on general principles, wipe buffers containing information
1089 that may be related to the possibly-pseudorandom values used during
1090 shredding. */
1091 static void
1092 clear_random_data (void)
1094 randint_all_free (randint_source);
1099 main (int argc, char **argv)
1101 bool ok = true;
1102 struct Options flags = { 0, };
1103 char **file;
1104 int n_files;
1105 int c;
1106 int i;
1107 char const *random_source = NULL;
1109 initialize_main (&argc, &argv);
1110 set_program_name (argv[0]);
1111 setlocale (LC_ALL, "");
1112 bindtextdomain (PACKAGE, LOCALEDIR);
1113 textdomain (PACKAGE);
1115 atexit (close_stdout);
1117 flags.n_iterations = DEFAULT_PASSES;
1118 flags.size = -1;
1120 while ((c = getopt_long (argc, argv, "fn:s:uvxz", long_opts, NULL)) != -1)
1122 switch (c)
1124 case 'f':
1125 flags.force = true;
1126 break;
1128 case 'n':
1130 uintmax_t tmp;
1131 if (xstrtoumax (optarg, NULL, 10, &tmp, NULL) != LONGINT_OK
1132 || MIN (UINT32_MAX, SIZE_MAX / sizeof (int)) < tmp)
1134 error (EXIT_FAILURE, 0, _("%s: invalid number of passes"),
1135 quotearg_colon (optarg));
1137 flags.n_iterations = tmp;
1139 break;
1141 case RANDOM_SOURCE_OPTION:
1142 if (random_source && !STREQ (random_source, optarg))
1143 error (EXIT_FAILURE, 0, _("multiple random sources specified"));
1144 random_source = optarg;
1145 break;
1147 case 'u':
1148 flags.remove_file = true;
1149 break;
1151 case 's':
1153 uintmax_t tmp;
1154 if (xstrtoumax (optarg, NULL, 0, &tmp, "cbBkKMGTPEZY0")
1155 != LONGINT_OK)
1157 error (EXIT_FAILURE, 0, _("%s: invalid file size"),
1158 quotearg_colon (optarg));
1160 flags.size = tmp;
1162 break;
1164 case 'v':
1165 flags.verbose = true;
1166 break;
1168 case 'x':
1169 flags.exact = true;
1170 break;
1172 case 'z':
1173 flags.zero_fill = true;
1174 break;
1176 case_GETOPT_HELP_CHAR;
1178 case_GETOPT_VERSION_CHAR (PROGRAM_NAME, AUTHORS);
1180 default:
1181 usage (EXIT_FAILURE);
1185 file = argv + optind;
1186 n_files = argc - optind;
1188 if (n_files == 0)
1190 error (0, 0, _("missing file operand"));
1191 usage (EXIT_FAILURE);
1194 randint_source = randint_all_new (random_source, SIZE_MAX);
1195 if (! randint_source)
1196 error (EXIT_FAILURE, errno, "%s", quotearg_colon (random_source));
1197 atexit (clear_random_data);
1199 for (i = 0; i < n_files; i++)
1201 char *qname = xstrdup (quotearg_colon (file[i]));
1202 if (STREQ (file[i], "-"))
1204 ok &= wipefd (STDOUT_FILENO, qname, randint_source, &flags);
1206 else
1208 /* Plain filename - Note that this overwrites *argv! */
1209 ok &= wipefile (file[i], qname, randint_source, &flags);
1211 free (qname);
1214 exit (ok ? EXIT_SUCCESS : EXIT_FAILURE);
1217 * vim:sw=2:sts=2: