1 /* shred.c - overwrite files and devices to make it harder to recover data
3 Copyright (C) 1999-2003 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 2, or (at your option)
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, write to the Free Software Foundation,
18 Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
20 Written by Colin Plumb. */
23 - use consistent non-capitalization in error messages
24 - add standard GNU copyleft comment
26 - Add -r/-R/--recursive
27 - Add -i/--interactive
30 - Deal with the amazing variety of gettimeofday() implementation bugs.
31 (Some systems use a one-arg form; still others insist that the timezone
32 either be NULL or be non-NULL. Whee.)
33 - Add an unlink-all option to emulate rm.
37 * Do a more secure overwrite of given files or devices, to make it harder
38 * for even very expensive hardware probing to recover the data.
40 * Although this process is also known as "wiping", I prefer the longer
41 * name both because I think it is more evocative of what is happening and
42 * because a longer name conveys a more appropriate sense of deliberateness.
44 * For the theory behind this, see "Secure Deletion of Data from Magnetic
45 * and Solid-State Memory", on line at
46 * http://www.cs.auckland.ac.nz/~pgut001/pubs/secure_del.html
48 * Just for the record, reversing one or two passes of disk overwrite
49 * is not terribly difficult with hardware help. Hook up a good-quality
50 * digitizing oscilloscope to the output of the head preamplifier and copy
51 * the high-res digitized data to a computer for some off-line analysis.
52 * Read the "current" data and average all the pulses together to get an
53 * "average" pulse on the disk. Subtract this average pulse from all of
54 * the actual pulses and you can clearly see the "echo" of the previous
57 * Real hard drives have to balance the cost of the media, the head,
58 * and the read circuitry. They use better-quality media than absolutely
59 * necessary to limit the cost of the read circuitry. By throwing that
60 * assumption out, and the assumption that you want the data processed
61 * as fast as the hard drive can spin, you can do better.
63 * If asked to wipe a file, this also unlinks it, renaming it to in a
64 * clever way to try to leave no trace of the original filename.
66 * The ISAAC code still bears some resemblance to the code written
67 * by Bob Jenkins, but he permits pretty unlimited use.
69 * This was inspired by a desire to improve on some code titled:
70 * Wipe V1.0-- Overwrite and delete files. S. 2/3/96
71 * but I've rewritten everything here so completely that no trace of
72 * the original remains.
75 * Bob Jenkins, for his good RNG work and patience with the FSF copyright
77 * Jim Meyering, for his work merging this into the GNU fileutils while
78 * still letting me feel a sense of ownership and pride. Getting me to
79 * tolerate the GNU brace style was quite a feat of diplomacy.
80 * Paul Eggert, for lots of useful discussion and code. I disagree with
81 * an awful lot of his suggestions, but they're disagreements worth having.
83 * Things to think about:
84 * - Security: Is there any risk to the race
85 * between overwriting and unlinking a file? Will it do anything
86 * drastically bad if told to attack a named pipe or socket?
89 /* The official name of this program (e.g., no `g' prefix). */
90 #define PROGRAM_NAME "shred"
92 #define AUTHORS "Colin Plumb"
103 #include <sys/types.h>
107 #include "closeout.h"
110 #include "inttostr.h"
111 #include "quotearg.h" /* For quotearg_colon */
112 #include "quote.h" /* For quotearg_colon */
115 # define O_NOCTTY 0 /* This is a very optional frill */
118 #define DEFAULT_PASSES 25 /* Default */
120 /* How many seconds to wait before checking whether to output another
121 verbose output line. */
122 #define VERBOSE_UPDATE 5
126 int force
; /* -f flag: chmod files if necessary */
127 size_t n_iterations
; /* -n flag: Number of iterations */
128 off_t size
; /* -s flag: size of file */
129 int remove_file
; /* -u flag: remove file after shredding */
130 int verbose
; /* -v flag: Print progress */
131 int exact
; /* -x flag: Do not round up file size */
132 int zero_fill
; /* -z flag: Add a final zero pass */
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 {"remove", no_argument
, NULL
, 'u'},
142 {"verbose", no_argument
, NULL
, 'v'},
143 {"zero", no_argument
, NULL
, 'z'},
144 {GETOPT_HELP_OPTION_DECL
},
145 {GETOPT_VERSION_OPTION_DECL
},
149 /* Global variable for error printing purposes */
150 char const *program_name
; /* Initialized before any possible use */
156 fprintf (stderr
, _("Try `%s --help' for more information.\n"),
160 printf (_("Usage: %s [OPTIONS] FILE [...]\n"), program_name
);
162 Overwrite the specified FILE(s) repeatedly, in order to make it harder\n\
163 for even very expensive hardware probing to recover the data.\n\
167 Mandatory arguments to long options are mandatory for short options too.\n\
170 -f, --force change permissions to allow writing if necessary\n\
171 -n, --iterations=N Overwrite N times instead of the default (%d)\n\
172 -s, --size=N shred this many bytes (suffixes like K, M, G accepted)\n\
175 -u, --remove truncate and remove file after overwriting\n\
176 -v, --verbose show progress\n\
177 -x, --exact do not round file sizes up to the next full block\n\
178 -z, --zero add a final overwrite with zeros to hide shredding\n\
179 - shred standard output\n\
181 fputs (HELP_OPTION_DESCRIPTION
, stdout
);
182 fputs (VERSION_OPTION_DESCRIPTION
, stdout
);
185 Delete FILE(s) if --remove (-u) is specified. The default is not to remove\n\
186 the files because it is common to operate on device files like /dev/hda,\n\
187 and those files usually should not be removed. When operating on regular\n\
188 files, most people use the --remove option.\n\
192 CAUTION: Note that shred relies on a very important assumption:\n\
193 that the filesystem overwrites data in place. This is the traditional\n\
194 way to do things, but many modern filesystem designs do not satisfy this\n\
195 assumption. The following are examples of filesystems on which shred is\n\
200 * log-structured or journaled filesystems, such as those supplied with\n\
201 AIX and Solaris (and JFS, ReiserFS, XFS, Ext3, etc.)\n\
203 * filesystems that write redundant data and carry on even if some writes\n\
204 fail, such as RAID-based filesystems\n\
206 * filesystems that make snapshots, such as Network Appliance's NFS server\n\
210 * filesystems that cache in temporary locations, such as NFS\n\
213 * compressed filesystems\n\
215 In addition, file system backups and remote mirrors may contain copies\n\
216 of the file that cannot be removed, and that will allow a shredded file\n\
217 to be recovered later.\n\
219 printf (_("\nReport bugs to <%s>.\n"), PACKAGE_BUGREPORT
);
225 # define fdatasync(fd) -1
229 * --------------------------------------------------------------------
230 * Bob Jenkins' cryptographic random number generator, ISAAC.
231 * Hacked by Colin Plumb.
233 * We need a source of random numbers for some of the overwrite data.
234 * Cryptographically secure is desirable, but it's not life-or-death
235 * so I can be a little bit experimental in the choice of RNGs here.
237 * This generator is based somewhat on RC4, but has analysis
238 * (http://ourworld.compuserve.com/homepages/bob_jenkins/randomnu.htm)
239 * pointing to it actually being better. I like it because it's nice
240 * and fast, and because the author did good work analyzing it.
241 * --------------------------------------------------------------------
244 #if defined __STDC__ && __STDC__
245 # define UINT_MAX_32_BITS 4294967295U
247 # define UINT_MAX_32_BITS 0xFFFFFFFF
250 #if ULONG_MAX == UINT_MAX_32_BITS
251 typedef unsigned long word32
;
253 # if UINT_MAX == UINT_MAX_32_BITS
254 typedef unsigned word32
;
256 # if USHRT_MAX == UINT_MAX_32_BITS
257 typedef unsigned short word32
;
259 # if UCHAR_MAX == UINT_MAX_32_BITS
260 typedef unsigned char word32
;
262 "No 32-bit type available!"
268 /* Size of the state tables to use. (You may change ISAAC_LOG) */
270 #define ISAAC_WORDS (1 << ISAAC_LOG)
271 #define ISAAC_BYTES (ISAAC_WORDS * sizeof (word32))
273 /* RNG state variables */
276 word32 mm
[ISAAC_WORDS
]; /* Main state array */
277 word32 iv
[8]; /* Seeding initial vector */
278 word32 a
, b
, c
; /* Extra index variables */
281 /* This index operation is more efficient on many processors */
283 (* (word32 *) ((char *) (mm) + ((x) & (ISAAC_WORDS - 1) * sizeof (word32))))
286 * The central step. This uses two temporaries, x and y. mm is the
287 * whole state array, while m is a pointer to the current word. off is
288 * the offset from m to the word ISAAC_WORDS/2 words away in the mm array,
289 * i.e. +/- ISAAC_WORDS/2.
291 #define isaac_step(mix, a, b, mm, m, off, r) \
293 a = ((a) ^ (mix)) + (m)[off], \
295 *(m) = y = ind (mm, x) + (a) + (b), \
296 *(r) = b = ind (mm, (y) >> ISAAC_LOG) + x \
300 * Refill the entire R array, and update S.
303 isaac_refill (struct isaac_state
*s
, word32 r
[/* ISAAC_WORDS */])
305 register word32 a
, b
; /* Caches of a and b */
306 register word32 x
, y
; /* Temps needed by isaac_step macro */
307 register word32
*m
= s
->mm
; /* Pointer into state array */
314 isaac_step (a
<< 13, a
, b
, s
->mm
, m
, ISAAC_WORDS
/ 2, r
);
315 isaac_step (a
>> 6, a
, b
, s
->mm
, m
+ 1, ISAAC_WORDS
/ 2, r
+ 1);
316 isaac_step (a
<< 2, a
, b
, s
->mm
, m
+ 2, ISAAC_WORDS
/ 2, r
+ 2);
317 isaac_step (a
>> 16, a
, b
, s
->mm
, m
+ 3, ISAAC_WORDS
/ 2, r
+ 3);
320 while ((m
+= 4) < s
->mm
+ ISAAC_WORDS
/ 2);
323 isaac_step (a
<< 13, a
, b
, s
->mm
, m
, -ISAAC_WORDS
/ 2, r
);
324 isaac_step (a
>> 6, a
, b
, s
->mm
, m
+ 1, -ISAAC_WORDS
/ 2, r
+ 1);
325 isaac_step (a
<< 2, a
, b
, s
->mm
, m
+ 2, -ISAAC_WORDS
/ 2, r
+ 2);
326 isaac_step (a
>> 16, a
, b
, s
->mm
, m
+ 3, -ISAAC_WORDS
/ 2, r
+ 3);
329 while ((m
+= 4) < s
->mm
+ ISAAC_WORDS
);
335 * The basic seed-scrambling step for initialization, based on Bob
336 * Jenkins' 256-bit hash.
338 #define mix(a,b,c,d,e,f,g,h) \
339 ( a ^= b << 11, d += a, \
340 b += c, b ^= c >> 2, e += b, \
341 c += d, c ^= d << 8, f += c, \
342 d += e, d ^= e >> 16, g += d, \
343 e += f, e ^= f << 10, h += e, \
344 f += g, f ^= g >> 4, a += f, \
345 g += h, g ^= h << 8, b += g, \
346 h += a, h ^= a >> 9, c += h, \
349 /* The basic ISAAC initialization pass. */
351 isaac_mix (struct isaac_state
*s
, word32
const seed
[/* ISAAC_WORDS */])
363 for (i
= 0; i
< ISAAC_WORDS
; i
+= 8)
374 mix (a
, b
, c
, d
, e
, f
, g
, h
);
396 #if 0 /* Provided for reference only; not used in this code */
398 * Initialize the ISAAC RNG with the given seed material.
399 * Its size MUST be a multiple of ISAAC_BYTES, and may be
400 * stored in the s->mm array.
402 * This is a generalization of the original ISAAC initialization code
403 * to support larger seed sizes. For seed sizes of 0 and ISAAC_BYTES,
407 isaac_init (struct isaac_state
*s
, word32
const *seed
, size_t seedsize
)
409 static word32
const iv
[8] =
411 0x1367df5a, 0x95d90059, 0xc3163e4b, 0x0f421ad8,
412 0xd92a4a78, 0xa51a3c49, 0xc4efea1b, 0x30609119};
416 /* The initialization of iv is a precomputed form of: */
417 for (i
= 0; i
< 7; i
++)
418 iv
[i
] = 0x9e3779b9; /* the golden ratio */
419 for (i
= 0; i
< 4; ++i
) /* scramble it */
420 mix (iv
[0], iv
[1], iv
[2], iv
[3], iv
[4], iv
[5], iv
[6], iv
[7]);
422 s
->a
= s
->b
= s
->c
= 0;
424 for (i
= 0; i
< 8; i
++)
429 /* First pass (as in reference ISAAC code) */
431 /* Second and subsequent passes (extension to ISAAC) */
432 while (seedsize
-= ISAAC_BYTES
)
435 for (i
= 0; i
< ISAAC_WORDS
; i
++)
437 isaac_mix (s
, s
->mm
);
442 /* The no seed case (as in reference ISAAC code) */
443 for (i
= 0; i
< ISAAC_WORDS
; i
++)
448 isaac_mix (s
, s
->mm
);
452 /* Start seeding an ISAAC structire */
454 isaac_seed_start (struct isaac_state
*s
)
456 static word32
const iv
[8] =
458 0x1367df5a, 0x95d90059, 0xc3163e4b, 0x0f421ad8,
459 0xd92a4a78, 0xa51a3c49, 0xc4efea1b, 0x30609119
464 /* The initialization of iv is a precomputed form of: */
465 for (i
= 0; i
< 7; i
++)
466 iv
[i
] = 0x9e3779b9; /* the golden ratio */
467 for (i
= 0; i
< 4; ++i
) /* scramble it */
468 mix (iv
[0], iv
[1], iv
[2], iv
[3], iv
[4], iv
[5], iv
[6], iv
[7]);
470 for (i
= 0; i
< 8; i
++)
472 /* We could initialize s->mm to zero, but why bother? */
474 /* s->c gets used for a data pointer during the seeding phase */
475 s
->a
= s
->b
= s
->c
= 0;
478 /* Add a buffer of seed material */
480 isaac_seed_data (struct isaac_state
*s
, void const *buf
, size_t size
)
486 avail
= sizeof s
->mm
- (size_t) s
->c
; /* s->c is used as a write pointer */
488 /* Do any full buffers that are necessary */
491 p
= (unsigned char *) s
->mm
+ s
->c
;
492 for (i
= 0; i
< avail
; i
++)
493 p
[i
] ^= ((unsigned char const *) buf
)[i
];
494 buf
= (char const *) buf
+ avail
;
496 isaac_mix (s
, s
->mm
);
498 avail
= sizeof s
->mm
;
501 /* And the final partial block */
502 p
= (unsigned char *) s
->mm
+ s
->c
;
503 for (i
= 0; i
< size
; i
++)
504 p
[i
] ^= ((unsigned char const *) buf
)[i
];
505 s
->c
= (word32
) size
;
509 /* End of seeding phase; get everything ready to produce output. */
511 isaac_seed_finish (struct isaac_state
*s
)
513 isaac_mix (s
, s
->mm
);
514 isaac_mix (s
, s
->mm
);
515 /* Now reinitialize c to start things off right */
518 #define ISAAC_SEED(s,x) isaac_seed_data (s, &(x), sizeof (x))
521 #if __GNUC__ >= 2 && (__i386__ || __alpha__)
523 * Many processors have very-high-resolution timer registers,
524 * The timer registers can be made inaccessible, so we have to deal with the
525 * possibility of SIGILL while we're working.
529 sigill_handler (int signum
)
532 longjmp (env
, 1); /* Trivial, just return an indication that it happened */
535 /* FIXME: find a better way.
536 This signal-handling code may well end up being ripped out eventually.
537 An example of how fragile it is, on an i586-sco-sysv5uw7.0.1 system, with
538 gcc-2.95.3pl1, the "rdtsc" instruction causes a segmentation violation.
539 So now, the code catches SIGSEGV. It'd probably be better to remove all
540 of that mess and find a better source of random data. Patches welcome. */
543 isaac_seed_machdep (struct isaac_state
*s
)
545 RETSIGTYPE (*old_handler
[2]) (int);
547 /* This is how one does try/except in C */
548 old_handler
[0] = signal (SIGILL
, sigill_handler
);
549 old_handler
[1] = signal (SIGSEGV
, sigill_handler
);
550 if (setjmp (env
)) /* ANSI: Must be entire controlling expression */
552 signal (SIGILL
, old_handler
[0]);
553 signal (SIGSEGV
, old_handler
[1]);
559 __asm__
__volatile__ ("rdtsc" : "=a" (t
[0]), "=d" (t
[1]));
563 __asm__
__volatile__ ("rpcc %0" : "=r" (t
));
566 /* Code not used because this instruction is available only on first-
567 generation PPCs and evokes a SIGBUS on some Linux 2.4 kernels. */
569 __asm__
__volatile__ ("mfspr %0,22" : "=r" (t
));
572 /* Code not used because this is not accessible from userland */
574 __asm__
__volatile__ ("mfc0\t%0,$9" : "=r" (t
));
577 /* This doesn't compile on all platforms yet. How to fix? */
579 __asm__
__volatile__ ("rd %%tick, %0" : "=r" (t
));
581 signal (SIGILL
, old_handler
[0]);
582 signal (SIGSEGV
, old_handler
[1]);
583 isaac_seed_data (s
, &t
, sizeof t
);
587 #else /* !(__i386__ || __alpha__) */
589 /* Do-nothing stub */
590 # define isaac_seed_machdep(s) (void) 0
592 #endif /* !(__i386__ || __alpha__) */
596 * Get seed material. 16 bytes (128 bits) is plenty, but if we have
597 * /dev/urandom, we get 32 bytes = 256 bits for complete overkill.
600 isaac_seed (struct isaac_state
*s
)
602 isaac_seed_start (s
);
604 { pid_t t
= getpid (); ISAAC_SEED (s
, t
); }
605 { pid_t t
= getppid (); ISAAC_SEED (s
, t
); }
606 { uid_t t
= getuid (); ISAAC_SEED (s
, t
); }
607 { gid_t t
= getgid (); ISAAC_SEED (s
, t
); }
611 hrtime_t t
= gethrtime ();
614 # if HAVE_CLOCK_GETTIME /* POSIX ns-resolution */
616 clock_gettime (CLOCK_REALTIME
, &t
);
618 # if HAVE_GETTIMEOFDAY
620 gettimeofday (&t
, (struct timezone
*) 0);
630 isaac_seed_machdep (s
);
634 int fd
= open ("/dev/urandom", O_RDONLY
| O_NOCTTY
);
639 isaac_seed_data (s
, buf
, 32);
643 fd
= open ("/dev/random", O_RDONLY
| O_NONBLOCK
| O_NOCTTY
);
646 /* /dev/random is more precious, so use less */
649 isaac_seed_data (s
, buf
, 16);
654 isaac_seed_finish (s
);
657 /* Single-word RNG built on top of ISAAC */
660 word32 r
[ISAAC_WORDS
];
662 struct isaac_state
*s
;
666 irand_init (struct irand_state
*r
, struct isaac_state
*s
)
673 * We take from the end of the block deliberately, so if we need
674 * only a small number of values, we choose the final ones which are
675 * marginally better mixed than the initial ones.
678 irand32 (struct irand_state
*r
)
682 isaac_refill (r
->s
, r
->r
);
683 r
->numleft
= ISAAC_WORDS
;
685 return r
->r
[--r
->numleft
];
689 * Return a uniformly distributed random number between 0 and n,
690 * inclusive. Thus, the result is modulo n+1.
692 * Theory of operation: as x steps through every possible 32-bit number,
693 * x % n takes each value at least 2^32 / n times (rounded down), but
694 * the values less than 2^32 % n are taken one additional time. Thus,
695 * x % n is not perfectly uniform. To fix this, the values of x less
696 * than 2^32 % n are disallowed, and if the RNG produces one, we ask
700 irand_mod (struct irand_state
*r
, word32 n
)
708 lim
= -n
% n
; /* == (2**32-n) % n == 2**32 % n */
718 * Fill a buffer with a fixed pattern.
720 * The buffer must be at least 3 bytes long, even if
721 * size is less. Larger sizes are filled exactly.
724 fillpattern (int type
, unsigned char *r
, size_t size
)
727 unsigned bits
= type
& 0xfff;
730 ((unsigned char *) r
)[0] = (bits
>> 4) & 255;
731 ((unsigned char *) r
)[1] = (bits
>> 8) & 255;
732 ((unsigned char *) r
)[2] = bits
& 255;
733 for (i
= 3; i
< size
/ 2; i
*= 2)
734 memcpy ((char *) r
+ i
, (char *) r
, i
);
736 memcpy ((char *) r
+ i
, (char *) r
, size
- i
);
738 /* Invert the first bit of every 512-byte sector. */
740 for (i
= 0; i
< size
; i
+= 512)
745 * Fill a buffer, R (of size SIZE_MAX), with random data.
746 * SIZE is rounded UP to a multiple of ISAAC_BYTES.
749 fillrand (struct isaac_state
*s
, word32
*r
, size_t size_max
, size_t size
)
751 size
= (size
+ ISAAC_BYTES
- 1) / ISAAC_BYTES
;
752 assert (size
<= size_max
);
762 * Generate a 6-character (+ nul) pass name string
763 * FIXME: allow translation of "random".
765 #define PASS_NAME_SIZE 7
767 passname (unsigned char const *data
, char name
[PASS_NAME_SIZE
])
770 sprintf (name
, "%02x%02x%02x", data
[0], data
[1], data
[2]);
772 memcpy (name
, "random", PASS_NAME_SIZE
);
776 * Do pass number k of n, writing "size" bytes of the given pattern "type"
777 * to the file descriptor fd. Qname, k and n are passed in only for verbose
778 * progress message purposes. If n == 0, no progress messages are printed.
780 * If *sizep == -1, the size is unknown, and it will be filled in as soon
784 dopass (int fd
, char const *qname
, off_t
*sizep
, int type
,
785 struct isaac_state
*s
, unsigned long k
, unsigned long n
)
788 off_t offset
; /* Current file posiiton */
789 time_t thresh
IF_LINT (= 0); /* Time to maybe print next status update */
790 time_t now
= 0; /* Current time */
791 size_t lim
; /* Amount of data to try writing */
792 size_t soff
; /* Offset into buffer for next write */
793 ssize_t ssize
; /* Return value from write */
794 #if ISAAC_WORDS > 1024
795 word32 r
[ISAAC_WORDS
* 3]; /* Multiple of 4K and of pattern size */
797 word32 r
[1024 * 3]; /* Multiple of 4K and of pattern size */
799 char pass_string
[PASS_NAME_SIZE
]; /* Name of current pass */
801 /* Printable previous offset into the file */
802 char previous_offset_buf
[LONGEST_HUMAN_READABLE
+ 1];
803 char const *previous_human_offset
IF_LINT (= 0);
805 if (lseek (fd
, (off_t
) 0, SEEK_SET
) == -1)
807 error (0, errno
, _("%s: cannot rewind"), qname
);
811 /* Constant fill patterns need only be set up once. */
815 if ((off_t
) lim
> size
&& size
!= -1)
819 fillpattern (type
, (unsigned char *) r
, lim
);
820 passname ((unsigned char *) r
, pass_string
);
824 passname (0, pass_string
);
827 /* Set position if first status update */
830 error (0, 0, _("%s: pass %lu/%lu (%s)..."), qname
, k
, n
, pass_string
);
831 thresh
= time (NULL
) + VERBOSE_UPDATE
;
832 previous_human_offset
= "";
838 /* How much to write this time? */
840 if ((off_t
) lim
> size
- offset
&& size
!= -1)
844 lim
= (size_t) (size
- offset
);
849 fillrand (s
, r
, sizeof r
, lim
);
850 /* Loop to retry partial writes. */
851 for (soff
= 0; soff
< lim
; soff
+= ssize
)
853 ssize
= write (fd
, (char *) r
+ soff
, lim
- soff
);
856 if ((ssize
== 0 || errno
== ENOSPC
)
859 /* Ah, we have found the end of the file */
860 *sizep
= size
= offset
+ soff
;
866 char buf
[INT_BUFSIZE_BOUND (uintmax_t)];
867 error (0, errnum
, _("%s: error writing at offset %s"),
868 qname
, umaxtostr ((uintmax_t) offset
+ soff
, buf
));
870 * I sometimes use shred on bad media, before throwing it
871 * out. Thus, I don't want it to give up on bad blocks.
872 * This code assumes 512-byte blocks and tries to skip
873 * over them. It works because lim is always a multiple
874 * of 512, except at the end.
876 if (errnum
== EIO
&& soff
% 512 == 0 && lim
>= soff
+ 512
879 if (lseek (fd
, (off_t
) (offset
+ soff
+ 512), SEEK_SET
)
885 error (0, errno
, "%s: lseek", qname
);
892 /* Okay, we have written "soff" bytes. */
894 if (offset
+ soff
< offset
)
896 error (0, 0, _("%s: file too large"), qname
);
902 /* Time to print progress? */
904 && ((offset
== size
&& *previous_human_offset
)
905 || thresh
<= (now
= time (NULL
))))
907 char offset_buf
[LONGEST_HUMAN_READABLE
+ 1];
908 char size_buf
[LONGEST_HUMAN_READABLE
+ 1];
909 int human_progress_opts
= (human_autoscale
| human_SI
910 | human_base_1024
| human_B
);
911 char const *human_offset
912 = human_readable (offset
, offset_buf
,
913 human_floor
| human_progress_opts
, 1, 1);
916 || !STREQ (previous_human_offset
, human_offset
))
919 error (0, 0, _("%s: pass %lu/%lu (%s)...%s"),
920 qname
, k
, n
, pass_string
, human_offset
);
923 uintmax_t off
= offset
;
924 int percent
= (size
== 0
926 : (off
<= TYPE_MAXIMUM (uintmax_t) / 100
928 : off
/ (size
/ 100)));
929 char const *human_size
930 = human_readable (size
, size_buf
,
931 human_ceiling
| human_progress_opts
,
934 human_offset
= human_size
;
935 error (0, 0, _("%s: pass %lu/%lu (%s)...%s/%s %d%%"),
936 qname
, k
, n
, pass_string
, human_offset
, human_size
,
940 strcpy (previous_offset_buf
, human_offset
);
941 previous_human_offset
= previous_offset_buf
;
942 thresh
= now
+ VERBOSE_UPDATE
;
945 * Force periodic syncs to keep displayed progress accurate
946 * FIXME: Should these be present even if -v is not enabled,
947 * to keep the buffer cache from filling with dirty pages?
948 * It's a common problem with programs that do lots of writes,
951 if (fdatasync (fd
) < 0 && fsync (fd
) < 0)
953 error (0, errno
, "%s: fsync", qname
);
960 /* Force what we just wrote to hit the media. */
961 if (fdatasync (fd
) < 0 && fsync (fd
) < 0)
963 error (0, errno
, "%s: fsync", qname
);
970 * The passes start and end with a random pass, and the passes in between
971 * are done in random order. The idea is to deprive someone trying to
972 * reverse the process of knowledge of the overwrite patterns, so they
973 * have the additional step of figuring out what was done to the disk
974 * before they can try to reverse or cancel it.
976 * First, all possible 1-bit patterns. There are two of them.
977 * Then, all possible 2-bit patterns. There are four, but the two
978 * which are also 1-bit patterns can be omitted.
979 * Then, all possible 3-bit patterns. Likewise, 8-2 = 6.
980 * Then, all possible 4-bit patterns. 16-4 = 12.
982 * The basic passes are:
983 * 1-bit: 0x000, 0xFFF
984 * 2-bit: 0x555, 0xAAA
985 * 3-bit: 0x249, 0x492, 0x924, 0x6DB, 0xB6D, 0xDB6 (+ 1-bit)
986 * 100100100100 110110110110
988 * 4-bit: 0x111, 0x222, 0x333, 0x444, 0x666, 0x777,
989 * 0x888, 0x999, 0xBBB, 0xCCC, 0xDDD, 0xEEE (+ 1-bit, 2-bit)
990 * Adding three random passes at the beginning, middle and end
991 * produces the default 25-pass structure.
993 * The next extension would be to 5-bit and 6-bit patterns.
994 * There are 30 uncovered 5-bit patterns and 64-8-2 = 46 uncovered
995 * 6-bit patterns, so they would increase the time required
996 * significantly. 4-bit patterns are enough for most purposes.
998 * The main gotcha is that this would require a trickier encoding,
999 * since lcm(2,3,4) = 12 bits is easy to fit into an int, but
1000 * lcm(2,3,4,5) = 60 bits is not.
1002 * One extension that is included is to complement the first bit in each
1003 * 512-byte block, to alter the phase of the encoded data in the more
1004 * complex encodings. This doesn't apply to MFM, so the 1-bit patterns
1005 * are considered part of the 3-bit ones and the 2-bit patterns are
1006 * considered part of the 4-bit patterns.
1009 * How does the generalization to variable numbers of passes work?
1012 * Have an ordered list of groups of passes. Each group is a set.
1013 * Take as many groups as will fit, plus a random subset of the
1014 * last partial group, and place them into the passes list.
1015 * Then shuffle the passes list into random order and use that.
1017 * One extra detail: if we can't include a large enough fraction of the
1018 * last group to be interesting, then just substitute random passes.
1020 * If you want more passes than the entire list of groups can
1021 * provide, just start repeating from the beginning of the list.
1026 -2, /* 2 random passes */
1027 2, 0x000, 0xFFF, /* 1-bit */
1028 2, 0x555, 0xAAA, /* 2-bit */
1029 -1, /* 1 random pass */
1030 6, 0x249, 0x492, 0x6DB, 0x924, 0xB6D, 0xDB6, /* 3-bit */
1031 12, 0x111, 0x222, 0x333, 0x444, 0x666, 0x777,
1032 0x888, 0x999, 0xBBB, 0xCCC, 0xDDD, 0xEEE, /* 4-bit */
1033 -1, /* 1 random pass */
1034 /* The following patterns have the frst bit per block flipped */
1035 8, 0x1000, 0x1249, 0x1492, 0x16DB, 0x1924, 0x1B6D, 0x1DB6, 0x1FFF,
1036 14, 0x1111, 0x1222, 0x1333, 0x1444, 0x1555, 0x1666, 0x1777,
1037 0x1888, 0x1999, 0x1AAA, 0x1BBB, 0x1CCC, 0x1DDD, 0x1EEE,
1038 -1, /* 1 random pass */
1043 * Generate a random wiping pass pattern with num passes.
1044 * This is a two-stage process. First, the passes to include
1045 * are chosen, and then they are shuffled into the desired
1049 genpattern (int *dest
, size_t num
, struct isaac_state
*s
)
1051 struct irand_state r
;
1056 size_t accum
, top
, swap
;
1064 /* Stage 1: choose the passes to use */
1067 d
= dest
; /* Destination for generated pass list */
1068 n
= num
; /* Passes remaining to fill */
1072 k
= *p
++; /* Block descriptor word */
1074 { /* Loop back to the beginning */
1078 { /* -k random passes */
1080 if ((size_t) k
>= n
)
1089 else if ((size_t) k
<= n
)
1090 { /* Full block of patterns */
1091 memcpy (d
, p
, k
* sizeof (int));
1096 else if (n
< 2 || 3 * n
< (size_t) k
)
1097 { /* Finish with random */
1102 { /* Pad out with k of the n available */
1105 if (n
== (size_t) k
-- || irand_mod (&r
, k
) < n
)
1116 top
= num
- randpasses
; /* Top of initialized data */
1117 /* assert (d == dest+top); */
1120 * We now have fixed patterns in the dest buffer up to
1121 * "top", and we need to scramble them, with "randpasses"
1122 * random passes evenly spaced among them.
1124 * We want one at the beginning, one at the end, and
1125 * evenly spaced in between. To do this, we basically
1126 * use Bresenham's line draw (a.k.a DDA) algorithm
1127 * to draw a line with slope (randpasses-1)/(num-1).
1128 * (We use a positive accumulator and count down to
1131 * So for each desired output value, we do the following:
1132 * - If it should be a random pass, copy the pass type
1133 * to top++, out of the way of the other passes, and
1134 * set the current pass to -1 (random).
1135 * - If it should be a normal pattern pass, choose an
1136 * entry at random between here and top-1 (inclusive)
1137 * and swap the current entry with that one.
1139 randpasses
--; /* To speed up later math */
1140 accum
= randpasses
; /* Bresenham DDA accumulator */
1141 for (n
= 0; n
< num
; n
++)
1143 if (accum
<= randpasses
)
1146 dest
[top
++] = dest
[n
];
1151 swap
= n
+ irand_mod (&r
, top
- n
- 1);
1153 dest
[n
] = dest
[swap
];
1156 accum
-= randpasses
;
1158 /* assert (top == num); */
1160 memset (&r
, 0, sizeof r
); /* Wipe this on general principles */
1164 * The core routine to actually do the work. This overwrites the first
1165 * size bytes of the given fd. Returns -1 on error, 0 on success.
1168 do_wipefd (int fd
, char const *qname
, struct isaac_state
*s
,
1169 struct Options
const *flags
)
1173 off_t size
; /* Size to write, size to read */
1174 unsigned long n
; /* Number of passes for printing purposes */
1177 n
= 0; /* dopass takes n -- 0 to mean "don't print progress" */
1179 n
= flags
->n_iterations
+ ((flags
->zero_fill
) != 0);
1181 if (fstat (fd
, &st
))
1183 error (0, errno
, "%s: fstat", qname
);
1187 /* If we know that we can't possibly shred the file, give up now.
1188 Otherwise, we may go into a infinite loop writing data before we
1189 find that we can't rewind the device. */
1190 if ((S_ISCHR (st
.st_mode
) && isatty (fd
))
1191 || S_ISFIFO (st
.st_mode
)
1192 || S_ISSOCK (st
.st_mode
))
1194 error (0, 0, _("%s: invalid file type"), qname
);
1198 /* Allocate pass array */
1199 passarray
= xmalloc (flags
->n_iterations
* sizeof (int));
1204 /* Accept a length of zero only if it's a regular file.
1205 For any other type of file, try to get the size another way. */
1206 if (S_ISREG (st
.st_mode
))
1211 error (0, 0, _("%s: file has negative size"), qname
);
1217 size
= lseek (fd
, (off_t
) 0, SEEK_END
);
1220 /* We are unable to determine the length, up front.
1221 Let dopass do that as part of its first iteration. */
1226 if (0 <= size
&& !(flags
->exact
))
1228 size
+= ST_BLKSIZE (st
) - 1 - (size
- 1) % ST_BLKSIZE (st
);
1230 /* If in rounding up, we've just overflowed, use the maximum. */
1232 size
= TYPE_MAXIMUM (off_t
);
1236 /* Schedule the passes in random order. */
1237 genpattern (passarray
, flags
->n_iterations
, s
);
1240 for (i
= 0; i
< flags
->n_iterations
; i
++)
1242 if (dopass (fd
, qname
, &size
, passarray
[i
], s
, i
+ 1, n
) < 0)
1244 memset (passarray
, 0, flags
->n_iterations
* sizeof (int));
1250 memset (passarray
, 0, flags
->n_iterations
* sizeof (int));
1253 if (flags
->zero_fill
)
1254 if (dopass (fd
, qname
, &size
, 0, s
, flags
->n_iterations
+ 1, n
) < 0)
1257 /* Okay, now deallocate the data. The effect of ftruncate on
1258 non-regular files is unspecified, so don't worry about any
1259 errors reported for them. */
1260 if (flags
->remove_file
&& ftruncate (fd
, (off_t
) 0) != 0
1261 && S_ISREG (st
.st_mode
))
1263 error (0, errno
, _("%s: error truncating"), qname
);
1270 /* A wrapper with a little more checking for fds on the command line */
1272 wipefd (int fd
, char const *qname
, struct isaac_state
*s
,
1273 struct Options
const *flags
)
1275 int fd_flags
= fcntl (fd
, F_GETFL
);
1279 error (0, errno
, "%s: fcntl", qname
);
1282 if (fd_flags
& O_APPEND
)
1284 error (0, 0, _("%s: cannot shred append-only file descriptor"), qname
);
1287 return do_wipefd (fd
, qname
, s
, flags
);
1290 /* --- Name-wiping code --- */
1292 /* Characters allowed in a file name - a safe universal set. */
1293 static char const nameset
[] =
1294 "0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ_+=%@#.";
1297 * This increments the name, considering it as a big-endian base-N number
1298 * with the digits taken from nameset. Characters not in the nameset
1299 * are considered to come before nameset[0].
1301 * It's not obvious, but this will explode if name[0..len-1] contains
1304 * This returns the carry (1 on overflow).
1307 incname (char *name
, unsigned len
)
1314 p
= strchr (nameset
, name
[--len
]);
1315 /* If the character is not found, replace it with a 0 digit */
1318 name
[len
] = nameset
[0];
1321 /* If this character has a successor, use it */
1327 /* Otherwise, set this digit to 0 and increment the prefix */
1328 name
[len
] = nameset
[0];
1329 return incname (name
, len
);
1333 * Repeatedly rename a file with shorter and shorter names,
1334 * to obliterate all traces of the file name on any system that
1335 * adds a trailing delimiter to on-disk file names and reuses
1336 * the same directory slot. Finally, unlink it.
1337 * The passed-in filename is modified in place to the new filename.
1338 * (Which is unlinked if this function succeeds, but is still present if
1339 * it fails for some reason.)
1341 * The main loop is written carefully to not get stuck if all possible
1342 * names of a given length are occupied. It counts down the length from
1343 * the original to 0. While the length is non-zero, it tries to find an
1344 * unused file name of the given length. It continues until either the
1345 * name is available and the rename succeeds, or it runs out of names
1346 * to try (incname wraps and returns 1). Finally, it unlinks the file.
1348 * The unlink is Unix-specific, as ANSI-standard remove has more
1349 * portability problems with C libraries making it "safe". rename
1352 * To force the directory data out, we try to open the directory and
1353 * invoke fdatasync on it. This is rather non-standard, so we don't
1354 * insist that it works, just fall back to a global sync in that case.
1355 * This is fairly significantly Unix-specific. Of course, on any
1356 * filesystem with synchronous metadata updates, this is unnecessary.
1359 wipename (char *oldname
, char const *qoldname
, struct Options
const *flags
)
1361 char *newname
, *base
; /* Base points to filename part of newname */
1365 int dir_fd
; /* Try to open directory to sync *it* */
1367 newname
= xstrdup (oldname
);
1369 error (0, 0, _("%s: removing"), qoldname
);
1371 /* Find the file name portion */
1372 base
= strrchr (newname
, '/');
1373 /* Temporary hackery to get a directory fd */
1377 dir_fd
= open (newname
, O_RDONLY
| O_NOCTTY
);
1382 dir_fd
= open (".", O_RDONLY
| O_NOCTTY
);
1384 base
= base
? base
+ 1 : newname
;
1385 len
= strlen (base
);
1389 memset (base
, nameset
[0], len
);
1394 if (lstat (newname
, &st
) < 0)
1396 if (rename (oldname
, newname
) == 0)
1399 || (fdatasync (dir_fd
) < 0 && fsync (dir_fd
) < 0))
1400 sync (); /* Force directory out */
1404 * People seem to understand this better than talking
1405 * about renaming oldname. newname doesn't need
1406 * quoting because we picked it. oldname needs to
1407 * be quoted only the first time.
1409 char const *old
= (first
? qoldname
: oldname
);
1410 error (0, 0, _("%s: renamed to %s"), old
, newname
);
1413 memcpy (oldname
+ (base
- newname
), base
, len
+ 1);
1418 /* The rename failed: give up on this length. */
1424 /* newname exists, so increment BASE so we use another */
1427 while (!incname (base
, len
));
1431 err
= unlink (oldname
);
1432 if (dir_fd
< 0 || (fdatasync (dir_fd
) < 0 && fsync (dir_fd
) < 0))
1435 if (!err
&& flags
->verbose
)
1436 error (0, 0, _("%s: removed"), qoldname
);
1441 * Finally, the function that actually takes a filename and grinds
1442 * it into hamburger.
1445 * Detail to note: since we do not restore errno to EACCES after
1446 * a failed chmod, we end up printing the error code from the chmod.
1447 * This is actually the error that stopped us from proceeding, so
1448 * it's arguably the right one, and in practice it'll be either EACCES
1449 * again or EPERM, which both give similar error messages.
1450 * Does anyone disagree?
1453 wipefile (char *name
, char const *qname
,
1454 struct isaac_state
*s
, struct Options
const *flags
)
1458 fd
= open (name
, O_WRONLY
| O_NOCTTY
);
1461 if (errno
== EACCES
&& flags
->force
)
1463 if (chmod (name
, S_IWUSR
) >= 0) /* 0200, user-write-only */
1464 fd
= open (name
, O_WRONLY
| O_NOCTTY
);
1466 else if ((errno
== ENOENT
|| errno
== ENOTDIR
)
1467 && strncmp (name
, "/dev/fd/", 8) == 0)
1469 /* We accept /dev/fd/# even if the OS doesn't support it */
1474 num
= strtoul (name
+ 8, &p
, 10);
1475 /* If it's completely decimal with no leading zeros... */
1476 if (errno
== 0 && !*p
&& num
<= INT_MAX
&&
1477 (('1' <= name
[8] && name
[8] <= '9')
1478 || (name
[8] == '0' && !name
[9])))
1480 return wipefd ((int) num
, qname
, s
, flags
);
1487 error (0, errno
, "%s", qname
);
1491 err
= do_wipefd (fd
, qname
, s
, flags
);
1492 if (close (fd
) != 0)
1494 error (0, 0, "%s: close", qname
);
1497 if (err
== 0 && flags
->remove_file
)
1499 err
= wipename (name
, qname
, flags
);
1501 error (0, 0, _("%s: cannot remove"), qname
);
1507 main (int argc
, char **argv
)
1509 struct isaac_state s
;
1511 struct Options flags
;
1517 program_name
= argv
[0];
1518 setlocale (LC_ALL
, "");
1519 bindtextdomain (PACKAGE
, LOCALEDIR
);
1520 textdomain (PACKAGE
);
1522 atexit (close_stdout
);
1526 memset (&flags
, 0, sizeof flags
);
1528 flags
.n_iterations
= DEFAULT_PASSES
;
1531 while ((c
= getopt_long (argc
, argv
, "fn:s:uvxz", long_opts
, NULL
)) != -1)
1545 if (xstrtoumax (optarg
, NULL
, 10, &tmp
, NULL
) != LONGINT_OK
1546 || (word32
) tmp
!= tmp
1547 || ((size_t) (tmp
* sizeof (int)) / sizeof (int) != tmp
))
1549 error (EXIT_FAILURE
, 0, _("%s: invalid number of passes"),
1550 quotearg_colon (optarg
));
1552 flags
.n_iterations
= (size_t) tmp
;
1557 flags
.remove_file
= 1;
1563 if (xstrtoumax (optarg
, NULL
, 0, &tmp
, "cbBkKMGTPEZY0")
1566 error (EXIT_FAILURE
, 0, _("%s: invalid file size"),
1567 quotearg_colon (optarg
));
1582 flags
.zero_fill
= 1;
1585 case_GETOPT_HELP_CHAR
;
1587 case_GETOPT_VERSION_CHAR (PROGRAM_NAME
, AUTHORS
);
1590 usage (EXIT_FAILURE
);
1594 file
= argv
+ optind
;
1595 n_files
= argc
- optind
;
1599 error (0, 0, _("missing file argument"));
1600 usage (EXIT_FAILURE
);
1603 for (i
= 0; i
< n_files
; i
++)
1605 char *qname
= xstrdup (quotearg_colon (file
[i
]));
1606 if (STREQ (file
[i
], "-"))
1608 if (wipefd (STDOUT_FILENO
, qname
, &s
, &flags
) < 0)
1613 /* Plain filename - Note that this overwrites *argv! */
1614 if (wipefile (file
[i
], qname
, &s
, &flags
) < 0)
1620 /* Just on general principles, wipe s. */
1621 memset (&s
, 0, sizeof s
);