patch - apply a diff file to an original
patch [options] [originalfile [patchfile]]
but usually just
patch -pnum <patchfile
patch takes a patch file patchfile containing a difference listing produced by the diff program and applies those differences to one or more original files, producing patched versions. Normally the patched versions are put in place of the originals. Backups can be made; see the -b or --backup option. The names of the files to be patched are usually taken from the patch file, but if there’s just one file to be patched it can specified on the command line as originalfile.
Upon startup, patch attempts to determine the type of the diff listing, unless overruled by a -c (--context), -e (--ed), -n (--normal), or -u (--unified) option. Context diffs (old-style, new-style, and unified) and normal diffs are applied by the patch program itself, while ed diffs are simply fed to the ed(1) editor via a pipe.
patch tries to skip any leading garbage, apply the diff, and then skip any trailing garbage. Thus you could feed an article or message containing a diff listing to patch, and it should work. If the entire diff is indented by a consistent amount, or if a context diff contains lines ending in CRLF or is encapsulated one or more times by prepending “- “ to lines starting with “-” as specified by Internet RFC 934, this is taken into account.
With context diffs, and to a lesser extent with normal diffs, patch can detect when the line numbers mentioned in the patch are incorrect, and attempts to find the correct place to apply each hunk of the patch. As a first guess, it takes the line number mentioned for the hunk, plus or minus any offset used in applying the previous hunk. If that is not the correct place, patch scans both forwards and backwards for a set of lines matching the context given in the hunk. First patch looks for a place where all lines of the context match. If no such place is found, and it’s a context diff, and the maximum fuzz factor is set to 1 or more, then another scan takes place ignoring the first and last line of context. If that fails, and the maximum fuzz factor is set to 2 or more, the first two and last two lines of context are ignored, and another scan is made. (The default maximum fuzz factor is 2.) If patch cannot find a place to install that hunk of the patch, it puts the hunk out to a reject file, which normally is the name of the output file plus a .rej suffix, or # if .rej would generate a file name that is too long (if even appending the single character # makes the file name too long, then # replaces the file name’s last character). (The rejected hunk comes out in ordinary context diff form regardless of the input patch’s form. If the input was a normal diff, many of the contexts are simply null.) The line numbers on the hunks in the reject file may be different than in the patch file: they reflect the approximate location patch thinks the failed hunks belong in the new file rather than the old one.
As each hunk is completed, you are told if the hunk failed, and if so which line (in the new file) patch thought the hunk should go on. If the hunk is installed at a different line from the line number specified in the diff you are told the offset. A single large offset may indicate that a hunk was installed in the wrong place. You are also told if a fuzz factor was used to make the match, in which case you should also be slightly suspicious. If the --verbose option is given, you are also told about hunks that match exactly.
If no original file origfile is specified on the command line, patch tries to figure out from the leading garbage what the name of the file to edit is, using the following rules.
First, patch takes an ordered list of candidate file names as follows:
? If the header is that of a context diff, patch takes the old and new file names in the header. A name is ignored if it does not have enough slashes to satisfy the -pnum or --strip=num option. The name /dev/null is also ignored.
? If there is an Index: line in the leading garbage and if either the old and new names are both absent or if patch is conforming to POSIX, patch takes the name in the Index: line.
? For the purpose of the following rules, the candidate file names are considered to be in the order (old, new, index), regardless of the order that they appear in the header.
Then patch selects a file name from the candidate list as follows:
? If some of the named files exist, patch selects the first name if conforming to POSIX, and the best name otherwise.
? If patch is not ignoring RCS, ClearCase, and SCCS (see the -g num or --get=num option), and no named files exist but an RCS, ClearCase, or SCCS master is found, patch selects the first named file with an RCS, ClearCase, or SCCS master.
? If no named files exist, no RCS, ClearCase, or SCCS master was found, some names are given, patch is not conforming to POSIX, and the patch appears to create a file, patch selects the best name requiring the creation of the fewest directories.
? If no file name results from the above heuristics, you are asked for the name of the file to patch, and patch selects that name.
To determine the best of a nonempty list of file names, patch first takes all the names with the fewest path name components; of those, it then takes all the names with the shortest basename; of those, it then takes all the shortest names; finally, it takes the first remaining name.
Additionally, if the leading garbage contains a Prereq: line, patch takes the first word from the prerequisites line (normally a version number) and checks the original file to see if that word can be found. If not, patch asks for confirmation before proceeding.
The upshot of all this is that you should be able to say, while in a news interface, something like the following:
| patch -d /usr/src/local/blurfl
and patch a file in the blurfl directory directly from the article containing the patch.
If the patch file contains more than one patch, patch tries to apply each of them as if they came from separate patch files. This means, among other things, that it is assumed that the name of the file to patch must be determined for each diff listing, and that the garbage before each diff listing contains interesting things such as file names and revision level, as mentioned previously.
-b or --backup
Make backup files. That is, when patching a file, rename or copy the original instead of removing it. When backing up a file that does not exist, an empty, unreadable backup file is created as a placeholder to represent the nonexistent file. See the -V or --version-control option for details about how backup file names are determined.
setting -p0 gives the entire file name unmodified, -p1 gives
without the leading slash, -p4 gives
and not specifying -p at all just gives you blurfl.c. Whatever you end up with is looked for either in the current directory, or the directory specified by the -d option.
? Take the first existing file from the list (old, new, index) when intuiting file names from diff headers.
? Do not remove files that are empty after patching.
? Do not ask whether to get files from RCS, ClearCase, or SCCS.
? Require that all options precede the files in the command line.
? Do not backup files when there is a mismatch.
Output names as-is.
shell Quote names for the shell if they contain shell metacharacters or would cause ambiguous output.
Quote names for the shell, even if they would normally not require quoting.
escape Quote as with c except omit the surrounding double-quote characters.
You can specify the default value of the --quoting-style option with the environment variable QUOTING_STYLE. If that environment variable is not set, the default value is shell.
If the first hunk of a patch fails, patch reverses the hunk to see if it can be applied that way. If it can, you are asked if you want to have the -R option set. If it can’t, the patch continues to be applied normally. (Note: this method cannot detect a reversed patch if it is a normal diff and if the first command is an append (i.e. it should have been a delete) since appends always succeed, due to the fact that a null context matches anywhere. Luckily, most patches add or change lines rather than delete them, so most reversed normal diffs begin with a delete, which fails, triggering the heuristic.)
The value of method is like the GNU Emacs ‘version-control’ variable; patch also recognizes synonyms that are more descriptive. The valid values for method are (unique abbreviations are accepted):
existing or nil
Make numbered backups of files that already have them, otherwise simple backups. This is the default.
numbered or t
Make numbered backups. The numbered backup file name for F is F.~N~ where N is the version number.
simple or never
Make simple backups. The -B or --prefix, -Y or --basename-prefix, and -z or --suffix options specify the simple backup file name. If none of these options are given, then a simple backup suffix is used; it is the value of the SIMPLE_BACKUP_SUFFIX environment variable if set, and is .orig otherwise.
With numbered or simple backups, if the backup file name is too long, the backup suffix ~ is used instead; if even appending ~ would make the name too long, then ~ replaces the last character of the file name.
The -Z or --set-utc and -T or --set-time options normally refrain from setting a file’s time if the file’s original time does not match the time given in the patch header, or if its contents do not match the patch exactly. However, if the -f or --force option is given, the file time is set regardless.
Due to the limitations of diff output format, these options cannot update the times of files whose contents have not changed. Also, if you use these options, you should remove (e.g. with make clean) all files that depend on the patched files, so that later invocations of make do not get confused by the patched files’ times.
This specifies whether patch gets missing or read-only files from RCS, ClearCase, or SCCS by default; see the -g or --get option.
VERSION_CONTROL or PATCH_VERSION_CONTROL
Selects version control style; see the -v or --version-control option.
controlling terminal; used to get answers to questions asked of the user
diff(1) , ed(1)
Marshall T. Rose and Einar A. Stefferud, Proposed Standard for Message Encapsulation, Internet RFC 934 <URL:ftp://ftp.isi.edu/innotes/rfc934.txt> (1985-01).
There are several things you should bear in mind if you are going to be sending out patches.
Create your patch systematically. A good method is the command diff -Naur old new where old and new identify the old and new directories. The names old and new should not contain any slashes. The diff command’s headers should have dates and times in Universal Time using traditional Unix format, so that patch recipients can use the -Z or --set-utc option. Here is an example command, using Bourne shell syntax:
LC_ALL=C TZ=UTC0 diff -Naur gcc-2.7 gcc-2.8
Tell your recipients how to apply the patch by telling them which directory to cd to, and which patch options to use. The option string -Np1 is recommended. Test your procedure by pretending to be a recipient and applying your patch to a copy of the original files.
You can save people a lot of grief by keeping a patchlevel.h file which is patched to increment the patch level as the first diff in the patch file you send out. If you put a Prereq: line in with the patch, it won’t let them apply patches out of order without some warning.
You can create a file by sending out a diff that compares /dev/null or an empty file dated the Epoch (1970-01-01 00:00:00 UTC) to the file you want to create. This only works if the file you want to create doesn’t exist already in the target directory. Conversely, you can remove a file by sending out a context diff that compares the file to be deleted with an empty file dated the Epoch. The file will be removed unless patch is conforming to POSIX and the -E or --remove-empty-files option is not given. An easy way to generate patches that create and remove files is to use GNU diff’s -N or --new-file option.
If the recipient is supposed to use the -pN option, do not send output that looks like this:
diff -Naur v2.0.29/prog/README prog/README
because the two file names have different numbers of slashes, and different versions of patch interpret the file names differently. To avoid confusion, send output that looks like this instead:
diff -Naur v2.0.29/prog/README v2.0.30/prog/README
Avoid sending patches that compare backup file names like README.orig, since this might confuse patch into patching a backup file instead of the real file. Instead, send patches that compare the same base file names in different directories, e.g. old/README and new/README.
Take care not to send out reversed patches, since it makes people wonder whether they already applied the patch.
Try not to have your patch modify derived files (e.g. the file configure where there is a line configure: configure.in in your makefile), since the recipient should be able to regenerate the derived files anyway. If you must send diffs of derived files, generate the diffs using UTC, have the recipients apply the patch with the -Z or --set-utc option, and have them remove any unpatched files that depend on patched files (e.g. with make clean).
While you may be able to get away with putting 582 diff listings into one file, it may be wiser to group related patches into separate files in case something goes haywire.
Diagnostics generally indicate that patch couldn’t parse your patch file.
If the --verbose option is given, the message Hmm... indicates that there is unprocessed text in the patch file and that patch is attempting to intuit whether there is a patch in that text and, if so, what kind of patch it is.
patch’s exit status is 0 if all hunks are applied successfully, 1 if some hunks cannot be applied, and 2 if there is more serious trouble. When applying a set of patches in a loop it behooves you to check this exit status so you don’t apply a later patch to a partially patched file.
Context diffs cannot reliably represent the creation or deletion of empty files, empty directories, or special files such as symbolic links. Nor can they represent changes to file metadata like ownership, permissions, or whether one file is a hard link to another. If changes like these are also required, separate instructions (e.g. a shell script) to accomplish them should accompany the patch.
patch cannot tell if the line numbers are off in an ed script, and can detect bad line numbers in a normal diff only when it finds a change or deletion. A context diff using fuzz factor 3 may have the same problem. Until a suitable interactive interface is added, you should probably do a context diff in these cases to see if the changes made sense. Of course, compiling without errors is a pretty good indication that the patch worked, but not always.
patch usually produces the correct results, even when it has to do a lot of guessing. However, the results are guaranteed to be correct only when the patch is applied to exactly the same version of the file that the patch was generated from.
The POSIX standard specifies behavior that differs from patch’s traditional behavior. You should be aware of these differences if you must interoperate with patch versions 2.1 and earlier, which do not conform to POSIX.
? In traditional patch, the -p option’s operand was optional, and a bare -p was equivalent to -p0. The -p option now requires an operand, and -p 0 is now equivalent to -p0. For maximum compatibility, use options like -p0 and -p1.
Also, traditional patch simply counted slashes when stripping path prefixes; patch now counts pathname components. That is, a sequence of one or more adjacent slashes now counts as a single slash. For maximum portability, avoid sending patches containing // in file names.
? In traditional patch, backups were enabled by default. This behavior is now enabled with the -b or --backup option.
Conversely, in POSIX patch, backups are never made, even when there is a mismatch. In GNU patch, this behavior is enabled with the --no-backup-if-mismatch option, or by conforming to POSIX with the --posix option or by setting the POSIXLY_CORRECT environment variable.
The -b suffix option of traditional patch is equivalent to the -b -z suffix options of GNU patch.
? Traditional patch used a complicated (and incompletely documented) method to intuit the name of the file to be patched from the patch header. This method did not conform to POSIX, and had a few gotchas. Now patch uses a different, equally complicated (but better documented) method that is optionally POSIX-conforming; we hope it has fewer gotchas. The two methods are compatible if the file names in the context diff header and the Index: line are all identical after prefix-stripping. Your patch is normally compatible if each header’s file names all contain the same number of slashes.
? When traditional patch asked the user a question, it sent the question to standard error and looked for an answer from the first file in the following list that was a terminal: standard error, standard output, /dev/tty, and standard input. Now patch sends questions to standard output and gets answers from /dev/tty. Defaults for some answers have been changed so that patch never goes into an infinite loop when using default answers.
? Traditional patch exited with a status value that counted the number of bad hunks, or with status 1 if there was real trouble. Now patch exits with status 1 if some hunks failed, or with 2 if there was real trouble.
? Limit yourself to the following options when sending instructions meant to be executed by anyone running GNU patch, traditional patch, or a patch that conforms to POSIX. Spaces are significant in the following list, and operands are required.
Please report bugs via email to <email@example.com>.
patch could be smarter about partial matches, excessively deviant offsets and swapped code, but that would take an extra pass.
If code has been duplicated (for instance with #ifdef OLDCODE ... #else ... #endif), patch is incapable of patching both versions, and, if it works at all, will likely patch the wrong one, and tell you that it succeeded to boot.
If you apply a patch you’ve already applied, patch thinks it is a reversed patch, and offers to un-apply the patch. This could be construed as a feature.
Copyright 1984, 1985, 1986, 1988 Larry Wall. Copyright 1989, 1990, 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998 Free Software Foundation, Inc.
Permission is granted to make and distribute verbatim copies of this manual provided the copyright notice and this permission notice are preserved on all copies.
Permission is granted to copy and distribute modified versions of this manual under the conditions for verbatim copying, provided that the entire resulting derived work is distributed under the terms of a permission notice identical to this one.
Permission is granted to copy and distribute translations of this manual into another language, under the above conditions for modified versions, except that this permission notice may be included in translations approved by the copyright holders instead of in the original English.
Larry Wall wrote the original version of patch. Paul Eggert removed patch’s arbitrary limits; added support for binary files, setting file times, and deleting files; and made it conform better to POSIX. Other contributors include Wayne Davison, who added unidiff support, and David MacKenzie, who added configuration and backup support.
Table of Contents