Otherwise, we run into problems in a typical autoconf-based build
system:
- config.status is created at some point between two seconds.
- config.status is run, generating Makefile by first writing to a file
in /tmp, and then mv-ing it to Makefile.
- If this mv happens before the beginning of the next second, Makefile
will be created with the same tv_sec as config.status, but with
tv_nsec = 0.
- When make runs, it sees that Makefile is older than config.status,
and re-runs config.status to generate Makefile.
In general, POSIX does not define /dev/std{in, out, err} because it
does not want to depend on the dev-filesystem.
For utilities, it thus introduced the '-'-keyword to denote standard
input (and output in some cases) and the programs have to deal with
it accordingly.
Sadly, the design of many tools doesn't allow strict shell-redirections
and many scripts don't even use this feature when possible.
Thus, we made the decision to implement it consistently across all
tools where it makes sense (namely those which read files).
Along the way, I spotted some behavioural bugs in libutil/crypt.c and
others where it was forgotten to fshut the files after use.
General convention is to use size_t to store sizes of all kinds.
Internally, the function uses double anyway, but at least this
doesn't clobber up the API any more and there's a chance in the
future to make this function a bit cleaner and not use this dirty
static buffer hack any more.
recurse() is getting smarter every day. I expect it to pass the Turing
test in a few months.
Along the way, it was reported that "rm -f" on nonexistant files reports
their missing as an internal recurse()-error.
So recurse() knows when to shut up, I added the SILENT flag to fix all
these things.
The restructuring of recurse() in the last few weeks actually broke
the recursion-flags in different tools.
As a long-term goal, the recursor should have a field "maxdepth"
which should be "1" for the non-Rflag-case. "0" stands for unlimited.
Basically, it's a conflict between POSIX and ISO C what do to when
input streams are passed to fflush().
POSIX mandates that the seeking-position should be synced, but ISO C
says it's undefined behaviour.
We love POSIX, but the standard-documents specify that in all conflict
cases, ISO C wins, so this breaks with EBADF on BSD's.
musl and glibc follow POSIX behaviour, which makes sense, but involves
numerous portability concerns.
To get around this, we just don't check fflush() and rely on the fact
that no implementation sets ferror on the file-stream in fflush if it
is an input stream, so every issue caught in fflush() is caught later
with ferror() and fclose().
Add a comment to fshut() because this stuff is so complicated, it
took us a day to figure out.
This has been a known issue for a long time. Example:
printf "word" > /dev/full
wouldn't report there's not enough space on the device.
This is due to the fact that every libc has internal buffers
for stdout which store fragments of written data until they reach
a certain size or on some callback to flush them all at once to the
kernel.
You can force the libc to flush them with fflush(). In case flushing
fails, you can check the return value of fflush() and report an error.
However, previously, sbase didn't have such checks and without fflush(),
the libc silently flushes the buffers on exit without checking the errors.
No offense, but there's no way for the libc to report errors in the exit-
condition.
GNU coreutils solve this by having onexit-callbacks to handle the flushing
and report issues, but they have obvious deficiencies.
After long discussions on IRC, we came to the conclusion that checking the
return value of every io-function would be a bit too much, and having a
general-purpose fclose-wrapper would be the best way to go.
It turned out that fclose() alone is not enough to detect errors. The right
way to do it is to fflush() + check ferror on the fp and then to a fclose().
This is what fshut does and that's how it's done before each return.
The return value is obviously affected, reporting an error in case a flush
or close failed, but also when reading failed for some reason, the error-
state is caught.
the !!( ... + ...) construction is used to call all functions inside the
brackets and not "terminating" on the first.
We want errors to be reported, but there's no reason to stop flushing buffers
when one other file buffer has issues.
Obviously, functionales come before the flush and ret-logic comes after to
prevent early exits as well without reporting warnings if there are any.
One more advantage of fshut() is that it is even able to report errors
on obscure NFS-setups which the other coreutils are unable to detect,
because they only check the return-value of fflush() and fclose(),
not ferror() as well.
It's not useful when 0 is returned anyway, so be sure that we have a
string with length > 0, this also solves some indexing-gotchas like
"len - 1" and so on.
Also, add checked getline()'s whenever it has been forgotten and
clean up the error-messages.
I've been wanting to do this for a while now, as tar(1) used to
be one of messiest and cruftiest tools.
First off, before walking through the audit, I'll talk about
what the DIRFIRST-flag for recurse() does.
It basically calls fn() on the first-level-dir before calling
it's subentries. It's necessary here, because else the order
of the tar-files would've been wrong (it would try to create
dir/file before creating dir/).
Now, to the audit:
1) Update manpage, fix mistake that compression is also available
for compressing. It's only available for extracting.
2) Define the major, minor and makedev macros from glibc by ourselves.
No need to rely on them, as they are common sense.
decomp()
3) Simple refactorization.
putoctal()
4) Add a truncation check for snprintf().
archive()
5) BUGFIX: Add checks to any checkable function, don't blindly call
them, this is harmful and there are 100 ways to exploit that.
6) Use estrlcpy() instead of snprintf() wherever possible, fix
alignment.
7) BUGFIX: Terminate the result-buffer of readlink(), check if
it even succeeded.
8) Fix sizeof()-formatting.
unarchive()
9) BUGFIX: Add checks to any checkable function, don't blindly call
them, this is harmful and there are 100 ways to exploit that.
10) BUGFIX: strtoul can happily return negative numbers. Add checks
for that and also if the full string has been processed.
11) Remove calls to perror(). We have eprintf, use it.
12) BUGFIX: "minor = strtoul(h->mode, 0, 8);". We need h->minor of
course.
13) Fix typo "usupported", remove fprintf-call.
print()
14) Check fread().
xt()
15) Get rid of snprintf-magic. Use estrlcat().
16) BUGFIX: check for ferror() on the tarfile.
usage()
17) Update it. The old usage() was like 1000 years old.
main()
18) Add DIRFIRST-flag to the recursor.
19) Don't print usage() when a mode is re-set. We allow this in
general.
20) Add function checks and fix error messages.
21) Add tarfilename-global for proper error-messages.
1) Rename cp_HLPflag -> cp_follow for consistency.
2) Use function-pointers for stat to clear up the code.
3) BUGFIX: TERMINATE THE RESULT BUFFER OF READLINK !!!
It's something I noticed earlier and it actually lead to some
pretty insane behaviour on our side using glibc (musl somehow
magically solves this).
Basically, symlinks used to contain the data of the file they
pointed to. I wondered for weeks where this came from and now
this has finally been solved.
4) BUGFIX: Do not unconditionally unlink target-files. Even GNU
coreutils do it wrong.
The basic idea is this:
If fflag == 0 --> don't touch target files if they exist.
If fflag == 1 --> unlink all and don't error out when we try
to unlink a file which doesn't exist.
5) Use estrlcpy and estrlcat instead of snprintf for path building.
6) Make it clearer what happens in preserve.
Okay, why yet another recurse()-refactor?
The last one added the recursor-struct, which simplified things
on the user-end, but there was still one thing that bugged me a lot:
Previously, all fn()'s were forced to (l)stat the paths themselves.
This does not work well when you try to keep up with H-, L- and P-
flags at the same time, as each utility-function would have to set
the right function-pointer for (l)stat every single time.
This is not desirable. Furthermore, recurse should be easy to use
and not involve trouble finding the right (l)stat-function to do it
right.
So, what we needed was a stat-argument for each fn(), so it is
directly accessible. This was impossible to do though when the
fn()'s are still directly called by the programs to "start" the
recurse.
Thus, the fundamental change is to make recurse() the function to
go, while designing the fn()'s in a way they can "live" with st
being NULL (we don't want a null-pointer-deref).
What you can see in this commit is the result of this work. Why
all this trouble instead of using nftw?
The special thing about recurse() is that you tell the function
when to recurse() in your fn(). You don't need special flags to
tell nftw() to skip the subtree, just to give an example.
The only single downside to this is that now, you are not allowed
to unconditionally call recurse() from your fn(). It has to be
a directory.
However, that is a cost I think is easily weighed up by the
advantages.
Another thing is the history: I added a procedure at the end of
the outmost recurse to free the history. This way we don't leak
memory.
A simple optimization on the side:
- if (h->dev == st.st_dev && h->ino == st.st_ino)
+ if (h->ino == st.st_ino && h->dev == st.st_dev)
First compare the likely difference in inode-numbers instead of
checking the unlikely condition that the device-numbers are
different.
Be more pedantic about the error-checking, fread can also return
values > 0 even though there has been a read-error.
We want to write the last incoming data and then bail.
pathconf() is just an insane interface to use. All sane operating-
systems set sane values for PATH_MAX. Due to the by-runtime-nature of
pathconf(), it actually weakens the programs depending on its values.
Given over 3 years it has still not been possible to implement a sane
and easy to use apathmax()-utility-function, and after discussing this
on IRC, we'll dump this garbage.
We are careful enough not to overflow PATH_MAX and even if, any user
is able to set another limit in config.mk if he so desires.
It has become a common idiom in sbase to check strlcat() and strlcpy()
using
if (strl{cat, cpy}(dst, src, siz) >= siz)
eprintf("path too long\n");
However, this was not carried out consistently and to this very day,
some tools employed unchecked calls to these functions, effectively
allowing silent truncations to happen, which in turn may lead to
security issues.
To finally put an end to this, the e*-functions detect truncation
automatically and the caller can lean back and enjoy coding without
trouble. :)
For loop detection, a history is mandatory. In the process of also
adding a flexible struct to recurse, the recurse-definition was moved
to fs.h.
The motivation behind the struct is to allow easy extensions to the
recurse-function without having to change the prototypes of all
functions in the process.
Adding flags is really simple as well now.
Using the recursor-struct, it's also easier to see which defaults
apply to a program (for instance, which type of follow, ...).
Another change was to add proper stat-lstat-usage in recurse. It
was wrong before.
While auditing du(1) I realized that there's no way the over 100 lines
of procedures in du() would pass the audit.
Instead, I decided to rewrite this section using recurse() from libutil.
However, the issue was that you'd need some kind of payload to count
the number of bytes in the subdirectories and use them in the higher
hierarchies.
The solution is to add a "void *data" data pointer to each recurse-
function-prototype, which we might also be able to use in other
recurse-applications.
recurse() itself had to be augmented with a recurse_samedev-flag, which
basically prevents recurse from leaving the current device.
Now, let's take a closer look at the audit:
1) Removing the now unnecessary util-functions push, pop, xrealpath,
rename print() to printpath(), localize some global variables.
2) Only pass the block count to nblks instead of the entire stat-
pointer.
3) Fix estrtonum to use the minimum of LLONG_MAX and SIZE_MAX.
4) Use idiomatic argv+argc-loop
5) Report proper exit-status.
After a short correspondence with Otto Moerbeek it turned out
mallocarray() is only in the OpenBSD-Kernel, because the kernel-
malloc doesn't have realloc.
Userspace applications should rather use reallocarray with an
explicit NULL-pointer.
Assuming reallocarray() will become available in c-stdlibs in the
next few years, we nip mallocarray() in the bud to allow an easy
transition to a system-provided version when the day comes.
A function used only in the OpenBSD-Kernel as of now, but it surely
provides a helpful interface when you just don't want to make sure
the incoming pointer to erealloc() is really NULL so it behaves
like malloc, making it a bit more safer.
Talking about *allocarray(): It's definitely a major step in code-
hardening. Especially as a system administrator, you should be
able to trust your core tools without having to worry about segfaults
like this, which can easily lead to privilege escalation.
How do the GNU coreutils handle this?
$ strings -n 4611686018427387903
strings: invalid minimum string length -1
$ strings -n 4611686018427387904
strings: invalid minimum string length 0
They silently overflow...
In comparison, sbase:
$ strings -n 4611686018427387903
mallocarray: out of memory
$ strings -n 4611686018427387904
mallocarray: out of memory
The first out of memory is actually a true OOM returned by malloc,
whereas the second one is a detected overflow, which is not marked
in a special way.
Now tell me which diagnostic error-messages are easier to understand.
Stateless and I stumbled upon this issue while discussing the
semantics of read, accepting a size_t but only being able to return
ssize_t, effectively lacking the ability to report successful
reads > SSIZE_MAX.
The discussion went along and we came to the topic of input-based
memory allocations. Basically, it was possible for the argument
to a memory-allocation-function to overflow, leading to a segfault
later.
The OpenBSD-guys came up with the ingenious reallocarray-function,
and I implemented it as ereallocarray, which automatically returns
on error.
Read more about it here[0].
A simple testcase is this (courtesy to stateless):
$ sbase-strings -n (2^(32|64) / 4)
This will segfault before this patch and properly return an OOM-
situation afterwards (thanks to the overflow-check in reallocarray).
[0]: http://www.openbsd.org/cgi-bin/man.cgi/OpenBSD-current/man3/calloc.3
The HLP-changes to sbase have been a great addition of functionality,
but they kind of "polluted" the enmasse() and recurse() prototypes.
As this will come in handy in the future, knowing at which "depth"
you are inside a recursing function is an important functionality.
Instead of having a special HLP-flag passed to enmasse, each sub-
function needs to provide it on its own and can calculate results
based on the current depth (for instance, 'H' implies 'P' at
depth > 0).
A special case is recurse(), because it actually depends on the
follow-type. A new flag "recurse_follow" brings consistency into
what used to be spread across different naming conventions (fflag,
HLP_flag, ...).
This also fixes numerous bugs with the behaviour of HLP in the
tools using it.
This has already been suggested by Evan Gates <evan.gates@gmail.com>
and he's totally right about it.
So, what's the problem?
I wrote a testing program asshole.c with
int
main(void)
{
execl("/path/to/sbase/echo", "echo", "test");
return 0;
}
and checked the results with glibc and musl. Note that the
sentinel NULL is missing from the end of the argument list.
glibc calculates an argc of 5, musl 4 (instead of 2) and thus
mess up things anyway.
The powerful arg.h also focuses on argv instead of argc as well,
but ignoring argc completely is also the wrong way to go.
Instead, a more idiomatic approach is to check *argv only and
decrement argc on the go.
While at it, I rewrote yes(1) in an argv-centric way as well.
All audited tools have been "fixed" and each following audited
tool will receive the same treatment.
Previously, it was not possible to use
sha1sum test.c | sha1sum -c
because the program would not differenciate between an empty
argument and a non-specified argument.
Moreover, why not allow this?
sha1sum -c hashlist1 hashlist2
Digging deeper I found that using function pointers and a
modification in the crypt-backend might simplify the program
a lot by passing the argument-list to both cryptmain and
cryptcheck.
Allowing more than one list-file to be specified is also
consistent with what the other implementations support,
so we not only have simpler code, we also do not silently
break if there's a script around passing multiple files to
check.
This is a useful behavior if you want to reorder the lines,
because otherwise you might end up with originally two lines
on one, e.g.
$ echo -ne "foo\nbar" | sort
barfoo
We should never mix FILE I/O with raw I/O. Going from raw I/O
to FILE I/O is fine but doing the opposite is extremely tricky and
only works under certain conditions (unbuffered stream + no call
to ungetc()).
The -d option is a GNU extension and is equivalent to its "-P
--preserve=links" options.
Since we don't implement the --preserve=links functionality anyway (it
means preserve hard links between files), just call it -P, which is
specified by POSIX.
Additionally, there is no need to check for cp_Pflag again before
copying the symlink itself because the only way the mode in the stat
will indicate a symlink is if we used lstat (which we only do if -P is
specified).
