When `makeset` got a string containing square brackets
followed by at least one extra character, e.g. "[abc]d",
it entered an infinite loop because it was assumed
`j` could not exceed `len` without having been equal to `len`.
It can, however, when `m == len` and subsequently `j = m + 1`.
See his description below. Thanks Michael!
---
A bug was introduced in bc4c293fe5 causing the
range length for the next set to be used instead of the first one. This causes
issues when choosing the replacement rune when the ranges are of different
lengths.
Current behavior:
$ echo 1234 | tr 'a-f1-4' '1-6a-d'
56ab
Correct behavior:
$ echo 1234 | tr 'a-f1-4' '1-6a-d'
abcd
This also fixes range expressions in the form [a-z], which get encoded as four
ranges '[', 'a'..'z', ']', causing all a-z characters to get mapped to ']'. This
form is occasionally used in shell scripts, including the syscalltbl.sh script
used to build linux.
---
Previously, this would not work properly and not be let through the
sanity check.
This is a dirty hack until the next iteration where I'll clean up the
data structures and make this saner.
If you look at GNU coreutils, they do not support the mappings
$ echo "1234abc" | tr "[:alnum:]" "[:upper:]"
$ echo "ABCabc" | tr -c "[:upper:]" "[l*]"
to only give a few examples. This commit broadens the scope of tr(1)
as far as humanly possible to map between classes and non-classes,
making tr a usable tool and actually fulfilling user expectations.
Posix really is of no help here as it still kind of assumes the
fixed ASCII table instead of complex Unicode code points or even
Grapheme clusters.
Okay, it took me a while and another look at the Posix spec to see that
I have been dealing with squeezing in a way too complicated way.
What just needed to be done is before doing the final write to deploy
the squeeze-check. We actually do not need this atomically complicated
squeeze check in every single edge-case. Now it should work properly.
Forgot that in case there is a second argument given with -s you
probably want to have your characters substituted.
I changed it so that shortly before "deploying" we check if the
"to be written"-Rune is equal to the last Rune, and proceed as
needed.
Get rid of the !!()-constructs and use ret where available (or introduce it).
In some cases, there would be an "abort" on the first fshut-error, but we want
to close all files and report all warnings and then quit, not just the warning
for the first file.
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.
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
Interface and function as proposed by cls.
The reasoning behind this function is that cls expressed his
interest to keep memory allocation out of libutf, which is a
very good motive.
This simplifies the function a lot and should also increase the
speed a bit, but the most important factor here is that there's
no malloc anywhere in libutf, making it a lot smaller and more
robust with a smaller attack-surface.
Look at the paste(1) and tr(1) changes for an idiomatic way to
allocate the right amount of space for the Rune-array.
Interface as proposed by cls, but internally rewritten after a few
considerations.
The code is much shorter and to the point, aligning itself with other
standard functions. It should also be much faster, which is not bad.
Equivalence classes are a hard matter and there's still no "standard"
way to solve the issue.
Previously, tr would just skip those classes, but it's much
better when it resolves a [=c=] to a normal c instead of treating
it as a literal.
Also, reflect recent changes in the manpage (octal escapes) and fix
the markup in some areas.
This is one aspect which I think has blown up the complexity of many
tr-implementations around today.
Instead of complicating the set-theory-based parser itself (he should
still be relying on one rune per char, not multirunes), I added a
preprocessor, which basically scans the code for upcoming '\'s, reads
what he finds, substitutes the real character onto '\'s index and shifts
the entire following array so there are no "holes".
What is left to reflect on is what to do with octal sequences.
I have a local implementation here, which works fine, but imho,
given tr is already so focused on UTF-8, we might as well ignore
POSIX at this point and rather implement the unicode UTF-8 code points,
which are way more contemporary and future-proof.
Reading in \uC3A4 as a an array of 0xC3 and 0xA4 is not the issue,
but I'm still struggling to find a way to turn it into a well-formed
byte sequence. Hit me with a mail if you have a simple solution for
that.
It's standard behaviour to map a whole class of matched objects
to the last element of a given simple set2 instead of just passing
it through.
Also, error out more strictly when the user gives us bogus sets.
tr(1) always used to be a saddening part of sbase, which was
inherently broken and crufted.
But to be fair, the POSIX-standard doesn't make it very simple.
Given the current version was unfixable and broken by design, I
sat down and rewrote tr(1) very close to the concept of set theory
and the POSIX-standard with a few exceptions:
- UTF-8: not allowed in POSIX, but in my opinion a must. This
finally allows you to work with UTF-8 streams without
problems or unexpected behaviour.
- Equivalence classes: Left out, even GNU coreutils ignore them
and depending on LC_COLLATE, which sucks.
- Character classes: No experiments or environment-variable-trickery.
Just plain definitions derived from the POSIX-
standard, working as expected.
I tested this thoroughly, but expect problems to show up in some
way given the wide range of input this program has to handle.
The only thing left on the TODO is to add support for literal
expressions ('\n', '\t', '\001', ...) and probably rethinking
the way [_*n] is unnecessarily restricted to string2.
It actually makes the binaries smaller, the code easier to read
(gems like "val == true", "val == false" are gone) and actually
predictable in the sense of that we actually know what we're
working with (one bitwise operator was quite adventurous and
should now be fixed).
This is also more consistent with the other suckless projects
around which don't use boolean types.
- Added support for character ranges ( a-z )
- Added support for complementary charset ( -c ), only in delete mode
- Added support for octal escape sequences
- Unicode now only works when there are no octal escape sequences,
otherwise behavior is not predictable at first sight.
- tr now supports null characters in the input
- Does not yet have support for character classes ( [:upper:] )
Pieter Kockx
Laslo Hunhold
sin
Michael Forney
Hiltjo Posthuma
Evan Gates
Adria Garriga