there are still some discussions going on about tweaking the code, but
at least thing brings us to the point of having something working in
the repository. hopefully the remaining major hashes (md5,sha) will
follow soon.
unfortunately, a large portion of programs which call crypt are not
prepared for its failure and do not check that the return value is
non-null before using it. thus, always "succeeding" but giving an
unmatchable hash is reportedly a better behavior than failing on
error.
it was suggested that we could do this the same way as other
implementations and put the null-to-unmatchable translation in the
wrapper rather than the individual crypt modules like crypt_des, but
when i tried to do it, i found it was making the logic in __crypt_r
for keeping track of which hash type we're working with and whether it
succeeded or failed much more complex, and potentially error-prone.
the way i'm doing it now seems to have essentially zero cost, anyway.
the new version is largely the work of Solar Designer, with minor
changes for integration with musl. compared to the old code, text size
is reduced by about 7k, stack space usage by about 70k, and
performance is greatly improved by avoiding expensive calculation of
constant tables on each run.
this version also adds support for extended des-based password hashes,
which allow for unlimited key (password) length and configurable
iteration counts.
i've also published the interface for crypt_r in a new crypt.h header.
especially since this is not a standard interface, i did not feel
compelled to match the glibc abi for the crypt_data structure. the
glibc structure is way too big to allocate on the stack; in fact it's
so big that the first usage may cause the main thread to exceed its
pre-committed stack size of 128k and thus could cause the program to
crash even on systems with overcommit disabled. the only legitimate
use of crypt_data for crypt_r is to store the hash string to return,
so i've reserved 256 bytes, which should be more than sufficient
(longest known password hashes are ~60 characters, and beyond that is
possibly even exceeding some implementations' passwd file field size
limit).
apparently some packages see stropts.h and want to be able to use
this. the implementation checks that the file descriptor is valid by
using fcntl/F_GETFD so it can report an error if not (as specified).
i did some testing trying to switch malloc to use the new internal
lock with priority inheritance, and my malloc contention test got
20-100 times slower. if priority inheritance futexes are this slow,
it's simply too high a price to pay for avoiding priority inversion.
maybe we can consider them somewhere down the road once the kernel
folks get their act together on this (and perferably don't link it to
glibc's inefficient lock API)...
as such, i've switch __lock to use malloc's implementation of
lightweight locks, and updated all the users of the code to use an
array with a waiter count for their locks. this should give optimal
performance in the vast majority of cases, and it's simple.
malloc is still using its own internal copy of the lock code because
it seems to yield measurably better performance with -O3 when it's
inlined (20% or more difference in the contention stress test).
I'm not sure if it's legal for wordexp to modify this field, but this
is the only easy/straightforward fix, and applications should not
care. if it's an issue, i can work out a different (but more complex)
solution later.
note that regardless of the name used, basename is always conformant.
it never takes on the bogus gnu behavior, unlike glibc where basename
is nonconformant when declared manually without including libgen.h.
setrlimit is supposed to be per-process, not per-thread, but again
linux gets it wrong. work around this in userspace. not only is it
needed for correctness; setxid also depends on the resource limits for
all threads being the same to avoid situations where temporarily
unlimiting the limit succeeds in some threads but fails in others.
we cannot report failure after forking, so the idea is to ensure prior
to fork that fd 0,1,2 exist. this will prevent dup2 from possibly
hitting a resource limit and failing in the child process. fcntl
rather than dup2 is used prior to forking to avoid race conditions.
clean and simple, but fails when the caller does not have permissions
to open the file for reading or when /proc is not available. i may
replace this with a full implementation later, possibly leaving this
version as an optimization to use when it works.
with datagram sockets, depending on fprintf not to flush the output
early was very fragile; the new version simply uses a small fixed-size
buffer. it could be updated to dynamic-allocate large buffers if
needed, but i can't envision any admin being happy about finding
64kb-long lines in their syslog...
sadly the C language does not specify any such implicit conversion, so
this is not a matter of just fixing warnings (as gcc treats it) but
actual errors. i would like to revisit a number of these changes and
possibly revise the types used to reduce the number of casts required.
