This addresses issue #40800 and a couple other problems I noticed while
trying to reproduce that one.
The original issue is just a missing cast to void* on the args of
__builtin___clear_cache(), and clang is picky about the implicit cast
between what it considers to be char of different signedness. Original
report is from MacOS but it's also reproducible on other clang targets.
The cmake-based original build system for equix and hashx was a handy
way to run tests, but it suffered from some warnings due to incorrect
application of include_directories().
And lastly, there were some return codes from hashx_exec() that get
ignored on equix when asserts are disabled. It bugged me too much to
just silence this with a (void) cast, since even though this is in the
realm of low-likelyhood programming errors and not true runtime errors, I
don't want to make it easy for the hashx_exec() wrappers to return
values that are dangerously wrong if an error is ignored. I made sure
that even if asserts are disabled, we return values that will cause the
solver and verifier to both fail to validate a potential solution.
Signed-off-by: Micah Elizabeth Scott <beth@torproject.org>
Here is the report:
*** CID 1531835: Resource leaks (RESOURCE_LEAK)
/src/test/test_crypto_slow.c: 683 in test_crypto_equix()
677
678 /* Solve phase: Make sure the test vector matches */
679 memset(&output, 0xa5, sizeof output);
680 equix_result result;
681 result = equix_solve(solve_ctx, challenge_literal,
682 challenge_len, &output);
>>> CID 1531835: Resource leaks (RESOURCE_LEAK)
>>> Variable "solve_ctx" going out of scope leaks the storage it points to.
Signed-off-by: David Goulet <dgoulet@torproject.org>
This is a change intended for 0.4.7 maintenance as well as main.
The CI builds use Debian Buster which is now end of life, and I was
experiencing inconsistent CI failures with accessing its security update
server. I wanted to update CI to a distro that isn't EOL, and Bullseye
is the current stable release of Debian.
This opened up a small can of worms that this commit also deals with.
In particular there's a docker engine bug that we work around by
removing the docker-specific apt cleanup script if it exists, and
there's a new incompatibility between tracing and sandbox support.
The tracing/sandbox incompatibility itself had two parts:
- The membarrier() syscall is used to deliver inter-processor
synchronization events, and the external "userspace-rcu"
data structure library would make assumptions that if membarrier
is available at initialization it always will be. This caused
segfaults in some cases when running trace + sandbox. Resolved this
by allowing membarrier entirely, in the sandbox.
- userspace-rcu also assumes it can block signals, and fails
hard if this can't be done. We already include a similar carveout
to allow this in the sandbox for fragile-hardening, so I extended
that to cover tracing as well.
Addresses issue #40799
Signed-off-by: Micah Elizabeth Scott <beth@torproject.org>
This exposes the new fallback behavior in hashx via a new AUTOBOOL
configuration option, available to both clients and services. The
default should be fine for nearly everyone, but it might be necessary
to enable or disable the compiler manually for diagnostic purposes.
Signed-off-by: Micah Elizabeth Scott <beth@torproject.org>
This change adapts the hs_pow layer and unit tests to API changes
in hashx and equix which modify the fault recovery responsibilities
and reporting behaivor.
This and the corresponding implementation changes in hashx and equix
form the fix for #40794, both solving the segfault and giving hashx a
way to report those failures up the call chain without them being
mistaken for a different error (unusable seed) that would warrant a
retry.
To handle these new late compiler failures with a minimum of fuss or
inefficiency, the failover is delegated to the internals of hashx and
tor needs only pass in a EQUIX_CTX_TRY_COMPILE flag to get the behavior
that tor was previously responsible for implementing.
Signed-off-by: Micah Elizabeth Scott <beth@torproject.org>
This change adapts Equi-X to the corresponding HashX API changes that
added HASHX_TRY_COMPILE. The new regularized HashX return codes are
reflected by revised corresponding Equi-X return codes.
Both solve and verify operations now return an error/success code, and a
new equix_solutions_buffer struct includes both the solution buffer
and information about the solution count and hashx implementation.
With this change, it's possible to discern between hash construction
failures (invalid seed) and some external error like an mprotect()
failure.
Signed-off-by: Micah Elizabeth Scott <beth@torproject.org>
This is an API breaking change to hashx, which modifies the error
handling strategy. The main goal here is to allow unproblematic
recovery from hashx_compile failures.
hashx_alloc can no longer fail for reasons other than memory
allocation. All platform-specific compile failures are now reported via
hashx_make(), in order to both allow later failure and avoid requiring
users of the API to maintain and test multiple failure paths.
Note that late failures may be more common in actual use than early
failures. Early failures represent architectures other than x86_64 and
aarch64. Late failures could represent a number of system configurations
where syscalls are restricted.
The definition of a hashx context no longer tries to overlay storage for
the different types of program, and instead allows one context to always
contain an interpretable description of the program as well as an optional
buffer for compiled code.
The hashx_type enum is now used to mean either a specific type of hash
function or a type of hashx context. You can allocate a context for use
only with interpreted or compiled functions, or you can use
HASHX_TRY_COMPILE to prefer the compiler with an automatic fallback on
the interpreter. After calling hashx_make(), the new hashx_query_type()
can be used if needed to determine which implementation was actually
chosen.
The error return types have been overhauled so that everyone uses the
hashx_result enum, and seed failures vs compile failures are always
clearly distinguishable.
Signed-off-by: Micah Elizabeth Scott <beth@torproject.org>
This is a minimal portion of the fix for tor issue #40794, in which
hashx segfaults due to denial of mprotect() syscalls at runtime.
Prior to this fix, hashx makes the assumption that if the JIT is
supported on the current architecture, it will also be usable at
runtime. This isn't true if mprotect fails on linux, which it may for
various reasons: the tor built-in sandbox, the shadow simulator, or
external security software that implements a syscall filter.
The necessary error propagation was missing internally in hashx,
causing us to obliviously call into code which was never made
executable. With this fix, hashx_make() will instead fail by returning
zero.
A proper fix will require API changes so that callers can discern
between different types of failures. Zero already means that a program
couldn't be constructed, which requires a different response: choosing a
different seed, vs switching implementations. Callers would also benefit
from a way to use one context (with its already-built program) to
run in either compiled or interpreted mode.
Signed-off-by: Micah Elizabeth Scott <beth@torproject.org>
The code style in equix and hashx sometimes uses bitwise operators
in place of logical ones in cases where it doesn't really matter
either way. This sometimes annoys our static analyzer tools.
Signed-off-by: Micah Elizabeth Scott <beth@torproject.org>
This is an additional test case for test_sandbox that runs a small
subset of test_crypto_equix() inside the syscall sandbox, where
mprotect() is filtered.
It's reasonable for the sandbox to disallow JIT. We could revise this
policy if we want, but it seems a good default for now. The problem
in issue 40794 is that both equix and hashx need improvements in their
API to handle failures after allocation time, and this failure occurs
while the hash function is being compiled.
With this commit only, the segfault from issue 40794 is reproduced.
Subsequent commits will fix the segfault and revise the API.
Signed-off-by: Micah Elizabeth Scott <beth@torproject.org>
