LibreSSL is now closer to OpenSSL 1.1 than OpenSSL 1.0. According to
https://undeadly.org/cgi?action=article;sid=20220116121253, this is the
intention of OpenBSD developers.
According to #40630, many special cases are needed to compile Tor against
LibreSSL 3.5 when using Tor's OpenSSL 1.0 compatibility mode, whereas only a
small number of #defines are required when using OpenSSL 1.1 compatibility
mode. One additional workaround is required for LibreSSL 3.4 compatibility.
Compiles and passes unit tests with LibreSSL 3.4.3 and 3.5.1.
For some syscalls the kernel ABI uses 32 bit signed integers. Whether
these 32 bit integer values are sign extended or zero extended to the
native 64 bit register sizes is undefined and dependent on the {arch,
compiler, libc} being used. Instead of trying to detect which cases
zero-extend and which cases sign-extend, this commit uses a masked
equality check on the lower 32 bits of the value.
The chown/chmod/rename syscalls have never existed on AArch64, and libc
implements the POSIX functions via the fchownat/fchmodat/renameat
syscalls instead.
Add new filter functions for fchownat/fchmodat/renameat, not made
architecture specific since the syscalls exists everywhere else too.
However, in order to limit seccomp filter space usage, we only insert
rules for one of {chown, chown32, fchownat} depending on the
architecture (resp. {chmod, fchmodat}, {rename, renameat}).
New glibc versions not sign-extending 32 bit negative constants seems to
not be a thing on AArch64. I suspect that this might not be the only
architecture where the sign-extensions is happening, and the correct fix
might be instead to use a proper 32 bit comparison for the first openat
parameter. For now, band-aid fix this so the sandbox can work again on
AArch64.
Not revalidating keys on every fork speeds up make test from about 45 seconds
to 10 seconds with OpenSSL 1.1.1n and from 6 minutes to 10 seconds with OpenSSL
3.0.2.
