#6338 introduces regression when building with LibreSSL (3.5.0).
```
cc -Wno-unused-result -Wsign-compare -Wunreachable-code -DNDEBUG -O2 -pipe -g -fPIC -O2 -pipe -g -O2 -pipe -g -O2 -pipe -fPIC -Isrc/borg/crypto -I/usr/local/include/python3.9 -c src/borg/crypto/low_level.c -o /tmp/ports/pobj/borgbackup-1.2.1/borg-eec359cf228caf00d9c72bde07bf939872e9d3fa/temp.openbsd-7.1-amd64-3.9/src/borg/crypto/low_level.o
src/borg/crypto/low_level.c:12439:48: error: use of undeclared identifier 'EVP_chacha20_poly1305'; did you mean 'EVP_aead_chacha20_poly1305'?
__pyx_v_self->__pyx_base.__pyx_base.cipher = EVP_chacha20_poly1305;
^~~~~~~~~~~~~~~~~~~~~
EVP_aead_chacha20_poly1305
/usr/include/openssl/evp.h:1161:17: note: 'EVP_aead_chacha20_poly1305' declared here
const EVP_AEAD *EVP_aead_chacha20_poly1305(void);
^
1 error generated.
```
Unfortunately `EVP_aead_chacha20_poly1305`, offered by LibreSSL, is not
a drop in replacement for `EVP_chacha20_poly1305`. More info on the
first can be found at https://man.openbsd.org/EVP_AEAD_CTX_init.3.
The key argument being sent to hashindex_get and hashindex_set by
multiple functions is a different signedness from what the functions
expect. This resolves the issue by changing the key type in the
unpack_user struct to unsigned char.
The value argument of hashindex_set is causing harmless pointer type
mismatches. This resolves the issue by changing the type to void*
which silences these types of warnings.
This resolves a compiler warning from the generated code that
resulted from a comparison of two local variables of different
signedness. The issue is resolved by changing the type of both
to int since this seems like the safest choice available.
All these are unsupported since long.
Newer versions of LibreSSL have gained chacha20-poly1305 support,
but still lack aes256-ocb support.
Also they have the HMAC_CTX_new/free api now.
docs: openssl >= 1.1.1 is required now
anything older is out of support anyway.
The generated source code was producing a compiler warning due to
the pointers differing in constness. The called function expects
a non-const pointer while the generated code produces a const pointer
via a cast. This changes the cast to drop 'const' to make the compiler
happy.
docs: clarify on-disk order and size of log entry fields
The order of the fields of a log entry on disk is CRC32 first, the docs had the size first.
I tried to make this list similar to the HashIndex struct description.
Thanks to Andrey Bienkowski (@hexagonrecursion) for reporting this and writing reproducer code.
Changes:
- use different, randomly (but recognizably) named temp files while writing (securely made by os.mkstemp())
- make sure temp files are cleaned up in normal and error conditions
- SyncFile can now get corresponding pair of path + open os-level fd
- cleaned up: fd now means os-level fd, f means python-file-like object
- fixed a caller of SaveFile
