It was impossible to enter a file:// URL because it got interpreted
as a 'f' command without any warning about invalid argument. The same
thing happened with relative addresses beginning with 'b' or 'f'.
Currently, followup requests to both INPUT and SENSITIVE INPUT responses are
omitted from history. However, the intended use of INPUT is for searching, and
there is SENSITIVE INPUT for sensitive information, so omitting regular INPUT
from history doesn't make much sense and only makes searching annoying as one
needs to re-enter the query when going back.
I wanted to add an optional [N] to 'd' and '|' to download/pipe the
Nth link instead of the current page. After getting the first one done,
I found out that '|' already has this option, so I fixed the relative
help line without changing that command.
All in all I don't know how useful this patch is, considering that
'[N]|' was already there. Feel free to discard it and just keep the
documentation change.
Cheers!
Previously, the last character of the provided input would not be read.
For example, `gmni -d termfmt gemini://godocs.io/-/search` would search
for "termfm".
It is common practice to quit programs reading from stdin by sending
EOF to their input. This commit ensures such an exit is functionally
equivalent to a regular exit via the 'q' command.
due to the wrong handling in download_resp() files may get
broken (last chunk missing) and file descriptors have not
been closed correctly.
Additionally we now allow downloading of local ressources
as well - for what its worth.
I've noticed headers are not displayed in some cases with -i/-I
specified. For example:
echo "printf" | gmni -i gemini://drewdevault.com/cgi-bin/man.sh
Output: empty
echo "printf" | gmni -IL gemini://drewdevault.com/cgi-bin/man.sh -
Output: 10 Search for a POSIX man page
INPUT header is here, but no REDIRECT header appeared.
The reason is headers processing is done after responses dispatch. So
some responses (redirect and input) are processed and dropped before.
Patch makes this logic a bit clearer imho: print response header before
any processing if mode is not OMIT_HEADERS and then process response body if mode
is not ONLY_HEADERS. It also deduplicates header printing as a bonus.
The host variable is freed too early. If a client certificate is not
found, the later error message in the
GEMINI_STATUS_CLASS_CLIENT_CERTIFICATE_REQUIRED case uses the freed host
variable to produce an incorrect openssl command. This fix just delays
the free to after the switch statement.
Test case:
gmnlm gemini://feeds.drewdevault.com
Prior:
The following OpenSSL command will generate a certificate for this host:
openssl req -x509 -newkey rsa:4096 \
-keyout /home/andrew/.local/share/gmni/certs/Ú-=öU.key \
-out /home/andrew/.local/share/gmni/certs/Ú-=öU.crt \
-days 36500 -nodes
Now:
The following OpenSSL command will generate a certificate for this host:
openssl req -x509 -newkey rsa:4096 \
-keyout /home/andrew/.local/share/gmni/certs/feeds.drewdevault.com.key \
-out /home/andrew/.local/share/gmni/certs/feeds.drewdevault.com.crt \
-days 36500 -nodes
Hi!
Static libraries don't need execution perms. I know this is very minor,
but best practice is to give everything the least amount of necessary
privileges. Thus, I propose changing the install command to use 644
perms.
Yours,
Ariadna
This is only supported with gmni for now - gmnlm support will come
later. A limitation with BearSSL prevents us from doing automated
certificate generation for now, unfortunately.
Previously the message was never displayed to users
leaving them with a simple "Error: certificate is untrusted".
This also fixes the display of line numbers in the message.
On systems using dsymutil the check for the "-g" flag was failing not
because the compiler didn't provide it but because of `/dev/null`
being used as output file.
Those changes fix the following compilation errors on OpenBSD:
src/tofu.c:128:28: error: format specifies type 'long' but the argument has type
'time_t' (aka 'long long') [-Werror,-Wformat]
"SHA-512", fingerprint, expires);
src/gmnlm.c:341:31: error: missing sentinel in function call
[-Werror,-Wsentinel]
execlp("sh", "sh", "-c", cmd);
^
, NULL
On some systems dirname uses a static string for its return value, so
we were calling mkdirs recursively on a string that was continuosly changing.
A check was also added after the `snprintf` to make sure there's no
information loss since there is no limit to the length of the string
returned by `get_data_pathfmt`.
Closes#48.
In file included from .../include/string.h:495,
from src/gmnlm.c:11:
In function ‘strncpy’,
inlined from ‘do_prompts’ at src/gmnlm.c:627:3:
...glibc-2.31-dev/include/bits/string_fortified.h:106:10:
error: ‘__builtin_strncpy’ specified bound 1024 equals
destination size [-Werror=stringop-truncation]
return __builtin___strncpy_chk (__dest, __src, __len, __bos (__dest));
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In function ‘strncpy’,
inlined from ‘do_prompts’ at src/gmnlm.c:612:3:
...glibc-2.31-dev/include/bits/string_fortified.h:106:10:
error: ‘__builtin_strncpy’ specified bound 1024 equals destination size
[-Werror=stringop-truncation]
return __builtin___strncpy_chk (__dest, __src, __len, __bos (__dest));
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Signed-off-by: William Casarin <jb55@jb55.com>
dirname has two main problems:
1. It could change in place the string that is passed to it.
2. It uses a static string for its return value, so one should copy it
somewhere else as soon as possible to avoid subsequent calls to
dirname to corrupt his data (see #48).
We avoid 1. passing a copy of `path` to dirname and 2. copying it's
return value into `dname`.
The '-D' option of `install` is not implemented on some systems that
don't use GNU install. Its uses were replaced with `mkdir -p` commands
to eventually provide the missing folders.
