INIT_OPTION used to initialize union option_value at compile time by
casting the default value to LIST_OF(struct option) *, which is the
type of the first member. On sparc64 and other big-endian systems
where sizeof(int) < sizeof(struct list_head *), this tended to leave
option->value.number as zero, thus messing up OPT_INT and OPT_BOOL
at least. OPT_LONG however tended to work right.
This would be easy to fix with C99 designated initializers,
but doc/hacking.txt says ELinks must be kept C89 compatible.
Another solution would be to make register_options() read the
value from option->value.tree (the first member), cast it back
to the right type, and write it to the appropriate member;
but that would still require somewhat dubious conversions
between integers, data pointers, and function pointers.
So here's a rather more invasive solution. Add struct option_init,
which is somewhat similar to struct option but has non-overlapping
members for different types of values, to ensure nothing is lost
in compile-time conversions. Move unsigned char *path from struct
option_info to struct option_init, and replace struct option_info
with a union that contains struct option_init and struct option.
Now, this union can be initialized with no portability problems,
and register_options() then moves the values from struct option_init
to their final places in struct option.
In my x86 ELinks build with plenty of options configured in, this
change bloated the text section by 340 bytes but compressed the data
section by 2784 bytes, presumably because union option_info is a
pointer smaller than struct option_info was.
(cherry picked from elinks-0.12 commit e5f6592ee2)
Conflicts:
src/protocol/fsp/fsp.c: All options had been removed in 0.13.GIT.
src/protocol/smb/smb2.c: Ditto.
Using the RFC 3546 server_name TLS extension with TLS 1.1 made
https://bugzilla.novell.com/ never respond to negotiation.
Disable TLS 1.1 with GnuTLS, like it has already been disabled with
OpenSSL. And if an SSL error is detected, disable TLS 1.2 as well.
src/network/ssl/ssl.c defined some constant arrays for passing to
GnuTLS functions. Those functions have been deprecated and their
calls are in #if 0 nowadays. Put #if 0 around the arrays as well.
The deprecated gnutls_certificate_verify_peers function returns -1
if it fails, or a set of gnutls_certificate_status_t bits; each bit
indicates some kind of problem, so the result is zero if the
certificate is OK.
The newer gnutls_certificate_verify_peers2 function returns -1 if it
fails, or 0 if it succeeds; and writes the status bits via a pointer.
When using that function, ELinks must check the status separately.
Do that.
Also, if GnuTLS is not being used, do not declare a status variable,
because leaving it unused would break the debug build:
/home/Kalle/src/elinks-0.13/src/network/ssl/socket.c: In function ‘ssl_want_read’:
/home/Kalle/src/elinks-0.13/src/network/ssl/socket.c:87: error: unused variable ‘status’
/home/Kalle/src/elinks-0.13/src/network/ssl/socket.c: In function ‘ssl_connect’:
/home/Kalle/src/elinks-0.13/src/network/ssl/socket.c:121: error: unused variable ‘status’
Avoid compilation error with GNUTLS 1.2.9:
/home/Kalle/src/elinks-0.12/src/network/ssl/ssl.c:258: error: implicit declaration of function ‘gnutls_priority_set_direct’
If the function is not available, use gnutls_set_default_priority instead.
Perhaps it'll work with bugzilla.novell.com, perhaps not.
- gnutls_handshake_set_private_extensions: Do not enable private cipher
suites that might not be supported by anything other than GNUTLS.
The GNUTLS 2.8.0 documentation notes that enabling these extensions
can cause interoperability problems.
- gnutls_set_default_priority: Explicitly disable OpenPGP certificates.
- gnutls_certificate_type_set_priority: Do not enable OpenPGP certificates.
The GNUTLS 2.8.0 documentation notes that OpenPGP certificate support
requires libgnutls-extra. Because libgnutls-extra 2.2.0 and later are
under GPLv3-or-later and thus not GPLv2 compatible, ELinks doesn't use
libgnutls-extra, so OpenPGP certificates didn't work anyway.
- gnutls_server_name_set: Do not tell the server the hostname from the URL.
This was supposed to let the server choose the appropriate certificate
for each name-based virtual host, but ELinks actually always sent just
"localhost", so it didn't work anyway. This will have to be revisited
when ELinks is changed to actually verify the subject name from the
server's certificate (ELinks bug 1024).
These changes should help ELinks negotiate SSL with bugzilla.novell.com.
[NEWS and commit message by me. --KON]
In load_uri, if there is no valid cache entry and no existing connection
for the requested URI, make one last check in case there is an
incomplete cache entry.
Documentation strings of most options used to contain a "\n" at the
end of each source line. When the option manager displayed these
strings, it treated each "\n" as a hard newline. On 80x24 terminals
however, the option description window has only 60 columes available
for the text (with the default setup.h), and the hard newlines were
further apart, so the option manager wrapped the text a second time,
resulting in rather ugly output where long lones are interleaved with
short ones. This could also cause the text to take up too much
vertical space and not fit in the window.
Replace most of those hard newlines with spaces so that the option
manager (or perhaps BFU) will take care of the wrapping. At the same
time, rewrap the strings in source code so that the source lines are
at most 79 columns wide.
In some options though, there is a list of possible values and their
meanings. In those lists, if the description of one value does not
fit in one line, then continuation lines should be indented. The
option manager and BFU are not currently able to do that. So, keep
the hard newlines in those lists, but rewrap them to 60 columns so
that they are less likely to require further wrapping at runtime.
Call stacks reported by valgrind:
==14702== at 0x80DD791: read_from_socket (socket.c:945)
==14702== by 0x8104D0C: read_more_http_data (http.c:1180)
==14702== by 0x81052FE: read_http_data (http.c:1388)
==14702== by 0x80DD69B: read_select (socket.c:910)
==14702== by 0x80D27AA: select_loop (select.c:307)
==14702== by 0x80D1ADE: main (main.c:358)
==14702== Address 0x4F4E598 is 56 bytes inside a block of size 81 free'd
==14702== at 0x402210F: free (vg_replace_malloc.c:233)
==14702== by 0x812BED8: debug_mem_free (memdebug.c:484)
==14702== by 0x80D7C82: done_connection (connection.c:479)
==14702== by 0x80D8A44: abort_connection (connection.c:769)
==14702== by 0x80D99CE: cancel_download (connection.c:1053)
==14702== by 0x8110EB6: abort_download (download.c:143)
==14702== by 0x81115BC: download_data_store (download.c:337)
==14702== by 0x8111AFB: download_data (download.c:446)
==14702== by 0x80D7B33: notify_connection_callbacks (connection.c:458)
==14702== by 0x80D781E: set_connection_state (connection.c:388)
==14702== by 0x80D7132: set_connection_socket_state (connection.c:234)
==14702== by 0x80DD78D: read_from_socket (socket.c:943)
read_from_socket() attempted to read socket->fd in order to set
handlers on it, but the socket had already been freed. Incidentally,
socket->fd was -1, which would have resulted in an assertion failure
if valgrind hadn't caught the bug first.
To fix this, add a list of weak references to sockets.
read_from_socket() registers a weak reference on entry and unregisters
it before exit. done_socket() breaks any weak references to the
specified socket. read_from_socket() then checks whether the weak
reference was broken, and doesn't access the socket any more if so.
This reverts src/{network,sched}/connection.c CVS revision 1.43,
which was made on 2003-07-03 and converted to Git commit
cae65f7941628109b51ffb2e2d05882fbbdc73ef in elinks-history.
It is pointless to check whether (c == d && c->id == d->id).
If c == d, then surely c->id == d->id, and I wouldn't be surprised
to see a compiler optimize that out.
Whereas, by taking the id as a parameter, connection_disappeared()
can check whether the pointer now points to a new struct connection
with a different id.
make_bittorrent_peer_connection() used to construct a struct uri on
the stack. This was hacky but worked nicely because the struct uri
was not really accessed after make_connection() returned. However,
since commit a83ff1f565, the struct uri
is also needed when the connection is being closed. Valgrind shows:
Invalid read of size 2
at 0x8100764: get_blacklist_entry (blacklist.c:33)
by 0x8100985: del_blacklist_entry (blacklist.c:64)
by 0x80DA579: complete_connect_socket (socket.c:448)
by 0x80DA84A: connected (socket.c:513)
by 0x80D0DDF: select_loop (select.c:297)
by 0x80D00C6: main (main.c:353)
Address 0xBEC3BFAE is just below the stack ptr. To suppress, use: --workaround-gcc296-bugs=yes
To fix this, allocate the struct uri on the heap instead, by
constructing a string and giving that to get_uri(). This string
cannot use the "bittorrent" URI scheme because parse_uri() does not
recognize the host and port fields in that. (The "bittorrent" scheme
has protocol_backend.free_syntax = 1 in order to support strings like
"bittorrent:http://beta.legaltorrents.com/get/159-noisome-beasts".)
Instead, define a new "bittorrent-peer" URI scheme for this purpose.
If the user attempts to use this URI scheme, its handler aborts the
connection with an error; but when make_bittorrent_peer_connection()
uses a bittorrent-peer URI, the handler is not called.
This change also lets get_uri() set the ipv6 flag if peer_info->ip is
an IPv6 address literal.
Reported by Witold Filipczyk.
fsp_open_session() has a bug where it does not set errno if getaddrinfo fails.
Before the bug 1013 fix, this caused an assertion failure.
After the bug 1013 fix, this caused a "Success" error message.
Now it instead causes "FSP server not found".
Replace almost all uses of enum connection_state with struct
connection_status. This removes the assumption that errno values used
by the system are between 0 and 100000. The GNU Hurd uses values like
ENOENT = 0x40000002 and EMIG_SERVER_DIED = -308.
This commit is derived from my attachments 450 and 467 to bug 1013.
It seems GnuTLS is not as good at negotiating a supported protocol as
OpenSSL is. ELinks tries to work around that by retrying with a
different protocol if the SSL library reports an error. However,
ELinks must not automatically retry POST requests where some data may
have already reached the server; POST is not a safe method in HTTP.
So instead, collect the name of the TLS-incapable server in a blacklist
when ELinks e.g. loads an HTML form from it; the actual POST can then
immediately use the protocol that worked.
It's a bit ugly that src/network/socket.c now uses
protocol/http/blacklist.h. It might be better to move the blacklist
files out of the http directory, and perhaps merge them with the
BitTorrent blacklisting code.
I am reverting all copiousoutput support because of bug 917.
This reverts commit 4dc4ea47f2.
Conflicts:
src/network/connection.h: After the original commit, the declaration
of copiousoutput_data had been changed to use the LIST_OF macro.
Also, connection.cgi had been added next to the connection.popen
member added by the original commit.
src/session/download.c: After the original commit, the definition of
copiousoutput_data had been changed to use the INIT_LIST_OF macro.
CGI scripts are distinguishable from normal files. I hope that this
fixes the bug 991. This commit also reverts the previous revert.
(cherry picked from commit 7ceba1e461)
Previously, each progress timer function registered with
start_update_progress() was directly used as the timer function of
progress.timer, so it was responsible of erasing the expired timer ID
from that member. Failing to do this could result in heap corruption.
The progress timer functions normally fulfilled the requirement by
calling update_progress(), but one such function upload_stat_timer()
had to erase the timer ID on its own too.
Now instead, there is a wrapper function progress_timeout(), which
progress.c sets as the timer function of progress.timer. This wrapper
erases the expired timer ID from progress.timer and then calls the
progress timer function registered with start_update_progress(). So
the progress timer function is no longer responsible of erasing the
timer ID and there's no risk that it could fail to do that in some
error situation.
This commit introduces a new risk though. Previously, if the struct
progress was freed while the timer was running, the (progress) timer
function would still be called, and it would be able to detect that
the progress pointer is NULL and recover from this situation. Now,
the timer function progress_timeout() has a pointer to the struct
progress and will dereference that pointer without being able to check
whether the structure has been freed. Fortunately, done_progress()
asserts that the timer is not running, so this should not occur.
Move connection.post_fd to http_post.post_fd.
Make connection.done point to the new done_http_connection(),
which calls the new done_http_post(), which closes post_fd.
So done_connection() no longer needs to do that.
Now that done_http_post() exists, a later commit can add dynamically
allocated data in struct http_post and ensure that it will be freed.
If ELinks is being linked with SSL library, use its random number
generator.
Otherwise, try /dev/urandom and /dev/prandom. If they do not work,
fall back to rand(), calling srand() only once. This fallback is
mostly interesting for the Hurd and Microsoft Windows.
BitTorrent piece selection and dom/test/html-mangle.c still use rand()
(but not srand()) directly. Those would not benefit from being
unpredictable, I think.
