The `socket_connect_one` function previously did an `exit(1)` when
encountering any errors with opening the socket. This would break
connecting to a host where multiple possible addrinfos are returned,
where the leading addrinfos are in fact impossible to connect to. E.g.
with a kernel configured without support for IPv6, the `getaddrinfo`
call may still return a hint containing an IPv6 address alongside
another hint with an IPv4 address. Creating the socket with the IPv6
address, which will cause an error, lead us to exiting early without
even trying remaining hints.
While one can argue that the user should have compiled without HAVE_IPV6
or used an appropriate DNS configuration, we can do better by simply
skipping over the current addrinfo causing an error. To do so, we split
out a new function `socket_connect_next`, which selects the next
available address info and subsequently calls `socket_connect_one`
again. When no hints remain, `sock_connect_one` will error out at that
point.
any structures may be invalid after callback invocation.
this has the side effect that the socket write callback now returns
void, like all other callbacks do.
the synchronous writing to the socket would have typically invoked the
write callback, which would flush further commands, thus recursing.
we take the easy way out and make it fully asynchronous, i.e., no data
is sent before (re-)entering the event loop.
this also has the effect that socket_write() cannot fail any more, and
any errors will be reported asynchronously. this is consistent with
socket_read(), and produces cleaner code.
this introduces a marginal performance regression: the maildir driver is
synchronous, so all messages (which fit into memory) will be read before
any data is sent. this is not considered relevant.
propagating many messages from a fast store (typically maildir or a
local IMAP server) to a slow asynchronous store could cause gigabytes of
data being buffered. avoid this by throttling fetches if the target
context reports memory usage above a configurable limit.
REFMAIL: 9737edb14457c71af4ed156c1be0ae59@mpcjanssen.nl
some servers actually bother to close down the SSL connection before
closing the socket.
this fixes the spurious "unhandled SSL error 6" messages.
REFMAIL: 20150120114805.GA17586@leeloo.kyriasis.com
the server can actually close the zlib stream before closing the socket,
so we need to accept it.
we don't do anything beyond that - the actual EOF will be signaled by
the socket, and if the server (erroneously) sends more data, zlib will
tell us about it.
REFMAIL: 1423048708-975-1-git-send-email-alex.bennee@linaro.org
zlib reports Z_BUF_ERROR when a flush is attempted without any activity
since the previous flush (if any). while this is harmless as such,
discerning the condition from genuine errors would be much harder than
avoiding the pointless flush in the first place.
REFMAIL: eb5681612f17be777bc8d138d31dd6d6@mpcjanssen.nl
the primary objective is reducing the number of small SSL packets (which
are always padded), but fewer syscalls in the non-SSL case should be
good as well.
instead of keeping the structures in an opaque array (which was a shadow
of the struct pollfd array if poll() was supported), make them directly
addressable.
this has the advantage that notifier-altering operations (mostly
en-/disabling) don't need to look up the structure by file handle each
time.
on the downside, data locality in the main loop is worse.
neither of these have any real effect on performance.
note that the structures are not allocated separately, but embedded into
the the parent structure (like sockets already were).
we should make no assumptions about the layout of OpenSSL's certificate
store, in particular when they are wrong. so copy the interesting part
instead of "deep-linking" into it.
this code is uglier than it should be, but OpenSSL's extensive use of
macros to manage data types would force us to include the ssl headers
into our headers otherwise, which would be even uglier.
REFMAIL: <545442CC.9020400@nodivisions.com>
instead of using a callback which messes with the certificate chain
verification, simply let OpenSSL ignore errors during that phase and
check the result only afterwards.
the combinations of the various options made quite a mess. additionally,
'RequireSSL no' is inherently insecure - "use SSL if available" is plain
stupid.
the old options are still accepted, but will elicit a warning.
it doesn't belong there - it's a property of imap_server_conf_t.
the port setup is now done while reading the config.
this makes socket.[hc] imap-agnostic.
always use getsockopt() to query the meaning of POLLERR, rather than
reporting "Unidentified socket error".
this is unlikely to have any effect when using select(), as that one
pretty much never signals exceptional conditions.
turns out that poll() may (and on linux does) signal POLLERR on
connection failure. this is unlike select(), which is specified to
signal write readiness in every case.
consequently, check whether we are connecting before checking for
POLLERR.
leave all the hard work to OpenSSL. this has several consequences:
- certificate chain validation actually works instead of throwing
around error 20
- the interactive approval is gone. i don't expect it to be useful
anyway, as mbsync is mostly a batch tool
- the code is much shorter
we did not check a valid certificate's subject at all so far.
this is no problem if the certificate file contains only exactly the
wanted host's certificate - before revision 04fdf7d1 (dec 2000, < v0.4),
this was even enforced (more or less - if the peer cert had been
signed directly by a root cert, it would be accepted as well).
however, when the file contains root certificates (like the system-wide
certificate file typically does), any host with a valid certificate
could pretend to be the wanted host.
unless an info message is explictly marked as a continuation, it must
terminate any pending line (typically the progress information) first.
debug output is not affected, as it is mutually exclusive with info
output, and no debug lines are left unterminated outside clear scopes.
- introduce sys_error() and use it instead of perror() and
error(strerror()) in all expected error conditions
- perror() is used only for "something's really wrong with the system"
kind of errors
- file names, etc. are quoted if they are not validated yet, so e.g. an
empty string becomes immediately obvious
- improve and unify language
- add missing newlines
- asynchronous sockets using an event loop
- connect & starttls have completion callback parameters
- callbacks for notification about filled input buffer and emptied
output buffer
- unsent imap command queue
- used when
- socket output buffer is non-empty
- number of commands in flight exceeds limit
- last sent command requires round-trip
- command has a dependency on completion of previous command
- trashnc is tri-state so only a single "scout" trash APPEND/COPY is
sent at first. a possibly resulting CREATE is injected in front of
the remaining trash commands, so they can succeed (or be cancel()d
if it fails).
- queue's presence necessitates imap_cancel implementation
