true asynchronicity is actually fairly useless, as it's unlikely that
both Stores in a Channel use IMAP, and both host resolutions take
particularly long - the main objective is imposing the Timeout setting.
however, we can't just use setjmp()+alarm(), as longjmp()ing out of
getaddrinfo() is undefined, as it may for example free() just at the
wrong time. so we go for the real thing.
this implementation just fork()s out a process which uses getaddrinfo()
(or gethostbyname()) per lookup. this isn't particularly scalable, but
as we don't expect a lot of lookups, it seems adequate.
this generally went unnoticed, as the tunnel usually terminates right
before we exit anyway. however, if multiple Channels are synced, it may
become visible.
this is a "shotgun" implementation, where the main loop just reaps all
unclaimed children.
arguably, it would be cleaner if each socket actually tracked its own
process, but getting synchronous kills+waits right is tricky, so we
continue to pretend that there is no process as far as the socket layer
is concerned.
poll()/select() are not restartable, so they need EINTR handling now
that SIGCHLD is actually delivered.
the mainloop-based refresh timer keeps spinning even if there is no
update. overload stats_steps to signal whether a refresh is needed.
amends 8fbc4323.
this got lost in d5a5da947.
this also simplifies a nested condition, where the logic has previously
been, but isn't applicable any more.
REFMAIL: 87fsjloz05.fsf@wavexx.thregr.org
make a complete list of contributors who hold copyright (and those who
don't).
the case of ted is particularly interesting - he recently disclaimed
significant contributions, but it turns out that this isn't true:
the rewrite in 130664b6 incorporated significant portions of his async
IMAP patch for debian.
speaking of debian, i deleted ted's and nicolas' mention as debian
maintainers, as debian/ is basically just a (poorly maintained) mirror,
and has an own copyright file.
remove items:
- the "Unidentified socket error" should be gone since 7ba7be111
- imap_commit_cmds() is implemented since cfaa4848d
- we will never use MULTIAPPEND and FETCH BODY with multiple messages,
as that would significantly complicate matters for no tangible benefit
... and add some new thoughts.
the idea to use strings of colons for quoting patterns doesn't work, as
colons in "regular" patterns could not be quoted.
that is, Path not ending with a slash.
pedantically, this is a bugfix, as the manual already suggested that
this is possible (and at least one user got the hint, though he was
disappointed).
the IMAP driver already supports this.
it was never documented, and i can't really think of a case where
someone would actually want it, as the MUA side of things would be just
weird. additionally, the case of Path being Inbox+'/' did not work
anyway. will reconsider and fix the latter case if someone complains.
reverts 98bd2b11.
there is no need to call maildir_list_{path,inbox}() from
maildir_list_recurse(), as maildir_list_store() will do that anyway -
if it's actually requested, that is. this means that this removes the
implicit listing when Inbox is nested into Path, or vice versa. this is
actually correct, as the Patterns matching would throw away the non-
requested boxes anyway.
this finally makes us compliant with IMAP4rev1. how fitting that the
meanwhile released IMAP4rev2 demoted UTF-7 to legacy status ...
based on a patch by Georgy Kibardin <georgy@kibardin.name>.
the new one isn't an exhaustive set anymore, but a delta to the default
(or previous setting). this has the advantage that new versions aren't
automatically disabled anymore as soon as mbsync learns about them if
the user has a setting they fail to adjust.
report them as errors (not warnings), let them cause a non-zero exit
code, and in the case of trashing, prevent the subsequent expunge. the
exception are messages that just disappeared below our feet.
the purpose of these checks is preventing triggering more driver calls,
while immediate followups to already completed calls should be still
made.
note that some of the checks are strictly speaking redundant, as
chaining into a next phase will return immediately anyway when the
other side isn't ready yet. but we keep them for clarity.
this is only a partial solution for reporting changes, but it's
adequate if the goal is merely triggering a bulk action, like
re-indexing the local store.
inspired by patches posted by Yuri D'Elia <wavexx@thregr.org>.
seconds are too coarse for what we intend. technically, centisecs would
be sufficient (and more honest, given that we prefer coarse timers,
which have only tick precision), but that's a somewhat untypical unit.
in the spirit of being liberal about what we accept.
this completely re-structures the header processing loop, with the nice
side effect of eliminating the gotos.
REFMAIL: 87bkyzhoov.fsf@curie.anarc.at
... instead of reporting conversion errors directly in copy_msg_convert().
this makes it easier to autotest properly.
this reverts bc15e571 in spirit, sidestepping the problem it tried to
solve instead.
don't print the status, as the user doesn't really care whether it's
NO or BAD. more importantly, "NO LOGIN failed" is a rather misleading
thing to report.
replace the "DOM-like" model with a "streaming" model, i.e., the tokens
are now immediately processed by stateful callbacks as they are found.
this avoids plenty of allocations and copies (though not of the message
contents, yet).
- short literals are now printed even with only -Dn.
this ensures that we see all parts of, for example, LIST responses.
we use xprintf() for that, so we don't mess up the -Dn output in case
the literal contains something unexpected (we don't care with -DN).
- omitted bytes are now indicated.
- missing trailing newlines are now "pretty-printed".
- long literals are now printed in chunks as we receive them. this
makes the output messier, but it's necessary for an upcoming change,
and, on the upside, we'll get better indication when the transfer
gets stuck in the middle of a message.
this introduces a "commit" callback to the xprintf kernel, to avoid
pointlessly assembling a temporary output string prior to printing it.
one could lift the buffer limitations by introducing a "segment"
callback instead, but that would slow down xvasprintf() due to the
higher callback rate, for no good reason.
