The line-based parsing can be a bit awkward when wanting to peek
ahead, (say, for folded header values), but it's so convenient
to be able to trust that a string terminator exists on every
line so it cleans up the code considerably.
Looks like we can copy in a hash-table implementation, (from cairo,
say), and then a few _ascii_ functions from glib, (we'll need to
switch a few current uses if things like isspace, etc. to locale-
independent versions as well). So not too hard to free ourselves
of glib for now, (until we add GMime back in later, of course).
Now completing the process of making this function "our own".
The documentation is deleted here, because we already have
the documentation we want in notmuch-private.h.
The original code expected this to be set by running configure.
We'll just manually set it here for now. This isn't as portable
as if we were doing some compile-time examination of the current
system, but I don't need portability now.
When someone comes along that wants to port notmuch to another
system, they will already have all the #ifdefs in place and
will simply need to add the appropriate machinery to set the
defines.
This change is gratuitous. For now, notmuch is still linking
against glib, so I don't have any requirement to remove this,
(unlike the last few changes where good taste really did
require the changes).
The motivation here is two-fold:
1. I'm considering switching away from all glib-based allocation
soon so that I can more easily verify that the memory management
is solid. I want valgrind to say "no leaks are possible" not
"there is tons of memory still allocated, but probably reachable
so who knows if there are leaks or not?". And glib seems to make
that impossible.
2. I don't think there's anything performance-sensitive about the
allocation here. (In fact, if there is, then the right answer
would be to do this parsing without any allocation whatsoever.)
While this is surely one of the most innocent typedefs, it still
annoys me to have basic types like 'int' re-defined like this.
It just makes it harder to copy the code between projects, with
very little benefit in readability.
For readability, predicate functions and variables should be
obviously Boolean-natured by their actual *names*.
That's got to be one of the hardest macro names to read, ever,
(it's phrased with an implicit negative in the condition,
rather than something simple like "assert").
Plus, it's evil, since it's a macro with a return in it.
And finally, it's actually *longer* than just typing "if"
and "return". So what's the point of this ugly idiom?
We can't rely on any gmime-internal headers, (and fortunately we
don't need to). We also aren't burdened with any autconf machinery
so don't reference any of that.
We're sucking in one gmime implementation file just to get the
piece that parses an RFC 822 date, because I don't want to go
through the pain of replicating that.
Since we're currently just trying to stitch together In-Reply-To
and References headers we don't need that much sophistication.
It's when we later add full-text searching that GMime will be
useful.
So for now, even though my own code here is surely very buggy
compared to GMime it's also a lot faster. And speed is what
we're after for the initial index creation.
This is the beginning of the notmuch library as well, with its
interface in notmuch.h. So far we've got create, open, close, and
add_message (all with a notmuch_database prefix).
The current add_message function has already been whittled down from
what we have in notmuch-index-message to add only references,
message-id, and thread-id to the index, (that is---just enough to do
thread-linkage but nothing for full-text searching).
The concept here is to do something quickly so that the user can get
some data into notmuch and start using it. (The most interesting stuff
is then thread-linkage and labels like inbox and unread.) We can
defer the full-text indexing of the body of the messages for later,
(such as in the background while the user is reading mail).
The initial thread-stitching step is still slower than I would like.
We may have to stop using libgmime for this step as its overhead is
not worth it for the simple case of just parsing the message-id,
references, and in-reply-to headers.
This was for some time testing, (to see how fast xapian could be
if we were strictly adding documents and not doing any other IO
or computation). The answer is that xapian is quite fast, (on
the order of 1000 documents per second).
Of course, there's not much that this program does yet. It's got
some structure for some sub-commands that don't do anything. And
it has a main command that prints some explanatory text and then
counts all the regular files in your mail archive.
These were more significant than the previous leak because these were
in the loop and leaking memory for every message being parsed. It
turns out that g_hash_table_new should probably be named
g_hash_table_new_and_leak_memory_please. The actually useful function
is g_hash_table_new_full which lets us pass a free function, (to free
keys when inserting duplicates into the hash table). And after all,
weeding out duplicates is the only reason we are using this hash table
in the first place.
It almost goes without saying, valgrind found these leaks.
This was a single object in main outside any loops, so there was
no impact on performance or anything, but obviously we still want
to patch this.
Of course, valgrind gets the credit for seeing this.
When looking for a trailing ':' to introduce a quotation we peek at
the last character before a newline. But for blank lines, that's not
where we want to look. And when the first line in our buffer is a
blank line, we're underrunning our buffer. The fix is easy---just
bail early on blank lines since they have no terms anyway.
Thanks to valgrind for pointing out this error.
Previously, we used as the thread-id the message-id of the first
message in the thread that we happened to find. In fact, this is a
totally arbitrary identifier, so it might as well be random. And an
advantage of actually using a random identifier is that we now have
fixed-length thead identifiers, (and the way is open to even allow
abbreviated identifiers like git does---though we're less likely to
show these identifiers to actual users).
Instead of always showing the overall rate, we wait until the end
to show that. Then, on incremental updates we show the rate over the
last increment. This makes it much easier to actually watch what's
happening, (and it's easy to see the efect of xapian's internal
10,000 document flush).
We could (and probably should) reparse and index all the headers from
the embedded message, but I'm not choosing to do that now---I'm just
indexing the body of the embedded message.
As can be seen here, there are not a lot of differences. I've verified
this by using sup-sync to import a month of mail from the sup mailing
list, and comparing the database term-by-term, value-by-value, and
data-by-data with that created by notmuch. There are no differences
other than those documented here.
Here's another change which I'm making for sup compatibility against
my better judgment. It seems that sup never indexes content from
mime parts with content-disposition of attachment. But these
attachments are often very indexable, (for example, the first one
I encountered was a small shell script).
So I'll have to think a bit about whether or not I want to revert
this commit. To do this properly we would really want to distinguish
between attachments that are indexable, (such as text), and those
that aren't, (such as binaries). I know the mime-type alone isn't
alwas sufficient here as even this little plaintext shell script
was attached as octet-stream.
And if we wanted to get really fancy we could run things like antiword
to generate text from non-text attachments and index their output.
Instead of using the recursive "foreach" method, we implement our
own recursive function. This allows us to ignore the signature
component of a multipart/signed message, (which we certainly
don't need to index).
Ignoring this whitespace seems like a good idea to me, but it's
interfering with my comparisons with sup since sup doesn't do this.
This might be a commit worth dropping in the future since it exists
only for pedantic consistency with sup and not for any reason of its
own.
Here's another instance where I "knew" gmime must have support for
some functionality, but not finding it, I rolled my own. Now that
I found g_mime_references_decode I'm glad to drop my ugly code.
This cleans up some old code that was very ugly, (separately opening
the mail file and seeking to the end of the headers to parse the
body). I knew gmime must have had support for transparently decoding
mime content, but I just couldn't find it previously.
Note: Multipart and MultipartSigned parts are not handled yet.
Things are quite happy now. The few differences I see with sup are:
1. sup forces email address domains to lowercase, (I don't think I care)
2. sup and notmuch disagree on ordering of multiple thread_id values
(another thing that's of no concern)
We are still doing one thing wrong when a message belongs to multiple
threads. We've got a nice comma-separated thread-value just like sup,
but then we're also putting in a comma-separated thread-term where
sup does multiple thread terms. That should be an easy fix.
Beyond that, sup and notmuch are still disagreeing on the term lists
for some messages, (I think attachment vs. inline content-disposition
is at least one piece of this). But there are likley still differences
in the heuristics for which chunks of the message body to index. I'll
be looking into this more.
Currently we're looking up all parents (based on In-reply-to and
References header) and using the list of all thread_id values
from those as our thread_id value. We're missing one step which
sup does which is to also look up any children in the database
that have reference our message ID. So we'll need to do that next.
