This updates the thread linking code to use ghost messages instead of
user metadata to link messages into threads.
In contrast with the old approach, this is actually correct.
Previously, thread merging updated only the thread IDs of message
documents, not thread IDs stored in user metadata. As originally
diagnosed by Mark Walters [1] and as demonstrated by the broken
T260-thread-order test, this can cause notmuch to fail to link
messages even though they're in the same thread. In principle the old
approach could have been fixed by updating the user metadata thread
IDs as well, but these are not indexed and hence this would have
required a full scan of all stored thread IDs. Ghost messages solve
this problem naturally by reusing the exact same thread ID and message
ID representation and indexing as regular messages.
Furthermore, thanks to this greater symmetry, ghost messages are also
algorithmically simpler. We continue to support the old user metadata
format, so this patch can't delete any code, but when we do remove
support for the old format, several functions can simply be deleted.
[1] id:8738h7kv2q.fsf@qmul.ac.uk
This updates the message abstraction to support ghost messages: it
adds a message flag that distinguishes regular messages from ghost
messages, and an internal function for initializing a newly created
(blank) message as a ghost message.
In the interest of robustness, avoid undefined behavior of
sortable_unserialise if the date value is missing. This shouldn't
happen now, but ghost messages will have blank date values.
This moves the code to retrieve and clear the metadata thread ID out
of _notmuch_database_link_message into its own function. This will
simplify future changes.
Previously, this was performed by notmuch_database_add_message. This
happens to be the only caller currently (which is why this was safe),
but we're about to introduce more callers, and it makes more sense to
put responsibility for ID compression in the lower-level function
rather than requiring each caller to handle it.
In Xapian, closing a database implicitly aborts any outstanding
transaction and commits changes. For historical reasons,
notmuch_database_close had grown to almost, but not quite duplicate
this behavior. Before closing the database, it would explicitly (and
unnecessarily) commit it. However, if there was an outstanding
transaction (ie atomic section), commit would throw a Xapian
exception, which notmuch_database_close would unnecessarily print to
stderr, even though notmuch_database_close would ultimately abort the
transaction anyway when it called close.
This patch simplifies notmuch_database_close to explicitly abort any
outstanding transaction and then just call Database::close. This
works for both read-only and read/write databases, takes care of
committing changes, unifies the exception handling path, and codifies
aborting outstanding transactions. This is currently the only way to
abort an atomic section (and may remain so, since it would be
difficult to roll back things we may have cached from rolled-back
modifications).
as stated in thread.cc:115
/* Construct an authors string from matched_authors_array and
* authors_array. The string contains matched authors first, then
* non-matched authors (with the two groups separated by '|'). Within
* each group, authors are listed in date order. */
this is, however, not reflected in the public API documentation in
notmuch.h:970. This patch a paragraph explaining how | separates the
group of authors of messages matching the query and those of messages
that do not, but are still contained in the thread.
48db8c8 introduced a disagreement between when
notmuch_database_needs_upgrade returned TRUE and when
notmuch_database_upgrade actually performed an upgrade. As a result,
if a database had a version less than 3, but no new features were
required, notmuch new would call notmuch_database_upgrade to perform
an upgrade, but notmuch_database_upgrade would return immediately
without updating the database version. Hence, the next notmuch new
would do the same, and so on.
Fix this by ensuring that the upgrade-required logic is identical
between the two.
Previously, there was no protection against a caller invoking an
operation on an old database version that would effectively corrupt
the database by treating it like a newer version.
According to notmuch.h, any caller that opens the database in
read/write mode is supposed to check if the database needs upgrading
and perform an upgrade if it does. This would protect against this,
but nobody (even the CLI) actually does this.
However, with features, it's easy to protect against incompatible
operations on a fine-grained basis. This lightweight change allows
callers to safely operate on old database versions, while preventing
specific operations that would corrupt the database with an
informative error message.
Commit 567bcbc2 introduced support for storing various headers in
document values. However, doing so in a backwards-compatible way
meant that genuinely empty header values could not be distinguished
from the old behavior of not storing the headers at all, so these
required parsing the original message.
Now that we have database features, new databases can declare that all
messages have header values, so if we have this feature flag, we can
use the stored header value even if it's the empty string.
This requires slight cleanup to notmuch_message_get_header, since the
code previously couldn't distinguish between empty headers and headers
that are never stored in the database (previously this distinction
didn't matter).
Previously, some parts of upgrade didn't report progress and for
others it was possible for the progress meter to restart at 0 part way
through the upgrade because each stage was reported separately.
Fix this by computing the total amount of work that needs to be done
up-front and updating completed work monotonically.
Rather than potentially making multiple passes over the same type of
data in the database, reorganize upgrade around each type of data that
may be upgraded. This eliminates code duplication, will make
multi-version upgrades faster, and will let us improve progress
reporting.
Previously, we had database version information hard-coded in the
upgrade code. Slightly re-organize the upgrade process around the set
of new database features to be enabled by the upgrade.
Previously, the upgrade was organized as two passes -- an upgrade
pass, and a separate cleanup pass -- so the database was always in a
valid state. This change substantially simplifies this code by
performing the upgrade in a transaction and combining both passes in
to one. This 1) eliminates a lot of duplicate code between the
passes, 2) speeds up the upgrade process, 3) makes progress reporting
more accurate, 4) eliminates the potential for stale data if the
upgrade is interrupted during the cleanup pass, and 5) makes it easier
to reason about the safety of the upgrade code.
Previously, our database schema was versioned by a single number.
Each database schema change had to occur "atomically" in Notmuch's
development history: before some commit, Notmuch used version N, after
that commit, it used version N+1. Hence, each new schema version
could introduce only one change, the task of developing a schema
change fell on a single person, and it all had to happen and be
perfect in a single commit series. This made introducing a new schema
version hard. We've seen only two schema changes in the history of
Notmuch.
This commit introduces database schema version 3; hopefully the last
schema version we'll need for a while. With this version, we switch
from a single version number to "features": a set of named,
independent aspects of the database schema.
Features should make backwards compatibility easier. For many things,
it should be easy to support databases both with and without a
feature, which will allow us to make upgrades optional and will enable
"unstable" features that can be developed and tested over time.
Features also make forwards compatibility easier. The features
recorded in a database include "compatibility flags," which can
indicate to an older version of Notmuch when it must support a given
feature to open the database for read or for write. This lets us
replace the old vague "I don't recognize this version, so something
might go wrong, but I promise to try my best" warnings upon opening a
database with an unknown version with precise errors. If a database
is safe to open for read/write despite unknown features, an older
version will know that and issue no message at all. If the database
is not safe to open for read/write because of unknown features, an
older version will know that, too, and can tell the user exactly which
required features it lacks support for.
According to RFC2822 References and In-Reply-To headers are supposed
to contain one or more Message-IDs, however older RFC822 allowed
almost any content. When both References and In-Reply-To headers ends
with something else that a Message-ID (see e.g. [1]), the thread
structure presented by notmuch is incorrect. The reason is that
notmuch treats this case as if the email contained no "replyto"
information (see _notmuch_database_link_message_to_parents).
This patch changes the parse_references() function to return the last
valid Message-ID encountered rather than NULL resulting from the last
hunk of text not being the Message-ID.
[1] https://lkml.org/lkml/headers/2014/5/19/864
Previously, we invalidated stored message metadata in
_notmuch_message_add_term and _notmuch_message_remove_term, but not in
_notmuch_message_gen_terms. This doesn't currently result in any bugs
because of our limited uses of _notmuch_message_gen_terms, but it may
could cause trouble in the future.
As noted in devel/STYLE, every private library function should start
with _notmuch. This patch corrects function naming that did not adhere
to this style in lib/notmuch-private.h. In particular, the old function
names that now begin with _notmuch are
notmuch_sha1_of_file
notmuch_sha1_of_string
notmuch_message_file_close
notmuch_message_file_get_header
notmuch_message_file_open
notmuch_message_get_author
notmuch_message_set_author
Signed-off-by: Charles Celerier <cceleri@cs.stanford.edu>
notmuch_database_close may fail in Xapian ->flush() or ->close(), so
report the status. Similarly for notmuch_database_destroy which calls
close.
This is required for notmuch insert to report error status if message
indexing failed.
This adds a 100 termpos gap between all phrases indexed by
_notmuch_message_gen_terms. This fixes a bug where terms from the end
of one header and the beginning of another header could match together
in a single phrase and a separate bug where term positions of
un-prefixed terms overlapped.
This fix only affects newly indexed messages. Messages that are
already indexed won't benefit from this fix without re-indexing, but
the fix won't make things any worse for existing messages.
Previously, we indexed the name and address parts of from/to headers
with two calls to _notmuch_message_gen_terms. In general, this
indicates that these parts are separate phrases. However, because of
an implementation quirk, the two calls to _notmuch_message_gen_terms
generated adjacent term positions for the prefixed terms, which
happens to be the right thing to do in this case, but the wrong thing
to do for all other calls. Furthermore, _notmuch_message_gen_terms
produced potentially overlapping term positions for the un-prefixed
copies of the terms, which is simply wrong.
This change indexes both the name and address in a single call to
_notmuch_message_gen_terms, indicating that they should be part of a
single phrase. This masks the problem with the un-prefixed terms
(fixing the two known-broken tests) and puts us in a position to fix
the unintentionally phrases generated by other calls to
_notmuch_message_gen_terms.
This is effectively a revert of
commit 6812136bf5
Author: Jani Nikula <jani@nikula.org>
Date: Mon Mar 31 00:21:48 2014 +0300
lib: drop support for single-message mbox files
The intention was to drop support for indexing new single-message mbox
files (and whether that was a good idea in the first place is
arguable). However this inadvertently broke support for reading
headers from previously indexed single-message mbox files, which is
far worse.
Distinguishing between the two cases would require more code than
simply bringing back support for single-message mbox files.
The notmuch library includes a full blown message header parser. Yet
the same message headers are parsed by gmime during indexing. Switch
to gmime parsing completely.
These are the main changes:
* Gmime stops header parsing at the first invalid header, and presumes
the message body starts from there. The current parser is quite
liberal in accepting broken headers. The change means we will be
much pickier about accepting invalid messages.
* The current parser converts tabs used in header folding to
spaces. Gmime preserve the tabs. Due to a broken python library used
in mailman, there are plenty of mailing lists that produce headers
with tabs in header folding, and we'll see plenty of tabs. (This
change has been mitigated in preparatory patches.)
* For pure header parsing, the current parser is likely faster than
gmime, which parses the whole message rather than just the
headers. Since we parse the message and its headers using gmime for
indexing anyway, this avoids and extra header parsing round when
adding new messages. In case of duplicate messages, we'll end up
parsing the full message although just headers would be
sufficient. All in all this should still speed up 'notmuch new'.
* Calls to notmuch_message_get_header() may be slightly slower than
previously for headers that are not indexed in the database, due to
parsing of the whole message. Within the notmuch code base, notmuch
reply is the only such user.
We've supported mbox files containing a single message for historical
reasons, but the support has been deprecated, with a warning message
while indexing, since Notmuch 0.15. Finally drop the support, and
consider all mbox files non-email.
In xapian terms, convert folder: prefix from probabilistic to boolean
prefix, matching the paths, relative from the maildir root, of the
message files, ignoring the maildir new and cur leaf directories.
folder:foo matches all message files in foo, foo/new, and foo/cur.
folder:foo/new does *not* match message files in foo/new.
folder:"" matches all message files in the top level maildir and its
new and cur subdirectories.
This change constitutes a database change: bump the database version
and add database upgrade support for folder: terms. The upgrade also
adds path: terms.
Finally, fix the folder search test for literal folder: search, as
some of the folder: matching capabilities are lost in the
probabilistic to boolean prefix change.
The path: prefix is a literal boolean prefix matching the paths,
relative from the maildir root, of the message files.
path:foo matches all message files in foo (but not in foo/new or
foo/cur).
path:foo/new matches all message files in foo/new.
path:"" matches all message files in the top level maildir.
path:foo/** matches all message files in foo and recursively in all
subdirectories of foo.
path:** matches all message files recursively, i.e. all messages.
Clarify that using the directory after destroying the corresponding
database is not permitted.
This is implicit in the description of notmuch_database_destroy, but
it doesn't hurt to be explicit, and we do express similar "ownership"
relationships at other places in the docs.
Currently if a Xapian exception happens in notmuch_message_get_header,
the exception is not caught leading to crash. In
notmuch_message_get_date the exception is caught, but an internal error
is raised, again leading to crash.
This patch fixes the error handling by making both functions catch the
Xapian exceptions, print an error and return NULL or 0.
The 'notmuch->exception_reported' is also set, as is done elsewhere,
even if I don't really get the idea of that field.
Signed-off-by: Tomi Valkeinen <tomi.valkeinen@iki.fi>
With some combination of clang and talloc, not using the return value
of talloc_steal() produces a warning. Ignore it, as talloc_steal() has
no failure modes per documentation.
This version of the library introduces LIBNOTMUCH_CHECK_VERSION and
the *_VERSION macros. Bumping the version number is also necessary to
make the comment on LIBNOTMUCH_CHECK_VERSION no longer a lie.
This makes it clear that these macros refer to the *library* version,
and not to the notmuch application-level release. Since there are no
consumers of these macros yet, this is now or never.
We have two distinct "library version" numbers: the soname version and
the version macros. We need both for different reasons: the version
macros enable easy compile-time version detection (and conditional
compilation), while the soname version enables runtime version
detection (which includes the version checking done by things like the
Python bindings).
However, currently, these two version numbers are different, which is
unnecessary and can lead to confusion (especially in things like
Debian, which include the soname version in the package name). This
patch makes them the same by bumping the version macros up to agree
with the soname version.
(We should probably keep the version number in just one place so they
can't get out of sync, but that can be done in another patch.)
Unfortunately old versions of GCC and clang do not provide byte order
macros, so we re-invent them.
If UTIL_BYTE_ORDER is not defined or defined to 0, we fall back to
macros supported by recent versions of GCC and clang
Previously PLATFORM_BYTE_ORDER and IS_LITTLE_ENDIAN were not defined,
so the little endian code was always compiled in.
This will have the effect that the "SHA1s" on big endian architectures
will change (i.e. become actual sha1s). So someone re-indexing their
database could conceivable lose tags on messages without a message-id
header.
