In id:87vcc2q5n2.fsf@nikula.org, Jani points out a memory leak in the
current version of the sup restore code. Among other things, this test
is intended to verify a fix for that leak.
The idea is run some code under valgrind --leak-check=full and report
a summary, leaving the user to peruse the log file if they want.
We go to some lengths to preserve the log files from accidental
overwriting; the full corpus takes about 3 hours to run under valgrind
on my machine.
The naming of the log directories may be slightly controversial; in
the unlikely event of two runs in less than a second, the log will be
overwritten. A previous version with mktemp+timestamp was dismissed as
overkill; just mktemp alone does not sort nicely.
One new test is included, to check notmuch new for memory leaks.
This is almost entirely renaming files, except for updating a few
references to those file names, and changing the makefile target.
A new set of memory tests will be run separately because they take
much longer.
Previously running search or pick from the pick buffer did not close
the message pane (if open). This meant that then new search ends up in
a very small window. Fix this so that the message pane is
shut. However, make it so that the pane is shut after the search
string is entered in case the user is basing the search on something
in the current message.
Apparently as of GMime 2.4, you don't need to call
internet_address_list_destroy anymore, but you still need to call
g_object_unref (from the GMime Changelog).
On the medium performance corpus, valgrind shows "possibly lost"
leakage in "notmuch new" dropping from 7M to 300k.
The message->headers hash table values get data returned by
g_mime_utils_header_decode_text ().
The pointer returned by g_mime_utils_header_decode_text is from the
following line in rfc2047_decode_tokens
return g_string_free (decoded, FALSE);
The docs for g_string_free say
Frees the memory allocated for the GString. If free_segment is TRUE
it also frees the character data. If it's FALSE, the caller gains
ownership of the buffer and must free it after use with g_free().
The remaining frees and allocations referencing to message->headers hash
values have been changed to use g_free and g_malloc functions.
This combines and completes the changes started by David Bremner.
The memory usage discipline of tag_op_list_t is never to free the
internal array of tag operations before freeing the whole list, so it
makes sense to take advantage of hierarchical de-allocation by talloc.
By not relying on the context passed into tag_parse_line, we can allow
tag_op_list_t structures to live longer than that context.
Previously notmuch_restore_command returned 0 if tag_message returned
a non-zero (failure) value. This is wrong, since non-zero status
indicates something mysterious went wrong with retrieving the message,
or applying it.
There was also a failure to check or propagate the return value from
tag_op_list_apply in tag_message.
Now that the invisibility display of parts is present we no longer
need to force the display of all multipart/alternatives: users can
toggle them for themselves when needed.
This adds a button action to show hidden parts. In this version "RET"
toggles the visibility of any part which puts content in the buffer
(as opposed to attachments such as application/pdf).
The button is used to hide parts when appropriate (eg text/html in
multipart/alternative).
This makes notmuch-show-insert-bodypart add an overlay for any
non-trivial part with a button header (currently the first text/plain
part does not have a button). At this point the overlay is available
to the button but there is no action using it yet.
In addition the argument HIDE is passed down to
notmuch-show-insert-part-overlays to request that the part be hidden
by default but this is not acted on yet.
This just make notmuch-show-insert-part-header save the basic button
text for parts as an attribute. This makes it simpler for the button
action (added in a later patch) to reword the label as appropriate (eg
append "(not shown)" or not as appropriate).
Previously, all visibility in show buffers for headers, message
bodies, and washed text was specified by generating one or more
symbols for each region and creating overlays with their 'invisible
property set to carefully crafted combinations of these symbols.
Visibility was controlled not by modifying the overlays directly, but
by adding and removing the generated symbols from a gigantic buffer
invisibilty spec.
This has myriad negative consequences. It's slow because Emacs'
display engine has to traverse the buffer invisibility list for every
overlay and, since every overlay has its own symbol, this makes
rendering O(N^2) in the number of overlays. It composes poorly
because symbol-type 'invisible properties are taken from the highest
priority overlay over a given character (which is often ambiguous!),
rather than being gathered from all overlays over a character. As a
result, we have to include symbols related to message hiding in the
wash code lest the wash overlays un-hide parts of hidden messages. It
also requires various workarounds for isearch to properly open
overlays, to set up buffer-invisibility-spec for
remove-from-invisibility-spec to work right, and to explicitly refresh
the display after updating the buffer invisibility spec.
None of this is necessary.
This patch converts show and wash to use simple boolean 'invisible
properties and to not use the buffer invisibility spec. Rather than
adding and removing generated symbols from the invisibility spec, the
code now directly toggles the 'invisible property of the appropriate
overlay. This speeds up rendering because the display engine only has
to check the boolean values of the overlays over a character. It
composes nicely because text will be invisible if *any* overlay over
it has 'invisible t, which means we can overlap invisibility overlays
with abandon. We no longer need any of the workarounds mentioned
above. And it fixes a minor bug for free: now, when isearch opens a
washed region, the button text will update to say "Click/Enter to
hide" rather than remaining unchanged.
Date range search may be implemented as a library change, but it's an
important user-facing change that affects all notmuch usage. Tag name
restrictions aren't as important, but they affect both the CLI
interface and the Emacs interface.
Removes Message.{format,print}_messages.
This code adds functionality at the python level that is unlikely to
be useful for anyone. Furthermore the python bindings strive to be a
thin wrapper around libnotmuch. The code has been marked as deprecated
in 0.14 and is now removed.
Signed-off-by: Justus Winter <4winter@informatik.uni-hamburg.de>
Add new format text0, which is otherwise the same as text, but use the
null character as separator instead of the newline character. This is
similar to find(1) -print0 option, and works together with the
xargs(1) -0 option.
Currently pick just uses notmuch-show to display messages in the
message pane: this means that they get indented just as show
would. However, since pick is only displaying one message at a time
there is no need to indent so override the indentation.
If a leaf part's body content is omitted, return the encoded length and
transfer encoding in --format=json output. This information may be used
by the consumer, e.g. to decide whether to download a large attachment
over a slow link.
Returning the _encoded_ content length is more efficient than returning
the _decoded_ content length. Returning the transfer encoding allows
the consumer to estimate the decoded content length.
Write a "charset" field for all omitted parts for which it is applicable,
not only text/html parts. Factor out the code to a separate function.
It will be extended with more fields next.
notmuch_json_show_sanitize replaced "filename" field values even in part
structures, where the value is predictable. Make it only normalize the
filename value if it is an absolute path (begins with slash), which is
true of the Maildir filenames that were intended to be normalized away.
Currently there is a period of pain whenever we make
backward-incompatible changes to the structured output format, which
discourages not only backward-incompatible improvements to the format,
but also backwards-compatible additions that may not be "perfect". In
the end, these problems limit experimentation and innovation.
This series of patches introduces a way for CLI callers to request a
specific format version on the command line and to determine if the
CLI does not supported the requested version (and perhaps present a
useful diagnostic to the user). Since the caller requests a format
version, it's also possible for the CLI to support multiple
incompatible versions simultaneously, unlike the alternate approach of
including version information in the output.
This patch lays the groundwork by introducing a versioning convention,
standard exit codes, and a utility function to check the requested
version and produce standardized diagnostic messages and exit
statuses.
This slightly changes the output of an existing test since we now
report non-zero exits with a pop-up buffer instead of at the end of
the search results.
This checks for non-zero exit status from JSON CLI calls and pops up
an error buffer with stderr and stdout. A consequence of this is that
show and reply now handle errors, rather than ignoring them.
This provides library functions for unified handling of errors from
the notmuch CLI. Follow-up patches will convert some scattered error
handling to use this and add error handling where we currently ignore
errors.
Update pick's archive message to respect notmuch-archive-tags. Also
split archive message into an archiving part and a separate
"then-next" part, to move more inline with show. Update the keybinding
so default behaviour is unchanged.
Previously, if the input stream consisted only of an error message,
notmuch-json-begin-compound would signal a (wrong-type-argument
number-or-marker-p nil) error when reaching the end of the error
message. This happened because notmuch-json-scan-to-value would think
that it reached a value and put the parser into the 'value state.
Even after notmuch-json-begin-compound signaled the syntax error, the
parser would remain in this state and when the resynchronization logic
reached the end of the buffer, the parser would fail because the
'value state indicates that characters are available.
This fixes this problem by restoring the parser's previous state if it
encounters a syntax error.
This makes the tag set a bit less trivial.
Note that if you use the small corpus, this is not so interesting (and
is also a bit noisy) since the messages will not be found. In the
future this could be checked for.
Conflicts:
performance-test/01-dump-restore
