Libtool Problem (continued)

Shortly after posting the last entry about the libtool problem I sent a message to the bug-libtool list, Scott helped to track down the problem.

With the help of the test script I wrote, he managed to track down the change on the libtool-2.0 branch that fixed the problem. Applying this same change to a 1.5.x release fixed the problem. He has uploaded a new Debian package with the change, and I’ve altered the jhbuild bootstrap module set to include the patch too. The copy of the patch included with JHBuild can be found here. Hopefully it will also be in a future 1.5.x release (assuming that there are any more).

Scott pointed out another case where people might run into the problem is when building binary packages for software. A packager usually builds the new version of the software into a temporary prefix (often by setting the $DESTDIR environment variable when calling make install). If the package includes a library with some applications that link to the library and there is an old version of the package installed on the system, libtool could end up linking with the library in /usr/lib, which could result in a build failure if some new APIs were added. The patch should fix this particular case too.

So if you release tarballs that make use of libtool, applying this patch may help out the people maintaining binary packages of the software for a distro too (assuming that they haven’t gone the scorched earth route and deleted all the .la files …).

25 October 2004

  • Post author:
  • Post category:Uncategorized

Drive Mount Applet

The new drive mount applet is now checked into the HEAD branch of gnome-applets, so will be in Gnome 2.10. There are a few things left to do, such as making it possible to open the file manager as well as unmounting/ejecting it. I did up a screenshot showing what it looks like as an applet.

Libtool

Finally managed to reproduce a particular libtool bug that people have reported on and off. It does show why some people decide that .la files are evil, since it doesn’t occur when people delete those files …

A reduced test case can be found here. The problem occurs when you have multiple copies of a library in your linker library search path with associated .la files. In the test case, there are the following libraries:

  • libfoo.so and libfoo.la in the directory /A. This is the library we want to link to.
  • libfoo.so and libfoo.la in the directory /B. We don’t want to link to this one, because it is old.
  • libbar.so and libbar.la in the directory /B.

Let’s say I then try to link an app that needs libbar and the new version of libfoo, and happen to use the following link line:

libtool --mode=link gcc -o main main.c -lbar -L/A -L/B -lfoo

In the absense of libtool, this would result in us linking against /B/libbar.so and /A/libfoo.so (since /A comes before /B in the search path).

However, libtool ends up doing something quite different. When it sees -lbar, it notices that there is a libbar.la in /B, expands that argument to the full path name of the actual library (/B/libbar.so), and prepends /B to the library search path. This means that when it gets round to processing -lfoo, it finds /B/libfoo.la instead of /A/libfoo.la, and links to the wrong library.

If this sounds like an obscure bug, note that it also happens if we replace /B with /usr/lib. In this case, we don’t even need the -L/usr/lib argument. So the following command results in linking with /usr/lib/libfoo.so instead of /A/libfoo.so:

libtool --mode=link gcc -o main main.c -lbar -L/A -lfoo

This sort of situation is quite common when trying to build some software into a separate prefix that is also provided by the OS, when you are relying on a few libraries installed in /usr/lib with .la files.

After putting together the test case I tested it out in the latest development release (1.9f), and it appears that the problem has been fixed. Given that the libtool developers are so close to a 2.0 release, I don’t know whether they would bother putting out another 1.5.x release to fix the problem.

So if you do run into the problem, some possible solutions are:

  1. Upgrade to libtool-1.9f. I’m not sure how good an idea this is if you are producing tarballs, since they will be packaged with the development release too.
  2. Remove all the .la files in /usr/lib. Some distributors seem to take this route (eg. Ximian/Novell and Red Hat).

20 October 2004

  • Post author:
  • Post category:Uncategorized

Even More Icon Theme Stuff

To make it a bit easier to correctly display themed icons, I added support to GtkImage, so that it is as easy as calling gtk_image_new_from_icon_name() or gtk_image_set_from_icon_name(). The patch is attached to bug #155688.

This code takes care of theme changes so the application developer doesn’t need to. Once this is in, it should be trivial to add themed icon support to various other widgets that use GtkImage (such as GtkAbout and GtkToolItem).

JHBuild

I started work on some extended documentation for JHBuild. At the moment, this just includes some information on setting it up and basic use. I’ll extend it to hold a reference to all JHBuild commands, some documentation on the module set file format and some frequently asked questions.

It would be good to include some information on setting up JHBuild’s tinderbox mode (like Luis has). Getting a few more tinderboxes running for Gnome on other platforms such Solaris/SPARC would be really useful — Luis’s build logs have already helped track down a few build failures, so having build information for a few more platforms would be very useful.

New Computer

I just got the last components for my new computer (an Athlon 64 system). It is a fair bit faster than the laptop I’ve been doing most of the development on, so should be quite nice once it is all set up.

It is amazing how much hardware has improved and gone down in price. The motherboard alone is packed with features I wouldn’t have expected for something costing AU$220:

  • Gigabit ethernet
  • An 802.11g wireless card (PCI)
  • An extra Promise SATA chip, bringing the number of SATA connectors up to 4, and the IDE connectors to 3.
  • Firewire
  • SPDIF output (both electrical and optical).
  • 6 headphone-style jacks on the back, so you can get 6 channel audio output without losing your line in and microphone jacks.

I also got a Raptor hard drive for the system. These drives seem to have up to twice the performance of most 7200rpm desktop drives, and make a big difference to the overall performance of the system.

It should be a nice system once I finish building it. Also, since it is an x86-64 system, it effectively provides two architectures to test stuff on.

Drive Mount Applet

  • Post author:
  • Post category:Uncategorized

I started to look at bringing the drive mount applet from gnome-applts up to scratch, since it hasn’t really had much work done on it other than porting to the 2.x development platform.

The applet is a classic example of Gnome 1.x user interface complexity. The applet shows a button that can be clicked to mount or unmount a particular mount point. For this simple functionality, it provides the following preferences:

  • The mount directory
  • The interval at which to check the mounted state
  • Which of a set of custom images to use to show the mounted state (eg. floppy drive, cdrom drive, etc).
  • The ability to specify custom mounted/unmounted images.
  • Whether to eject the disk after unmounting.
  • Whether to use a second method for checking the mounted state which might work better with automounters.

The applet also had a few problems, such as not being able to unmount a disk if there was a trash directory on it (since fam would have a dnotify watch active on that directory).

By using some of the newer gnome-vfs APIs, I think it should be possible to remove all the preferences:

  1. The GnomeVFSVolumeManager API can be used to get a list of attached drives and volumes, and receive notification when volumes are mounted or new drives are connected or disconnected. This allows one applet to display the status of all user (un)mountable drives/volumes in one applet.
  2. By using the gnome-vfs APIs to unmount or eject a volume, a volume_pre_unmount signal to interested applications before attempting to unmount it. When Nautilus receives this signal, it drops its FAM watches on the trash directory for that volume, and closes all associated windows. This means that Nautilus/FAM won’t cause a “volume is busy” error (although some other apps might hold files opne on the volume).
  3. The gnome-vfs APIs tell us what the volume type is. This way we can automatically do an unmount+eject for cdrom, zip and jaz drives like Nautilus does rather than having to provide a preference.
  4. Gnome-vfs also tells us what icon name should be used for a particular drive or volume. This way the applet can just pull the icons from the current icon theme rather than having to maintain separate ones.

My code is at the point where it produces nice screenshots, and has the above features. The screenshot shows it as a separate window, but adding the applet wrapper stuff isn’t too difficult (it is easier to test as a standalone app). It is less than half the size of the old applet too, which is promising.

11 October 2004

  • Post author:
  • Post category:Uncategorized

Federal Election

Looks like we are going to have at least another three years with The Rodent. It also looks like they will have a majority in the senate, which will reduce the senate’s effectiveness as a house of review.

We might not have John Howard for the entire term though, since he is of retirement age. NineMSN seems to think that Peter Costello is already the leader.

It also looks like The Democrats senators up for reelection got completely wiped out, with much of their support going to The Greens.

Gnome-VFS

Robert Collins wrote an interesting critique on the gnome-vfs API. I don’t agree with all the points, and there are some reasons why the API isn’t as elegant as it could be. Below are some responses.

Initialisation

One thing that gnome_vfs_init() does is to call g_thread_init(). Before this function is called, the locking APIs in glib are no-ops. You really want this function called early on if the app is going to use threads, otherwise you will end up with inconsistencies (eg. a lock() call might be a no-op, but the unlock() call might not be if g_thread_init() is called in between).

The other issue is that gnome_vfs_init() can fail. If it is called automatically, then any function that might invoke the initialisation routine now has a new failure mode. I don’t know whether this is a real problem or not though.

Calling Style – Inconsistent Ordering

One big difference between the out parameters in gnome_vfs_open() and gnome_vfs_read() is that the first function is essentially a constructor for a file handle, while the second is a method for a file handle that fills in a provided buffer.

I’ll agree that the calling conventions are not as nice as they could be though. If they were being designed today, I suspect that they would look more like this:

GnomeVFSHandle  *gnome_vfs_open (const gchar     *text_uri,
                                 GnomeVFSOpenMode open_mode,
                                 GError         **error);
GnomeVFSFileSize gnome_vfs_read (GnomeVFSHandle  *handle,
                                 gpointer         buffer,
                                 GnomeVFSFileSize bytes,
                                 GError         **error);

Unfortunately, the GError API was not developped til the 2.0 series, while these parts of the gnome-vfs API persist from the 1.x days.

Calling Style – Inconsistent Method Naming

I agree that the gnome_vfs_truncate() function name is inconsistent. My guess as to why they chose gnome_vfs_truncate and gnome_vfs_truncate_handle() was to match the underlying truncate() and ftruncate() C library calls. This was probably a case of balancing consistent APIs with ease of transition from libc APIs to gnome-vfs APIs.

Returning data to pointers

I agree that the existing calling convention is not as nice as it could be. As I said earlier, it would probably have been designed to use GError if it was being developed today.

The GError API has a number of benefits over errno style ones, including:

  1. Automatically threadsafe. The place where the error is reported is on the stack. A global variable is a problem for multi threaded apps on systems without Linux 2.6 style thread local storage (you need to do tricks like making errno into a function that returns the appropriate variable for the current thread).
  2. In Robert’s example, the actual error information is looked up from a file handle. What do you do if there is no file handle involved in the function call? Also, wouldn’t gnome_vfs_open() return a NULL file handle on error?
  3. For the cases where there is a file handle to look up the error info on, what happens if two threads are working with the file handle at the same time?
  4. GError is consistent with other Gnome APIs 🙂

The GError API also makes it easy to pass error data up a number of call frames similar to exceptions. If your function has a GError argument, you can simply pass that same error object to other functions when you call them. If those functions fail on an error, simply return immediately, and the caller can handle the error.

Streams Interface

There actually is a stream interface available in one of the libraries both GTK and gnome-vfs depend on: GIOChannel. I guess it would be nice if gnome-vfs provided a GIOChannel implementation for VFS file handles. The main thing that would be needed here would be the io_create_watch() implementation, which would probably require exposing a file descriptor to poll on (this could probably be implemented using a pipe pretty easily).

Doing this as GObject interfaces isn’t really an option, since GIOChannel is implemented in libglib which is below libgobject, and gnome-vfs file handles aren’t GObjects. I know that at one point Ian was planning to change the various handles to GObjects, but this didn’t happen. It would probably be possible to do this kind of change while only requiring changes to VFS methods, so it can’t be completely ruled out.

Directory Interface

You can asynchronously load a directory listing using gnome_vfs_async_load_directory(). I don’t blame you for missing it — the organisation of the APIs in the various headers is a bit confusing.

Language Bindings

There are a lot of things a language binding can do to make gnome-vfs nicer to use. Some of these things include:

  • If the language provides exceptions, convert error GnomeVFSResult‘s to exceptions and change the calling conventions to something more sane.
  • If the language allows for runtime type checking or multiple dispatch, don’t wrap the gnome_vfs_foo() and gnome_vfs_foo_uri() functions separately. Instead, just check if a string or a GnomeVFSURI was passed in and do the right thing.
  • If the language has a standard file handle interface or convention, try to implement it in the binding.

The Python bindings do some of these things, and definitely make things easier to use.