23.2.2.11 Writing a Frame Filter

There are three basic elements that a frame filter must implement: it must correctly implement the documented interface (see Frame Filter API), it must register itself with GDB, and finally, it must decide if it is to work on the data provided by GDB. In all cases, whether it works on the iterator or not, each frame filter must return an iterator. A bare-bones frame filter follows the pattern in the following example.

import gdb
class FrameFilter():
    def __init__(self):
        # Frame filter attribute creation.
        #
        # 'name' is the name of the filter that GDB will display.
        #
        # 'priority' is the priority of the filter relative to other
        # filters.
        #
        # 'enabled' is a boolean that indicates whether this filter is
        # enabled and should be executed.
        self.name = "Foo"
        self.priority = 100
        self.enabled = True
        # Register this frame filter with the global frame_filters
        # dictionary.
        gdb.frame_filters[self.name] = self
    def filter(self, frame_iter):
        # Just return the iterator.
        return frame_iter

The frame filter in the example above implements the three requirements for all frame filters. It implements the API, self registers, and makes a decision on the iterator (in this case, it just returns the iterator untouched).

The first step is attribute creation and assignment, and as shown in the comments the filter assigns the following attributes: name, priority and whether the filter should be enabled with the enabled attribute.

The second step is registering the frame filter with the dictionary or dictionaries that the frame filter has interest in. As shown in the comments, this filter just registers itself with the global dictionary gdb.frame_filters. As noted earlier, gdb.frame_filters is a dictionary that is initialized in the gdb module when GDB starts. What dictionary a filter registers with is an important consideration. Generally, if a filter is specific to a set of code, it should be registered either in the objfile or progspace dictionaries as they are specific to the program currently loaded in GDB. The global dictionary is always present in GDB and is never unloaded. Any filters registered with the global dictionary will exist until GDB exits. To avoid filters that may conflict, it is generally better to register frame filters against the dictionaries that more closely align with the usage of the filter currently in question. See Python Auto-loading, for further information on auto-loading Python scripts.

GDB takes a hands-off approach to frame filter registration, therefore it is the frame filter’s responsibility to ensure registration has occurred, and that any exceptions are handled appropriately. In particular, you may wish to handle exceptions relating to Python dictionary key uniqueness. It is mandatory that the dictionary key is the same as frame filter’s name attribute. When a user manages frame filters (see Frame Filter Management), the names GDB will display are those contained in the name attribute.

The final step of this example is the implementation of the filter method. As shown in the example comments, we define the filter method and note that the method must take an iterator, and also must return an iterator. In this bare-bones example, the frame filter is not very useful as it just returns the iterator untouched. However this is a valid operation for frame filters that have the enabled attribute set, but decide not to operate on any frames.

In the next example, the frame filter operates on all frames and utilizes a frame decorator to perform some work on the frames. See Frame Decorator API, for further information on the frame decorator interface.

This example works on inlined frames. It highlights frames which are inlined by tagging them with an “[inlined]” tag. By applying a frame decorator to all frames with the Python itertools imap method, the example defers actions to the frame decorator. Frame decorators are only processed when GDB prints the backtrace.

This introduces a new decision making topic: whether to perform decision making operations at the filtering step, or at the printing step. In this example’s approach, it does not perform any filtering decisions at the filtering step beyond mapping a frame decorator to each frame. This allows the actual decision making to be performed when each frame is printed. This is an important consideration, and well worth reflecting upon when designing a frame filter. An issue that frame filters should avoid is unwinding the stack if possible. Some stacks can run very deep, into the tens of thousands in some cases. To search every frame to determine if it is inlined ahead of time may be too expensive at the filtering step. The frame filter cannot know how many frames it has to iterate over, and it would have to iterate through them all. This ends up duplicating effort as GDB performs this iteration when it prints the frames.

In this example decision making can be deferred to the printing step. As each frame is printed, the frame decorator can examine each frame in turn when GDB iterates. From a performance viewpoint, this is the most appropriate decision to make as it avoids duplicating the effort that the printing step would undertake anyway. Also, if there are many frame filters unwinding the stack during filtering, it can substantially delay the printing of the backtrace which will result in large memory usage, and a poor user experience.

class InlineFilter():
    def __init__(self):
        self.name = "InlinedFrameFilter"
        self.priority = 100
        self.enabled = True
        gdb.frame_filters[self.name] = self
    def filter(self, frame_iter):
        frame_iter = itertools.imap(InlinedFrameDecorator,
                                    frame_iter)
        return frame_iter

This frame filter is somewhat similar to the earlier example, except that the filter method applies a frame decorator object called InlinedFrameDecorator to each element in the iterator. The imap Python method is light-weight. It does not proactively iterate over the iterator, but rather creates a new iterator which wraps the existing one.

Below is the frame decorator for this example.

class InlinedFrameDecorator(FrameDecorator):
    def __init__(self, fobj):
        super(InlinedFrameDecorator, self).__init__(fobj)
    def function(self):
        frame = fobj.inferior_frame()
        name = str(frame.name())
        if frame.type() == gdb.INLINE_FRAME:
            name = name + " [inlined]"
        return name

This frame decorator only defines and overrides the function method. It lets the supplied FrameDecorator, which is shipped with GDB, perform the other work associated with printing this frame.

The combination of these two objects create this output from a backtrace:

#0  0x004004e0 in bar () at inline.c:11
#1  0x00400566 in max [inlined] (b=6, a=12) at inline.c:21
#2  0x00400566 in main () at inline.c:31

So in the case of this example, a frame decorator is applied to all frames, regardless of whether they may be inlined or not. As GDB iterates over the iterator produced by the frame filters, GDB executes each frame decorator which then makes a decision on what to print in the function callback. Using a strategy like this is a way to defer decisions on the frame content to printing time.

Eliding Frames

It might be that the above example is not desirable for representing inlined frames, and a hierarchical approach may be preferred. If we want to hierarchically represent frames, the elided frame decorator interface might be preferable.

This example approaches the issue with the elided method. This example is quite long, but very simplistic. It is out-of-scope for this section to write a complete example that comprehensively covers all approaches of finding and printing inlined frames. However, this example illustrates the approach an author might use.

This example comprises of three sections.

class InlineFrameFilter():
    def __init__(self):
        self.name = "InlinedFrameFilter"
        self.priority = 100
        self.enabled = True
        gdb.frame_filters[self.name] = self
    def filter(self, frame_iter):
        return ElidingInlineIterator(frame_iter)

This frame filter is very similar to the other examples. The only difference is this frame filter is wrapping the iterator provided to it (frame_iter) with a custom iterator called ElidingInlineIterator. This again defers actions to when GDB prints the backtrace, as the iterator is not traversed until printing.

The iterator for this example is as follows. It is in this section of the example where decisions are made on the content of the backtrace.

class ElidingInlineIterator:
    def __init__(self, ii):
        self.input_iterator = ii
    def __iter__(self):
        return self
    def next(self):
        frame = next(self.input_iterator)
        if frame.inferior_frame().type() != gdb.INLINE_FRAME:
            return frame
        try:
            eliding_frame = next(self.input_iterator)
        except StopIteration:
            return frame
        return ElidingFrameDecorator(eliding_frame, [frame])

This iterator implements the Python iterator protocol. When the next function is called (when GDB prints each frame), the iterator checks if this frame decorator, frame, is wrapping an inlined frame. If it is not, it returns the existing frame decorator untouched. If it is wrapping an inlined frame, it assumes that the inlined frame was contained within the next oldest frame, eliding_frame, which it fetches. It then creates and returns a frame decorator, ElidingFrameDecorator, which contains both the elided frame, and the eliding frame.

class ElidingInlineDecorator(FrameDecorator):
    def __init__(self, frame, elided_frames):
        super(ElidingInlineDecorator, self).__init__(frame)
        self.frame = frame
        self.elided_frames = elided_frames
    def elided(self):
        return iter(self.elided_frames)

This frame decorator overrides one function and returns the inlined frame in the elided method. As before it lets FrameDecorator do the rest of the work involved in printing this frame. This produces the following output.

#0  0x004004e0 in bar () at inline.c:11
#2  0x00400529 in main () at inline.c:25
    #1  0x00400529 in max (b=6, a=12) at inline.c:15

In that output, max which has been inlined into main is printed hierarchically. Another approach would be to combine the function method, and the elided method to both print a marker in the inlined frame, and also show the hierarchical relationship.