Example - Using Scripting and Python to Debug in LLDB


LLDB has been structured from the beginning to be scriptable in two ways -- a Unix Python session can initiate/run a debug session non-interactively using LLDB; and within the LLDB debugger tool, Python scripts can be used to help with many tasks, including inspecting program data, iterating over containers and determining if a breakpoint should stop execution or continue. This document will show how to do some of these things by going through an example, explaining how to use Python scripting to find a bug in a program that searches for text in a large binary tree.

The Test Program and Input

We have a simple C program (dictionary.c) that reads in a text file, and stores all the words from the file in a Binary Search Tree, sorted alphabetically. It then enters a loop prompting the user for a word, searching for the word in the tree (using Binary Search), and reporting to the user whether or not it found the word in the tree.

The input text file we are using to test our program contains the text for William Shakespeare's famous tragedy "Romeo and Juliet".

The Bug

When we try running our program, we find there is a problem. While it successfully finds some of the words we would expect to find, such as "love" or "sun", it fails to find the word "Romeo", which MUST be in the input text file:

% ./dictionary Romeo-and-Juliet.txt
Dictionary loaded.
Enter search word: love
Enter search word: sun
Enter search word: Romeo
Enter search word: ^D

Is the word in our tree: Using Depth First Search

Our first job is to determine if the word "Romeo" actually got inserted into the tree or not. Since "Romeo and Juliet" has thousands of words, trying to examine our binary search tree by hand is completely impractical. Therefore we will write a Python script to search the tree for us. We will write a recursive Depth First Search function that traverses the entire tree searching for a word, and maintaining information about the path from the root of the tree to the current node. If it finds the word in the tree, it returns the path from the root to the node containing the word. This is what our DFS function in Python would look like, with line numbers added for easy reference in later explanations:

 1: def DFS (root, word, cur_path):
 2:     root_word_ptr = root.GetChildMemberWithName ("word")
 3:     left_child_ptr = root.GetChildMemberWithName ("left")
 4:     right_child_ptr = root.GetChildMemberWithName ("right")
 5:     root_word = root_word_ptr.GetSummary()
 6:     end = len (root_word) - 1
 7:     if root_word[0] == '"' and root_word[end] == '"':
 8:         root_word = root_word[1:end]
 9:     end = len (root_word) - 1
10:     if root_word[0] == '\'' and root_word[end] == '\'':
11:        root_word = root_word[1:end]
12:     if root_word == word:
13:         return cur_path
14:     elif word < root_word:
15:         if left_child_ptr.GetValue() == None:
16:             return ""
17:         else:
18:             cur_path = cur_path + "L"
19:             return DFS (left_child_ptr, word, cur_path)
20:     else:
21:         if right_child_ptr.GetValue() == None:
22:             return ""
23:         else:
24:             cur_path = cur_path + "R"
25:             return DFS (right_child_ptr, word, cur_path)

Accessing & Manipulating Program Variables in Python

Before we can call any Python function on any of our program's variables, we need to get the variable into a form that Python can access. To show you how to do this we will look at the parameters for the DFS function. The first parameter is going to be a node in our binary search tree, put into a Python variable. The second parameter is the word we are searching for (a string), and the third parameter is a string representing the path from the root of the tree to our current node.

The most interesting parameter is the first one, the Python variable that needs to contain a node in our search tree. How can we take a variable out of our program and put it into a Python variable? What kind of Python variable will it be? The answers are to use the LLDB API functions, provided as part of the LLDB Python module. Running Python from inside LLDB, LLDB will automatically give us our current frame object as a Python variable, "lldb.frame". This variable has the type "SBFrame" (see the LLDB API for more information about SBFrame objects). One of the things we can do with a frame object, is to ask it to find and return its local variable. We will call the API function "FindVariable" on the lldb.frame object to give us our dictionary variable as a Python variable:

root = lldb.frame.FindVariable ("dictionary")

The line above, executed in the Python script interpreter in LLDB, asks the current frame to find the variable named "dictionary" and return it. We then store the returned value in the Python variable named "root". This answers the question of HOW to get the variable, but it still doesn't explain WHAT actually gets put into "root". If you examine the LLDB API, you will find that the SBFrame method "FindVariable" returns an object of type SBValue. SBValue objects are used, among other things, to wrap up program variables and values. There are many useful methods defined in the SBValue class to allow you to get information or children values out of SBValues. For complete information, see the header file SBValue.h. The SBValue methods that we use in our DFS function are GetChildMemberWithName(), GetSummary(), and GetValue().

Explaining Depth First Search Script in Detail

"DFS" Overview. Before diving into the details of this code, it would be best to give a high-level overview of what it does. The nodes in our binary search tree were defined to have type tree_node *, which is defined as:

typedef struct tree_node
  const char *word;
  struct tree_node *left;
  struct tree_node *right;
} tree_node;

Lines 2-11 of DFS are getting data out of the current tree node and getting ready to do the actual search; lines 12-25 are the actual depth-first search. Lines 2-4 of our DFS function get the word, left and right fields out of the current node and store them in Python variables. Since root_word_ptr is a pointer to our word, and we want the actual word, line 5 calls GetSummary() to get a string containing the value out of the pointer. Since GetSummary() adds quotes around its result, lines 6-11 strip surrounding quotes off the word.

Line 12 checks to see if the word in the current node is the one we are searching for. If so, we are done, and line 13 returns the current path. Otherwise, line 14 checks to see if we should go left (search word comes before the current word). If we decide to go left, line 15 checks to see if the left pointer child is NULL ("None" is the Python equivalent of NULL). If the left pointer is NULL, then the word is not in this tree and we return an empty path (line 16). Otherwise, we add an "L" to the end of our current path string, to indicate we are going left (line 18), and then recurse on the left child (line 19). Lines 20-25 are the same as lines 14-19, except for going right rather than going left.

One other note: Typing something as long as our DFS function directly into the interpreter can be difficult, as making a single typing mistake means having to start all over. Therefore we recommend doing as we have done: Writing your longer, more complicated script functions in a separate file (in this case tree_utils.py) and then importing it into your LLDB Python interpreter.

Seeing the DFS Script in Action

At this point we are ready to use the DFS function to see if the word "Romeo" is in our tree or not. To actually use it in LLDB on our dictionary program, you would do something like this:

% lldb
(lldb) process attach -n "dictionary"
Architecture set to: x86_64.
Process 521 stopped
* thread #1: tid = 0x2c03, 0x00007fff86c8bea0 libSystem.B.dylib`read$NOCANCEL + 8, stop reason = signal SIGSTOP
frame #0: 0x00007fff86c8bea0 libSystem.B.dylib`read$NOCANCEL + 8
(lldb) breakpoint set -n find_word
Breakpoint created: 1: name = 'find_word', locations = 1, resolved = 1
(lldb) continue
Process 521 resuming
Process 521 stopped
* thread #1: tid = 0x2c03, 0x0000000100001830 dictionary`find_word + 16
at dictionary.c:105, stop reason = breakpoint 1.1
frame #0: 0x0000000100001830 dictionary`find_word + 16 at dictionary.c:105
102 int
103 find_word (tree_node *dictionary, char *word)
104 {
-> 105 if (!word || !dictionary)
106 return 0;
108 int compare_value = strcmp (word, dictionary->word);
(lldb) script
Python Interactive Interpreter. To exit, type 'quit()', 'exit()' or Ctrl-D.
>>> import tree_utils
>>> root = lldb.frame.FindVariable ("dictionary")
>>> current_path = ""
>>> path = tree_utils.DFS (root, "Romeo", current_path)
>>> print path
>>> ^D

The first bit of code above shows starting lldb, attaching to the dictionary program, and getting to the find_word function in LLDB. The interesting part (as far as this example is concerned) begins when we enter the script command and drop into the embedded interactive Python interpreter. We will go over this Python code line by line. The first line

import tree_utils

imports the file where we wrote our DFS function, tree_utils.py, into Python. Notice that to import the file we leave off the ".py" extension. We can now call any function in that file, giving it the prefix "tree_utils.", so that Python knows where to look for the function. The line

root = lldb.frame.FindVariable ("dictionary")

gets our program variable "dictionary" (which contains the binary search tree) and puts it into the Python variable "root". See Accessing & Manipulating Program Variables in Python above for more details about how this works. The next line is

current_path = ""

This line initializes the current_path from the root of the tree to our current node. Since we are starting at the root of the tree, our current path starts as an empty string. As we go right and left through the tree, the DFS function will append an 'R' or an 'L' to the current path, as appropriate. The line

path = tree_utils.DFS (root, "Romeo", current_path)

calls our DFS function (prefixing it with the module name so that Python can find it). We pass in our binary tree stored in the variable root, the word we are searching for, and our current path. We assign whatever path the DFS function returns to the Python variable path.

Finally, we want to see if the word was found or not, and if so we want to see the path through the tree to the word. So we do

print path

From this we can see that the word "Romeo" was indeed found in the tree, and the path from the root of the tree to the node containing "Romeo" is left-left-right-right-left.

What next? Using Breakpoint Command Scripts...

We are halfway to figuring out what the problem is. We know the word we are looking for is in the binary tree, and we know exactly where it is in the binary tree. Now we need to figure out why our binary search algorithm is not finding the word. We will do this using breakpoint command scripts.

The idea is as follows. The binary search algorithm has two main decision points: the decision to follow the right branch; and, the decision to follow the left branch. We will set a breakpoint at each of these decision points, and attach a Python breakpoint command script to each breakpoint. The breakpoint commands will use the global path Python variable that we got from our DFS function. Each time one of these decision breakpoints is hit, the script will compare the actual decision with the decision the front of the path variable says should be made (the first character of the path). If the actual decision and the path agree, then the front character is stripped off the path, and execution is resumed. In this case the user never even sees the breakpoint being hit. But if the decision differs from what the path says it should be, then the script prints out a message and does NOT resume execution, leaving the user sitting at the first point where a wrong decision is being made.

Side Note: Python Breakpoint Command Scripts are NOT What They Seem

What do we mean by that? When you enter a Python breakpoint command in LLDB, it appears that you are entering one or more plain lines of Python. BUT LLDB then takes what you entered and wraps it into a Python FUNCTION (just like using the "def" Python command). It automatically gives the function an obscure, unique, hard-to-stumble-across function name, and gives it two parameters: frame and bp_loc. When the breakpoint gets hit, LLDB wraps up the frame object where the breakpoint was hit, and the breakpoint location object for the breakpoint that was hit, and puts them into Python variables for you. It then calls the Python function that was created for the breakpoint command, and passes in the frame and breakpoint location objects.

So, being practical, what does this mean for you when you write your Python breakpoint commands? It means that there are two things you need to keep in mind: 1. If you want to access any Python variables created outside your script, you must declare such variables to be global. If you do not declare them as global, then the Python function will treat them as local variables, and you will get unexpected behavior. 2. All Python breakpoint command scripts automatically have a frame and a bp_loc variable. The variables are pre-loaded by LLDB with the correct context for the breakpoint. You do not have to use these variables, but they are there if you want them.

The Decision Point Breakpoint Commands

This is what the Python breakpoint command script would look like for the decision to go right:

global path
if path[0] == 'R':
    path = path[1:]
    thread = frame.GetThread()
    process = thread.GetProcess()
    print "Here is the problem; going right, should go left!"

Just as a reminder, LLDB is going to take this script and wrap it up in a function, like this:

def some_unique_and_obscure_function_name (frame, bp_loc):
    global path
    if path[0] == 'R':
        path = path[1:]
        thread = frame.GetThread()
        process = thread.GetProcess()
        print "Here is the problem; going right, should go left!"

LLDB will call the function, passing in the correct frame and breakpoint location whenever the breakpoint gets hit. There are several things to notice about this function. The first one is that we are accessing and updating a piece of state (the path variable), and actually conditioning our behavior based upon this variable. Since the variable was defined outside of our script (and therefore outside of the corresponding function) we need to tell Python that we are accessing a global variable. That is what the first line of the script does. Next we check where the path says we should go and compare it to our decision (recall that we are at the breakpoint for the decision to go right). If the path agrees with our decision, then we strip the first character off of the path.

Since the decision matched the path, we want to resume execution. To do this we make use of the frame parameter that LLDB guarantees will be there for us. We use LLDB API functions to get the current thread from the current frame, and then to get the process from the thread. Once we have the process, we tell it to resume execution (using the Continue() API function).

If the decision to go right does not agree with the path, then we do not resume execution. We allow the breakpoint to remain stopped (by doing nothing), and we print an informational message telling the user we have found the problem, and what the problem is.

Actually Using the Breakpoint Commands

Now we will look at what happens when we actually use these breakpoint commands on our program. Doing a source list -n find_word shows us the function containing our two decision points. Looking at the code below, we see that we want to set our breakpoints on lines 113 and 115:

(lldb) source list -n find_word
File: /Volumes/Data/HD2/carolinetice/Desktop/LLDB-Web-Examples/dictionary.c.
102 int
103 find_word (tree_node *dictionary, char *word)
104 {
105   if (!word || !dictionary)
106     return 0;
108   int compare_value = strcmp (word, dictionary->word);
110   if (compare_value == 0)
111     return 1;
112   else if (compare_value < 0)
113     return find_word (dictionary->left, word);
114   else
115     return find_word (dictionary->right, word);
116 }

So, we set our breakpoints, enter our breakpoint command scripts, and see what happens:

(lldb) breakpoint set -l 113
Breakpoint created: 2: file ='dictionary.c', line = 113, locations = 1, resolved = 1
(lldb) breakpoint set -l 115
Breakpoint created: 3: file ='dictionary.c', line = 115, locations = 1, resolved = 1
(lldb) breakpoint command add -s python 2
Enter your Python command(s). Type 'DONE' to end.
> global path
> if (path[0] == 'L'):
>     path = path[1:]
>     thread = frame.GetThread()
>     process = thread.GetProcess()
>     process.Continue()
> else:
>     print "Here is the problem. Going left, should go right!"
(lldb) breakpoint command add -s python 3
Enter your Python command(s). Type 'DONE' to end.
> global path
> if (path[0] == 'R'):
>     path = path[1:]
>     thread = frame.GetThread()
>     process = thread.GetProcess()
>     process.Continue()
> else:
>     print "Here is the problem. Going right, should go left!"
(lldb) continue
Process 696 resuming
Here is the problem. Going right, should go left!
Process 696 stopped
* thread #1: tid = 0x2d03, 0x000000010000189f dictionary`find_word + 127 at dictionary.c:115, stop reason = breakpoint 3.1
  frame #0: 0x000000010000189f dictionary`find_word + 127 at dictionary.c:115
    112   else if (compare_value < 0)
    113     return find_word (dictionary->left, word);
    114   else
 -> 115     return find_word (dictionary->right, word);
    116 }
    118 void

After setting our breakpoints, adding our breakpoint commands and continuing, we run for a little bit and then hit one of our breakpoints, printing out the error message from the breakpoint command. Apparently at this point in the tree, our search algorithm decided to go right, but our path says the node we want is to the left. Examining the word at the node where we stopped, and our search word, we see:

(lldb) expr dictionary->word
(const char *) $1 = 0x0000000100100080 "dramatis"
(lldb) expr word
(char *) $2 = 0x00007fff5fbff108 "romeo"

So the word at our current node is "dramatis", and the word we are searching for is "romeo". "romeo" comes after "dramatis" alphabetically, so it seems like going right would be the correct decision. Let's ask Python what it thinks the path from the current node to our word is:

(lldb) script print path

According to Python we need to go left-left-right-right-left from our current node to find the word we are looking for. Let's double check our tree, and see what word it has at that node:

(lldb) expr dictionary->left->left->right->right->left->word
(const char *) $4 = 0x0000000100100880 "Romeo"

So the word we are searching for is "romeo" and the word at our DFS location is "Romeo". Aha! One is uppercase and the other is lowercase: We seem to have a case conversion problem somewhere in our program (we do).

This is the end of our example on how you might use Python scripting in LLDB to help you find bugs in your program.

Source Files for The Example

The complete code for the Dictionary program (with case-conversion bug), the DFS function and other Python script examples (tree_utils.py) used for this example are available via following file links:

tree_utils.py - Example Python functions using LLDB's API, including DFS
dictionary.c - Sample dictionary program, with bug

The text for "Romeo and Juliet" can be obtained from the Gutenberg Project (http://www.gutenberg.org).