External Mergesort: part two

Introduction

In this last article, we examined how external mergesort works in theory, let’s implement it now.

First, you can find all the source codes of this implementation in this Github repo. In this repo, I implemented both two-way and multi-way solutions, which are tracked in different branches, please be aware of this. But for the sake of simplicity, I will focus on the generalized multi-way mergesort solution in this article.

Data Preparation

Before diving into the code, let’s define the problem we need to solve here. I will generate an input file containing several millions of seven digits random numbers, from 1,000,000 to 9,999,999. The random numbers can be duplicated, and each number is stored in one new line of the input file. The input file can be prepared with the following Bash script which calls GNU shuf :

The generated input file is roughly 60 MB in size. For modern computers, it can be loaded to memory easily. But since we are working external memory algorithm, so let’s assume we are running the algorithm on an old computer, which has only 100000 Byte memory. Based on this assumed restriction, we need to sort the numbers in the input file and save the result in a new file.

Implementation

Let’s define some global constants in this algorithm:

MEMORY_LIMIT denotes the 100000 bytes memory limit; In C, we can use the type unsigned int to store an integer in the range from 1,000,000 to 9,999,999. So RECORD_SIZE means each record(or integer) will take up 4 bytes of memory.

And by default, the algorithm will use the two-way merge, but the user can pass an argument to run a multi-way merge too.

Sort phase

The sort phase is implemented inside the separationSort function as follows:

The logic is not difficult. The function takes the input file descriptor as a parameter and reads each line(via the getline method) in a loop until reaches the end of the file. The numbers will be read into the memory buffer before hitting the memory limit. When the memory buffer is full(100000 bytes), the numbers are sorted with the function mergeSort.

The function mergeSort is defined inside the file internal_sort.c, which implements the classic internal merge sorting algorithm. I will not cover it in this article, since you can find many documents about it online. If you don’t know about it, please spend some time learning about it. Of course, you can replace it with other sorting algorithms like quick sort too. I leave this for the readers.

After sorting, the numbers are saved in the temporary files in the directory of ./tmp/pass0. The filename is just the run number.

We can verify the result of the sort phase as follows:

You can see each file contains up to 25000 (equal to MEMORY_LIMIT/RECORD_SIZE) numbers and 312 files are created in pass0.

Note that I will not examine the details about how to make a new directory and how to open a new file to read or write. You can learn such Linux file I/O concepts by yourself.

Merge phase

The exMerge function controls the passes in the merge phase starting from pass1.

The variable fileNum stores the run number in each pass. And the variable ways denotes the number of multi-way. Thus, the run number of the next pass should be calculated as fileNum / ways.

The detailed merging logic is inside the function exMergeSort, which takes two parameters. pass means the current pass number(starting from 1), while nums means how many runs(or sub-files) in the last pass need to be merged.

The above code creates a temp directory for each pass and a temp file for each run.

Next, we create an array of the input files for each run. And the input file is inside the temp directory of the last pass. Each run merges multiple files in the for loop. The only trick logic here is that the remaining files in the last run may be less than the number of ways declared, we need to handle that properly (line 5 of the below code block).

Next, we need to read one number from every input file until only one file is not run out. Find the smallest one and save it in the temp output run file. And for the last remaining file, remember to put the rest numbers into the output file too.

The above code bock utilizes several methods like getRecord, validateRecords, getMinRecordIndex and getLastRemainRecordIndex as follows, and these functions are easy to understand.

In detail, you can refer to the source code of this github repo. Next, let’s evaluate the performance of this algorithm by tuning the number of ways for merging.

Performance Evaluation

We’ll use the Linux time utility to measure the running time of this algorithm.

The result of two-way mergesort is:

while six-way mergesort can complete with a shorter runtime.

External Mergesort: part one

Introduction

This article will examine one interesting question which I came across when reading the book Programming Pearls. This question simply goes like this: How to sort a large disk file? The disk file has so much data that it cannot all fit into the main memory. I considered this algorithm question for a while; but noticed that all the classic sorting algorithms, like Quick sort and merge sort, can’t solve it easily. Then I did some research about it, and I will share what I learned in this article. I believe you can solve this problem as well, after reading this article.

Background of External Algorithm

Traditionally, computer scientists analyze the running time of an algorithm by counting the number of executed instructions, which is usually expressed as a function of the input size n, like the well-known Big O notation. This kind of algorithm complexity analysis is based on the Random access machine(RAM) model, which defines the following assumptions:

• A machine with an unbounded amount of available memory;
• Any desired memory location can be accessed in unit time;
• Every instruction takes the same amount of time.

This model works well when the amount of memory that the algorithm needs to use is smaller than the amount of memory in the computer running the code.

But in the real world, some applications need to process data that are too large to fit into a computer’s main memory at once. And the algorithms used to solve such problems are called external memory algorithms or disk-based algorithms; since the input data are stored on an external memory storage device.

Instead of the RAM model, the external memory algorithm is analyzed based on the external memory model. The model consists of a CPU processor with an internal memory of bounded size, connected to an unbounded external memory. One I/O operation consists of moving a block of contiguous elements from external to internal memory(this is called page cache and is managed by the kernel.).

Compared with the main memory, the access speed of external memory is slower by several orders of magnitude, even though modern storage techniques such as SSD are already adopted. Thus, for external algorithms, the bottleneck of the performance is disk IO speed instead of the CPU cycles.

In this section, we examined the computational model of the external memory algorithms. Next, let’s review one typical type of external memory algorithm: external sorting.

External Mergesort

Similar to the traditional internal sorting algorithms, several different solutions are available for external sorting. In this article, I will focus on external mergesort.

External mergesort is a straightforward generalization of internal mergesort. The algorithm starts by repeatedly loading M input items into the memory(since the memory buffer size is limited, can only store M input items at once), sorting them, and writing them back out to disk. This step divides the input file into some(or many, if the input file is very large) sorted runs, each with M items sorted. Then the algorithm repeatedly merges multiple sorted runs, until only a single sorted run containing all the input data remains.

Let’s use the following example model to analyze this algorithm. First of all, assume the memory buffer is limited, and the size is one page. And the input file size is 8 pages. External mergesort can be divided into two phases: sort phase and merge phase.

Sort phase:

• Divide the entire input file into 8 groups, each group size is one page(memory buffer capacity).
• Load each page into the memory, sort the items in the memory buffer(with an internal sorting algorithm), and save the sorted items in a temporary sub-file. Each sub-file is called a run.

At the end of the sort phase, 8 temporary sorted 1-page runs will be created. This step can be marked as pass 0.

Merge phase:

The 8 sorted runs in pass 0 will be merged into a single sorted file with 3 more passes.

• pass 1: Perform 4 runs for the merge.
• Run 1: Merge the first two small 1-page runs into a big 2-page run. This merging step goes as follows:
• Read the first two sorted sub-files (one item from each file).
• Find the smaller item, output it in the new sub-file, and the appropriate input sub-file is advanced. Repeat this cycle until the input sub-file is completed. This routine’s logic is the same as the internal mergesort algorithm.
• Run 2: Merge the next two 1-page runs into a 2-page run.
• Run 3 and 4: follow the same process.
• At the end of pass 1, 4 temporary sorted 2-page runs will be created.
• pass 2: Perform 2 runs for the merge.
• At the end of pass 2, 2 temporary sorted 4-page runs will be created.
• pass 3: Perform 1 run for the merge.
• At the end of pass 3, the final 8-page run containing all the sorted items will be created.

Note: the above process may seem complicated at the first sight, but the logic is nearly the same as the internal merge sort. The only difference is internal merging is based on memory buffers, while external merging is based on disk files, and needs reading the item from disk to memory.

Since we keep merging two small sub-files into a big one with doubled size, the above algorithm can be called two-way external merge sorting. We can generalize the idea to multi-way external merge sorting, which merges M runs into one.

Next, let’s analyze its complexity. Suppose the input file has N items and each page consists of B items. And M denotes the number of ways used in the merge phase, thus the number of passes should be: logM(N/B) + 1, where plus one means the first pass in the sort phase. And each pass, each item is read and written once from and to the disk file. So the total number of disk I/O is: 2N*(logM(N/B) + 1)

Summary

In this article, we examined the abstract computational model of the external memory algorithm and analyzed the details of the external mergesort algorithm. Next article, let’s implement the code and evaluate the performance.

Understand userland heap memory allocation: part three - free chunk

Introduction

In the last article, we investigated how the allocated chunks are aligned and stored in the heap. This article continues to examine how to free a chunk of memory and how the freed chunks are stored in the heap.

Hands-on demo

Let’s continue debugging the demo code shown in the last article:

Previously, we allocated a chunk of memory and put data in it. The next line will free this chunk. Before we run the instruction and show the demo result, let’s discuss the theory first.

The freed chunk will not be returned to the kernel immediately after the free is called. Instead, the heap allocator keeps track of the freed chunks in a linked list data structure. So the freed chunks in the linked list can be reused when the application requests new allocations again. This can decrease the performance overhead by avoiding too many system calls.

The allocator could store all the freed chunks together in a long linked list, this would work but the performance would be slow. Instead, the glibc maintains a series of freed linked lists called bins, which can speed up the allocations and frees. We will examine how bins work later.

It is worth noting that each free chunk needs to store pointers to other chunks to form the linked list. That’s what we discussed in the last section, there’re two points in the malloc_chunk structure: fd and bk, right? Since the user data region of the freed chunk is free for use by the allocator, so it repurposes the user data region as the place to store the pointer.

Based on the above description, the following picture illustrates the exact structure of a freed chunk:

Now step over one line in gdb and check chunks in the heap as follows:

You can see the changes: the allocated chunk is marked as a Free chunk (tcache) and pointer fd is set(which indicates this freed chunk is inserted into a linked list).

The tcache is one kind of bins provided by glibc. The gdb pwndbg plugin allows you to check the content of bins by running command bins as follows:

Note that the freed chunk(at 0x5555555592a0) is inserted into tcache bins as the liked list header.

Note that there 5 types of bins: small bins, large bins, unsorted bins, fast bins and tcache bins. If you don’t know, don’t worry I will examine them in the following section.

According to the definition, after the second malloc(100) is called, the allocator should reuse the freed chunk in the bins. The following image can prove this:

The freed chunk at 0x555555559290 is in use again and all bins are empty after the chunk is removed from the linked list. All right!

Recycling memory with bins

Next, I want to spend a little bit of time examining why we need bins and how bins optimize chunk allocation and free.

If the allocator keeps track of all the freed chunks in a long linked list. The time complexity is O(N) for the allocator to find a freed chunk with fit size by traversing from the head to the tail. If the allocator wants to keep the chunks in order, then at least O(NlogN) time is needed to sort the list by size. This slow process would have a bad impact on the overall performance of programs. That’s the reason why we need bins to optimize this process. In summary, the optimization is done on the following two aspects:

• High-performance data structure
• Per-thread cache without lock contention

High-performance data structure

Take the small bins and large bins as a reference, they are defined as follows:

They are defined together in an array of linked lists and each linked list(or bin) stores chunks that are all the same fixed size. From bins[2] to bins[63] are the small bins, which track freed chunks less than 1024 bytes while the large bins are for bigger chunks. small bins and large bins can be represented as a double-linked list shown below:

The glibc provides a function to calculate the index of the corresponding small(or large) bin in the array based on the requested size. Since the index operation of the array is in O(1) time. Moreover, each bin contains chunks of the same size, so it can also take O(1) time to insert or remove one chunk into or from the list. As a result, the entire allocation time is optimized to O(1).

bins are LIFO(Last In First Out) data structure. The insert and remove operations can be illustrated as follows:

Moreover, for small bins and large bins, if the neighbors of the current chunk are free, they are merged into a larger one. That’s the reason we need a double-linked list to allow running fast traverse both forward and backward.

Unlike small bins and large bins, fast bins and tcache bins chunks are never merged with their neighbors. In practice, the glibc allocator doesn’t set the P special flag at the start of the next chunk. This can avoid the overhead of merging chunks so that the freed chunk can be immediately reused if the same size chunk is requested. Moreover, since fast bins and tcache bins are never merged, they are implemented as a single-linked list.

This can be proved by running the second free method in the demo code and checking the chunks in the heap as follows:

First, the top chunk’s size is still 0x20d01 rather than 0x20d00, which indicates the P bit is equal to 1. Second, the Free chunk only has one pointer: fd. If it’s in a double-linked list, both fd and bk should point to a valid address.

Per-thread cache without lock contention

The letter t in tcache bins represents the thread, which is used to optimize the performance of multi-thread programs. In multi-thread programming, the most common way solution to prevent the race condition issue is using the lock or mutex. Similarly, The glibc maintains a lock in the data structure for each heap. But this design comes with a performance cost: lock contention, which happens when one thread attempts to acquire a lock held by another thread. This means the thread can’t do any tasks.

tcache bins are per-thread bins. This means if the thread has a chunk on its tcache bins, it can serve the allocation without waiting for the heap lock!

Summary

In this article, we examined how the userland heap allocaor works by debugging into the heap memory with gdb. The discussion is fully based on the glibc implementation. The design and behavior of the glibc heap allocator are complex but interesting, what we covered here just touches the tip of the iceberg. You can explore more by yourself.

Moreover, I plan to write a simple version of a heap allocator for learning and teaching purpose. Please keep watching my blog!

Understand userland heap memory allocation: part two - allocate chunk

Introduction

The previous article gave a general overview of memory management. The story goes on. In this section, let’s break into the heap memory to see how it works basically.

Memory allocator

We need to first understand some terminology in the memory management field:

• mutator: the program that modifies the objects in the heap, which is simply the user application. But I will use the term mutator in this article.
• allocator: the mutator doesn’t allocate memory by itself, it delegates this generic job to the allocator. At the code level, the allocator is generally implemented as a library. The detailed allocation behavior is fully determined by the implementations, in this article I will focus on the memory allocator in the library of glibc.

The relationship between the mutator and allocator is shown in the following diagram:

There is a third component in the memory management field: the garbage collector(GC). GC reclaims memories automatically. Since this article is talking about manual heap memory allocation in system programming, we will ignore GC for now. GC is a very interesting technical challenge, I will examine it in the future. Please keep watching my blog!

Hands-on demo

We will use gdb and pwndbg(which is a gdb plugin) and break into the heap memory to see how it works. The gdb provides the functionality to extend it via Python plugins. pwndbg is the most widely used.

The demo code is as follows:

The demo code above just allocates some memory, set the content of the memory and releases it later. And then allocate other chunks of memory again. Very simple, all right?

First, set a breakpoint at line 7(the first malloc call) and run the program in gdb. Then run vmmap command from pwndbg, which can get the process memory layout as follows:

Note that there is no heap segment yet before the first malloc call is made. After step over one line in gdb, check the layout again:

Now the heap segment is created with the size of 132KB(21000 in hexadecimal). As described above, the kernel maps 132KB of physical memory to this process’s virtual memory and marks this 132KB block of physical memory as used to isolate other processes. This mapping routine is done via system calls like brk, sbrk and mmap. Please investigate these system calls yourself.

132KB is much bigger than the 100B(the size passed to malloc). This behavior can answer one question at the beginning of this article. The system calls aren’t necessary to be triggered each time when malloc is called. This design is aimed to decrease performance overhead. Now the 132KB heap memory is maintained by the allocator. Next time the application calls malloc again, the allocator will allocate memory for it.

Next, step one more line in gdb to assign value(“AAAABBBBCCCCDDDD”) to the allocated block. Let’s check the content of this 132KB heap segment with heap command as follows:

There are 3 chunks. Let’s examine these chunks one by one.

The top chunk contains all the remaining memories which have not been allocated yet. In our case, the kernel maps 132KB of physical memory to this process. And 100B memory is allocated by calling malloc(100), so the remaining memories are in the top chunk. The top chunk stays on the border of the heap segment, and it can grow and shrink as the process allocates more memory or release unused memory.

Then let’s look at the chunk with the size of 0x291. The allocator uses this chunk to store heap management structures. It is not important for our analysis, just skip it.

What we care about is the chunk in the middle with a size of 0x71. It should be the block we requested and contains the string “AAAABBBBCCCCDDDD”. We can verify this point by checking its content:

gdb’s x command can display the memory contents at a given address using the specified format. x/40wx 0x555555559290 prints 40 words(each word is 32 bits) of memories starting from 0x555555559290 in the hexadecimal format.

We can see that the string “AAAABBBBCCCCDDDD” is there. So our guess is correct. But the question is why the size of this chunk is 0x71. To understand this, we need to first analyze how the allocator stores chunk. A chunk of memory is represented by the following structure:

• prev_size: the size of the previous chunk only when the previous chunk is free, otherwise when the previous chunk is in use it stores the user data of the previous chunk.
• size: the size of the current chunk.
• fd: pointer to the next free chunk only when the current chunk is free, otherwise when the current chunk is in use it stores the user data.
• bk: pointer to the previous free chunk. Behaves in the same way as pointer fd.

Based on the above description, the following picture illustrates the exact structure of an allocated chunk:

• chunk: indicates the real starting address of the object in the heap memory.
• mem: indicates the returned address by malloc.

The memory in between is reserved for the metadata mentioned above: prev_size and size. On a 64-bit system, they’re (type of INTERNAL_SIZE_T) 8 bytes in length.

For the size field, it is worth noting:

• It includes both the size of metadata and the size of the actual user data.
• It is usually aligned to a multiple of 16 bytes. You can investigate the purpose of memory alignment by yourself.
• It contains three special flags(A|M|P) at the three least significant bits. We can ignore the other two bits for now, but the last bit indicates whether the previous chunk is in use(set to 1) or not(set to 0).

According to this, let’s review the content of this chunk again:

I add marks on the image to help you understand. Let’s do some simple calculations. 100 + 8 = 108, 100 is the size of memory we requested, 8 is the size of metadata(for size field). Then 108 is aligned to 112 as a multiple of 16 bytes. Finally, since the special flag P is set to 1, then we get 112 + 1 = 113(0x71)(that’s the reason why the size is 0x71 instead of 0x70).

In this section, we break into the heap segment and see how an allocated chunk works. Next, we’ll check how to free a chunk.

Understand userland heap memory allocation: part one - overview

Introduction

In my eyes, compared with developing applications with high-level programming languages, one of the biggest differences for system programming with low-level languages like C and C++, is you have to manage the memory by yourself. So you call APIs like malloc, and free to allocate the memory based on your need and release the memory when the resource is no longer needed. It is not only one of the most frequent causes of bugs in system programming; but also can lead to many security issues.

It’s not difficult to understand the correct usage of APIs like malloc, and free. But have you ever wondered how they work, for example:

• When you call malloc, does it trigger system calls and delegate the task to the kernel or there are some other mechanisms?
• When you call malloc(10) and try to allocate 10 bytes of heap memory, how many bytes of memory do you get? 10 bytes or more?
• When the memory is allocated, where exactly the heap objects are located?
• When you call free, is the memory directly returned to the kernel?

Note that memory is a super complex topic, so I can’t cover everything about it in one article (In fact, what is covered in this article is very limited). This article will focus on userland memory(heap) allocation.

Process memory management overview

Process virtual memory

Every time we start a program, a memory area for that program is reserved, and that’s process virtual memory as shown in the following image:

You can note that each process has one invisible memory segment containing kernel codes and data structures. This invisible memory segment is important; since it’s directly related to virtual memory, which is employed by the kernel for memory management. Before we dive into the other different segments, let’s understand virtual memory first.

Virtual memory technique

Why do we need virtual memory? Virtual memory is a service provided by the kernel in the form of abstraction. Without virtual memory, applications would need to manage their physical memory space, coordinating with every other process running on the computer. Virtual memory leaves that management to the kernel by creating the maps that allow translation between virtual and physical memory. The kernel creates an illusion that each process occupies the entire physical memory space. We can also realize process isolation based on virtual memory to enhance security.

Virtual memory is out of this article’s scope, if you’re interested, please take a look at the core techniques: paging and swapping.

Static vs Dynamic memory allocation

Next, let’s take a close look at the process memory layout above and understand where they are from. Generally speaking, there are two ways via which memories can be allocated for storing data: static and dynamic. Static memory allocation happens at compile time, while dynamic memory allocation occurs at runtime.

When a program started, the executable file(on the Linux system, it’s called an ELF file) will be loaded into the memory as a Process Image. This ELF file contains the following segments:

• .TEXT: contains the executable part of the program with all the machine codes.
• .DATA: contains initialized static and global variables.
• .BSS: is short for block started by symbol contains uninitialized static and global variables.

The ELF file will be loaded by the kernel and create a process image. And these static data will be mapped into the corresponding segments of the virtual memory. The ELF loader is also an interesting topic, I will write another article about it in the future. Please keep watching my blog!

The memory-mapped region segment is used for storing the shared libraries.

Finally, stack and heap segments are produced at runtime dynamically, which are used to store and operate on temporary variables that are used during the execution of the program. Previously, I once wrote an article about the stack, please refer to it if you want to know the details.

The only remaining segment we didn’t mention yet is the heap, which is this article’s focus!

You can check the memory layout of one process by examining this file /proc/{pid}/maps as below:

Note that the above investigation doesn’t consider multiple threads. The memory layout of the process with multi-threads will be more complex, please refer to other online documents.

In this section, we had a rough overview of memory management from top to bottom. Hope you can see the big picture and know where we are. Next, let’s dig into the heap segment and see how it works.

cPacketSniffer

Background

In this post, I want to introduce my new project: cPacketSniffer. I worked on it for the past two months. Finally, I worked it out and feel very proud of putting it here!

Motivation

Simply speaking, I want to sharpen my techniques in network programming and Linux system programming. Both of these two topics can lead you to the bottom of computers or software. Feynman said “There is plenty of room at the bottom”, I think this physics law can apply to software as well.

Acknowledgement

It’s very lucky for me to come across this site “Network programming in Linux”, which developed a network packet capturing tool with C++. After confirming that the documents and source code on this site is completed and clear, I decided to refactor it with C language. That’s the starting point for my project cPacketSniffer.

Features

As a network packets sniffer, cPacketSniffer provides the following features:

• Integrate with libpcap to support: filtering captured packets, capturing packets offline, capturing packets on specific devices and capturing packets in promiscuous mode.
• Analyze network packets at low layers of TCP/IP stack, including Ethernet, ARP, ICMP, IP(IPv4), TCP, UDP, etc. Also one protocol in the application layer: TFTP.
• Detect network security attacks:
• ARP spoofing detection.
• Ping flood detection.
• Analyze and track network traffics:
• TCP session tracking and traffic analysis.
• TFTP session tracking and traffic analysis.

The following images demonstrate some typical usages of cPacketSniffer:

Packet Analysis:

ARP Spoofing Detection:

PING Flood Detection:

TCP Session Tracking:

Besides the above network programming-related functionalities, it also covers the following points:

• Develop a generic data structure in C.
• Error handling in C.
• Data encapsulation (object-oriented style programming) in C.
• Manual memory management in C.
• etc.

This article will not cover these points in detail, I will write articles on these topics separately in the future. Please keep watching my blog!

Future work

Now cPacketSniffer can work as a network packet sniffer based on the design. Moreover, it can also serve as a testbed to try experimental features. Next step I plan to try the following ideas:

• Implement the network intrusion detection function.
• Improve the performance with advanced data structures, like binary search trees.
• Memory and cache performance tuning.
• Automatic memory management by Garbage Collection.
• Integrate ncurses for Text-based user interface.

Write a Linux firewall from scratch based on Netfilter: part three - Netfilter module

Background

In the previous article, we examined how to write a Kernel module and load it dynamically into a running Linux system. Based on this understanding, let’s continue our journey to write a Netfilter module as our mini-firewall.

Netfilter architecture.

Basics of Netfilter hooks

The Netfilter framework provides a bunch of hooks in the Linux kernel. As network packets pass through the protocol stack in the kernel, they will traverse these hooks as well. And Netfilter allows you to write modules and register callback functions with these hooks. When the hooks are triggered, the callback functions will be called. This is the basic idea behind Netfilter architecture. Not difficult to understand, right?

Currently, Netfilter provides the following 5 hooks for IPv4:

• NF_INET_PRE_ROUTING: is triggered right after the packet has been received on a network card. This hook is triggered before the routing decision was made. Then the kernel determines whether this packet is destined for the current host or not. Based on the condition, the following two hooks will be triggered.
• NF_INET_LOCAL_IN: is triggered for network packets that are destined for the current host.
• NF_INET_FORWARD: is triggered for network packets that should be forwarded.
• NF_INET_POST_ROUTING: is triggered for network packets that have been routed and before being sent out to the network card.
• NF_INET_LOCAL_OUT: is triggered for network packets generated by the processes on the current host.

The hook function you defined in the module can mangle or filter the packets, but it eventually must return a status code to Netfilter. There are several possible values for the code, but for now, you only need to understand two of them:

• NF_ACCEPT: this means the hook function accepts the packet and it can go on the network stack trip.
• NF_DROP: this means the packet is dropped and no further parts of the network stack will be traversed.

Netfilter allows you to register multiple callback functions to the same hook with different priorities. If the first hook function accepts the packet, then the packet will be passed to the next functions with low priority. If the packet is dropped by one callback function, then the next functions(if existing) will not be traversed.

As you see, Netfilter has a big scope and I can’t cover every detail in the articles. So the mini-firewall developed here will work on the hook NF_INET_PRE_ROUTING, which means it works by controlling the inbound network traffic. But the way of registering the hook and handling the packet can be applied to all other hooks.

Note: there is another remarkable question: what’s the difference between Netfilter and eBPF? If you don’t know eBPF, please refer to my previous article. Both of them are important network features in the Linux kernel. The important thing is Netfilter and eBPF hooks are located in different layers of the Kernel. As I drew in the above diagram, eBPF is located in a lower layer.

Kernel code of Netfilter hooks

To have a clear understanding of how the Netfilter framework is implemented inside the protocol stack, let’s dig a little bit deeper and take a look at the kernel source code (Don’t worry, only shows several simple functions). Let’s use the hook NF_INET_PRE_ROUTING as an example; since the mini-firewall will be written based on it.

When an IPv4 packet is received, its handler function ip_rcv will be called as follows:

In this handler function, you can see the hook is passed to the function NF_HOOK. Based on the name NF_HOOK, you can guess that it is for triggering the Netfilter hooks. Right? Let’s continue to examine how NF_HOOK is implemented as follows:

The function NF_HOOK contains two steps:

• First, runs the hook’s callback functions by calling the underlying function nf_hook.
• Second, invokes the function okfn (passed to NF_HOOK as the argument), if the packet passes through the hook functions and doesn’t drop.

For the hook NF_INET_LOCAL_IN, the function ip_rcv_finish will be invoked after the hook functions pass. Its job is to pass the packet on to the next protocol handler(TCP or UDP) in the protocol stack to continue its journey!

The other 4 hooks all use the same function NF_HOOK to trigger the callback functions. The following table shows where the hooks are embedded in the kernel, I leave them to the readers.

Hook File Function
NF_INET_PRE_ROUTING /kernel-src/net/ipv4/ip_input.c ip_rcv()
NF_INET_LOCAL_IN /kernel-src/net/ipv4/ip_input.c ip_local_deliver()
NF_INET_FORWARD /kernel-src/net/ipv4/ip_forward.c ip_forward()
NF_INET_POST_ROUTING /kernel-src/net/ipv4/ip_output.c ip_build_and_send_pkt()
NF_INET_LOCAL_OUT /kernel-src/net/ipv4/ip_output.c ip_output()

Next, Let’s review the Netfilter’s APIs to create and register the hook function.

Netfilter API

It’s straightforward to create a Netfilter module, which involves three steps:

• Define the hook function.
• Register the hook function in the kernel module initialization process.
• Unregister the hook function in the kernel module clean-up process.

Let’s go through them quickly one by one.

Define a hook function

The hook function name can be whatever you want, but it must follow the signature below:

The hook function can mangle or filter the packet whose data is stored in the sk_buff structure (we can ignore the other two parameters; since we don’t use them in our mini-firewall). As we mentioned above, the callback function must return a Netfilter status code which is an integer. For instance, the accepted and dropped status is defined as follows:

Register and unregister a hook function

To register a hook function, we should wrap the defined hook function with related information, such as which hook you want to bind to, the protocol family and the priority of the hook function, into a structure struct nf_hook_ops and pass it to the function nf_register_net_hook.

Most of the fields are very straightforward to understand. The one need to emphasize is the field hooknum, which is just the Netfilter hooks discussed above. They are defined as enumerators as follows:

Next, let’s take a look at the functions to register and unregister hook functions goes as follows:

The first parameter struct net is related to the network namespace, we can ignore it for now and use a default value.

Next, let’s implement our mini-firewall based on these APIs. All right?

Implement mini-firewall

First, we need to clarify the requirements for our mini-firewall. We’ll implement two network traffic control rules in the mini-firewall as follows:

• Network protocol rule: drops the ICMP protocol packets.
• IP address rule: drops the packets from one specific IP address.

The completed code implementation is in this Github repo.

Drop ICMP protocol packets

ICMP is a network protocol widely used in the real world. The popular diagnostic tools like ping and traceroute run the ICMP protocol. We can filter out the ICMP packets based on the protocol type in the IP headers with the following hook function:

The logic in the above hook function is easy to understand. First, we retrieve the IP headers from the network packet. And then according to the protocol type field in the headers, we decided to accept TCP and UDP packets but drop the ICMP packets. The only technique we need to pay attention to is the function ip_hdr, which is the kernel function defined as follows:

The function ip_hdr delegates the task to the function skb_network_header. It gets IP headers based on the following two data:

• head: is the pointer to the packet;
• network_header: is the offset between the pointer to the packet and the pointer to the network layer protocol header. In detail, you can refer to this document.

Next, we can register the above hook function as follows:

The above logic is self-explaining. I will not spend too much time here.

Next, it’s time to demo how our mini-firewall works.

Demo time

Before we load the mini-firewall module, the ping command can work as expected:

In contrast, after the mini-firewall module is built and loaded (based on the commands we discussed previously):

You can see all the packets are lost; because it is dropped by our mini-firewall. We can verify this by running the command dmesg:

But other protocol packets can still run through the firewall. For instance, the command wget 142.250.4.103 can return normally as follows:

Next, let’s try to ban the traffic from this IP address.

Drop packets source from one specific IP address

As we mentioned above, multiple callback functions are allowed to be registered on the same Netfilter hook. So we will define the second hook function with a different priority. The logic of this hook function goes like this: we can get the source IP address from the IP headers and make the drop or accept decision according to it. The code goes as follows

This hook function uses two interesting techniques:

• ntohl: is a kernel function, which is used to convert the value from network byte order to host byte order. Byte order is related to the computer science concept of Endianness. Endianness defines the order or sequence of bytes of a word of digital data in computer memory. A big-endian system stores the most significant byte of a word at the smallest memory address. A little-endian system, in contrast, stores the least-significant byte at the smallest address. Network protocol uses the big-endian system. But different OS and platforms run various Endianness system. So it may need such conversion based on the host machine.

• IPADDRESS: is a macro, which generates the standard IP address format(four 8-bit fields separated by periods) from a 32-bit integer. It uses the technique of the equivalence of arrays and pointers in C. I will write another article to examine what it is and how it works. Please keep watching my updates!

Next, we can register this hook function in the same way discussed above. The only remarkable point is this callback function should have a different priority as follows:

Let’s see how it works with a demo.

Demo time

After re-build and re-load the module, we can get:

The wget 142.250.4.103 can’t return response. Because it is dropped by our mini-firewall. Great!

More space to expand

You can find the full code implementation here. But I have to say, our mini-firewall only touches the surface of what Netfilter can provide. You can keep expanding the functionalities. For example, currently, the rules are hardcoded, why not make it possible to config the rules dynamically. There are many cool ideas worth trying. I leave it for the readers.

Summary

In this article, we implement the mini-firewall step by step and examined many detailed techniques. Not only code; but we also verify the behavior of the mini-firewall by running real demos.

Write a Linux firewall from scratch based on Netfilter: part two - hello world module

Background

In the last article, we examined the basics of Netfilter and Linux kernel modules in theory. Starting from this article, we will make our hands dirty and start implementing our mini-firewall. We will walk through the whole process step by step. In this article, let’s write our first Linux kernel module using a simple hello world demo. Then let’s learn how to build the module(which is very different from compiling an application in the user space) and how to load it in the kernel. After understanding how to write a module, in the next article, let’s write the initial version of our mini-firewall module using Netfilter’s hook architecture. All right. Let’s start the journey.

Make the first Kernel module

First, I have to admit that Linux Kernel module development is a kind of large and complex technology topic. And there are many great online resources about it. This series of articles is focusing on developing the mini-firewall based on Netfilter, so we can’t cover all the aspects of the Kernel module itself. In future articles, I’ll examine more in-depth knowledge of kernel modules.

Write the module

You can write the hello world Kernel module with a single C source code file hello.c as follows:

We can write a Kernel module in such an easy and simple way because the Linux Kernel does the magic for you. Remember the design philosophy of Linux(Unix): Design for simplicity; add complexity only where you must.

Let’s examine several technical points worth to remark as follows:

First, Kernel modules must have at least two functions: a “start” function which is called when the module is loaded into the kernel, and an “end” function which is called just before it is removed from the kernel. Before kernel 2.3.13, the names of these two functions are hardcoded as init_module() and cleanup_module(). But in the new versions, you can use whatever name you like for the start and end functions of a module by using the module_init and module_exit macros. The macros are defined in include/linux/module.h and include/linux/init.h. You can refer there for detailed information.

Typically, module_init either registers a handler for something with the kernel (for example, the mini-firewall developed in this article), or it replaces one of the kernel functions with its own code (usually code to do something and then call the original function). The module_exit function is supposed to undo whatever module_init did, so the module can be unloaded safely.

Second, printk function provides similar behaviors to printf, which accepts the format string as the first argument. The printk function prototype goes as follows:

printk function allows a caller to specify log level to indicate the type and importance of the message being sent to the kernel message log. For example, in the above code, the log level KERN_INFO is specified by prepending to the format string. In C programming, this syntax is called string literal concatenation. (In other high-level programming languages, string concatenation is generally done with + operator). For the function printk and log level, you can find more information in include/linux/kern_levels.h and include/linux/printk.h.

Note: The path to header files for Linux kernel module development is different from the one you often used for the application development. Don’t try to find the header file inside /usr/include/linux, instead please use the following path /lib/modules/uname -r/build/include/linux (uname -r command returns your kernel version).

Next, let’s build this hello-world kernel module.

Build the module

The way to build a kernel module is a little different from how to build a user-space application. The efficient solution to build kernel image and its modules is Kernel Build System(Kbuild).

Kbuild is a complex topic and I won’t explain it in too much detail here. Simply speaking, Kbuild allows you to create highly customized kernel binary images and modules. Technically, each subdirectory contains a Makefile compiling only the source code files in its directory. And a top-level Makefile recursively executes each subdirectory’s Makefile to generate the binary objects. And you can control which subdirectories are included by defining config files. In detail, you can refer to other documents.

The following is the Makefile for the hello world module:

The make -C dir command changes to directory dir before reading the makefiles or doing anything else. The top-level Makefile in /lib/modules/\$(shell uname -r)/build will be used. You can find that command make M=dir modules is used to make all modules in specified dir.

And in the module-level Makefile, the obj-m syntax tells kbuild system to build module_name.o from module_name.c, and after linking, will result in the kernel module module_name.ko. In our case, the module name is hello.

The build process goes as follows:

After the build, you can get several new files in the same directory:

The file ends with .ko is the kernel module. You can ignore other files now, I will write another article later to have a deep discussion about the kernel module system.

With the file command, you can note that the kernel module is an ELF(Executable and Linkable Format) format file. ELF files are typically the output of a compiler or linker and are a binary format.

Next step, let’s try to install and remove the module dynamically. You need to know the following three commands:

• lsmod: shows the list of kernel modules currently loaded.
• insmod: inserts a module into the Linux Kernel by running sudo insmod module_name.ko
• rmmod: removes a module from the Linux Kernel by running sudo rmmod module_name

Since the hello world module is quite simple, you can easily install and remove the module as you wish. I will not show the detailed commands here and leave it to the readers.

Note: It doesn’t mean that you can easily install and remove any kernel module without any issues. If the module you are loading has bugs, the entire system can crash.

Debug the module

Next step, let’s prove that the hello world module is installed and removed as expected. We will use dmesg command. dmesg (diagnostic messages) can print the messages in the kernel ring buffer.

First, a ring buffer is a data structure that uses a single, fixed-size buffer as if it were connected end-to-end. The kernel ring buffer is a ring buffer that records messages related to the operation of the kernel. As we mentioned above, the kernel logs printed by the printk function will be sent to the kernel ring buffer.

We can find the messages produced by our module with command dmesg | grep world as follows:

Now you can see that the hello world is loaded into the kernel correctly. And it can be removed dynamically as well. Great.

Summary

In this article, we examine how to write a kernel module, how to build it and how to install it into the kernel dynamically. Next article we can work on the mini-firewall as a Netfilter module.

Write a Linux firewall from scratch based on Netfilter: part one- Netfilter and Kernel Modules

Background

Firewalls are an important tool that can be configured to protect your servers and infrastructure. Firewalls’ main functionalities are filtering data, redirecting traffic, and protecting against network attacks. There are both hardware-based firewalls and software-based firewalls. I will not discuss too much about the background here, since you can find many online documents about it.

Have you ever thought of implementing a simple firewall from scratch? Sounds crazy? But with the power of Linux, you can do that. After you read this series of articles, you will find that actually, it is quite simple.

You may once use various firewalls on Linux such as iptables, nftables, UFW, etc. All of these firewall tools are user-space utility programs, and they are all relying on Netfilter. Netfilter is the Linux kernel subsystem that allows various networking-related operations to be implemented. Netfilter allows you to develop your firewall using the Linux Kernel Module. If you don’t know the techniques such as the Linux Kernel module and Netfilter, don’t worry. In this article, let’s write a Linux firewall from scratch based on Netfilter. You can learn the following interesting points:

• Linux kernel module development.
• Linux kernel network programming.
• Netfilter module development.

Netfilter and Kernel modules

Basics of Netfilter

Netfilter can be considered to be the third generation of firewall on Linux. Before Netfilterwas introduced in Linux Kernel 2.4, there are two older generations of firewalls on Linux as follows:

• The first generation was a port of an early version of BSD UNIX’s ipfw to Linux 1.1.
• The second generation was ipchains developed in the 2.2 series of Linux Kernel.

As we mentioned above, Netfilter was designed to provide the infrastructure inside the Linux kernel for various networking operations. So firewall is just one of the multiple functionalities provided by Netfilter as follows:

• Packet filtering: is in charge of filtering the packets based on the rules. It is also the topic of this article.
• NAT (Network address translation): is in charge of translating the IP address of network packets. NAT is an important protocol, which has become a popular and essential tool in conserving global address space in the face of IPv4 address exhaustion. If you don’t know NAT protocol, you can refer to other documents. I will examine it in other future articles.
• Packet mangling: is in charge of modifying the packet content(In fact, NAT is one kind of packet mangling, which modifies the source or destination IP address). For example, MSS (Maximum Segment Size) value of TCP SYN packets can be altered to allow large-size packets transported over the network.

Note: this article will focus on building a simple firewall to filter packets based on Netfilter. So the NAT and Packet Mangling parts are not in the scope of this article.

Packet filtering can only be done inside the Linux kernel (Netfilter’s code is in the kernel as well), if we want to write a mini firewall, it has to run in the kernel space. Right? Does it mean we need to add our code into the kernel and recompile the kernel? Imagine you have to recompile the kernel each time you want to add a new packet filtering rule. That’s a bad idea. The good news is that Netfilter allows you to add extensions using the Linux kernel modules.

Basics of Linux Kernel modules

Although Linux is a monolithic kernel, it can be extended using kernel modules. Modules can be inserted into the kernel and removed on demand. Linux isolates the kernel but allows you to add specific functionality on the fly through modules. In this way, Linux keeps a balance between stability and usability.

I want to examine one confusing point about the kernel module here: what is the difference between driver and module:

• A driver is a bit of code that runs in the kernel to talk to some hardware device. It drives the hardware. Standard practice is to build drivers as kernel modules where possible, rather than link them statically to the kernel since that gives more flexibility.
• A kernel module may not be a device driver at all.

Summary

In the first post of this series, we examine the basics of Netfilter and Linux kernel modules. In the next post, let’s start implementing the mini firewall.

Write a Linux packet sniffer from scratch: part two- BPF

Introduction

In the previous article, we examined how to develop a network sniffer with PF_SOCKET socket in Linux platform. The sniffer developed in the last article captures all the network packets. But a powerful network sniffer like tcpdump should provide the packet filtering functionality. For instance, the sniffer can only capture the TCP segment(and skip the UPD), or it can only capture the packets from a specific source IP address. In this article, let’s continue to explore how to do that.

Background of BPF

Berkeley Packet Filter(BPF) is the essential underlying technology for packet capture in Unix-like operating systems.
Search BPF as the keyword online, and the result is very confusing. It turns out that BPF keeps evolving, and there are several associated concepts such as BPF cBPF eBPF and LSF. So let us examine those concepts along the timeline:

• In 1992, BPF was first introduced to the BSD Unix system for filtering unwanted network packets. The proposal of BPF was from researchers in Lawrence Berkeley Laboratory, who also developed the libpcap and tcpdump.

• In 1997, Linux Socket Filter(LSF) was developed based on BPF and introduced in Linux kernel version 2.1.75. Note that LSF and BPF have some distinct differences, but in the Linux context, when we speak of BPF or LSF, we mean the same packet filtering mechanism in the Linux kernel. We’ll examine the detailed theory and design of BPF in the following sections.

• Originally, BPF was designed as a network packet filter. But in 2013, BPF was widely extended, and it can be used for non-networking purposes such as performance analysis and troubleshooting. Nowadays, the extended BPF is called eBPF, and the original and obsolete version is renamed to classic BPF (cBPF). Note that what we examine in this article is cBPF, and eBPF is not inside the scope of this article. eBPF is the hottest technology in today’s software world, and I’ll talk about it in the future.

Where to place BPF

The first question to answer is where should we place the filter. The last article examines the path of a received packet as follows:

The best solution to this question is to put the filter as early as possible in the path. Since copying a large amount of data from kernel space to the user space produces a huge overhead, which can influence the system performance a lot. So BPF is a kernel feature. The filter should be triggered immediately when a packet is received at the network interface.As the original BPF paper said To minimize memory traffic, the major bottleneck in most modern system, the packet should be filtered ‘in place’ (e.g., where the network interface DMA engine put it) rather than copied to some other kernel buffer before filtering.
Let’s verify this behavior by examining the kernel source code as follows (Note the kernel code shown in this article is based on version 2.6, which contains the cBPF implementation.):

packet_create function handles the socket creation when the application calls the socket system call. In lines 11 and 14, it attaches the hook function to the socket. The hook function executes when the packet is received.

The following code block shows the hook function packet_rcv:

packet_rcv function calls run_filter, which is just the BPF logic part(Currently, you can regard it as a black box. In the next section, we’ll examine the details). Based on the return value of run_filter the packet can be filtered out or put into the queue.

So far, you can understand BPF(or the packet filtering) is working inside kernel space. But the packet sniffer is a user-space application. The next question is how to link the filtering rules in user space to the filtering handler in kernel space.

To answer this question, we have to understand BPF itself. It’s right time to understand this great piece of work.

BPF machine

As I mentioned above, BPF was introduced in this original paper written by researchers from Berkeley. I strongly recommend you read this great paper based on my own experience. In the beginning, I felt crazy to read it, so I read other related documents and tried to understand BPF. But most documents only cover one portion of the entire system, so it is difficult to piece all the information together. Finally, I read the original paper and connected all parts together. As the saying goes, sometimes taking time is actually a shortcut.

Virtual CPU

A packet filter is simply a boolean-valued function on a packet. If the value of the function is true the kernel copies the packet for the application; if it is false the packet is ignored.

In order to be as flexible as possible and not to limit the application to a set of predefined conditions, the BPF is actually implemented as a register-based virtual machine (for the difference between stack-based and register-based virtual machine, you can refer to this article) running a user-defined program.

You can regard the BPF as a virtual CPU. And it consists of an accumulator, an index register(x), a scratch memory store, and an implicit program counter. If you’re not familiar with these concepts, I add some simple illustrations as follows:

• An accumulator is a type of register included in a CPU. It acts as a temporary storage location holding an intermediate value in mathematical and logical calculations. For example, in the operation of “1+2+3”, the accumulator would hold the value 1, then the value 3, then the value 6. The benefit of an accumulator is that it does not need to be explicitly referenced.
• An index register in a computer’s CPU is a processor register or assigned memory location used for modifying operand addresses during the run of a program.
• A program counter is a CPU register in the computer processor which has the address of the next instruction to be executed from memory.

In the BPF machine, the accumulator is used for arithmetic operations, while the index register provides offsets into the packet or the scratch memory areas.

Instructions set and addressing mode

Same as the physical CPU, the BPF provides a small set of arithmetic, logical and jump instructions as follows, these instructions run on the BPF virtual machine(or CPU):

The first column opcodes lists the BPF instructions written in an assembly language style. For example, ld, ldh and ldb means to copy the indicated value into the accumulator. ldx means to copy the indicated value into the index register. jeq means jump to the target instruction if the accumulator equals the indicated value. ret means return the indicated value. You can check the functionality of the instructions set in detail in the paper.

This kind of assembly-like style is more readable to humans. But when we develop an application (like the sniffer written in this article), we use binary code directly as the BPF instruction. This kind of binary format is called BPF Bytecode. I’ll examine the way to convert this assembly language to bytecode later.

The second column addr modes lists the addressing modes allowed for each instruction. The semantics of the addressing modes are listed in the following table:

For instance, [k] means the data at byte offset k in the packet. #k means the literal value stored in k. You can read the paper in detail to check the meaning of other address modes.

Example BPF program

Now let’s try to understand the following small BPF program based on the knowledge above:

The BPF program consists of an array of BPF instructions. For example, the above BPF program contains four instructions.

The first instruction ldh loads a half-word(16-bit) value into the accumulator from offset 12 in the Ethernet packet. According to the Ethernet frame format shown below, the value is just the Ethernet type field. The Ethernet type is used to indicate which protocol is encapsulated in the frame’s payload (for example, 0x0806 for ARP, 0x0800 for IPv4, and 0x86DD for IPv6).

The second instruction jeq compares the accumulator (currently stores Ethernet type field) to 0x800(stands for IPv4). If the comparison fails, zero is returned, and the packet is rejected. If it is successful, a non-zero value is returned, and the packet is accepted. So the small BPF program filters and accepts all IP packets. You can find other BPF programs in the original paper. Go to read it, and you can feel the flexibility of BPF as well as the beauty of the design.

Kernel implementation of BPF

Next, let’s examine how kernel implements BPF. As mentioned above, the hook function packet_rcv calls run_filter to handle the filtering logic. run_filter is defined as follows:

You can find that the real filtering logic is inside sk_run_filter:

Same as we mentioned, sk_run_filter is simply a boolean-valued function on a packet. It maintains the accumulator, the index register, etc. as local variables. And process the array of BPF filter instructions in a for loop. Each instruction will update the value of local variables. In this way, it simulates a virtual CPU. Interesting, right?

BPF JIT

Since each network packet must go through the filtering function, it becomes the performance bottleneck of the entire system.

A just-in-time (JIT) compiler was introduced into the kernel in 2011 to speed up BPF bytecode execution.

• What is a JIT compiler? A JIT compiler runs after the program has started and compiles the code(usually bytecode or some type of VM instructions) on the fly(or just in time) into a form that’s usually faster, typically the host CPU’s native instruction set. This is in contrast to a traditional compiler that compiles all the code to machine language before the program is first run.

In the BPF case, the JIT compiler translates BPF bytecode into a host system’s assembly code directly, which can optimize the performance a lot. I’ll not show details about JIT in this article. You can refer to the kernel code.

Set BPF in sniffer

Next, let’s add BPF into our packet sniffer. As we mentioned above in the application level, the BPF instructions should use bytecode format with the following data structure:

How can we convert the BPF assembly language into bytecode? There are two solutions. First, there is a small helper tool called bpf_asm(which is provided along with the Linux kernel), and you can regard it as the BPF assembly language interpreter. But it is not recommended to application developers.

Second, we can use tcpdump, which provides the converting functionality. You can find the following information from the tcpdump man page:

• -d: Dump the compiled packet-matching code in a human-readable form to standard output and stop.

• -dd: Dump packet-matching code as a C program fragment.

• -ddd: Dump packet-matching code as decimal numbers (preceded with a count).

tcpdump ip means we want to capture all the IP packets. With options -d, -dd and -ddd, the output goes as follows:

Option -d prints the BPF instructions in assembly language (same as the example BPF program shown above). Options -dd prints the bytecode as a C program fragment. So tcpdump is the most convenient tool when you want to get the BPF bytecode.

The BPF filter bytecode (wrapped in the structure sock_fprog) can be passed to the kernel through setsockopt system call as follows:

setsockopt system call triggers two kernel functions: sock_setsockopt and sk_attach_filter (I’ll not show the details for these two functions), which binds the filters to the socket. And in run_filter kernel function (mentioned above), it can get the filters from the socket and execute the filters on the packet.

So far, every piece is connected. The puzzle of BPF is solved. The BPF machine allows the user-space applications to inject customized BPF programs straight into a kernel. Once loaded and verified, BPF programs execute in kernel context. These BPF programs operate inside kernel memory space with access to all the internal kernel states available to it. For example, the cBPF machine which uses the network packet data. But this power can be extended as eBPF, which can be used in many other varied applications. As someone said In some way, eBPF does to the kernel what Javascript does to the websites: it allows all sorts of new application to be created. In the future, I plan to examine eBPF in depth.

Process the packet

We examined the BPF filtering theory on the kernel level a lot in the above section. But for our tiny sniffer, the last step we need to do is process the network packet.

• First, the recvfrom system call reads the packet from the socket. And we put the system call in a while loop to keep reading the incoming packets.

• Then, we print the source and destination MAC address in the packet(the packet we got is a raw Ethernet frame in Layer 2, right?). And if what this Ethernet frame contains is an IP4 packet, then we print out the source and destination IP address. To understand more about it, you can study the header format of various network protocols. I will not cover in details here.

You can find the complete source code of the sniffer in this Github repo.

Summary

In this article, we examine how to add filters to our sniffer. First, we analyze why the filter should be running inside kernel space instead of the application space. Then, this article examines the BPF machine design and implementation in detail based on the paper. We reviewed the kernel source code to understand how to implement the BPF virtual machine. As I mentioned above, the original BPF(cBPF) was extended to eBPF now. But the understanding of the BPF virtual machine is very helpful to eBPF as well.