We will continue to look at winpcap today. The sending process, the network tap and filtering process are all comparable between BPF and NPF but the windows operating system is faster at handling hardware interrupts and in all the operations made by the NIC driver and NDIS code. In order to test the overall flow through the system, an application that calls the packet capture mechanism but discards all the packets was used. This test evaluated the entire WinPCap architecture, including copy process from the interface driver to the kernel buffer and then to the user buffer. There were no filters in this test and all packets received by the tap were delivered to the application. There was no packet loss. The delayed write capability that lets the kernel to wait for a minimum amount of data and to copy a large block of data to user space in a single system call tremendously helped with this test.
Another test aimed to dump all the packets to a file. When the network is overloaded, the systems suffer noticeable losses when the cpu time is not available i.e a new packet arrives when the tap was processing an earlier one.  or when there is no space in the kernel buffer. In this test, there was a non-negligible worsening when the whole packet is dumped to file.
An adhoc program was used to test the monitoring capabilities of WinPCap. The test confirmed that the CPU load is considerably low and that the results match the ones earlier. The additional cost of the monitoring code degrades the user level application results since it requires a non-negligible amount of memory.

We will next look at some portability considerations and performance concerns with WinPCap. Porting of libpcap to winpcap was made easier because BPF and NPF have similar interfaces. There are some system calls in Unix that don't have a direct mapping to the winsock library and so these were written with windows dependent code. The porting resides mostly in the WPCap module. This as we discussed uses the packet module methods instead of the NPF. WinPCap has some differences from libPCap because of Windows dependent code. For example,  Win32 applications cannot use the select function on NPF device in order to know if there are packets that needs to be read so winpcap implements a new event.
To test the performance, two machines were used one as a sender and another as a receiver. The sender generates the traffic and the receiver captures the packets. The packets were made such that it would generate the maximum amount of  packets per second since this is the worst case for a network analyzer. Packet size of 88 bytes showed the maximum number of packets. The minimum packet size on Ethernet is 64 bytes. Ethernet load was full at packet size of 400 bytes.
Tests evaluated the performance of both sending process as well as filtering process. Packets are received by the network tap and checked by the filter. No packets are expected to match the filter so this utilizes the NPF as much as possible. Results show that windows flavors have similar behavior and almost all the packets are received and filtered by NPF.

We continue to review the WinPCap today. We talked about the statistics mode. Next we talk about the packets injection mode. Both BPF and NPF have write capabilities that allow the user sending raw packets to the network. However, libpcap did not support this, thus BPF was never used for this purpose. Unix applications uses raw sockets instead. Win32 has very limited support for raw sockets. This is where WinPCap provides a standard and consistent set of functions for packet injection. The packet data is copied to the kernel buffer and then sent to the network over the NDIS.
There is a well-known libnet packet assembly library that was used to add a layer for packet construction and injection on top of WinPCap.
We now talk about the packet.dll which is one of the modules that comes with WinPCap and provides a common system independent API to enable programs to capture packets  on a variety of windows operating system flavors, architectures, editions and versions. Packet.dll includes several additional functionalities such as low level operations on the network adapters and the dynamic loading of the drivers and some hardware counters.
Note that packet.dll interfaces only with the Network packet filter and does not partiticpate in statistics mode directly off the NDIS.
Besides the NPF and the packet module, the third module that is not OS dependent is the WPcap.dll. This has high level functions such as filter generation and user level buffering, plus advanced features such as statistics and packet injection. This set of API is more granular than the low level APIs exposed by the packet module where there was almost a one to one mapping between the APIs and the kernel mode calls.  The higher level functions of WPCap module are more user-friendly and a single call can translate to several packet module calls.
Win32 network architecture relies on NDIS which is the Network Driver Interface Specification which handles the interaction between the NIC drivers and protocol drivers. Internally, NDIS creates an artificial world of its own abstracting the protocol drivers from the complexity of the hardware. This implies that the same protocol stack can now work with a variety of technologies.
While BPF requires NIC device drivers, NPF does not have the same luxury as a BPF driver specification. Instead NPF locates the network tap as protocol driver on top of the NDIS structures. This pseudo device works the same way as virtual private networking protocol drivers. Due to its location, NPF has the restriction that it cannot tap the lower level packets that do not reach the NDIS upper layer as for example point to point protocol including link control protocols, network control protocols including auth and encryption etc. As a comparision, the Packet Socket in a Linux kernel also suffers from a similar problem.
Note that NPF system calls are all blocking in nature.  Besides, NDIS is performant in that it doesn't copy the entire packet if no protocol drivers require it. Thus NPF provides a clean and isolated plugin for an efficient packet capture on commercial systems.

The other significant difference between a WinPCap NPF and a BSD BPF is that the former has a ring buffer for every user application. that copies data from the kernel mode to the user mode  The copy operation to the user mode buffer happens via a single read call thereby reducing the transitions between the user mode and the kernel mode. This kind of buffer allows the storing of network bursts because it makes more memory available than the BPF.
The kernel buffer is also larger than in BPF. If the application is not able to read as fast  as the driver captures for a limited time interval, the capturing process is penalized. The size of the user buffer is important because it determines the maximum amount of data that can be copied from the kernel space in a single system call.  A smaller buffer is generally suitable for real-time applications since it guarantees that the kernel will copy the memory as soon as the application makes it available. Thus NPF is more configurable in that it allows users to choose between efficiency and responsiveness.
Another configuration parameter is the timeout between read values. By default the timeout is 1 sec and the minimum amount of data is 16K. This is referred to as delayed write.
 One of the core issues of any network analysis and packet capturing is that this is a very CPU intensive task and network packets can overwhelm the CPU. This situation is obviously even worse on faster networks. The typical approach to improving speed include filtering engines and I-copy architectures, which avoid copying packets between kernel space and user space by mapping the kernel buffer in the application's memory. The advantages of this shared buffer may be limited if a user still makes one system call for every packet which results in high number of context switches.
WinPCap introduces the notion that the monitoring not only needs no copying but also pushes it down to the kernel avoiding both data transfer and processing at user mode.
Applications need not call the libpcap APIs to get the data. They can also use the statistics mode of the NPF. Statistics mode avoids packet copies and it implements a 0-copy mechanism - statistic is performed when packet is still in  NIC driver's memory, then the packet is discarded. Moreover, the number of context switches is kept the lowest because the results are returned to the user by a single system call.  The syntax for requesting this statistics is same as in libpcap but doesn't have to go through the libpcap. These are some of the differences.

We will continue to review WinPCap architecture today. We mentioned that the WinPCap has three main components similar to the BSD capturing components. These are:
 the libpcap library based programmability component of WinPCap.
 a Berkeley Packet Filter with its kernel level buffer that keeps all the data.
 a Network Tap that snoops all packets flowing through the network.

We will review these components in detail.
A packet satisfying the filter is copied to the kernel buffer. The kernel buffer is subdivided into two small buffers store and hold that are used to keep the captured packets  These are runtime allocations of blocks of memory. The store buffer is used to keep the data coming from the network adapter and the second one is used to copy the packets to the user buffer. If the store buffer is full and the user buffer is empty, BPF swaps them.  In this way, user level applications does not interfere with the adapters device driver.

Since BPF has proved to be a powerful and stable architecture, the basic structure of WinPCap has these three components as well. However, WinPCap has significant differences in the structure and in the behavior of the capture stack. It can be seen as the evolution of BPF.
The filtering process starts with the libpcap compatible component that is able to accept a user-defined filter and compiles them into a set of pseudo instructions. The kernel module then has to execute these instructions against all incoming packets.

The WinPCap NPF as opposed to BPF uses a circular ring buffer as kernel buffer. It follows  a sliding window protocol. and is a bit more challenging to maintain than that in BPF because data copies can have varying sizes and the bytes copied is updated while the transferring to user mode happens and not after. The kernel has higher priority and can pre-empt the user mode copying. This implementation allows more memory to be used by kernel as compared to only half of the memory used in the BPF swapping buffers. The entire kernel buffer is copied by means of a single read and reducing the number of system calls and therefore the number of context switches between user and kernel mode.

We are reading from the paper on an architecture for high performance network analysis by Risso and  Degioanni

Next we look at the WinPCap architecture. WinPCap adds a similar functionality to what libpcap or tcpdump does to flavors of Unix. There have been other modules in Windows and some with available APIs and each one with a kernel mode driver, however they suffer of severe limitations. However, Netmon API is not freely available and its extensibility is limited. And it did not support sending packets. In this architecture we review some of these functionalities as well.
WinPCap was the first open system for packet capture on Win32 and it fills an important gap between Unix and Windows. Furthermore WinPCap puts performance at the first place.WinPCap consists of a kernel mode component to select packets and a user mode library to deliver them to the applications. The library also provides a low level network access and allows programmers to avoid kernel level programming. WinPCap includes an optimized kernel mode driver called Netgroup packet filter and a set of user-level libraries that are libpcap compatible. From the outset, libpcap compatibility was important to WinPCap so that unix applications could be ported over.
We now look at the BSD capturing components. Getting and sending data over the low level network interfaces was an important objective of BSD. There are three components to BSD. The first block Berkeley Packet Filter is the kernel level component for packet used to store packets coming from the kernel.The Network Tap is designed to snoop all packets flowing through the network and it reads the interface through the interface driver. It is followed by the filter which analyzes incoming packets The Libpcap library is the third component. A packet satisfying the filter for Network Tap is copied to the kernel buffer in the BSD. The user has direct access to each of these three layers. The user accesses the Network Interface Card driver with other protocol stacks to send and receive data.  The user code can directly access the BPF. Lastly the applications can write user code calls to libpcap.

We now look at the optimization techniques with the Berkeley Packet filter in TcpDump. Optimizing the code generation was required because users could specify compound queries. In such queries the generated code was redundant and highly inefficient. As a comparison, if we had taken a simple query such as tcp src port 100, then the resulting code would have been a linear decision tree with stteps evaluating down from IP, frag, TCP, sport, 100, true with each progress down only on true.
IF we wanted the same for both directions, we would have two such trees  with an OR in between  and evaluating one below the other. If the query had been both in same directions such as with tcp port 100 and 200, then we would have both trees one evaluating to the other.  and this combination evaluating one below the other. In this case, we would have had a significant code bloat. This is highly redundant and inefficient.  This is where the optimization techniques are used which have their roots in compiler design. One such technique is called the dominator technique. What this does, it eliminates common subexpressions.  So if the exit from one is the same as entry to another, we can replace that sub-expression with a third. While this traditional technique does optimize the code, it does not address the branching because data could flow to this sub-expression from both . If we look at the edge relationships instead of node relationships, we can do even more optimization. When we assume a particular sub-expression has already been evaluated by the expression above, then we can bypass that sub-expression and directly move to tthe outcome of the sub-expression. This creates new opportunities and when we repeat the cycle for all sub-expressions, at each stage, we could eliminated redundancies. An interesting observation here is that this exercise for optimizing code could also help us detect unreachable code and simplify the graph. Now we take the example above and remove the redundancy of opcode nodes and instead replace the edged to move directly to the outcome of the sub-expressions above. In this case we had the linear decision tree of IP, frag, TCP, common and we remove three sets of those copies and adding edges directly from them to the outcome. We also add edges from src port 100 to dest port 100 and dest port 200 as well as an edge from dest port 100 to src port 200 and completing the branches from the remaining nodes to the outcomes. We only have two outcomes true or false in the leaf level and all the nodes will be connected to it via edges. This covers the optimization in the Berkeley packet filter.