In Part I of this two-part series on the Linux Packet Filter, Gianluca describes a packet's journey through the kernel.
Network geeks among you may remember my article, “Linux Socket Filter: Sniffing Bytes over the Network”, in the June 2001 issue of LJ, regarding the use of the packet filter built inside the Linux kernel. In that article I provided an overview of the functionality of the packet filter itself; this time, I delve into the depths of the kernel mechanisms that allow the filter to work and share some insights on Linux packet processing internals.
In the previous article, some arguments regarding kernel packet processing were raised. It is worthwhile to recall briefly the most important of them:
Packet reception is first dealt with at the network card's driver level, more precisely in the interrupt service routine. The service routine looks up the protocol type inside the received frame and queues it appropriately for later processing.
During reception and protocol processing, packets might be discarded if the machine is congested. Furthermore, as they travel upward toward user land, packets lose network lower-level information.
At the socket level, just before reaching user land, the kernel checks whether an open socket for the given packet exists. If it does not, the packet is discarded.
Then the Linux kernel implements a generic-purpose protocol, called PF_PACKET, which allows you to create a socket that receives packets directly from the network card driver. Hence, any other protocols' handling is skipped, and any packets can be received.
An Ethernet card usually passes only the packets destined to itself to the kernel, discarding all the others. Nevertheless, it is possible to configure the card in such a way that all the packets flowing through the network are captured, independent of their MAC address (promiscuous mode).
Finally, you can attach a filter to a socket, so that only packets matching your filter's rules are accepted and passed to the socket. Combined with PF_PACKET sockets, this mechanism allows you to sniff selected packets efficiently from your LAN.
Even though we built our sniffer using PF_PACKET sockets, the Linux socket filter (LSF) is not limited to those. In fact, the filter also can be used on plain TCP and UDP sockets to filter out unwanted packets—of course, this use of the filter is much less common.
In the following, I sometimes refer either to a socket or to a sock structure. As far as this article is concerned, both forms indicate the same object, and the latter corresponds to the kernel's internal representation of the former. Actually, the kernel holds both a socket structure and a sock structure, but the difference between the two is not relevant here.
Another data structure that will recur quite often is the sk_buff (short for socket buffer), which represents a packet inside the kernel. The structure is arranged in such a way that addition and removal of header and trailer information to the packet data can be done in a relatively inexpensive way: no data actually needs to be copied since everything is done by just shifting pointers.
Before going on, it may be useful to clear up possible ambiguities. Despite having a similar name, the Linux socket filter has a completely different purpose with respect to the Netfilter framework introduced into the kernel in early 2.3 versions. Even if Netfilter allows you to bring packets up to user space and feed them to your programs, the focus there is to handle network address translation (NAT), packet mangling, connection tracking, packet filtering for security purposes and so on. If you just need to sniff packets and filter them according to certain rules, the most straightforward tool is LSF.
Now we are going to follow the trip of a packet from its very ingress into the computer to its delivery to user land at the socket level. We first consider the general case of a plain (i.e., not PF_PACKET) socket. Our analysis at link layer level is based on Ethernet, since this is the most widespread and representative LAN technology. Cases of other link layer technologies do not present significant differences.
As we mentioned in the previous article, the Ethernet card is hard-wired with a particular link layer (or MAC) address and is always listening for packets on its interface. When it sees a packet whose MAC address matches either its own address or the link layer broadcast address (i.e., FF:FF:FF:FF:FF:FF for Ethernet) it starts reading it into memory.
Upon completion of packet reception, the network card generates an interrupt request. The interrupt service routine that handles the request is the card driver itself, which runs with interrupts disabled and typically performs the following operations:
Allocates a new sk_buff structure, defined in include/linux/skbuff.h, which represents the kernel's view of a packet.
Fetches packet data from the card buffer into the freshly allocated sk_buff, possibly using DMA.
Invokes netif_rx(), the generic network reception handler.
When netif_rx() returns, re-enables interrupts and terminates the service routine.
The netif_rx() function prepares the kernel for the next reception step; it puts the sk_buff into the incoming packets queue for the current CPU and marks the NET_RX softirq (softirq is explained below) for execution via the __cpu_raise_softirq() call. Two points are worth noticing at this stage. First, if the queue is full the packet is discarded and lost forever. Second, we have one queue for each CPU; together with the new deferred kernel processing model (softirqs instead of bottom halves), this allows for concurrent packet reception in SMP machines.
If you want to see a real-world Ethernet driver in action, you can refer to the simple NE 2000 card PCI driver, located in drivers/net/8390.c; the interrupt service routine called ei_interrupt(), calls ei_receive(), which in turn, performs the following procedure:
Allocates a new sk_buff structure via the dev_alloc_skb() call.
Reads the packet from the card buffer (ei_block_input() call) and sets skb->protocol accordingly.
Repeats the procedure for a maximum of ten consecutive packets.
A slightly more complex example is provided by the 3COM driver, located in 3c59x.c, which uses DMA to transfer the packet from the card memory to the sk_buff.
Let's take a closer look at the netif_rx() function. As mentioned before, this function has the task of receiving a packet from a network driver and queuing it for upper-layer processing. It acts as a single gathering point for all the packets collected by the different network card drivers, providing input to the upper protocols' processing.
Since this function runs in interrupt context (that is, its execution flow follows the interrupt service path) with other interrupts disabled, it has to be quick and short. It cannot perform lengthy checks or other complex tasks since the system is potentially losing packets while netif_rx() runs. So, what this function does is basically select the packet queue from an array called softnet_data, whose index is based on the CPU currently running. It then checks the status of the queue, identifying one of five possible congestion levels: NET_RX_SUCCESS (no congestion), NET_RX_CN_LOW, NET_RX_CN_MOD, NET_RX_CN_HIGH (low, moderate and high congestion, respectively) or NET_RX_DROP (packet dropped due to critical congestion).
Should the critical congestion level be reached, netif_rx() engages a throttling policy that allows the queue to go back to a noncongested status, avoiding service disruption due to kernel overload. Among other benefits, this helps avert possible DOS attacks.
Under normal conditions, the packet is finally queued (__skb_queue_tail()), and __cpu_raise_softirq(cpuid, NET_IF_SOFTIRQ) is called. The latter function has the effect of scheduling a softirq for execution.
The netif_rx() function terminates, returning a value indicating the current congestion level to the caller. At this point, interrupt context processing is done, and the packet is ready to be taken care of by upper-layer protocols. This processing is deferred to a later time, when interrupts will have been re-enabled and execution timing will not be as critical. The deferred execution mechanism has changed radically from kernel versions 2.2 (where it was based on bottom halves) to versions 2.4 (where it is based on softirqs).
Explaining in detail about bottom halves (BHs) and their evolution is out of the scope of this article. But, some points are worth recalling briefly.
First off, their design was based on the principle that the kernel should perform as few computations as possible while in interrupt context. Thus, when long operations were to be done in response to an interrupt, the corresponding driver would mark the appropriate BH for execution, without actually doing anything complex. Then, at a later time, the kernel would have checked the BH mask to determine whether some BHs were marked for execution and execute them before any application-level task.
BHs worked quite well, with one important drawback: due to their structure, their execution was serialized strictly among CPUs. That is, the same BH could not be executed by more than one CPU at the same time. This obviously prevented any kind of kernel parallelism on SMP machines and seriously affected performance. softirqs represent the 2.4-age evolution of BHs and, together with tasklets, belong to the family of kernel software interrupts, pieces of code that can be executed by the kernel when requested, without strict response-time guarantees.
The major difference with respect to BHs is that the same softirq may be run on more than one CPU at a time. Serialization, if required, now must be obtained explicitly by using kernel spinlocks.
softirq's processing core is performed in the do_softirq() routine, located in kernel/softirq.c. This function checks a bit mask, and if the bit corresponding to a given softirq is set, it calls the appropriate handling routine. In the case of NET_RX_SOFTIRQ, the one we are interested in at this time, the relevant function is net_rx_action(), located in net/core/dev.c. The do_softirq() function may get called from three distinct places inside the kernel: do_IRQ(), in kernel/irq.c, which is the generic interrupt handler; system calls' exit point, in kernel/entry.S; and schedule(), in kernel/sched.c, which is the main process scheduling function.
In other words, execution of a softirq may happen either when a hardware interrupt has been processed, when an application-level process invokes a system call or when a new process is scheduled for execution. This way, softirqs are drained frequently enough that none of them will lie waiting for their turn for too long.
The trigger mechanism also was exactly the same for the old-style bottom halves.
We've seen the packet come in through the network interface and get queued for later processing. Then, we've considered how this processing is resumed by a call to the net_rx_action() function. It's now time to see what this function does. Basically, its operation is pretty simple: it just dequeues the first packet (sk_buff) from the current CPU's queue and runs through the two lists of packet handlers, calling the relevant processing functions.
Some more words are worth spending on those lists and how they are built. The two lists are called ptype_all and ptype_base and contain, respectively, protocol handlers for generic packets and for specific packet types. Protocol handlers register themselves, either at kernel startup time or when a particular socket type is created, declaring which protocol type they can handle; the involved function is dev_add_pack() in net/core/dev.c, which adds a packet type structure (see include/linux/netdevice.h) containing a pointer to the function that will be called when a packet of that type is received. Upon registration, each handler's structure is either put in the ptype_all list (for the ETH_P_ALL type) or hashed into the ptype_base list (for other ETH_P_* types).
So, what the NET_RX softirq does is call in sequence each protocol handler function registered to handle the packet's protocol type. Generic handlers (that is, ptype_all protocols) are called first, regardless of the packet's protocol; specific handlers follow. As we will see, the PF_PACKET protocol is registered in one of the two lists, depending on the socket type chosen by the application. On the other hand, the normal IP handler is registered in the second list, hashed with the key ETH_P_IP.
If the queue contains more than one packet, net_rx_action() loops on the packets until either a maximum number of packets has been processed (netdev_max_backlog) or too much time has been spent here (the time limit is 1 jiffy, i.e., 10ms on most kernels). If net_rx_action() breaks the loop leaving a non-empty queue, the NET_RX_SOFTIRQ is enabled again to allow for the processing to be resumed at a later time.
The IP protocol receive function, namely ip_rcv() (in net/ipv4/ip_input.c), is pointed to by the packet type structure registered within the kernel at startup time (ip_init(), in net/ipv4/ip_output.c). Obviously, the registered protocol type for IP is ETH_P_IP.
Thus, ip_rcv() gets called from within net_rx_action() during the processing of a softirq, whenever a packet with type ETH_P_IP is dequeued. This function performs all the initial checks on the IP packet, which mainly involve verifying its integrity (IP checksum, IP header fields and minimum significant packet length). If the packet looks correct, ip_rcv_finish() is called. As a side note, the call to this function passes through the Netfilter prerouting control point, which is practically implemented with the NF_HOOK macro.
ip_rcv_finish(), still in ip_input.c, mainly deals with the routing functionality of IP. It checks whether the packet should be forwarded to another machine or if it is destined to the local host. In the former case, routing is performed, and the packet is sent out via the appropriate interface; otherwise, local delivery is performed. All the magic is realized by the ip_route_input() function, called at the very beginning of ip_rcv_finish(), which determines the next processing step by setting the appropriate function pointer in skb->dst->input. In the case of locally bound packets, this pointer is the address of the ip_local_deliver() function. ip_rcv_finish() terminates with a call to skb->dst->input().
At this point, packets definitely are traveling toward the upper-layer protocols. Control is passed to ip_local_deliver(); this function just deals with IP fragments' reassembly (in case the IP datagram is fragmented) and then goes over to the ip_local_deliver_finish() function. Just before calling it, another Netfilter hook (the ip-local-ip) is executed.
The latter is the last call involving IP-level processing; ip_local_deliver_finish() carries out the tasks still pending to complete the upper part of layer 3. IP header data is trimmed so that the packet is ready to be transferred to the layer 4 protocol. A check is done to assess whether the packet belongs to a raw IP socket, in which case the corresponding handler (raw_v4_input()) is called.
Raw IP is a protocol that allows applications to forge and receive their own IP packets directly, without incurring actual layer 4 processing. Its main use is for network tools that need to send particular packets to perform their tasks. Well-known examples of such tools are ping and traceroute, which use raw IP to build packets with specific header values. Another possible application of raw IP is, for example, realizing custom network protocols at the user level (such as RSVP, the resource reservation protocol). Raw IP may be considered a standard equivalent of the PF_PACKET protocol family, just shifted up one open systems interconnection (OSI) level.
Most commonly, though, packets will be headed toward a further kernel protocol handler. In order to determine which one it is, the Protocol field inside the IP header is examined. The method used by the kernel at this point is very similar to the one adopted by the net_rx_action() function; a hash is defined, called inet_protos, which contains all the registered post-IP protocol handlers. The hash key is, of course, derived from the IP header's protocol field. The inet_protos hash is filled in at kernel startup time by inet_init() (in net/ipv4/af_inet.c), which repeatedly calls inet_add_protocol() to register TCP, UDP, ICMP and IGMP handlers (the latter only if multicast is enabled). The complete protocol table is defined in net/ipv4/protocol.c.
For each protocol, a handler function is defined: tcp_v4_rcv(), udp_rcv(), icmp_rcv() and igmp_rcv() are the obvious names corresponding to the above-mentioned protocols. One of these functions is thus called to proceed with packet processing. The function's return value is used to determine whether an ICMP Destination Unreachable message has to be returned to the sender. This is the case when the upper-level protocols do not recognize the packet as belonging to an existing socket. As you will recall from the previous article, one of the issues in sniffing network data was to have a socket able to receive packets independent of their port/address values. Here (and in the just-mentioned *_rcv() functions) is the point where that limitation arises from.
At this point, the packet is more than halfway through its journey. Since space is limited in our beloved magazine, we will leave the packet in the capable hands of upper-layer 3 protocols until next month. What still remains to be explored is layer 4 processing (TCP and UDP), PF_PACKETs handling and, of course, the socket filter hooks and implementation. Be patient!