Recovering from failures without impact on data traffic in a shared bus architecture

Abstract

Methods of detecting and recovering from communication failures within an operating network switching device that is switching packets in a communication network, and associated structures. The communication failures addressed involve communications between the packet processors and a host CPU over a shared communications bus, e.g., PCI bus. The affected packet processor(s)—which may be all or a subset of the packet processors of the network switch—may be recovered without affecting hardware packet forwarding through the affected packet processors. This maximizes the up time of the network switching device. Other packet processor(s), if any, of the network switching device, which are not affected by the communication failure, may continue their normal packet forwarding, i.e., hardware forwarding that does not involve communications with the host CPU as well as forwarding or other operations that do involve communications with the host CPU.

Description

BACKGROUND OF THE INVENTION

The present invention relates generally to communications error recovery, and in particular to recovery from communication errors in networking equipment having a shared control bus architecture.

Network switching equipment typically comprises a set of physical ports, one or more packet processors that receive data packets from and send packets to a physical port (or to a plurality of physical ports) of the network switching device via PHY and MAC circuitry, and a controller (typically CPU-based). The packet processors may be ASICs or FPGAs. The physical ports and packet processors (each of which has its own input and output ports) may be arrayed over a plurality of linecards, as in a chassis type unit, or on a single circuit board, as in a stackable unit. A received data packet can be processed and forwarded by a packet processor to another physical port coupled to an output port of the same packet processor. Alternatively, the received data packet can be forwarded by a receiving packet processor through a switching fabric to another packet processor to be further processed and forwarded via one of its output ports to a physical port of the switching device

During normal operation, there is occasional data communication between the respective packet processors and the CPU-based controller. The CPU-based controller may be on a separate management circuit board, such as in a typical chassis type network switching device, or on the same circuit board, as in a typical stackable type network switching device. In some network switching devices, there may be a dedicated controller for each packet processor or for a plural subset of the packet processors of the switching device In a cost optimized system, on the other hand, there may be one CPU-based controller for the whole system, e.g., for the entire set of packet processors of the network switching device. This centralized architecture is advantageous from a system-cost point of view, since there is only one CPU-based controller for all of the packet processors, but it poses some challenges, since there is a shared control bus coupling the CPU-based controller to the packet processors. The complex intercoupling of data structures between the controller and the packet processor that form the basis of controller-packet processor communication is susceptible to malfunctions that can render the system unusable.

BRIEF SUMMARY OF THE INVENTION

In a data switching system, error recovery from data communication errors according to the present invention includes detecting occurrence of an error condition arising in a network switching device during data communication between a packet processor(s) and a controller in a shared bus architecture. In one embodiment, the controller, which is running a software program, detects the error. In response to detecting the error, data communication on the shared bus between the affected packet processor(s) and the controller is ceased pending a recovery step. Recovery of the data structures of the packet processor using data structures stored in the controller is performed, and communication between the affected packet processor(s) and the controller is resumed. Meanwhile, packet processing and forwarding by packet processor(s) (if any) are unaffected by the error continues, even to the extent such packet processing and forwarding requires communication with the controller. Further, packet processing and forwarding of packets by the packet processor(s) affected by the communication error also continues during the period of error detection and recovery, to the extent such processing and forwarding does not require communication between the affected packet processor(s) and the CPU controller, as would be the case, for instance, where the CPU controller had already programmed the memory(ies) accessed by the affected packet processor(s) with forwarding information necessary to process and forward particular received packets. This exemplary aspect allows for greater reliability of the network switch, because in prior shared bus systems, no packet forwarding was performed by the affected packet processor(s), or perhaps by the entire network switch, during the period of error detection and recovery, i.e., received packets were not forwarded, regardless of whether the memory(ies) had previously be programmed with forwarding information.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a generalized schematic diagram of an illustrative example of a shared bus architecture for packet processors and host CPU embodied according to the present invention.

FIG. 1A is a generalized schematic diagram of a line card including packet processors.

FIG. 2 is an illustration of the data structures used between the host CPU and each of the packet processors.

FIG. 2A is a illustration of a packet descriptor.

FIG. 3 shows the bits in a PCI-defined status register relevant to a communication error.

FIG. 4 illustrates the detection sequence for a communication error on the PCI bus.

FIG. 5 illustrates the flow for error recovery according to the present invention.

FIG. 6 illustrates the flow for error recovery while pre-programmed, a.k.a., hardware, forwarding continues.

DETAILED DESCRIPTION OF THE ILLUSTRATIVE EMBODIMENTS

FIG. 1 shows at high level a network switching device 100 (e.g., L2/L3 switch, wireless network switch, L4-7 switch, and the like) embodied according to the present invention. The switching device 100 can comprise one or more system management controllers, although the embodiment shown in FIG. 1 shows only one system management controller (CPU 102 or host CPU 102). CPU 102 controls switching device 100, including the one or more packet processors of switching device 100. Typically, the CPU 102 comprises a general data processor, which executes software stored in a memory of switching device 100.

Switching device 100 of FIG. 1 includes a plurality of packet processors 106, which may be arrayed on a single circuit board, or amongst a plurality of circuit boards. Sometimes such circuit boards are called linecards. For example, referring to FIG. 1A, three such packet processors 106a-106c are shown on a single line card 107 in an uppermost portion of FIG. 1A, but the number of packet processors may be more or less than three. The logic that constitutes a packet processor 106 is typically embodied in an ASIC (application-specific integrated circuit) or an FPGA (field-programmable gate array). An example packet processor may be one of the Prestera family of packet processors from Marvell Semiconductor, Inc. Such custom circuits are built for high speed performance and may have various functionality and speeds, such as 10/100, 1G, or 10G Ethernet capability. Each packet processor 106a-106c has one or more input ports (represented by a heavy arrow) and one or more output ports (represented by a heavy arrow). The host CPU 102 and the packet processors 106 are generally connected via a connector module 104 that includes a shared bus architecture to facilitate communication between the host CPU 102 and the packet processors 106a-106c.

One of the functions provided by the host CPU 102 is to program a memory 109 (e.g., a CAM and/or RAM, internal to the packet processor or external) with learned or user-inputted forwarding data so that, upon access of the memory by the packet processor 106 in response to a packet received via an in port of the switching device 100, the packet processor 106 will be able process the packet using the forwarding data to forward the received packet, typically at line rate, without the need for assistance from or communications with, the CPU 102. This type of pre-programmed forwarding by the packet processor 106 is sometimes called hardware forwarding. After processing in the packet processor 106, a received packet may be transmitted through the out port of the packet processor 106 to an out put port of the switching device 100, or to the in port of another of the packet processors 106, for further processing before the packet ultimately is sent to an output port of the switching device 100 (or dropped) Absent such pre-programmed forwarding data, a packet processor 106 may communicate with the host CPU 102 over the shared control bus, for assistance in forwarding the received packet. The CPU 102 may process the packet, and forward the packet to the same or another packet processor 106 for further processing and forwarding, e.g., VLAN flooding, or drop the packet, in accordance with the CPU's instruction set. Such communications (and other communications between the packet processors 106 and the host CPU 102) occur via the connection module 104 of FIG. 1. In this case, connection module 104 has a shared bus architecture and provides a communication channel coupling all of the packet processors 106a-c to CPU 102. In this example, the shared bus architecture is a PCI architecture, and includes a plurality of buses 114 and two bridges 124 that are each coupled to a pair of the packet processors 106. Typically, the PCI architecture is composed of a set of interconnected integrated circuits. It should be apparent that other shared bus architectures besides a PCI architecture can be adapted in accordance with the present invention. Further, practitioners will appreciate that there could be a variety of circuits, e.g., traffic manager circuits, and a switching fabric in switching device 100 to facilitate packet forwarding by and among the packet processors 106 and the CPU 102.

The present invention involves, among other things, detecting and resolving an error arising in the shared communication channel between the host CPU 102 and packet processor(s) 106. The error may affect the communications between the CPU 102 and all, or less than all (one or more), of the packet processors 106a-c. Such an error may be due to an unintended mechanical or electrical disturbance, e.g., a power spike, or the like, or may be due to an intended action, such as a hotswap action. In a hotswap, a circuit board (e.g., a line card including some packet processors 106, or a management card including CPU 102 (and perhaps some packet processors 106 as well), is hot removed from the switching device 100, and hot replaced with a like circuit board without a powering down of the switching device 100. Hotswap capability can improve the up time of the switching device 100.

I. Communication over the Control Plane

In an architecture such as depicted in FIG. 1, data communications between the host CPU 102 and packet processors 106a-106c comprise transferring data over the PCI bus. A data structure is used to manage the communication between the host CPU 102 and the packet processors 106. The reference to “control plane” refers to the data communication channel between the host CPU 102 and a respective packet processor 106a-106c.

FIGS. 2 and 2A show an example of such a data structure. The example data structure, a packet descriptor ring 122, comprises packet descriptors 132 organized into a structure commonly referred to as a circular linked list. Each packet descriptor 132 comprises descriptive information portion 132a (FIG. 2A) and a link portion 132b. The link portion 132b contains an address of the next packet descriptor 132 in the packet descriptor ring 122. The link portion of the last packet descriptor 132 (identified in FIG. 2 by “#n”) contains an address of the first packet descriptor 132 (identified in FIG. 2 by “#1”). One such packet descriptor ring 122 is associated with each packet processor 106x. In the particular implementation described herein, the packet descriptor ring 122 is provided in local memory 112 of the host CPU 102 (the figure also shows a processing unit 114 of the host CPU, for example, a conventional data processing unit or a microcontroller). It will be understood that a suitable program code 136 can be provided in local memory 112 to operate the CPU 114.

The descriptive information portion 132a of a packet descriptor 132 comprises data defined by the PCI standard and vendor-specific data. For example in one implementation, the packet descriptor 132 is a sixteen-byte data structure which includes a buffer address 232 (FIG. 2A) portion that stores an address (pointer) to a location in an area of the memory 112 called the packet buffer 134. There is an area in packet buffer 134 that is pointed to by each packet descriptor 132 to store data associated with that packet descriptor 132. The buffer address 232 stores the beginning address of the portion of the packet buffer 134 associated with the packet descriptor. The descriptive information portion 132a may include an additional data field 234. The data contained in the descriptive information portion 132a can vary from one kind of shared bus architecture (e.g. PCI bus) to another. In fact, the data contained in the descriptive information portion 132a may vary even among different implementations of the shared bus or vendors of the shared bus (e.g., PCI bus). For example, a vendor of the chip set that implements a PCI bus architecture may include vendor-specific information in addition to the information required by the PCI standard. The descriptive information portion 132a of each packet descriptor is intended to contain that additional data.

FIG. 2 shows that each packet processor 106x (e.g., the packet processors 106a-c of FIG. 1) includes, among other circuits, a pair of registers: a next descriptor register 116a and a next buffer register 116b, collectively referred to as the packet processor registers 116. As indicated in FIG. 2, each packet processor 106x can access at least a portion of the local memory 112 of the host CPU 102. For example, in the disclosed illustrative embodiment, each packet processor 106x uses a direct memory access (DMA) technique to access its respective packet descriptor ring 122 in the local memory 112, and to access the packet buffers 134 pointed by the constituent packet descriptors 132 of the data structure.

Upon start up of switching device 100, the processing unit 114 of the host CPU 102 configures its memory 112 with packet descriptors 132 and a packet buffer 134. The packet buffer 134 can be initialized with zeroes or some other suitable initial value. Likewise, the additional data portion 234 of each packet descriptor 132 can be filled with a suitable initial value (e.g., zero). This initialization process can be accomplished in software (e.g., in program code 136), or by simply loading a portion of the memory 112 with an image of a predefined configuration of packet descriptors and a packet buffer.

Initializing the packet descriptors 132 includes, for each packet descriptor, at least storing in the buffer address 232 an address of the beginning of a portion of storage in the packet buffer 134. The packet descriptors 132 are linked to create a packet descriptor ring 122 for each packet processor 106x. This includes linking a group of the packet descriptors 132 to each other in a linked list. For example, the first N packet descriptors 132 can be linked in a ring structure to define a packet descriptor ring 122 for a first packet processor 106x. The next N packet descriptors 132 can be linked in another ring structure to define the next packet descriptor ring 122 for the second packet processor 106x, and so on. In this way, an initialized packet descriptor ring 122 for each packet processor 106x is created and initialized.

At system startup, the processing unit 114 also initializes the packet processor registers 116 of each packet processor 106x. This includes loading the next descriptor address register 116a with the address of the first packet descriptor 132 in the packet descriptor ring 122 corresponding to the packet processor 106x. The next buffer address register 116b is loaded with the buffer address contained in the first packet descriptor. As will be explained below, the packet processor 106x uses this address information to transfer data into its corresponding packet descriptor through a DMA operation.

Upon successful completion of the transfer of the control data, status fields in the packet descriptors 132 that were used for that transfer (note that more than one packet descriptor 132 may be consumed) are updated to indicate this fact. A protocol, described below, between the host CPU 102 and the packet processors 106 synchronizes the usage of the packet descriptors 132 (and packet buffers) between them.

An example of a synchronization mechanism between the host CPU 102 and the packet processors 106x will now be described. Each packet descriptor 132 contains an ownership bit field that is used for this purpose. The initiator—which could be either the host CPU 102 or packet processor 106x—only uses the packet descriptor 132 which is owned by that packet processor.

Consider a data transfer from a packet processor 106x to the host CPU 102. The packet processor 106x uses the packet descriptor 132 that is currently owned by it, identified by the address contained in the next descriptor address register 116a. The packet processor 106x transfers control data to the packet buffer 134 in the CPU memory 112 through a DMA operation, identified by the address contained in the next buffer address register 116b. After the packet processor 106x completes the data transfer into CPU memory, the ownership bit field in the packet descriptor 132 is changed to indicate that the packet processor relinquishes the ownership of that descriptor and now it is owned by the host CPU 102. The host CPU 102, when it sees that it now owns the packet descriptor 132, becomes aware of the availability of control data in the packet buffer and processes the received control data. When the processing is finished, the host CPU 102 changes the ownership bit of the packet descriptor 132 back to that of the packet processor 106x. This indicates that the descriptor is now available for the packet processor 106x for further data transfer. The same is applicable when host CPU 102 initiates a data transfer to the packet processor 106x.

I. Control Plane Failure

A communication error in a shared bus architecture communication channel between a packet processor 106 and the host CPU 102 may occur, for example, during a read or write operation to the packet descriptor ring 122 that is shared between the host CPU 102 and the affected packet processor(s) 106. Such an error could have effects such as: (a) the packet processor receives an invalid packet descriptor address or an invalid buffer address; or (b) the packet descriptor ring 122 in CPU memory 112 becomes corrupted. In both of the above scenarios, the communication between the host CPU 102 and the affected packet processor 106—which may be any subset or all of packet processors 106 of FIG. 1—over the control plane 104 will be broken down, leading (in prior art systems) to an inoperative system. Such may require the prior art switching device to be reset, resulting in no packet forwarding during the reset period of time, an undesirable situation.

The sections which follow disclose exemplary structures and methods to detect exemplary errors, and an exemplary recovery procedure to re-establish control plane communication between the host CPU 102 and the affected ones of packet processors 106. Such structures and methods allow the switching device 100 to continue switching packets through the network during the detection and recovery processes. For instance, packet processor(s) 106 unaffected by the communications error may continue processing forwarding all packet traffic and communicating with the host CPU 102 as necessary over the shared communication channel, while affected packet processor(s) 106 may continue forwarding any packet traffic that does not require communications with CPU 102 (e.g., where forwarding data for the particular received packets already is programmed in the memory 109 accessed by the packet processor 106 during packet processing and forwarding.) This technique maximizes the up time and throughput of the network switching device 100 in the case of such an error.

For the sake of example, we will discuss two error situations that could occur in a shared-bus network switching device 100 operating in a network to which it is coupled when a communication failure occurs in the control plane (e.g., due to hot-swapping or some electrical glitch), namely (1) loss of address/data integrity over the PCI bus, resulting in a parity error; and (2) PCI timeout or failure, yielding incorrect data to the master. Following is an explanation on how the above two exemplary scenarios are detected by the host CPU 102 in accordance with its program code, in accordance with an illustrative embodiment of the present invention.

(1) Loss of Address/Data Integrity over the PCI Bus, Resulting in Parity Error

An electrical disturbance or other anomalous condition involving the PCI bus can cause the data that is present on the bus at that time to become corrupted, causing the loss of integrity. The PCI standard specifies mechanisms to detect this type of error condition using, e.g., parity information generated during both the address and data phases of a transaction. A detailed description of the parity generation mechanism on the PCI bus can be found in the PCI standard specification. What is relevant for the discussion here is that, when a parity error is detected for the data on the PCI bus, the devices involved in the erroneous transaction (host CPU 102, packet processor(s) 106) set some error bits in the PCI command/status register. This mechanism is defined per the PCI standard as well. Thus, any PCI device will have this mechanism implemented. For clarity of further discussion, some bit-fields of the 32 bit PCI configuration command/status register (offset 0x4) relevant to error detection are shown in FIG. 3.

(2) PCI Timeout or Failure

A timeout or other failure during a PCI transaction may occur, for instance, due to hotswapping a linecard while a packet processor(s) 106 of the linecard is communicating with the host CPU 102. For instance, such an electrical disturbance occurring during communication between a packet processor 106 and the host CPU 102 can result in a failure yielding incorrect data to the device (packet processor 106 or the host CPU 102) that originated the operation. In the context of a switching system 100, this could result in the affected packet processor(s) 106 getting an invalid packet descriptor address or an invalid packet buffer address if the affected packet processor(s) was in the middle of a read operation during the hotswap operation. An invalid address obtained like this will cause the packet processor 106 to be unable to properly transfer data to the CPU. For example, an invalid packet descriptor address may cause the packet processor 106 to access an incorrect packet descriptor 132 within its associated packet descriptor ring 122, or worse yet may cause the packet processor 106 to access a packet descriptor 132 belonging to another packet processor. Similarly, an incorrect packet buffer address will cause the wrong area in the packet buffer 134 to be accessed, resulting in writing of data to an area in the packet buffer that belongs to another packet processor 106.

III. Control Plane Failure Detection by the Host CPU

FIG. 5 illustrates a detection and recovery process within the present invention, and for this example is explained in terms of the disclosed illustrative PCI architecture embodiment. The processing is performed in the host CPU 102 by way of program code that is executed by the processing unit 143. Typically, the program code is stored in the CPU memory 112, but can be stored in a ROM (read-only memory) or a flash device, and so on. It will be apparent from the discussion that follows that the present invention can be practiced, for instance, using any shared bus architecture.

Referring to FIG. 5, the process begins with detection of a communication error between the host CPU 102 and a packet processor 106, step 502. In an embodiment of the present invention, the underlying bus architecture for communication between the host CPU 102 and the packet processors 106 is the PCI shared bus.

Referring to FIG. 4, a data phase parity error will be described as an example of a communication error. This type of error can arise during a PCI write (or read) transaction. FIG. 4 shows segments of connector module 104 (FIG. 1) involved in the transaction, which include the CPU-to-PCI bridge segment and PCI bridge-to-packet processor segment. If an error occurs during a transaction, which could be in either direction, the PCI error status is propagated to both ends. The host CPU 102, the PCI bridge 124, and the packet processor 106a all receive the PCI error status. Thus, the host CPU 102 detects the PCI error for both communication scenarios: (1) where the host CPU 102 sends data to packet processor 106 and (2) where the packet processor 106 sends data to the host CPU 102.

Upon detection of a data phase parity error, the PCI device (i.e., host CPU 102 or packet processor 106) that checked the parity is responsible for asserting the Detected Parity Error bit (bit31) (FIG. 3) in its PCI configuration status register. It also asserts the PERR# signal if the Parity Error Response bit in its PCI Configuration command register (bit6) is set to one. A parity error on the address phase of a transaction causes the PCI device to set the SERR# signal, if the SERR# enable bit is set to one in the PCI configuration command register. In this example, the packet processors 106a-106c and the host CPU 102 are the PCI devices.

The PERR# signal or the SERR# signal, when asserted, could generate an interrupt to the host CPU 102. In order to get these signals, the host CPU 102 interrupts must be enabled. This is done in the software which configures the host CPU's PIC (Programmable Interrupt Control) registers appropriately during the hardware initialization (startup) phase.

Upon detection of the occurrence of an error, an attempt is initiated to identify the devices that are involved with the error, step 504. In the disclosed example embodiment, packet processors 106a-106c communicate with the host CPU 102 over the PCI shared bus. Thus, the host CPU 102 will always be one of the devices involved in a detected communication error. The following description of step 504, in this particular embodiment, then amounts to a discussion about identifying the packet processor(s) 106a-106c that experienced the communication error.

A communication error can manifest itself in the PCI shared bus by raising an interrupt signal, as described above. When the host CPU 102 is interrupted by the interrupt signal due to a PCI error detection, a corresponding interrupt service routine is executed. The interrupt running in the host CPU polls each of the packet processors 106a-106c (which may be on one or more linecards) residing on the PCI bus on which the interrupt originated. In an embodiment, the host CPU 102 reads the PCI configuration status register (FIG. 3) of each packet processor 106 and inspects the “Detected Parity Error” bit (bit 31). A packet processor 106 that detected the PCI parity error will have its “Detected Parity Error” bit set to one. The host CPU 102 in this way identifies the packet processor 106 that experienced the communication error. As mentioned, there may be a situation where more than one packet processor experiences a communication error with the host CPU. Thus, in an embodiment of the present invention, the host CPU 102 polls each and every one of the plural packet processors 106 to identify all packet processor(s) 106 experiencing a communication error.

The discussion above deals with detecting a communication error between a packet processor 106 and the host CPU where the packet processor 106 is sending data to the host CPU 102. However, the error handling herein is applicable for data transfer in either direction between host CPU 102 and the packet processor. When host CPU 102 detects a PCI error that results when data is transferred over the shared control plane from the packet processor 106 to the host CPU 102 or vice-versa, the recovery procedure is invoked to rectify the error. As will be explained below, the present invention allows for error recovery while at the same time leaving unaffected pre-programmed hardware forwarding of received packets by the affected (and unaffected) packet processor(s) 106 to continue. Further, packet processors 106 not affected by the error may continue control plane communications with CPU 102 as necessary to process and forward received packets.

As another example of steps 502 and 504, a PCI timeout or failure may be detected as follows. The read and write operation from host CPU 102 to the packet processor 106 is implemented using known DMA (direct memory access) techniques. Underneath the DMA, it is a PCI transaction as the host CPU 102 and packet processor 106 are connected through a PCI interface. The actions involved in such an operation from the CPU-side are:

1. CPU configures the DMA registers;

2. CPU initiates a timer;

3. CPU initiates the DMA operation;

4. CPU checks the DMA status for error/success/busy;

5. CPU repeats the above “check” step until the returned status is success, or error, or until the timer expires.

In the case of successful operations, the host CPU 102 detects the DMA completion by the channel being not busy and the DMA status does not indicate an error. In the case of a PCI failure, the host CPU 102 detects the DMA engine returning the error back. In the case of a PCI timeout, the DMA channel never returns and the host CPU 102 initiated timer expires, thus indicating the PCI timeout discussed above.

In a case for example of a failure or timeout during a PCI read transaction (where the packet processor 106 attempts to read data from the host CPU), the operation typically returns with 0xFFFFFFFF to the packet processor. The host CPU 102 can detect this error condition by reading the relevant hardware register in the packet processor. The relevant register can vary, depending on the shared bus architecture used to connect the packet processor and CPU. If they are connected through PCI, this register would be, e.g., the PCI status register. In a more generic context, the packet processor vendor could implement a specific register to reflect the status of a transaction it made to read from CPU memory.

Continuing, the host CPU 102 polls the relevant register of each packet processor 106 present in the switching device 100 in order to identify the one or more packet processors 106 that encountered a communication error over the shared communication bus. If the error condition is detected, then the host CPU 102 invokes the recovery method discussed below for each of the one or more affected packet processors 106 in order to recover. In one embodiment, this process of polling the relevant registers of the packet processors can be selectively performed. For instance, it may be programmed to occur only when a hotswap is detected (e.g., loss of heartbeat signal or loss of power signal or some other signaling method). Alternatively, the polling can be performed periodically.

Again, the figures and discussion herein, while sometimes directed to a read operation, for example, are applicable in the other direction as well, because the PCI status gets propagated to both ends—the CPU 102, the PCI bridge, and the packet processor 106 all get that information. Thus, the CPU 102 detects the PCI error for both communication scenarios, where the CPU 102 sends data to the packet processor 106, and where the packet processor 106 sends data to the CPU 102.

Note that, while the above examples have the CPU 102 detect the control plane communication error by the generation of an interrupt to the CPU 102, or by periodic or event-driven (e.g., hotswap) polling of registers, other methods of detecting the error condition, and identifying the affected packet processor(s) 106 may be utilized. This may vary, for instance, by the type of shared bus architecture used in the network switching device 100. As mentioned, the example of a PCI bus is exemplary only.

Continuing with FIG. 5, the discussion will now be directed to recovery steps performed in response to detecting the error. As discussed above, the consequences of communication errors in the control plane 104 resulting from a PCI failure during a read/write transaction involving by a packet processor 106 may include:

• • 1. the affected packet processor gets an invalid packet descriptor address or invalid packet buffer address
  • 2. a packet descriptor ring (whether the one associated with the packet processor, or one associated with another packet processor) in CPU memory gets corrupted
  • 3. the packet buffer(s) get corrupted

When a communication error is detected and the packet processor(s) 106 that faced the error is identified by the CPU 102 via step 504, the host CPU 102 then performs the following actions according to its program code to recover from it:

• • step 506—Communication between the identified packet processor 106 and the host CPU 102 is disabled. In an embodiment, a specific hardware configuration register in the packet processor is modified for this purpose. The modification could be, for example, to set or reset a certain bit or bits in the packet processor register which would correspond to enabling or disabling the transfer of packets to or from the CPU memory. This would be a register which is dependent on the specific manufacturer of the packet processor. It is noted that data communication between the host CPU 102 and other functional unaffected packet processors 106 will not be ceased, and will continue. It is further noted that packet forwarding taking place among the packet processors 106 likewise continues to the extent it is pre-programmed or involves unaffected portions of the control plane. Thus, although communication between the host CPU and the identified affected packet processor(s) 106 is halted per this step, data communication continues in those parts of the switching device 100 that have not experienced a communication error.
  • step 508—The data structures in the CPU memory 112 allocated to the identified affected packet processor 106 are recovered and reinitialized. Thus, in an embodiment of the present invention, only the packet descriptor ring 122 associated with the identified packet processor 106 is recreated by resetting the link portions 132b of its constituent packet descriptors 132 as described during the startup process. The additional data portions 132a are likewise reset. The portions of the packet buffer 134 associated with the packet descriptors 132 are reset; e.g., zero-filled. The buffer address 232 of each packet descriptor is reset to the starting address of its corresponding portion of the packet buffer 134. The data structures used for packet communication that were allocated for the identified affected packet processor 106 are thus recovered, although any data that was in transit between the host CPU 102 and the affected packet processor 106 will be lost. Consequently, repeated communication failures experienced by a packet processor 106 will not result in the accumulation of lost resources (e.g., memory allocated to data structures). This aspect of the present invention prevents the possibility of running out of resources (e.g., allocated memory) if a packet processor 106 experiences repeated communication failures.
  • step 510—The hardware registers in the identified packet processor 106 are updated by the CPU 102 with the address of the first packet descriptor in a re-initialized packet descriptor ring and its buffer address to be used in a next packet transfer.
  • Step 512—The data transfer between the now-restored packet processor 106 and the host CPU 102 is re-enabled.

Thus, in accordance with the present invention, a communication error in a switching device 100 that occurs between a packet processor (s) 106 and the host CPU 102 can be recovered without having to power down the entire switching device 100. This is advantageous because packet switching can continue during the detection and recovery process. Affected packet processor(s) 106 can process and forward packet traffic to the extent necessary forwarding data already is programmed in the memory accessed by the packet processor (or other information or programming of the packet processor allows it to process the packet without data communication with the CPU 102). This maximizes the up time of the switching device 100 during such an error condition.

In the case where more than one packet processor 106 experiences a communication failure with the host CPU 102, the host CPU 102 can identify (more? of FIG. 5) in step 514 each additional failed packet processor 106 and repeat the foregoing recovery steps 506-512. It is noted that the host CPU 102 can detect communication error in packet processors 106 residing among different linecards 107 in a chassis type system. Thus, for example, in step 504 the host CPU 102 might identify an affected packet processor 106 on a first linecard 107 and, in step 514 might identify an affected packet processor 106 on a second linecard. The recovery step 506-512 can be performed for the failed packet processor 106 on the first linecard, and then repeated for the failed packet processor 106 on the second linecard.

In an alternative embodiment, step 514 may be performed after step 506, and then after all affected packet processors 106 are identified, then steps 508-512 may be performed for each of the affected packet processors 106.

The switching device of the present invention recovers from certain communication errors by reestablishing the communication between the host CPU and the affected packet processor in manner that maintains normal packet switching functionality to the greatest possible extent. Since the procedure disrupts only the data transfer between the host CPU 102 and the affected packet processor 106, the data transfer that occurs between host CPU 102 and the other unaffected packet processor(s) 106 and the data transfer that is switched by the packet processors 106 themselves without CPU intervention is not impacted.

FIG. 6. illustrates a process 600 performed in a network switch 100 in accordance with the present invention. Thus, in a step 602, a packet processor 106 receives data at one of its ports from a physical port of a network switch (via a PHY and MAC interface, for example). In the course of processing the packet, a determination is made (step 603) by the receiving packet processor 106 as to whether the packet processor 106 can hardware forward the received packet based on, e.g., pre-programmed forwarding data in the memory 109 accessed by the packet processor. If yes, then the packet processor 106 will forward the received data packet accordingly, in a step 610; i.e., the host CPU 102 is not involved. The flow of steps 602 and 610, since there is no interaction over the shared communication bus with the host CPU 102, can occur at wire speed because the received data packet is hardware forwarded. By contrast, if the destination is not known, then the packet processor 106 will communicate in step 604 with the host CPU 102 over the control plane in order that, e.g., the host CPU 102 might learn from, drop, or assist in the forwarding of, the received packet. In step 608, in the no error condition, the packet is processed by the CPU 102, and further processed and forwarded by the same or a different packet processor 106 accordance with the processing by the CPU. If, during or after this control plane communication, an error is detected by the host CPU 102 in a step 605, then the flow proceeds via step 606 to the processing shown in FIG. 5. During the detection and recovery process, steps 602 and 610 may continue to be performed by the packet processor(s) that are undergoing error recovery via step 606. Other packet processors 106 unaffected by the control plane communications error may continue with steps 601 and 610 using pre-programmed forwarding data, or with steps 604 and 608 where data communication with the host CPU 102 over the shared control plane is necessary to process and forward a particular packet.

The above embodiments are exemplary only, and may be modified without departing from the invention in is broader aspects.

Claims

1. A method in a network device, the method comprising: storing, in a memory associated with a host processor of the network device, a set of data structures used for transferring data, on a shared bus, between the host processor and a plurality of packet processors of the network device, each packet processor in the plurality of packet processors configured to forward, from the network device, one or more packets received by the network device;detecting, by the host processor, an error condition indicative of a communication error between the host processor and a first packet processor from the plurality of packet processors;in response to detection of the error condition: identifying, by the host processor, from the plurality of packet processors, the first packet processor affected by the error condition; andperforming, by the host processor, a set of recovery actions for recovering from the error condition, the set of recovery actions including disabling communication between the host processor and the first packet processor; andwhile the set of recovery actions is being performed, communicating data on the shared bus, between the host processor and at least one packet processor from the plurality of packet processors other than the first packet processor and forwarding, by the first packet processor at least one packet received by the network device using forwarding information programmed prior to the host processor detecting the error condition.

2. The method of claim 1 wherein the shared bus is a shared PCI (peripheral component interconnect) bus.

3. The method of claim 1 wherein the detecting comprises detecting setting of one or more error bits due to a parity error detected on the shared bus.

4. The method of claim 1 wherein the detecting comprises generating an interrupt to the host processor.

5. The method of claim 1 wherein the detecting comprises polling a register.

6. The method of claim 5 wherein the polling is programmed to occur periodically.

7. The method of claim 5 wherein the polling is programmed to occur depending on occurrence of a hotswap.

8. The method of claim 1 wherein the error condition is a timeout.

9. The method of claim 1 wherein the forwarding, by the first packet processor, of the at least one packet includes forwarding the at least one packet to a second packet processor in the plurality of packet processors or forwarding the at least one packet to an output port of the network device.

10. The method of claim 1 wherein the set of recovery actions further comprises restoring, by the host processor, to an initial state a data structure from the set of data structures used for communication, on the shared bus, between the host processor and the first packet processor.

11. The method of claim 10 wherein restoring the data structure includes resetting data values of the data structure with initial data values.

12. The method of claim 1 wherein the set of recovery actions further comprises re-enabling communication between the host processor and the first packet processor.

13. A network device comprising: a control processor with an associated memory; anda plurality of packet processors, each packet processor in the plurality of packet processors configured to forward, from the network device, one or more packets received by the network device;the control processor configured to: store, in the associated memory, a set of data structures used for transferring data, on a shared bus, between the control processor and the plurality of packet processors;in response to detection of an error condition relating to communication between the control processor and a first packet processor from the plurality of packet processors, identify from the plurality of packet processors, the first packet processor affected by the error condition and perform a set of recovery actions for recovering from the error condition, the set of recovery actions including disabling communication with the first packet processor;the first packet processor from the plurality of packet processors configured to: while the set of recovery actions is being performed by the control processor, forward at least one packet received by the network device using forwarding information programmed prior to the host processor detecting the error condition; anda second packet processor from the plurality of packet processors configured to: while the set of recovery actions is being performed by the control processor, continue to communicate on the shared bus with the control processor.

14. The device of claim 13 wherein the shared bus is a shared PCI (peripheral component interconnect) bus.

15. The device of claim 13 wherein the control processor is further configured to detect the error condition by detecting setting of one or more error bits due to a parity error detected on the shared bus.

16. The device of claim 13 wherein the control processor is further configured to detect the error condition by polling a register.

17. The device of claim 16 wherein the polling is programmed to occur periodically.

18. The device of claim 16 wherein the polling is programmed to occur depending on occurrence of a hotswap.

19. The device of claim 13 wherein the error condition is a timeout.

20. The device of claim 13 wherein the first packet processor is configured to forward the at least one data packet by forwarding the at least one data packet to another packet processor in the plurality of packet processors or forwarding the at least one data packet to an output port of the network device.

21. The device of claim 13 wherein: the first packet processor includes at least one data register; andthe control processor is further configured to store initialization data to the at least one data register in performing the set of recovery actions.

22. The device of claim 13 wherein the control processor is further configured to re-enable communication with the first packet processor in performing the set of recovery actions.