1. Field of the Invention
The present invention relates to communication systems, in particular, to an accelerated processor architecture for network communications.
2. Description of the Related Art
Network processors are generally used for analyzing and processing packet data for routing and switching packets in a variety of applications, such as network surveillance, video transmission, protocol conversion, voice processing, and internet traffic routing. Early types of network processors were based on software-based approaches with general-purpose processors, either singly or in a multi-core implementation, but such software-based approaches are slow. Further, increasing the number of general-purpose processors had diminishing performance improvements, or might actually slow down overall network processor throughput. Newer designs add hardware accelerators to offload certain tasks from the general-purpose processors, such as encryption/decryption, packet data inspections, and the like. These newer network processor designs are traditionally implemented with either i) a non-pipelined architecture or ii) a fixed pipeline architecture.
In a typical non-pipelined architecture, general-purpose processors are responsible for each action taken by acceleration functions. A non-pipelined architecture provides great flexibility in that the general-purpose processors can make decisions on a dynamic, packet-by-packet basis, thus providing data packets only to the accelerators or other processors that are required to process each packet. However, significant software overhead is involved in those cases where multiple accelerator actions might occur in sequence.
In a typical fixed-pipeline architecture, packet data flows through the general-purpose processors and/or accelerators in a fixed sequence regardless of whether a particular processor or accelerator is required to process a given packet. This fixed sequence might add significant overhead to packet processing and has limited flexibility to handle new protocols, limiting the advantage provided by the using accelerators.
Read latency and overall read throughput to storage devices with sequential access penalties, particularly memories external to a system on chip (SoC), can be performance bottlenecks for the SoC. For example, an external memory might include two or more substructures (e.g., multiple banks of DRAM). In such a system, a latency penalty might be incurred for sequential read requests to the same memory substructure. Several mechanisms have been developed for addressing this bottleneck. One mechanism queues read operations or requests (“read requests”) destined for each individual memory substructure and then selects read requests for non-busy substructures from one or more queues. Queuing works well when these read requests are spread evenly among the memory substructures, but fails if all the read requests target a particular substructure. Another mechanism duplicates the entire data structure multiple times with a number of copies and then selects a non-busy substructure as the target of the read request. This mechanism works well and overcomes some of the shortcomings of the other mechanism, but the amount of data stored by the memory is reduced by i) the inverse of the number of copies regardless of whether or not all of the data benefited from the duplication, or ii) the memory required increases as a multiple of the number of copies required.
This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.
Described embodiments provide a packet classifier for a network processor that generates tasks corresponding to each received packet. The packet classifier includes a scheduler to generate contexts corresponding to tasks received by the packet classifier from a plurality of processing modules of the network processor. A multi-thread instruction engine processes threads of instructions, each thread of instructions corresponding to a context received from the scheduler. A thread status manager maintains a thread status table having N entries to track up to N active threads. Each status entry includes a valid status indicator, a sequence value, and a thread indicator. A sequence counter generates a sequence value for each thread and is incremented when processing of a thread is started, and is decremented when a thread is completed, by the multi-thread instruction engine. Instructions are processed by the multi-thread instruction engine in the order in which the threads were started.
Other aspects, features, and advantages of the present invention will become more fully apparent from the following detailed description, the appended claims, and the accompanying drawings in which like reference numerals identify similar or identical elements.
Described embodiments provide a packet classifier for a network processor that generates tasks corresponding to each received packet. The packet classifier includes a scheduler to generate contexts corresponding to tasks received by the packet classifier from a plurality of processing modules of the network processor. A multi-thread instruction engine processes threads of instructions, each thread of instructions corresponding to a context received from the scheduler. A thread status manager maintains a thread status table having N entries to track up to N active threads. Each status entry includes a valid status indicator, a sequence value, and a thread indicator. A sequence counter generates a sequence value for each thread and is incremented when processing of a thread is started, and is decremented when a thread is completed, by the multi-thread instruction engine. Instructions are processed by the multi-thread instruction engine in the order in which the threads were started.
Table 1 defines a list of acronyms employed throughout this specification as an aid to understanding the described embodiments of the present invention:
I/O interface 104 might typically be implemented as hardware that connects network processor 100 to one or more external devices through I/O communication link 102. I/O communication link 102 might generally be employed for communication with one or more external devices, such as a computer system or a networking device, that interface with network processor 100. I/O communication link 102 might be a custom-designed communication link, or might conform to a standard communication protocol such as, for example, a Small Computer System Interface (“SCSI”) protocol bus, a Serial Attached SCSI (“SAS”) protocol bus, a Serial Advanced Technology Attachment (“SATA”) protocol bus, a Universal Serial Bus (“USB”), an Ethernet link, an IEEE 802.11 link, an IEEE 802.15 link, an IEEE 802.16 link, a Peripheral Component Interconnect Express (“PCI-E”) link, a Serial Rapid I/O (“SRIO”) link, or any other interface link. Received packets are preferably placed in a buffer in shared memory 112 by transfer between I/O interface 104 and shared memory 112 through switch 110.
In embodiments of the present invention, shared memory 112 is a conventional memory operating as a cache that might be allocated and/or subdivided. For example, shared memory 112 might include one or more FIFO queues that might be dynamically allocated to the various μP cores 106 and hardware accelerators 108. External memory interface 114 couples shared memory 112 to external memory 116 to provide off-chip storage of data not needed by the various μP cores 106 and hardware accelerators 108 to free space in shared memory 112. The μP cores and hardware accelerators might interact with each other as described in related U.S. patent application Ser. Nos. 12/782,379, 12/782,393, and 12/782,411, all filed May 18, 2010, for example, by one or more communication bus rings that pass “tasks” from a source core to a destination core. As described herein, tasks are instructions to the destination core to perform certain functions, and a task might contain address pointers to data stored in shared memory 112.
Network processor 100 might typically receive data packets from one or more source devices, perform processing operations for the received data packets, and transmit data packets out to one or more destination devices. As shown in
The various μP cores 106 and hardware accelerators 108 of network processor 100 might include several exemplary types of processors or accelerators. For example, the various μP cores 106 and hardware accelerators 108 might include, for example, a Modular Packet Processor (MPP), a Packet Assembly Block (PAB), a Modular Traffic Manager (MTM), a Memory Management Block (MMB), a Stream Editor (SED), a Security Protocol Processor (SPP), a Regular Expression (RegEx) engine, and other special-purpose modules.
The MTM is a software-driven accelerator that provides packet scheduling for up to six levels of scheduling hierarchy. The MTM might support millions of queues and schedulers (enabling per flow queuing if desired). The MTM might provide support for shaping and scheduling with smooth deficit weighed round robin (SDWRR) for every queue and scheduler. The MTM might also support multicasting. Each copy of a packet is scheduled independently and traverses down different virtual pipelines enabling multicast with independent encapsulations or any other processing. The MTM might also contain a special purpose processor that can be used for fine-grained control of scheduling decisions. The MTM might be used to make discard decisions as well as scheduling and shaping decisions.
The SED is a software-driven accelerator that allows for editing of packets. The SED performs packet editing functions that might include adding and modifying packet headers as well as fragmenting or segmenting data (e.g., IP fragmentation). The SED receives packet data as well as parameters from tasks and a task specified per-flow state. The output of the SED becomes the outgoing packet data and can also update task parameters.
The RegEx engine is a packet search engine for state-based cross-packet pattern matching. The RegEx engine is multi-threaded accelerator. An exemplary RegEx engine might be implemented such as described in U.S. Pat. No. 7,439,652 or U.S. Patent Application Publication No. 2008/0270342, both of which are incorporated by reference herein in their entireties.
The SPP provides encryption/decryption capabilities and is a command-driven hardware accelerator, preferably having the flexibility to handle protocol variability and changing standards with the ability to add security protocols with firmware upgrades. The ciphers and integrity (hash) functions might be implemented in hardware. The SPP has a multiple ordered task queue mechanism, discussed in more detail below, that is employed for load balancing across the threads.
The MMB allocates and frees memory resources in shared memory 112. Memory is allocated for such applications as task FIFO storage, packet data storage, hash-table collision handling, timer event management, and traffic manager queues. The MMB provides reference counts to each block of memory within shared memory 112. Multiple reference counts allow for more efficient storage of information, such as multicast traffic (data to be sent to multiple destinations) or for retransmission. Multiple reference counts remove the need for replicating the data each time the data is needed. The MMB preferably tracks the memory allocations using a stack-based approach since a memory block recently released is preferably the next block to be allocated for a particular task, reducing cache trashing and cache tracking overhead.
The PAB is a command driven hardware accelerator providing a holding buffer with packet assembly, transmit, retransmit, and delete capabilities. An incoming task to the PAB can specify to insert/extract data from anywhere in any assembly buffer. Gaps are supported in any buffer. Locations to insert and extract can be specified to the bit level. Exemplary traditional packet reassembly functions might be supported, such as IP defragmentation. The PAB might also support generalized holding buffer and sliding window protocol transmit/retransmit buffering, providing an offload for features like TCP origination, termination, and normalization.
The MPP is a multi-threaded special purpose processor that provides tree based longest prefix and access control list classification. The MPP also has a hardware hash-based classification capability with full hardware management of hash-table additions, deletions, and collision handling. Optionally associated with each hash entry is a timer that might be used under software control for tasks such as connection timeout and retransmission timing. The MPP contains a statistics and state management engine, which when combined with the hash table and timer facilities, provides support for state-based protocol processing. The MPP might support millions of flows, limited only by the amount of DRAM capacity assigned to the functions. The MPP architecture might be able to store all per thread states in memory instead of in register files.
As described herein, MPP 200 might generally be employed as a packet classification engine in network processor 100. In general, packet classification categorizes packets into classes, for example, based on port number or protocol. Each resulting packet class might be treated differently to control packet flow, for example, each packet class might be subject to a different rate limit or prioritized differently relative to other packet classes. Classification is achieved by various means. Matching bit patterns of data to those of known protocols is a simple, yet widely-used technique. More advanced traffic classification techniques rely on statistical analysis of attributes such as byte frequencies, packet sizes and packet inter-arrival times. Upon classifying a traffic flow using a particular protocol, a predetermined policy can be applied to it and other flows to either guarantee a certain quality (as with VoIP or media streaming service) or to provide best-effort delivery.
As shown in
Semaphore Engine (SEM) 222 implements semaphore logic in MPP 200, and might support up to 1024 logical semaphores, which might correspond to 4 physical semaphores, each corresponding to 256 logical semaphores. Semaphores are used to manage atomic access to a hardware resource of network processor 100 and MPP 200. For example, for a context thread to utilize an instance of a hardware resource, the context thread might have to reserve a semaphore for that resource. A context might be allowed to have up to 4 outstanding physical semaphores. Semaphores are allocated and released by SEM 222 based on function calls received by function bus 212. SEM 222 might support ordered and unordered semaphore calls.
Hash table operations might be carried out by Hash Engine (HE) 220. HE 220 implements hash engine functionality in MPP 200. HE 220 receives instructions from Function Bus Interface (FBI) 216 over function bus 212. HE 220 executes the function calls in the order in which it receives them on the function bus, for example by employing order queues. HE 220 might include order logic to store function calls for up to 64 contexts. Hash tables implemented by HE 220 are stored in system memory 112, via memory interface 224. Embodiments of HE 220 might implement up to 1024 independent hash tables. Each hash table might be allocated dedicated static memory at system startup of network processor 100, but might also be dynamically allocated additional memory over time as network processor 100 operates. In some embodiments, additional memory is allocated dynamically to a hash table in 256 B blocks.
State Engine (SENG) 218 might perform functions of a finite state machine (FSM) that operates on received packets. For example, SENG 218 might perform statistics counts and run traffic shaper scripts. SENG 218 might store statistics data in system memory 112, via memory interface 224, and might employ a data cache to reduce accesses to system memory 112 when there are multiple accesses to the same location of system memory.
MPP 200 might generally be implemented as a multi-threaded engine capable of executing parallel functions. The multi-threading operation is performed by multiple contexts in MTIE 214. Some embodiments of MPP 200 might employ more than one MTIE 214 to support additional context processing. For example, MPP 200 might preferably include 4 MTIE cores, each capable of processing 32 contexts, for a total of 128 contexts. These contexts might be supported by 256 task identifiers (TIDs), meaning that contexts for up to 256 tasks might be concurrently active in MPP 200.
MPP 200 might typically receive input tasks via a task ring such as described in U.S. patent application Ser. No. 12/782,379 filed May 18, 2010. Additionally, MPP 200 might receive a timer event via a timer ring. Receiving a task or receiving a timer event results in a context being generated in MPP 200 corresponding to the received task or timer event. Upon receiving a task, MPP 200 reads the task from system memory 112, for example via communication interface 202 and memory interface 224. Communication interface 202 issues a task start request to MTIE core 214 via scheduler (SCH) 204. A typical task might include 32 bytes of parameter data, and a typical timer event might include 13 bytes of parameter data.
SCH 204 tracks MPP contexts and maintains a list of free contexts. Upon receiving a task start request, if a free context is available, SCH 204 issues a context start indication to one or more other modules of MPP 200 such that the various modules, if necessary, might initialize themselves to process the context. SCH 204 also maintains task template to root address table 228. Root address table 228 specifies the instruction entry point (e.g., the address of first instruction in flow memory 230) for a given task template. Root address table 228 might typically be loaded on initial configuration of MPP 200.
Upon receiving the context start indication from SCH 204, MTIE 214 initializes its internal context memory and loads the task parameters of the received task. MTIE 214 also loads the root address to use for the context from root address table 228, such that MTIE 214 can determine what processing to perform for the received input task. Upon receiving the context start indication from SCH 204, Data Buffer Controller 206 initiates a data read operation to read the packet data corresponding to the context from at least one of system memory 112 and external memory 116. HE 220, FBI 216 and PIC 210 reset various valid bits for error detection for the context.
After the context start indication is issued, SCH 204 issues a context schedule indication to MTIE 214. In response to the context schedule indication, MTIE 214 starts executing a first command stored at the location specified in root address table 228. The command might be stored in at least one of root tree memory 232, flow memory 230, and external tree memory 234. While executing the specified commands, MTIE 214 fetches tree instructions from either root tree memory 232 or external tree memory 234. MTIE 214 also fetches flow instructions from flow memory 230. Some embodiments might include a 16 KB flow memory for each MTIE core of MPP 200, and some embodiments might further allow the flow memory for multiple MTIE cores to be shared to increase the size of the flow memory for all MTIE cores.
Upon reaching a point in context processing that requires processing by a module of MPP 200 external to MTIE 214, MTIE 214 sends the context along with the corresponding function call and arguments to FBI 216. Once the context is delivered to FBI 216, the context might become inactive in MTIE 214 as, in general, a given context might only be active in one module of MPP 200 at any one time. FBI 216 provides the function call to the designated unit for execution via function bus 212. Although function bus 212 is shown in
Data Buffer Controller (DBC) 206 might implement the data buffer function. DBC 206 fetches PDU data for MTIE 214 from memory external to MPP 200 (e.g., one of system memory 112 or external memory 116). DBC 206 might issue a read indication signal and a read done indication signal to FBI 216 to schedule the read requests. DBC 206 might have up to 2 read requests pending at any time for a given context. FBI 216 might prevent context termination if DBC 206 has pending reads for the context.
For functions that are defined as ordered, FBI 216 sends out function calls in the order in which the contexts are started in MPP 200. For functions that are not defined as ordered, FBI 216 might send out function calls in the order they are received by FBI 216. FBI 216 might typically queue contexts so that generally newer contexts wait for the generally oldest context to be executed. FBI 216 also determines the routing of each function call to a destination module and determines whether the function returns any data upon completion. If a function call does not return data, then FBI 216 sends the context to SCH 204 when the destination module returns an indication that it has started processing the function call. If the function call does return data, then FBI 216 sends the context to SCH 204 after the data is returned to FBI 216 by the destination module. Upon receiving the data, FBI 216 sends the data to MTIE 214, and MTIE 214 writes the data to an internal memory (not shown). Once the returned data is written to memory, the context is provided to SCH 204. Additionally, FBI 216 might determine if a function call is a “terminating” function call that ends context processing by MPP 200. Terminating function calls might typically be issued by Pre-Queue Modifier 208 directly to SCH 204. When a terminating function call is processed, MPP 200 generates an output task that is communicated, for example, over a ring communication bus to a next module of network processor 100 for subsequent processing after MPP 200.
MPP 200 might track a virtual flow identifier (vflow ID) and an index (vflow Index) with each output task, indicative of what one(s) of cores 106 or accelerators 108 operate on a data packet after MPP 200 has finished its processing. Communication interface 202 generates an output task based on the vflow ID and vflow Index and the output task is transmitted, for example via a task ring, to the subsequent destination module. An input task might result in the generation of multiple output tasks. As described herein, MPP 200 maintains task order between input and output, such that output tasks are generated in the order in which the input tasks are received by MPP 200, and thus also the order in which the corresponding contexts are started in MPP 200.
SCH 204 starts a new context when new tasks are received by MPP 200. SCH 204 receives a Task ID (TID) that identifies the received task and starts a context by allocating a context number to associate with that task. The TID and context number might be passed on to other modules of MPP 200 when the context is started. A context is associated with this TID and context number until SCH 204 receives an indication that processing of the context is terminated. In general, a new context is started for a received task if the following conditions are true: (1) there are available contexts; and (2) a Task Start FIFO buffer has enough available entries for at least one complete task. To start a new context, SCH 204 reads task information from one or more Task Start FIFO buffer locations. The Task Start FIFO buffers might be FIFO buffers stored in an internal memory of SCH 204. SCH 204 starts a context by allocating a new context number and setting a status bit of the context, indicating that this context is ready to be scheduled. SCH 204 stores the task information in a Per-Context Memory (PCM) of SCH 204. The PCM might be addressed by context number. In some embodiments, the PCM is implemented as a two-port memory with one port dedicated to write context information, and one port dedicated to read context information. The context information might also be provided to other modules of MPP 200 when the context is started, allowing the modules to initialize any per-context memories for the new context.
As will be described, SCH 204 maintains a Classification Completion List (CCL). The CCL stores pointers to the contexts and control data, such as context start order, context number, and thread identifiers (THID), for each context. When a new terminating function is issued by PQM 208 to SCH 204, the terminating function is appended to the CCL after any older CCL entries for the corresponding context. The next newest context, for example the next context in the CCL linked list, is then started. When a context becomes the oldest context in MPP 200, SCH 204 reads the CCL contents and sends them to PQM 208 to form instructions to communication interface 202 to generate a corresponding output task that is, for example, based on a vflow ID, a vflow Index, and the actual packet data. SCH 204 might determine which context is the oldest if the context is the head entry of the CCL linked list. Alternatively, if SCH 204 employs more than one output queue, a CCL linked list might exist for each output queue, and, thus, SCH 204 might select the oldest context from one of the output queues, and sends that context to PQM 208. Since an ordering requirement between OQs is not necessary, any non-empty OQ might be selected (for example, using a round robin algorithm) to begin transmission.
The CCL location is freed for another context and the output task is sent to the next destination module of network processor 100. When a context is terminated, that context is not reallocated until all other modules of MPP 200 have acknowledged to SCH 204 that they are done processing the context. Thus, as described herein, SCH 204 provides context start and context complete information to other modules of MPP 200, and provides context scheduling information to MTIE 214. As will be described, MTIE 214 might also provide instruction breakpoint information to SCH 204.
In situations where one or more system resources are running low, SCH 204 might stop scheduling contexts that consume the resources. Thus, SCH 204 might place a context in a “parked mode”. While a context is parked, SCH 204 will not schedule it to MTIE 214. SCH 204 might place a context in parked mode for any of the following cases. For case (1), the context is placed in a parked mode when free locations in the Classification Completion List (CCL) are below a minimum threshold, thus becoming at risk of not being able to satisfy all active contexts. In this condition, any context that allocates a new CCL location, and is not a terminating function, is parked by SCH 204. A context parked for this reason remains parked until free locations in the CCL are above the minimum threshold. For case (2), the context is placed in a parked mode when PQM 208 instruction memory is below a minimum threshold and at risk of not being able to satisfy all the active contexts. In this condition, any context that uses PQM instruction memory is parked by SCH 204. A context parked for this reason remains parked until free PQM instruction memory is above the minimum threshold. In some embodiments, contexts parked for either cases (1) or (2) might remain parked until the tests for both cases (1) and (2) are satisfied, for example, that free locations in the CCL are above the minimum threshold and free PQM instruction memory is above the minimum threshold. For case (3), the context is placed in a parked mode when SCH 204 parks a context due to an instruction breakpoint, which might be performed for diagnostic purposes. Thus, a context might be parked due to system resources being below a minimum (e.g., one or both of free locations in the CCL and free PQM instruction memory) or a context might be parked because of an instruction breakpoint.
The instruction breakpoint mechanism allows stepping through software code using a configuration-specified instruction breakpoint. As will be described, when a MTIE 214 executes an instruction that has a breakpoint set, and a breakpoint mode is enabled, MTIE 214 signals SCH 204 to park the context. Multiple contexts might be parked in this manner in a single clock cycle, since each of the one or more MTIE modules has an independent interface to SCH 204. Upon reaching an instruction having a breakpoint, MTIE 214 might send the context to SCH 204 with a corresponding breakpoint indication set. Upon receiving a context with the breakpoint indication set, SCH 204 might request all of the one or more MTIE modules to send all the active contexts to SCH 204 and put the contexts in instruction breakpoint park mode. Once SCH 204 has received control over all active contexts, SCH 204 might generate an interrupt, for example, to one of the various μP cores 106 of network processor 100.
Through debug interface 236, a module external to MPP 200, for example the one of μP cores 106 that received the interrupt, might interrogate the state of MTIE 214, SCH 204, and other modules of MPP 200. Once software is done with its examination of the hardware, it might returns to the hardware to a running state by clearing the scheduler control register. After the one of μP cores 106 that received the interrupt is finished interrogating the state of MPP 200, the interrupt might be cleared to return MPP 200 to a running state. When returned to a running state, SCH 204 clears the instruction breakpoint park for all contexts, allowing them to be rescheduled to MTIE 214. When not in breakpoint mode, for each clock cycle, SCH 204 attempts to pick a context to schedule to MTIE 214, based on the status of the contexts, for example, contexts with a “ready” status and that are not parked. When SCH 204 is in breakpoint mode, no contexts are rescheduled, and no new contexts are started.
MTIE 214 includes flow memory 230. Flow memory 230 might be 24 bits wide and 16 KB in size. The first (e.g., lowest addressed) flow instructions might be stored in the flow instruction cache of flow memory 230, while subsequent instructions (e.g., higher addressed flow instructions) might be mapped by a base register of MTIE 214 and stored in external tree memory 234. In exemplary embodiments, MPP 200 might include 1, 2, 4, 6 or 8 MTIE cores. In embodiments with multiple MTIE cores, the flow memory of one or more cores might be joined together to increase the flow memory size for the overall group of MTIE cores. In general, flow memory 230 might have a lower read latency versus external tree memory 234.
MTIE 214 includes root tree memory 232. Root tree memory 232 might include 1K of memory and might contain the first 1024 tree instructions for zero latency instruction access. In general, root tree memory 232 might have a lower read latency versus external tree memory 234. To improve the read latency of external tree memory 234, data might be duplicated across one or more locations of external tree memory 234. For example, as will be described, one logical address might be mapped to one or more physical addresses in external tree memory 234. The contents of the tree memory might be duplicated across one or more physical memory banks of external tree memory 234 to reduce memory contention for frequently accessed instructions.
Described embodiments reduce the average latency of read requests to a memory that is read by one or more requestors, where the memory might include two or more substructures (e.g., multiple banks of DRAM). In such a system, a latency penalty might be incurred for read requests to the same substructure sequentially, and the average latency of read requests might be reduced by having multiple copies of the same data in multiple memory substructures. In such an embodiment, a requestor might initiate a read request to the substructure that holds a copy of the data that will incur the smallest latency. This decision might be based on knowledge of prior requests and which data is duplicated. In described embodiments, the address of the read request is used to lookup the availability of duplicated data from a programmable table based on address ranges.
Embodiments of the present invention further provide that the data to be duplicated can be chosen as less than all the data, based on usage statistics of the data and the size of available memory. For example, heavily used data might be duplicated in all of the memory substructures to minimize access time, while infrequently used data would have fewer copies or not be duplicated at all, allowing more overall data to be stored in the memory. Thus, the level of data duplication is configurable based on the requirements of a given implementation of network processor 100.
When MTIE 214 requires data from one of root tree memory 232 and external tree memory 234, MTIE 214 sends a read request external tree memory 234. Read requests might be temporarily stored in tree memory FIFO 302. Comparator 306 compares at least a portion of the requested address against the entries of lookup table 304. Comparator 306 returns the value Table_Hit, which is a match indication for the table index whose address range included the request address. For example, when the requested address is less than or equal to the ending address of the address range (IEND [N]) and is greater than or equal to the base address of the address range (IBASE[N]), and when the corresponding valid indication (Valid[N]) is set. For example, in described embodiments, MTIE 214 might perform the address range comparison: Table_Hit[N−1:0]=(Valid[N] && (IBASE[N]<=Request Address<=IEND[N])). The information about how the data in that address range is duplicated (DF[N]) and any other information (SBASE[N]) required to transform the request address into an actual structure address is read by SubStructure Selector and Address Former 314 from the matching entry of lookup table 304 defined by the Table_Hit value, and selected by multiplexer 310.
MTIE 214 might also maintain a corresponding “busy” indicator for each memory substructure, for example, SubStructure Status BitMask 312, which includes a bitmask of SubStructure_Busy[M−1:0]. Based on the requested address, address translation information is read from the table and, based on the address translation information and the “busy” state of the memory substructure that includes the requested address, SubStructure Selector and Address Former 314 might determine a memory address to read that would result in the minimum latency. SubStructure Status BitMask 312 is updated for the substructure that receives the request, allowing its ability to accept future requests to be tracked.
SubStructure Status BitMask 312 might include one bit per memory bank. SubStructure Status BitMask 312 might set an indicator, such as a flag bit, for one cycle after a bank was accessed to indicate that the corresponding memory bank is busy for one clock cycle to process a read request. The indicator for the corresponding memory bank might clear the following clock cycle to indicate that the memory bank is available to accept new read requests. For substructures that have more than a one clock cycle latency penalty between requests, their status could be tracked with a counter, shift register or some similar mechanism to indicate their busy status over multiple clock cycles. For dynamic substructures that require periodic refresh cycles, the refresh status of the structures might also be tracked and used as input to at least one of SubStructure Selector and Address Former 314 or SubStructure Status BitMask 312.
If there are back-to-back read requests, SubStructure Selector and Address Former 314 might provide the first read request to either memory bank. SubStructure Selector and Address Former 314 might provide the second read request to the bank that was not selected for the first read request. For example, if the first request went to bank 0, it would set SubStructure_Busy[0] to indicate bank 0's busy status. For the second request in the next clock cycle, SubStructure Selector and Address Former 314 reads SubStructure_Busy[0] indicating that bank 0 is busy, sends the second request to bank 1, and sets SubStructure_Busy[1] to indicate bank 1's busy status. For a third request in the next clock cycle, SubStructure_Busy[0] indicates that bank 0 is available, and SubStructure_Busy[1] indicates that bank 1 is busy, and the third read request is sent to bank 0, and so on, for subsequent read requests. Request addresses 0-7 map to structure addresses 0-7 and 8-15.
Table[0]: IBASE=0, IEND=7, SBASE=0, DF=0 (locations 0-7 not duplicated), and
Table[1]: IBASE=8, IEND=11, SBASE=4, DF=1 (locations 8-11 duplicated).
For data addresses that are not duplicated (e.g., DF=0), the storage address is the request address (e.g., 0-7). For data addresses that are duplicated, the storage address might be formed by SubStructure Selector and Address Former 314 by concatenating the request address with one or more values from lookup table 304. SubStructure Selector and Address Former 314 might form the storage address by performing the concatenation: (Requested Logical Address−IBASE+SBASE) and, if data duplication is enabled (e.g., DF=1), and shifting the result by a number corresponding to the number of memory banks with the data duplication. As shown in
For example, in the exemplary case shown in
Alternatively, in the exemplary case shown in
Table 2 defines terms used herein as an aid to understanding the described embodiment:
MPP 200 might typically employ multi-threaded processing to interface with high latency memory systems. As input packets arrive, MPP 200 starts a thread by sending a Thread Start (TS) indication to SCH 204. A new thread might start execution even when an older thread has not completed execution. A newer thread might complete execution before an older thread has completed execution. SCH 204 might include multiple output queues (OQ), and each thread might specify its corresponding OQ before starting output transmission. SCH 204 maintains “wire order” on a particular OQ, meaning that each OQ transmits the packets for a given thread contiguously in the order in which the threads were started, regardless of any interleaving of the input packets between threads. Embodiments of the present invention allow efficient implementation of wire order transmission in a multi threaded, multi OQ system. Described embodiments provide SCH 204 to efficiently transmit threads in the order in which they were started, and to select them from multiple OQs.
As described herein, MPP 200 transmits packets in wire order. Tables 3-5 show an exemplary condition for processing packets of 3 threads in a system employing two output queues (OQ0 and OQ1). As shown in Tables 3-5, below, an ordering requirement is not necessarily required between OQ0 and OQ1. In these tables, the OQ is shown as being specified in the TS indication, but the OQ corresponding to a thread might be specified at any time up until or coincident with MPP 200 receiving the first packet for a given thread.
Tables 3-5 show an exemplary case of a last overall packet received for various active threads. When the last packet for a particular thread is transmitted, it is an indication for the next thread in that particular OQ (if any) to begin transmission.
One embodiment of MPP 200 might transmit the thread for the first input packet to arrive, and continue transmitting each input packet as it arrives, enqueuing all other input packets (those for other threads) into a relatively large queue. After the last input packet is received for the given thread, SCH 204 begins processing the next oldest entry in the queue (deleting it from the head of the queue), and traverses the queue from oldest to newest, extracting (and transmitting) any entries that pertain to the next thread. If any input packets were received for that thread while SCH 204 was traversing the queue, SCH 204 enqueues that input packet into the large queue. If SCH 204 reached an entry which was the last entry for that thread, SCH 204 would then begin transmitting a new thread starting with the (next) oldest queue entry, starting back at the head of the queue. SCH 204 continues this algorithm unless or until the queue was empty. If SCH 204 traversed the entire queue without finding the last entry for the thread, SCH 204 stops transmitting until the last input packet for that thread was received. This embodiment might be relatively inefficient since the entire queue would need to be repeatedly traversed; if there were a large number of threads in progress, this could take a very long time. This embodiment requires a large amount of memory for SCH 204 to support a large number of simultaneously-active threads.
As shown in
As described, CL 802 is a linked list which stores information necessary to transmit a particular input packet associated with a particular thread. Each CL entry includes a link pointer to another CL entry (either the next CL entry for that thread, or the first entry of the next thread in the same OQ). Each CL entry also stores the thread identifier (THID) of the thread and an indication if the entry is the last CL entry for the thread (not necessarily the last in the OQ). The entries for a given thread stored in CL 802 are linked to each other. Threads that have specified their OQ have their smaller linked lists within CL 802 linked together.
SCH 204 maintains a Head Pointer (shown as 804(1)-804(N)) and a Tail Pointer (shown as 805(1)-805(N)) in Per-Thread Table (PTT) 808 for each OQ 704(1)-704(N). The HP points to the oldest CL entry for a given OQ. The oldest CL entry is the next entry to be transmitted for that queue. The TP points to the newest (last) CL entry for the given OQ.
Oldest Unspecified List (OUL) 702 is a list used by SCH 204 to track the order in which the TS indications were received for each thread. The oldest thread in the list is removed after it has specified its OQ. OUL 702 is an ordered list of THIDs for which SCH 204 has received a TS. The oldest entry is not read from OUL 702 until it has specified its OQ.
Queue Table (QT) 806 is a table used by SCH 204 to track the OQ specified for each THID. QT 806 is a per-THID table that records the OQ number specified for a given THID, and a valid bit indicating whether or not that THID has yet specified its OQ number. PTT 808 records, for each THID, the head pointer (oldest) and tail pointer (newest) entry for that thread within CL 802. At a given point in time, these smaller linked lists may or may not be linked to other linked lists within CL 802, depending on whether or not the thread has been moved out of the OUL.
When SCH 204 receives an indication of the start of a thread, SCH 204 records the TS indicator in OUL 702. Entries in OUL 702 are written in the order in which the threads are started, and read in the same order. Before, or coincident with when SCH 204 receives the first input packet for a thread, SCH 204 receives an indication of which OQ the thread is to use. SCH 204 records this OQ number in QT 806 and sets the valid bit for that QT entry. When SCH 204 receives an input packet for a thread, it updates PTT 808. A new CL entry is allocated for the input packet, and the corresponding HP and TP of PTT 808 for that THID are updated to link in the new CL location. If this is the first packet for the thread, PTT 808 HP and TP are both set to point to the new CL entry. If there are already one or more CL entries for the thread, the oldest CL entry link is pointed to the new CL entry, and PTT 808 TP is set to point to the new CL entry. The information necessary to transmit the packet is also written to CL 802, as well as the indication of whether or not the packet is the last one for this thread.
While a thread is in OUL 702, OUL 702 might receive input packets. If the thread is not the oldest OUL entry, and the oldest entry has not yet specified its OQ (that is, the valid bit in the QT is still 0), the thread must remain in OUL 702. The corresponding entry of PTT 808 for the thread is updated, but the thread is not yet “moved” out of OUL 702 (e.g., not linked to an OQ). When the oldest thread in OUL 702 has specified its OQ, the thread is moved into CL 802 in the specified one of OQs 704(1)-704(N).
As shown in
When there are any non-empty OQs, transmitting threads might be permitted to start. Since an ordering requirement between OQs is not necessary, any non-empty OQ might be selected (for example, using a round robin algorithm) to begin transmission. Once an OQ is selected, the selected OQ is the only OQ to transmit until the end of the thread is reached, which can be determined by examining the “Last” bit stored in the CL. To transmit a thread, SCH 204 selects a non-empty OQ and begins reading locations from CL 802 using the OQ HP for the selected queue. If SCH 204, when it selects an OQ to transmit, is in the middle of a current transmission, SCH 204 stays in this mode until it reads a CL entry which has the Last bit set.
Before transmitting, SCH 204 examines the oldest entry in the OQ (the CL entry pointed to by the OQ HP). If the next entry to be read has a different THID than the last entry read, and the previous entry did not have the Last bit set, SCH 204 stops transmitting until the next (and possibly last) packet for the thread is received. In this case SCH 204 enters “Bypass Mode”, and records the THID of the thread which SCH 204 is in the middle of transmitting. SCH 204 also enters “Bypass Mode” if the OQ becomes empty after reading a location which did not have the Last bit set. Otherwise, if SCH 204 reads and transmits an entry which has the Last bit set, then it is no longer in the middle of transmitting a thread and may select any non-empty OQ for the next thread to transmit.
While SCH 204 is in Bypass Mode, if it receives a new input packet it examines the THID for the packet. If the THID matches the THID for which it is in Bypass Mode (the bypass THID), then the packet information is passed right to the output, bypassing the CL. SCH 204 remains in this mode until such an input packet is received which has the Last bit set. If input packets are received which do not match the bypass THID, SCH 204 handles the input packet in a normal manner by adding the input packet to OUL 702 and/or CL 802. A particular THID is not necessarily reused by MPP 200 until the THID has at least been moved from OUL 702 to CL 802. At that time, the valid bit in the QT is reset to 0.
In the case where an empty thread is linked into an OQ, and a CL entry might be allocated but not yet used, the next (first) input packet for that thread might use the CL entry. One possible alternative implementation would be to not move the oldest OUL location into its OQ until the first input packet is received for that thread; with that alternative, there would never be the case of moving an empty thread to an OQ.
Embodiments of the present invention provide hardware instruction break point capability in a multi-threaded processing environment. A dedicated instruction break point flag is added to each instruction word that allows the execution engine to halt execution of the running thread and return it to the scheduler. The scheduler then signals the execution engine to return all remaining running threads to the scheduler and enter an idle state. Through a debug interface, the instruction break point status of each thread in the scheduler can be queried and the thread state memories in the execution engine can be accessed for analysis.
A typical software instruction break point might replace a given instruction with a special debugging instruction. Upon execution of the break point instruction, the running thread is halted. The debug instruction is a part of the instruction set that the underlying execution engine decodes and executes similarly to any other instruction of the instruction set. Additionally, inter-thread communication might be required to bring the execution engine to an orderly idle state before debugging begins. Embodiments of the present invention provide a hardware instruction break point that adds a dedicated instruction break point flag to each instruction word of the instruction set. If the instruction break point is enabled and the instruction break point flag is set, the execution engine executes an implicit no op instruction and returns the running thread to the scheduler. The scheduler then signals the execution engine to return all remaining running threads and enter an idle state. Multiple running threads might reach the same or different instruction break points at the same time. Through a debug interface, the instruction break point status of each thread might be queried and the thread state memories in the execution engine might be read.
A dedicated instruction break point flag in the instruction word is used to indicate to execution engine MTIE 214 that a running thread is to be returned to SCH 204 to be parked due to the breakpoint. MTIE 214 might include a configuration register to enable the instruction break point flag. Upon receiving a thread including an instruction break point, SCH 204 signals MTIE 214 to return all remaining running threads to SCH 204 to be parked, thus putting MTIE 214 in an idle state.
As described herein, in a multi-threaded processing system such as network processor 100, each thread executes a flow of instructions based upon task assignment. Typically, an instruction set for such a multi-threaded processing system is small and each thread is allocated state memories such as instruction pointer, argument pointer, stack, global registers, and the like. As shown in
SCH 204 waits at step 1322 for the breakpoint to be released, for example, via a signal received from the debug interface. Through the debug interface, the thread instruction breakpoint status in SCH 204 might be accessed by devices external to network processor 100 via, for example, a Joint Test Action Group (JTAG) interface, a Serial Wire Debug (SWD) interface, a Serial Peripheral Interface (SPI) or a Universal Asynchronous Receiver/Transmitter (UART). Thread state memories in MTIE 214 might similarly be accessed for analysis. Once the breakpoint is released by, for example, a device external to network processor 100 via the debug interface, at step 1324 the parked threads are returned from SCH 204 to MTIE 214 to resume instruction execution. At step 1324, when SCH 204 returns parked threads to MTIE 214 to resume instruction execution, SCH 204 also returns an indication of which instruction(s) first reached the breakpoint. At step 1310, MTIE 214 then executes the instruction that first reached the breakpoint once it is returned from SCH 204 without requiring the corresponding breakpoint flag to be cleared first.
Processing of the thread might continue as described above until the thread is completed. At step 1326, if the thread is not complete, MTIE 214 might retrieve the next thread instruction at step 1304. If the thread is complete, at step 1328, MTIE 214 returns the thread status to SCH 204. At step 1330, SCH 204 retires the competed thread and thread processing of the corresponding thread is complete. When multiple threads are active, processing continues for each thread until each thread is completed.
SCH 204 might include one bit vector per each context. Via the debug interface, a breakpoint might be set on a particular address in the instruction memory (e.g., flow memory 230) of MTIE 214. When that particular address is accessed by MTIE 214 to read and process that instruction, MTIE 214 recognizes the breakpoint and returns the thread to SCH 204, just as if the thread had completed normally. SCH 204 then halts all threads in MTIE 204 by requesting MTIE 214 return any remaining threads to SCH 204. Thus, embodiments of the present invention provide a scheduler module to halt threads from one or more processor of an SoC.
Embodiments of the present invention provide that threads in a multithreaded system might be allocated (started) in any order and de-allocated (terminated) in any order, and that processes associated with the threads are handled in the order in which the threads were started. Embodiments of the present invention define a per-thread state structure, how the structure is managed when threads are allocated or de-allocated and how per-thread status information is used to find the oldest thread. This per-thread status structure allows for: i) tracking active threads in thread start order; ii) single cycle update of per-thread status on a thread de-allocate; and iii) single cycle lookup of the next oldest thread.
As described herein, network processor 100 might execute multiple threads in parallel with functions for the various threads issued without particular ordering. Synchronizing processing of these events or functions in the order the threads were started might be desirable. Specific events or functions that need to be ordered might be defined within submodules of network processor 100 such that only the threads associated with these functions are ordered. For example, functions destined for different modules might be defined to be ordered by FBI 216. A list of active threads might be maintained in the order the threads were started and this active thread list might be used for scheduling events or functions associated with the thread. Embodiments of the present invention allow for management of active threads in thread start order and updates the active thread list on a thread de-allocate event. Further, embodiments of the present invention provide simplified lookup of the oldest active thread.
Some design implementations typically use linked list structures maintained in memory for tracking active threads. Removal of an active thread from middle of the linked list due to a thread de-allocate event requires 2 clock cycles: one clock cycle to read the link from memory and a second clock cycle to write the value to different memory location. Since this operation takes two clock cycles, the operation requires additional complexity, such as FIFOs and hold logic, for processing back-to-back thread de-allocate events. Another approach implements event order lists or memory structures with a scalable number of read ports, meaning that each read port has dedicated RAM for optimal performance. The number of read ports is a function of how many independent events need to be synchronized, so, to prevent backup of threads in cases where oldest thread is not de-allocated for a long time, the ordered list size might be large.
Embodiments of the present invention define i) a data structure for tracking currently active threads by thread start order, ii) allocate and de-allocate events to update the thread status information, and iii) a sequence value to identify next oldest thread in the list. As shown in
MPP 200 might maintain a global sequence counter that is incremented each time a new thread is allocated. When a thread is allocated, thread status data structure 1400 is updated such that the sequence field (e.g., the corresponding one of 1404(1)-1404(N)) for the thread is updated with the sequence number. The valid bit (e.g., the corresponding one of 1402(1)-1402(N)) is set to 1. When the thread is de-allocated, the structure corresponding to the thread is updated. For any thread structure with a sequence value greater or equal to the sequence value of the de-allocated thread, the sequence value is decremented. The valid bit is cleared for the de-allocated thread. The global sequence counter is decremented.
When a thread is de-allocated, the sequence value and thread value associated with this thread is read from thread status data structure 1400. These values might be broadcast to modules of MPP 200, for example, as shown in
Thread status data structure 1400 might be updated by TSM 1500 through comparison logic (not shown) to determine if the incoming sequence matches the sequence associated with this thread. Structures with no matches output a value of 0 for the thread. The sequence values for each valid thread are mutually exclusive; therefore, for any sequence, at most there is generally only one match. All the output thread values are logic ORed together by OR gate 1506 to generate a thread value. Rd port 1504 is used by ESMs 1502(1)-1502(Y) to find the oldest thread in thread status data structure 1400. As described, the oldest thread is assigned sequence value of 0, until this thread is de-allocated, at which point each active thread has its corresponding sequence value decremented, where the thread with resulting sequence value of 0 is the oldest thread. As shown in
As shown in
With more than one active thread in the system, each ESM might lookup the next oldest thread information by advancing the local sequence value and using it to request thread value via Rd Port 1506. Each ESM updates its local sequence value appropriately when a thread de-allocate request is provided on de-allocate interface 1510. ESMs 1502(1)-1502(Y) use the sequence value to adjust their local sequence values accordingly. ESMs with a local sequence value greater than or equal to the de-allocate sequence value decrement their local sequence values.
As shown, ESM11604 receives a fund event request from thread 3 at time T=6. Threads scheduled by ESM01602 are shown as threads 0′, 1′, 2′ and 4′. ESM01602 schedules func0 events on threads 0, 1 and 2 however, ESM01602 cannot schedule a func0 event for thread 4 until thread 3 is de-allocated, or thread 3 requests a func0 event, such that thread 4 becomes the oldest unscheduled thread for ESM01602. Threads scheduled by ESM01602 might be employed to request func1 events. As shown in the example of
As described with regard to
Thus, as described herein, embodiments of the present invention provide a packet classifier for a network processor that generates tasks corresponding to each received packet. The packet classifier includes a scheduler to generate contexts corresponding to tasks received by the packet classifier from a plurality of processing modules of the network processor. A multi-thread instruction engine processes threads of instructions, each thread of instructions corresponding to a context received from the scheduler. A thread status manager maintains a thread status table having N entries to track up to N active threads. Each status entry includes a valid status indicator, a sequence value, and a thread indicator. A sequence counter generates a sequence value for each thread and is incremented when processing of a thread is started, and is decremented when a thread is completed, by the multi-thread instruction engine. Instructions are processed by the multi-thread instruction engine in the order in which the threads were started.
While the exemplary embodiments of the present invention have been described with respect to processing blocks in a software program, including possible implementation as a digital signal processor, micro-controller, or general purpose computer, the present invention is not so limited. As would be apparent to one skilled in the art, various functions of software might also be implemented as processes of circuits. Such circuits might be employed in, for example, a single integrated circuit, a multi-chip module, a single card, or a multi-card circuit pack.
The present invention can be embodied in the form of methods and apparatuses for practicing those methods. The present invention can also be embodied in the form of program code embodied in tangible media, such as magnetic recording media, optical recording media, solid state memory, floppy diskettes, CD-ROMs, hard drives, or any other non-transitory machine-readable storage medium, wherein, when the program code is loaded into and executed by a machine, such as a computer, the machine becomes an apparatus for practicing the invention. The present invention can also be embodied in the form of program code, for example, whether stored in a non-transitory machine-readable storage medium, loaded into and/or executed by a machine, or transmitted over some transmission medium or carrier, such as over electrical wiring or cabling, through fiber optics, or via electromagnetic radiation, wherein, when the program code is loaded into and executed by a machine, such as a computer, the machine becomes an apparatus for practicing the invention. When implemented on a general-purpose processor, the program code segments combine with the processor to provide a unique device that operates analogously to specific logic circuits. The present invention can also be embodied in the form of a bitstream or other sequence of signal values electrically or optically transmitted through a medium, stored magnetic-field variations in a magnetic recording medium, etc., generated using a method and/or an apparatus of the present invention.
It should be understood that the steps of the exemplary methods set forth herein are not necessarily required to be performed in the order described, and the order of the steps of such methods should be understood to be merely exemplary. Likewise, additional steps might be included in such methods, and certain steps might be omitted or combined, in methods consistent with various embodiments of the present invention.
As used herein in reference to an element and a standard, the term “compatible” means that the element communicates with other elements in a manner wholly or partially specified by the standard, and would be recognized by other elements as sufficiently capable of communicating with the other elements in the manner specified by the standard. The compatible element does not need to operate internally in a manner specified by the standard.
Also for purposes of this description, the terms “couple,” “coupling,” “coupled,” “connect,” “connecting,” or “connected” refer to any manner known in the art or later developed in which energy is allowed to be transferred between two or more elements, and the interposition of one or more additional elements is contemplated, although not required. Conversely, the terms “directly coupled,” “directly connected,” etc., imply the absence of such additional elements. Signals and corresponding nodes or ports might be referred to by the same name and are interchangeable for purposes here.
It will be further understood that various changes in the details, materials, and arrangements of the parts which have been described and illustrated in order to explain the nature of this invention might be made by those skilled in the art without departing from the scope of the invention as expressed in the following claims.
This application claims the benefit of the filing date of U.S. provisional application Nos. 61/313,399 filed Mar. 12, 2010 and 61/313,219 filed Mar. 12, 2010, the teachings of which are incorporated herein in their entireties by reference. This application is a continuation-in-part, and claims the benefit of the filing date, of U.S. patent application Ser. No. 12/782,379 filed May 18, 2010, Ser. No. 12/782,393 filed May 18, 2010, and Ser. No. 12/782,411 filed May 18, 2010, the teachings of which are incorporated herein in their entireties by reference. The subject matter of this application is related to U.S. patent application Ser. No. 12/430,438 filed Apr. 27, 2009, Ser. No. 12/729,226 filed Mar. 22, 2010, Ser. No. 12/729,231 filed Mar. 22, 2010, Ser. No. 12/963,895 filed Dec. 9, 2010, Ser. No. 12/971,742 filed Dec. 17, 2010, ______ having attorney docket number L09-0547 (300.251) filed on common date herewith, ______ having attorney docket number L09-0734 (300.293) filed on common date herewith, and ______ having attorney docket number L10-0127 (300.271) filed on common date herewith, the teachings of which are incorporated herein in their entireties by reference.
Number | Date | Country | |
---|---|---|---|
61313399 | Mar 2010 | US | |
61313219 | Mar 2010 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 12782379 | May 2010 | US |
Child | 12974477 | US | |
Parent | 12782393 | May 2010 | US |
Child | 12782379 | US | |
Parent | 12782411 | May 2010 | US |
Child | 12782393 | US |