Patent Application
20060190945
The present invention relates in general to the field of multithreaded microprocessors, and particularly to execution of multiprocessor operating systems thereon.
Microprocessor designers employ many techniques to increase microprocessor performance. Most microprocessors operate using a clock signal running at a fixed frequency. Each clock cycle the circuits of the microprocessor perform their respective functions. According to Hennessy and Patterson (see Computer Architecture: A Quantitative Approach, 3rd Edition), the true measure of a microprocessor's performance is the time required to execute a program or collection of programs. From this perspective, the performance of a microprocessor is a function of its clock frequency, the average number of clock cycles required to execute an instruction (or alternately stated, the average number of instructions executed per clock cycle), and the number of instructions executed in the program or collection of programs. Semiconductor scientists and engineers are continually making it possible for microprocessors to run at faster clock frequencies, chiefly by reducing transistor size, resulting in faster switching times. The number of instructions executed is largely fixed by the task to be performed by the program, although it is also affected by the instruction set architecture of the microprocessor. Large performance increases have been realized by architectural and organizational notions that improve the instructions per clock cycle, in particular by notions of parallelism.
One notion of parallelism that has improved the clock frequency of microprocessors is pipelining, which overlaps execution of multiple instructions within pipeline stages of the microprocessor. In an ideal situation, each clock cycle one instruction moves down the pipeline to a new stage, which performs a different function on the instruction. Thus, although each individual instruction takes multiple clock cycles to complete, the multiple cycles of the individual instructions overlap. Because the circuitry of each individual pipeline stage is only required to perform a small function relative to the sum of the functions required to be performed by a non-pipelined processor, the clock cycle of the pipelined processor may be reduced. The performance improvements of pipelining may be realized to the extent that the instructions in the program permit it, namely to the extent that an instruction does not depend upon its predecessors in order to execute and can therefore execute in parallel with its predecessors, which is commonly referred to as instruction-level parallelism. Another way in which instruction-level parallelism is exploited by contemporary microprocessors is the issuing of multiple instructions for execution per clock cycle. These microprocessors are commonly referred to as superscalar microprocessors.
What has been discussed above pertains to parallelism at the individual instruction-level. However, the performance improvement that may be achieved through exploitation of instruction-level parallelism is limited. Various constraints imposed by limited instruction-level parallelism and other performance-constraining issues have recently renewed an interest in exploiting parallelism at the level of blocks, or sequences, or streams of instructions, commonly referred to as thread-level parallelism. A thread is simply a sequence, or stream, of program instructions. A multithreaded microprocessor concurrently executes multiple threads according to some scheduling policy that dictates the fetching and issuing of instructions of the various threads, such as interleaved, blocked, or simultaneous multithreading. A multithreaded microprocessor typically allows the multiple threads to share the functional units of the microprocessor (e.g., instruction fetch and decode units, caches, branch prediction units, and load/store, integer, floating-point, SIMD, etc. execution units) in a concurrent fashion. However, multithreaded microprocessors include multiple sets of resources, or contexts, for storing the unique state of each thread, such as multiple program counters and general purpose register sets, to facilitate the ability to quickly switch between threads to fetch and issue instructions. In other words, because each thread context has its own program counter and general purpose register set, the multithreading microprocessor does not have to save and restore these resources when switching between threads, thereby potentially reducing the average number of clock cycles per instruction.
One example of a performance-constraining issue addressed by multithreading microprocessors is the fact that accesses to memory outside the microprocessor that must be performed due to a cache miss typically have a relatively long latency. It is common for the memory access time of a contemporary microprocessor-based computer system to be between one and two orders of magnitude greater than the cache hit access time. Instructions dependent upon the data missing in the cache are stalled in the pipeline waiting for the data to come from memory. Consequently, some or all of the pipeline stages of a single-threaded microprocessor may be idle performing no useful work for many clock cycles. Multithreaded microprocessors may solve this problem by issuing instructions from other threads during the memory fetch latency, thereby enabling the pipeline stages to make forward progress performing useful work, somewhat analogously to, but at a finer level of granularity than, an operating system performing a task switch on a page fault. Other examples of performance-constraining issues addressed by multithreading microprocessors are pipeline stalls and their accompanying idle cycles due to a data dependence; or due to a long latency instruction such as a divide instruction, floating-point instruction, or the like; or due to a limited hardware resource conflict. Again, the ability of a multithreaded microprocessor to issue instructions from independent threads to pipeline stages that would otherwise be idle may significantly reduce the time required to execute the program or collection of programs comprising the threads.
Multiprocessing is a technique related to multithreading that exploits thread-level parallelism, albeit at a higher system level, to execute a program or collection of programs faster. In a conventional multiprocessor system, multiple processors, or CPUs, share a memory system and I/O devices. A multiprocessor (MP) operating system facilitates the simultaneous execution of a program or collection of programs on the multiprocessor system. For example, the system may include multiple Pentium IV processors all sharing a memory and I/O subsystem running an MP operating system—such as Linux SMP, an MP-capable version of Windows, Sun Solaris, etc., and executing one or more application programs concurrently.
Multithreading microprocessors exploit thread-level parallelism at an even lower level than multiprocessor systems by sharing instruction fetch, issue, and execution resources, as described above, in addition to sharing a memory system and I/O devices. An MP operating system may run on a multithreading microprocessor if the multithreading microprocessor presents multiple processors, or CPUs, in an architected manner recognized by the MP operating system. Perhaps the most highly publicized example is the Hyper-Threading (HT) Technology employed in the Intel® Xeon® multithreading microprocessor. An HT Xeon includes effectively the same execution resources (e.g., caches, execution units, branch predictors) as a non-HT Xeon processor, but replicates the architectural state to present multiple distinct logical processors to an MP OS. That is, the MP operating system recognizes each logical processor as a separate processor, or CPU, each presenting the architecture of a single processor. The cost of replicating the architectural state for the additional logical processor in the Xeon in terms of additional chip size and power consumption is almost 5%.
One aspect of the architecture presented by each of the multiple processors to the MP operating system is the ability to handle a list of architected exceptions. Generally speaking, an exception is an error or other unusual condition or event that occurs during the execution of a program. In response to an exception, the processor saves the state of the currently executing program and begins fetching and executing instructions at a predefined address, thereby transferring execution to an alternate program, commonly referred to as an exception handler located at the predefined address. The predefined address may be common to all exceptions in the list of architected exception types or may be unique to some or all of the exception types. The exception handler, when appropriate, may restore the state and resume execution of the previously executing program. Examples of common exceptions include a page fault, a divide by zero, a faulty address generated by the program, a bus error encountered by the processor when attempting to read a memory location, or an invalid instruction exception caused by an invalid instruction opcode or invalid instruction operand.
Another common exception type is an interrupt, or interrupt request. Interrupts are typically grouped as hardware interrupts and software interrupts. A software interrupt is generated when the currently executing program executes an architected software interrupt instruction, which causes an exception that transfers control to the architected interrupt vector associated with the software interrupt to invoke an interrupt service routine, or handler. A hardware interrupt is a signal received by the processor from a device to request service by the processor. Examples of interrupting devices are disk drives, direct memory access controllers, and timers. In response to the interrupt request, the processor transfers control to an architected interrupt vector associated with the interrupt request to invoke an interrupt service routine, or handler.
One function which MP operating systems need to be able to perform is for one processor, or CPU, to interrupt the operation of another specific one of the processors, and in some cases to interrupt all the processors in the system. These operations are sometimes referred to as inter-processor interrupts (IPIs). Commonly in a multiprocessor system, each processor includes an interrupt controller, which enables each processor to direct an interrupt specifically to each of the other processors. The HT Xeon processors, for example, include a replicated Advanced Programmable Interrupt Controller (APIC) for each logical processor, which enables each logical processor to send a hardware interrupt specifically to each of the other logical processors.
An example of the use of an IPI is in preemptive time-sharing operating systems, which receive periodic timer interrupts, in response to which the operating system may perform a task switch on one or more of the processors to schedule a different task or process to execute on the processors. In Linux SMP, for example, the timer handling routine running on the processor that receives the timer interrupt not only schedules the tasks on its own processor, but also directs an interrupt to each of the other processors to cause them to schedule their tasks. Each processor has an architected interrupt mechanism, which the timer interrupt-receiving processor uses to direct an IPI to each of the other processors in the multiprocessor system.
Another multithreading microprocessor core architecture which takes a somewhat different approach than, for example, the Intel HT architecture is the MIPS® Multithreading (MT) Application-Specific Extension (ASE) of the MIPS Instruction Set Architecture (ISA) and MIPS Privileged Resource Architecture (PRA). The MIPS MT ASE allows two distinct, but not mutually-exclusive, multithreading capabilities. A single MIPS MT ASE microprocessor core comprises one or more Virtual Processing Elements (VPEs), and each VPE comprises one or more thread contexts (TCs). This architecture is described in the document MIPS32® Architecture for Programmers Volume IV-f: The MIPS® MT Application-Specific Extension (ASE) to the MIPS32® Architecture, Document Number: MD00378, Revision 1.00, Sep. 28, 2005, available from MIPS Technologies, 1225 Charleston Road, Mountain View, Calif. 94043-1353, which is hereby incorporated by reference in its entirety for all purposes. Embodiments of the architecture are also described in the above-referenced U.S. patent applications.
In the MIPS MT ASE architecture, an N-VPE processor core presents to an SMP operating system an N-way symmetric multiprocessor. In particular, it presents to the SMP operating system N MIPS32® Architecture processors. Thus, SMP operating systems configured to run on a conventional multiprocessor system having N MIPS32 processors without the MT ASE capability will run on a single MIPS32 core with the MT ASE capabilities with little or no modifications to the SMP operating system. In particular, each VPE presents an architected exception domain to the SMP operating system including an architected list of exceptions that the VPE will handle. The list includes interrupts that one VPE may direct to another specific VPE in the multithreading microprocessor, somewhat similar to the HT Xeon approach.
As mentioned above, each VPE comprises at least one thread context, and may comprise multiple thread contexts. A thread context in the MIPS MT ASE comprises a program counter representation, a set of general purpose registers, a set of multiplier result registers, and some of the MIPS PRA Coprocessor 0 state, such as state describing the execution privilege level and address space identifier (ASID) of each thread context. The thread contexts are relatively lightweight compared to VPEs with respect to storage elements required to store state and are therefore less expensive than VPEs in terms of chip area and power consumption. Advantageously, the lightweight feature of MIPS MT ASE thread contexts makes them inherently more scalable than VPEs, and potentially than Intel HT logical processors, for example.
In particular, in the interest of providing lightweight thread contexts and the concomitant advantages, such as improved scalability, within the MIPS MT ASE, the domain for exception handling is at the VPE level, not the thread context level. In particular, a VPE handles asynchronous exceptions, such as interrupts, opportunistically. That is, when an asynchronous exception is raised to the VPE, the VPE selects one of the eligible (i.e., not marked as exempt from servicing asynchronous exceptions) thread contexts to execute the exception handler. Thus, although there is an architected means for a thread context to direct an asynchronous exception to a VPE, the thread context cannot specify to the VPE which thread context should handle the exception within the VPE in a MIPS MT ASE processor, i.e., the exception architecture does not provide an explicit way for the thread context to direct an asynchronous exception to a specific other thread context. This is a problem, particularly with MP operating systems, such as Linux SMP, that rely on the ability of one CPU to direct an inter-processor interrupt to another CPU in response to a timer interrupt request in order to accomplish preemptive multitasked process scheduling.
In accordance with the goal of providing high scalability of MIPS MT thread contexts, not only is the interrupt controller not replicated for each thread context, i.e., the exception domain is at the VPE level rather than at the thread context level, but other resources in a MIPS MT processor core also may not be replicated for each thread context. For example, each thread context may not have its own translation lookaside buffer (TLB) or floating point coprocessor.
Various MP operating systems have been developed to run on multiprocessor systems in which the multiple processors are MIPS architecture processors. As mentioned above, an SMP operating system running on a conventional multiprocessor system having N non-MT MIPS processors that views the system as having N CPUs will likewise view a single MIPS32 MT N-VPE microprocessor core as having N CPUs and run with little or no modifications to the SMP operating system. However, the existing MP operating systems do not have knowledge of the MIPS MT VPE/TC hierarchy, and in particular because a MIPS MT thread context is much more lightweight than a MIPS MT VPE and does not provide all the architectural state of a VPE, the existing MP operating systems do not view a MIPS MT core having M thread contexts as an M CPU system. However, it would be advantageous to enable the MP operating systems to view a MIPS MT core having M thread contexts as an M CPU system, particularly due to the highly scalable nature of MIPS MT thread contexts to a relatively large number of threads.
Therefore, what is needed is a means to have each lightweight thread context—to which is replicated less than the full architected CPU state anticipated by an existing MP operating system, such as a MIPS MT ASE thread context—appear as an architected CPU to the MP operating system, such as Linux SMP or other MP derivatives of UNIX-style operating systems.
The present invention describes modifications to existing SMP operating systems that makes highly scalable, lightweight thread contexts within a multithreaded processor that would normally by themselves be unable to run an image, or instance, of the operating system, to function as a physical CPU for the purposes of the operating system resource management.
In one aspect, the present invention provides a multiprocessing system. The system includes a multithreading microprocessor, including a plurality of thread contexts (TCs), each comprising a first control indicator for controlling whether the TC is exempt from servicing interrupt requests to an exception domain for the plurality of TCs, and a virtual processing element (VPE), comprising the exception domain, configured to receive the interrupt requests, wherein the interrupt requests are non-specific to the plurality of TCs, wherein the VPE is configured to select a non-exempt one of the plurality of TCs to service each of the interrupt requests, the VPE further comprising a second control indicator for controlling whether the VPE is enabled to select one of the plurality of TCs to service the interrupt requests. The system also includes a multiprocessor operating system (OS), configured to initially set the second control indicator to enable the VPE to service the interrupts, and further configured to schedule execution of threads on the plurality of TCs, wherein each of the threads is configured to individually disable itself from servicing the interrupts by setting the first control indicator, rather than by clearing the second control indicator.
In another aspect, the present invention provides a multiprocessing system. The system includes a multithreading microprocessor, includes a plurality of thread contexts (TCs), and a virtual processing element (VPE), configured to receive a plurality of interrupt requests, wherein the interrupt requests are non-specific to the plurality of TCs, wherein the VPE is configured to select one of the plurality of TCs to execute a thread to service an active one of the interrupt requests and to set a control bit to disable the VPE from servicing any of the interrupt requests, the VPE comprising a register for masking individual ones of the interrupt requests globally from being serviced by any of the plurality of TCs of the VPE. The system also includes a multiprocessor operating system (OS), comprising the interrupt service thread, for execution on the selected TC, configured to clear the control bit, and to write to the register to globally mask the active one of the interrupt requests, prior to clearing the control bit, in order to prevent the VPE from selecting a different one of the plurality of TCs to redundantly service the active interrupt request between clearing the control bit and clearing a source of the active interrupt request.
In another aspect, the present invention provides a multiprocessing system. The system includes a multithreading microprocessor, including a plurality of thread contexts (TCs), each comprising a first register, and a virtual processing element (VPE), comprising a second register for indicating whether an interrupt request to the VPE is active, and a third register for masking the interrupt request from being serviced, wherein the interrupt request is non-specific to the plurality of TCs, wherein the VPE is configured to select one of the plurality of TCs to execute code to service the interrupt request if active, unless masked in the third register. The system also includes a multiprocessor operating system (OS), comprising the code, wherein the code is configured to save the second register contents to the first register of the one of the plurality of TCs selected by the VPE to service the interrupt request, and to mask the interrupt request from being serviced, prior to enabling the VPE to service the interrupt request in order to prevent the VPE from selecting a different one of the plurality of TCs to redundantly service the interrupt request until the code clears a source of the interrupt request.
An advantage of the present invention is that it allows an SMP operating system, configured as if it were running on a relatively large number of symmetric CPUs, to run on a multithreaded processor, because each “CPU” is associated with a thread context that is very lightweight in terms of chip area and power consumption and therefore highly scalable. The thread contexts are lightweight because they do not each comprise the entire architectural state associated with an independent symmetric CPU; rather, the thread contexts have some architectural state replicated to each of them (such as a program counter and general purpose register set), but also share much of the architectural state between them (such as a TLB and interrupt control logic), which requires modifications to the SMP operating system to enable the number of operating system CPUs be equal to the number of thread contexts. Consequently, an existing body of coarse-grain multithreading technology embodied in SMP operating systems, such as multithreading telematics, robotics, or multimedia applications, may be exploited on such a highly scalable processor core.
For a better understanding of exception processing, translation lookaside buffer (TLB) operation, and floating point unit (FPU) coprocessor operation on MIPS architecture processors in general, the reader is referred to MIPS RISC Architecture, by Gerry Kane and Joe Heinrich, published by Prentice Hall, and to See MIPS Run, by Dominic Sweetman, published by Morgan Kaufman Publishers.
Embodiments of the present invention are described herein in the context of a processor core that includes the MIPS® MT Application-Specific Extension (ASE) to the MIPS32® Architecture; however, the present invention is not limited to a processor core with said architecture. Rather, the present invention may be implemented in any processor system which includes a plurality of thread contexts for concurrently executing a corresponding plurality of threads, but which does not include an interrupt input for each of the plurality of thread contexts that would allow one thread context to direct an inter-processor interrupt specifically to another thread context.
Referring now to
A thread context 104 comprises a collection of storage elements, such as registers or latches, and/or bits in the storage elements of the microprocessor 100 that describe the state of execution of a thread, and which enable an operating system to manage the resources of the thread context 104. That is, the thread context describes the state of its respective thread, which is unique to the thread, rather than state shared with other threads of execution executing concurrently on the microprocessor 100. A thread—also referred to herein as a thread of execution, or instruction stream—is a sequence of instructions. The microprocessor 100 is a multithreading microprocessor. That is, the microprocessor 100 is configured to concurrently execute multiple threads of execution. By storing the state of each thread in the multiple thread contexts 104, the microprocessor 100 is configured to quickly switch between threads to fetch and issue instructions. The elements of a thread context 104 of various embodiments are described below with respect to the remaining Figures. Advantageously, the present microprocessor 100 is configured to execute the MFTR instruction 300 of
The VPE context 106 includes a collection of storage elements, such as registers or latches, and/or bits in the storage elements of the microprocessor 100 that describe the state of execution of a VPE 102, which enable an operating system to manage the resources of the VPE 102, such as virtual memory, caches, exceptions, and other configuration and status information. Consequently, a microprocessor 100 with N VPEs 102 may appear to an operating system as an N-way symmetric multiprocessor. However, as also described herein, a microprocessor 100 with M thread contexts 104 may appear to an operating system as an M-way symmetric multiprocessor, such as shown with respect to
The VPEs 102 share various of the microprocessor 100 resources, such as the instruction cache 202, instruction fetcher 204, instruction decoder 206, instruction issuer 208, instruction scheduler 216, execution units 212, and data cache 242 of
In one respect, a VPE 102 may be viewed as an exception domain. That is, when an asynchronous exception (such as a hardware or software interrupt) is generated, or when an instruction of one of the thread contexts 104 of a VPE 102 generates a synchronous exception (such as an address error, bus error, or invalid instruction exception), multithreading is suspended on the VPE 102 (i.e., only instructions of the instruction stream associated with the thread context 104 servicing the exception are fetched and issued), and each VPE context 106 includes the state necessary to service the exception. Once the exception is serviced, the exception handler may selectively re-enable multithreading on the VPE 102. When an asynchronous exception such as an interrupt is raised to the VPE 102, the VPE 102 selects one of the eligible (i.e., not marked as exempt from servicing asynchronous exceptions as indicated by the IXMT bit 518 of
Referring now to
The microprocessor 100 includes a scheduler 216 for scheduling execution of the various threads being concurrently executed by the microprocessor 100. The scheduler 216 is coupled to the VMP context 108 and VPE contexts 106 of
The microprocessor 100 includes an instruction cache 202 for caching program instructions fetched from a system memory of a system including the microprocessor 100, such as the MFTR/MTTR 300/400 instructions. In one embodiment, the microprocessor 100 provides virtual memory capability, and the fetch unit 204 includes a translation lookaside buffer (TLB) for caching virtual to physical memory page translations. In one embodiment, each thread, or program, or task, executing on the microprocessor 100 is assigned a unique task ID, or address space ID (ASID), which is used to perform memory accesses and in particular memory address translations, and a thread context 104 also includes storage for an ASID associated with the thread. In one embodiment, the various threads executing on the microprocessor 100 share the instruction cache 202 and TLB, as discussed in more detail below.
The microprocessor 100 also includes a fetch unit 204, coupled to the instruction cache 202, for fetching program instructions, such as MFTR/MTTR 300/400 instructions, from the instruction cache 202 and system memory. The fetch unit 204 fetches instructions at an instruction fetch address provided by a multiplexer 244. The multiplexer 244 receives a plurality of instruction fetch addresses from the corresponding plurality of program counters 222. Each of the program counters 222 stores a current instruction fetch address for a different program thread. The embodiment of
The microprocessor 100 also includes a decode unit 206, coupled to the fetch unit 204, for decoding program instructions fetched by the fetch unit 204, such as MFTR/MTTR 300/400 instructions. The decode unit 206 decodes the opcode, operand, and other fields of the instructions. In one embodiment, the various threads executing on the microprocessor 100 share the decode unit 206.
The microprocessor 100 also includes execution units 212 for executing instructions. The execution units 212 may include but are not limited to one or more integer units for performing integer arithmetic, Boolean operations, shift operations, rotate operations, and the like; floating point units for performing floating point operations; load/store units for performing memory accesses and in particular accesses to a data cache 242 coupled to the execution units 212; and a branch resolution unit for resolving the outcome and target address of branch instructions. In one embodiment, the data cache 242 includes a translation lookaside buffer (TLB) for caching virtual to physical memory page translations, which is shared by the various thread contexts, as described in more detail below. In addition to the operands received from the data cache 242, the execution units 212 also receive operands from registers of the general purpose register sets 224. In particular, an execution unit 212 receives operands from a register set 224 of the thread context 104 allocated to the thread to which the instruction belongs. A multiplexer 248 selects operands from the appropriate register set 224 for provision to the execution units 212. In addition, the multiplexer 248 receives data from each of the other per-thread contexts 226 and program counters 222, for selective provision to the execution units 212 based on the thread context 104 of the instruction being executed by the execution unit 212. In one embodiment, the various execution units 212 may concurrently execute instructions from multiple concurrent threads.
The microprocessor 100 also includes an instruction issue unit 208, coupled to the scheduler 216 and coupled between the decode unit 206 and the execution units 212, for issuing instructions to the execution units 212 as instructed by the scheduler 216 and in response to information about the instructions decoded by the decode unit 206. In particular, the instruction issue unit 208 insures that instructions are not issued to the execution units 212 if they have data dependencies on other instructions previously issued to the execution units 212. In one embodiment, an instruction queue is imposed between the decode unit 206 and the instruction issue unit 208 for buffering instructions awaiting issue to the execution units 212 for reducing the likelihood of starvation of the execution units 212. In one embodiment, the various threads executing on the microprocessor 100 share the instruction issue unit 208.
The microprocessor 100 also includes a write-back unit 214, coupled to the execution units 212, for writing back results of instructions into the general purpose register sets 224, program counters 222, and other thread contexts 226. A demultiplexer 246 receives the instruction result from the write-back unit 214 and stores the instruction result into the appropriate register set 224, program counters 222, and other thread contexts 226 associated with the instruction's thread. The instruction results are also provided for storage into the VPE contexts 106 and the VMP context 108.
Referring now to
Bits 11-15 are an rd field 308, which specifies an rd register 322, or destination register 322, within the general purpose register set 224 of
Bits 16-20, 6-10, 5, 4, and 0-2 are an rt field 306, rx field 312, u field 314, h field 316, and sel field 318, respectively, which collectively are used to specify a source register 324 of a thread context 104 distinct from the issuing thread context, referred to herein as the target thread context 104. The use of the rt field 306, rx field 312, u field 314, h field 316, and sel field 318 to specify the source register 324 is described in detail in table 350 of
In one embodiment, the microprocessor 100 includes one or more processor control coprocessors, referred to in the MIPS PRA as Coprocessor 0, or CP0, or Cop0, denoted 602 in
The source register 324 is further specified by a TargTC operand 332. The TargTC 332 operand specifies the target thread context 104 containing the source register 324. In one embodiment, the TargTC operand 332 is stored in the VPEControl Register 504 of
Referring now to
Referring now to
The EVP bit 513 of
As discussed above, TargTC field 332 of
The TCU0 . . . TCU3 bits 581 of the TCStatus Register 508 of
The TCContext Register 595 of
The RNST bits 582 of the TCStatus Register 508 indicate the state of the thread context 104, namely whether the thread context 104 is running or blocked, and if blocked the reason for blockage. The RNST 582 value is only stable when read by an MFTR instruction 300 if the target thread context 104 is in a halted state, which is described below; otherwise, the RNST 582 value may change asynchronously and unpredictably. When a thread context 104 is in the running state, the microprocessor 100 will fetch and issue instructions from the thread of execution specified by the thread context 104 program counter 222 according to the scheduler 216 scheduling policy.
Independently of whether a thread context 104 is free or activated, a thread context 104 may be halted if the H bit 599 of the TCHalt Register 509 of
When a thread context 104 is in a halted state, the TCRestart Register 594 may be read to obtain the address 549 of the instruction at which the microprocessor 100 will resume execution of the thread context 104 when the thread context 104 is restarted. In the case of branch and jump instructions with architectural branch delay slots, the restart address 549 will advance beyond the address of the branch or jump instruction only after the instruction in the delay slot has been retired. If the thread context 104 is halted between the execution of a branch instruction and the associated delay slot instruction, the branch delay slot is indicated by the TDS bit 584 of the TCStatus Register 508.
Conversely, the TCRestart register 594 can be written while its thread context 104 is halted to change the address at which the thread context 104 will restart. Furthermore, a first thread context 104 running an operating system thread may restart a second thread context 104 by writing a 0 to the H bit 599 of the TCHalt Register 509 of the second thread context 104. Clearing the H bit 599 of an activated thread context 104 allows the thread context 104 to be scheduled, and begin fetching and executing instructions at its restart address 549 specified in its TCRestart register 594.
In the MIPS PRA, the Coprocessor 0 EPC Register 598 of
In a MIPS MT ASE microprocessor 100, the EPC Register 598 is instantiated for each VPE 102 in the microprocessor 100. When an exception is raised to a VPE 102, the VPE 102 selects one of its thread contexts 104 to service the exception. All thread contexts 104 of the VPE 102, other than the thread context 104 selected to service the exception, are stopped and suspended until the EXL bit 576 and ERL bit 575 of the Status Register 571 are cleared. When a synchronous exception is raised due to the execution of an instruction contained in a thread of execution, the microprocessor 100 selects the thread context 104 running the thread containing the offending instruction to service the exception. That is, the general purpose registers 224, program counter 222, and other per-thread context 226 of the offending thread context 104 are used to service the synchronous exception. When an asynchronous exception is raised, such as an interrupt, the microprocessor 100 selects one of the eligible thread contexts 104 bound to the VPE 102 to service the asynchronous exception. The VPE 102 to which a thread context 104 is bound (as indicated by the CurVPE field 558 of the TCBind Register 556) is the exception domain for the thread context 104. In particular, a VPE 102 selects a thread context 104 bound to it, i.e., within its exception domain, to service an exception. Additionally, a thread context 104 utilizes the resources related to handling exceptions (such as the Coprocessor 0 EPC Register 598 and Status Register 571) of the exception domain, or VPE 102, to which the thread context 104 is bound when servicing an exception. The method for choosing the eligible thread context 104 to service an asynchronous exception is implementation-dependent and may be adapted to satisfy the particular application in which the microprocessor 100 is employed. However, as discussed herein, the MIPS MT ASE does not provide the capability for the asynchronous exception to specify which of the thread contexts 104 must service the asynchronous exception. The microprocessor 100 saves the restart address of the thread context 104 selected to service the exception in the EPC Register 598 of the VPE 102 to which the selected thread context 104 is bound. Additionally, a thread context 104 may be made ineligible for being selected to service an asynchronous exception by setting the IXMT bit 518 in its TCStatus Register 508.
In one embodiment, the program counter 222 of
The Coprocessor 0 Status Register 571 of
Referring now to
Referring now to
Referring now to
At block 802, the instruction issuer 208 of
At decision block 803, the execution unit 212 examines the TKSU bits 589 of the TCStatus Register 508 to determine whether the privilege level of the issuing thread context 104 is at kernel privilege level. If so, flow proceeds to decision block 804; otherwise, flow proceeds to block 805.
At block 805, the execution unit 212 raises an exception to the MFTR instruction 300 since the issuing thread context 104 does not have sufficient privilege level to execute the MFTR instruction 300. Flow ends at block 805.
At decision block 804, the execution unit 212 determines whether the target thread context 104 is halted by examining the value of the H bit 599 of the TCHalt Register 509 of
At decision block 808, the execution unit 212 examines the TargTC 332 value of the issuing VPE 102 VPEControl Register 504 to determine whether the TargTC 332 value is valid. In one embodiment, the TargTC 332 value is not valid if the issuing VPE is not the master VPE 102, as indicated by a clear value in the MVP bit 553 of the VPEConf0 Register 505 of
At decision block 812, the execution unit 212 examines the TCU bits 581 in the TCStatus Register 508 of
At decision block 814, the execution unit 212 determines whether the source register 324 specified by the MFTR instruction 300 is instantiated. If so, flow proceeds to block 824; otherwise, flow proceeds to block 816.
At block 816, the results of the MFTR instruction 300 are invalid. That is, the microprocessor 100 attempts to perform block 824; however, the source, destination, and values of the data transfer are invalid. Flow ends at block 816.
At block 824, the execution unit 212 copies the contents of the source register 324 of the target thread context 104 to the destination register 322 of the issuing thread context 104. In one embodiment, the microprocessor 100, after reading the source register 324, updates the source register 324 with an update value. In one embodiment, the read/update is performed atomically. In one embodiment, the update value is provided in the GPR 224 specified by the rd field 308 in the MFTR instruction 300. Flow ends at block 824.
Referring now to
At block 902, the instruction issuer 208 of
At decision block 903, the execution unit 212 examines the TKSU bits 589 of the TCStatus Register 508 to determine whether the privilege level of the issuing thread context 104 is at kernel privilege level. If so, flow proceeds to decision block 904; otherwise, flow proceeds to block 905.
At block 905, the execution unit 212 raises an exception to the MTTR instruction 400 since the issuing thread context 104 does not have sufficient privilege level to execute the MTTR instruction 400. Flow ends at block 905.
At decision block 904, the execution unit 212 determines whether the target thread context 104 is halted by examining the value of the H bit 599 of the TCHalt Register 509 of
At decision block 908, the execution unit 212 examines the TargTC 332 value of the issuing VPE 102 VPEControl Register 504 to determine whether the TargTC 332 value is valid. In one embodiment, the TargTC 332 value is not valid if the issuing VPE is not the master VPE 102, as indicated by a clear value in the MVP bit 553 of the VPEConf0 Register 505 of
At decision block 912, the execution unit 212 examines the TCU bits 581 in the TCStatus Register 508 of
At decision block 914, the execution unit 212 determines whether the destination register 422 specified by the MTTR instruction 400 is instantiated. If so, flow proceeds to block 924; otherwise, flow proceeds to block 916.
At block 916, the microprocessor 100 performs no operation because there is no valid destination register to which the source data may be written. Flow ends at block 916.
At block 924, the execution unit 212 copies the contents of the source register 424 of the issuing thread context 104 to the destination register 422 of the target thread context 104. Flow ends at block 924.
Referring now to
Within the flowchart, reference is made to a thread A running on a thread context A 104 and a thread B running on a thread context B 104. Thread A running on thread context A 104 directs a software-emulated inter-processor interrupt (IPI) to thread context B 104, by employing MFTR instructions 300 and MTTR instructions 400. In the example of the flowchart, thread context A 104 and thread context B 104 are bound to the same VPE 102. Although the flowchart of
As described above, when an asynchronous hardware interrupt (such as a periodic timer interrupt used for operating system task scheduling purposes) is requested in a MIPS MT ASE processor, the VPE 102 that received the hardware interrupt request selects an eligible thread context (in this example, thread context A 104) to handle the exception. In the MIPS architecture, when a hardware interrupt request is made, control is transferred to a general exception vector of the operating system. The general exception vector decodes the cause of the exception and invokes the appropriate interrupt request handler (in this example, thread A), such as the timer handler.
The Linux SMP kernel for the MIPS architecture assumes that every processor, or CPU, in the SMP system will get a periodic interrupt, and divides the work performed by the timer interrupt handler into a local clock interrupt function that executes on all CPUs, and a system clock interrupt function that executes only on one CPU of the SMP system. In the MIPS processor architecture, each VPE 102 includes one timer in Coprocessor 0 shared by all thread contexts 104 bound to the VPE 102 (see the Count/Compare register pairs described in MIPS32® Architecture for Programmers Volume III: The MIPS32® Privileged Resource Architecture, Document Number: MD00090, Revision 2.50, Jul. 1, 2005). In one embodiment of SMTC Linux, only one of the timers of one of the VPEs 102 is invoked as the single timer for all CPUs of the SMP system. In another embodiment, the timer of each of the VPEs 102 is invoked for all CPUs of that VPE 102. The thread context 104 selected to service the asynchronous timer interrupt executes the system clock interrupt function and then broadcasts, or directs, an IPI to all the other thread contexts 104 of the VPE 102. The directed IPI is a local clock interrupt type IPI which instructs the receiving thread contexts 104 to execute only the local clock interrupt function. Although the SMTC Linux timer interrupt handler directs an IPI message to each thread context 104 known to the operating system as a processor, the flowchart of
At block 1002, at source code line 38, thread A running on thread context A 104 halts thread B running on thread context B 104 by executing an MTTR instruction 400 instruction to clear the H bit 599 of the TCHalt Register 509 of
At block 1004, at lines 95-100 (via the call the post_direct_ipi at line 64), thread A creates a new stack frame on the kernel stack of thread context B 104. In one embodiment, the new stack frame is effectively created by the assignment of a value to the kernel stack pointer of thread context B 104, and storing values on the new stack frame comprises storing values at predetermined offsets from the kernel stack pointer value. It is also noted that if the target thread context 104 is exempted from taking interrupts (as indicated by a set IXMT bit 518 of
At block 1006, at line 82, thread A reads the TCStatus Register 508 of thread context B 104 via the function read_tc_c0_tcstatus, which includes an MFTR instruction 300. The TCStatus Register 508 includes the thread context B 104 execution privilege level and interrupt exemption status, among other things. Thread A, at line 104, also saves the TCStatus Register 508 value to the stack frame created at block 1004. Flow proceeds to block 1008.
At block 1008, at line 83, thread A reads the restart address 549 of thread B from TCRestart register 594 of thread context B 104 via the function read_tc_c0_tcrestart, which includes an MFTR instruction 300. Thread A, at line 102, also saves the restart address 549 to the stack frame created at block 1004. Flow proceeds to block 1012.
At block 1012, at lines 106 and 108, thread A saves the address of the operating system IPI handler and a reference to an IPI message on the stack frame created at block 1004. In the embodiment of the source code listing, advantageously, the code manipulates the target thread context B 104 and stack frame such that a common IPI handler may be invoked to support SMTC operation. The common IPI handler is invoked to handle both software emulated interrupts described herein and actual hardware interrupts, i.e., interrupts for which target thread context 104 B is the thread context 104 selected by the VPE 102 to handle the hardware interrupt request, such as may be invoked at block 1114 of
At block 1014, at lines 110-112, thread A writes the TCStatus Register 508 of thread context B 104 via the function the function write_tc_c0_tcstatus, which includes an MTTR instruction 400, to set the execution privilege level of thread context B 104 to kernel mode and disables, or exempts, thread context B 104 from receiving interrupts. Conceptually, thread A would set the EXL bit 576 in Coprocessor 0 Status Register 571 in order to emulate an exception. However, when EXL 576 is set, multithreading is disables on the VPE 102, i.e., only one thread context 104 is allowed to run when EXL 576 is set. And thread A needs thread context B 104 to run when unhalted below at block 1018. Therefore, the setting of EXL 576 must be left up to thread context B 104 by smtc_ipi_vector at block 1022 below. Thus, until then, thread A temporarily accomplishes a similar effect to setting EXL 576 by setting IXMT 518 and TKSU 589 to kernel mode in the thread context B 104 TCStatus Register 508. Flow proceeds to block 1016.
At block 1016, at line 115, thread A writes the restart address 549 of thread B in the TCRestart register 594 of thread context B 104 via the function the function write_tc_c0_tcrestart, which includes an MTTR instruction 400, with the address of smtc_ipi vector. Flow proceeds to block 1018.
At block 1018, at line 65, thread A un-halts, or restarts, thread context B 104 to cause smtc_ipi_vector to begin running on thread context B 104. Flow proceeds to block 1022.
At block 1022, at lines 163-165, the smtc_ipi_vector sets EXL 576, which has the effect of disabling interrupts and setting the execution privilege level to kernel mode for all thread contexts 104 bound to the VPE 102. It is noted that at line 160 the smtc_ipi_vector disables multithreading on the VPE 102 before setting EXL 576. Additionally, if multithreading was enabled prior to line 160, the code restores multithreading at lines 168-170. It is also noted that if thread context B 104 was in user mode when halted at block 1002, the smtc_ipi_vector sets the CU0 bit 572 of the Status Register 571. Flow proceeds to block 1024.
At block 1024, at lines 196 and 198, the smtc_ipi_vector restores the thread context B 104 pre-halted TCStatus Register 508 value that was saved at block 1006, and in particular restores its execution privilege level and interrupt exemption state. Flow proceeds to block 1026.
At block 1026, at lines 200-201, the smtc_ipi_vector loads the EPC Register 598 with the thread context B 104 pre-halted TCRestart register 594 value saved at block 1008. Consequently, when the standard Linux SMP return from interrupt code subsequently executes an ERET instruction at block 1036, thread B will be restarted on thread context B 104 at the address at which it was halted at block 1002. Thus, by setting EXL 576 at block 1022 and populating the EPC Register 598 at block 1026, the smtc_ipi_vector effectively emulates what the microprocessor 100 hardware would do if thread context B 104 had been selected to service the asynchronous interrupt (rather than thread context A 104). Flow proceeds to block 1028.
At block 1028, at line 203, the smtc_ipi_vector saves all of the general purpose registers 224 to the stack frame created at block 1004. Flow proceeds to block 1032.
At block 1032, at line 204 via the CLI macro, the smtc_ipi_vector sets itself to kernel mode execution privilege level and exempts itself from servicing interrupts. It is noted that this is performed only for thread context B 104, not for the entire VPE 102. It is noted that the CLI macro is a standard Linux macro which is modified to support SMTC by setting kernel mode execution privilege level and exempting from interrupt servicing (via the IXMT bit 518) only the invoking thread context 104, rather than the entire VPE 102 (as the non-SMTC code does by clearing the IE bit 577 of the Status Register 571 of
At block 1034, at lines 205-210, the smtc_ipi_vector calls the common IPI handler (which is ipi_decode, as populated at line 108) with the IPI message reference saved on the stack frame at block 1012 as an argument. Flow proceeds to block 1036.
At block 1036, at line 212, after the operating system IPI handler returns, the smtc_ipi_vector jumps to the standard operating system return from interrupt code (which in Linux SMP is ret_from_irq), which eventually executes an ERET instruction to return execution on thread context B 104 to thread B with its pre-halted execution privilege level and interrupt exemption state. Prior to executing the ERET instruction, the return from interrupt code restores the EPC Register 598 with the restart address value saved at block 1008 and restores the Status Register 571 KSU bits 574 with the value saved at block 1006. Flow ends at block 1036.
Referring now to
At block 1102, a timer generates an interrupt request to a VPE 102, which are the exception domains of the microprocessor 100. In one embodiment, the timer interrupt request is an asynchronous hardware interrupt generated by the MIPS PRA Count/Compare register pairs of one of the VPEs 102 of microprocessor 100, and the Count/Compare register pairs of the other VPEs 102 are all disabled. Flow proceeds to block 1104.
At block 1104, the interrupted VPE 102 selects an eligible thread context 104 bound to itself to service the timer interrupt request. As described above, in the MIPS MT ASE, a thread context 104 is eligible if its IXMT bit 518 is clear and the curVPE field 558 of the TCBind Register 556 of
At block 1106, the VPE 102 suspends execution of the threads executing on all thread contexts 104 bound to the VPE 102 except for the thread context 104 selected at block 1104. In particular, the VPE 102 ceases to issue instructions to the execution pipeline of the threads. Flow proceeds to block 1108.
At block 1108, the VPE 102 saves the restart address of the selected thread context 104 into the EPC Register 598, sets the EXL bit 576 of the Status Register 571, and populates the MIPS PRA Cause register 536, all of the VPE's 102 Coprocessor 0 VPE context 106. Flow proceeds to block 1112.
At block 1112, the VPE 102 causes the selected thread context 104 to execute a general exception handler at the general exception vector according to the MIPS PRA. The general exception handler decodes the cause of the exception via the MIPS PRA Cause register 536 and Status Register 571 and determines the exception was an asynchronous hardware interrupt generated by the timer. Consequently, the general exception handler calls the timer interrupt service routine, which among other functions, schedules processes according to the preemptive multitasking algorithm of the operating system. In one embodiment, the timer interrupt routine may call a separate routine dedicated to scheduling processes. Flow proceeds to block 1114.
At block 1114, the timer interrupt service routine determines whether a new process, or task, should be scheduled on the selected thread context 104 according to the SMP OS multitasking scheduling algorithm. If so, the timer interrupt service routine schedules a new process to run on the selected thread context 104; otherwise, the timer interrupt service routine leaves the current process to run on the selected thread context 104. It is noted that a thread and a process herein are not necessarily synonymous. A process is an entity managed by the SMP operating system, and typically comprises entire programs, such as application programs or portions of the operating system itself; whereas a thread is simply a stream of instructions, which of course may be a stream of instructions of an operating system process, or task. Flow proceeds to block 1116.
At block 1116, the timer interrupt service routine issues a software-emulated inter-processor interrupt to each other thread context 104 in the microprocessor 100, according to
At block 1118, the timer interrupt service routine calls the operating system return from interrupt code, which executes an ERET instruction. If a new process was scheduled to run at block 114, then the ERET causes the newly scheduled process to run; otherwise, the ERET causes the process that was interrupted by the timer interrupt request to continue running. Flow proceeds to block 1122.
At block 1122, each thread context 104 that was the target of a software-emulated inter-processor interrupt performed at block 1116 eventually calls the inter-processor interrupt service routine, according to block 1034 of
As may be observed from
Referring now to
An ASID is an address space identifier, which identifies a unique memory map. A memory map comprises a mapping, or association, or binding, between a virtual address space and a set of physical page addresses. Most commonly, the operating system creates a memory map when it creates a new process, or task. Each process created by the operating system has a memory map. Additionally, the operating system has its own memory map. Multiple processes may share a memory map. Consequently, two CPUs using a shared memory map will result in the same virtual address accessing the same physical memory, or generating identical page fault exceptions. An example in a UNIX-like operating system of two processes sharing a memory map is when a process makes a fork( ) system call (not to be confused with the MIPS MT ASE FORK instruction). In this case, a new process is created which shares the same memory map as its parent process until such time as one of the processes performs a store to memory which would change the contents of the memory. Additionally, and perhaps more commonly, a multithreaded process may have multiple threads running in the same address space using the same memory map. Still further, multiple processes may specifically designate particular memory pages that they share.
In some embodiments, a memory map comprises a simple contiguous array of page table entries, with each entry specifying a virtual to physical page address translation and other relevant page attribute information. However, because a linear page table may require a significant amount of contiguous memory per process (such as in an embedded application with relatively small pages such as 4 KB pages with a relatively large address space), other memory map schemes may be employed. For example, a multi-level page/segment table structure may be employed in which a memory map is described by a segment table which in turn points to a set of page table entries, some of which (in particular, those which correspond to unpopulated parts of the address space) may be common to multiple memory maps.
The ASID cache 1204 is a kernel variable maintained in the system memory for each of the CPUs. The operating system uses the ASID cache 1204 to assign a new ASID to a newly created memory map, or to assign a new ASID for the respective CPU to an existing memory map that was previously used on another CPU. The operating system initializes each ASID cache 1204 value to zero. Each time the instance of the operating system executing on a respective CPU assigns a new ASID value from the ASID cache 1204, the operating system monotonically increments the ASID cache 1204 value of the respective CPU. This process continues until the ASID cache 1204 value wraps back to zero and the cycle continues.
Generally speaking, the TLB 1202 is a small cache memory in which each entry includes a tag portion and a data portion. The tag portion includes a virtual page address, or virtual page number (VPN), portion that is concatenated with an ASID portion. When the CPU generates a virtual memory address to make a memory access, such as when a load or store instruction is executed, the virtual memory address is concatenated with the ASID of the process making the memory access, and the result is compared with the TLB 1202 tags to see if a match occurs. The ASID of the process making the memory access is supplied by the ASID field 538 of the EntryHi Register 526 of
In the prior art system 1200 of
Referring now to
The system 100 of
The system 100 of
Each of the CPUs in the system 1200 of
Referring now to
Referring now to
At block 1502, the operating system begins its initialization sequence. Flow proceeds to block 1504.
At block 1504, the initialization sequence invokes the SMP Linux cpu_probe( ) routine only for TC 0104, which corresponds to SMTC Linux CPU number 0 (the primary, or boot, CPU/TC 104), in order to populate the cpu_data array entry 1408 at index 0. Flow proceeds to block 1506.
At block 1506, the initialization sequence copies the cpu_data array entry 1408 at index 0 to all the other entries in the cpu_data array, i.e., to the entry for each of the other CPUs/TCs 104. Flow proceeds to block 1508.
At block 1508, the initialization sequence updates the TC_ID field 1404 and VPE_ID field 1406 of the cpu_data array entry 1408 for each of the CPUs/TCs 104 based on their CurTC field 557 and CurVPE field 558 values, respectively. It is noted that prior to the step at block 1508, the binding of thread contexts 104 to VPEs 102 has been performed, i.e., the CurVPE field 558 for each thread context 104 has been populated. In one embodiment, the operating system performs the binding of thread contexts 104 to VPEs 102. In another embodiment, the binding of thread contexts 104 to VPEs 102 may be performed when the microprocessor 100 is synthesized or fabricated. Additionally, the initialization sequence updates the cpu_data array entry 1408 for each of the CPUs/TCs 104 to indicate whether it has permission to access the FPU 1306. The TCU1 bit 581 of the TCStatus Register 508 of
At block 1512, the initialization sequence invokes the per_cpu_trap_init( ) routine only for one thread context 104 for each VPE 102 since the VPE 102 is the exception domain for the thread contexts 104 bound to it; that is, each thread context 104 is not its own exception domain, particularly since asynchronous exceptions may not be directed specifically to a particular thread context 104, as discussed above. This is in contrast to conventional SMP Linux in which the per_cpu_trap_init( ) routine is invoked once per CPU, since each CPU in the conventional system 1200 is an exception domain. Flow ends at block 1512.
Referring now to
At block 1602, a thread executing on one of the thread contexts 104 includes a floating point instruction. However, the thread context 104 does not have permission to access the FPU 1306. Therefore, a floating point exception is taken so that a floating point instruction emulation may be performed. Flow proceeds to block 1604.
At block 1604, the operating system increments a count associated with the thread for which the floating point emulation was performed. Flow proceeds to decision block 1606.
At decision block 1606, the operating system determines whether the count has exceeded a threshold parameter. If not, flow ends; otherwise, flow proceeds to block 1608.
At block 1608, the operating system sets a cpus_allowed mask, which is a kernel variable, to cause the operating system to schedule the thread on a thread context 104 that has permission to access the FPU 1306 during a subsequent time slice. A time slice is a time quantum used by the operating system to schedule processes or threads and is typically an integer multiple of the timer interrupt time quantum. Flow ends at block 1608.
Flow of the second flowchart of
At block 1612, a time slice of the operating system completes and the operating system performs its thread scheduling. Flow proceeds to decision block 1614.
At decision block 1614, for each running thread, the operating system determines whether the thread executed any floating point instructions during the time slice. In one embodiment, the thread has not executed any floating point instructions during the time slice if the CU1 bit 572 in the Status Register 571 of
At block 1616, the operating system clears the cpus_allowed mask to enable the operating system to schedule the thread on a thread context 104 that does not have permission to access the FPU 1306 during a subsequent time slice. Flow ends at block 1616.
Advantageously, the method described in the flowcharts of
Referring now to
At block 1702, the operating system begins its initialization sequence. Flow proceeds to block 1704.
At block 1704, the operating system sets the IE bit 577 in the Status Register 571 of
Flow of the second flowchart of
At block 1712, a thread executing on a thread context 104 invokes an interrupt disable routine, such as the CLI macro at source code lines 215-250, on a CPU/TC 104 executing the thread. Flow proceeds to block 1714.
At block 1714, the interrupt disable routine sets the IXMT bit 518 of the TCStatus Register 508 of
Flow of the third flowchart of
At block 1722, a thread executing on a thread context 104 invokes an interrupt enable routine, for example a Linux STI macro, on a CPU/TC 104 executing the thread. Flow proceeds to block 1724.
At block 1724, the interrupt enable routine clears the IXMT bit 518 of the TCStatus Register 508 of
Referring now to
At block 1802, an interrupt request is activated. In response, the VPE 102 receiving the interrupt request sets the EXL bit 576 in the Status Register 571 of
At block 1804, the interrupt vector code saves the contents of the Cause Register 536 of
At block 1806, the interrupt vector code masks off the currently active interrupt sources indicated in the Cause Register 536 by setting the corresponding IM bits 573 in the Status Register 571 of
At block 1808, the interrupt vector code clears the EXL bit 576, which ceases to disable the VPE 102 from taking interrupts which were activated at block 1802. Flow proceeds to block 1812.
At block 1812, the interrupt vector code decodes the interrupt sources based on the Cause Register 536 contents and transfers control to the appropriate interrupt handlers registered to handle interrupts for the specific types of active interrupt sources. Flow proceeds to block 1814.
At block 1814, the interrupt source-specific interrupt handler clears the interrupt source and services the interrupt source. Flow proceeds to block 1816.
At block 1816, the interrupt handler invokes the common return from interrupt code to restore the context and return from the interrupt. Flow proceeds to block 1818.
At block 1818, the return from interrupt code reads the TCContext Register 595 and unmasks the interrupt sources indicated therein as previously having been inactive by clearing the corresponding IM bits 573 in the Status Register 571. Flow ends at block 1818.
It is noted that a kernel variable in memory could be used instead of the TCContext Register 595 to save the Cause Register 536 contents. However, using the TCContext Register 595 is more efficient, and is particularly appropriate in an embodiment in which the value must be saved and restored on a context switch.
In addition to the modifications described in
Referring now to
At block 1902, the VPE 102 invokes the operating system TLB miss handler in response to a TLB miss exception. It is noted that in a MIPS Architecture processor, the operating system is responsible for handling TLB 1302 misses. That is, the operating system is responsible for updating the TLB 1302 with the appropriate virtual to physical page translation information if the information is missing in the TLB 1302. This is in contrast to some processor architectures in which the processor hardware automatically fills the TLB on a TLB miss. Flow proceeds to block 1904.
At block 1904, the TLB miss handler reads the TCBind Register 556 of
Flow of the second flowchart of
At block 1912, a thread invokes the operating system get_kernel_sp( ) or set_kernel_sp( ) routine to get or set, respectively, the kernel stack pointer value for the CPU/TC 104 executing the thread. Flow proceeds to block 1914.
At block 1914, the invoked routine reads the TCBind Register 556 of
It is noted that conventional SMP Linux for MIPS uses the PTEBase field 542 of the Coprocessor 0 Context Register 527 of
Referring now to
At block 2002, a thread executing on a CPU/TC 104 invokes a TLB 1302 maintenance routine. Flow proceeds to block 2004.
At block 2004, the routine disables interrupts. In one embodiment, the routine disables interrupts only on the executing thread context 104, such as via a CLI described above. In another embodiment, the routine disables interrupts on the entire VPE 102 to which the thread context 104 is bound by clearing the IE bit 577 of the Status Register 571 of
At block 2006, the routine inhibits multi-VPE 102 operation, i.e., inhibits concurrent execution of threads other than the thread executing the routine. That is, the routine prevents the instruction scheduler 216 from dispatching to the execution units 212 instructions from any of the VPEs 102 of the system 100 other than the VPE 102 to which the thread context 104 executing the routine is bound and from dispatching from any of the thread contexts 104 bound to the VPE 102 except the thread context 104 executing the routine. In one embodiment, the routine executes a MIPS MT ASE DVPE instruction to disable multi-VPE operation. Flow proceeds to block 2008.
At block 2008, the routine performs the specified TLB 1302 maintenance required by the TLB 1302 maintenance routine. Flow proceeds to block 2012.
At block 2012, the routine restores the multi-VPE operation state that existed on the system 100 prior to performing the step at block 2006. In one embodiment, the routine executes a MIPS MT ASE EVPE instruction to enable multi-VPE operation if that was the previous state. Flow proceeds to block 2014.
At block 2014, the routine restores the interrupt enable state that existed on the VPE 102 prior to performing the step at block 2004. In one embodiment, the routine clears the IXMT bit 518 in the TCStatus Register 508 of
Referring now to
As mentioned above, in a conventional MIPS SMP Linux system 1200 of
To ensure that the same ASID is not assigned to two different memory maps concurrently in use on two different CPUs/TCs 104, SMTC Linux shares a common ASID cache 1304 across all CPUs/TCs 104, and serializes use and update of the shared ASID cache 1304 by suspending thread scheduling during the read-modify-write operation of the ASID cache 1304 that is performed when obtaining a new ASID value from the ASID cache 1304. Flow begins at block 2102.
At block 2102, a thread executing on a thread context 104 requires a new ASID for a memory map for a particular CPU/TC 104. The most common situations in which a new ASID is required for a memory map are when a new memory map is being created or when an ASID generation rollover occurs, as described below. In particular, a thread is being scheduled to run on a CPU/TC 104, i.e., the thread is being swapped in to the CPU/TC 104 by the operating system. Among other things, the operating system loads the general purpose registers 224 of
At block 2104, the operating system gains exclusive access to the shared ASID cache 1304. In one embodiment, the step at block 2104 is performed by disabling interrupts and disabling multi-VPE operation as described with respect to blocks 2004 and 2006, respectively, of
At block 2106, the operating system increments the current ASID cache 1304 value to obtain the new ASID value. An example of the step performed at block 2106 is found at lines 282 and 285 of the source code listing. Flow proceeds to decision block 2108.
At decision block 2108, the operating system determines whether the ASID cache 1304 value rolled over to a new generation when it was incremented at block 2106. The ASID cache 1304 rolls over to a new generation as follows. The ASID cache 1304 value is maintained as a 32-bit value. However, the TASID bits 528 of the TCStatus Register 508 of
At block 2112, the operating system updates a live ASID table. In addition, when an ASID generation rollover occurs, the operating system updates the new ASID to the first ASID generation value and flushes the shared TLB 1302. A live ASID is an ASID that is in use by another CPU/TC 104. The live ASID table indicates, for each ASID, which CPUs/TCs 104, if any, are currently using the ASID. The operating system updates the live ASID table by reading the TASID field 528 of the TCStatus Register 508 of
At decision block 2114, the operating system determines whether the new ASID is equal to a live ASID. An example of the step performed at decision block 2114 is found at line 313 of the source code listing. If the new ASID is equal to a live ASID, flow returns to block 2106 so that the operating system can attempt to obtain a new non-live ASID; otherwise, flow proceeds to block 2116.
At block 2116, the operating system assigns the new ASID to the memory map for all CPUs/TCs 104 in the system 100. As discussed above, in one embodiment, SMTC Linux uses the asid_cache storage space in the original fields 1402 effectively as a single ASID cache 1304 by updating each asid_cache field in each cpu_data array entry 1408 even when generating a new ASID value for only a single CPU/TC 104; however, other embodiments are contemplated in which a single kernel variable is used to store the single ASID cache 1304. The operating system advantageously assigns the new ASID to the memory map for all CPUs/TCs 104 in order to make more efficient use of the shared TLB 1302, i.e., to avoid the following situation. Assume two processes share a common memory map and execute on different CPUs/TCs 104. In a conventional SMP Linux system 1200, the memory map would be assigned a different ASID for each CPU, since each CPU has its own ASID cache 1204. However, in the shared TLB 1302 system 100, the first time each CPU/TC 104 accessed a shared memory page, the operating system would allocate an entry in the shared TLB 1302 for the page translation since the ASID value differed for each CPU/TC 104, i.e., two TLB 1302 entries would be consumed for the same shared physical page, which would be an inefficient use of the shared TLB 1302 entries. A similar inefficiency could occur when a process was migrated from one CPU/TC 104 to another. Thus, to avoid this situation and make more efficient use of the shared TLB 1302, SMTC Linux assigns the new ASID to the memory map not only for the CPU/TC 104 for which it was obtained, but also causes the new ASID to be assigned to and used by all CPUs/TCs 104 that reference the memory map. Stated alternatively, when the operating system assigns a new ASID to a memory map, if a process uses the memory map, then all threads of the process which use the memory map use the new ASID on all CPUs/TCs 104 that execute the threads. In particular, when a thread using a memory map is swapped into any thread context 104 after a new ASID is assigned to the memory map, the new ASID, rather than an old ASID identifying the memory map, gets loaded into the TASID field 528 of the TCStatus Register 508 of
At block 2118, the operating system relinquishes exclusive access to the shared ASID cache 1304. In one embodiment, the step at block 2118 is performed by restoring interrupts and multi-VPE operation to their previous states, as described with respect to blocks 2012 and 2014, respectively, of
Although the present invention and its objects, features, and advantages have been described in detail, other embodiments are encompassed by the invention. For example, although embodiments have been described in which the modified SMP OS is Linux, other SMP operating systems are contemplated for adaptation to run on a multithreading microprocessor having non-independent lightweight thread contexts that share processor state with one another, such as MIPS MT ASE thread contexts, each of which is an independent CPU to the SMP OS. For example, other variants of the UNIX operating system, such as SUN Solaris, HP UX, Mac OS X, Open VMS, and others may be adapted to view the thread contexts as a CPU. Still further, other SMP operating systems such as SMP-capable variants of the Microsoft Windows operating system may be adapted to view the thread contexts as a CPU. Furthermore, although the invention has been described with respect to modifications to an existing SMP operating system, the invention is not limited to existing operating systems, but rather new operating systems may be developed which employ the steps described to employ non-independent lightweight thread contexts that share processor state with one another, such as MIPS MT ASE thread contexts, as independent CPUs to the new SMP OS.
While various embodiments of the present invention have been described above, it should be understood that they have been presented by way of example, and not limitation. It will be apparent to persons skilled in the relevant computer arts that various changes in form and detail can be made therein without departing from the scope of the invention. For example, in addition to using hardware (e.g., within or coupled to a Central Processing Unit (“CPU”), microprocessor, microcontroller, digital signal processor, processor core, System on Chip (“SOC”), or any other device), implementations may also be embodied in software (e.g., computer readable code, program code, instructions and/or data disposed in any form, such as source, object or machine language) disposed, for example, in a computer usable (e.g., readable) medium configured to store the software. Such software can enable, for example, the function, fabrication, modeling, simulation, description and/or testing of the apparatus and methods described herein. For example, this can be accomplished through the use of general programming languages (e.g., C, C++), GDSII databases, hardware description languages (HDL) including Verilog HDL, VHDL, and so on, or other available programs and databases. Such software can be disposed in any known computer usable medium such as semiconductor, magnetic disk, or optical disc (e.g., CD-ROM, DVD-ROM, etc.). The software can also be disposed as a computer data signal embodied in a computer usable (e.g., readable) transmission medium (e.g., carrier wave or any other medium including digital, optical, or analog-based medium). Embodiments of the present invention may include methods of providing operating system software described herein by providing the software and subsequently transmitting the software as a computer data signal over a communication network including the Internet and intranets, such as shown in
This application is a continuation-in-part (CIP) of the following co-pending Non-Provisional U.S. patent applications, which are hereby incorporated by reference in their entirety for all purposes: Serial No.(Docket No.)Filing DateTitle11/313,272Dec. 20, 2005SOFTWARE EMULATION OF(MIPS.0214-00-US)DIRECTED EXCEPTIONS IN AMULTITHREADINGPROCESSOR11/313,296Dec. 20, 2005PREEMPTIVE MULTITASKING(MIPS.0214-01-US)EMPLOYING SOFTWAREEMULATION OF DIRECTEDEXCEPTIONS IN A MULTI-THREADING PROCESSOR Each of the two above co-pending Non-Provisional U.S. patent applications is a continuation-in-part (CIP) of the following co-pending Non-Provisional U.S. patent application, which is hereby incorporated by reference in its entirety for all purposes: Serial No.(Docket No.)Filing DateTitle10/929,097Aug. 27, 2005APPARATUS, METHOD, AND(MIPS.0194-00-US)INSTRUCTION FOR SOFT-WARE MANAGEMENT OFMULTIPLE COMPUTATIONALCONTEXTS IN A MULTI-THREADED MICRO-PROCESSOR Co-pending Non-Provisional U.S. patent application Ser. No. 10/929,097 (MIPS.0194-00-US) is a continuation-in-part (CIP) of the following co-pending Non-Provisional U.S. patent applications, which are hereby incorporated by reference in their entirety for all purposes: Serial No.(Docket No.)Filing DateTitle10/684,350Oct. 10, 2003MECHANISMS FOR ASSURING(MIPS.0188-01-US)QUALITY OF SERVICE FORPROGRAMS EXECUTING ON AMULTITHREADEDPROCESSOR10/684,348Oct. 10, 2003INTEGRATED MECHANISM(MIPS.0189-00-US)FOR SUSPENSION ANDDEALLOCATION OFCOMPUTATIONAL THREADSOF EXECUTION IN APROCESSOR Each of co-pending Non-Provisional U.S. patent application Ser. No. 10/684,350 (MIPS.0188-01-US) and Ser. NO. 10/684,348 (MIPS.0189-00-US) claims the benefit of the following expired U.S. Provisional Applications, which are hereby incorporated by reference in their entirety for all purposes: Serial No.(Docket No.)Filing DateTitle60/499,180Aug. 28, 2003MULTITHREADING(MIPS.0188-00-US)APPLICATION SPECIFICEXTENSION60/502,358Sep. 12, 2003MULTITHREADING(MIPS.0188-02-US)APPLICATION SPECIFICEXTENSION TO A PROCESSORARCHITECTURE60/502,359Sep. 12, 2003MULTITHREADING(MIPS.0188-03-US)APPLICATION SPECIFICEXTENSION TO A PROCESSORARCHITECTURE
| Number | Date | Country | |
|---|---|---|---|
| 60499180 | Aug 2003 | US | |
| 60502358 | Sep 2003 | US | |
| 60502359 | Sep 2003 | US | |
| 60499180 | Aug 2003 | US | |
| 60502358 | Sep 2003 | US | |
| 60502359 | Sep 2003 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 11313272 | Dec 2005 | US |
| Child | 11330915 | Jan 2006 | US |
| Parent | 11313296 | Dec 2005 | US |
| Child | 11330915 | Jan 2006 | US |
| Parent | 10929097 | Aug 2004 | US |
| Child | 11313272 | US | |
| Parent | 10929097 | Aug 2004 | US |
| Child | 11313296 | US | |
| Parent | 10684350 | Oct 2003 | US |
| Child | 10929097 | Aug 2004 | US |
| Parent | 10684348 | Oct 2003 | US |
| Child | 10929097 | Aug 2004 | US |
