Method and apparatus for communication in a multi-processor computer system6332180Abstract Multiple instances of operating systems execute cooperatively in a single multiprocessor computer wherein all processors and resources are electrically connected together. The single physical machine with multiple physical processors and resources is subdivided by software into multiple partitions, each with the ability to run a distinct copy, or instance, of an operating system. At different times, different operating system instances may be loaded on a given partition. Resources, such as CPUs and memory, can be dynamically assigned to different partitions and used by instances of operating systems running within the machine by modifying the configuration. The partitions themselves can also be changed without rebooting the system by modifying the configuration tree. Inter-instance communication can occur in at least two different ways. In a first communication method, a single bit, which is part of a bitvector in shared memory, is modified by a first instance to indicate to a second instance that a predetermined event associated with that bit has occurred. In a second method, packetized data from the first instance is stored in shared memory, and the first instance notifies the second instance of the presence of the data. The second instance then retrieves the data and acts upon it accordingly. Claims What is claimed is: Description FIELD OF THE INVENTION
typedef struct_gct_node {
unsigned char type;
unsigned char subtype;
uint16 size;
GCT_HANDLE owner;
GCT_HANDLE current_owner;
GCT_ID id;
union {
uint64 node_flags;
struct {
unsigned node_hardware : 1;
unsigned node_hotswap : 1;
unsigned node_unavailable : 1;
unsigned node_hw_template : 1;
unsigned node_initialized : 1;
unsigned node_cpu_primary : 1;
#define NODE_HARDWARE 0x001
#define NODE_HOTSWAP 0x002
#define NODE_UNAVAILABLE 0x004
#define NODE_HW_TEMPLATE 0x008
#define NODE_INITIALIZED 0x010
#define NODE_PRIMARY 0x020
} flag_bits;
} flag_union;
GCT_HANDLE config;
GCT_HANDLE affinity;
GCT_HANDLE parent;
GCT_HANDLE next_sib;
GCT_HANDLE prev_sib;
GCT_HANDLE child;
GCT_HANDLE reserved
uint32 magic
} GCT_NODE;
In the above definition the type definitions "uint" are unsigned integers with the appropriate bit lengths. As previously mentioned, nodes are located and identified by a handle (identified by the typedef GCT_HANDLE in the definition above). An illustrative handle is a signed 32-bit offset from the base of the configuration tree to the node. The value is unique across all partitions in the computer system. That is, a handle obtained on one partition must be valid to lookup a node, or as an input to a console callback, on all partitions. The magic field contains a predetermined bit pattern which indicates that the node is actually a valid node. The tree root node represents the entire system. Its handle is always zero. That is, it is always located at the first physical location in the memory allocated for the configuration tree following the config header. It has the following definition:
typedef struct_gct_root_node {
GCT_NODE hd;
uint64 lock;
uint64 transient_level;
uint64 current_level;
uint64 console_req;
uint64 min_alloc;
uint64 min_align;
uint64 base_alloc;
uint64 base_align;
uint64 max_phys_address;
uint64 mem_size;
uint64 platform_type;
int32 pIatform_name;
GCT_HANDLE primary_instance;
GCT_HANDLE first_free;
GCT_HANDLE high_limit;
GCT_HANDLE lookaside;
GCT_HANDLE available;
uint32 max_partition;
int32 partitions;
int32 communities;
uint32 max_platform_partition;
uint32 max_fragments;
uint32 max_desc;
char APMP_id[16];
char APMP_id_pad[4];
int32 bindings;
} GCT_ROOT_NODE;
The fields in the root node are defined as follows: lock This field is used as a simple lock by software wishing to inhibit changes to the structure of the tree, and the software configuration. When this value is -1 (all bits on ) the tree is unlocked; when the value is >=0 the tree is locked. This field is modified using atomic operations. The caller of the lock routine passes a partition ID which is written to the lock field. This can be used to assist in fault tracing, and recovery during crashes. transient_level This field is incremented at the start of a tree update. current_level This field is updated at the completion of a tree update. console_req This field specifies the memory required in bytes for the console in the base memory segment of a partition. min_alloc This field holds the minimum size of a memory fragment, and the allocation unit (fragments size must be a multiple of the allocation ). It must be a power of 2. min_align This field holds the alignment requirements for a memory fragment. It must be a power of 2. base_alloc This field specifies the minimum memory in bytes (including console_req) needed for the base memory segment for a partition. This is where the console, console structures, and operating system will be loaded for a partition. It must be greater or equal to minAlloc and a multiple of minAlloc. base_align This field holds the alignment requirement for the base memory segment of a partition. It must be a power of 2, and have an alignment of at least min_align. max_phys_address The field holds the calculated largest physical address that could exist on the system, including memory subsystems that are not currently powered on and available. mem_size This field holds the total memory currently in system. platform_type This field stores the type of platform taken from a field in the HWRPB. platform_name This field holds an integer offset from the base of the tree root node to a string representing the name of the platform. primary_instance This field stores the partition ID of the first operating system instance. first_free This field holds the offset from the tree root node to the first free byte of memory pool used for new nodes. high_limit This field holds the highest address at which a valid node can be located within the configuration tree. It is used by callbacks to validate that a handle is legal. lookaside This field is the handle of a linked list of nodes that have been deleted, and that may be reclaimed. When a community or partition are deleted, the node is linked into this list, and creation of a new partition or community will look at this list before allocating from free pool. available This field holds the number of bytes remaining in the free pool pointed to by the first_free field. max_partitions This field holds the maximum number of partitions computed by the platform based on the amount of hardware resources currently available. partitions This field holds an offset from the base of the root node to an array of handles. Each partition ID is used as an index into this array, and the partition node handle is stored at the indexed location. When a new partition is created, this array is examined to find the first partition ID which does not have a corresponding partition node handle and this partition ID is used as the ID for the new partition. communities This field also holds an offset from the base of the root node to an array of handles. Each community ID is used an index into this array, and a community node handle is stored in the array. When a new community is created, this array is examined to find the first community ID which does not have a corresponding community node handle and this community ID is used as the ID for the new community. There cannot be more communities than partitions, so the array is sized based on the maximum number of partitions. max_platform_partition This field holds the maximum number of partitions that can simultaneously exist on the platform, even if additional hardware is added (potentially inswapped). max_fragments This field holds a platform defined maximum number of fragments into which a memory descriptor can be divided. It is used to size the array of fragments in the memory descriptor node. max_desc This field holds the maximum number of memory descriptors for the platform. APMP_id This field holds a system ID set by system software and saved in non-volatile RAM. APMP_id_pad This field holds padding bytes for the APMP ID. bindings This field holds an offset to an array of "bindings" Each binding entry describes a type of hardware node, the type of node the parent must be, the configuration binding, and the affinity binding for a node type. Bindings are used by software to determine how node types are related and configuration and affinity rules. A community provides the basis for the sharing of resources between partitions. While a hardware component may be assigned to any partition in a community, the actual sharing of a device, such as memory, occurs only within a community. The community node 310 contains a pointer to a control section, called an APMP database, which allows the operating system instances to control access and membership in the community for the purpose of sharing memory and communications between instances. The APMP database and the creation of communities are discussed in detail below. The configuration ID for the community is a signed 16-bit integer value assigned by the console program. The ID value will never be greater than the maximum number of partitions that can be created on the platform. A partition node, such as node 312 or 314, represents a collection of hardware that is capable of running an independent copy of the console program, and an independent copy of an operating system. The configuration ID for this node is a signed 16-bit integer value assigned by the console. The ID will never be greater than the maximum number of partitions that can be created on the platform. The node has the definition:
typedef struct _gct_partition_node {
GCT_NODE hd;
uint64 hwrpb;
uint64 incarnation;
uint64 priority;
int32 os_type;
uint32 partition_reserved_1;
uint64 instance_name_format;
char instance_name[128];
} GCT_PARTITION_NODE;
The defined fields have the definitions: hwrpb This field holds the physical address of the hardware restart parameter block for this partition. To minimize changes to the HWRPB, the HWRPB does not contain a pointer to the partition, or the partition ID. Instead, the partition nodes contain a pointer to the HWRPB. System software can then determine the partition ID of the partition in which it is running by searching the partition nodes for the partition which contains the physical address of its HWRPB. incarnation This field holds a value which is incremented each time the primary CPU of the partition executes a boot or restart operation on the partition. priority This field holds a partition priority. os_type This field holds a value which indicates the type of operating system that will be loaded in the partition. partition_reserved.sub.-- 1 This field is reserved for future use. instance_name_format This field holds a value that describes the format of the instance name string. instance_name This field holds a formatted string which is interpreted using the instance_name_format field. The value in this field provides a high-level path name to the operating system instance executing in the partition. This field is loaded by system software and is not saved across power cycles. The field is cleared at power up and at partition creation and deletion. A System Building Block node, such as node 322 or 324, represents an arbitrary piece of hardware, or conceptual grouping used by system platforms with modular designs such as that illustrated in FIG. 2. A QBB (Quad Building Block) is a specific example of an SBB and corresponds to units such as units 100, 102, 104 and 106 in FIG. 1. Children of the SBB nodes 322 and 324 include input/output processor nodes 326 and 340. CPU nodes, such as nodes 328-332 and 342-346, are assumed to be capable of operation as a primary CPU for SMP operation. In the rare case where a CPU is not primary capable, it will have a SUBTYPE code indicating that it cannot be used as a primary CPU in SMP operation. This information is critical when configuring resources to create a new partition. The CPU node will also carry information on where the CPU is currently executing. The primary for a partition will have the NODE_CPU_PRIMARY flag set in the NODE_FLAGS field. The CPU node has the following definition: typedef struct _gct_cpu_node {GCT_NODE hd; }GCT_CPU_NODE; A memory subsystem node, such as node 334 or 348, is a "pseudo" node that groups together nodes representing the physical memory controllers and the assignments of the memory that the controllers provide. The children of this node consist of one or more memory controller nodes (such as nodes 336 and 350) which the console has configured to operate together (interleaved), and one or more memory descriptor nodes (such as nodes 338 and 352) which describe physically contiguous ranges of memory. A memory controller node (such as nodes 336 or 350) is used to express a physical hardware component, and its owner is typically the partition which will handle errors, and initialization. Memory controllers cannot be assigned to communities, as they require a specific operating system instance for initialization, testing and errors. However, a memory description, defined by a memory descriptor node, may be split into "fragments" to allow different partitions or communities to own specific memory ranges within the memory descriptor. Memory is unlike other hardware resources in that it may be shared concurrently, or broken into "private" areas. Each memory descriptor node contains a list of subset ranges that allow the memory to be divided among partitions, as well as shared between partitions (owned by a community). A memory descriptor node (such as nodes 338 or 352) is defined as: typedef struct_gct_mem_desc node {GCT_NODE hd; GCT_MEM_INFO mem_info; int32 mem_frag; }GCT_MEM_DESC_NODE; The mem_info structure has the following definition:
typedef struct_gct_mem_info {
uint64 base_pa;
uint64 base_size;
uint32 desc_count;
uint32 info_fill;
}GCT_MEM_INFO:
The mem_frag field holds an offset from the base of the memory descriptor node to an array of GCT_MEM_DESC structures which have the definition:
typedef struct_gct_mem_desc {
uint64 pa;
unit64 size;
GCT_HANDLE mem_owner;
GCT_HANDLE mem_current_owner;
union {
uint32 mem_flags;
struct {
unsigned mem_console : 1;
unsigned mem_private : 1;
unsigned mem_shared : 1;
unsigned base : 1;
#define CGT_MEM_CONSOLE 0x1
#define CGT_MEM_PRIVATE 0x2
#define CGT_MEM_SHARED 0x4
#define CGT_MEM_CONSOLE 0x8
}flag_bits;
} flag_union;
uint32 mem_fill;
} GCT_MEM_DESC;
The number of fragments in a memory description node (nodes 338 or 352) is limited by platform firmware. This creates an upper bound on memory division, and limits unbounded growth of the configuration tree. Software can determine the maximum number of fragments from the max_fragments field in the tree root node 302 (discussed above), or by calling an appropriate console callback function to return the value. Each fragment can be assigned to any partition, provided that the config binding, and the ownership of the memory descriptor and memory subsystem nodes allow it. Each fragment contains a base physical address, size, and owner field, as well as flags indicating the type of usage. To allow shared memory access, the memory subsystem parent node, and the memory descriptor node must be owned by a community. The fragments within the memory descriptor may then be owned by the community (shared) or by any partition within the community. Fragments can have minimum allocation sizes and alignments provided in the tree root node 302. The base memory for a partition (the fragments where the console and operating system will be loaded) may have a greater allocation and alignment than other fragments (see the tree root node definition above). If the owner field of the memory descriptor node is a partition, then the fragments can only be owned by that partition. FIG. 4 illustrates the configuration tree shown in FIG. 3 when it is viewed from a perspective of ownership. The console program for a partition relinquishes ownership and control of the partition resources to the operating system instance running in that partition when the primary CPU for that partition starts execution. The concept of "ownership" determines how the hardware resources and CPUs are assigned to software partitions and communities. The configuration tree has ownership pointers illustrated in FIG. 4 which determine the mapping of hardware devices to software such as partitions (exclusive access) and communities (shared access). An operating system instance uses the information in the configuration tree to determine to which hardware resources it has access and reconfiguration control. Passive hardware resources which have no owner are unavailable for use until ownership is established. Once ownership is established by altering the configuration tree, the operating system instances may begin using the resources. When an instance makes an initial request, ownership can be changed by causing the owning operating system to stop using a resource or by a console program taking action to stop using a resource in a partition where no operating system instance is executing. The configuration tree is then altered to transfer ownership of the resource to another operating system instance. The action required to cause an operating system to stop using a hardware resource is operating system specific, and may require a reboot of the operating system instances affected by the change. To manage the transition of a resource from an owned and active state, to a unowned and inactive state, two fields are provided in each node of the tree. The owner field represents the owner of a resource and is loaded with the handle of the owning software partition or community. At power up of an APMP system, the owner fields of the hardware nodes are loaded from the contents of non-volatile RAM to establish an initial configuration. To change the owner of a resource, the handle value is modified in the owner field of the hardware component, and in the owner fields of any descendants of the hardware component which are bound to the component by their config handles. The current_owner field represents the current user of the resource. When the owner and current_owner fields hold the same non-zero value, the resource is owned and active. Only the owner of a resource can de-assign the resource (set the owner field to zero). A resource that has null owner and current_owner fields is unowned, and inactive. Only resources which have null owner and current_owner fields may be assigned to a new partition or community. When a resource is de-assigned, the owner may decide to deassign the owner field, or both the owner and current_owner fields. The decision is based on the ability of the owning operating system instance running in the partition to discontinue the use of the resource prior to de-assigning ownership. In the case where a reboot is required to relinquish ownership, the owner field is cleared, but the current_owner field is not changed. When the owning operating system instance reboots, the console program can clear any current_owner fields for resources that have no owner during initialization. During initialization, the console program will modify the current_owner field to match the owner field for any node of which it is the owner, and for which the current_owner field is null. System software should only use hardware of which it is the current owner. In the case of a de-assignment of a resource which is owned by a community, it is the responsibility of system software to manage the transition between states. In some embodiments, a resource may be loaned to another partition. In this condition, the owner and current_owner fields are both valid, but not equal. The following table summarizes the possible resource states and the values of the owner and current_owner fields:
TABLE 1
owner field value current_owner field value Resource State
none none unowned, and inactive
none valid unowned, but still active
valid none owned, not yet active
valid equal to owner owned and active
valid is not equal to owner loaned
Because CPUs are active devices, and sharing of CPUs means that a CPU could be executing in the context of a partition which may not be its "owner", ownership of a CPU is different from ownership of a passive resource. The CPU node in the configuration tree provides two fields that indicate which partition a CPU is nominally "owned" by, and in which partition the CPU is currently executing. The owner field contains a value which indicates the nominal ownership of the CPU, or more specifically, the partition in which the CPU will initially execute at system power up. Until an initial ownership is established (that is, if the owner field is unassigned), CPUs are placed into a HWRPB context decided by the master console, but the HWRPB available bit for the CPU will not be set in any HWRPB. This combination prevents the CPU from joining any operating system instance in SMP operation. When ownership of a CPU is established (the owner field is filled in with a valid partition handle), the CPU will migrate, if necessary, to the owning partition, set the available bit in the HWRPB associated with that partition, and request to join SMP operation of the instance running in that partition, or join the console program in SMP mode. The combination of the present and available bits in the HWRPB tell the operating system instance that the CPU is available for use in SMP operation, and the operating system instance may use these bits to build appropriate per-CPU data structures, and to send a message to the CPU to request it to join SMP operation. When a CPU sets the available bit in an HWRPB, it also enters a value into the current_owner field in its corresponding CPU node in the configuration tree. The current_owner field value is the handle of the partition in which the CPU has set the active HWRPB bit and is capable of joining SMP operation. The current_owner field for a CPU is only set by the console program. When a CPU migrates from one partition to another partition, or is halted into an unassigned state, the current_owner field is cleared (or changed to the new partition handle value) at the same time that the available bit is cleared in the HWRPB. The current_owner field should not be written to directly by system software, and only reflects which HWRPB has the available bit set for the CPU. During runtime, an operating system instance can temporarily "loan" a CPU to another partition without changing the nominal ownership of the CPU. The traditional SMP concept of ownership using the HWRPB present and available bits is used to reflect the current execution context of the CPU by modifying the HWRPB and the configuration tree in atomic operations. The current_owner field can further be used by system software in one of the partitions to determine in which partition the CPU is to currently executing (other instances can determine the location of a particular CPU by examining the configuration tree.) It is also possible to de-assign a CPU and return it into a state in which the available bit is not set in any HWRPB, and the current_owner field in the configuration tree node for the CPU is cleared. This is accomplished by halting the execution of the CPU and causing the console program to clear the owner field in the configuration tree node, as well as the current_owner field and the available HWRPB bit. The CPU will then execute in console mode and poll the owner field waiting for a valid partition handle to be written to it. System software can then establish a new owner, and the CPU begin execution in the new partition. Illustrative ownership pointers are illustrated in FIG. 4 by arrows. Each of the nodes in FIG. 4 that corresponds to a similar node in FIG. 3 is given a corresponding number. For example, the software root node denoted in FIG. 3 as node 306 is denoted as node 406 in FIG. 4. As shown in FIG. 4, the community 410 is "owned" by the software root 406. Likewise, the system building blocks 1 and 2 (422 and 425) are owned by the community 410. Similarly, partitions 412 and 414 are also owned by the community 410. Partition 412 owns CPUs 428-432 and the I/O processor 426. The memory controller 436 is also a part of partition 1 (412). In a like manner, partition 2 (414) owns CPUs 442-446, I/O processor 440 and memory controller 450. The common or shared memory in the system is comprised of memory subsystems 434 and 448 and memory descriptors 438 and 452. These are owned by the community 410. Thus, FIG. 4 describes the layout of the system as it would appear to the operating system instances. Operating System Characteristics As previously mentioned, the illustrative computer system can operate with several different operating systems in different partitions. However, conventional operating systems may need to be modified in some aspects in order to make them compatible with the inventive system, depending on how the system is configured. Some sample modifications for the illustrative embodiment are listed below: 1. Instances may need to be modified to include a mechanism for choosing a "primary" CPU in the partition to run the console and be a target for communication from other instances. The selection of a primary CPU can be done in a conventional manner using arbitration mechanisms or other conventional devices. 2. Each instance may need modifications that allow it to communicate and cooperate with the console program which is responsible for creating a configuration data block that describes the resources available to the partition in which the instance is running. For example, the instance should not probe the underlying hardware to determine what resources are available for usage by the instance. Instead, if it is passed a configuration data block that describes what resources that instance is allowed to access, it will need to work with the specified resources. 3. An instance may need to be capable of starting at an arbitrary physical address and may not be able to reserve any specific physical address in order to avoid conflicting with other operating systems running at that particular address. 4. An instance may need to be capable of supporting multiple arbitrary physical holes in its address space, if it is part of a system configuration in which memory is shared between partitions. In addition, an instance may need to deal with physical holes in its address space in order to support "hot inswap" of memory. 5. An instance may need to pass messages and receive notifications that new resources are available to partitions and instances. More particularly, a protocol is needed to inform an instance to search for a new resource. Otherwise, the instance may never realize that the resource has arrived and is ready for use. 6. An instance may need to be capable of running entirely within its "private memory" if it is used in a system where instances do not share memory. Alternatively, an instance may need to be capable of using physical "shared memory" for communicating or sharing data with other instances running within the computer if the instance is part of a system in which memory is shared. In such a shared memory system, an instance may need to be capable of mapping physical "shared memory" as identified in the configuration tree into its virtual address space, and the virtual address spaces of the "processes" running within that operating system instance. 7. Each instance may need some mechanism to contact another CPU in the computer system in order to communicate with it. 8. An instance may also need to be able to recognize other CPUs that are compatible with its operations, even if the CPUs are not currently assigned to its partition. For example, the instance may need to be able to ascertain CPU parameters, such as console revision number and clock speed, to determine whether it could run with that CPU, if the CPU was re-assigned to the partition in which the instance is running. Changing the Configuration Tree Each console program provides a number of callback functions to allow the associated operating system instance to change the configuration of the APMP system, for example, by creating a new community or partition, or altering the ownership of memory fragments. In addition, other callback functions provide the ability to remove a community, or partition, or to start operation on a newly-created partition. However, callback functions do not cause any changes to take place on the running operating system instances. Any changes made to the configuration tree must be acted upon by each instance affected by the change. The type of action that must take place in an instance when the configuration tree is altered is a function of the type of change, and the operating system instance capabilities. For example, moving an input/output processor from one partition to another may require both partitions to reboot. Changing the memory allocation of fragments, on the other hand, might be handled by an operating system instance without the need for a reboot. Configuration of an APMP system entails the creation of communities and partitions, and the assignment of unassigned components. When a component is moved from one partition to another, the current owner removes itself as owner of the resource and then indicates the new owner of the resource. The new owner can then use the resource. When an instance running in a partition releases a component, the instance must no longer access the component. This simple procedure eliminates the complex synchronization needed to allow blind stealing of a component from an instance, and possible race conditions in booting an instance during a reconfiguration. Once initialized, configuration tree nodes will never be deleted or moved, that is, their handles will always be valid. Thus, hardware node addresses may be cached by software. Callback functions which purport to delete a partition or a community do not actually delete the associated node, or remove it from the tree, but instead flag the node as UNAVAILABLE, and clear the ownership fields of any hardware resource that was owned by the software component. In order to synchronize changes to the configuration tree, the root node of the tree maintains two counters (transient_level and current_level). The transient_level counter is incremented at the start of an update to the tree, and the current_level counter is incremented when the update is complete. Software may use these counters to determine when a change has occurred, or is occurring to the tree. When an update is completed by a console, an interrupt can be generated to all CPUs in the APMP system. This interrupt can be used to cause system software to update its state based on changes to the tree. Creation of an APMP Computer System FIG. 5 is a flowchart that illustrates an overview of the formation of the illustrative adaptively-partitioned, multi-processor (APMP) computer system. The routine starts in step 500 and proceeds to step 502 where a master console program is started. If the APMP computer system is being created on power up, the CPU on which the master console runs is chosen by a predetermined mechanism, such as arbitration, or another hardware mechanism. If the APMP computer system is being created on hardware that is already running, a CPU in the first partition that tries to join the (non-existent) system runs the master console program, as discussed below. Next, in step 504, the master console program probes the hardware and creates the configuration tree in step 506 as discussed above. If there is more than one partition in the APMP system on power up, each partition is initialized and its console program is started (step 508). Finally, an operating system instance is booted in at least one of the partitions as indicated in step 510. The first operating system instance to boot creates an APMP database and fills in the entries as described below. APMP databases store information relating to the state of active operating system instances in the system. The routine then finishes in step 512. It should be noted that an instance is not required to participate in an APMP system. The instance can choose not to participate or to participate at a time that occurs well after boot. Those instances which do participate form a "sharing set." The first instance which decides to join a sharing set must create it. There can be multiple sharing sets operating on a single APMP system and each sharing set has its own APMP database. Deciding to Create a New APMP System or to Join an Existing APMP System An operating system instance running on a platform which is also running the APMP computer system does not necessarily have to be a member of the APMP computer system. The instance can attempt to become a member of the APMP system at any time after booting. This may occur either automatically at boot, or after an operator-command explicitly initiates joining. After the operating system is loaded at boot time, the operating system initialization routine is invoked and examines a stored parameter to see whether it specifies immediate joining and, if so, the system executes a joining routine which is part of the APMP computer system. An operator command would result in an execution of the same routine. APMP Database An important data structure supporting the inventive software allocation of resources is the APMP database which keeps track of operating system instances which are members of a sharing set. The first operating system instance attempting to set up the APMP computer system initializes an APMP database, thus creating, or instantiating, the inventive software resource allocations for the initial sharing set. Later instances wishing to become part of the sharing set join by registering in the APMP database associated with that sharing set. The APMP database is a shared data structure containing the centralized information required for the management of shared resources of the sharing set. An APMP database is also initialized when the APMP computer system is reformed in response to an unrecoverable error. More specifically, each APMP database is a three-part structure. The first part is a fixed-size header portion including basic synchronization structures for creation of the APMP computer system, address-mapping information for the database and offsets to the service-specific segments that make up the second portion. The second portion is an array of data blocks with one block assigned to each potential instance. The data blocks are called "node blocks." The third portion is divided into segments used by each of the computer system sub-facilities. Each sub-facility is responsible for the content of, and synchronizing access to, its own segment. The initial, header portion of an APMP database is the first part of the APMP database mapped by a joining operating system instance. Portions of the header are accessed before the instance has joined the sharing set, and, in fact, before the instance knows that the APMP computer system exists. The header section contains: 1. a membership and creation synchronization quadword 2. a computer system software version 3. state information, creation time, incarnation count, etc. 4. a pointer (offset) to a membership mask 5. crashing instance, crash acknowledge bits, etc. 6. validation masks, including a bit for each service 7. memory mapping information (page frame number information) for the entire APMP database 8. offset/length pairs describing each of the service segments (lengths in bytes rounded to pages and offsets full pages) including: shared memory services cpu communications services membership services (if required) locking services The array of node blocks is indexed by a system partition id (one per instance possible on the current platform) and each block contains: instance software version interrupt reason mask instance state instance incarnation instance heartbeat instance membership timestamp little brother instance id and inactive-time; big brother instance id instance validation done bit. An APMP database is stored in shared memory. The initial fixed portion of N physically contiguous pages occupies the first N pages of one of two memory ranges allocated by the first instance to join during initial partitioning of the hardware. The instance directs the console to store the starting physical addresses of these ranges in the configuration tree. The purpose of allocating two ranges is to permit failover in case of hardware memory failure. Memory management is responsible for mapping the physical memory into virtual address space for the APMP database. The detailed actions taken by an operating system instance are illustrated in FIG. 6. More specifically, when an operating system instance wishes to become a member of a sharing set, it must be prepared to create the APMP computer system if it is the first instance attempting to "join" a non-existent system. In order for the instance to determine whether an APMP system already exists, the instance must be able to examine the state of shared memory as described above. Further, it must be able to synchronize with other instances which may be attempting to join the APMP system and the sharing set at the same time to prevent conflicting creation attempts. The master console creates the configuration tree as discussed above. Subsequently, a region of memory is initialized by the first, or primary, operating system instance to boot, and this memory region can be used for an APMP database. Mapping the APMP Database Header The goal of the initial actions taken by all operating system instances is to map the header portion of the APMP database and initialize primitive inter-instance interrupt handling to lay the groundwork for a create or join decision. The routine used is illustrated in FIG. 6 which begins in step 600. The first action taken by each instance (step 602) is to engage memory management to map the initial segment of the APMP database as described above. At this time, the array of node blocks in the second database section is also mapped. Memory management maps the initial and second segments of the APMP database into the primary operating system address space and returns the start address and length. The instance then informs the console to store the location and size of the segments in the configuration tree. Next, in step 604, the initial virtual address of the APMP database is used to allow the initialization routine to zero interrupt reason masks in the node block assigned to the current instance. A zero initial value is then stored to the heartbeat field for the instance in the node block, and other node block fields. In some cases, the instance attempting to create a new APMP computer system was previously a member of an APMP system and did not withdraw from the APMP system. If this instance is rebooting before the other instances have removed it, then its bit will still be "on" in the system membership mask. Other unusual or error cases can also lead to "garbage" being stored in the system membership mask. Next, in step 608, the virtual address (VA) of the APMP database is stored in a private cell which is examined by an inter-processor interrupt handler. The handler examines this cell to determine whether to test the per-instance interrupt reason mask in the APMP database header for work to do. If this cell is zero, the APMP database is not mapped and nothing further is done by the handler. As previously discussed, the entire APMP database, including this mask, is initialized so that the handler does nothing before the address is stored. In addition, a clock interrupt handler can examine the same private cell to determine whether to increment the instance-specific heartbeat field for this instance in the appropriate node block. If the private cell is zero, the interrupt handler does not increment the heartbeat field. At this point, the routine is finished (step 610) and the APMP database header is accessible and the joining instance is able to examine the header and decide whether the APMP computer system does not exist and, therefore, the instance must create it, or whether the instance will be joining an already-existing APMP system. Once the APMP header is mapped, the header is examined to determine whether an APMP computer system is up and functioning, and, if not, whether the current instance should initialize the APMP database and create the APMP computer system. The problem of joining an existing APMP system becomes more difficult, for example, if the APMP computer system was created at one time, but now has no members, or if the APMP system is being reformed after an error. In this case, the state of the APMP database memory is not known in advance, and a simple memory test is not sufficient. An instance that is attempting to join a possibly existing APMP system must be able to determine whether an APMP system exists or not and, if it does not, the instance must be able to create a new APMP system without interference from other instances. This interference could arise from threads running either on the same instance or on another instance. In order to prevent such interference, the create/join decision is made by first locking the APMP database and then examining the APMP header to determine whether there is a functioning APMP computer system. If there is a properly functioning APMP system, then the instance joins the system and releases the lock on the APMP database. Alternatively, if there is no APMP system, or if the there is an APMP system, but it is non-functioning, then the instance creates a new APMP system, with itself as a member and releases the lock on the APMP database. If there appears to be an APMP system in transition, then the instance waits until the APMP system is again operational or dead, and then proceeds as above. If a system cannot be created, then joining fails. Creating a new APMP Computer System Assuming that a new APMP system must be created, the creator instance is responsible for allocating the rest of the APMP database, initializing the header and invoking system services. Assuming the APMP database is locked as described above, the following steps are taken by the creator instance to initialize the APMP system (these steps are shown in FIGS. 7A and 7B):
Step 702 the creator instance sets the APMP system state and
its node block state to "initializing."
Step 704 the creator instance calls a size routine for each
system service with the address of its length field in
the header.
Step 706 the resulting length fields are summed and the creator
instance calls memory management to allocate space for
the entire APMP database by creating a new mapping and
deleting the old mapping.
Step 708 the creator instance fills in the offsets to the beginnings
of each system service segment.
Step 710 the initialization routine for each service is called
with the virtual addresses of the APMP database, the
service segment and the segment length.
Step 712 the creator instance initializes a membership mask to make
itself the sole member and increments an incarnation count.
It then sets creation time, software version, and other
creation parameters.
Step 714 the instance then sets itself as its own big and little brother
(for heartbeat monitoring purposes as described below).
Step 716 the instance then fills in its instance state as "member" and
the APMP system state as "operational."
Step 718 finally, the instance releases the APMP database lock.
The routine then ends in step 720.
Joining an Existing APMP Computer System Assuming an instance has the APMP database locked, the following steps are taken by the instance to become a member of an existing APMP system (shown in FIGS. 8A and 8B):
Step 802 the instance checks to make sure that its instance name
is unique. If another current member has the instance's
proposed name, joining is aborted.
Step 804 the instance sets the APMP system state and its node
block state to "instance joining"
Step 806 the instance calls a memory management routine to
map the variable portion of the APMP database into
its local address space.
Step 808 the instance calls system joining routines for each system
service with the virtual addresses of the APMP database and
its segment and its segment length.
Step 810 if all system service joining routines report success, then
the instance joining routine continues. If any system service
join routine fails, the instance joining process must start
over and possibly create a new APMP computer system.
Step 812 assuming that success was achieved in step 810, the instance
adds itself to the system membership mask.
Step 814 the instance selects a big brother to monitor its instance health
as set forth below.
Step 816 the instance fills in its instance state as "member" and sets a
local membership flag.
Step 818 the instance releases the configuration database lock.
The routine then ends in step 820.
The loss of an instance, either through inactivity timeout or a crash, is detected by means of a "heartbeat" mechanism implemented in the APMP database. Instances will attempt to do minimal checking and cleanup and notify the rest of the APMP system during an instance crash. When this is not possible, system services will detect the disappearance of an instance via a software heartbeat mechanism. In particular, a "heartbeat" field is allocated in the APMP database for each active instance. This field is written to by the corresponding instance at time intervals that are less than a predetermined value, for example, every two milliseconds. Any instance may examine the heartbeat field of any other instance to make a direct determination for some specific purpose. An instance reads the heartbeat field of another instance by reading its heartbeat field twice separated by a two millisecond time duration. If the heartbeat is not incremented between the two reads, the instance is considered inactive (gone, halted at control-P, or hung at or above clock interrupt priority level.) If the instance remains inactive for a predetermined time, then the instance is considered dead or disinterested. In addition, a special arrangement is used to monitor all instances because it is not feasible for every instance to watch every other instance, especially as the APMP system becomes large. This arrangement uses a "big brother--little brother" scheme. More particularly, when an instance joins the APMP system, before releasing the lock on the APMP database, it picks one of the current members to be its big brother and watch over the joining instance. The joining instance first assumes big brother duties for its chosen big brother's current little brother, and then assigns itself as the new little brother of the chosen instance. Conversely, when an instance exits the APMP computer system while still in operation so that it is able to perform exit processing, and while it is holding the lock on the APMP database, it assigns its big brother duties to its current big brother before it stops incrementing its heartbeat. Every clock tick, after incrementing its own heartbeat, each instance reads its little brother's heartbeat and compares it to the value read at the last clock tick. If the new value is greater, or the little brother's ID has changed, the little brother is considered active. However, if the little brother ID and its heartbeat value are the same, the little brother is considered inactive, and the current instance begins watching its little brother's little brother as well. This accumulation of responsibility continues to a predetermined maximum and insures that the failure of one instance does not result in missing the failure of its little brother. If the little brother begins incrementing its heartbeat again, all additional responsibilities are dropped. If a member instance is judged dead, or disinterested, and it has not notified the APMP computer system of its intent to shut down or crash, the instance is removed from the APMP system. This may be done, for example, by setting the "bugcheck" bit in the instance primitive interrupt mask and sending an IP interrupt to all CPU's of the instance. As a rule, shared memory may only be accessed below the hardware priority of the IP interrupt. This insures that if the CPUs in the instance should attempt to execute at a priority below that of the IP interrupt, the IP interrupt will occur first and thus the CPU will see the "bugcheck" bit before any lower priority threads can execute. This insures the operating system instance will crash and not touch shared resources such as memory which may have been reallocated for other purposes when the instances were judged dead. As an additional or alternative mechanism, a console callback (should one exist) can be invoked to remove the instance. In addition, in accordance with a preferred embodiment, whenever an instance disappears or drops out of the APMP computer system without warning, the remaining instances perform some sanity checks to determine whether they can continue. These checks include verifying that all pages in the APMP database are still accessible, i.e. that there was not a memory failure. Assignment of Resources After Joining A CPU can have at most one owner partition at any given time in the power-up life of an APMP system. However, the reflection of that ownership and the entity responsible for controlling it can change as a result of configuration and state transitions undergone by the resource itself, the partition it resides within, and the instance running in that partition. CPU ownership is indicated in a number of ways, in a number of structures dictated by the entity that is managing the resource at the time. In the most basic case, the CPU can be in an unassigned state, available to all partitions that reside in the same sharing set as the CPU. Eventually that CPU is assigned to a specific partition, which may or may not be running an operating system instance. In either case, the partition reflects its ownership to all other partitions through the configuration tree structure, and to all operating system instances that may run in that partition through the AVAILABLE bit in the HWRPB per-CPU flags field. If the owning partition has no operating system instance running on it, its console is responsible for responding to, and initiating, transition events on the resources within it. The console decides if the resource is in a state that allows it to migrate to another partition or to revert back to the unassigned state. If, however, there is an instance currently running in the partition, the console relinquishes responsibility for initiating resource transitions and is responsible for notifying the running primary of the instance when a configuration change has taken place. It is still the facilitator of the underlying hardware transition, but control of resource transitions is elevated one level up to the operating system instance. The transfer of responsibility takes place when the primary CPU executes its first instruction outside of console mode in a system boot. Operating system instances can maintain ownership state information in any number of ways that promote the most efficient usage of the information internally. For example, a hierarchy of state bit vectors can be used which reflect the instance-specific information both internally and globally (to other members sharing an APMP database). The internal representations are strictly for the use of the instance. They are built up at boot time from the underlying configuration tree and HWRPB information, but are maintained as strict software constructs for the life of the operating system instance. They represent the software view of the partition resources available to the instance, and may--through software rule sets--further restrict the configuration to a subset of that indicated by the physical constructs. Nevertheless, all resources in the partition are owned and managed by the instance--using the console mechanisms to direct state transitions--until that operating system invocation is no longer a viable entity. That state is indicated by halting the primary CPU once again back into console mode with no possibility of returning without a reboot. Ownership of CPU resources never extends beyond the instance. The state information of each individual instance is duplicated in an APMP database for read-only decision-making purposes, but no other instance can force a state transition event for another's CPU resource. Each instance is responsible for understanding and controlling its own resource set; it may receive external requests for its resources, but only it can make the decision to allow the resources to be transferred. When each such CPU becomes operational, it does not set its AVAILABLE bit in the per-CPU flags. When the AVAILABLE bit is not set, no instance will attempt to start, nor expect the CPU to join in SMP operation. Instead, the CPU, in console mode, polls the owner field in the configuration tree waiting for a valid partition to be assigned. Once a valid partition is assigned as the owner by the primary console, the CPU will begin operation in that partition. During runtime, the current_owner field reflects the partition where a CPU is executing. The AVAILABLE bit in the per-CPU flags field in the HWRPB remains the ultimate indicator of whether a CPU is actually available, or executing, for SMP operation with an operating system instance, and has the same meaning as in conventional SMP systems. It should be noted that an instance need not be a member of a sharing set to participate in many of the reconfiguration features of an APMP computer system. An instance can transfer its resources to another instance in the APMP system so that an instance which is not a part of a sharing set can transfer a resource to an instance which is part of the sharing set. Similarly, the instance which is not a part of the sharing set can receive a resource from an instance which is part of the sharing set. Communications Between Operating System Instances Numerous situations arise during system operation that require information to be conveyed from one operating system to an operating system running on another partition. In the present invention, there are two preferred ways to communicate between operating system instances. The first is through "single-bit communication," and the second is through the transmission of packetized data. Each of these communication types makes use of the system shared memory and is described in more detail below. Single bit communication takes advantage of the fact that within the shared memory structure of the present invention are sets of bits used for notification of predetermined system events. For each partition, one set of notification bits (referred to herein as the "notification bitvector") is allocated from the shared memory. In the preferred embodiment, the bits of each notification bitvector are normally set to zero. The instance running on a particular partition, when receiving an interprocessor interrupt (IPINT), examines its notification bitvector and, if any of the bits are nonzero, the instance takes an action according to the bit that has been set. Since each bit of the bitvector represents a different event, identification of the modified bit provides notification of the corresponding event. This allows an instance running on one partition (the "source" instance) to signal an event to an instance running on another partition (the "target" instance) by modifying the appropriate bit of the bitvector associated with the target instance, and then sending the primary CPU of the target instance an IPINT to cause it to examine its notification bitvector. An overview of the steps involved in this type of communication is shown in FIGS. 9A and 9B. When needing to communicate an event to a target instance, a source instance follows the steps shown in FIG. 9A. The routine starts in step 900, and proceeds to step 902, in which the source validates the node ID of the appropriate target node by examining the configuration tree. The source instance then determines a "local" virtual address for the target node, that address identifying the notification bitvector of the target instance (step 904). Obtaining a local virtual address involves obtaining the address of the target node in terms of the addressing of the source instance. Since the memory array for each of the instances has the same physical layout, an index to the desired location in the bitvector of the target instance can be determined to put the address of the location in terms of the source instance's virtual addressing. In the preferred embodiment, the bits of each bitvector are organized in the same order, so that no additional address determinations are necessary to find the desired bit within the target instance bitvector. Once the local address is obtained, the source instance modifies the appropriate notification bit in the bitvector (step 908), and generates an IPINT to the target instance via the inter-instance interrupt handler (step 908). The actions of the target node for single-bit communication are depicted in FIG. 9B. When the IPINT is received, the interrupt causes the target instance to determine its own local virtual address to its notification bitvector (step 910). Once the address is determined, the entire bitvector is copied to local memory, and the shared memory bits are reset to zero (step 912). By using the locally copied version of the bitvector, the instance may conduct operations in response to nonzero bits without disabling its actual shared memory bitvector. Thus, while the instance is processing the copied version of the bitvector, other notifications from source instances are still recorded, ensuring that the target instance will still process the new notifications during the next examination of the shared memory bitvector. After resetting the shared memory bitvector, the target instance examines the locally stored copy of the bitvector, and determines which, if any, of the bits are nonzero (step 914). Since each of the bits represents a different event, the target instance responds to each of the nonzero bits one at a time. For each of bits that have been set, the instance initiates the appropriate processing routine necessary to respond to the event signaled by the nonzero bit being examined. The call function for the appropriate routine is then located and executed by using the position within the bitvector of the bit being examined as an index (step 916). The routine then ends in step 918. The second form of communication within the system uses the transmission of data packets. Since some inter-instance communications require not just the signaling of a predetermined event, but the transmission of specific data as well, a mechanism is provided for allowing such communication. A source instance makes use of shared memory space for storing the data to be transmitted, which is thereafter located and copied by the target instance. A general overview of the details of this packet communication process are shown in the flowchart of FIGS. 10A and 10B. The process starts in step 1000, and then proceeds to step 1002, in which the source instance (i.e. the instance having data to communicate) determines the amount of memory space necessary for storing the data to be transmitted. This memory space is then allocated by the source instance within the shared memory of the APMP database (step 1004), and the data is copied to the allocated memory space (step 1006). After copying the data, the source instance determines a local virtual address for the listhead of the packet queue of the target instance (step 1008). This is accomplished in the same manner as described above for determining the local virtual address for the notification bitvector. The source instance then places the packet in the queue listhead (step 1009). The packet processing of the target instance may be in any known manner, and is the details of such are not germane to the present invention. However, in this embodiment, the target instance has a packet queue w | ||||||