Registration of resources for commit procedures5276876
Abstract
A computer system provides registration of resource for synchronization point processing. Resources include file systems, data bases, and protected conversations. A resource becomes protected by the system when it is registered in the application's execution environment. The execution environment includes but is not exclusive to the interactive environment. Accordingly the invention provides an efficient and uniform method for identification of protected resources. The registration facility is separate from the application. Thus, the users of applications, application program developers, system administrators and operators need not have to consider or design a registration facility but only involve the one provided. A protected resource manager, through its adapter, can modify, at any time, the registration information, either by adding missing information or by changing the existing information. A resource registers only for the work unit in which it wishes to participate. An application could have several work units. A resource can register for two work units in the application or two resources can register for a single work unit or both.
Claims
What is claimed is:
1. A computer system for executing an application program in an execution environment, said application program making a first set of work requests to read or update a first plurality of resources coupled to said execution environment, a second set of work requests to read or update a second plurality of said resources, a first request to commit said first set of work requests, and a second request to commit said second set of work requests, a period between a work request of the first set and the first commit request being at least partially concurrent with a period between a work request of the second set and the second commit request, said system comprising:
means for providing the execution environment;
means, executing in said execution environment, for coordinating a commit procedure for said first set of work requests and coordinating a commit procedure for said second set of work requests; and
a plurality of resource manager interface means, including means for coupling to respective managers of said resources, executing in said execution environment, and coupled to the coordinating means, each of said resource manager interface means for said first plurality of resources responding to the respective work request of said first set by registering the respective resource in a first work unit with said coordinating means and each of said resource manager interface means for said second plurality of resources responding to the respective work request of said second set by registering the respective resource in a second work unit with said coordinating means, said first and second work units being at least partially concurrent with each other; and wherein
said coordinating means responds to said first request by said application program to commit said first set of work requests by determining from said registrations that said first plurality of resources are included in said first work unit and transmitting commit requests for said first plurality of resources and responds to said second request by said application program to commit said second work unit by determining from said registrations that said second plurality of resources are included in said second work unit and transmitting commit requests for said second plurality of resources.
2. A computer system as set forth in claim 1 further comprising:
means for providing a second execution environment; and
communication facility means for coupling the first said execution environment to the second execution environment; and wherein
one of said resources comprises means for making a conversation between said application program executing in said first execution environment and a second application program executing in said second execution environment;
a work request of said first set requests that said conversation be established, said conversation causing an update of another resource coupled to said second execution environment; and
said coordinating means responds to said first commit request from the application program executing in said first execution environment by transmitting a commit request for said second application program to commit said update of said other resource.
3. A computer system as set forth in claim 1 wherein said coordinating means transmits each of the commit requests via the respective resource manager interface means.
4. A computer system as set forth in claim 1 wherein said first set of work requests includes a plurality of requests to update one of said resources, and said resource manager interface means for said one resource includes means for avoiding registration with said coordinating means for said one resource except for a first one of said plurality of work requests for said one resource.
5. A computer system as set forth in claim 1 wherein the resource manager interface means is responsive to a loss of coupling to the respective resource manager of a previously registered resource by deleting or invalidating the registration with said coordinating means.
6. A computer system as set forth in claim 1 wherein the resource manager interface means, includes means, responsive to a request from the respective resource manager, to unregister the respective resource from the respective work unit in said coordinating means.
7. A computer system as set forth in claim 1 wherein said coordinating means directs a two phase commit procedure for said first work unit.
8. A computer system as set forth in claim 1 further comprising an operating system for controlling execution of said application program and wherein said plurality of resource manager interface means and said coordinating means are all part of said operating system.
9. A method for executing an application program in an execution environment, said application program making a first set of work requests to read or update a first plurality of resources coupled to said execution environment, a second set of work requests to read or update a second plurality of said resources, a first request to commit said first set of work requests, and a second request to commit said second set of work requests, a period between a work request of the first set and the first commit request being at least partially concurrent with a period between a work request of the second set and the second commit request, said method comprising the steps of:
providing the execution environment;
establishing in said execution environment a coordinator for the first and second commit requests;
responding, in said execution environment to the work requests of said first set by registering said first plurality of resources in a first work unit with said coordinator;
responding to the work requests of said second set by registering said second plurality of resources in a second work unit with said coordinator, said first and second work units being at least partially concurrent with each other; and
said coordinator responding to said first request by said application program to commit said first set of work requests by determining from said registrations that said first plurality of resources are included in said first work unit and transmitting commit requests for said first plurality of resources;
said coordinator responding to said second request by said application program to commit said second work unit by determining from said registrations that said second plurality of resources are included in said second work unit and transmitting commit requests for said second plurality of resources.
10. A method as set forth in claim 9 further comprising the steps of:
providing a second execution environment; and
coupling the first said execution environment to the second execution environment; and wherein
one of said resources comprises means for making a conversation between said application program executing in said first execution environment and a second application program executing in said second execution environment;
a work request of said first set requests that said conversation be established, said conversation causing an update of another resource coupled to said second execution environment; and
said coordinator responds to said first commit request from the application program executing in said first execution environment by transmitting a commit request for said second application program to commit said update of said other resource.
11. A method as set forth in claim 9 wherein said coordinator transmits each of the commit requests for the respective resource via a respective resource manager interface means which resides within said execution environment.
12. A method as set forth in claim 9 wherein said first set of work requests includes a plurality of requests to update one of said resources, and registration with said coordinator for said one resource is avoided except for a first one of said plurality of work requests for said one resource.
13. A method as set forth in claim 9 further comprising the step of deleting or invalidating registration of a previously registered resource with said coordinator when coupling to said resource is lost.
14. A method as set forth in claim 9 further comprising the step of unregistering a previously registered resource from the respective work unit in response to a request from a respective resource manager to unregister the resource from the respective work unit.
15. A method as set forth in claim 9 wherein said coordinator directs a two phase commit procedure for said first work unit.
16. A method as set forth in claim 9 wherein said coordinator is part of an operating system which controls execution of said application program.
Description
BACKGROUND OF THE INVENTION
The invention relates generally to computer operating systems, and deals more particularly with a computer operating system which provides registration of resources for commit procedures to minimize overhead.
This patent application is related to U.S. patent applications:
U.S. patent application Ser. No. 07/525,430, entitled "LOG NAME EXCHANGE FOR RECOVERY OF PROTECTED RESOURCES" filed May 16, 1990 by M. K. Ainsworth et al.; and
U.S. patent application Ser. No. 07/526,471, entitled "OPTIMIZATION OF COMMIT PROCEDURES" filed May 16, 1990 by A. Coleman et al.; and
U.S. patent application Ser. No. 07/525,938, entitled "RECOVERY FACILITY FOR INCOMPLETE SNYC POINTS FOR DISTRIBUTED APPLICATION" filed May 16, 1990 by M. K. Ainsworth et al.; and
U.S. patent application Ser. No. 07/525,427, entitled "COORDINATED SYNC POINT MANAGEMENT OF PROTECTED RESOURCES" filed May 16, 1990 by A. Coleman; and
U.S. patent application Ser. No. 07/525,429, entitled "ASYNCHRONOUS RESYNCHRONIZATION OF A COMMIT PROCEDURE" filed May 16, 1990 by K. Britton; and
U.S. Pat. No. 5,165,031, entitled "COORDINATED HANDLING OF ERROR CODES AND INFORMATION DESCRIBING ERRORS IN A COMMIT PROCEDURE" by E. A. Pruul et al.; and
U.S. patent application Ser. No. 07/525,426, entitled "LOCAL AND GLOBAL COMMIT SCOPES TAILORED TO WORK UNITS" filed May 16, 1990 by B. A. Maslak et al.
The operating system of the present invention can be used in a network of computer systems. Each such computer system can comprise a central, host computer and a multiplicity of virtual machines or other types of execution environments. The host computer for the virtual machines includes a system control program to schedule access by each virtual machine to a data processor of the host, and help to manage the resources of the host, including a large memory, such that each virtual machine appears to be a separate computer. Each virtual machine can also converse with the other virtual machines to send messages or files via the host. Each virtual machine has its own CMS portion of the system control program to interact with (i.e., receive instructions from and provide prompts for) the user of the virtual machine. There may be resources such as shared file system (SFS) and shared SQL relational databases which are accessible by any user virtual machine and the host.
Each such system is considered to be one real machine. It is common to interconnect two or more such real machines in a network, and transfer data via conversations between virtual machines of different real machines. Such a transfer is made via communication facilities such as AVS Gateway and VTAM facilities ("AVS Gateway" and "VTAM" are trademarks of IBM Corp. of Armonk, N.Y.).
An application can change a database or file resource by first making a work request defining the changes. In response, provisional changes according to the work request are made in shadow files while the original database or file is unchanged. At this time, the shadow files are not valid. Then, the application can request that the changes be committed to validate the shadow file changes, and thereby, substitute the shadow file changes for the original file. A one-phase commit procedure can be utilized. The one-phase commit procedure consists of a command to commit the change of the resource as contained in the shadow file. When resources such as SFS or SQL resources are changed, the commits to the resources can be completed in separate one-phase commit procedures. In the vast majority of cases, all resources will be committed in the separate procedures without error or interruption. However, if a problem arises during any one-phase commit procedure some of the separate commits may have completed while others have not, causing inconsistencies. The cost of rebuilding non-critical resources after the problem may be tolerable in view of the efficiency of the one-phase commit procedure.
However, a two-phase commit procedure is required to protect critical resources and critical conversations. For example, assume a first person's checking account is represented in a first database and a second person's savings account is represented in a second database. If the first person writes a check to the second person and the second person deposits the check in his/her savings account, the two-phase commit procedure ensures that if the first person's checking account is debited then the second person's savings account is credited or else neither account is changed. The checking and savings accounts are considered protected, critical resources because it is very important that data transfers involving the checking and savings accounts be handled reliably. An application program can initiate the two-phase commit procedure with a single command, which procedure consists of the following steps, or phases:
(1) During a prepare phase, each participant (debit and credit) resource is polled by the sync point manager to determine if the resource is ready to commit all changes. Each resource promises to complete the resource update if all resources successfully complete the prepare phase i.e. are ready to be updated.
(2) During a commit phase, the sync point manager directs all resources to finalize the updates or back them out if any resource could not complete the prepare phase successfully.
Some means is required in a computer system to determine which resources participate in a commit procedure and to determine what type of commit procedure, one phase or two phase, is appropriate. In the prior art, protected resources have been associated with a special or additional environment to participate in a commit procedure. However, this association has the following disadvantages. The end user, application developer and administrator must consider the additional environment, and the end users must switch from the personal computing environment to the special environment. Also, data is dedicated to the special environment and not readily available to other environments.
In an IBM System 38 prior art system, applications had a verb that could enable or disable commitment control for an application. Thus, an application developer could inadvertently create an application that updates multiple protected resources without using a two-phase commit procedure.
The CICS/VS prior art system distributed a separate module that was included with the application. The module intercepted application program interface requests to a SQL/DS resource manager and registered for the resource manager. Thus, application developers and administrators had to manage a sync point manager module.
A synchronization point design in CICS/VS system requires that a resource manager notify the system through an enable function that the resource manager is active for the application environment. This enabling required an operator interaction with each resource manager at the start of a CICS/VS transaction environment, and is global i.e. for all applications. Applications could not access the resource manager for any resources (protected or not protected) until after the enable function executed.
In CICS/VS systems, the CICS/VS verb to commit or back out changes to protected resources required a preprocess by the CICS/VS system. The application developer needed to execute the preprocess of the entire application before the compile. Thus, an extra step needed to be added to the application development process.
The CICS/VS prior art system supports modification, by a resource adapter, of registration for an application. A resource manager could state whether or not (i.e. suspension) it wanted to participate in a sync point.
The prior art IBM Shared File System (SFS), contained within the prior art IBM VM/SP Release 6 control program, supports only a one-phase commit procedure for only one resource, a shared file system resource. However, the one-phase commit procedure can atomically commit updates to more than one file on the same resource. Also, the SFS system can atomically commit multiple SFS resource managers where no more than one resource manager is in write mode. This is because a one-phase commit procedure is adequate for this case.
The commit scope, known as a CMS work unit, is controlled by the application which defines which files are read and written in any particular work unit. The work unit is limited to a single resource manager. The application may define multiple work units and use them simultaneously, allowing the application to have multiple work scopes. A commit of one work unit has no effect on any other work units.
Another prior art system, SQL/DS system, has support similar to SFS. It allows multiple work units to a single resource manager, but a work unit can access only one resource manager, even for read-only.
A general object of the present invention is to provide an efficient system and process for identification of resources requiring a two-phase commit procedure.
Another general object of the present invention is to minimize the burden on the users of applications, application program developers, resource manager developers, manager developers, system administrators and operators in providing the efficient method for identifying protected resources.
Another general object of the present invention is to minimize overhead in sync point processing.
SUMMARY
The invention resides in a computer system which provides registration of resources for synchronization point processing. Resources include file systems, data bases, and protected conversations. A resource becomes protected by the system when it is registered in the application's execution environment. The execution environment includes but is not exclusive to the interactive environment. Accordingly the invention provides an efficient and uniform method for identification of protected resources. The registration facility is separate from the application. Thus, the users of applications, application program developers, system administrators and operators need not have to consider or design a registration facility but only involve the one provided. Uniformity means one flexible way to start, change, update, and end participation in synchronization point. According to one feature of the invention, a protected resource manager, through its adapter, can modify at any time, the registration information, either by adding missing information or by changing the existing information. The system allows registration with the information available to the resource manager at that time or to be completed later. The resource adapter is therefore more efficient in design. System performance is improved both by enabling synchronization point optimizations and by reducing the amount of information exchanged at synchronization point. The flexibility of the invention allows the resource adapter the degrees of freedom necessary to meet its product goals. According to another feature of the invention, a resource registers only for the work unit in which it wishes to participate. An application could have several work units. A resource can register for two work units in the application or two resources can register for a single work unit or both. The resources can be homogeneous or heterogeneous. No restraints, pre-processing and/or past processing of the application's source code, are imposed. Another aspect of flexibility is to allow protected resources to withdraw whenever appropriate from application synchronization point processing and thus conserve both resource manager and system resources. Withdrawal can occur at any time, thus simplifying the design of the resource adapter. However, the application can still use the resource adapter.
BRIEF DESCRIPTION OF THE FIGURES
FIG. 1 is a block diagram of a computer system which incorporates all commit and recovery functions in each execution environment, according to the prior art.
FIG. 2 is a block diagram of a computer network including two interconnected computer systems according to the present invention. Each of the systems supports multiple execution environments with a common recovery facility and log.
FIG. 3 is a flowchart of a two-phase commit procedure for resources used by an application running in an execution environment of FIG. 2.
FIG. 4 is a flowchart of recovery processing that is implemented when an interruption occurs during the two-phase commit procedure described in FIG. 3.
FIGS. 5(A) and 5(B) are a flowchart of a two-phase commit procedure for resources used by partner applications running in two distributed application environments connected by a protected conversation supporting sync point facilities of FIG. 2.
FIG. 6 is a block diagram illustrating plural work units defining different commit scopes within a single application environment of FIG. 2, and a commit scope transversing more than one system of FIG. 2.
FIG. 7 is a flowchart illustrating the use of local work units and a global logical unit of work by one application environment of FIG. 2 to define the scope of and facilitate commit processing.
FIG. 8 is a flowchart illustrating the use of local work units and the global logical unit of work of FIG. 7 by another related application environment of FIG. 2 to define the scope of and facilitate commit processing.
FIG. 9 is a timing diagram of a protected conversation in the global logical unit of work of FIGS. 7 and 8.
FIG. 10 is a block diagram that illustrates automatic and generic registration of resources within the systems of FIG. 2.
FIG. 11 is a flowchart illustrating a procedure for registering resources in a sync point manager of FIG. 6 for a suitable type of commit procedure and the steps of the commit procedure.
FIG. 12 is a block diagram illustrating registration on a work unit basis within the systems of FIG. 2.
FIG. 13 is time flow diagram of bank transactions illustrating registration on a work unit basis.
FIG. 14 is a flowchart illustrating a procedure for registering resources, changing registration information for resources and unregistering resources in the sync point manager.
FIG. 15 is a flowchart illustrating the procedure used by resource adapters, protected conversation adapters, and the sync point manager to unregister resources.
FIG. 16 is a flowchart illustrating processing by the sync point manager in response to a sync point request, and optimizations by the sync point manager in selecting one-phase or two-phase commit procedures.
FIG. 17 is a flowchart illustrating the two-phase commit procedure.
FIG. 18 is a flow diagram illustrating three distributed application programs participating in a two-phase commit procedure.
FIG. 19 is a block diagram illustrating the components and procedure for exchanging log names to support recovery of a failed commit procedure when a protected conversation is made between an application in one system and a partner application in another system of FIG. 2.
FIGS. 20(A) and 20(B) are flowcharts of communications facility processing associated with FIG. 19 for an initial event and a subsequent conversation event, respectively.
FIG. 21 is a flowchart of recovery facility processing associated with FIG. 19 that results when a local communications facility requests that the recovery facility exchange log names for a path.
FIG. 22 is a flowchart of recovery facility processing associated with FIG. 19 that results from receiving an exchange of log names request from another recovery facility.
FIG. 23 is a block diagram illustrating the components and procedure for exchanging log names with a local resource manager in a section of FIG. 2.
FIG. 24 is a block diagram illustrating the components and procedure for exchanging log names using a system of FIG. 2 and a remote resource manager.
FIG. 25 is a block diagram illustrating the contents of a recovery facility of FIG. 2.
FIGS. 26 and 27 are flowcharts illustrating the processing for exchange of log names between a participating resource manager and the recovery facility.
FIG. 28 is a block diagram illustrating portability of the sync point log and capability for activating back up recovery facilities.
FIG. 29 is a block diagram which illustrates participation by the resource adapter and sync point manager of FIG. 2 in passing an error flag and information that defines a problem in a commit procedure to an application program.
FIG. 30 is a flowchart illustrating a procedure for using the components of FIG. 29 to pass the error information to the application program.
FIG. 31 is a control block structure for sharing the pages used by error blocks associated with FIG. 29 in order to reduce system working storage.
FIG. 32 is a block diagram of components of FIG. 2 that participate in the generation and management of the error flags and information of FIG. 29.
FIG. 33 is a block diagram illustrating three systems including commit cycles that encompass more than one of the systems commit scopes incorporating resource managers that reside in the same and different systems as an initiating application and communications paths employed during commit processing as well as paths used for sync point recovery processing.
FIG. 34 is a block diagram illustrating three participating application and application environments from FIG. 33 and the resource managers that they employ, forming a tree of sync point participants.
FIG. 35 is a high level flowchart illustrating the recovery facility procedures for pre-sync point agreements and procedures for recovery from a sync point failure.
FIG. 36 is a flowchart illustrating in more detail the recovery facility procedures for recovery from a sync point failure.
FIG. 37 is a block diagram illustrating the contents of logs 72 of FIG. 2 and control structures required to control the procedures represented by FIG. 35.
FIG. 38 is a flowchart providing detail for FIG. 35, steps 299 and 300.
FIG. 39 is a flowchart providing detail for FIG. 35, steps 301 and 302.
FIG. 40 is a flowchart providing detail for FIG. 36, step 311.
FIG. 41 is a flowchart providing detail for FIG. 36, step 312.
FIG. 42 is a flowchart providing detail for FIG. 36, step 313.
FIG. 43 is a flowchart providing detail for FIG. 36, step 314.
FIG. 44 is a flowchart providing detail for FIG. 36, step 315.
FIG. 45 is a flowchart providing detail for FIG. 36, step 304.
FIG. 46 is a flowchart providing detail for FIG. 36, step 317.
FIG. 47 is a flowchart providing detail for FIG. 36, step 318.
FIG. 48 is a flowchart providing detail for FIG. 36, step 319.
FIG. 49 is a flowchart providing detail for FIG. 36, step 306.
FIGS. 50(A) and 50(B) are block diagrams which illustrate application 56A and application 56D requesting asynchronous resynchronization should an error occur during sync point processing.
FIG. 51 is a flow graph illustrating the steps of the asynchronous, resynchronization-in-progress process involving an additional system 50C.
FIG. 52 is a flow graph illustrating the steps of the asynchronous, resynchronization-in-progress process involving a failed backout order originating from system 50C.
FIG. 53 is a flow graph illustrating the steps of the asynchronous, resynchronization-in-progress process involving a failed backout order originating from system 50A.
FIG. 53A is a flow graph illustrating the steps of asynchronous, resynchronization-in-progress process involving a failed prepare call originating from system 50A.
FIG. 54 is a block diagram of another embodiment of the invention as an alternate to FIG. 2.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
Referring to the drawings in detail wherein like reference numerals indicate like elements throughout the several views, FIG. 1 illustrates an LU6.2 syncpoint tower model or architecture according to the Prior Art. This architecture is defined as one execution environment. In the illustrated example, three application programs 14, 16, and 18 are run in execution environment 12 in a time-shared manner. Resource Managers 26 and 27, DB/2 or CICS File Control (DB/2 and CICS are trademarks of IBM Corp.), control access to resources 22 and 24, respectively. It should be noted that if a DB/2 (CICS/MVS operating system) or a SQL/DS (CICS/VSE operating system) resource manager were located outside of environment 12, then environment 12 would include a resource adapter to interface to the resource manager according to the prior art. In this prior art architecture, application program 14 makes a work request invoking resources 22 and 24 to syncpoint manager 20 before requesting committal of resources involved in the work request.
Next, application program 14 requests a commit from syncpoint manager 20 to commit the data updates of the previous work request. In response, syncpoint manager 20 implements a two-phase commit procedure by polling resource managers 26 and 27 to determine if they are ready to commit the resources and if so, to subsequently order the commit. At each phase (and each step of each phase) of the two-phase commit procedure, the syncpoint manager transfers syncpoint information to log 30 indicating the state of the two-phase commit procedure. If a failure occurs during the two-phase commit procedure, the syncpoint manager will implement a synchronization point recovery procedure to bring the resources to a consistent state. The syncpoint manager relies on the synchronization point information in log 30 to determine how far the two-phase commit procedure had progressed before interruption.
Syncpoint manager 20 and the two-phase commit procedure are also used when any one of the applications 14, 16 or 18 attempts to communicate via protected conversation manager 40 using a protected conversation to an application partner in a separate environment in the same system (not shown) or to an application partner within another system (not shown) which is interconnected via a communication facility. According to the prior art synchronization point architecture, this other system/other environment is functionally identical to the execution environment 12 and includes another syncpoint manager functionally identical to 20, another synchronization point log functionally identical to 30, another protected conversation manager functionally identical to 40 and other resource managers functionally identical to 26 and 27. This other environment provides coordination and recovery functions which are separate from those of execution environment 12.
COORDINATED SYNC POINT MANAGEMENT OF PROTECTED RESOURCES
FIG. 2 illustrates a syncpoint architecture according to the Present Invention. The invention includes a distributed computer operating system which supports distributed and non-distributed applications executing within their own execution environment such as a UNIX environment, OS/2 environment, DOS environment in OS/2 operating system, CMS environment in VM operating system, AIX environment in VM operating system, CICS in VM operating system, and MUSIC environment in VM operating system. A distributed application is distinguished by using a resource in another execution environment or having a communications conversation--a special type of resource--with an application partner in another execution environment. The execution environment for the resource manager of application partner may be in the same system or a different one; it can be in the same type environment or a foreign environment. A distributed application execution environment comprises one or more systems supporting applications in their own environments that might not have all the resources required; those resources are distributed elsewhere and are acquired with the aid of a communication facility. The complete environment of a distributed application appears to be full function because the distributed application involves resources that are in other environments--especially the recovery facility and communication facility.
The present invention comprises one or more systems (real machines or central electronic complexes (CECs)) 50A, D. In the illustrated embodiment, system 50A comprises a plurality of identical, distributed application environments 52A, B, and C, a conversation manager 53A and execution environment control programs 61A, B, and C which are part of a system control program 55A, and a recovery facility 70A. By way of example and not limitation, each of the environments 52A, B, and C can be an enhanced version of a VM virtual machine, recovery facility 70A can reside in another enhanced version of a VM virtual machine and system control program 55A can be an enhanced version of a VM operating system for virtual machines 52A, B, and C. Applications running in distributed application environments 52A-C in real machine 50A can communicate with application partners running in similar distributed application environments running in real machine 50D or other systems (not shown) via communication facilities 57A and D. By way of example, communication facility 57A comprises Virtual Telecommunications Access Method ("VTAM") facility and APPC/VM VTAM Support (AVS) gateway facility. Each distributed application environment 52 comprises a single syncpoint manager (SPM) 60A and a plurality of protected resource adapters 62A-B and 64A. A syncpoint manager allows a group of related updates to be committed or backed out in such a way that the changes appear to be atomic. The updates performed between syncpoints (i.e.commit/backout) are called a logical unit of work and the related updates are identified through a unique name assigned by the syncpoint manager via the recovery facility called a logical unit of work identifier. The logical unit of work can involve multiple protected resources accessed by an application in the same distributed application environment and can also involve protected resources accessed by a partner application in other application environments via a conversation which is one type of protected resource.
A conversation is a path established in an architected manner between two partner applications. The use of the conversation by each application is determined by the applications' design and the conversation paradigm used. When a conversation is to be included in the syncpoint process, it is called a protected conversation. Protected resources become part of the logical unit of work by contacting the syncpoint manager through a process called registration as described below in Registration of Resources for Commit Procedure. Each protected resource adapter provides an interface to a resource manager both for an application and for the syncpoint manager. (Alternatively, the protected resource adapter can be merged with the resource manager if the resource manager resides in the same execution environment as the application.)
In the illustrated embodiment, protected resources are files and conversations. In other embodiments of the present invention, protected resources could be database tables, queues, remote procedure cells, and others. Protected resource adapters 62A and B handle interfaces on behalf of application 56A for resource managers 63A and B, respectively, which manage files 78A and B. Resource managers 63A and B are located in the same system. Alternatively, they could reside in a different system in a communication network. In the illustrated embodiment, conversations are managed by a conversation manager which manages the conversations or paths from an application to other partner applications running in different distributed application environments in the same system, or different distributed application environments in different systems in a communication network. If the protected conversation is between two application partners running in different application environments in the same system, e.g. between application partners running in 52A and 52B, then the conversation manager is totally contained in the system control program 55A of system 50A, and communication is made between the application partners via each protected conversation adapter 64A and 64B (not shown). If the protected conversation is between different application environments in different systems, e.g. between application partners running in 52A and 52D, then communication is made between the conversation managers 53A and 53D in systems 50A and 50D via communication facilities 57A and 57D. In this embodiment, such communications utilize a peer to peer communication format. Conversation managers 53A, D use an intra-environment format to communicate facilities 57A, D. Communication facilities 57A, D translate the intra-environment format to an architected inter-system communication standard format and vice versa. By way of example this architected intersystem communication standard format can be of a type defined by IBM's System Network Architecture, LU 6.2 protocol.
Recovery facility 70A serves all distributed application environments 52A,B, and C within real machine 50A. It contains log 72A, it processes handle logging for the syncpoint managers 60A,B, and C it provides recovery for failing syncpoints for all distributed application environments 52A, B, and C. The same is true for recovery facility 70D and its log 72D, and syncpoint manager 60D on system 50D.
When application 56A within distributed application environment 52A desires to update files 78A and 78B, application 56A makes two separate update requests via a file application program interface witin application 56A. The requests involve protected resource adapters (henceforth called protected file adapter for this type of resource) 62A and 62B respectively for files 78A and 78B (step 500 of FIG. 3). Based on resource manager specific implementation, the protected file adapter knows the file is protected. If not already registered with the syncpoint manager for the work unit, protected file adapters 62A and 62B register with syncpoint manager 60A the fact that they want to be involved in all Commit/Backout request for this work unit (step 502). A "work unit" is a grouping of all resources, directly accessible and visible by the application, that participate in a sync point. It is generally associated with a logical unit of work identifier. For a further explanation of work units, see Local and Global Commit Scopes Tailored to Work Units below. Then protected file adapters 62A and 62B contact their respective resources managers 63A and 63B to update files 78A and 78B (Step 504). Return is made to application 56A. Next application 56A requests a syncpoint 58A, i.e. a commit in this case, to syncpoint manager 60A (Step 506). In response, syncpoint manager 60A initiates a two-phase commit procedure (step 508) to be carried out for both of its registered resources, files 78A and 78B, represented by protected file adapters 62A and 62B and their respective resource managers 63A and 63B. In step 508, syncpoint manager 60A calls each of its registered resources at the adapter exit syncpoint exit entry point, given to the syncpoint manager by each resource adapter during registration, with a phase one "prepare" call.
During the course of executing its two-phase commit procedures, syncpoint manager 60A issues a request to recovery facility 70A to force log ("force log" means to make sure the information was written to the actual physical device before returning to syncpoint manager 60A) on log 72A phase one manager information (Step 508). This information includes the logical unit of work identifier, the syncpoint manager state and the names and other pertinent information about each registered protected resource adapter participating in the commit request. This information was given to syncpoint manager 60A when file adapters 62A and 62B registered. Syncpoint manager 60A's state is determined by the rules of the two-phase commit paradigm being followed. For example, the two-phase commit paradigm is a type described by System Network Architecture LU 6.2 Reference: Peer Protocols, SC31-6808, Chapter 5.3 Presentation Services-Sync Point verbs published by the IBM Corporation. If a failure occurs during the syncpoint processing, the syncpoint manager state is used to determine the outcome (Commit or Backout) of the logical unit of work. As per the rules of the two-phase commit paradigm used by this embodiment, the syncpoint manager phase one state is, Initiator, Syncpoint Manager Pending. If the first phase of the two-phase commit procedure is not interrupted and is completed (decision block 512), syncpoint manager 60A issues a second request to recovery facility 70A to force log 72A to its phase two state. Based on the replies from the protected file adapters and resource managers and the rules of the two-phase commit paradigm being used, syncpoint manager 60A knows its second phase decision. In this embodiment, the paradigm is as follow. If one or more protected resources adapters respond "backout" to the phase one request, the phase two decision is "backout"; if all respond "request commit", the decision is "commit". In the example illustrated in FIG. 3, protected file adapters 62A and 62B responded "request commit" (Step 510) and the phase two state is logged by syncpoint manager 60A as Initiator Committed. It should be noted that in this example, file managers 63A and 63B after replying "request commit" through their respective file adapters 62A and 62B to the phase one request are in a state of "indoubt", that is they can commit or backout the file updates based on the phase two decision from syncpoint manager 60A.
After logging, syncpoint manager 60A then issues the phase two call with the decision of commit to protected file adapters 62A and 62B (Step 513). When the file managers 63A and 63B receive the phase two commit decision, each proceeds to do whatever processing is necessary to commit the data, i.e. make the updates permanent (Step 516). When a successful reply is received from protected file adapters 62A and 62B on behalf of their respective resource managers and there is no interruption in syncpoint processing (decision block 514), syncpoint manager 60A calls recovery facility 70A to write to log 72A the state of "forget" for this logical unit of work (Step 515). This does not have to be a force log write which means the log record is written to a data buffer and return can be made to syncpoint manager 60A. The buffer can be written to the physical media at a later point in time. Based on the two phase commit paradigm used in this embodiment, syncpoint manager 60A updates the logical unit of work identifier (increments it by one) which guarantees uniqueness for the next logical unit of work done by application 56A. The syncpoint manager then returns to application 56 (Step 515A).
The two-phase commit paradigms have rules for recovery processing, such that recovery facility 70A knows how to complete an interrupted syncpoint (Step 517 and FIG. 4). If syncpoint manager 60A's process was interrupted, decision block 514 leads to step 517 in which syncpoint manager 60A contacts recovery facility 70A. In step 517 recovery facility 70A receives the logical unit of work identifier and information about the associated failed resource or resources from syncpoint manager 60A. Recovery facility 70A then finds the correct log entry (Step 518 of FIG. 4). The log information, in combination with the two phase commit paradigm being used, allows recovery facility 70A's procedures to complete the interrupted syncpoint processing (Step 519). Based on the two-phase commit paradigm being used in this illustrated example, if the syncpoint state entry for the logical unit of work identifier on log 72A is Initiator, Syncpoint Manager Pending, each failed resource manager 63A or 63B will be told to backout; otherwise, each will be told the syncpoint manager phase two state which is on the log, i.e. commit or backout (Step 520). Once the recovery state is determined, recovery facility 70A will start recovery processes with each failed protected resource manager as described below in Log Name Exchange For Recovery of Protected Resources and in Recovery Facility For Incomplete Sync Points For Distributed Application. This processing consists of exchanging log names and a comparison of states whereby the recovery process of recovery facility 70A tells the failed resource manager 63A or 63B what to do, i.e. commit or backout, and the resource manager 63A or 63B tells the recovery process what it did. The recovery process of recovery facility 70A knows how to contact the failed resource based on information written by syncpoint manager 60A during its phase one logging activity. If the failed resource manager can be contacted (decision block 521) recovery takes place immediately (Step 522). After recovery takes place with each failed resource (decision block 523) return can be made to syncpoint manager 60A (Step 523A). Syncpoint manager 60A will then return to the application 56A (Step 515A). If the failed resource manager could not be contacted, decision block 521 leads to decision block 524 in which recovery facility 70A checks to see if it must complete the recovery processing before returning to application 56A. This decision is based on information contained in the log record for the logical unit of work written by the syncpoint manager during phase one logging. If it must complete recovery, the recovery process keeps trying to contact the failed resource (Step 525); if it can complete the recovery at a later point in time, i.e. wait for recovery was previously selected, recovery facility 70A returns to syncpoint manager 60A with the intent of the recovery processing (i.e. commit or backout) and indication that the recovery will be completed later (Step 526) as described below in Asynchronous Resynchronization of a Commit Procedure. When all resources are recovered (Step 525A), syncpoint manager 60A returns to application 56A (Step 515) with this information.
FIG. 2 also illustrates that application 56A can be part of a distributed application. This means there is at least one partner application that can work with application 56A to complete its processing. To establish a distributed application, application 56A initiates a protected conversation which starts partner application 56D in system 50D by invoking the application program conversation initiate interface and indicates the conversation is to be protected (FIG. 5a, Step 530). This request is handled by protected conversation adapter 64A. Protected conversation adapter 64A asks syncpoint manager 60A for the logical unit of work identifier and includes it along with a unique conversation identifier in the information sent to the remote system 50D. Protected conversation adapter 64A then sends the request to the conversation manager 53A which sends it to communications facility 57A. Protected conversation adapter 64A gets an indication that the conversation initiate request was (or will be) sent from communications facility 57A to communications facility 57D. At this time protected conversation adapter 64A registers with syncpoint manager 60A (Step 532). Asynchronously to this registration process, the conversation initiate request is transmitted to communication facility 57D, and then to conversation manager 53D, and then to protected conversation adapter 64D (Step 532 of FIG. 5A). Protected conversation adapter 64D retrieves the logical unit of work identifier and unique conversation identifier and registers with syncpoint manager 60D on behalf of the conversation manager (Step 532). Protected conversation adapter 64D at this time also gives syncpoint manager 60D the logical unit of work identifier it received on the conversation initiate request. Protected work done by application 56D will be associated with this logical unit of work originally started by application 56A (Step 532). The logical unit of work identifier will also be assigned to a new work unit for application 56D and application 56D is started.
Thus, applications 56A and 56D are partner applications, and together they are called a distributed application. The protected conversation allows application 56A and 56D to send and receive data in a peer to peer manner. This means each side, application 56A or application 56D, can originate the send or receive which is determined by the application writer and the paradigm being used by the communication manager. As described above, a protected conversation is registered with both syncpoint managers by protected conversation adapters 64A and 64D, respectively. During syncpoint processing for the application that issued the first commit, a protected conversation adapter represents a resource to the syncpoint manager that must respond if it can commit (first phase) and whether or not it successfully performed the work requested (second phase). To the other protected conversation adapter receiving the first phase call from its partner protected conversation adapter, the protected conversation is a partner syncpoint manager over which it will receive phase one and phase two orders. Its local syncpoint manager acts like a resource manager, that is the protected conversation adapter will get the results of what the syncpoint manager's resources did (phase one and phase two). It should be noted that the syncpoint paradigm used provides rules for which application partner can issue the first commit. In this example, any application partner can issue the commit first and this is determined by the distributed application design.
Application 56A gets control with the indication that the request to start was successfully sent by communication facility 57A. At this point application 56A is able to send requests to application 56D and application 56A sends a request to application 56D over the established conversation. In this illustrated example, this request eventually causes application 56D to invoke a file application program interface to update file 78D. As described above, the update request causes protected file adapter 62D to register with syncpoint manager 60D under the same work unit (previously assigned for application 56D (Step 532) when application 56D was started) (Step 533). Also in step 533, application 56D sends a reply to application 56A over the conversation indicating that it completed its work. Next, application 56A issues update requests for files 78A and 78B. As previously described, protected file adapters 62A and 62B had previously registered with syncpoint manager 60A and they each contact resource managers 63A and 63B to perform the updates (Steps 533 and 533A).
Application 56A now issues a commit 58A to syncpoint manager 60A (Step 534). As described above, syncpoint manager 60A contacts recovery facility 70A for its phase one logging and issues a phase one "prepare" call to each registered resource (Steps 534A and 535A). Protected file adapters 62A and 62B behave as described above. When protected conversation adapter 64A receives the phase one "prepare" call, it sends an intersystem architected "prepare" call over the protected conversation it represents, i.e. the one originally established by application 56A to application 56D (Step 535). Protected conversation adapter 64D recognizes this "prepare" call and gives application 56D, which had issued a conversation message receive call, a return code requesting it to issue a commit (Step 536). Application 56D then issues a commit 58D to syncpoint manager 60D (Step 537). As described above, syncpoint manager 60D contacts its recovery facility, in this case recovery facility 70D to force log 72D with phase one information (Step 538). Because application 56A issued the original commit request which caused application 56D to subsequently issue a commit, and based on the two-phase commit paradigm used in this embodiment, syncpoint manager 60D's phase one state is "Initiator Cascade, Syncpoint Manager Pending" (Step 538). Syncpoint manager 60D contacts protected file adapter 62D with a phase one "prepare" call (Step 538). Protected file adapter 62D and its associated resource manager 63D perform phase one processing as previously described and returns a reply of "request commit".
In this example, there were no interruptions and decision block 539 leads to step 540 in which syncpoint manager 60D contacts recovery facility 70D to force log 72D to a state of "Agent, Indoubt". This state means that if an interruption subsequently occurs such that syncpoint manager 60D does not receive the phase two decision from syncpoint manager 60A, it would have to wait for recovery processing from recovery facility 70A to complete its syncpoint processing. Syncpoint manager 60D then contact protected conversation adapter 64D with a reply of "request commit". Protected conversation adapter 64D then sends an intersystem architected "request commit" reply to protected conversation adapter 64A (Step 541) which in turn replies "request commit" to syncpoint manager 60A (Step 542). As described above, syncpoint manager 60A received "request commit" from protected file adapters 62A and 62B (Step 535A). Since there are no interruptions in the illustrated example, decision block 543 leads to step 544 in which syncpoint manager 60A contacts the recovery facility 70A to force log 72A to a phase two state of "Initiator, committed" (Step 544). Syncpoint manager 60A then calls each registered protected resource adapter with the phase two decision of "Committed" (FIG. 5b, Step 545). Protected file adapters 62A and 62B process the commit decision as described above (Step 545A). When protected conversation adapter 64A receives the commit decision, it sends an intersystem architected "committed" call over the protected conversation it represents, i.e. the one originally established by application 56A to application 56D (Step 546). Protected conversation adapter 64D receives the "commit" call and replies to syncpoint manager 60D the phase two decision of "commit" (Step 547).
As described above syncpoint manager 60D contacts recovery facility 70D to force log 72D to the phase two state. Because application 56A issued the original commit request which caused application 56D to subsequently issue a commit, and based on the two-phase commit paradigm used in this embodiment, syncpoint manager 60D's phase two state is "Initiator Cascade, Committed" (Step 548). Syncpoint manager 60D contacts protected file adapter 62D with the phase two commit decision. (Step 549). Protected file adapter 62D and its associated resource manager 63D perform commit processing as previously described and returns a reply of "forget". Since there were no interruptions (decision block 550), syncpoint manager 60D contacts resource facility 70D to log in log 72D a state of "Forget" for the syncpoint log record for this logical unit of work (Step 551). "Forget" means that syncpoint processing is complete and the log record can be erased. Syncpoint manager 60D then contacts protected conversation adapter 64D with a reply of "forget". Based on the two-phase commit paradigm used in this embodiment, syncpoint manager 60D increments the logical unit of work identifier by one and returns to application 56D with an indication that the commit completed successfully. (Step 552). Updating the logical unit of work identifier guarantees uniqueness for the next logical unit of work done by the distributed application.
Next, protected conversation adapter 64D sends an intersystem architected "forget" reply to protected conversation adapter 64A which in turn replies "forget" to syncpoint manager 60A (Step 553). As described above syncpoint manager 60A also receives "forget" replies from protected file adapters 62A and 62B (Step 545A). Assuming there are no interruptions, decision block 554 leads to step 555 in which syncpoint manager 60A contacts recovery facility 70A to log in log 72A a state of "forget" for this logical unit of work. Again based on the paradigm of the two-phase commit process being used, syncpoint manager 60A then increments the logical unit of work identifier by one (Step 556). This change guarantees a new unique logical unit of work identifier for the distributed application. Syncpoint manager 60A then notifies application 56A that the Commit request completed successfully. If during the two-phase commit procedure, the syncpoint processing was interrupted in either syncpoint manager 60A, syncpoint manager 60D or both, recovery facility 70A and recovery facility 70D would implement a recovery operation which is represented in the logical flow by steps 557,558 and 559,560 and is more fully described below in Log Name Exchange For Recovery of Protected Resources, Recovery Facility For Incomplete Sync Points For Distributed Application, and Asynchronous Resynchronization of a Commit Procedure.
FIG. 54 is an alternate embodiment to that illustrated in FIG. 2 and can best be described by comparison to FIG. 2. In both FIG. 2 and FIG. 54, application environments, system facilities, and resource managers are distributed. However, in FIG. 2 one physical device, system 50A, contains multiple application environments, 52A,B,C, two resource managers 63A,B, recovery facility 70A and communication facility 57A. FIG. 2 shows that System Control Program 55A contains the conversation manager 53A and the Syncpoint Manager 60A,B,C. System 50A of FIG. 2 can be a mainframe computer and configurations of this type are often called centralized computing. Also, FIG. 2 shows application environments in system 50A connected to application environments in system 50D through a communication network. In contrast. FIG. 54 shows each application environment, system facility and resource manager in a separate physical machine. This configuration is called distributed computing. In this environment systems 90A,B,C, 110E, 114F, and 120G are programmable workstations similar in function but not necessarily similar in size and power to systems 50A,D of FIG. 2. The systems of FIG. 54 are connected by a communication network which, for example, is a local area network (LAN). Application environments 92A,B, and C of FIG. 54 are functionally equivalent to application environments 52A,B, and C of FIG. 2. However, each application environment 92A,B, and C is contained in a separate programmable workstation. Each system control program 95A,B, and C of FIG. 54 is functionally equivalent to system control program 55A of FIG. 2. Each system control program 95A, B, and C contains (a) a Syncpoint Manager 100A, B, or C which is functionally equivalent to Syncpoint Managers 60A,B, and C, (b) execution environment control programs 91A, B, and C which are functionally equivalent to execution environment control programs 61A, B, and C, (c) protected conversation adapters (PCA) 104A, B, and C which are functionally equivalent to PCA 64A, B, and C, (d) resource adapters (RA) 102A,B,C and 103 A,B,C which are functionally equivalent to resource adapters 62A, B, and (e) conversation managers 93A,B,C which are functionally equivalent to conversation managers 53A,B,C and communication facilities 97A,B,C each of which is functionally equivalent to communication facility 57A. However, in the example of FIG. 54, the communication facility is part of each system control program 95A, B, and C and not in its own execution environment. Also in FIG. 54, resource managers 112E and 113F and their respective files/logs 115E,116E and 117F,118F are functionally equivalent to resource managers 63A and 63B and their respective files/logs 78A, 800A and 78B, 800B of FIG. 2. However, resource managers 112E and 113F are each on separate programmable workstations. Recovery facility 121 G and its log 122G in FIG. 54 are functionally equivalent to recovery facility 70A and its log 72A in FIG. 2. However, recovery facility 121G is in a programmable workstation. System 50D of FIG. 54 is the same as system 50D of FIG. 2 and is included to show the versatility of the network. A description of syncpoint processing in this environment can be obtained by substituting the correct numbers from FIG. 54 for the corresponding numbers from FIG. 2 as just described into the syncpoint processing description above. Thus, there are a wide range of computer systems and networks in which the present invention can reside.
It is possible in system 50A, FIG. 2, for recovery facility 70A to become unavailable for a variety of reasons. Accordingly, system 50A provides back-ups. For example, if recovery facility 70A is part of an execution environment which also controls a resource manager and the resource manager encounters a disabling failure, then recovery facility 70A will also become inoperational. In the example illustrated in FIG. 28, system 50A includes more than one execution environment dedicated to a resource manager, and each execution environment containing the resource manager also contains a recovery facility program, although only one recovery facility in a system may be active at one time.
Specifically, FIG. 28 illustrates that in system 50A there are three identical execution environments 52E, 52F and 52G each containing a resource manager (program) 260A, 260B and 260C, respectively. Preferrably, each resource manager 260A, 260B and 260C is an enhanced version of the Shared File System (SFS) resource manager of the VM/SP Release 6 operating system (`VM` is a trademark of the IBM Corp. of Armonk, N.Y.) and associated resources 262A, 262B and 262C, respectively. In addition, each execution environment 52E, 52F and 52G also contains a program 70A, B and C to provide the function of recovery facility 70A illustrated in FIG. 23. An advantage of locating each recovery facility in an execution environment which includes the shared file system is that the shared file system includes services, i.e. communication and tasking services, that the recovery facility can use. The communication services handle communication protocols, interrupt processing, and message management. In system 50A FIG. 28, recovery facility 70A is initially identified to the system control program as the recovery facility associated with recovery facility log 72A when the execution environment 52E is initialized. This is accomplished by specification of a parameter as input to the execution environment 52E's initialization process. Execution environment 52E identifies itself to the system control program as the recovery facility and as the target of all communication in system 50A for the sync.sub.-- point.sub.-- log resource identifier. (Refer to section `Log Name Exchange for Recovery of Protected Resources` for description of term sync.sub.-- point.sub.-- log resource identifier.) This sync.sub.-- point.sub.-- log resource identifier must be unique in system 50A and can be associated with only one execution environment at any time. In the illustrated embodiment, execution environment 52E defines a nonvolatile storage area which contains recovery facility log 72A so that specification of execution environment 52E automatically implies log 72A as the resource recovery log, absent an overruling specification of another storage area.
However, if execution environment 52E is not available, the user can activate recovery facility 70B or 70C as a backup and move log 72A to execution environment 52F or 52G by specifying the aforesaid parameter at initialization of execution environment 52F or 52G and specifying to the execution environment the location of recovery facility log 72A. The user specifies the location of log 72A by giving the system control program the necessary commands from the chosen execution environment 52F or 52G to identify the location of the non-volatile storage area that contains recovery facility log 72A.
All the information that is needed by the recovery facility to complete resynchronization after a syncpoint failure is contained in recovery facility log 72A, and no information required for the syncpoint recovery is contained in the execution environment, resource manager, or associated non-volatile storage. Therefore, any execution environment with the resource manager that contains the recovery facility program can act as the recovery facility 70A as long as the active recovery facility has access to log 72A. The back-up transfer of the recovery facility function to execution environment 52F is indicated by communication path 272B, and the back-up transfer of the recovery facility function to execution environment 52G is indicated by communication path 272C.
Communication between any of the syncpoint managers 60A, 60B, or 60C in any application environment with the recovery facility 70 is accomplished by using the sync.sub.-- point.sub.-- log resource identifier when initiating a conversation through the system control program to the recovery facility.
LOCAL AND GLOBAL COMMIT SCOPES TAILORED TO WORK UNITS
The foregoing flowcharts of FIGS. 5A,B illustrate an example where a single logical unit of work or commit scope extends to two application partners in different systems, for example, to resources and applications in more than one execution environment in different systems, and the commit procedure is coordinated between the two application partners. The following describes in detail this process as well as the ability of System 50A to provide separate work units or commit scopes for the same application in the same execution environment. Thus, all systems 50 can tailor commit scopes to the precise resources which are involved in one or more related work units.
As noted above, a "work unit" is the scope of resources that are directly accessible by one application and participate in a common syncpoint. For example (in FIG. 2), the resources coupled to resource adapters 62A and 62B and protected conversation adapter 64A are all directly accessible by application 56A and therefore, could all have the same work unit. They would all have the same work unit if they all were involved in related work requests made by application 56A. The work unit identifiers are selected by the system control program 55 and are unique within each execution environment. In the illustrated embodiment, the system control program 55A comprises a conversation manager 53A, and an execution environment control program 61 for each execution environment 52. By way of example and not limitation, execution environment control program 61A can be an enhanced CMS component of the VM/SP Release 6 operating system ("VM" is a trademark of IBM Corp. of Armonk, N.Y.). This execution environment control program controls the execution of application 56A and, as noted above, assigns the work unit identifications. Thus, the work unit identifications are unique within each execution environment. The application uses the same work unit for multiple, related work requests and different work units for unrelated work requests. A "logical unit of work" identifier is a globally unique (network wide) identifier for all resources that are involved in related work requests and encompasses all the related work requests. The logical unit of work identifiers are assigned by the recovery facility 70 of the system in which the work request originated and in this embodiment comprises:
(1) A network identifier which identifies a group of interconnected systems;
(2) A system identifier which identifies one communication facility within the network;
(3) An instance number that provides a locally unique element to the LUWID (for example, a timestamp may be used); and
(4) A sequence number which identifies a particular syncpoint instance.
By way of example, this is of the type defined by System Network Architecture LU 6.2 Reference: Peer Protocols, SC31-6808 Chapter 5.3 Presentation Services-Sync Point verbs. The syncpoint manager 60 requests the logical unit of work identifier (LUWID) from the recovery facility when a protected conversation is involved in the work unit or when a two-phase commit procedure will be required, even if the work request does not require a protected conversation. The LUWID may be requested by the resource adapter by calling the syncpoint manager, or by the syncpoint manager by requesting an LUWID at the beginning of commit processing if one has not been acquired yet and it is needed for the commit. As described in more detail below, a work unit is associated with a LUWID when protected resources such as a protected conversation or multiple protected resources are involved in the work unit. A work unit can include a mixture of multiple files and multiple file repositories, other protected resources and other participating resource managers, and protected conversations between different parts of a distributed application. In the case of a protected conversation, a single logical unit of work extends between two or more application partners, even though each application partner assigns a different work unit (within each execution environment) to the same protected conversation and to other resources directly accessed by this application. Thus, each application partner associated with a protected conversation assigns and uses its own work unit locally, but the work units of the two or more application partners refer to the same distributed logical unit of work. It should be noted that each execution environment is ignorant of the work unit identifications assigned by the other execution environment, and it is possible by coincidence only that work units in different execution environments have the same identifier. Work units with the extended scope described above, rather than LUWIDs, are used to define local commit scopes because existing applications can benefit from the extended function with a minimum of change. Changing from work units to LUWIDs would be cumbersome and would require existing applications to change.
FIGS. 6-9 illustrate, by example, a process for establishing different work units and logical units of work for the same application 56A, and another logical unit of work which extends to multiple resources associated with a plurality of application partners 56A and 56D running in different systems 50A and 50D, respectively. In the illustrated example in FIG. 7, application 56A is initiated and obtains a work unit identifier X from execution environment control program 61A (Step 928). The execution environment control program is responsible for selecting a unique work unit identifier within each execution environment. Then, application 56A makes a work request to resource adapter 62A within execution environment 52A to update a file located in resource 78A specifying that the work request is to be made under work unit X, or by default, the work request is assigned to be under a "current work unit" designated by application 56A (Step 930). If the resource adapter requests the LUWID for work unit X (Decision Block 935), then syncpoint manager 60A requests a LUWID from recovery facility 70A to encompass work unit X if one is not already assigned and associates it with work unit X. Then the syncpoint manager returns the LUWID to the resource adapter (Step 936). In the illustrated example in FIG. 6, resource 78A (accessed via resource adapter 62A) is not a protected conversation so Decision Block 937 (FIG. 7) leads to Step 939 in which the resources are updated. If resource adapter 62A was not previously registered for work unit X (Decision block 933), then resource adapter 62A registers with syncpoint manager 60A (Step 934). In the foregoing example, application 56A does not desire to perform additional work under the same work unit (Decision Block 940), and does not desire to do new unrelated work (Decision Block 941), so the next step is for application 56A to issue a commit (Step 942). In response, syncpoint manager 60A initiates the one-phase commit procedure (Step 944). However, it should be noted that application 56A is not required to issue the commit for work unit X before beginning some other unrelated work request (Decision Block 941). In this particular case, the syncpoint manager is performing a one-phase commit procedure and so, does not need a LUWID.
In the illustrated example, application 56A next begins the following process to do work independently of work unit X. Application 56A requests a new work unit from execution environment control program 61A, and execution environment control program 61A returns work unit Y (Step 928). Next, application 56A makes a request to update resource 78B via resource adapter 62B under work unit Y (Step 930). If the resource adapter requests the LUWID for work unit Y (Decision Block 935), syncpoint manager 60A obtains from recovery facility 70A a LUWID and associates it with work unit Y (Step 936). At this time, the logical unit of work for work unit Y extends only to resource manager 63B. Next, an update to resource 78B is implemented (Step 939). Since resource adapter 62B has not yet registered for work unit Y, it registers with syncpoint manager 60A (Step 934).
Next, application 56A desires to do additional work under the same work unit Y (Decision Block 940) e.g. to make changes to data in other resources. In the example illustrated in FIG. 6, the other resource is a protected conversation, and the protected conversation is used to access resources in system 50D via distributed application partner 56D. In the illustrated example, this is the beginning of a new protected conversation. Thus, application 56A initiates a new protected conversation with application 56D under work unit Y (Step 930). Because protected conversation adapter 64A requests the LUWID for work unit Y, the syncpoint manager invokes the recovery facility if a LUWID has not yet been assigned and associated with the work unit, and returns the LUWID to the protected conversation adapter (Step 936). (The protected conversation adapter will need the LUWID when the conversation is initiated (Step 947).) Decision Block 937 leads to Decision Block 946. Because this is a new protected conversation, conversation manager 53A initiates a protected conversation and sends the LUWID associated with work unit Y to a communication facility (Step 947). In the illustrated example, where application partner 56D resides in a different system, communication facility 57A is utilized. However, it should be noted that if the application partner resided in another execution environment, for example 52B, within the same system 50A, then the communication function is provided by conversation manager 53A of system control program 55A, without involvement of communication facility 57A. When protected conversation adapter 64A receives control back from conversation manager 53A and the protected conversation initiation request was indicated as successful, protected conversation adapter 64A registers with syncpoint manager 60A (Step 948) and gives control back to application 56A. At this time application 56A sends a message to application 56D requesting the update of resource 78D (Step 949). However, the message is buffered in system 50D until application 56D is initiated. After the message is sent, application 56A has no more work to do (Decision Blocks 940 and 941) and issues a commit on work unit Y (Step 942). Syncpoint manager 60A initiates a two-phase commit procedure (Step 944).
When system control program 55D receives the conversation initiation request from communication facility 57A via communication facility 57D (Step 960 in FIG. 8), system control program 55D initiates execution environment 52D (Step 962). Protected conversation adapter 64D obtains new work unit Z for execution environment 52D in which application 56D will run from execution environment control program 61D. This work unit is unique within execution environment 52D. Also, protected conversation adapter 64D tells the syncpoint manager to associate the LUWID received with the initiated conversation to the new work unit, and then registers with syncpoint manager 60D under the new work unit (Step 966). (The flow of the conversation initiation request in Step 947 is from protected conversation adapter 64A to conversation manager 53A, to communication facility 57A, to communication facility 57D, to conversation manager 53D, and to protected conversation adapter 64D.) Application 56D is then started.
Next, application 56D makes a work request in Step 930D, and in the illustrated example, the first work request is to receive a message on the conversation. Because the protected conversation already has the LUWID, Decision Block 935D leads to Decision Block 937D. Because this is a protected conversation but not a new outbound protected conversation (i.e., not an initiation of a new protected conversation), Decision Blocks 937D and 946D lead to Step 949D in which the message is received by application 56D.
In the illustrated example from FIG. 6, the protected conversation causes application 56D to perform additional work e.g. update a file within resource 78D (via resource adapter 62D) and therefore Decision Block 940D leads to Step 930D in which application 56D makes a work request to update resource 78D using work unit Z. If the resource adapter requests the LUWID (Decision Block 935D), the syncpoint manager returns the LUWID to the resource adapter (Step 936D). It was not necessary to invoke the recovery facility to assign the LUWID since it was already assigned and associated with the work unit in Step 966. Because this work request does not involve a protected conversation resource, Decision Block 937D leads to Step 939D in which resource 78D is updated according to the work request. Because resource adapter 62D was not previously registered, Decision Block 933D leads to step 934D in which resource adapter 62D is registered with syncpoint manager 60D. Application 56D now needs to determine when application 56A requests the commit of the work. This is accomplished by application 56D by doing a receive (work request) on the protected conversation. Application 56D will get a return code of Take.sub.-- Syncpoint when application 56A has issued the commit. Therefore, Decision Block 940D leads to Step 930D in which application 56D issues a receive on the protected conversation under work unit Z. Since protected resource adapter 64D does not need the LUWID, (Decision Block 935D) and the work request involves a protected conversation (Decision Block 937D) and the protected conversation is not a new outbound conversation (Decision Block 946D), the receive is done (Step 949D). Since application 56D has no additional work to do on work unit Z, Decision Block 940D will lead to Decision Block 941D. When application 56A has issued the commit (Decision Block 941D), application 56D will get a Take.sub.-- Syncpoint return code on the receive, and issue a commit (Step 942D). Next, Syncpoint Manager 60D will initiate the commit procedure (Step 944D). In the illustrated example, this concludes the work request associated with work unit Z, and Decision Block 950D leads to the end of application 56D. At this time, application 56A receives control back from syncpoint manager 60A and ends.
FIG. 9 (and FIGS. 3-5 above) illustrate the timing of the commits in execution environments 52A and 52D according to the example used in this invention. When the protected conversation is in a send state relative to execution environment 52A, application 56A issues a commit for work unit Y, as previously described in Step 942 (FIG. 7). When execution environment 52D is in receive state for the protected conversation, it receives a message along with a return code of Take.sub.-- Syncpoint from execution environment 52A. It should be noted that after receipt of the Take.sub.-- Syncpoint return code, application 56D should issue a commit as soon as possible because this return code indicates that application 56A has issued the commit and is waiting for execution environment 52D to issue the corresponding commit. Thus, after receipt of the message on the protected conversation and the return code, application 56D completes work on other protected resources associated with the work unit in System 50D to get those other resources into a consistent state. After this is done, such that all resources in System 50D associated with the work unit Z are consistent, application 56D issues the commit. Next, syncpoint manager 60A and 60D implement respective two-phase commit procedures for resources directly accessed by the respective applications 56A and 56D. Even though separate commits are invoked to commit those resources which are directly accessed by the respective applications, during the two-phase commit processing each syncpoint manager reports syncpoint status information to the other syncpoint manager. For a more detailed description of syncpoint processing, see Coordinated Sync Point Management of Protected Resources.
REGISTRATION OF RESOURCES FOR COMMIT PROCEDURE
FIG. 10 schematically illustrates automatic and generic registration of resources, where registration is a facility that identifies protected resources to synchronization point manager (SPM) 60. In each application execution environment 52, the resource adapter 62/64 and the SPM 60 participate in registration on behalf of the application 56. In the illustrated embodiment, the resource manager 63 and the resource 78 are located outside of this environment.
In FIG. 10, the application 56 is shown as having two parts, a work request and a commit request. Both parts usually execute in the same application execution environment. However, a broken line between the two parts is shown in the figure to indicate that the application may be distributed and that the two request types may originate from different environments.
Assume that an end user starts application 56 by invoking the start facility of the system control program. The start facility builds the application execution environment 52, and loads and transfers control to the application 56. When the application 56 starts to execute, there are no resources 78 yet registered with SPM 60.
When the application 56 in FIG. 2 makes a work request (steps 500/530 in FIGS. 3/5(A)) to use a resource 78, this request invokes a specific adapter 62/64 associated with the resource 78. The general function of the adapter 62/64 is to connect the application 56 to the resource manager 63. In system 50 the resource adapter 62/64 is extended to include a registration sub-routine that automatically registers in the sync point manger 60, and an adapter sync point exit entry point that supports the two-phase commit procedure.
The work request entry point indicates code lines in the adapter 62/64 that pass the work request (ex. to open a file, insert records into a data base, initiate a conversation, etc.) from the application 56 to the resource manager 63. These code lines also interact with the registration sub-routine in the adapter 62/64 to do automatic registration. Registration informs SPM 60 that the resource 78 is part of a work unit. Also, registration identifies the resource manager 63 to SPM 60. This consists specifically of telling SPM 60 the adapter sync point exit entry point, and the resource manager's object.sub.-- recovery resource identifier.
The adapter sync point exit entry point indicates code lines within the resource adapter 62/64 to be used by the SPM 60's two-phase commit facility when a commit request is made (Steps 506/534 in FIGS. 3/5A). The object.sub.-- recovery resource identifier is the identifier used by the recovery facility 70, described in the below section entitled "Log Name Exchange for Protected Resources" (Step 225 of FIG. 26), to initiate a conversation with the resource manager 63 in the event of a failure during the SPM 60 two-phase commit process.
The process initiated by a work request to any resource adapter 62/64 to handle automatic registration for the application 56 is resource dependent. The resource 78 to be used can be inherently protected regardless of the nature of the work request, and if it has not yet registered, the adapter 62/64 uses its registration sub-routine to automatically register the resource with SPM 60 for the application 56. Alternately the adapter 62/64 may not know if the resource 78 is protected. The resource manager 63 may have this knowledge. In this case, the adapter 62/64 may register and pass the work request to the resource manager 63. The resource manager 63 may do the work request and return to the adapter 62/64 with an indicator whether the resource 78 requires or does not require protection. If protection is not required, the adapter 62/64 may use its registration sub-routine to unregister with SPM 60. Or the adapter 62/64 may determine inherently from the work request or from the resource manager 63 that the resource will not be changed by the application 56; that is, the resource is used only for read. For this case, the adapter 62/64 may use the registration facility of SPM 60 to change the registration to read-only. Finally, the adapter 62/64 may determine that the resource 78 is a read-only resource or an unprotected resource that should be made available to other applications as soon as possible. In this case, the adapter may remain registered in order to obtain the prepare order during a two-phase commit procedure. The resource adapter 62/64 can then use the order as a cue to unlock the resource 78. In this case the adapter 62/64 may respond "prepared" and "commit" to the orders from SPM 60.
By supporting unregistration and change of registration, as described in more detail below, the adapter 62/64 can give information to SPM 60 that allows for optimizing the two-phase commit procedure (also, as described below). When the application 56 issues a commit request, the SPM 60 may realize that only one resource is registered as having been changed (either no other resource is registered, or all other resources are registered as read-only). For this case the SPM 60 may use the more efficient one-phase commit process.
Now consider the foregoing general control flow as applied to a specific example where application 56A of FIG. 2 is executing and makes a work request for a protected conversation with a partner application 56D (Step 530 of FIG. 5A). The request is processed by protected conversation adapter 64A which is one type of resource adapter. This adapter uses its registration sub-routine to invoke the registration facility of SPM 60A (Step 532). Next the adapter 64A uses communication facility 57A, which acts as a resource manager, to initialize the partner application 56D. As illustrated in FIG. 2, the conversation manager 53A is capable of starting a partner application on the same system 50A, or of communicating with a counterpart communication facility 57D on another system 50D via communication facility 57A to start an application within system 50D. In the latter case, the partner application runs on system 50D and the communication facility 57D starts the partner application 56D by invoking the system control program 55D's start facility. This facility builds the new application execution environment 52D for the partner application 56D. Since the start facility knows that it is building a partner application 56D, it knows that the communications facility 57D will be used in the protected conversation with the originating application 56A. Thus, the start facility temporarily acts as the partner application 56D and invokes the resource adapter 64D for protected conversations. Then, adapter 64D registers the protected conversation with the SPM 60D. Thus, the partner application 56D's protected conversation with the originating application 56A is registered prior to the invocation of the partner (alternatively, the registration could be delayed until the partner application 56D uses the conversation with the application 56A). Thus, in FIG. 2, the SPM 60A within execution environment 52A of the application 56A and the SPM 60D within the execution environment 52D of the partner application 56D are each informed of the protected conversation resource.
At this point in the discussion in FIG. 2, the application 56A and the partner application 56D are each executing in their own execution environments 52A and 52D under respective work units, and each may use one or more protected resources 78A or 78D. Each may, for example, use protected files. When the application 56A makes a request to use a file resource 78A, the file resource adapter 62A is invoked. The adapter uses its registration sub-routine to invoke the SPM 60A registration facility. Then the adapter invokes the file resource manager 63A. Thus, again, application 56A's usage of a protected resource 78A is automatically registered. Analogous registrations can be made in execution environment 52D for one or more resources such as resource 78D.
From the above examples we see that this embodiment of registration is generic because registration does not depend on resource type. In FIG. 10, any resource manager 63, that wants to support a protected resources 78 may add the registration subroutine to its resource adapter 62/64. No changes would be required to the system 50 sync point support.
In FIG. 10, the application 56 may also use non-protected resources. For example, the application may want to create a non-protected partner application that periodically displays messages about the work being done, where the display need not be synchronized with the actual completion of work. For this case, the application 56 makes a work request to have a non-protected conversation. The control flow is much the same as for a protected conversation in the above example. The only difference is that the resource adapter 64 knows from information in the work request that the conversation is not protected and in the illustrated embodiment, does not register with the SPM 60. Thus, the non-protected conversation will not participate in the synchronization point processing of SPM 60.
In FIG. 10, given the registration process described above, whenever the application 56 issues a commit request, the SPM 60 has a complete list of protected resources that need to be synchronized. See the foregoing section entitled "Coordinated Sync Point Management of Protected Resources", where the two-phase commit procedure in SPM 60 is described. This shows how SPM 60 uses the adapter sync point exit entry points in the resource adapter 62/64 to use the sync point support in the resource managers 63. Although not shown in FIG. 10, the application 56 may issue a back out request. For this case, the SPM 60 gives a back out order to the adapter sync point exit entry point in the resource adapter 62/64.
At the end of the synchronization point process, each SPM 60 does not destroy the application 56's registration list. It does, however, invoke the resource adapter's exit one more time for post synchronization processing. For this invocation, the adapter may decide to modify its registration. For performance reasons, the adapter may keep the resource registered until the application 56 ends. On the other hand, if the adapter knows that the resource 78 will no longer be used (for example, a protected conversation may end before the application 56 ends), the adapter may use its registration entry point 62 to unregister with SPM 60.
The control flows above assumed distributed resource managers 63. Thus, any request to use a resource 78 always went to the appropriate resource adapter 62/64 which, in turn, invoked the registration facility in SPM 60 and the work request in the distributed resource manager 63. However, for the case where the resource manager 63 is not distributed, the adapter need not get involved with a work request. For this case, since resource manager 63 and SPM 60 are in the same application execution environment 52, the resource manager 63 may directly invoke the registration facility in SPM 60.
In the illustrated example of FIG. 12, application 56A makes multiple work requests. They are processed by system 50A concurrently and involve more than one resource manager and resource. Specifically for the example, application 56A makes eight work requests for two work units, C and D, that are processed concurrently by system 50A. The commit points, shown in FIG. 13, are at times 19 and 44 for work unit C and at time 33 for work unit D. The time units in FIG. 13 are logical clock units denoting sequence (not physical clock units). In the illustration of FIG. 13, events occurring at the same time implies that their order is not important.
A work unit is an application's understanding, or scope, of which resources participate in a synchronization point. An application can specify for which work unit changes to protected resources are made. An application can also specify under what work unit protected conversations are initiated. System 50A permits multiple work units in the application execution environment (52A in FIG. 12). Specifically, applications, sync point manager 60A, and protected adapters (e.g., SQL Resource Adapter in FIG. 12) can support multiple concurrent work units. System 50A also permits tying together the work units of two application execution environments via a protected conversation. Each work unit can have a series of synchronization points. A synchronization point request to a work unit does not affect activity on other work units in an application's environment.
Consider the following example illustrated in FIGS. 12 and 13. Mr. Jones of Hometown wishes to make a transfer to his son's trust fund. The security department for Mr. Jones' bank keeps track of all people involved in any transaction including both customers and employees. The security log and financial records are not in a mutual "all or nothing" embrace but the two work units may need to be processed concurrently--one reason could be that response time would be too slow if the two work units were processed serially.
In the illustrated example, the work request for work unit C at time 1 involves resource manager 63A which controls the security log in the bank's headquarters in Chicago. Unprotected conversation 1 is used by resource adapter 62A to communicate with resource manager 63A. The work request for work unit D at time 1 also involves resource manager 63A in Chicago for Mr Jones' trust fund while the request at time 7 is to resource manager 63B in Hometown where Mr. Jones' other financial records are kept. Unprotected conversation 2 is used by resource adapter 62A to communicate with resource manager 63A and unprotected conversation 3 is used by resource adapter 62B to communicate with resource manager 63B.
When application 56A writes its first record, a "start security event" message, using work unit C, (Step 612 in FIG. 14) resource manager 63A registers via its resource adapter 62A in application execution environment 52A. Sync point manager 60A builds a registry entry for resource manager 63A in FIG. 12 table 126 under work unit C (Step 614). This entry contains the parameter list to pass to the exit for resource adapter 62A which includes the routine name of the exit and a special and private value that resource adapter 62A passed on registration. The resource adapter exit can use the special value to locate its control block for conversation 1.
Consequently, when application 56A requests a commit at time 19 for work unit C, sync point manager 60A reads table 126 to determine which resource adapter exits should be notified to initiate the commit procedure. In the illustrated embodiment, at time 19 when commit is requested for work unit C, synchronization point manager 60A calls the exit routine for resource adapter 62A to initiate a one-phase commit procedure since only one protected resource is registered; resource adapter 62A's exit routine knows to use conversation 1 to communicate with resource manager 63A since it receives from synchronization point manager 60A the special value saved in table 126 during registration.
Registration is subsequently avoided (Step 613) at time 26 when logging the employee id of the bank clerk handling Mr. Jones' transaction. Re-registration is not required because sync point manager 60A already knows from the work unit registration table 126, that resource manager 63A is participating in work unit C. Consequently, the processing of each work request for work unit C after the first work request and the subsequent commit at time 44 is expedited. Also, at each synchronization point for work unit C, only resource adapter 62A and resource manager 63A are notified; there is no time wasted notifying other resource adapters or other resource managers.
When application 56A makes work requests at times 1 and 7 under Work Unit D, both resource adapters 62A and 62B register with sync point manager 60A which adds registry entries 63A and 63B to table 127.
When the first security log commit is done at time 19, the trust fund update at time 17 is not affected in any way. When the trust fund and financial records are committed at time 33, the clerk-id message is not affected either. Note that resource manager 63A in Chicago is not confused since it is communicating on two separate conversations, 1 and 2, to application 56A.
The development of a resource adapter is simplified because system 50A knows which work units are active for the resource manager, relieving the resource adapter of that task. Since the design is simple the resource adapter exit performs well; it has everything it needs and simply sends sync point manager 60A's actions to its resource manager. Another performance perspective is that sync point manager 60A can optimize synchronization point procedures because it knows for which work units the resource manager is active, avoiding the overhead of calling resource adapters or resource managers for resources which are not involved in synchronization points.
In system 50A, there may be occasions when the type of work request made on a protected resource, such as a shared file or database, changes the state of the resource such that the registration information should be changed. This is important because an original work request may be a read-only request and require only a one-phase commit procedure, but a subsequent related work request under the same work unit may be a write request and require a two-phase commit procedure in order to coordinate the multiple protected resources involved.
As another example illustrated in FIG. 3, an application 56A typically makes one or more read requests on a file before making a write request in order to locate a particular record in the file to update. Such read operations can be implemented using a one-phase commit procedure in which case, upon receipt of the read work request by resource adapter 62A (Step 500), the resource adapter registers with syncpoint manager 60A for read mode (Step 502). It should be noted that during subsequent read operations, the resource adapter 62A need not interact with syncpoint manager 60A because there is no change in the type of commit procedure that is required. However, when application 56A subsequently makes a write request to resource adapter 62A under the same work unit (Step 504), resource adapter 62A changes its registration status with syncpoint manager 60A to write mode. As described in more detail below, the rather time-consuming two-phase commit procedure will be used if more than one protected resource is registered for write mode on the same work unit.
This example of registration change is illustrated in detail by the flow chart of FIG. 11. When the work request in step 580 is the first one for the protected resource and the request is read-only, decision block 581 leads to decision block 582. It should be noted that the resource adapter 62A keeps an internal indicator for each resource under each work unit for which it has already registered. This indicator is tested in decision block 581. The resource is not a protected conversation, therefore decision block 582 leads to decision block 583. Because the work is read-only, decision block 583 leads to step 585. In step 585, the corresponding resource adapter 62A registers as a read-only resource. When the next work request to step 580 is to write into, or update, the same resource under the same work unit, decision block 581 leads to decision block 584 because the resource adapter 62A previously registered in step 585, albeit for read mode. Decision block 584 leads to decision block 586 because the resource is not a protected conversation, and decision block 586 leads to decision block 588 because the request is for update mode. Next, decision block 588 leads to step 590 where the resource adapter 62A (which had previously registered in step 585 for read mode) changes its registration within syncpoint manager 60A to write mode. It should be noted that according to FIG. 11, if the first work request under a work unit for the resource is write mode, then the resource adapter 62A registers for write mode in step 592.
There is also the situation of a resource manager 63 which has completed a sync point and has had no further requests since completing that sync point. Its resource adapter 62 is allowed to modify its registration status to "suspended", at the completion of a sync point procedure, so that the sync point manager 60 will know that resource manager 63 is currently not participating in any sync points for the work unit. The suspension of a write mode resource may permit sync point manager 60 to optimize a subsequent commit procedure (one-phase commit) for the remaining resources when, for example, there is only one other write mode resource in the work unit. If the suspended resource adapter 62 receives a new work request for the work unit, it can reactivate its registration through the same registration modification function.
The designs of certain resource managers require that their resource adapters register early in their interaction with the application in order to be notified of distributed sync point activities. However, they may not have a complete set of registration information at that time. For example, the protected conversation adapter 64A needs to register at the point that it initiates a protected conversation with a partner application 56D because it needs to know if a sync point occurs, yet it will not have all registration information until the conversation partner accepts the conversation, an event which may occur much later. This information can be added later under the foregoing change of registration process illustrated in step 590.
System 50 provides additional time-saving techniques in the registration process. When each resource adapter 62 registers a first time with syncpoint manager 60, it registers information in addition to the identification of the resource manager 63 and the resource adapter exit routine name for sync point processing. Much of this additional information usually does not change when the registration changes. Consequently, this additional information is not re-registered when the registration changes in step 590 for a resource adapter 62. The following is a list of some of the additional information which the resource adapter 62 registers only once with the syncpoint manager and which does not change when other registration information changes:
1. Resource and network identifiers which describe where the resource manager and resource are located in the system and the network;
2. Product identifier which indicates the product and thus the type of resource--e.g., shared file, database, protected conversation et.; and
3. Additional data which is required for resynchronization.
Because this additional information is not re-registered each time, the registration process is expedited.
There are a variety of occasions when an application can or will no longer use a protected resource. Examples include such events as end of application, termination of a resource manager, or unavailability of the path to the resource manager. There may be application/resource manager protocols which allow the application to declare a resource to no longer be in use. The application execution environment may support protocols which make it appropriate to unregister resources prior to end of application. Protected conversations may also terminate due to application action or due to an error condition such as a path failure. Upon any such occasion, it is preferable for the resource adapter or protected conversation adapter to unregister all applicable instances of the resource from the syncpoint manager because such unregistration will make subsequent syncpoint processing more efficient (fewer resources to consider and probably less memory consumed) (step 618 of FIG. 14). In addition, the resource adapter or protected conversation adapter can delete any control information about the registered resource and thus be more efficient in its subsequent processing.
FIG. 15 shows the flow of unregistration activity when a resource adapter 62 or a protected conversation adapter 64 discovers that a resource 78 or protected conversation is not available (step 904) or that the application has ended (step 903). Note that the adapter would typically discover that the resource is not available while processing an application work request (step 902). The adapter would determine from its own resource registration status information what registered resources should be unregistered (step 906). For each such registered resource, the adapter would call the syncpoint manager 60 to unregister the resource (step 907). Note that the adapter must identify the resource and the work unit to the syncpoint manager 60.
In FIG. 15, for each call to syncpoint manager 60 (step 910), the syncpoint manager 60 uses the adapter-supplied work unit identifier to locate the work unit resource table (step 911). Within this work unit resource table, the syncpoint manager 60 uses the adapter-supplied resource identifier to locate the desired resource entry (step 912). The syncpoint manager 60 then flags the resource entry as unregistered (step 913) and returns to the calling adapter (step 914 back to step 907). However, the syncpoint manager 60 cannot yet erase the unregistered resource entry because the resource entry logically contains error information which must be preserved until the next synchronization point (see "Coordinated Handling of Error Codes and Information Describing Errors in a Commit Procedure").
The adapter can now delete its control information (or otherwise mark it as unregistered) about the unregistered resource (step 908). Note that an event which causes unregistration may cause multiple resource registrations to be deleted (for example, a resource may be registered for multiple work units). Thus, steps 906, 907, and 908 can be a program loop to handle each applicable previously registered resource. At this point, the adapter can return to its caller (step 909). If the work request has failed due to an unavailable resource, the adapter can report the error condition to the application by whatever mechanism the resource adapter has chosen to return error information to its application users.
The resource adapter may have other processing considerations as a result of the unavailable resource or the application termination. For example, if the unavailable resource condition will cause the backout of resource updates, the adapter will need to notify the application and/or the syncpoint manager 60 that the next syncpoint on the applicable work unit(s) must be a backout. This condition during syncpoint processing requires the adapter to notify syncpoint manager 60 of the resource status (which is backing out). There may be other resource, environment, or implementation dependencies.
Syncpoint manager 60 is now concerned with handling the flagged unregistered resources (from step 913) so that they are ignored for normal operation and so that they are eventually erased. Syncpoint manager 60 can erase flagged unregistered resource entries at the beginning of the next syncpoint for the affected work unit. FIG. 16 describes the syncpoint process flow within syncpoint manager 60. When the next syncpoint process reads the registered resource table (step 622), it can erase any flagged unregistered resource entries in that table (an action not shown in FIG. 16). Because step 622 builds all syncpoint resource participation lists for the duration of the current syncpoint process, resource unregistrations and modifications of resource registry entries by adapters will not affect the current syncpoint process. At this point, the total unregistration process is complete.
OPTIMIZATION OF COMMIT PROCEDURES
Each participating resource manager is capable of performing the two-phase commit procedure, such as the two-phase commit procedure described by System Network Architecture LU 6.2: Peer Protocols, SC31-6808, Chapter 5.3 Presentation Services - Sync Point verbs, and may or may not be capable of performing the one-phase commit procedure. The two-phase commit procedure is important to protect resources; however, the two-phase commit procedure is a relatively complex and time consuming process compared to the one-phase commit procedure. For example, as described in more detail below, the two-phase commit procedure requires the time-consuming step of logging information about the sync point participants in the recovery facility log 72 (FIG. 2), whereas the one-phase commit procedure does not require such logging. Also, the two-phase commit procedure requires two invocations of the resource adapter coordination exit to perform the commit, whereas the one-phase commit procedure requires only one such invocation to commit data. A "resource adapter coordination exit" is the mechanism for the sync point manager 60 (FIG. 2) to provide information to the associated resource manager. The sync point manager utilizes the two-phase commit procedure only when necessary to make the system operate as expeditiously as possible. In summary, the sync point manager utilizes the two-phase commit procedure whenever a protected conversation is involved, or at least two resources are in update mode, or one or more participating resource managers is not capable of performing the one-phase commit procedure. Whenever all resources are capable of performing the one-phase commit procedure and no more than one resource is in update mode, the sync point manager utilizes the one-phase commit procedure. Also, if any resource is in read-only mode such that the data in the resource is read and not updated and the resource manager is capable of performing the one-phase commit procedure, then a one-phase commit procedure is used for this resource regardless of the type of commit procedure used for the other resources. A key component of this optimization is the resource manager's ability and resource adapter's ability to determine prior to the synchronization point its state defined by the work request, that is, whether the resource is in read-only mode or in update mode. When a resource is in read-only mode, it means that the application has only read data from the resource. When a resource is in update mode, this means that the application has changed the data in the resource.
The optimization process begins as follows. Application 56 (FIG. 2) makes a work request to a resource (step 612 of FIG. 14). If this is the first work request for a particular work unit (decision block 613 in FIG. 14), the resource adapter 62 (FIG. 2) associated with the resource registers with the synchronization point manager the fact that it is now an active, participating resource for the work unit (step 615 in FIG. 14). One of the pieces of information about the resource that must be provided at registration time (step 616 in FIG. 14) is whether the associated resource manager is capable of performing the one-phase commit procedure, e.g., is the resource a database manager which under certain circumstances could perform a one-phase commit procedure. Also during registration, the resource adapter records with the sync point manager whether the work request made by the application placed the resource in the read-only mode or update mode (step 616 in FIG. 14).
After the initial registration of a resource, subsequent work requests made by the application against that resource may change the state of the resource. That is, the resource may change from read-only to update mode. When these changes occur, the resource adapter must inform the sync point manager about these changes, and the registration information is updated to reflect the new state (step 619 in FIG. 14).
If the work request from the application is for a protected conversation, the registration entry for the protected conversation adapter will always show that the protected conversation adapter is in update mode and that it is not capable of performing a one-phase commit procedure. Since the protected conversation adapter represents a communication path to another application execution environment, which may involve a plurality of resources, it is not possible for the protected conversation adapter to determine accurately if it represents a communication path to read-only mode resources or to update mode resources. Therefore, the presence of a communication path to another application execution environment requires the two-phase commit procedure, to provide the necessary protection of the critical resources. The protected conversation adapter insures that the two-phase commit procedure will be used by registering as an update mode resource that is not capable of performing the one-phase commit procedure.
After the application has completed all its work, it will attempt to either commit or back out the data at the resources. To accomplish this, the application issues a sync point request to the sync point manager. To start processing the sync point request, (step 620 in FIG. 16) the sync point manager reads the work unit table to find the entry for the affected work unit (step 621 in FIG. 16). For more information on work units, see Local and Global Commit Scopes Tailored To Work Unit. Once the correct work unit entry is located, the sync point manager reads the information in that entry about the resources registered for that work unit and creates three lists of resources (step 622 in FIG. 16).
Each of these lists has a different meaning. The read-only list contains those resources whose data has only been read by the application. The update list contains those resources whose data has been changed by the application and those resources that are in read-only state but whose resource manager is not capable of performing the one-phase commit procedure. The initiator list contains the list of communication partners that have sent a message that they want to synchronize updates to resources. Each resource may appear in only one of the lists.
In practice, the registration for each resource includes two flags which are read by the sync point manager and used to determine if a resource should be entered into the update list or the read-only list. The first flag is on when the resource is in read-only mode, and is off when the resource is in update mode. The second flag is on when the resource supports both the one-phase commit procedure and the two-phase commit procedure, and is off when the resource is capable of performing only the two-phase commit procedure. In practice, the registration for each resource also includes a field that contains information about whether this resource adapter received a message from a communication partner indicating that it wants to synchronize updates to resources. The sync point manager reads this field and uses the data to determine if the resource should be entered into the initiator list.
Once the lists of resources have been built, the sync point manager examines the sync point request type (decision block 623 in FIG. 16). If the sync point request is to back out, the sync point manager performs backout processing as follows. First, all the resource adapters in the update list, if any, are told to back out the changes to their resource (step 626 in FIG. 16). Then, all the resource adapters in the read-only list, if any, are told to back out the effects on their resource (step 627 in FIG. 16). It should be noted that the processing of a "backout" for a read-only resource is defined by the resource implementation, since there are no changes to the actual data in the resource to be backed out. For example, processing for a backout of a read-only file in a shared file resource manager 63 (FIG. 2), could include closing the file and discarding any file positioning information previously maintained for the application's use. After the read-only resources are told to back out, then all the resource adapters in the initiator list, if any, are told that this application execution environment backed out the changes for this synchronization point (step 628 in FIG. 16).
If instead the sync point request is to commit (decision block 623 in FIG. 16), then the sync point manager starts the optimization process for the commit. The first step in the optimization process is to determine if the initiator list is not empty (decision block 624 in FIG. 16). If the initiator list is not empty, this means that this application execution environment is a cascaded initiator in the sync point tree, and that the full two-phase commit procedure must be used for this commit. This is necessary because neither application execution environment knows the full scope of the sync point tree, that is, how many resources are active and in update mode for this synchronization point. Since the number is not known, the two-phase commit procedure must be used, to provide the necessary protection of these critical resources.
If the initiator list is empty (decision block 624 in FIG. 16), the next step is to determine if more than one resource is in the update list (decision block 625 in FIG. 16). If this is true, then the full two-phase commit procedure must be used for this commit. The two-phase commit procedure provides more protection for the update mode resources, because no resource commits its changes until all resources have voted that they can commit their changes.
If there are less than two resources in the update list (decision block 625 in FIG. 16), the next step is to determine if there are zero or one resources in the update list 640 (FIG. 16). If there are zero resources in the update list, then the one-phase commit procedure will be used to commit the read-only resources. Likewise, if there is exactly one resource in the update list, and its resource manager is capable of performing the one-phase commit procedure, then the one-phase commit procedure will be used.
The one-phase commit procedure starts by the sync point manager telling the resource adapters in the update list, if any, to commit their changes (step 641 in FIG. 16). It should be noted that the one-phase commit of data by the resource manager is achieved by only one invocation of the resource adapter, in contrast with the two invocations needed during the two-phase commit procedure. Since there can be only zero or one resources in update mode in the entire synchronization point, there is no chance of data inconsistency caused by different decisions for different resources. Also note that during the one-phase commit procedure, there is no writing to the recovery facility log 72 (FIG. 2), as opposed to the required logging that is part of the two-phase commit procedure (steps 644, 648, 651, 658, 659 of FIG. 17). The one-phase commit procedure ends with the sync point manager telling the resource adapters in the read-only list, if any, to commit their changes (step 642 in FIG. 16). It should be noted that a "commit" of a read-only resource is defined by the resource implementation, since there are no actual changes to the data to be committed. For example, some shared file resource managers 63 (FIG. 2) provide read consistency, so when an application reads a file in a shared file resource manager, the application is provided with a consistent image of the file, that is, changes made to the file by other application environments will not interfere with the reading of the contents of the file, as they existed at the time the file was opened. When the application opens the file with the intent of read, the image is created by the resource manager, which is considered to be a read-only resource. When the application is done reading the file, it closes the file and attempts a commit. When the shared file resource manager performs the commit as a read-only resource, it could discard the image maintained for the application's use. Now, if the application opens the file again, it will see an image of the file which contains all committed updates made by other applications.
If the sync point request results in a two-phase commit procedure according to the outcome of decision blocks 624, 625, or 640 of FIG. 16, the sync point manager 60 (FIG. 2) still optimizes the commit of the read-only resources. There are several parts to this optimization for the read-only resources. First, (step 644 of FIG. 17) information about the read-only resources is not written to the recovery facility log 72 (FIG. 2). Information about the read-only resources does not have to be logged at the recovery facility 70 (FIG. 2) because the read-only resources will never log the state of "In-doubt" on their own logs. This means that the resource manager will never attempt to resynchronize with the recovery facility 70 (FIG. 2), so the recovery facility does not need any knowledge about the resource. Second, the read-only resources are not involved in the first phase of the commit, which is sending prepare to all resource adapters in the update list (step 645 of FIG. 17). The actions of a read-only resource cannot affect the protection of the resources, since in terms of data consistency, a backout is equivalent to a commit for a read-only resource.
The only time that the read-only resources are involved in the two-phase commit procedure is when they are told the final direction of the commit, that is, they are told whether to commit their changes (step 653 of FIG. 17) or told to back out their changes (step 655 of FIG. 17).
The following is an example of a two-phase commit procedure involving three different application execution environments, which are part of a system such as System 50 (FIG. 2). Each application execution environment is executing a different application. Application A and Application B are communicating via a protected conversation; Application B and Application C are communicating via a protected conversation. The two-phase commit procedure is started when Application A attempts to commit by issuing a commit request B1 (FIG. 18) to the sync point manager which is currently running in the same execution environment as Application A. Phase one starts when the sync point manager writes the SPM Pending log record to the recovery facility log B2 (FIG. 18). The SPM Pending log record contains the logical unit of work identifier for the synchronization point and information about the synchronization point participants, in this case, the SPM Pending record shows one participant, Application B.
After the SPM Pending log record is successfully written to the recovery facility log, the sync point manager sends a prepare message via the protected conversation adapters to Application B. Application B is notified that its conversation partner Application A wishes to synchronize resources, and Application B subsequently issues a commit request B3 (FIG. 18) to the sync point manager which is currently running in the same execution environment as Application B.
For the sync point manager at B, the first phase of the two-phase commit procedure starts by writing the SPM Pending record to the recovery facility log B4 (FIG. 18). The SPM Pending record contains the logical unit of work identifier for the synchronization point and information about the synchronization point participants. In this case, the SPM Pending log record contains information about Application A, showing it as the synchronization point initiator, and Application C as a synchronization point participant. Once the SPM Pending log record is successfully written to the recovery facility log, the sync point manager sends a prepare message via the protected conversation adapters to Application C. Application C is notified that its conversation partner Application B wishes to synchronize resources, a |
