CSL Notebook Topic To CSL Date August 10, 1984 From Mike Kupfer Location PARC Subject Internal structure of Alpine Locks Organization CSL XEROX Release as [Indigo]Release52>Doc>LockInternals.press Came from [Indigo]Release52>Doc>LockInternals.tioga Last edited by Kupfer, August 11, 1984 2:37:26 pm PDT Abstract The Alpine Lock Manager controls a complicated web of variant records. This memo describes the variant record structure and gives an example taken from a running Alpine server. It may be necessary to read the memo once, look at the example, and then read the memo again. Introduction The Alpine file service, as designed, supports atomic transactions that span multiple servers. This document briefly explains the high-level interface to Alpine locks, and it explains the complicated data structure that each Alpine server's Lock Manager maintains. Lock interface Alpine names a lock with a LockID (defined in AlpineInternal.mesa). Unlike, say, a transaction ID, a LockID always refers to exactly one object, such as a file or a page, even across events such as server crashes. A client of the Lock interface is some other module in Alpine, acting on behalf of a transaction. This client may request that the lock manager Set a lock in a given mode. If there is already a conflicting lock set, the client may specify whether it wishes to wait until the lock is available, or whether it wishes to be informed immediately that someone else has the lock. A lock may be held by more than one transaction, and a transaction may hold more than one lock. Given the lock, the transaction may Release it when it no longer needs it. The transaction manager also has the power to abort a transaction and take all of its locks away, usually so that the locks can be given to other transactions. What the Lock Manager needs Suppose a transaction requests the lock MyLock. The Lock Manager must quickly find this lock in a potentially large jungle of other locks, determine if anyone else has MyLock, and determine whether the two transactions can both have MyLock at the same time. If there is a conflict and the requesting transaction wishes to wait, the Lock Manager must place the request in a queue, waiting for the lock's eventual availability. Suppose that a transaction completes (either commits or aborts). The Lock Manager must find all of the locks owned by the transaction and free them. If another transaction is waiting for one of these locks, the Lock Manager should give the lock to this second transaction. Finally, consider the maintenance work that the Lock Manager must do. It should be able to detect a circular chain of requests (deadlock), and it should be able to scan waiting requests to find cases of livelock (aka starvation). It should be able to perform these tasks quickly and with little overhead, so that problems are corrected quickly and with little performance loss. The Solution The Alpine Lock Manager uses a single data structure to meet these varied needs. This data structure consists of many variant records, all of which are tied together by assorted REF links into circular and simple singly-linked lists. LockInternal.Object The variant record that the Alpine Lock Manager uses is defined in LockInternal.mesa. A Lock Object always contains a requestList REF and a body record. The two main variants of a Lock Object -- request and header -- are distinguished by varying the body. We shall refer to these two variants as "request objects" and "header objects," respectively. The requestList REF links objects associated with one lock into a circular list, as we will explain below. Header Object Recall that an Alpine LockID uniquely identifies one object that is to be locked (eg., a page or a file). When a transaction requests a lock, the Lock Manager enters a request object for the transaction on a circular list associated with the desired LockID. This circular list is tied together with the requestList REF, as stated above. A header object also sits in the circular list, and it identifies which lock is being requested. The Lock Manager refers to header objects via a hash table, using the lockID as the hash key. The next field is used by the hash package to connect header objects that hash to the same location. Request Object A request object always contains a transaction handle, a transList REF, and mode information. The transaction handle tells what transaction the request object is associated with. We will describe the transList REF below. The mode tells what locking mode has been requested. The rest of a request object is yet another variant record, consisting of either a granted, a waiting, or a transHeader variant. waiting variant When a transaction must wait for a lock, the Lock Manager creates a "waiting" variant for the transaction. All of the "waiting" request objects are tied in a singly-linked, non-circular list through the transList REF. This list is used, for example, by the Lock Watchdog to find lock requests that have timed out. granted variant When the Lock Manager grants a request, it makes a "granted" variant for the transaction. This object sits in a cicular list containing all granted requests for the transaction. The list itself is bound together via the transList REF. transHeader variant In addition to the transaction's granted requests, the transList-threaded circular list contains a transHeader. This object identifies the transaction and contains a count of the (granted) request objects in the list. An Example The accompanying diagram (filed in [Indigo]Release52>Doc>LockExample.griffin) was derived by tracing pointers on a running Alpine server (Monitor.alpine). It shows the lock objects for two transactions, numbers 402 and 1002. There are four locks involved, as shown by the four header objects: A, B, C, and D. The transactions A transaction's locks are associated with its worker(s), so "self," in the upper left corner, is a Worker.Handle for transaction 402. It points to the circular list of all the locks that 402 has been granted. Because the requestList REF is present in all Lock objects but doesn't have a useful function in a transHeader, it is NIL in the transHeader. The transHeader for 1002 is at the bottom of the diagram, in the middle. Transaction 402 has been granted all four locks (A, B, C, and D). Transaction 1002 has been granted one lock (D), and is waiting for another (C). Transaction 1002 is waiting because it is requesting a mode that is incompatible with the mode that 402 has locked C with. The locks Lock A has been granted to one transaction (402). It is the only lock in its hash bucket, so its "next" field is NIL. Lock B has been granted to 402. B's and C's lockID's hash to the same value, so B is linked to C via "next". Lock C has been granted to 402, and 1002 is waiting for it. The "waiting" request object for 1002 is the last one in the Lock Manager's list of waiting objects. Lock D has been granted to both 402 and 1002. ส– "Cedar" style˜Icenter•Mark centerHeaderšะbl˜ImemoHeadšฯsœžœ˜!Lšžœžœ˜ Lšž œ#ž œ˜