Transactions and resource tracking ================================== Assorted notes: - A transaction is tied to a thread which has created it. A transaction can create a sub-transaction, so every thread keeps a stack of running transactions in its per-thread data. - Every transaction keeps a resource pool containing resources, which have been created during the transaction. Whenever the transaction is running, this pool is set as current. You are allowed to switch to a different pool, but please do so carefully. + When a transaction ends, the pool is destroyed and the previous active pool is popped off the transaction stack. The fate of the resources inside the pool depends on the operation used to end the transaction: * commit -- all resources are detached from the pool * rollback -- all resources are freed * fold -- instead of destroying the pool, it is added as a subpool to the parent transaction (which must exist) - Each transaction also includes a memory pool, from which all temporary structures (including all resources created by the transaction) are allocated. Feel free to allocate your temporary data from this pool, too; they will be freed when the transaction is committed or rolled back. When the transaction ends with a fold, this pool gets included inside the parent transaction's pool. + (To be true, there is actually a shared transaction pool per thread and the transaction logic uses @mp_push() and @mp_pop() to keep a stack of per-transaction data.) - Transactions are usually used together with exceptions (which are similar to how exceptions work in other languages, but they differ in subtle details, so please read carefully). When a failure condition of some kind is detected, an exception is raised. It involves creating an exception object and jumping out of the transaction by a `longjmp()`. The exception object (`struct exception`) contains an identification of the error and possibly additional data. + Usually, creation of an transaction and handling of exceptions is done using helper macros (it is not strictly necessary, but highly recommended): + TRANS_TRY { // Code that runs inside the transaction. } TRANS_CATCH(x) { // When an exception is raised, execution continues here. } TRANS_END; The code inside the transaction ends with an implicit @trans_commit(). If you want to end the transaction in a different way, you can do so, but you need to use a `break` statement to skip the implicit commit. The exception handling code gets a local variable `x` pointing to the exception object. When the exception is handled (for example, an error message is logged), @trans_caught() is called automatically, which rolls back the transaction and frees all its resources. Again, you can use the `break` statement to skip this. Alternatively, when you are in a nested transaction, you can throw a different exception or re-throw the original one. This raises an exception in the context of the parent transaction. In this case, the child transaction is not rolled back, but its pools are folded as sub-pools of the parent transaction and kept until @trans_caught() is called finally. * When an exception is thrown outside a transaction, it is converted to a plain @die(). * Memory management and lifetime of various objects and pools deserve special attention, as usually when non-local jumps are taking place. When an exception is raised, the exception structure is allocated from the memory pool of the current transaction. When the exception is propagated through the stack of transactions, no transaction is ever rolled back -- all of them are folded and their pools remain accessible until @trans_caught() is called at the end. Therefore exceptions can carry pointers to the objects which have failed without a risk of the object becoming invalid. However, you need to avoid pointing to on-stack data like local variables of functions, because these are of course destroyed during the `longjmp()`. FIXME: Interaction between exceptions, pools and other libucw modules.