PostgreSQL allocates memory within memory contexts, which provide a convenient method of managing allocations made in many different places that need to live for differing amounts of time. Destroying a context releases all the memory that was allocated in it. Thus, it is not necessary to keep track of individual objects to avoid memory leaks --- only a relatively small number of contexts have to be managed. palloc and related functions allocate memory from the "current" context.
SPI_connect creates a new memory context and makes it current. SPI_finish restores the previous current memory context and destroys the context created by SPI_connect. These actions ensure that transient memory allocations made inside your procedure are reclaimed at procedure exit, avoiding memory leakage.
However, if your procedure needs to return an allocated memory object (such as a value of a pass-by-reference data type), you can't allocate the return object using palloc, at least not while you are connected to SPI. If you try, the object will be deallocated during SPI_finish, and your procedure will not work reliably!
To solve this problem, use SPI_palloc to allocate your return object. SPI_palloc allocates space from "upper Executor" memory --- that is, the memory context that was current when SPI_connect was called, which is precisely the right context for return values of your procedure.
If called while not connected to SPI, SPI_palloc acts the same as plain palloc.
Before a procedure connects to the SPI manager, the current memory context is the upper Executor context, so all allocations made by the procedure via palloc or by SPI utility functions are made in this context.
After SPI_connect is called, the current context is the procedure's private context made by SPI_connect. All allocations made via palloc/repalloc or by SPI utility functions (except for SPI_copytuple, SPI_copytupledesc, SPI_copytupleintoslot, SPI_modifytuple, and SPI_palloc) are made in this context.
When a procedure disconnects from the SPI manager (via SPI_finish) the current context is restored to the upper Executor context, and all allocations made in the procedure memory context are freed and can't be used any more!
All functions described in this section may be used by both connected and unconnected procedures. In an unconnected procedure, they act the same as the underlying ordinary backend functions (palloc etc).