Possible SIGSEGV durning aset memory tracking

The crash path runs through src/backend/utils/mmgr/aset.c: AllocSetContextCreateInternal, AllocSetDelete and AllocSetTransferAccounting all cast the parent context to AllocSet without checking its type. Reproduce with the test extension linked in the issue, starting from the backtrace through MemoryContextSetParent in mcxt.c. Done means the reproduction no longer crashes and the parent-type checks are in place, though the issue leaves open whether to add validation or move accounting into mcxt.c, which is a design decision for maintainers.

Memory context allocation mostly tracked at aset.c and the code relied on the assumption that context tree is an AllocSet context tree, previously that was correct, but now the next query:

-- table test contains only i int column
select * from test order by i;

may allocate a Generation context (from tuplesort_begin_batch) in the context tree and this commit adds additional context validation.
But seems it's not enough:

• AllocSetContextCreateInternal doesn't validate parent context for it's type: here, and here, so after invalid cast there would be invalid pointers
• AllocSetDelete doesn't validate parent context type, it's only perform unconditional cast - after invalid cast there may be memory corruption
• AllocSetTransferAccounting after finding parent it's type isn't checked

Suppose that at some point there would be Generation context under the AccountingRoot (aset) context and over leaf Aset context - there would be performed an invalid type cast with further invalid writes which leads to SIGSEGV. Such cases may seems insubstantial, but postgres code base may mix context in further releases.

Here is a test extension which leads to SIGSEGV:

#include "postgres.h"

#include "access/htup_details.h"
#include "commands/async.h"
#include "commands/trigger.h"
#include "executor/spi.h"
#include "lib/stringinfo.h"
#include "utils/rel.h"
#include "utils/memutils.h"
#include "utils/syscache.h"

PG_MODULE_MAGIC;

PG_FUNCTION_INFO_V1(test_ctx);

Datum
test_ctx(PG_FUNCTION_ARGS)
{
	MemoryContext newAset = AllocSetContextCreate(CurrentMemoryContext,
												 "test preparent aset",
												 ALLOCSET_DEFAULT_SIZES);

	// create and remove context, so further context would be created
	// from freelist
	MemoryContext deleteAse = AllocSetContextCreate(CurrentMemoryContext,
												 "to drop aset",
												 ALLOCSET_DEFAULT_SIZES);
	MemoryContextDelete(deleteAse);

	MemoryContext newGen = GenerationContextCreate(newAset, "test mid",
													SLAB_DEFAULT_BLOCK_SIZE,
													SLAB_DEFAULT_BLOCK_SIZE,
													SLAB_DEFAULT_BLOCK_SIZE
	);

	MemoryContext asetLast = AllocSetContextCreate(newGen,
												"last aset",
												 ALLOCSET_DEFAULT_SIZES);

	MemoryContextDelete(newGen);
	elog(INFO, "last ctx addr %p", (void*)asetLast);
	PG_RETURN_VOID();
}

And here is a link to full test extension

The backtrace

#0  __pthread_kill_implementation (threadid=<optimized out>, signo=11, no_tid=0) at ./nptl/pthread_kill.c:44
#1  __pthread_kill_internal (threadid=<optimized out>, signo=11) at ./nptl/pthread_kill.c:89
#2  __GI___pthread_kill (threadid=<optimized out>, signo=signo@entry=11) at ./nptl/pthread_kill.c:100
#3  0x00007e41fb245b7e in __GI_raise (sig=11) at ../sysdeps/posix/raise.c:26
#4  0x00005865986bae91 in StandardHandlerForSigillSigsegvSigbus_OnMainThread (processName=0x586598e5dabb "Master process", postgres_signal_arg=11) at elog.c:5547
#5  0x0000586598487a83 in CdbProgramErrorHandler (postgres_signal_arg=11) at postgres.c:3990
#6  0x0000586598c8e953 in wrapper_handler (postgres_signal_arg=11) at pqsignal.c:90
#7  <signal handler called>
#8  AllocSetTransferAccounting (context=0x5865b8655080, new_parent=0x0) at aset.c:1874
#9  0x000058659870899e in MemoryContextSetParent (context=0x5865b8655080, new_parent=0x0) at mcxt.c:596
#10 0x00005865987086f5 in MemoryContextDeleteImpl (context=0x5865b8655080, sfile=0x586598ec3e7f "mcxt.c", func=0x0, sline=497) at mcxt.c:468
#11 0x000058659870879d in MemoryContextDeleteChildren (context=0x5865b864f040) at mcxt.c:497
#12 0x00005865987086c8 in MemoryContextDeleteImpl (context=0x5865b864f040, sfile=0x7e41fbdcf021 "ctx.c", func=0x0, sline=40) 

The problem occurs on the debug build with --enable-cassert

For now I suppose the problem should be closed by additional validation, but seems more correct way is to move the logic of memory tracking from aset to mctx and also extend Generation context methods to support tracking.