I notice that we have in md.c
TRACE_POSTGRESQL_SMGR_MD_READ_DONE(forknum, blocknum, reln->smgr_rnode.spcNode, reln->smgr_rnode.dbNode,
reln->smgr_rnode.relNode,relpath(reln->smgr_rnode, forknum), nbytes, BLCKSZ);
TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum, reln->smgr_rnode.spcNode, reln->smgr_rnode.dbNode,
reln->smgr_rnode.relNode,relpath(reln->smgr_rnode, forknum), nbytes, BLCKSZ);
relpath() returns a palloc'd string, which will not get freed, which
means that any serious use of these probe points will shortly blow out
the backend's memory. So far as I can see the path argument is
redundant with the other probe arguments and we might as well just
remove it --- objections?
There's another problem in tuplesort.c:
TRACE_POSTGRESQL_SORT_DONE(state->tapeset, (state->tapeset ? LogicalTapeSetBlocks(state->tapeset) :
(state->allowedMem- state->availMem + 1023) / 1024));
This is called after state->tapeset has been freed, which means that the
LogicalTapeSetBlocks call will very likely give a wrong answer,
especially in assert-enabled builds but maybe even in a regular one.
Seems that we need to have been quality-checking the dtrace patches
a bit more carefully.
regards, tom lane