Some ideas for improvements Generate a copy of the library without shadowing stdlib Confirm that sample_request has the correct distribution Test tracked_alloc's ability to detect tracked allocs Confirm that tracked_alloc never hands out overlapping addresses Consider injecting faults? Help detect heap corruption Let tracked allocations float by a couple pages away from 1G alignment Insert guard pages before and after tracked allocations Shift allocations to make them flush with the trailing guard page Supplement guard pages with noise filling to detect short overflows Scan DieHard / DieHarder for more tricks Improve statistical sampling Track the actual number of bytes before each sampled allocation Using that be easier than correcting for large allocations.
Generate from an Exponential without libm's log Parse a configuration file for, e.g., the sample period Open a (abstract) unix socket for online configuration Vary sample rate based on dynamic context (e.g., hash(caller)) Real argument parsing in poireau.py Consider a small command shell in poireau.py Let poireau.py ingest stack trace patterns for known leaks Report a random number + pid in tracepoints Hook malloc-adjacent utilities like malloc_usable_size Hook posix_memalign, etc. Hook jemalloc/tcmalloc/etc. extensions Investigate need for fork hooks Especially if we start listening on a unix socket.
Dynamically size the allocation tracking table Automatically find a good age threshold Online streamgraph visualisation Flamegraph of average byte * second (time averaged snapshot) Sample on pure call count? Detect allocation / deallocation pairs in different threads Fully disable allocation tracking until explicitly enabled We want to PRELOAD with minimal overhead until we actually need the
tracking allocator.
Explore ways to work better with static linking List of candidates for related tool names