Add finalizer to ZStream and make TranscodingStreams.finalize a noop

[nhz2]

Fixes #48 and fixes #88 by porting JuliaIO/CodecBzip2.jl#43 to this package.

According to https://www.zlib.net/manual.html#Basic

If zlib is used in a multi-threaded application, zalloc and zfree must be thread safe. In that case, zlib is thread-safe.

So, there shouldn't be any thread safety issues with the finalizer here.

[codecov]

Codecov Report

Attention: Patch coverage is 96.66667% with 1 line in your changes missing coverage. Please review.

Project coverage is 84.84%. Comparing base (f29c86f) to head (f50e6f9).

Files with missing lines	Patch %	Lines
src/libz.jl	91.66%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #97      +/-   ##
==========================================
+ Coverage   83.51%   84.84%   +1.33%     
==========================================
  Files           4        4              
  Lines         188      198      +10     
==========================================
+ Hits          157      168      +11     
+ Misses         31       30       -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[mkitti]

Is there a reason we use ccall here rather than Libc.malloc ?

[nhz2]

This is to make sure the memory being used is visible to the GC https://discourse.julialang.org/t/c-struct-garbage-collection-not-run-frequently-enough/116919/14

If I replace this with Libc.malloc Julia uses more memory in the memory leak test, because GC isn't run frequently enough.

[mkitti]

There is some reference within the C code suggesting that calloc may be warranted here.

https://github.com/madler/zlib/blob/develop/deflate.c#L397

The comments here suggest that malloc may be an appropriate optimization on modern systems.

madler/zlib#517

[nhz2]

Yes, it looks like calloc is only needed if pointers are 16 bits.

[mkitti]

Per my Slack comment, consider adding and encouraging a do syntax to explicitly finalize the compressor rather than relying on garbage collection to work. GC will only invoke the finalizers under memory pressure.

[nhz2]

Unlike with files, or GPU arrays, the only resource being used here is CPU memory, so unless I am missing something it is fine if the freeing is delayed until there is memory pressure.

The issue with the do syntax is that it may be hard to prove that the compressor doesn't escape the block, and there are some use cases where the compressor needs to escape the block. For example, the memory could be reused by another compressor with the same options: Arrow.jl has some pools to do this https://github.com/apache/arrow-julia/blob/main/src/Arrow.jl#L88-L92. Also, it would also be interesting to try and use https://github.com/MasonProtter/Bumper.jl if there are multiple compressors that all have the same lifetime, so each compressor doesn't need to be individually freed.

[mkitti]

There is a long thread about this by Jeff here:
JuliaLang/julia#11207

The point of the do syntax is not to prevent leakage. The point is to syntactically pair construction and destruction.

I think the use of the finalizer here is fine as a last resort measure, but we should not depend on it. Memory will appear to leak until something actually activates the GC, which may be never.

[mkitti]

What happens if this gets called twice?

[nhz2]

Nothing.

InflateEnd starts by checking if zstream.state is null and returns Z_STREAM_ERROR if it is https://github.com/madler/zlib/blob/8a844d434f0eef87d972ae6406b00968f7c52944/inflate.c#L1268-L1269

InflateEnd also ends by setting zstream.state to null. https://github.com/madler/zlib/blob/8a844d434f0eef87d972ae6406b00968f7c52944/inflate.c#L1273

The same happens in DeflateEnd https://github.com/madler/zlib/blob/8a844d434f0eef87d972ae6406b00968f7c52944/deflate.c#L1266-L1283

Also, since the ZStream constructor initializes .state to C_NULL, the finalizer can run on uninitialized streams without doing anything.

[mkitti]

Since zfree is just a pass through to jl_free perhaps you should just provide the function pointer directly:

cglobal(:jl_free).

[nhz2]

They have different signatures. zfree takes a pointer to an allocator context and a pointer to free, while jl_free just takes a pointer to free.

[nhz2]

IIUC, the problem in JuliaLang/julia#11207 that is relevant here is that finalizers can slow down when there are many different objects with finalizers.

I can try and benchmark with ZipArchives.jl where I create new compressors for each new file that needs to be compressed.

When this PR is merged, I can safely reuse the compressor if the new file has the same compression level as the previous file without worrying about memory leaks or use after free, so I will benchmark that as well.

GC never running sounds like a good thing, no need to waste cycles freeing memory if a process is short-lived and never uses much memory.