Project

General

Profile

Actions

Feature #20470

closed

Extract Ruby's Garbage Collector

Added by peterzhu2118 (Peter Zhu) 8 months ago. Updated 6 months ago.

Status:
Closed
Assignee:
-
Target version:
-
[ruby-core:117765]

Description

Extract Ruby's Garbage Collector

Background

As described in [Feature #20351], we are working on the ability to plug alternative garbage collector implementations into Ruby. Our goal is to allow developers and researchers to create and experiment with new implementations of garbage collectors in Ruby in a simplified way. This will also allow experimentation with different GC implementations in production systems so users can choose the best GC implementation for their workloads.

Implementation

GitHub PR: 10721

In this patch, we have split the current gc.c file into two files: gc.c and gc_impl.c.

gc.c now only contains code not specific to Ruby GC. This includes code to mark objects (which the GC implementation may choose not to use) and wrappers for internal APIs that the implementation may need to use (e.g. locking the VM).

gc_impl.c now contains the implementation of Ruby's GC. This includes marking, sweeping, compaction, and statistics. Most importantly, gc_impl.c only uses public APIs in Ruby and a limited set of functions exposed in gc.c. This allows us to build gc_impl.c independently of Ruby and plug Ruby's GC into itself.

Demonstration

After checking out the branch, we can first configure with --with-shared-gc:

$ ./configure --with-shared-gc
...
$ make -j
...

Let's now change the slot size of the GC to 64 bytes:

$ sed -i 's/\(#define BASE_SLOT_SIZE\).*/\1 64/' gc_impl.c

We can compile gc_impl.c independently using the following commands for clang or gcc (you may have to change the last -I to match your architecture and platform):

$ clang -Iinclude -I. -I.ext/include/arm64-darwin23 -undefined dynamic_lookup -g -O3 -dynamiclib -o libgc.dylib gc_impl.c
$ gcc -Iinclude -I. -I.ext/include/x86_64-linux -Wl,-undefined,dynamic_lookup -fPIC -g -O3 -shared -o libgc.so gc_impl.c

We can see that by default, the slot size is 40 bytes and objects are 40 bytes in size:

$ ./ruby -e "puts GC.stat_heap(0, :slot_size)"
40
$ ./ruby -robjspace -e "puts ObjectSpace.dump(Object.new)"
{"address":"0x1054a23f0", "type":"OBJECT", "shape_id":3, "slot_size":40, "class":"0x10528fd38", "embedded":true, "ivars":0, "memsize":40, "flags":{"wb_protected":true}}

We can now load our new GC using the RUBY_GC_LIBRARY_PATH environment variable (note that you may have to change the path to the DSO):

$ RUBY_GC_LIBRARY_PATH=./libgc.dylib ./ruby -e "puts GC.stat_heap(0, :slot_size)"
64
$ RUBY_GC_LIBRARY_PATH=./libgc.dylib ./ruby -robjspace -e "puts ObjectSpace.dump(Object.new)"
{"address":"0x1038de440", "type":"OBJECT", "shape_id":3, "slot_size":64, "class":"0x10355fc00", "embedded":true, "ivars":0, "memsize":64, "flags":{"wb_protected":true}}

Benchmark

Benchmarks were ran on commit c78cebb on Ubuntu 22.04 using yjit-bench on commit cc5a76e.

Compiling gc_impl branch without --with-shared-gc (i.e. how the default Ruby is built), the benchmarks show little to no decrease in performance, with most of it being 0% to 1% slower:

--------------  -----------  ----------  ------------  ----------  ---------------  --------------
bench           master (ms)  stddev (%)  gc_impl (ms)  stddev (%)  gc_impl 1st itr  master/gc_impl
activerecord    73.9         0.3         74.6          0.3         1.00             0.99          
chunky-png      911.1        0.2         937.2         0.2         0.97             0.97          
erubi-rails     1582.4       0.1         1583.5        0.0         1.00             1.00          
hexapdf         2716.2       1.1         2760.2        0.7         1.00             0.98          
liquid-c        68.9         0.5         68.6          0.4         1.00             1.00          
liquid-compile  67.9         0.1         68.2          0.2         0.99             1.00          
liquid-render   172.8        0.1         174.9         0.1         0.99             0.99          
lobsters        1033.9       0.4         1036.0        0.3         1.08             1.00          
mail            135.1        0.2         136.5         0.2         0.99             0.99          
psych-load      2250.8       0.1         2274.9        0.3         0.99             0.99          
railsbench      2499.2       0.2         2502.9        0.1         1.00             1.00          
rubocop         178.3        0.5         179.8         0.4         1.00             0.99          
ruby-lsp        116.8        0.1         118.5         0.2         1.00             0.99          
sequel          75.4         0.2         76.2          0.3         0.99             0.99          
--------------  -----------  ----------  ------------  ----------  ---------------  --------------

Compiling gc_impl branch with --with-shared-gc and loading Ruby's current GC using RUBY_GC_LIBRARY_PATH, the benchmarks are still fairly good with performance decrease of only around 1% to 2%:

--------------  -----------  ----------  ------------  ----------  ---------------  --------------
bench           master (ms)  stddev (%)  gc_impl (ms)  stddev (%)  gc_impl 1st itr  master/gc_impl
activerecord    74.2         0.2         75.4          0.5         0.98             0.98          
chunky-png      916.3        0.3         933.2         0.1         0.98             0.98          
erubi-rails     1597.6       0.1         1586.3        0.2         1.01             1.01          
hexapdf         2731.4       0.5         2776.8        0.7         1.00             0.98          
liquid-c        68.5         0.1         68.9          0.4         0.97             0.99          
liquid-compile  67.4         0.4         68.3          0.2         0.95             0.99          
liquid-render   171.8        0.1         175.6         0.2         0.97             0.98          
lobsters        1031.9       0.3         1041.4        0.3         0.94             0.99          
mail            135.5        0.4         136.7         0.1         0.99             0.99          
psych-load      2246.0       0.1         2281.3        0.1         0.99             0.98          
railsbench      2490.9       0.0         2490.0        0.1         1.01             1.00          
rubocop         179.8        2.3         180.0         0.4         0.94             1.00          
ruby-lsp        117.3        0.1         118.5         0.1         0.99             0.99          
sequel          75.8         0.5         76.3          0.2         0.99             0.99          
--------------  -----------  ----------  ------------  ----------  ---------------  --------------

Limitations

We recognize that our current implementation does not yet offer the flexibility required for a generic plug-in GC. Specifically, the set of APIs that the plug-in GC has to implement is relatively large, at around 70 functions. Additionally, some of these functions are specific to the current GC.

We would like to emphasize that the API is NOT stable and is subject to change. We will be working on improving this API and reducing the surface area. This will be future work and we're not working on it in this phase.

Future plans

  • Refactor and improve gc_impl.c.
  • Implement alternate GC implementations, such as the Epsilon GC and MMTk to prove that this API allows for alternate implementations of the GC.
  • Reduce and improve the API of the GC implementation.
  • Benchmark and improve performance of the DSO API.

Updated by nobu (Nobuyoshi Nakada) 7 months ago

My concern is that a part of this feature, adding a new environment variable that loads a shared object implicitly, can cause security issues.
LD_PRELOAD, DYLD_INSERT_LIBRARIES and so on have the same risk, so dynamic linkers remove such variables in suid processes at least.
However it would be difficult to let all necessary people to know a new variable.
So I'd be negative to allow this in production.

Updated by peterzhu2118 (Peter Zhu) 7 months ago

I think the risk is fairly low, since the user has to compile Ruby with --with-shared-gc to enable this feature. Since they have to manually enable this feature at compile time, they should be aware of the possible security issues that comes with it.

If we want to completely mitigate this risk, we could instead use a command line argument rather than an environment variable.

Updated by katei (Yuta Saito) 7 months ago ยท Edited

+1 for extracting GC implementation of gc.c into a separate gc_impl.c file.

My motivation: Some of the use cases of Ruby on Wasm (e.g. edge computing platform) do not actually need to collect garbage objects during execution because such a process does not last so long. Splitting out the collector implementation allows us to have light-weight implementation like no-op GC in JVM. It will be a good fit for such use cases to reduce runtime overhead and program size, and also be helpful for performance analysis too.

Updated by wks (Kunshan Wang) 6 months ago

I am working on the mmtk-ruby project, and I think this patch is definitely on the right track. This is similar to JEP 304: Garbage Collector Interface. OpenJDK introduced a GC interface which made some other GC algorithms (including Epsilong, ZGC and Shenandoah) much easier to implement in OpenJDK.

In my opinion, whether the GC is extracted into a dedicated .so is secondary. The most important thing is the interface between the VM and the GC. A good VM-GC interface shall clearly define the responsibilities of the VM and the GC, helping the VM developers avoid making too much assumption about one particular GC implementation and hindering the adoption of other more advanced GC algorithms.

Depending on the programming language, the interface may include functions (including call-back functions), abstract classes, interfaces (as in Java), traits, and so on. The GC and the VM each implements some functions, classes, etc. for the other party to use. This patch defines the interface informally as the set of exported functions, and it doesn't even have a dedicated header file for gc_impl.c, or a "struct of function pointers" (which is roughly the counterpart of a C++ abstract class or a Java interface) for gc_impl.c to implement. I believe this is just the first step of attempting to isolating the GC from the VM, and is not the final form of this patch.

The contribution

The most important contribution of this patch is clarifying the responsibilities of the VM and the GC. For example,

  • The GC is responsible for allocating the memory for the objects. The VM function newobj_of calls rb_gc_impl_new_obj to allocate an object, and the GC takes care of details such as size classes (size pools) and mutator-local (ractor-local in Ruby) caches, if the particular GC implementation has the notion of size classes at all. (FYI, bump-pointer allocators allocate objects of different sizes into the same contiguous region.)
  • The VM is responsible for root scanning and object scanning. At the start of a GC, the GC calls rb_gc_mark_roots to identify roots. When the GC visits an object, it calls rb_gc_mark_children to identify the children of an object, and calls rb_gc_update_object_references to forward the reference fields if it is a copying GC. These functions were previously private functions in gc.c. Now they are exported, and ready to be called by different GC implementations.

"Binding": what the current gc_impl.c really is

The gc_impl.c in this patch is said to "contain the implementation of Ruby's GC". As a developer of MMTk, I find that the current gc_impl.c not only contains the core of the GC (responsible for object allocation, marking, sweeping, compaction, ...), but also contains parts that are related to GC, but are also closely related to the VM. That includes

  • The initialization of object header (flags and klass) and object fields (v1, v2, v3) and other metadata (including WB-protected flag) in rb_gc_impl_new_obj
    • FYI, the MMTk core is only responsible for allocating an object of a given size/alignment. The VM initializes the type metadata and fields.
  • The implementation of finalization, including "zombies" (including the linked list objspace->heap_pages.deferred_final), the finalizer_table, and the mechanism to call a function (rb_gc_obj_free) when an object is dead.
    • FYI, the MMTk core only provides an API to (1) tell if an object is reachable, (2) forward a reference to a reachable object, and (3) resurrect an unreachable object (and its descendents. The VM decides what to do when an object dies.
  • The implementation of the bidirectional object-to-ID mapping.
    • FYI, the VM can implement it as a weak table, and the GC is only called to determine if a table entry points to a dead object.

This is not a bad thing. Putting those VM-related things in gc_impl.c will allow those things to be implemented in a way specific to a particular GC implementation. For example, if a GC implementation does not use lazy sweeping, then we can eagerly clear the entries of the finalizer_table that involve dead objects during a GC, instead of clearing them when the object is lazily swept (in obj_free). As another example, if a GC implementation never moves any object, or if it is able to pin an object if its ID is observed, it can simply use the object's address as the ID instead of maintaining any hash tables.

In MMTk, we call such a component a "VM binding". It is a bi-directional component that bridges the GC implementation (such as MMTk) and the VM (such as CRuby, OpenJDK, JikesRVM, etc.). We strongly recommend CRuby to use the term "binding", too. While MMTk calls it "VM binding", CRuby can call it "GC binding".

Further splitting gc_impl.c into a binding and an actual implementation.

One possible improvement over this patch is further splitting the current gc_impl.c into two parts:

  1. The GC implementation that the Ruby VM can be completely agnostic of, and
  2. the bridge (binding) between the VM and the GC implementation.

The first part includes the code that allocates raw bytes from heap pages, manages bitmaps (such as the marking bits), does the marking, transitive closure, sweeping, compacting, etc. The second part implements functions that are called by the VM, and wraps functions provided by the Ruby VM for the first part to call.

For example, gc_mark_stacked_objects implements the tracing loop that computes the transitive closure from roots. It is an implementation detail of the GC, and the VM does not need to know it. It belongs to the first part.

As another example, rb_gc_impl_new_obj is called by the VM (newobj_of), and it calls the low-level allocator (newobj_alloc) to allocate raw objects. It also initializes Ruby fields. Therefore, rb_gc_impl_new_obj is part of the binding.

As another example, the function gc_mark_children is called by gc_mark_stacked_objects from the GC. It wraps the rb_gc_mark_children provided by the VM, but also calls the GC-specific gc_mark_set_parent. It sits between the GC and the VM, therefore belongs to the binding.

Ideally, if the actual GC implementation of CRuby is fully isolated, it can be used for other VMs like Lua, and we just need to make another binding for the Lua VM.

Conversely, when supporting other GC implementations in CRuby, such as MMTk and Lua's minimalist GC, we only need to develop the second part (the binding). For MMTk, the first part (the actual GC implementation) is in the mmtk-core written in Rust, and the second part (the binding) will be implemented by mmtk-ruby. If we are curious enough to try to port Lua's GC into CRuby, we may just copy the source code into CRuby's source tree because it's so small (<2000 LOC) and it's written in C, too.

For performance reasons, a binding usually reimplements part of the GC in a VM-specific way (such as letting the JIT compiler emit the bump-pointer allocation fast path which theoretically belongs to the GC), and reimplements part of the VM in a GC-specific way (such as rewriting the object-scanning functions in Rust so that they can be inlined into MMTk's tracing loop, the single hottest loop during a GC). I am expecting CRuby's GC binding to do the same for performance optimization, especially when using YJIT, in which case the overhead of GC will become more significant and the allocation fast paths will become more important once the mutator gets faster. Evidently, the GC interface should also involve the YJIT compiler so that allocation fast paths and write barrier fast paths can be emitted in GC-specific machine instruction sequences.

Problems

Assumption of mark-sweep-compact

This patch also reveals problems in the GC interface. The most obvious problem is the strong assumption of mark-sweep-compact GC algorithm. There are rb_gc_mark_children and rb_gc_update_object_references. The former is called during marking, and the second is called during compaction. This approach is not suitable for evacuating GC algorithms, such as SemiSpace, Immix, etc. When an evacuating GC visits a field, the object it points to is immediately moved (if this is the first time the object is visited), and the field is updated immediately to point to the new location. As a workaround, we may call both rb_gc_mark_children and rb_gc_update_object_references when visiting each object. And we may rewrite the object-scanning function in the implementation langauge of the GC (Rust for MMTk) for performance reasons. But it is impractical for the GC binding to rewrite it for every type, especially complex types implemented in C (such as classes and iseqs), and it is impossible to rewrite the "marking" and "compacting" functions written by third-party C extensions.

To solve the problem, we should stop abusing rb_gc_mark_children for implementing rb_objspace_reachable_objects_from, but instead make rb_objspace_reachable_objects_from a generic field-visiting function with a field-visitor callback that may potentially update fields, and use that to implement rb_gc_mark_children. Such a generic field-visiting (and updating) function is what the GC really needs to scan an object. Matthew Valentine-House has an experimental branch for this: https://github.com/eightbitraptor/ruby/tree/mvh-test-push-fpointer-to-ractor

Limit of object size

Another problem worth mentioning is the limit of object size the GC can allocate. The current GC implementation of CRuby cannot allocate objects larger than 640 bytes (the largest size pool), and that's why the VM still calls the function rb_gc_size_allocatable_p implemented in gc_impl.c. Most other GC implementations, including the tiny GC implementation in Lua, don't have such limitation. Such a limitation is also the source of the embedded-vs-heap design for many built-in types, including T_OBJECT, T_ARRAY, T_STRING, etc. When switching to another GC implementation, CRuby cannot make the best use of the capability to allocate objects larger than 640 bytes unless the VM make changes accordingly. For example, do not ask the GC if an object of a given size can be allocated (because the GC can allocate any size). Instead, the VM should determine the optimal size of a String, Array or Object, either using predefined thresholds, or using profile data at run time.

Updated by peterzhu2118 (Peter Zhu) 6 months ago

Thank you for the comment @wks (Kunshan Wang).

As you've mentioned, this is just the very beginning of this feature. There are shortcomings and shortcuts taken to make this first iteration simpler. If this feature is accepted, we plan on improving this feature in various ways including:

  • Improving allocation so that initialization of the object's contents does not occur inside of the GC implementation, but rather done by the VM.
  • Support allocation of arbitrary sized objects from the GC. Currently, many types (e.g. string, array, object, hashes) have assumptions about the upper and lower bound of object sizes. It might be a bit difficult to drop assumptions about lower bound due to performance reasons (it would be slower to check if a size can be allocated rather than just assuming that the size can be allocated), but we should at least drop assumptions about the upper bound.
  • Remove the dependency on size classes in the VM. We currently need this because the object shape tree requires size classes to build the root shapes, so we need to change the implementation of object shapes to not depend on this.

There are many other suggestions in your comment, some of which may be non-trivial to implement, but we'll definitely keep them in mind and try to achieve as many as possible.

Actions #6

Updated by peterzhu2118 (Peter Zhu) 6 months ago

  • Status changed from Open to Closed

Applied in changeset git|51bd816517941798c63e587a5a9f3caf69cd510e.


[Feature #20470] Split GC into gc_impl.c

This commit splits gc.c into two files:

  • gc.c now only contains code not specific to Ruby GC. This includes
    code to mark objects (which the GC implementation may choose not to
    use) and wrappers for internal APIs that the implementation may need
    to use (e.g. locking the VM).

  • gc_impl.c now contains the implementation of Ruby's GC. This includes
    marking, sweeping, compaction, and statistics. Most importantly,
    gc_impl.c only uses public APIs in Ruby and a limited set of functions
    exposed in gc.c. This allows us to build gc_impl.c independently of
    Ruby and plug Ruby's GC into itself.

Actions

Also available in: Atom PDF

Like1
Like0Like0Like0Like0Like0Like0