Skip to content

Latest commit

 

History

History
266 lines (230 loc) · 14.8 KB

porting.md

File metadata and controls

266 lines (230 loc) · 14.8 KB

Conservative Garbage Collector Porting Directions

The collector is designed to be relatively easy to port, but is not portable code per se. The collector inherently has to perform operations, such as scanning the stack(s), that are not possible in portable C code.

All of the following assumes that the collector is being ported to a byte-addressable 32- or 64-bit machine. Currently all successful ports to 64-bit machines involve LP64 and LLP64 targets (notably Win64). You are hereby discouraged from attempting a port to non-byte-addressable, or 8-bit, or 16-bit machines.

The difficulty of porting the collector varies greatly depending on the needed functionality. In the simplest case, only some small additions are needed for the include/private/gcconfig.h file. This is described in the following section. Later sections discuss some of the optional features, which typically involve more porting effort.

Note that the collector makes heavy use of ifdefs. Unlike some other software projects, we have concluded repeatedly that this is preferable to system dependent files, with code duplicated between the files. However, to keep this manageable, we do strongly believe in indenting ifdefs correctly (for historical reasons usually without the leading sharp sign). (Separate source files are of course fine if they do not result in code duplication.)

Adding Platforms to gcconfig.h

If neither thread support, nor tracing of dynamic library data is required, these are often the only changes you will need to make.

The include/private/gcconfig.h file consists of three sections:

  1. A section that defines GC-internal macros that identify the architecture (e.g. IA64 or I386) and operating system (e.g. LINUX or MSWIN32). This is usually done by testing predefined macros. By defining our own macros instead of using the predefined ones directly, we can impose a bit more consistency, and somewhat isolate ourselves from compiler differences. It is relatively straightforward to add a new entry here. But please try to be consistent with the existing code. In particular, 64-bit variants of 32-bit architectures general are not treated as a new architecture. Instead we explicitly test for 64-bit-ness in the few places in which it matters. (The notable exception here is I386 and X86_64. This is partially historical, and partially justified by the fact that there are arguably more substantial architecture and ABI differences here than for RISC variants.) On GNU-based systems, cpp -dM empty_source_file.c seems to generate a set of predefined macros. On some other systems, the "verbose" compiler option may do so, or the manual page may list them.

  2. A section that defines a small number of platform-specific macros, which are then used directly by the collector. For simple ports, this is where most of the effort is required. We describe the macros below. This section contains a subsection for each architecture (enclosed in a suitable ifdef. Each subsection usually contains some architecture-dependent defines, followed by several sets of OS-dependent defines, again enclosed in ifdefs.

  3. A section that fills in defaults for some macros left undefined in the preceding section, and defines some other macros that rarely need adjustment for new platforms. You will typically not have to touch these. If you are porting to an OS that was previously completely unsupported, it is likely that you will need to add another clause to the definition of GET_MEM.

The following macros must be defined correctly for each architecture and operating system:

  • MACH_TYPE - Defined to a string that represents the machine architecture. Usually just the macro name used to identify the architecture, but enclosed in quotes.
  • OS_TYPE - Defined to a string that represents the operating system name. Usually just the macro name used to identify the operating system, but enclosed in quotes.
  • CPP_WORDSZ - The address (also referred simply as "word") size in bits as a constant suitable for preprocessor tests, i.e. without casts or sizeof expressions. For platforms supporting both 32- and 64-bit ABIs, this should be conditionally defined depending on the current ABI.
  • CPP_PTRSZ - Similar to CPP_WORDSZ but for the size of a pointer (also referred as "pointer-sized" word) in bits. On most platforms, its value is equal to that of CPP_WORDSZ.
  • ALIGNMENT - Defined to be the largest N such that all pointer are guaranteed to be aligned on N-byte boundaries. Defining it to be 1 will always work, but perform poorly. For all modern 32-bit platforms, this is 4. For all modern 64-bit platforms, this is 8. (Whether or not x86 qualifies as a modern architecture here is compiler- and OS-dependent.)
  • DATASTART - The beginning of the main data segment. The collector will trace all memory between DATASTART and DATAEND for root pointers. On some platforms, this can be defined to a constant address, though experience has shown that to be risky. Ideally the linker will define a symbol (e.g. _data) whose address is the beginning of the data segment. Sometimes the value can be computed using the GC_SysVGetDataStart function. Not used if either the next macro is defined, or if dynamic loading is supported, and the dynamic loading support defines a function GC_register_main_static_data which returns false.
  • SEARCH_FOR_DATA_START - If this is defined DATASTART will be defined to a dynamically computed value which is obtained by starting with the address of _end and walking backwards until non-addressable memory is found. This often works on Posix-like platforms. It makes it harder to debug client programs, since startup involves generating and catching a segmentation fault, which tends to confuse users.
  • DATAEND - Set to the end of the main data segment. Defaults to _end, where that is declared as an array. This works in some cases, since the linker introduces a suitable symbol.
  • DATASTART2, DATAEND2 - Some platforms have two discontiguous main data segments, e.g. for initialized and uninitialized data. If so, these two macros should be defined to the limits of the second main data segment.
  • STACK_GROWS_UP - Should be defined if the stack (or thread stacks) grow towards higher addresses. (This appears to be true only on PA-RISC. If your architecture has more than one stack per thread, and is not supported yet, you will need to do more work. Grep for "IA64" in the source for an example.)
  • STACKBOTTOM - Defined to be the cold end of the stack, which is usually (i.e. when the stacks grow down) the highest address in the stack. It must bound the region of the stack that contains pointers into the GC heap. With thread support, this must be the cold end of the main stack, which typically cannot be found in the same way as the other thread stacks. If this is not defined and none of the following three macros is defined, client code must explicitly set GC_stackbottom to an appropriate value before calling GC_INIT or any other GC_ routine.
  • LINUX_STACKBOTTOM - May be defined instead of STACKBOTTOM. If defined, then the cold end of the stack will be determined, we usually read it from /proc.
  • HEURISTIC1 - May be defined instead of STACKBOTTOM. STACK_GRAN should generally also be redefined. The cold end of the stack is determined by taking an address inside GC_inits frame, and rounding it up to the next multiple of STACK_GRAN. This works well if the stack bottom is always aligned to a large power of two. (STACK_GRAN is predefined to 0x1000000, which is rarely optimal.)
  • HEURISTIC2 - May be defined instead of STACKBOTTOM. The cold end of the stack is determined by taking an address inside GC_inits frame, incrementing it repeatedly in small steps (decrement if STACK_GROWS_UP), and reading the value at each location. We remember the value when the first Segmentation violation or Bus error is signaled, round that to the nearest plausible page boundary, and use that as the stack bottom.
  • DYNAMIC_LOADING - Should be defined if dyn_load.c has been updated for this platform and tracing of dynamic library roots is supported.
  • GWW_VDB, MPROTECT_VDB, PROC_VDB, SOFT_VDB - May be defined if the corresponding virtual dirty bit implementation in os_dep.c is usable on this platform. This allows incremental/generational garbage collection. (GWW_VDB uses the Win32 GetWriteWatch function to read dirty bits, MPROTECT_VDB identifies modified pages by write protecting the heap and catching faults. PROC_VDB and SOFT_VDB use the /proc pseudo-files to read dirty bits.)
  • PREFETCH, GC_PREFETCH_FOR_WRITE - The collector uses PREFETCH(x) to preload the cache with the data at x address. This defaults to a no-op.
  • CLEAR_DOUBLE - If it is defined, then CLEAR_DOUBLE(x) is used as a fast way to clear two "pointer-sized" words at GC_malloc-aligned address x. By default, NULL pointers storing is used instead.
  • HEAP_START - May be defined as the initial address hint for mmap-based allocation.

Additional requirements for a basic port

In some cases, you may have to add additional platform-specific code to other files. A likely candidate is the implementation of GC_with_callee_saves_pushed in mach_dep.c. This ensure that register contents that the collector must trace from are copied to the stack. Typically this can be done portably, but on some platforms it may require assembly code, or just tweaking of conditional compilation tests.

If your platform supports getcontext then defining the macro UNIX_LIKE for your OS in include/private/gcconfig.h (if it is not defined there yet) is likely to solve the problem. Otherwise, if you are using gcc, _builtin_unwind_init will be used, and should work fine. If that is not applicable either, the implementation will try to use setjmp. This will work if your setjmp implementation saves all possibly pointer-valued registers into the buffer, as opposed to trying to unwind the stack at longjmp time. The setjmp_test test tries to determine this, but often does not get it right. Registers tracing handled with an assembly code is generally to be avoided.

Most commonly os_dep.c will not require attention, but see below.

Thread support

Supporting threads requires that the collector be able to find and suspend all threads potentially accessing the garbage-collected heap, and locate any state associated with each thread that must be traced.

The functionality needed for thread support is generally implemented in one or more files specific to the particular thread interface. For example, somewhat portable pthread support is implemented in pthread_support.c and pthread_stop_world.c. The essential functionality consists of:

  • GC_stop_world - Stops all threads which may access the garbage collected heap, other than the caller;
  • GC_start_world - Restart other threads;
  • GC_push_all_stacks - Push the contents of all thread stacks (or, at least, of pointer-containing regions in the thread stacks) onto the mark stack.

These very often require that the garbage collector maintain its own data structures to track active threads.

In addition, LOCK and UNLOCK must be implemented in include/private/gc_locks.h file.

The easiest case is probably a new pthreads platform on which threads can be stopped with signals. In this case, the changes involve:

  1. Introducing a suitable GC_xxx_THREADS macro, which should be automatically defined by gc_config_macros.h in the right cases. It should also result in a definition of GC_PTHREADS, as for the existing cases.
  2. Ensuring that the atomic_ops package at least minimally supports the platform. If incremental GC is needed, or if pthread locks do not perform adequately as the allocator lock, you will probably need to ensure that a sufficient atomic_ops port exists for the platform to provided an atomic test and set operation. The latest GC code can use GCC atomic intrinsics instead of atomic_ops package (see include/private/gc_atomic_ops.h).
  3. Making any needed adjustments to pthread_stop_world.c and pthread_support.c. Ideally none should be needed. In fact, not all of this is as well standardized as one would like, and outright bugs requiring workarounds are common. Non-preemptive threads packages will probably require further work. Similarly thread-local allocation and parallel marking requires further work in pthread_support.c, and may require better atomic_ops support for the designed platform.

Dynamic library support

So long as DATASTART and DATAEND are defined correctly, the collector will trace memory reachable from file scope or static variables defined as part of the main executable. This is sufficient if either the program is statically linked, or if pointers to the garbage-collected heap are never stored in non-stack variables defined in dynamic libraries.

If dynamic library data sections must also be traced, then:

  • DYNAMIC_LOADING must be defined in the appropriate section of include/private/gcconfig.h.
  • An appropriate versions of the functions GC_register_dynamic_libraries should be defined in dyn_load.c. This function should invoke GC_cond_add_roots(region_start, region_end, TRUE) on each dynamic library data section.

Implementations that scan for writable data segments are error prone, particularly in the presence of threads. They frequently result in race conditions when threads exit and stacks disappear. They may also accidentally trace large regions of graphics memory, or mapped files. On at least one occasion they have been known to try to trace device memory that could not safely be read in the manner the GC wanted to read it.

It is usually safer to walk the dynamic linker data structure, especially if the linker exports an interface to do so. But beware of poorly documented locking behavior in this case.

Incremental GC support

For incremental and generational collection to work, os_dep.c must contain a suitable virtual dirty bit implementation, which allows the collector to track which heap pages (assumed to be a multiple of the collector's block size) have been written during a certain time interval. The collector provides several implementations, which might be adapted. The default (DEFAULT_VDB) is a placeholder which treats all pages as having been written. This ensures correctness, but renders incremental and generational collection essentially useless.

Stack traces for debug support

If stack traces in objects are needed for debug support, GC_save_callers and GC_print_callers must be implemented.

Disclaimer

This is an initial pass at porting guidelines. Some things have no doubt been overlooked.