|
|
|
---
|
|
|
|
author: "Jaroslav Sykora (jsa)"
|
|
|
|
---
|
|
|
|
|
|
|
|
# Using the Private Heap Manager in a foreign code (PHM-XT)
|
|
|
|
|
|
|
|
This document describes the XT variant of the SAC Private Heap Manager (PHM).
|
|
|
|
The XT variant of the PHM library enables the heap manager to run in a general multithreaded environment.
|
|
|
|
The XT variant is *only* used in sac4c-related code: i.e. when the SAC code is linked as a library
|
|
|
|
to a foreign C code. The classical standalone SAC applications continue to use the original MT variant
|
|
|
|
of the PHM library.
|
|
|
|
|
|
|
|
>[!important]
|
|
|
|
> The PHM code used in SAC standalone multithreaded apps is exactly the same as used to.
|
|
|
|
> Hence certain people don't have to get annoyed.
|
|
|
|
> The new code is used only when compiling for external use via sac4c.
|
|
|
|
> This is a completely new feature which improves on the state of the matter, as previously there was no PHM in sac4c code at all.
|
|
|
|
|
|
|
|
|
|
|
|
The sac2c build system compiles the following variants of the PHM library: **SEQ**, **MT**, **XT**.
|
|
|
|
For each of them there is a **production** build and a **diagnostic** (diag) build.
|
|
|
|
|
|
|
|
List of the PHM variants:
|
|
|
|
|
|
|
|
| Variant |Description |Standalone apps |External [sac4c] |
|
|
|
|
|---------|------------------------|----------------|-----------------|
|
|
|
|
|SEQ |Purely sequential |yes (no change) |no |
|
|
|
|
|MT |Standalone multithreaded|yes (no change) |no! |
|
|
|
|
|XT |Externally multithreaded|if needed |yes! |
|
|
|
|
|
|
|
|
|
|
|
|
Physically, the following files are created:
|
|
|
|
|
|
|
|
```
|
|
|
|
libsacphm$(TARGET).seq.[so|a]
|
|
|
|
libsacphm$(TARGET).seq.diag.[so|a]
|
|
|
|
libsacphm$(TARGET).mt.[so|a]
|
|
|
|
libsacphm$(TARGET).mt.diag.[so|a]
|
|
|
|
libsacphm$(TARGET).xt.[so|a]
|
|
|
|
libsacphm$(TARGET).xt.diag.[so|a]
|
|
|
|
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
## Modifications
|
|
|
|
|
|
|
|
The normal standalone **MT** variant of the PHM library makes the following assumptions about the runtime environment:
|
|
|
|
|
|
|
|
1. Only SAC MT runtime ever creates new threads.
|
|
|
|
2. Memory allocated in thread X will be eventually freed by the **same** thread X.
|
|
|
|
3. The PHM library is explicitly initialized prior to creating any threads.
|
|
|
|
|
|
|
|
The assumptions listed above no longer hold when the PHM library is used in a general environment.
|
|
|
|
|
|
|
|
The PHM-XT variant enables the following new features in the code to overcome the problematic
|
|
|
|
assumptions:
|
|
|
|
|
|
|
|
4. **SAC_DO_HM_DISCOVER_THREADS=1**
|
|
|
|
There may be other threads in the process than those created by SAC runtime.
|
|
|
|
Discover and assign IDs autonomously.
|
|
|
|
5. **SAC_DO_HM_XTHR_FREE=1**
|
|
|
|
The free()'ing thread may not be the same as the malloc()'ing thread.
|
|
|
|
6. **SAC_DO_HM_AUTONOMOUS=1**
|
|
|
|
Initialise the library on its own, no dependencies on the other SAC runtime.
|
|
|
|
|
|
|
|
The new features are orthogonal and therefore will be described separately.
|
|
|
|
|
|
|
|
|
|
|
|
### SAC_DO_HM_DISCOVER_THREADS: Discover and assign thread IDs
|
|
|
|
|
|
|
|
The PHM requires that each concurrent thread has its own set of memory pools ('*arenas*').
|
|
|
|
This allows threads to allocate and free memory in their arenas without any interference
|
|
|
|
from other threads (which would imply synchronization via locking).
|
|
|
|
On start up a set of arenas for a given maximal number of threads is created.
|
|
|
|
When any thread later requires a memory allocation via a malloc() call the PHM library must use
|
|
|
|
the set of arenas reserved for that particular thread.
|
|
|
|
|
|
|
|
Threads are automatically assigned unique ID numbers in the range $0$ to $max\_threads-1$,
|
|
|
|
where $max\_threads$ is given on library initialization and it must be sufficiently large.
|
|
|
|
When threads terminate their unique IDs are released and may be reused later.
|
|
|
|
|
|
|
|
The PHM library provides the function in `libsacphm/heap/thread_ids.c`
|
|
|
|
|
|
|
|
```c
|
|
|
|
unsigned int SAC_HM_CurrentThreadId(void)
|
|
|
|
```
|
|
|
|
|
|
|
|
to determine the unique thread ID for the current thread.
|
|
|
|
The function uses the Thread Local Storage (TLS) facility from the pthread library
|
|
|
|
to store the thread ID of the current thread.
|
|
|
|
If this is the first call in the given thread the value in the TLS slot will be NULL,
|
|
|
|
and hence a fresh thread ID is assigned and stored in the TLS slot of the thread.
|
|
|
|
Later calls in the given thread will return the same ID.
|
|
|
|
When thread terminates a destructor hook function for the TLS slot is executed,
|
|
|
|
allowing the PHM library to release the thread ID.
|
|
|
|
|
|
|
|
Thread IDs are needed for every memory allocation (but not for memory freeing).
|
|
|
|
Code using the legacy malloc() interface cannot directly provide the thread ID,
|
|
|
|
hence the TLS slot is used to determine it automatically.
|
|
|
|
In native SAC code the thread ID is redundantly kept in thread's contextual data structure (the 'bees')
|
|
|
|
and passed directly to the PHM on every call as a normal parameter,
|
|
|
|
thus avoiding the TLS overhead.
|
|
|
|
|
|
|
|
When the SAC_DO_HM_DISCOVER_THREADS feature is disabled, such as in the PHM-MT variant,
|
|
|
|
the SAC MT layer which normally creates the worker threads (bees) also directly assigns
|
|
|
|
them the thread IDs.
|
|
|
|
|
|
|
|
|
|
|
|
### SAC_DO_HM_XTHR_FREE: Cross-thread free()
|
|
|
|
|
|
|
|
The normal PHM-MT variant assumes that calls to free() are executed in the same thread
|
|
|
|
as the corresponding malloc() that has obtained the memory.
|
|
|
|
This assumption allows the heap manager to manipulate the arena data structures
|
|
|
|
of the memory block directly, without any locking.
|
|
|
|
|
|
|
|
This limitation is overcome in the PHM-XT variant by splitting the memory freeing procedure into two steps:
|
|
|
|
|
|
|
|
7. The memory being freed by the application is marked as 'unused' in its arena.
|
|
|
|
As the target arena may be foreign to the current thread, this operation
|
|
|
|
requires an atomic instruction, but is otherwise cheap.
|
|
|
|
|
|
|
|
8. Later on when a potentially different thread that owns the arena tries to allocate in it
|
|
|
|
the unused blocks of the arena are reintegrated into the free list.
|
|
|
|
|
|
|
|
|
|
|
|
Technically, the arena structure is extended with a pointer to the list of unused blocks:
|
|
|
|
|
|
|
|
```c
|
|
|
|
typedef struct arena_t {
|
|
|
|
...
|
|
|
|
volatile SAC_HM_header_t *unused_list;
|
|
|
|
...
|
|
|
|
};
|
|
|
|
```
|
|
|
|
|
|
|
|
The 'unused_list' is a linked list of blocks that were released by free().
|
|
|
|
The blocks are linked via the 'nextfree' field in the SAC_HM_header_t structure.
|
|
|
|
As the 'unused_list' field is potentially accessed asynchronously from different threads,
|
|
|
|
the linked list must be manipulated using only atomic operations (this is the only field
|
|
|
|
in the arena structure which entertains asynchronous accesses).
|
|
|
|
Luckily, we only need two simple atomic operations on the list: to push a new element
|
|
|
|
at the beginning of the list, and to grab the whole list for a local processing.
|
|
|
|
Both operations can be efficiently implemented by the compare-and-swap atomic instruction
|
|
|
|
that is built in the GCC [see `gcc atomic instructions`_, function __sync_bool_compare_and_swap()](http://gcc.gnu.org/onlinedocs/gcc-4.1.1/gcc/Atomic-Builtins.html).
|
|
|
|
|
|
|
|
The pertinent functions to look for in the code are:
|
|
|
|
|
|
|
|
do_free_small_unused_chunks(), do_free_large_unused_chunks(),
|
|
|
|
push_smallchunk_to_arena_unused_list(), push_largechunk_to_arena_unused_list(),
|
|
|
|
grab_arena_unused_list().
|
|
|
|
|
|
|
|
|
|
|
|
### SAC_DO_HM_AUTONOMOUS: Autonomous initialization
|
|
|
|
|
|
|
|
The normal PHM-MT variant requires that the initialization routine SAC_HM_Setup(int num_threads) be run by the application prior to creating any threads.
|
|
|
|
Furthermore, a couple of constant global variables must be defined in the main program
|
|
|
|
to communicate other parameters that are required by the PHM even before the SAC_HM_Setup() call may be made.
|
|
|
|
|
|
|
|
Both the above requirements are technically unpleasant in a general use.
|
|
|
|
For the time being we got away with them by simply statically hard-coding the parameters in the source code.
|
|
|
|
Simply put:
|
|
|
|
|
|
|
|
```c
|
|
|
|
#if SAC_DO_HM_DISCOVER_THREADS || SAC_DO_HM_AUTONOMOUS
|
|
|
|
/* Assumed number of threads: this is only used in sac4c XT variant.
|
|
|
|
* Statically compiled in to allow static memory allocs. */
|
|
|
|
#define SAC_HM_ASSUME_THREADS_MAX 512
|
|
|
|
|
|
|
|
#define SAC_SET_INITIAL_MASTER_HEAPSIZE (1024*1024)
|
|
|
|
#define SAC_SET_INITIAL_WORKER_HEAPSIZE (1024*1024)
|
|
|
|
#define SAC_SET_INITIAL_UNIFIED_HEAPSIZE (1024*1024)
|
|
|
|
|
|
|
|
#endif
|
|
|
|
``` |
|
|
\ No newline at end of file |
