Unused slabs which get recycled are no longer reinitialized from scratch if their chunk size matches what the allocator needed. If the chunk size matches, the list of available chunks is left unchanged, and just the various counters are reset.
Added __get_slab_stats() function.
Added support for global __slab_purge_threshold tuning variable.
Added a short test program for the slab allocator.
The malloc-test program was linked against the wrong object file in GNUmakefile.68k. Fixed.
Added more consistency checking to the slab allocator, which is built if DEBUG is defined in "stdlib_slab.c".
Memory allocations are no longer guaranteed to be aligned to 64 bit word boundaries. In fact, this has not even worked reliably in the past 10 years.
Memory allocation request sizes are now rounded to multiples of 32 bit words (the size of an address pointer) instead to the size of a 64 bit word.
Reduced the memory footprint of the memory allocation management data structures by reusing the most significant bit of the memory allocation size. This allows many more allocations to fit into the 32 byte chunk slabs, but limits the maximum memory allocation size to a little less than 2 GBytes.
Added integer overflow checks to the memory management code.
Reduced the memory management overhead further. This cuts an additional 8 bytes per allocation, unless neither the slab allocator nor memory pools are available. With this reduction the slab allocator is able to use 16 byte chunks, which cover memory allocation requests of 1..8 bytes.
Fixed a bug caused by returning an allocation back to a slab which passed the wrong pointer.
If the first slab in the list of slabs which share the same chunk size has no more room, it means that all other slabs following it have no room either. This speeds up the test to find a slab with free space, which can now abort and directly proceed to allocate memory for a new slab.
If an empty slab's decay count hits zero, it is moved to the front of the empty slab list to be reclaimed more quickly.
Allocations made from the slab now carry a pointer back to the slab which they are a part of. This speeds up deallocation but has the downside of making the smallest usable slab chunk size 64 bytes, which is double what used to be the minimum before.
The maximum slab size is now 2^17 bytes (= 131072). If you request a slab size larger than this, you will get slab sizes of 131072 bytes instead.
Enabling the memory management debugging code no longer produces compiler errors.
Added a slab allocator which replaces the use of memory pools or the plain AllocMem() operations, respectively. In order to activate the slab allocator, choose a slab size (e.g. 2048 bytes or 4096 bytes) and declare a global variable like this:
ULONG __slab_max_size = 2048;
Memory allocations smaller than the slab size will be made from "slabs", i.e. large chunks of memory of the given size. Larger allocations will be managed separately.
If a new slab is allocated, or an empty slab is reused, the list of empty slabs is now purged, if possible. Empty slabs which cannot be immediately purged "decay" and will be collected the next time a new slab is allocated or an empty slab is reused.
If a chunk is allocated and the slab which the chunk came from is fully utilized, the slab is moved to the end of its list. This should avoid searching this slab for free space before all other slabs of the same chunk size have been tested first. Slabs with free space are always placed closer to the head of the list.
A slab-allocator can now be used as the back-end for all memory allocations by setting __slab_max_size to a value > 0. This controls the slab size used for allocations and should be a power of 2, such as 2048 or 4096 bytes.
This change still needs testing.
Changes in the naming of certain dos.library functions in the
V53 header files are now accounted for. This also affects
unistd/stat/utime code which uses the same header file.
- Replaced ISO Latin 1 code #160 with a plain blank space (ASCII code #32) where necessary
git-svn-id: file:///Users/olsen/Code/migration-svn-zu-git/logical-line-staging/clib2/trunk@15305 87f5fb63-7c3d-0410-a384-fd976d0f7a62
- The fscanf() family failed to parse and convert %f parameters correctly
if the respective number did not begin with a digit, but a decimal
point. Hence ".7" would not be processed, but "0.7" would.
c.lib 1.206 (24.4.2015)
- Reworked the __putc() and putc() macros to reference the 'c' input
parameter only once, and to be free of side-effects when tinkering
with the buffer position.
- isatty() had the __fd_lock() call in the wrong place, which could have
led to cleanup problems later.
- The close action in the stdio, socket and termios hook code now
also zaps the fd pointer itself after cleaning up the file descriptor
table entry.
- Removed the remains of all the stack extension and stack overflow/underflow
checking code. It never actually worked. The bit that does work is the stack
usage measurement code, plus the bit that sets up the the custom stack
according to local setting or by calling a query function.
git-svn-id: file:///Users/olsen/Code/migration-svn-zu-git/logical-line-staging/clib2/trunk@15215 87f5fb63-7c3d-0410-a384-fd976d0f7a62
checking code. It never actually worked. The bit that does work is the stack
usage measurement code, plus the bit that sets up the the custom stack
according to local setting or by calling a query function.
git-svn-id: file:///Users/olsen/Code/migration-svn-zu-git/logical-line-staging/clib2/trunk@15212 87f5fb63-7c3d-0410-a384-fd976d0f7a62
parameter only once, and to be free of side-effects when tinkering
with the buffer position.
- isatty() had the __fd_lock() call in the wrong place, which could have
led to cleanup problems later.
- The close action in the stdio, socket and termios hook code now
also zaps the fd pointer itself after cleaning up the file descriptor
table entry.
git-svn-id: file:///Users/olsen/Code/migration-svn-zu-git/logical-line-staging/clib2/trunk@15211 87f5fb63-7c3d-0410-a384-fd976d0f7a62
