haproxy/doc/internals/api/memory.txt
Willy Tarreau 33d72568dd MINOR: tools: also implement ha_aligned_alloc_typed()
This one is a macro and will allocate a properly aligned and sized
object. This will help make sure that the alignment promised to the
compiler is respected.

When memstats is used, the type name is passed as a string into the
.extra field so that it can be displayed in "debug dev memstats". Two
tiny mistakes related to memstats macros were also fixed (calloc
instead of malloc for zalloc), and the doc was also added to document
how to use these calls.
2025-08-13 17:37:08 +02:00

87 lines
4.8 KiB
Plaintext

2025-08-13 - Memory allocation in HAProxy 3.3
The vast majority of dynamic memory allocations are performed from pools. Pools
are optimized to store pre-calibrated objects of the right size for a given
usage, try to favor locality and hot objects as much as possible, and are
heavily instrumented to detect and help debug a wide class of bugs including
buffer overflows, use-after-free, etc.
For objects of random sizes, or those used only at configuration time, pools
are not suited, and the regular malloc/free family is available, in addition of
a few others.
The standard allocation calls are intercepted at the code level (#define) when
the code is compiled with -DDEBUG_MEM_STATS. For this reason, these calls are
redefined as macros in "bug.h", and one must not try to use the pointers to
such functions, as this may break DEBUG_MEM_STATS. This provides fine-grained
stats about allocation/free per line of source code using locally implemented
counters that can be consulted by "debug dev memstats". The calls are
categorized into one of "calloc", "free", "malloc", "realloc", "strdup",
"p_alloc", "p_free", the latter two designating pools. Extra calls such as
memalign() and similar are also intercepted and counted as malloc.
Due to the nature of this replacement, DEBUG_MEM_STATS cannot see operations
performed in libraries or dependencies.
In addition to DEBUG_MEM_STATS, when haproxy is built with USE_MEMORY_PROFILING
the standard functions are wrapped by new ones defined in "activity.c", which
also hold counters by call place. These ones are able to trace activity in
libraries because the functions check the return pointer to figure where the
call was made. The approach is different and relies on a large hash table. The
files, function names and line numbers are not know, but by passing the pointer
to dladdr(), we can often resolve most of these symbols. These operations are
consulted via "show profiling memory". It must first be enabled either in the
global config "profiling.memory on" or the CLI using "set profiling memory on".
Memory profiling can also track pool allocations and frees thanks to knowing
the size of the element and knowing a place where to store it. Some future
evolutions might consider making this possible as well for pure malloc/free
too by leveraging malloc_usable_size() a bit more.
Finally, 3.3 brought aligned allocations. These are made available via a new
family of functions around ha_aligned_alloc() that simply map to either
posix_memalign(), memalign() or _aligned_malloc() for CYGWIN, depending on
which one is available. This latter one requires to pass the pointer to
_aligned_free() instead of free(), so for this reason, all aligned allocations
have to be released using ha_aligned_free(). Since this mostly happens on
configuration elements, in practice it's not as inconvenient as it can sound.
These functions are in reality macros handled in "bug.h" like the previous
ones in order to deal with DEBUG_MEM_STATS. All "alloc" variants are reported
in memstats as "malloc". All "zalloc" variants are reported in memstats as
"calloc".
The currently available allocators are the following:
- void *ha_aligned_alloc(size_t align, size_t size)
- void *ha_aligned_zalloc(size_t align, size_t size)
Equivalent of malloc() but aligned to <align> bytes. The alignment MUST be
at least as large as one word and MUST be a power of two. The "zalloc"
variant also zeroes the area on success. Both return NULL on failure.
- void *ha_aligned_alloc_safe(size_t align, size_t size)
- void *ha_aligned_zalloc_safe(size_t align, size_t size)
Equivalent of malloc() but aligned to <align> bytes. The alignment is
automatically adjusted to the nearest larger power of two that is at least
as large as a word. The "zalloc" variant also zeroes the area on
success. Both return NULL on failure.
- (type *)ha_aligned_alloc_typed(size_t count, type)
(type *)ha_aligned_zalloc_typed(size_t count, type)
This macro returns an area aligned to the required alignment for type
<type>, large enough for <count> objects of this type, and the result is a
pointer of this type. The goal is to ease allocation of known structures
whose alignment is not necessarily known to the developer (and to avoid
encouraging to hard-code alignment). The cast in return also provides a
last-minute control in case a wrong type is mistakenly used due to a poor
copy-paste or an extra "*" after the type. When DEBUG_MEM_STATS is in use,
the type is stored as a string in the ".extra" field so that it can be
displayed in "debug dev memstats". The "zalloc" variant also zeroes the
area on success. Both return NULL on failure.
- void ha_aligned_free(void *ptr)
Frees the area pointed to by ptr. It is the equivalent of free() but for
objects allocated using one of the functions above.