User Manual jemalloc @jemalloc_version@ Jason Evans Author JEMALLOC 3 jemalloc jemalloc general purpose memory allocation functions LIBRARY This manual describes jemalloc @jemalloc_version@. More information can be found at the jemalloc website. SYNOPSIS #include <jemalloc/jemalloc.h> Standard API void *malloc size_t size void *calloc size_t number size_t size int posix_memalign void **ptr size_t alignment size_t size void *aligned_alloc size_t alignment size_t size void *realloc void *ptr size_t size void free void *ptr Non-standard API void *mallocx size_t size int flags void *rallocx void *ptr size_t size int flags size_t xallocx void *ptr size_t size size_t extra int flags size_t sallocx void *ptr int flags void dallocx void *ptr int flags void sdallocx void *ptr size_t size int flags size_t nallocx size_t size int flags int mallctl const char *name void *oldp size_t *oldlenp void *newp size_t newlen int mallctlnametomib const char *name size_t *mibp size_t *miblenp int mallctlbymib const size_t *mib size_t miblen void *oldp size_t *oldlenp void *newp size_t newlen void malloc_stats_print void (*write_cb) void *, const char * void *cbopaque const char *opts size_t malloc_usable_size const void *ptr void (*malloc_message) void *cbopaque const char *s const char *malloc_conf; DESCRIPTION Standard API The malloc() function allocates size bytes of uninitialized memory. The allocated space is suitably aligned (after possible pointer coercion) for storage of any type of object. The calloc() function allocates space for number objects, each size bytes in length. The result is identical to calling malloc() with an argument of number * size, with the exception that the allocated memory is explicitly initialized to zero bytes. The posix_memalign() function allocates size bytes of memory such that the allocation's base address is a multiple of alignment, and returns the allocation in the value pointed to by ptr. The requested alignment must be a power of 2 at least as large as sizeof(void *). The aligned_alloc() function allocates size bytes of memory such that the allocation's base address is a multiple of alignment. The requested alignment must be a power of 2. Behavior is undefined if size is not an integral multiple of alignment. The realloc() function changes the size of the previously allocated memory referenced by ptr to size bytes. The contents of the memory are unchanged up to the lesser of the new and old sizes. If the new size is larger, the contents of the newly allocated portion of the memory are undefined. Upon success, the memory referenced by ptr is freed and a pointer to the newly allocated memory is returned. Note that realloc() may move the memory allocation, resulting in a different return value than ptr. If ptr is NULL, the realloc() function behaves identically to malloc() for the specified size. The free() function causes the allocated memory referenced by ptr to be made available for future allocations. If ptr is NULL, no action occurs. Non-standard API The mallocx(), rallocx(), xallocx(), sallocx(), dallocx(), sdallocx(), and nallocx() functions all have a flags argument that can be used to specify options. The functions only check the options that are contextually relevant. Use bitwise or (|) operations to specify one or more of the following: MALLOCX_LG_ALIGN(la) Align the memory allocation to start at an address that is a multiple of (1 << la). This macro does not validate that la is within the valid range. MALLOCX_ALIGN(a) Align the memory allocation to start at an address that is a multiple of a, where a is a power of two. This macro does not validate that a is a power of 2. MALLOCX_ZERO Initialize newly allocated memory to contain zero bytes. In the growing reallocation case, the real size prior to reallocation defines the boundary between untouched bytes and those that are initialized to contain zero bytes. If this macro is absent, newly allocated memory is uninitialized. MALLOCX_TCACHE(tc) Use the thread-specific cache (tcache) specified by the identifier tc, which must have been acquired via the tcache.create mallctl. This macro does not validate that tc specifies a valid identifier. MALLOCX_TCACHE_NONE Do not use a thread-specific cache (tcache). Unless MALLOCX_TCACHE(tc) or MALLOCX_TCACHE_NONE is specified, an automatically managed tcache will be used under many circumstances. This macro cannot be used in the same flags argument as MALLOCX_TCACHE(tc). MALLOCX_ARENA(a) Use the arena specified by the index a. This macro has no effect for regions that were allocated via an arena other than the one specified. This macro does not validate that a specifies an arena index in the valid range. The mallocx() function allocates at least size bytes of memory, and returns a pointer to the base address of the allocation. Behavior is undefined if size is 0. The rallocx() function resizes the allocation at ptr to be at least size bytes, and returns a pointer to the base address of the resulting allocation, which may or may not have moved from its original location. Behavior is undefined if size is 0. The xallocx() function resizes the allocation at ptr in place to be at least size bytes, and returns the real size of the allocation. If extra is non-zero, an attempt is made to resize the allocation to be at least (size + extra) bytes, though inability to allocate the extra byte(s) will not by itself result in failure to resize. Behavior is undefined if size is 0, or if (size + extra > SIZE_T_MAX). The sallocx() function returns the real size of the allocation at ptr. The dallocx() function causes the memory referenced by ptr to be made available for future allocations. The sdallocx() function is an extension of dallocx() with a size parameter to allow the caller to pass in the allocation size as an optimization. The minimum valid input size is the original requested size of the allocation, and the maximum valid input size is the corresponding value returned by nallocx() or sallocx(). The nallocx() function allocates no memory, but it performs the same size computation as the mallocx() function, and returns the real size of the allocation that would result from the equivalent mallocx() function call, or 0 if the inputs exceed the maximum supported size class and/or alignment. Behavior is undefined if size is 0. The mallctl() function provides a general interface for introspecting the memory allocator, as well as setting modifiable parameters and triggering actions. The period-separated name argument specifies a location in a tree-structured namespace; see the section for documentation on the tree contents. To read a value, pass a pointer via oldp to adequate space to contain the value, and a pointer to its length via oldlenp; otherwise pass NULL and NULL. Similarly, to write a value, pass a pointer to the value via newp, and its length via newlen; otherwise pass NULL and 0. The mallctlnametomib() function provides a way to avoid repeated name lookups for applications that repeatedly query the same portion of the namespace, by translating a name to a Management Information Base (MIB) that can be passed repeatedly to mallctlbymib(). Upon successful return from mallctlnametomib(), mibp contains an array of *miblenp integers, where *miblenp is the lesser of the number of components in name and the input value of *miblenp. Thus it is possible to pass a *miblenp that is smaller than the number of period-separated name components, which results in a partial MIB that can be used as the basis for constructing a complete MIB. For name components that are integers (e.g. the 2 in arenas.bin.2.size), the corresponding MIB component will always be that integer. Therefore, it is legitimate to construct code like the following: The malloc_stats_print() function writes summary statistics via the write_cb callback function pointer and cbopaque data passed to write_cb, or malloc_message() if write_cb is NULL. The statistics are presented in human-readable form unless J is specified as a character within the opts string, in which case the statistics are presented in JSON format. This function can be called repeatedly. General information that never changes during execution can be omitted by specifying g as a character within the opts string. Note that malloc_message() uses the mallctl*() functions internally, so inconsistent statistics can be reported if multiple threads use these functions simultaneously. If is specified during configuration, m and a can be specified to omit merged arena and per arena statistics, respectively; b, l, and h can be specified to omit per size class statistics for bins, large objects, and huge objects, respectively. Unrecognized characters are silently ignored. Note that thread caching may prevent some statistics from being completely up to date, since extra locking would be required to merge counters that track thread cache operations. The malloc_usable_size() function returns the usable size of the allocation pointed to by ptr. The return value may be larger than the size that was requested during allocation. The malloc_usable_size() function is not a mechanism for in-place realloc(); rather it is provided solely as a tool for introspection purposes. Any discrepancy between the requested allocation size and the size reported by malloc_usable_size() should not be depended on, since such behavior is entirely implementation-dependent. TUNING Once, when the first call is made to one of the memory allocation routines, the allocator initializes its internals based in part on various options that can be specified at compile- or run-time. The string specified via , the string pointed to by the global variable malloc_conf, the name of the file referenced by the symbolic link named /etc/malloc.conf, and the value of the environment variable MALLOC_CONF, will be interpreted, in that order, from left to right as options. Note that malloc_conf may be read before main() is entered, so the declaration of malloc_conf should specify an initializer that contains the final value to be read by jemalloc. and malloc_conf are compile-time mechanisms, whereas /etc/malloc.conf and MALLOC_CONF can be safely set any time prior to program invocation. An options string is a comma-separated list of option:value pairs. There is one key corresponding to each opt.* mallctl (see the section for options documentation). For example, abort:true,narenas:1 sets the opt.abort and opt.narenas options. Some options have boolean values (true/false), others have integer values (base 8, 10, or 16, depending on prefix), and yet others have raw string values. IMPLEMENTATION NOTES Traditionally, allocators have used sbrk 2 to obtain memory, which is suboptimal for several reasons, including race conditions, increased fragmentation, and artificial limitations on maximum usable memory. If sbrk 2 is supported by the operating system, this allocator uses both mmap 2 and sbrk 2, in that order of preference; otherwise only mmap 2 is used. This allocator uses multiple arenas in order to reduce lock contention for threaded programs on multi-processor systems. This works well with regard to threading scalability, but incurs some costs. There is a small fixed per-arena overhead, and additionally, arenas manage memory completely independently of each other, which means a small fixed increase in overall memory fragmentation. These overheads are not generally an issue, given the number of arenas normally used. Note that using substantially more arenas than the default is not likely to improve performance, mainly due to reduced cache performance. However, it may make sense to reduce the number of arenas if an application does not make much use of the allocation functions. In addition to multiple arenas, unless is specified during configuration, this allocator supports thread-specific caching for small and large objects, in order to make it possible to completely avoid synchronization for most allocation requests. Such caching allows very fast allocation in the common case, but it increases memory usage and fragmentation, since a bounded number of objects can remain allocated in each thread cache. Memory is conceptually broken into equal-sized chunks, where the chunk size is a power of two that is greater than the page size. Chunks are always aligned to multiples of the chunk size. This alignment makes it possible to find metadata for user objects very quickly. User objects are broken into three categories according to size: small, large, and huge. Multiple small and large objects can reside within a single chunk, whereas huge objects each have one or more chunks backing them. Each chunk that contains small and/or large objects tracks its contents as runs of contiguous pages (unused, backing a set of small objects, or backing one large object). The combination of chunk alignment and chunk page maps makes it possible to determine all metadata regarding small and large allocations in constant time. Small objects are managed in groups by page runs. Each run maintains a bitmap to track which regions are in use. Allocation requests that are no more than half the quantum (8 or 16, depending on architecture) are rounded up to the nearest power of two that is at least sizeof(double). All other object size classes are multiples of the quantum, spaced such that there are four size classes for each doubling in size, which limits internal fragmentation to approximately 20% for all but the smallest size classes. Small size classes are smaller than four times the page size, large size classes are smaller than the chunk size (see the opt.lg_chunk option), and huge size classes extend from the chunk size up to the largest size class that does not exceed PTRDIFF_MAX. Allocations are packed tightly together, which can be an issue for multi-threaded applications. If you need to assure that allocations do not suffer from cacheline sharing, round your allocation requests up to the nearest multiple of the cacheline size, or specify cacheline alignment when allocating. The realloc(), rallocx(), and xallocx() functions may resize allocations without moving them under limited circumstances. Unlike the *allocx() API, the standard API does not officially round up the usable size of an allocation to the nearest size class, so technically it is necessary to call realloc() to grow e.g. a 9-byte allocation to 16 bytes, or shrink a 16-byte allocation to 9 bytes. Growth and shrinkage trivially succeeds in place as long as the pre-size and post-size both round up to the same size class. No other API guarantees are made regarding in-place resizing, but the current implementation also tries to resize large and huge allocations in place, as long as the pre-size and post-size are both large or both huge. In such cases shrinkage always succeeds for large size classes, but for huge size classes the chunk allocator must support splitting (see arena.<i>.chunk_hooks). Growth only succeeds if the trailing memory is currently available, and additionally for huge size classes the chunk allocator must support merging. Assuming 2 MiB chunks, 4 KiB pages, and a 16-byte quantum on a 64-bit system, the size classes in each category are as shown in . Size classes Category Spacing Size Small lg [8] 16 [16, 32, 48, 64, 80, 96, 112, 128] 32 [160, 192, 224, 256] 64 [320, 384, 448, 512] 128 [640, 768, 896, 1024] 256 [1280, 1536, 1792, 2048] 512 [2560, 3072, 3584, 4096] 1 KiB [5 KiB, 6 KiB, 7 KiB, 8 KiB] 2 KiB [10 KiB, 12 KiB, 14 KiB] Large 2 KiB [16 KiB] 4 KiB [20 KiB, 24 KiB, 28 KiB, 32 KiB] 8 KiB [40 KiB, 48 KiB, 54 KiB, 64 KiB] 16 KiB [80 KiB, 96 KiB, 112 KiB, 128 KiB] 32 KiB [160 KiB, 192 KiB, 224 KiB, 256 KiB] 64 KiB [320 KiB, 384 KiB, 448 KiB, 512 KiB] 128 KiB [640 KiB, 768 KiB, 896 KiB, 1 MiB] 256 KiB [1280 KiB, 1536 KiB, 1792 KiB] Huge 256 KiB [2 MiB] 512 KiB [2560 KiB, 3 MiB, 3584 KiB, 4 MiB] 1 MiB [5 MiB, 6 MiB, 7 MiB, 8 MiB] 2 MiB [10 MiB, 12 MiB, 14 MiB, 16 MiB] 4 MiB [20 MiB, 24 MiB, 28 MiB, 32 MiB] 8 MiB [40 MiB, 48 MiB, 56 MiB, 64 MiB] ... ... 512 PiB [2560 PiB, 3 EiB, 3584 PiB, 4 EiB] 1 EiB [5 EiB, 6 EiB, 7 EiB]
MALLCTL NAMESPACE The following names are defined in the namespace accessible via the mallctl*() functions. Value types are specified in parentheses, their readable/writable statuses are encoded as rw, r-, -w, or --, and required build configuration flags follow, if any. A name element encoded as <i> or <j> indicates an integer component, where the integer varies from 0 to some upper value that must be determined via introspection. In the case of stats.arenas.<i>.*, <i> equal to arenas.narenas can be used to access the summation of statistics from all arenas. Take special note of the epoch mallctl, which controls refreshing of cached dynamic statistics. version (const char *) r- Return the jemalloc version string. epoch (uint64_t) rw If a value is passed in, refresh the data from which the mallctl*() functions report values, and increment the epoch. Return the current epoch. This is useful for detecting whether another thread caused a refresh. config.cache_oblivious (bool) r- was specified during build configuration. config.debug (bool) r- was specified during build configuration. config.fill (bool) r- was specified during build configuration. config.lazy_lock (bool) r- was specified during build configuration. config.malloc_conf (const char *) r- Embedded configure-time-specified run-time options string, empty unless was specified during build configuration. config.munmap (bool) r- was specified during build configuration. config.prof (bool) r- was specified during build configuration. config.prof_libgcc (bool) r- was not specified during build configuration. config.prof_libunwind (bool) r- was specified during build configuration. config.stats (bool) r- was specified during build configuration. config.tcache (bool) r- was not specified during build configuration. config.tls (bool) r- was not specified during build configuration. config.utrace (bool) r- was specified during build configuration. config.valgrind (bool) r- was specified during build configuration. config.xmalloc (bool) r- was specified during build configuration. opt.abort (bool) r- Abort-on-warning enabled/disabled. If true, most warnings are fatal. The process will call abort 3 in these cases. This option is disabled by default unless is specified during configuration, in which case it is enabled by default. opt.dss (const char *) r- dss (sbrk 2) allocation precedence as related to mmap 2 allocation. The following settings are supported if sbrk 2 is supported by the operating system: disabled, primary, and secondary; otherwise only disabled is supported. The default is secondary if sbrk 2 is supported by the operating system; disabled otherwise. opt.lg_chunk (size_t) r- Virtual memory chunk size (log base 2). If a chunk size outside the supported size range is specified, the size is silently clipped to the minimum/maximum supported size. The default chunk size is 2 MiB (2^21). opt.narenas (unsigned) r- Maximum number of arenas to use for automatic multiplexing of threads and arenas. The default is four times the number of CPUs, or one if there is a single CPU. opt.purge (const char *) r- Purge mode is “ratio” (default) or “decay”. See opt.lg_dirty_mult for details of the ratio mode. See opt.decay_time for details of the decay mode. opt.lg_dirty_mult (ssize_t) r- Per-arena minimum ratio (log base 2) of active to dirty pages. Some dirty unused pages may be allowed to accumulate, within the limit set by the ratio (or one chunk worth of dirty pages, whichever is greater), before informing the kernel about some of those pages via madvise 2 or a similar system call. This provides the kernel with sufficient information to recycle dirty pages if physical memory becomes scarce and the pages remain unused. The default minimum ratio is 8:1 (2^3:1); an option value of -1 will disable dirty page purging. See arenas.lg_dirty_mult and arena.<i>.lg_dirty_mult for related dynamic control options. opt.decay_time (ssize_t) r- Approximate time in seconds from the creation of a set of unused dirty pages until an equivalent set of unused dirty pages is purged and/or reused. The pages are incrementally purged according to a sigmoidal decay curve that starts and ends with zero purge rate. A decay time of 0 causes all unused dirty pages to be purged immediately upon creation. A decay time of -1 disables purging. The default decay time is 10 seconds. See arenas.decay_time and arena.<i>.decay_time for related dynamic control options. opt.stats_print (bool) r- Enable/disable statistics printing at exit. If enabled, the malloc_stats_print() function is called at program exit via an atexit 3 function. If is specified during configuration, this has the potential to cause deadlock for a multi-threaded process that exits while one or more threads are executing in the memory allocation functions. Furthermore, atexit() may allocate memory during application initialization and then deadlock internally when jemalloc in turn calls atexit(), so this option is not universally usable (though the application can register its own atexit() function with equivalent functionality). Therefore, this option should only be used with care; it is primarily intended as a performance tuning aid during application development. This option is disabled by default. opt.junk (const char *) r- [] Junk filling. If set to alloc, each byte of uninitialized allocated memory will be initialized to 0xa5. If set to free, all deallocated memory will be initialized to 0x5a. If set to true, both allocated and deallocated memory will be initialized, and if set to false, junk filling be disabled entirely. This is intended for debugging and will impact performance negatively. This option is false by default unless is specified during configuration, in which case it is true by default unless running inside Valgrind. opt.quarantine (size_t) r- [] Per thread quarantine size in bytes. If non-zero, each thread maintains a FIFO object quarantine that stores up to the specified number of bytes of memory. The quarantined memory is not freed until it is released from quarantine, though it is immediately junk-filled if the opt.junk option is enabled. This feature is of particular use in combination with Valgrind, which can detect attempts to access quarantined objects. This is intended for debugging and will impact performance negatively. The default quarantine size is 0 unless running inside Valgrind, in which case the default is 16 MiB. opt.redzone (bool) r- [] Redzones enabled/disabled. If enabled, small allocations have redzones before and after them. Furthermore, if the opt.junk option is enabled, the redzones are checked for corruption during deallocation. However, the primary intended purpose of this feature is to be used in combination with Valgrind, which needs redzones in order to do effective buffer overflow/underflow detection. This option is intended for debugging and will impact performance negatively. This option is disabled by default unless running inside Valgrind. opt.zero (bool) r- [] Zero filling enabled/disabled. If enabled, each byte of uninitialized allocated memory will be initialized to 0. Note that this initialization only happens once for each byte, so realloc() and rallocx() calls do not zero memory that was previously allocated. This is intended for debugging and will impact performance negatively. This option is disabled by default. opt.utrace (bool) r- [] Allocation tracing based on utrace 2 enabled/disabled. This option is disabled by default. opt.xmalloc (bool) r- [] Abort-on-out-of-memory enabled/disabled. If enabled, rather than returning failure for any allocation function, display a diagnostic message on STDERR_FILENO and cause the program to drop core (using abort 3). If an application is designed to depend on this behavior, set the option at compile time by including the following in the source code: This option is disabled by default. opt.tcache (bool) r- [] Thread-specific caching (tcache) enabled/disabled. When there are multiple threads, each thread uses a tcache for objects up to a certain size. Thread-specific caching allows many allocations to be satisfied without performing any thread synchronization, at the cost of increased memory use. See the opt.lg_tcache_max option for related tuning information. This option is enabled by default unless running inside Valgrind, in which case it is forcefully disabled. opt.lg_tcache_max (size_t) r- [] Maximum size class (log base 2) to cache in the thread-specific cache (tcache). At a minimum, all small size classes are cached, and at a maximum all large size classes are cached. The default maximum is 32 KiB (2^15). opt.prof (bool) r- [] Memory profiling enabled/disabled. If enabled, profile memory allocation activity. See the opt.prof_active option for on-the-fly activation/deactivation. See the opt.lg_prof_sample option for probabilistic sampling control. See the opt.prof_accum option for control of cumulative sample reporting. See the opt.lg_prof_interval option for information on interval-triggered profile dumping, the opt.prof_gdump option for information on high-water-triggered profile dumping, and the opt.prof_final option for final profile dumping. Profile output is compatible with the jeprof command, which is based on the pprof that is developed as part of the gperftools package. See HEAP PROFILE FORMAT for heap profile format documentation. opt.prof_prefix (const char *) r- [] Filename prefix for profile dumps. If the prefix is set to the empty string, no automatic dumps will occur; this is primarily useful for disabling the automatic final heap dump (which also disables leak reporting, if enabled). The default prefix is jeprof. opt.prof_active (bool) r- [] Profiling activated/deactivated. This is a secondary control mechanism that makes it possible to start the application with profiling enabled (see the opt.prof option) but inactive, then toggle profiling at any time during program execution with the prof.active mallctl. This option is enabled by default. opt.prof_thread_active_init (bool) r- [] Initial setting for thread.prof.active in newly created threads. The initial setting for newly created threads can also be changed during execution via the prof.thread_active_init mallctl. This option is enabled by default. opt.lg_prof_sample (size_t) r- [] Average interval (log base 2) between allocation samples, as measured in bytes of allocation activity. Increasing the sampling interval decreases profile fidelity, but also decreases the computational overhead. The default sample interval is 512 KiB (2^19 B). opt.prof_accum (bool) r- [] Reporting of cumulative object/byte counts in profile dumps enabled/disabled. If this option is enabled, every unique backtrace must be stored for the duration of execution. Depending on the application, this can impose a large memory overhead, and the cumulative counts are not always of interest. This option is disabled by default. opt.lg_prof_interval (ssize_t) r- [] Average interval (log base 2) between memory profile dumps, as measured in bytes of allocation activity. The actual interval between dumps may be sporadic because decentralized allocation counters are used to avoid synchronization bottlenecks. Profiles are dumped to files named according to the pattern <prefix>.<pid>.<seq>.i<iseq>.heap, where <prefix> is controlled by the opt.prof_prefix option. By default, interval-triggered profile dumping is disabled (encoded as -1). opt.prof_gdump (bool) r- [] Set the initial state of prof.gdump, which when enabled triggers a memory profile dump every time the total virtual memory exceeds the previous maximum. This option is disabled by default. opt.prof_final (bool) r- [] Use an atexit 3 function to dump final memory usage to a file named according to the pattern <prefix>.<pid>.<seq>.f.heap, where <prefix> is controlled by the opt.prof_prefix option. Note that atexit() may allocate memory during application initialization and then deadlock internally when jemalloc in turn calls atexit(), so this option is not universally usable (though the application can register its own atexit() function with equivalent functionality). This option is disabled by default. opt.prof_leak (bool) r- [] Leak reporting enabled/disabled. If enabled, use an atexit 3 function to report memory leaks detected by allocation sampling. See the opt.prof option for information on analyzing heap profile output. This option is disabled by default. thread.arena (unsigned) rw Get or set the arena associated with the calling thread. If the specified arena was not initialized beforehand (see the arenas.initialized mallctl), it will be automatically initialized as a side effect of calling this interface. thread.allocated (uint64_t) r- [] Get the total number of bytes ever allocated by the calling thread. This counter has the potential to wrap around; it is up to the application to appropriately interpret the counter in such cases. thread.allocatedp (uint64_t *) r- [] Get a pointer to the the value that is returned by the thread.allocated mallctl. This is useful for avoiding the overhead of repeated mallctl*() calls. thread.deallocated (uint64_t) r- [] Get the total number of bytes ever deallocated by the calling thread. This counter has the potential to wrap around; it is up to the application to appropriately interpret the counter in such cases. thread.deallocatedp (uint64_t *) r- [] Get a pointer to the the value that is returned by the thread.deallocated mallctl. This is useful for avoiding the overhead of repeated mallctl*() calls. thread.tcache.enabled (bool) rw [] Enable/disable calling thread's tcache. The tcache is implicitly flushed as a side effect of becoming disabled (see thread.tcache.flush). thread.tcache.flush (void) -- [] Flush calling thread's thread-specific cache (tcache). This interface releases all cached objects and internal data structures associated with the calling thread's tcache. Ordinarily, this interface need not be called, since automatic periodic incremental garbage collection occurs, and the thread cache is automatically discarded when a thread exits. However, garbage collection is triggered by allocation activity, so it is possible for a thread that stops allocating/deallocating to retain its cache indefinitely, in which case the developer may find manual flushing useful. thread.prof.name (const char *) r- or -w [] Get/set the descriptive name associated with the calling thread in memory profile dumps. An internal copy of the name string is created, so the input string need not be maintained after this interface completes execution. The output string of this interface should be copied for non-ephemeral uses, because multiple implementation details can cause asynchronous string deallocation. Furthermore, each invocation of this interface can only read or write; simultaneous read/write is not supported due to string lifetime limitations. The name string must be nil-terminated and comprised only of characters in the sets recognized by isgraph 3 and isblank 3. thread.prof.active (bool) rw [] Control whether sampling is currently active for the calling thread. This is an activation mechanism in addition to prof.active; both must be active for the calling thread to sample. This flag is enabled by default. tcache.create (unsigned) r- [] Create an explicit thread-specific cache (tcache) and return an identifier that can be passed to the MALLOCX_TCACHE(tc) macro to explicitly use the specified cache rather than the automatically managed one that is used by default. Each explicit cache can be used by only one thread at a time; the application must assure that this constraint holds. tcache.flush (unsigned) -w [] Flush the specified thread-specific cache (tcache). The same considerations apply to this interface as to thread.tcache.flush, except that the tcache will never be automatically discarded. tcache.destroy (unsigned) -w [] Flush the specified thread-specific cache (tcache) and make the identifier available for use during a future tcache creation. arena.<i>.purge (void) -- Purge all unused dirty pages for arena <i>, or for all arenas if <i> equals arenas.narenas. arena.<i>.decay (void) -- Trigger decay-based purging of unused dirty pages for arena <i>, or for all arenas if <i> equals arenas.narenas. The proportion of unused dirty pages to be purged depends on the current time; see opt.decay_time for details. arena.<i>.reset (void) -- Discard all of the arena's extant allocations. This interface can only be used with arenas created via arenas.extend. None of the arena's discarded/cached allocations may accessed afterward. As part of this requirement, all thread caches which were used to allocate/deallocate in conjunction with the arena must be flushed beforehand. This interface cannot be used if running inside Valgrind, nor if the quarantine size is non-zero. arena.<i>.dss (const char *) rw Set the precedence of dss allocation as related to mmap allocation for arena <i>, or for all arenas if <i> equals arenas.narenas. See opt.dss for supported settings. arena.<i>.lg_dirty_mult (ssize_t) rw Current per-arena minimum ratio (log base 2) of active to dirty pages for arena <i>. Each time this interface is set and the ratio is increased, pages are synchronously purged as necessary to impose the new ratio. See opt.lg_dirty_mult for additional information. arena.<i>.decay_time (ssize_t) rw Current per-arena approximate time in seconds from the creation of a set of unused dirty pages until an equivalent set of unused dirty pages is purged and/or reused. Each time this interface is set, all currently unused dirty pages are considered to have fully decayed, which causes immediate purging of all unused dirty pages unless the decay time is set to -1 (i.e. purging disabled). See opt.decay_time for additional information. arena.<i>.chunk_hooks (chunk_hooks_t) rw Get or set the chunk management hook functions for arena <i>. The functions must be capable of operating on all extant chunks associated with arena <i>, usually by passing unknown chunks to the replaced functions. In practice, it is feasible to control allocation for arenas created via arenas.extend such that all chunks originate from an application-supplied chunk allocator (by setting custom chunk hook functions just after arena creation), but the automatically created arenas may have already created chunks prior to the application having an opportunity to take over chunk allocation. The chunk_hooks_t structure comprises function pointers which are described individually below. jemalloc uses these functions to manage chunk lifetime, which starts off with allocation of mapped committed memory, in the simplest case followed by deallocation. However, there are performance and platform reasons to retain chunks for later reuse. Cleanup attempts cascade from deallocation to decommit to purging, which gives the chunk management functions opportunities to reject the most permanent cleanup operations in favor of less permanent (and often less costly) operations. The chunk splitting and merging operations can also be opted out of, but this is mainly intended to support platforms on which virtual memory mappings provided by the operating system kernel do not automatically coalesce and split, e.g. Windows. typedef void *(chunk_alloc_t) void *chunk size_t size size_t alignment bool *zero bool *commit unsigned arena_ind A chunk allocation function conforms to the chunk_alloc_t type and upon success returns a pointer to size bytes of mapped memory on behalf of arena arena_ind such that the chunk's base address is a multiple of alignment, as well as setting *zero to indicate whether the chunk is zeroed and *commit to indicate whether the chunk is committed. Upon error the function returns NULL and leaves *zero and *commit unmodified. The size parameter is always a multiple of the chunk size. The alignment parameter is always a power of two at least as large as the chunk size. Zeroing is mandatory if *zero is true upon function entry. Committing is mandatory if *commit is true upon function entry. If chunk is not NULL, the returned pointer must be chunk on success or NULL on error. Committed memory may be committed in absolute terms as on a system that does not overcommit, or in implicit terms as on a system that overcommits and satisfies physical memory needs on demand via soft page faults. Note that replacing the default chunk allocation function makes the arena's arena.<i>.dss setting irrelevant. typedef bool (chunk_dalloc_t) void *chunk size_t size bool committed unsigned arena_ind A chunk deallocation function conforms to the chunk_dalloc_t type and deallocates a chunk of given size with committed/decommited memory as indicated, on behalf of arena arena_ind, returning false upon success. If the function returns true, this indicates opt-out from deallocation; the virtual memory mapping associated with the chunk remains mapped, in the same commit state, and available for future use, in which case it will be automatically retained for later reuse. typedef bool (chunk_commit_t) void *chunk size_t size size_t offset size_t length unsigned arena_ind A chunk commit function conforms to the chunk_commit_t type and commits zeroed physical memory to back pages within a chunk of given size at offset bytes, extending for length on behalf of arena arena_ind, returning false upon success. Committed memory may be committed in absolute terms as on a system that does not overcommit, or in implicit terms as on a system that overcommits and satisfies physical memory needs on demand via soft page faults. If the function returns true, this indicates insufficient physical memory to satisfy the request. typedef bool (chunk_decommit_t) void *chunk size_t size size_t offset size_t length unsigned arena_ind A chunk decommit function conforms to the chunk_decommit_t type and decommits any physical memory that is backing pages within a chunk of given size at offset bytes, extending for length on behalf of arena arena_ind, returning false upon success, in which case the pages will be committed via the chunk commit function before being reused. If the function returns true, this indicates opt-out from decommit; the memory remains committed and available for future use, in which case it will be automatically retained for later reuse. typedef bool (chunk_purge_t) void *chunk size_tsize size_t offset size_t length unsigned arena_ind A chunk purge function conforms to the chunk_purge_t type and optionally discards physical pages within the virtual memory mapping associated with chunk of given size at offset bytes, extending for length on behalf of arena arena_ind, returning false if pages within the purged virtual memory range will be zero-filled the next time they are accessed. typedef bool (chunk_split_t) void *chunk size_t size size_t size_a size_t size_b bool committed unsigned arena_ind A chunk split function conforms to the chunk_split_t type and optionally splits chunk of given size into two adjacent chunks, the first of size_a bytes, and the second of size_b bytes, operating on committed/decommitted memory as indicated, on behalf of arena arena_ind, returning false upon success. If the function returns true, this indicates that the chunk remains unsplit and therefore should continue to be operated on as a whole. typedef bool (chunk_merge_t) void *chunk_a size_t size_a void *chunk_b size_t size_b bool committed unsigned arena_ind A chunk merge function conforms to the chunk_merge_t type and optionally merges adjacent chunks, chunk_a of given size_a and chunk_b of given size_b into one contiguous chunk, operating on committed/decommitted memory as indicated, on behalf of arena arena_ind, returning false upon success. If the function returns true, this indicates that the chunks remain distinct mappings and therefore should continue to be operated on independently. arenas.narenas (unsigned) r- Current limit on number of arenas. arenas.initialized (bool *) r- An array of arenas.narenas booleans. Each boolean indicates whether the corresponding arena is initialized. arenas.lg_dirty_mult (ssize_t) rw Current default per-arena minimum ratio (log base 2) of active to dirty pages, used to initialize arena.<i>.lg_dirty_mult during arena creation. See opt.lg_dirty_mult for additional information. arenas.decay_time (ssize_t) rw Current default per-arena approximate time in seconds from the creation of a set of unused dirty pages until an equivalent set of unused dirty pages is purged and/or reused, used to initialize arena.<i>.decay_time during arena creation. See opt.decay_time for additional information. arenas.quantum (size_t) r- Quantum size. arenas.page (size_t) r- Page size. arenas.tcache_max (size_t) r- [] Maximum thread-cached size class. arenas.nbins (unsigned) r- Number of bin size classes. arenas.nhbins (unsigned) r- [] Total number of thread cache bin size classes. arenas.bin.<i>.size (size_t) r- Maximum size supported by size class. arenas.bin.<i>.nregs (uint32_t) r- Number of regions per page run. arenas.bin.<i>.run_size (size_t) r- Number of bytes per page run. arenas.nlruns (unsigned) r- Total number of large size classes. arenas.lrun.<i>.size (size_t) r- Maximum size supported by this large size class. arenas.nhchunks (unsigned) r- Total number of huge size classes. arenas.hchunk.<i>.size (size_t) r- Maximum size supported by this huge size class. arenas.extend (unsigned) r- Extend the array of arenas by appending a new arena, and returning the new arena index. prof.thread_active_init (bool) rw [] Control the initial setting for thread.prof.active in newly created threads. See the opt.prof_thread_active_init option for additional information. prof.active (bool) rw [] Control whether sampling is currently active. See the opt.prof_active option for additional information, as well as the interrelated thread.prof.active mallctl. prof.dump (const char *) -w [] Dump a memory profile to the specified file, or if NULL is specified, to a file according to the pattern <prefix>.<pid>.<seq>.m<mseq>.heap, where <prefix> is controlled by the opt.prof_prefix option. prof.gdump (bool) rw [] When enabled, trigger a memory profile dump every time the total virtual memory exceeds the previous maximum. Profiles are dumped to files named according to the pattern <prefix>.<pid>.<seq>.u<useq>.heap, where <prefix> is controlled by the opt.prof_prefix option. prof.reset (size_t) -w [] Reset all memory profile statistics, and optionally update the sample rate (see opt.lg_prof_sample and prof.lg_sample). prof.lg_sample (size_t) r- [] Get the current sample rate (see opt.lg_prof_sample). prof.interval (uint64_t) r- [] Average number of bytes allocated between interval-based profile dumps. See the opt.lg_prof_interval option for additional information. stats.cactive (size_t *) r- [] Pointer to a counter that contains an approximate count of the current number of bytes in active pages. The estimate may be high, but never low, because each arena rounds up when computing its contribution to the counter. Note that the epoch mallctl has no bearing on this counter. Furthermore, counter consistency is maintained via atomic operations, so it is necessary to use an atomic operation in order to guarantee a consistent read when dereferencing the pointer. stats.allocated (size_t) r- [] Total number of bytes allocated by the application. stats.active (size_t) r- [] Total number of bytes in active pages allocated by the application. This is a multiple of the page size, and greater than or equal to stats.allocated. This does not include stats.arenas.<i>.pdirty, nor pages entirely devoted to allocator metadata. stats.metadata (size_t) r- [] Total number of bytes dedicated to metadata, which comprise base allocations used for bootstrap-sensitive internal allocator data structures, arena chunk headers (see stats.arenas.<i>.metadata.mapped), and internal allocations (see stats.arenas.<i>.metadata.allocated). stats.resident (size_t) r- [] Maximum number of bytes in physically resident data pages mapped by the allocator, comprising all pages dedicated to allocator metadata, pages backing active allocations, and unused dirty pages. This is a maximum rather than precise because pages may not actually be physically resident if they correspond to demand-zeroed virtual memory that has not yet been touched. This is a multiple of the page size, and is larger than stats.active. stats.mapped (size_t) r- [] Total number of bytes in active chunks mapped by the allocator. This is a multiple of the chunk size, and is larger than stats.active. This does not include inactive chunks, even those that contain unused dirty pages, which means that there is no strict ordering between this and stats.resident. stats.retained (size_t) r- [] Total number of bytes in virtual memory mappings that were retained rather than being returned to the operating system via e.g. munmap 2. Retained virtual memory is typically untouched, decommitted, or purged, so it has no strongly associated physical memory (see chunk hooks for details). Retained memory is excluded from mapped memory statistics, e.g. stats.mapped. stats.arenas.<i>.dss (const char *) r- dss (sbrk 2) allocation precedence as related to mmap 2 allocation. See opt.dss for details. stats.arenas.<i>.lg_dirty_mult (ssize_t) r- Minimum ratio (log base 2) of active to dirty pages. See opt.lg_dirty_mult for details. stats.arenas.<i>.decay_time (ssize_t) r- Approximate time in seconds from the creation of a set of unused dirty pages until an equivalent set of unused dirty pages is purged and/or reused. See opt.decay_time for details. stats.arenas.<i>.nthreads (unsigned) r- Number of threads currently assigned to arena. stats.arenas.<i>.pactive (size_t) r- Number of pages in active runs. stats.arenas.<i>.pdirty (size_t) r- Number of pages within unused runs that are potentially dirty, and for which madvise... MADV_DONTNEED or similar has not been called. stats.arenas.<i>.mapped (size_t) r- [] Number of mapped bytes. stats.arenas.<i>.retained (size_t) r- [] Number of retained bytes. See stats.retained for details. stats.arenas.<i>.metadata.mapped (size_t) r- [] Number of mapped bytes in arena chunk headers, which track the states of the non-metadata pages. stats.arenas.<i>.metadata.allocated (size_t) r- [] Number of bytes dedicated to internal allocations. Internal allocations differ from application-originated allocations in that they are for internal use, and that they are omitted from heap profiles. This statistic is reported separately from stats.metadata and stats.arenas.<i>.metadata.mapped because it overlaps with e.g. the stats.allocated and stats.active statistics, whereas the other metadata statistics do not. stats.arenas.<i>.npurge (uint64_t) r- [] Number of dirty page purge sweeps performed. stats.arenas.<i>.nmadvise (uint64_t) r- [] Number of madvise... MADV_DONTNEED or similar calls made to purge dirty pages. stats.arenas.<i>.purged (uint64_t) r- [] Number of pages purged. stats.arenas.<i>.small.allocated (size_t) r- [] Number of bytes currently allocated by small objects. stats.arenas.<i>.small.nmalloc (uint64_t) r- [] Cumulative number of allocation requests served by small bins. stats.arenas.<i>.small.ndalloc (uint64_t) r- [] Cumulative number of small objects returned to bins. stats.arenas.<i>.small.nrequests (uint64_t) r- [] Cumulative number of small allocation requests. stats.arenas.<i>.large.allocated (size_t) r- [] Number of bytes currently allocated by large objects. stats.arenas.<i>.large.nmalloc (uint64_t) r- [] Cumulative number of large allocation requests served directly by the arena. stats.arenas.<i>.large.ndalloc (uint64_t) r- [] Cumulative number of large deallocation requests served directly by the arena. stats.arenas.<i>.large.nrequests (uint64_t) r- [] Cumulative number of large allocation requests. stats.arenas.<i>.huge.allocated (size_t) r- [] Number of bytes currently allocated by huge objects. stats.arenas.<i>.huge.nmalloc (uint64_t) r- [] Cumulative number of huge allocation requests served directly by the arena. stats.arenas.<i>.huge.ndalloc (uint64_t) r- [] Cumulative number of huge deallocation requests served directly by the arena. stats.arenas.<i>.huge.nrequests (uint64_t) r- [] Cumulative number of huge allocation requests. stats.arenas.<i>.bins.<j>.nmalloc (uint64_t) r- [] Cumulative number of allocations served by bin. stats.arenas.<i>.bins.<j>.ndalloc (uint64_t) r- [] Cumulative number of allocations returned to bin. stats.arenas.<i>.bins.<j>.nrequests (uint64_t) r- [] Cumulative number of allocation requests. stats.arenas.<i>.bins.<j>.curregs (size_t) r- [] Current number of regions for this size class. stats.arenas.<i>.bins.<j>.nfills (uint64_t) r- [ ] Cumulative number of tcache fills. stats.arenas.<i>.bins.<j>.nflushes (uint64_t) r- [ ] Cumulative number of tcache flushes. stats.arenas.<i>.bins.<j>.nruns (uint64_t) r- [] Cumulative number of runs created. stats.arenas.<i>.bins.<j>.nreruns (uint64_t) r- [] Cumulative number of times the current run from which to allocate changed. stats.arenas.<i>.bins.<j>.curruns (size_t) r- [] Current number of runs. stats.arenas.<i>.lruns.<j>.nmalloc (uint64_t) r- [] Cumulative number of allocation requests for this size class served directly by the arena. stats.arenas.<i>.lruns.<j>.ndalloc (uint64_t) r- [] Cumulative number of deallocation requests for this size class served directly by the arena. stats.arenas.<i>.lruns.<j>.nrequests (uint64_t) r- [] Cumulative number of allocation requests for this size class. stats.arenas.<i>.lruns.<j>.curruns (size_t) r- [] Current number of runs for this size class. stats.arenas.<i>.hchunks.<j>.nmalloc (uint64_t) r- [] Cumulative number of allocation requests for this size class served directly by the arena. stats.arenas.<i>.hchunks.<j>.ndalloc (uint64_t) r- [] Cumulative number of deallocation requests for this size class served directly by the arena. stats.arenas.<i>.hchunks.<j>.nrequests (uint64_t) r- [] Cumulative number of allocation requests for this size class. stats.arenas.<i>.hchunks.<j>.curhchunks (size_t) r- [] Current number of huge allocations for this size class. HEAP PROFILE FORMAT Although the heap profiling functionality was originally designed to be compatible with the pprof command that is developed as part of the gperftools package, the addition of per thread heap profiling functionality required a different heap profile format. The jeprof command is derived from pprof, with enhancements to support the heap profile format described here. In the following hypothetical heap profile, [...] indicates elision for the sake of compactness. The following matches the above heap profile, but most tokens are replaced with <description> to indicate descriptions of the corresponding fields. / : : [: ] [...] : : [: ] [...] : : [: ] [...] @ [...] [...] : : [: ] : : [: ] : : [: ] [...] MAPPED_LIBRARIES: /maps>]]> DEBUGGING MALLOC PROBLEMS When debugging, it is a good idea to configure/build jemalloc with the and options, and recompile the program with suitable options and symbols for debugger support. When so configured, jemalloc incorporates a wide variety of run-time assertions that catch application errors such as double-free, write-after-free, etc. Programs often accidentally depend on uninitialized memory actually being filled with zero bytes. Junk filling (see the opt.junk option) tends to expose such bugs in the form of obviously incorrect results and/or coredumps. Conversely, zero filling (see the opt.zero option) eliminates the symptoms of such bugs. Between these two options, it is usually possible to quickly detect, diagnose, and eliminate such bugs. This implementation does not provide much detail about the problems it detects, because the performance impact for storing such information would be prohibitive. However, jemalloc does integrate with the most excellent Valgrind tool if the configuration option is enabled. DIAGNOSTIC MESSAGES If any of the memory allocation/deallocation functions detect an error or warning condition, a message will be printed to file descriptor STDERR_FILENO. Errors will result in the process dumping core. If the opt.abort option is set, most warnings are treated as errors. The malloc_message variable allows the programmer to override the function which emits the text strings forming the errors and warnings if for some reason the STDERR_FILENO file descriptor is not suitable for this. malloc_message() takes the cbopaque pointer argument that is NULL unless overridden by the arguments in a call to malloc_stats_print(), followed by a string pointer. Please note that doing anything which tries to allocate memory in this function is likely to result in a crash or deadlock. All messages are prefixed by <jemalloc>: . RETURN VALUES Standard API The malloc() and calloc() functions return a pointer to the allocated memory if successful; otherwise a NULL pointer is returned and errno is set to ENOMEM. The posix_memalign() function returns the value 0 if successful; otherwise it returns an error value. The posix_memalign() function will fail if: EINVAL The alignment parameter is not a power of 2 at least as large as sizeof(void *). ENOMEM Memory allocation error. The aligned_alloc() function returns a pointer to the allocated memory if successful; otherwise a NULL pointer is returned and errno is set. The aligned_alloc() function will fail if: EINVAL The alignment parameter is not a power of 2. ENOMEM Memory allocation error. The realloc() function returns a pointer, possibly identical to ptr, to the allocated memory if successful; otherwise a NULL pointer is returned, and errno is set to ENOMEM if the error was the result of an allocation failure. The realloc() function always leaves the original buffer intact when an error occurs. The free() function returns no value. Non-standard API The mallocx() and rallocx() functions return a pointer to the allocated memory if successful; otherwise a NULL pointer is returned to indicate insufficient contiguous memory was available to service the allocation request. The xallocx() function returns the real size of the resulting resized allocation pointed to by ptr, which is a value less than size if the allocation could not be adequately grown in place. The sallocx() function returns the real size of the allocation pointed to by ptr. The nallocx() returns the real size that would result from a successful equivalent mallocx() function call, or zero if insufficient memory is available to perform the size computation. The mallctl(), mallctlnametomib(), and mallctlbymib() functions return 0 on success; otherwise they return an error value. The functions will fail if: EINVAL newp is not NULL, and newlen is too large or too small. Alternatively, *oldlenp is too large or too small; in this case as much data as possible are read despite the error. ENOENT name or mib specifies an unknown/invalid value. EPERM Attempt to read or write void value, or attempt to write read-only value. EAGAIN A memory allocation failure occurred. EFAULT An interface with side effects failed in some way not directly related to mallctl*() read/write processing. The malloc_usable_size() function returns the usable size of the allocation pointed to by ptr. ENVIRONMENT The following environment variable affects the execution of the allocation functions: MALLOC_CONF If the environment variable MALLOC_CONF is set, the characters it contains will be interpreted as options. EXAMPLES To dump core whenever a problem occurs: ln -s 'abort:true' /etc/malloc.conf To specify in the source a chunk size that is 16 MiB: SEE ALSO madvise 2, mmap 2, sbrk 2, utrace 2, alloca 3, atexit 3, getpagesize 3 STANDARDS The malloc(), calloc(), realloc(), and free() functions conform to ISO/IEC 9899:1990 (ISO C90). The posix_memalign() function conforms to IEEE Std 1003.1-2001 (POSIX.1).