numa API

https://zhuanlan.zhihu.com/p/240461637

The libnuma library offers a simple programming interface to the NUMA (Non Uniform Memory Access) policy supported by the Linux kernel. On a NUMA architecture some memory areas have different latency or bandwidth than others.

Available policies are page interleaving (i.e., allocate in a round-robin fashion from all, or a subset, of the nodes on the system), preferred node allocation (i.e., preferably allocate on a particular node), local allocation (i.e., allocate on the node on which the task is currently executing), or allocation only on specific nodes (i.e., allocate on some subset of the available nodes). It is also possible to bind tasks to specific nodes.

Numa memory allocation policy may be specified as a per-task attribute, that is inherited by children tasks and processes, or as an attribute of a range of process virtual address space. Numa memory policies specified for a range of virtual address space are shared by all tasks in the process. Further more, memory policies specified for a range of a shared memory attached using shmat(2) or mmap(2) from shmfs/hugetlbfs are shared by all processes that attach to that region. Memory policies for shared disk backed file mappings are currently ignored.

The default memory allocation policy for tasks and all memory range is local allocation. This assumes that no ancestor has installed a non-default policy.

For setting a specific policy globally for all memory allocations in a process and its children it is easiest to start it with the numactl(8) utility. For more finegrained policy inside an application this library can be used.

All numa memory allocation policy only takes effect when a page is actually faulted into the address space of a process by accessing it. The numa_alloc_* functions take care of this automatically.

A node is defined as an area where all memory has the same speed as seen from a particular CPU. A node can contain multiple CPUs. Caches are ignored for this definition.

Most functions in this library are only concerned about numa nodes and their memory. The exceptions to this are: int numa_node_to_cpus(int node, struct bitmask *mask), int numa_node_of_cpu(int cpu), void numa_bind(struct bitmask *nodemask), int numa_run_on_node(int node), int numa_run_on_node_mask(struct bitmask *nodemask) and struct bitmask *numa_get_run_node_mask(void). These functions deal with the CPUs associated with numa nodes. See the descriptions below for more information.

Some of these functions accept or return a pointer to struct bitmask. A struct bitmask controls a bit map of arbitrary length containing a bit representation of nodes. The predefined variable struct bitmask *numa_all_nodes_ptr points to a bit mask that has all available nodes set; struct bitmask *numa_no_nodes_ptr points to the empty set.

Before any other calls in this library can be used int numa_available(void) must be called. If it returns -1, all other functions in this library are undefined.
int numa_max_possible_node(void)

returns the number of the highest possible node in a system. In other words, the size of a kernel type nodemask_t (in bits) minus 1. This number can be gotten by calling int int numa_num_possible_nodes() and subtracting 1.
int numa_num_possible_nodes()

returns the size of kernel’s node mask (kernel type nodemask_t). In other words, large enough to represent the maximum number of nodes that the kernel can handle. This will match the kernel’s MAX_NUMNODES value. This count is derived from /proc/self/status, field Mems_allowed.
int numa_max_node(void)

returns the highest node number available on the current system. (See the node numbers in /sys/devices/system/node/ ). Also see int numa_num_configured_nodes().
int numa_num_configured_nodes()

returns the number of memory nodes in the system. This count includes any nodes that are currently disabled. This count is derived from the node numbers in /sys/devices/system/node. (Depends on the kernel being configured with /sys (CONFIG_SYSFS)).
struct bitmask *numa_get_mems_allowed(void)

returns the mask of nodes from which the process is allowed to allocate memory in it’s current cpuset context. Any nodes that are not included in the returned bitmask will be ignored in any of the following libnuma memory policy calls.
int numa_num_configured_cpus(void)

returns the number of cpus in the system. This count includes any cpus that are currently disabled. This count is derived from the cpu numbers in /sys/devices/system/cpu. If the kernel is configured without /sys (CONFIG_SYSFS=n) then it falls back to using the number of online cpus.
struct bitmask *numa_all_nodes_ptr

points to a bitmask that is allocated by the library with bits representing all nodes on which the calling task may allocate memory. This set may be up to all nodes on the system, or up to the nodes in the current cpuset. The bitmask is allocated by a call to struct bitmask *numa_allocate_nodemask() using size int numa_max_possible_node(void). The set of nodes to record is derived from /proc/self/status, field “Mems_allowed”. The user should not alter this bitmask.
struct bitmask *numa_no_nodes_ptr

points to a bitmask that is allocated by the library and left all zeroes. The bitmask is allocated by a call to struct bitmask *numa_allocate_nodemask() using size int numa_max_possible_node(void). The user should not alter this bitmask.
struct bitmask *numa_all_cpus_ptr

points to a bitmask that is allocated by the library with bits representing all cpus on which the calling task may execute. This set may be up to all cpus on the system, or up to the cpus in the current cpuset. The bitmask is allocated by a call to struct bitmask *numa_allocate_cpumask() using size numa_num_possible_cpus(). The set of cpus to record is derived from /proc/self/status, field “Cpus_allowed”. The user should not alter this bitmask.
int numa_num_task_cpus()

returns the number of cpus that the calling task is allowed to use. This count is derived from the map /proc/self/status, field “Cpus_allowed”. Also see the bitmask struct bitmask *numa_all_cpus_ptr.
int numa_num_task_nodes()

returns the number of nodes on which the calling task is allowed to allocate memory. This count is derived from the map /proc/self/status, field “Mems_allowed”. Also see the bitmask struct bitmask *numa_all_nodes_ptr.
int numa_parse_bitmap(char *line , struct bitmask *mask)

parses line , which is a character string such as found in /sys/devices/system/node/nodeN/cpumap into a bitmask structure. The string contains the hexadecimal representation of a bit map. The bitmask may be allocated with struct bitmask *numa_allocate_cpumask(). Returns 0 on success. Returns -1 on failure. This function is probably of little use to a user application, but it is used by libnuma internally.
struct bitmask *numa_parse_nodestring(char *string)

parses a character string list of nodes into a bit mask. The bit mask is allocated by struct bitmask *numa_allocate_nodemask(). The string is a comma-separated list of node numbers or node ranges. A leading ! can be used to indicate “not” this list (in other words, all nodes except this list), and a leading + can be used to indicate that the node numbers in the list are relative to the task’s cpuset. The string can be “all” to specify all ( int numa_num_task_nodes() ) nodes. Node numbers are limited by the number in the system. See int numa_max_node(void) and int numa_num_configured_nodes().
Examples: 1-5,7,10 !4-5 +0-3
If the string is of 0 length, bitmask struct bitmask *numa_no_nodes_ptr is returned. Returns 0 if the string is invalid.
struct bitmask *numa_parse_cpustring(char *string)

parses a character string list of cpus into a bit mask. The bit mask is allocated by struct bitmask *numa_allocate_cpumask(). The string is a comma-separated list of cpu numbers or cpu ranges. A leading ! can be used to indicate “not” this list (in other words, all cpus except this list), and a leading + can be used to indicate that the cpu numbers in the list are relative to the task’s cpuset. The string can be “all” to specify all ( int numa_num_task_cpus() ) cpus. Cpu numbers are limited by the number in the system. See int numa_num_task_cpus() and int numa_num_configured_cpus(void).
Examples: 1-5,7,10 !4-5 +0-3
Returns 0 if the string is invalid.
long numa_node_size(int node, long *freep)

returns the memory size of a node. If the argument freep is not NULL, it used to return the amount of free memory on the node. On error it returns -1.
long long numa_node_size64(int node, long long *freep)

works the same as long numa_node_size(int node, long *freep) except that it returns values as long long instead of long. This is useful on 32-bit architectures with large nodes.
int numa_preferred(void)

returns the preferred node of the current task. This is the node on which the kernel preferably allocates memory, unless some other policy overrides this.
void numa_set_preferred(int node)

sets the preferred node for the current task to node. The system will attempt to allocate memory from the preferred node, but will fall back to other nodes if no memory is available on the the preferred node. Passing a node of -1 argument specifies local allocation and is equivalent to calling void numa_set_localalloc(void).
struct bitmask *numa_get_interleave_mask(void)

returns the current interleave mask if the task’s memory allocation policy is page interleaved. Otherwise, this function returns an empty mask.
void numa_set_interleave_mask(struct bitmask *nodemask)

sets the memory interleave mask for the current task to nodemask. All new memory allocations are page interleaved over all nodes in the interleave mask. Interleaving can be turned off again by passing an empty mask (numa_no_nodes). The page interleaving only occurs on the actual page fault that puts a new page into the current address space. It is also only a hint: the kernel will fall back to other nodes if no memory is available on the interleave target.
void numa_interleave_memory(void *start, size_t size, struct bitmask *nodemask)

interleaves size bytes of memory page by page from start on nodes specified in nodemask. The size argument will be rounded up to a multiple of the system page size. If nodemask contains nodes that are externally denied to this process, this call will fail. This is a lower level function to interleave allocated but not yet faulted in memory. Not yet faulted in means the memory is allocated using mmap(2) or shmat(2), but has not been accessed by the current process yet. The memory is page interleaved to all nodes specified in nodemask. Normally void *numa_alloc_interleaved(size_t size) should be used for private memory instead, but this function is useful to handle shared memory areas. To be useful the memory area should be several megabytes at least (or tens of megabytes of hugetlbfs mappings) If the void numa_set_strict(int strict) flag is true then the operation will cause a numa_error if there were already pages in the mapping that do not follow the policy.
void numa_bind(struct bitmask *nodemask)

binds the current task and its children to the nodes specified in nodemask. They will only run on the CPUs of the specified nodes and only be able to allocate memory from them. This function is equivalent to calling numa_run_on_node_mask(nodemask) followed by numa_set_membind(nodemask). If tasks should be bound to individual CPUs inside nodes consider using numa_node_to_cpus and the sched_setaffinity(2) syscall.
void numa_set_localalloc(void)

sets the memory allocation policy for the calling task to local allocation. In this mode, the preferred node for memory allocation is effectively the node where the task is executing at the time of a page allocation.
void numa_set_membind(struct bitmask *nodemask)

sets the memory allocation mask. The task will only allocate memory from the nodes set in nodemask. Passing an empty nodemask or a nodemask that contains nodes other than those in the mask returned by struct bitmask *numa_get_mems_allowed(void) will result in an error.
struct bitmask *numa_get_membind(void)

returns the mask of nodes from which memory can currently be allocated. If the returned mask is equal to numa_all_nodes, then memory allocation is allowed from all nodes.
void *numa_alloc_onnode(size_t size, int node)

allocates memory on a specific node. The size argument will be rounded up to a multiple of the system page size. if the specified node is externally denied to this process, this call will fail. This function is relatively slow compared to the malloc(3), family of functions. The memory must be freed with void numa_free(void *start, size_t size). On errors NULL is returned.
void *numa_alloc_local(size_t size)

allocates size bytes of memory on the local node. The size argument will be rounded up to a multiple of the system page size. This function is relatively slow compared to the malloc(3) family of functions. The memory must be freed with void numa_free(void *start, size_t size). On errors NULL is returned.
void *numa_alloc_interleaved(size_t size)

allocates size bytes of memory page interleaved on all nodes. This function is relatively slow and should only be used for large areas consisting of multiple pages. The interleaving works at page level and will only show an effect when the area is large. The allocated memory must be freed with void numa_free(void *start, size_t size). On error, NULL is returned.
void *numa_alloc_interleaved_subset(size_t size, struct bitmask *nodemask)

attempts to allocate size bytes of memory page interleaved on all nodes. The size argument will be rounded up to a multiple of the system page size. The nodes on which a process is allowed to allocate memory may be constrained externally. If this is the case, this function may fail. This function is relatively slow compare to malloc(3), family of functions and should only be used for large areas consisting of multiple pages. The interleaving works at page level and will only show an effect when the area is large. The allocated memory must be freed with void numa_free(void *start, size_t size). On error, NULL is returned.
void *numa_alloc(size_t size)

allocates size bytes of memory with the current NUMA policy. The size argument will be rounded up to a multiple of the system page size. This function is relatively slow compare to the malloc(3) family of functions. The memory must be freed with void numa_free(void *start, size_t size). On errors NULL is returned.
void *numa_realloc(void *old_addr, size_t old_size, size_t new_size)

changes the size of the memory area pointed to by old_addr from old_size to new_size. The memory area pointed to by old_addr must have been allocated with one of the numa_alloc* functions. The new_size will be rounded up to a multiple of the system page size. The contents of the memory area will be unchanged to the minimum of the old and new sizes; newly allocated memory will be uninitialized. The memory policy (and node bindings) associated with the original memory area will be preserved in the resized area. For example, if the initial area was allocated with a call to void *numa_alloc_onnode(size_t size, int node), then the new pages (if the area is enlarged) will be allocated on the same node. However, if no memory policy was set for the original area, then void *numa_realloc(void *old_addr, size_t old_size, size_t new_size) cannot guarantee that the new pages will be allocated on the same node. On success, the address of the resized area is returned (which might be different from that of the initial area), otherwise NULL is returned and errno is set to indicate the error. The pointer returned by void *numa_realloc(void *old_addr, size_t old_size, size_t new_size) is suitable for passing to void numa_free(void *start, size_t size).

void numa_free(void start, size_t size) frees size bytes of memory starting at start, allocated by the numa_alloc_ functions above. The size argument will be rounded up to a multiple of the system page size.
int numa_run_on_node(int node)

runs the current task and its children on a specific node. They will not migrate to CPUs of other nodes until the node affinity is reset with a new call to int numa_run_on_node_mask(struct bitmask *nodemask). Passing -1 permits the kernel to schedule on all nodes again. On success, 0 is returned; on error -1 is returned, and errno is set to indicate the error.
int numa_run_on_node_mask(struct bitmask *nodemask)

runs the current task and its children only on nodes specified in nodemask. They will not migrate to CPUs of other nodes until the node affinity is reset with a new call to int numa_run_on_node_mask(struct bitmask *nodemask) or int numa_run_on_node(int node). Passing numa_all_nodes permits the kernel to schedule on all nodes again. On success, 0 is returned; on error -1 is returned, and errno is set to indicate the error.
struct bitmask *numa_get_run_node_mask(void)

returns a mask of CPUs on which the current task is allowed to run.
void numa_tonode_memory(void *start, size_t size, int node)

put memory on a specific node. The constraints described for void numa_interleave_memory(void *start, size_t size, struct bitmask *nodemask) apply here too.
void numa_tonodemask_memory(void *start, size_t size, struct bitmask *nodemask)

put memory on a specific set of nodes. The constraints described for void numa_interleave_memory(void *start, size_t size, struct bitmask *nodemask) apply here too.
void numa_setlocal_memory(void *start, size_t size)

locates memory on the current node. The constraints described for void numa_interleave_memory(void *start, size_t size, struct bitmask *nodemask) apply here too.
void numa_police_memory(void *start, size_t size)

locates memory with the current NUMA policy. The constraints described for void numa_interleave_memory(void *start, size_t size, struct bitmask *nodemask) apply here too.
int numa_distance(int node1, int node2)

reports the distance in the machine topology between two nodes. The factors are a multiple of 10. It returns 0 when the distance cannot be determined. A node has distance 10 to itself. Reporting the distance requires a Linux kernel version of 2.6.10 or newer.
void numa_set_bind_policy(int strict)

specifies whether calls that bind memory to a specific node should use the preferred policy or a strict policy. The preferred policy allows the kernel to allocate memory on other nodes when there isn’t enough free on the target node. strict will fail the allocation in that case. Setting the argument to specifies strict, 0 preferred. Note that specifying more than one node non strict may only use the first node in some kernel versions.
void numa_set_strict(int strict)

sets a flag that says whether the functions allocating on specific nodes should use use a strict policy. Strict means the allocation will fail if the memory cannot be allocated on the target node. Default operation is to fall back to other nodes. This doesn’t apply to interleave and default.
int numa_get_interleave_node(void)

is used by libnuma internally. It is probably not useful for user applications. It uses the MPOL_F_NODE flag of the get_mempolicy system call, which is not intended for application use (its operation may change or be removed altogether in future kernel versions). See get_mempolicy(2).
numa_pagesize()

returns the number of bytes in page. This function is simply a fast alternative to repeated calls to the getpagesize system call. See getpagesize(2).
int numa_sched_getaffinity(pid_t pid, struct bitmask *mask)

retrieves a bitmask of the cpus on which a task may run. The task is specified by pid. Returns the return value of the sched_getaffinity system call. See sched_getaffinity(2). The bitmask must be at least the size of the kernel’s cpu mask structure. Use struct bitmask *numa_allocate_cpumask() to allocate it. Test the bits in the mask by calling int numa_bitmask_isbitset(const struct bitmask *bmp, unsigned int n).
int numa_sched_setaffinity(pid_t pid, struct bitmask *mask)

sets a task’s allowed cpu’s to those cpu’s specified in mask. The task is specified by pid. Returns the return value of the sched_setaffinity system call. See sched_setaffinity(2). You may allocate the bitmask with struct bitmask *numa_allocate_cpumask(). Or the bitmask may be smaller than the kernel’s cpu mask structure. For example, call struct bitmask *numa_bitmask_alloc(unsigned int n) using a maximum number of cpus from int numa_num_configured_cpus(void). Set the bits in the mask by calling struct bitmask *numa_bitmask_setbit(struct bitmask *bmp, unsigned int n).
int numa_node_to_cpus(int node, struct bitmask *mask)

converts a node number to a bitmask of CPUs. The user must pass a bitmask structure with a mask buffer long enough to represent all possible cpu’s. Use struct bitmask *numa_allocate_cpumask() to create it. If the bitmask is not long enough errno will be set to ERANGE and -1 returned. On success 0 is returned.
int numa_node_of_cpu(int cpu)

returns the node that a cpu belongs to. If the user supplies an invalid cpu errno will be set to EINVAL and -1 will be returned.
numa_allocate_cpumask ()

returns a bitmask of a size equal to the kernel’s cpu mask (kernel type cpumask_t). In other words, large enough to represent NR_CPUS cpus. This number of cpus can be gotten by calling numa_num_possible_cpus(). The bitmask is zero-filled.

numa_free_cpumask frees a cpumask previously allocate by numa_allocate_cpumask.

struct bitmask *numa_allocate_nodemask() returns a bitmask of a size equal to the kernel’s node mask (kernel type nodemask_t). In other words, large enough to represent MAX_NUMNODES nodes. This number of nodes can be gotten by calling int int numa_num_possible_nodes(). The bitmask is zero-filled.
void numa_free_nodemask()

frees a nodemask previous allocated by struct bitmask *numa_allocate_nodemask().
struct bitmask *numa_bitmask_alloc(unsigned int n)

allocates a bitmask structure and its associated bit mask. The memory allocated for the bit mask contains enough words (type unsigned long) to contain n bits. The bit mask is zero-filled. The bitmask structure points to the bit mask and contains the n value.
struct bitmask *numa_bitmask_clearall(struct bitmask *bmp)

sets all bits in the bit mask to 0. The bitmask structure points to the bit mask and contains its size ( bmp ->size). The value of bmp is always returned. Note that struct bitmask *numa_bitmask_alloc(unsigned int n) creates a zero-filled bit mask.
struct bitmask *numa_bitmask_clearbit(struct bitmask *bmp, unsigned int n)

sets a specified bit in a bit mask to 0. Nothing is done if the n value is greater than the size of the bitmask (and no error is returned). The value of bmp is always returned.
int numa_bitmask_equal(const struct bitmask *bmp1, const struct bitmask *bmp2)

returns 1 if two bitmasks are equal. It returns 0 if they are not equal. If the bitmask structures control bit masks of different sizes, the “missing” trailing bits of the smaller bit mask are considered to be 0.
void numa_bitmask_free(struct bitmask *bmp)

deallocates the memory of both the bitmask structure pointed to by bmp and the bit mask. It is an error to attempt to free this bitmask twice.
int numa_bitmask_isbitset(const struct bitmask *bmp, unsigned int n)

returns the value of a specified bit in a bit mask. If the n value is greater than the size of the bit map, 0 is returned.
unsigned int numa_bitmask_nbytes(struct bitmask *bmp)

returns the size (in bytes) of the bit mask controlled by bmp. The bit masks are always full words (type unsigned long), and the returned size is the actual size of all those words.
struct bitmask *numa_bitmask_setall(struct bitmask *bmp)

sets all bits in the bit mask to 1. The bitmask structure points to the bit mask and contains its size ( bmp ->size). The value of bmp is always returned.
struct bitmask *numa_bitmask_setbit(struct bitmask *bmp, unsigned int n)

sets a specified bit in a bit mask to 1. Nothing is done if n is greater than the size of the bitmask (and no error is returned). The value of bmp is always returned.
void copy_bitmask_to_nodemask(struct bitmask *bmp, nodemask_t *nodemask)

copies the body (the bit map itself) of the bitmask structure pointed to by bmp to the nodemask_t structure pointed to by the nodemask pointer. If the two areas differ in size, the copy is truncated to the size of the receiving field or zero-filled.
void copy_nodemask_to_bitmask(nodemask_t *nodemask, struct bitmask *bmp)

copies the nodemask_t structure pointed to by the nodemask pointer to the body (the bit map itself) of the bitmask structure pointed to by the bmp pointer. If the two areas differ in size, the copy is truncated to the size of the receiving field or zero-filled.
void copy_bitmask_to_bitmask(struct bitmask *bmpfrom, struct bitmask *bmpto)

copies the body (the bit map itself) of the bitmask structure pointed to by the bmpfrom pointer to the body of the bitmask structure pointed to by the bmpto pointer. If the two areas differ in size, the copy is truncated to the size of the receiving field or zero-filled.
unsigned int numa_bitmask_weight(const struct bitmask *bmp )

returns a count of the bits that are set in the body of the bitmask pointed to by the bmp argument.
int numa_move_pages(int pid, unsigned long count, void **pages, const int *nodes, int *status, int flags)

moves a list of pages in the address space of the currently executing or current process. It simply uses the move_pages system call.
pid - ID of task. If not valid, use the current task.
count - Number of pages.
pages - List of pages to move.
nodes - List of nodes to which pages can be moved.
status - Field to which status is to be returned.
flags - MPOL_MF_MOVE or MPOL_MF_MOVE_ALL
See move_pages(2).
int numa_migrate_pages(int pid, struct bitmask *fromnodes, struct bitmask *tonodes)

simply uses the migrate_pages system call to cause the pages of the calling task, or a specified task, to be migated from one set of nodes to another. See migrate_pages(2). The bit masks representing the nodes should be allocated with struct bitmask *numa_allocate_nodemask() , or with struct bitmask *numa_bitmask_alloc(unsigned int n) using an n value returned from int int numa_num_possible_nodes(). A task’s current node set can be gotten by calling struct bitmask *numa_get_membind(void). Bits in the tonodes mask can be set by calls to struct bitmask *numa_bitmask_setbit(struct bitmask *bmp, unsigned int n).
void numa_error(char *where)

is a libnuma internal function that can be overridden by the user program. This function is called with a char * argument when a libnuma function fails. Overriding the library internal definition makes it possible to specify a different error handling strategy when a libnuma function fails. It does not affect int numa_available(void). The void numa_error(char *where) function defined in libnuma prints an error on stderr and terminates the program if extern int numa_exit_on_error is set to a non-zero value. The default value of extern int numa_exit_on_error is zero.
void numa_warn(int number, char *where, …)

is a libnuma internal function that can be also overridden by the user program. It is called to warn the user when a libnuma function encounters a non-fatal error. The default implementation prints a warning to stderr. The first argument is a unique number identifying each warning. After that there is a printf(3)-style format string and a variable number of arguments. numa_warn exits the program when extern int numa_exit_on_warn is set to a non-zero value. The default value of extern int numa_exit_on_warn is zero.