Scroll to navigation

LIBCXL(3) libcxl Manual LIBCXL(3)


libcxl - A library to interact with CXL devices through sysfs(5) and ioctl(2) interfaces


#include <cxl/libcxl.h>
cc ... -lcxl


libcxl provides interfaces to interact with CXL devices in Linux, using sysfs interfaces for most kernel interactions, and the ioctl() interface for command submission.

The starting point for all library interfaces is a cxl_ctx object, returned by linklibcxl:cxl_new[3]. CXL Type 3 memory devices and other CXL device objects are descendants of the cxl_ctx object, and can be iterated via an object an iterator API of the form cxl_<object>_foreach(<parent object>, <object iterator>).


The object representing a CXL memory expander (Type 3 device) is struct cxl_memdev. Library interfaces related to these devices have the prefix cxl_memdev_. These interfaces are mostly associated with sysfs interactions (unless otherwise noted in their respective documentation sections). They are typically used to retrieve data published by the kernel, or to send data or trigger kernel operations for a given device.

MEMDEV: Enumeration

struct cxl_memdev *cxl_memdev_get_first(struct cxl_ctx *ctx);
struct cxl_memdev *cxl_memdev_get_next(struct cxl_memdev *memdev);
struct cxl_ctx *cxl_memdev_get_ctx(struct cxl_memdev *memdev);
const char *cxl_memdev_get_host(struct cxl_memdev *memdev)
struct cxl_memdev *cxl_endpoint_get_memdev(struct cxl_endpoint *endpoint);
#define cxl_memdev_foreach(ctx, memdev) \

for (memdev = cxl_memdev_get_first(ctx); \
memdev != NULL; \
memdev = cxl_memdev_get_next(memdev))

CXL memdev instances are enumerated from the global library context struct cxl_ctx. By default a memdev only offers a portal to submit memory device commands, see the port, decoder, and endpoint APIs to determine what if any CXL Memory Resources are reachable given a specific memdev.

The host of a memdev is the PCIe Endpoint device that registered its CXL capabilities with the Linux CXL core.

MEMDEV: Attributes

int cxl_memdev_get_id(struct cxl_memdev *memdev);
unsigned long long cxl_memdev_get_serial(struct cxl_memdev *memdev);
const char *cxl_memdev_get_devname(struct cxl_memdev *memdev);
int cxl_memdev_get_major(struct cxl_memdev *memdev);
int cxl_memdev_get_minor(struct cxl_memdev *memdev);
unsigned long long cxl_memdev_get_pmem_size(struct cxl_memdev *memdev);
unsigned long long cxl_memdev_get_ram_size(struct cxl_memdev *memdev);
const char *cxl_memdev_get_firmware_verison(struct cxl_memdev *memdev);
size_t cxl_memdev_get_label_size(struct cxl_memdev *memdev);
int cxl_memdev_nvdimm_bridge_active(struct cxl_memdev *memdev);
int cxl_memdev_get_numa_node(struct cxl_memdev *memdev);

A memdev is given a kernel device name of the form "mem%d" where an id (cxl_memdev_get_id()) is dynamically allocated as devices are discovered. Note that there are no guarantees that ids / kernel device names for memdevs are stable from one boot to the next, devices are enumerated asynchronously. If a stable identifier is use cxl_memdev_get_serial() which returns a value according to the Device Serial Number Extended Capability in the PCIe 5.0 Base Specification.

The character device node for command submission can be found by default at /dev/cxl/mem%d, or created with a major / minor returned from cxl_memdev_get_{major,minor}().

The pmem_size and ram_size attributes return the current provisioning of DPA (Device Physical Address / local capacity) in the device.

cxl_memdev_get_numa_node() returns the affinitized CPU node number if available or -1 otherwise.

MEMDEV: Control

int cxl_memdev_disable_invalidate(struct cxl_memdev *memdev);
int cxl_memdev_enable(struct cxl_memdev *memdev);

When a memory device is disabled it unregisters its associated endpoints and potentially intervening switch ports if there are no other memdevs pinning that port active. That means that any existing port objects that the library has previously returned are in valid and need to be re-read. Callers must be careful to re-retrieve port objects after cxl_memdev_disable_invalidate(). Any usage of a previously obtained port object after a cxl_memdev_disable_invalidate() call is a use-after-free programming error. It follows that after cxl_memdev_enable() new ports may appear in the topology that were not previously enumerable.


cxl_memdev_disable_invalidate() will force disable the memdev
regardless of whether the memory provided by the device is in active use by the operating system. Callers take responisbility for assuring that it is safe to disable the memory device. Otherwise, this call can be as destructive as ripping a DIMM out of a running system. Like all other libcxl calls that mutate the system state or divulge security sensitive information this call requires root / CAP_SYS_ADMIN.

MEMDEV: Commands

struct cxl_cmd *cxl_cmd_new_raw(struct cxl_memdev *memdev, int opcode);
struct cxl_cmd *cxl_cmd_new_identify(struct cxl_memdev *memdev);
struct cxl_cmd *cxl_cmd_new_get_health_info(struct cxl_memdev *memdev);
struct cxl_cmd *cxl_cmd_new_read_label(struct cxl_memdev *memdev,

unsigned int offset, unsigned int length); struct cxl_cmd *cxl_cmd_new_write_label(struct cxl_memdev *memdev, void *buf,
unsigned int offset, unsigned int length); int cxl_memdev_zero_label(struct cxl_memdev *memdev, size_t length,
size_t offset); int cxl_memdev_read_label(struct cxl_memdev *memdev, void *buf, size_t length,
size_t offset); int cxl_memdev_write_label(struct cxl_memdev *memdev, void *buf, size_t length,
size_t offset); struct cxl_cmd *cxl_cmd_new_get_partition(struct cxl_memdev *memdev); struct cxl_cmd *cxl_cmd_new_set_partition(struct cxl_memdev *memdev,
unsigned long long volatile_size);

A cxl_cmd is a reference counted object which is used to perform Mailbox commands as described in the CXL Specification. A cxl_cmd object is tied to a cxl_memdev. Associated library interfaces have the prefix cxl_cmd_. Within this sub-class of interfaces, there are:

cxl_cmd_new_*() interfaces that allocate a new cxl_cmd object for a given command type targeted at a given memdev. As part of the command instantiation process the library validates that the command is supported by the memory device, otherwise it returns NULL to indicate no support. The libcxl command id is translated by the kernel into a CXL standard opcode. See the potential command ids in /usr/include/linux/cxl_mem.h.

cxl_cmd_<name>_set_<field> interfaces that set specific fields in a cxl_cmd

cxl_cmd_submit which submits the command via ioctl()

cxl_cmd_<name>_get_<field> interfaces that get specific fields out of the command response

cxl_cmd_get_* interfaces to get general command related information.

cxl_cmd_new_raw() supports so called RAW commands where the command id is RAW and it carries an unmodified CXL memory device command payload associated with the opcode argument. Given the kernel does minimal input validation on these commands typically raw commands are not supported by the kernel outside debug build scenarios. libcxl is limited to supporting commands that appear in the CXL standard / public specifications.

cxl_memdev{read,write,zero}_label() are helpers for marshaling multiple label access commands over an arbitrary extent of the device’s label area.

cxl_cmd_partition_set_mode() supports selecting NEXTBOOT or IMMEDIATE mode. When CXL_SETPART_IMMEDIATE mode is set, it is the caller’s responsibility to avoid immediate changes to partitioning when the device is in use. When CXL_SETPART_NEXTBOOT mode is set, the change in partitioning shall become the “next” configuration, to become active on the next device reset.


The CXL Memory space is CPU and Device coherent. The address ranges that support coherent access are described by platform firmware and communicated to the operating system via a CXL root object struct cxl_bus.

BUS: Enumeration

struct cxl_bus *cxl_bus_get_first(struct cxl_ctx *ctx);
struct cxl_bus *cxl_bus_get_next(struct cxl_bus *bus);
struct cxl_ctx *cxl_bus_get_ctx(struct cxl_bus *bus);
struct cxl_bus *cxl_memdev_get_bus(struct cxl_memdev *memdev);
struct cxl_bus *cxl_port_get_bus(struct cxl_port *port);
struct cxl_bus *cxl_endpoint_get_bus(struct cxl_endpoint *endpoint);
#define cxl_bus_foreach(ctx, bus)                                           \

for (bus = cxl_bus_get_first(ctx); bus != NULL; \
bus = cxl_bus_get_next(bus))

When a memdev is active it has established a CXL port hierarchy between itself and the root of its associated CXL topology. The cxl_{memdev,endpoint}_get_bus() helpers walk that topology to retrieve the associated bus object.

BUS: Attributes

const char *cxl_bus_get_provider(struct cxl_bus *bus);
const char *cxl_bus_get_devname(struct cxl_bus *bus);
int cxl_bus_get_id(struct cxl_bus *bus);

The provider name of a bus is a persistent name that is independent of discovery order. The possible provider names are ACPI.CXL and cxl_test. The devname and id attributes, like other objects, are just the kernel device names that are subject to change based on discovery order.

BUS: Control

int cxl_bus_disable_invalidate(struct cxl_bus *bus);

An entire CXL topology can be torn down with this API. Like other _invalidate APIs callers must assume that all library objects have been freed. This one goes one step further and also frees the @bus argument. This may crash the system and is only useful in kernel driver development scenarios.


CXL ports track the PCIe hierarchy between a platform firmware CXL root object, through CXL / PCIe Host Bridges, CXL / PCIe Root Ports, and CXL / PCIe Switch Ports.

PORT: Enumeration

struct cxl_port *cxl_bus_get_port(struct cxl_bus *bus);
struct cxl_port *cxl_port_get_first(struct cxl_port *parent);
struct cxl_port *cxl_port_get_next(struct cxl_port *port);
struct cxl_port *cxl_port_get_parent(struct cxl_port *port);
struct cxl_ctx *cxl_port_get_ctx(struct cxl_port *port);
const char *cxl_port_get_host(struct cxl_port *port);
struct cxl_port *cxl_decoder_get_port(struct cxl_decoder *decoder);
struct cxl_port *cxl_port_get_next_all(struct cxl_port *port,

const struct cxl_port *top); struct cxl_port *cxl_dport_get_port(struct cxl_dport *dport); #define cxl_port_foreach(parent, port) \
for (port = cxl_port_get_first(parent); port != NULL; \
port = cxl_port_get_next(port)) #define cxl_port_foreach_all(top, port) \
for (port = cxl_port_get_first(top); port != NULL; \
port = cxl_port_get_next_all(port, top))

A bus object encapsulates a CXL port object. Use cxl_bus_get_port() to use generic port APIs on root objects.

Ports are hierarchical. All but the a root object have another CXL port as a parent object retrievable via cxl_port_get_parent().

The root port of a hiearchy can be retrieved via any port instance in that hierarchy via cxl_port_get_bus().

The host of a port is the corresponding device name of the PCIe Root Port, or Switch Upstream Port with CXL capabilities.

The cxl_port_foreach_all() helper does a depth first iteration of all ports beneath the top port argument.

PORT: Control

int cxl_port_disable_invalidate(struct cxl_port *port); int cxl_port_enable(struct cxl_port *port); --- cxl_port_disable_invalidate() is a violent operation that disables entire sub-tree of CXL Memory Device and Ports, only use it for test / debug scenarios, or ensuring that all impacted devices are deactivated first.

PORT: Attributes

const char *cxl_port_get_devname(struct cxl_port *port);
int cxl_port_get_id(struct cxl_port *port);
int cxl_port_is_enabled(struct cxl_port *port);
bool cxl_port_is_root(struct cxl_port *port);
bool cxl_port_is_switch(struct cxl_port *port);
bool cxl_port_is_endpoint(struct cxl_port *port);
int cxl_port_get_depth(struct cxl_port *port);
bool cxl_port_hosts_memdev(struct cxl_port *port, struct cxl_memdev *memdev);
int cxl_port_get_nr_dports(struct cxl_port *port);

The port type is communicated via cxl_port_is_<type>(). An enabled port is one that has succeeded in discovering the CXL component registers in the host device and has enumerated its downstream ports. In order for a memdev to be enabled for CXL memory operation all CXL ports in its ancestry must also be enabled including a root port, an arbitrary number of intervening switch ports, and a terminal endpoint port.

cxl_port_hosts_memdev() returns true if the port’s host appears in the memdev host’s device topology ancestry.


A CXL dport object represents a CXL / PCIe Switch Downstream Port, or a CXL / PCIe host bridge.

DPORT: Enumeration

struct cxl_dport *cxl_dport_get_first(struct cxl_port *port);
struct cxl_dport *cxl_dport_get_next(struct cxl_dport *dport);
struct cxl_dport *cxl_port_get_dport_by_memdev(struct cxl_port *port,

struct cxl_memdev *memdev); #define cxl_dport_foreach(port, dport) \
for (dport = cxl_dport_get_first(port); dport != NULL; \
dport = cxl_dport_get_next(dport))

DPORT: Attributes

const char *cxl_dport_get_devname(struct cxl_dport *dport);
const char *cxl_dport_get_physical_node(struct cxl_dport *dport);
int cxl_dport_get_id(struct cxl_dport *dport);
bool cxl_dport_maps_memdev(struct cxl_dport *dport, struct cxl_memdev *memdev);

The id of a dport is the hardware idenfifier used by an upstream port to reference a downstream port. The physical node of a dport is only available for platform firmware defined downstream ports and alias the companion object, like a PCI host bridge, in the PCI device hierarchy.

The cxl_dport_maps_memdev() helper checks if a dport is an ancestor of a given memdev.


CXL endpoint objects encapsulate the set of host-managed device-memory (HDM) decoders in a physical memory device. The endpoint is the last hop in a decoder chain that translate SPA to DPA (system-physical-address to device-local-physical-address).

ENDPOINT: Enumeration

struct cxl_endpoint *cxl_endpoint_get_first(struct cxl_port *parent);
struct cxl_endpoint *cxl_endpoint_get_next(struct cxl_endpoint *endpoint);
struct cxl_ctx *cxl_endpoint_get_ctx(struct cxl_endpoint *endpoint);
struct cxl_port *cxl_endpoint_get_parent(struct cxl_endpoint *endpoint);
struct cxl_port *cxl_endpoint_get_port(struct cxl_endpoint *endpoint);
const char *cxl_endpoint_get_host(struct cxl_endpoint *endpoint);
struct cxl_endpoint *cxl_memdev_get_endpoint(struct cxl_memdev *memdev);
struct cxl_endpoint *cxl_port_to_endpoint(struct cxl_port *port);
#define cxl_endpoint_foreach(port, endpoint)                                 \

for (endpoint = cxl_endpoint_get_first(port); endpoint != NULL; \
endpoint = cxl_endpoint_get_next(endpoint))

ENDPOINT: Attributes

const char *cxl_endpoint_get_devname(struct cxl_endpoint *endpoint);
int cxl_endpoint_get_id(struct cxl_endpoint *endpoint);
int cxl_endpoint_is_enabled(struct cxl_endpoint *endpoint);


Decoder objects are associated with the "HDM Decoder Capability" published in Port devices and CXL capable PCIe endpoints. The kernel additionally models platform firmware described CXL memory ranges (like the ACPI CEDT.CFMWS) as static decoder objects. They route System Physical Addresses through a port topology to an endpoint decoder that does the final translation from SPA to DPA (system-physical-address to device-local-physical-address).

DECODER: Enumeration

struct cxl_decoder *cxl_decoder_get_first(struct cxl_port *port);
struct cxl_decoder *cxl_decoder_get_next(struct cxl_decoder *decoder);
struct cxl_ctx *cxl_decoder_get_ctx(struct cxl_decoder *decoder);
struct cxl_decoder *cxl_target_get_decoder(struct cxl_target *target);
#define cxl_decoder_foreach(port, decoder)                                  \

for (decoder = cxl_decoder_get_first(port); decoder != NULL; \
decoder = cxl_decoder_get_next(decoder))

The definition of a CXL port in libcxl is an object that hosts one or more CXL decoder objects.

DECODER: Attributes

unsigned long long cxl_decoder_get_resource(struct cxl_decoder *decoder);
unsigned long long cxl_decoder_get_size(struct cxl_decoder *decoder);
unsigned long long cxl_decoder_get_dpa_resource(struct cxl_decoder *decoder);
unsigned long long cxl_decoder_get_dpa_size(struct cxl_decoder *decoder);
int cxl_decoder_set_dpa_size(struct cxl_decoder *decoder, unsigned long long size);
const char *cxl_decoder_get_devname(struct cxl_decoder *decoder);
int cxl_decoder_get_id(struct cxl_decoder *decoder);
int cxl_decoder_get_nr_targets(struct cxl_decoder *decoder);
struct cxl_region *cxl_decoder_get_region(struct cxl_decoder *decoder);
enum cxl_decoder_target_type {

CXL_DECODER_TTYPE_ACCELERATOR, }; cxl_decoder_get_target_type(struct cxl_decoder *decoder); enum cxl_decoder_mode {
CXL_DECODER_MODE_RAM, }; enum cxl_decoder_mode cxl_decoder_get_mode(struct cxl_decoder *decoder); int cxl_decoder_set_mode(struct cxl_decoder *decoder, enum cxl_decoder_mode mode); bool cxl_decoder_is_pmem_capable(struct cxl_decoder *decoder); bool cxl_decoder_is_volatile_capable(struct cxl_decoder *decoder); bool cxl_decoder_is_mem_capable(struct cxl_decoder *decoder); bool cxl_decoder_is_accelmem_capable(struct cxl_decoder *decoder); bool cxl_decoder_is_locked(struct cxl_decoder *decoder);

The kernel protects the enumeration of the physical address layout of the system. Without CAP_SYS_ADMIN cxl_decoder_get_resource() returns ULLONG_MAX to indicate that the address information was not retrievable. Otherwise, cxl_decoder_get_resource() returns the currently programmed value of the base of the decoder’s decode range. A zero-sized decoder indicates a disabled decoder.

Root level decoders only support limited set of memory types in their address range. The cxl_decoder_is_<memtype>_capable() helpers identify what is supported. Switch level decoders, in contrast are capable of routing any memory type, i.e. they just forward along the memory type support from their parent port. Endpoint decoders follow the capabilities of their host memory device.

The capabilities of a decoder are not to be confused with their type / mode. The type ultimately depends on the endpoint. For example an accelerator requires all decoders in its ancestry to be set to CXL_DECODER_TTYPE_ACCELERATOR, and conversely plain memory expander devices require CXL_DECODER_TTYPE_EXPANDER.

Platform firmware may setup the CXL decode hierarchy before the OS boots, and may additionally require that the OS not change the decode settings. This property is indicated by the cxl_decoder_is_locked() API.

When a decoder is associated with a region cxl_decoder_get_region() returns that region object. Note that it is only applicable to switch and endpoint decoders as root decoders have a 1:N relationship with regions. Use cxl_region_foreach() for the similar functionality for root decoders.


A root or switch level decoder takes an SPA (system-physical-address) as input and routes it to a downstream port. Which downstream port depends on the downstream port’s position in the interleave. A struct cxl_target object represents the properties of a given downstream port relative to its interleave configuration.

TARGET: Enumeration

struct cxl_target *cxl_decoder_get_target_by_memdev(struct cxl_decoder *decoder,

struct cxl_memdev *memdev); struct cxl_target * cxl_decoder_get_target_by_position(struct cxl_decoder *decoder, int position); struct cxl_target *cxl_target_get_first(struct cxl_decoder *decoder); struct cxl_target *cxl_target_get_next(struct cxl_target *target); #define cxl_target_foreach(decoder, target) \
for (target = cxl_target_get_first(decoder); target != NULL; \
target = cxl_target_get_next(target))

Target objects can only be enumerated if the decoder has been configured, for switch decoders. For root decoders they are always available since the root decoder target mapping is static. The cxl_decoder_get_target_by_memdev() helper walks the topology to validate if the given memory device is capable of receiving cycles from this upstream decoder. It does not validate if the memory device is currently configured to participate in that decode.

TARGET: Attributes

int cxl_target_get_position(struct cxl_target *target);
unsigned long cxl_target_get_id(struct cxl_target *target);
const char *cxl_target_get_devname(struct cxl_target *target);
bool cxl_target_maps_memdev(struct cxl_target *target,

struct cxl_memdev *memdev); const char *cxl_target_get_physical_node(struct cxl_target *target);

The position of a decoder along with the interleave granularity dictate which address in the decoder’s resource range map to which port.

The target id is an identifier that the CXL port uses to reference this downstream port. For CXL / PCIe downstream switch ports the id is defined by the PCIe Link Capability Port Number field. For root decoders the id is specified by platform firmware specific mechanism. For ACPI.CXL defined root ports the id comes from the CEDT.CHBS / ACPI0016 _UID.

The device name of a target is the name of the host device for the downstream port. For CXL / PCIe downstream ports the devname is downstream switch port PCI device. For CXL root ports the devname is a platform firmware object for the host bridge like a ACPI0016 device instance.

The cxl_target_maps_memdev() helper is the companion of cxl_decoder_get_target_by_memdev() to determine which downstream ports / targets are capable of mapping which memdevs.

Some platform firmware implementations define an alias / companion device to represent the root of a PCI device hierarchy. The cxl_target_get_physical_node() helper returns the device name of that companion object in the PCI hierarchy.


A CXL region is composed of one or more slices of CXL memdevs, with configurable interleave settings - both the number of interleave ways, and the interleave granularity. In terms of hierarchy, it is the child of a CXL root decoder. A root decoder (recall that this corresponds to an ACPI CEDT.CFMWS window), may have multiple child regions, but a region is strictly tied to one root decoder.

The slices that compose a region are called mappings. A mapping is a tuple of memdev, endpoint decoder, and the position.

REGION: Enumeration

struct cxl_region *cxl_region_get_first(struct cxl_decoder *decoder);
struct cxl_region *cxl_region_get_next(struct cxl_region *region);
#define cxl_region_foreach(decoder, region)                                    \

for (region = cxl_region_get_first(decoder); region != NULL; \
region = cxl_region_get_next(region)) #define cxl_region_foreach_safe(decoder, region, _region) \
for (region = cxl_region_get_first(decoder), \
_region = region ? cxl_region_get_next(region) : NULL; \
region != NULL; \
region = _region, \
_region = _region ? cxl_region_get_next(_region) : NULL)

REGION: Attributes

int cxl_region_get_id(struct cxl_region *region);
const char *cxl_region_get_devname(struct cxl_region *region);
void cxl_region_get_uuid(struct cxl_region *region, uuid_t uu);
unsigned long long cxl_region_get_size(struct cxl_region *region);
unsigned long long cxl_region_get_resource(struct cxl_region *region);
unsigned int cxl_region_get_interleave_ways(struct cxl_region *region);
unsigned int cxl_region_get_interleave_granularity(struct cxl_region *region);
struct cxl_decoder *cxl_region_get_target_decoder(struct cxl_region *region,

int position); int cxl_region_set_size(struct cxl_region *region, unsigned long long size); int cxl_region_set_uuid(struct cxl_region *region, uuid_t uu); int cxl_region_set_interleave_ways(struct cxl_region *region,
unsigned int ways); int cxl_region_set_interleave_granularity(struct cxl_region *region,
unsigned int granularity); int cxl_region_set_target(struct cxl_region *region, int position,
struct cxl_decoder *decoder); int cxl_region_clear_target(struct cxl_region *region, int position); int cxl_region_clear_all_targets(struct cxl_region *region); int cxl_region_decode_commit(struct cxl_region *region); int cxl_region_decode_reset(struct cxl_region *region);

A region’s resource attribute is the Host Physical Address at which the region’s address space starts. The region’s address space is a subset of the parent root decoder’s address space.

The interleave ways is the number of component memdevs participating in the region.

The interleave granularity depends on the root decoder’s granularity, and must follow the interleave math rules defined in the CXL spec.

Regions have a list of targets 0..N, which are programmed with the name of an endpoint decoder under each participating memdev.

The decode_commit and decode_reset attributes reserve and free DPA space on a given memdev by allocating an endpoint decoder, and programming it based on the region’s interleave geometry.


Copyright © 2016 - 2022, Intel Corporation. License GPLv2: GNU GPL version 2 <>. This is free software: you are free to change and redistribute it. There is NO WARRANTY, to the extent permitted by law.



2022-08-24 libcxl