The SPDK repository is, first and foremost, a collection of high-performance storage-centric software libraries. With this in mind, much care has been taken to ensure that these libraries have consistent and robust naming and versioning conventions. The libraries themselves are also divided across two directories (lib
and module
) inside of the SPDK repository in a deliberate way to prevent mixing of SPDK event framework dependent code and lower level libraries. This document is aimed at explaining the structure, naming conventions, versioning scheme, and use cases of the libraries contained in these two directories.
The SPDK libraries are divided into two directories. The lib
directory contains the base libraries that compose SPDK. Some of these base libraries define plug-in systems. Instances of those plug-ins are called modules and are located in the module
directory. For example, the spdk_sock
library is contained in the lib
directory while the implementations of socket abstractions, sock_posix
and sock_uring
are contained in the module
directory.
The libraries in the lib
directory can be readily divided into four categories:
spdk_log
contains macro definitions that provide a consistent logging paradigm and spdk_json
is a general purpose JSON parsing library.spdk_nvmf
and spdk_vhost
each define the storage protocols after which they are named.spdk_bdev
provides a general block device abstraction layer, spdk_lvol
provides a logical volume abstraction, spdk_blobfs
provides a filesystem abstraction, and spdk_ftl
provides a flash translation layer abstraction.spdk_jsonrpc
) and thread abstractions (see spdk_thread
). The most notable library in this category is the spdk_env_dpdk
library which provides a shim for the underlying Data Plane Development Kit (DPDK) environment and provides services like memory management.The one library in the lib
directory that doesn't fit into the above classification is the spdk_event
library. This library defines a framework used by the applications contained in the app
and example
directories. Much care has been taken to keep the SPDK libraries independent from this framework. The libraries in lib
are engineered to allow plugging directly into independent application frameworks such as Seastar or libuv with minimal effort.
Currently there are two exceptions in the lib
directory which still rely on spdk_event
, spdk_vhost
and spdk_iscsi
. There are efforts underway to remove all remaining dependencies these libraries have on the spdk_event
library.
Much like the spdk_event
library, the spdk_env_dpdk
library has been architected in such a way that it can be readily replaced by an alternate environment shim. More information on replacing the spdk_env_dpdk
module and the underlying dpdk
environment can be found in the environment section.
The component libraries in the module
directory represent specific implementations of the base libraries in the lib
directory. As with the lib
directory, much care has been taken to avoid dependencies on the spdk_event
framework except for those libraries which directly implement the spdk_event
module plugin system.
There are seven sub-directories in the module
directory which each hold a different class of libraries. These sub-directories can be divided into two types.
lib
directory and are registered with that library at runtime by way of a specific constructor function. The parent library in the lib
directory then manages the module directly. These types of libraries each implement a function table defined by their parent library. The following table shows these directories and their corresponding parent libraries:module directory | parent library | dependent on event library |
---|---|---|
module/accel | spdk_accel | no |
module/bdev | spdk_bdev | no |
module/event | spdk_event | yes |
module/sock | spdk_sock | no |
lib
directory but are not explicitly registered to that library via a constructor. The libraries in the blob
, blobfs
, and env_dpdk
directories fall into this category. None of the libraries in this category depend explicitly on the spdk_event
library.The SPDK libraries follow strict conventions for naming functions, logging, versioning, and header files.
All public SPDK header files exist in the include
directory of the SPDK repository. These headers are divided into two sub-directories.
include/spdk
contains headers intended to be used by consumers of the SPDK libraries. All of the functions, variables, and types in these functions are intended for public consumption. Multiple headers in this directory may depend upon the same underlying library and work together to expose different facets of the library. The spdk_bdev
library, for example, is exposed in three different headers. bdev_module.h
defines the interfaces a bdev module library would need to implement, bdev.h
contains general block device functions that would be used by an application consuming block devices exposed by SPDK, and bdev_zone.h
exposes zoned bdev specific functions. Many of the other libraries exhibit a similar behavior of splitting headers between consumers of the library and those wishing to register a module with that library.
include/spdk_internal
, as its name suggests contains header files intended to be consumed only by other libraries inside of the SPDK repository. These headers are typically used for sharing lower level functions between two libraries that both require similar functions. For example spdk_internal/nvme_tcp.h
contains low level tcp functions used by both the spdk_nvme
and spdk_nvmf
libraries. These headers are NOT intended for general consumption.
Other header files contained directly in the lib
and module
directories are intended to be consumed only by source files of their corresponding library. Any symbols intended to be used across libraries need to be included in a header in the include/spdk_internal
directory.
All public types and functions in SPDK libraries begin with the prefix spdk_
. They are also typically further namespaced using the spdk library name. The rest of the function or type name describes its purpose.
There are no internal library functions that begin with the spdk_
prefix. This naming convention is enforced by the SPDK continuous Integration testing. Functions not intended for use outside of their home library should be namespaced with the name of the library only.
SPDK libraries can be built as both static and shared object files. To facilitate building libraries as shared objects, each one has a corresponding map file (e.g. spdk_nvmf
relies on spdk_nvmf.map
). SPDK libraries not exporting any symbols rely on a blank map file located at mk/spdk_blank.map
.
SPDK shared objects follow a semantic versioning pattern with a major and minor version. Any changes which break backwards compatibility (symbol removal or change) will cause a shared object major increment and backwards compatible changes will cause a minor version increment; i.e. an application that relies on libspdk_nvmf.so.3.0
will be compatible with libspdk_nvmf.so.3.1
but not with libspdk_nvmf.so.4.0
.
Shared object versions are incremented only once between each release cycle. This means that at most, the major version of each SPDK shared library will increment only once between each SPDK release.
There are currently no guarantees in SPDK of ABI compatibility between two major SPDK releases.
The point releases of an LTS release will be ABI compatible with the corresponding LTS major release.
Shared objects are versioned independently of one another. This means that libspdk_nvme.so.3.0
and libspdk_bdev.so.3.0
do not necessarily belong to the same release. This also means that shared objects with the same suffix are not necessarily compatible with each other. It is important to source all of your SPDK libraries from the same repository and version to ensure inter-library compatibility.
Shared objects in SPDK are created on a per-library basis. There is a top level libspdk.so
object which is a linker script. It simply contains references to all of the other spdk shared objects.
There are essentially two ways of linking to SPDK libraries.
In the second instance, please note that applications need only link to the libraries upon which they directly depend. All SPDK libraries have their dependencies specified at object compile time. This means that when linking to spdk_net
, one does not also have to specify spdk_log
, spdk_util
, spdk_json
, spdk_jsonrpc
, and spdk_rpc
. However, this dependency inclusion does not extend to the application itself; i.e. if an application directly uses symbols from both spdk_bdev
and spdk_log
, both libraries will need to be supplied to the linker when linking the application even though spdk_log
is a dependency of spdk_bdev
.
Please also note that when linking to SPDK libraries, both the spdk_env shim library and the env library itself need to be supplied to the linker. In the examples above, these are spdk_env_dpdk
and dpdk
respectively. This was intentional and allows one to easily swap out both the environment and the environment shim.
SPDK depends on an environment abstraction that provides crucial pinned memory management and PCIe bus management operations. The interface for this environment abstraction is defined in the include/env.h
header file. The default implementation of this environment is located in spdk_env_dpdk
. This abstraction in turn relies upon the DPDK libraries. This two part implementation was deliberate and allows for easily swapping out the dpdk version upon which the spdk libraries rely without making modifications to the spdk source directly.
Any environment can replace the spdk_env_dpdk
environment by implementing the include/env.h
header file. The environment can either be implemented wholesale in a single library or as a two-part shim/implementation library system.
SPDK static objects are compiled by default even when no parameters are supplied to the build system. Unlike SPDK shared objects, the filename does not contain any versioning semantics. Linking against static objects is similar to shared objects but will always require the use of -Wl,--whole-archive
as argument. This is due to the use of constructor functions in SPDK such as those to register NVMe transports.
Due to the lack of versioning semantics, it is not recommended to install static libraries system wide. Instead the path to these static libraries should be added as argument at compile time using -L/path/to/static/libs
. The use of static objects instead of shared objects can also be forced through -Wl,-Bstatic
, otherwise some compilers might prefer to use the shared objects if both are available.