An fio plugin was added that can route I/O to the bdev layer. See the plugin documentation for more information.
spdk_bdev_unmap() was modified to take an offset and a length in bytes as arguments instead of requiring the user to provide an array of SCSI unmap descriptors. This limits unmaps to a single contiguous range.
spdk_bdev_write_zeroes() was introduced as an alternative to spdk_bdev_unmap(). It ensures that all unmapped blocks will be zeroed out. This function is currently only supported by NVMe block devices.
The AIO bdev now allows the user to override the auto-detected block size.
The NVMe driver now recognizes the NVMe 1.3 Namespace Optimal I/O Boundary field. NVMe 1.3 devices may report an optimal I/O boundary, which the driver will take into account when splitting I/O requests.
The HotplugEnable option in [Nvme] sections of the configuration file is now
"No" by default. It was previously "Yes".
The NVMe library now includes a function spdk_nvme_ns_get_ctrlr which returns the NVMe Controller associated with a given namespace.
The NVMe-oF target no longer requires any in capsule data buffers to run, and the feature is now entirely optional. Previously, at least 4KiB in capsule data buffers were required.
A new default value, SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, was added to provide additional clarity when constructing spdk_mempools. Previously, -1 could be passed and the library would choose a reasonable default, but this new value makes it explicit that the default is being used.
spdk_bs_io_readv_blob() and spdk_bs_io_writev_blob() were added to enable scattered payloads.
The ability to set a thread name, previously only used by the reactor code, is
now part of the spdk_thread_allocate() API. Users may specify a thread name
which will show up in tools like gdb.
A configure script has been added to simplify the build configuration process.
The existing CONFIG file and make CONFIG_... options are also still supported.
Run ./configure --help for information about available configuration options.
A DPDK submodule has been added to make building SPDK easier. If no --with-dpdk
option is specified to configure, the SPDK build system will automatically build a
known-good configuration of DPDK with the minimal options enabled. See the Building
section of README.md for more information.
A Vagrant setup has been added to make it easier to develop and use SPDK on systems without suitable NVMe hardware. See the Vagrant section of README.md for more information.
The vhost library and example app have been updated to support the vhost-blk protocol in addition to the existing vhost-scsi protocol. See the vhost documentation for more details.
A GPT virtual block device has been added, which automatically exposes GPT partitions with a special SPDK-specific partition type as bdevs. See the GPT bdev documentation for more information.
The NVMe driver has been updated to support recent Intel SSDs, including the Intel® Optane™ SSD DC P4800X series.
A workaround has been added for devices that failed to recognize register writes during controller reset.
The NVMe driver now allocates request tracking objects on a per-queue basis. The
number of requests allowed on an I/O queue may be set during spdk_nvme_probe() by
modifying io_queue_requests in the opts structure.
The SPDK NVMe fio_plugin has been updated to support multiple threads (numjobs).
spdk_nvme_ctrlr_alloc_io_qpair() has been modified to allow the user to override controller-level options for each individual I/O queue pair. Existing callers with qprio == 0 can be updated to:
... = spdk_nvme_ctrlr_alloc_io_qpair(ctrlr, NULL, 0);
Callers that need to specify a non-default qprio should be updated to:
struct spdk_nvme_io_qpair_opts opts;
spdk_nvme_ctrlr_get_default_io_qpair_opts(ctrlr, &opts, sizeof(opts));
opts.qprio = SPDK_NVME_QPRIO_...;
... = spdk_nvme_ctrlr_alloc_io_qpair(ctrlr, &opts, sizeof(opts));
The environment abstraction layer has been updated to include several new functions
in order to wrap additional DPDK functionality. See include/spdk/env.h for the
current set of functions.
Support for SPDK performance analysis has been added to Intel® VTune™ Amplifier 2018.
This analysis provides:
- I/O performance monitoring (calculating standard I/O metrics like IOPS, throughput, etc.)
- Tuning insights on the interplay of I/O and compute devices by estimating how many cores would be reasonable to provide for SPDK to keep up with a current storage workload.
See the VTune Amplifier documentation for more information.
The blobstore is a persistent, power-fail safe block allocator designed to be used as the local storage system backing a higher-level storage service. See the blobstore documentation for more details.
BlobFS adds basic filesystem functionality like filenames on top of the blobstore. This release also includes a RocksDB Env implementation using BlobFS in place of the kernel filesystem. See the BlobFS documentation for more details.
A userspace implementation of the QEMU vhost-scsi protocol has been added. The vhost target is capable of exporting SPDK bdevs to QEMU-based VMs as virtio devices. See the vhost documentation for more details.
The overhead of the main reactor event loop was reduced by optimizing the number of calls to spdk_get_ticks() per iteration.
The NVMe library will now automatically split readv/writev requests with scatter-gather lists that do not map to valid PRP lists when the NVMe controller does not natively support SGLs.
The identify and perf NVMe examples were modified to add a consistent format for
specifying remote NVMe over Fabrics devices via the -r option.
This is implemented using the new spdk_nvme_transport_id_parse() function.
The [Nvme] section of the configuration file was modified to remove the BDF directive
and replace it with a TransportID directive. Both local (PCIe) and remote (NVMe-oF)
devices can now be specified as the backing block device. A script to generate an
entire [Nvme] section based on the local NVMe devices attached was added at
scripts/gen_nvme.sh.
The [Nvme] section of the configuration file was modified to remove the BDF directive
and replace it with a TransportID directive. Both local (PCIe) and remote (NVMe-oF)
devices can now be specified as the backing block device. A script to generate an
entire [Nvme] section based on the local NVMe devices attached was added at
scripts/gen_nvme.sh.
The NVMe library has been changed to create its own request memory pool rather than
requiring the user to initialize the global request_mempool variable. Apps can be
updated by simply removing the initialization of request_mempool. Since the NVMe
library user no longer needs to know the size of the internal NVMe request
structure to create the pool, the spdk_nvme_request_size() function was also removed.
The spdk_nvme_ns_cmd_deallocate() function was renamed and extended to become
spdk_nvme_ns_cmd_dataset_management(), which allows access to all of the NVMe
Dataset Management command's parameters. Existing callers can be updated to use
spdk_nvme_ns_cmd_dataset_management() with SPDK_NVME_DSM_ATTR_DEALLOCATE as the
type parameter.
The NVMe library SGL callback prototype has been changed to return virtual addresses
rather than physical addresses. Callers of spdk_nvme_ns_cmd_readv() and
spdk_nvme_ns_cmd_writev() must update their next_sge_fn callbacks to match.
The NVMe library now supports NVMe over Fabrics devices in addition to the existing
support for local PCIe-attached NVMe devices. For an example of how to enable
NVMe over Fabrics support in an application, see examples/nvme/identify and
examples/nvme/perf.
Hot insert/remove support for NVMe devices has been added. To enable NVMe hotplug
support, an application should call the spdk_nvme_probe() function on a regular
basis to probe for new devices (reported via the existing probe_cb callback) and
removed devices (reported via a new remove_cb callback). Hotplug is currently
only supported on Linux with the uio_pci_generic driver, and newly-added NVMe
devices must be bound to uio_pci_generic by an external script or tool.
Multiple processes may now coordinate and use a single NVMe device simultaneously using DPDK Multi-process Support.
The nvmf_tgt configuration file format has been updated significantly to enable
new features. See the example configuration file etc/spdk/nvmf.conf.in for
more details on the new and changed options.
The NVMe over Fabrics target now supports virtual mode subsystems, which allow the
user to export devices from the SPDK block device abstraction layer as NVMe over
Fabrics subsystems. Direct mode (raw NVMe device access) is also still supported,
and a single nvmf_tgt may export both types of subsystems simultaneously.
The bdev layer now supports scatter/gather read and write I/O APIs, and the NVMe
blockdev driver has been updated to support scatter/gather. Apps can use the
new scatter/gather support via the spdk_bdev_readv() and spdk_bdev_writev()
functions.
The bdev status returned from each I/O has been extended to pass through NVMe or SCSI status codes directly in cases where the underlying device can provide a more specific status code.
A Ceph RBD (RADOS Block Device) blockdev driver has been added. This allows the
iscsi_tgt and nvmf_tgt apps to export Ceph RBD volumes as iSCSI LUNs or
NVMe namespaces.
libpciaccess has been removed as a dependency and DPDK PCI enumeration is
used instead. Prior to DPDK 16.07 enumeration by class code was not supported,
so for earlier DPDK versions, only Intel SSD DC P3x00 devices will be discovered
by the NVMe library.
The env environment abstraction library has been introduced, and a default
DPDK-based implementation is provided as part of SPDK. The goal of the env
layer is to enable use of alternate user-mode memory allocation and PCI access
libraries. See doc/porting.md for more details.
The build process has been modified to produce all of the library files in the
build/lib directory. This is intended to simplify the use of SPDK from external
projects, which can now link to SPDK libraries by adding the build/lib directory
to the library path via -L and linking the SPDK libraries by name (for example,
-lspdk_nvme -lspdk_log -lspdk_util).
nvmf_tgt and iscsi_tgt now have a JSON-RPC interface, which allows the user
to query and modify the configuration at runtime. The RPC service is disabled by
default, since it currently does not provide any authentication or security
mechanisms; it should only be enabled on systems with controlled user access
behind a firewall. An example RPC client implemented in Python is provided in
scripts/rpc.py.
This release adds a userspace iSCSI target. The iSCSI target is capable of exporting NVMe devices over a network using the iSCSI protocol. The application is located in app/iscsi_tgt and a documented configuration file can be found at etc/spdk/spdk.conf.in.
This release also significantly improves the existing NVMe over Fabrics target.
- The configuration file format was changed, which will require updates to
any existing nvmf.conf files (see
etc/spdk/nvmf.conf.in):SubsystemGroupwas renamed toSubsystem.AuthFilewas removed (it was unimplemented).nvmf_tgtwas updated to correctly recognize NQN (NVMe Qualified Names) when naming subsystems. The default node name was changed to reflect this; it is now "nqn.2016-06.io.spdk".PortandHostsections were merged into theSubsystemsection- Global options to control max queue depth, number of queues, max I/O size, and max in-capsule data size were added.
- The Nvme section was removed. Now a list of devices is specified by bus/device/function directly in the Subsystem section.
- Subsystems now have a Mode, which can be Direct or Virtual. This is an attempt to future-proof the interface, so the only mode supported by this release is "Direct".
- Many bug fixes and cleanups were applied to the
nvmf_tgtapp and library. - The target now supports discovery.
This release also adds one new feature and provides some better examples and tools for the NVMe driver.
- The Weighted Round Robin arbitration method is now supported. This allows
the user to specify different priorities on a per-I/O-queue basis. To
enable WRR, set the
arb_mechanismfield duringspdk_nvme_probe(). - A simplified "Hello World" example was added to show the proper way to use
the NVMe library API; see
examples/nvme/hello_world/hello_world.c. - A test for measuring software overhead was added. See
test/lib/nvme/overhead.
This release adds a userspace NVMf (NVMe over Fabrics) target, conforming to the newly-released NVMf 1.0/NVMe 1.2.1 specification. The NVMf target exports NVMe devices from a host machine over the network via RDMA. Currently, the target is limited to directly exporting physical NVMe devices, and the discovery subsystem is not supported.
This release includes a general API cleanup, including renaming all declarations
in public headers to include a spdk prefix to prevent namespace clashes with
user code.
- NVMe
- The
nvme_attach()API was reworked into a new probe/attach model, which moves device detection into the NVMe library. The new API also allows parallel initialization of NVMe controllers, providing a major reduction in startup time when using multiple controllers. - I/O queue allocation was changed to be explicit in the API. Each function
that generates I/O requests now takes a queue pair (
spdk_nvme_qpair *) argument, and I/O queues may be allocated usingspdk_nvme_ctrlr_alloc_io_qpair(). This allows more flexible assignment of queue pairs than the previous model, which only allowed a single queue per thread and limited the total number of I/O queues to the lowest number supported on any attached controller. - Added support for the Write Zeroes command.
examples/nvme/perfcan now report I/O command latency from the the controller's viewpoint using the Intel vendor-specific read/write latency log page.- Added namespace reservation command support, which can be used to coordinate sharing of a namespace between multiple hosts.
- Added hardware SGL support, which enables use of scattered buffers that don't conform to the PRP list alignment and length requirements on supported NVMe controllers.
- Added end-to-end data protection support, including the ability to write and
read metadata in extended LBA (metadata appended to each block of data in the
buffer) and separate metadata buffer modes.
See
spdk_nvme_ns_cmd_write_with_md()andspdk_nvme_ns_cmd_read_with_md()for details.
- The
- IOAT
- The DMA block fill feature is now exposed via the
ioat_submit_fill()function. This is functionally similar tomemset(), except the memory is filled with an 8-byte repeating pattern instead of a single byte like memset.
- The DMA block fill feature is now exposed via the
- PCI
- Added support for using DPDK for PCI device mapping in addition to the existing libpciaccess option. Using the DPDK PCI support also allows use of the Linux VFIO driver model, which means that SPDK userspace drivers will work with the IOMMU enabled. Additionally, SPDK applications may be run as an unprivileged user with access restricted to a specific set of PCIe devices.
- The PCI library API was made more generic to abstract away differences between the underlying PCI access implementations.
This release adds a user-space driver with support for the Intel I/O Acceleration Technology (I/OAT, also known as "Crystal Beach") DMA offload engine.
- IOAT
- New user-space driver supporting DMA memory copy offload
- Example programs
ioat/perfandioat/verify - Kernel-mode DMA engine test driver
kperffor performance comparison
- NVMe
- Per-I/O flags for Force Unit Access (FUA) and Limited Retry
- Public API for retrieving log pages
- Reservation register/acquire/release/report command support
- Scattered payload support - an alternate API to provide I/O buffers via a sequence of callbacks
- Declarations and
nvme/identifysupport for Intel SSD DC P3700 series vendor-specific log pages and features
- Updated to support DPDK 2.2.0
This is the initial open source release of the Storage Performance Development Kit (SPDK).
Features:
- NVMe user-space driver
- NVMe example programs
examples/nvme/perftests performance (IOPS) using the NVMe user-space driverexamples/nvme/identifydisplays NVMe controller information in a human-readable format
- Linux and FreeBSD support