PMDK Changelog

The master branch changelog in text format for the current development version can be found at https://github.com/pmem/pmdk/blob/master/ChangeLog

Conventions

This changelog uses the following conventions

  • Each release has a publication date along with the name of the publisher

  • Changes are prefixed with a tag denoting the area, if any, they relate to. For example: 'doc' (documentation), 'obj' (libpmemobj), 'pool' (libpmempool), etc

  • Bugs and new features may link to PMDK issue repository using '(#nnn)' or external bug repositories such as the Red Hat Bugzilla (RHBZ).

Version 1.7.0

This release:

  • Introduces new APIs in libpmemobj for managing space used by transactions.

    (see pmemobj_tx_log_append_buffer man page for details)

  • Introduces new APIs in librpmem, splitting rpmem_persist into rpmem_flush

    and rpmem_drain, allowing applications to use the flush + drain model

    already known from libpmem. (libpmemobj does not use this feature yet)

  • Optimizes large libpmemobj transactions by significantly reducing

    the amount of memory modified at the commit phase.

  • Optimizes tracking of libpmemobj reservations.

  • Adds new flags for libpmemobj's pmemobj_tx_xadd_range[_direct] API: POBJ_XADD_NO_SNAPSHOT and POBJ_XADD_ASSUME_INITIALIZED, allowing applications to optimize how memory is tracked by the library.

To support some of the above changes the libpmemobj on-media layout had to be changed, which means that old pools have to be converted using pmdk-convert >= 1.7.

Other changes:

  • obj: fix merging of ranges when NOFLUSH flag is used (#1100)

  • rpmem: fix closing of ssh connection (#995, #1060)

  • obj: abort transaction on pmemobj_tx_publish failure

Internal changes:

  • test: fault injection tests for pmemblk, pmemlog, and pmemobj

  • test: improved Python testing framework

  • test: support real pmem in bad blocks tests

  • common: allow not building examples and benchmarks

Version 1.6.0

Released this on Mar 26 · 304 commits to master since this release

This release:

  • Enables unsafe shutdown and bad block detection on Linux on systems with libndctl >= 63. It is expected that systems with libndctl >= 63 has necessary kernel support (Linux >= 4.20). However, due to bugs in libndctl = 63 and Linux = 4.20, it is recommended to use libndctl >= 64.1 and Linux >= 5.0.4. On systems with libndctl < 63, PMDK uses old superuser-only interfaces. Support for old or new interfaces is chosen at BUILD time.

  • Introduces arena control interface in pmemobj, allowing applications to tweak performance and scalability of heap operations. See pmemobj_ctl_get man page ("heap" namespace) for details.

  • Introduces copy_on_write mode, which allows testing applications using pmemobj with pmreorder. See pmemobj_ctl_get man page ("copy_on_write" namespace) for details.

Other changes:

  • allocate file space when creating a pool on existing file (pmem/issues#167)

  • initial support for testing using fault injection

  • initial Python test framework

  • improve performance of pmemobj_pool_by_ptr

Bug fixes:

  • common: work around tmpfs bug during pool creation (pmem/issues#1018)

  • pool: race-free pmempool create --max-size

  • obj: don't modify remote pools in pmemobj_check

Version 1.5.1

Tue Feb 19 2018 Marcin Ślusarz <marcin.slusarz@intel.com>

This release fixes minor bugs and improves compatibility with newer tool chains.

Notable bug fixes:

  • common: make detection of device-dax instances more robust

  • obj: fix pmemobj_check for pools with some sizes

  • obj: don't use anon struct in an union (public header)

  • obj: fix pmemobj_tx_lock error handling

  • obj: don't use braces in an expression with clang (public header)

  • obj: suppress pmemcheck warnings for statistics

  • pmreorder: fix markers nontype issue

Version 1.5

Fri Oct 26 2018 Marcin Ślusarz <marcin.slusarz@intel.com>

This release has had two major focus areas - performance and RAS (Reliability, Availability and Serviceability). Beyond that, it introduces new APIs, new tools and many other improvements. As a side effect of performance optimizations, the libpmemobj on-media layout had to be changed, which means that old pools have to be converted using pmdk-convert. libpmemcto experiment has been finished and removed from the tree.

For more details, please see http://pmem.io/2018/10/22/release-1-5.html.

New features:

  • common: unsafe shutdown detection (SDS)

  • common: detection and repair of uncorrectable memory errors (bad blocks)

  • pool: new "feature" subcommand for enabling and disabling detection of

    unsafe shutdown and uncorrectable memory errors

  • common: auto flush detection on Windows (on Linux since 1.4)

  • pmreorder: new tool for verification of persistent memory algorithms

  • obj: new on media layout

  • pmem/obj: new flexible memcpy|memmove|memset API

  • obj: new flushing APIs: pmemobj_xpersist, pmemobj_xflush

    (PMEMOBJ_F_RELAXED)

  • rpmem: new flag RPMEM_PERSIST_RELAXED for rpmem_persist

  • obj: lazily initialized volatile variables (pmemobj_volatile)

    (EXPERIMENTAL)

  • obj: allocation classes with alignment

  • obj: new action APIs: pmemobj_defer_free, POBJ_XRESERVE_NEW,

    POBJ_XRESERVE_ALLOC

  • blk/log: new "ctl" API

Optimizations:

  • obj: major performance improvements for AEP NVDIMMs

  • obj: better space utilization for small allocations

  • common: call msync only on one page for deep drain

Other changes:

Bug fixes:

  • obj: fix type numbers for pmemobj_list_insert_new

  • pmem: fix inconsistency in pmem_is_pmem

  • common: fix windows mmap destruction

  • daxio: fix checking and adjusting length

  • common: fix long paths support on Windows

Version 1.4

Thu Mar 29 2018 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

This is the first release of PMDK under a new name. The NVML project has been renamed to PMDK (Persistent Memory Development Kit). This is only the project/repo name change and it does not affect the names of the PMDK packages. See this blog article for more details on the reasons and impact of the name change: http://pmem.io/2017/12/11/NVML-is-now-PMDK.html

New features

  • common: support for concatenated Device-DAX devices

    with 2M/1G alignment

  • common: add support for MAP_SYNC flag

  • common: always enable Valgrind instrumentation (#292)

  • common: pool set options / headerless pools

  • pmem: add support for "deep flush" operation

  • rpmem: add rpmem_deep_persist

  • doc: split man pages and add per-function aliases (#385)

Optimizations

  • pmem: skip CPU cache flushing when eADR is available

    (no Windows support yet)

  • pmem: add AVX512F support in pmem_memcpy/memset (#656)

Bug fixes

  • common: fix library dependencies (#767, RHBZ #1539564)

  • common: use rpm-config CFLAGS/LDFLAGS when building packages

    (#768, RHBZ #1539564)

  • common: do not unload librpmem on close (#776)

  • common: fix NULL check in os_fopen (#813)

  • common: fix missing version in .pc files

  • obj: fix cancel of huge allocations (#726)

  • obj: fix error handling in pmemobj_open (#750)

  • obj: validate pe_offset in pmemobj_list_* APIs (#772)

  • obj: fix add_range with size == 0 (#781)

  • log: add check for negative iovcnt (#690)

  • rpmem: limit maximum number of lanes (#609)

  • rpmem: change order of memory registration (#655)

  • rpmem: fix removing remote pools (#721)

  • pool: fix error handling (#643)

  • pool: fix sync with switched parts (#730)

  • pool: fix sync with missing replica (#731)

  • pool: fix detection of Device DAX size (#805)

  • pool: fail pmempool_sync if there are no replicas (#816)

  • benchmark: fix calculating standard deviation (#318)

  • doc: clarify pmem_is_pmem behavior (#719)

  • doc: clarify pmemobj_root behavior (#733)

Experimental features

  • common: port PMDK to FreeBSD

  • common: add experimental support for aarch64

  • obj: introduce allocation classes

  • obj: introduce two-phase heap ops (reserve/publish) (#380, #415)

  • obj: provide basic heap statistics (#676)

  • obj: implement run-time pool extending (#382)

  • cto: add close-to-open persistence library (#192)

Disabled Features

The following features are disabled by default, until ndctl v60.0 is available:

  • daxio: add utility to perform I/O on Device-DAX

  • RAS: unsafe shutdown detection/handling

Version 1.3.1

Wed Dec 20 2017 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

Bug fixes

  • rpmem: fix issues reported by Coverity

  • rpmem: fix read error handling

  • rpmem: add fip monitor (#597)

  • test: add rpmemd termination handling test

  • cpp: fix pop.persist function in obj_cpp_ptr

  • rpmem: return failure for a failed allocation

  • rpmem: fix potential memory leak

  • common: fix available rm options msg (#651)

  • pool: fix pmempool_get_max_size

  • obj: fix potential deadlock during realloc (#635, #636, #637)

  • obj: initialize TLS data

  • rpmem: fix cleanup if fork() failed (#634)

  • obj: fix bogus OOM after exhausting first zone

Version 1.3

Thu Jul 13 2017 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

This release introduces some useful features and optimizations in libpmemobj. Most of them are experimental and controlled by the new pmemobj_ctl APIs. For details, please check the feature requests identified by the issue numbers listed next to the items below.

Other important changes are related to performance tuning and stabilization of librpmem library, which is used by libpmemobj to get remote access to persistent memory and to provide basic data replication over RDMA. The librpmem is still considered experimental.

NVML for Windows is feature complete (except for libvmmalloc). This release includes the support for Unicode, long paths and the NVML installer.

New features

  • common: add support for concatenated DAX Devices

  • common: add Unicode support on Windows

  • common: add long path support on Windows

  • common: add NVML installer for Windows

  • pmem: make pmem_is_pmem() true for Device DAX only

  • obj: add pmemobj_wcsdup()/pmemobj_tx_wcsdup() APIs

  • obj: export non-inlined pmemobj_direct()

  • obj: add PMEMOBJ_NLANES env variable

  • cpp: introduce the allocator

  • cpp: add wstring version of C++ entry points

  • vmem: add vmem_wcsdup() API entry

  • pool: add pmempool_rm() function (#307)

  • pool: add --force flag for create command (#529)

  • benchmark: add a minimal execution time option

  • benchmark: add thread affinity option

  • benchmark: print 99% and 99.9% percentiles

  • doc: separate Linux/Windows version of web-based man pages

Optimizations

  • obj: cache _pobj_cached_pool in pmemobj_direct()

  • obj: optimize thread utilization of buckets

  • obj: stop grabbing a lock when querying pool ptr

  • rpmem: use multiple endpoints

Bug fixes

  • common: fix issues reported by static code analyzers

  • pmem: fix mmap() implementation on Windows

  • pmem: fix mapping addr/length alignment on Windows

  • pmem: fix PMEM_MMAP_HINT implementation on Windows

  • pmem: fix pmem_is_pmem() on invalid memory ranges

  • pmem: fix wrong is_pmem returned by pmem_map_file()

  • pmem: fix mprotect() for private mappings on Windows

  • pmem: modify pmem_is_pmem() behavior for len==0

  • obj: add failsafe to prevent allocs in constructor

  • cpp: fix swap implementation

  • cpp: fix sync primitives' constructors

  • cpp: fix wrong pointer type in the allocator

  • cpp: return persistent_ptr::swap to being public

  • pool: treat invalid answer as 'n'

  • pool: unify flags value for dry run

  • pool: transform for remote replicas

  • rpmem: persistency method detection

  • benchmark: fix time measurement

Experimental features & optimizations

  • obj: pmemobjctl - statistics and control submodule (#194, #211)

  • obj: zero-overhead allocations - customizable alloc header (#347)

  • obj: flexible run size index (#377)

  • obj: dynamic range cache (#378)

  • obj: asynchronous post-commit (#381)

  • obj: configurable object cache (#515)

  • obj: add cache size and threshold tx params

  • obj: add CTL var for suppressing expensive checks

  • rpmem: add rpmem_set_attr() API entry

  • rpmem: switch to libfabric v1.4.2

Version 1.2.3

Thu May 18 2017 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

Bug fixes

  • test: extend timeout for selected tests

  • test: reduce number of operations in obj_tx_mt

  • test: define cfree() as free() in vmmalloc_calloc

Other changes

  • common: move Docker images to new repo

Version 1.2.2

Sat Apr 15 2017 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

Bug fixes

  • pmempool: fix mapping type in pool_params_parse

  • test: limit number of arenas in vmem_stats

  • test: do not run pool_lock test as root

  • common: fix pkg-config files

  • common: fix building packages for Debian

Version 1.2.1

Tue Feb 21 2017 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

This NVML release changes the behavior of pmem_is_pmem() on Linux. The pmem_is_pmem() function will now return true only if the entire range is mapped directly from Device DAX (/dev/daxX.Y) without an intervening file system, and only if the corresponding file mapping was created with pmem_map_file(). See libpmem(3) for details.

Bug fixes

  • jemalloc: fix test compilation on Fedora 26 (rawhide)

  • test: fix cpp test compilation on Fedora 26 (rawhide)

  • common: use same queue.h on linux and windows

  • common: queue.h clang static analyzer fix

  • common: fix path handling in build-dpkg.sh

  • test: fix match files in pmempool_transform/TEST8

Version 1.2 - Windows Technical Preview #1

Fri Dec 30 2016 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

This is the first Technical Preview release of NVML for Windows. It is based on NVML 1.2 version, but not all the 1.2 features are ported to Windows. In particular, Device DAX and remote access to persistent memory (librpmem) are not supported by design.

NOTE: This release has not gone through the full validation cycle, but only through some basic tests on Travis and AppVeyor. Thus, it cannot be assumed "Production quality" and should not be used in production environments.

New Windows Support

Besides several minor improvements and bug fixes, all the other changes since NVML 1.2 release were related to Windows support:

  • win: port libvmem (and jemalloc)

  • win: benchmarks Windows port

  • win: fix mapping files of unaligned length

  • win: clean up possible race condition in mmap_init()

  • win: enable QueryVirtualMemoryInformation() in pmem_is_pmem()

  • test: check open handles at START/DONE

  • test: port all the remaining unit tests

    (scope, pmem_map, obj_debug, util_poolset, pmempool_*)

  • win: add resource files for versioning

Known Issues

Known issues and limitations of Windows version of NVML:

  • Unicode support is missing. The UTF/USC-encoded file paths

    or pool set files may not be handled correctly.

  • The libvmmalloc library is not ported yet.

  • The on-media format of pmem pools is not portable at the moment.

  • The pmem pools created using Windows version of NVM libraries

    cannot be open on Linux and vice versa.

  • Despite the fact the current version of NVML would work

    with any recent version of Windows OS, to take full

    advantage of PMEM and NVML features and to benefit from

    the PMEM performance, the recommended platforms needs be

    equipped with the real NVDIMMs hardware and should support

    the native, Microsoft's implementation of DAX-enabled file

    system (i.e. Windows Server 2016 or later).

    In case of using NVML with older versions of Windows or with

    the custom implementation of PMEM/DAX drivers, the performance

    might not be satisfactory. Please, contact the provider

    of PMEM/DAX drivers for your platform to get the customized

    version of NVML in such case.

Version 1.2

Thu Dec 15 2016 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

This NVML release causes a "flag day" for libpmemobj. The pmemobj pools built under NVML 1.1 are incompatible with pools built under NVML 1.2 and later. This is because an issue was discovered with the alignment of locks (#358) and, although rare, the issue potentially impacts program correctness, making the fix mandatory.

The major version number of the pmemobj pool layout and the version of the libpmemobj API is changed to prevent the use of the potentially incorrect layout.

Key Changes

  • Add Device DAX support, providing that "optimized flush" mechanism

    defined in SNIA NVM Programming Model can safely be used, even

    if PMEM-aware file system supporting that model is not available, or if the user does not want to use the file system for some reason.

  • Add a package for libpmemobj C++ bindings.

  • C++ API is no longer considered experimental.

  • Web-based documentation for C++ API is available on http://pmem.io.

  • Add "sync" and "transform" commands to pmempool utility.

    The "sync" command allows to recover missing or corrupted part(s)

    of a pool set from a healthy replica, while the "transform" command

    is a convenient way for modifying the structure of an existing

    pool set, i.e. by adding or removing replicas.

  • Add experimental support for remote access to persistent memory and

    basic remote data replication over RDMA (librpmem). Experimental

    support for remote replicas is also provided by libpmemobj library.

New features

  • common: add Device DAX support (#197)

  • obj: add C++ bindings package (libpmemobj++-devel)

  • obj: add TOID_OFFSETOF macro

  • pmempool: add "sync" and "transform" commands (#172, #196)

Bug fixes

  • obj: force alignment of pmem lock structures (#358)

  • blk: cast translation entry to uint64_t when calculating data offset

  • obj: fix Valgrind instrumentation of chunk headers and cancelled

    allocations

  • obj: set error message when user called pmemobj_tx_abort()

  • obj: fix status returned by pmemobj_list_insert() (#226)

  • obj: defer allocation of global structures

Optimizations

  • obj: fast path for pmemobj_pool_by_ptr() when inside a transaction

  • obj: simplify and optimize allocation class generation

Experimental features

  • rpmem: add support for remote access to persistent memory and basic

    remote data replication over RDMA

  • libpmempool: add pmempool_sync() and pmempool_transform() (#196)

  • obj: introduce pmemobj_oid()

  • obj: add pmemobj_tx_xalloc()/pmemobj_tx_xadd_range() APIs and

    the corresponding macros

  • obj: add transaction stage transition callbacks

Version 1.1

Thu Jun 23 2016 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

This NVML release introduces a new version of libpmemobj pool layout.

Internal undo log structure has been modified to improve performance of pmemobj transactions. Memory pools created with older versions of the libpmemobj library must be converted to the new format using "pmempool convert" command. See pmempool-convert(1) for details.

A new "libpmempool" library is available, providing support for off-line pool management and diagnostics. Initially it provides only "check" and "repair" operations for log and blk memory pools, and for BTT devices.

Other changes

  • pmem: deprecate PCOMMIT

  • blk: match BTT Flog initialization with Linux NVDIMM BTT

  • mem: defer pmem_is_pmem() initialization (#158)

  • obj: add TOID_TYPEOF macro

Bug fixes

  • doc: update description of valid file size units (#133)

  • pmempool: fix --version short option in man page (#135)

  • pmempool: print usage when running rm without arg (#136)

  • cpp: clarify polymorphism in persistent_ptr (#150)

  • obj: let the before flag be any non-zero value (#151)

  • obj: fix compare array pptr to nullptr (#152)

  • obj: cpp pool.get_root() fix (#156)

  • log/blk: set errno if replica section is specified (#161)

  • cpp: change exception message (#163)

  • doc: remove duplicated words in man page (#164)

  • common: always append EXTRA_CFLAGS after our CFLAGS

Experimental features

  • Implementation of C++ bindings for libpmempobj is complete.

  • Web-based documentation for C++ API is available on http://pmem.io.

    Note that C++ API is still considered experimental. Do not use it

    in production environments.

  • Porting NVML to Windows is in progress. There are MS Visual Studio

    solution/projects available, allowing to compile libpmem, libpmemlog,

    libpmemblk and libpmemobj on Windows, but the libraries are not fully

    functional and most of the test are not enabled yet.

Version 1.0

Thu Apr 07 2016 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

The API of six libraries (libpmem, libpmemblk, libpmemlog, libpmemobj, libvmem, libvmmalloc) is complete and stable.

The on-media layout of persistent memory pools will be maintained from this point, and if changed it will be backward compatible.

Man pages are all complete.

This release has been validated to "Production quality".

For the purpose of new features planned for next releases of NVML there have been some API modifications made:

  • pmem: pmem_map replaced with pmem_map_file

  • log/blk: 'off_t' substituted with 'long long'

  • obj: type numbers extended to 64-bit

  • obj: new entry points and macros added:

    • pmemobj_tx_errno

    • pmemobj_tx_lock

    • pmemobj_mutex_timedlock,

    • TX_ADD_DIRECT

    • TX_ADD_FIELD_DIRECT

    • TX_SET_DIRECT

Other key changes since version 0.4 include

  • common: updated/fixed installation scripts

  • common: eliminated dependency on libuuid

  • pmem: CPU features/ISA detection using CPUID

  • obj: improved error handling

  • obj: atomic allocation fails if constructor returns error

  • obj: multiple performance optimizations

  • obj: object store refactoring

  • obj: additional examples and benchmarks

This release also introduces a prototype implementation of C++ bindings for libpmemobj. Note that C++ API is still experimental and should not be used in production environments.

Version 0.4

Fri Dec 04 2015 Krzysztof Czurylo <krzysztof.czurylo@intel.com>

This NVML version primarily focuses on improving code quality and reliability. In addition to a couple of bug fixes, the changes include:

  • benchmarks for libpmemobj, libpmemblk and libvmem

  • additional pmemobj tests and examples

  • pool mapping address randomization

  • added pmempool "rm" command

  • eliminated libpmem dependency on libpthread

  • enabled extra warnings

  • minor performance improvements

Man pages are all complete.

This release is considered "Beta quality" by the team, having been thoroughly validated, including significant performance analysis.

The pmempool command does not yet support "check" and "repair" operations for pmemobj type pools.

Version 0.3

Sun Sep 13 2015 Andy Rudoff <andy.rudoff@intel.com>

NVML is now feature complete, adding support for:

  • pool sets

  • pmemobj local replication (active/passive)

  • experimental valgrind support

  • pmempool support for all pool types

Man pages are all complete.

This release is considered "Alpha quality" by the team, having gone through significant validation but only some performance analysis at this point.

Version 0.2

Tue Jun 30 2015 Andy Rudoff <andy.rudoff@intel.com>

NVML now consists of six libraries:

  • libpmem (basic flushing, etc)

  • libpmemblk, libpmemlog, libpmemobj (transactions)

  • libvmem, libvmmalloc (volatile use of pmem)

The "pmempool" command is available for managing pmem files.

Man pages for all the above are complete.

The only things documented in man pages but not implemented are:

  • pmem sets (ability to spread a pool over a set of files)

  • replication (coming for libpmemobj)

The pmempool command does not yet support pmemobj type pools.

Version 0.1

Thu Sep 11 2014 Andy Rudoff <andy.rudoff@intel.com>

  • Initial development done in 0.1 builds