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
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).
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
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
Tue Feb 19 2018 Marcin Ślusarz <[email protected]>
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
Fri Oct 26 2018 Marcin Ślusarz <[email protected]>
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:
cto: removed
obj: remove actions limit
common: new dependency on libndctl on Linux
pmempool: "convert" subcommand is now a wrapper around pmdk-convert
(please see https://github.com/pmem/pmdk-convert)
obj: C++ bindings have been moved to a new repository (please see https://github.com/pmem/libpmemobj-cpp)
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
Thu Mar 29 2018 Krzysztof Czurylo <[email protected]>
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
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: 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
Wed Dec 20 2017 Krzysztof Czurylo <[email protected]>
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: initialize TLS data
rpmem: fix cleanup if fork() failed (#634)
obj: fix bogus OOM after exhausting first zone
Thu Jul 13 2017 Krzysztof Czurylo <[email protected]>
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: 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
Thu May 18 2017 Krzysztof Czurylo <[email protected]>
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
Sat Apr 15 2017 Krzysztof Czurylo <[email protected]>
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
Tue Feb 21 2017 Krzysztof Czurylo <[email protected]>
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 <[email protected]>
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.
Thu Dec 15 2016 Krzysztof Czurylo <[email protected]>
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
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
Thu Jun 23 2016 Krzysztof Czurylo <[email protected]>
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.
Thu Apr 07 2016 Krzysztof Czurylo <[email protected]>
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.
Fri Dec 04 2015 Krzysztof Czurylo <[email protected]>
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.
Sun Sep 13 2015 Andy Rudoff <[email protected]>
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.
Tue Jun 30 2015 Andy Rudoff <[email protected]>
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.
Thu Sep 11 2014 Andy Rudoff <[email protected]>
Initial development done in 0.1 builds
