Skip to content

5.0.0
Compare
Choose a tag to compare
@jasone jasone released this 13 Jun 20:18
· 1462 commits to dev since this release
5.0.0
5018fe3

Unlike all previous jemalloc releases, this release does not use naturally aligned "chunks" for virtual memory management, and instead uses page-aligned "extents". This change has few externally visible effects, but the internal impacts are... extensive. Many other internal changes combine to make this the most cohesively designed version of jemalloc so far, with ample opportunity for further enhancements.

Continuous integration is now an integral aspect of development thanks to the efforts of @davidtgoldblatt, and the dev branch tends to remain reasonably stable on the tested platforms (Linux, FreeBSD, macOS, and Windows). As a side effect the official release frequency may decrease over time.

New features:

  • Implement optional per-CPU arena support; threads choose which arena to use based on current CPU rather than on fixed thread-->arena associations. (@interwq)
  • Implement two-phase decay of unused dirty pages. Pages transition from dirty-->muzzy-->clean, where the first phase transition relies on madvise(... MADV_FREE) semantics, and the second phase transition discards pages such that they are replaced with demand-zeroed pages on next access. (@jasone)
  • Increase decay time resolution from seconds to milliseconds. (@jasone)
  • Implement opt-in per CPU background threads, and use them for asynchronous decay-driven unused dirty page purging. (@interwq)
  • Add mutex profiling, which collects a variety of statistics useful for diagnosing overhead/contention issues. (@interwq)
  • Add C++ new/delete operator bindings. (@djwatson)
  • Support manually created arena destruction, such that all data and metadata are discarded. Add MALLCTL_ARENAS_DESTROYED for accessing merged stats associated with destroyed arenas. (@jasone)
  • Add MALLCTL_ARENAS_ALL as a fixed index for use in accessing merged/destroyed arena statistics via mallctl. (@jasone)
  • Add opt.abort_conf to optionally abort if invalid configuration options are detected during initialization. (@interwq)
  • Add opt.stats_print_opts, so that e.g. JSON output can be selected for the stats dumped during exit if opt.stats_print is true. (@jasone)
  • Add --with-version=VERSION for use when embedding jemalloc into another project's git repository. (@jasone)
  • Add --disable-thp to support cross compiling. (@jasone)
  • Add --with-lg-hugepage to support cross compiling. (@jasone)
  • Add mallctl interfaces (various authors):
    • background_thread
    • opt.abort_conf
    • opt.retain
    • opt.percpu_arena
    • opt.background_thread
    • opt.{dirty,muzzy}_decay_ms
    • opt.stats_print_opts
    • arena.<i>.initialized
    • arena.<i>.destroy
    • arena.<i>.{dirty,muzzy}_decay_ms
    • arena.<i>.extent_hooks
    • arenas.{dirty,muzzy}_decay_ms
    • arenas.bin.<i>.slab_size
    • arenas.nlextents
    • arenas.lextent.<i>.size
    • arenas.create
    • stats.background_thread.{num_threads,num_runs,run_interval}
    • stats.mutexes.{ctl,background_thread,prof,reset}.{num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds,num_owner_switch}
    • stats.arenas.<i>.{dirty,muzzy}_decay_ms
    • stats.arenas.<i>.uptime
    • stats.arenas.<i>.{pmuzzy,base,internal,resident}
    • stats.arenas.<i>.{dirty,muzzy}_{npurge,nmadvise,purged}
    • stats.arenas.<i>.bins.<j>.{nslabs,reslabs,curslabs}
    • stats.arenas.<i>.bins.<j>.mutex.{num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds,num_owner_switch}
    • stats.arenas.<i>.lextents.<j>.{nmalloc,ndalloc,nrequests,curlextents}
    • stats.arenas.i.mutexes.{large,extent_avail,extents_dirty,extents_muzzy,extents_retained,decay_dirty,decay_muzzy,base,tcache_list}.{num_ops,num_spin_acq,num_wait,max_wait_time,total_wait_time,max_num_thds,num_owner_switch}

Portability improvements:

  • Improve reentrant allocation support, such that deadlock is less likely if e.g. a system library call in turn allocates memory. (@davidtgoldblatt, @interwq)
  • Support static linking of jemalloc with glibc. (@djwatson)

Optimizations and refactors:

  • Organize virtual memory as "extents" of virtual memory pages, rather than as naturally aligned "chunks", and store all metadata in arbitrarily distant locations. This reduces virtual memory external fragmentation, and will interact better with huge pages (not yet explicitly supported). (@jasone)
  • Fold large and huge size classes together; only small and large size classes remain. (@jasone)
  • Unify the allocation paths, and merge most fast-path branching decisions. (@davidtgoldblatt, @interwq)
  • Embed per thread automatic tcache into thread-specific data, which reduces conditional branches and dereferences. Also reorganize tcache to increase fast-path data locality. (@interwq)
  • Rewrite atomics to closely model the C11 API, convert various synchronization from mutex-based to atomic, and use the explicit memory ordering control to resolve various hypothetical races without increasing synchronization overhead. (@davidtgoldblatt)
  • Extensively optimize rtree via various methods:
    • Add multiple layers of rtree lookup caching, since rtree lookups are now part of fast-path deallocation. (@interwq)
    • Determine rtree layout at compile time. (@jasone)
    • Make the tree shallower for common configurations. (@jasone)
    • Embed the root node in the top-level rtree data structure, thus avoiding one level of indirection. (@jasone)
    • Further specialize leaf elements as compared to internal node elements, and directly embed extent metadata needed for fast-path deallocation. (@jasone)
    • Ignore leading always-zero address bits (architecture-specific). (@jasone)
  • Reorganize headers (ongoing work) to make them hermetic, and disentangle various module dependencies. (@davidtgoldblatt)
  • Convert various internal data structures such as size class metadata from boot-time-initialized to compile-time-initialized. Propagate resulting data structure simplifications, such as making arena metadata fixed-size. (@jasone)
  • Simplify size class lookups when constrained to size classes that are multiples of the page size. This speeds lookups, but the primary benefit is complexity reduction in code that was the source of numerous regressions. (@jasone)
  • Lock individual extents when possible for localized extent operations, rather than relying on a top-level arena lock. (@davidtgoldblatt, @jasone)
  • Use first fit layout policy instead of best fit, in order to improve packing. (@jasone)
  • If munmap(2) is not in use, use an exponential series to grow each arena's virtual memory, so that the number of disjoint virtual memory mappings remains low. (@jasone)
  • Implement per arena base allocators, so that arenas never share any virtual memory pages. (@jasone)
  • Automatically generate private symbol name mangling macros. (@jasone)

Incompatible changes:

  • Replace chunk hooks with an expanded/normalized set of extent hooks. (@jasone)
  • Remove ratio-based purging. (@jasone)
  • Remove --disable-tcache. (@jasone)
  • Remove --disable-tls. (@jasone)
  • Remove --enable-ivsalloc. (@jasone)
  • Remove --with-lg-size-class-group. (@jasone)
  • Remove --with-lg-tiny-min. (@jasone)
  • Remove --disable-cc-silence. (@jasone)
  • Remove --enable-code-coverage. (@jasone)
  • Remove --disable-munmap (replaced by opt.retain). (@jasone)
  • Remove Valgrind support. (@jasone)
  • Remove quarantine support. (@jasone)
  • Remove redzone support. (@jasone)
  • Remove mallctl interfaces (various authors):
    • config.munmap
    • config.tcache
    • config.tls
    • config.valgrind
    • opt.lg_chunk
    • opt.purge
    • opt.lg_dirty_mult
    • opt.decay_time
    • opt.quarantine
    • opt.redzone
    • opt.thp
    • arena.<i>.lg_dirty_mult
    • arena.<i>.decay_time
    • arena.<i>.chunk_hooks
    • arenas.initialized
    • arenas.lg_dirty_mult
    • arenas.decay_time
    • arenas.bin.<i>.run_size
    • arenas.nlruns
    • arenas.lrun.<i>.size
    • arenas.nhchunks
    • arenas.hchunk.<i>.size
    • arenas.extend
    • stats.cactive
    • stats.arenas.<i>.lg_dirty_mult
    • stats.arenas.<i>.decay_time
    • stats.arenas.<i>.metadata.{mapped,allocated}
    • stats.arenas.<i>.{npurge,nmadvise,purged}
    • stats.arenas.<i>.huge.{allocated,nmalloc,ndalloc,nrequests}
    • stats.arenas.<i>.bins.<j>.{nruns,reruns,curruns}
    • stats.arenas.<i>.lruns.<j>.{nmalloc,ndalloc,nrequests,curruns}
    • stats.arenas.<i>.hchunks.<j>.{nmalloc,ndalloc,nrequests,curhchunks}

Bug fixes:

  • Improve interval-based profile dump triggering to dump only one profile when a single allocation's size exceeds the interval. (@jasone)
  • Use prefixed function names (as controlled by --with-jemalloc-prefix) when pruning backtrace frames in jeprof. (@jasone)
Assets 3