[PATCH v5 00/50] Improvements to memory use
From: Ian Rogers
Date: Mon Nov 27 2023 - 17:09:26 EST
Fix memory leaks detected by address/leak sanitizer affecting LBR
call-graphs, perf mem and BPF offcpu.
Make branch_type_stat in callchain_list optional as it is large and
not always necessary - in particular it isn't used by perf top.
Make the allocations of zstd streams, kernel symbols and event copies
lazier in order to save memory in cases like perf record.
Handle the thread exit event and have it remove the thread from the
threads set in machine. Don't do this for perf report as it causes a
regression for task lists, which assume threads are never removed from
the machine's set, and offcpu events, that may sythensize samples for
threads that have exited.
Avoid using 8kb buffers for filename__read_str which is excessive for
reading CPU maps. Add io_dir as an allocation free readdir
replacement, opendir allocating 32kb by default and the code uses it
recursively.
Shrink perf map using a two value byte to replace two function
pointers. Modify the implementation of maps to not use an rbtree as
the container for maps, instead use a sorted array. Improve locking
and reference counting issues.
Similar to maps separate out and reimplement threads to use a hashmap
for lower memory consumption and faster look up. The fixes a
regression in memory usage where reference count checking switched to
using non-invasive tree nodes. Reduce its default size by 32 times
and improve locking discipline. Also, fix regressions where tids had
become unordered to make `perf report --tasks` and
`perf trace --summary` output easier to read.
Better encapsulate the dsos abstraction. Remove the linked list and
rbtree used for faster iteration and log(n) lookup to a sorted array
for similar performance but half the memory usage per dso. Improve
reference counting and locking discipline, adding reference count
checking to dso. Experimented with, but abandoned, a hashmap
implementation due to the need for extra storage and the keys not
being stable.
The overall effect is to reduce memory consumption significantly for
perf top - with call graphs enabled running longer before 1GB of
memory is consumed. For a perf record of 'true', the memory
consumption goes from 39912kb max resident to 20096kb max resident -
nearly halved. perf inject with -b of a system wide perf record of
'true' reduces the max resident by roughly 4.5% (3.4% in v4 due to
branch_type_stat changes being merged). This is while improving
correctness with locking discipline and reference count checking.
Patch organization (v5):
- 50 patches is a lot, the patches aren't divided as they merge conflict and
later patches, for example in dsos, rely on the changes and fixes to maps.
- the dso reference count checking patch is larger due to switch use of dso to
be by accessors, to encapsulate the reference count checker macros. The
reference count checking changes within this largely mechanical change amount
to a few lines and so weren't separated.
- the first patch contains a build fix if the rwsem error checking is
enabled missed from v3.
- the next patches are an assortment of memory size fixes.
- the next patches are the refactoring of maps.
- the next patches are the refactoring of threads.
- the next patches are the refactoring of dsos.
- finally reference count checking is added to dso and some lock/reference
count issues are resolved. This is done after changing the data structures,
for example, as the single pointer on an array is easier to add reference
count checking to compared to the 5 previous pointers.
v5: 3 patches were merged. 2nd patch addressed feedback from
namhyung@xxxxxxxxxx and Guilherme Amadio <amadio@xxxxxxxxxx>. 4th
patch rename function to getdelim as suggested by
namhyung@xxxxxxxxxx. 5 patch adds the missing sysfs mountpoint as
suggested by namhyung@xxxxxxxxxx. 49th patch fix a missed put in
the dso_data tests.
v4: Rebased as 11 changes moved to perf-tools-next. Address comments
from v3 such as error checking on zstd streams. Improve the
dsos/dso in ways similar to threads and maps, with the addition of
reference count checking on dso.
v3: Additional memory/speed improvements, in particular for maps and
threads. Address review comments from namhyung@xxxxxxxxxx and
adrian.hunter@xxxxxxxxx.
v2: Add additional memory fixes on top of initial LBR and rc check
fixes.
Ian Rogers (50):
perf comm: Use regular mutex
libperf: Lazily allocate/size mmap event copy
perf mmap: Lazily initialize zstd streams
tools api fs: Switch filename__read_str to use io.h
tools api fs: Avoid reading whole file for a 1 byte bool
tools lib api: Add io_dir an allocation free readdir alternative
perf maps: Switch modules tree walk to io_dir__readdir
perf record: Be lazier in allocating lost samples buffer
perf pmu: Switch to io_dir__readdir
perf header: Switch mem topology to io_dir__readdir
perf events: Remove scandir in thread synthesis
perf map: Simplify map_ip/unmap_ip and make map size smaller
perf maps: Move symbol maps functions to maps.c
perf thread: Add missing RC_CHK_EQUAL
perf maps: Add maps__for_each_map to call a function on each entry
perf maps: Add remove maps function to remove a map based on callback
perf debug: Expose debug file
perf maps: Refactor maps__fixup_overlappings
perf maps: Do simple merge if given map doesn't overlap
perf maps: Rename clone to copy from
perf maps: Add maps__load_first
perf maps: Add find next entry to give entry after the given map
perf maps: Reduce scope of map_rb_node and maps internals
perf maps: Fix up overlaps during fixup_end
perf maps: Switch from rbtree to lazily sorted array for addresses
perf maps: Get map before returning in maps__find
perf maps: Get map before returning in maps__find_by_name
perf maps: Get map before returning in maps__find_next_entry
perf maps: Hide maps internals
perf maps: Locking tidy up of nr_maps
perf dso: Reorder variables to save space in struct dso
perf report: Sort child tasks by tid
perf trace: Ignore thread hashing in summary
perf machine: Move fprintf to for_each loop and a callback
perf threads: Move threads to its own files
perf threads: Switch from rbtree to hashmap
perf threads: Reduce table size from 256 to 8
perf dsos: Attempt to better abstract dsos internals
perf dsos: Tidy reference counting and locking
perf dsos: Add dsos__for_each_dso
perf dso: Move dso functions out of dsos
perf dsos: Switch more loops to dsos__for_each_dso
perf dsos: Switch backing storage to array from rbtree/list
perf dsos: Remove __dsos__addnew
perf dsos: Remove __dsos__findnew_link_by_longname_id
perf dsos: Switch hand code to bsearch
perf dso: Add reference count checking and accessor functions
perf dso: Reference counting related fixes
perf dso: Use container_of to avoid a pointer in dso_data
perf env: Avoid recursively taking env->bpf_progs.lock
tools/lib/api/Makefile | 2 +-
tools/lib/api/fs/fs.c | 80 +-
tools/lib/api/io.h | 11 +-
tools/lib/api/io_dir.h | 75 +
tools/lib/perf/include/internal/mmap.h | 3 +-
tools/lib/perf/mmap.c | 21 +-
tools/perf/arch/x86/tests/dwarf-unwind.c | 1 +
tools/perf/arch/x86/util/event.c | 103 +-
tools/perf/builtin-annotate.c | 6 +-
tools/perf/builtin-buildid-cache.c | 2 +-
tools/perf/builtin-buildid-list.c | 18 +-
tools/perf/builtin-inject.c | 96 +-
tools/perf/builtin-kallsyms.c | 2 +-
tools/perf/builtin-mem.c | 4 +-
tools/perf/builtin-record.c | 57 +-
tools/perf/builtin-report.c | 243 ++--
tools/perf/builtin-script.c | 8 +-
tools/perf/builtin-top.c | 4 +-
tools/perf/builtin-trace.c | 41 +-
tools/perf/tests/code-reading.c | 8 +-
tools/perf/tests/dso-data.c | 67 +-
tools/perf/tests/hists_common.c | 6 +-
tools/perf/tests/hists_cumulate.c | 4 +-
tools/perf/tests/hists_output.c | 2 +-
tools/perf/tests/maps.c | 64 +-
tools/perf/tests/symbols.c | 2 +-
tools/perf/tests/thread-maps-share.c | 8 +-
tools/perf/tests/vmlinux-kallsyms.c | 181 +--
tools/perf/ui/browsers/annotate.c | 6 +-
tools/perf/ui/browsers/hists.c | 8 +-
tools/perf/ui/browsers/map.c | 4 +-
tools/perf/util/Build | 1 +
tools/perf/util/annotate.c | 44 +-
tools/perf/util/auxtrace.c | 2 +-
tools/perf/util/block-info.c | 2 +-
tools/perf/util/bpf-event.c | 17 +-
tools/perf/util/bpf-event.h | 12 +-
tools/perf/util/bpf_lock_contention.c | 10 +-
tools/perf/util/build-id.c | 136 +-
tools/perf/util/build-id.h | 2 -
tools/perf/util/callchain.c | 4 +-
tools/perf/util/comm.c | 10 +-
tools/perf/util/compress.h | 6 +-
tools/perf/util/data-convert-json.c | 2 +-
tools/perf/util/db-export.c | 6 +-
tools/perf/util/debug.c | 22 +-
tools/perf/util/debug.h | 1 +
tools/perf/util/dlfilter.c | 12 +-
tools/perf/util/dso.c | 468 ++++---
tools/perf/util/dso.h | 544 ++++++--
tools/perf/util/dsos.c | 529 ++++---
tools/perf/util/dsos.h | 40 +-
tools/perf/util/env.c | 53 +-
tools/perf/util/env.h | 4 +
tools/perf/util/event.c | 12 +-
tools/perf/util/header.c | 47 +-
tools/perf/util/hist.c | 4 +-
tools/perf/util/intel-pt.c | 22 +-
tools/perf/util/machine.c | 652 +++------
tools/perf/util/machine.h | 32 +-
tools/perf/util/map.c | 93 +-
tools/perf/util/map.h | 83 +-
tools/perf/util/maps.c | 1239 +++++++++++++----
tools/perf/util/maps.h | 95 +-
tools/perf/util/mmap.c | 5 +-
tools/perf/util/mmap.h | 1 -
tools/perf/util/pmu.c | 48 +-
tools/perf/util/pmus.c | 30 +-
tools/perf/util/probe-event.c | 62 +-
tools/perf/util/rb_resort.h | 5 -
.../scripting-engines/trace-event-python.c | 21 +-
tools/perf/util/session.c | 21 +
tools/perf/util/session.h | 2 +
tools/perf/util/sort.c | 19 +-
tools/perf/util/srcline.c | 65 +-
tools/perf/util/symbol-elf.c | 138 +-
tools/perf/util/symbol.c | 521 ++-----
tools/perf/util/symbol.h | 1 -
tools/perf/util/symbol_fprintf.c | 4 +-
tools/perf/util/synthetic-events.c | 156 ++-
tools/perf/util/thread.c | 48 +-
tools/perf/util/thread.h | 6 -
tools/perf/util/threads.c | 186 +++
tools/perf/util/threads.h | 35 +
tools/perf/util/unwind-libunwind-local.c | 50 +-
tools/perf/util/unwind-libunwind.c | 9 +-
tools/perf/util/vdso.c | 89 +-
tools/perf/util/zstd.c | 63 +-
88 files changed, 4101 insertions(+), 2827 deletions(-)
create mode 100644 tools/lib/api/io_dir.h
create mode 100644 tools/perf/util/threads.c
create mode 100644 tools/perf/util/threads.h
--
2.43.0.rc1.413.gea7ed67945-goog