Re: [PATCH v5 0/7] dso/dsos memory savings and clean up

[Ian Rogers]

On Mon, Apr 29, 2024 at 12:17 PM Arnaldo Carvalho de Melo
<acme@xxxxxxxxxx> wrote:
>
> On Mon, Apr 29, 2024 at 11:46:07AM -0700, Ian Rogers wrote:
> > 7 more patches from:
> > https://lore.kernel.org/lkml/20240202061532.1939474-1-irogers@xxxxxxxxxx/
> > a near half year old adventure in trying to lower perf's dynamic
> > memory use. Bits like the memory overhead of opendir are on the
> > sidelines for now, too much fighting over how
> > distributions/C-libraries present getdents. These changes are more
> > good old fashioned replace an rb-tree with a sorted array and add
> > reference count tracking.
> >
> > The changes migrate dsos code, the collection of dso structs, more
> > into the dsos.c/dsos.h files. As with maps and threads, this is done
> > so the internals can be changed - replacing a linked list (for fast
> > iteration) and an rb-tree (for fast finds) with a lazily sorted
> > array. The complexity of operations remain roughly the same, although
> > iterating an array is likely faster than iterating a linked list, the
> Th> memory usage is at least reduced by half.
> >
> > As fixing the memory usage necessitates changing operations like find,
> > modify these operations so that they increment the reference count to
> > avoid races like a find in dsos and a remove. Similarly tighten up
> > lock usage so that operations working on dsos state hold the
> > appropriate lock. Note, since this series is partially applied in the
> > perf-tools-next tree currently some memory leaks have been introduced.
> >
> > v5. Rebase, adding use of accessors to dso as necessary. Previous
> >     versions were all rebases or dropping merged patches.
>
> So, on an Intel machine:
>
> ⬢[acme@toolbox perf-tools-next]$ git log --oneline -8
> fb401385575211e6 (HEAD -> perf-tools-next) perf dso: Use container_of to avoid a pointer in 'struct dso_data'
> 0fe118d129ab1c77 perf dso: Reference counting related fixes
> 35e44adf6103a407 perf dso: Add reference count checking and accessor functions
> 35673675ebbbac5d perf dsos: Switch hand code to bsearch()
> 654d60f2f5c737cd perf dsos: Remove __dsos__findnew_link_by_longname_id()
> 94b0ba802e090b66 perf dsos: Remove __dsos__addnew()
> 47692286dd856469 perf dsos: Switch backing storage to array from rbtree/list
> 8c618b58c89ce4c2 (perf-tools-next.korg/perf-tools-next, acme.korg/perf-tools-next) perf test: Reintroduce -p/--parallel and make -S/--sequential the default
> ⬢[acme@toolbox perf-tools-next]$
>
> When I'm at:
>
> 8c618b58c89ce4c2 (perf-tools-next.korg/perf-tools-next, acme.korg/perf-tools-next) perf test: Reintroduce -p/--parallel and make -S/--sequential the default
>
> root@x1:~# perf -v
> perf version 6.9.rc5.g8c618b58c89c
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : Ok
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : Ok
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : Ok
> root@x1:~# time perf test "lock contention"
>  87: kernel lock contention analysis test                            : Ok
>
> real    0m9.143s
> user    0m5.201s
> sys     0m4.812s
> root@x1:~#
>
> Moving to the first patch in this series:
>
> ⬢[acme@toolbox perf-tools-next]$ git log --oneline -2
> 47692286dd856469 (HEAD) perf dsos: Switch backing storage to array from rbtree/list
> 8c618b58c89ce4c2 (perf-tools-next.korg/perf-tools-next, acme.korg/perf-tools-next) perf test: Reintroduce -p/--parallel and make -S/--sequential the default
> ⬢[acme@toolbox perf-tools-next]$ alias m
> alias m='rm -rf ~/libexec/perf-core/ ; make -k CORESIGHT=1 O=/tmp/build/$(basename $PWD)/ -C tools/perf install-bin && perf test pythond'
> ⬢[acme@toolbox perf-tools-next]$ rm -rf /tmp/build/$(basename $PWD)/ ; mkdir -p /tmp/build/$(basename $PWD)/ ; m
> <SNIP>
>
> root@x1:~# perf -v
> perf version 6.9.rc5.g47692286dd85
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : FAILED!
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : FAILED!
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : FAILED!
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : FAILED!
> root@x1:~# perf test -vvv "lock contention"
>  87: kernel lock contention analysis test:
> --- start ---
> test child forked, pid 2279518
> Testing perf lock record and perf lock contention
> [Fail] Recorded result count is not 1: 2
> ---- end(-1) ----
>  87: kernel lock contention analysis test                            : FAILED!
> root@x1:~#
>
> This breaks bisectability, but then lets see if at the end of the series
> it works:
>
> ⬢[acme@toolbox perf-tools-next]$ git log --oneline -8
> fb401385575211e6 (HEAD -> perf-tools-next) perf dso: Use container_of to avoid a pointer in 'struct dso_data'
> 0fe118d129ab1c77 perf dso: Reference counting related fixes
> 35e44adf6103a407 perf dso: Add reference count checking and accessor functions
> 35673675ebbbac5d perf dsos: Switch hand code to bsearch()
> 654d60f2f5c737cd perf dsos: Remove __dsos__findnew_link_by_longname_id()
> 94b0ba802e090b66 perf dsos: Remove __dsos__addnew()
> 47692286dd856469 perf dsos: Switch backing storage to array from rbtree/list
> 8c618b58c89ce4c2 (perf-tools-next.korg/perf-tools-next, acme.korg/perf-tools-next) perf test: Reintroduce -p/--parallel and make -S/--sequential the default
> ⬢[acme@toolbox perf-tools-next]$ rm -rf /tmp/build/$(basename $PWD)/ ; mkdir -p /tmp/build/$(basename $PWD)/ ; m
>
> root@x1:~# perf -v
> perf version 6.9.rc5.gfb4013855752
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : FAILED!
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : FAILED!
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : FAILED!
> root@x1:~# perf test "lock contention"
>  87: kernel lock contention analysis test                            : FAILED!
> root@x1:~# perf test -vvv "lock contention"
>  87: kernel lock contention analysis test:
> --- start ---
> test child forked, pid 2289060
> Testing perf lock record and perf lock contention
> Testing perf lock contention --use-bpf
> Testing perf lock record and perf lock contention at the same time
> [Fail] Recorded result count is not 1: 2
> ---- end(-1) ----
>  87: kernel lock contention analysis test                            : FAILED!
> root@x1:~#
>
> ⬢[acme@toolbox perf-tools-next]$ gcc --version | head -1
> gcc (GCC) 13.2.1 20240316 (Red Hat 13.2.1-7)
> ⬢[acme@toolbox perf-tools-next]$ rpm -q gcc
> gcc-13.2.1-7.fc39.x86_64
> ⬢[acme@toolbox perf-tools-next]$
>
> root@x1:~# grep -m1 'model name' /proc/cpuinfo
> model name      : 13th Gen Intel(R) Core(TM) i7-1365U
> root@x1:~# free
>                total        used        free      shared  buff/cache   available
> Mem:        32507912     9081644     3531432     1987616    22554868    23426268
> Swap:        8388604      314112     8074492
> root@x1:~#

Thanks, will check. Will likely need a v6 to fix. v5 was just
addressing the rebase issues.

Ian

> - Arnaldo