Re: [PATCH v2 0/6] maps memory improvements and fixes

[Namhyung Kim]

On Sat, Feb 10, 2024 at 10:08 AM Ian Rogers <irogers@xxxxxxxxxx> wrote:
>
> On Fri, Feb 9, 2024 at 6:46 PM Ian Rogers <irogers@xxxxxxxxxx> wrote:
> >
> > On Thu, Feb 8, 2024 at 9:44 AM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
> > >
> > > Hi Ian,
> > >
> > > On Wed, Feb 7, 2024 at 2:37 PM Ian Rogers <irogers@xxxxxxxxxx> wrote:
> > > >
> > > > First 6 patches from:
> > > > https://lore.kernel.org/lkml/20240202061532.1939474-1-irogers@xxxxxxxxxx/
> > > >
> > > > v2. Fix NO_LIBUNWIND=1 build issue.
> > > >
> > > > Ian Rogers (6):
> > > >   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
> > >
> > > Now I see a perf test failure on the vmlinux test:
> > >
> > > $ sudo ./perf test -v vmlinux
> > >   1: vmlinux symtab matches kallsyms                                 :
> > > --- start ---
> > > test child forked, pid 4164115
> > > /proc/{kallsyms,modules} inconsistency while looking for
> > > "[__builtin__kprobes]" module!
> > > /proc/{kallsyms,modules} inconsistency while looking for
> > > "[__builtin__kprobes]" module!
> > > /proc/{kallsyms,modules} inconsistency while looking for
> > > "[__builtin__ftrace]" module!
> > > Looking at the vmlinux_path (8 entries long)
> > > Using /usr/lib/debug/boot/vmlinux-6.5.13-1rodete2-amd64 for symbols
> > > perf: Segmentation fault
> > > Obtained 16 stack frames.
> > > ./perf(+0x1b7dcd) [0x55c40be97dcd]
> > > ./perf(+0x1b7eb7) [0x55c40be97eb7]
> > > /lib/x86_64-linux-gnu/libc.so.6(+0x3c510) [0x7f33d7a5a510]
> > > ./perf(+0x1c2e9c) [0x55c40bea2e9c]
> > > ./perf(+0x1c43f6) [0x55c40bea43f6]
> > > ./perf(+0x1c4649) [0x55c40bea4649]
> > > ./perf(+0x1c46d3) [0x55c40bea46d3]
> > > ./perf(+0x1c7303) [0x55c40bea7303]
> > > ./perf(+0x1c70b5) [0x55c40bea70b5]
> > > ./perf(+0x1c73e6) [0x55c40bea73e6]
> > > ./perf(+0x11833e) [0x55c40bdf833e]
> > > ./perf(+0x118f78) [0x55c40bdf8f78]
> > > ./perf(+0x103d49) [0x55c40bde3d49]
> > > ./perf(+0x103e75) [0x55c40bde3e75]
> > > ./perf(+0x1044c0) [0x55c40bde44c0]
> > > ./perf(+0x104de0) [0x55c40bde4de0]
> > > test child interrupted
> > > ---- end ----
> > > vmlinux symtab matches kallsyms: FAILED!
> >
> > Ah, tripped over a latent bug summarized in this part of an asan stack trace:
> > ```
> > freed by thread T0 here:
> >    #0 0x7fa13bcd74b5 in __interceptor_realloc
> > ../../../../src/libsanitizer/asan/asan_malloc_linux.cpp:85
> >    #1 0x561d66377713 in __maps__insert util/maps.c:353
> >    #2 0x561d66377b89 in maps__insert util/maps.c:413
> >    #3 0x561d6652911d in dso__process_kernel_symbol util/symbol-elf.c:1460
> >    #4 0x561d6652aaae in dso__load_sym_internal util/symbol-elf.c:1675
> >    #5 0x561d6652b6dc in dso__load_sym util/symbol-elf.c:1771
> >    #6 0x561d66321a4e in dso__load util/symbol.c:1914
> >    #7 0x561d66372cd9 in map__load util/map.c:353
> >    #8 0x561d663730e7 in map__find_symbol_by_name_idx util/map.c:397
> >    #9 0x561d663731e7 in map__find_symbol_by_name util/map.c:410
> >    #10 0x561d66378208 in maps__find_symbol_by_name_cb util/maps.c:524
> >    #11 0x561d66377f49 in maps__for_each_map util/maps.c:471
> >    #12 0x561d663784a0 in maps__find_symbol_by_name util/maps.c:546
> >    #13 0x561d662093e8 in machine__find_kernel_symbol_by_name util/machine.h:243
> >    #14 0x561d6620abbd in test__vmlinux_matches_kallsyms
> > tests/vmlinux-kallsyms.c:330
> > ...
> > ```
> > dso__process_kernel_symbol rewrites the kernel maps here:
> > https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/symbol-elf.c#n1378
> > which resizes the maps_by_address array causing the maps__for_each_map
> > iteration in frame 11 to be iterating over a stale/freed value.
> >
> > The most correct solutions would be to clone the maps_by_address array
> > prior to iteration, or reference count maps_by_address and its size.
> > Neither of these solutions particularly appeal, so just reloading the
> > maps_by_address and size on each iteration also fixes the problem, but
> > possibly causes some maps to be skipped/repeated. I think this is
> > acceptable correctness for the performance.

Can we move map__load() out of maps__for_each_map() ?
I think the callback should just return the map and break the loop.
And it can call the map__load() out of the read lock.

>
> An aside, shouldn't taking a write lock to modify the maps deadlock
> with holding the read lock for iteration? Well no because
> perf_singlethreaded is true for the test:
> https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/rwsem.c#n17
> Another perf_singlethreaded considered evil :-) Note, just getting rid
> of perf_singlethreaded means latent bugs like this will pop up and
> will need resolution.

Yeah, maybe.  How about turning it on in the test code?

Thanks,
Namhyung