2024-04-29 22:57:48

by Namhyung Kim

[permalink] [raw]
Subject: [PATCH] perf maps: Remove check_invariants() from maps__lock()

I found that the debug build was a lot slowed down by the maps lock code
since it checks the invariants whenever it gets the pointer to the lock.
This means it checks twice the invariants before and after the access.

Instead, let's move the checking code within the lock area but after any
modification and remove it from the read paths. This would remove (more
than) the half of the maps lock overhead.

The time for perf report with a huge data file (200k+ of MMAP2 events).

Non-debug Before After
--------- -------- --------
2m 43s 6m 45s 4m 21s

Signed-off-by: Namhyung Kim <[email protected]>
---
tools/perf/util/maps.c | 16 +++++++++++-----
1 file changed, 11 insertions(+), 5 deletions(-)

diff --git a/tools/perf/util/maps.c b/tools/perf/util/maps.c
index 725300896f38..61eb742d91e3 100644
--- a/tools/perf/util/maps.c
+++ b/tools/perf/util/maps.c
@@ -211,11 +211,6 @@ void maps__set_unwind_libunwind_ops(struct maps *maps, const struct unwind_libun

static struct rw_semaphore *maps__lock(struct maps *maps)
{
- /*
- * When the lock is acquired or released the maps invariants should
- * hold.
- */
- check_invariants(maps);
return &RC_CHK_ACCESS(maps)->lock;
}

@@ -358,6 +353,7 @@ static int map__strcmp(const void *a, const void *b)
static int maps__sort_by_name(struct maps *maps)
{
int err = 0;
+
down_write(maps__lock(maps));
if (!maps__maps_by_name_sorted(maps)) {
struct map **maps_by_name = maps__maps_by_name(maps);
@@ -384,6 +380,7 @@ static int maps__sort_by_name(struct maps *maps)
maps__set_maps_by_name_sorted(maps, true);
}
}
+ check_invariants(maps);
up_write(maps__lock(maps));
return err;
}
@@ -502,6 +499,7 @@ int maps__insert(struct maps *maps, struct map *map)

down_write(maps__lock(maps));
ret = __maps__insert(maps, map);
+ check_invariants(maps);
up_write(maps__lock(maps));
return ret;
}
@@ -536,6 +534,7 @@ void maps__remove(struct maps *maps, struct map *map)
{
down_write(maps__lock(maps));
__maps__remove(maps, map);
+ check_invariants(maps);
up_write(maps__lock(maps));
}

@@ -602,6 +601,7 @@ void maps__remove_maps(struct maps *maps, bool (*cb)(struct map *map, void *data
else
i++;
}
+ check_invariants(maps);
up_write(maps__lock(maps));
}

@@ -942,6 +942,8 @@ int maps__copy_from(struct maps *dest, struct maps *parent)
map__put(new);
}
}
+ check_invariants(dest);
+
up_read(maps__lock(parent));
up_write(maps__lock(dest));
return err;
@@ -1097,6 +1099,7 @@ void maps__fixup_end(struct maps *maps)
map__set_end(maps_by_address[n - 1], ~0ULL);

RC_CHK_ACCESS(maps)->ends_broken = false;
+ check_invariants(maps);

up_write(maps__lock(maps));
}
@@ -1147,6 +1150,8 @@ int maps__merge_in(struct maps *kmaps, struct map *new_map)
map__start(kmaps_maps_by_address[first_after_]) >= map__end(new_map)) {
/* No overlap so regular insert suffices. */
int ret = __maps__insert(kmaps, new_map);
+
+ check_invariants(kmaps);
up_write(maps__lock(kmaps));
return ret;
}
@@ -1184,6 +1189,7 @@ int maps__merge_in(struct maps *kmaps, struct map *new_map)
map__zput(kmaps_maps_by_address[i]);

free(kmaps_maps_by_address);
+ check_invariants(kmaps);
up_write(maps__lock(kmaps));
return 0;
}
--
2.44.0.769.g3c40516874-goog



2024-04-30 02:10:17

by Ian Rogers

[permalink] [raw]
Subject: Re: [PATCH] perf maps: Remove check_invariants() from maps__lock()

On Mon, Apr 29, 2024 at 3:57 PM Namhyung Kim <[email protected]> wrote:
>
> I found that the debug build was a lot slowed down by the maps lock code
> since it checks the invariants whenever it gets the pointer to the lock.
> This means it checks twice the invariants before and after the access.
>
> Instead, let's move the checking code within the lock area but after any
> modification and remove it from the read paths. This would remove (more
> than) the half of the maps lock overhead.
>
> The time for perf report with a huge data file (200k+ of MMAP2 events).
>
> Non-debug Before After
> --------- -------- --------
> 2m 43s 6m 45s 4m 21s

Thanks Namhyung, I think the change makes sense. There is an issue
that a user of a map may mutate it in a way that breaks the maps
sorting, symbol code would be a likely culprit of such a thing. A fix
for such a breakage would be to just set the unsorted flag on maps.
We'd be less likely to spot such an issue after this change, but it is
simple enough to add the function call when needed. Another option
would be to add another make flag like PARSER_DEBUG for these kind of,
"be paranoid," type things. Anyway, let's try to fix those problems
when they exist.

Reviewed-by: Ian Rogers <[email protected]>

Thanks,
Ian

> Signed-off-by: Namhyung Kim <[email protected]>
> ---
> tools/perf/util/maps.c | 16 +++++++++++-----
> 1 file changed, 11 insertions(+), 5 deletions(-)
>
> diff --git a/tools/perf/util/maps.c b/tools/perf/util/maps.c
> index 725300896f38..61eb742d91e3 100644
> --- a/tools/perf/util/maps.c
> +++ b/tools/perf/util/maps.c
> @@ -211,11 +211,6 @@ void maps__set_unwind_libunwind_ops(struct maps *maps, const struct unwind_libun
>
> static struct rw_semaphore *maps__lock(struct maps *maps)
> {
> - /*
> - * When the lock is acquired or released the maps invariants should
> - * hold.
> - */
> - check_invariants(maps);
> return &RC_CHK_ACCESS(maps)->lock;
> }
>
> @@ -358,6 +353,7 @@ static int map__strcmp(const void *a, const void *b)
> static int maps__sort_by_name(struct maps *maps)
> {
> int err = 0;
> +
> down_write(maps__lock(maps));
> if (!maps__maps_by_name_sorted(maps)) {
> struct map **maps_by_name = maps__maps_by_name(maps);
> @@ -384,6 +380,7 @@ static int maps__sort_by_name(struct maps *maps)
> maps__set_maps_by_name_sorted(maps, true);
> }
> }
> + check_invariants(maps);
> up_write(maps__lock(maps));
> return err;
> }
> @@ -502,6 +499,7 @@ int maps__insert(struct maps *maps, struct map *map)
>
> down_write(maps__lock(maps));
> ret = __maps__insert(maps, map);
> + check_invariants(maps);
> up_write(maps__lock(maps));
> return ret;
> }
> @@ -536,6 +534,7 @@ void maps__remove(struct maps *maps, struct map *map)
> {
> down_write(maps__lock(maps));
> __maps__remove(maps, map);
> + check_invariants(maps);
> up_write(maps__lock(maps));
> }
>
> @@ -602,6 +601,7 @@ void maps__remove_maps(struct maps *maps, bool (*cb)(struct map *map, void *data
> else
> i++;
> }
> + check_invariants(maps);
> up_write(maps__lock(maps));
> }
>
> @@ -942,6 +942,8 @@ int maps__copy_from(struct maps *dest, struct maps *parent)
> map__put(new);
> }
> }
> + check_invariants(dest);
> +
> up_read(maps__lock(parent));
> up_write(maps__lock(dest));
> return err;
> @@ -1097,6 +1099,7 @@ void maps__fixup_end(struct maps *maps)
> map__set_end(maps_by_address[n - 1], ~0ULL);
>
> RC_CHK_ACCESS(maps)->ends_broken = false;
> + check_invariants(maps);
>
> up_write(maps__lock(maps));
> }
> @@ -1147,6 +1150,8 @@ int maps__merge_in(struct maps *kmaps, struct map *new_map)
> map__start(kmaps_maps_by_address[first_after_]) >= map__end(new_map)) {
> /* No overlap so regular insert suffices. */
> int ret = __maps__insert(kmaps, new_map);
> +
> + check_invariants(kmaps);
> up_write(maps__lock(kmaps));
> return ret;
> }
> @@ -1184,6 +1189,7 @@ int maps__merge_in(struct maps *kmaps, struct map *new_map)
> map__zput(kmaps_maps_by_address[i]);
>
> free(kmaps_maps_by_address);
> + check_invariants(kmaps);
> up_write(maps__lock(kmaps));
> return 0;
> }
> --
> 2.44.0.769.g3c40516874-goog
>

2024-05-02 19:39:35

by Arnaldo Carvalho de Melo

[permalink] [raw]
Subject: Re: [PATCH] perf maps: Remove check_invariants() from maps__lock()

On Mon, Apr 29, 2024 at 07:09:54PM -0700, Ian Rogers wrote:
> On Mon, Apr 29, 2024 at 3:57 PM Namhyung Kim <[email protected]> wrote:
> >
> > I found that the debug build was a lot slowed down by the maps lock code
> > since it checks the invariants whenever it gets the pointer to the lock.
> > This means it checks twice the invariants before and after the access.
> >
> > Instead, let's move the checking code within the lock area but after any
> > modification and remove it from the read paths. This would remove (more
> > than) the half of the maps lock overhead.
> >
> > The time for perf report with a huge data file (200k+ of MMAP2 events).
> >
> > Non-debug Before After
> > --------- -------- --------
> > 2m 43s 6m 45s 4m 21s

Yeah, I had a debug build to check the size of 'struct hist_entry' with
pahole and noticed that invariant checking in 'perf top'.

> Thanks Namhyung, I think the change makes sense. There is an issue
> that a user of a map may mutate it in a way that breaks the maps
> sorting, symbol code would be a likely culprit of such a thing. A fix
> for such a breakage would be to just set the unsorted flag on maps.
> We'd be less likely to spot such an issue after this change, but it is
> simple enough to add the function call when needed. Another option
> would be to add another make flag like PARSER_DEBUG for these kind of,
> "be paranoid," type things. Anyway, let's try to fix those problems
> when they exist.
>
> Reviewed-by: Ian Rogers <[email protected]>

Thanks, applied to perf-tools-next,

- Arnaldo