[PATCH 3/4] perf symbols: Update kcore map before merging in remaining symbols

James Clark posted 4 patches 1 year, 9 months ago
[PATCH 3/4] perf symbols: Update kcore map before merging in remaining symbols
Posted by James Clark 1 year, 9 months ago
When loading kcore, the main vmlinux map is updated in the same loop
that merges the remaining maps. If a map that overlaps is merged in
before kcore, the list can become unsortable when the main map addresses
are updated. This will later trigger the check_invariants() assert:

  $ perf record
  $ perf report

  util/maps.c:96: check_invariants: Assertion `map__end(prev) <=
    map__start(map) || map__start(prev) == map__start(map)' failed.
  Aborted

Fix it by moving the main map update prior to the loop so that
maps__merge_in() can split it if necessary.

Fixes: 659ad3492b91 ("perf maps: Switch from rbtree to lazily sorted array for addresses")
Signed-off-by: James Clark <james.clark@arm.com>
---
 tools/perf/util/symbol.c | 40 +++++++++++++++++++++-------------------
 1 file changed, 21 insertions(+), 19 deletions(-)

diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c
index 2d95f22d713d..e98dfe766da3 100644
--- a/tools/perf/util/symbol.c
+++ b/tools/perf/util/symbol.c
@@ -1289,7 +1289,7 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
 {
 	struct maps *kmaps = map__kmaps(map);
 	struct kcore_mapfn_data md;
-	struct map *replacement_map = NULL;
+	struct map *map_ref, *replacement_map = NULL;
 	struct machine *machine;
 	bool is_64_bit;
 	int err, fd;
@@ -1367,6 +1367,24 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
 	if (!replacement_map)
 		replacement_map = list_entry(md.maps.next, struct map_list_node, node)->map;
 
+	/*
+	 * Update addresses of vmlinux map. Re-insert it to ensure maps are
+	 * correctly ordered. Do this before using maps__merge_in() for the
+	 * remaining maps so vmlinux gets split if necessary.
+	 */
+	map_ref = map__get(map);
+	maps__remove(kmaps, map_ref);
+
+	map__set_start(map_ref, map__start(replacement_map));
+	map__set_end(map_ref, map__end(replacement_map));
+	map__set_pgoff(map_ref, map__pgoff(replacement_map));
+	map__set_mapping_type(map_ref, map__mapping_type(replacement_map));
+
+	err = maps__insert(kmaps, map_ref);
+	map__put(map_ref);
+	if (err)
+		goto out_err;
+
 	/* Add new maps */
 	while (!list_empty(&md.maps)) {
 		struct map_list_node *new_node = list_entry(md.maps.next, struct map_list_node, node);
@@ -1374,24 +1392,8 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
 
 		list_del_init(&new_node->node);
 
-		if (RC_CHK_EQUAL(new_map, replacement_map)) {
-			struct map *map_ref;
-
-			/* Ensure maps are correctly ordered */
-			map_ref = map__get(map);
-			maps__remove(kmaps, map_ref);
-
-			map__set_start(map_ref, map__start(new_map));
-			map__set_end(map_ref, map__end(new_map));
-			map__set_pgoff(map_ref, map__pgoff(new_map));
-			map__set_mapping_type(map_ref, map__mapping_type(new_map));
-
-			err = maps__insert(kmaps, map_ref);
-			map__put(map_ref);
-			map__put(new_map);
-			if (err)
-				goto out_err;
-		} else {
+		/* skip if replacement_map, already inserted above */
+		if (!RC_CHK_EQUAL(new_map, replacement_map)) {
 			/*
 			 * Merge kcore map into existing maps,
 			 * and ensure that current maps (eBPF)
-- 
2.34.1
Re: [PATCH 3/4] perf symbols: Update kcore map before merging in remaining symbols
Posted by Namhyung Kim 1 year, 9 months ago
On Tue, May 7, 2024 at 7:13 AM James Clark <james.clark@arm.com> wrote:
>
> When loading kcore, the main vmlinux map is updated in the same loop
> that merges the remaining maps. If a map that overlaps is merged in
> before kcore, the list can become unsortable when the main map addresses
> are updated. This will later trigger the check_invariants() assert:
>
>   $ perf record
>   $ perf report
>
>   util/maps.c:96: check_invariants: Assertion `map__end(prev) <=
>     map__start(map) || map__start(prev) == map__start(map)' failed.
>   Aborted
>
> Fix it by moving the main map update prior to the loop so that
> maps__merge_in() can split it if necessary.

Looks like you and Leo are working on the same problem.

https://lore.kernel.org/r/20240505202805.583253-1-leo.yan@arm.com/

>
> Fixes: 659ad3492b91 ("perf maps: Switch from rbtree to lazily sorted array for addresses")
> Signed-off-by: James Clark <james.clark@arm.com>
> ---
>  tools/perf/util/symbol.c | 40 +++++++++++++++++++++-------------------
>  1 file changed, 21 insertions(+), 19 deletions(-)
>
> diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c
> index 2d95f22d713d..e98dfe766da3 100644
> --- a/tools/perf/util/symbol.c
> +++ b/tools/perf/util/symbol.c
> @@ -1289,7 +1289,7 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
>  {
>         struct maps *kmaps = map__kmaps(map);
>         struct kcore_mapfn_data md;
> -       struct map *replacement_map = NULL;
> +       struct map *map_ref, *replacement_map = NULL;
>         struct machine *machine;
>         bool is_64_bit;
>         int err, fd;
> @@ -1367,6 +1367,24 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
>         if (!replacement_map)
>                 replacement_map = list_entry(md.maps.next, struct map_list_node, node)->map;
>
> +       /*
> +        * Update addresses of vmlinux map. Re-insert it to ensure maps are
> +        * correctly ordered. Do this before using maps__merge_in() for the
> +        * remaining maps so vmlinux gets split if necessary.
> +        */
> +       map_ref = map__get(map);
> +       maps__remove(kmaps, map_ref);

A nitpick.  It'd be natural to use 'map' instead of 'map_ref'
(even if they are the same) since IIUC we want to remove
the old 'map' and update 'map_ref' then add it back.

> +
> +       map__set_start(map_ref, map__start(replacement_map));
> +       map__set_end(map_ref, map__end(replacement_map));
> +       map__set_pgoff(map_ref, map__pgoff(replacement_map));
> +       map__set_mapping_type(map_ref, map__mapping_type(replacement_map));

So here, replacement_map should not be NULL right?

Thanks,
Namhyung

> +
> +       err = maps__insert(kmaps, map_ref);
> +       map__put(map_ref);
> +       if (err)
> +               goto out_err;
> +
>         /* Add new maps */
>         while (!list_empty(&md.maps)) {
>                 struct map_list_node *new_node = list_entry(md.maps.next, struct map_list_node, node);
> @@ -1374,24 +1392,8 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
>
>                 list_del_init(&new_node->node);
>
> -               if (RC_CHK_EQUAL(new_map, replacement_map)) {
> -                       struct map *map_ref;
> -
> -                       /* Ensure maps are correctly ordered */
> -                       map_ref = map__get(map);
> -                       maps__remove(kmaps, map_ref);
> -
> -                       map__set_start(map_ref, map__start(new_map));
> -                       map__set_end(map_ref, map__end(new_map));
> -                       map__set_pgoff(map_ref, map__pgoff(new_map));
> -                       map__set_mapping_type(map_ref, map__mapping_type(new_map));
> -
> -                       err = maps__insert(kmaps, map_ref);
> -                       map__put(map_ref);
> -                       map__put(new_map);
> -                       if (err)
> -                               goto out_err;
> -               } else {
> +               /* skip if replacement_map, already inserted above */
> +               if (!RC_CHK_EQUAL(new_map, replacement_map)) {
>                         /*
>                          * Merge kcore map into existing maps,
>                          * and ensure that current maps (eBPF)
> --
> 2.34.1
>
Re: [PATCH 3/4] perf symbols: Update kcore map before merging in remaining symbols
Posted by James Clark 1 year, 9 months ago

On 08/05/2024 05:10, Namhyung Kim wrote:
> On Tue, May 7, 2024 at 7:13 AM James Clark <james.clark@arm.com> wrote:
>>
>> When loading kcore, the main vmlinux map is updated in the same loop
>> that merges the remaining maps. If a map that overlaps is merged in
>> before kcore, the list can become unsortable when the main map addresses
>> are updated. This will later trigger the check_invariants() assert:
>>
>>   $ perf record
>>   $ perf report
>>
>>   util/maps.c:96: check_invariants: Assertion `map__end(prev) <=
>>     map__start(map) || map__start(prev) == map__start(map)' failed.
>>   Aborted
>>
>> Fix it by moving the main map update prior to the loop so that
>> maps__merge_in() can split it if necessary.
> 
> Looks like you and Leo are working on the same problem.
> 
> https://lore.kernel.org/r/20240505202805.583253-1-leo.yan@arm.com/
> 

Oops I should have checked the list. It looks like we can still take his
fix as well though, with an updated comment.

>>
>> Fixes: 659ad3492b91 ("perf maps: Switch from rbtree to lazily sorted array for addresses")
>> Signed-off-by: James Clark <james.clark@arm.com>
>> ---
>>  tools/perf/util/symbol.c | 40 +++++++++++++++++++++-------------------
>>  1 file changed, 21 insertions(+), 19 deletions(-)
>>
>> diff --git a/tools/perf/util/symbol.c b/tools/perf/util/symbol.c
>> index 2d95f22d713d..e98dfe766da3 100644
>> --- a/tools/perf/util/symbol.c
>> +++ b/tools/perf/util/symbol.c
>> @@ -1289,7 +1289,7 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
>>  {
>>         struct maps *kmaps = map__kmaps(map);
>>         struct kcore_mapfn_data md;
>> -       struct map *replacement_map = NULL;
>> +       struct map *map_ref, *replacement_map = NULL;
>>         struct machine *machine;
>>         bool is_64_bit;
>>         int err, fd;
>> @@ -1367,6 +1367,24 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
>>         if (!replacement_map)
>>                 replacement_map = list_entry(md.maps.next, struct map_list_node, node)->map;
>>
>> +       /*
>> +        * Update addresses of vmlinux map. Re-insert it to ensure maps are
>> +        * correctly ordered. Do this before using maps__merge_in() for the
>> +        * remaining maps so vmlinux gets split if necessary.
>> +        */
>> +       map_ref = map__get(map);
>> +       maps__remove(kmaps, map_ref);
> 
> A nitpick.  It'd be natural to use 'map' instead of 'map_ref'
> (even if they are the same) since IIUC we want to remove
> the old 'map' and update 'map_ref' then add it back.
> 

Using map makes sense, I can update that.

>> +
>> +       map__set_start(map_ref, map__start(replacement_map));
>> +       map__set_end(map_ref, map__end(replacement_map));
>> +       map__set_pgoff(map_ref, map__pgoff(replacement_map));
>> +       map__set_mapping_type(map_ref, map__mapping_type(replacement_map));
> 
> So here, replacement_map should not be NULL right?
> 

Yes it shouldn't be. It would only be NULL if md.maps is empty, but
there's already an exit condition for that above.

Some of the other code also assumes node->map is always set, so it can't
be NULL that way either.

> Thanks,
> Namhyung
> 
>> +
>> +       err = maps__insert(kmaps, map_ref);
>> +       map__put(map_ref);
>> +       if (err)
>> +               goto out_err;
>> +
>>         /* Add new maps */
>>         while (!list_empty(&md.maps)) {
>>                 struct map_list_node *new_node = list_entry(md.maps.next, struct map_list_node, node);
>> @@ -1374,24 +1392,8 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
>>
>>                 list_del_init(&new_node->node);
>>
>> -               if (RC_CHK_EQUAL(new_map, replacement_map)) {
>> -                       struct map *map_ref;
>> -
>> -                       /* Ensure maps are correctly ordered */
>> -                       map_ref = map__get(map);
>> -                       maps__remove(kmaps, map_ref);
>> -
>> -                       map__set_start(map_ref, map__start(new_map));
>> -                       map__set_end(map_ref, map__end(new_map));
>> -                       map__set_pgoff(map_ref, map__pgoff(new_map));
>> -                       map__set_mapping_type(map_ref, map__mapping_type(new_map));
>> -
>> -                       err = maps__insert(kmaps, map_ref);
>> -                       map__put(map_ref);
>> -                       map__put(new_map);
>> -                       if (err)
>> -                               goto out_err;
>> -               } else {
>> +               /* skip if replacement_map, already inserted above */
>> +               if (!RC_CHK_EQUAL(new_map, replacement_map)) {
>>                         /*
>>                          * Merge kcore map into existing maps,
>>                          * and ensure that current maps (eBPF)
>> --
>> 2.34.1
>>
Re: [PATCH 3/4] perf symbols: Update kcore map before merging in remaining symbols
Posted by Leo Yan 1 year, 9 months ago
On 5/8/2024 10:14 AM, James Clark wrote:

[...]

>> Looks like you and Leo are working on the same problem.
>>
>> https://lore.kernel.org/r/20240505202805.583253-1-leo.yan@arm.com/
>>
> 
> Oops I should have checked the list. It looks like we can still take his
> fix as well though, with an updated comment.

Sorry for duplicate work. I will resend my patch separately with refined
comment, as suggested by Adrian.

[...]

>>> @@ -1289,7 +1289,7 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
>>>  {
>>>         struct maps *kmaps = map__kmaps(map);
>>>         struct kcore_mapfn_data md;
>>> -       struct map *replacement_map = NULL;
>>> +       struct map *map_ref, *replacement_map = NULL;
>>>         struct machine *machine;
>>>         bool is_64_bit;
>>>         int err, fd;
>>> @@ -1367,6 +1367,24 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
>>>         if (!replacement_map)
>>>                 replacement_map = list_entry(md.maps.next, struct map_list_node, node)->map;

As the 'replacement' map is mainly used to adjust the kernel's sections
between '_stext' and '_end', some arches might don't share the same issue with
Arm64.  So it is a bit redundant for assignment 'replacement_map' if it is
NULL, we can consider to remove the above two lines.

>>>
>>> +       /*
>>> +        * Update addresses of vmlinux map. Re-insert it to ensure maps are
>>> +        * correctly ordered. Do this before using maps__merge_in() for the
>>> +        * remaining maps so vmlinux gets split if necessary.
>>> +        */
>>> +       map_ref = map__get(map);
>>> +       maps__remove(kmaps, map_ref);
>>
>> A nitpick.  It'd be natural to use 'map' instead of 'map_ref'
>> (even if they are the same) since IIUC we want to remove
>> the old 'map' and update 'map_ref' then add it back.
>>
> 
> Using map makes sense, I can update that.
> 
>>> +
>>> +       map__set_start(map_ref, map__start(replacement_map));
>>> +       map__set_end(map_ref, map__end(replacement_map));
>>> +       map__set_pgoff(map_ref, map__pgoff(replacement_map));
>>> +       map__set_mapping_type(map_ref, map__mapping_type(replacement_map));
>>
>> So here, replacement_map should not be NULL right?
>>
> 
> Yes it shouldn't be. It would only be NULL if md.maps is empty, but
> there's already an exit condition for that above.
> 
> Some of the other code also assumes node->map is always set, so it can't
> be NULL that way either.

Thus, we can consider to check condition for 'replacement' map is NULL or not.

    if (replacement_map) {
	list_del_init(&new_node->node);

	map_ref = map__get(map);
	maps__remove(kmaps, map_ref);
	...
	map__put(new_map);
	if (err)
		goto out_err;
	free(new_node);
    }

[...]

>>> @@ -1374,24 +1392,8 @@ static int dso__load_kcore(struct dso *dso, struct map *map,
>>>
>>>                 list_del_init(&new_node->node);
>>>
>>> -               if (RC_CHK_EQUAL(new_map, replacement_map)) {
>>> -                       struct map *map_ref;
>>> -
>>> -                       /* Ensure maps are correctly ordered */
>>> -                       map_ref = map__get(map);
>>> -                       maps__remove(kmaps, map_ref);
>>> -
>>> -                       map__set_start(map_ref, map__start(new_map));
>>> -                       map__set_end(map_ref, map__end(new_map));
>>> -                       map__set_pgoff(map_ref, map__pgoff(new_map));
>>> -                       map__set_mapping_type(map_ref, map__mapping_type(new_map));
>>> -
>>> -                       err = maps__insert(kmaps, map_ref);
>>> -                       map__put(map_ref);
>>> -                       map__put(new_map);
>>> -                       if (err)
>>> -                               goto out_err;
>>> -               } else {
>>> +               /* skip if replacement_map, already inserted above */
>>> +               if (!RC_CHK_EQUAL(new_map, replacement_map)) {

With above change, we don't need check 'replacement_map' at here.

Just extend a bit for considering a more clean fixing, we need to sort all
ranges in 'md.maps', this would be benefit for two things:

- We can fix up any map regions, not only limit to the 'replacement_map'. With
  sorting maps in 'md.maps', we can totally remove the code for
  'replacement_map'.
- We can report the potential issue caused by overlapping in the first place
  rather than the assert log in check_invariants(). This is easier for later
  debugging.

But current patch is good enough for me, I don't have strong opinion for this.

Thanks,
Leo

>>>                         /*
>>>                          * Merge kcore map into existing maps,
>>>                          * and ensure that current maps (eBPF)
>>> --
>>> 2.34.1
>>>