Skip to contentSkip to conversation

Bail out in __stack_depot_save() if the stack_table is not allocated and delete the kmemleak_initialized judgment in set_track_prepare()

PATCH In review Xiaolei Wang 3 participants 11 messages Aug 10, 2023 – Aug 15, 2023

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover Bail out in __stack_depot_save() if the stack_table is not allocated and delete the kmemleak_initialized judgment in set_track_prepare()

patch1 solves the null pointer situation when __stack_depot_save() is called when the stack_table is not initialized.

patch2 solved there is no call trace for the memory leak which object is created before the kmemleak_late_init()

Xiaolei Wang (2):
lib/stackdepot: Bail out in __stack_depot_save() if the stack_table is
not allocated
mm/kmemleak: No need to check kmemleak_initialized in
set_track_prepare()

 lib/stackdepot.c | 2 +-
 mm/kmemleak.c    | 2 --
 2 files changed, 1 insertion(+), 3 deletions(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH 1/2 lib/stackdepot: Bail out in __stack_depot_save() if the stack_table is not allocated

The __stack_depot_save() may be used by some subsystems even before the stack depot is initialized. So add a check of stack_table in __stack_depot_save() to make sure no oops in this case.

Fixes 56a61617dd22 ("mm: use stack_depot for recording kmemleak's backtrace")
---
 lib/stackdepot.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
lib/stackdepot.c +1 -1
index 2f5aa851834e..a0651d013a0d 100644
@@ -376,7 +376,7 @@ depot_stack_handle_t __stack_depot_save(unsigned long *entries,
 	 */
 	nr_entries = filter_irq_stacks(entries, nr_entries);
 
-	if (unlikely(nr_entries == 0) || stack_depot_disabled)
+	if (unlikely(nr_entries == 0) || stack_depot_disabled || unlikely(!stack_table))
 		goto fast_exit;
 
 	hash = hash_stack(entries, nr_entries);
Signature · 1 line
2.25.1
#2 lore raw

Re: [PATCH 1/2] lib/stackdepot: Bail out in __stack_depot_save() if the stack_table is not allocated

Quoted 2 lines · Xiaolei Wang

The __stack_depot_save() may be used by some subsystems even before the stack depot is initialized.

Does that currently happen, or only after patch 2/2 it starts happening via kmemleak?

Quoted 4 lines

So add a check of stack_table in __stack_depot_save() to make sure no oops in this case.

Fixes 56a61617dd22 ("mm: use stack_depot for recording kmemleak's backtrace")

In case it's only after 2/2 I don't think this is truly "Fixes"?

Quoted 18 lines
---
 lib/stackdepot.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
lib/stackdepot.c +1 -1
index 2f5aa851834e..a0651d013a0d 100644
@@ -376,7 +376,7 @@ depot_stack_handle_t __stack_depot_save(unsigned long *entries,
 	 */
 	nr_entries = filter_irq_stacks(entries, nr_entries);
 
-	if (unlikely(nr_entries == 0) || stack_depot_disabled)
+	if (unlikely(nr_entries == 0) || stack_depot_disabled || unlikely(!stack_table))
 		goto fast_exit;
 
 	hash = hash_stack(entries, nr_entries);
#3 lore raw

Re: [PATCH 1/2] lib/stackdepot: Bail out in __stack_depot_save() if the stack_table is not allocated

Quoted 8 lines · Vlastimil Babka

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

Quoted 2 lines · Xiaolei Wang

The __stack_depot_save() may be used by some subsystems even before the stack depot is initialized.

Does that currently happen, or only after patch 2/2 it starts happening via kmemleak?

Yes, currently it happens after patch 2/2 it starts happening via kmemleak,

The reason why I take it as the first patch is because I think this can avoid

exceptions when there is only patch2, such as when we use git bisect to debug

Quoted 6 lines
Quoted 5 lines

So add a check of stack_table in __stack_depot_save() to make sure no oops in this case.

Fixes 56a61617dd22 ("mm: use stack_depot for recording kmemleak's backtrace")

In case it's only after 2/2 I don't think this is truly "Fixes"?

Yes, it is indeed for patch2 at present, it is to fix the situation that kmemleak

has no backtrace in the boot phase, and at the same time, it can also prevent

patch2 from entering the stable kernel and causing panic

thanks

xiaolei

Quoted 19 lines
Quoted 19 lines
---
  lib/stackdepot.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
lib/stackdepot.c +1 -1
index 2f5aa851834e..a0651d013a0d 100644
@@ -376,7 +376,7 @@ depot_stack_handle_t __stack_depot_save(unsigned long *entries,
        */
       nr_entries = filter_irq_stacks(entries, nr_entries);
 
-     if (unlikely(nr_entries == 0) || stack_depot_disabled)
+     if (unlikely(nr_entries == 0) || stack_depot_disabled || unlikely(!stack_table))
               goto fast_exit;
 
       hash = hash_stack(entries, nr_entries);
#4 lore raw

PATCH 2/2 mm/kmemleak: No need to check kmemleak_initialized in set_track_prepare()

In reply to Xiaolei Wang

The kmemleak_late_init() is defined as a late_initcall. The current implementation of set_track_prepare() depends on the kmemleak init. That also means there is no call trace for the memory leak which object is created before the kmemleak_late_init().

In a previous patch, we have fixed a bug in stack_depot_save() so that it can be invoked even before stack depot is initialized. So there is no reason to check the kmemleak_initialized in set_track_prepare(). So delete the kmemleak_initialized judgment in set_track_prepare()

unreferenced object 0xc674ca80 (size 64):
comm "swapper/0", pid 1, jiffies 4294938337 (age 204.880s) hex dump (first 32 bytes):

    80 55 75 c6 80 54 75 c6 00 55 75 c6 80 52 75 c6 .Uu..Tu..Uu..Ru.
    00 53 75 c6 00 00 00 00 00 00 00 00 00 00 00 00 .Su..........
Fixes 56a61617dd22 ("mm: use stack_depot for recording kmemleak's backtrace")
---
 mm/kmemleak.c | 2 --
 1 file changed, 2 deletions(-)
mm/kmemleak.c +0 -2
index a2d34226e3c8..c9f2f816db19 100644
@@ -610,8 +610,6 @@ static noinline depot_stack_handle_t set_track_prepare(void)
 	unsigned long entries[MAX_TRACE];
 	unsigned int nr_entries;
 
-	if (!kmemleak_initialized)
-		return 0;
 	nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
 	trace_handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
 
Signature · 1 line
2.25.1
#5 lore raw

Re: [PATCH 2/2] mm/kmemleak: No need to check kmemleak_initialized in set_track_prepare()

Quoted 4 lines · Xiaolei Wang

The kmemleak_late_init() is defined as a late_initcall. The current implementation of set_track_prepare() depends on the kmemleak init. That also means there is no call trace for the memory leak which object is created before the kmemleak_late_init().

So if I understand correctly, we have the following sequence of events durin boot

...
A: stack_depot is initialized
...
B: kmemleak is initialized
...

before this patchset, we can miss allocations before B, aftewards only before A (which can't be helped), so we now have between A and B.

That's nice, but it's weird that can record kmemleak when !kmemleak_initialized. Why can't it be initialized sooner in that case?

Quoted 30 lines

In a previous patch, we have fixed a bug in stack_depot_save() so that it can be invoked even before stack depot is initialized. So there is no reason to check the kmemleak_initialized in set_track_prepare(). So delete the kmemleak_initialized judgment in set_track_prepare()

unreferenced object 0xc674ca80 (size 64):
comm "swapper/0", pid 1, jiffies 4294938337 (age 204.880s) hex dump (first 32 bytes):

    80 55 75 c6 80 54 75 c6 00 55 75 c6 80 52 75 c6 .Uu..Tu..Uu..Ru.
    00 53 75 c6 00 00 00 00 00 00 00 00 00 00 00 00 .Su..........
Fixes 56a61617dd22 ("mm: use stack_depot for recording kmemleak's backtrace")
---
 mm/kmemleak.c | 2 --
 1 file changed, 2 deletions(-)
mm/kmemleak.c +0 -2
index a2d34226e3c8..c9f2f816db19 100644
@@ -610,8 +610,6 @@ static noinline depot_stack_handle_t set_track_prepare(void)
 	unsigned long entries[MAX_TRACE];
 	unsigned int nr_entries;
 
-	if (!kmemleak_initialized)
-		return 0;
 	nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
 	trace_handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
 
#6 lore raw

Re: [PATCH 2/2] mm/kmemleak: No need to check kmemleak_initialized in set_track_prepare()

Quoted 15 lines · Vlastimil Babka
Quoted 4 lines · Xiaolei Wang

The kmemleak_late_init() is defined as a late_initcall. The current implementation of set_track_prepare() depends on the kmemleak init. That also means there is no call trace for the memory leak which object is created before the kmemleak_late_init().

So if I understand correctly, we have the following sequence of events durin boot

...
A: stack_depot is initialized
...
B: kmemleak is initialized
...

Quoted 5 lines

before this patchset, we can miss allocations before B, aftewards only before A (which can't be helped), so we now have between A and B.

That's nice, but it's weird that can record kmemleak when !kmemleak_initialized. Why can't it be initialized sooner in that case?

Looking closer, I think what you want could be achieved by kmemleak_init() setting a variable that is checked in kmemleak_initialized() instead of the kmemleak_initialized that's set too late.

I think this should work because:
- I assume kmemleak can't record anything before kmemleak_init()
- stack depot early init is requested one way or the other
- mm_core_init() calls stack_depot_early_init() before kmemleak_init()

But I also wonder how kmemleak can even reach set_track_prepare() before kmemleak_init(), maybe that's the issue?

Quoted 31 lines
Quoted 30 lines

In a previous patch, we have fixed a bug in stack_depot_save() so that it can be invoked even before stack depot is initialized. So there is no reason to check the kmemleak_initialized in set_track_prepare(). So delete the kmemleak_initialized judgment in set_track_prepare()

unreferenced object 0xc674ca80 (size 64):
comm "swapper/0", pid 1, jiffies 4294938337 (age 204.880s) hex dump (first 32 bytes):

    80 55 75 c6 80 54 75 c6 00 55 75 c6 80 52 75 c6 .Uu..Tu..Uu..Ru.
    00 53 75 c6 00 00 00 00 00 00 00 00 00 00 00 00 .Su..........
Fixes 56a61617dd22 ("mm: use stack_depot for recording kmemleak's backtrace")
---
 mm/kmemleak.c | 2 --
 1 file changed, 2 deletions(-)
mm/kmemleak.c +0 -2
index a2d34226e3c8..c9f2f816db19 100644
@@ -610,8 +610,6 @@ static noinline depot_stack_handle_t set_track_prepare(void)
 	unsigned long entries[MAX_TRACE];
 	unsigned int nr_entries;
 
-	if (!kmemleak_initialized)
-		return 0;
 	nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
 	trace_handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
 
#7 lore raw

Re: [PATCH 2/2] mm/kmemleak: No need to check kmemleak_initialized in set_track_prepare()

Quoted 27 lines · Vlastimil Babka

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

Quoted 14 lines · Vlastimil Babka
Quoted 4 lines · Xiaolei Wang

The kmemleak_late_init() is defined as a late_initcall. The current implementation of set_track_prepare() depends on the kmemleak init. That also means there is no call trace for the memory leak which object is created before the kmemleak_late_init().

So if I understand correctly, we have the following sequence of events durin boot

...
A: stack_depot is initialized
...
B: kmemleak is initialized
...

Quoted 5 lines

before this patchset, we can miss allocations before B, aftewards only before A (which can't be helped), so we now have between A and B.

That's nice, but it's weird that can record kmemleak when !kmemleak_initialized. Why can't it be initialized sooner in that case?

Looking closer, I think what you want could be achieved by kmemleak_init() setting a variable that is checked in kmemleak_initialized() instead of the kmemleak_initialized that's set too late.

Quoted 7 lines

I think this should work because:
- I assume kmemleak can't record anything before kmemleak_init()
- stack depot early init is requested one way or the other
- mm_core_init() calls stack_depot_early_init() before kmemleak_init()

But I also wonder how kmemleak can even reach set_track_prepare() before kmemleak_init(), maybe that's the issue?

Before kmemleak_init, many places also need to allocate kmemleak_object,

and also need to save stack in advance, but kmemleak_object is allocated

in the form of an array, after kmemleak_init 'object_cache = KMEM_CACHE(kmemleak_object, SLAB_NOLEAKTRACE);'

I think there is still some memory not recorded on the backtrace before

stack_depot_early_init(), does anyone have a better suggestion?

thanks

xiaolei

Quoted 31 lines
Quoted 31 lines
Quoted 31 lines

In a previous patch, we have fixed a bug in stack_depot_save() so that it can be invoked even before stack depot is initialized. So there is no reason to check the kmemleak_initialized in set_track_prepare(). So delete the kmemleak_initialized judgment in set_track_prepare()

unreferenced object 0xc674ca80 (size 64):

   comm "swapper/0", pid 1, jiffies 4294938337 (age 204.880s)
   hex dump (first 32 bytes):
     80 55 75 c6 80 54 75 c6 00 55 75 c6 80 52 75 c6 .Uu..Tu..Uu..Ru.
     00 53 75 c6 00 00 00 00 00 00 00 00 00 00 00 00 .Su..........
Fixes 56a61617dd22 ("mm: use stack_depot for recording kmemleak's backtrace")
---
  mm/kmemleak.c | 2 --
  1 file changed, 2 deletions(-)
mm/kmemleak.c +0 -2
index a2d34226e3c8..c9f2f816db19 100644
@@ -610,8 +610,6 @@ static noinline depot_stack_handle_t set_track_prepare(void)
      unsigned long entries[MAX_TRACE];
      unsigned int nr_entries;
 
-    if (!kmemleak_initialized)
-            return 0;
      nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
      trace_handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
 
#8 lore raw

Re: [PATCH 2/2] mm/kmemleak: No need to check kmemleak_initialized in set_track_prepare()

Quoted 37 lines · wang xiaolei
Quoted 27 lines · Vlastimil Babka

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

Quoted 14 lines · Vlastimil Babka
Quoted 4 lines · Xiaolei Wang

The kmemleak_late_init() is defined as a late_initcall. The current implementation of set_track_prepare() depends on the kmemleak init. That also means there is no call trace for the memory leak which object is created before the kmemleak_late_init().

So if I understand correctly, we have the following sequence of events durin boot

...
A: stack_depot is initialized
...
B: kmemleak is initialized
...

Quoted 5 lines

before this patchset, we can miss allocations before B, aftewards only before A (which can't be helped), so we now have between A and B.

That's nice, but it's weird that can record kmemleak when !kmemleak_initialized. Why can't it be initialized sooner in that case?

Looking closer, I think what you want could be achieved by kmemleak_init() setting a variable that is checked in kmemleak_initialized() instead of the kmemleak_initialized that's set too late.

Quoted 7 lines

I think this should work because:
- I assume kmemleak can't record anything before kmemleak_init()
- stack depot early init is requested one way or the other
- mm_core_init() calls stack_depot_early_init() before kmemleak_init()

But I also wonder how kmemleak can even reach set_track_prepare() before kmemleak_init(), maybe that's the issue?

Quoted 6 lines

Before kmemleak_init, many places also need to allocate kmemleak_object,

and also need to save stack in advance, but kmemleak_object is allocated

in the form of an array, after kmemleak_init 'object_cache = KMEM_CACHE(kmemleak_object, SLAB_NOLEAKTRACE);'

Hm I see, kmemleak has this static mempool so it really can record some objects very early.

Quoted 3 lines

I think there is still some memory not recorded on the backtrace before

stack_depot_early_init(), does anyone have a better suggestion?

No we can't record the backtrace earlier. But I don't think it's a problem in practice. AFAIU kmemleak needs to record these very early allocations so if they point to further objects, those are not suspected as orphans. But the early allocations themselves also are very unlikely to be leaks, so does it really matter that we don't have a backtrace for their allocation? Because the backtrace is the only thing that's missing - the object is otherwise recorded even if set_track_prepare() returns 0.

Quoted 35 lines

thanks

xiaolei

Quoted 31 lines
Quoted 31 lines
Quoted 31 lines

In a previous patch, we have fixed a bug in stack_depot_save() so that it can be invoked even before stack depot is initialized. So there is no reason to check the kmemleak_initialized in set_track_prepare(). So delete the kmemleak_initialized judgment in set_track_prepare()

unreferenced object 0xc674ca80 (size 64):

   comm "swapper/0", pid 1, jiffies 4294938337 (age 204.880s)
   hex dump (first 32 bytes):
     80 55 75 c6 80 54 75 c6 00 55 75 c6 80 52 75 c6 .Uu..Tu..Uu..Ru.
     00 53 75 c6 00 00 00 00 00 00 00 00 00 00 00 00 .Su..........
Fixes 56a61617dd22 ("mm: use stack_depot for recording kmemleak's backtrace")
---
  mm/kmemleak.c | 2 --
  1 file changed, 2 deletions(-)
mm/kmemleak.c +0 -2
index a2d34226e3c8..c9f2f816db19 100644
@@ -610,8 +610,6 @@ static noinline depot_stack_handle_t set_track_prepare(void)
      unsigned long entries[MAX_TRACE];
      unsigned int nr_entries;
 
-    if (!kmemleak_initialized)
-            return 0;
      nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
      trace_handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
 
maintainer
#9 lore raw

Re: [PATCH 2/2] mm/kmemleak: No need to check kmemleak_initialized in set_track_prepare()

Quoted 16 lines · Vlastimil Babka
Quoted 15 lines · wang xiaolei
Quoted 11 lines · Vlastimil Babka

Looking closer, I think what you want could be achieved by kmemleak_init() setting a variable that is checked in kmemleak_initialized() instead of the kmemleak_initialized that's set too late.

I think this should work because:
- I assume kmemleak can't record anything before kmemleak_init()
- stack depot early init is requested one way or the other
- mm_core_init() calls stack_depot_early_init() before kmemleak_init()

But I also wonder how kmemleak can even reach set_track_prepare() before kmemleak_init(), maybe that's the issue?

Before kmemleak_init, many places also need to allocate kmemleak_object,

Quoted 7 lines
Quoted 4 lines

and also need to save stack in advance, but kmemleak_object is allocated

in the form of an array, after kmemleak_init 'object_cache = KMEM_CACHE(kmemleak_object, SLAB_NOLEAKTRACE);'

Hm I see, kmemleak has this static mempool so it really can record some objects very early.

Indeed, otherwise we'd get a lot of false positives.

Quoted 11 lines
Quoted 3 lines

I think there is still some memory not recorded on the backtrace before

stack_depot_early_init(), does anyone have a better suggestion?

No we can't record the backtrace earlier. But I don't think it's a problem in practice. AFAIU kmemleak needs to record these very early allocations so if they point to further objects, those are not suspected as orphans. But the early allocations themselves also are very unlikely to be leaks, so does it really matter that we don't have a backtrace for their allocation? Because the backtrace is the only thing that's missing - the object is otherwise recorded even if set_track_prepare() returns 0.

It's not a functional problem, just a reporting one. There are rare early leaks (usually false positives) so identifying them would help. That said, I think set_track_prepare() is too conservative in waiting for kmemleak_initialized to be set in kmemleak_late_init(). That's a late_initcall() meant for the scanning thread etc. not the core kmemleak functionality (which is on from early boot).

We can instead use a different variable to check in set_track_prepare(), e.g. object_cache. stack_depot_early_init() is called prior to kmemleak_init(), so it should be fine.

If "kmemleak_initialized" is confusing, we could rename it to "kmemleak_late_initialized" or "kmemleak_fully_initialized". I'm not too fussed about this as long as we add some comment on why we check object_cache instead of kmemleak_initialized.

Signature · 1 line
Catalin
#10 lore raw

Re: [PATCH 2/2] mm/kmemleak: No need to check kmemleak_initialized in set_track_prepare()

Quoted 44 lines · Catalin Marinas

CAUTION: This email comes from a non Wind River email account!
Do not click links or open attachments unless you recognize the sender and know the content is safe.

Quoted 13 lines · Vlastimil Babka
Quoted 12 lines · wang xiaolei
Quoted 11 lines · Vlastimil Babka

Looking closer, I think what you want could be achieved by kmemleak_init() setting a variable that is checked in kmemleak_initialized() instead of the kmemleak_initialized that's set too late.

I think this should work because:
- I assume kmemleak can't record anything before kmemleak_init()
- stack depot early init is requested one way or the other
- mm_core_init() calls stack_depot_early_init() before kmemleak_init()

But I also wonder how kmemleak can even reach set_track_prepare() before kmemleak_init(), maybe that's the issue?

Quoted 8 lines
Quoted 6 lines

Before kmemleak_init, many places also need to allocate kmemleak_object,

and also need to save stack in advance, but kmemleak_object is allocated

in the form of an array, after kmemleak_init 'object_cache = KMEM_CACHE(kmemleak_object, SLAB_NOLEAKTRACE);'

Hm I see, kmemleak has this static mempool so it really can record some objects very early.

Indeed, otherwise we'd get a lot of false positives.

Quoted 10 lines
Quoted 3 lines

I think there is still some memory not recorded on the backtrace before

stack_depot_early_init(), does anyone have a better suggestion?

No we can't record the backtrace earlier. But I don't think it's a problem in practice. AFAIU kmemleak needs to record these very early allocations so if they point to further objects, those are not suspected as orphans. But the early allocations themselves also are very unlikely to be leaks, so does it really matter that we don't have a backtrace for their allocation? Because the backtrace is the only thing that's missing - the object is otherwise recorded even if set_track_prepare() returns 0.

It's not a functional problem, just a reporting one. There are rare early leaks (usually false positives) so identifying them would help. That said, I think set_track_prepare() is too conservative in waiting for kmemleak_initialized to be set in kmemleak_late_init(). That's a late_initcall() meant for the scanning thread etc. not the core kmemleak functionality (which is on from early boot).

Quoted 8 lines

We can instead use a different variable to check in set_track_prepare(), e.g. object_cache. stack_depot_early_init() is called prior to kmemleak_init(), so it should be fine.

If "kmemleak_initialized" is confusing, we could rename it to "kmemleak_late_initialized" or "kmemleak_fully_initialized". I'm not too fussed about this as long as we add some comment on why we check object_cache instead of kmemleak_initialized.

Ok, I will send v2 version, use object_cache instead of kmemleak_initialized

to check in set_track_prepare, and update kmemleak_initialized to kmemleak_late_initialized

thanks

xiaolei

Quoted 3 lines

--
Catalin