[1/2] gdb: call frame unwinders' dealloc_cache methods through destroying the frame cache

  Currently, some frame resources are deallocated by iterating on the
frame chain (starting from the sentinel), calling dealloc_cache.  The
problem is that user-created frames are not part of that chain, so we
never call dealloc_cache for them.

I propose to make it so the dealloc_cache callbacks are called when the
frames are removed from the frame_stash hash table, by registering a
deletion function to the hash table.  This happens when
frame_stash_invalidate is called by reinit_frame_cache.  This way, all
frames registered in the cache will get their unwinder's dealloc_cache
callbacks called.

Note that at the moment, the sentinel frames are not registered in the
cache, so we won't call dealloc_cache for them.  However, it's just a
theoritical problem, because the sentinel frame unwinder does not
provide this callback.  Also, a subsequent patch will change things so
that sentinel frames are registered to the cache.

I moved the obstack_free / obstack_init pair below the
frame_stash_invalidate call in reinit_frame_cache, because I assumed
that some dealloc_cache would need to access some data on that obstack,
so it would be better to free it after clearing the hash table.

Change-Id: If4f9b38266b458c4e2f7eb43e933090177c22190
---
 gdb/frame.c | 39 ++++++++++++++++++++++++---------------
 1 file changed, 24 insertions(+), 15 deletions(-)
  

On 1/30/23 21:02, Simon Marchi via Gdb-patches wrote:
> Currently, some frame resources are deallocated by iterating on the
> frame chain (starting from the sentinel), calling dealloc_cache.  The
> problem is that user-created frames are not part of that chain, so we
> never call dealloc_cache for them.
> 
> I propose to make it so the dealloc_cache callbacks are called when the
> frames are removed from the frame_stash hash table, by registering a
> deletion function to the hash table.  This happens when
> frame_stash_invalidate is called by reinit_frame_cache.  This way, all
> frames registered in the cache will get their unwinder's dealloc_cache
> callbacks called.
> 
> Note that at the moment, the sentinel frames are not registered in the
> cache, so we won't call dealloc_cache for them.  However, it's just a
> theoritical problem, because the sentinel frame unwinder does not
> provide this callback.  Also, a subsequent patch will change things so
> that sentinel frames are registered to the cache.
> 
> I moved the obstack_free / obstack_init pair below the
> frame_stash_invalidate call in reinit_frame_cache, because I assumed
> that some dealloc_cache would need to access some data on that obstack,
> so it would be better to free it after clearing the hash table.
> 

For me this causes:
...
(gdb) PASS: gdb.btrace/record_goto.exp: instruction-history from 19 forwards
record goto 27^M
/data/vries/gdb/src/gdb/record-btrace.c:1654: internal-error: 
bfcache_new: Assertion `*slot == NULL' failed.^M
A problem internal to GDB has been detected,^M
further debugging may prove unreliable.^M
----- Backtrace -----^M
FAIL: gdb.btrace/record_goto.exp: record goto 27 (GDB internal error)
...

Note that I've been having some problems with btrace tests, possible 
related to cpu/kernel combination (PRs 30073 and 30075), so this may be 
difficult to reproduce, I'm not sure.

Thanks,
- Tom

> Change-Id: If4f9b38266b458c4e2f7eb43e933090177c22190
> ---
>   gdb/frame.c | 39 ++++++++++++++++++++++++---------------
>   1 file changed, 24 insertions(+), 15 deletions(-)
> 
> diff --git a/gdb/frame.c b/gdb/frame.c
> index a08a8f47ebc4..fed961b2a8df 100644
> --- a/gdb/frame.c
> +++ b/gdb/frame.c
> @@ -259,6 +259,22 @@ frame_addr_hash_eq (const void *a, const void *b)
>     return f_entry->this_id.value == f_element->this_id.value;
>   }
>   
> +/* Deletion function for the frame cache hash table.  */
> +
> +static void
> +frame_info_del (void *frame_v)
> +{
> +  frame_info *frame = (frame_info *) frame_v;
> +
> +  if (frame->prologue_cache != nullptr
> +      && frame->unwind->dealloc_cache != nullptr)
> +    frame->unwind->dealloc_cache (frame, frame->prologue_cache);
> +
> +  if (frame->base_cache != nullptr
> +      && frame->base->unwind->dealloc_cache != nullptr)
> +    frame->base->unwind->dealloc_cache (frame, frame->base_cache);
> +}
> +
>   /* Internal function to create the frame_stash hash table.  100 seems
>      to be a good compromise to start the hash table at.  */
>   
> @@ -268,7 +284,7 @@ frame_stash_create (void)
>     frame_stash = htab_create (100,
>   			     frame_addr_hash,
>   			     frame_addr_hash_eq,
> -			     NULL);
> +			     frame_info_del);
>   }
>   
>   /* Internal function to add a frame to the frame_stash hash table.
> @@ -2048,26 +2064,19 @@ reinit_frame_cache (void)
>   {
>     ++frame_cache_generation;
>   
> -  /* Tear down all frame caches.  */
> -  for (frame_info *fi = sentinel_frame; fi != NULL; fi = fi->prev)
> -    {
> -      if (fi->prologue_cache && fi->unwind->dealloc_cache)
> -	fi->unwind->dealloc_cache (fi, fi->prologue_cache);
> -      if (fi->base_cache && fi->base->unwind->dealloc_cache)
> -	fi->base->unwind->dealloc_cache (fi, fi->base_cache);
> -    }
> -
> -  /* Since we can't really be sure what the first object allocated was.  */
> -  obstack_free (&frame_cache_obstack, 0);
> -  obstack_init (&frame_cache_obstack);
> -
>     if (sentinel_frame != NULL)
>       annotate_frames_invalid ();
>   
> -  sentinel_frame = NULL;		/* Invalidate cache */
>     invalidate_selected_frame ();
> +
> +  /* Invalidate cache.  */
> +  sentinel_frame = NULL;
>     frame_stash_invalidate ();
>   
> +  /* Since we can't really be sure what the first object allocated was.  */
> +  obstack_free (&frame_cache_obstack, 0);
> +  obstack_init (&frame_cache_obstack);
> +
>     for (frame_info_ptr &iter : frame_info_ptr::frame_list)
>       iter.invalidate ();
>
  

On 2/9/23 08:40, Tom de Vries wrote:
> On 1/30/23 21:02, Simon Marchi via Gdb-patches wrote:
>> Currently, some frame resources are deallocated by iterating on the
>> frame chain (starting from the sentinel), calling dealloc_cache.  The
>> problem is that user-created frames are not part of that chain, so we
>> never call dealloc_cache for them.
>>
>> I propose to make it so the dealloc_cache callbacks are called when the
>> frames are removed from the frame_stash hash table, by registering a
>> deletion function to the hash table.  This happens when
>> frame_stash_invalidate is called by reinit_frame_cache.  This way, all
>> frames registered in the cache will get their unwinder's dealloc_cache
>> callbacks called.
>>
>> Note that at the moment, the sentinel frames are not registered in the
>> cache, so we won't call dealloc_cache for them.  However, it's just a
>> theoritical problem, because the sentinel frame unwinder does not
>> provide this callback.  Also, a subsequent patch will change things so
>> that sentinel frames are registered to the cache.
>>
>> I moved the obstack_free / obstack_init pair below the
>> frame_stash_invalidate call in reinit_frame_cache, because I assumed
>> that some dealloc_cache would need to access some data on that obstack,
>> so it would be better to free it after clearing the hash table.
>>
> 
> For me this causes:
> ...
> (gdb) PASS: gdb.btrace/record_goto.exp: instruction-history from 19 
> forwards
> record goto 27^M
> /data/vries/gdb/src/gdb/record-btrace.c:1654: internal-error: 
> bfcache_new: Assertion `*slot == NULL' failed.^M
> A problem internal to GDB has been detected,^M
> further debugging may prove unreliable.^M
> ----- Backtrace -----^M
> FAIL: gdb.btrace/record_goto.exp: record goto 27 (GDB internal error)
> ...
> 
> Note that I've been having some problems with btrace tests, possible 
> related to cpu/kernel combination (PRs 30073 and 30075), so this may be 
> difficult to reproduce, I'm not sure.
> 

I also managed to reproduce this on openSUSE Tumbleweed, which doesn't 
show the problems with btrace tests, so I'm hoping this is easy to 
reproduce.

Thanks,
- Tom

> Thanks,
> - Tom
> 
>> Change-Id: If4f9b38266b458c4e2f7eb43e933090177c22190
>> ---
>>   gdb/frame.c | 39 ++++++++++++++++++++++++---------------
>>   1 file changed, 24 insertions(+), 15 deletions(-)
>>
>> diff --git a/gdb/frame.c b/gdb/frame.c
>> index a08a8f47ebc4..fed961b2a8df 100644
>> --- a/gdb/frame.c
>> +++ b/gdb/frame.c
>> @@ -259,6 +259,22 @@ frame_addr_hash_eq (const void *a, const void *b)
>>     return f_entry->this_id.value == f_element->this_id.value;
>>   }
>> +/* Deletion function for the frame cache hash table.  */
>> +
>> +static void
>> +frame_info_del (void *frame_v)
>> +{
>> +  frame_info *frame = (frame_info *) frame_v;
>> +
>> +  if (frame->prologue_cache != nullptr
>> +      && frame->unwind->dealloc_cache != nullptr)
>> +    frame->unwind->dealloc_cache (frame, frame->prologue_cache);
>> +
>> +  if (frame->base_cache != nullptr
>> +      && frame->base->unwind->dealloc_cache != nullptr)
>> +    frame->base->unwind->dealloc_cache (frame, frame->base_cache);
>> +}
>> +
>>   /* Internal function to create the frame_stash hash table.  100 seems
>>      to be a good compromise to start the hash table at.  */
>> @@ -268,7 +284,7 @@ frame_stash_create (void)
>>     frame_stash = htab_create (100,
>>                    frame_addr_hash,
>>                    frame_addr_hash_eq,
>> -                 NULL);
>> +                 frame_info_del);
>>   }
>>   /* Internal function to add a frame to the frame_stash hash table.
>> @@ -2048,26 +2064,19 @@ reinit_frame_cache (void)
>>   {
>>     ++frame_cache_generation;
>> -  /* Tear down all frame caches.  */
>> -  for (frame_info *fi = sentinel_frame; fi != NULL; fi = fi->prev)
>> -    {
>> -      if (fi->prologue_cache && fi->unwind->dealloc_cache)
>> -    fi->unwind->dealloc_cache (fi, fi->prologue_cache);
>> -      if (fi->base_cache && fi->base->unwind->dealloc_cache)
>> -    fi->base->unwind->dealloc_cache (fi, fi->base_cache);
>> -    }
>> -
>> -  /* Since we can't really be sure what the first object allocated 
>> was.  */
>> -  obstack_free (&frame_cache_obstack, 0);
>> -  obstack_init (&frame_cache_obstack);
>> -
>>     if (sentinel_frame != NULL)
>>       annotate_frames_invalid ();
>> -  sentinel_frame = NULL;        /* Invalidate cache */
>>     invalidate_selected_frame ();
>> +
>> +  /* Invalidate cache.  */
>> +  sentinel_frame = NULL;
>>     frame_stash_invalidate ();
>> +  /* Since we can't really be sure what the first object allocated 
>> was.  */
>> +  obstack_free (&frame_cache_obstack, 0);
>> +  obstack_init (&frame_cache_obstack);
>> +
>>     for (frame_info_ptr &iter : frame_info_ptr::frame_list)
>>       iter.invalidate ();
>
  

On 2/9/23 07:42, Tom de Vries via Gdb-patches wrote:
> On 2/9/23 08:40, Tom de Vries wrote:
>> On 1/30/23 21:02, Simon Marchi via Gdb-patches wrote:
>>> Currently, some frame resources are deallocated by iterating on the
>>> frame chain (starting from the sentinel), calling dealloc_cache.  The
>>> problem is that user-created frames are not part of that chain, so we
>>> never call dealloc_cache for them.
>>>
>>> I propose to make it so the dealloc_cache callbacks are called when the
>>> frames are removed from the frame_stash hash table, by registering a
>>> deletion function to the hash table.  This happens when
>>> frame_stash_invalidate is called by reinit_frame_cache.  This way, all
>>> frames registered in the cache will get their unwinder's dealloc_cache
>>> callbacks called.
>>>
>>> Note that at the moment, the sentinel frames are not registered in the
>>> cache, so we won't call dealloc_cache for them.  However, it's just a
>>> theoritical problem, because the sentinel frame unwinder does not
>>> provide this callback.  Also, a subsequent patch will change things so
>>> that sentinel frames are registered to the cache.
>>>
>>> I moved the obstack_free / obstack_init pair below the
>>> frame_stash_invalidate call in reinit_frame_cache, because I assumed
>>> that some dealloc_cache would need to access some data on that obstack,
>>> so it would be better to free it after clearing the hash table.
>>>
>>
>> For me this causes:
>> ...
>> (gdb) PASS: gdb.btrace/record_goto.exp: instruction-history from 19 forwards
>> record goto 27^M
>> /data/vries/gdb/src/gdb/record-btrace.c:1654: internal-error: bfcache_new: Assertion `*slot == NULL' failed.^M
>> A problem internal to GDB has been detected,^M
>> further debugging may prove unreliable.^M
>> ----- Backtrace -----^M
>> FAIL: gdb.btrace/record_goto.exp: record goto 27 (GDB internal error)
>> ...
>>
>> Note that I've been having some problems with btrace tests, possible related to cpu/kernel combination (PRs 30073 and 30075), so this may be difficult to reproduce, I'm not sure.
>>
> 
> I also managed to reproduce this on openSUSE Tumbleweed, which doesn't show the problems with btrace tests, so I'm hoping this is easy to reproduce.
> 
> Thanks,
> - Tom

Thanks for reporting.  I managed to reproduce on gcc13 (I only have
computers with AMD cpus at the moment, so I can't run the btrace tests).

I sent a patch, the explications of the cause and proposed solution are
there:

  https://inbox.sourceware.org/gdb-patches/20230209195037.100368-1-simon.marchi@efficios.com/T/#u

Simon