[PATCH 1/6] : bug fix, remove partial zero out

[PATCH 1/6] : bug fix, remove partial zero out

[Lai Jiangshan]

partial-zero-out a struct is very dangerous, we should zero out
field by field directly when need.

partial-zero-out for struct trace_iterator exists when ftrace
was first introduced into mainline kernel. But in this few years,
the code of ftrace is changed a lot, and:

1) partial-zero-out for struct trace_iterator has a bug now,
   cpumask_var_t started should not be zeroed out.

2) I viewed the codes and found that fields below
   "/* The below is zeroed out in pipe_read */"
   don't need to be zeroed out or initialized now.

So, we remove the code of "partial zero out"

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
---
diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h
index 3ca9485..c6d0e1a 100644
--- a/include/linux/ftrace_event.h
+++ b/include/linux/ftrace_event.h
@@ -54,7 +54,6 @@ struct trace_iterator {
 	struct ring_buffer_iter	*buffer_iter[NR_CPUS];
 	unsigned long		iter_flags;
 
-	/* The below is zeroed out in pipe_read */
 	struct trace_seq	seq;
 	struct trace_entry	*ent;
 	int			leftover;
diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
index 5314c90..27fecf8 100644
--- a/kernel/trace/trace.c
+++ b/kernel/trace/trace.c
@@ -3124,12 +3124,6 @@ waitagain:
 	if (cnt >= PAGE_SIZE)
 		cnt = PAGE_SIZE - 1;
 
-	/* reset all but tr, trace, and overruns */
-	memset(&iter->seq, 0,
-	       sizeof(struct trace_iterator) -
-	       offsetof(struct trace_iterator, seq));
-	iter->pos = -1;
-
 	trace_event_read_lock();
 	trace_access_lock(iter->cpu_file);
 	while (find_next_entry_inc(iter) != NULL) {
@@ -4398,12 +4392,7 @@ static void __ftrace_dump(bool disable_tracing)
 
 		cnt++;
 
-		/* reset all but tr, trace, and overruns */
-		memset(&iter.seq, 0,
-		       sizeof(struct trace_iterator) -
-		       offsetof(struct trace_iterator, seq));
 		iter.iter_flags |= TRACE_FILE_LAT_FMT;
-		iter.pos = -1;
 
 		if (find_next_entry_inc(&iter) != NULL) {
 			int ret;

Re: [PATCH 1/6] : bug fix, remove partial zero out

[Frederic Weisbecker]

On Tue, Jan 19, 2010 at 03:33:56PM +0800, Lai Jiangshan wrote:
> partial-zero-out a struct is very dangerous, we should zero out
> field by field directly when need.
> 
> partial-zero-out for struct trace_iterator exists when ftrace
> was first introduced into mainline kernel. But in this few years,
> the code of ftrace is changed a lot, and:
> 
> 1) partial-zero-out for struct trace_iterator has a bug now,
>    cpumask_var_t started should not be zeroed out.
> 
> 2) I viewed the codes and found that fields below
>    "/* The below is zeroed out in pipe_read */"
>    don't need to be zeroed out or initialized now.
> 
> So, we remove the code of "partial zero out"
> 
> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
> ---
> diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h
> index 3ca9485..c6d0e1a 100644
> --- a/include/linux/ftrace_event.h
> +++ b/include/linux/ftrace_event.h
> @@ -54,7 +54,6 @@ struct trace_iterator {
>  	struct ring_buffer_iter	*buffer_iter[NR_CPUS];
>  	unsigned long		iter_flags;
>  
> -	/* The below is zeroed out in pipe_read */
>  	struct trace_seq	seq;
>  	struct trace_entry	*ent;
>  	int			leftover;
> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> index 5314c90..27fecf8 100644
> --- a/kernel/trace/trace.c
> +++ b/kernel/trace/trace.c
> @@ -3124,12 +3124,6 @@ waitagain:
>  	if (cnt >= PAGE_SIZE)
>  		cnt = PAGE_SIZE - 1;
>  
> -	/* reset all but tr, trace, and overruns */
> -	memset(&iter->seq, 0,
> -	       sizeof(struct trace_iterator) -
> -	       offsetof(struct trace_iterator, seq));
> -	iter->pos = -1;
> -



I'm not sure exaclty why we needed to zero the seq here.
We already reset it in trace_seq_init().

We might do it again on waitagain. I lost track how we could
ever need to goto waitagain. It was about a tricky bug to fix
but I'm don't remember exactly the details.

That said, if trace_seq_to_user returns -EBUSY, we
re-init the seq buffer, so it should be fine I guess.

But concerning the need of setting iter->pos to -1, I'm not
sure we need to remove it. Shouldn't it be set to 0 btw?

Steve?

Re: [PATCH 1/6] : bug fix, remove partial zero out

[Lai Jiangshan]

Frederic Weisbecker wrote:
> On Tue, Jan 19, 2010 at 03:33:56PM +0800, Lai Jiangshan wrote:
>> partial-zero-out a struct is very dangerous, we should zero out
>> field by field directly when need.
>>
>> partial-zero-out for struct trace_iterator exists when ftrace
>> was first introduced into mainline kernel. But in this few years,
>> the code of ftrace is changed a lot, and:
>>
>> 1) partial-zero-out for struct trace_iterator has a bug now,
>>    cpumask_var_t started should not be zeroed out.
>>
>> 2) I viewed the codes and found that fields below
>>    "/* The below is zeroed out in pipe_read */"
>>    don't need to be zeroed out or initialized now.
>>
>> So, we remove the code of "partial zero out"
>>
>> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
>> ---
>> diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h
>> index 3ca9485..c6d0e1a 100644
>> --- a/include/linux/ftrace_event.h
>> +++ b/include/linux/ftrace_event.h
>> @@ -54,7 +54,6 @@ struct trace_iterator {
>>  	struct ring_buffer_iter	*buffer_iter[NR_CPUS];
>>  	unsigned long		iter_flags;
>>  
>> -	/* The below is zeroed out in pipe_read */
>>  	struct trace_seq	seq;
>>  	struct trace_entry	*ent;
>>  	int			leftover;
>> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
>> index 5314c90..27fecf8 100644
>> --- a/kernel/trace/trace.c
>> +++ b/kernel/trace/trace.c
>> @@ -3124,12 +3124,6 @@ waitagain:
>>  	if (cnt >= PAGE_SIZE)
>>  		cnt = PAGE_SIZE - 1;
>>  
>> -	/* reset all but tr, trace, and overruns */
>> -	memset(&iter->seq, 0,
>> -	       sizeof(struct trace_iterator) -
>> -	       offsetof(struct trace_iterator, seq));
>> -	iter->pos = -1;
>> -
> 
> 
> 
> I'm not sure exaclty why we needed to zero the seq here.
> We already reset it in trace_seq_init().
> 
> We might do it again on waitagain. I lost track how we could
> ever need to goto waitagain. It was about a tricky bug to fix
> but I'm don't remember exactly the details.
> 
> That said, if trace_seq_to_user returns -EBUSY, we
> re-init the seq buffer, so it should be fine I guess.

Yes, -EBUSY is strange here.
but any way, trace_seq_init() is called.

> 
> But concerning the need of setting iter->pos to -1, I'm not
> sure we need to remove it. Shouldn't it be set to 0 btw?
> 

->pos is not used here, ->idx is just increased here,
so we don't need to initialize them.

 

Re: [PATCH 1/6] : bug fix, remove partial zero out

[Frederic Weisbecker]

On Tue, Jan 26, 2010 at 11:31:58AM +0800, Lai Jiangshan wrote:
> Frederic Weisbecker wrote:
> > On Tue, Jan 19, 2010 at 03:33:56PM +0800, Lai Jiangshan wrote:
> >> partial-zero-out a struct is very dangerous, we should zero out
> >> field by field directly when need.
> >>
> >> partial-zero-out for struct trace_iterator exists when ftrace
> >> was first introduced into mainline kernel. But in this few years,
> >> the code of ftrace is changed a lot, and:
> >>
> >> 1) partial-zero-out for struct trace_iterator has a bug now,
> >>    cpumask_var_t started should not be zeroed out.
> >>
> >> 2) I viewed the codes and found that fields below
> >>    "/* The below is zeroed out in pipe_read */"
> >>    don't need to be zeroed out or initialized now.
> >>
> >> So, we remove the code of "partial zero out"
> >>
> >> Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
> >> ---
> >> diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h
> >> index 3ca9485..c6d0e1a 100644
> >> --- a/include/linux/ftrace_event.h
> >> +++ b/include/linux/ftrace_event.h
> >> @@ -54,7 +54,6 @@ struct trace_iterator {
> >>  	struct ring_buffer_iter	*buffer_iter[NR_CPUS];
> >>  	unsigned long		iter_flags;
> >>  
> >> -	/* The below is zeroed out in pipe_read */
> >>  	struct trace_seq	seq;
> >>  	struct trace_entry	*ent;
> >>  	int			leftover;
> >> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> >> index 5314c90..27fecf8 100644
> >> --- a/kernel/trace/trace.c
> >> +++ b/kernel/trace/trace.c
> >> @@ -3124,12 +3124,6 @@ waitagain:
> >>  	if (cnt >= PAGE_SIZE)
> >>  		cnt = PAGE_SIZE - 1;
> >>  
> >> -	/* reset all but tr, trace, and overruns */
> >> -	memset(&iter->seq, 0,
> >> -	       sizeof(struct trace_iterator) -
> >> -	       offsetof(struct trace_iterator, seq));
> >> -	iter->pos = -1;
> >> -
> > 
> > 
> > 
> > I'm not sure exaclty why we needed to zero the seq here.
> > We already reset it in trace_seq_init().
> > 
> > We might do it again on waitagain. I lost track how we could
> > ever need to goto waitagain. It was about a tricky bug to fix
> > but I'm don't remember exactly the details.
> > 
> > That said, if trace_seq_to_user returns -EBUSY, we
> > re-init the seq buffer, so it should be fine I guess.
> 
> Yes, -EBUSY is strange here.
> but any way, trace_seq_init() is called.
> 
> > 
> > But concerning the need of setting iter->pos to -1, I'm not
> > sure we need to remove it. Shouldn't it be set to 0 btw?
> > 
> 
> ->pos is not used here, ->idx is just increased here,
> so we don't need to initialize them.


Ok.