Skip to contentSkip to conversation

tracing: silence GCC 9 array bounds warning

PATCH Superseded Miguel Ojeda 3 participants 7 messages May 17, 2019 – May 19, 2019 Tracing

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 tracing: silence GCC 9 array bounds warning

Starting with GCC 9, -Warray-bounds detects cases when memset is called starting on a member of a struct but the size to be cleared ends up writing over further members.

Such a call happens in the trace code to clear, at once, all members after and including seq on struct trace_iterator:

In function 'memset',
inlined from 'ftrace_dump' at kernel/trace/trace.c:8914:3:
./include/linux/string.h:344:9: warning: '__builtin_memset' offset [8505, 8560] from the object at 'iter' is out of the bounds of referenced subobject 'seq' with type 'struct trace_seq' at offset 4368 [-Warray-bounds]
344 | return __builtin_memset(p, c, size);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~

In order to avoid GCC complaining about it, we compute the address ourselves by adding the offsetof distance instead of referring directly to the member.

Since there are two places doing this clear (trace.c and trace_kdb.c), take the chance to move the workaround into a single place in the internal header.

---
 kernel/trace/trace.c     |  7 +------
 kernel/trace/trace.h     | 14 ++++++++++++++
 kernel/trace/trace_kdb.c |  7 +------
 3 files changed, 16 insertions(+), 12 deletions(-)
kernel/trace/trace.c +1 -6
index ca1ee656d6d8..37990532351b 100644
@@ -8627,12 +8627,7 @@ void ftrace_dump(enum ftrace_dump_mode oops_dump_mode)
 
 		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;
+		trace_iterator_reset(&iter);
 
 		if (trace_find_next_entry_inc(&iter) != NULL) {
 			int ret;
kernel/trace/trace.h +14 -0
index d80cee49e0eb..80ad656f43eb 100644
@@ -1964,4 +1964,18 @@ static inline void tracer_hardirqs_off(unsigned long a0, unsigned long a1) { }
 
 extern struct trace_iterator *tracepoint_print_iter;
 
+/* reset all but tr, trace, and overruns */
+static __always_inline void trace_iterator_reset(struct trace_iterator * iter)
+{
+	/*
+	 * Equivalent to &iter->seq, but avoids GCC 9 complaining about
+	 * overwriting more members than just iter->seq (-Warray-bounds)
+	 */
+	memset((char *)(iter) + offsetof(struct trace_iterator, seq), 0,
+	       sizeof(struct trace_iterator) -
+	       offsetof(struct trace_iterator, seq));
+	iter->iter_flags |= TRACE_FILE_LAT_FMT;
+	iter->pos = -1;
+}
+
 #endif /* _LINUX_KERNEL_TRACE_H */
kernel/trace/trace_kdb.c +1 -6
index 810d78a8d14c..0a2a166ee716 100644
@@ -41,12 +41,7 @@ static void ftrace_dump_buf(int skip_lines, long cpu_file)
 
 	kdb_printf("Dumping ftrace buffer:\n");
 
-	/* 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;
+	trace_iterator_reset(&iter);
 
 	if (cpu_file == RING_BUFFER_ALL_CPUS) {
 		for_each_tracing_cpu(cpu) {
Signature · 1 line
2.17.1
maintainer
#1 lore raw
Quoted 16 lines · On Fri, 17 May 2019 11:25:02 +0200 Miguel Ojeda…

Starting with GCC 9, -Warray-bounds detects cases when memset is called starting on a member of a struct but the size to be cleared ends up writing over further members.

Such a call happens in the trace code to clear, at once, all members after and including seq on struct trace_iterator:

In function 'memset',
inlined from 'ftrace_dump' at kernel/trace/trace.c:8914:3:
./include/linux/string.h:344:9: warning: '__builtin_memset' offset [8505, 8560] from the object at 'iter' is out of the bounds of referenced subobject 'seq' with type 'struct trace_seq' at offset 4368 [-Warray-bounds]
344 | return __builtin_memset(p, c, size);
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~

Quoted 7 lines

In order to avoid GCC complaining about it, we compute the address ourselves by adding the offsetof distance instead of referring directly to the member.

Since there are two places doing this clear (trace.c and trace_kdb.c), take the chance to move the workaround into a single place in the internal header.

Hi Miguel,

Linus mentioned this too.

https://lore.kernel.org/lkml/CAHk-=wihYB8w__YQjgYjYZsVniu5CtkTcFycmCGdqVg8GUje7g@mail.gmail.com/T/#u

I was going to do a helper function, and put it in the queue for the next merge window (as it isn't really a bug, just gcc complaining a little more aggressively). But since you already did the patch, I'll use yours. But I have some nits about it below.

Add here:

Quoted 13 lines
---
 kernel/trace/trace.c     |  7 +------
 kernel/trace/trace.h     | 14 ++++++++++++++
 kernel/trace/trace_kdb.c |  7 +------
 3 files changed, 16 insertions(+), 12 deletions(-)
kernel/trace/trace.c +0 -0
index ca1ee656d6d8..37990532351b 100644

kernel/trace/trace.c void ftrace_dump(enum ftrace_dump_mode · line 8634

Quoted 8 lines
kernel/trace/trace.c +0 -5
@@ -8627,12 +8627,7 @@ void ftrace_dump(enum ftrace_dump_mode
 oops_dump_mode) 
 		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;

Setting the LAT_FMT isn't something a function called "reset" should do.

Quoted 21 lines
kernel/trace/trace.c +1 -1
-		iter.pos = -1;
+		trace_iterator_reset(&iter);
 
 		if (trace_find_next_entry_inc(&iter) != NULL) {
 			int ret;
kernel/trace/trace.h +4 -0
index d80cee49e0eb..80ad656f43eb 100644
@@ -1964,4 +1964,18 @@ static inline void
 tracer_hardirqs_off(unsigned long a0, unsigned long a1) { } 
 extern struct trace_iterator *tracepoint_print_iter;
 
+/* reset all but tr, trace, and overruns */
+static __always_inline void trace_iterator_reset(struct
 trace_iterator * iter) +{
+	/*
+	 * Equivalent to &iter->seq, but avoids GCC 9 complaining

about

kernel/trace/trace.h +1 -0
+	 * overwriting more members than just iter->seq

(-Warray-bounds)

kernel/trace/trace.h

Quoted 2 lines
kernel/trace/trace.h +2 -0
+	 */
+	memset((char *)(iter) + offsetof(struct trace_iterator,

Why (char *)? Please use (void *).

kernel/trace/trace.h static inline void tracer_hardirqs_off(unsigned long a0, unsigned long a1) { } · line 1976

Quoted 3 lines

seq), 0,

kernel/trace/trace.h +2 -0
+	       sizeof(struct trace_iterator) -
+	       offsetof(struct trace_iterator, seq));

Make a variable for offset and reuse that (see Linus's email).

kernel/trace/trace.h static inline void tracer_hardirqs_off(unsigned long a0, unsigned long a1) { } · line 1977

Quoted 1 line
kernel/trace/trace.h +1 -0
+	iter->iter_flags |= TRACE_FILE_LAT_FMT;

Again, leave the LAT_FMT change in the other locations.

Please send a v2 version with these updates.

Thanks!

-- Steve

Quoted 22 lines
kernel/trace/trace.h +3 -0
+	iter->pos = -1;
+}
+
 #endif /* _LINUX_KERNEL_TRACE_H */
kernel/trace/trace_kdb.c +1 -6
index 810d78a8d14c..0a2a166ee716 100644
@@ -41,12 +41,7 @@ static void ftrace_dump_buf(int skip_lines, long
 cpu_file) 
 	kdb_printf("Dumping ftrace buffer:\n");
 
-	/* 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;
+	trace_iterator_reset(&iter);
 
 	if (cpu_file == RING_BUFFER_ALL_CPUS) {
 		for_each_tracing_cpu(cpu) {
#3 lore raw
In reply to Miguel Ojeda

kernel/trace/trace.h static inline void tracer_hardirqs_off(unsigned long a0, unsigned long a1) { } · line 1976

Quoted 4 lines · Miguel Ojeda
kernel/trace/trace.h +3 -0
+       memset((char *)(iter) + offsetof(struct trace_iterator, seq), 0,
+              sizeof(struct trace_iterator) -
+              offsetof(struct trace_iterator, seq));

Honestly, the above is nasty.

Whenever you have to split an expression or statement over several lines, you should ask yourself why it's so complicated.

This can be trivially rewritten as

        const unsigned int offset = offsetof(struct trace_iterator, seq);
        memset(offset + (void *)iter, 0, sizeof(*iter) - offset);

and now we have two much simpler expressions that are each much easier to read and don't need multiple lines.

In particular using that "offset" variable means that you can trivially see that "oh, we're starting the memset at that offset, and we're using the 'sizeof() - offset' thing to limit the size". It's a lot clearer to use the same trivial 'offset' thing in both the offset and the size, than to have a more complex expression that it duplicated twice and is not at all as visually obvious that it's the exact same thing simply because it's bigger and more complex.

Also, the while 'offset' is a variable, any compiler will immediately see that it's a constant value, so it's not like this will affect the generated code at all. Unless you compile with something crazy like '-O0', which is not a supported configuration exactly because we expect compilers to not be terminally stupid.

So you get simpler and clearer source, with no downsides.

Linus

#2 lore raw
In reply to Steven Rostedt

Hi Steven,

Quoted 6 lines · Steven Rostedt

Hi Miguel,

Linus mentioned this too.

https://lore.kernel.org/lkml/CAHk-=wihYB8w__YQjgYjYZsVniu5CtkTcFycmCGdqVg8GUje7g@mail.gmail.com/T/#u

Ah, I didn't see that. We were discussing here [1] backporting to 4.19 some cleanups we did for GCC 9 a few months ago, so I was testing that Linus' master was still clean a few hours ago. When I found that couple of warnings I quickly made a patch before I forgot. I should probably have added RFC :-)

[1] https://lore.kernel.org/lkml/CANiq72=fsL5m2_e+bNovFCHy3=YVf53EKGtGE_sWvsAD=ONHuQ@mail.gmail.com/

Quoted 1 line

Setting the LAT_FMT isn't something a function called "reset" should do.

I added the surrounding code heuristically since it was in a single block with the comment above and since it was repeated in both places, but I had no idea on the semantics. :-)

Quoted 3 lines
Quoted 1 line

+ memset((char *)(iter) + offsetof(struct trace_iterator,

Why (char *)? Please use (void *).

We are adding a byte-sized offset, so char * is the one that makes sense. Doing it with void * works, although it is a GNU extension. But as you prefer :-)

Cheers,
Miguel

#4 lore raw
In reply to Linus Torvalds
Quoted 4 lines · Linus Torvalds
Quoted 1 line · Miguel Ojeda
Quoted 8 lines

kernel/trace/trace.h static inline void tracer_hardirqs_off(unsigned long a0, unsigned long a1) { } · line 1976

Quoted 3 lines
kernel/trace/trace.h +3 -0
+       memset((char *)(iter) + offsetof(struct trace_iterator, seq), 0,
+              sizeof(struct trace_iterator) -
+              offsetof(struct trace_iterator, seq));

Honestly, the above is nasty.

Whenever you have to split an expression or statement over several lines, you should ask yourself why it's so complicated.

Will do -- I was trying to keep the code as closely to the original as possible (I simply replaced the &iter.seq expression :-)

By the way, how do you all feel about moving this as a generic facility to zero out the suffix/prefix of an structure? In particular, since we won't have the LAT* stuff according to Steven.

Quoted 3 lines

Also, the while 'offset' is a variable, any compiler will immediately see that it's a constant value, so it's not like this will affect the generated code at all.

I like C++'s constexpr (for variable defs), maybe one day we will get it on C; it is useful to cleanly annotate compile-time values like this.

Quoted 3 lines

Unless you compile with something crazy like
'-O0', which is not a supported configuration exactly because we expect compilers to not be terminally stupid.

Fun fact: it seems clang folds some of these even under -O0. In godbolt I see it folding the third argument completely. The first one isn't, but it is computed on the function prologue, leaving the 'offset' variable unused.

Cheers,
Miguel

maintainer
#5 lore raw
In reply to Miguel Ojeda
Quoted 3 lines · On Fri, 17 May 2019 21:09:21 +0200 Miguel Ojeda…

By the way, how do you all feel about moving this as a generic facility to zero out the suffix/prefix of an structure? In particular, since we won't have the LAT* stuff according to Steven.

Is this done in other places? If so, how many.

Note, the LAT* update doesn't belong in the iterator reset function, but the pos = -1 still does.

Thanks!

-- Steve