Skip to contentSkip to conversation

Fix blktrace unaligned memory access

PATCH In review Jens Axboe 3 participants 6 messages Jul 10, 2009 – Jul 14, 2009

Subsystems (2)
lore
Share & tools
lore

Jens Axboe

author
#0 lore raw

PATCH 1/1 Fix blktrace unaligned memory access

Hi,

It seems that relay_reserve() (or the ring_buffer_event_data(), that one still needs some love) can return unaligned memory, there's no way around that when you have pdu lengths that aren't a nice size. This can cause unaligned access warnings on platforms that care about alignment.

This is an RFC, perhaps we can fix this in some other way. This one takes the simple approach, use an on-stack copy and memcpy() that to the destination.

kernel/trace/blktrace.c +14 -10
index 39af8af..9aba4ec 100644
@@ -62,7 +62,7 @@ static void blk_unregister_tracepoints(void);
 static void trace_note(struct blk_trace *bt, pid_t pid, int action,
 		       const void *data, size_t len)
 {
-	struct blk_io_trace *t;
+	struct blk_io_trace *ptr, __t, *t = &__t;
 	struct ring_buffer_event *event = NULL;
 	int pc = 0;
 	int cpu = smp_processor_id();
@@ -75,15 +75,15 @@ static void trace_note(struct blk_trace *bt, pid_t pid, int action,
 						  0, pc);
 		if (!event)
 			return;
-		t = ring_buffer_event_data(event);
+		ptr = t = ring_buffer_event_data(event);
 		goto record_it;
 	}
 
 	if (!bt->rchan)
 		return;
 
-	t = relay_reserve(bt->rchan, sizeof(*t) + len);
-	if (t) {
+	ptr = relay_reserve(bt->rchan, sizeof(*t) + len);
+	if (ptr) {
 		t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
 		t->time = ktime_to_ns(ktime_get());
 record_it:
@@ -92,7 +92,9 @@ record_it:
 		t->pid = pid;
 		t->cpu = cpu;
 		t->pdu_len = len;
-		memcpy((void *) t + sizeof(*t), data, len);
+		if (t == &__t)
+			memcpy(ptr, t, sizeof(*t));
+		memcpy((void *) ptr + sizeof(*ptr), data, len);
 
 		if (blk_tracer)
 			trace_buffer_unlock_commit(blk_tr, event, 0, pc);
@@ -178,7 +180,7 @@ static void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
 {
 	struct task_struct *tsk = current;
 	struct ring_buffer_event *event = NULL;
-	struct blk_io_trace *t;
+	struct blk_io_trace *ptr, __t, *t = &__t;
 	unsigned long flags = 0;
 	unsigned long *sequence;
 	pid_t pid;
@@ -209,7 +211,7 @@ static void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
 						  0, pc);
 		if (!event)
 			return;
-		t = ring_buffer_event_data(event);
+		ptr = t = ring_buffer_event_data(event);
 		goto record_it;
 	}
 
@@ -223,8 +225,8 @@ static void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
 	if (unlikely(tsk->btrace_seq != blktrace_seq))
 		trace_note_tsk(bt, tsk);
 
-	t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
-	if (t) {
+	ptr = relay_reserve(bt->rchan, sizeof(*ptr) + pdu_len);
+	if (ptr) {
 		sequence = per_cpu_ptr(bt->sequence, cpu);
 
 		t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
@@ -247,8 +249,10 @@ record_it:
 		t->error = error;
 		t->pdu_len = pdu_len;
 
+		if (t == &__t)
+			memcpy(ptr, t, sizeof(*t));
 		if (pdu_len)
-			memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
+			memcpy((void *) ptr + sizeof(*ptr), pdu_data, pdu_len);
 
 		if (blk_tracer) {
 			trace_buffer_unlock_commit(blk_tr, event, 0, pc);
Signature · 1 line
Jens Axboe
#1 lore raw
Quoted 7 lines · Jens Axboe

Hi,

It seems that relay_reserve() (or the ring_buffer_event_data(), that one still needs some love) can return unaligned memory, there's no way around that when you have pdu lengths that aren't a nice size. This can cause unaligned access warnings on platforms that care about alignment.

Seems relay_reserve() does nothing for alignment..On the other hand, ring_buffer_event_data() returns a ptr which is 32bit-aligned, but this still means it can cause unaligned accesses on 64bits arch, while I think it's fixable in ring buffer, it's certainly not an easy job.

Quoted 4 lines

This is an RFC, perhaps we can fix this in some other way. This one takes the simple approach, use an on-stack copy and memcpy() that to the destination.

or get_unaligned() ?

Quoted 85 lines
kernel/trace/blktrace.c +14 -10
index 39af8af..9aba4ec 100644
@@ -62,7 +62,7 @@ static void blk_unregister_tracepoints(void);
 static void trace_note(struct blk_trace *bt, pid_t pid, int action,
 		       const void *data, size_t len)
 {
-	struct blk_io_trace *t;
+	struct blk_io_trace *ptr, __t, *t = &__t;
 	struct ring_buffer_event *event = NULL;
 	int pc = 0;
 	int cpu = smp_processor_id();
@@ -75,15 +75,15 @@ static void trace_note(struct blk_trace *bt, pid_t pid, int action,
 						  0, pc);
 		if (!event)
 			return;
-		t = ring_buffer_event_data(event);
+		ptr = t = ring_buffer_event_data(event);
 		goto record_it;
 	}
 
 	if (!bt->rchan)
 		return;
 
-	t = relay_reserve(bt->rchan, sizeof(*t) + len);
-	if (t) {
+	ptr = relay_reserve(bt->rchan, sizeof(*t) + len);
+	if (ptr) {
 		t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
 		t->time = ktime_to_ns(ktime_get());
 record_it:
@@ -92,7 +92,9 @@ record_it:
 		t->pid = pid;
 		t->cpu = cpu;
 		t->pdu_len = len;
-		memcpy((void *) t + sizeof(*t), data, len);
+		if (t == &__t)
+			memcpy(ptr, t, sizeof(*t));
+		memcpy((void *) ptr + sizeof(*ptr), data, len);
 
 		if (blk_tracer)
 			trace_buffer_unlock_commit(blk_tr, event, 0, pc);
@@ -178,7 +180,7 @@ static void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
 {
 	struct task_struct *tsk = current;
 	struct ring_buffer_event *event = NULL;
-	struct blk_io_trace *t;
+	struct blk_io_trace *ptr, __t, *t = &__t;
 	unsigned long flags = 0;
 	unsigned long *sequence;
 	pid_t pid;
@@ -209,7 +211,7 @@ static void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
 						  0, pc);
 		if (!event)
 			return;
-		t = ring_buffer_event_data(event);
+		ptr = t = ring_buffer_event_data(event);
 		goto record_it;
 	}
 
@@ -223,8 +225,8 @@ static void __blk_add_trace(struct blk_trace *bt, sector_t sector, int bytes,
 	if (unlikely(tsk->btrace_seq != blktrace_seq))
 		trace_note_tsk(bt, tsk);
 
-	t = relay_reserve(bt->rchan, sizeof(*t) + pdu_len);
-	if (t) {
+	ptr = relay_reserve(bt->rchan, sizeof(*ptr) + pdu_len);
+	if (ptr) {
 		sequence = per_cpu_ptr(bt->sequence, cpu);
 
 		t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
@@ -247,8 +249,10 @@ record_it:
 		t->error = error;
 		t->pdu_len = pdu_len;
 
+		if (t == &__t)
+			memcpy(ptr, t, sizeof(*t));
 		if (pdu_len)
-			memcpy((void *) t + sizeof(*t), pdu_data, pdu_len);
+			memcpy((void *) ptr + sizeof(*ptr), pdu_data, pdu_len);
 
 		if (blk_tracer) {
 			trace_buffer_unlock_commit(blk_tr, event, 0, pc);
author
#2 lore raw
Quoted 9 lines · Li Zefan
Quoted 7 lines · Jens Axboe

Hi,

It seems that relay_reserve() (or the ring_buffer_event_data(), that one still needs some love) can return unaligned memory, there's no way around that when you have pdu lengths that aren't a nice size. This can cause unaligned access warnings on platforms that care about alignment.

Quoted 4 lines

Seems relay_reserve() does nothing for alignment..On the other hand, ring_buffer_event_data() returns a ptr which is 32bit-aligned, but this still means it can cause unaligned accesses on 64bits arch, while I think it's fixable in ring buffer, it's certainly not an easy job.

Right, it's a bit nasty...

Quoted 6 lines
Quoted 4 lines

This is an RFC, perhaps we can fix this in some other way. This one takes the simple approach, use an on-stack copy and memcpy() that to the destination.

or get_unaligned() ?

put_unaligned(), you mean? The big question is then which is faster, using put_unaligned() or doing the memcpy() of the structure...

Signature · 1 line
Jens Axboe
#3 lore raw
Quoted 9 lines · Jens Axboe
Quoted 8 lines · Li Zefan
Quoted 7 lines · Jens Axboe

Hi,

It seems that relay_reserve() (or the ring_buffer_event_data(), that one still needs some love) can return unaligned memory, there's no way around that when you have pdu lengths that aren't a nice size. This can cause unaligned access warnings on platforms that care about alignment.

Quoted 7 lines
Quoted 4 lines

Seems relay_reserve() does nothing for alignment..On the other hand, ring_buffer_event_data() returns a ptr which is 32bit-aligned, but this still means it can cause unaligned accesses on 64bits arch, while I think it's fixable in ring buffer, it's certainly not an easy job.

Right, it's a bit nasty...

Lai Jiangshan noticed this issue long ago and had some ideas in mind how to fix ring buffer, but never try it out for it will probably be frustrating..

Quoted 9 lines
Quoted 5 lines
Quoted 4 lines

This is an RFC, perhaps we can fix this in some other way. This one takes the simple approach, use an on-stack copy and memcpy() that to the destination.

or get_unaligned() ?

put_unaligned(), you mean? The big question is then which is faster, using put_unaligned() or doing the memcpy() of the structure...

Ah, I meant put_unaligned().

I think the patch you posted can be a workaround at least for now, and can be improved by detecting HAVE_EFFICIENT_UNALIGNED_ACCESS

author
#4 lore raw
Quoted 28 lines · Li Zefan
Quoted 9 lines · Jens Axboe
Quoted 8 lines · Li Zefan
Quoted 7 lines · Jens Axboe

Hi,

It seems that relay_reserve() (or the ring_buffer_event_data(), that one still needs some love) can return unaligned memory, there's no way around that when you have pdu lengths that aren't a nice size. This can cause unaligned access warnings on platforms that care about alignment.

Quoted 7 lines
Quoted 4 lines

Seems relay_reserve() does nothing for alignment..On the other hand, ring_buffer_event_data() returns a ptr which is 32bit-aligned, but this still means it can cause unaligned accesses on 64bits arch, while I think it's fixable in ring buffer, it's certainly not an easy job.

Right, it's a bit nasty...

Lai Jiangshan noticed this issue long ago and had some ideas in mind how to fix ring buffer, but never try it out for it will probably be frustrating..

Quoted 6 lines
Quoted 5 lines
Quoted 4 lines

This is an RFC, perhaps we can fix this in some other way. This one takes the simple approach, use an on-stack copy and memcpy() that to the destination.

or get_unaligned() ?

Quoted 8 lines
Quoted 3 lines

put_unaligned(), you mean? The big question is then which is faster, using put_unaligned() or doing the memcpy() of the structure...

Ah, I meant put_unaligned().

I think the patch you posted can be a workaround at least for now, and can be improved by detecting HAVE_EFFICIENT_UNALIGNED_ACCESS

We either do one or the other, I don't want to clutter the code with both methods implemented and some ifdef checking and deciding. I'll check the unaligned put here on x86/ppc/sparc and then decide which one is faster.

Signature · 1 line
Jens Axboe
maintainer
#5 lore raw
Quoted 11 lines · Jens Axboe
Quoted 10 lines · Li Zefan
Quoted 9 lines · Jens Axboe
Quoted 8 lines · Li Zefan
Quoted 7 lines · Jens Axboe

Hi,

It seems that relay_reserve() (or the ring_buffer_event_data(), that one still needs some love) can return unaligned memory, there's no way around that when you have pdu lengths that aren't a nice size. This can cause unaligned access warnings on platforms that care about alignment.

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

Seems relay_reserve() does nothing for alignment..On the other hand, ring_buffer_event_data() returns a ptr which is 32bit-aligned, but this still means it can cause unaligned accesses on 64bits arch, while I think it's fixable in ring buffer, it's certainly not an easy job.

Correct, this is only needed if the data is bigger than 32 bits.

Quoted 7 lines
Quoted 7 lines
Quoted 3 lines

Right, it's a bit nasty...

Lai Jiangshan noticed this issue long ago and had some ideas in mind how to fix ring buffer, but never try it out for it will probably be frustrating..

Perhaps we can make an attribute for the ring buffer itself that would allow the user to give an alignment. Would anything ever need a bigger than 64bit (8 byte) alignment? We can set a flag that says, make all entries 8 bytes aligned.

-- Steve