Skip to contentSkip to conversation

PERF_EVENT_IOC_SET_FILTER has different values based on bitness

PATCH David Ahern 5 participants 10 messages Oct 7, 2014 – Oct 9, 2014 perf events

lore
Share & tools
lore
#0 lore raw

32-bit perf binaries are not able to set filters on 64-bit kernels.

   $ perf record -e  net:netif_receive_skb --filter 'name == "eth1"
   Error: failed to set filter with 25 (Inappropriate ioctl for device)

The reason is that the definition of PERF_EVENT_IOC_SET_FILTER contains a pointer:

#define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)

the size of which of course differs for 32-bit and 64-bit. This has been there since the original commit (6fb2915df7f07) back in 2009.

Thoughts on how to fix this? Changing the definition of SET_FILTER breaks existing setups so that rules it out. What about something like this:

#define PERF_EVENT_IOC_SET_FILTER_32 _IOW('$', 6, u32)

and then

kernel/events/core.c +1 -0
index 963bf139e2b2..c805132ac1cf 100644
@@ -3714,6 +3714,7 @@ static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
    }
 
    case PERF_EVENT_IOC_SET_FILTER:
+   case PERF_EVENT_IOC_SET_FILTER_32:
        return perf_event_set_filter(event, (void __user *)arg);
 
    default:

David

#1 lore raw
Quoted 34 lines · David Ahern

32-bit perf binaries are not able to set filters on 64-bit kernels.

    $ perf record -e  net:netif_receive_skb --filter 'name == "eth1"
    Error: failed to set filter with 25 (Inappropriate ioctl for device)

The reason is that the definition of PERF_EVENT_IOC_SET_FILTER contains a pointer:

#define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)

the size of which of course differs for 32-bit and 64-bit. This has been there since the original commit (6fb2915df7f07) back in 2009.

Thoughts on how to fix this? Changing the definition of SET_FILTER breaks existing setups so that rules it out. What about something like this:

#define PERF_EVENT_IOC_SET_FILTER_32 _IOW('$', 6, u32)

and then

kernel/events/core.c +1 -0
index 963bf139e2b2..c805132ac1cf 100644
@@ -3714,6 +3714,7 @@ static long perf_ioctl(struct file *file, unsigned
 int cmd, unsigned long arg)
     }
 
     case PERF_EVENT_IOC_SET_FILTER:
+   case PERF_EVENT_IOC_SET_FILTER_32:
         return perf_event_set_filter(event, (void __user *)arg);
 
Quoted 2 lines

default:

Oh, PERF_EVENT_IOC_ID has the same problem:

#define PERF_EVENT_IOC_ID _IOR('$', 7, __u64 *)

David

#2 lore raw
Quoted 13 lines · David Ahern
Quoted 13 lines
kernel/events/core.c +1 -0
index 963bf139e2b2..c805132ac1cf 100644
@@ -3714,6 +3714,7 @@ static long perf_ioctl(struct file *file, unsigned
 int cmd, unsigned long arg)
     }
 
     case PERF_EVENT_IOC_SET_FILTER:
+   case PERF_EVENT_IOC_SET_FILTER_32:
         return perf_event_set_filter(event, (void __user *)arg);
 
Quoted 6 lines
Quoted 2 lines

default:

Oh, PERF_EVENT_IOC_ID has the same problem:

#define PERF_EVENT_IOC_ID _IOR('$', 7, __u64 *)

The right way is to add a compat_perf_ioctl()

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only
#3 lore raw
Quoted 14 lines · Andi Kleen
Quoted 13 lines · David Ahern
Quoted 13 lines
kernel/events/core.c +1 -0
index 963bf139e2b2..c805132ac1cf 100644
@@ -3714,6 +3714,7 @@ static long perf_ioctl(struct file *file, unsigned
 int cmd, unsigned long arg)
      }
 
      case PERF_EVENT_IOC_SET_FILTER:
+   case PERF_EVENT_IOC_SET_FILTER_32:
          return perf_event_set_filter(event, (void __user *)arg);
 
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines

default:

Oh, PERF_EVENT_IOC_ID has the same problem:

#define PERF_EVENT_IOC_ID _IOR('$', 7, __u64 *)

The right way is to add a compat_perf_ioctl()

Sure, looked into that way as well. But SET_FILTER and IOC_ID will still compile to the same values for a 64-bit kernel.

David

#4 lore raw
Quoted 20 lines · David Ahern
Quoted 19 lines · Andi Kleen
Quoted 18 lines · David Ahern
Quoted 13 lines
kernel/events/core.c +1 -0
index 963bf139e2b2..c805132ac1cf 100644
@@ -3714,6 +3714,7 @@ static long perf_ioctl(struct file *file, unsigned
 int cmd, unsigned long arg)
      }
 
      case PERF_EVENT_IOC_SET_FILTER:
+   case PERF_EVENT_IOC_SET_FILTER_32:
          return perf_event_set_filter(event, (void __user *)arg);
 
Quoted 2 lines

default:

Oh, PERF_EVENT_IOC_ID has the same problem:

Quoted 6 lines
Quoted 3 lines
Quoted 1 line

#define PERF_EVENT_IOC_ID _IOR('$', 7, __u64 *)

The right way is to add a compat_perf_ioctl()

Sure, looked into that way as well. But SET_FILTER and IOC_ID will still compile to the same values for a 64-bit kernel.

Sure you have to add/use the new defines too.

-Andi

Signature · 1 line
[email protected] -- Speaking for myself only
maintainer
#5 lore raw
In reply to David Ahern
Quoted 6 lines · David Ahern

32-bit perf binaries are not able to set filters on 64-bit kernels.

   $ perf record -e  net:netif_receive_skb --filter 'name == "eth1"
   Error: failed to set filter with 25 (Inappropriate ioctl for device)
Quoted 8 lines

The reason is that the definition of PERF_EVENT_IOC_SET_FILTER contains a pointer:

#define PERF_EVENT_IOC_SET_FILTER _IOW('$', 6, char *)

the size of which of course differs for 32-bit and 64-bit. This has been there since the original commit (6fb2915df7f07) back in 2009.

Should be fixed for a while now

---
commit b3f207855f57b9c8f43a547a801340bb5cbc59e5

Author: Pawel Moll <[email protected]>
Date:   Fri Jun 13 16:03:32 2014 +0100

perf: Handle compat ioctl

When running a 32-bit userspace on a 64-bit kernel (eg. i386 application on x86_64 kernel or 32-bit arm userspace on arm64 kernel) some of the perf ioctls must be treated with special care, as they have a pointer size encoded in the command.

For example, PERF_EVENT_IOC_ID in 32-bit world will be encoded as 0x80042407, but 64-bit kernel will expect 0x80082407. In result the ioctl will fail returning -ENOTTY.

This patch solves the problem by adding code fixing up the size as compat_ioctl file operation.

kernel/events/core.c +22 -1
index 1cf24b3e42ec..f9c1ed002dbc 100644
@@ -41,6 +41,7 @@
 #include <linux/cgroup.h>
 #include <linux/module.h>
 #include <linux/mman.h>
+#include <linux/compat.h>
 
 #include "internal.h"
 
@@ -3717,6 +3718,26 @@ static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 	return 0;
 }
 
+#ifdef CONFIG_COMPAT
+static long perf_compat_ioctl(struct file *file, unsigned int cmd,
+				unsigned long arg)
+{
+	switch (_IOC_NR(cmd)) {
+	case _IOC_NR(PERF_EVENT_IOC_SET_FILTER):
+	case _IOC_NR(PERF_EVENT_IOC_ID):
+		/* Fix up pointer size (usually 4 -> 8 in 32-on-64-bit case */
+		if (_IOC_SIZE(cmd) == sizeof(compat_uptr_t)) {
+			cmd &= ~IOCSIZE_MASK;
+			cmd |= sizeof(void *) << IOCSIZE_SHIFT;
+		}
+		break;
+	}
+	return perf_ioctl(file, cmd, arg);
+}
+#else
+# define perf_compat_ioctl NULL
+#endif
+
 int perf_event_task_enable(void)
 {
 	struct perf_event *event;
@@ -4222,7 +4243,7 @@ static const struct file_operations perf_fops = {
 	.read			= perf_read,
 	.poll			= perf_poll,
 	.unlocked_ioctl		= perf_ioctl,
-	.compat_ioctl		= perf_ioctl,
+	.compat_ioctl		= perf_compat_ioctl,
 	.mmap			= perf_mmap,
 	.fasync			= perf_fasync,
 };
#6 lore raw
Quoted 9 lines · Peter Zijlstra

Should be fixed for a while now

---
commit b3f207855f57b9c8f43a547a801340bb5cbc59e5

Author: Pawel Moll <[email protected]>
Date:   Fri Jun 13 16:03:32 2014 +0100

perf: Handle compat ioctl

that works. Can we get this queued to the stable releases (at least as far back as v3.4)?

Thanks,
David

#7 lore raw

Em Wed, Oct 08, 2014 at 09:17:38AM -0600, David Ahern escreveu:

Quoted 8 lines · Peter Zijlstra
Quoted 7 lines

Should be fixed for a while now

---
commit b3f207855f57b9c8f43a547a801340bb5cbc59e5

Author: Pawel Moll <[email protected]>
Date:   Fri Jun 13 16:03:32 2014 +0100
Quoted 5 lines
Quoted 2 lines

perf: Handle compat ioctl

that works. Can we get this queued to the stable releases (at least as far back as v3.4)?

I wonder how that works best, being on the receiving side of such requests from time to time.

I guess that to scale, that would be better done by:

1. Reporter tests if the patch applies (and works) on the desired targets.

2. Reporter sends the request, with the above test results, to [email protected], following whatever conventions are to get the attention of the stable release maintainers.

- Arnaldo

#8 lore raw
Quoted 11 lines · Arnaldo Carvalho de Melo

I wonder how that works best, being on the receiving side of such requests from time to time.

I guess that to scale, that would be better done by:

1. Reporter tests if the patch applies (and works) on the desired targets.

2. Reporter sends the request, with the above test results, to [email protected], following whatever conventions are to get the attention of the stable release maintainers.

My understanding is that subsystem maintainers do the stable requests.

A patch adjustment is needed for v3.4 and v3.10. I am willing to do the necessary patch mods for 3.4, 3.10, and 3.14.

David

#9 lore raw
Quoted 9 lines · David Ahern
Quoted 8 lines · Arnaldo Carvalho de Melo

I wonder how that works best, being on the receiving side of such requests from time to time.

I guess that to scale, that would be better done by:

1. Reporter tests if the patch applies (and works) on the desired targets.

Quoted 6 lines
Quoted 3 lines

2. Reporter sends the request, with the above test results, to [email protected], following whatever conventions are to get the attention of the stable release maintainers.

My understanding is that subsystem maintainers do the stable requests.

At least for perf bits, the usual workflow is that pretty much anyone can make requests for backports that we maintainers missed, and subsystem maintainers look over them and object if they don't like the suggestion. (Obviously you need to Cc:
maintainers.)

In other words, feel free to forward (tested!) backport commit IDs to -stable, with maintainers Cc:-ed, and feel free to provide conflict resolution as well, in cases where they don't apply cleanly.

Quoted 2 lines

A patch adjustment is needed for v3.4 and v3.10. I am willing to do the necessary patch mods for 3.4, 3.10, and 3.14.

That would be useful.

Thanks,

Ingo