On Mon, Oct 18, 2021 at 04:58:03PM +0200, Takashi Iwai wrote:
On Mon, 18 Oct 2021 16:43:00 +0200, Rich Felker wrote:
On Sun, Oct 10, 2021 at 09:53:38AM +0200, Takashi Iwai wrote:
On Fri, 08 Oct 2021 14:07:39 +0200, Rich Felker wrote:
On Fri, Oct 08, 2021 at 01:11:34PM +0200, Takashi Iwai wrote:
On Fri, 08 Oct 2021 11:24:39 +0200, Arnd Bergmann wrote:
On Fri, Oct 8, 2021 at 10:43 AM Takashi Iwai tiwai@suse.de wrote: > On Thu, 07 Oct 2021 18:51:58 +0200, Rich Felker wrote: > > On Thu, Oct 07, 2021 at 06:18:52PM +0200, Takashi Iwai wrote: > > @@ -557,11 +558,15 @@ struct __snd_pcm_sync_ptr { > #if defined(__BYTE_ORDER) ? __BYTE_ORDER == __BIG_ENDIAN : defined(__BIG_ENDIAN) > typedef char __pad_before_uframe[sizeof(__u64) - sizeof(snd_pcm_uframes_t)]; > typedef char __pad_after_uframe[0]; > +typedef char __pad_before_u32[4]; > +typedef char __pad_after_u32[0]; > #endif > > #if defined(__BYTE_ORDER) ? __BYTE_ORDER == __LITTLE_ENDIAN : defined(__LITTLE_ENDIAN) > typedef char __pad_before_uframe[0]; > typedef char __pad_after_uframe[sizeof(__u64) - sizeof(snd_pcm_uframes_t)]; > +typedef char __pad_before_u32[0]; > +typedef char __pad_after_u32[4]; > #endif
I think these should remain unchanged, the complex expression was intentionally done so the structures are laid out the same way on 64-bit architectures, so that the kernel can use the __SND_STRUCT_TIME64 path internally on both 32-bit and 64-bit architectures.
That was explicitly defined, but OK, this isn't necessarily defined here.
> @@ -2970,8 +2981,17 @@ static int snd_pcm_sync_ptr(struct snd_pcm_substream *substream, > memset(&sync_ptr, 0, sizeof(sync_ptr)); > if (get_user(sync_ptr.flags, (unsigned __user *)&(_sync_ptr->flags))) > return -EFAULT; > - if (copy_from_user(&sync_ptr.c.control, &(_sync_ptr->c.control), sizeof(struct snd_pcm_mmap_control))) > - return -EFAULT; > + if (buggy_control) { > + if (copy_from_user(&sync_ptr.c.control_api_2_0_15, > + &(_sync_ptr->c.control_api_2_0_15), > + sizeof(sync_ptr.c.control_api_2_0_15))) > + return -EFAULT; > + } else { > + if (copy_from_user(&sync_ptr.c.control, > + &(_sync_ptr->c.control), > + sizeof(sync_ptr.c.control))) > + return -EFAULT; > + }
The problem I see with this is that it might break musl's ability to emulate the new interface on top of the old (time32) one for linux-4.x and older kernels, as the conversion function is no longer stateless but has to know the negotiated interface version.
It's probably fine as long as we can be sure that the 2.0.16+ API version only gets negotiated if both the kernel and user sides support it, and musl only emulates the 2.0.15 API version from the current kernels.
I've tried to understand this part of musl's convert_ioctl_struct(), but I just can't figure out whether it does the conversion based the on the layout that is currently used in the kernel, or based on the layout we should have been using, and would use with the above fix. Rich, can you help me here?
So, at this moment, I'm not sure whether we should correct the struct at all. This will lead to yet more breakage, and basically the struct itself *works* -- the only bug is in 32bit compat handling in the kernel (again).
The below is a revised kernel patch (again untested), just correcting the behavior of 32bit compat mode. 32bit apps on 32bit kernel work fine as is, as well as 64bit apps on 64bit kernel.
I'm perfectly okay with this if Arnd is! It's probably the least invasive and has the least long-term maintenance cost and fallout on other projects.
OK, I'll submit a proper patch now, to be included in the next PR for 5.15-rc. For further fixes, let's think carefully.
Am I correct in my understanding that the fix of keeping the "broken" definition (and having the 64-bit kernel honor it for 32-bit binaries) has been accepted?
Yes, as it was already set in stone, we accept the broken definition as is.
Since musl's translation for pre-time64 kernels seems to have been using the "non-broken" definition, I think completing the fix requires a change in musl too.
Hm, musl translator contains the own definition of ioctl?
If so, we may reconsider about renumbering ioctls altogether.
No, I don't think so. The musl translator is to translate between the time64 ioctl structures and the old time32 ones for the sake of executing on an old kernel. Up til now, it has been broken comparably to how 32-bit binaries running in compat mode on a 64-bit kernel were broken: the code in musl translated the time64 structure to (and back from) the time32 one assuming the intended padding. But the application was using the actual kernel uapi struct where the padding was (and still is) illogical. Thus, nothing was built with the wrong ABI; it's only the musl-internal translation logic that was wrong (and only pre-time64 kernels are affected).
The attached patch should fix it, I think.
Rich