New compat_alloc_user_space() GPLness and fglrx
commit c41d68a513c71e35a14f66d71782d27a79a81ea6
compat: Make compat_alloc_user_space() incorporate the access_ok()
promotes compat_alloc_user_space() to EXPORT_SYMBOL_GPL, and moves it from asm/compat.h to linux/compat.h.
This breaks AMD/ATI's proprietary fglrx GPU driver, which wants to use compat_alloc_user_space(), and expects to find it in asm/compat.h.
Fixing this on the fglrx side is easy enough, one can simply audit the two or three call sites in the fglrx glue and make sure they can use arch_compat_alloc_user_space() safely.
But such a fix doesn't look right. IMHO, either we don't want any non-GPL module touching *_compat_alloc_user_space and compat_alloc_user_space, or it would be better to let them use compat_alloc_user_space() in the interest of overall security...
Signature · 4 lines
"One disk to rule them all, One disk to find them. One disk to bring them all and in the darkness grind them. In the Land of Redmond where the shadows lie." -- The Silicon Valley Tarot Henrique Holschuh
Quoted 6 lines · On Saturday 18 September 2010 03:42:32 Henrique de Moraes Holschuh wrote:
promotes
compat_alloc_user_space()toEXPORT_SYMBOL_GPL, and moves it fromasm/compat.htolinux/compat.h.This breaks AMD/ATI's proprietary fglrx GPU driver, which wants to use
compat_alloc_user_space(), and expects to find it inasm/compat.h.
Quoted 8 lines
Fixing this on the fglrx side is easy enough, one can simply audit the two or three call sites in the fglrx glue and make sure they can use
arch_compat_alloc_user_space()safely.But such a fix doesn't look right. IMHO, either we don't want any non-GPL module touching
*_compat_alloc_user_spaceandcompat_alloc_user_space, or it would be better to let them usecompat_alloc_user_space()in the interest of overall security...
Right, don't use arch_compat_alloc_user_space(), that would be worse.
compat_alloc_user_space is still a horrible hack and shouldn't be used at all if possible. Just fix the code to either not require compat ioctl handling or split out the common code from the native ioctl function and pass it a kernel data structure, which you copy from user space in different ways depending on where you come from.
Arnd
Quoted 16 lines · Arnd Bergmann
Quoted 6 lines · On Saturday 18 September 2010 03:42:32 Henrique de Moraes Holschuh wrote:
promotes
compat_alloc_user_space()toEXPORT_SYMBOL_GPL, and moves it fromasm/compat.htolinux/compat.h.This breaks AMD/ATI's proprietary fglrx GPU driver, which wants to use
compat_alloc_user_space(), and expects to find it inasm/compat.h.Quoted 8 lines
Fixing this on the fglrx side is easy enough, one can simply audit the two or three call sites in the fglrx glue and make sure they can use
arch_compat_alloc_user_space()safely.But such a fix doesn't look right. IMHO, either we don't want any non-GPL module touching
*_compat_alloc_user_spaceandcompat_alloc_user_space, or it would be better to let them usecompat_alloc_user_space()in the interest of overall security...
Quoted 7 lines
Right, don't use
arch_compat_alloc_user_space(), that would be worse.
compat_alloc_user_spaceis still a horrible hack and shouldn't be used at all if possible. Just fix the code to either not require compat ioctl handling or split out the common code from the native ioctl function and pass it a kernel data structure, which you copy from user space in different ways depending on where you come from.
Well, that means ATI needs to mess with the insides of their binary blob... Anyone knows of a channel to talk to them? Someone must tell them about this, I am just an unwilling user of fglrx (R600 support in the open drivers is nowhere near what I need yet).
Meanwhile, what is the solution? Keep using (the newly named) arch_compat_alloc_user_space()?
Signature · 4 lines
"One disk to rule them all, One disk to find them. One disk to bring them all and in the darkness grind them. In the Land of Redmond where the shadows lie." -- The Silicon Valley Tarot Henrique Holschuh
Quoted 7 lines · Henrique de Moraes Holschuh
Well, that means ATI needs to mess with the insides of their binary blob... Anyone knows of a channel to talk to them? Someone must tell them about this, I am just an unwilling user of fglrx (R600 support in the open drivers is nowhere near what I need yet).
Meanwhile, what is the solution? Keep using (the newly named)
arch_compat_alloc_user_space()?
Since you don't know how that is used in the driver, that could very well keep the security hole open as long as the driver is loaded.
You still have the obvious options that you won't like:
1. stay on an older (broken) kernel
2. don't use fglrx until this is fixed
3. copy the new compat_alloc_user_space as an inline function into the wrapper module (possibly making it undistributable)
Arnd