Le 05/12/2020 à 10:56, Christophe Leroy a écrit :
Le 05/12/2020 à 09:48, Christoph Hellwig a écrit :
On Sat, Dec 05, 2020 at 08:43:06AM +0000, Christophe Leroy wrote:
Since commit c33165253492 ("powerpc: use non-set_fs based maccess
routines"), userspace access is not granted anymore when using
copy_from_kernel_nofault()
However, kthread_probe_data() uses copy_from_kernel_nofault()
to check validity of pointers. When the pointer is NULL,
it points to userspace, leading to a KUAP fault and triggering
the following big hammer warning many times when you request
a sysrq "show task":
To avoid that, copy_from_kernel_nofault_allowed() is used to check
whether the address is a valid kernel address. But the default
version of it returns true for any address.
Provide a powerpc version of copy_from_kernel_nofault_allowed()
that returns false when the address is below TASK_USER_MAX,
so that copy_from_kernel_nofault() will return -ERANGE.
Looks good. I wonder if we should just default to the TASK_SIZE_MAX
check in copy_from_kernel_nofault_allowed for architectures that select
CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE?
Yes maybe that would be better.
Can you cook a patch an get it into 5.10 ?
In fact it doesn't seem so easy because only s390, powerpc and x86 have TASK_SIZE_MAX while
CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE is selected by arm, arm64, powerpc and x86
So maybe for 5.10 we take the powerpc fix ?
Christophe
