drivers: char: mem: Check {read,write}_kmem() addresses
authorRobin Murphy <robin.murphy@arm.com>
Wed, 1 Jun 2016 18:21:42 +0000 (19:21 +0100)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 31 Aug 2016 13:21:18 +0000 (15:21 +0200)
commit148a1bc84398039e2b96ff78678c4d9a67f81452
tree7c477963ee32e086a8d374b70b0015b93ffc5601
parentd61f3088ea79c1f11fc8fe869d6f90a4cc97a452
drivers: char: mem: Check {read,write}_kmem() addresses

Arriving at read_kmem() with an offset representing a bogus kernel
address (e.g. 0 from a simple "cat /dev/kmem") leads to copy_to_user
faulting on the kernel-side read.

x86_64 happens to get away with this since the optimised implementation
uses "rep movs*", thus the user write (which is allowed to fault) and
the kernel read are the same instruction, the kernel-side fault falls
into the user-side fixup handler and the chain of events which
transpires ends up returning an error as one might expect, even if it's
an inappropriate -EFAULT. On other architectures, though, the read is
not covered by the fixup entry for the write, and we get a big scary
"Unable to hande kernel paging request..." dump.

The more typical use-case of mmap_kmem() has always (within living
memory at least) returned -EIO for addresses which don't satisfy
pfn_valid(), so let's make that consistent across {read,write}_kem()
too.

Reported-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/char/mem.c