Re: [PATCH 2/3] RISC-V: use memcpy for kexec_file mode

[ebiederm@xmission.com (Eric W. Biederman)]

Björn Töpel <bjorn.topel@gmail.com> writes:

> On Sat, 30 Oct 2021 at 05:51, Eric W. Biederman <ebiederm@xmission.com> wrote:
>>
>> Liao Chang <liaochang1@huawei.com> writes:
>>
>> > The pointer to buffer loading kernel binaries is in kernel space for
>> > kexec_fil mode, When copy_from_user copies data from pointer to a block
>> > of memory, it checkes that the pointer is in the user space range, on
>> > RISCV-V that is:
>> >
>> > static inline bool __access_ok(unsigned long addr, unsigned long size)
>> > {
>> >       return size <= TASK_SIZE && addr <= TASK_SIZE - size;
>> > }
>> >
>> > and TASK_SIZE is 0x4000000000 for 64-bits, which now causes
>> > copy_from_user to reject the access of the field 'buf' of struct
>> > kexec_segment that is in range [CONFIG_PAGE_OFFSET - VMALLOC_SIZE,
>> > CONFIG_PAGE_OFFSET), is invalid user space pointer.
>> >
>> > This patch fixes this issue by skipping access_ok(), use mempcy() instead.
>>
>> I am a bit confused.
>>
>> Why is machine_kexec ever calling copy_from_user?  That seems wrong in
>> all cases.
>>
>
> It's not machine_kexec -- it's machine_kexec_prepare, which pulls out
> the FDT from the image. It looks like MIPS does it similarly.
>
> (Caveat: I might be confused as well! ;-))

True it is machine_kexec_prepare.  But still.  copy_from_user does not
belong in there.  It is not passed a userspace pointer.

This looks more like a case for kmap to read a table from the firmware.

Even if it someone made sense it definitely does not make sense to
make it a conditional copy_from_user.  That way lies madness.

The entire change is a smell that there is some abstraction that is
going wrong, and that abstraction needs to get fixed.

Eric