Re: [Patch V3] x86/ldt: correct fpu emulation access to ldt

[Juergen Gross <jgross@suse.com>]

On 08/06/2015 05:15 PM, Andy Lutomirski wrote:
> On Thu, Aug 6, 2015 at 2:21 AM, Juergen Gross <jgross@suse.com> wrote:
>> Commit 37868fe113ff ("x86/ldt: Make modify_ldt synchronous") introduced
>> a new struct ldt_struct anchored at mm->context.ldt.
>>
>> Adapt the x86 fpu emulation to use that new structure.
>>
>> Cc: <stable@vger.kernel.org> # 37868fe113ff: x86/ldt: Make modify_ldt synchronous
>> Cc: <stable@vger.kernel.org> # a5b9e5a2f14f: x86/ldt: Make modify_ldt optional
>> Signed-off-by: Juergen Gross <jgross@suse.com>
>> ---
>>   arch/x86/math-emu/fpu_entry.c   |  3 +--
>>   arch/x86/math-emu/fpu_system.h  | 21 ++++++++++++++++++---
>>   arch/x86/math-emu/get_address.c |  3 +--
>>   3 files changed, 20 insertions(+), 7 deletions(-)
>>
>> diff --git a/arch/x86/math-emu/fpu_entry.c b/arch/x86/math-emu/fpu_entry.c
>> index f37e84a..203318a 100644
>> --- a/arch/x86/math-emu/fpu_entry.c
>> +++ b/arch/x86/math-emu/fpu_entry.c
>> @@ -29,7 +29,6 @@
>>
>>   #include <asm/uaccess.h>
>>   #include <asm/traps.h>
>> -#include <asm/desc.h>
>>   #include <asm/user.h>
>>   #include <asm/fpu/internal.h>
>>
>> @@ -181,7 +180,7 @@ void math_emulate(struct math_emu_info *info)
>>                          math_abort(FPU_info, SIGILL);
>>                  }
>>
>> -               code_descriptor = LDT_DESCRIPTOR(FPU_CS);
>> +               code_descriptor = *FPU_get_ldt_descriptor(FPU_CS);
>>                  if (SEG_D_SIZE(code_descriptor)) {
>>                          /* The above test may be wrong, the book is not clear */
>>                          /* Segmented 32 bit protected mode */
>> diff --git a/arch/x86/math-emu/fpu_system.h b/arch/x86/math-emu/fpu_system.h
>> index 9ccecb6..d4a49d7 100644
>> --- a/arch/x86/math-emu/fpu_system.h
>> +++ b/arch/x86/math-emu/fpu_system.h
>> @@ -16,9 +16,24 @@
>>   #include <linux/kernel.h>
>>   #include <linux/mm.h>
>>
>> -/* s is always from a cpu register, and the cpu does bounds checking
>> - * during register load --> no further bounds checks needed */
>> -#define LDT_DESCRIPTOR(s)      (((struct desc_struct *)current->mm->context.ldt)[(s) >> 3])
>> +#include <asm/desc.h>
>> +#include <asm/mmu_context.h>
>> +
>> +static inline struct desc_struct *FPU_get_ldt_descriptor(unsigned seg)
>> +{
>> +       static struct desc_struct zero_desc;
>> +       struct desc_struct *ret = &zero_desc;
>> +
>> +#ifdef CONFIG_MODIFY_LDT_SYSCALL
>> +       seg >>= 3;
>> +       mutex_lock(&current->mm->context.lock);
>> +       if (current->mm->context.ldt && seg < current->mm->context.ldt->size)
>> +               ret = current->mm->context.ldt->entries + seg;
>> +       mutex_unlock(&current->mm->context.lock);
>> +#endif
>
> Is there a good reason to return a pointer instead of returning struct
> desc_struct directly?  I think that, if you return a pointer, the
> locking is still wrong.  context.ldt can change at any point during
> which IRQs are enabled (unless you hold the mutex), so I don't think
> the mutex is sufficient -- the pointer can become invalid even after
> this function returns.

Aah, of course. Sorry about that.

I just wanted to avoid returning a 8 byte structure on 32 bit. I'll send
V4...


Thanks,

Juergen