Re: [Qemu-devel] [PATCH v2] mips: Decide to map PAGE_EXEC in map_address

[Jakub Jermar]

Hi Philippe,

On 5/16/19 6:31 PM, Philippe Mathieu-Daudé wrote:
> Hi Jakub,
> 
> On 5/16/19 3:10 PM, Jakub Jermar wrote:
>> Hi,
>>
>> On 5/3/19 12:02 PM, Jakub Jermar wrote:
>>> Hi,
>>>
>>> On 4/23/19 4:58 PM, Jakub Jermar wrote:
>>>> Hi Philippe!
>>>>
>>>> On 4/23/19 3:48 PM, Philippe Mathieu-Daudé wrote:
>>>>> Hi Jakub,
>>>>>
>>>>> On 4/23/19 1:00 PM, Jakub Jermář wrote:
>>>>>> This commit addresses QEMU Bug #1825311:
>>>>>>
>>>>>>   mips_cpu_handle_mmu_fault renders all accessed pages executable
>>>>>>
>>>>>> It allows finer-grained control over whether the accessed page should be
>>>>>> executable by moving the decision to the underlying map_address
>>>>>> function, which has more information for this.
>>>>>>
>>>>>> As a result, pages that have the XI bit set in the TLB and are accessed
>>>>>> for read/write, don't suddenly end up being executable.
>>>>>>
>>>>>
>>>>> Fixes: https://bugs.launchpad.net/qemu/+bug/1825311
>>>>>
>>>>>> Signed-off-by: Jakub Jermář <address@hidden>
>>>>>> ---
>>>>>>  target/mips/helper.c | 17 ++++++++++-------
>>>>>>  1 file changed, 10 insertions(+), 7 deletions(-)
>>>>>>
>>>>>> diff --git a/target/mips/helper.c b/target/mips/helper.c
>>>>>> index c44cdca3b5..132d073fbe 100644
>>>>>> --- a/target/mips/helper.c
>>>>>> +++ b/target/mips/helper.c
>>>>>> @@ -43,7 +43,7 @@ int no_mmu_map_address (CPUMIPSState *env, hwaddr 
>>>>>> *physical, int *prot,
>>>>>>                          target_ulong address, int rw, int access_type)
>>>>>>  {
>>>>>>      *physical = address;
>>>>>> -    *prot = PAGE_READ | PAGE_WRITE;
>>>>>> +    *prot = PAGE_READ | PAGE_WRITE | PAGE_EXEC;
>>>>>>      return TLBRET_MATCH;
>>>>>>  }
>>>>>>  
>>>>>> @@ -61,7 +61,7 @@ int fixed_mmu_map_address (CPUMIPSState *env, hwaddr 
>>>>>> *physical, int *prot,
>>>>>>      else
>>>>>>          *physical = address;
>>>>>>  
>>>>>> -    *prot = PAGE_READ | PAGE_WRITE;
>>>>>> +    *prot = PAGE_READ | PAGE_WRITE | PAGE_EXEC;
>>>>>>      return TLBRET_MATCH;
>>>>>>  }
>>>>>>  
>>>>>> @@ -101,6 +101,9 @@ int r4k_map_address (CPUMIPSState *env, hwaddr 
>>>>>> *physical, int *prot,
>>>>>>                  *prot = PAGE_READ;
>>>>>>                  if (n ? tlb->D1 : tlb->D0)
>>>>>>                      *prot |= PAGE_WRITE;
>>>>>> +                if (!(n ? tlb->XI1 : tlb->XI0)) {
>>>>>> +                    *prot |= PAGE_EXEC;
>>>>>> +                }
>>>>>
>>>>> This was indeed missed in commit 2fb58b73746e.
> 
> Aleksandar, if this patch is OK with you, can you amend this comment,
> and add the "Fixes:" tag too when applying? Thanks!
> 
> Reviewed-by: Philippe Mathieu-Daudé <address@hidden>
> 
>>>>>
>>>>>>                  return TLBRET_MATCH;
>>>>>>              }
>>>>>>              return TLBRET_DIRTY;
>>>>>> @@ -182,7 +185,7 @@ static int get_seg_physical_address(CPUMIPSState 
>>>>>> *env, hwaddr *physical,
>>>>>>      } else {
>>>>>>          /* The segment is unmapped */
>>>>>>          *physical = physical_base | (real_address & segmask);
>>>>>> -        *prot = PAGE_READ | PAGE_WRITE;
>>>>>> +        *prot = PAGE_READ | PAGE_WRITE | PAGE_EXEC;
>>>>>>          return TLBRET_MATCH;
>>>>>>      }
>>>>>>  }
>>>>>> @@ -913,8 +916,8 @@ int mips_cpu_handle_mmu_fault(CPUState *cs, vaddr 
>>>>>> address, int size, int rw,
>>>>>>      }
>>>>>>      if (ret == TLBRET_MATCH) {
>>>>>>          tlb_set_page(cs, address & TARGET_PAGE_MASK,
>>>>>> -                     physical & TARGET_PAGE_MASK, prot | PAGE_EXEC,
>>>>>> -                     mmu_idx, TARGET_PAGE_SIZE);
>>>>>> +                     physical & TARGET_PAGE_MASK, prot, mmu_idx,
>>>>>> +                     TARGET_PAGE_SIZE);
>>>>>>          ret = 0;
>>>>>>      } else if (ret < 0)
>>>>>>  #endif
>>>>>> @@ -936,8 +939,8 @@ int mips_cpu_handle_mmu_fault(CPUState *cs, vaddr 
>>>>>> address, int size, int rw,
>>>>>>                                             address, rw, access_type, 
>>>>>> mmu_idx);
>>>>>>                  if (ret == TLBRET_MATCH) {
>>>>>>                      tlb_set_page(cs, address & TARGET_PAGE_MASK,
>>>>>> -                            physical & TARGET_PAGE_MASK, prot | 
>>>>>> PAGE_EXEC,
>>>>>> -                            mmu_idx, TARGET_PAGE_SIZE);
>>>>>> +                            physical & TARGET_PAGE_MASK, prot, mmu_idx,
>>>>>> +                            TARGET_PAGE_SIZE);
>>>>>>                      ret = 0;
>>>>>>                      return ret;
>>>>>>                  }
>>>>>>
>>>>>
>>>>> Your patch looks correct, but I'd like to test it.
>>>>> Do you have a reproducer?
>>>>> Can you describe the command line you used?
>>>>
>>>> I've just attached a reproducer image and script to the bug. It's a
>>>> 32-bit little-endian test binary running on top of the L4Re microkernel.
> 
> I can't get the "TAP" output you described on launchpad.

Rarely the entire guest gets stuck early during boot at the following line:

Calibrating timer loop...

This is an unrelated problem, not sure why that happens.

If that is the case, just rerun the reproducer until you get either two
lines beginning with "ok" or two lines beginning with "not ok" between
TAP TEST START and TAP TEST FINISHED. We want to see two "ok" lines with
the patched version and "not ok" lines with the unpatched version.

Also note that you might have to wait a couple of seconds in the case of
the unpatched version as the individual tests terminate via timeouts then.

Otherwise please let me know what is the actual output you get there.

> 
>>>> Let me know if you also need a 64-bit version.
> 
> 64-bit version is welcomed.

Okay, I uploaded the 64-bit reproducer to the bug report. It works in a
similar way like the 32-bit version.

Cheers,
Jakub

> 
>>>> I tested both 32 and 64-bit versions of the reproducer and also checked
>>>> to see that the the other images I have lying around here (Linux 2.6.32
>>>> big endian and HelenOS master little-endian, both 32-bit for 4Kc)
>>>> continue to run without regressions.
> 
> Yes, definitively an improvement:
> Tested-by: Philippe Mathieu-Daudé <address@hidden>
> 
> Regards,
> 
> Phil.
> 

-- 
Kernkonzept GmbH at Dresden, Germany, HRB 31129, CEO Dr.-Ing. Michael
Hohmuth