Re: [Qemu-devel] [PATCH 1/3] hw/s390x/ipl: Fix alignment problems of S390IPLState members

[Thomas Huth]

On 2018-09-25 17:37, David Hildenbrand wrote:
> On 25/09/2018 17:20, Thomas Huth wrote:
>> The IplParameterBlock and QemuIplParameters structures are declared
>> with QEMU_PACKED, so the compiler assumes that the structures do not
>> need to be aligned in memory. Since the are listed after a "bool"
>> within the S390IPLState, the IplParameterBlock and QemuIplParameters
>> are also indeed mis-aligned in memory. This causes problems on Sparc
>> during migration, since we use VMSTATE_UINT16 in vmstate_iplb to access
>> the devno member for example, and the corresponding migration functions
>> (like qemu_get_be16s) then try to access a 16-bit value from a mis-
>> aligned memory address.
>> The easiest solution to fix this problem is to move the packed structures
>> to the beginning of the S390IPLState. Also add some additional comments
>> here to prevent that this problem will be introduced again in the future.
>>
>> Signed-off-by: Thomas Huth <address@hidden>
>> ---
>>  hw/s390x/ipl.h | 6 ++++--
>>  1 file changed, 4 insertions(+), 2 deletions(-)
>>
>> diff --git a/hw/s390x/ipl.h b/hw/s390x/ipl.h
>> index 4e87b89..f72a82f 100644
>> --- a/hw/s390x/ipl.h
>> +++ b/hw/s390x/ipl.h
>> @@ -132,15 +132,17 @@ typedef struct QemuIplParameters QemuIplParameters;
>>  struct S390IPLState {
>>      /*< private >*/
>>      DeviceState parent_obj;
>> +    /* Packed structs first (to make sure we've got a proper alignment): */
> 
> Is that a sad smiley at the end? ;)
> 
>> +    IplParameterBlock iplb;
>> +    QemuIplParameters qipl;
>> +    /* Other private members without packed attribute: */
>>      uint64_t start_addr;
>>      uint64_t compat_start_addr;
>>      uint64_t bios_start_addr;
>>      uint64_t compat_bios_start_addr;
>>      bool enforce_bios;
>> -    IplParameterBlock iplb;
>>      bool iplb_valid;
>>      bool netboot;
>> -    QemuIplParameters qipl;
>>      /* reset related properties don't have to be migrated or reset */
>>      enum s390_reset reset_type;
>>      int reset_cpu_index;
>>
> 
> Instead of the comment, compile time asserts using offsetof?

Good idea, let's do that instead!

 Thomas