On 31.05.2023 22:06, Oleksii wrote:
> On Tue, 2023-05-30 at 18:00 +0200, Jan Beulich wrote:
>>> +static uint32_t read_instr(unsigned long pc)
>>> +{
>>> + uint16_t instr16 = *(uint16_t *)pc;
>>> +
>>> + if ( GET_INSN_LENGTH(instr16) == 2 )
>>> + return (uint32_t)instr16;
>>> + else
>>> + return *(uint32_t *)pc;
>>> +}
>>
>> As long as this function is only used on Xen code, it's kind of okay.
>> There you/we control whether code can change behind our backs. But as
>> soon as you might use this on guest code, the double read is going to
>> be a problem
> Will it be enough to add a comment that read_instr() should be used
> only on Xen code? Or it is needed to introduce some lock?
A comment will do for now. A lock would be problematic: It won't help
when the function is used on non-Xen code, and since you use this in
exception handling you may deadlock unless you carefully use a
recursive lock.
>> (I think; I wonder how hardware is supposed to deal with
>> the situation: Maybe they indeed fetch in 16-bit quantities?).
> I thought that it reads amount of bytes corresponded to i-cache size
> and then the pipeline tracks whether an instruction is 16 or 32 bit.
And what if an insn spans a cacheline boundary?
Jan
