On 19/10/2017 11:34, Tulio Magno Quites Machado Filho wrote:
> Forwarding some comments from Adhemerval sent to libc-alpha [1]...
>
> Adhemerval Zanella <adhemerval.zane...@linaro.org> writes:
>> Florian Weimer <fwei...@redhat.com> writes:
>>
>>> On 10/12/2017 12:17 PM, Michael Ellerman wrote:
>>>> + pr_info("Enabling TM (Transactional Memory) with Suspend Disabled\n");
>>>> + cur_cpu_spec->cpu_features |= CPU_FTR_TM;
>>>> + cur_cpu_spec->cpu_user_features2 |= PPC_FEATURE2_HTM_NO_SUSPEND;
>>>> + tm_suspend_disabled = true;
>>>
>>> This doesn't look right because if suspend is not available, you need to
>>> clear the original PPC_FEATURE2_HTM flag because the semantics are not
>>> right, so that applications can use fallback code. Otherwise,
>>> applications may incorrectly select the HTM code and break if running on
>>> a system which supports HTM, but without the suspend state.
>>>
>>> The new flag should say that HTM is supported, but without the suspend
>>> state, and it should be always set if PPC_FEATURE2_HTM is set.
>>
>> Will it also change TEXARS with the abort information?
>
> It should, with a permanent error cause so that old applications entering
> suspended state can adopt another technique.
> Michael, could you clarify if this is indeed happening, please?
>
>> I completely agree with Florian here, this is as *ABI* change
>> and the kernel need to advertise a different TM ABI instead
>> of as an extension.
>
> Adhemerval, could you elaborate which problems you're foreseeing, please?
>
Pretty much the same Florian already stated: an application can not any
more assume for instance:
tsr. 0
mfcr r9,128
andis. r10,r9,0x4000
be cr0,L(suspend)
andis. r10,r9,0x2000
be cr0,L(transactional)
However thinking more about it I am not sure if this should be really a
problem: on default HTM mode the program must handle self-induced failures
as the tbegin. failure path and I assume trying to suspend/resume in this
case will trigger this. For instance:
if (__builtin_tbegin (0))
{
/* some transactional stuff. */
__builtin_tsuspend ();
/* non transactional stuff. */
__builtin_tresume ();
/* more transactional stuff. */
}
else
{
/* fall-out code. */
}
So I assume for these chips without suspend/resume support the example
code will always run the fall-out code.
