mttcg: warning to user about memory ordering

mttcg: warning to user about memory ordering

[Pierrick Bouvier]

Hello,

When using multi thread tcg, a warning is emitted to user if models for 
guest and host are not compatible (like qemu-system-x86_64 running on 
aarch64 host)

accel/tcg/tcg-all.c:
161             if (!check_tcg_memory_orders_compatible()) {
162                 warn_report("Guest expects a stronger memory ordering "
163                             "than the host provides");
164                 error_printf("This may cause strange/hard to debug 
errors\n");
165             }

However, in tcg/tcg-op.c, all load/stores are emitting memory barriers 
using tcg_gen_req_mo, which lets think it's already implemented for any 
host/guest combination (was added in b32dc3).

In your opinion, are there still corner cases that are not covered 
correctly? If no, should we drop this scary warning?

Regards,
Pierrick

Re: mttcg: warning to user about memory ordering

[Richard Henderson]

On 2/22/23 06:28, Pierrick Bouvier wrote:
> Hello,
> 
> When using multi thread tcg, a warning is emitted to user if models for guest and host are 
> not compatible (like qemu-system-x86_64 running on aarch64 host)
> 
> accel/tcg/tcg-all.c:
> 161             if (!check_tcg_memory_orders_compatible()) {
> 162                 warn_report("Guest expects a stronger memory ordering "
> 163                             "than the host provides");
> 164                 error_printf("This may cause strange/hard to debug errors\n");
> 165             }
> 
> However, in tcg/tcg-op.c, all load/stores are emitting memory barriers using 
> tcg_gen_req_mo, which lets think it's already implemented for any host/guest combination 
> (was added in b32dc3).
> 
> In your opinion, are there still corner cases that are not covered correctly? If no, 
> should we drop this scary warning?

See

https://lore.kernel.org/qemu-devel/20210316220735.2048137-1-richard.henderson@linaro.org/
https://gitlab.com/rth7680/qemu/-/commits/tcg-mb-small/

It never got merged because it didn't actually solve the s390x on aarch64 problem.  I 
assume there's something I've missed.


r~

Re: mttcg: warning to user about memory ordering

[Pierrick Bouvier]

On 2/22/23 20:14, Richard Henderson wrote:
> On 2/22/23 06:28, Pierrick Bouvier wrote:
>> Hello,
>>
>> When using multi thread tcg, a warning is emitted to user if models for guest and host are
>> not compatible (like qemu-system-x86_64 running on aarch64 host)
>>
>> accel/tcg/tcg-all.c:
>> 161             if (!check_tcg_memory_orders_compatible()) {
>> 162                 warn_report("Guest expects a stronger memory ordering "
>> 163                             "than the host provides");
>> 164                 error_printf("This may cause strange/hard to debug errors\n");
>> 165             }
>>
>> However, in tcg/tcg-op.c, all load/stores are emitting memory barriers using
>> tcg_gen_req_mo, which lets think it's already implemented for any host/guest combination
>> (was added in b32dc3).
>>
>> In your opinion, are there still corner cases that are not covered correctly? If no,
>> should we drop this scary warning?
> 
> See
> 
> https://lore.kernel.org/qemu-devel/20210316220735.2048137-1-richard.henderson@linaro.org/
> https://gitlab.com/rth7680/qemu/-/commits/tcg-mb-small/
> 
> It never got merged because it didn't actually solve the s390x on aarch64 problem.  I
> assume there's something I've missed.
> 

Thanks for the pointers.

> 
> r~