Intel TSX: xbegin always returns 0

Question

I am writing some programs which contains RTM transactional regions. I have checked my machine supports HLE and RTM using cpuid instruction. However, I notice that the xbegin instruction always returns zero (the value in rax register), which indicates that the transaction is aborted. So I write a quite simple demo using xbegin and xend instruction, and do not insert any instructions between them.

The code described above is:

int main() {
    asm volatile(".byte 0xc7, 0xf8, 0x03, 0x00, 0x00, 0x00" ::"r"(-1ul)); // xbegin machine code
    asm volatile(".byte 0x0f, 0x01, 0xd5"); // xend machine code
    return 0;
}

The corresponding assembly code is:

00000000000005fa <main>:
 5fa:   48 c7 c0 ff ff ff ff    mov    $0xffffffffffffffff,%rax
 601:   c7 f8 03 00 00 00       xbeginq 60a <main+0x10>
 607:   0f 01 d5                xend   
 60a:   b8 00 00 00 00          mov    $0x0,%eax
 60f:   c3                      retq   

In this code, I set the fallback address in xbegin with the address of the next instruction of xend. So whether the transaction aborts or not, the code can finish execution normally. Then I use perf to profile the execution of this code, and I find that this transaction aborts due to HLE-unfriendly instructions. I don't know the reason.

$ sudo perf stat -e rtm_retired.start ./demo     
                                                                                                                                               
 Performance counter stats for './demo':
                              
                 1      rtm_retired.start

       0.000614539 seconds time elapsed                                               
                                                                                                 
       0.000634000 seconds user
       0.000000000 seconds sys 

$ sudo perf stat -e rtm_retired.commit ./demo                                                                                        
                                                                                                                                                                             
 Performance counter stats for './demo':                                                                                                                                     
                                                                                                                                                                             
                 0      rtm_retired.commit                                                                                                                                   
                                                                                                                                                                             
       0.000657118 seconds time elapsed                                                                                                                                      
                                                                                                                                                                             
       0.000616000 seconds user         
       0.000000000 seconds sys

$ sudo perf stat -e rtm_retired.aborted_unfriendly ./demo

 Performance counter stats for './demo':

                 1      rtm_retired.aborted_unfriendly                                   

       0.000617362 seconds time elapsed

       0.000609000 seconds user
       0.000000000 seconds sys

My machine configuration:

CPU: Intel(R) Core(TM) i7-6700 CPU @ 3.40GHz
microcode: 0xf0
OS: Linux 5.4.0-150-generic, Ubuntu 18.04.1

Update: I have noticed an article, which means that Intel has disabled TSX through a microcode update. They added an new CPUID bit to indicate this change. I have checked this bit and found that it is set on my processor. So the RTM regions cannot be executed successfully.

Answer 1

I have noticed an article, which means that Intel has disabled TSX through a microcode update. They added an new CPUID bit to indicate this change. I have checked this bit and found that it is set on my processor. So the RTM regions cannot be executed successfully.