Re: [PATCH 03/15] sparc: Use Linux kABI for syscall return

[Adhemerval Zanella <adhemerval.zanella@linaro.org>]

On 11/02/2020 16:24, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>> On 11/02/2020 08:15, Florian Weimer wrote:
>>> * Adhemerval Zanella:
>>>
>>>> +	if (__glibc_unlikely (__g1 != 0)) 				\
>>>
>>> This change is inconsistent with the other updates, which use __g1 ==
>>> -1.  Is this deliberate?
>>>
>>> Thanks,
>>> Florian
>>>
>>
>> In fact __SYSCALL_STRING already sets the 'o0' to a negative value if
>> the 'xcc' condition is set (indicating that the syscall has failed).
>> The 'g1' check is superfluous, it will be always true since 'g1' will
>> be either 0 or 1. 
>>
>> And both the set and check of 'g1' result is also superfluous, since 'o0' 
>> will already hold all the required information.
>>
>> Below is an updated patch, checked on sparc64-linux-gnu and 
>> sparcv9-linux-gnu.
> 
> I see, nice additional cleanup.
> 
>> It changes the sparc internal_syscall* macros to return a negative
>> value instead the 'g1' register value on 'err' macro argument.
>                ^ of                     ^ in the?
> 
> 

Ack.

>> The __SYSCALL_STRING macro is also changed to no set the 'g1'
>                                                 ^ not
>> value, since 'o1' already holds all the required information
>> to check if syscall has failed.
>>
>> The macro INTERNAL_SYSCALL_DECL is no longer required, and the
>> INTERNAL_SYSCALL_ERROR_P follows the other Linux kABIS.  The
>                           ^ macro?                 ^ kABIs

Ack.

>                 (or drop the “the“ on the preceding line)
> 
>> redefinition of INTERNAL_VSYSCALL_CALL is also no longer
>> required.
>>
>> Checked on sparc64-linux-gnu and sparcv9-linux-gnu. It fixes
>> the sporadic issues on sparc32 where clock_nanosleep does not
>> act as cancellation entrypoint.
> 
> I double-checked this against the kernel sources, and entry.S has
> this:
> 
> ret_sys_call:
>         ld      [%curptr + TI_FLAGS], %l6
>         cmp     %o0, -ERESTART_RESTARTBLOCK
>         ld      [%sp + STACKFRAME_SZ + PT_PSR], %g3
>         set     PSR_C, %g2
>         bgeu    1f
> 
> But ERESTART_RESTARTBLOCK is not 4095, so glibc with this change will
> now treat certain internal kernel error codes as errors, while they
> were previously reported as success.  This looks like a kernel bug, in
> that ERESTART_RESTARTBLOCK was not updated when more error codes were
> added.  On the other hand, these error codes should never leak into
> userspace.

My understanding is such errors should not be visible by the application,
as indicated by include/linux/errno.h comment. And it seems to be the
case for sparc, at least on:

arch/sparc/kernel/signal_64.c
477 static void do_signal(struct pt_regs *regs, unsigned long orig_i0)
[...]
514         restart_syscall = 0;
515         if (pt_regs_is_syscall(regs) &&
516             (regs->tstate & (TSTATE_XCARRY | TSTATE_ICARRY))) {
517                 restart_syscall = 1; 
518                 orig_i0 = regs->u_regs[UREG_G6];
519         }
520
521         if (has_handler) {
522                 if (restart_syscall)
523                         syscall_restart(orig_i0, regs, &ksig.ka.sa);                               
524                 signal_setup_done(setup_rt_frame(&ksig, regs), &ksig, 0);                          
525         } else {
526                 if (restart_syscall) {                                                             
527                         switch (regs->u_regs[UREG_I0]) {                                           
528                         case ERESTARTNOHAND:
529                         case ERESTARTSYS:
530                         case ERESTARTNOINTR:                                                       
531                                 /* replay the system call when we are done */                      
532                                 regs->u_regs[UREG_I0] = orig_i0;                                   
533                                 regs->tpc -= 4;
534                                 regs->tnpc -= 4;                                                   
535                                 pt_regs_clear_syscall(regs);                                       
536                                 /* fall through */                                                 
537                         case ERESTART_RESTARTBLOCK:                                                
538                                 regs->u_regs[UREG_G1] = __NR_restart_syscall;                      
539                                 regs->tpc -= 4;                                                    
540                                 regs->tnpc -= 4;                                                   
541                                 pt_regs_clear_syscall(regs);                                       
542                         }
543                 }
544             

If signal has a handler, syscall_restart will either set EINTR or
previous 'o0' value.  Otherwise if syscall should be restarted,
either it will be trying again (line 538) or previous error code
would be set.

> 
> To me, your patch looks good.
>