Re: [PATCH] add attribute none to pthread_setspecific (BZ #27714)

[Martin Sebor via Libc-alpha <libc-alpha@sourceware.org>]

On 4/28/21 1:32 AM, Florian Weimer wrote:
> * Martin Sebor:
> 
>> diff --git a/nptl/tst-tsd3.c b/nptl/tst-tsd3.c
>> index 0dd39ccb2b..45c7e4e1ea 100644
>> --- a/nptl/tst-tsd3.c
>> +++ b/nptl/tst-tsd3.c
>> @@ -37,7 +37,8 @@ destr1 (void *arg)
>>       {
>>         puts ("set key2");
>>   
>> -      if (pthread_setspecific (key2, (void *) 1l) != 0)
>> +      /* Use an arbirary but valid pointer to avoid GCC warnings.  */
>> +      if (pthread_setspecific (key2, (void *) &left) != 0)
>>   	{
>>   	  puts ("destr1: setspecific failed");
>>   	  exit (1);
>> @@ -53,7 +54,8 @@ destr2 (void *arg)
>>       {
>>         puts ("set key1");
>>   
>> -      if (pthread_setspecific (key1, (void *) 1l) != 0)
>> +      /* Use an arbirary but valid pointer to avoid GCC warnings.  */
>> +      if (pthread_setspecific (key1, (void *) &left) != 0)
>>   	{
>>   	  puts ("destr2: setspecific failed");
>>   	  exit (1);
> 
> Sorry, this is clearly a bug in attribute access (none).  No access
> should mean no access, not access to one byte, as the warning currently
> implies.

It's not a bug.  Access none was introduced as a check whether
the object has the expected size without assuming it's written
to or read from (each access mode has its own checks).  Pointers
to void are treated as char*.  This was documented in the GCC 10
manual when attribute access was first added and hasn't changed
with GCC 11.

The motivating use case for access none was a Linux kernel API
(I think check_copy_size) used to verify that a pointer points
to an object with the expected size.

> 
> Please fix this for GCC 11.2 and adjust the glibc version check for the
> none variant of the attribute.

I'll see if I can come up with a solution for
the pthread_setspecific use case.

Martin