[PATCH v6 07/13] ARC: Linux Syscall Interface

Adhemerval Zanella adhemerval.zanella@linaro.org
Thu Jun 4 11:06:06 GMT 2020



On 03/06/2020 17:17, Vineet Gupta wrote:
> On 6/3/20 1:04 PM, Adhemerval Zanella via Libc-alpha wrote:
>>
>>
>> On 03/06/2020 16:46, Vineet Gupta wrote:
>>> On 5/29/20 9:49 AM, Adhemerval Zanella via Libc-alpha wrote:
>>>>> +	; ----- child starts here ---------
>>>>> +
>>>>> +	; Setup TP register (only recent kernels v4.19+ do that)
>>>>> +	and.f	0, r12, CLONE_SETTLS
>>>>> +	mov.nz	r25, r9
>>>> Do you still need to set it since the minimum supported kernel
>>>> for ARC is 5.1 ?
>>>
>>> Right.
>>>
>>>> It should be safe for internal glibc usage, since for both pthread
>>>> and posix_spawn it blocks all signals including SIGCANCEL and SIGXID.
>>>> However this is still small race window if this is called directly 
>>>> with pthread cancellation or g*uid in multithread.
>>>
>>> I'm not sure what you mean above. Do you mean not doing this in glibc and even if
>>> kernel support didn't exist should be safe internally ?
>>
>> At least for internal clone usage with CLONE_VM within glibc we explicit
>> disable all signals (posix_spawn and pthread_create).
>>
>>>
>>> fwiw as mentioned above kernel sets up TP for clone (SETTLS). I detested doing
>>> that for a long time, give ABI implications but ended up doing it anyways due to
>>> an actual race hit when running uClibc tst-kill6 [1]
>>
>> We explicit disable all signals during the create_thread call in pthread_create
>> (b3cae39dcbfa2432b3f3aa28854d8ac57f0de1b8), so it should not happen on glibc
>> anymore.  However it is still an issue if application calls clone itself.
> 
> The scenario there was pthread_create() and parent getting scheduled before child
> and immediately doing pthread_kill() causing child signal handler to be invoked
> and signal handler doing pthread_self() which was broken as TP was not setup.
> 
> With commit above, parent pthread_kill() will block and will only run when child
> is scheduled and unblocks the signals !

It could happen with other scenarios when signal plus tp accesses were involved as
well: pthread_cancel the thread (since the sigcancel_handler acesses the tp
with THREAD_GETMEM macro), setg* functions (since sighandler_setxid), or any
functions with a syscall failure that sets the errno.



More information about the Libc-alpha mailing list