Re: [PATCH v8 2/8] Add rseq extensible ABI support

[Michael Jeanson <mjeanson@efficios.com>]

On 2024-02-16 15 h 37, DJ Delorie wrote:
> Michael Jeanson <mjeanson@efficios.com> writes:
>>> The TLS doesn't get resized, does it?  Or am I thinking of the DTV?
>>
>> The terminology is a bit confusing and could be clarified, my
>> understanding is that the 'static TLS block' is the area where the 'TLS
>> blocks' from shared objects and the main executable are allocated and
>> the DTV is the array of descriptors that point to each of these 'TLS
>> blocks' inside the 'static TLS block'.
> 
> I might be thinking of where TLS data for dlopen()'d objects go, after
> the static TLS block is filled.
> 
>>> Isn't __rseq_size always at least 32 ?
>>
>> It's at least 32 when the registration succeeded, however if the
>> registration failed or was disabled by tunable then it's set to 0.
> 
> Are we sure?
> 
>    /* Default to the rseq ABI minimum sizes, this will reduce TLS usage to 32
>       bytes when rseq is disabled by tunables.  */
>    size_t rseq_size = TLS_DL_RSEQ_MIN_SIZE;
>    size_t rseq_align = TLS_DL_RSEQ_MIN_ALIGN;
>    bool do_rseq = true;
>    do_rseq = TUNABLE_GET_FULL (glibc, pthread, rseq, int, NULL);
>    if (do_rseq)
>      {
>        rseq_align = GLRO(dl_tls_rseq_align);
>        /* Make sure the rseq area size is at least the minimum ABI size and a
>           multiple of the requested aligment. */
>        rseq_size = roundup (MAX (GLRO(dl_tls_rseq_feature_size),
> 			      TLS_DL_RSEQ_MIN_SIZE), rseq_align);
>      }
> 
> 
> static inline bool
> rseq_register_current_thread (struct pthread *self, bool do_rseq)
> {
>    if (do_rseq)
>      {
>        /* The kernel expects 'rseq_area->rseq_cs == NULL' on registration, zero
>           the whole rseq area.  */
>        memset(rseq_get_area(), 0, GLRO (dl_tls_rseq_size));
>        int ret = INTERNAL_SYSCALL_CALL (rseq, rseq_get_area(),
>                                         GLRO (dl_tls_rseq_size),
>                                         0, RSEQ_SIG);
>        if (!INTERNAL_SYSCALL_ERROR_P (ret))
>          return true;
>      }
>    RSEQ_SETMEM (rseq_get_area(), cpu_id, RSEQ_CPU_ID_REGISTRATION_FAILED);
>    return false;
> }
When rseq registration is disabled by tunable we still need to allocate 
an rseq block because application code is allowed to check for the 
registration status by reading 'rseq->cpu_id'. So in this case we 
allocate the minimum ABI size of 32. Internally this is recorded in 
'dl_tls_rseq_size' but the application visible '_rseq_size' is left at 0 
(this happens in sysdeps/nptl/dl-tls_init_tp.c).

I'll try making this more clear with additional comments and finding a 
name for 'rseq_size' that is not so close to '__rseq_size'.