NPTL branch temporarily broken

[Khem Raj]

arm is still seing crashes in sysvinit. I can boot into busybox shell.

On Fri, Dec 12, 2008 at 5:32 AM, Carmelo AMOROSO <carmelo.amoroso at st.com> wrote:
> Carmelo AMOROSO wrote:
>>
>> Denys Vlasenko wrote:
>>>
>>> On Wednesday 10 December 2008 10:06, Carmelo AMOROSO wrote:
>>>>
>>>> Carmelo AMOROSO wrote:
>>>>>
>>>>> Hi Khem,
>>>>> I'm seeing similar problem on sh4-nptl as happened to you as you told
>>>>> me yesterday.. Something recently merged is causing issues.
>>>>> I'm doing investigation and keep you all informed.
>>>>>
>>>>> Further I'm just writing a detailed report on current status of merge,
>>>>> that I can say is almost complete (once figured out what is not
>>>>> actually working ;-) ).
>>>>>
>>>>> I'm currently 100% focused on uclibc.
>>>>>
>>>>> Cheers,
>>>>> Carmelo
>>>>
>>>> Khem,
>>>> I think to have found what is causing problems on nptl branch
>>>> (anyway I don't know why yet).
>>>>
>>>> The cause is the change merged from trunk in the file
>>>> libc/sysdeps/linux/common/bits/sigset.h (rev 24243)
>>>>
>>>> that includes all changes occurred in the same file on trunk
>>>> in rev 24183 / 24220 / 24221 by vda.
>>>>
>>>> Please may you try on ARM reverting that file
>>>> and see if it helps. On sh4-nptl it's working better now
>>>> Below the boot log on my box... not more PANIC on INIT
>>>> even if it's blocking on udev
>>>>
>>>> This seems confirming my idea on signal handling changes.
>>>> I'll try to merge on my private branch all changes that are already
>>>> in the nptl-branch but thos on signal handling.
>>>
>>> The change is that sigset_t is not 1024 bits, but 64 bits
>>> on almost all architectures. Only MIPS has 128.
>>>
>>> This affects sigaction and sigprocmask, but they already
>>> are translating libc struct sigaction / sigset_t to kernel's.
>>>
>>> With this change sigset_t actually should match kernel one,
>>> albeit struct sigaction is still different a bit.
>>>
>>> Anyway, the symptom of breakage would be, for example,
>>> in libc/sysdeps/linux/arm/sigaction.c:
>>>
>>>
>>> int __libc_sigaction (int sig, const struct sigaction *act, struct
>>> sigaction *oact)
>>> {
>>>        int result;
>>>        struct kernel_sigaction kact, koact;
>>>        enum {
>>>                SIGSET_MIN_SIZE = sizeof(kact.sa_mask) <
>>> sizeof(act->sa_mask)
>>>                                ? sizeof(kact.sa_mask) :
>>> sizeof(act->sa_mask)
>>>        };
>>>
>>>
>>> *** first thing to check is whether sizeof(kact.sa_mask) ==
>>> sizeof(act->sa_mask).
>>> I tried to make it work even if it does not, but it's always better if
>>> they match.
>>> add line
>>> struct BUG { int bug[sizeof(kact.sa_mask) == sizeof(act->sa_mask) ? 1 :
>>> -1]; };
>>> and if compile fails, sizeof's are not ==.
>>> Let me know that.
>>>
>>>
>>>        if (act) {
>>>                kact.k_sa_handler = act->sa_handler;
>>>                memcpy (&kact.sa_mask, &act->sa_mask, SIGSET_MIN_SIZE);
>>>                kact.sa_flags = act->sa_flags;
>>> # ifdef HAVE_SA_RESTORER
>>>                if (kact.sa_flags & SA_RESTORER) {
>>>                        kact.sa_restorer = act->sa_restorer;
>>>                } else {
>>>                        kact.sa_restorer = choose_restorer
>>> (kact.sa_flags);
>>>                        kact.sa_flags |= SA_RESTORER;
>>>                }
>>> # endif
>>>        }
>>>
>>>
>>>
>>> *** next:
>>>
>>>        /* NB: kernel (as of 2.6.25) will return EINVAL
>>>         * if sizeof(kact.sa_mask) does not match kernel's
>>> sizeof(sigset_t) */
>>>        result = __syscall_rt_sigaction(sig,
>>>                        act ? __ptrvalue (&kact) : NULL,
>>>                        oact ? __ptrvalue (&koact) : NULL,
>>>                        sizeof(kact.sa_mask));
>>>
>>> Basically, if it throws EINVAL, always, then sizeof(kact.sa_mask) is
>>> botched.
>>> (Kernel simply compares it with constant, it must be correct to work,
>>> no "greater or equal" slack there).
>>> I think adding if (errno == EINVAL) write(2, "EINVAL!\n", 8);
>>> will tell you whether it happens.
>>>
>>>
>>> sigprocmask is simpler:
>>>
>>> int sigprocmask(int how, const sigset_t * set, sigset_t * oldset)
>>> {
>>>        if (set &&
>>> ...
>>>                ) {
>>>                __set_errno(EINVAL);
>>>                return -1;
>>>        }
>>>        return __rt_sigprocmask(how, set, oldset, _NSIG / 8);
>>>
>>> Again, _NSIG / 8 must be correct here, or it will fail always.
>>>
>>>
>>>
>>> Hope this helps
>>>
>>> --
>>> vda
>>>
>> Hi Denys,
>> I
>> ll try to understand the impacts of this change on ntpl,
>> following your hints.
>> For now I preferred to revert back the "offending" changes to have the
>> nptl branch working again.
>>
>> Keep you informed.
>>
>> Carmelo
>>
> Another step forward.
> Merged some other piece of signal handling (except for
> libc/signal/sigaction.c) and nptl-branch is still working.
> Running test-suite on sh4 without regressions.
>
> As usual, Khem if you can, give a try on arm.
> TIA,
> Carmelo
>>
>> _______________________________________________
>> uClibc mailing list
>> uClibc at uclibc.org
>> http://busybox.net/cgi-bin/mailman/listinfo/uclibc
>>
>
>