Re: armv7-on-aarch64 stuck at urdlck

From: mmel@freebsd.org <meloun.michal_at_gmail.com>
Date: Wed, 24 Jul 2024 17:34:47 UTC

On 24.07.2024 17:47, Konstantin Belousov wrote:
> On Wed, Jul 24, 2024 at 01:07:39PM +0000, John F Carr wrote:
>>
>>
>>> On Jul 24, 2024, at 06:50, Konstantin Belousov <kib@freebsd.org> wrote:
>>>
>>> On Wed, Jul 24, 2024 at 12:34:57PM +0200, mmel@freebsd.org wrote:
>>>>
>>>>
>>>> On 24.07.2024 12:24, Konstantin Belousov wrote:
>>>>> On Tue, Jul 23, 2024 at 08:11:13PM +0000, John F Carr wrote:
>>>>>> On Jul 23, 2024, at 13:46, Michal Meloun <meloun.michal@gmail.com> wrote:
>>>>>>>
>>>>>>> On 23.07.2024 11:36, Konstantin Belousov wrote:
>>>>>>>> On Tue, Jul 23, 2024 at 09:53:41AM +0200, Michal Meloun wrote:
>>>>>>>>> The good news is that I'm finally able to generate a working/locking
>>>>>>>>> test case.  The culprit (at least for me) is if "-mcpu" is used when
>>>>>>>>> compiling libthr (e.g. indirectly injected via CPUTYPE in /etc/make.conf).
>>>>>>>>> If it is not used, libthr is broken (regardless of -O level or debug/normal
>>>>>>>>> build), but -mcpu=cortex-a15 will always produce a working libthr.
>>>>>>>> I think this is very significant progress.
>>>>>>>> Do you plan to drill down more to see what is going on?
>>>>>>>
>>>>>>> So the problem is now clear, and I fear it may apply to other architectures as well.
>>>>>>> dlopen_object() (from rtld_elf),
>>>>>>> https://cgit.freebsd.org/src/tree/libexec/rtld-elf/rtld.c#n3766,
>>>>>>> holds the rtld_bind_lock write lock for almost the entire time a new library is loaded.
>>>>>>> If the code uses a yet unresolved symbol to load the library, the rtl_bind() function attempts to get read lock of  rtld_bind_lock and a deadlock occurs.
>>>>>>>
>>>>>>> In this case, it round_up() in _thr_stack_fix_protection,
>>>>>>> https://cgit.freebsd.org/src/tree/lib/libthr/thread/thr_stack.c#n136.
>>>>>>> Issued by __aeabi_uidiv (since not all armv7 processors support HW divide).
>>>>>>>
>>>>>>> Unfortunately, I'm not sure how to fix it.  The compiler can emit __aeabi_<> in any place, and I'm not sure if it can resolve all the symbols used by rtld_eld and libthr beforehand.
>>>>>>>
>>>>>>>
>>>>>>> Michal
>>>>>>>
>>>>>>
>>>>>> In this case (but not for all _aeabi_ functions) we can avoid division
>>>>>> as long as page size is a power of 2.
>>>>>>
>>>>>> The function is
>>>>>>
>>>>>>    static inline size_t
>>>>>>    round_up(size_t size)
>>>>>>    {
>>>>>>     if (size % _thr_page_size != 0)
>>>>>>     size = ((size / _thr_page_size) + 1) *
>>>>>>         _thr_page_size;
>>>>>>     return size;
>>>>>>    }
>>>>>>
>>>>>> The body can be condensed to
>>>>>>
>>>>>>    return (size + _thr_page_size - 1) & ~(_thr_page_size - 1);
>>>>>>
>>>>>> This is shorter in both lines of code and instruction bytes.
>>>>>
>>>>> Lets not allow this to be lost.  Could anybody confirm that the patch
>>>>> below fixes the issue?
>>>>>
>>>>> commit d560f4f6690a48476565278fd07ca131bf4eeb3c
>>>>> Author: Konstantin Belousov <kib@FreeBSD.org>
>>>>> Date:   Wed Jul 24 13:17:55 2024 +0300
>>>>>
>>>>>      rtld: avoid division in __thr_map_stacks_exec()
>>>>>      The function is called by rtld with the rtld bind lock write-locked,
>>>>>      when fixing the stack permission during dso load.  Not every ARMv7 CPU
>>>>>      supports the div, which causes the recursive entry into rtld to resolve
>>>>>      the  __aeabi_uidiv symbol, causing self-lock.
>>>>>      Workaround the problem by using roundup2() instead of open-coding less
>>>>>      efficient formula.
>>>>>      Diagnosed by:   mmel
>>>>>      Based on submission by: John F Carr <jfc@mit.edu>
>>>>>      Sponsored by:   The FreeBSD Foundation
>>>>>      MFC after:      1 week
>>>>>
>>> Just realized that it is wrong.  Stack size is user-controlled and it does
>>> not need to be power of two.
>>
>> Your change is correct.  _thr_page_size is set to getpagesize(),
>> which is a power of 2.   The call to roundup2 takes a user-provided
>> size and rounds it up to a multiple of the system page size.
>>
>> I tested the change and it works.  My change also works and
>> should compile to identical code.  I forgot there was a standard
>> function to do the rounding.
> Right, my bad, thank you for correcting my thinko.
> 
>>
>>> For final resolving of deadlocks, after a full day of digging, I'm very much
>>>> incline  of adding -znow to the linker flags for libthr.so (and maybe also
>>>> for ld-elf.so). The runtime cost of resolving all symbols at startup is very
>>>> low. Direct pre-solving in _thr_rtld_init() is problematic for the _aeabi_*
>>>> symbols, since they don't have an official C prototypes, and some are not
>>>> compatible with C calling conventions.
>>> I do not like it. `-z now' changes (breaks) the ABI and makes some symbols
>>> not preemtible.
>>>
>>> In the worst case, we would need a call to the asm routine which causes the
>>> resolution of the _eabi_* symbols on arm.
>>>
>>
>> It would also be possible to link libthr with libgcc.a and use a linker map
>> to hide the _eabi_ symbols.
> In principle yes, but if the ARM ABI states that _eabi symbols must be used,
> and exported from libc, then this is also some form of ABI breakage.

I hope that https://reviews.freebsd.org/D46104 is acceptable :)