ARM board lockups/hangs triggered by locks and mutexes

[Rafał Miłecki]

On 14.08.2023 11:04, Geert Uytterhoeven wrote:
> Hi Rafal,
> 
> On Mon, Aug 7, 2023 at 1:11 PM Rafał Miłecki <zajec5 at gmail.com> wrote:
>> On 4.08.2023 13:07, Rafał Miłecki wrote:
>>> I triple checked that. Dropping a single unused function breaks kernel /
>>> device stability on BCM53573!
>>>
>>> AFAIK the only thing below diff actually affects is location of symbols
>>> (I actually verified that by comparing System.map before and after -
>>> over 22'000 of relocated symbols).
>>>
>>> Can some unfortunate location of symbols cause those hangs/lockups?
>>
>> I performed another experiment. First I dropped mtd_check_of_node() to
>> bring kernel back to the stable state.
>>
>> Then I started adding useless code to the mtdchar_unlocked_ioctl(). I
>> ended up adding just enough to make sure all post-mtd symbols in
>> System.map got the same offset as in case of backporting
>> mtd_check_of_node().
>>
>> I started experiencing lockups/hangs again.
>>
>> I repeated the same test with adding dumb code to the brcm_nvram_probe()
>> and verifying symbols offsets following brcm_nvram_probe one.
>>
>> I believe this confirms that this problem is about offset or alignment
>> of some specific symbol(s). The remaining question is what symbols and
>> how to fix or workaround that.
> 
> I had similar experiences on other ARM platforms many years ago:
> bisection lead to something completely bogus, and it turned out
> adding a single line of innocent code made the system lock-up or crash
> unexpectedly.  It was definitely related to alignment, as adding the
> right extra amount of innocent code would fix the problem. Until some
> later change changing alignment again...
> I never found the real cause, but the problems went away over time.
> I am not sure I did enable all required errata config options, so I
> may have missed some...

I already experiented some weird performance variations on Broadcom's
Northstar platform that was related to symbols layout & cache hit/miss
ratio. For that reason I use -falign-functions=32 for that whole
OpenWrt's "bcm53xx" target (it covers Northstar and BCM53573). So
this aspect should be ruled out already in my case.