Re: [PATCH RFC] MIPS: Remove detect_memory_region()

[Jinyang He]

On 02/26/2021 02:52 PM, Jiaxun Yang wrote:

> 在 2021/2/26 上午9:37, Jinyang He 写道:
> > On 02/24/2021 11:40 PM, Jiaxun Yang wrote:
> >
> > > On Wed, Feb 24, 2021, at 9:02 PM, Jinyang He wrote:
> > > > detect_memory_region() was committed by Commit 4d9f77d25268 ("MIPS: 
> > > > add
> > > > detect_memory_region()"). Then it was equipped by Commit dd63b00804a5
> > > > ("MIPS: ralink: make use of the new memory detection code") and
> > > > Commit 9b75733b7b5e ("MIPS: ath79: make use of the new memory 
> > > > detection
> > > > code"). Its code is based on early ath79 platform code.
> > > >
> > > > What puzzles me is that how memcmp() detect the memory region. If 
> > > > `break`
> > > > was touched, the function could make sense. That means memcmp() should
> > > > return zero. Otherwise, the loop will be end by size > sz_max.
> > > >
> > > > I have tested detect_memory_region() on Loongson64 3A3000. On our 
> > > > design,
> > > > kseg0 low 256MB maps real memory and kseg0 high 256MB maps IO/PCI. The
> > > > function runs and last stopped on kseg1 where is uncached. In this 
> > > > process
> > > > memcmp also returned non-zero when detected kseg0 high 256MB. Then 
> > > > I did
> > > > another thing. memcpy first and test memcmp then (after &_end). It 
> > > > works
> > > > well on 3A3000 but badly on 3A4000. Maybe because kseg0 high 256MB 
> > > > maps
> > > > IO/PCI and it is dangerous to write like write memory.
> > > >
> > > > At last, read memory from where is not memory region may always 
> > > > return 0.
> > > > (Or trigger exception.) This function have been used several years and
> > > > seems no error occur. Maybe it's a fallback way.
> > > That is not true for other platforms like ath79 or mtk.
> > > They'll wrap around or return 0xffffffff for out of boundary accessing.
> > >
> > > Loongson does not apply to this case as it have special "Address 
> > > Window"
> > > design to accurately describe address regions.
> > > Any access beyond described windows will be handled by MC and return 
> > > 0 or random stuff.
> > >
> > > Again, please don't make changes because you can.
> > >
> > > Thanks.
> > >
> > > - Jiaxun
> >
> > Hi, Jiaxun,
> >
> > Thank you for answering this puzzle for me in detail.
> >
> > Assume that the machine has 8MB real memory and dm address is (base + 
> > 3M).
> > When size = 8MB, there will be a phenomenon of `wrap around`, the actual
> > content of (dm + 8M + 3M) is content of (dm + 3M), so it will trigger
> > `break`, right? At this time, the kernel will add 8M to the memory.
>
> Hi Jingyang,
>
> How can you boot kernel with 8M memory in present days ;-)
> (Ohh with respect to Nintendo64 developer who had proven it's possible)
>
> For what I can say, detect_memory_region exists because many devices
> doesn't have a way to pass memory size information from bootloader to
> kernel. Or their bootloader even don't care about memory size.
>
> Kernel needs it to get memory size correctly. Although it seems fragile.
>
> That's life, we must accept imperfect past and don't repeat it in future.
>
> Thanks.
>
> - Jiaxun
That's just a assume. Because it looks fresh to me.
Thank you very much. :-D

> > Thanks,
> > Jinyang