Re: [PATCH] MIPS: fix memory reservation for non-usermem setups

From: Mike Rapoport
Date: Wed Apr 14 2021 - 09:47:17 EST


On Mon, Apr 12, 2021 at 11:45:45PM -0700, Ilya Lipnitskiy wrote:
> Hi Thomas,
>
> On Tue, Apr 6, 2021 at 6:18 AM Thomas Bogendoerfer
> <tsbogend@xxxxxxxxxxxxxxxx> wrote:
> >
> > On Sat, Apr 03, 2021 at 07:02:13PM -0700, Ilya Lipnitskiy wrote:
> > > Hi Mike,
> > >
> > > On Tue, Mar 16, 2021 at 11:33 PM Mike Rapoport <rppt@xxxxxxxxxx> wrote:
> > > >
> > > > Hi Ilya,
> > > >
> > > > On Tue, Mar 16, 2021 at 10:10:09PM -0700, Ilya Lipnitskiy wrote:
> > > > > Hi Thomas,
> > > > >
> > > > > On Fri, Mar 12, 2021 at 7:19 AM Thomas Bogendoerfer
> > > > > <tsbogend@xxxxxxxxxxxxxxxx> wrote:
> > > > > >
> > > > > > On Sun, Mar 07, 2021 at 11:40:30AM -0800, Ilya Lipnitskiy wrote:
> > > > > > > From: Tobias Wolf <dev-NTEO@xxxxxxxxx>
> > > > > > >
> > > > > > > Commit 67a3ba25aa95 ("MIPS: Fix incorrect mem=X@Y handling") introduced a new
> > > > > > > issue for rt288x where "PHYS_OFFSET" is 0x0 but the calculated "ramstart" is
> > > > > > > not. As the prerequisite of custom memory map has been removed, this results
> > > > > > > in the full memory range of 0x0 - 0x8000000 to be marked as reserved for this
> > > > > > > platform.
> > > > > >
> > > > > > and where is the problem here ?
> > > > > Turns out this was already attempted to be upstreamed - not clear why
> > > > > it wasn't merged. Context:
> > > > > https://lore.kernel.org/linux-mips/6504517.U6H5IhoIOn@loki/
> > > > >
> > > > > I hope the thread above helps you understand the problem.
> > > >
> > > > The memory initialization was a bit different then. Do you still see the
> > > > same problem?
> > > Thanks for asking. I obtained a RT2880 device and gave it a try. It
> > > hangs at boot without this patch, however selecting
> >
> > can you provide debug logs with memblock=debug for both good and bad
> > kernels ? I'm curious what's the reason for failing allocation...
>
> Sorry for taking a while to respond. See attached.
> FWIW, it seems these are the lines that stand out in hang.log:
> [ 0.000000] memblock_reserve: [0x00000000-0x07ffffff] setup_arch+0x214/0x5d8
> [ 0.000000] Wasting 1048576 bytes for tracking 32768 unused pages
> ...
> [ 0.000000] reserved[0x0] [0x00000000-0x087137aa], 0x087137ab
> bytes flags: 0x0
>
> Ilya

> ---------------------------CONTINUTES-BOOTING-NORMALLY-----------------------

> [ 0.000000] MEMBLOCK configuration:
> [ 0.000000] memory size = 0x02000000 reserved size = 0x0875a542
> [ 0.000000] memory.cnt = 0x1
> [ 0.000000] memory[0x0] [0x08000000-0x09ffffff], 0x02000000 bytes flags: 0x0
> [ 0.000000] reserved.cnt = 0x5
> [ 0.000000] reserved[0x0] [0x00000000-0x087137aa], 0x087137ab bytes flags: 0x0
> [ 0.000000] reserved[0x1] [0x087137b0-0x087137b3], 0x00000004 bytes flags: 0x0
> [ 0.000000] reserved[0x2] [0x087137c0-0x08715276], 0x00001ab7 bytes flags: 0x0
> [ 0.000000] reserved[0x3] [0x08715278-0x0871a533], 0x000052bc bytes flags: 0x0
> [ 0.000000] reserved[0x4] [0x0871a540-0x0875a55f], 0x00040020 bytes flags: 0x0

...

> [ 0.000000] Memory: 25168K/32768K available (4299K kernel code, 575K rwdata, 952K rodata, 1204K init, 205K bss, 7600K reserved, 0K cma-reserved)
> ----------------------------------------HANGS-FOREVER-HERE---------------------------------

I'd say that with ARCH_PFN_OFFSET set to 0 and actual memory start address
at 0x08000000 any attempt to do pfn_to_page()/page_to_pfn()/page_address()
will give an incorrect result and will crash the system.

No idea why the crash is silent, though :)

--
Sincerely yours,
Mike.