Hi,
It seems the patch below can solve many problems after switched to NO_BOOTMEM,
because the memory allocation behavior is more similar as before.
diff --git a/arch/mips/kernel/setup.c b/arch/mips/kernel/setup.c
index 070234b..7a449d9 100644
--- a/arch/mips/kernel/setup.c
+++ b/arch/mips/kernel/setup.c
@@ -839,6 +839,7 @@ static void __init arch_mem_init(char **cmdline_p)
/* call board setup routine */
plat_mem_setup();
+ memblock_set_bottom_up(true);
/*
* Make sure all kernel memory is in the maps. The "UP" and
diff --git a/arch/mips/kernel/traps.c b/arch/mips/kernel/traps.c
index 0f852e1..15e103c 100644
--- a/arch/mips/kernel/traps.c
+++ b/arch/mips/kernel/traps.c
@@ -2260,10 +2260,8 @@ void __init trap_init(void)
unsigned long size = 0x200 + VECTORSPACING*64;
phys_addr_t ebase_pa;
- memblock_set_bottom_up(true);
ebase = (unsigned long)
memblock_alloc_from(size, 1 << fls(size), 0);
- memblock_set_bottom_up(false);
/*
* Try to ensure ebase resides in KSeg0 if possible.
@@ -2307,6 +2305,7 @@ void __init trap_init(void)
if (board_ebase_setup)
board_ebase_setup();
per_cpu_trap_init(true);
+ memblock_set_bottom_up(false);
/*
* Copy the generic exception handlers to their final destination.
------------------ Original ------------------
From: "Mike Rapoport"<r...@linux.ibm.com>;
Date: Fri, Nov 9, 2018 02:01 AM
To: "Thomas Bogendoerfer"<tbogendoer...@suse.de>;
Cc: "Ralf Baechle"<r...@linux-mips.org>; "Paul Burton"<paul.bur...@mips.com>;
"James Hogan"<jho...@kernel.org>; "Huacai Chen"<che...@lemote.com>;
"linux-mips"<linux-m...@linux-mips.org>;
"linux-kernel"<linux-kernel@vger.kernel.org>; "rppt"<r...@linux.vnet.ibm.com>;
Subject: Re: [[PATCH]] mips: Fix switch to NO_BOOTMEM for SGI-IP27/loongons3
NUMA
On November 8, 2018 6:52:17 PM GMT+02:00, Thomas Bogendoerfer
<tbogendoer...@suse.de> wrote:
>On Thu, 8 Nov 2018 18:18:23 +0200
>Mike Rapoport <r...@linux.ibm.com> wrote:
>
>> On Thu, Nov 08, 2018 at 03:44:28PM +0100, Thomas Bogendoerfer wrote:
>> > Commit bcec54bf3118 ("mips: switch to NO_BOOTMEM") broke SGI-IP27
>> > and NUMA enabled loongson3 by doing memblock_set_current_limit()
>> > before max_low_pfn has been evaluated. Both platforms need to do
>the
>> > memblock_set_current_limit() in platform specific code. For
>> > consistency the call to memblock_set_current_limit() is moved
>> > to the common bootmem_init(), where max_low_pfn is calculated
>> > for non NUMA enabled platforms.
>> [..]
>>
>> As for SGI-IP27, the initialization of max_low_pfn as late as in
>> paging_init() seems to be broken because it's value is used in
>> arch_mem_init() and in finalize_initrd() anyway.
>
>well, the patch is tested on real hardware and the first caller of
>a memblock_alloc* function is in a function called by
>free_area_init_nodes().
Then, apparently, I've missed something else.
The Onyx2 I worked on is dead for a couple of years now ;-)
>> AFAIU, both platforms set max_low_pfn to last available pfn, so it
>seems we
>> can simply do
>>
>> max_low_pfn = PFN_PHYS(memblock_end_of_DRAM())
>>
Should have been PHYS_PFN, sorry.
>> in the prom_meminit() function for both platforms and drop the loop
>> evaluating max_low_pfn in paging_init().
>
>sounds like a better plan. I'll prepare a new patch.
>
>Thomas.
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.
