Fragmented physical memory on powerpc/32

[Mike Rapoport]

On September 14, 2022 10:43:52 AM GMT+01:00, Christophe Leroy <christophe.leroy at csgroup.eu> wrote:
>
>
>Le 14/09/2022 à 11:32, Mike Rapoport a écrit :
>> On Tue, Sep 13, 2022 at 02:36:13PM +0200, Christophe Leroy wrote:
>>>
>>>
>>> Le 13/09/2022 à 08:11, Christophe Leroy a écrit :
>>>>
>>>>
>>>> Le 12/09/2022 à 23:16, Pali Rohár a écrit :
>>>>>>
>>>>>> My guess would be that something went wrong in the linear map
>>>>>> setup, but it
>>>>>> won't hurt running with "memblock=debug" added to the kernel
>>>>>> command line
>>>>>> to see if there is anything suspicious there.
>>>>>
>>>>> Here is boot log on serial console with memblock=debug command line:
>>>>>
>>>> ...
>>>>>
>>>>> Do you need something more for debug?
>>>>
>>>> Can you send me the 'vmlinux' used to generate the above Oops so that I
>>>> can see exactly where we are in function mem_init().
>>>>
>>>> And could you also try without CONFIG_HIGHMEM just in case.
>>>>
>>>
>>> I looked at the vmlinux you sent me, the problem is in the loop for highmem
>>> in mem_init(). It crashes in the call to free_highmem_page()
>>>
>>> #ifdef CONFIG_HIGHMEM
>>> 	{
>>> 		unsigned long pfn, highmem_mapnr;
>>>
>>> 		highmem_mapnr = lowmem_end_addr >> PAGE_SHIFT;
>>> 		for (pfn = highmem_mapnr; pfn < max_mapnr; ++pfn) {
>>> 			phys_addr_t paddr = (phys_addr_t)pfn << PAGE_SHIFT;
>>> 			struct page *page = pfn_to_page(pfn);
>>> 			if (!memblock_is_reserved(paddr))
>>> 				free_highmem_page(page);
>>> 		}
>>> 	}
>>> #endif /* CONFIG_HIGHMEM */
>>>
>>>
>>> As far as I can see in the memblock debug lines, the holes don't seem to be
>>> marked as reserved by memblock. So it is above valid ? Other architectures
>>> seem to do differently.
>>>
>>> Can you try by replacing !memblock_is_reserved(paddr) by
>>> memblock_is_memory(paddr) ?
>> 
>> The holes should not be marked as reserved, we just need to loop over the
>> memory ranges rather than over pfns. Then the holes will be taken into
>> account.
>> 
>> I believe arm and xtensa got this right:
>> 
>> (from arch/arm/mm/init.c)
>> 
>> static void __init free_highpages(void)
>> {
>> #ifdef CONFIG_HIGHMEM
>> 	unsigned long max_low = max_low_pfn;
>> 	phys_addr_t range_start, range_end;
>> 	u64 i;
>> 
>> 	/* set highmem page free */
>> 	for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE,
>> 				&range_start, &range_end, NULL) {
>> 		unsigned long start = PFN_UP(range_start);
>> 		unsigned long end = PFN_DOWN(range_end);
>> 
>> 		/* Ignore complete lowmem entries */
>> 		if (end <= max_low)
>> 			continue;
>> 
>> 		/* Truncate partial highmem entries */
>> 		if (start < max_low)
>> 			start = max_low;
>> 
>> 		for (; start < end; start++)
>> 			free_highmem_page(pfn_to_page(start));
>> 	}
>> #endif
>> }
>> 
>
>
>And what about the way MIPS does it ?
>
>static inline void __init mem_init_free_highmem(void)
>{
>#ifdef CONFIG_HIGHMEM
>	unsigned long tmp;
>
>	if (cpu_has_dc_aliases)
>		return;
>
>	for (tmp = highstart_pfn; tmp < highend_pfn; tmp++) {
>		struct page *page = pfn_to_page(tmp);
>
>		if (!memblock_is_memory(PFN_PHYS(tmp)))
>			SetPageReserved(page);
>		else
>			free_highmem_page(page);
>	}
>#endif
>}

This iterates over all PFNs in the highmem range and skips those in holes.
for_each_free_mem_range() skips the holes altogether.

Largely, I think we need to move, say, arm's version to mm/ and use it everywhere, except, perhaps, x86.

>Christophe
-- 
Sincerely yours,
Mike