Re: [PATCH v1 2/4] mm/memory_hotplug: Make unregister_memory_section() never fail

[David Hildenbrand <david@redhat.com>]

On 17.04.19 14:45, Oscar Salvador wrote:
> On Tue, 2019-04-09 at 12:01 +0200, David Hildenbrand wrote:
>> Failing while removing memory is mostly ignored and cannot really be
>> handled. Let's treat errors in unregister_memory_section() in a nice
>> way, warning, but continuing.
>>
>> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
>> Cc: "Rafael J. Wysocki" <rafael@kernel.org>
>> Cc: Ingo Molnar <mingo@kernel.org>
>> Cc: Andrew Banman <andrew.banman@hpe.com>
>> Cc: "mike.travis@hpe.com" <mike.travis@hpe.com>
>> Cc: David Hildenbrand <david@redhat.com>
>> Cc: Andrew Morton <akpm@linux-foundation.org>
>> Cc: Oscar Salvador <osalvador@suse.de>
>> Cc: Michal Hocko <mhocko@suse.com>
>> Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
>> Cc: Qian Cai <cai@lca.pw>
>> Cc: Wei Yang <richard.weiyang@gmail.com>
>> Cc: Arun KS <arunks@codeaurora.org>
>> Cc: Mathieu Malaterre <malat@debian.org>
>> Signed-off-by: David Hildenbrand <david@redhat.com>
>> ---
>>  drivers/base/memory.c  | 16 +++++-----------
>>  include/linux/memory.h |  2 +-
>>  mm/memory_hotplug.c    |  4 +---
>>  3 files changed, 7 insertions(+), 15 deletions(-)
>>
>> diff --git a/drivers/base/memory.c b/drivers/base/memory.c
>> index 0c9e22ffa47a..f180427e48f4 100644
>> --- a/drivers/base/memory.c
>> +++ b/drivers/base/memory.c
>> @@ -734,15 +734,18 @@ unregister_memory(struct memory_block *memory)
>>  {
>>  	BUG_ON(memory->dev.bus != &memory_subsys);
>>  
>> -	/* drop the ref. we got in remove_memory_section() */
>> +	/* drop the ref. we got via find_memory_block() */
>>  	put_device(&memory->dev);
>>  	device_unregister(&memory->dev);
>>  }
>>  
>> -static int remove_memory_section(struct mem_section *section)
>> +void unregister_memory_section(struct mem_section *section)
>>  {
>>  	struct memory_block *mem;
>>  
>> +	if (WARN_ON_ONCE(!present_section(section)))
>> +		return;
>> +
>>  	mutex_lock(&mem_sysfs_mutex);
>>  
>>  	/*
>> @@ -763,15 +766,6 @@ static int remove_memory_section(struct
>> mem_section *section)
>>  
>>  out_unlock:
>>  	mutex_unlock(&mem_sysfs_mutex);
>> -	return 0;
>> -}
>> -
>> -int unregister_memory_section(struct mem_section *section)
>> -{
>> -	if (!present_section(section))
>> -		return -EINVAL;
>> -
>> -	return remove_memory_section(section);
>>  }
>>  #endif /* CONFIG_MEMORY_HOTREMOVE */
>>  
>> diff --git a/include/linux/memory.h b/include/linux/memory.h
>> index a6ddefc60517..e1dc1bb2b787 100644
>> --- a/include/linux/memory.h
>> +++ b/include/linux/memory.h
>> @@ -113,7 +113,7 @@ extern int
>> register_memory_isolate_notifier(struct notifier_block *nb);
>>  extern void unregister_memory_isolate_notifier(struct notifier_block
>> *nb);
>>  int hotplug_memory_register(int nid, struct mem_section *section);
>>  #ifdef CONFIG_MEMORY_HOTREMOVE
>> -extern int unregister_memory_section(struct mem_section *);
>> +extern void unregister_memory_section(struct mem_section *);
>>  #endif
>>  extern int memory_dev_init(void);
>>  extern int memory_notify(unsigned long val, void *v);
>> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
>> index 696ed7ee5e28..b0cb05748f99 100644
>> --- a/mm/memory_hotplug.c
>> +++ b/mm/memory_hotplug.c
>> @@ -527,9 +527,7 @@ static int __remove_section(struct zone *zone,
>> struct mem_section *ms,
>>  	if (!valid_section(ms))
>>  		return ret;
>>  
>> -	ret = unregister_memory_section(ms);
>> -	if (ret)
>> -		return ret;
>> +	unregister_memory_section(ms);
> 
> So, technically unregister_memory_section() can __only__ fail in case
> the section is not present, returning -EINVAL.
> 
> Now, I was checking how the pair valid/present sections work.
> Unless I am mistaken, we only mark sections as memmap those sections
> that are present.
> 
> This can come from two paths:
> 
> - Early boot:
>   * memblocks_present
>      memory_present           - mark sections as present
>     sparse_init               - iterates only over present sections
>      sparse_init_nid
>       sparse_init_one_section - mark section as valid
> 
> - Hotplug path:
>   * sparse_add_one_section
>      section_mark_present     - mark section as present
>      sparse_init_one_section  - mark section as valid
>    
> 
> During early boot, sparse_init iterates __only__ over present sections,
> so only those are marked valid as well, and during hotplug, the section
> is both marked present and valid.
> 
> All in all, I think that we are safe if we remove the present_section
> check in your new unregister_memory_section(), as a valid_section
> cannot be non-present, and we do already catch those early in
> __remove_section().

Yes, dropping the check completely would be the next step.

> 
> Then, the only thing missing to be completely error-less in that
> codepath is to make unregister_mem_sect_under_nodes() void return-type.
> Not that it matters a lot as we are already ignoring its return code,
> but I find that quite disturbing and wrong.
> 
> So, would you like to take this patch in your patchset in case you re-
> submit?

I already had a look at that approach but would like to defer it, until
I eventually factor out creation/deletion of memory devices (see my
other RFC, still some work has to be done).

Reasons:

1. We only support unplug of system ram memory belonging to exactly one
node. No need to iterate over all nodes. mem->node should later tell us
exactly what needs to be removed. No need to iterate at all.

2. unregister_mem_sect_under_nodes() should be replaced by
"unregister_mem_block_under_nodes()". We only support unplug of whole
memory blocks. Ripping out random sections from a memory block is not
going to work and is not supported.

IOW, this is some leftover when people turned memory block devices to
span multiple sections. Once we clean that up, this problem goes away.

-- 

Thanks,

David / dhildenb