Date:   Wed, 12 Jun 2019 08:53:33 +0200
From:   David Hildenbrand <david@...hat.com>
To:     Anshuman Khandual <anshuman.khandual@....com>,
        Andrew Morton <akpm@...ux-foundation.org>
Cc:     linux-mm@...ck.org, linux-kernel@...r.kernel.org,
        linux-arm-kernel@...ts.infradead.org, catalin.marinas@....com,
        will.deacon@....com, ard.biesheuvel@....com, osalvador@...e.de,
        mhocko@...e.com, mark.rutland@....com
Subject: Re: [PATCH V5 - Rebased] mm/hotplug: Reorder memblock_[free|remove]()
 calls in try_remove_memory()

On 12.06.19 06:02, Anshuman Khandual wrote:
> 
> 
> On 06/12/2019 03:49 AM, Andrew Morton wrote:
>> On Tue, 11 Jun 2019 16:56:13 +0530 Anshuman Khandual <anshuman.khandual@....com> wrote:
>>
>>> Memory hot remove uses get_nid_for_pfn() while tearing down linked sysfs
>>> entries between memory block and node. It first checks pfn validity with
>>> pfn_valid_within() before fetching nid. With CONFIG_HOLES_IN_ZONE config
>>> (arm64 has this enabled) pfn_valid_within() calls pfn_valid().
>>>
>>> pfn_valid() is an arch implementation on arm64 (CONFIG_HAVE_ARCH_PFN_VALID)
>>> which scans all mapped memblock regions with memblock_is_map_memory(). This
>>> creates a problem in memory hot remove path which has already removed given
>>> memory range from memory block with memblock_[remove|free] before arriving
>>> at unregister_mem_sect_under_nodes(). Hence get_nid_for_pfn() returns -1
>>> skipping subsequent sysfs_remove_link() calls leaving node <-> memory block
>>> sysfs entries as is. Subsequent memory add operation hits BUG_ON() because
>>> of existing sysfs entries.
>>>
>>> [   62.007176] NUMA: Unknown node for memory at 0x680000000, assuming node 0
>>> [   62.052517] ------------[ cut here ]------------
>>> [   62.053211] kernel BUG at mm/memory_hotplug.c:1143!
>>> [   62.053868] Internal error: Oops - BUG: 0 [#1] PREEMPT SMP
>>> [   62.054589] Modules linked in:
>>> [   62.054999] CPU: 19 PID: 3275 Comm: bash Not tainted 5.1.0-rc2-00004-g28cea40b2683 #41
>>> [   62.056274] Hardware name: linux,dummy-virt (DT)
>>> [   62.057166] pstate: 40400005 (nZcv daif +PAN -UAO)
>>> [   62.058083] pc : add_memory_resource+0x1cc/0x1d8
>>> [   62.058961] lr : add_memory_resource+0x10c/0x1d8
>>> [   62.059842] sp : ffff0000168b3ce0
>>> [   62.060477] x29: ffff0000168b3ce0 x28: ffff8005db546c00
>>> [   62.061501] x27: 0000000000000000 x26: 0000000000000000
>>> [   62.062509] x25: ffff0000111ef000 x24: ffff0000111ef5d0
>>> [   62.063520] x23: 0000000000000000 x22: 00000006bfffffff
>>> [   62.064540] x21: 00000000ffffffef x20: 00000000006c0000
>>> [   62.065558] x19: 0000000000680000 x18: 0000000000000024
>>> [   62.066566] x17: 0000000000000000 x16: 0000000000000000
>>> [   62.067579] x15: ffffffffffffffff x14: ffff8005e412e890
>>> [   62.068588] x13: ffff8005d6b105d8 x12: 0000000000000000
>>> [   62.069610] x11: ffff8005d6b10490 x10: 0000000000000040
>>> [   62.070615] x9 : ffff8005e412e898 x8 : ffff8005e412e890
>>> [   62.071631] x7 : ffff8005d6b105d8 x6 : ffff8005db546c00
>>> [   62.072640] x5 : 0000000000000001 x4 : 0000000000000002
>>> [   62.073654] x3 : ffff8005d7049480 x2 : 0000000000000002
>>> [   62.074666] x1 : 0000000000000003 x0 : 00000000ffffffef
>>> [   62.075685] Process bash (pid: 3275, stack limit = 0x00000000d754280f)
>>> [   62.076930] Call trace:
>>> [   62.077411]  add_memory_resource+0x1cc/0x1d8
>>> [   62.078227]  __add_memory+0x70/0xa8
>>> [   62.078901]  probe_store+0xa4/0xc8
>>> [   62.079561]  dev_attr_store+0x18/0x28
>>> [   62.080270]  sysfs_kf_write+0x40/0x58
>>> [   62.080992]  kernfs_fop_write+0xcc/0x1d8
>>> [   62.081744]  __vfs_write+0x18/0x40
>>> [   62.082400]  vfs_write+0xa4/0x1b0
>>> [   62.083037]  ksys_write+0x5c/0xc0
>>> [   62.083681]  __arm64_sys_write+0x18/0x20
>>> [   62.084432]  el0_svc_handler+0x88/0x100
>>> [   62.085177]  el0_svc+0x8/0xc
>>
>> This seems like a serious problem.  Once which should be fixed in 5.2
>> and perhaps the various -stable kernels as well.
> 
> But the problem does not exist in the current kernel as yet till the reworked
> versions of the other two patches in this series get merged. This patch was
> after arm64 hot-remove enablement in V1 (https://lkml.org/lkml/2019/4/3/28)
> but after some discussions it was decided to be moved before hot-remove from
> V2 (https://lkml.org/lkml/2019/4/14/5) onwards as a prerequisite patch instead.
> 
>>
>>> Re-ordering memblock_[free|remove]() with arch_remove_memory() solves the
>>> problem on arm64 as pfn_valid() behaves correctly and returns positive
>>> as memblock for the address range still exists. arch_remove_memory()
>>> removes applicable memory sections from zone with __remove_pages() and
>>> tears down kernel linear mapping. Removing memblock regions afterwards
>>> is safe because there is no other memblock (bootmem) allocator user that
>>> late. So nobody is going to allocate from the removed range just to blow
>>> up later. Also nobody should be using the bootmem allocated range else
>>> we wouldn't allow to remove it. So reordering is indeed safe.
>>>
>>> ...
>>>
>>>
>>> - Rebased on linux-next (next-20190611)
>>
>> Yet the patch you've prepared is designed for 5.3.  Was that
>> deliberate, or should we be targeting earlier kernels?
> 
> It was deliberate for 5.3 as a preparation for upcoming reworked arm64 hot-remove.
> 

We should probably add to the patch description something like "This is
a preparation for arm64 memory hotremove. The described issue is not
relevant on other architectures."

-- 

Thanks,

David / dhildenb

Powered by blists - more mailing lists