Re: [PATCH v8 0/5] support reserving crashkernel above 4G on arm64 kdump

[John Donnelly]

> On Jun 3, 2020, at 8:20 AM, chenzhou <chenzhou10@xxxxxxxxxx> wrote:
> 
> Hi,
> 
> 
> On 2020/6/3 19:47, Prabhakar Kushwaha wrote:
>> Hi Chen,
>> 
>> On Tue, Jun 2, 2020 at 8:12 PM John Donnelly <john.p.donnelly@xxxxxxxxxx> wrote:
>>> 
>>> 
>>>> On Jun 2, 2020, at 12:38 AM, Prabhakar Kushwaha <prabhakar.pkin@xxxxxxxxx> wrote:
>>>> 
>>>> On Tue, Jun 2, 2020 at 3:29 AM John Donnelly <john.p.donnelly@xxxxxxxxxx> wrote:
>>>>> Hi .  See below !
>>>>> 
>>>>>> On Jun 1, 2020, at 4:02 PM, Bhupesh Sharma <bhsharma@xxxxxxxxxx> wrote:
>>>>>> 
>>>>>> Hi John,
>>>>>> 
>>>>>> On Tue, Jun 2, 2020 at 1:01 AM John Donnelly <John.P.donnelly@xxxxxxxxxx> wrote:
>>>>>>> Hi,
>>>>>>> 
>>>>>>> 
>>>>>>> On 6/1/20 7:02 AM, Prabhakar Kushwaha wrote:
>>>>>>>> Hi Chen,
>>>>>>>> 
>>>>>>>> On Thu, May 21, 2020 at 3:05 PM Chen Zhou <chenzhou10@xxxxxxxxxx> wrote:
>>>>>>>>> This patch series enable reserving crashkernel above 4G in arm64.
>>>>>>>>> 
>>>>>>>>> There are following issues in arm64 kdump:
>>>>>>>>> 1. We use crashkernel=X to reserve crashkernel below 4G, which will fail
>>>>>>>>> when there is no enough low memory.
>>>>>>>>> 2. Currently, crashkernel=Y@X can be used to reserve crashkernel above 4G,
>>>>>>>>> in this case, if swiotlb or DMA buffers are required, crash dump kernel
>>>>>>>>> will boot failure because there is no low memory available for allocation.
>>>>>>>>> 
>>>>>>>> We are getting "warn_alloc" [1] warning during boot of kdump kernel
>>>>>>>> with bootargs as [2] of primary kernel.
>>>>>>>> This error observed on ThunderX2  ARM64 platform.
>>>>>>>> 
>>>>>>>> It is observed with latest upstream tag (v5.7-rc3) with this patch set
>>>>>>>> and https://urldefense.com/v3/__https://lists.infradead.org/pipermail/kexec/2020-May/025128.html__;!!GqivPVa7Brio!LnTSARkCt0V0FozR0KmqooaH5ADtdXvs3mPdP3KRVqALmvSK2VmCkIPIhsaxbiIAAlzu$
>>>>>>>> Also **without** this patch-set
>>>>>>>> "https://urldefense.com/v3/__https://www.spinics.net/lists/arm-kernel/msg806882.html__;!!GqivPVa7Brio!LnTSARkCt0V0FozR0KmqooaH5ADtdXvs3mPdP3KRVqALmvSK2VmCkIPIhsaxbjC6ujMA$";
>>>>>>>> 
>>>>>>>> This issue comes whenever crashkernel memory is reserved after 0xc000_0000.
>>>>>>>> More details discussed earlier in
>>>>>>>> https://urldefense.com/v3/__https://www.spinics.net/lists/arm-kernel/msg806882.html__;!!GqivPVa7Brio!LnTSARkCt0V0FozR0KmqooaH5ADtdXvs3mPdP3KRVqALmvSK2VmCkIPIhsaxbjC6ujMA$  without any
>>>>>>>> solution
>>>>>>>> 
>>>>>>>> This patch-set is expected to solve similar kind of issue.
>>>>>>>> i.e. low memory is only targeted for DMA, swiotlb; So above mentioned
>>>>>>>> observation should be considered/fixed. .
>>>>>>>> 
>>>>>>>> --pk
>>>>>>>> 
>>>>>>>> [1]
>>>>>>>> [   30.366695] DMI: Cavium Inc. Saber/Saber, BIOS
>>>>>>>> TX2-FW-Release-3.1-build_01-2803-g74253a541a mm/dd/yyyy
>>>>>>>> [   30.367696] NET: Registered protocol family 16
>>>>>>>> [   30.369973] swapper/0: page allocation failure: order:6,
>>>>>>>> mode:0x1(GFP_DMA), nodemask=(null),cpuset=/,mems_allowed=0
>>>>>>>> [   30.369980] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.7.0-rc3+ #121
>>>>>>>> [   30.369981] Hardware name: Cavium Inc. Saber/Saber, BIOS
>>>>>>>> TX2-FW-Release-3.1-build_01-2803-g74253a541a mm/dd/yyyy
>>>>>>>> [   30.369984] Call trace:
>>>>>>>> [   30.369989]  dump_backtrace+0x0/0x1f8
>>>>>>>> [   30.369991]  show_stack+0x20/0x30
>>>>>>>> [   30.369997]  dump_stack+0xc0/0x10c
>>>>>>>> [   30.370001]  warn_alloc+0x10c/0x178
>>>>>>>> [   30.370004]  __alloc_pages_slowpath.constprop.111+0xb10/0xb50
>>>>>>>> [   30.370006]  __alloc_pages_nodemask+0x2b4/0x300
>>>>>>>> [   30.370008]  alloc_page_interleave+0x24/0x98
>>>>>>>> [   30.370011]  alloc_pages_current+0xe4/0x108
>>>>>>>> [   30.370017]  dma_atomic_pool_init+0x44/0x1a4
>>>>>>>> [   30.370020]  do_one_initcall+0x54/0x228
>>>>>>>> [   30.370027]  kernel_init_freeable+0x228/0x2cc
>>>>>>>> [   30.370031]  kernel_init+0x1c/0x110
>>>>>>>> [   30.370034]  ret_from_fork+0x10/0x18
>>>>>>>> [   30.370036] Mem-Info:
>>>>>>>> [   30.370064] active_anon:0 inactive_anon:0 isolated_anon:0
>>>>>>>> [   30.370064]  active_file:0 inactive_file:0 isolated_file:0
>>>>>>>> [   30.370064]  unevictable:0 dirty:0 writeback:0 unstable:0
>>>>>>>> [   30.370064]  slab_reclaimable:34 slab_unreclaimable:4438
>>>>>>>> [   30.370064]  mapped:0 shmem:0 pagetables:14 bounce:0
>>>>>>>> [   30.370064]  free:1537719 free_pcp:219 free_cma:0
>>>>>>>> [   30.370070] Node 0 active_anon:0kB inactive_anon:0kB
>>>>>>>> active_file:0kB inactive_file:0kB unevictable:0kB isolated(anon):0kB
>>>>>>>> isolated(file):0kB mapped:0kB dirty:0kB writeback:0kB shmem:0kB
>>>>>>>> shmem_thp: 0kB shmem_pmdmapped: 0kB anon_thp: 0kB writeback_tmp:0kB
>>>>>>>> unstable:0kB all_unreclaimable? no
>>>>>>>> [   30.370073] Node 1 active_anon:0kB inactive_anon:0kB
>>>>>>>> active_file:0kB inactive_file:0kB unevictable:0kB isolated(anon):0kB
>>>>>>>> isolated(file):0kB mapped:0kB dirty:0kB writeback:0kB shmem:0kB
>>>>>>>> shmem_thp: 0kB shmem_pmdmapped: 0kB anon_thp: 0kB writeback_tmp:0kB
>>>>>>>> unstable:0kB all_unreclaimable? no
>>>>>>>> [   30.370079] Node 0 DMA free:0kB min:0kB low:0kB high:0kB
>>>>>>>> reserved_highatomic:0KB active_anon:0kB inactive_anon:0kB
>>>>>>>> active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB
>>>>>>>> present:128kB managed:0kB mlocked:0kB kernel_stack:0kB pagetables:0kB
>>>>>>>> bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB
>>>>>>>> [   30.370084] lowmem_reserve[]: 0 250 6063 6063
>>>>>>>> [   30.370090] Node 0 DMA32 free:256000kB min:408kB low:664kB
>>>>>>>> high:920kB reserved_highatomic:0KB active_anon:0kB inactive_anon:0kB
>>>>>>>> active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB
>>>>>>>> present:269700kB managed:256000kB mlocked:0kB kernel_stack:0kB
>>>>>>>> pagetables:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB
>>>>>>>> [   30.370094] lowmem_reserve[]: 0 0 5813 5813
>>>>>>>> [   30.370100] Node 0 Normal free:5894876kB min:9552kB low:15504kB
>>>>>>>> high:21456kB reserved_highatomic:0KB active_anon:0kB inactive_anon:0kB
>>>>>>>> active_file:0kB inactive_file:0kB unevictable:0kB writepending:0kB
>>>>>>>> present:8388608kB managed:5953112kB mlocked:0kB kernel_stack:21672kB
>>>>>>>> pagetables:56kB bounce:0kB free_pcp:876kB local_pcp:176kB free_cma:0kB
>>>>>>>> [   30.370104] lowmem_reserve[]: 0 0 0 0
>>>>>>>> [   30.370107] Node 0 DMA: 0*4kB 0*8kB 0*16kB 0*32kB 0*64kB 0*128kB
>>>>>>>> 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 0kB
>>>>>>>> [   30.370113] Node 0 DMA32: 0*4kB 0*8kB 0*16kB 0*32kB 0*64kB 0*128kB
>>>>>>>> 0*256kB 0*512kB 0*1024kB 1*2048kB (M) 62*4096kB (M) = 256000kB
>>>>>>>> [   30.370119] Node 0 Normal: 2*4kB (M) 3*8kB (ME) 2*16kB (UE) 3*32kB
>>>>>>>> (UM) 1*64kB (U) 2*128kB (M) 2*256kB (ME) 3*512kB (ME) 3*1024kB (ME)
>>>>>>>> 3*2048kB (UME) 1436*4096kB (M) = 5893600kB
>>>>>>>> [   30.370129] Node 0 hugepages_total=0 hugepages_free=0
>>>>>>>> hugepages_surp=0 hugepages_size=1048576kB
>>>>>>>> [   30.370130] 0 total pagecache pages
>>>>>>>> [   30.370132] 0 pages in swap cache
>>>>>>>> [   30.370134] Swap cache stats: add 0, delete 0, find 0/0
>>>>>>>> [   30.370135] Free swap  = 0kB
>>>>>>>> [   30.370136] Total swap = 0kB
>>>>>>>> [   30.370137] 2164609 pages RAM
>>>>>>>> [   30.370139] 0 pages HighMem/MovableOnly
>>>>>>>> [   30.370140] 612331 pages reserved
>>>>>>>> [   30.370141] 0 pages hwpoisoned
>>>>>>>> [   30.370143] DMA: failed to allocate 256 KiB pool for atomic
>>>>>>>> coherent allocation
>>>>>>> 
>>>>>>> During my testing I saw the same error and Chen's  solution corrected it .
>>>>>> Which combination you are using on your side? I am using Prabhakar's
>>>>>> suggested environment and can reproduce the issue
>>>>>> with or without Chen's crashkernel support above 4G patchset.
>>>>>> 
>>>>>> I am also using a ThunderX2 platform with latest makedumpfile code and
>>>>>> kexec-tools (with the suggested patch
>>>>>> <https://urldefense.com/v3/__https://lists.infradead.org/pipermail/kexec/2020-May/025128.html__;!!GqivPVa7Brio!J6lUig58-Gw6TKZnEEYzEeSU36T-1SqlB1kImU00xtX_lss5Tx-JbUmLE9TJC3foXBLg$ >).
>>>>>> 
>>>>>> Thanks,
>>>>>> Bhupesh
>>>>> 
>>>>> I did this activity 5 months ago and I have moved on to other activities. My DMA failures were related to PCI devices that could not be enumerated because  low-DMA space was not  available when crashkernel was moved above 4G; I donât recall the exact platform.
>>>>> 
>>>>> 
>>>>> 
>>>>> For this failure ,
>>>>> 
>>>>>>>> DMA: failed to allocate 256 KiB pool for atomic
>>>>>>>> coherent allocation
>>>>> 
>>>>> Is due to :
>>>>> 
>>>>> 
>>>>> 3618082c
>>>>> ("arm64 use both ZONE_DMA and ZONE_DMA32")
>>>>> 
>>>>> With the introduction of ZONE_DMA to support the Raspberry DMA
>>>>> region below 1G, the crashkernel is placed in the upper 4G
>>>>> ZONE_DMA_32 region. Since the crashkernel does not have access
>>>>> to the ZONE_DMA region, it prints out call trace during bootup.
>>>>> 
>>>>> It is due to having this CONFIG item  ON  :
>>>>> 
>>>>> 
>>>>> CONFIG_ZONE_DMA=y
>>>>> 
>>>>> Turning off ZONE_DMA fixes a issue and Raspberry PI 4 will
>>>>> use the device tree to specify memory below 1G.
>>>>> 
>>>>> 
>>>> Disabling ZONE_DMA is temporary solution.  We may need proper solution
>>> 
>>> Perhaps the Raspberry platform configuration dependencies need separated  from âserver classâ Arm  equipment ?  Or auto-configured on boot ?  Consult an expert ;-)
>>> 
>>> 
>>> 
>>>>> I would like to see Chenâs feature added , perhaps as EXPERIMENTAL,  so we can get some configuration testing done on it.   It corrects having a DMA zone in low memory while crash-kernel is above 4GB.  This has been going on for a year now.
>>>> I will also like this patch to be added in Linux as early as possible.
>>>> 
>>>> Issue mentioned by me happens with or without this patch.
>>>> 
>>>> This patch-set can consider fixing because it uses low memory for DMA
>>>> & swiotlb only.
>>>> We can consider restricting crashkernel within the required range like below
>>>> 
>>>> diff --git a/kernel/crash_core.c b/kernel/crash_core.c
>>>> index 7f9e5a6dc48c..bd67b90d35bd 100644
>>>> --- a/kernel/crash_core.c
>>>> +++ b/kernel/crash_core.c
>>>> @@ -354,7 +354,7 @@ int __init reserve_crashkernel_low(void)
>>>>                       return 0;
>>>>       }
>>>> 
>>>> -       low_base = memblock_find_in_range(0, 1ULL << 32, low_size, CRASH_ALIGN);
>>>> +       low_base = memblock_find_in_range(0,0xc0000000, low_size, CRASH_ALIGN);
>>>>       if (!low_base) {
>>>>               pr_err("Cannot reserve %ldMB crashkernel low memory,
>>>> please try smaller size.\n",
>>>>                      (unsigned long)(low_size >> 20));
>>>> 
>>>> 
>>>    I suspect  0xc0000000  would need to be a CONFIG item  and not hard-coded.
>>> 
>> if you consider this as valid change,  can you please incorporate as
>> part of your patch-set.
> 
> After commit 1a8e1cef7 ("arm64: use both ZONE_DMA and ZONE_DMA32")ïthe 0-4G memory is splited
> to DMA [mem 0x0000000000000000-0x000000003fffffff] and DMA32 [mem 0x0000000040000000-0x00000000ffffffff] on arm64.
> 
> From the above discussion, on your platform, the low crashkernel fall in DMA32 region, but your environment needs to access DMA
> region, so there is the call trace.
> 
> I have a question, why do you choose 0xc0000000 here?
> 
> Besides, this is common code, we also need to consider about x86.
> 

 + nsaenzjulienne@xxxxxxx 

  Exactly .  This is why it needs to be a CONFIG option for  Raspberry ..,  or device tree option. 


  We could revert 1a8e1cef7 since it broke  Arm kdump too.


> 
> Thanks,
> Chen Zhou
>