Re: [PATCH] mm: don't rely on system state to detect hot-plug operations

From: Laurent Dufour
Date: Wed Sep 09 2020 - 05:22:29 EST

Le 09/09/2020 à 11:09, Michal Hocko a écrit :
On Wed 09-09-20 09:48:59, Laurent Dufour wrote:
Le 09/09/2020 à 09:40, Michal Hocko a écrit :
[reposting because the malformed cc list confused my email client]

On Tue 08-09-20 19:08:35, Laurent Dufour wrote:
In register_mem_sect_under_node() the system_state’s value is checked to
detect whether the operation the call is made during boot time or during an
hot-plug operation. Unfortunately, that check is wrong on some
architecture, and may lead to sections being registered under multiple
nodes if node's memory ranges are interleaved.

Why is this check arch specific?

I was wrong the check is not arch specific.

This can be seen on PowerPC LPAR after multiple memory hot-plug and
hot-unplug operations are done. At the next reboot the node's memory ranges
can be interleaved

What is the exact memory layout?

For instance:
[ 0.000000] Early memory node ranges
[ 0.000000] node 1: [mem 0x0000000000000000-0x000000011fffffff]
[ 0.000000] node 2: [mem 0x0000000120000000-0x000000014fffffff]
[ 0.000000] node 1: [mem 0x0000000150000000-0x00000001ffffffff]
[ 0.000000] node 0: [mem 0x0000000200000000-0x000000048fffffff]
[ 0.000000] node 2: [mem 0x0000000490000000-0x00000007ffffffff]

Include this into the changelog.

and since the call to link_mem_sections() is made in
topology_init() while the system is in the SYSTEM_SCHEDULING state, the
node's id is not checked, and the sections registered multiple times.

So a single memory section/memblock belongs to two numa nodes?

If the node id is not checked in register_mem_sect_under_node(), yes that the case.

I do not follow. register_mem_sect_under_node is about user interface.
This is independent on the low level memory representation - aka memory
section. I do not think we can handle a section in multiple zones/nodes.
Memblock in multiple zones/nodes is a different story and interleaving
physical memory layout can indeed lead to it. This is something that we
do not allow for runtime hotplug but have to somehow live with that - at
least not crash.

register_mem_sect_under_node() is called at boot time and when memory is hot added. In the later case the assumption is made that all the pages of the added block are in the same node. And that's a valid assumption. However at boot time the call is made using the node's whole range, lowest address to highest address for that node. In the case there are interleaved ranges, this means the interleaved sections are registered for each nodes which is not correct.

that case, the system is able to boot but later hot-plug operation may lead
to this panic because the node's links are correctly broken:

Correctly broken? Could you provide more details on the inconsistency

laurent@ltczep3-lp4:~$ ls -l /sys/devices/system/memory/memory21
total 0
lrwxrwxrwx 1 root root 0 Aug 24 05:27 node1 -> ../../node/node1
lrwxrwxrwx 1 root root 0 Aug 24 05:27 node2 -> ../../node/node2
-rw-r--r-- 1 root root 65536 Aug 24 05:27 online
-r--r--r-- 1 root root 65536 Aug 24 05:27 phys_device
-r--r--r-- 1 root root 65536 Aug 24 05:27 phys_index
drwxr-xr-x 2 root root 0 Aug 24 05:27 power
-r--r--r-- 1 root root 65536 Aug 24 05:27 removable
-rw-r--r-- 1 root root 65536 Aug 24 05:27 state
lrwxrwxrwx 1 root root 0 Aug 24 05:25 subsystem -> ../../../../bus/memory
-rw-r--r-- 1 root root 65536 Aug 24 05:25 uevent
-r--r--r-- 1 root root 65536 Aug 24 05:27 valid_zones

OK, so there are two nodes referenced here. Not terrible from the user
point of view. Such a memory block will refuse to offline or online

No the memory block is still owned by one node, only the sysfs representation is wrong. So the memory block can be hot unplugged, but only one node's link will be cleaned, and a '/syss/devices/system/node#/memory21' link will remain and that will be detected later when that memory block is hot plugged again.

Which physical memory range you are trying to add here and what is the
node affinity?

None is added, the root cause of the issue is happening at boot time.

Let me clarify my question. The crash has clearly happened during the
hotplug add_memory_resource - which is clearly not a boot time path.
I was askin for more information about why this has failed. It is quite
clear that sysfs machinery has failed and that led to BUG_ON but we are
mising an information on why. What was the physical memory range to be
added and why sysfs failed?

The BUG_ON is detecting a bad state generated earlier, at boot time because register_mem_sect_under_node() didn't check for the block's node id.

------------[ cut here ]------------
kernel BUG at /Users/laurent/src/linux-ppc/mm/memory_hotplug.c:1084!
Oops: Exception in kernel mode, sig: 5 [#1]
Modules linked in: rpadlpar_io rpaphp pseries_rng rng_core vmx_crypto gf128mul binfmt_misc ip_tables x_tables xfs libcrc32c crc32c_vpmsum autofs4
CPU: 8 PID: 10256 Comm: drmgr Not tainted 5.9.0-rc1+ #25
NIP: c000000000403f34 LR: c000000000403f2c CTR: 0000000000000000
REGS: c0000004876e3660 TRAP: 0700 Not tainted (5.9.0-rc1+)
MSR: 800000000282b033 <SF,VEC,VSX,EE,FP,ME,IR,DR,RI,LE> CR: 24000448 XER: 20040000
CFAR: c000000000846d20 IRQMASK: 0
GPR00: c000000000403f2c c0000004876e38f0 c0000000012f6f00 ffffffffffffffef
GPR04: 0000000000000227 c0000004805ae680 0000000000000000 00000004886f0000
GPR08: 0000000000000226 0000000000000003 0000000000000002 fffffffffffffffd
GPR12: 0000000088000484 c00000001ec96280 0000000000000000 0000000000000000
GPR16: 0000000000000000 0000000000000000 0000000000000004 0000000000000003
GPR20: c00000047814ffe0 c0000007ffff7c08 0000000000000010 c0000000013332c8
GPR24: 0000000000000000 c0000000011f6cc0 0000000000000000 0000000000000000
GPR28: ffffffffffffffef 0000000000000001 0000000150000000 0000000010000000
NIP [c000000000403f34] add_memory_resource+0x244/0x340
LR [c000000000403f2c] add_memory_resource+0x23c/0x340
Call Trace:
[c0000004876e38f0] [c000000000403f2c] add_memory_resource+0x23c/0x340 (unreliable)
[c0000004876e39c0] [c00000000040408c] __add_memory+0x5c/0xf0
[c0000004876e39f0] [c0000000000e2b94] dlpar_add_lmb+0x1b4/0x500
[c0000004876e3ad0] [c0000000000e3888] dlpar_memory+0x1f8/0xb80
[c0000004876e3b60] [c0000000000dc0d0] handle_dlpar_errorlog+0xc0/0x190
[c0000004876e3bd0] [c0000000000dc398] dlpar_store+0x198/0x4a0
[c0000004876e3c90] [c00000000072e630] kobj_attr_store+0x30/0x50
[c0000004876e3cb0] [c00000000051f954] sysfs_kf_write+0x64/0x90
[c0000004876e3cd0] [c00000000051ee40] kernfs_fop_write+0x1b0/0x290
[c0000004876e3d20] [c000000000438dd8] vfs_write+0xe8/0x290
[c0000004876e3d70] [c0000000004391ac] ksys_write+0xdc/0x130
[c0000004876e3dc0] [c000000000034e40] system_call_exception+0x160/0x270
[c0000004876e3e20] [c00000000000d740] system_call_common+0xf0/0x27c
Instruction dump:
48442e35 60000000 0b030000 3cbe0001 7fa3eb78 7bc48402 38a5fffe 7ca5fa14
78a58402 48442db1 60000000 7c7c1b78 <0b030000> 7f23cb78 4bda371d 60000000
---[ end trace 562fd6c109cd0fb2 ]---

The BUG_ON on failure is absolutely horrendous. There must be a better
way to handle a failure like that. The failure means that
sysfs_create_link_nowarn has failed. Please describe why that is the

This patch addresses the root cause by not relying on the system_state
value to detect whether the call is due to a hot-plug operation or not. An
additional parameter is added to link_mem_sections() to tell the context of
the call and this parameter is propagated to register_mem_sect_under_node()
throuugh the walk_memory_blocks()'s call.

This looks like a hack to me and it deserves a better explanation. The
existing code is a hack on its own and it is inconsistent with other
boot time detection. We are using (system_state < SYSTEM_RUNNING) at other
places IIRC. Would it help to use the same here as well? Maybe we want to
wrap that inside a helper (early_memory_init()) and use it at all

I agree, this looks like a hack to check for the system_state value.
I'll follow the David's proposal and introduce an enum detailing when the
node id check has to be done or not.

I am not sure an enum is going to make the existing situation less
messy. Sure we somehow have to distinguish boot init and runtime hotplug
because they have different constrains. I am arguing that a) we should
have a consistent way to check for those and b) we shouldn't blow up
easily just because sysfs infrastructure has failed to initialize.

For the point a, using the enum allows to know in register_mem_sect_under_node() if the link operation is due to a hotplug operation or done at boot time.

For the point b, one option would be ignore the link error in the case the link is already existing, but that BUG_ON() had the benefit to highlight the root issue.