mm, memory_hotplug: do not clear numa_node association after hot_remove

RFC PATCH In review Michal Hocko 3 participants 10 messages Nov 8, 2018 – Dec 20, 2018

Subsystems (2)

Memory management Memory hot(un)plug

lore

Share & tools

lore

Michal Hocko

author

2018-11-08 10:04 UTC #0 lore raw

RFC PATCH 1/1 mm, memory_hotplug: do not clear numa_node association after hot_remove

From: Michal Hocko <[email protected]>

Per-cpu numa_node provides a default node for each possible cpu. The association gets initialized during the boot when the architecture specific code explores cpu->NUMA affinity. When the whole NUMA node is removed though we are clearing this association

try_offline_node
check_and_unmap_cpu_on_node
unmap_cpu_on_node
numa_clear_node
numa_set_node(cpu, NUMA_NO_NODE)

This means that whoever calls cpu_to_node for a cpu associated with such a node will get NUMA_NO_NODE. This is problematic for two reasons. First it is fragile because __alloc_pages_node would simply blow up on an out-of-bound access. We have encountered this when loading kvm module

BUG: unable to handle kernel paging request at 00000000000021c0
IP: [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
PGD 800000ffe853e067 PUD 7336bbc067 PMD 0
Oops: 0000 [#1] SMP
[...]
CPU: 88 PID: 1223749 Comm: modprobe Tainted: G        W          4.4.156-94.64-default #1
task: ffff88727eff1880 ti: ffff887354490000 task.ti: ffff887354490000
RIP: 0010:[<ffffffff8119ccb3>]  [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
RSP: 0018:ffff887354493b40  EFLAGS: 00010202
RAX: 00000000000021c0 RBX: 0000000000000000 RCX: 0000000000000000
RDX: 0000000000000000 RSI: 0000000000000002 RDI: 00000000014000c0
RBP: 00000000014000c0 R08: ffffffffffffffff R09: 0000000000000000
R10: ffff88fffc89e790 R11: 0000000000014000 R12: 0000000000000101
R13: ffffffffa0772cd4 R14: ffffffffa0769ac0 R15: 0000000000000000
FS:  00007fdf2f2f1700(0000) GS:ffff88fffc880000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00000000000021c0 CR3: 00000077205ee000 CR4: 0000000000360670
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Stack:
 0000000000000086 014000c014d20400 ffff887354493bb8 ffff882614d20f4c
 0000000000000000 0000000000000046 0000000000000046 ffffffff810ac0c9
 ffff88ffe78c0000 ffffffff0000009f ffffe8ffe82d3500 ffff88ff8ac55000
Call Trace:
 [<ffffffffa07476cd>] alloc_vmcs_cpu+0x3d/0x90 [kvm_intel]
 [<ffffffffa0772c0c>] hardware_setup+0x781/0x849 [kvm_intel]
 [<ffffffffa04a1c58>] kvm_arch_hardware_setup+0x28/0x190 [kvm]
 [<ffffffffa04856fc>] kvm_init+0x7c/0x2d0 [kvm]
 [<ffffffffa0772cf2>] vmx_init+0x1e/0x32c [kvm_intel]
 [<ffffffff8100213a>] do_one_initcall+0xca/0x1f0
 [<ffffffff81193886>] do_init_module+0x5a/0x1d7
 [<ffffffff81112083>] load_module+0x1393/0x1c90
 [<ffffffff81112b30>] SYSC_finit_module+0x70/0xa0
 [<ffffffff8161cbc3>] entry_SYSCALL_64_fastpath+0x1e/0xb7
DWARF2 unwinder stuck at entry_SYSCALL_64_fastpath+0x1e/0xb7

on an older kernel but the code is basically the same in the current Linus tree as well. alloc_vmcs_cpu could use alloc_pages_nodemask which would recognize NUMA_NO_NODE and use alloc_pages_node which would translate it to numa_mem_id but that is wrong as well because it would use a cpu affinity of the local CPU which might be quite far from the original node. It is also reasonable to expect that cpu_to_node will provide a sane value and there might be many more callers like that.

The second problem is that __register_one_node relies on cpu_to_node to properly associate cpus back to the node when it is onlined. We do not want to lose that link as there is no arch independent way to get it from the early boot time AFAICS.

Drop the whole check_and_unmap_cpu_on_node machinery and keep the association to fix both issues. The NODE_DATA(nid) is not deallocated so it will stay in place and if anybody wants to allocate from that node then a fallback node will be used.

Thanks to Vlastimil Babka for his live system debugging skills that helped debugging the issue.

Debugged-by Vlastimil Babka <[email protected]>

Reported-by Miroslav Benes <[email protected]>

Fixes e13fe8695c57 ("cpu-hotplug,memory-hotplug: clear cpu_to_node() when offlining the node")

Cc Wen Congyang <[email protected]>

Cc Tang Chen <[email protected]>

Signed-off-by Michal Hocko <[email protected]>

---

Hi,
please note that I am sending this as an RFC even though this has been confirmed to fix the oops in kvm_intel module because I cannot simply tell that there are no other side effect that I do not see from the code reading. I would appreciate some background from people who have introduced this code e13fe8695c57 ("cpu-hotplug,memory-hotplug: clear cpu_to_node() when offlining the node") because the changelog doesn't really explain the motivation much.

 mm/memory_hotplug.c | 30 +-----------------------------
 1 file changed, 1 insertion(+), 29 deletions(-)

mm/memory_hotplug.c +1 ~~-29~~

index 2b2b3ccbbfb5..87aeafac54ee 100644
@@ -1753,34 +1753,6 @@ static int check_cpu_on_node(pg_data_t *pgdat)
 	return 0;
 }
 
-static void unmap_cpu_on_node(pg_data_t *pgdat)
-{
-#ifdef CONFIG_ACPI_NUMA
-	int cpu;
-
-	for_each_possible_cpu(cpu)
-		if (cpu_to_node(cpu) == pgdat->node_id)
-			numa_clear_node(cpu);
-#endif
-}
-
-static int check_and_unmap_cpu_on_node(pg_data_t *pgdat)
-{
-	int ret;
-
-	ret = check_cpu_on_node(pgdat);
-	if (ret)
-		return ret;
-
-	/*
-	 * the node will be offlined when we come here, so we can clear
-	 * the cpu_to_node() now.
-	 */
-
-	unmap_cpu_on_node(pgdat);
-	return 0;
-}
-
 /**
  * try_offline_node
  * @nid: the node ID

@@ -1813,7 +1785,7 @@ void try_offline_node(int nid)
 		return;
 	}
 
-	if (check_and_unmap_cpu_on_node(pgdat))
+	if (check_cpu_on_node(pgdat))
 		return;
 
 	/*

Signature · 1 line

2.19.1

Michal Hocko

author

2018-11-08 10:29 UTC #1 lore raw

[Removing Wen Congyang and Tang Chen from the CC list because their
emails bounce. It seems that we will never learn about their motivation]

Quoted 144 lines · Michal Hocko

From: Michal Hocko <[email protected]>
Per-cpu numa_node provides a default node for each possible cpu. The association gets initialized during the boot when the architecture specific code explores cpu->NUMA affinity. When the whole NUMA node is removed though we are clearing this association
try_offline_node
check_and_unmap_cpu_on_node
unmap_cpu_on_node
numa_clear_node
numa_set_node(cpu, NUMA_NO_NODE)
This means that whoever calls cpu_to_node for a cpu associated with such a node will get NUMA_NO_NODE. This is problematic for two reasons. First it is fragile because __alloc_pages_node would simply blow up on an out-of-bound access. We have encountered this when loading kvm module
BUG: unable to handle kernel paging request at 00000000000021c0
IP: [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
PGD 800000ffe853e067 PUD 7336bbc067 PMD 0
Oops: 0000 [#1] SMP
[...]
CPU: 88 PID: 1223749 Comm: modprobe Tainted: G        W          4.4.156-94.64-default #1
task: ffff88727eff1880 ti: ffff887354490000 task.ti: ffff887354490000
RIP: 0010:[<ffffffff8119ccb3>]  [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
RSP: 0018:ffff887354493b40  EFLAGS: 00010202
RAX: 00000000000021c0 RBX: 0000000000000000 RCX: 0000000000000000
RDX: 0000000000000000 RSI: 0000000000000002 RDI: 00000000014000c0
RBP: 00000000014000c0 R08: ffffffffffffffff R09: 0000000000000000
R10: ffff88fffc89e790 R11: 0000000000014000 R12: 0000000000000101
R13: ffffffffa0772cd4 R14: ffffffffa0769ac0 R15: 0000000000000000
FS:  00007fdf2f2f1700(0000) GS:ffff88fffc880000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00000000000021c0 CR3: 00000077205ee000 CR4: 0000000000360670
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Stack:
 0000000000000086 014000c014d20400 ffff887354493bb8 ffff882614d20f4c
 0000000000000000 0000000000000046 0000000000000046 ffffffff810ac0c9
 ffff88ffe78c0000 ffffffff0000009f ffffe8ffe82d3500 ffff88ff8ac55000
Call Trace:
 [<ffffffffa07476cd>] alloc_vmcs_cpu+0x3d/0x90 [kvm_intel]
 [<ffffffffa0772c0c>] hardware_setup+0x781/0x849 [kvm_intel]
 [<ffffffffa04a1c58>] kvm_arch_hardware_setup+0x28/0x190 [kvm]
 [<ffffffffa04856fc>] kvm_init+0x7c/0x2d0 [kvm]
 [<ffffffffa0772cf2>] vmx_init+0x1e/0x32c [kvm_intel]
 [<ffffffff8100213a>] do_one_initcall+0xca/0x1f0
 [<ffffffff81193886>] do_init_module+0x5a/0x1d7
 [<ffffffff81112083>] load_module+0x1393/0x1c90
 [<ffffffff81112b30>] SYSC_finit_module+0x70/0xa0
 [<ffffffff8161cbc3>] entry_SYSCALL_64_fastpath+0x1e/0xb7
DWARF2 unwinder stuck at entry_SYSCALL_64_fastpath+0x1e/0xb7
on an older kernel but the code is basically the same in the current Linus tree as well. alloc_vmcs_cpu could use alloc_pages_nodemask which would recognize NUMA_NO_NODE and use alloc_pages_node which would translate it to numa_mem_id but that is wrong as well because it would use a cpu affinity of the local CPU which might be quite far from the original node. It is also reasonable to expect that cpu_to_node will provide a sane value and there might be many more callers like that.
The second problem is that __register_one_node relies on cpu_to_node to properly associate cpus back to the node when it is onlined. We do not want to lose that link as there is no arch independent way to get it from the early boot time AFAICS.
Drop the whole check_and_unmap_cpu_on_node machinery and keep the association to fix both issues. The NODE_DATA(nid) is not deallocated so it will stay in place and if anybody wants to allocate from that node then a fallback node will be used.
Thanks to Vlastimil Babka for his live system debugging skills that helped debugging the issue.
Debugged-by Vlastimil Babka <[email protected]>
Reported-by Miroslav Benes <[email protected]>
Fixes e13fe8695c57 ("cpu-hotplug,memory-hotplug: clear cpu_to_node() when offlining the node")
Cc Wen Congyang <[email protected]>
Cc Tang Chen <[email protected]>
Signed-off-by Michal Hocko <[email protected]>
---
Hi,
please note that I am sending this as an RFC even though this has been confirmed to fix the oops in kvm_intel module because I cannot simply tell that there are no other side effect that I do not see from the code reading. I would appreciate some background from people who have introduced this code e13fe8695c57 ("cpu-hotplug,memory-hotplug: clear cpu_to_node() when offlining the node") because the changelog doesn't really explain the motivation much.
 mm/memory_hotplug.c | 30 +-----------------------------
 1 file changed, 1 insertion(+), 29 deletions(-)
mm/memory_hotplug.c +1 ~~-29~~
index 2b2b3ccbbfb5..87aeafac54ee 100644
@@ -1753,34 +1753,6 @@ static int check_cpu_on_node(pg_data_t *pgdat)
 	return 0;
 }
 
-static void unmap_cpu_on_node(pg_data_t *pgdat)
-{
-#ifdef CONFIG_ACPI_NUMA
-	int cpu;
-
-	for_each_possible_cpu(cpu)
-		if (cpu_to_node(cpu) == pgdat->node_id)
-			numa_clear_node(cpu);
-#endif
-}
-
-static int check_and_unmap_cpu_on_node(pg_data_t *pgdat)
-{
-	int ret;
-
-	ret = check_cpu_on_node(pgdat);
-	if (ret)
-		return ret;
-
-	/*
-	 * the node will be offlined when we come here, so we can clear
-	 * the cpu_to_node() now.
-	 */
-
-	unmap_cpu_on_node(pgdat);
-	return 0;
-}
-
 /**
  * try_offline_node
  * @nid: the node ID
@@ -1813,7 +1785,7 @@ void try_offline_node(int nid)
 		return;
 	}
 
-	if (check_and_unmap_cpu_on_node(pgdat))
+	if (check_cpu_on_node(pgdat))
 		return;
 
 	/*
Signature · 1 line
2.19.1

Signature · 2 lines

Michal Hocko
SUSE Labs

Anshuman Khandual

2018-11-09 03:42 UTC #2 lore raw

Quoted 62 lines · Michal Hocko

[Removing Wen Congyang and Tang Chen from the CC list because their
emails bounce. It seems that we will never learn about their motivation]
Quoted 58 lines · Michal Hocko
From: Michal Hocko <[email protected]>
Per-cpu numa_node provides a default node for each possible cpu. The association gets initialized during the boot when the architecture specific code explores cpu->NUMA affinity. When the whole NUMA node is removed though we are clearing this association
try_offline_node
check_and_unmap_cpu_on_node
unmap_cpu_on_node
numa_clear_node
numa_set_node(cpu, NUMA_NO_NODE)
This means that whoever calls cpu_to_node for a cpu associated with such a node will get NUMA_NO_NODE. This is problematic for two reasons. First it is fragile because __alloc_pages_node would simply blow up on an out-of-bound access. We have encountered this when loading kvm module
BUG: unable to handle kernel paging request at 00000000000021c0
IP: [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
PGD 800000ffe853e067 PUD 7336bbc067 PMD 0
Oops: 0000 [#1] SMP
[...]
CPU: 88 PID: 1223749 Comm: modprobe Tainted: G        W          4.4.156-94.64-default #1
task: ffff88727eff1880 ti: ffff887354490000 task.ti: ffff887354490000
RIP: 0010:[<ffffffff8119ccb3>]  [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
RSP: 0018:ffff887354493b40  EFLAGS: 00010202
RAX: 00000000000021c0 RBX: 0000000000000000 RCX: 0000000000000000
RDX: 0000000000000000 RSI: 0000000000000002 RDI: 00000000014000c0
RBP: 00000000014000c0 R08: ffffffffffffffff R09: 0000000000000000
R10: ffff88fffc89e790 R11: 0000000000014000 R12: 0000000000000101
R13: ffffffffa0772cd4 R14: ffffffffa0769ac0 R15: 0000000000000000
FS:  00007fdf2f2f1700(0000) GS:ffff88fffc880000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00000000000021c0 CR3: 00000077205ee000 CR4: 0000000000360670
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Stack:
 0000000000000086 014000c014d20400 ffff887354493bb8 ffff882614d20f4c
 0000000000000000 0000000000000046 0000000000000046 ffffffff810ac0c9
 ffff88ffe78c0000 ffffffff0000009f ffffe8ffe82d3500 ffff88ff8ac55000
Call Trace:
 [<ffffffffa07476cd>] alloc_vmcs_cpu+0x3d/0x90 [kvm_intel]
 [<ffffffffa0772c0c>] hardware_setup+0x781/0x849 [kvm_intel]
 [<ffffffffa04a1c58>] kvm_arch_hardware_setup+0x28/0x190 [kvm]
 [<ffffffffa04856fc>] kvm_init+0x7c/0x2d0 [kvm]
 [<ffffffffa0772cf2>] vmx_init+0x1e/0x32c [kvm_intel]
 [<ffffffff8100213a>] do_one_initcall+0xca/0x1f0
 [<ffffffff81193886>] do_init_module+0x5a/0x1d7
 [<ffffffff81112083>] load_module+0x1393/0x1c90
 [<ffffffff81112b30>] SYSC_finit_module+0x70/0xa0
 [<ffffffff8161cbc3>] entry_SYSCALL_64_fastpath+0x1e/0xb7
DWARF2 unwinder stuck at entry_SYSCALL_64_fastpath+0x1e/0xb7
on an older kernel but the code is basically the same in the current Linus tree as well. alloc_vmcs_cpu could use alloc_pages_nodemask which would recognize NUMA_NO_NODE and use alloc_pages_node which would translate it to numa_mem_id but that is wrong as well because it would use a cpu affinity of the local CPU which might be quite far from the original node.

But then the original node is getting/already off-lined. The allocation is going to come from a different node. alloc_pages_node() at least steer the allocation alway from VM_BUG_ON() because of NUMA_NO_NODE by replacing it with numa_mem_id().

If node fallback order is important for this allocation then could not it use __alloc_pages_nodemask() directly giving preference for its zonelist node and nodemask. Just curious.

Quoted 2 lines

Quoted 2 lines
It is also reasonable to expect that cpu_to_node will provide a sane value and there might be many more callers like that.

AFAICS there are two choices here. Either mark them NUMA_NO_NODE for all cpus of a node going offline or keep the existing mapping in case the node comes back again.

Quoted 5 lines

Quoted 5 lines
The second problem is that __register_one_node relies on cpu_to_node to properly associate cpus back to the node when it is onlined. We do not want to lose that link as there is no arch independent way to get it from the early boot time AFAICS.

Retaining the links seems to be right unless unmap_cpu_on_node() is sort of a weak callback letting arch to decide.

Quoted 3 lines

Quoted 3 lines
Drop the whole check_and_unmap_cpu_on_node machinery and keep the association to fix both issues. The NODE_DATA(nid) is not deallocated

Though retaining the link is a problem in itself but the allocation related crash could be solved by exploring __alloc_pages_nodemask() options.

Quoted 2 lines

Quoted 2 lines
so it will stay in place and if anybody wants to allocate from that node then a fallback node will be used.

Right, NODE_DATA(nid) is an advantage of retaining the link.

Michal Hocko

author

2018-11-09 07:59 UTC #3 lore raw

Quoted 66 lines · Anshuman Khandual

Quoted 62 lines · Michal Hocko
[Removing Wen Congyang and Tang Chen from the CC list because their
emails bounce. It seems that we will never learn about their motivation]
Quoted 58 lines · Michal Hocko
From: Michal Hocko <[email protected]>
Per-cpu numa_node provides a default node for each possible cpu. The association gets initialized during the boot when the architecture specific code explores cpu->NUMA affinity. When the whole NUMA node is removed though we are clearing this association
try_offline_node
check_and_unmap_cpu_on_node
unmap_cpu_on_node
numa_clear_node
numa_set_node(cpu, NUMA_NO_NODE)
This means that whoever calls cpu_to_node for a cpu associated with such a node will get NUMA_NO_NODE. This is problematic for two reasons. First it is fragile because __alloc_pages_node would simply blow up on an out-of-bound access. We have encountered this when loading kvm module
BUG: unable to handle kernel paging request at 00000000000021c0
IP: [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
PGD 800000ffe853e067 PUD 7336bbc067 PMD 0
Oops: 0000 [#1] SMP
[...]
CPU: 88 PID: 1223749 Comm: modprobe Tainted: G        W          4.4.156-94.64-default #1
task: ffff88727eff1880 ti: ffff887354490000 task.ti: ffff887354490000
RIP: 0010:[<ffffffff8119ccb3>]  [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
RSP: 0018:ffff887354493b40  EFLAGS: 00010202
RAX: 00000000000021c0 RBX: 0000000000000000 RCX: 0000000000000000
RDX: 0000000000000000 RSI: 0000000000000002 RDI: 00000000014000c0
RBP: 00000000014000c0 R08: ffffffffffffffff R09: 0000000000000000
R10: ffff88fffc89e790 R11: 0000000000014000 R12: 0000000000000101
R13: ffffffffa0772cd4 R14: ffffffffa0769ac0 R15: 0000000000000000
FS:  00007fdf2f2f1700(0000) GS:ffff88fffc880000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00000000000021c0 CR3: 00000077205ee000 CR4: 0000000000360670
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Stack:
 0000000000000086 014000c014d20400 ffff887354493bb8 ffff882614d20f4c
 0000000000000000 0000000000000046 0000000000000046 ffffffff810ac0c9
 ffff88ffe78c0000 ffffffff0000009f ffffe8ffe82d3500 ffff88ff8ac55000
Call Trace:
 [<ffffffffa07476cd>] alloc_vmcs_cpu+0x3d/0x90 [kvm_intel]
 [<ffffffffa0772c0c>] hardware_setup+0x781/0x849 [kvm_intel]
 [<ffffffffa04a1c58>] kvm_arch_hardware_setup+0x28/0x190 [kvm]
 [<ffffffffa04856fc>] kvm_init+0x7c/0x2d0 [kvm]
 [<ffffffffa0772cf2>] vmx_init+0x1e/0x32c [kvm_intel]
 [<ffffffff8100213a>] do_one_initcall+0xca/0x1f0
 [<ffffffff81193886>] do_init_module+0x5a/0x1d7
 [<ffffffff81112083>] load_module+0x1393/0x1c90
 [<ffffffff81112b30>] SYSC_finit_module+0x70/0xa0
 [<ffffffff8161cbc3>] entry_SYSCALL_64_fastpath+0x1e/0xb7
DWARF2 unwinder stuck at entry_SYSCALL_64_fastpath+0x1e/0xb7
on an older kernel but the code is basically the same in the current Linus tree as well. alloc_vmcs_cpu could use alloc_pages_nodemask which would recognize NUMA_NO_NODE and use alloc_pages_node which would translate it to numa_mem_id but that is wrong as well because it would use a cpu affinity of the local CPU which might be quite far from the original node.

Quoted 8 lines

But then the original node is getting/already off-lined. The allocation is going to come from a different node. alloc_pages_node() at least steer the allocation alway from VM_BUG_ON() because of NUMA_NO_NODE by replacing it with numa_mem_id().
If node fallback order is important for this allocation then could not it use __alloc_pages_nodemask() directly giving preference for its zonelist node and nodemask. Just curious.

How does the caller get the right node to allocate from? We do have the proper zone list for the offline node so why not use it?

Quoted 6 lines

Quoted 2 lines
Quoted 2 lines
It is also reasonable to expect that cpu_to_node will provide a sane value and there might be many more callers like that.
AFAICS there are two choices here. Either mark them NUMA_NO_NODE for all cpus of a node going offline or keep the existing mapping in case the node comes back again.

Or update the mapping to the closeses node. I have chosen to keep the mapping because it is the easiest and the most natural one.

Quoted 5 lines

Quoted 4 lines
Quoted 4 lines
The second problem is that __register_one_node relies on cpu_to_node to properly associate cpus back to the node when it is onlined. We do not want to lose that link as there is no arch independent way to get it from the early boot time AFAICS.

Quoted 8 lines

Retaining the links seems to be right unless unmap_cpu_on_node() is sort of a weak callback letting arch to decide.
Quoted 3 lines
Quoted 3 lines
Drop the whole check_and_unmap_cpu_on_node machinery and keep the association to fix both issues. The NODE_DATA(nid) is not deallocated
Though retaining the link is a problem in itself but the allocation related crash could be solved by exploring __alloc_pages_nodemask() options.

Yes that is the case but this looks like a very fragile fix to me. If you are getting a node number from cpu_to_node then you shouldn't really think about obscurities like which allocation function to use, right? You should just get a valid node number.

Do you see any problems with the patch as is?

Signature · 2 lines

Michal Hocko
SUSE Labs

Anshuman Khandual

2018-11-09 11:04 UTC #4 lore raw

Quoted 72 lines · Michal Hocko

Quoted 71 lines · Anshuman Khandual
Quoted 62 lines · Michal Hocko
[Removing Wen Congyang and Tang Chen from the CC list because their
emails bounce. It seems that we will never learn about their motivation]
Quoted 58 lines · Michal Hocko
From: Michal Hocko <[email protected]>
Per-cpu numa_node provides a default node for each possible cpu. The association gets initialized during the boot when the architecture specific code explores cpu->NUMA affinity. When the whole NUMA node is removed though we are clearing this association
try_offline_node
check_and_unmap_cpu_on_node
unmap_cpu_on_node
numa_clear_node
numa_set_node(cpu, NUMA_NO_NODE)
This means that whoever calls cpu_to_node for a cpu associated with such a node will get NUMA_NO_NODE. This is problematic for two reasons. First it is fragile because __alloc_pages_node would simply blow up on an out-of-bound access. We have encountered this when loading kvm module
BUG: unable to handle kernel paging request at 00000000000021c0
IP: [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
PGD 800000ffe853e067 PUD 7336bbc067 PMD 0
Oops: 0000 [#1] SMP
[...]
CPU: 88 PID: 1223749 Comm: modprobe Tainted: G        W          4.4.156-94.64-default #1
task: ffff88727eff1880 ti: ffff887354490000 task.ti: ffff887354490000
RIP: 0010:[<ffffffff8119ccb3>]  [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
RSP: 0018:ffff887354493b40  EFLAGS: 00010202
RAX: 00000000000021c0 RBX: 0000000000000000 RCX: 0000000000000000
RDX: 0000000000000000 RSI: 0000000000000002 RDI: 00000000014000c0
RBP: 00000000014000c0 R08: ffffffffffffffff R09: 0000000000000000
R10: ffff88fffc89e790 R11: 0000000000014000 R12: 0000000000000101
R13: ffffffffa0772cd4 R14: ffffffffa0769ac0 R15: 0000000000000000
FS:  00007fdf2f2f1700(0000) GS:ffff88fffc880000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00000000000021c0 CR3: 00000077205ee000 CR4: 0000000000360670
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Stack:
 0000000000000086 014000c014d20400 ffff887354493bb8 ffff882614d20f4c
 0000000000000000 0000000000000046 0000000000000046 ffffffff810ac0c9
 ffff88ffe78c0000 ffffffff0000009f ffffe8ffe82d3500 ffff88ff8ac55000
Call Trace:
 [<ffffffffa07476cd>] alloc_vmcs_cpu+0x3d/0x90 [kvm_intel]
 [<ffffffffa0772c0c>] hardware_setup+0x781/0x849 [kvm_intel]
 [<ffffffffa04a1c58>] kvm_arch_hardware_setup+0x28/0x190 [kvm]
 [<ffffffffa04856fc>] kvm_init+0x7c/0x2d0 [kvm]
 [<ffffffffa0772cf2>] vmx_init+0x1e/0x32c [kvm_intel]
 [<ffffffff8100213a>] do_one_initcall+0xca/0x1f0
 [<ffffffff81193886>] do_init_module+0x5a/0x1d7
 [<ffffffff81112083>] load_module+0x1393/0x1c90
 [<ffffffff81112b30>] SYSC_finit_module+0x70/0xa0
 [<ffffffff8161cbc3>] entry_SYSCALL_64_fastpath+0x1e/0xb7
DWARF2 unwinder stuck at entry_SYSCALL_64_fastpath+0x1e/0xb7
on an older kernel but the code is basically the same in the current Linus tree as well. alloc_vmcs_cpu could use alloc_pages_nodemask which would recognize NUMA_NO_NODE and use alloc_pages_node which would translate it to numa_mem_id but that is wrong as well because it would use a cpu affinity of the local CPU which might be quite far from the original node.
But then the original node is getting/already off-lined. The allocation is going to come from a different node. alloc_pages_node() at least steer the allocation alway from VM_BUG_ON() because of NUMA_NO_NODE by replacing it with numa_mem_id().

Quoted 6 lines

Quoted 3 lines
If node fallback order is important for this allocation then could not it use __alloc_pages_nodemask() directly giving preference for its zonelist node and nodemask. Just curious.
How does the caller get the right node to allocate from? We do have the proper zone list for the offline node so why not use it?

I get your point. NODE_DATA() for the off lined node is still around and so does the proper zone list for allocation, so why the caller should work around the problem by building it's preferred nodemask_t etc. No problem, I was just curious.

Quoted 10 lines

Quoted 7 lines
Quoted 3 lines
Quoted 3 lines
It is also reasonable to expect that cpu_to_node will provide a sane value and there might be many more callers like that.
AFAICS there are two choices here. Either mark them NUMA_NO_NODE for all cpus of a node going offline or keep the existing mapping in case the node comes back again.
Or update the mapping to the closeses node. I have chosen to keep the mapping because it is the easiest and the most natural one.

Agreed.

Quoted 12 lines

Quoted 12 lines
Quoted 5 lines
Quoted 5 lines
The second problem is that __register_one_node relies on cpu_to_node to properly associate cpus back to the node when it is onlined. We do not want to lose that link as there is no arch independent way to get it from the early boot time AFAICS.
Retaining the links seems to be right unless unmap_cpu_on_node() is sort of a weak callback letting arch to decide.
Quoted 3 lines
Quoted 3 lines
Drop the whole check_and_unmap_cpu_on_node machinery and keep the association to fix both issues. The NODE_DATA(nid) is not deallocated

Quoted 7 lines

Quoted 2 lines
Though retaining the link is a problem in itself but the allocation related crash could be solved by exploring __alloc_pages_nodemask() options.
Yes that is the case but this looks like a very fragile fix to me. If you are getting a node number from cpu_to_node then you shouldn't really think about obscurities like which allocation function to use, right? You should just get a valid node number.

Probably not worth it for the caller.

Quoted 2 lines

Do you see any problems with the patch as is?

No, this patch does remove an erroneous node-cpu map update which help solve a real crash.

Michal Hocko

author

2018-11-09 11:07 UTC #5 lore raw

Quoted 79 lines · Anshuman Khandual

Quoted 76 lines · Michal Hocko
Quoted 74 lines · Anshuman Khandual
Quoted 62 lines · Michal Hocko
[Removing Wen Congyang and Tang Chen from the CC list because their
emails bounce. It seems that we will never learn about their motivation]
Quoted 58 lines · Michal Hocko
From: Michal Hocko <[email protected]>
Per-cpu numa_node provides a default node for each possible cpu. The association gets initialized during the boot when the architecture specific code explores cpu->NUMA affinity. When the whole NUMA node is removed though we are clearing this association
try_offline_node
check_and_unmap_cpu_on_node
unmap_cpu_on_node
numa_clear_node
numa_set_node(cpu, NUMA_NO_NODE)
This means that whoever calls cpu_to_node for a cpu associated with such a node will get NUMA_NO_NODE. This is problematic for two reasons. First it is fragile because __alloc_pages_node would simply blow up on an out-of-bound access. We have encountered this when loading kvm module
BUG: unable to handle kernel paging request at 00000000000021c0
IP: [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
PGD 800000ffe853e067 PUD 7336bbc067 PMD 0
Oops: 0000 [#1] SMP
[...]
CPU: 88 PID: 1223749 Comm: modprobe Tainted: G        W          4.4.156-94.64-default #1
task: ffff88727eff1880 ti: ffff887354490000 task.ti: ffff887354490000
RIP: 0010:[<ffffffff8119ccb3>]  [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
RSP: 0018:ffff887354493b40  EFLAGS: 00010202
RAX: 00000000000021c0 RBX: 0000000000000000 RCX: 0000000000000000
RDX: 0000000000000000 RSI: 0000000000000002 RDI: 00000000014000c0
RBP: 00000000014000c0 R08: ffffffffffffffff R09: 0000000000000000
R10: ffff88fffc89e790 R11: 0000000000014000 R12: 0000000000000101
R13: ffffffffa0772cd4 R14: ffffffffa0769ac0 R15: 0000000000000000
FS:  00007fdf2f2f1700(0000) GS:ffff88fffc880000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00000000000021c0 CR3: 00000077205ee000 CR4: 0000000000360670
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Stack:
 0000000000000086 014000c014d20400 ffff887354493bb8 ffff882614d20f4c
 0000000000000000 0000000000000046 0000000000000046 ffffffff810ac0c9
 ffff88ffe78c0000 ffffffff0000009f ffffe8ffe82d3500 ffff88ff8ac55000
Call Trace:
 [<ffffffffa07476cd>] alloc_vmcs_cpu+0x3d/0x90 [kvm_intel]
 [<ffffffffa0772c0c>] hardware_setup+0x781/0x849 [kvm_intel]
 [<ffffffffa04a1c58>] kvm_arch_hardware_setup+0x28/0x190 [kvm]
 [<ffffffffa04856fc>] kvm_init+0x7c/0x2d0 [kvm]
 [<ffffffffa0772cf2>] vmx_init+0x1e/0x32c [kvm_intel]
 [<ffffffff8100213a>] do_one_initcall+0xca/0x1f0
 [<ffffffff81193886>] do_init_module+0x5a/0x1d7
 [<ffffffff81112083>] load_module+0x1393/0x1c90
 [<ffffffff81112b30>] SYSC_finit_module+0x70/0xa0
 [<ffffffff8161cbc3>] entry_SYSCALL_64_fastpath+0x1e/0xb7
DWARF2 unwinder stuck at entry_SYSCALL_64_fastpath+0x1e/0xb7
on an older kernel but the code is basically the same in the current Linus tree as well. alloc_vmcs_cpu could use alloc_pages_nodemask which would recognize NUMA_NO_NODE and use alloc_pages_node which would translate it to numa_mem_id but that is wrong as well because it would use a cpu affinity of the local CPU which might be quite far from the original node.
But then the original node is getting/already off-lined. The allocation is going to come from a different node. alloc_pages_node() at least steer the allocation alway from VM_BUG_ON() because of NUMA_NO_NODE by replacing it with numa_mem_id().
If node fallback order is important for this allocation then could not it use __alloc_pages_nodemask() directly giving preference for its zonelist node and nodemask. Just curious.

Quoted 6 lines

Quoted 2 lines
How does the caller get the right node to allocate from? We do have the proper zone list for the offline node so why not use it?
I get your point. NODE_DATA() for the off lined node is still around and so does the proper zone list for allocation, so why the caller should work around the problem by building it's preferred nodemask_t etc. No problem, I was just curious.

I thought I've made it cler in the changelog. If not, I am open to suggestions on how to make it more clear.

Signature · 2 lines

Michal Hocko
SUSE Labs

Andrew Morton

maintainer

2018-12-20 22:48 UTC #6 lore raw

In reply to Anshuman Khandual

Quoted 5 lines · On Fri, 9 Nov 2018 16:34:29 +0530 Anshuman Khandual…

Quoted 2 lines
Do you see any problems with the patch as is?
No, this patch does remove an erroneous node-cpu map update which help solve a real crash.

I think I'll take that as an ack.

Michal Hocko

author

2018-11-14 07:14 UTC #7 lore raw

In reply to Michal Hocko

It seems there were no objections here. So can we have it in linux-next for a wider testing a possibly target the next merge window?

Quoted 144 lines · Michal Hocko

From: Michal Hocko <[email protected]>
Per-cpu numa_node provides a default node for each possible cpu. The association gets initialized during the boot when the architecture specific code explores cpu->NUMA affinity. When the whole NUMA node is removed though we are clearing this association
try_offline_node
check_and_unmap_cpu_on_node
unmap_cpu_on_node
numa_clear_node
numa_set_node(cpu, NUMA_NO_NODE)
This means that whoever calls cpu_to_node for a cpu associated with such a node will get NUMA_NO_NODE. This is problematic for two reasons. First it is fragile because __alloc_pages_node would simply blow up on an out-of-bound access. We have encountered this when loading kvm module
BUG: unable to handle kernel paging request at 00000000000021c0
IP: [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
PGD 800000ffe853e067 PUD 7336bbc067 PMD 0
Oops: 0000 [#1] SMP
[...]
CPU: 88 PID: 1223749 Comm: modprobe Tainted: G        W          4.4.156-94.64-default #1
task: ffff88727eff1880 ti: ffff887354490000 task.ti: ffff887354490000
RIP: 0010:[<ffffffff8119ccb3>]  [<ffffffff8119ccb3>] __alloc_pages_nodemask+0x93/0xb70
RSP: 0018:ffff887354493b40  EFLAGS: 00010202
RAX: 00000000000021c0 RBX: 0000000000000000 RCX: 0000000000000000
RDX: 0000000000000000 RSI: 0000000000000002 RDI: 00000000014000c0
RBP: 00000000014000c0 R08: ffffffffffffffff R09: 0000000000000000
R10: ffff88fffc89e790 R11: 0000000000014000 R12: 0000000000000101
R13: ffffffffa0772cd4 R14: ffffffffa0769ac0 R15: 0000000000000000
FS:  00007fdf2f2f1700(0000) GS:ffff88fffc880000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 00000000000021c0 CR3: 00000077205ee000 CR4: 0000000000360670
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Stack:
 0000000000000086 014000c014d20400 ffff887354493bb8 ffff882614d20f4c
 0000000000000000 0000000000000046 0000000000000046 ffffffff810ac0c9
 ffff88ffe78c0000 ffffffff0000009f ffffe8ffe82d3500 ffff88ff8ac55000
Call Trace:
 [<ffffffffa07476cd>] alloc_vmcs_cpu+0x3d/0x90 [kvm_intel]
 [<ffffffffa0772c0c>] hardware_setup+0x781/0x849 [kvm_intel]
 [<ffffffffa04a1c58>] kvm_arch_hardware_setup+0x28/0x190 [kvm]
 [<ffffffffa04856fc>] kvm_init+0x7c/0x2d0 [kvm]
 [<ffffffffa0772cf2>] vmx_init+0x1e/0x32c [kvm_intel]
 [<ffffffff8100213a>] do_one_initcall+0xca/0x1f0
 [<ffffffff81193886>] do_init_module+0x5a/0x1d7
 [<ffffffff81112083>] load_module+0x1393/0x1c90
 [<ffffffff81112b30>] SYSC_finit_module+0x70/0xa0
 [<ffffffff8161cbc3>] entry_SYSCALL_64_fastpath+0x1e/0xb7
DWARF2 unwinder stuck at entry_SYSCALL_64_fastpath+0x1e/0xb7
on an older kernel but the code is basically the same in the current Linus tree as well. alloc_vmcs_cpu could use alloc_pages_nodemask which would recognize NUMA_NO_NODE and use alloc_pages_node which would translate it to numa_mem_id but that is wrong as well because it would use a cpu affinity of the local CPU which might be quite far from the original node. It is also reasonable to expect that cpu_to_node will provide a sane value and there might be many more callers like that.
The second problem is that __register_one_node relies on cpu_to_node to properly associate cpus back to the node when it is onlined. We do not want to lose that link as there is no arch independent way to get it from the early boot time AFAICS.
Drop the whole check_and_unmap_cpu_on_node machinery and keep the association to fix both issues. The NODE_DATA(nid) is not deallocated so it will stay in place and if anybody wants to allocate from that node then a fallback node will be used.
Thanks to Vlastimil Babka for his live system debugging skills that helped debugging the issue.
Debugged-by Vlastimil Babka <[email protected]>
Reported-by Miroslav Benes <[email protected]>
Fixes e13fe8695c57 ("cpu-hotplug,memory-hotplug: clear cpu_to_node() when offlining the node")
Cc Wen Congyang <[email protected]>
Cc Tang Chen <[email protected]>
Signed-off-by Michal Hocko <[email protected]>
---
Hi,
please note that I am sending this as an RFC even though this has been confirmed to fix the oops in kvm_intel module because I cannot simply tell that there are no other side effect that I do not see from the code reading. I would appreciate some background from people who have introduced this code e13fe8695c57 ("cpu-hotplug,memory-hotplug: clear cpu_to_node() when offlining the node") because the changelog doesn't really explain the motivation much.
 mm/memory_hotplug.c | 30 +-----------------------------
 1 file changed, 1 insertion(+), 29 deletions(-)
mm/memory_hotplug.c +1 ~~-29~~
index 2b2b3ccbbfb5..87aeafac54ee 100644
@@ -1753,34 +1753,6 @@ static int check_cpu_on_node(pg_data_t *pgdat)
 	return 0;
 }
 
-static void unmap_cpu_on_node(pg_data_t *pgdat)
-{
-#ifdef CONFIG_ACPI_NUMA
-	int cpu;
-
-	for_each_possible_cpu(cpu)
-		if (cpu_to_node(cpu) == pgdat->node_id)
-			numa_clear_node(cpu);
-#endif
-}
-
-static int check_and_unmap_cpu_on_node(pg_data_t *pgdat)
-{
-	int ret;
-
-	ret = check_cpu_on_node(pgdat);
-	if (ret)
-		return ret;
-
-	/*
-	 * the node will be offlined when we come here, so we can clear
-	 * the cpu_to_node() now.
-	 */
-
-	unmap_cpu_on_node(pgdat);
-	return 0;
-}
-
 /**
  * try_offline_node
  * @nid: the node ID
@@ -1813,7 +1785,7 @@ void try_offline_node(int nid)
 		return;
 	}
 
-	if (check_and_unmap_cpu_on_node(pgdat))
+	if (check_cpu_on_node(pgdat))
 		return;
 
 	/*
Signature · 1 line
2.19.1

Signature · 2 lines

Michal Hocko
SUSE Labs

Andrew Morton

maintainer

2018-11-14 23:18 UTC #8 lore raw

Quoted 3 lines · On Wed, 14 Nov 2018 08:14:42 +0100 Michal Hocko <[email protected]> wrote:

It seems there were no objections here. So can we have it in linux-next for a wider testing a possibly target the next merge window?

top-posting sucks!

I already have this queued for 4.21-rc1.

Michal Hocko

author

2018-11-15 07:35 UTC #9 lore raw

Quoted 7 lines · Andrew Morton

Quoted 3 lines · On Wed, 14 Nov 2018 08:14:42 +0100 Michal Hocko <[email protected]> wrote:
It seems there were no objections here. So can we have it in linux-next for a wider testing a possibly target the next merge window?
top-posting sucks!

I thought it would make your life easier in this case. Will do it differently next time.

Quoted 1 line

I already have this queued for 4.21-rc1.

Thanks! I must have missed the mm-commit email.

Signature · 2 lines

Michal Hocko
SUSE Labs