[Resending with the mailing lists CCed - sorry for spamming]
Hi Andrew,
Oscar has reported two issue when playing with the memory hotplug
[1][2]. The first one seems more serious and patch 1 should address it.
In short we are overly optimistic about zone movable not containing any
non-movable pages and after 72b39cfc4d75 ("mm, memory_hotplug: do not
fail offlining too early") this can lead to a seemingly stuck (still
interruptible by a signal) memory offline.
Patch 2 fixes an over-eager warning which is not harmful but surely
annoying.
I know we are late in the release cycle but I guess both would be
candidates for rc7. They are simple enough and they should be
"obviously" correct. If you would like more time for them for testing
then I am perfectly fine postponing to the next merge window of course.
[1] http://lkml.kernel.org/r/[email protected]
[2] http://lkml.kernel.org/r/[email protected]
Michal Hocko (2):
mm, memory_hotplug: make has_unmovable_pages more robust
mm: do not warn on offline nodes unless the specific node is explicitly requested
Diffstat
include/linux/gfp.h | 2 +-
mm/page_alloc.c | 16 ++++++++++------
2 files changed, 11 insertions(+), 7 deletions(-)
From: Michal Hocko <[email protected]>
Oscar has reported:
: Due to an unfortunate setting with movablecore, memblocks containing bootmem
: memory (pages marked by get_page_bootmem()) ended up marked in zone_movable.
: So while trying to remove that memory, the system failed in do_migrate_range
: and __offline_pages never returned.
:
: This can be reproduced by running
: qemu-system-x86_64 -m 6G,slots=8,maxmem=8G -numa node,mem=4096M -numa node,mem=2048M
: and movablecore=4G kernel command line
:
: linux kernel: BIOS-provided physical RAM map:
: linux kernel: BIOS-e820: [mem 0x0000000000000000-0x000000000009fbff] usable
: linux kernel: BIOS-e820: [mem 0x000000000009fc00-0x000000000009ffff] reserved
: linux kernel: BIOS-e820: [mem 0x00000000000f0000-0x00000000000fffff] reserved
: linux kernel: BIOS-e820: [mem 0x0000000000100000-0x00000000bffdffff] usable
: linux kernel: BIOS-e820: [mem 0x00000000bffe0000-0x00000000bfffffff] reserved
: linux kernel: BIOS-e820: [mem 0x00000000feffc000-0x00000000feffffff] reserved
: linux kernel: BIOS-e820: [mem 0x00000000fffc0000-0x00000000ffffffff] reserved
: linux kernel: BIOS-e820: [mem 0x0000000100000000-0x00000001bfffffff] usable
: linux kernel: NX (Execute Disable) protection: active
: linux kernel: SMBIOS 2.8 present.
: linux kernel: DMI: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.0.0-prebuilt.qemu-project.org
: linux kernel: Hypervisor detected: KVM
: linux kernel: e820: update [mem 0x00000000-0x00000fff] usable ==> reserved
: linux kernel: e820: remove [mem 0x000a0000-0x000fffff] usable
: linux kernel: last_pfn = 0x1c0000 max_arch_pfn = 0x400000000
:
: linux kernel: SRAT: PXM 0 -> APIC 0x00 -> Node 0
: linux kernel: SRAT: PXM 1 -> APIC 0x01 -> Node 1
: linux kernel: ACPI: SRAT: Node 0 PXM 0 [mem 0x00000000-0x0009ffff]
: linux kernel: ACPI: SRAT: Node 0 PXM 0 [mem 0x00100000-0xbfffffff]
: linux kernel: ACPI: SRAT: Node 0 PXM 0 [mem 0x100000000-0x13fffffff]
: linux kernel: ACPI: SRAT: Node 1 PXM 1 [mem 0x140000000-0x1bfffffff]
: linux kernel: ACPI: SRAT: Node 0 PXM 0 [mem 0x1c0000000-0x43fffffff] hotplug
: linux kernel: NUMA: Node 0 [mem 0x00000000-0x0009ffff] + [mem 0x00100000-0xbfffffff] -> [mem 0x0
: linux kernel: NUMA: Node 0 [mem 0x00000000-0xbfffffff] + [mem 0x100000000-0x13fffffff] -> [mem 0
: linux kernel: NODE_DATA(0) allocated [mem 0x13ffd6000-0x13fffffff]
: linux kernel: NODE_DATA(1) allocated [mem 0x1bffd3000-0x1bfffcfff]
:
: zoneinfo shows that the zone movable is placed into both numa nodes:
: Node 0, zone Movable
: pages free 160140
: min 1823
: low 2278
: high 2733
: spanned 262144
: present 262144
: managed 245670
: Node 1, zone Movable
: pages free 448427
: min 3827
: low 4783
: high 5739
: spanned 524288
: present 524288
: managed 515766
Note how only Node 0 has a hutplugable memory region which would rule
it out from the early memblock allocations (most likely memmap). Node1
will surely contain memmaps on the same node and those would prevent
offlining to succeed. So this is arguably a configuration issue.
Although one could argue that we should be more clever and rule early
allocations from the zone movable. This would be correct but probably
not worth the effort considering what a hack movablecore is.
Anyway, We could do better for those cases though. We rely on
start_isolate_page_range resp. has_unmovable_pages to do their job. The
first one isolates the whole range to be offlined so that we do not
allocate from it anymore and the later makes sure we are not stumbling
over non-migrateable pages.
has_unmovable_pages is overly optimistic, however. It doesn't check all
the pages if we are withing zone_movable because we rely that those
pages will be always migrateable. As it turns out we are still not
perfect there. While bootmem pages in zonemovable sound like a clear bug
which should be fixed let's remove the optimization for now and warn if
we encounter unmovable pages in zone_movable in the meantime. That
should help for now at least.
Btw. this wasn't a real problem until 72b39cfc4d75 ("mm, memory_hotplug:
do not fail offlining too early") because we used to have a small number
of retries and then failed. This turned out to be too fragile though.
Reported-by: Oscar Salvador <[email protected]>
Tested-by: Oscar Salvador <[email protected]>
Signed-off-by: Michal Hocko <[email protected]>
---
mm/page_alloc.c | 16 ++++++++++------
1 file changed, 10 insertions(+), 6 deletions(-)
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 3c6f4008ea55..b9a45753244d 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -7629,11 +7629,12 @@ bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
unsigned long pfn, iter, found;
/*
- * For avoiding noise data, lru_add_drain_all() should be called
- * If ZONE_MOVABLE, the zone never contains unmovable pages
+ * TODO we could make this much more efficient by not checking every
+ * page in the range if we know all of them are in MOVABLE_ZONE and
+ * that the movable zone guarantees that pages are migratable but
+ * the later is not the case right now unfortunatelly. E.g. movablecore
+ * can still lead to having bootmem allocations in zone_movable.
*/
- if (zone_idx(zone) == ZONE_MOVABLE)
- return false;
/*
* CMA allocations (alloc_contig_range) really need to mark isolate
@@ -7654,7 +7655,7 @@ bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
page = pfn_to_page(check);
if (PageReserved(page))
- return true;
+ goto unmovable;
/*
* Hugepages are not in LRU lists, but they're movable.
@@ -7704,9 +7705,12 @@ bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
* page at boot.
*/
if (found > count)
- return true;
+ goto unmovable;
}
return false;
+unmovable:
+ WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
+ return true;
}
#if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
--
2.17.0
From: Michal Hocko <[email protected]>
Oscar has noticed that we splat
linux kernel: WARNING: CPU: 0 PID: 64 at ./include/linux/gfp.h:467 vmemmap_alloc_block+0x4e/0xc9
[...]
linux kernel: CPU: 0 PID: 64 Comm: kworker/u4:1 Tainted: G W E 4.17.0-rc5-next-20180517-1-default+ #66
linux kernel: Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.0.0-prebuilt.qemu-project.org 04/01/2014
linux kernel: Workqueue: kacpi_hotplug acpi_hotplug_work_fn
linux kernel: RIP: 0010:vmemmap_alloc_block+0x4e/0xc9
linux kernel: Code: fb ff 8d 69 01 75 07 65 8b 1d 9d cb 93 7e 81 fb ff 03 00 00 76 02 0f 0b 48 63 c3 48 0f a3 05 c8 b1 b4 00 0f 92 c0 84 c0 75 02 <0f> 0b 31 c9 89 da 89 ee bf c0 06 40 01 e8 0f d1 ad ff 48 85 c0 74
linux kernel: RSP: 0018:ffffc90000d03bf0 EFLAGS: 00010246
linux kernel: RAX: 0000000000000000 RBX: 0000000000000001 RCX: 0000000000000008
linux kernel: RDX: 0000000000000000 RSI: 0000000000000001 RDI: 00000000000001ff
linux kernel: RBP: 0000000000000009 R08: 0000000000000001 R09: ffffc90000d03ae8
linux kernel: R10: 0000000000000001 R11: 0000000000000000 R12: ffffea0006000000
linux kernel: R13: ffffea0005e00000 R14: ffffea0006000000 R15: 0000000000000001
linux kernel: FS: 0000000000000000(0000) GS:ffff88013fc00000(0000) knlGS:0000000000000000
linux kernel: CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
linux kernel: CR2: 00007fa92a698018 CR3: 00000001184ce000 CR4: 00000000000006f0
linux kernel: DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
linux kernel: DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
linux kernel: Call Trace:
linux kernel: vmemmap_populate+0xf2/0x2ae
linux kernel: sparse_mem_map_populate+0x28/0x35
linux kernel: sparse_add_one_section+0x4c/0x187
linux kernel: __add_pages+0xe7/0x1a0
linux kernel: add_pages+0x16/0x70
linux kernel: add_memory_resource+0xa3/0x1d0
linux kernel: add_memory+0xe4/0x110
linux kernel: acpi_memory_device_add+0x134/0x2e0
linux kernel: acpi_bus_attach+0xd9/0x190
linux kernel: acpi_bus_scan+0x37/0x70
linux kernel: acpi_device_hotplug+0x389/0x4e0
linux kernel: acpi_hotplug_work_fn+0x1a/0x30
linux kernel: process_one_work+0x146/0x340
linux kernel: worker_thread+0x47/0x3e0
linux kernel: kthread+0xf5/0x130
linux kernel: ? max_active_store+0x60/0x60
linux kernel: ? kthread_bind+0x10/0x10
linux kernel: ret_from_fork+0x35/0x40
linux kernel: ---[ end trace 2e2241f4e2f2f018 ]---
====
when adding memory to a node that is currently offline.
The VM_WARN_ON is just too loud without a good reason. In this
particular case we are doing
alloc_pages_node(node, GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN, order)
so we do not insist on allocating from the given node (it is more a
hint) so we can fall back to any other populated node and moreover we
explicitly ask to not warn for the allocation failure.
Soften the warning only to cases when somebody asks for the given node
explicitly by __GFP_THISNODE.
Reported-by: Oscar Salvador <[email protected]>
Tested-by: Oscar Salvador <[email protected]>
Signed-off-by: Michal Hocko <[email protected]>
---
include/linux/gfp.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/include/linux/gfp.h b/include/linux/gfp.h
index 036846fc00a6..7f860ea29ec6 100644
--- a/include/linux/gfp.h
+++ b/include/linux/gfp.h
@@ -464,7 +464,7 @@ static inline struct page *
__alloc_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
{
VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES);
- VM_WARN_ON(!node_online(nid));
+ VM_WARN_ON((gfp_mask & __GFP_THISNODE) && !node_online(nid));
return __alloc_pages(gfp_mask, order, nid);
}
--
2.17.0
Reviewed-by: Pavel Tatashin <[email protected]>
Reviewed-by: Pavel Tatashin <[email protected]>
On 05/23/2018 06:25 PM, Michal Hocko wrote:
> when adding memory to a node that is currently offline.
>
> The VM_WARN_ON is just too loud without a good reason. In this
> particular case we are doing
> alloc_pages_node(node, GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN, order)
>
> so we do not insist on allocating from the given node (it is more a
> hint) so we can fall back to any other populated node and moreover we
> explicitly ask to not warn for the allocation failure.
>
> Soften the warning only to cases when somebody asks for the given node
> explicitly by __GFP_THISNODE.
node hint passed here eventually goes into __alloc_pages_nodemask()
function which then picks up the applicable zonelist irrespective of
the GFP flag __GFP_THISNODE. Though we can go into zones of other
nodes if the present node (whose zonelist got picked up) does not
have any memory in it's zones. So warning here might not be without
any reason. But yes, if the request has __GFP_NOWARN it makes sense
not to print any warning.
On Wed 23-05-18 19:15:51, Anshuman Khandual wrote:
> On 05/23/2018 06:25 PM, Michal Hocko wrote:
> > when adding memory to a node that is currently offline.
> >
> > The VM_WARN_ON is just too loud without a good reason. In this
> > particular case we are doing
> > alloc_pages_node(node, GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN, order)
> >
> > so we do not insist on allocating from the given node (it is more a
> > hint) so we can fall back to any other populated node and moreover we
> > explicitly ask to not warn for the allocation failure.
> >
> > Soften the warning only to cases when somebody asks for the given node
> > explicitly by __GFP_THISNODE.
>
> node hint passed here eventually goes into __alloc_pages_nodemask()
> function which then picks up the applicable zonelist irrespective of
> the GFP flag __GFP_THISNODE.
__GFP_THISNODE should enforce the given node without any fallbacks
unless something has changed recently.
> Though we can go into zones of other
> nodes if the present node (whose zonelist got picked up) does not
> have any memory in it's zones. So warning here might not be without
> any reason.
I am not sure I follow. Are you suggesting a different VM_WARN_ON?
--
Michal Hocko
SUSE Labs
On 05/23/2018 07:36 PM, Michal Hocko wrote:
> On Wed 23-05-18 19:15:51, Anshuman Khandual wrote:
>> On 05/23/2018 06:25 PM, Michal Hocko wrote:
>>> when adding memory to a node that is currently offline.
>>>
>>> The VM_WARN_ON is just too loud without a good reason. In this
>>> particular case we are doing
>>> alloc_pages_node(node, GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN, order)
>>>
>>> so we do not insist on allocating from the given node (it is more a
>>> hint) so we can fall back to any other populated node and moreover we
>>> explicitly ask to not warn for the allocation failure.
>>>
>>> Soften the warning only to cases when somebody asks for the given node
>>> explicitly by __GFP_THISNODE.
>>
>> node hint passed here eventually goes into __alloc_pages_nodemask()
>> function which then picks up the applicable zonelist irrespective of
>> the GFP flag __GFP_THISNODE.
>
> __GFP_THISNODE should enforce the given node without any fallbacks
> unless something has changed recently.
Right. I was just saying requiring given preferred node to be online
whose zonelist (hence allocation zone fallback order) is getting picked
up during allocation and warning when that is not online still makes
sense. We should only hide the warning if the allocation request has
__GFP_NOWARN.
>
>> Though we can go into zones of other
>> nodes if the present node (whose zonelist got picked up) does not
>> have any memory in it's zones. So warning here might not be without
>> any reason.
>
> I am not sure I follow. Are you suggesting a different VM_WARN_ON?
I am just suggesting this instead.
diff --git a/include/linux/gfp.h b/include/linux/gfp.h
index 036846fc00a6..7f860ea29ec6 100644
--- a/include/linux/gfp.h
+++ b/include/linux/gfp.h
@@ -464,7 +464,7 @@ static inline struct page *
__alloc_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
{
VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES);
- VM_WARN_ON(!node_online(nid));
+ VM_WARN_ON(!(gfp_mask & __GFP_NOWARN) && !node_online(nid));
return __alloc_pages(gfp_mask, order, nid);
}
On Thu 24-05-18 08:52:14, Anshuman Khandual wrote:
> On 05/23/2018 07:36 PM, Michal Hocko wrote:
> > On Wed 23-05-18 19:15:51, Anshuman Khandual wrote:
> >> On 05/23/2018 06:25 PM, Michal Hocko wrote:
> >>> when adding memory to a node that is currently offline.
> >>>
> >>> The VM_WARN_ON is just too loud without a good reason. In this
> >>> particular case we are doing
> >>> alloc_pages_node(node, GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN, order)
> >>>
> >>> so we do not insist on allocating from the given node (it is more a
> >>> hint) so we can fall back to any other populated node and moreover we
> >>> explicitly ask to not warn for the allocation failure.
> >>>
> >>> Soften the warning only to cases when somebody asks for the given node
> >>> explicitly by __GFP_THISNODE.
> >>
> >> node hint passed here eventually goes into __alloc_pages_nodemask()
> >> function which then picks up the applicable zonelist irrespective of
> >> the GFP flag __GFP_THISNODE.
> >
> > __GFP_THISNODE should enforce the given node without any fallbacks
> > unless something has changed recently.
>
> Right. I was just saying requiring given preferred node to be online
> whose zonelist (hence allocation zone fallback order) is getting picked
> up during allocation and warning when that is not online still makes
> sense.
Why? We have a fallback and that is expected to be used. How does
offline differ from depleted node from the semantical point of view?
> We should only hide the warning if the allocation request has
> __GFP_NOWARN.
>
> >
> >> Though we can go into zones of other
> >> nodes if the present node (whose zonelist got picked up) does not
> >> have any memory in it's zones. So warning here might not be without
> >> any reason.
> >
> > I am not sure I follow. Are you suggesting a different VM_WARN_ON?
>
> I am just suggesting this instead.
>
> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> index 036846fc00a6..7f860ea29ec6 100644
> --- a/include/linux/gfp.h
> +++ b/include/linux/gfp.h
> @@ -464,7 +464,7 @@ static inline struct page *
> __alloc_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
> {
> VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES);
> - VM_WARN_ON(!node_online(nid));
> + VM_WARN_ON(!(gfp_mask & __GFP_NOWARN) && !node_online(nid));
>
> return __alloc_pages(gfp_mask, order, nid);
> }
I have considered that but I fail to see why should we warn about
regular GFP_KERNEL allocations as mentioned above. Just consider an
allocation for the preffered node. Do you want to warn just because that
node went offline?
--
Michal Hocko
SUSE Labs
On 05/24/2018 01:30 PM, Michal Hocko wrote:
> On Thu 24-05-18 08:52:14, Anshuman Khandual wrote:
>> On 05/23/2018 07:36 PM, Michal Hocko wrote:
>>> On Wed 23-05-18 19:15:51, Anshuman Khandual wrote:
>>>> On 05/23/2018 06:25 PM, Michal Hocko wrote:
>>>>> when adding memory to a node that is currently offline.
>>>>>
>>>>> The VM_WARN_ON is just too loud without a good reason. In this
>>>>> particular case we are doing
>>>>> alloc_pages_node(node, GFP_KERNEL|__GFP_RETRY_MAYFAIL|__GFP_NOWARN, order)
>>>>>
>>>>> so we do not insist on allocating from the given node (it is more a
>>>>> hint) so we can fall back to any other populated node and moreover we
>>>>> explicitly ask to not warn for the allocation failure.
>>>>>
>>>>> Soften the warning only to cases when somebody asks for the given node
>>>>> explicitly by __GFP_THISNODE.
>>>>
>>>> node hint passed here eventually goes into __alloc_pages_nodemask()
>>>> function which then picks up the applicable zonelist irrespective of
>>>> the GFP flag __GFP_THISNODE.
>>>
>>> __GFP_THISNODE should enforce the given node without any fallbacks
>>> unless something has changed recently.
>>
>> Right. I was just saying requiring given preferred node to be online
>> whose zonelist (hence allocation zone fallback order) is getting picked
>> up during allocation and warning when that is not online still makes
>> sense.
>
> Why? We have a fallback and that is expected to be used. How does
> offline differ from depleted node from the semantical point of view?
Hmm, right. I agree. Offlined and depleted nodes are same from memory
allocation semantics point of view. It will proceed picking up next
available zones on the zonelist in the fallback order exactly in the
same fashion either way.
>
>> We should only hide the warning if the allocation request has
>> __GFP_NOWARN.
>>
>>>
>>>> Though we can go into zones of other
>>>> nodes if the present node (whose zonelist got picked up) does not
>>>> have any memory in it's zones. So warning here might not be without
>>>> any reason.
>>>
>>> I am not sure I follow. Are you suggesting a different VM_WARN_ON?
>>
>> I am just suggesting this instead.
>>
>> diff --git a/include/linux/gfp.h b/include/linux/gfp.h
>> index 036846fc00a6..7f860ea29ec6 100644
>> --- a/include/linux/gfp.h
>> +++ b/include/linux/gfp.h
>> @@ -464,7 +464,7 @@ static inline struct page *
>> __alloc_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
>> {
>> VM_BUG_ON(nid < 0 || nid >= MAX_NUMNODES);
>> - VM_WARN_ON(!node_online(nid));
>> + VM_WARN_ON(!(gfp_mask & __GFP_NOWARN) && !node_online(nid));
>>
>> return __alloc_pages(gfp_mask, order, nid);
>> }
>
> I have considered that but I fail to see why should we warn about
> regular GFP_KERNEL allocations as mentioned above. Just consider an
> allocation for the preffered node. Do you want to warn just because that
> node went offline?
As you have mentioned before, the semantics is similar when the node is
offlined compared to when its depleted. Right. I tend to agree with your
approach of not warning in such situations.