Re: kernel panic in reading /proc/kpageflags when enabling RAM-simulated PMEM
From: Naoya Horiguchi
Date: Thu Jun 07 2018 - 02:24:06 EST
On Wed, Jun 06, 2018 at 09:24:05AM +0000, Horiguchi Naoya(堀口 直也) wrote:
> On Wed, Jun 06, 2018 at 09:06:30AM +0000, Horiguchi Naoya(堀口 直也) wrote:
> > On Wed, Jun 06, 2018 at 10:53:19AM +0200, Oscar Salvador wrote:
> > > On Wed, Jun 06, 2018 at 10:04:08AM +0200, Oscar Salvador wrote:
> > > > On Wed, Jun 06, 2018 at 05:16:24AM +0000, Naoya Horiguchi wrote:
> > > > > On Tue, Jun 05, 2018 at 07:35:01AM +0000, Horiguchi Naoya(堀口 直也) wrote:
> > > > > > On Mon, Jun 04, 2018 at 06:18:36PM -0700, Matthew Wilcox wrote:
> > > > > > > On Tue, Jun 05, 2018 at 12:54:03AM +0000, Naoya Horiguchi wrote:
> > > > > > > > Reproduction precedure is like this:
> > > > > > > > - enable RAM based PMEM (with a kernel boot parameter like memmap=1G!4G)
> > > > > > > > - read /proc/kpageflags (or call tools/vm/page-types with no arguments)
> > > > > > > > (- my kernel config is attached)
> > > > > > > >
> > > > > > > > I spent a few days on this, but didn't reach any solutions.
> > > > > > > > So let me report this with some details below ...
> > > > > > > >
> > > > > > > > In the critial page request, stable_page_flags() is called with an argument
> > > > > > > > page whose ->compound_head was somehow filled with '0xffffffffffffffff'.
> > > > > > > > And compound_head() returns (struct page *)(head - 1), which explains the
> > > > > > > > address 0xfffffffffffffffe in the above message.
> > > > > > >
> > > > > > > Hm. compound_head shares with:
> > > > > > >
> > > > > > > struct list_head lru;
> > > > > > > struct list_head slab_list; /* uses lru */
> > > > > > > struct { /* Partial pages */
> > > > > > > struct page *next;
> > > > > > > unsigned long _compound_pad_1; /* compound_head */
> > > > > > > unsigned long _pt_pad_1; /* compound_head */
> > > > > > > struct dev_pagemap *pgmap;
> > > > > > > struct rcu_head rcu_head;
> > > > > > >
> > > > > > > None of them should be -1.
> > > > > > >
> > > > > > > > It seems that this kernel panic happens when reading kpageflags of pfn range
> > > > > > > > [0xbffd7, 0xc0000), which coresponds to a 'reserved' range.
> > > > > > > >
> > > > > > > > [ 0.000000] user-defined physical RAM map:
> > > > > > > > [ 0.000000] user: [mem 0x0000000000000000-0x000000000009fbff] usable
> > > > > > > > [ 0.000000] user: [mem 0x000000000009fc00-0x000000000009ffff] reserved
> > > > > > > > [ 0.000000] user: [mem 0x00000000000f0000-0x00000000000fffff] reserved
> > > > > > > > [ 0.000000] user: [mem 0x0000000000100000-0x00000000bffd6fff] usable
> > > > > > > > [ 0.000000] user: [mem 0x00000000bffd7000-0x00000000bfffffff] reserved
> > > > > > > > [ 0.000000] user: [mem 0x00000000feffc000-0x00000000feffffff] reserved
> > > > > > > > [ 0.000000] user: [mem 0x00000000fffc0000-0x00000000ffffffff] reserved
> > > > > > > > [ 0.000000] user: [mem 0x0000000100000000-0x000000013fffffff] persistent (type 12)
> > > > > > > >
> > > > > > > > So I guess 'memmap=' parameter might badly affect the memory initialization process.
> > > > > > > >
> > > > > > > > This problem doesn't reproduce on v4.17, so some pre-released patch introduces it.
> > > > > > > > I hope this info helps you find the solution/workaround.
> > > > > > >
> > > > > > > Can you try bisecting this? It could be one of my patches to reorder struct
> > > > > > > page, or it could be one of Pavel's deferred page initialisation patches.
> > > > > > > Or something else ;-)
> > > > > >
> > > > > > Thank you for the comment. I'm trying bisecting now, let you know the result later.
> > > > > >
> > > > > > And I found that my statement "not reproduce on v4.17" was wrong (I used
> > > > > > different kvm guests, which made some different test condition and misguided me),
> > > > > > this seems an older (at least < 4.15) bug.
> > > > >
> > > > > (Cc: Pavel)
> > > > >
> > > > > Bisection showed that the following commit introduced this issue:
> > > > >
> > > > > commit f7f99100d8d95dbcf09e0216a143211e79418b9f
> > > > > Author: Pavel Tatashin <pasha.tatashin@xxxxxxxxxx>
> > > > > Date: Wed Nov 15 17:36:44 2017 -0800
> > > > >
> > > > > mm: stop zeroing memory during allocation in vmemmap
> > > > >
> > > > > This patch postpones struct page zeroing to later stage of memory initialization.
> > > > > My kernel config disabled CONFIG_DEFERRED_STRUCT_PAGE_INIT so two callsites of
> > > > > __init_single_page() were never reached. So in such case, struct pages populated
> > > > > by vmemmap_pte_populate() could be left uninitialized?
> > > > > And I'm not sure yet how this issue becomes visible with memmap= setting.
> > > >
> > > > I think that this becomes visible because memmap=x!y creates a persistent memory region:
> > > >
> > > > parse_memmap_one
> > > > {
> > > > ...
> > > > } else if (*p == '!') {
> > > > start_at = memparse(p+1, &p);
> > > > e820__range_add(start_at, mem_size, E820_TYPE_PRAM);
> > > > ...
> > > > }
> > > >
> > > > and this region it is not added neither in memblock.memory nor in memblock.reserved.
> > > > Ranges in memblock.memory get zeroed in memmap_init_zone(), while memblock.reserved get zeroed
> > > > in free_low_memory_core_early():
> > > >
> > > > static unsigned long __init free_low_memory_core_early(void)
> > > > {
> > > > ...
> > > > for_each_reserved_mem_region(i, &start, &end)
> > > > reserve_bootmem_region(start, end);
> > > > ...
> > > > }
> > > >
> > > >
> > > > Maybe I am mistaken, but I think that persistent memory regions should be marked as reserved.
> > > > A comment in do_mark_busy() suggests this:
> > > >
> > > > static bool __init do_mark_busy(enum e820_type type, struct resource *res)
> > > > {
> > > >
> > > > ...
> > > > /*
> > > > * Treat persistent memory like device memory, i.e. reserve it
> > > > * for exclusive use of a driver
> > > > */
> > > > ...
> > > > }
> > > >
> > > >
> > > > I wonder if something like this could work and if so, if it is right (i haven't tested it yet):
> > > >
> > > > diff --git a/arch/x86/kernel/e820.c b/arch/x86/kernel/e820.c
> > > > index 71c11ad5643e..3c9686ef74e5 100644
> > > > --- a/arch/x86/kernel/e820.c
> > > > +++ b/arch/x86/kernel/e820.c
> > > > @@ -1247,6 +1247,11 @@ void __init e820__memblock_setup(void)
> > > > if (end != (resource_size_t)end)
> > > > continue;
> > > >
> > > > + if (entry->type == E820_TYPE_PRAM || entry->type == E820_TYPE_PMEM) {
> > > > + memblock_reserve(entry->addr, entry->size);
> > > > + continue;
> > > > + }
> > > > +
> > > > if (entry->type != E820_TYPE_RAM && entry->type != E820_TYPE_RESERVED_KERN)
> > > > continue;
> > >
> > > It does not seem to work, so the reasoning might be incorrect.
> >
> > Thank you for the comment.
> >
> > One note is that the memory region with "broken struct page" is a typical
> > reserved region, not a pmem region. Strangely reading offset 0xbffd7 of
> > /proc/kpageflags is OK if pmem region does not exist, but NG if pmem region exists.
> > Reading the offset like 0x100000 (on pmem region) does not cause the crash,
> > so pmem region seems properly set up.
> >
> > [ 0.000000] user-defined physical RAM map:
> > [ 0.000000] user: [mem 0x0000000000000000-0x000000000009fbff] usable
> > [ 0.000000] user: [mem 0x000000000009fc00-0x000000000009ffff] reserved
> > [ 0.000000] user: [mem 0x00000000000f0000-0x00000000000fffff] reserved
> > [ 0.000000] user: [mem 0x0000000000100000-0x00000000bffd6fff] usable
> > [ 0.000000] user: [mem 0x00000000bffd7000-0x00000000bfffffff] reserved ===> "broken struct page" region
> > [ 0.000000] user: [mem 0x00000000feffc000-0x00000000feffffff] reserved
> > [ 0.000000] user: [mem 0x00000000fffc0000-0x00000000ffffffff] reserved
> > [ 0.000000] user: [mem 0x0000000100000000-0x000000013fffffff] persistent (type 12) => pmem region
> > [ 0.000000] user: [mem 0x0000000140000000-0x000000023fffffff] usable
> >
>
> I have another note:
>
> > My kernel config disabled CONFIG_DEFERRED_STRUCT_PAGE_INIT so two callsites of
> > __init_single_page() were never reached. So in such case, struct pages populated
> > by vmemmap_pte_populate() could be left uninitialized?
>
> I quickly checked whether enabling CONFIG_DEFERRED_STRUCT_PAGE_INIT affect
> the issue. And found that the kernel panic happens even with this config enabled.
> So I'm still confused...
Let me share some new facts:
I gave accidentally an inconvenient memmap layout like 'memmap=1G!4G' in
2 NUMA node with 8 GB memory.
While I didn't intended this, but 4GB is the address starting some memory
block when no "memmap=" option is provided.
(messages from free_area_init_nodes() for no "memmap=" case
[ 0.000000] Early memory node ranges
[ 0.000000] node 0: [mem 0x0000000000001000-0x000000000009efff]
[ 0.000000] node 0: [mem 0x0000000000100000-0x00000000bffd6fff]
[ 0.000000] node 0: [mem 0x0000000100000000-0x000000013fffffff] // <---
[ 0.000000] node 1: [mem 0x0000000140000000-0x000000023fffffff]
When "memmap=1G!4G" is given, the range [0x0000000100000000-0x000000013fffffff]
disappears and kernel messages are like below:
(messages from free_area_init_nodes() for "memmap=1G!4G" case
[ 0.000000] Early memory node ranges
[ 0.000000] node 0: [mem 0x0000000000001000-0x000000000009efff]
[ 0.000000] node 0: [mem 0x0000000000100000-0x00000000bffd6fff]
[ 0.000000] node 1: [mem 0x0000000140000000-0x000000023fffffff]
This makes kernel think that the end pfn of node 0 is 0 0xbffd7
instead of 0x140000, which affects the memory initialization process.
memmap_init_zone() calls __init_single_page() for each page within a zone,
so if zone->spanned_pages are underestimated, some pages are left uninitialized.
If I provide 'memmap=1G!7G', the kernel panic does not reproduce and
kernel messages are like below.
(messages from free_area_init_nodes() for "memmap=1G!7G" case
[ 0.000000] Early memory node ranges
[ 0.000000] node 0: [mem 0x0000000000001000-0x000000000009efff]
[ 0.000000] node 0: [mem 0x0000000000100000-0x00000000bffd6fff]
[ 0.000000] node 0: [mem 0x0000000100000000-0x000000013fffffff]
[ 0.000000] node 1: [mem 0x0000000140000000-0x00000001bfffffff]
[ 0.000000] node 1: [mem 0x0000000200000000-0x000000023fffffff]
I think that in order to fix this, we need some conditions and/or prechecks
for memblock layout, does it make sense? Or any other better approaches?
Thanks,
Naoya Horiguchi