Re: [PATCH] mm: page_alloc: ignore init_on_free=1 for page alloc
From: Sergei Trofimovich
Date: Fri Mar 26 2021 - 13:26:13 EST
On Fri, 26 Mar 2021 15:17:00 +0100
Vlastimil Babka <vbabka@xxxxxxx> wrote:
> On 3/26/21 12:26 PM, Sergei Trofimovich wrote:
> > init_on_free=1 does not guarantee that free pages contain only zero bytes.
> >
> > Some examples:
> > 1. page_poison=on takes presedence over init_on_alloc=1 / ini_on_free=1
>
> Yes, and it spits out a message that you enabled both and poisoning takes
> precedence. It was that way even before my changes IIRC, but not consistent.
Yeah. I probably should not have included this case as page_poison=on actually
made my machine boot just fine. My main focus was to understand why I an seeing
the crash on kernel with init_on_alloc=1 init_on_free=1 and most debugging options
on.
My apologies! I'll try to find where this extra poisoning comes from.
Making a step back and explaining my setup:
Initially it's an ia64 box that manages to consistently corrupt memory
on socket free; https://lkml.org/lkml/2021/2/23/653
To get better understanding where corruption comes from I enabled
A Lot of VM, pagealloc and slab debugging options. Full config:
https://dev.gentoo.org/~slyfox/configs/guppy-config-5.12.0-rc4-00016-g427684abc9fd-dirty
I boot machine as:
[ 0.000000] Kernel command line: BOOT_IMAGE=/boot/vmlinuz-5.12.0-rc4-00016-g427684abc9fd-dirty root=/dev/sda3 ro slab_nomerge memblock=debug debug_pagealloc=1 hardened_usercopy=1 page_owner=on page_poison=0 init_on_alloc=1 init_on_free=1 debug_guardpage_minorder=0
My boot log:
https://dev.gentoo.org/~slyfox/bugs/ia64-boot-bug/2021-03-26-init_on_alloc-fail
Caveats in reading boot log:
- kernel crashes too early: stack unwinder does not have working kmalloc() yet
- kernel crashes in MCE handler: normally it should not. It's an unrelated bug
that makes backtrace useless. I'll try to fix it later, but it will not be fast.
- I added a bunch of printk()s around the crash.
The important pernel boot failure part is:
[ 0.000000] put_kernel_page: pmd=e000000100000000
[ 0.000000] pmd:(____ptrval____): aaaaaaaaaaaaaaaa aaaaaaaaaaaaaaaa aaaaaaaaaaaaaaaa aaaaaaaaaaaaaaaa ................................
Note 1: I do not really enable page_poison at runtime and was misleading you
in previous emails. (I initially assumed kernel_poison_pages() poisons pages
unconditionally but you all explained it does not). Something else manages to
poison my pmd(s?).
Note 2: I have many other debugging options enabled that might trigger
poisoning.
> > 2. free_pages_prepare() always poisons pages:
> >
> > if (want_init_on_free())
> > kernel_init_free_pages(page, 1 << order);
> > kernel_poison_pages(page, 1 << order
>
> kernel_poison_pages() includes a test if poisoning is enabled. And in that case
> want_init_on_free() shouldn't be. see init_mem_debugging_and_hardening()
I completely missed that! Thank you! Will try to trace real cause of poisoning.
> > I observed use of poisoned pages as the crash on ia64 booted with
> > init_on_free=1 init_on_alloc=1 (CONFIG_PAGE_POISONING=y config).
> > There pmd page contained 0xaaaaaaaa poison pages and led to early crash.
>
> Hm but that looks lika a sign that ia64 pmd allocation should use __GFP_ZERO and
> doesn't. It shouldn't rely on init_on_alloc or init_on_free being enabled.
ia64 does use __GFP_ZERO (I even tried to add it manually to pmd_alloc_one()
before I realized all _PGTABLEs imply __GFP_ZERO).
I'll provide the call chain I arrived at for completeness:
- [ia64 boots]
- mem_init() (defined at arch/ia64/mm/init.c)
-> setup_gate() (defined at arch/ia64/mm/init.c)
-> put_kernel_page() (defined at arch/ia64/mm/init.c)
-> [NOTE: from now on it's all generic code, not ia64-speficic]
-> pmd_alloc() (defined at include/linux/mm.h)
-> __pmd_alloc() (defined at mm/memory.c)
-> [under #ifndef __PAGETABLE_PMD_FOLDED] pmd_alloc_one() (defined at include/asm-generic/pgalloc.h)
-> pmd_alloc_one() [defined at include/asm-generic/pgalloc.h):
static inline pmd_t *pmd_alloc_one(struct mm_struct *mm, unsigned long addr)
{
struct page *page;
gfp_t gfp = GFP_PGTABLE_USER;
if (mm == &init_mm)
gfp = GFP_PGTABLE_KERNEL;
page = alloc_pages(gfp, 0);
if (!page)
return NULL;
if (!pgtable_pmd_page_ctor(page)) {
__free_pages(page, 0);
return NULL;
}
return (pmd_t *)page_address(page);
}
In our case it is a GFP_PGTABLE_KERNEL with __GFP_ZERO and result is
poisoned page instead of zeroed page.
If I interpret the above correctly it means that something (probably
memalloc_free_pages() ?) puts initial free pages as poisoned and later
alloc_pages() assumes they are memset()-zero. But I don't see why.
> > The change drops the assumption that init_on_free=1 guarantees free
> > pages to contain zeros.
>
> The change assumes that page_poison=on also leaves want_init_on_free() enabled,
> but it doesn't.
>
> > Alternative would be to make interaction between runtime poisoning and
> > sanitizing options and build-time debug flags like CONFIG_PAGE_POISONING
> > more coherent. I took the simpler path.
>
> So that was done in 5.11 and the decisions can be seen in
> init_mem_debugging_and_hardening(). There might be of course a bug, or later
> changes broke something. Which was the version that you observed a bug?
>
> > Tested the fix on rx3600.
> >
> > CC: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> > CC: linux-mm@xxxxxxxxx
> > Signed-off-by: Sergei Trofimovich <slyfox@xxxxxxxxxx>
> > ---
> > mm/page_alloc.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index cfc72873961d..d57d9b4f7089 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -2301,7 +2301,7 @@ inline void post_alloc_hook(struct page *page, unsigned int order,
> > kernel_unpoison_pages(page, 1 << order);
> > set_page_owner(page, order, gfp_flags);
> >
> > - if (!want_init_on_free() && want_init_on_alloc(gfp_flags))
> > + if (want_init_on_alloc(gfp_flags))
> > kernel_init_free_pages(page, 1 << order);
> > }
> >
> >
>
--
Sergei