Re: [PATCH v3] mm/kfence: reset PG_slab and memcg_data before freeing __kfence_pool

From: Hyeonggon Yoo
Date: Thu May 05 2022 - 07:33:46 EST


On Thu, May 05, 2022 at 06:54:18PM +0800, Muchun Song wrote:
> On Thu, May 05, 2022 at 07:13:37PM +0900, Hyeonggon Yoo wrote:
> > When kfence fails to initialize kfence pool, it frees the pool.
> > But it does not reset PG_slab flag and memcg_data of struct page.
> >
> > Below is a BUG because of this. Let's fix it by resetting PG_slab
> > and memcg_data before free.
> >
> > [ 0.089149] BUG: Bad page state in process swapper/0 pfn:3d8e06
> > [ 0.089149] page:ffffea46cf638180 refcount:0 mapcount:0 mapping:0000000000000000 index:0x0 pfn:0x3d8e06
> > [ 0.089150] memcg:ffffffff94a475d1
> > [ 0.089150] flags: 0x17ffffc0000200(slab|node=0|zone=2|lastcpupid=0x1fffff)
> > [ 0.089151] raw: 0017ffffc0000200 ffffea46cf638188 ffffea46cf638188 0000000000000000
> > [ 0.089152] raw: 0000000000000000 0000000000000000 00000000ffffffff ffffffff94a475d1
> > [ 0.089152] page dumped because: page still charged to cgroup
> > [ 0.089153] Modules linked in:
> > [ 0.089153] CPU: 0 PID: 0 Comm: swapper/0 Tainted: G B W 5.18.0-rc1+ #965
> > [ 0.089154] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-2 04/01/2014
> > [ 0.089154] Call Trace:
> > [ 0.089155] <TASK>
> > [ 0.089155] dump_stack_lvl+0x49/0x5f
> > [ 0.089157] dump_stack+0x10/0x12
> > [ 0.089158] bad_page.cold+0x63/0x94
> > [ 0.089159] check_free_page_bad+0x66/0x70
> > [ 0.089160] __free_pages_ok+0x423/0x530
> > [ 0.089161] __free_pages_core+0x8e/0xa0
> > [ 0.089162] memblock_free_pages+0x10/0x12
> > [ 0.089164] memblock_free_late+0x8f/0xb9
> > [ 0.089165] kfence_init+0x68/0x92
> > [ 0.089166] start_kernel+0x789/0x992
> > [ 0.089167] x86_64_start_reservations+0x24/0x26
> > [ 0.089168] x86_64_start_kernel+0xa9/0xaf
> > [ 0.089170] secondary_startup_64_no_verify+0xd5/0xdb
> > [ 0.089171] </TASK>
> >
> > Fixes: 0ce20dd84089 ("mm: add Kernel Electric-Fence infrastructure")
> > Fixes: 8f0b36497303 ("mm: kfence: fix objcgs vector allocation")
> > Signed-off-by: Hyeonggon Yoo <42.hyeyoo@xxxxxxxxx>
> > Reviewed-by: Marco Elver <elver@xxxxxxxxxx>
> > Reviewed-by: Muchun Song <songmuchun@xxxxxxxxxxxxx>
> > ---
> >
> > v2 -> v3:
> > - Add Reviewed-by: tags from Marco and Muchun. Thanks!
> > - Initialize folio where it is defined.
> >
> > mm/kfence/core.c | 8 ++++++++
> > 1 file changed, 8 insertions(+)
> >
> > diff --git a/mm/kfence/core.c b/mm/kfence/core.c
> > index a203747ad2c0..b7d3a9667f00 100644
> > --- a/mm/kfence/core.c
> > +++ b/mm/kfence/core.c
> > @@ -642,6 +642,14 @@ static bool __init kfence_init_pool_early(void)
> > * fails for the first page, and therefore expect addr==__kfence_pool in
> > * most failure cases.
> > */
> > + for (char *p = (char *)addr; p < __kfence_pool + KFENCE_POOL_SIZE; p += PAGE_SIZE) {
> > + struct folio *folio = virt_to_folio(p);
> > +
>
> After more thinking, I think it is better to use 'struct slab *'
> to define a local variable since we already use this struct
> throughout slab core. What do you think?
>

I think that may not be better.

In the code we're freeing folios (so not going to reuse it again in slab/kfence).
And it may not be Slab depending on why kfence_init_pool() failed.

> Thanks.
>
> > + __folio_clear_slab(folio);
> > +#ifdef CONFIG_MEMCG
> > + folio->memcg_data = 0;
> > +#endif
> > + }
> > memblock_free_late(__pa(addr), KFENCE_POOL_SIZE - (addr - (unsigned long)__kfence_pool));
> > __kfence_pool = NULL;
> > return false;
> > --
> > 2.32.0
> >
> >

--
Thanks,
Hyeonggon