Re: [PATCH] mm: kmemleak: Skip check in kmemleak_*_phys when pfn bound is not ready

[Catalin Marinas <catalin.marinas@arm.com>]

Hi Patrick,

On Mon, May 30, 2022 at 09:32:18PM +0800, Patrick Wang wrote:
> On 2022/5/30 10:27, Yee Lee wrote:
> > On Fri, 2022-05-27 at 21:39 +0800, patrick wang wrote:
> > > On Fri, May 27, 2022 at 11:25 AM <yee.lee@mediatek.com> wrote:
> > > > From: Yee Lee <yee.lee@mediatek.com>
> > > > 
> > > > In some archs (arm64), memblock allocates memory in boot time when
> > > > the pfn boundary (max_pfn/min_pfn) is not ready. The lowmen checks in
> > > > kmemleak_*_phys() drop those blocks and cause some false leak alarms
> > > > on common kernel objects.
> > > > 
> > > > Kmemleak output: (Qemu/arm64)
> > > > unreferenced object 0xffff0000c0170a00 (size 128):
> > > >    comm "swapper/0", pid 1, jiffies 4294892404 (age 126.208s)
> > > >    hex dump (first 32 bytes):
> > > >      62 61 73 65 00 00 00 00 00 00 00 00 00 00 00 00  base............
> > > >      00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
> > > >    backtrace:
> > > >      [<(____ptrval____)>] __kmalloc_track_caller+0x1b0/0x2e4
> > > >      [<(____ptrval____)>] kstrdup_const+0x8c/0xc4
> > > >      [<(____ptrval____)>] kvasprintf_const+0xbc/0xec
> > > >      [<(____ptrval____)>] kobject_set_name_vargs+0x58/0xe4
> > > >      [<(____ptrval____)>] kobject_add+0x84/0x100
> > > >      [<(____ptrval____)>] __of_attach_node_sysfs+0x78/0xec
> > > >      [<(____ptrval____)>] of_core_init+0x68/0x104
> > > >      [<(____ptrval____)>] driver_init+0x28/0x48
> > > >      [<(____ptrval____)>] do_basic_setup+0x14/0x28
> > > >      [<(____ptrval____)>] kernel_init_freeable+0x110/0x178
> > > >      [<(____ptrval____)>] kernel_init+0x20/0x1a0
> > > >      [<(____ptrval____)>] ret_from_fork+0x10/0x20
> > > > 
> > > > This patch relaxs the boundary checking in kmemleak_*_phys api
> > > > if max_low_pfn is uninitialzed.
> > > > 
> > > > Fixes: 23c2d4 (mm: kmemleak: take a full lowmem check in kmemleak_*_phy)

BTW, please use at least 12 characters for the git sha1, the above is
ambiguous.

> > > > Signed-off-by: Yee Lee <yee.lee@mediatek.com>
> > > > ---
> > > >   mm/kmemleak.c | 8 ++++----
> > > >   1 file changed, 4 insertions(+), 4 deletions(-)
> > > > 
> > > > diff --git a/mm/kmemleak.c b/mm/kmemleak.c
> > > > index a182f5ddaf68..6b2af544aa0f 100644
> > > > --- a/mm/kmemleak.c
> > > > +++ b/mm/kmemleak.c
> > > > @@ -1132,7 +1132,7 @@ EXPORT_SYMBOL(kmemleak_no_scan);
> > > >   void __ref kmemleak_alloc_phys(phys_addr_t phys, size_t size, int min_count,
> > > >                                 gfp_t gfp)
> > > >   {
> > > > -       if (PHYS_PFN(phys) >= min_low_pfn && PHYS_PFN(phys) < max_low_pfn)
> > > > +       if (!max_low_pfn || (PHYS_PFN(phys) >= min_low_pfn && PHYS_PFN(phys) < max_low_pfn))
> > > 
> > > Just skip checking will bring the crash possibility back. Seems
> > > it's beyond these interfaces' handle scope for this situation,
> > > since "min_low_pfn" and "max_low_pfn" are depending on arches.
> > 
> > Yes, for the cases beyond the pfn guard, users have to take care the
> > boundary by themselves.
> 
> Could we record these early calling and deal with them when it's
> ready? Is this appropriate?
> 
> I have an implementation based on this approach. Could you please
> help to have a test on your machine as well? And someone to take
> a look or review?

We had something similar until 5.4, removed by commit c5665868183f ("mm:
kmemleak: use the memory pool for early allocations"). It was a bit
painful as we never had the right buffer, so I'm not keen on adding it
back.

> From 82cae75dfaa78f414faf85bb49133e95159c041a Mon Sep 17 00:00:00 2001
> From: Patrick Wang <patrick.wang.shcn@gmail.com>
> Date: Mon, 30 May 2022 18:38:23 +0800
> Subject: [PATCH] mm: kmemleak: record early operations and handle later
> 
> The kmemleak_*_phys() interface uses "min_low_pfn" and
> "max_low_pfn" to check address. But on some architectures,
> kmemleak_*_phys() is called before those two variables
> initialized. Record these early operations and handle them
> when kmemleak_*_phys() are ready.

Could we instead record everything (no checks) but later avoid scanning
if below min or above max_low_pfn? We can add an OBJECT_PHYS flag to all
objects allocated via kmemleak_*_phys() and always check the
virt_to_phys() boundaries at scan time. It may actually help with this
problem as well:

https://lore.kernel.org/r/9dd08bb5-f39e-53d8-f88d-bec598a08c93@gmail.com

-- 
Catalin