Re: [PATCH] remoteproc: qcom: Don't memcpy_toio more than is provided

From: Bjorn Andersson <bjorn.andersson@linaro.org>
To: Doug Anderson <dianders@chromium.org>
Cc: Stephen Boyd <swboyd@chromium.org>,
	linux-kernel@vger.kernel.org, linux-arm-msm@vger.kernel.org
Subject: Re: [PATCH] remoteproc: qcom: Don't memcpy_toio more than is provided
Date: Mon, 6 Dec 2021 16:24:28 -0800	[thread overview]
Message-ID: <Ya6pvF8VIwokzqjs@ripper> (raw)
In-Reply-To: <CAD=FV=XgV5=5mMuB5KbX38EFtZ6qxGGPMN6HFng_KJEc1KERxw@mail.gmail.com>

On Mon 06 Dec 16:01 PST 2021, Doug Anderson wrote:

> Hi,
> 
> On Tue, Nov 16, 2021 at 10:55 PM Stephen Boyd <swboyd@chromium.org> wrote:
> >
> > If the string passed into qcom_pil_info_store() isn't as long as
> > PIL_RELOC_NAME_LEN we'll try to copy the string assuming the length is
> > PIL_RELOC_NAME_LEN to the io space and go beyond the bounds of the
> > string. Let's only copy as many byes as the string is long, ignoring the
> > NUL terminator.
> >
> > This fixes the following KASAN error:
> >
> >  BUG: KASAN: global-out-of-bounds in __memcpy_toio+0x124/0x140
> >  Read of size 1 at addr ffffffd35086e386 by task rmtfs/2392
> >
> >  CPU: 2 PID: 2392 Comm: rmtfs Tainted: G        W         5.16.0-rc1-lockdep+ #10
> >  Hardware name: Google Lazor (rev3+) with KB Backlight (DT)
> >  Call trace:
> >   dump_backtrace+0x0/0x410
> >   show_stack+0x24/0x30
> >   dump_stack_lvl+0x7c/0xa0
> >   print_address_description+0x78/0x2bc
> >   kasan_report+0x160/0x1a0
> >   __asan_report_load1_noabort+0x44/0x50
> >   __memcpy_toio+0x124/0x140
> >   qcom_pil_info_store+0x298/0x358 [qcom_pil_info]
> >   q6v5_start+0xdf0/0x12e0 [qcom_q6v5_mss]
> >   rproc_start+0x178/0x3a0
> >   rproc_boot+0x5f0/0xb90
> >   state_store+0x78/0x1bc
> >   dev_attr_store+0x70/0x90
> >   sysfs_kf_write+0xf4/0x118
> >   kernfs_fop_write_iter+0x208/0x300
> >   vfs_write+0x55c/0x804
> >   ksys_pwrite64+0xc8/0x134
> >   __arm64_compat_sys_aarch32_pwrite64+0xc4/0xdc
> >   invoke_syscall+0x78/0x20c
> >   el0_svc_common+0x11c/0x1f0
> >   do_el0_svc_compat+0x50/0x60
> >   el0_svc_compat+0x5c/0xec
> >   el0t_32_sync_handler+0xc0/0xf0
> >   el0t_32_sync+0x1a4/0x1a8
> >
> >  The buggy address belongs to the variable:
> >   .str.59+0x6/0xffffffffffffec80 [qcom_q6v5_mss]
> >
> >  Memory state around the buggy address:
> >   ffffffd35086e280: 00 00 00 00 02 f9 f9 f9 f9 f9 f9 f9 00 00 00 00
> >   ffffffd35086e300: 00 02 f9 f9 f9 f9 f9 f9 00 00 00 06 f9 f9 f9 f9
> >  >ffffffd35086e380: 06 f9 f9 f9 05 f9 f9 f9 00 00 00 00 00 06 f9 f9
> >                     ^
> >   ffffffd35086e400: f9 f9 f9 f9 01 f9 f9 f9 04 f9 f9 f9 00 00 01 f9
> >   ffffffd35086e480: f9 f9 f9 f9 00 00 00 00 00 00 00 01 f9 f9 f9 f9
> >
> > Fixes: 549b67da660d ("remoteproc: qcom: Introduce helper to store pil info in IMEM")
> > Signed-off-by: Stephen Boyd <swboyd@chromium.org>
> > ---
> >  drivers/remoteproc/qcom_pil_info.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/drivers/remoteproc/qcom_pil_info.c b/drivers/remoteproc/qcom_pil_info.c
> > index 7c007dd7b200..aca21560e20b 100644
> > --- a/drivers/remoteproc/qcom_pil_info.c
> > +++ b/drivers/remoteproc/qcom_pil_info.c
> > @@ -104,7 +104,7 @@ int qcom_pil_info_store(const char *image, phys_addr_t base, size_t size)
> >         return -ENOMEM;
> >
> >  found_unused:
> > -       memcpy_toio(entry, image, PIL_RELOC_NAME_LEN);
> > +       memcpy_toio(entry, image, strnlen(image, PIL_RELOC_NAME_LEN));
> 
> The above seems slightly sketchy...
> 
> Let's say:
> 
> image = "modem" (5 characters plus null termination)
> PIL_RELOC_NAME_LEN = 8
> 
> ...so strnlen(image, 8) = 5, right?
> ...so we'll copy characters _not_ including the NULL termination.
> 
> I guess that's OK as long as we're certain that the destination was
> zero-initted, but maybe it would be better/safer to write:
> 
> memcpy_toio(entry, image, strnlen(image, PIL_RELOC_NAME_LEN - 1) + 1);
> 

Yes, I agree that it looks a little bit sketchy.

But we have to assume that either there's remnants from the boot stages
or perhaps remnants of stale data from before a reboot. Therefor I
concluded as I wrote this that I had to memset() the entire region in
qcom_pil_info_init(), so this is taken care of.

The region being zero-initialized is also a requirement for the
conditional jump to "found_unused" based on !buf[0]...

Regards,
Bjorn