Re: [PATCH] media: vsp1: dl: Fix NULL pointer dereference on unbind
Hi Kieran,
On Mon, May 25, 2020 at 02:19:02PM +0100, Kieran Bingham wrote:
> Hi Eugeniu,
>
> Yeouch. Looks like I really missed a trick there!
Not a big deal. The good part is that it can be proactively fixed and
shared across the community.
>
> We should probably update the $SUBJECT to match what is performed in the
> patch, which is perhaps more like:
>
> "media: vsp1: dl: Store VSP reference when creating cmd pools"
To be honest, I am not a big fan of WHAT summary lines.
Rather, I prefer the WHY summary lines (and I think everyone should).
>
> On 23/05/2020 09:13, Eugeniu Rosca wrote:
>
> And then we can explain here:
>
> In commit f3b98e3c4d2e16 ("media: vsp1: Provide support for extended
> command pools"), the vsp pointer used for referencing the VSP1 device
> structure from a command pool during vsp1_dl_ext_cmd_pool_destroy() was
> not populated.
>
> Correctly assign the pointer to prevent the following
> null-pointer-dereference when removing the device:
That sounds good and I can push this improved description as v2.
> > Fixes: f3b98e3c4d2e16 ("media: vsp1: Provide support for extended command
> > pools")
> > Cc: sta...@vger.kernel.org # v4.19+
> > Signed-off-by: Eugeniu Rosca
>
> Reviewed-by: Kieran Bingham
>
> > ---
> >
> > How about adding a new unit test perfoming unbind/rebind to
> > http://git.ideasonboard.com/renesas/vsp-tests.git, to avoid
> > such issues in future?
>
> Yes, now I wish I had done so back at 4.19! I hope this wasn't too
> painful to diagnose and fix, and thank you for being so thorough in your
> report!
>
>
> > Locally, below command has been used to identify the problem:
> >
> > for f in $(find /sys/bus/platform/devices/ -name "*vsp*" -o -name "*fdp*");
> > do \
> > b=$(basename $f); \
> > echo $b > $f/driver/unbind; \
> > done
> >
>
> I've created a test to add to vsp-tests, which I'll post next, thank you
> for the suggestion.
>
> Before your patch is applied, I experience the same crash you have seen,
> and after your patch - I can successfully unbind/bind all of the VSP1
> instances.
>
> So I think you can have this too:
>
> Tested-by: Kieran Bingham
Awesome. Thanks!
--
Best regards,
Eugeniu Rosca
Re: [PATCH] media: vsp1: dl: Fix NULL pointer dereference on unbind
Hi Eugeniu,
Yeouch. Looks like I really missed a trick there!
We should probably update the $SUBJECT to match what is performed in the
patch, which is perhaps more like:
"media: vsp1: dl: Store VSP reference when creating cmd pools"
On 23/05/2020 09:13, Eugeniu Rosca wrote:
And then we can explain here:
In commit f3b98e3c4d2e16 ("media: vsp1: Provide support for extended
command pools"), the vsp pointer used for referencing the VSP1 device
structure from a command pool during vsp1_dl_ext_cmd_pool_destroy() was
not populated.
Correctly assign the pointer to prevent the following
null-pointer-dereference when removing the device:
> v4.19 commit f3b98e3c4d2e16 ("media: vsp1: Provide support for extended
> command pools") introduced below issue [*], consistently reproduced.
>
> In order to fix it, inspire from the sibling/predecessor v4.18-rc1
> commit 5de0473982aab2 ("media: vsp1: Provide a body pool"), which saves
> the vsp1 instance address in vsp1_dl_body_pool_create().
>
> [*] h3ulcb-kf #>
> echo fea28000.vsp > /sys/bus/platform/devices/fea28000.vsp/driver/unbind
> Unable to handle kernel NULL pointer dereference at virtual address
> 0028
> Mem abort info:
>ESR = 0x9606
>EC = 0x25: DABT (current EL), IL = 32 bits
>SET = 0, FnV = 0
>EA = 0, S1PTW = 0
> Data abort info:
>ISV = 0, ISS = 0x0006
>CM = 0, WnR = 0
> user pgtable: 4k pages, 48-bit VAs, pgdp=0007318be000
> [0028] pgd=0007333a1003, pud=0007333a6003,
> pmd=
> Internal error: Oops: 9606 [#1] PREEMPT SMP
> Modules linked in:
> CPU: 1 PID: 486 Comm: sh Not tainted
> 5.7.0-rc6-arm64-renesas-00118-ge644645abf47 #185
> Hardware name: Renesas H3ULCB Kingfisher board based on r8a77951 (DT)
> pstate: 4005 (nZcv daif -PAN -UAO)
> pc : vsp1_dlm_destroy+0xe4/0x11c
> lr : vsp1_dlm_destroy+0xc8/0x11c
> sp : 800012963b60
> x29: 800012963b60 x28: 0006f83fc440
> x27: x26: 0006f5e13e80
> x25: 0006f5e13ed0 x24: 0006f5e13ed0
> x23: 0006f5e13ed0 x22: dead0122
> x21: 0006f5e3a080 x20: 0006f5df2938
> x19: 0006f5df2980 x18: 0003
> x17: x16: 0016
> x15: 0003 x14: 000393c0
> x13: 800011a5ec18 x12: 800011d8d000
> x11: 0006f83fcc68 x10: 800011a53d70
> x9 : 8000111f3000 x8 :
> x7 : 00210d00 x6 :
> x5 : 800010872e60 x4 : 0004
> x3 : 78068000 x2 : 800012781000
> x1 : 2c00 x0 :
> Call trace:
> vsp1_dlm_destroy+0xe4/0x11c
> vsp1_wpf_destroy+0x10/0x20
> vsp1_entity_destroy+0x24/0x4c
> vsp1_destroy_entities+0x54/0x130
> vsp1_remove+0x1c/0x40
> platform_drv_remove+0x28/0x50
> __device_release_driver+0x178/0x220
> device_driver_detach+0x44/0xc0
> unbind_store+0xe0/0x104
> drv_attr_store+0x20/0x30
> sysfs_kf_write+0x48/0x70
> kernfs_fop_write+0x148/0x230
> __vfs_write+0x18/0x40
> vfs_write+0xdc/0x1c4
> ksys_write+0x68/0xf0
> __arm64_sys_write+0x18/0x20
> el0_svc_common.constprop.0+0x70/0x170
> do_el0_svc+0x20/0x80
> el0_sync_handler+0x134/0x1b0
> el0_sync+0x140/0x180
> Code: b4c2 f9403a60 d2800084 a9400663 (f9401400)
> ---[ end trace 3875369841fb288a ]---
>
> Fixes: f3b98e3c4d2e16 ("media: vsp1: Provide support for extended command
> pools")
> Cc: sta...@vger.kernel.org # v4.19+
> Signed-off-by: Eugeniu Rosca
Reviewed-by: Kieran Bingham
> ---
>
> How about adding a new unit test perfoming unbind/rebind to
> http://git.ideasonboard.com/renesas/vsp-tests.git, to avoid
> such issues in future?
Yes, now I wish I had done so back at 4.19! I hope this wasn't too
painful to diagnose and fix, and thank you for being so thorough in your
report!
> Locally, below command has been used to identify the problem:
>
> for f in $(find /sys/bus/platform/devices/ -name "*vsp*" -o -name "*fdp*");
> do \
> b=$(basename $f); \
> echo $b > $f/driver/unbind; \
> done
>
I've created a test to add to vsp-tests, which I'll post next, thank you
for the suggestion.
Before your patch is applied, I experience the same crash you have seen,
and after your patch - I can successfully unbind/bind all of the VSP1
instances.
So I think you can have this too:
Tested-by: Kieran Bingham
> ---
> drivers/media/platform/vsp1/vsp1_dl.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/drivers/media/platform/vsp1/vsp1_dl.c
> b/drivers/media/platform/vsp1/vsp1_dl.c
> index d7b43037e500..e07b135613eb 100644
> --- a/drivers/media/platform/vsp1/vsp1_dl.c
> +++ b/drivers/media/platform/vsp1/vsp1_dl.c
> @@ -431,6 +431,8 @@ vsp1_dl_cmd_pool_create(struct vsp1_device *vsp1, enum
> vsp1_extcmd_type type,
> if (!pool)
> return NULL;
>
> + pool->vsp1 = vsp1;
> +
> spin_lock_init(&pool->lock);
> INIT_LIST_HEAD(&pool->free);
>
>
[PATCH] media: vsp1: dl: Fix NULL pointer dereference on unbind
v4.19 commit f3b98e3c4d2e16 ("media: vsp1: Provide support for extended
command pools") introduced below issue [*], consistently reproduced.
In order to fix it, inspire from the sibling/predecessor v4.18-rc1
commit 5de0473982aab2 ("media: vsp1: Provide a body pool"), which saves
the vsp1 instance address in vsp1_dl_body_pool_create().
[*] h3ulcb-kf #>
echo fea28000.vsp > /sys/bus/platform/devices/fea28000.vsp/driver/unbind
Unable to handle kernel NULL pointer dereference at virtual address
0028
Mem abort info:
ESR = 0x9606
EC = 0x25: DABT (current EL), IL = 32 bits
SET = 0, FnV = 0
EA = 0, S1PTW = 0
Data abort info:
ISV = 0, ISS = 0x0006
CM = 0, WnR = 0
user pgtable: 4k pages, 48-bit VAs, pgdp=0007318be000
[0028] pgd=0007333a1003, pud=0007333a6003,
pmd=
Internal error: Oops: 9606 [#1] PREEMPT SMP
Modules linked in:
CPU: 1 PID: 486 Comm: sh Not tainted
5.7.0-rc6-arm64-renesas-00118-ge644645abf47 #185
Hardware name: Renesas H3ULCB Kingfisher board based on r8a77951 (DT)
pstate: 4005 (nZcv daif -PAN -UAO)
pc : vsp1_dlm_destroy+0xe4/0x11c
lr : vsp1_dlm_destroy+0xc8/0x11c
sp : 800012963b60
x29: 800012963b60 x28: 0006f83fc440
x27: x26: 0006f5e13e80
x25: 0006f5e13ed0 x24: 0006f5e13ed0
x23: 0006f5e13ed0 x22: dead0122
x21: 0006f5e3a080 x20: 0006f5df2938
x19: 0006f5df2980 x18: 0003
x17: x16: 0016
x15: 0003 x14: 000393c0
x13: 800011a5ec18 x12: 800011d8d000
x11: 0006f83fcc68 x10: 800011a53d70
x9 : 8000111f3000 x8 :
x7 : 00210d00 x6 :
x5 : 800010872e60 x4 : 0004
x3 : 78068000 x2 : 800012781000
x1 : 2c00 x0 :
Call trace:
vsp1_dlm_destroy+0xe4/0x11c
vsp1_wpf_destroy+0x10/0x20
vsp1_entity_destroy+0x24/0x4c
vsp1_destroy_entities+0x54/0x130
vsp1_remove+0x1c/0x40
platform_drv_remove+0x28/0x50
__device_release_driver+0x178/0x220
device_driver_detach+0x44/0xc0
unbind_store+0xe0/0x104
drv_attr_store+0x20/0x30
sysfs_kf_write+0x48/0x70
kernfs_fop_write+0x148/0x230
__vfs_write+0x18/0x40
vfs_write+0xdc/0x1c4
ksys_write+0x68/0xf0
__arm64_sys_write+0x18/0x20
el0_svc_common.constprop.0+0x70/0x170
do_el0_svc+0x20/0x80
el0_sync_handler+0x134/0x1b0
el0_sync+0x140/0x180
Code: b4c2 f9403a60 d2800084 a9400663 (f9401400)
---[ end trace 3875369841fb288a ]---
Fixes: f3b98e3c4d2e16 ("media: vsp1: Provide support for extended command
pools")
Cc: sta...@vger.kernel.org # v4.19+
Signed-off-by: Eugeniu Rosca
---
How about adding a new unit test perfoming unbind/rebind to
http://git.ideasonboard.com/renesas/vsp-tests.git, to avoid
such issues in future?
Locally, below command has been used to identify the problem:
for f in $(find /sys/bus/platform/devices/ -name "*vsp*" -o -name "*fdp*"); do \
b=$(basename $f); \
echo $b > $f/driver/unbind; \
done
---
drivers/media/platform/vsp1/vsp1_dl.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/drivers/media/platform/vsp1/vsp1_dl.c
b/drivers/media/platform/vsp1/vsp1_dl.c
index d7b43037e500..e07b135613eb 100644
--- a/drivers/media/platform/vsp1/vsp1_dl.c
+++ b/drivers/media/platform/vsp1/vsp1_dl.c
@@ -431,6 +431,8 @@ vsp1_dl_cmd_pool_create(struct vsp1_device *vsp1, enum
vsp1_extcmd_type type,
if (!pool)
return NULL;
+ pool->vsp1 = vsp1;
+
spin_lock_init(&pool->lock);
INIT_LIST_HEAD(&pool->free);
--
2.26.2
