free_pardevice() is called by parport_unregister_device() and already frees
pp->pdev->name, don't try to do it again.
This bug causes kernel crashes.
I found and verified this with KASAN and some added pr_emerg()s:
[ 60.316568] pp_release: pp->pdev->name == ffff88039cb264c0
[ 60.316692] free_pardevice: freeing par_dev->name at ffff88039cb264c0
[ 60.316706] pp_release: kfree(ffff88039cb264c0)
[ 60.316714] ==========================================================
[ 60.316722] BUG: Double free or freeing an invalid pointer
[ 60.316731] Unexpected shadow byte: 0xFB
[ 60.316801] Object at ffff88039cb264c0, in cache kmalloc-32 size: 32
[ 60.316813] Allocated:
[ 60.316824] PID = 1695
[ 60.316869] Freed:
[ 60.316880] PID = 1695
[ 60.316935] ==========================================================
CCing Andy Lutomirski because I think this is what broke vmapped stacks
for me - after applying this patch, vmapped stacks worked for me.
Previously, I got oopses (and lockups) caused by area->pages[0] being
0x400000000 in __vunmap(), with area->pages being allocated in the kmalloc
area.
Signed-off-by: Jann Horn <[email protected]>
---
drivers/char/ppdev.c | 3 ---
1 file changed, 3 deletions(-)
diff --git a/drivers/char/ppdev.c b/drivers/char/ppdev.c
index d23368874710..6af1ce04b3da 100644
--- a/drivers/char/ppdev.c
+++ b/drivers/char/ppdev.c
@@ -748,10 +748,7 @@ static int pp_release(struct inode *inode, struct file *file)
}
if (pp->pdev) {
- const char *name = pp->pdev->name;
-
parport_unregister_device(pp->pdev);
- kfree(name);
pp->pdev = NULL;
pr_debug(CHRDEV "%x: unregistered pardevice\n", minor);
}
--
2.1.4
On Sun, Oct 30, 2016 at 8:44 AM, Jann Horn <[email protected]> wrote:
> free_pardevice() is called by parport_unregister_device() and already frees
> pp->pdev->name, don't try to do it again.
>
> This bug causes kernel crashes.
>
> I found and verified this with KASAN and some added pr_emerg()s:
>
> [ 60.316568] pp_release: pp->pdev->name == ffff88039cb264c0
> [ 60.316692] free_pardevice: freeing par_dev->name at ffff88039cb264c0
> [ 60.316706] pp_release: kfree(ffff88039cb264c0)
> [ 60.316714] ==========================================================
> [ 60.316722] BUG: Double free or freeing an invalid pointer
> [ 60.316731] Unexpected shadow byte: 0xFB
> [ 60.316801] Object at ffff88039cb264c0, in cache kmalloc-32 size: 32
> [ 60.316813] Allocated:
> [ 60.316824] PID = 1695
> [ 60.316869] Freed:
> [ 60.316880] PID = 1695
> [ 60.316935] ==========================================================
>
> CCing Andy Lutomirski because I think this is what broke vmapped stacks
> for me - after applying this patch, vmapped stacks worked for me.
> Previously, I got oopses (and lockups) caused by area->pages[0] being
> 0x400000000 in __vunmap(), with area->pages being allocated in the kmalloc
> area.
That's an odd symptom. I assume that what's happening is that the
pages array is being freed early by the extra kfree in here and then
getting corrupted.
--Andy
On Sun, Oct 30, 2016 at 09:29:10AM -0700, Andy Lutomirski wrote:
> On Sun, Oct 30, 2016 at 8:44 AM, Jann Horn <[email protected]> wrote:
> > free_pardevice() is called by parport_unregister_device() and already frees
> > pp->pdev->name, don't try to do it again.
> >
> > This bug causes kernel crashes.
> >
> > I found and verified this with KASAN and some added pr_emerg()s:
> >
> > [ 60.316568] pp_release: pp->pdev->name == ffff88039cb264c0
> > [ 60.316692] free_pardevice: freeing par_dev->name at ffff88039cb264c0
> > [ 60.316706] pp_release: kfree(ffff88039cb264c0)
> > [ 60.316714] ==========================================================
> > [ 60.316722] BUG: Double free or freeing an invalid pointer
> > [ 60.316731] Unexpected shadow byte: 0xFB
> > [ 60.316801] Object at ffff88039cb264c0, in cache kmalloc-32 size: 32
> > [ 60.316813] Allocated:
> > [ 60.316824] PID = 1695
> > [ 60.316869] Freed:
> > [ 60.316880] PID = 1695
> > [ 60.316935] ==========================================================
> >
> > CCing Andy Lutomirski because I think this is what broke vmapped stacks
> > for me - after applying this patch, vmapped stacks worked for me.
> > Previously, I got oopses (and lockups) caused by area->pages[0] being
> > 0x400000000 in __vunmap(), with area->pages being allocated in the kmalloc
> > area.
>
> That's an odd symptom. I assume that what's happening is that the
> pages array is being freed early by the extra kfree in here and then
> getting corrupted.
Well, as far as I can tell, there are two ways to reach that.
Obvious first way, but a pretty tight race:
Task A: free(name)
Task B: allocate area->pages in same place
Task A: second free(name), releasing area->pages
Second way (if the SLAB allocator, which I'm using, is used):
Task A: free(name), appends the object to ac->entry in ___cache_free()
Task A: second free(name), appends the object to ac->entry again
Task B: ____cache_alloc() returns object from array cache
Task C: ____cache_alloc() returns same object again
So then the same memory would be used by two separate objects?
On Sunday 30 October 2016 09:14 PM, Jann Horn wrote:
> free_pardevice() is called by parport_unregister_device() and already frees
> pp->pdev->name, don't try to do it again.
>
> This bug causes kernel crashes.
>
> I found and verified this with KASAN and some added pr_emerg()s:
>
> [ 60.316568] pp_release: pp->pdev->name == ffff88039cb264c0
> [ 60.316692] free_pardevice: freeing par_dev->name at ffff88039cb264c0
> [ 60.316706] pp_release: kfree(ffff88039cb264c0)
> [ 60.316714] ==========================================================
> [ 60.316722] BUG: Double free or freeing an invalid pointer
> [ 60.316731] Unexpected shadow byte: 0xFB
> [ 60.316801] Object at ffff88039cb264c0, in cache kmalloc-32 size: 32
> [ 60.316813] Allocated:
> [ 60.316824] PID = 1695
> [ 60.316869] Freed:
> [ 60.316880] PID = 1695
> [ 60.316935] ==========================================================
>
> CCing Andy Lutomirski because I think this is what broke vmapped stacks
> for me - after applying this patch, vmapped stacks worked for me.
> Previously, I got oopses (and lockups) caused by area->pages[0] being
> 0x400000000 in __vunmap(), with area->pages being allocated in the kmalloc
> area.
I think the above should not be a part of the commit message.
It looks like a valid bug. As we have now started using the device model
from this release, so device_unregister() will finally call
free_pardevice() and that will free the name.
>
> Signed-off-by: Jann Horn <[email protected]>
> ---
Acked-by: Sudip Mukherjee <[email protected]>
> drivers/char/ppdev.c | 3 ---
> 1 file changed, 3 deletions(-)
>
> diff --git a/drivers/char/ppdev.c b/drivers/char/ppdev.c
> index d23368874710..6af1ce04b3da 100644
> --- a/drivers/char/ppdev.c
> +++ b/drivers/char/ppdev.c
> @@ -748,10 +748,7 @@ static int pp_release(struct inode *inode, struct file *file)
> }
>
> if (pp->pdev) {
> - const char *name = pp->pdev->name;
> -
> parport_unregister_device(pp->pdev);
> - kfree(name);
> pp->pdev = NULL;
> pr_debug(CHRDEV "%x: unregistered pardevice\n", minor);
> }
>
On Sun, Oct 30, 2016 at 11:24:27PM +0530, Sudip Mukherjee wrote:
> On Sunday 30 October 2016 09:14 PM, Jann Horn wrote:
> >free_pardevice() is called by parport_unregister_device() and already frees
> >pp->pdev->name, don't try to do it again.
> >
> >This bug causes kernel crashes.
> >
> >I found and verified this with KASAN and some added pr_emerg()s:
> >
> >[ 60.316568] pp_release: pp->pdev->name == ffff88039cb264c0
> >[ 60.316692] free_pardevice: freeing par_dev->name at ffff88039cb264c0
> >[ 60.316706] pp_release: kfree(ffff88039cb264c0)
> >[ 60.316714] ==========================================================
> >[ 60.316722] BUG: Double free or freeing an invalid pointer
> >[ 60.316731] Unexpected shadow byte: 0xFB
> >[ 60.316801] Object at ffff88039cb264c0, in cache kmalloc-32 size: 32
> >[ 60.316813] Allocated:
> >[ 60.316824] PID = 1695
> >[ 60.316869] Freed:
> >[ 60.316880] PID = 1695
> >[ 60.316935] ==========================================================
> >
> >CCing Andy Lutomirski because I think this is what broke vmapped stacks
> >for me - after applying this patch, vmapped stacks worked for me.
> >Previously, I got oopses (and lockups) caused by area->pages[0] being
> >0x400000000 in __vunmap(), with area->pages being allocated in the kmalloc
> >area.
>
> I think the above should not be a part of the commit message.
Ah, yes.
@maintainers: Feel free to remove that from the commit message. Or should I
resend?
On Sun, Oct 30, 2016 at 07:00:36PM +0100, Jann Horn wrote:
> On Sun, Oct 30, 2016 at 11:24:27PM +0530, Sudip Mukherjee wrote:
> > On Sunday 30 October 2016 09:14 PM, Jann Horn wrote:
> > >free_pardevice() is called by parport_unregister_device() and already frees
> > >pp->pdev->name, don't try to do it again.
> > >
> > >This bug causes kernel crashes.
> > >
> > >I found and verified this with KASAN and some added pr_emerg()s:
> > >
> > >[ 60.316568] pp_release: pp->pdev->name == ffff88039cb264c0
> > >[ 60.316692] free_pardevice: freeing par_dev->name at ffff88039cb264c0
> > >[ 60.316706] pp_release: kfree(ffff88039cb264c0)
> > >[ 60.316714] ==========================================================
> > >[ 60.316722] BUG: Double free or freeing an invalid pointer
> > >[ 60.316731] Unexpected shadow byte: 0xFB
> > >[ 60.316801] Object at ffff88039cb264c0, in cache kmalloc-32 size: 32
> > >[ 60.316813] Allocated:
> > >[ 60.316824] PID = 1695
> > >[ 60.316869] Freed:
> > >[ 60.316880] PID = 1695
> > >[ 60.316935] ==========================================================
> > >
> > >CCing Andy Lutomirski because I think this is what broke vmapped stacks
> > >for me - after applying this patch, vmapped stacks worked for me.
> > >Previously, I got oopses (and lockups) caused by area->pages[0] being
> > >0x400000000 in __vunmap(), with area->pages being allocated in the kmalloc
> > >area.
> >
> > I think the above should not be a part of the commit message.
>
> Ah, yes.
>
> @maintainers: Feel free to remove that from the commit message. Or should I
> resend?
Please resend, never make a maintainer hand-edit a patch changelog :)
thanks,
greg k-h