From: Eric Biggers <[email protected]>
WRITE_ONCE() isn't the correct way to publish a pointer to a data
structure, since it doesn't include a write memory barrier. Therefore
other tasks may see that the pointer has been set but not see that the
pointed-to memory has finished being initialized yet. Instead a
primitive with "release" semantics is needed.
Use smp_store_release() for this.
The use of READ_ONCE() on the read side is still potentially correct if
there's no control dependency, i.e. if all memory being "published" is
transitively reachable via the pointer itself. But this pairing is
somewhat confusing and error-prone. So just upgrade the read side to
smp_load_acquire() so that it clearly pairs with smp_store_release().
Cc: Dan Williams <[email protected]>
Cc: Arnd Bergmann <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Kees Cook <[email protected]>
Cc: Matthew Wilcox <[email protected]>
Cc: Russell King <[email protected]>
Cc: Andrew Morton <[email protected]>
Cc: Greg Kroah-Hartman <[email protected]>
Fixes: 3234ac664a87 ("/dev/mem: Revoke mappings when a driver claims the region")
Signed-off-by: Eric Biggers <[email protected]>
---
drivers/char/mem.c | 10 +++++++---
1 file changed, 7 insertions(+), 3 deletions(-)
diff --git a/drivers/char/mem.c b/drivers/char/mem.c
index 934c92dcb9ab..687d4af6945d 100644
--- a/drivers/char/mem.c
+++ b/drivers/char/mem.c
@@ -814,7 +814,8 @@ static struct inode *devmem_inode;
#ifdef CONFIG_IO_STRICT_DEVMEM
void revoke_devmem(struct resource *res)
{
- struct inode *inode = READ_ONCE(devmem_inode);
+ /* pairs with smp_store_release() in devmem_init_inode() */
+ struct inode *inode = smp_load_acquire(&devmem_inode);
/*
* Check that the initialization has completed. Losing the race
@@ -1028,8 +1029,11 @@ static int devmem_init_inode(void)
return rc;
}
- /* publish /dev/mem initialized */
- WRITE_ONCE(devmem_inode, inode);
+ /*
+ * Publish /dev/mem initialized.
+ * Pairs with smp_load_acquire() in revoke_devmem().
+ */
+ smp_store_release(&devmem_inode, inode);
return 0;
}
base-commit: f8456690ba8eb18ea4714e68554e242a04f65cff
--
2.27.0
On Wed, Jul 15, 2020 at 11:07 PM Eric Biggers <[email protected]> wrote:
>
> From: Eric Biggers <[email protected]>
>
> WRITE_ONCE() isn't the correct way to publish a pointer to a data
> structure, since it doesn't include a write memory barrier. Therefore
> other tasks may see that the pointer has been set but not see that the
> pointed-to memory has finished being initialized yet. Instead a
> primitive with "release" semantics is needed.
>
> Use smp_store_release() for this.
>
> The use of READ_ONCE() on the read side is still potentially correct if
> there's no control dependency, i.e. if all memory being "published" is
> transitively reachable via the pointer itself. But this pairing is
> somewhat confusing and error-prone. So just upgrade the read side to
> smp_load_acquire() so that it clearly pairs with smp_store_release().
>
> Cc: Dan Williams <[email protected]>
> Cc: Arnd Bergmann <[email protected]>
> Cc: Ingo Molnar <[email protected]>
> Cc: Kees Cook <[email protected]>
> Cc: Matthew Wilcox <[email protected]>
> Cc: Russell King <[email protected]>
> Cc: Andrew Morton <[email protected]>
> Cc: Greg Kroah-Hartman <[email protected]>
> Fixes: 3234ac664a87 ("/dev/mem: Revoke mappings when a driver claims the region")
> Signed-off-by: Eric Biggers <[email protected]>
Makes sense:
Acked-by: Dan Williams <[email protected]>