si->inuse_pages could be accessed concurrently as noticed by KCSAN,
write to 0xffff98b00ebd04dc of 4 bytes by task 82262 on cpu 92:
swap_range_free+0xbe/0x230
swap_range_free at mm/swapfile.c:719
swapcache_free_entries+0x1be/0x250
free_swap_slot+0x1c8/0x220
__swap_entry_free.constprop.19+0xa3/0xb0
free_swap_and_cache+0x53/0xa0
unmap_page_range+0x7e0/0x1ce0
unmap_single_vma+0xcd/0x170
unmap_vmas+0x18b/0x220
exit_mmap+0xee/0x220
mmput+0xe7/0x240
do_exit+0x598/0xfd0
do_group_exit+0x8b/0x180
get_signal+0x293/0x13d0
do_signal+0x37/0x5d0
prepare_exit_to_usermode+0x1b7/0x2c0
ret_from_intr+0x32/0x42
read to 0xffff98b00ebd04dc of 4 bytes by task 82499 on cpu 46:
try_to_unuse+0x86b/0xc80
try_to_unuse at mm/swapfile.c:2185
__x64_sys_swapoff+0x372/0xd40
do_syscall_64+0x91/0xb05
entry_SYSCALL_64_after_hwframe+0x49/0xbe
The plain reads in try_to_unuse() are outside si->lock critical section
which result in data races that could be dangerous to be used in a loop.
Fix them by adding READ_ONCE().
Signed-off-by: Qian Cai <[email protected]>
---
mm/swapfile.c | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/mm/swapfile.c b/mm/swapfile.c
index a65622eec66f..36fd1536a83d 100644
--- a/mm/swapfile.c
+++ b/mm/swapfile.c
@@ -2137,7 +2137,7 @@ int try_to_unuse(unsigned int type, bool frontswap,
swp_entry_t entry;
unsigned int i;
- if (!si->inuse_pages)
+ if (!READ_ONCE(si->inuse_pages))
return 0;
if (!frontswap)
@@ -2153,7 +2153,7 @@ int try_to_unuse(unsigned int type, bool frontswap,
spin_lock(&mmlist_lock);
p = &init_mm.mmlist;
- while (si->inuse_pages &&
+ while (READ_ONCE(si->inuse_pages) &&
!signal_pending(current) &&
(p = p->next) != &init_mm.mmlist) {
@@ -2182,7 +2182,7 @@ int try_to_unuse(unsigned int type, bool frontswap,
mmput(prev_mm);
i = 0;
- while (si->inuse_pages &&
+ while (READ_ONCE(si->inuse_pages) &&
!signal_pending(current) &&
(i = find_next_to_unuse(si, i, frontswap)) != 0) {
@@ -2224,7 +2224,7 @@ int try_to_unuse(unsigned int type, bool frontswap,
* been preempted after get_swap_page(), temporarily hiding that swap.
* It's easy and robust (though cpu-intensive) just to keep retrying.
*/
- if (si->inuse_pages) {
+ if (READ_ONCE(si->inuse_pages)) {
if (!signal_pending(current))
goto retry;
retval = -EINTR;
--
1.8.3.1
On Mon, Feb 24, 2020 at 04:15:03PM -0500, Qian Cai wrote:
>si->inuse_pages could be accessed concurrently as noticed by KCSAN,
>
> write to 0xffff98b00ebd04dc of 4 bytes by task 82262 on cpu 92:
> swap_range_free+0xbe/0x230
> swap_range_free at mm/swapfile.c:719
> swapcache_free_entries+0x1be/0x250
> free_swap_slot+0x1c8/0x220
> __swap_entry_free.constprop.19+0xa3/0xb0
> free_swap_and_cache+0x53/0xa0
> unmap_page_range+0x7e0/0x1ce0
> unmap_single_vma+0xcd/0x170
> unmap_vmas+0x18b/0x220
> exit_mmap+0xee/0x220
> mmput+0xe7/0x240
> do_exit+0x598/0xfd0
> do_group_exit+0x8b/0x180
> get_signal+0x293/0x13d0
> do_signal+0x37/0x5d0
> prepare_exit_to_usermode+0x1b7/0x2c0
> ret_from_intr+0x32/0x42
>
> read to 0xffff98b00ebd04dc of 4 bytes by task 82499 on cpu 46:
> try_to_unuse+0x86b/0xc80
> try_to_unuse at mm/swapfile.c:2185
> __x64_sys_swapoff+0x372/0xd40
> do_syscall_64+0x91/0xb05
> entry_SYSCALL_64_after_hwframe+0x49/0xbe
>
>The plain reads in try_to_unuse() are outside si->lock critical section
>which result in data races that could be dangerous to be used in a loop.
>Fix them by adding READ_ONCE().
>
>Signed-off-by: Qian Cai <[email protected]>
>---
> mm/swapfile.c | 8 ++++----
> 1 file changed, 4 insertions(+), 4 deletions(-)
>
>diff --git a/mm/swapfile.c b/mm/swapfile.c
>index a65622eec66f..36fd1536a83d 100644
>--- a/mm/swapfile.c
>+++ b/mm/swapfile.c
>@@ -2137,7 +2137,7 @@ int try_to_unuse(unsigned int type, bool frontswap,
> swp_entry_t entry;
> unsigned int i;
>
>- if (!si->inuse_pages)
>+ if (!READ_ONCE(si->inuse_pages))
> return 0;
>
> if (!frontswap)
>@@ -2153,7 +2153,7 @@ int try_to_unuse(unsigned int type, bool frontswap,
>
> spin_lock(&mmlist_lock);
> p = &init_mm.mmlist;
>- while (si->inuse_pages &&
>+ while (READ_ONCE(si->inuse_pages) &&
The change is not wrong. But since it is not protected by the lock, some
status in swap_info_struct could still be modified after we test this
inuse_pages is not zero. Would this be some problem?
--
Wei Yang
Help you, Help me