| Message ID | 1582578903-29294-1-git-send-email-cai@lca.pw (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | mm/swapfile: fix data races in try_to_unuse() | expand |
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 <cai@lca.pw> >--- > 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?
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;
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 <cai@lca.pw> --- mm/swapfile.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)