coredump: zap_threads() must skip kernel threads
Subsystems (2)
PATCH 3/3 coredump: zap_threads() must skip kernel threads
The main loop in zap_threads() must skip kthreads which may use the same mm. Otherwise we "kill" this thread erroneously (for example, it can not fork or exec after that), and the coredumping task stucks in the TASK_UNINTERRUPTIBLE state forever because of the wrong ->core_waiters count.
fs/exec.c renamed from fs/exec.c~3_CD_FIX_RACE_USE_MM +3 -1
@@ -1568,11 +1568,13 @@ static inline int zap_threads(struct tasfor_each_process (g) {if (g == tsk->group_leader)continue ; +if (g->flags &PF_KTHREAD ) +continue ; p = g; do { if (p->mm) { - if (p->mm == mm) { + if (unlikely(p->mm == mm)) { lock_task_sighand(p, &flags); zap_process(p); unlock_task_sighand(p, &flags);
Quoted 14 lines · On Sun, 1 Jun 2008 19:30:45 +0400 Oleg Nesterov <[email protected]> wrote:
The main loop in
zap_threads()must skip kthreads which may use the same mm. Otherwise we "kill" this thread erroneously (for example, it can not fork or exec after that), and the coredumping task stucks in theTASK_UNINTERRUPTIBLEstate forever because of the wrong ->core_waiters count.Signed-off-by Oleg Nesterov <[email protected]>fs/exec.c renamed from fs/exec.c~3_CD_FIX_RACE_USE_MM +1
-0@@ -1568,11 +1568,13 @@ static inline int zap_threads(struct tasfor_each_process (g) {if (g == tsk->group_leader)continue ; +if (g->flags &PF_KTHREAD )
fs/exec.c static inline int zap_threads(struct tas · line 1580
Quoted 10 lines
fs/exec.c renamed from fs/exec.c~3_CD_FIX_RACE_USE_MM +2
-1+continue ; p = g; do { if (p->mm) { - if (p->mm == mm) { + if (unlikely(p->mm == mm)) { lock_task_sighand(p, &flags); zap_process(p); unlock_task_sighand(p, &flags);
This is a bugfix, yes?
How does it get triggered?
Do you think the bug is sufficiently serious to fix it in 2.6.26? In 2.6.25.x? If so, it would be better if this patch were not dependent upon the preceding ones, which do not appear to be 2.6.26 or -stable material.
Quoted 3 lines
This is a bugfix, yes?
How does it get triggered?
Yes, I think it fixes a bug. The trigger would be an aio request doing some work (inside aio_kick_handler) simultaneous with some thread in the requester's mm doing a core dump (inside zap_threads).
Quoted 4 lines
Do you think the bug is sufficiently serious to fix it in 2.6.26? In 2.6.25.x? If so, it would be better if this patch were not dependent upon the preceding ones, which do not appear to be 2.6.26 or -stable material.
It has probably never been seen for real, but might be possible to produce with an exploit that works hard to hit the race. I'm not sure off hand what all the bad effects would be, mainly those of SIGKILL'ing the workqueue thread (keventd I guess). The core-dumping threads will be stuck in uninterruptible waits and never be killable.
Oleg's cleanups make the fix much nicer because there is an easy persistent flag to check without races. Probably the most isolated fix for this is something like the bit below (wholly untested). This is hairy enough that I think Oleg's 1/3 + 2/3 would be preferable even for -stable.
Thanks,
Roland
fs/exec.c +17 -2
@@ -1545,8 +1545,23 @@ static inline int zap_threads(struct tas p = g; do { - if (p->mm) { - if (p->mm == mm) { + struct mm_struct *pmm = p->mm; + if (pmm) { + /* + * We must ignore a kernel thread (aio) + * using PF_BORROWED_MM. But we need + * task_lock() to avoid races with use_mm() + * or unuse_mm(). + */ + if (pmm == mm) { + task_lock(p); + if (p->flags & PF_BORROWED_MM) + pmm = NULL; + else + pmm = p->mm; + task_unlock(p); + } + if (pmm == mm) { /* * p->sighand can't disappear, but * may be changed by de_thread()
Quoted 19 lines · On Tue, 3 Jun 2008 14:49:58 -0700 (PDT) Roland McGrath <[email protected]>…
Quoted 3 lines
This is a bugfix, yes?
How does it get triggered?
Yes, I think it fixes a bug. The trigger would be an aio request doing some work (inside
aio_kick_handler) simultaneous with some thread in the requester's mm doing a core dump (insidezap_threads).Quoted 4 lines
Do you think the bug is sufficiently serious to fix it in 2.6.26? In 2.6.25.x? If so, it would be better if this patch were not dependent upon the preceding ones, which do not appear to be 2.6.26 or -stable material.
It has probably never been seen for real, but might be possible to produce with an exploit that works hard to hit the race. I'm not sure off hand what all the bad effects would be, mainly those of SIGKILL'ing the workqueue thread (keventd I guess). The core-dumping threads will be stuck in uninterruptible waits and never be killable.
Quoted 4 lines
Oleg's cleanups make the fix much nicer because there is an easy persistent flag to check without races. Probably the most isolated fix for this is something like the bit below (wholly untested). This is hairy enough that I think Oleg's 1/3 + 2/3 would be preferable even for -stable.
OK, thanks.
I'll tentatively queue these three for 2.6.26 and will leave 2.6.25.x alone. The bug seems sufficiently obscure?
(This required a bit of massaging of coredump-zap_threads-must-skip-kernel-threads.patch in fs/exec.c due, I assume, to dependencies on other things which we have queued for 2.6.27).