[PATCH v2] x86/mce: Fix endless loop when run task works after #MC

[Ding Hui <dinghui@xxxxxxxxxxxxxx>]

Recently we encounter multi #MC on the same task when it's
task_work_run() has not been called, current->mce_kill_me was
added to task_works list more than once, that make a circular
linked task_works, so task_work_run() will do a endless loop.

More seriously, the SIGBUS signal can not be delivered to the
userspace task which tigger the #MC and I met #MC flood.

I borrowed mce_kill_me.func to check whether current->mce_kill_me
has been added to task_works, prevent duplicate addition and override
current->mce_xxx. When work function be called, the task_works
must has been taken, so it is safe to be cleared in callback.

Fixes: commit 5567d11c21a1 ("x86/mce: Send #MC singal from task work")
Cc: <stable@xxxxxxxxxxxxxxx> #v5.8+
Signed-off-by: Ding Hui <dinghui@xxxxxxxxxxxxxx>

---
v1->v2:
do not call kill_me_now in kill_me_maybe, to avoid mce_kill_me.func
race condition
do not override current->mce_xxx before callback

 arch/x86/kernel/cpu/mce/core.c | 38 ++++++++++++++++++++++------------
 1 file changed, 25 insertions(+), 13 deletions(-)

diff --git a/arch/x86/kernel/cpu/mce/core.c b/arch/x86/kernel/cpu/mce/core.c
index 22791aadc085..95d244a95486 100644
--- a/arch/x86/kernel/cpu/mce/core.c
+++ b/arch/x86/kernel/cpu/mce/core.c
@@ -1250,6 +1250,7 @@ static void __mc_scan_banks(struct mce *m, struct pt_regs *regs, struct mce *fin
 
 static void kill_me_now(struct callback_head *ch)
 {
+	WRITE_ONCE(ch->func, NULL);
 	force_sig(SIGBUS);
 }
 
@@ -1258,15 +1259,22 @@ static void kill_me_maybe(struct callback_head *cb)
 	struct task_struct *p = container_of(cb, struct task_struct, mce_kill_me);
 	int flags = MF_ACTION_REQUIRED;
 	int ret;
+	u64 mce_addr = p->mce_addr;
+	__u64 mce_kflags = p->mce_kflags;
+	bool mce_ripv = p->mce_ripv;
+	bool mce_whole_page = p->mce_whole_page;
 
-	pr_err("Uncorrected hardware memory error in user-access at %llx", p->mce_addr);
+	/* reset func after save p->mce_xxx */
+	WRITE_ONCE(cb->func, NULL);
 
-	if (!p->mce_ripv)
+	pr_err("Uncorrected hardware memory error in user-access at %llx", mce_addr);
+
+	if (!mce_ripv)
 		flags |= MF_MUST_KILL;
 
-	ret = memory_failure(p->mce_addr >> PAGE_SHIFT, flags);
-	if (!ret && !(p->mce_kflags & MCE_IN_KERNEL_COPYIN)) {
-		set_mce_nospec(p->mce_addr >> PAGE_SHIFT, p->mce_whole_page);
+	ret = memory_failure(mce_addr >> PAGE_SHIFT, flags);
+	if (!ret && !(mce_kflags & MCE_IN_KERNEL_COPYIN)) {
+		set_mce_nospec(mce_addr >> PAGE_SHIFT, mce_whole_page);
 		sync_core();
 		return;
 	}
@@ -1283,23 +1291,27 @@ static void kill_me_maybe(struct callback_head *cb)
 		force_sig_mceerr(BUS_MCEERR_AR, p->mce_vaddr, PAGE_SHIFT);
 	} else {
 		pr_err("Memory error not recovered");
-		kill_me_now(cb);
+		force_sig(SIGBUS);
 	}
 }
 
 static void queue_task_work(struct mce *m, int kill_current_task)
 {
-	current->mce_addr = m->addr;
-	current->mce_kflags = m->kflags;
-	current->mce_ripv = !!(m->mcgstatus & MCG_STATUS_RIPV);
-	current->mce_whole_page = whole_page(m);
+	struct callback_head ch;
 
 	if (kill_current_task)
-		current->mce_kill_me.func = kill_me_now;
+		ch.func = kill_me_now;
 	else
-		current->mce_kill_me.func = kill_me_maybe;
+		ch.func = kill_me_maybe;
+
+	if (!cmpxchg(&current->mce_kill_me.func, NULL, ch.func)) {
+		current->mce_addr = m->addr;
+		current->mce_kflags = m->kflags;
+		current->mce_ripv = !!(m->mcgstatus & MCG_STATUS_RIPV);
+		current->mce_whole_page = whole_page(m);
 
-	task_work_add(current, &current->mce_kill_me, TWA_RESUME);
+		task_work_add(current, &current->mce_kill_me, TWA_RESUME);
+	}
 }
 
 /*
-- 
2.17.1