mm/mmu_notifier: silence PROVE_RCU_LIST warnings
authorQian Cai <cai@lca.pw>
Sun, 22 Mar 2020 01:22:34 +0000 (18:22 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sun, 22 Mar 2020 01:56:06 +0000 (18:56 -0700)
It is safe to traverse mm->notifier_subscriptions->list either under
SRCU read lock or mm->notifier_subscriptions->lock using
hlist_for_each_entry_rcu().  Silence the PROVE_RCU_LIST false positives,
for example,

  WARNING: suspicious RCU usage
  -----------------------------
  mm/mmu_notifier.c:484 RCU-list traversed in non-reader section!!

  other info that might help us debug this:

  rcu_scheduler_active = 2, debug_locks = 1
  3 locks held by libvirtd/802:
   #0: ffff9321e3f58148 (&mm->mmap_sem#2){++++}, at: do_mprotect_pkey+0xe1/0x3e0
   #1: ffffffff91ae6160 (mmu_notifier_invalidate_range_start){+.+.}, at: change_p4d_range+0x5fa/0x800
   #2: ffffffff91ae6e08 (srcu){....}, at: __mmu_notifier_invalidate_range_start+0x178/0x460

  stack backtrace:
  CPU: 7 PID: 802 Comm: libvirtd Tainted: G          I       5.6.0-rc6-next-20200317+ #2
  Hardware name: HP ProLiant BL460c Gen8, BIOS I31 11/02/2014
  Call Trace:
    dump_stack+0xa4/0xfe
    lockdep_rcu_suspicious+0xeb/0xf5
    __mmu_notifier_invalidate_range_start+0x3ff/0x460
    change_p4d_range+0x746/0x800
    change_protection+0x1df/0x300
    mprotect_fixup+0x245/0x3e0
    do_mprotect_pkey+0x23b/0x3e0
    __x64_sys_mprotect+0x51/0x70
    do_syscall_64+0x91/0xae8
    entry_SYSCALL_64_after_hwframe+0x49/0xb3

Signed-off-by: Qian Cai <cai@lca.pw>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Link: http://lkml.kernel.org/r/20200317175640.2047-1-cai@lca.pw
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/mmu_notifier.c

index ef3973a..06852b8 100644 (file)
@@ -307,7 +307,8 @@ static void mn_hlist_release(struct mmu_notifier_subscriptions *subscriptions,
         * ->release returns.
         */
        id = srcu_read_lock(&srcu);
-       hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist)
+       hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist,
+                                srcu_read_lock_held(&srcu))
                /*
                 * If ->release runs before mmu_notifier_unregister it must be
                 * handled, as it's the only way for the driver to flush all
@@ -370,7 +371,8 @@ int __mmu_notifier_clear_flush_young(struct mm_struct *mm,
 
        id = srcu_read_lock(&srcu);
        hlist_for_each_entry_rcu(subscription,
-                                &mm->notifier_subscriptions->list, hlist) {
+                                &mm->notifier_subscriptions->list, hlist,
+                                srcu_read_lock_held(&srcu)) {
                if (subscription->ops->clear_flush_young)
                        young |= subscription->ops->clear_flush_young(
                                subscription, mm, start, end);
@@ -389,7 +391,8 @@ int __mmu_notifier_clear_young(struct mm_struct *mm,
 
        id = srcu_read_lock(&srcu);
        hlist_for_each_entry_rcu(subscription,
-                                &mm->notifier_subscriptions->list, hlist) {
+                                &mm->notifier_subscriptions->list, hlist,
+                                srcu_read_lock_held(&srcu)) {
                if (subscription->ops->clear_young)
                        young |= subscription->ops->clear_young(subscription,
                                                                mm, start, end);
@@ -407,7 +410,8 @@ int __mmu_notifier_test_young(struct mm_struct *mm,
 
        id = srcu_read_lock(&srcu);
        hlist_for_each_entry_rcu(subscription,
-                                &mm->notifier_subscriptions->list, hlist) {
+                                &mm->notifier_subscriptions->list, hlist,
+                                srcu_read_lock_held(&srcu)) {
                if (subscription->ops->test_young) {
                        young = subscription->ops->test_young(subscription, mm,
                                                              address);
@@ -428,7 +432,8 @@ void __mmu_notifier_change_pte(struct mm_struct *mm, unsigned long address,
 
        id = srcu_read_lock(&srcu);
        hlist_for_each_entry_rcu(subscription,
-                                &mm->notifier_subscriptions->list, hlist) {
+                                &mm->notifier_subscriptions->list, hlist,
+                                srcu_read_lock_held(&srcu)) {
                if (subscription->ops->change_pte)
                        subscription->ops->change_pte(subscription, mm, address,
                                                      pte);
@@ -476,7 +481,8 @@ static int mn_hlist_invalidate_range_start(
        int id;
 
        id = srcu_read_lock(&srcu);
-       hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist) {
+       hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist,
+                                srcu_read_lock_held(&srcu)) {
                const struct mmu_notifier_ops *ops = subscription->ops;
 
                if (ops->invalidate_range_start) {
@@ -528,7 +534,8 @@ mn_hlist_invalidate_end(struct mmu_notifier_subscriptions *subscriptions,
        int id;
 
        id = srcu_read_lock(&srcu);
-       hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist) {
+       hlist_for_each_entry_rcu(subscription, &subscriptions->list, hlist,
+                                srcu_read_lock_held(&srcu)) {
                /*
                 * Call invalidate_range here too to avoid the need for the
                 * subsystem of having to register an invalidate_range_end
@@ -582,7 +589,8 @@ void __mmu_notifier_invalidate_range(struct mm_struct *mm,
 
        id = srcu_read_lock(&srcu);
        hlist_for_each_entry_rcu(subscription,
-                                &mm->notifier_subscriptions->list, hlist) {
+                                &mm->notifier_subscriptions->list, hlist,
+                                srcu_read_lock_held(&srcu)) {
                if (subscription->ops->invalidate_range)
                        subscription->ops->invalidate_range(subscription, mm,
                                                            start, end);
@@ -714,7 +722,8 @@ find_get_mmu_notifier(struct mm_struct *mm, const struct mmu_notifier_ops *ops)
 
        spin_lock(&mm->notifier_subscriptions->lock);
        hlist_for_each_entry_rcu(subscription,
-                                &mm->notifier_subscriptions->list, hlist) {
+                                &mm->notifier_subscriptions->list, hlist,
+                                lockdep_is_held(&mm->notifier_subscriptions->lock)) {
                if (subscription->ops != ops)
                        continue;