summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorKoundinya Veluri <kouvel@users.noreply.github.com>2017-10-05 12:06:07 -0700
committerGitHub <noreply@github.com>2017-10-05 12:06:07 -0700
commit7e43a10cdb136cc1213b1f5b2db406b9830108e4 (patch)
treec120a96eb32248246054a78b9bfad3a082a77da1
parentb085bfacb5eed105e3e55681f75c5413d91c1f86 (diff)
downloadcoreclr-7e43a10cdb136cc1213b1f5b2db406b9830108e4.tar.gz
coreclr-7e43a10cdb136cc1213b1f5b2db406b9830108e4.tar.bz2
coreclr-7e43a10cdb136cc1213b1f5b2db406b9830108e4.zip
Fix new deadlock possibility in ReaderWriterLockSlim from a recent change (#14337)
Fix new deadlock possibility in ReaderWriterLockSlim from a recent change Fixed the following scenario that would lead to deadlock (rare, but frequently reproducible if code is changed): - Thread T0 signals the write waiter event or the upgradeable read waiter event to wake a waiter - There are no threads waiting on the event, but T1 is in WaitOnEvent() after exiting the spin lock and before actually waiting on the event (that is, it's recorded that there is one waiter for the event). It remains in this region for a while, in the repro case it typically gets context-switched out. - T2 acquires the RW lock in some fashion that blocks T0 or T3 from acquiring the RW lock - T0 or T3 fails to acquire the RW lock enough times for it to enter WaitOnEvent for the same event as T1 - T0 or T3 resets the event - T2 releases the RW lock and does not wake a waiter because the reset at the previous step lost a signal but _waiterStates was not updated to reflect that - T1 and other threads begin waiting on the event, but there's no longer any thread that would wake them
-rw-r--r--src/mscorlib/shared/System/Threading/ReaderWriterLockSlim.cs73
1 files changed, 48 insertions, 25 deletions
diff --git a/src/mscorlib/shared/System/Threading/ReaderWriterLockSlim.cs b/src/mscorlib/shared/System/Threading/ReaderWriterLockSlim.cs
index c34d6ea444..3c4aad603a 100644
--- a/src/mscorlib/shared/System/Threading/ReaderWriterLockSlim.cs
+++ b/src/mscorlib/shared/System/Threading/ReaderWriterLockSlim.cs
@@ -55,7 +55,7 @@ namespace System.Threading
{
private static readonly int ProcessorCount = Environment.ProcessorCount;
- //Specifying if locked can be reacquired recursively.
+ //Specifying if the lock can be reacquired recursively.
private readonly bool _fIsReentrant;
// Lock specification for _spinLock: This lock protects exactly the local fields associated with this
@@ -967,7 +967,51 @@ namespace System.Threading
Debug.Assert(_spinLock.IsHeld);
#endif
+ WaiterStates waiterSignaledState = WaiterStates.None;
+ EnterSpinLockReason enterMyLockReason;
+ switch (enterLockType)
+ {
+ case EnterLockType.UpgradeableRead:
+ waiterSignaledState = WaiterStates.UpgradeableReadWaiterSignaled;
+ goto case EnterLockType.Read;
+
+ case EnterLockType.Read:
+ enterMyLockReason = EnterSpinLockReason.EnterAnyRead;
+ break;
+
+ case EnterLockType.Write:
+ waiterSignaledState = WaiterStates.WriteWaiterSignaled;
+ enterMyLockReason = EnterSpinLockReason.EnterWrite;
+ break;
+
+ default:
+ Debug.Assert(enterLockType == EnterLockType.UpgradeToWrite);
+ enterMyLockReason = EnterSpinLockReason.UpgradeToWrite;
+ break;
+ }
+
+ // It was not possible to acquire the RW lock because some other thread was holding some type of lock. The other
+ // thread, when it releases its lock, will wake appropriate waiters. Along with resetting the wait event, clear the
+ // waiter signaled bit for this type of waiter if applicable, to indicate that a waiter of this type is no longer
+ // signaled.
+ //
+ // If the waiter signaled bit is not updated upon event reset, the following scenario would lead to deadlock:
+ // - Thread T0 signals the write waiter event or the upgradeable read waiter event to wake a waiter
+ // - There are no threads waiting on the event, but T1 is in WaitOnEvent() after exiting the spin lock and before
+ // actually waiting on the event (that is, it's recorded that there is one waiter for the event). It remains in
+ // this region for a while, in the repro case it typically gets context-switched out.
+ // - T2 acquires the RW lock in some fashion that blocks T0 or T3 from acquiring the RW lock
+ // - T0 or T3 fails to acquire the RW lock enough times for it to enter WaitOnEvent for the same event as T1
+ // - T0 or T3 resets the event
+ // - T2 releases the RW lock and does not wake a waiter because the reset at the previous step lost a signal but
+ // _waiterStates was not updated to reflect that
+ // - T1 and other threads begin waiting on the event, but there's no longer any thread that would wake them
+ if (waiterSignaledState != WaiterStates.None && (_waiterStates & waiterSignaledState) != WaiterStates.None)
+ {
+ _waiterStates &= ~waiterSignaledState;
+ }
waitEvent.Reset();
+
numWaiters++;
HasNoWaiters = false;
@@ -986,40 +1030,19 @@ namespace System.Threading
}
finally
{
- WaiterStates waiterSignaledState = WaiterStates.None;
- EnterSpinLockReason enterMyLockReason;
- switch (enterLockType)
- {
- case EnterLockType.UpgradeableRead:
- waiterSignaledState = WaiterStates.UpgradeableReadWaiterSignaled;
- goto case EnterLockType.Read;
-
- case EnterLockType.Read:
- enterMyLockReason = EnterSpinLockReason.EnterAnyRead;
- break;
-
- case EnterLockType.Write:
- waiterSignaledState = WaiterStates.WriteWaiterSignaled;
- enterMyLockReason = EnterSpinLockReason.EnterWrite;
- break;
-
- default:
- Debug.Assert(enterLockType == EnterLockType.UpgradeToWrite);
- enterMyLockReason = EnterSpinLockReason.UpgradeToWrite;
- break;
- }
_spinLock.Enter(enterMyLockReason);
--numWaiters;
- if (waitSuccessful && waiterSignaledState != WaiterStates.None)
+ if (waitSuccessful &&
+ waiterSignaledState != WaiterStates.None &&
+ (_waiterStates & waiterSignaledState) != WaiterStates.None)
{
// Indicate that a signaled waiter of this type has woken. Since non-read waiters are signaled to wake one
// at a time, we avoid waking up more than one waiter of that type upon successive enter/exit loops until
// the signaled thread actually wakes up. For example, if there are multiple write waiters and one thread is
// repeatedly entering and exiting a write lock, every exit would otherwise signal a different write waiter
// to wake up unnecessarily when only one woken waiter may actually succeed in entering the write lock.
- Debug.Assert((_waiterStates & waiterSignaledState) != WaiterStates.None);
_waiterStates &= ~waiterSignaledState;
}