Changeset 163661 in webkit

Timestamp:

Feb 7, 2014 4:24:13 PM (10 years ago)

Author:

mark.lam@apple.com

Message:

Fix bug in stack limit adjustments in JSLock.
<​https://webkit.org/b/128406>

Reviewed by Geoffrey Garen.

1. JSLock::unlock() was only clearing the VM::stackPointerAtEntry when m_vm->stackPointerAtVMEntry == entryStackPointer. FYI, entryStackPointer is a field in JSLock.

When DropAllLocks::~DropAllLocks() will call JSLock::grabAllLocks()
to relock the JSLock, JSLock::grabAllLocks() will set a new
entryStackPointer value. Thereafter, DropAllLocks::~DropAllLocks() will
restore the saved VM::stackPointerAtEntry, which will now defer from
the JSLock's entryStackPointer value.

It turns out that when m_vm->stackPointerAtVMEntry was initialized,
it was set to whatever value entryStackPointer is set to. At no time
do we ever expect the 2 values to differ. The only time it differs is
when this bug manifests.

The fix is to remove the entryStackPointer field in JSLock and its uses
altogether.

2. DropAllLocks was unconditionally clearing VM::stackPointerAtEntry in its constructor instead of letting JSLock::unlock() do the clearing.

However, DropAllLocks will not actually drop locks if it isn't required
to (e.g. when alwaysDropLocks is DontAlwaysDropLocks), and when we've
already drop locks once (i.e. JSLock::m_lockDropDepth is not 0).

We should not have cleared VM::stackPointerAtEntry here if we don't
actually drop the locks.

• runtime/JSLock.cpp:

(JSC::JSLock::unlock):
(JSC::JSLock::DropAllLocks::DropAllLocks):

Location:

trunk/Source/JavaScriptCore

Files:

• ChangeLog (modified) (1 diff)
• runtime/JSLock.cpp (modified) (3 diffs)

trunk/Source/JavaScriptCore/ChangeLog

@@ +1 @@
+2014-02-07  Mark Lam  <mark.lam@apple.com>
+
+        Fix bug in stack limit adjustments in JSLock.
+        <https://webkit.org/b/128406>
+
+        Reviewed by Geoffrey Garen.
+
+        1. JSLock::unlock() was only clearing the VM::stackPointerAtEntry when
+           m_vm->stackPointerAtVMEntry == entryStackPointer. FYI,
+           entryStackPointer is a field in JSLock.
+
+           When DropAllLocks::~DropAllLocks() will call JSLock::grabAllLocks()
+           to relock the JSLock, JSLock::grabAllLocks() will set a new
+           entryStackPointer value. Thereafter, DropAllLocks::~DropAllLocks() will
+           restore the saved VM::stackPointerAtEntry, which will now defer from
+           the JSLock's entryStackPointer value.
+
+           It turns out that when m_vm->stackPointerAtVMEntry was initialized,
+           it was set to whatever value entryStackPointer is set to. At no time
+           do we ever expect the 2 values to differ. The only time it differs is
+           when this bug manifests.
+
+           The fix is to remove the entryStackPointer field in JSLock and its uses
+           altogether.
+
+        2. DropAllLocks was unconditionally clearing VM::stackPointerAtEntry in
+           its constructor instead of letting JSLock::unlock() do the clearing.
+
+           However, DropAllLocks will not actually drop locks if it isn't required
+           to (e.g. when alwaysDropLocks is DontAlwaysDropLocks), and when we've
+           already drop locks once (i.e. JSLock::m_lockDropDepth is not 0).
+
+           We should not have cleared VM::stackPointerAtEntry here if we don't
+           actually drop the locks.
+
+        * runtime/JSLock.cpp:
+        (JSC::JSLock::unlock):
+        (JSC::JSLock::DropAllLocks::DropAllLocks):
+
 2014-02-07  Joseph Pecoraro  <pecoraro@apple.com>
 

trunk/Source/JavaScriptCore/runtime/JSLock.cpp

@@ -143 +143 @@
 
     if (!m_lockCount) {
-        if (m_vm && m_vm->stackPointerAtVMEntry == entryStackPointer) {
+        if (m_vm) {
             m_vm->stackPointerAtVMEntry = nullptr;
             m_vm->updateStackLimitWithReservedZoneSize(wtfThreadData().savedReservedZoneSize());
@@ -313 +313 @@
     threadData.setSavedReservedZoneSize(m_vm->reservedZoneSize());
 
-    m_vm->stackPointerAtVMEntry = nullptr;
-
     if (alwaysDropLocks)
         m_lockCount = m_vm->apiLock().dropAllLocksUnconditionally(spinLock);
@@ -338 +336 @@
     threadData.setSavedReservedZoneSize(m_vm->reservedZoneSize());
 
-    m_vm->stackPointerAtVMEntry = nullptr;
-
     if (alwaysDropLocks)
         m_lockCount = m_vm->apiLock().dropAllLocksUnconditionally(spinLock);