// Copyright (c) 2011 The Chromium Authors. All rights reserved. // Use of this source code is governed by a BSD-style license that can be // found in the LICENSE file. #import "chrome/browser/chrome_browser_application_mac.h" #import "base/logging.h" #import "base/metrics/histogram.h" #import "base/memory/scoped_nsobject.h" #include "base/sys_info.h" #import "base/sys_string_conversions.h" #import "chrome/app/breakpad_mac.h" #import "chrome/browser/app_controller_mac.h" #import "chrome/browser/ui/cocoa/objc_method_swizzle.h" #import "chrome/browser/ui/cocoa/objc_zombie.h" // The implementation of NSExceptions break various assumptions in the // Chrome code. This category defines a replacement for // -initWithName:reason:userInfo: for purposes of forcing a break in // the debugger when an exception is raised. -raise sounds more // obvious to intercept, but it doesn't catch the original throw // because the objc runtime doesn't use it. @interface NSException (NSExceptionSwizzle) - (id)chromeInitWithName:(NSString*)aName reason:(NSString*)aReason userInfo:(NSDictionary *)someUserInfo; @end static IMP gOriginalInitIMP = NULL; @implementation NSException (NSExceptionSwizzle) - (id)chromeInitWithName:(NSString*)aName reason:(NSString*)aReason userInfo:(NSDictionary *)someUserInfo { // Method only called when swizzled. DCHECK(_cmd == @selector(initWithName:reason:userInfo:)); // Parts of Cocoa rely on creating and throwing exceptions. These are not // worth bugging-out over. It is very important that there be zero chance that // any Chromium code is on the stack; these must be created by Apple code and // then immediately consumed by Apple code. static NSString* const kAcceptableNSExceptionNames[] = { // If an object does not support an accessibility attribute, this will // get thrown. NSAccessibilityException, nil }; BOOL found = NO; for (int i = 0; kAcceptableNSExceptionNames[i]; ++i) { if (aName == kAcceptableNSExceptionNames[i]) { found = YES; } } if (!found) { // Update breakpad with the exception info. static NSString* const kNSExceptionKey = @"nsexception"; NSString* value = [NSString stringWithFormat:@"%@ reason %@", aName, aReason]; SetCrashKeyValue(kNSExceptionKey, value); // Force crash for selected exceptions to generate crash dumps. BOOL fatal = NO; if (aName == NSInternalInconsistencyException) { NSString* const kNSMenuItemArrayBoundsCheck = @"Invalid parameter not satisfying: (index >= 0) && " @"(index < [_itemArray count])"; if ([aReason isEqualToString:kNSMenuItemArrayBoundsCheck]) { fatal = YES; } NSString* const kNoWindowCheck = @"View is not in any window"; if ([aReason isEqualToString:kNoWindowCheck]) { fatal = YES; } } // Mostly "unrecognized selector sent to (instance|class)". A // very small number of things like nil being passed to an // inappropriate receiver. if (aName == NSInvalidArgumentException) { fatal = YES; } // Dear reader: Something you just did provoked an NSException. // NSException is implemented in terms of setjmp()/longjmp(), // which does poor things when combined with C++ scoping // (destructors are skipped). Chrome should be NSException-free, // please check your backtrace and see if you can't file a bug // with a repro case. if (fatal) { LOG(FATAL) << "Someone is trying to raise an exception! " << base::SysNSStringToUTF8(value); } else { // Make sure that developers see when their code throws // exceptions. DLOG(ERROR) << "Someone is trying to raise an exception! " << base::SysNSStringToUTF8(value); NOTREACHED(); } } // Forward to the original version. return gOriginalInitIMP(self, _cmd, aName, aReason, someUserInfo); } @end namespace chrome_browser_application_mac { // Maximum number of known named exceptions we'll support. There is // no central registration, but I only find about 75 possibilities in // the system frameworks, and many of them are probably not // interesting to track in aggregate (those relating to distributed // objects, for instance). const size_t kKnownNSExceptionCount = 25; const size_t kUnknownNSException = kKnownNSExceptionCount; size_t BinForException(NSException* exception) { // A list of common known exceptions. The list position will // determine where they live in the histogram, so never move them // around, only add to the end. static NSString* const kKnownNSExceptionNames[] = { // Grab-bag exception, not very common. CFArray (or other // container) mutated while being enumerated is one case seen in // production. NSGenericException, // Out-of-range on NSString or NSArray. Quite common. NSRangeException, // Invalid arg to method, unrecognized selector. Quite common. NSInvalidArgumentException, // malloc() returned null in object creation, I think. Turns out // to be very uncommon in production, because of the OOM killer. NSMallocException, // This contains things like windowserver errors, trying to draw // views which aren't in windows, unable to read nib files. By // far the most common exception seen on the crash server. NSInternalInconsistencyException, nil }; // Make sure our array hasn't outgrown our abilities to track it. DCHECK_LE(arraysize(kKnownNSExceptionNames), kKnownNSExceptionCount); NSString* name = [exception name]; for (int i = 0; kKnownNSExceptionNames[i]; ++i) { if (name == kKnownNSExceptionNames[i]) { return i; } } return kUnknownNSException; } void RecordExceptionWithUma(NSException* exception) { UMA_HISTOGRAM_ENUMERATION("OSX.NSException", BinForException(exception), kUnknownNSException); } void RegisterBrowserCrApp() { [BrowserCrApplication sharedApplication]; }; void Terminate() { [NSApp terminate:nil]; } void CancelTerminate() { [NSApp cancelTerminate:nil]; } } // namespace chrome_browser_application_mac namespace { // Do-nothing wrapper so that we can arrange to only swizzle // -[NSException raise] when DCHECK() is turned on (as opposed to // replicating the preprocess logic which turns DCHECK() on). BOOL SwizzleNSExceptionInit() { gOriginalInitIMP = ObjcEvilDoers::SwizzleImplementedInstanceMethods( [NSException class], @selector(initWithName:reason:userInfo:), @selector(chromeInitWithName:reason:userInfo:)); return YES; } } // namespace @implementation BrowserCrApplication + (void)initialize { // Whitelist releases that are compatible with objc zombies. int32 major_version = 0, minor_version = 0, bugfix_version = 0; base::SysInfo::OperatingSystemVersionNumbers( &major_version, &minor_version, &bugfix_version); if (major_version == 10 && (minor_version == 5 || minor_version == 6)) { // Turn all deallocated Objective-C objects into zombies, keeping // the most recent 10,000 of them on the treadmill. ObjcEvilDoers::ZombieEnable(YES, 10000); } } - init { CHECK(SwizzleNSExceptionInit()); return [super init]; } //////////////////////////////////////////////////////////////////////////////// // HISTORICAL COMMENT (by viettrungluu, from // http://codereview.chromium.org/1520006 with mild editing): // // A quick summary of the state of things (before the changes to shutdown): // // Currently, we are totally hosed (put in a bad state in which Cmd-W does the // wrong thing, and which will probably eventually lead to a crash) if we begin // quitting but termination is aborted for some reason. // // I currently know of two ways in which termination can be aborted: // (1) Common case: a window has an onbeforeunload handler which pops up a // "leave web page" dialog, and the user answers "no, don't leave". // (2) Uncommon case: popups are enabled (in Content Settings, i.e., the popup // blocker is disabled), and some nasty web page pops up a new window on // closure. // // I don't know of other ways in which termination can be aborted, but they may // exist (or may be added in the future, for that matter). // // My CL [see above] does the following: // a. Should prevent being put in a bad state (which breaks Cmd-W and leads to // crash) under all circumstances. // b. Should completely handle (1) properly. // c. Doesn't (yet) handle (2) properly and puts it in a weird state (but not // that bad). // d. Any other ways of aborting termination would put it in that weird state. // // c. can be fixed by having the global flag reset on browser creation or // similar (and doing so might also fix some possible d.'s as well). I haven't // done this yet since I haven't thought about it carefully and since it's a // corner case. // // The weird state: a state in which closing the last window quits the browser. // This might be a bit annoying, but it's not dangerous in any way. //////////////////////////////////////////////////////////////////////////////// // |-terminate:| is the entry point for orderly "quit" operations in Cocoa. This // includes the application menu's quit menu item and keyboard equivalent, the // application's dock icon menu's quit menu item, "quit" (not "force quit") in // the Activity Monitor, and quits triggered by user logout and system restart // and shutdown. // // The default |-terminate:| implementation ends the process by calling exit(), // and thus never leaves the main run loop. This is unsuitable for Chrome since // Chrome depends on leaving the main run loop to perform an orderly shutdown. // We support the normal |-terminate:| interface by overriding the default // implementation. Our implementation, which is very specific to the needs of // Chrome, works by asking the application delegate to terminate using its // |-tryToTerminateApplication:| method. // // |-tryToTerminateApplication:| differs from the standard // |-applicationShouldTerminate:| in that no special event loop is run in the // case that immediate termination is not possible (e.g., if dialog boxes // allowing the user to cancel have to be shown). Instead, this method sets a // flag and tries to close all browsers. This flag causes the closure of the // final browser window to begin actual tear-down of the application. // Termination is cancelled by resetting this flag. The standard // |-applicationShouldTerminate:| is not supported, and code paths leading to it // must be redirected. - (void)terminate:(id)sender { AppController* appController = static_cast<AppController*>([NSApp delegate]); if ([appController tryToTerminateApplication:self]) { [[NSNotificationCenter defaultCenter] postNotificationName:NSApplicationWillTerminateNotification object:self]; } // Return, don't exit. The application is responsible for exiting on its own. } - (void)cancelTerminate:(id)sender { AppController* appController = static_cast<AppController*>([NSApp delegate]); [appController stopTryingToTerminateApplication:self]; } - (BOOL)sendAction:(SEL)anAction to:(id)aTarget from:(id)sender { // The Dock menu contains an automagic section where you can select // amongst open windows. If a window is closed via JavaScript while // the menu is up, the menu item for that window continues to exist. // When a window is selected this method is called with the // now-freed window as |aTarget|. Short-circuit the call if // |aTarget| is not a valid window. if (anAction == @selector(_selectWindow:)) { // Not using -[NSArray containsObject:] because |aTarget| may be a // freed object. BOOL found = NO; for (NSWindow* window in [self windows]) { if (window == aTarget) { found = YES; break; } } if (!found) { return NO; } } // When a Cocoa control is wired to a freed object, we get crashers // in the call to |super| with no useful information in the // backtrace. Attempt to add some useful information. static NSString* const kActionKey = @"sendaction"; // If the action is something generic like -commandDispatch:, then // the tag is essential. NSInteger tag = 0; if ([sender isKindOfClass:[NSControl class]]) { tag = [sender tag]; if (tag == 0 || tag == -1) { tag = [sender selectedTag]; } } else if ([sender isKindOfClass:[NSMenuItem class]]) { tag = [sender tag]; } NSString* actionString = NSStringFromSelector(anAction); NSString* value = [NSString stringWithFormat:@"%@ tag %d sending %@ to %p", [sender className], tag, actionString, aTarget]; ScopedCrashKey key(kActionKey, value); return [super sendAction:anAction to:aTarget from:sender]; } // NSExceptions which are caught by the event loop are logged here. // NSException uses setjmp/longjmp, which can be very bad for C++, so // we attempt to track and report them. - (void)reportException:(NSException *)anException { // If we throw an exception in this code, we can create an infinite // loop. If we throw out of the if() without resetting // |reportException|, we'll stop reporting exceptions for this run. static BOOL reportingException = NO; DCHECK(!reportingException); if (!reportingException) { reportingException = YES; chrome_browser_application_mac::RecordExceptionWithUma(anException); // http://crbug.com/45928 is a bug about needing to double-close // windows sometimes. One theory is that |-isHandlingSendEvent| // gets latched to always return |YES|. Since scopers are used to // manipulate that value, that should not be possible. One way to // sidestep scopers is setjmp/longjmp (see above). The following // is to "fix" this while the more fundamental concern is // addressed elsewhere. [self clearIsHandlingSendEvent]; // Store some human-readable information in breakpad keys in case // there is a crash. Since breakpad does not provide infinite // storage, we track two exceptions. The first exception thrown // is tracked because it may be the one which caused the system to // go off the rails. The last exception thrown is tracked because // it may be the one most directly associated with the crash. static NSString* const kFirstExceptionKey = @"firstexception"; static BOOL trackedFirstException = NO; static NSString* const kLastExceptionKey = @"lastexception"; // TODO(shess): It would be useful to post some stacktrace info // from the exception. // 10.6 has -[NSException callStackSymbols] // 10.5 has -[NSException callStackReturnAddresses] // 10.5 has backtrace_symbols(). // I've tried to combine the latter two, but got nothing useful. // The addresses are right, though, maybe we could train the crash // server to decode them for us. NSString* value = [NSString stringWithFormat:@"%@ reason %@", [anException name], [anException reason]]; if (!trackedFirstException) { SetCrashKeyValue(kFirstExceptionKey, value); trackedFirstException = YES; } else { SetCrashKeyValue(kLastExceptionKey, value); } reportingException = NO; } [super reportException:anException]; } @end